xref: /illumos-gate/usr/src/uts/common/fs/zfs/vdev.c (revision be6fd75a69ae679453d9cda5bff3326111e6d1ca)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
25  * Copyright (c) 2013 by Delphix. All rights reserved.
26  */
27 
28 #include <sys/zfs_context.h>
29 #include <sys/fm/fs/zfs.h>
30 #include <sys/spa.h>
31 #include <sys/spa_impl.h>
32 #include <sys/dmu.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/vdev_impl.h>
35 #include <sys/uberblock_impl.h>
36 #include <sys/metaslab.h>
37 #include <sys/metaslab_impl.h>
38 #include <sys/space_map.h>
39 #include <sys/zio.h>
40 #include <sys/zap.h>
41 #include <sys/fs/zfs.h>
42 #include <sys/arc.h>
43 #include <sys/zil.h>
44 #include <sys/dsl_scan.h>
45 
46 /*
47  * Virtual device management.
48  */
49 
50 static vdev_ops_t *vdev_ops_table[] = {
51 	&vdev_root_ops,
52 	&vdev_raidz_ops,
53 	&vdev_mirror_ops,
54 	&vdev_replacing_ops,
55 	&vdev_spare_ops,
56 	&vdev_disk_ops,
57 	&vdev_file_ops,
58 	&vdev_missing_ops,
59 	&vdev_hole_ops,
60 	NULL
61 };
62 
63 /* maximum scrub/resilver I/O queue per leaf vdev */
64 int zfs_scrub_limit = 10;
65 
66 /*
67  * Given a vdev type, return the appropriate ops vector.
68  */
69 static vdev_ops_t *
70 vdev_getops(const char *type)
71 {
72 	vdev_ops_t *ops, **opspp;
73 
74 	for (opspp = vdev_ops_table; (ops = *opspp) != NULL; opspp++)
75 		if (strcmp(ops->vdev_op_type, type) == 0)
76 			break;
77 
78 	return (ops);
79 }
80 
81 /*
82  * Default asize function: return the MAX of psize with the asize of
83  * all children.  This is what's used by anything other than RAID-Z.
84  */
85 uint64_t
86 vdev_default_asize(vdev_t *vd, uint64_t psize)
87 {
88 	uint64_t asize = P2ROUNDUP(psize, 1ULL << vd->vdev_top->vdev_ashift);
89 	uint64_t csize;
90 
91 	for (int c = 0; c < vd->vdev_children; c++) {
92 		csize = vdev_psize_to_asize(vd->vdev_child[c], psize);
93 		asize = MAX(asize, csize);
94 	}
95 
96 	return (asize);
97 }
98 
99 /*
100  * Get the minimum allocatable size. We define the allocatable size as
101  * the vdev's asize rounded to the nearest metaslab. This allows us to
102  * replace or attach devices which don't have the same physical size but
103  * can still satisfy the same number of allocations.
104  */
105 uint64_t
106 vdev_get_min_asize(vdev_t *vd)
107 {
108 	vdev_t *pvd = vd->vdev_parent;
109 
110 	/*
111 	 * If our parent is NULL (inactive spare or cache) or is the root,
112 	 * just return our own asize.
113 	 */
114 	if (pvd == NULL)
115 		return (vd->vdev_asize);
116 
117 	/*
118 	 * The top-level vdev just returns the allocatable size rounded
119 	 * to the nearest metaslab.
120 	 */
121 	if (vd == vd->vdev_top)
122 		return (P2ALIGN(vd->vdev_asize, 1ULL << vd->vdev_ms_shift));
123 
124 	/*
125 	 * The allocatable space for a raidz vdev is N * sizeof(smallest child),
126 	 * so each child must provide at least 1/Nth of its asize.
127 	 */
128 	if (pvd->vdev_ops == &vdev_raidz_ops)
129 		return (pvd->vdev_min_asize / pvd->vdev_children);
130 
131 	return (pvd->vdev_min_asize);
132 }
133 
134 void
135 vdev_set_min_asize(vdev_t *vd)
136 {
137 	vd->vdev_min_asize = vdev_get_min_asize(vd);
138 
139 	for (int c = 0; c < vd->vdev_children; c++)
140 		vdev_set_min_asize(vd->vdev_child[c]);
141 }
142 
143 vdev_t *
144 vdev_lookup_top(spa_t *spa, uint64_t vdev)
145 {
146 	vdev_t *rvd = spa->spa_root_vdev;
147 
148 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
149 
150 	if (vdev < rvd->vdev_children) {
151 		ASSERT(rvd->vdev_child[vdev] != NULL);
152 		return (rvd->vdev_child[vdev]);
153 	}
154 
155 	return (NULL);
156 }
157 
158 vdev_t *
159 vdev_lookup_by_guid(vdev_t *vd, uint64_t guid)
160 {
161 	vdev_t *mvd;
162 
163 	if (vd->vdev_guid == guid)
164 		return (vd);
165 
166 	for (int c = 0; c < vd->vdev_children; c++)
167 		if ((mvd = vdev_lookup_by_guid(vd->vdev_child[c], guid)) !=
168 		    NULL)
169 			return (mvd);
170 
171 	return (NULL);
172 }
173 
174 void
175 vdev_add_child(vdev_t *pvd, vdev_t *cvd)
176 {
177 	size_t oldsize, newsize;
178 	uint64_t id = cvd->vdev_id;
179 	vdev_t **newchild;
180 
181 	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
182 	ASSERT(cvd->vdev_parent == NULL);
183 
184 	cvd->vdev_parent = pvd;
185 
186 	if (pvd == NULL)
187 		return;
188 
189 	ASSERT(id >= pvd->vdev_children || pvd->vdev_child[id] == NULL);
190 
191 	oldsize = pvd->vdev_children * sizeof (vdev_t *);
192 	pvd->vdev_children = MAX(pvd->vdev_children, id + 1);
193 	newsize = pvd->vdev_children * sizeof (vdev_t *);
194 
195 	newchild = kmem_zalloc(newsize, KM_SLEEP);
196 	if (pvd->vdev_child != NULL) {
197 		bcopy(pvd->vdev_child, newchild, oldsize);
198 		kmem_free(pvd->vdev_child, oldsize);
199 	}
200 
201 	pvd->vdev_child = newchild;
202 	pvd->vdev_child[id] = cvd;
203 
204 	cvd->vdev_top = (pvd->vdev_top ? pvd->vdev_top: cvd);
205 	ASSERT(cvd->vdev_top->vdev_parent->vdev_parent == NULL);
206 
207 	/*
208 	 * Walk up all ancestors to update guid sum.
209 	 */
210 	for (; pvd != NULL; pvd = pvd->vdev_parent)
211 		pvd->vdev_guid_sum += cvd->vdev_guid_sum;
212 }
213 
214 void
215 vdev_remove_child(vdev_t *pvd, vdev_t *cvd)
216 {
217 	int c;
218 	uint_t id = cvd->vdev_id;
219 
220 	ASSERT(cvd->vdev_parent == pvd);
221 
222 	if (pvd == NULL)
223 		return;
224 
225 	ASSERT(id < pvd->vdev_children);
226 	ASSERT(pvd->vdev_child[id] == cvd);
227 
228 	pvd->vdev_child[id] = NULL;
229 	cvd->vdev_parent = NULL;
230 
231 	for (c = 0; c < pvd->vdev_children; c++)
232 		if (pvd->vdev_child[c])
233 			break;
234 
235 	if (c == pvd->vdev_children) {
236 		kmem_free(pvd->vdev_child, c * sizeof (vdev_t *));
237 		pvd->vdev_child = NULL;
238 		pvd->vdev_children = 0;
239 	}
240 
241 	/*
242 	 * Walk up all ancestors to update guid sum.
243 	 */
244 	for (; pvd != NULL; pvd = pvd->vdev_parent)
245 		pvd->vdev_guid_sum -= cvd->vdev_guid_sum;
246 }
247 
248 /*
249  * Remove any holes in the child array.
250  */
251 void
252 vdev_compact_children(vdev_t *pvd)
253 {
254 	vdev_t **newchild, *cvd;
255 	int oldc = pvd->vdev_children;
256 	int newc;
257 
258 	ASSERT(spa_config_held(pvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
259 
260 	for (int c = newc = 0; c < oldc; c++)
261 		if (pvd->vdev_child[c])
262 			newc++;
263 
264 	newchild = kmem_alloc(newc * sizeof (vdev_t *), KM_SLEEP);
265 
266 	for (int c = newc = 0; c < oldc; c++) {
267 		if ((cvd = pvd->vdev_child[c]) != NULL) {
268 			newchild[newc] = cvd;
269 			cvd->vdev_id = newc++;
270 		}
271 	}
272 
273 	kmem_free(pvd->vdev_child, oldc * sizeof (vdev_t *));
274 	pvd->vdev_child = newchild;
275 	pvd->vdev_children = newc;
276 }
277 
278 /*
279  * Allocate and minimally initialize a vdev_t.
280  */
281 vdev_t *
282 vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
283 {
284 	vdev_t *vd;
285 
286 	vd = kmem_zalloc(sizeof (vdev_t), KM_SLEEP);
287 
288 	if (spa->spa_root_vdev == NULL) {
289 		ASSERT(ops == &vdev_root_ops);
290 		spa->spa_root_vdev = vd;
291 		spa->spa_load_guid = spa_generate_guid(NULL);
292 	}
293 
294 	if (guid == 0 && ops != &vdev_hole_ops) {
295 		if (spa->spa_root_vdev == vd) {
296 			/*
297 			 * The root vdev's guid will also be the pool guid,
298 			 * which must be unique among all pools.
299 			 */
300 			guid = spa_generate_guid(NULL);
301 		} else {
302 			/*
303 			 * Any other vdev's guid must be unique within the pool.
304 			 */
305 			guid = spa_generate_guid(spa);
306 		}
307 		ASSERT(!spa_guid_exists(spa_guid(spa), guid));
308 	}
309 
310 	vd->vdev_spa = spa;
311 	vd->vdev_id = id;
312 	vd->vdev_guid = guid;
313 	vd->vdev_guid_sum = guid;
314 	vd->vdev_ops = ops;
315 	vd->vdev_state = VDEV_STATE_CLOSED;
316 	vd->vdev_ishole = (ops == &vdev_hole_ops);
317 
318 	mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_DEFAULT, NULL);
319 	mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL);
320 	mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL);
321 	for (int t = 0; t < DTL_TYPES; t++) {
322 		space_map_create(&vd->vdev_dtl[t], 0, -1ULL, 0,
323 		    &vd->vdev_dtl_lock);
324 	}
325 	txg_list_create(&vd->vdev_ms_list,
326 	    offsetof(struct metaslab, ms_txg_node));
327 	txg_list_create(&vd->vdev_dtl_list,
328 	    offsetof(struct vdev, vdev_dtl_node));
329 	vd->vdev_stat.vs_timestamp = gethrtime();
330 	vdev_queue_init(vd);
331 	vdev_cache_init(vd);
332 
333 	return (vd);
334 }
335 
336 /*
337  * Allocate a new vdev.  The 'alloctype' is used to control whether we are
338  * creating a new vdev or loading an existing one - the behavior is slightly
339  * different for each case.
340  */
341 int
342 vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id,
343     int alloctype)
344 {
345 	vdev_ops_t *ops;
346 	char *type;
347 	uint64_t guid = 0, islog, nparity;
348 	vdev_t *vd;
349 
350 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
351 
352 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) != 0)
353 		return (SET_ERROR(EINVAL));
354 
355 	if ((ops = vdev_getops(type)) == NULL)
356 		return (SET_ERROR(EINVAL));
357 
358 	/*
359 	 * If this is a load, get the vdev guid from the nvlist.
360 	 * Otherwise, vdev_alloc_common() will generate one for us.
361 	 */
362 	if (alloctype == VDEV_ALLOC_LOAD) {
363 		uint64_t label_id;
364 
365 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ID, &label_id) ||
366 		    label_id != id)
367 			return (SET_ERROR(EINVAL));
368 
369 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
370 			return (SET_ERROR(EINVAL));
371 	} else if (alloctype == VDEV_ALLOC_SPARE) {
372 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
373 			return (SET_ERROR(EINVAL));
374 	} else if (alloctype == VDEV_ALLOC_L2CACHE) {
375 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
376 			return (SET_ERROR(EINVAL));
377 	} else if (alloctype == VDEV_ALLOC_ROOTPOOL) {
378 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
379 			return (SET_ERROR(EINVAL));
380 	}
381 
382 	/*
383 	 * The first allocated vdev must be of type 'root'.
384 	 */
385 	if (ops != &vdev_root_ops && spa->spa_root_vdev == NULL)
386 		return (SET_ERROR(EINVAL));
387 
388 	/*
389 	 * Determine whether we're a log vdev.
390 	 */
391 	islog = 0;
392 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &islog);
393 	if (islog && spa_version(spa) < SPA_VERSION_SLOGS)
394 		return (SET_ERROR(ENOTSUP));
395 
396 	if (ops == &vdev_hole_ops && spa_version(spa) < SPA_VERSION_HOLES)
397 		return (SET_ERROR(ENOTSUP));
398 
399 	/*
400 	 * Set the nparity property for RAID-Z vdevs.
401 	 */
402 	nparity = -1ULL;
403 	if (ops == &vdev_raidz_ops) {
404 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY,
405 		    &nparity) == 0) {
406 			if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY)
407 				return (SET_ERROR(EINVAL));
408 			/*
409 			 * Previous versions could only support 1 or 2 parity
410 			 * device.
411 			 */
412 			if (nparity > 1 &&
413 			    spa_version(spa) < SPA_VERSION_RAIDZ2)
414 				return (SET_ERROR(ENOTSUP));
415 			if (nparity > 2 &&
416 			    spa_version(spa) < SPA_VERSION_RAIDZ3)
417 				return (SET_ERROR(ENOTSUP));
418 		} else {
419 			/*
420 			 * We require the parity to be specified for SPAs that
421 			 * support multiple parity levels.
422 			 */
423 			if (spa_version(spa) >= SPA_VERSION_RAIDZ2)
424 				return (SET_ERROR(EINVAL));
425 			/*
426 			 * Otherwise, we default to 1 parity device for RAID-Z.
