xref: /titanic_51/usr/src/uts/common/fs/zfs/vdev.c (revision 5a45682c3e7b01faa1761ab8d86f0bed4cc1d363)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
25  * Copyright (c) 2011 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 	 * The 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 (EINVAL);
354 
355 	if ((ops = vdev_getops(type)) == NULL)
356 		return (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 (EINVAL);
368 
369 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
370 			return (EINVAL);
371 	} else if (alloctype == VDEV_ALLOC_SPARE) {
372 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
373 			return (EINVAL);
374 	} else if (alloctype == VDEV_ALLOC_L2CACHE) {
375 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
376 			return (EINVAL);
377 	} else if (alloctype == VDEV_ALLOC_ROOTPOOL) {
378 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
379 			return (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 (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 (ENOTSUP);
395 
396 	if (ops == &vdev_hole_ops && spa_version(spa) < SPA_VERSION_HOLES)
397 		return (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 (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 (ENOTSUP);
415 			if (nparity > 2 &&
416 			    spa_version(spa) < SPA_VERSION_RAIDZ3)
417 				return (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 (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) {
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 	ASSERT3U(vd->vdev_stat.vs_space, ==, 0);
596 	ASSERT3U(vd->vdev_stat.vs_dspace, ==, 0);
597 	ASSERT3U(vd->vdev_stat.vs_alloc, ==, 0);
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 	tvd->vdev_mg = svd->vdev_mg;
668 	tvd->vdev_ms = svd->vdev_ms;
669 
670 	svd->vdev_mg = NULL;
671 	svd->vdev_ms = NULL;
672 
673 	if (tvd->vdev_mg != NULL)
674 		tvd->vdev_mg->mg_vd = tvd;
675 
676 	tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
677 	tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
678 	tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
679 
680 	svd->vdev_stat.vs_alloc = 0;
681 	svd->vdev_stat.vs_space = 0;
682 	svd->vdev_stat.vs_dspace = 0;
683 
684 	for (t = 0; t < TXG_SIZE; t++) {
685 		while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
686 			(void) txg_list_add(&tvd->vdev_ms_list, msp, t);
687 		while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
688 			(void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
689 		if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
690 			(void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
691 	}
692 
693 	if (list_link_active(&svd->vdev_config_dirty_node)) {
694 		vdev_config_clean(svd);
695 		vdev_config_dirty(tvd);
696 	}
697 
698 	if (list_link_active(&svd->vdev_state_dirty_node)) {
699 		vdev_state_clean(svd);
700 		vdev_state_dirty(tvd);
701 	}
702 
703 	tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
704 	svd->vdev_deflate_ratio = 0;
705 
706 	tvd->vdev_islog = svd->vdev_islog;
707 	svd->vdev_islog = 0;
708 }
709 
710 static void
711 vdev_top_update(vdev_t *tvd, vdev_t *vd)
712 {
713 	if (vd == NULL)
714 		return;
715 
716 	vd->vdev_top = tvd;
717 
718 	for (int c = 0; c < vd->vdev_children; c++)
719 		vdev_top_update(tvd, vd->vdev_child[c]);
720 }
721 
722 /*
723  * Add a mirror/replacing vdev above an existing vdev.
724  */
725 vdev_t *
726 vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
727 {
728 	spa_t *spa = cvd->vdev_spa;
729 	vdev_t *pvd = cvd->vdev_parent;
730 	vdev_t *mvd;
731 
732 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
733 
734 	mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
735 
736 	mvd->vdev_asize = cvd->vdev_asize;
737 	mvd->vdev_min_asize = cvd->vdev_min_asize;
738 	mvd->vdev_ashift = cvd->vdev_ashift;
739 	mvd->vdev_state = cvd->vdev_state;
740 	mvd->vdev_crtxg = cvd->vdev_crtxg;
741 
742 	vdev_remove_child(pvd, cvd);
743 	vdev_add_child(pvd, mvd);
744 	cvd->vdev_id = mvd->vdev_children;
745 	vdev_add_child(mvd, cvd);
746 	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
747 
748 	if (mvd == mvd->vdev_top)
749 		vdev_top_transfer(cvd, mvd);
750 
751 	return (mvd);
752 }
753 
754 /*
755  * Remove a 1-way mirror/replacing vdev from the tree.
756  */
757 void
758 vdev_remove_parent(vdev_t *cvd)
759 {
760 	vdev_t *mvd = cvd->vdev_parent;
761 	vdev_t *pvd = mvd->vdev_parent;
762 
763 	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
764 
765 	ASSERT(mvd->vdev_children == 1);
766 	ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
767 	    mvd->vdev_ops == &vdev_replacing_ops ||
768 	    mvd->vdev_ops == &vdev_spare_ops);
769 	cvd->vdev_ashift = mvd->vdev_ashift;
770 
771 	vdev_remove_child(mvd, cvd);
772 	vdev_remove_child(pvd, mvd);
773 
774 	/*
775 	 * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
776 	 * Otherwise, we could have detached an offline device, and when we
777 	 * go to import the pool we'll think we have two top-level vdevs,
778 	 * instead of a different version of the same top-level vdev.