427 			 */
428 			nparity = 1;
429 		}
430 	} else {
431 		nparity = 0;
432 	}
433 	ASSERT(nparity != -1ULL);
434 
435 	vd = vdev_alloc_common(spa, id, guid, ops);
436 
437 	vd->vdev_islog = islog;
438 	vd->vdev_nparity = nparity;
439 
440 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &vd->vdev_path) == 0)
441 		vd->vdev_path = spa_strdup(vd->vdev_path);
442 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &vd->vdev_devid) == 0)
443 		vd->vdev_devid = spa_strdup(vd->vdev_devid);
444 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH,
445 	    &vd->vdev_physpath) == 0)
446 		vd->vdev_physpath = spa_strdup(vd->vdev_physpath);
447 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &vd->vdev_fru) == 0)
448 		vd->vdev_fru = spa_strdup(vd->vdev_fru);
449 
450 	/*
451 	 * Set the whole_disk property.  If it's not specified, leave the value
452 	 * as -1.
453 	 */
454 	if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
455 	    &vd->vdev_wholedisk) != 0)
456 		vd->vdev_wholedisk = -1ULL;
457 
458 	/*
459 	 * Look for the 'not present' flag.  This will only be set if the device
460 	 * was not present at the time of import.
461 	 */
462 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT,
463 	    &vd->vdev_not_present);
464 
465 	/*
466 	 * Get the alignment requirement.
467 	 */
468 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT, &vd->vdev_ashift);
469 
470 	/*
471 	 * Retrieve the vdev creation time.
472 	 */
473 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_CREATE_TXG,
474 	    &vd->vdev_crtxg);
475 
476 	/*
477 	 * If we're a top-level vdev, try to load the allocation parameters.
478 	 */
479 	if (parent && !parent->vdev_parent &&
480 	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
481 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
482 		    &vd->vdev_ms_array);
483 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
484 		    &vd->vdev_ms_shift);
485 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE,
486 		    &vd->vdev_asize);
487 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVING,
488 		    &vd->vdev_removing);
489 	}
490 
491 	if (parent && !parent->vdev_parent && alloctype != VDEV_ALLOC_ATTACH) {
492 		ASSERT(alloctype == VDEV_ALLOC_LOAD ||
493 		    alloctype == VDEV_ALLOC_ADD ||
494 		    alloctype == VDEV_ALLOC_SPLIT ||
495 		    alloctype == VDEV_ALLOC_ROOTPOOL);
496 		vd->vdev_mg = metaslab_group_create(islog ?
497 		    spa_log_class(spa) : spa_normal_class(spa), vd);
498 	}
499 
500 	/*
501 	 * If we're a leaf vdev, try to load the DTL object and other state.
502 	 */
503 	if (vd->vdev_ops->vdev_op_leaf &&
504 	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE ||
505 	    alloctype == VDEV_ALLOC_ROOTPOOL)) {
506 		if (alloctype == VDEV_ALLOC_LOAD) {
507 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL,
508 			    &vd->vdev_dtl_smo.smo_object);
509 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE,
510 			    &vd->vdev_unspare);
511 		}
512 
513 		if (alloctype == VDEV_ALLOC_ROOTPOOL) {
514 			uint64_t spare = 0;
515 
516 			if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE,
517 			    &spare) == 0 && spare)
518 				spa_spare_add(vd);
519 		}
520 
521 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE,
522 		    &vd->vdev_offline);
523 
524 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_RESILVERING,
525 		    &vd->vdev_resilvering);
526 
527 		/*
528 		 * When importing a pool, we want to ignore the persistent fault
529 		 * state, as the diagnosis made on another system may not be
530 		 * valid in the current context.  Local vdevs will
531 		 * remain in the faulted state.
532 		 */
533 		if (spa_load_state(spa) == SPA_LOAD_OPEN) {
534 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED,
535 			    &vd->vdev_faulted);
536 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED,
537 			    &vd->vdev_degraded);
538 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED,
539 			    &vd->vdev_removed);
540 
541 			if (vd->vdev_faulted || vd->vdev_degraded) {
542 				char *aux;
543 
544 				vd->vdev_label_aux =
545 				    VDEV_AUX_ERR_EXCEEDED;
546 				if (nvlist_lookup_string(nv,
547 				    ZPOOL_CONFIG_AUX_STATE, &aux) == 0 &&
548 				    strcmp(aux, "external") == 0)
549 					vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
550 			}
551 		}
552 	}
553 
554 	/*
555 	 * Add ourselves to the parent's list of children.
556 	 */
557 	vdev_add_child(parent, vd);
558 
559 	*vdp = vd;
560 
561 	return (0);
562 }
563 
564 void
565 vdev_free(vdev_t *vd)
566 {
567 	spa_t *spa = vd->vdev_spa;
568 
569 	/*
570 	 * vdev_free() implies closing the vdev first.  This is simpler than
571 	 * trying to ensure complicated semantics for all callers.
572 	 */
573 	vdev_close(vd);
574 
575 	ASSERT(!list_link_active(&vd->vdev_config_dirty_node));
576 	ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
577 
578 	/*
579 	 * Free all children.
580 	 */
581 	for (int c = 0; c < vd->vdev_children; c++)
582 		vdev_free(vd->vdev_child[c]);
583 
584 	ASSERT(vd->vdev_child == NULL);
585 	ASSERT(vd->vdev_guid_sum == vd->vdev_guid);
586 
587 	/*
588 	 * Discard allocation state.
589 	 */
590 	if (vd->vdev_mg != NULL) {
591 		vdev_metaslab_fini(vd);
592 		metaslab_group_destroy(vd->vdev_mg);
593 	}
594 
595 	ASSERT0(vd->vdev_stat.vs_space);
596 	ASSERT0(vd->vdev_stat.vs_dspace);
597 	ASSERT0(vd->vdev_stat.vs_alloc);
598 
599 	/*
600 	 * Remove this vdev from its parent's child list.
601 	 */
602 	vdev_remove_child(vd->vdev_parent, vd);
603 
604 	ASSERT(vd->vdev_parent == NULL);
605 
606 	/*
607 	 * Clean up vdev structure.
608 	 */
609 	vdev_queue_fini(vd);
610 	vdev_cache_fini(vd);
611 
612 	if (vd->vdev_path)
613 		spa_strfree(vd->vdev_path);
614 	if (vd->vdev_devid)
615 		spa_strfree(vd->vdev_devid);
616 	if (vd->vdev_physpath)
617 		spa_strfree(vd->vdev_physpath);
618 	if (vd->vdev_fru)
619 		spa_strfree(vd->vdev_fru);
620 
621 	if (vd->vdev_isspare)
622 		spa_spare_remove(vd);
623 	if (vd->vdev_isl2cache)
624 		spa_l2cache_remove(vd);
625 
626 	txg_list_destroy(&vd->vdev_ms_list);
627 	txg_list_destroy(&vd->vdev_dtl_list);
628 
629 	mutex_enter(&vd->vdev_dtl_lock);
630 	for (int t = 0; t < DTL_TYPES; t++) {
631 		space_map_unload(&vd->vdev_dtl[t]);
632 		space_map_destroy(&vd->vdev_dtl[t]);
633 	}
634 	mutex_exit(&vd->vdev_dtl_lock);
635 
636 	mutex_destroy(&vd->vdev_dtl_lock);
637 	mutex_destroy(&vd->vdev_stat_lock);
638 	mutex_destroy(&vd->vdev_probe_lock);
639 
640 	if (vd == spa->spa_root_vdev)
641 		spa->spa_root_vdev = NULL;
642 
643 	kmem_free(vd, sizeof (vdev_t));
644 }
645 
646 /*
647  * Transfer top-level vdev state from svd to tvd.
648  */
649 static void
650 vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
651 {
652 	spa_t *spa = svd->vdev_spa;
653 	metaslab_t *msp;
654 	vdev_t *vd;
655 	int t;
656 
657 	ASSERT(tvd == tvd->vdev_top);
658 
659 	tvd->vdev_ms_array = svd->vdev_ms_array;
660 	tvd->vdev_ms_shift = svd->vdev_ms_shift;
661 	tvd->vdev_ms_count = svd->vdev_ms_count;
662 
663 	svd->vdev_ms_array = 0;
664 	svd->vdev_ms_shift = 0;
665 	svd->vdev_ms_count = 0;
666 
667 	if (tvd->vdev_mg)
668 		ASSERT3P(tvd->vdev_mg, ==, svd->vdev_mg);
669 	tvd->vdev_mg = svd->vdev_mg;
670 	tvd->vdev_ms = svd->vdev_ms;
671 
672 	svd->vdev_mg = NULL;
673 	svd->vdev_ms = NULL;
674 
675 	if (tvd->vdev_mg != NULL)
676 		tvd->vdev_mg->mg_vd = tvd;
677 
678 	tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
679 	tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
680 	tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
681 
682 	svd->vdev_stat.vs_alloc = 0;
683 	svd->vdev_stat.vs_space = 0;
684 	svd->vdev_stat.vs_dspace = 0;
685 
686 	for (t = 0; t < TXG_SIZE; t++) {
687 		while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
688 			(void) txg_list_add(&tvd->vdev_ms_list, msp, t);
689 		while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
690 			(void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
691 		if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
692 			(void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
693 	}
694 
695 	if (list_link_active(&svd->vdev_config_dirty_node)) {
696 		vdev_config_clean(svd);
697 		vdev_config_dirty(tvd);
698 	}
699 
700 	if (list_link_active(&svd->vdev_state_dirty_node)) {
701 		vdev_state_clean(svd);
702 		vdev_state_dirty(tvd);
703 	}
704 
705 	tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
706 	svd->vdev_deflate_ratio = 0;
707 
708 	tvd->vdev_islog = svd->vdev_islog;
709 	svd->vdev_islog = 0;
710 }
711 
712 static void
713 vdev_top_update(vdev_t *tvd, vdev_t *vd)
714 {
715 	if (vd == NULL)
716 		return;
717 
718 	vd->vdev_top = tvd;
719 
720 	for (int c = 0; c < vd->vdev_children; c++)
721 		vdev_top_update(tvd, vd->vdev_child[c]);
722 }
723 
724 /*
725  * Add a mirror/replacing vdev above an existing vdev.
726  */
727 vdev_t *
728 vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
729 {
730 	spa_t *spa = cvd->vdev_spa;
731 	vdev_t *pvd = cvd->vdev_parent;
732 	vdev_t *mvd;
733 
734 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
735 
736 	mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
737 
738 	mvd->vdev_asize = cvd->vdev_asize;
739 	mvd->vdev_min_asize = cvd->vdev_min_asize;
740 	mvd->vdev_max_asize = cvd->vdev_max_asize;
741 	mvd->vdev_ashift = cvd->vdev_ashift;
742 	mvd->vdev_state = cvd->vdev_state;
743 	mvd->vdev_crtxg = cvd->vdev_crtxg;
744 
745 	vdev_remove_child(pvd, cvd);
746 	vdev_add_child(pvd, mvd);
747 	cvd->vdev_id = mvd->vdev_children;
748 	vdev_add_child(mvd, cvd);
749 	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
750 
751 	if (mvd == mvd->vdev_top)
752 		vdev_top_transfer(cvd, mvd);
753 
754 	return (mvd);
755 }
756 
757 /*
758  * Remove a 1-way mirror/replacing vdev from the tree.
759  */
760 void
761 vdev_remove_parent(vdev_t *cvd)
762 {
763 	vdev_t *mvd = cvd->vdev_parent;
764 	vdev_t *pvd = mvd->vdev_parent;
765 
766 	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
767 
768 	ASSERT(mvd->vdev_children == 1);
769 	ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
770 	    mvd->vdev_ops == &vdev_replacing_ops ||
771 	    mvd->vdev_ops == &vdev_spare_ops);
772 	cvd->vdev_ashift = mvd->vdev_ashift;
773 
774 	vdev_remove_child(mvd, cvd);
775 	vdev_remove_child(pvd, mvd);
776 
777 	/*
778 	 * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
779 	 * Otherwise, we could have detached an offline device, and when we
780 	 * go to import the pool we'll think we have two top-level vdevs,
781 	 * instead of a different version of the same top-level vdev.
782 	 */
783 	if (mvd->vdev_top == mvd) {
784 		uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
785 		cvd->vdev_orig_guid = cvd->vdev_guid;
786 		cvd->vdev_guid += guid_delta;
787 		cvd->vdev_guid_sum += guid_delta;
788 	}
789 	cvd->vdev_id = mvd->vdev_id;
790 	vdev_add_child(pvd, cvd);
791 	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
792 
793 	if (cvd == cvd->vdev_top)
794 		vdev_top_transfer(mvd, cvd);
795 
796 	ASSERT(mvd->vdev_children == 0);
797 	vdev_free(mvd);
798 }
799 
800 int
801 vdev_metaslab_init(vdev_t *vd, uint64_t txg)
802 {
803 	spa_t *spa = vd->vdev_spa;
804 	objset_t *mos = spa->spa_meta_objset;
805 	uint64_t m;
806 	uint64_t oldc = vd->vdev_ms_count;
807 	uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
808 	metaslab_t **mspp;
809 	int error;
810 
811 	ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER));
812 
813 	/*
814 	 * This vdev is not being allocated from yet or is a hole.
815 	 */
816 	if (vd->vdev_ms_shift == 0)
817 		return (0);
818 
819 	ASSERT(!vd->vdev_ishole);
820 
821 	/*
822 	 * Compute the raidz-deflation ratio.  Note, we hard-code
823 	 * in 128k (1 << 17) because it is the current "typical" blocksize.
824 	 * Even if SPA_MAXBLOCKSIZE changes, this algorithm must never change,
825 	 * or we will inconsistently account for existing bp's.
826 	 */
827 	vd->vdev_deflate_ratio = (1 << 17) /
828 	    (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
829 
830 	ASSERT(oldc <= newc);
831 
832 	mspp = kmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
833 
834 	if (oldc != 0) {
835 		bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp));
836 		kmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
837 	}
838 
839 	vd->vdev_ms = mspp;
840 	vd->vdev_ms_count = newc;
841 
842 	for (m = oldc; m < newc; m++) {
843 		space_map_obj_t smo = { 0, 0, 0 };
844 		if (txg == 0) {
845 			uint64_t object = 0;
846 			error = dmu_read(mos, vd->vdev_ms_array,
847 			    m * sizeof (uint64_t), sizeof (uint64_t), &object,
848 			    DMU_READ_PREFETCH);
849 			if (error)
850 				return (error);
851 			if (object != 0) {
852 				dmu_buf_t *db;
853 				error = dmu_bonus_hold(mos, object, FTAG, &db);
854 				if (error)
855 					return (error);
856 				ASSERT3U(db->db_size, >=, sizeof (smo));
857 				bcopy(db->db_data, &smo, sizeof (smo));
858 				ASSERT3U(smo.smo_object, ==, object);
859 				dmu_buf_rele(db, FTAG);
860 			}
861 		}
862 		vd->vdev_ms[m] = metaslab_init(vd->vdev_mg, &smo,
863 		    m << vd->vdev_ms_shift, 1ULL << vd->vdev_ms_shift, txg);
864 	}
865 
866 	if (txg == 0)
867 		spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER);
868 
869 	/*
870 	 * If the vdev is being removed we don't activate
871 	 * the metaslabs since we want to ensure that no new
872 	 * allocations are performed on this device.
873 	 */
874 	if (oldc == 0 && !vd->vdev_removing)
875 		metaslab_group_activate(vd->vdev_mg);
876 
877 	if (txg == 0)
878 		spa_config_exit(spa, SCL_ALLOC, FTAG);
879 
880 	return (0);
881 }
882 
883 void
884 vdev_metaslab_fini(vdev_t *vd)
885 {
886 	uint64_t m;
887 	uint64_t count = vd->vdev_ms_count;
888 
889 	if (vd->vdev_ms != NULL) {
890 		metaslab_group_passivate(vd->vdev_mg);
891 		for (m = 0; m < count; m++)
892 			if (vd->vdev_ms[m] != NULL)
893 				metaslab_fini(vd->vdev_ms[m]);
894 		kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
895 		vd->vdev_ms = NULL;
896 	}
897 }
898 
899 typedef struct vdev_probe_stats {
900 	boolean_t	vps_readable;
901 	boolean_t	vps_writeable;
902 	int		vps_flags;
903 } vdev_probe_stats_t;
904 
905 static void
906 vdev_probe_done(zio_t *zio)
907 {
908 	spa_t *spa = zio->io_spa;
909 	vdev_t *vd = zio->io_vd;
910 	vdev_probe_stats_t *vps = zio->io_private;
911 
912 	ASSERT(vd->vdev_probe_zio != NULL);
913 
914 	if (zio->io_type == ZIO_TYPE_READ) {
915 		if (zio->io_error == 0)
916 			vps->vps_readable = 1;
917 		if (zio->io_error == 0 && spa_writeable(spa)) {
918 			zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
919 			    zio->io_offset, zio->io_size, zio->io_data,
920 			    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
921 			    ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
922 		} else {
923 			zio_buf_free(zio->io_data, zio->io_size);
924 		}
925 	} else if (zio->io_type == ZIO_TYPE_WRITE) {
926 		if (zio->io_error == 0)
927 			vps->vps_writeable = 1;
928 		zio_buf_free(zio->io_data, zio->io_size);
929 	} else if (zio->io_type == ZIO_TYPE_NULL) {
930 		zio_t *pio;
931 
932 		vd->vdev_cant_read |= !vps->vps_readable;
933 		vd->vdev_cant_write |= !vps->vps_writeable;
934 
935 		if (vdev_readable(vd) &&
936 		    (vdev_writeable(vd) || !spa_writeable(spa))) {
937 			zio->io_error = 0;
938 		} else {
939 			ASSERT(zio->io_error != 0);
940 			zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
941 			    spa, vd, NULL, 0, 0);
942 			zio->io_error = SET_ERROR(ENXIO);
943 		}
944 
945 		mutex_enter(&vd->vdev_probe_lock);
946 		ASSERT(vd->vdev_probe_zio == zio);
947 		vd->vdev_probe_zio = NULL;
948 		mutex_exit(&vd->vdev_probe_lock);
949 
950 		while ((pio = zio_walk_parents(zio)) != NULL)
951 			if (!vdev_accessible(vd, pio))
952 				pio->io_error = SET_ERROR(ENXIO);
953 
954 		kmem_free(vps, sizeof (*vps));
955 	}
956 }
957 
958 /*
959  * Determine whether this device is accessible by reading and writing
960  * to several known locations: the pad regions of each vdev label
961  * but the first (which we leave alone in case it contains a VTOC).
962  */
963 zio_t *
964 vdev_probe(vdev_t *vd, zio_t *zio)
965 {
966 	spa_t *spa = vd->vdev_spa;
967 	vdev_probe_stats_t *vps = NULL;
968 	zio_t *pio;
969 
970 	ASSERT(vd->vdev_ops->vdev_op_leaf);
971 
972 	/*
973 	 * Don't probe the probe.
974 	 */
975 	if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
976 		return (NULL);
977 
978 	/*
979 	 * To prevent 'probe storms' when a device fails, we create
980 	 * just one probe i/o at a time.  All zios that want to probe
981 	 * this vdev will become parents of the probe io.
982 	 */
983 	mutex_enter(&vd->vdev_probe_lock);
984 
985 	if ((pio = vd->vdev_probe_zio) == NULL) {
986 		vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
987 
988 		vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
989 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
990 		    ZIO_FLAG_TRYHARD;
991 
992 		if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
993 			/*
994 			 * vdev_cant_read and vdev_cant_write can only
995 			 * transition from TRUE to FALSE when we have the
996 			 * SCL_ZIO lock as writer; otherwise they can only
997 			 * transition from FALSE to TRUE.  This ensures that
998 			 * any zio looking at these values can assume that
999 			 * failures persist for the life of the I/O.  That's
1000 			 * important because when a device has intermittent
1001 			 * connectivity problems, we want to ensure that
1002 			 * they're ascribed to the device (ENXIO) and not
1003 			 * the zio (EIO).
1004 			 *
1005 			 * Since we hold SCL_ZIO as writer here, clear both
1006 			 * values so the probe can reevaluate from first
1007 			 * principles.
1008 			 */
1009 			vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
1010 			vd->vdev_cant_read = B_FALSE;
1011 			vd->vdev_cant_write = B_FALSE;
1012 		}
1013 
1014 		vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
1015 		    vdev_probe_done, vps,
1016 		    vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
1017 
1018 		/*
1019 		 * We can't change the vdev state in this context, so we
1020 		 * kick off an async task to do it on our behalf.
1021 		 */
1022 		if (zio != NULL) {
1023 			vd->vdev_probe_wanted = B_TRUE;
1024 			spa_async_request(spa, SPA_ASYNC_PROBE);
1025 		}
1026 	}
1027 
1028 	if (zio != NULL)
1029 		zio_add_child(zio, pio);
1030 
1031 	mutex_exit(&vd->vdev_probe_lock);
1032 
1033 	if (vps == NULL) {
1034 		ASSERT(zio != NULL);
1035 		return (NULL);
1036 	}
1037 
1038 	for (int l = 1; l < VDEV_LABELS; l++) {
1039 		zio_nowait(zio_read_phys(pio, vd,
1040 		    vdev_label_offset(vd->vdev_psize, l,
1041 		    offsetof(vdev_label_t, vl_pad2)),
1042 		    VDEV_PAD_SIZE, zio_buf_alloc(VDEV_PAD_SIZE),
1043 		    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1044 		    ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1045 	}
1046 
1047 	if (zio == NULL)
1048 		return (pio);
1049 
1050 	zio_nowait(pio);
1051 	return (NULL);
1052 }
1053 
1054 static void
1055 vdev_open_child(void *arg)
1056 {
1057 	vdev_t *vd = arg;
1058 
1059 	vd->vdev_open_thread = curthread;
1060 	vd->vdev_open_error = vdev_open(vd);
1061 	vd->vdev_open_thread = NULL;
1062 }
1063 
1064 boolean_t
1065 vdev_uses_zvols(vdev_t *vd)
1066 {
1067 	if (vd->vdev_path && strncmp(vd->vdev_path, ZVOL_DIR,
1068 	    strlen(ZVOL_DIR)) == 0)
1069 		return (B_TRUE);
1070 	for (int c = 0; c < vd->vdev_children; c++)
1071 		if (vdev_uses_zvols(vd->vdev_child[c]))
1072 			return (B_TRUE);
1073 	return (B_FALSE);
1074 }
1075 
1076 void
1077 vdev_open_children(vdev_t *vd)
1078 {
1079 	taskq_t *tq;
1080 	int children = vd->vdev_children;
1081 
1082 	/*
1083 	 * in order to handle pools on top of zvols, do the opens
1084 	 * in a single thread so that the same thread holds the
1085 	 * spa_namespace_lock
1086 	 */
1087 	if (vdev_uses_zvols(vd)) {
1088 		for (int c = 0; c < children; c++)
1089 			vd->vdev_child[c]->vdev_open_error =
1090 			    vdev_open(vd->vdev_child[c]);
1091 		return;
1092 	}
1093 	tq = taskq_create("vdev_open", children, minclsyspri,
1094 	    children, children, TASKQ_PREPOPULATE);
1095 
1096 	for (int c = 0; c < children; c++)
1097 		VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c],
1098 		    TQ_SLEEP) != NULL);
1099 
1100 	taskq_destroy(tq);
1101 }
1102 
1103 /*
1104  * Prepare a virtual device for access.
1105  */
1106 int
1107 vdev_open(vdev_t *vd)
1108 {
1109 	spa_t *spa = vd->vdev_spa;
1110 	int error;
1111 	uint64_t osize = 0;
1112 	uint64_t max_osize = 0;
1113 	uint64_t asize, max_asize, psize;
1114 	uint64_t ashift = 0;
1115 
1116 	ASSERT(vd->vdev_open_thread == curthread ||
1117 	    spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1118 	ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1119 	    vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1120 	    vd->vdev_state == VDEV_STATE_OFFLINE);
1121 
1122 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1123 	vd->vdev_cant_read = B_FALSE;
1124 	vd->vdev_cant_write = B_FALSE;
1125 	vd->vdev_min_asize = vdev_get_min_asize(vd);
1126 
1127 	/*
1128 	 * If this vdev is not removed, check its fault status.  If it's
1129 	 * faulted, bail out of the open.
1130 	 */
1131 	if (!vd->vdev_removed && vd->vdev_faulted) {
1132 		ASSERT(vd->vdev_children == 0);
1133 		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1134 		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1135 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1136 		    vd->vdev_label_aux);
1137 		return (SET_ERROR(ENXIO));
1138 	} else if (vd->vdev_offline) {
1139 		ASSERT(vd->vdev_children == 0);
1140 		vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1141 		return (SET_ERROR(ENXIO));
1142 	}
1143 
1144 	error = vd->vdev_ops->vdev_op_open(vd, &osize, &max_osize, &ashift);
1145 
1146 	/*
1147 	 * Reset the vdev_reopening flag so that we actually close
1148 	 * the vdev on error.
1149 	 */
1150 	vd->vdev_reopening = B_FALSE;
1151 	if (zio_injection_enabled && error == 0)
1152 		error = zio_handle_device_injection(vd, NULL, ENXIO);
1153 
1154 	if (error) {
1155 		if (vd->vdev_removed &&
1156 		    vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1157 			vd->vdev_removed = B_FALSE;
1158 
1159 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1160 		    vd->vdev_stat.vs_aux);
1161 		return (error);
1162 	}
1163 
1164 	vd->vdev_removed = B_FALSE;
1165 
1166 	/*
1167 	 * Recheck the faulted flag now that we have confirmed that
1168 	 * the vdev is accessible.  If we're faulted, bail.
1169 	 */
1170 	if (vd->vdev_faulted) {
1171 		ASSERT(vd->vdev_children == 0);
1172 		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1173 		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1174 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1175 		    vd->vdev_label_aux);
1176 		return (SET_ERROR(ENXIO));
1177 	}
1178 
1179 	if (vd->vdev_degraded) {
1180 		ASSERT(vd->vdev_children == 0);
1181 		vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1182 		    VDEV_AUX_ERR_EXCEEDED);
1183 	} else {
1184 		vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1185 	}
1186 
1187 	/*
1188 	 * For hole or missing vdevs we just return success.
1189 	 */
1190 	if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1191 		return (0);
1192 
1193 	for (int c = 0; c < vd->vdev_children; c++) {
1194 		if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1195 			vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1196 			    VDEV_AUX_NONE);
1197 			break;
1198 		}
1199 	}
1200 
1201 	osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1202 	max_osize = P2ALIGN(max_osize, (uint64_t)sizeof (vdev_label_t));
1203 
1204 	if (vd->vdev_children == 0) {
1205 		if (osize < SPA_MINDEVSIZE) {
1206 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1207 			    VDEV_AUX_TOO_SMALL);
1208 			return (SET_ERROR(EOVERFLOW));
1209 		}
1210 		psize = osize;
1211 		asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1212 		max_asize = max_osize - (VDEV_LABEL_START_SIZE +
1213 		    VDEV_LABEL_END_SIZE);
1214 	} else {
1215 		if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1216 		    (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1217 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1218 			    VDEV_AUX_TOO_SMALL);
1219 			return (SET_ERROR(EOVERFLOW));
1220 		}
1221 		psize = 0;
1222 		asize = osize;
1223 		max_asize = max_osize;
1224 	}
1225 
1226 	vd->vdev_psize = psize;
1227 
1228 	/*
1229 	 * Make sure the allocatable size hasn't shrunk.
1230 	 */
1231 	if (asize < vd->vdev_min_asize) {
1232 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1233 		    VDEV_AUX_BAD_LABEL);
1234 		return (SET_ERROR(EINVAL));
1235 	}
1236 
1237 	if (vd->vdev_asize == 0) {
1238 		/*
1239 		 * This is the first-ever open, so use the computed values.
1240 		 * For testing purposes, a higher ashift can be requested.
1241 		 */
1242 		vd->vdev_asize = asize;
1243 		vd->vdev_max_asize = max_asize;
1244 		vd->vdev_ashift = MAX(ashift, vd->vdev_ashift);
1245 	} else {
1246 		/*
1247 		 * Detect if the alignment requirement has increased.
1248 		 * We don't want to make the pool unavailable, just
1249 		 * issue a warning instead.