779 	 */
780 	if (mvd->vdev_top == mvd) {
781 		uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
782 		cvd->vdev_orig_guid = cvd->vdev_guid;
783 		cvd->vdev_guid += guid_delta;
784 		cvd->vdev_guid_sum += guid_delta;
785 	}
786 	cvd->vdev_id = mvd->vdev_id;
787 	vdev_add_child(pvd, cvd);
788 	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
789 
790 	if (cvd == cvd->vdev_top)
791 		vdev_top_transfer(mvd, cvd);
792 
793 	ASSERT(mvd->vdev_children == 0);
794 	vdev_free(mvd);
795 }
796 
797 int
798 vdev_metaslab_init(vdev_t *vd, uint64_t txg)
799 {
800 	spa_t *spa = vd->vdev_spa;
801 	objset_t *mos = spa->spa_meta_objset;
802 	uint64_t m;
803 	uint64_t oldc = vd->vdev_ms_count;
804 	uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
805 	metaslab_t **mspp;
806 	int error;
807 
808 	ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER));
809 
810 	/*
811 	 * This vdev is not being allocated from yet or is a hole.
812 	 */
813 	if (vd->vdev_ms_shift == 0)
814 		return (0);
815 
816 	ASSERT(!vd->vdev_ishole);
817 
818 	/*
819 	 * Compute the raidz-deflation ratio.  Note, we hard-code
820 	 * in 128k (1 << 17) because it is the current "typical" blocksize.
821 	 * Even if SPA_MAXBLOCKSIZE changes, this algorithm must never change,
822 	 * or we will inconsistently account for existing bp's.
823 	 */
824 	vd->vdev_deflate_ratio = (1 << 17) /
825 	    (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
826 
827 	ASSERT(oldc <= newc);
828 
829 	mspp = kmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
830 
831 	if (oldc != 0) {
832 		bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp));
833 		kmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
834 	}
835 
836 	vd->vdev_ms = mspp;
837 	vd->vdev_ms_count = newc;
838 
839 	for (m = oldc; m < newc; m++) {
840 		space_map_obj_t smo = { 0, 0, 0 };
841 		if (txg == 0) {
842 			uint64_t object = 0;
843 			error = dmu_read(mos, vd->vdev_ms_array,
844 			    m * sizeof (uint64_t), sizeof (uint64_t), &object,
845 			    DMU_READ_PREFETCH);
846 			if (error)
847 				return (error);
848 			if (object != 0) {
849 				dmu_buf_t *db;
850 				error = dmu_bonus_hold(mos, object, FTAG, &db);
851 				if (error)
852 					return (error);
853 				ASSERT3U(db->db_size, >=, sizeof (smo));
854 				bcopy(db->db_data, &smo, sizeof (smo));
855 				ASSERT3U(smo.smo_object, ==, object);
856 				dmu_buf_rele(db, FTAG);
857 			}
858 		}
859 		vd->vdev_ms[m] = metaslab_init(vd->vdev_mg, &smo,
860 		    m << vd->vdev_ms_shift, 1ULL << vd->vdev_ms_shift, txg);
861 	}
862 
863 	if (txg == 0)
864 		spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER);
865 
866 	/*
867 	 * If the vdev is being removed we don't activate
868 	 * the metaslabs since we want to ensure that no new
869 	 * allocations are performed on this device.
870 	 */
871 	if (oldc == 0 && !vd->vdev_removing)
872 		metaslab_group_activate(vd->vdev_mg);
873 
874 	if (txg == 0)
875 		spa_config_exit(spa, SCL_ALLOC, FTAG);
876 
877 	return (0);
878 }
879 
880 void
881 vdev_metaslab_fini(vdev_t *vd)
882 {
883 	uint64_t m;
884 	uint64_t count = vd->vdev_ms_count;
885 
886 	if (vd->vdev_ms != NULL) {
887 		metaslab_group_passivate(vd->vdev_mg);
888 		for (m = 0; m < count; m++)
889 			if (vd->vdev_ms[m] != NULL)
890 				metaslab_fini(vd->vdev_ms[m]);
891 		kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
892 		vd->vdev_ms = NULL;
893 	}
894 }
895 
896 typedef struct vdev_probe_stats {
897 	boolean_t	vps_readable;
898 	boolean_t	vps_writeable;
899 	int		vps_flags;
900 } vdev_probe_stats_t;
901 
902 static void
903 vdev_probe_done(zio_t *zio)
904 {
905 	spa_t *spa = zio->io_spa;
906 	vdev_t *vd = zio->io_vd;
907 	vdev_probe_stats_t *vps = zio->io_private;
908 
909 	ASSERT(vd->vdev_probe_zio != NULL);
910 
911 	if (zio->io_type == ZIO_TYPE_READ) {
912 		if (zio->io_error == 0)
913 			vps->vps_readable = 1;
914 		if (zio->io_error == 0 && spa_writeable(spa)) {
915 			zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
916 			    zio->io_offset, zio->io_size, zio->io_data,
917 			    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
918 			    ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
919 		} else {
920 			zio_buf_free(zio->io_data, zio->io_size);
921 		}
922 	} else if (zio->io_type == ZIO_TYPE_WRITE) {
923 		if (zio->io_error == 0)
924 			vps->vps_writeable = 1;
925 		zio_buf_free(zio->io_data, zio->io_size);
926 	} else if (zio->io_type == ZIO_TYPE_NULL) {
927 		zio_t *pio;
928 
929 		vd->vdev_cant_read |= !vps->vps_readable;
930 		vd->vdev_cant_write |= !vps->vps_writeable;
931 
932 		if (vdev_readable(vd) &&
933 		    (vdev_writeable(vd) || !spa_writeable(spa))) {
934 			zio->io_error = 0;
935 		} else {
936 			ASSERT(zio->io_error != 0);
937 			zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
938 			    spa, vd, NULL, 0, 0);
939 			zio->io_error = ENXIO;
940 		}
941 
942 		mutex_enter(&vd->vdev_probe_lock);
943 		ASSERT(vd->vdev_probe_zio == zio);
944 		vd->vdev_probe_zio = NULL;
945 		mutex_exit(&vd->vdev_probe_lock);
946 
947 		while ((pio = zio_walk_parents(zio)) != NULL)
948 			if (!vdev_accessible(vd, pio))
949 				pio->io_error = ENXIO;
950 
951 		kmem_free(vps, sizeof (*vps));
952 	}
953 }
954 
955 /*
956  * Determine whether this device is accessible by reading and writing
957  * to several known locations: the pad regions of each vdev label
958  * but the first (which we leave alone in case it contains a VTOC).