1250 		 */
1251 		if (ashift > vd->vdev_top->vdev_ashift &&
1252 		    vd->vdev_ops->vdev_op_leaf) {
1253 			cmn_err(CE_WARN,
1254 			    "Disk, '%s', has a block alignment that is "
1255 			    "larger than the pool's alignment\n",
1256 			    vd->vdev_path);
1257 		}
1258 		vd->vdev_max_asize = max_asize;
1259 	}
1260 
1261 	/*
1262 	 * If all children are healthy and the asize has increased,
1263 	 * then we've experienced dynamic LUN growth.  If automatic
1264 	 * expansion is enabled then use the additional space.
1265 	 */
1266 	if (vd->vdev_state == VDEV_STATE_HEALTHY && asize > vd->vdev_asize &&
1267 	    (vd->vdev_expanding || spa->spa_autoexpand))
1268 		vd->vdev_asize = asize;
1269 
1270 	vdev_set_min_asize(vd);
1271 
1272 	/*
1273 	 * Ensure we can issue some IO before declaring the
1274 	 * vdev open for business.
1275 	 */
1276 	if (vd->vdev_ops->vdev_op_leaf &&
1277 	    (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1278 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1279 		    VDEV_AUX_ERR_EXCEEDED);
1280 		return (error);
1281 	}
1282 
1283 	/*
1284 	 * If a leaf vdev has a DTL, and seems healthy, then kick off a
1285 	 * resilver.  But don't do this if we are doing a reopen for a scrub,
1286 	 * since this would just restart the scrub we are already doing.
1287 	 */
1288 	if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1289 	    vdev_resilver_needed(vd, NULL, NULL))
1290 		spa_async_request(spa, SPA_ASYNC_RESILVER);
1291 
1292 	return (0);
1293 }
1294 
1295 /*
1296  * Called once the vdevs are all opened, this routine validates the label
1297  * contents.  This needs to be done before vdev_load() so that we don't
1298  * inadvertently do repair I/Os to the wrong device.
1299  *
1300  * If 'strict' is false ignore the spa guid check. This is necessary because
1301  * if the machine crashed during a re-guid the new guid might have been written
1302  * to all of the vdev labels, but not the cached config. The strict check
1303  * will be performed when the pool is opened again using the mos config.
1304  *
1305  * This function will only return failure if one of the vdevs indicates that it
1306  * has since been destroyed or exported.  This is only possible if
1307  * /etc/zfs/zpool.cache was readonly at the time.  Otherwise, the vdev state
1308  * will be updated but the function will return 0.
1309  */
1310 int
1311 vdev_validate(vdev_t *vd, boolean_t strict)
1312 {
1313 	spa_t *spa = vd->vdev_spa;
1314 	nvlist_t *label;
1315 	uint64_t guid = 0, top_guid;
1316 	uint64_t state;
1317 
1318 	for (int c = 0; c < vd->vdev_children; c++)
1319 		if (vdev_validate(vd->vdev_child[c], strict) != 0)
1320 			return (SET_ERROR(EBADF));
1321 
1322 	/*
1323 	 * If the device has already failed, or was marked offline, don't do
1324 	 * any further validation.  Otherwise, label I/O will fail and we will
1325 	 * overwrite the previous state.
1326 	 */
1327 	if (vd->vdev_ops->vdev_op_leaf && vdev_readable(vd)) {
1328 		uint64_t aux_guid = 0;
1329 		nvlist_t *nvl;
1330 		uint64_t txg = spa_last_synced_txg(spa) != 0 ?
1331 		    spa_last_synced_txg(spa) : -1ULL;
1332 
1333 		if ((label = vdev_label_read_config(vd, txg)) == NULL) {
1334 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1335 			    VDEV_AUX_BAD_LABEL);
1336 			return (0);
1337 		}
1338 
1339 		/*
1340 		 * Determine if this vdev has been split off into another
1341 		 * pool.  If so, then refuse to open it.
1342 		 */
1343 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID,
1344 		    &aux_guid) == 0 && aux_guid == spa_guid(spa)) {
1345 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1346 			    VDEV_AUX_SPLIT_POOL);
1347 			nvlist_free(label);
1348 			return (0);
1349 		}
1350 
1351 		if (strict && (nvlist_lookup_uint64(label,
1352 		    ZPOOL_CONFIG_POOL_GUID, &guid) != 0 ||
1353 		    guid != spa_guid(spa))) {
1354 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1355 			    VDEV_AUX_CORRUPT_DATA);
1356 			nvlist_free(label);
1357 			return (0);
1358 		}
1359 
1360 		if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl)
1361 		    != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID,
1362 		    &aux_guid) != 0)
1363 			aux_guid = 0;
1364 
1365 		/*
1366 		 * If this vdev just became a top-level vdev because its
1367 		 * sibling was detached, it will have adopted the parent's
1368 		 * vdev guid -- but the label may or may not be on disk yet.
1369 		 * Fortunately, either version of the label will have the
1370 		 * same top guid, so if we're a top-level vdev, we can
1371 		 * safely compare to that instead.
1372 		 *
1373 		 * If we split this vdev off instead, then we also check the
1374 		 * original pool's guid.  We don't want to consider the vdev
1375 		 * corrupt if it is partway through a split operation.
1376 		 */
1377 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID,
1378 		    &guid) != 0 ||
1379 		    nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID,
1380 		    &top_guid) != 0 ||
1381 		    ((vd->vdev_guid != guid && vd->vdev_guid != aux_guid) &&
1382 		    (vd->vdev_guid != top_guid || vd != vd->vdev_top))) {
1383 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1384 			    VDEV_AUX_CORRUPT_DATA);
1385 			nvlist_free(label);
1386 			return (0);
1387 		}
1388 
1389 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1390 		    &state) != 0) {
1391 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1392 			    VDEV_AUX_CORRUPT_DATA);
1393 			nvlist_free(label);
1394 			return (0);
1395 		}
1396 
1397 		nvlist_free(label);
1398 
1399 		/*
1400 		 * If this is a verbatim import, no need to check the
1401 		 * state of the pool.
1402 		 */
1403 		if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) &&
1404 		    spa_load_state(spa) == SPA_LOAD_OPEN &&
1405 		    state != POOL_STATE_ACTIVE)
1406 			return (SET_ERROR(EBADF));
1407 
1408 		/*
1409 		 * If we were able to open and validate a vdev that was
1410 		 * previously marked permanently unavailable, clear that state
1411 		 * now.
1412 		 */
1413 		if (vd->vdev_not_present)
1414 			vd->vdev_not_present = 0;
1415 	}
1416 
1417 	return (0);
1418 }
1419 
1420 /*
1421  * Close a virtual device.
1422  */
1423 void
1424 vdev_close(vdev_t *vd)
1425 {
1426 	spa_t *spa = vd->vdev_spa;
1427 	vdev_t *pvd = vd->vdev_parent;
1428 
1429 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1430 
1431 	/*
1432 	 * If our parent is reopening, then we are as well, unless we are
1433 	 * going offline.
1434 	 */
1435 	if (pvd != NULL && pvd->vdev_reopening)
1436 		vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline);
1437 
1438 	vd->vdev_ops->vdev_op_close(vd);
1439 
1440 	vdev_cache_purge(vd);
1441 
1442 	/*
1443 	 * We record the previous state before we close it, so that if we are
1444 	 * doing a reopen(), we don't generate FMA ereports if we notice that
1445 	 * it's still faulted.
1446 	 */
1447 	vd->vdev_prevstate = vd->vdev_state;
1448 
1449 	if (vd->vdev_offline)
1450 		vd->vdev_state = VDEV_STATE_OFFLINE;
1451 	else
1452 		vd->vdev_state = VDEV_STATE_CLOSED;
1453 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1454 }
1455 
1456 void
1457 vdev_hold(vdev_t *vd)
1458 {
1459 	spa_t *spa = vd->vdev_spa;
1460 
1461 	ASSERT(spa_is_root(spa));
1462 	if (spa->spa_state == POOL_STATE_UNINITIALIZED)
1463 		return;
1464 
1465 	for (int c = 0; c < vd->vdev_children; c++)
1466 		vdev_hold(vd->vdev_child[c]);
1467 
1468 	if (vd->vdev_ops->vdev_op_leaf)
1469 		vd->vdev_ops->vdev_op_hold(vd);
1470 }
1471 
1472 void
1473 vdev_rele(vdev_t *vd)
1474 {
1475 	spa_t *spa = vd->vdev_spa;
1476 
1477 	ASSERT(spa_is_root(spa));
1478 	for (int c = 0; c < vd->vdev_children; c++)
1479 		vdev_rele(vd->vdev_child[c]);
1480 
1481 	if (vd->vdev_ops->vdev_op_leaf)
1482 		vd->vdev_ops->vdev_op_rele(vd);
1483 }
1484 
1485 /*
1486  * Reopen all interior vdevs and any unopened leaves.  We don't actually
1487  * reopen leaf vdevs which had previously been opened as they might deadlock
1488  * on the spa_config_lock.  Instead we only obtain the leaf's physical size.
1489  * If the leaf has never been opened then open it, as usual.
1490  */
1491 void
1492 vdev_reopen(vdev_t *vd)
1493 {
1494 	spa_t *spa = vd->vdev_spa;
1495 
1496 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1497 
1498 	/* set the reopening flag unless we're taking the vdev offline */
1499 	vd->vdev_reopening = !vd->vdev_offline;
1500 	vdev_close(vd);
1501 	(void) vdev_open(vd);
1502 
1503 	/*
1504 	 * Call vdev_validate() here to make sure we have the same device.
1505 	 * Otherwise, a device with an invalid label could be successfully
1506 	 * opened in response to vdev_reopen().
1507 	 */
1508 	if (vd->vdev_aux) {
1509 		(void) vdev_validate_aux(vd);
1510 		if (vdev_readable(vd) && vdev_writeable(vd) &&
1511 		    vd->vdev_aux == &spa->spa_l2cache &&
1512 		    !l2arc_vdev_present(vd))
1513 			l2arc_add_vdev(spa, vd);
1514 	} else {
1515 		(void) vdev_validate(vd, B_TRUE);
1516 	}
1517 
1518 	/*
1519 	 * Reassess parent vdev's health.
1520 	 */
1521 	vdev_propagate_state(vd);
1522 }
1523 
1524 int
1525 vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1526 {
1527 	int error;
1528 
1529 	/*
1530 	 * Normally, partial opens (e.g. of a mirror) are allowed.
1531 	 * For a create, however, we want to fail the request if
1532 	 * there are any components we can't open.
1533 	 */
1534 	error = vdev_open(vd);
1535 
1536 	if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
1537 		vdev_close(vd);
1538 		return (error ? error : ENXIO);
1539 	}
1540 
1541 	/*
1542 	 * Recursively initialize all labels.
1543 	 */
1544 	if ((error = vdev_label_init(vd, txg, isreplacing ?
1545 	    VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
1546 		vdev_close(vd);
1547 		return (error);
1548 	}
1549 
1550 	return (0);
1551 }
1552 
1553 void
1554 vdev_metaslab_set_size(vdev_t *vd)
1555 {
1556 	/*
1557 	 * Aim for roughly 200 metaslabs per vdev.
1558 	 */
1559 	vd->vdev_ms_shift = highbit(vd->vdev_asize / 200);
1560 	vd->vdev_ms_shift = MAX(vd->vdev_ms_shift, SPA_MAXBLOCKSHIFT);
1561 }
1562 
1563 void
1564 vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
1565 {
1566 	ASSERT(vd == vd->vdev_top);
1567 	ASSERT(!vd->vdev_ishole);
1568 	ASSERT(ISP2(flags));
1569 	ASSERT(spa_writeable(vd->vdev_spa));
1570 
1571 	if (flags & VDD_METASLAB)
1572 		(void) txg_list_add(&vd->vdev_ms_list, arg, txg);
1573 
1574 	if (flags & VDD_DTL)
1575 		(void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
1576 
1577 	(void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
1578 }
1579 
1580 /*
1581  * DTLs.
1582  *
1583  * A vdev's DTL (dirty time log) is the set of transaction groups for which
1584  * the vdev has less than perfect replication.  There are four kinds of DTL:
1585  *
1586  * DTL_MISSING: txgs for which the vdev has no valid copies of the data
1587  *
1588  * DTL_PARTIAL: txgs for which data is available, but not fully replicated
1589  *
1590  * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
1591  *	scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
1592  *	txgs that was scrubbed.
1593  *
1594  * DTL_OUTAGE: txgs which cannot currently be read, whether due to
1595  *	persistent errors or just some device being offline.
1596  *	Unlike the other three, the DTL_OUTAGE map is not generally
1597  *	maintained; it's only computed when needed, typically to
1598  *	determine whether a device can be detached.
1599  *
1600  * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
1601  * either has the data or it doesn't.
1602  *
1603  * For interior vdevs such as mirror and RAID-Z the picture is more complex.
1604  * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
1605  * if any child is less than fully replicated, then so is its parent.
1606  * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
1607  * comprising only those txgs which appear in 'maxfaults' or more children;
1608  * those are the txgs we don't have enough replication to read.  For example,
1609  * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
1610  * thus, its DTL_MISSING consists of the set of txgs that appear in more than
1611  * two child DTL_MISSING maps.
1612  *
1613  * It should be clear from the above that to compute the DTLs and outage maps
1614  * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
1615  * Therefore, that is all we keep on disk.  When loading the pool, or after
1616  * a configuration change, we generate all other DTLs from first principles.
1617  */
1618 void
1619 vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1620 {
1621 	space_map_t *sm = &vd->vdev_dtl[t];
1622 
1623 	ASSERT(t < DTL_TYPES);
1624 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1625 	ASSERT(spa_writeable(vd->vdev_spa));
1626 
1627 	mutex_enter(sm->sm_lock);
1628 	if (!space_map_contains(sm, txg, size))
1629 		space_map_add(sm, txg, size);
1630 	mutex_exit(sm->sm_lock);
1631 }
1632 
1633 boolean_t
1634 vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1635 {
1636 	space_map_t *sm = &vd->vdev_dtl[t];
1637 	boolean_t dirty = B_FALSE;
1638 
1639 	ASSERT(t < DTL_TYPES);
1640 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1641 
1642 	mutex_enter(sm->sm_lock);
1643 	if (sm->sm_space != 0)
1644 		dirty = space_map_contains(sm, txg, size);
1645 	mutex_exit(sm->sm_lock);
1646 
1647 	return (dirty);
1648 }
1649 
1650 boolean_t
1651 vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
1652 {
1653 	space_map_t *sm = &vd->vdev_dtl[t];
1654 	boolean_t empty;
1655 
1656 	mutex_enter(sm->sm_lock);
1657 	empty = (sm->sm_space == 0);
1658 	mutex_exit(sm->sm_lock);
1659 
1660 	return (empty);
1661 }
1662 
1663 /*
1664  * Reassess DTLs after a config change or scrub completion.
1665  */
1666 void
1667 vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
1668 {
1669 	spa_t *spa = vd->vdev_spa;
1670 	avl_tree_t reftree;
1671 	int minref;
1672 
1673 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
1674 
1675 	for (int c = 0; c < vd->vdev_children; c++)
1676 		vdev_dtl_reassess(vd->vdev_child[c], txg,
1677 		    scrub_txg, scrub_done);
1678 
1679 	if (vd == spa->spa_root_vdev || vd->vdev_ishole || vd->vdev_aux)
1680 		return;
1681 
1682 	if (vd->vdev_ops->vdev_op_leaf) {
1683 		dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1684 
1685 		mutex_enter(&vd->vdev_dtl_lock);
1686 		if (scrub_txg != 0 &&
1687 		    (spa->spa_scrub_started ||
1688 		    (scn && scn->scn_phys.scn_errors == 0))) {
1689 			/*
1690 			 * We completed a scrub up to scrub_txg.  If we
1691 			 * did it without rebooting, then the scrub dtl
1692 			 * will be valid, so excise the old region and
1693 			 * fold in the scrub dtl.  Otherwise, leave the
1694 			 * dtl as-is if there was an error.
1695 			 *
1696 			 * There's little trick here: to excise the beginning
1697 			 * of the DTL_MISSING map, we put it into a reference
1698 			 * tree and then add a segment with refcnt -1 that
1699 			 * covers the range [0, scrub_txg).  This means
1700 			 * that each txg in that range has refcnt -1 or 0.
1701 			 * We then add DTL_SCRUB with a refcnt of 2, so that
1702 			 * entries in the range [0, scrub_txg) will have a
1703 			 * positive refcnt -- either 1 or 2.  We then convert
1704 			 * the reference tree into the new DTL_MISSING map.
1705 			 */
1706 			space_map_ref_create(&reftree);
1707 			space_map_ref_add_map(&reftree,
1708 			    &vd->vdev_dtl[DTL_MISSING], 1);
1709 			space_map_ref_add_seg(&reftree, 0, scrub_txg, -1);
1710 			space_map_ref_add_map(&reftree,
1711 			    &vd->vdev_dtl[DTL_SCRUB], 2);
1712 			space_map_ref_generate_map(&reftree,
1713 			    &vd->vdev_dtl[DTL_MISSING], 1);
1714 			space_map_ref_destroy(&reftree);
1715 		}
1716 		space_map_vacate(&vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
1717 		space_map_walk(&vd->vdev_dtl[DTL_MISSING],
1718 		    space_map_add, &vd->vdev_dtl[DTL_PARTIAL]);
1719 		if (scrub_done)
1720 			space_map_vacate(&vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
1721 		space_map_vacate(&vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
1722 		if (!vdev_readable(vd))
1723 			space_map_add(&vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
1724 		else
1725 			space_map_walk(&vd->vdev_dtl[DTL_MISSING],
1726 			    space_map_add, &vd->vdev_dtl[DTL_OUTAGE]);
1727 		mutex_exit(&vd->vdev_dtl_lock);
1728 
1729 		if (txg != 0)
1730 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
1731 		return;
1732 	}
1733 
1734 	mutex_enter(&vd->vdev_dtl_lock);
1735 	for (int t = 0; t < DTL_TYPES; t++) {
1736 		/* account for child's outage in parent's missing map */
1737 		int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
1738 		if (t == DTL_SCRUB)
1739 			continue;			/* leaf vdevs only */
1740 		if (t == DTL_PARTIAL)
1741 			minref = 1;			/* i.e. non-zero */
1742 		else if (vd->vdev_nparity != 0)
1743 			minref = vd->vdev_nparity + 1;	/* RAID-Z */
1744 		else
1745 			minref = vd->vdev_children;	/* any kind of mirror */
1746 		space_map_ref_create(&reftree);
1747 		for (int c = 0; c < vd->vdev_children; c++) {
1748 			vdev_t *cvd = vd->vdev_child[c];
1749 			mutex_enter(&cvd->vdev_dtl_lock);
1750 			space_map_ref_add_map(&reftree, &cvd->vdev_dtl[s], 1);
1751 			mutex_exit(&cvd->vdev_dtl_lock);
1752 		}
1753 		space_map_ref_generate_map(&reftree, &vd->vdev_dtl[t], minref);
1754 		space_map_ref_destroy(&reftree);
1755 	}
1756 	mutex_exit(&vd->vdev_dtl_lock);
1757 }
1758 
1759 static int
1760 vdev_dtl_load(vdev_t *vd)
1761 {
1762 	spa_t *spa = vd->vdev_spa;
1763 	space_map_obj_t *smo = &vd->vdev_dtl_smo;
1764 	objset_t *mos = spa->spa_meta_objset;
1765 	dmu_buf_t *db;
1766 	int error;
1767 
1768 	ASSERT(vd->vdev_children == 0);
1769 
1770 	if (smo->smo_object == 0)
1771 		return (0);
1772 
1773 	ASSERT(!vd->vdev_ishole);
1774 
1775 	if ((error = dmu_bonus_hold(mos, smo->smo_object, FTAG, &db)) != 0)
1776 		return (error);
1777 
1778 	ASSERT3U(db->db_size, >=, sizeof (*smo));
1779 	bcopy(db->db_data, smo, sizeof (*smo));
1780 	dmu_buf_rele(db, FTAG);
1781 
1782 	mutex_enter(&vd->vdev_dtl_lock);
1783 	error = space_map_load(&vd->vdev_dtl[DTL_MISSING],
1784 	    NULL, SM_ALLOC, smo, mos);
1785 	mutex_exit(&vd->vdev_dtl_lock);
1786 
1787 	return (error);
1788 }
1789 
1790 void
1791 vdev_dtl_sync(vdev_t *vd, uint64_t txg)
1792 {
1793 	spa_t *spa = vd->vdev_spa;
1794 	space_map_obj_t *smo = &vd->vdev_dtl_smo;
1795 	space_map_t *sm = &vd->vdev_dtl[DTL_MISSING];
1796 	objset_t *mos = spa->spa_meta_objset;
1797 	space_map_t smsync;
1798 	kmutex_t smlock;
1799 	dmu_buf_t *db;
1800 	dmu_tx_t *tx;
1801 
1802 	ASSERT(!vd->vdev_ishole);
1803 
1804 	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1805 
1806 	if (vd->vdev_detached) {
1807 		if (smo->smo_object != 0) {
1808 			int err = dmu_object_free(mos, smo->smo_object, tx);
1809 			ASSERT0(err);
1810 			smo->smo_object = 0;
1811 		}
1812 		dmu_tx_commit(tx);
1813 		return;
1814 	}
1815 
1816 	if (smo->smo_object == 0) {
1817 		ASSERT(smo->smo_objsize == 0);
1818 		ASSERT(smo->smo_alloc == 0);
1819 		smo->smo_object = dmu_object_alloc(mos,
1820 		    DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
1821 		    DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
1822 		ASSERT(smo->smo_object != 0);
1823 		vdev_config_dirty(vd->vdev_top);
1824 	}
1825 
1826 	mutex_init(&smlock, NULL, MUTEX_DEFAULT, NULL);
1827 
1828 	space_map_create(&smsync, sm->sm_start, sm->sm_size, sm->sm_shift,
1829 	    &smlock);
1830 
1831 	mutex_enter(&smlock);
1832 
1833 	mutex_enter(&vd->vdev_dtl_lock);
1834 	space_map_walk(sm, space_map_add, &smsync);
1835 	mutex_exit(&vd->vdev_dtl_lock);
1836 
1837 	space_map_truncate(smo, mos, tx);
1838 	space_map_sync(&smsync, SM_ALLOC, smo, mos, tx);
1839 	space_map_vacate(&smsync, NULL, NULL);
1840 
1841 	space_map_destroy(&smsync);
1842 
1843 	mutex_exit(&smlock);
1844 	mutex_destroy(&smlock);
1845 
1846 	VERIFY(0 == dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
1847 	dmu_buf_will_dirty(db, tx);
1848 	ASSERT3U(db->db_size, >=, sizeof (*smo));
1849 	bcopy(smo, db->db_data, sizeof (*smo));
1850 	dmu_buf_rele(db, FTAG);
1851 
1852 	dmu_tx_commit(tx);
1853 }
1854 
1855 /*
1856  * Determine whether the specified vdev can be offlined/detached/removed
1857  * without losing data.
1858  */
1859 boolean_t
1860 vdev_dtl_required(vdev_t *vd)
1861 {
1862 	spa_t *spa = vd->vdev_spa;
1863 	vdev_t *tvd = vd->vdev_top;
1864 	uint8_t cant_read = vd->vdev_cant_read;
1865 	boolean_t required;
1866 
1867 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1868 
1869 	if (vd == spa->spa_root_vdev || vd == tvd)
1870 		return (B_TRUE);
1871 
1872 	/*
1873 	 * Temporarily mark the device as unreadable, and then determine
1874 	 * whether this results in any DTL outages in the top-level vdev.
1875 	 * If not, we can safely offline/detach/remove the device.
1876 	 */
1877 	vd->vdev_cant_read = B_TRUE;
1878 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1879 	required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
1880 	vd->vdev_cant_read = cant_read;
1881 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1882 
1883 	if (!required && zio_injection_enabled)
1884 		required = !!zio_handle_device_injection(vd, NULL, ECHILD);
1885 
1886 	return (required);
1887 }
1888 
1889 /*
1890  * Determine if resilver is needed, and if so the txg range.
1891  */
1892 boolean_t
1893 vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
1894 {
1895 	boolean_t needed = B_FALSE;
1896 	uint64_t thismin = UINT64_MAX;
1897 	uint64_t thismax = 0;
1898 
1899 	if (vd->vdev_children == 0) {
1900 		mutex_enter(&vd->vdev_dtl_lock);
1901 		if (vd->vdev_dtl[DTL_MISSING].sm_space != 0 &&
1902 		    vdev_writeable(vd)) {
1903 			space_seg_t *ss;
1904 
1905 			ss = avl_first(&vd->vdev_dtl[DTL_MISSING].sm_root);
1906 			thismin = ss->ss_start - 1;
1907 			ss = avl_last(&vd->vdev_dtl[DTL_MISSING].sm_root);
1908 			thismax = ss->ss_end;
1909 			needed = B_TRUE;
1910 		}
1911 		mutex_exit(&vd->vdev_dtl_lock);
1912 	} else {
1913 		for (int c = 0; c < vd->vdev_children; c++) {
1914 			vdev_t *cvd = vd->vdev_child[c];
1915 			uint64_t cmin, cmax;
1916 
1917 			if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
1918 				thismin = MIN(thismin, cmin);
1919 				thismax = MAX(thismax, cmax);
1920 				needed = B_TRUE;
1921 			}
1922 		}
1923 	}
1924 
1925 	if (needed && minp) {
1926 		*minp = thismin;
1927 		*maxp = thismax;
1928 	}
1929 	return (needed);
1930 }
1931 
1932 void
1933 vdev_load(vdev_t *vd)
1934 {
1935 	/*
1936 	 * Recursively load all children.
1937 	 */
1938 	for (int c = 0; c < vd->vdev_children; c++)
1939 		vdev_load(vd->vdev_child[c]);
1940 
1941 	/*
1942 	 * If this is a top-level vdev, initialize its metaslabs.
1943 	 */
1944 	if (vd == vd->vdev_top && !vd->vdev_ishole &&
1945 	    (vd->vdev_ashift == 0 || vd->vdev_asize == 0 ||
1946 	    vdev_metaslab_init(vd, 0) != 0))
1947 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1948 		    VDEV_AUX_CORRUPT_DATA);
1949 
1950 	/*
1951 	 * If this is a leaf vdev, load its DTL.
1952 	 */
1953 	if (vd->vdev_ops->vdev_op_leaf && vdev_dtl_load(vd) != 0)
1954 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1955 		    VDEV_AUX_CORRUPT_DATA);
1956 }
1957 
1958 /*
1959  * The special vdev case is used for hot spares and l2cache devices.  Its
1960  * sole purpose it to set the vdev state for the associated vdev.  To do this,
1961  * we make sure that we can open the underlying device, then try to read the
1962  * label, and make sure that the label is sane and that it hasn't been
1963  * repurposed to another pool.
1964  */
1965 int
1966 vdev_validate_aux(vdev_t *vd)
1967 {
1968 	nvlist_t *label;
1969 	uint64_t guid, version;
1970 	uint64_t state;
1971 
1972 	if (!vdev_readable(vd))
1973 		return (0);
1974 
1975 	if ((label = vdev_label_read_config(vd, -1ULL)) == NULL) {
1976 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1977 		    VDEV_AUX_CORRUPT_DATA);
1978 		return (-1);
1979 	}
1980 
1981 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
1982 	    !SPA_VERSION_IS_SUPPORTED(version) ||
1983 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
1984 	    guid != vd->vdev_guid ||
1985 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
1986 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1987 		    VDEV_AUX_CORRUPT_DATA);
1988 		nvlist_free(label);
1989 		return (-1);
1990 	}
1991 
1992 	/*
1993 	 * We don't actually check the pool state here.  If it's in fact in
1994 	 * use by another pool, we update this fact on the fly when requested.