959  */
960 zio_t *
961 vdev_probe(vdev_t *vd, zio_t *zio)
962 {
963 	spa_t *spa = vd->vdev_spa;
964 	vdev_probe_stats_t *vps = NULL;
965 	zio_t *pio;
966 
967 	ASSERT(vd->vdev_ops->vdev_op_leaf);
968 
969 	/*
970 	 * Don't probe the probe.
971 	 */
972 	if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
973 		return (NULL);
974 
975 	/*
976 	 * To prevent 'probe storms' when a device fails, we create
977 	 * just one probe i/o at a time.  All zios that want to probe
978 	 * this vdev will become parents of the probe io.
979 	 */
980 	mutex_enter(&vd->vdev_probe_lock);
981 
982 	if ((pio = vd->vdev_probe_zio) == NULL) {
983 		vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
984 
985 		vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
986 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
987 		    ZIO_FLAG_TRYHARD;
988 
989 		if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
990 			/*
991 			 * vdev_cant_read and vdev_cant_write can only
992 			 * transition from TRUE to FALSE when we have the
993 			 * SCL_ZIO lock as writer; otherwise they can only
994 			 * transition from FALSE to TRUE.  This ensures that
995 			 * any zio looking at these values can assume that
996 			 * failures persist for the life of the I/O.  That's
997 			 * important because when a device has intermittent
998 			 * connectivity problems, we want to ensure that
999 			 * they're ascribed to the device (ENXIO) and not
1000 			 * the zio (EIO).
1001 			 *
1002 			 * Since we hold SCL_ZIO as writer here, clear both
1003 			 * values so the probe can reevaluate from first
1004 			 * principles.
1005 			 */
1006 			vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
1007 			vd->vdev_cant_read = B_FALSE;
1008 			vd->vdev_cant_write = B_FALSE;
1009 		}
1010 
1011 		vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
1012 		    vdev_probe_done, vps,
1013 		    vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
1014 
1015 		/*
1016 		 * We can't change the vdev state in this context, so we
1017 		 * kick off an async task to do it on our behalf.
1018 		 */
1019 		if (zio != NULL) {
1020 			vd->vdev_probe_wanted = B_TRUE;
1021 			spa_async_request(spa, SPA_ASYNC_PROBE);
1022 		}
1023 	}
1024 
1025 	if (zio != NULL)
1026 		zio_add_child(zio, pio);
1027 
1028 	mutex_exit(&vd->vdev_probe_lock);
1029 
1030 	if (vps == NULL) {
1031 		ASSERT(zio != NULL);
1032 		return (NULL);
1033 	}
1034 
1035 	for (int l = 1; l < VDEV_LABELS; l++) {
1036 		zio_nowait(zio_read_phys(pio, vd,
1037 		    vdev_label_offset(vd->vdev_psize, l,
1038 		    offsetof(vdev_label_t, vl_pad2)),
1039 		    VDEV_PAD_SIZE, zio_buf_alloc(VDEV_PAD_SIZE),
1040 		    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1041 		    ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1042 	}
1043 
1044 	if (zio == NULL)
1045 		return (pio);
1046 
1047 	zio_nowait(pio);
1048 	return (NULL);
1049 }
1050 
1051 static void
1052 vdev_open_child(void *arg)
1053 {
1054 	vdev_t *vd = arg;
1055 
1056 	vd->vdev_open_thread = curthread;
1057 	vd->vdev_open_error = vdev_open(vd);
1058 	vd->vdev_open_thread = NULL;
1059 }
1060 
1061 boolean_t
1062 vdev_uses_zvols(vdev_t *vd)
1063 {
1064 	if (vd->vdev_path && strncmp(vd->vdev_path, ZVOL_DIR,
1065 	    strlen(ZVOL_DIR)) == 0)
1066 		return (B_TRUE);
1067 	for (int c = 0; c < vd->vdev_children; c++)
1068 		if (vdev_uses_zvols(vd->vdev_child[c]))
1069 			return (B_TRUE);
1070 	return (B_FALSE);
1071 }
1072 
1073 void
1074 vdev_open_children(vdev_t *vd)
1075 {
1076 	taskq_t *tq;
1077 	int children = vd->vdev_children;
1078 
1079 	/*
1080 	 * in order to handle pools on top of zvols, do the opens
1081 	 * in a single thread so that the same thread holds the
1082 	 * spa_namespace_lock
1083 	 */
1084 	if (vdev_uses_zvols(vd)) {
1085 		for (int c = 0; c < children; c++)
1086 			vd->vdev_child[c]->vdev_open_error =
1087 			    vdev_open(vd->vdev_child[c]);
1088 		return;
1089 	}
1090 	tq = taskq_create("vdev_open", children, minclsyspri,
1091 	    children, children, TASKQ_PREPOPULATE);
1092 
1093 	for (int c = 0; c < children; c++)
1094 		VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c],
1095 		    TQ_SLEEP) != NULL);
1096 
1097 	taskq_destroy(tq);
1098 }
1099 
1100 /*
1101  * Prepare a virtual device for access.