1995 	 */
1996 	nvlist_free(label);
1997 	return (0);
1998 }
1999 
2000 void
2001 vdev_remove(vdev_t *vd, uint64_t txg)
2002 {
2003 	spa_t *spa = vd->vdev_spa;
2004 	objset_t *mos = spa->spa_meta_objset;
2005 	dmu_tx_t *tx;
2006 
2007 	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2008 
2009 	if (vd->vdev_dtl_smo.smo_object) {
2010 		ASSERT0(vd->vdev_dtl_smo.smo_alloc);
2011 		(void) dmu_object_free(mos, vd->vdev_dtl_smo.smo_object, tx);
2012 		vd->vdev_dtl_smo.smo_object = 0;
2013 	}
2014 
2015 	if (vd->vdev_ms != NULL) {
2016 		for (int m = 0; m < vd->vdev_ms_count; m++) {
2017 			metaslab_t *msp = vd->vdev_ms[m];
2018 
2019 			if (msp == NULL || msp->ms_smo.smo_object == 0)
2020 				continue;
2021 
2022 			ASSERT0(msp->ms_smo.smo_alloc);
2023 			(void) dmu_object_free(mos, msp->ms_smo.smo_object, tx);
2024 			msp->ms_smo.smo_object = 0;
2025 		}
2026 	}
2027 
2028 	if (vd->vdev_ms_array) {
2029 		(void) dmu_object_free(mos, vd->vdev_ms_array, tx);
2030 		vd->vdev_ms_array = 0;
2031 		vd->vdev_ms_shift = 0;
2032 	}
2033 	dmu_tx_commit(tx);
2034 }
2035 
2036 void
2037 vdev_sync_done(vdev_t *vd, uint64_t txg)
2038 {
2039 	metaslab_t *msp;
2040 	boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg));
2041 
2042 	ASSERT(!vd->vdev_ishole);
2043 
2044 	while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
2045 		metaslab_sync_done(msp, txg);
2046 
2047 	if (reassess)
2048 		metaslab_sync_reassess(vd->vdev_mg);
2049 }
2050 
2051 void
2052 vdev_sync(vdev_t *vd, uint64_t txg)
2053 {
2054 	spa_t *spa = vd->vdev_spa;
2055 	vdev_t *lvd;
2056 	metaslab_t *msp;
2057 	dmu_tx_t *tx;
2058 
2059 	ASSERT(!vd->vdev_ishole);
2060 
2061 	if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0) {
2062 		ASSERT(vd == vd->vdev_top);
2063 		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2064 		vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
2065 		    DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
2066 		ASSERT(vd->vdev_ms_array != 0);
2067 		vdev_config_dirty(vd);
2068 		dmu_tx_commit(tx);
2069 	}
2070 
2071 	/*
2072 	 * Remove the metadata associated with this vdev once it's empty.
2073 	 */
2074 	if (vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing)
2075 		vdev_remove(vd, txg);
2076 
2077 	while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
2078 		metaslab_sync(msp, txg);
2079 		(void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
2080 	}
2081 
2082 	while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
2083 		vdev_dtl_sync(lvd, txg);
2084 
2085 	(void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
2086 }
2087 
2088 uint64_t
2089 vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
2090 {
2091 	return (vd->vdev_ops->vdev_op_asize(vd, psize));
2092 }
2093 
2094 /*
2095  * Mark the given vdev faulted.  A faulted vdev behaves as if the device could
2096  * not be opened, and no I/O is attempted.
2097  */
2098 int
2099 vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2100 {
2101 	vdev_t *vd, *tvd;
2102 
2103 	spa_vdev_state_enter(spa, SCL_NONE);
2104 
2105 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2106 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2107 
2108 	if (!vd->vdev_ops->vdev_op_leaf)
2109 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2110 
2111 	tvd = vd->vdev_top;
2112 
2113 	/*
2114 	 * We don't directly use the aux state here, but if we do a
2115 	 * vdev_reopen(), we need this value to be present to remember why we
2116 	 * were faulted.
2117 	 */
2118 	vd->vdev_label_aux = aux;
2119 
2120 	/*
2121 	 * Faulted state takes precedence over degraded.
2122 	 */
2123 	vd->vdev_delayed_close = B_FALSE;
2124 	vd->vdev_faulted = 1ULL;
2125 	vd->vdev_degraded = 0ULL;
2126 	vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
2127 
2128 	/*
2129 	 * If this device has the only valid copy of the data, then
2130 	 * back off and simply mark the vdev as degraded instead.
2131 	 */
2132 	if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) {
2133 		vd->vdev_degraded = 1ULL;
2134 		vd->vdev_faulted = 0ULL;
2135 
2136 		/*
2137 		 * If we reopen the device and it's not dead, only then do we
2138 		 * mark it degraded.
2139 		 */
2140 		vdev_reopen(tvd);
2141 
2142 		if (vdev_readable(vd))
2143 			vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
2144 	}
2145 
2146 	return (spa_vdev_state_exit(spa, vd, 0));
2147 }
2148 
2149 /*
2150  * Mark the given vdev degraded.  A degraded vdev is purely an indication to the
2151  * user that something is wrong.  The vdev continues to operate as normal as far
2152  * as I/O is concerned.
2153  */
2154 int
2155 vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2156 {
2157 	vdev_t *vd;
2158 
2159 	spa_vdev_state_enter(spa, SCL_NONE);
2160 
2161 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2162 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2163 
2164 	if (!vd->vdev_ops->vdev_op_leaf)
2165 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2166 
2167 	/*
2168 	 * If the vdev is already faulted, then don't do anything.
2169 	 */
2170 	if (vd->vdev_faulted || vd->vdev_degraded)
2171 		return (spa_vdev_state_exit(spa, NULL, 0));
2172 
2173 	vd->vdev_degraded = 1ULL;
2174 	if (!vdev_is_dead(vd))
2175 		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
2176 		    aux);
2177 
2178 	return (spa_vdev_state_exit(spa, vd, 0));
2179 }
2180 
2181 /*
2182  * Online the given vdev.  If 'unspare' is set, it implies two things.  First,
2183  * any attached spare device should be detached when the device finishes
2184  * resilvering.  Second, the online should be treated like a 'test' online case,
2185  * so no FMA events are generated if the device fails to open.
2186  */
2187 int
2188 vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
2189 {
2190 	vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
2191 
2192 	spa_vdev_state_enter(spa, SCL_NONE);
2193 
2194 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2195 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2196 
2197 	if (!vd->vdev_ops->vdev_op_leaf)
2198 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2199 
2200 	tvd = vd->vdev_top;
2201 	vd->vdev_offline = B_FALSE;
2202 	vd->vdev_tmpoffline = B_FALSE;
2203 	vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
2204 	vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
2205 
2206 	/* XXX - L2ARC 1.0 does not support expansion */
2207 	if (!vd->vdev_aux) {
2208 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2209 			pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
2210 	}
2211 
2212 	vdev_reopen(tvd);
2213 	vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
2214 
2215 	if (!vd->vdev_aux) {
2216 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2217 			pvd->vdev_expanding = B_FALSE;
2218 	}
2219 
2220 	if (newstate)
2221 		*newstate = vd->vdev_state;
2222 	if ((flags & ZFS_ONLINE_UNSPARE) &&
2223 	    !vdev_is_dead(vd) && vd->vdev_parent &&
2224 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2225 	    vd->vdev_parent->vdev_child[0] == vd)
2226 		vd->vdev_unspare = B_TRUE;
2227 
2228 	if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
2229 
2230 		/* XXX - L2ARC 1.0 does not support expansion */
2231 		if (vd->vdev_aux)
2232 			return (spa_vdev_state_exit(spa, vd, ENOTSUP));
2233 		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
2234 	}
2235 	return (spa_vdev_state_exit(spa, vd, 0));
2236 }
2237 
2238 static int
2239 vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
2240 {
2241 	vdev_t *vd, *tvd;
2242 	int error = 0;
2243 	uint64_t generation;
2244 	metaslab_group_t *mg;
2245 
2246 top:
2247 	spa_vdev_state_enter(spa, SCL_ALLOC);
2248 
2249 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2250 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2251 
2252 	if (!vd->vdev_ops->vdev_op_leaf)
2253 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2254 
2255 	tvd = vd->vdev_top;
2256 	mg = tvd->vdev_mg;
2257 	generation = spa->spa_config_generation + 1;
2258 
2259 	/*
2260 	 * If the device isn't already offline, try to offline it.
2261 	 */
2262 	if (!vd->vdev_offline) {
2263 		/*
2264 		 * If this device has the only valid copy of some data,
2265 		 * don't allow it to be offlined. Log devices are always
2266 		 * expendable.
2267 		 */
2268 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2269 		    vdev_dtl_required(vd))
2270 			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2271 
2272 		/*
2273 		 * If the top-level is a slog and it has had allocations
2274 		 * then proceed.  We check that the vdev's metaslab group
2275 		 * is not NULL since it's possible that we may have just
2276 		 * added this vdev but not yet initialized its metaslabs.
2277 		 */
2278 		if (tvd->vdev_islog && mg != NULL) {
2279 			/*
2280 			 * Prevent any future allocations.
2281 			 */
2282 			metaslab_group_passivate(mg);
2283 			(void) spa_vdev_state_exit(spa, vd, 0);
2284 
2285 			error = spa_offline_log(spa);
2286 
2287 			spa_vdev_state_enter(spa, SCL_ALLOC);
2288 
2289 			/*
2290 			 * Check to see if the config has changed.
2291 			 */
2292 			if (error || generation != spa->spa_config_generation) {
2293 				metaslab_group_activate(mg);
2294 				if (error)
2295 					return (spa_vdev_state_exit(spa,
2296 					    vd, error));
2297 				(void) spa_vdev_state_exit(spa, vd, 0);
2298 				goto top;
2299 			}
2300 			ASSERT0(tvd->vdev_stat.vs_alloc);
2301 		}
2302 
2303 		/*
2304 		 * Offline this device and reopen its top-level vdev.
2305 		 * If the top-level vdev is a log device then just offline
2306 		 * it. Otherwise, if this action results in the top-level
2307 		 * vdev becoming unusable, undo it and fail the request.
2308 		 */
2309 		vd->vdev_offline = B_TRUE;
2310 		vdev_reopen(tvd);
2311 
2312 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2313 		    vdev_is_dead(tvd)) {
2314 			vd->vdev_offline = B_FALSE;
2315 			vdev_reopen(tvd);
2316 			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2317 		}
2318 
2319 		/*
2320 		 * Add the device back into the metaslab rotor so that
2321 		 * once we online the device it's open for business.
2322 		 */
2323 		if (tvd->vdev_islog && mg != NULL)
2324 			metaslab_group_activate(mg);
2325 	}
2326 
2327 	vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
2328 
2329 	return (spa_vdev_state_exit(spa, vd, 0));
2330 }
2331 
2332 int
2333 vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
2334 {
2335 	int error;
2336 
2337 	mutex_enter(&spa->spa_vdev_top_lock);
2338 	error = vdev_offline_locked(spa, guid, flags);
2339 	mutex_exit(&spa->spa_vdev_top_lock);
2340 
2341 	return (error);
2342 }
2343 
2344 /*
2345  * Clear the error counts associated with this vdev.  Unlike vdev_online() and
2346  * vdev_offline(), we assume the spa config is locked.  We also clear all
2347  * children.  If 'vd' is NULL, then the user wants to clear all vdevs.
2348  */
2349 void
2350 vdev_clear(spa_t *spa, vdev_t *vd)
2351 {
2352 	vdev_t *rvd = spa->spa_root_vdev;
2353 
2354 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2355 
2356 	if (vd == NULL)
2357 		vd = rvd;
2358 
2359 	vd->vdev_stat.vs_read_errors = 0;
2360 	vd->vdev_stat.vs_write_errors = 0;
2361 	vd->vdev_stat.vs_checksum_errors = 0;
2362 
2363 	for (int c = 0; c < vd->vdev_children; c++)
2364 		vdev_clear(spa, vd->vdev_child[c]);
2365 
2366 	/*
2367 	 * If we're in the FAULTED state or have experienced failed I/O, then
2368 	 * clear the persistent state and attempt to reopen the device.  We
2369 	 * also mark the vdev config dirty, so that the new faulted state is
2370 	 * written out to disk.
2371 	 */
2372 	if (vd->vdev_faulted || vd->vdev_degraded ||
2373 	    !vdev_readable(vd) || !vdev_writeable(vd)) {
2374 
2375 		/*
2376 		 * When reopening in reponse to a clear event, it may be due to
2377 		 * a fmadm repair request.  In this case, if the device is
2378 		 * still broken, we want to still post the ereport again.
2379 		 */
2380 		vd->vdev_forcefault = B_TRUE;
2381 
2382 		vd->vdev_faulted = vd->vdev_degraded = 0ULL;
2383 		vd->vdev_cant_read = B_FALSE;
2384 		vd->vdev_cant_write = B_FALSE;
2385 
2386 		vdev_reopen(vd == rvd ? rvd : vd->vdev_top);
2387 
2388 		vd->vdev_forcefault = B_FALSE;
2389 
2390 		if (vd != rvd && vdev_writeable(vd->vdev_top))
2391 			vdev_state_dirty(vd->vdev_top);
2392 
2393 		if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
2394 			spa_async_request(spa, SPA_ASYNC_RESILVER);
2395 
2396 		spa_event_notify(spa, vd, ESC_ZFS_VDEV_CLEAR);
2397 	}
2398 
2399 	/*
2400 	 * When clearing a FMA-diagnosed fault, we always want to
2401 	 * unspare the device, as we assume that the original spare was
2402 	 * done in response to the FMA fault.
2403 	 */
2404 	if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
2405 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2406 	    vd->vdev_parent->vdev_child[0] == vd)
2407 		vd->vdev_unspare = B_TRUE;
2408 }
2409 
2410 boolean_t
2411 vdev_is_dead(vdev_t *vd)
2412 {
2413 	/*
2414 	 * Holes and missing devices are always considered "dead".
2415 	 * This simplifies the code since we don't have to check for
2416 	 * these types of devices in the various code paths.
2417 	 * Instead we rely on the fact that we skip over dead devices
2418 	 * before issuing I/O to them.
2419 	 */
2420 	return (vd->vdev_state < VDEV_STATE_DEGRADED || vd->vdev_ishole ||
2421 	    vd->vdev_ops == &vdev_missing_ops);
2422 }
2423 
2424 boolean_t
2425 vdev_readable(vdev_t *vd)
2426 {
2427 	return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
2428 }
2429 
2430 boolean_t
2431 vdev_writeable(vdev_t *vd)
2432 {
2433 	return (!vdev_is_dead(vd) && !vd->vdev_cant_write);
2434 }
2435 
2436 boolean_t
2437 vdev_allocatable(vdev_t *vd)
2438 {
2439 	uint64_t state = vd->vdev_state;
2440 
2441 	/*
2442 	 * We currently allow allocations from vdevs which may be in the
2443 	 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
2444 	 * fails to reopen then we'll catch it later when we're holding
2445 	 * the proper locks.  Note that we have to get the vdev state
2446 	 * in a local variable because although it changes atomically,
2447 	 * we're asking two separate questions about it.