1102  */
1103 int
1104 vdev_open(vdev_t *vd)
1105 {
1106 	spa_t *spa = vd->vdev_spa;
1107 	int error;
1108 	uint64_t osize = 0;
1109 	uint64_t asize, psize;
1110 	uint64_t ashift = 0;
1111 
1112 	ASSERT(vd->vdev_open_thread == curthread ||
1113 	    spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1114 	ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1115 	    vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1116 	    vd->vdev_state == VDEV_STATE_OFFLINE);
1117 
1118 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1119 	vd->vdev_cant_read = B_FALSE;
1120 	vd->vdev_cant_write = B_FALSE;
1121 	vd->vdev_min_asize = vdev_get_min_asize(vd);
1122 
1123 	/*
1124 	 * If this vdev is not removed, check its fault status.  If it's
1125 	 * faulted, bail out of the open.
1126 	 */
1127 	if (!vd->vdev_removed && vd->vdev_faulted) {
1128 		ASSERT(vd->vdev_children == 0);
1129 		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1130 		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1131 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1132 		    vd->vdev_label_aux);
1133 		return (ENXIO);
1134 	} else if (vd->vdev_offline) {
1135 		ASSERT(vd->vdev_children == 0);
1136 		vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1137 		return (ENXIO);
1138 	}
1139 
1140 	error = vd->vdev_ops->vdev_op_open(vd, &osize, &ashift);
1141 
1142 	/*
1143 	 * Reset the vdev_reopening flag so that we actually close
1144 	 * the vdev on error.
1145 	 */
1146 	vd->vdev_reopening = B_FALSE;
1147 	if (zio_injection_enabled && error == 0)
1148 		error = zio_handle_device_injection(vd, NULL, ENXIO);
1149 
1150 	if (error) {
1151 		if (vd->vdev_removed &&
1152 		    vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1153 			vd->vdev_removed = B_FALSE;
1154 
1155 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1156 		    vd->vdev_stat.vs_aux);
1157 		return (error);
1158 	}
1159 
1160 	vd->vdev_removed = B_FALSE;
1161 
1162 	/*
1163 	 * Recheck the faulted flag now that we have confirmed that
1164 	 * the vdev is accessible.  If we're faulted, bail.
1165 	 */
1166 	if (vd->vdev_faulted) {
1167 		ASSERT(vd->vdev_children == 0);
1168 		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1169 		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1170 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1171 		    vd->vdev_label_aux);
1172 		return (ENXIO);
1173 	}
1174 
1175 	if (vd->vdev_degraded) {
1176 		ASSERT(vd->vdev_children == 0);
1177 		vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1178 		    VDEV_AUX_ERR_EXCEEDED);
1179 	} else {
1180 		vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1181 	}
1182 
1183 	/*
1184 	 * For hole or missing vdevs we just return success.
1185 	 */
1186 	if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1187 		return (0);
1188 
1189 	for (int c = 0; c < vd->vdev_children; c++) {
1190 		if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1191 			vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1192 			    VDEV_AUX_NONE);
1193 			break;
1194 		}
1195 	}
1196 
1197 	osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1198 
1199 	if (vd->vdev_children == 0) {
1200 		if (osize < SPA_MINDEVSIZE) {
1201 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1202 			    VDEV_AUX_TOO_SMALL);
1203 			return (EOVERFLOW);
1204 		}
1205 		psize = osize;
1206 		asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1207 	} else {
1208 		if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1209 		    (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1210 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1211 			    VDEV_AUX_TOO_SMALL);
1212 			return (EOVERFLOW);
1213 		}
1214 		psize = 0;
1215 		asize = osize;
1216 	}
1217 
1218 	vd->vdev_psize = psize;
1219 
1220 	/*
1221 	 * Make sure the allocatable size hasn't shrunk.
1222 	 */
1223 	if (asize < vd->vdev_min_asize) {
1224 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1225 		    VDEV_AUX_BAD_LABEL);
1226 		return (EINVAL);
1227 	}
1228 
1229 	if (vd->vdev_asize == 0) {
1230 		/*
1231 		 * This is the first-ever open, so use the computed values.
1232 		 * For testing purposes, a higher ashift can be requested.
1233 		 */
1234 		vd->vdev_asize = asize;
1235 		vd->vdev_ashift = MAX(ashift, vd->vdev_ashift);
1236 	} else {
1237 		/*
1238 		 * Make sure the alignment requirement hasn't increased.
1239 		 */
1240 		if (ashift > vd->vdev_top->vdev_ashift) {
1241 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1242 			    VDEV_AUX_BAD_LABEL);
1243 			return (EINVAL);
1244 		}
1245 	}
1246 
1247 	/*
1248 	 * If all children are healthy and the asize has increased,
1249 	 * then we've experienced dynamic LUN growth.  If automatic
1250 	 * expansion is enabled then use the additional space.
1251 	 */
1252 	if (vd->vdev_state == VDEV_STATE_HEALTHY && asize > vd->vdev_asize &&
1253 	    (vd->vdev_expanding || spa->spa_autoexpand))
1254 		vd->vdev_asize = asize;
1255 
1256 	vdev_set_min_asize(vd);
1257 
1258 	/*
1259 	 * Ensure we can issue some IO before declaring the
1260 	 * vdev open for business.
1261 	 */
1262 	if (vd->vdev_ops->vdev_op_leaf &&
1263 	    (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1264 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1265 		    VDEV_AUX_ERR_EXCEEDED);
1266 		return (error);
1267 	}
1268 
1269 	/*
1270 	 * If a leaf vdev has a DTL, and seems healthy, then kick off a
1271 	 * resilver.  But don't do this if we are doing a reopen for a scrub,
1272 	 * since this would just restart the scrub we are already doing.
1273 	 */
1274 	if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1275 	    vdev_resilver_needed(vd, NULL, NULL))
1276 		spa_async_request(spa, SPA_ASYNC_RESILVER);
1277 
1278 	return (0);
1279 }
1280 
1281 /*
1282  * Called once the vdevs are all opened, this routine validates the label
1283  * contents.  This needs to be done before vdev_load() so that we don't
1284  * inadvertently do repair I/Os to the wrong device.
1285  *
1286  * This function will only return failure if one of the vdevs indicates that it
1287  * has since been destroyed or exported.  This is only possible if
1288  * /etc/zfs/zpool.cache was readonly at the time.  Otherwise, the vdev state
1289  * will be updated but the function will return 0.
1290  */
1291 int
1292 vdev_validate(vdev_t *vd)
1293 {
1294 	spa_t *spa = vd->vdev_spa;
1295 	nvlist_t *label;
1296 	uint64_t guid = 0, top_guid;
1297 	uint64_t state;
1298 
1299 	for (int c = 0; c < vd->vdev_children; c++)
1300 		if (vdev_validate(vd->vdev_child[c]) != 0)
1301 			return (EBADF);
1302 
1303 	/*
1304 	 * If the device has already failed, or was marked offline, don't do
1305 	 * any further validation.  Otherwise, label I/O will fail and we will
1306 	 * overwrite the previous state.
1307 	 */
1308 	if (vd->vdev_ops->vdev_op_leaf && vdev_readable(vd)) {
1309 		uint64_t aux_guid = 0;
1310 		nvlist_t *nvl;
1311 
1312 		if ((label = vdev_label_read_config(vd)) == NULL) {
1313 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1314 			    VDEV_AUX_BAD_LABEL);
1315 			return (0);
1316 		}
1317 
1318 		/*
1319 		 * Determine if this vdev has been split off into another
1320 		 * pool.  If so, then refuse to open it.
1321 		 */
1322 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID,
1323 		    &aux_guid) == 0 && aux_guid == spa_guid(spa)) {
1324 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1325 			    VDEV_AUX_SPLIT_POOL);
1326 			nvlist_free(label);
1327 			return (0);
1328 		}
1329 
1330 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_GUID,
1331 		    &guid) != 0 || guid != spa_guid(spa)) {
1332 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1333 			    VDEV_AUX_CORRUPT_DATA);
1334 			nvlist_free(label);
1335 			return (0);
1336 		}
1337 
1338 		if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl)
1339 		    != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID,
1340 		    &aux_guid) != 0)
1341 			aux_guid = 0;
1342 
1343 		/*
1344 		 * If this vdev just became a top-level vdev because its
1345 		 * sibling was detached, it will have adopted the parent's
1346 		 * vdev guid -- but the label may or may not be on disk yet.
1347 		 * Fortunately, either version of the label will have the
1348 		 * same top guid, so if we're a top-level vdev, we can
1349 		 * safely compare to that instead.
1350 		 *
1351 		 * If we split this vdev off instead, then we also check the
1352 		 * original pool's guid.  We don't want to consider the vdev
1353 		 * corrupt if it is partway through a split operation.
1354 		 */
1355 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID,
1356 		    &guid) != 0 ||
1357 		    nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID,
1358 		    &top_guid) != 0 ||
1359 		    ((vd->vdev_guid != guid && vd->vdev_guid != aux_guid) &&
1360 		    (vd->vdev_guid != top_guid || vd != vd->vdev_top))) {
1361 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1362 			    VDEV_AUX_CORRUPT_DATA);
1363 			nvlist_free(label);
1364 			return (0);
1365 		}
1366 
1367 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1368 		    &state) != 0) {
1369 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1370 			    VDEV_AUX_CORRUPT_DATA);
1371 			nvlist_free(label);
1372 			return (0);
1373 		}
1374 
1375 		nvlist_free(label);
1376 
1377 		/*
1378 		 * If this is a verbatim import, no need to check the
1379 		 * state of the pool.
1380 		 */
1381 		if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) &&
1382 		    spa_load_state(spa) == SPA_LOAD_OPEN &&
1383 		    state != POOL_STATE_ACTIVE)
1384 			return (EBADF);
1385 
1386 		/*
1387 		 * If we were able to open and validate a vdev that was
1388 		 * previously marked permanently unavailable, clear that state
1389 		 * now.