2448 	 */
2449 	return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
2450 	    !vd->vdev_cant_write && !vd->vdev_ishole);
2451 }
2452 
2453 boolean_t
2454 vdev_accessible(vdev_t *vd, zio_t *zio)
2455 {
2456 	ASSERT(zio->io_vd == vd);
2457 
2458 	if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
2459 		return (B_FALSE);
2460 
2461 	if (zio->io_type == ZIO_TYPE_READ)
2462 		return (!vd->vdev_cant_read);
2463 
2464 	if (zio->io_type == ZIO_TYPE_WRITE)
2465 		return (!vd->vdev_cant_write);
2466 
2467 	return (B_TRUE);
2468 }
2469 
2470 /*
2471  * Get statistics for the given vdev.
2472  */
2473 void
2474 vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
2475 {
2476 	vdev_t *rvd = vd->vdev_spa->spa_root_vdev;
2477 
2478 	mutex_enter(&vd->vdev_stat_lock);
2479 	bcopy(&vd->vdev_stat, vs, sizeof (*vs));
2480 	vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
2481 	vs->vs_state = vd->vdev_state;
2482 	vs->vs_rsize = vdev_get_min_asize(vd);
2483 	if (vd->vdev_ops->vdev_op_leaf)
2484 		vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
2485 	vs->vs_esize = vd->vdev_max_asize - vd->vdev_asize;
2486 	mutex_exit(&vd->vdev_stat_lock);
2487 
2488 	/*
2489 	 * If we're getting stats on the root vdev, aggregate the I/O counts
2490 	 * over all top-level vdevs (i.e. the direct children of the root).
2491 	 */
2492 	if (vd == rvd) {
2493 		for (int c = 0; c < rvd->vdev_children; c++) {
2494 			vdev_t *cvd = rvd->vdev_child[c];
2495 			vdev_stat_t *cvs = &cvd->vdev_stat;
2496 
2497 			mutex_enter(&vd->vdev_stat_lock);
2498 			for (int t = 0; t < ZIO_TYPES; t++) {
2499 				vs->vs_ops[t] += cvs->vs_ops[t];
2500 				vs->vs_bytes[t] += cvs->vs_bytes[t];
2501 			}
2502 			cvs->vs_scan_removing = cvd->vdev_removing;
2503 			mutex_exit(&vd->vdev_stat_lock);
2504 		}
2505 	}
2506 }
2507 
2508 void
2509 vdev_clear_stats(vdev_t *vd)
2510 {
2511 	mutex_enter(&vd->vdev_stat_lock);
2512 	vd->vdev_stat.vs_space = 0;
2513 	vd->vdev_stat.vs_dspace = 0;
2514 	vd->vdev_stat.vs_alloc = 0;
2515 	mutex_exit(&vd->vdev_stat_lock);
2516 }
2517 
2518 void
2519 vdev_scan_stat_init(vdev_t *vd)
2520 {
2521 	vdev_stat_t *vs = &vd->vdev_stat;
2522 
2523 	for (int c = 0; c < vd->vdev_children; c++)
2524 		vdev_scan_stat_init(vd->vdev_child[c]);
2525 
2526 	mutex_enter(&vd->vdev_stat_lock);
2527 	vs->vs_scan_processed = 0;
2528 	mutex_exit(&vd->vdev_stat_lock);
2529 }
2530 
2531 void
2532 vdev_stat_update(zio_t *zio, uint64_t psize)
2533 {
2534 	spa_t *spa = zio->io_spa;
2535 	vdev_t *rvd = spa->spa_root_vdev;
2536 	vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
2537 	vdev_t *pvd;
2538 	uint64_t txg = zio->io_txg;
2539 	vdev_stat_t *vs = &vd->vdev_stat;
2540 	zio_type_t type = zio->io_type;
2541 	int flags = zio->io_flags;
2542 
2543 	/*
2544 	 * If this i/o is a gang leader, it didn't do any actual work.
2545 	 */
2546 	if (zio->io_gang_tree)
2547 		return;
2548 
2549 	if (zio->io_error == 0) {
2550 		/*
2551 		 * If this is a root i/o, don't count it -- we've already
2552 		 * counted the top-level vdevs, and vdev_get_stats() will
2553 		 * aggregate them when asked.  This reduces contention on
2554 		 * the root vdev_stat_lock and implicitly handles blocks
2555 		 * that compress away to holes, for which there is no i/o.
2556 		 * (Holes never create vdev children, so all the counters
2557 		 * remain zero, which is what we want.)
2558 		 *
2559 		 * Note: this only applies to successful i/o (io_error == 0)
2560 		 * because unlike i/o counts, errors are not additive.
2561 		 * When reading a ditto block, for example, failure of
2562 		 * one top-level vdev does not imply a root-level error.
2563 		 */
2564 		if (vd == rvd)
2565 			return;
2566 
2567 		ASSERT(vd == zio->io_vd);
2568 
2569 		if (flags & ZIO_FLAG_IO_BYPASS)
2570 			return;
2571 
2572 		mutex_enter(&vd->vdev_stat_lock);
2573 
2574 		if (flags & ZIO_FLAG_IO_REPAIR) {
2575 			if (flags & ZIO_FLAG_SCAN_THREAD) {
2576 				dsl_scan_phys_t *scn_phys =
2577 				    &spa->spa_dsl_pool->dp_scan->scn_phys;
2578 				uint64_t *processed = &scn_phys->scn_processed;
2579 
2580 				/* XXX cleanup? */
2581 				if (vd->vdev_ops->vdev_op_leaf)
2582 					atomic_add_64(processed, psize);
2583 				vs->vs_scan_processed += psize;
2584 			}
2585 
2586 			if (flags & ZIO_FLAG_SELF_HEAL)
2587 				vs->vs_self_healed += psize;
2588 		}
2589 
2590 		vs->vs_ops[type]++;
2591 		vs->vs_bytes[type] += psize;
2592 
2593 		mutex_exit(&vd->vdev_stat_lock);
2594 		return;
2595 	}
2596 
2597 	if (flags & ZIO_FLAG_SPECULATIVE)
2598 		return;
2599 
2600 	/*
2601 	 * If this is an I/O error that is going to be retried, then ignore the
2602 	 * error.  Otherwise, the user may interpret B_FAILFAST I/O errors as
2603 	 * hard errors, when in reality they can happen for any number of
2604 	 * innocuous reasons (bus resets, MPxIO link failure, etc).
2605 	 */
2606 	if (zio->io_error == EIO &&
2607 	    !(zio->io_flags & ZIO_FLAG_IO_RETRY))
2608 		return;
2609 
2610 	/*
2611 	 * Intent logs writes won't propagate their error to the root
2612 	 * I/O so don't mark these types of failures as pool-level
2613 	 * errors.
2614 	 */
2615 	if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
2616 		return;
2617 
2618 	mutex_enter(&vd->vdev_stat_lock);
2619 	if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
2620 		if (zio->io_error == ECKSUM)
2621 			vs->vs_checksum_errors++;
2622 		else
2623 			vs->vs_read_errors++;
2624 	}
2625 	if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
2626 		vs->vs_write_errors++;
2627 	mutex_exit(&vd->vdev_stat_lock);
2628 
2629 	if (type == ZIO_TYPE_WRITE && txg != 0 &&
2630 	    (!(flags & ZIO_FLAG_IO_REPAIR) ||
2631 	    (flags & ZIO_FLAG_SCAN_THREAD) ||
2632 	    spa->spa_claiming)) {
2633 		/*
2634 		 * This is either a normal write (not a repair), or it's
2635 		 * a repair induced by the scrub thread, or it's a repair
2636 		 * made by zil_claim() during spa_load() in the first txg.
2637 		 * In the normal case, we commit the DTL change in the same
2638 		 * txg as the block was born.  In the scrub-induced repair
2639 		 * case, we know that scrubs run in first-pass syncing context,
2640 		 * so we commit the DTL change in spa_syncing_txg(spa).
2641 		 * In the zil_claim() case, we commit in spa_first_txg(spa).
2642 		 *
2643 		 * We currently do not make DTL entries for failed spontaneous
2644 		 * self-healing writes triggered by normal (non-scrubbing)
2645 		 * reads, because we have no transactional context in which to
2646 		 * do so -- and it's not clear that it'd be desirable anyway.
2647 		 */
2648 		if (vd->vdev_ops->vdev_op_leaf) {
2649 			uint64_t commit_txg = txg;
2650 			if (flags & ZIO_FLAG_SCAN_THREAD) {
2651 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2652 				ASSERT(spa_sync_pass(spa) == 1);
2653 				vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
2654 				commit_txg = spa_syncing_txg(spa);
2655 			} else if (spa->spa_claiming) {
2656 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2657 				commit_txg = spa_first_txg(spa);
2658 			}
2659 			ASSERT(commit_txg >= spa_syncing_txg(spa));
2660 			if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
2661 				return;
2662 			for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2663 				vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
2664 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
2665 		}
2666 		if (vd != rvd)
2667 			vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
2668 	}
2669 }
2670 
2671 /*
2672  * Update the in-core space usage stats for this vdev, its metaslab class,
2673  * and the root vdev.
2674  */
2675 void
2676 vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
2677     int64_t space_delta)
2678 {
2679 	int64_t dspace_delta = space_delta;
2680 	spa_t *spa = vd->vdev_spa;
2681 	vdev_t *rvd = spa->spa_root_vdev;
2682 	metaslab_group_t *mg = vd->vdev_mg;
2683 	metaslab_class_t *mc = mg ? mg->mg_class : NULL;
2684 
2685 	ASSERT(vd == vd->vdev_top);
2686 
2687 	/*
2688 	 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
2689 	 * factor.  We must calculate this here and not at the root vdev
2690 	 * because the root vdev's psize-to-asize is simply the max of its
2691 	 * childrens', thus not accurate enough for us.
2692 	 */
2693 	ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
2694 	ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
2695 	dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
2696 	    vd->vdev_deflate_ratio;
2697 
2698 	mutex_enter(&vd->vdev_stat_lock);
2699 	vd->vdev_stat.vs_alloc += alloc_delta;
2700 	vd->vdev_stat.vs_space += space_delta;
2701 	vd->vdev_stat.vs_dspace += dspace_delta;
2702 	mutex_exit(&vd->vdev_stat_lock);
2703 
2704 	if (mc == spa_normal_class(spa)) {
2705 		mutex_enter(&rvd->vdev_stat_lock);
2706 		rvd->vdev_stat.vs_alloc += alloc_delta;
2707 		rvd->vdev_stat.vs_space += space_delta;
2708 		rvd->vdev_stat.vs_dspace += dspace_delta;
2709 		mutex_exit(&rvd->vdev_stat_lock);
2710 	}
2711 
2712 	if (mc != NULL) {
2713 		ASSERT(rvd == vd->vdev_parent);
2714 		ASSERT(vd->vdev_ms_count != 0);
2715 
2716 		metaslab_class_space_update(mc,
2717 		    alloc_delta, defer_delta, space_delta, dspace_delta);
2718 	}
2719 }
2720 
2721 /*
2722  * Mark a top-level vdev's config as dirty, placing it on the dirty list
2723  * so that it will be written out next time the vdev configuration is synced.
2724  * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
2725  */
2726 void
2727 vdev_config_dirty(vdev_t *vd)
2728 {
2729 	spa_t *spa = vd->vdev_spa;
2730 	vdev_t *rvd = spa->spa_root_vdev;
2731 	int c;
2732 
2733 	ASSERT(spa_writeable(spa));
2734 
2735 	/*
2736 	 * If this is an aux vdev (as with l2cache and spare devices), then we
2737 	 * update the vdev config manually and set the sync flag.
2738 	 */
2739 	if (vd->vdev_aux != NULL) {
2740 		spa_aux_vdev_t *sav = vd->vdev_aux;
2741 		nvlist_t **aux;
2742 		uint_t naux;
2743 
2744 		for (c = 0; c < sav->sav_count; c++) {
2745 			if (sav->sav_vdevs[c] == vd)
2746 				break;
2747 		}
2748 
2749 		if (c == sav->sav_count) {
2750 			/*
2751 			 * We're being removed.  There's nothing more to do.
2752 			 */
2753 			ASSERT(sav->sav_sync == B_TRUE);
2754 			return;
2755 		}
2756 
2757 		sav->sav_sync = B_TRUE;
2758 
2759 		if (nvlist_lookup_nvlist_array(sav->sav_config,
2760 		    ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
2761 			VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
2762 			    ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
2763 		}
2764 
2765 		ASSERT(c < naux);
2766 
2767 		/*
2768 		 * Setting the nvlist in the middle if the array is a little
2769 		 * sketchy, but it will work.
2770 		 */
2771 		nvlist_free(aux[c]);
2772 		aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0);
2773 
2774 		return;
2775 	}
2776 
2777 	/*
2778 	 * The dirty list is protected by the SCL_CONFIG lock.  The caller
2779 	 * must either hold SCL_CONFIG as writer, or must be the sync thread
2780 	 * (which holds SCL_CONFIG as reader).  There's only one sync thread,
2781 	 * so this is sufficient to ensure mutual exclusion.
2782 	 */
2783 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2784 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2785 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
2786 
2787 	if (vd == rvd) {
2788 		for (c = 0; c < rvd->vdev_children; c++)
2789 			vdev_config_dirty(rvd->vdev_child[c]);
2790 	} else {
2791 		ASSERT(vd == vd->vdev_top);
2792 
2793 		if (!list_link_active(&vd->vdev_config_dirty_node) &&
2794 		    !vd->vdev_ishole)
2795 			list_insert_head(&spa->spa_config_dirty_list, vd);
2796 	}
2797 }
2798 
2799 void
2800 vdev_config_clean(vdev_t *vd)
2801 {
2802 	spa_t *spa = vd->vdev_spa;
2803 
2804 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2805 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2806 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
2807 
2808 	ASSERT(list_link_active(&vd->vdev_config_dirty_node));
2809 	list_remove(&spa->spa_config_dirty_list, vd);
2810 }
2811 
2812 /*
2813  * Mark a top-level vdev's state as dirty, so that the next pass of
2814  * spa_sync() can convert this into vdev_config_dirty().  We distinguish
2815  * the state changes from larger config changes because they require
2816  * much less locking, and are often needed for administrative actions.