1390 		 */
1391 		if (vd->vdev_not_present)
1392 			vd->vdev_not_present = 0;
1393 	}
1394 
1395 	return (0);
1396 }
1397 
1398 /*
1399  * Close a virtual device.
1400  */
1401 void
1402 vdev_close(vdev_t *vd)
1403 {
1404 	spa_t *spa = vd->vdev_spa;
1405 	vdev_t *pvd = vd->vdev_parent;
1406 
1407 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1408 
1409 	/*
1410 	 * If our parent is reopening, then we are as well, unless we are
1411 	 * going offline.
1412 	 */
1413 	if (pvd != NULL && pvd->vdev_reopening)
1414 		vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline);
1415 
1416 	vd->vdev_ops->vdev_op_close(vd);
1417 
1418 	vdev_cache_purge(vd);
1419 
1420 	/*
1421 	 * We record the previous state before we close it, so that if we are
1422 	 * doing a reopen(), we don't generate FMA ereports if we notice that
1423 	 * it's still faulted.
1424 	 */
1425 	vd->vdev_prevstate = vd->vdev_state;
1426 
1427 	if (vd->vdev_offline)
1428 		vd->vdev_state = VDEV_STATE_OFFLINE;
1429 	else
1430 		vd->vdev_state = VDEV_STATE_CLOSED;
1431 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1432 }
1433 
1434 void
1435 vdev_hold(vdev_t *vd)
1436 {
1437 	spa_t *spa = vd->vdev_spa;
1438 
1439 	ASSERT(spa_is_root(spa));
1440 	if (spa->spa_state == POOL_STATE_UNINITIALIZED)
1441 		return;
1442 
1443 	for (int c = 0; c < vd->vdev_children; c++)
1444 		vdev_hold(vd->vdev_child[c]);
1445 
1446 	if (vd->vdev_ops->vdev_op_leaf)
1447 		vd->vdev_ops->vdev_op_hold(vd);
1448 }
1449 
1450 void
1451 vdev_rele(vdev_t *vd)
1452 {
1453 	spa_t *spa = vd->vdev_spa;
1454 
1455 	ASSERT(spa_is_root(spa));
1456 	for (int c = 0; c < vd->vdev_children; c++)
1457 		vdev_rele(vd->vdev_child[c]);
1458 
1459 	if (vd->vdev_ops->vdev_op_leaf)
1460 		vd->vdev_ops->vdev_op_rele(vd);
1461 }
1462 
1463 /*
1464  * Reopen all interior vdevs and any unopened leaves.  We don't actually
1465  * reopen leaf vdevs which had previously been opened as they might deadlock
1466  * on the spa_config_lock.  Instead we only obtain the leaf's physical size.
1467  * If the leaf has never been opened then open it, as usual.
1468  */
1469 void
1470 vdev_reopen(vdev_t *vd)
1471 {
1472 	spa_t *spa = vd->vdev_spa;
1473 
1474 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1475 
1476 	/* set the reopening flag unless we're taking the vdev offline */
1477 	vd->vdev_reopening = !vd->vdev_offline;
1478 	vdev_close(vd);
1479 	(void) vdev_open(vd);
1480 
1481 	/*
1482 	 * Call vdev_validate() here to make sure we have the same device.
1483 	 * Otherwise, a device with an invalid label could be successfully
1484 	 * opened in response to vdev_reopen().
1485 	 */
1486 	if (vd->vdev_aux) {
1487 		(void) vdev_validate_aux(vd);
1488 		if (vdev_readable(vd) && vdev_writeable(vd) &&
1489 		    vd->vdev_aux == &spa->spa_l2cache &&
1490 		    !l2arc_vdev_present(vd))
1491 			l2arc_add_vdev(spa, vd);
1492 	} else {
1493 		(void) vdev_validate(vd);
1494 	}
1495 
1496 	/*
1497 	 * Reassess parent vdev's health.
1498 	 */
1499 	vdev_propagate_state(vd);
1500 }
1501 
1502 int
1503 vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1504 {
1505 	int error;
1506 
1507 	/*
1508 	 * Normally, partial opens (e.g. of a mirror) are allowed.
1509 	 * For a create, however, we want to fail the request if
1510 	 * there are any components we can't open.
1511 	 */
1512 	error = vdev_open(vd);
1513 
1514 	if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
1515 		vdev_close(vd);
1516 		return (error ? error : ENXIO);
1517 	}
1518 
1519 	/*
1520 	 * Recursively initialize all labels.
1521 	 */
1522 	if ((error = vdev_label_init(vd, txg, isreplacing ?
1523 	    VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
1524 		vdev_close(vd);
1525 		return (error);
1526 	}
1527 
1528 	return (0);
1529 }
1530 
1531 void
1532 vdev_metaslab_set_size(vdev_t *vd)
1533 {
1534 	/*
1535 	 * Aim for roughly 200 metaslabs per vdev.
1536 	 */
1537 	vd->vdev_ms_shift = highbit(vd->vdev_asize / 200);
1538 	vd->vdev_ms_shift = MAX(vd->vdev_ms_shift, SPA_MAXBLOCKSHIFT);
1539 }
1540 
1541 void
1542 vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
1543 {
1544 	ASSERT(vd == vd->vdev_top);
1545 	ASSERT(!vd->vdev_ishole);
1546 	ASSERT(ISP2(flags));
1547 	ASSERT(spa_writeable(vd->vdev_spa));
1548 
1549 	if (flags & VDD_METASLAB)
1550 		(void) txg_list_add(&vd->vdev_ms_list, arg, txg);
1551 
1552 	if (flags & VDD_DTL)
1553 		(void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
1554 
1555 	(void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
1556 }
1557 
1558 /*
1559  * DTLs.