2817  */
2818 void
2819 vdev_state_dirty(vdev_t *vd)
2820 {
2821 	spa_t *spa = vd->vdev_spa;
2822 
2823 	ASSERT(spa_writeable(spa));
2824 	ASSERT(vd == vd->vdev_top);
2825 
2826 	/*
2827 	 * The state list is protected by the SCL_STATE lock.  The caller
2828 	 * must either hold SCL_STATE as writer, or must be the sync thread
2829 	 * (which holds SCL_STATE as reader).  There's only one sync thread,
2830 	 * so this is sufficient to ensure mutual exclusion.
2831 	 */
2832 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2833 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2834 	    spa_config_held(spa, SCL_STATE, RW_READER)));
2835 
2836 	if (!list_link_active(&vd->vdev_state_dirty_node) && !vd->vdev_ishole)
2837 		list_insert_head(&spa->spa_state_dirty_list, vd);
2838 }
2839 
2840 void
2841 vdev_state_clean(vdev_t *vd)
2842 {
2843 	spa_t *spa = vd->vdev_spa;
2844 
2845 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2846 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2847 	    spa_config_held(spa, SCL_STATE, RW_READER)));
2848 
2849 	ASSERT(list_link_active(&vd->vdev_state_dirty_node));
2850 	list_remove(&spa->spa_state_dirty_list, vd);
2851 }
2852 
2853 /*
2854  * Propagate vdev state up from children to parent.
2855  */
2856 void
2857 vdev_propagate_state(vdev_t *vd)
2858 {
2859 	spa_t *spa = vd->vdev_spa;
2860 	vdev_t *rvd = spa->spa_root_vdev;
2861 	int degraded = 0, faulted = 0;
2862 	int corrupted = 0;
2863 	vdev_t *child;
2864 
2865 	if (vd->vdev_children > 0) {
2866 		for (int c = 0; c < vd->vdev_children; c++) {
2867 			child = vd->vdev_child[c];
2868 
2869 			/*
2870 			 * Don't factor holes into the decision.
2871 			 */
2872 			if (child->vdev_ishole)
2873 				continue;
2874 
2875 			if (!vdev_readable(child) ||
2876 			    (!vdev_writeable(child) && spa_writeable(spa))) {
2877 				/*
2878 				 * Root special: if there is a top-level log
2879 				 * device, treat the root vdev as if it were
2880 				 * degraded.
2881 				 */
2882 				if (child->vdev_islog && vd == rvd)
2883 					degraded++;
2884 				else
2885 					faulted++;
2886 			} else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
2887 				degraded++;
2888 			}
2889 
2890 			if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
2891 				corrupted++;
2892 		}
2893 
2894 		vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
2895 
2896 		/*
2897 		 * Root special: if there is a top-level vdev that cannot be
2898 		 * opened due to corrupted metadata, then propagate the root
2899 		 * vdev's aux state as 'corrupt' rather than 'insufficient
2900 		 * replicas'.
2901 		 */
2902 		if (corrupted && vd == rvd &&
2903 		    rvd->vdev_state == VDEV_STATE_CANT_OPEN)
2904 			vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
2905 			    VDEV_AUX_CORRUPT_DATA);
2906 	}
2907 
2908 	if (vd->vdev_parent)
2909 		vdev_propagate_state(vd->vdev_parent);
2910 }
2911 
2912 /*
2913  * Set a vdev's state.  If this is during an open, we don't update the parent
2914  * state, because we're in the process of opening children depth-first.
2915  * Otherwise, we propagate the change to the parent.
2916  *
2917  * If this routine places a device in a faulted state, an appropriate ereport is
2918  * generated.
2919  */
2920 void
2921 vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
2922 {
2923 	uint64_t save_state;
2924 	spa_t *spa = vd->vdev_spa;
2925 
2926 	if (state == vd->vdev_state) {
2927 		vd->vdev_stat.vs_aux = aux;
2928 		return;
2929 	}
2930 
2931 	save_state = vd->vdev_state;
2932 
2933 	vd->vdev_state = state;
2934 	vd->vdev_stat.vs_aux = aux;
2935 
2936 	/*
2937 	 * If we are setting the vdev state to anything but an open state, then
2938 	 * always close the underlying device unless the device has requested
2939 	 * a delayed close (i.e. we're about to remove or fault the device).
2940 	 * Otherwise, we keep accessible but invalid devices open forever.
2941 	 * We don't call vdev_close() itself, because that implies some extra
2942 	 * checks (offline, etc) that we don't want here.  This is limited to
2943 	 * leaf devices, because otherwise closing the device will affect other
2944 	 * children.
2945 	 */
2946 	if (!vd->vdev_delayed_close && vdev_is_dead(vd) &&
2947 	    vd->vdev_ops->vdev_op_leaf)
2948 		vd->vdev_ops->vdev_op_close(vd);
2949 
2950 	/*
2951 	 * If we have brought this vdev back into service, we need
2952 	 * to notify fmd so that it can gracefully repair any outstanding
2953 	 * cases due to a missing device.  We do this in all cases, even those
2954 	 * that probably don't correlate to a repaired fault.  This is sure to
2955 	 * catch all cases, and we let the zfs-retire agent sort it out.  If
2956 	 * this is a transient state it's OK, as the retire agent will
2957 	 * double-check the state of the vdev before repairing it.
2958 	 */
2959 	if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf &&
2960 	    vd->vdev_prevstate != state)
2961 		zfs_post_state_change(spa, vd);
2962 
2963 	if (vd->vdev_removed &&
2964 	    state == VDEV_STATE_CANT_OPEN &&
2965 	    (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
2966 		/*
2967 		 * If the previous state is set to VDEV_STATE_REMOVED, then this
2968 		 * device was previously marked removed and someone attempted to
2969 		 * reopen it.  If this failed due to a nonexistent device, then
2970 		 * keep the device in the REMOVED state.  We also let this be if
2971 		 * it is one of our special test online cases, which is only
2972 		 * attempting to online the device and shouldn't generate an FMA
2973 		 * fault.
2974 		 */
2975 		vd->vdev_state = VDEV_STATE_REMOVED;
2976 		vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
2977 	} else if (state == VDEV_STATE_REMOVED) {
2978 		vd->vdev_removed = B_TRUE;
2979 	} else if (state == VDEV_STATE_CANT_OPEN) {
2980 		/*
2981 		 * If we fail to open a vdev during an import or recovery, we
2982 		 * mark it as "not available", which signifies that it was
2983 		 * never there to begin with.  Failure to open such a device
2984 		 * is not considered an error.
2985 		 */
2986 		if ((spa_load_state(spa) == SPA_LOAD_IMPORT ||
2987 		    spa_load_state(spa) == SPA_LOAD_RECOVER) &&
2988 		    vd->vdev_ops->vdev_op_leaf)
2989 			vd->vdev_not_present = 1;
2990 
2991 		/*
2992 		 * Post the appropriate ereport.  If the 'prevstate' field is
2993 		 * set to something other than VDEV_STATE_UNKNOWN, it indicates
2994 		 * that this is part of a vdev_reopen().  In this case, we don't
2995 		 * want to post the ereport if the device was already in the
2996 		 * CANT_OPEN state beforehand.
2997 		 *
2998 		 * If the 'checkremove' flag is set, then this is an attempt to
2999 		 * online the device in response to an insertion event.  If we
3000 		 * hit this case, then we have detected an insertion event for a
3001 		 * faulted or offline device that wasn't in the removed state.
3002 		 * In this scenario, we don't post an ereport because we are
3003 		 * about to replace the device, or attempt an online with
3004 		 * vdev_forcefault, which will generate the fault for us.
3005 		 */
3006 		if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
3007 		    !vd->vdev_not_present && !vd->vdev_checkremove &&
3008 		    vd != spa->spa_root_vdev) {
3009 			const char *class;
3010 
3011 			switch (aux) {
3012 			case VDEV_AUX_OPEN_FAILED:
3013 				class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
3014 				break;
3015 			case VDEV_AUX_CORRUPT_DATA:
3016 				class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
3017 				break;
3018 			case VDEV_AUX_NO_REPLICAS:
3019 				class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
3020 				break;
3021 			case VDEV_AUX_BAD_GUID_SUM:
3022 				class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
3023 				break;
3024 			case VDEV_AUX_TOO_SMALL:
3025 				class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
3026 				break;
3027 			case VDEV_AUX_BAD_LABEL:
3028 				class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
3029 				break;
3030 			default:
3031 				class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
3032 			}
3033 
3034 			zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
3035 		}
3036 
3037 		/* Erase any notion of persistent removed state */
3038 		vd->vdev_removed = B_FALSE;
3039 	} else {
3040 		vd->vdev_removed = B_FALSE;
3041 	}
3042 
3043 	if (!isopen && vd->vdev_parent)
3044 		vdev_propagate_state(vd->vdev_parent);
3045 }
3046 
3047 /*
3048  * Check the vdev configuration to ensure that it's capable of supporting
3049  * a root pool. Currently, we do not support RAID-Z or partial configuration.
3050  * In addition, only a single top-level vdev is allowed and none of the leaves
3051  * can be wholedisks.
3052  */
3053 boolean_t
3054 vdev_is_bootable(vdev_t *vd)
3055 {
3056 	if (!vd->vdev_ops->vdev_op_leaf) {
3057 		char *vdev_type = vd->vdev_ops->vdev_op_type;
3058 
3059 		if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
3060 		    vd->vdev_children > 1) {
3061 			return (B_FALSE);
3062 		} else if (strcmp(vdev_type, VDEV_TYPE_RAIDZ) == 0 ||
3063 		    strcmp(vdev_type, VDEV_TYPE_MISSING) == 0) {
3064 			return (B_FALSE);
3065 		}
3066 	} else if (vd->vdev_wholedisk == 1) {
3067 		return (B_FALSE);
3068 	}
3069 
3070 	for (int c = 0; c < vd->vdev_children; c++) {
3071 		if (!vdev_is_bootable(vd->vdev_child[c]))
3072 			return (B_FALSE);
3073 	}
3074 	return (B_TRUE);
3075 }
3076 
3077 /*
3078  * Load the state from the original vdev tree (ovd) which
3079  * we've retrieved from the MOS config object. If the original
3080  * vdev was offline or faulted then we transfer that state to the
3081  * device in the current vdev tree (nvd).
3082  */
3083 void
3084 vdev_load_log_state(vdev_t *nvd, vdev_t *ovd)
3085 {
3086 	spa_t *spa = nvd->vdev_spa;
3087 
3088 	ASSERT(nvd->vdev_top->vdev_islog);
3089 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3090 	ASSERT3U(nvd->vdev_guid, ==, ovd->vdev_guid);
3091 
3092 	for (int c = 0; c < nvd->vdev_children; c++)
3093 		vdev_load_log_state(nvd->vdev_child[c], ovd->vdev_child[c]);
3094 
3095 	if (nvd->vdev_ops->vdev_op_leaf) {
3096 		/*
3097 		 * Restore the persistent vdev state
3098 		 */
3099 		nvd->vdev_offline = ovd->vdev_offline;
3100 		nvd->vdev_faulted = ovd->vdev_faulted;
3101 		nvd->vdev_degraded = ovd->vdev_degraded;
3102 		nvd->vdev_removed = ovd->vdev_removed;
3103 	}
3104 }
3105 
3106 /*
3107  * Determine if a log device has valid content.  If the vdev was
3108  * removed or faulted in the MOS config then we know that
3109  * the content on the log device has already been written to the pool.
3110  */
3111 boolean_t
3112 vdev_log_state_valid(vdev_t *vd)
3113 {
3114 	if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted &&
3115 	    !vd->vdev_removed)
3116 		return (B_TRUE);
3117 
3118 	for (int c = 0; c < vd->vdev_children; c++)
3119 		if (vdev_log_state_valid(vd->vdev_child[c]))
3120 			return (B_TRUE);
3121 
3122 	return (B_FALSE);
3123 }
3124 
3125 /*
3126  * Expand a vdev if possible.
3127  */
3128 void
3129 vdev_expand(vdev_t *vd, uint64_t txg)
3130 {
3131 	ASSERT(vd->vdev_top == vd);
3132 	ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
3133 
3134 	if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count) {
3135 		VERIFY(vdev_metaslab_init(vd, txg) == 0);
3136 		vdev_config_dirty(vd);
3137 	}
3138 }
3139 
3140 /*
3141  * Split a vdev.
3142  */
3143 void
3144 vdev_split(vdev_t *vd)
3145 {
3146 	vdev_t *cvd, *pvd = vd->vdev_parent;
3147 
3148 	vdev_remove_child(pvd, vd);
3149 	vdev_compact_children(pvd);
3150 
3151 	cvd = pvd->vdev_child[0];
3152 	if (pvd->vdev_children == 1) {
3153 		vdev_remove_parent(cvd);
3154 		cvd->vdev_splitting = B_TRUE;
3155 	}
3156 	vdev_propagate_state(cvd);
3157 }
3158 
3159 void
3160 vdev_deadman(vdev_t *vd)
3161 {
3162 	for (int c = 0; c < vd->vdev_children; c++) {
3163 		vdev_t *cvd = vd->vdev_child[c];
3164 
3165 		vdev_deadman(cvd);
3166 	}
3167 
3168 	if (vd->vdev_ops->vdev_op_leaf) {
3169 		vdev_queue_t *vq = &vd->vdev_queue;
3170 
3171 		mutex_enter(&vq->vq_lock);
3172 		if (avl_numnodes(&vq->vq_pending_tree) > 0) {
3173 			spa_t *spa = vd->vdev_spa;
3174 			zio_t *fio;
3175 			uint64_t delta;
3176 
3177 			/*
3178 			 * Look at the head of all the pending queues,
3179 			 * if any I/O has been outstanding for longer than
3180 			 * the spa_deadman_synctime we panic the system.
3181 			 */
3182 			fio = avl_first(&vq->vq_pending_tree);
3183 			delta = ddi_get_lbolt64() - fio->io_timestamp;
3184 			if (delta > NSEC_TO_TICK(spa_deadman_synctime(spa))) {
3185 				zfs_dbgmsg("SLOW IO: zio timestamp %llu, "
3186 				    "delta %llu, last io %llu",
3187 				    fio->io_timestamp, delta,
3188 				    vq->vq_io_complete_ts);
3189 				fm_panic("I/O to pool '%s' appears to be "
3190 				    "hung.", spa_name(spa));
3191 			}
3192 		}
3193 		mutex_exit(&vq->vq_lock);
3194 	}
3195 }
3196