1560  *
1561  * A vdev's DTL (dirty time log) is the set of transaction groups for which
1562  * the vdev has less than perfect replication.  There are four kinds of DTL:
1563  *
1564  * DTL_MISSING: txgs for which the vdev has no valid copies of the data
1565  *
1566  * DTL_PARTIAL: txgs for which data is available, but not fully replicated
1567  *
1568  * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
1569  *	scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
1570  *	txgs that was scrubbed.
1571  *
1572  * DTL_OUTAGE: txgs which cannot currently be read, whether due to
1573  *	persistent errors or just some device being offline.
1574  *	Unlike the other three, the DTL_OUTAGE map is not generally
1575  *	maintained; it's only computed when needed, typically to
1576  *	determine whether a device can be detached.
1577  *
1578  * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
1579  * either has the data or it doesn't.
1580  *
1581  * For interior vdevs such as mirror and RAID-Z the picture is more complex.
1582  * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
1583  * if any child is less than fully replicated, then so is its parent.
1584  * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
1585  * comprising only those txgs which appear in 'maxfaults' or more children;
1586  * those are the txgs we don't have enough replication to read.  For example,
1587  * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
1588  * thus, its DTL_MISSING consists of the set of txgs that appear in more than
1589  * two child DTL_MISSING maps.
1590  *
1591  * It should be clear from the above that to compute the DTLs and outage maps
1592  * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
1593  * Therefore, that is all we keep on disk.  When loading the pool, or after
1594  * a configuration change, we generate all other DTLs from first principles.
1595  */
1596 void
1597 vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1598 {
1599 	space_map_t *sm = &vd->vdev_dtl[t];
1600 
1601 	ASSERT(t < DTL_TYPES);
1602 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1603 	ASSERT(spa_writeable(vd->vdev_spa));
1604 
1605 	mutex_enter(sm->sm_lock);
1606 	if (!space_map_contains(sm, txg, size))
1607 		space_map_add(sm, txg, size);
1608 	mutex_exit(sm->sm_lock);
1609 }
1610 
1611 boolean_t
1612 vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1613 {
1614 	space_map_t *sm = &vd->vdev_dtl[t];
1615 	boolean_t dirty = B_FALSE;
1616 
1617 	ASSERT(t < DTL_TYPES);
1618 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1619 
1620 	mutex_enter(sm->sm_lock);
1621 	if (sm->sm_space != 0)
1622 		dirty = space_map_contains(sm, txg, size);
1623 	mutex_exit(sm->sm_lock);
1624 
1625 	return (dirty);
1626 }
1627 
1628 boolean_t
1629 vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
1630 {
1631 	space_map_t *sm = &vd->vdev_dtl[t];
1632 	boolean_t empty;
1633 
1634 	mutex_enter(sm->sm_lock);
1635 	empty = (sm->sm_space == 0);
1636 	mutex_exit(sm->sm_lock);
1637 
1638 	return (empty);
1639 }
1640 
1641 /*
1642  * Reassess DTLs after a config change or scrub completion.
1643  */
1644 void
1645 vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
1646 {
1647 	spa_t *spa = vd->vdev_spa;
1648 	avl_tree_t reftree;
1649 	int minref;
1650 
1651 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
1652 
1653 	for (int c = 0; c < vd->vdev_children; c++)
1654 		vdev_dtl_reassess(vd->vdev_child[c], txg,
1655 		    scrub_txg, scrub_done);
1656 
1657 	if (vd == spa->spa_root_vdev || vd->vdev_ishole || vd->vdev_aux)
1658 		return;
1659 
1660 	if (vd->vdev_ops->vdev_op_leaf) {
1661 		dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1662 
1663 		mutex_enter(&vd->vdev_dtl_lock);
1664 		if (scrub_txg != 0 &&
1665 		    (spa->spa_scrub_started ||
1666 		    (scn && scn->scn_phys.scn_errors == 0))) {
1667 			/*
1668 			 * We completed a scrub up to scrub_txg.  If we
1669 			 * did it without rebooting, then the scrub dtl
1670 			 * will be valid, so excise the old region and
1671 			 * fold in the scrub dtl.  Otherwise, leave the
1672 			 * dtl as-is if there was an error.
1673 			 *
1674 			 * There's little trick here: to excise the beginning
1675 			 * of the DTL_MISSING map, we put it into a reference
1676 			 * tree and then add a segment with refcnt -1 that
1677 			 * covers the range [0, scrub_txg).  This means
1678 			 * that each txg in that range has refcnt -1 or 0.
1679 			 * We then add DTL_SCRUB with a refcnt of 2, so that
1680 			 * entries in the range [0, scrub_txg) will have a
1681 			 * positive refcnt -- either 1 or 2.  We then convert
1682 			 * the reference tree into the new DTL_MISSING map.
1683 			 */
1684 			space_map_ref_create(&reftree);
1685 			space_map_ref_add_map(&reftree,
1686 			    &vd->vdev_dtl[DTL_MISSING], 1);
1687 			space_map_ref_add_seg(&reftree, 0, scrub_txg, -1);
1688 			space_map_ref_add_map(&reftree,
1689 			    &vd->vdev_dtl[DTL_SCRUB], 2);
1690 			space_map_ref_generate_map(&reftree,
1691 			    &vd->vdev_dtl[DTL_MISSING], 1);
1692 			space_map_ref_destroy(&reftree);
1693 		}
1694 		space_map_vacate(&vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
1695 		space_map_walk(&vd->vdev_dtl[DTL_MISSING],
1696 		    space_map_add, &vd->vdev_dtl[DTL_PARTIAL]);
1697 		if (scrub_done)
1698 			space_map_vacate(&vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
1699 		space_map_vacate(&vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
1700 		if (!vdev_readable(vd))
1701 			space_map_add(&vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
1702 		else
1703 			space_map_walk(&vd->vdev_dtl[DTL_MISSING],
1704 			    space_map_add, &vd->vdev_dtl[DTL_OUTAGE]);
1705 		mutex_exit(&vd->vdev_dtl_lock);
1706 
1707 		if (txg != 0)
1708 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
1709 		return;
1710 	}
1711 
1712 	mutex_enter(&vd->vdev_dtl_lock);
1713 	for (int t = 0; t < DTL_TYPES; t++) {
1714 		/* account for child's outage in parent's missing map */
1715 		int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
1716 		if (t == DTL_SCRUB)
1717 			continue;			/* leaf vdevs only */
1718 		if (t == DTL_PARTIAL)
1719 			minref = 1;			/* i.e. non-zero */
1720 		else if (vd->vdev_nparity != 0)
1721 			minref = vd->vdev_nparity + 1;	/* RAID-Z */
1722 		else
1723 			minref = vd->vdev_children;	/* any kind of mirror */
1724 		space_map_ref_create(&reftree);
1725 		for (int c = 0; c < vd->vdev_children; c++) {
1726 			vdev_t *cvd = vd->vdev_child[c];
1727 			mutex_enter(&cvd->vdev_dtl_lock);
1728 			space_map_ref_add_map(&reftree, &cvd->vdev_dtl[s], 1);
1729 			mutex_exit(&cvd->vdev_dtl_lock);
1730 		}
1731 		space_map_ref_generate_map(&reftree, &vd->vdev_dtl[t], minref);
1732 		space_map_ref_destroy(&reftree);
1733 	}
1734 	mutex_exit(&vd->vdev_dtl_lock);
1735 }
1736 
1737 static int
1738 vdev_dtl_load(vdev_t *vd)
1739 {
1740 	spa_t *spa = vd->vdev_spa;
1741 	space_map_obj_t *smo = &vd->vdev_dtl_smo;
1742 	objset_t *mos = spa->spa_meta_objset;
1743 	dmu_buf_t *db;
1744 	int error;
1745 
1746 	ASSERT(vd->vdev_children == 0);
1747 
1748 	if (smo->smo_object == 0)
1749 		return (0);
1750 
1751 	ASSERT(!vd->vdev_ishole);
1752 
1753 	if ((error = dmu_bonus_hold(mos, smo->smo_object, FTAG, &db)) != 0)
1754 		return (error);
1755 
1756 	ASSERT3U(db->db_size, >=, sizeof (*smo));
1757 	bcopy(db->db_data, smo, sizeof (*smo));
1758 	dmu_buf_rele(db, FTAG);
1759 
1760 	mutex_enter(&vd->vdev_dtl_lock);
1761 	error = space_map_load(&vd->vdev_dtl[DTL_MISSING],
1762 	    NULL, SM_ALLOC, smo, mos);
1763 	mutex_exit(&vd->vdev_dtl_lock);
1764 
1765 	return (error);
1766 }
1767 
1768 void
1769 vdev_dtl_sync(vdev_t *vd, uint64_t txg)
1770 {
1771 	spa_t *spa = vd->vdev_spa;
1772 	space_map_obj_t *smo = &vd->vdev_dtl_smo;
1773 	space_map_t *sm = &vd->vdev_dtl[DTL_MISSING];
1774 	objset_t *mos = spa->spa_meta_objset;
1775 	space_map_t smsync;
1776 	kmutex_t smlock;
1777 	dmu_buf_t *db;
1778 	dmu_tx_t *tx;
1779 
1780 	ASSERT(!vd->vdev_ishole);
1781 
1782 	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1783 
1784 	if (vd->vdev_detached) {
1785 		if (smo->smo_object != 0) {
1786 			int err = dmu_object_free(mos, smo->smo_object, tx);
1787 			ASSERT3U(err, ==, 0);
1788 			smo->smo_object = 0;
1789 		}
1790 		dmu_tx_commit(tx);
1791 		return;
1792 	}
1793 
1794 	if (smo->smo_object == 0) {
1795 		ASSERT(smo->smo_objsize == 0);
1796 		ASSERT(smo->smo_alloc == 0);
1797 		smo->smo_object = dmu_object_alloc(mos,
1798 		    DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
1799 		    DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
1800 		ASSERT(smo->smo_object != 0);
1801 		vdev_config_dirty(vd->vdev_top);
1802 	}
1803 
1804 	mutex_init(&smlock, NULL, MUTEX_DEFAULT, NULL);
1805 
1806 	space_map_create(&smsync, sm->sm_start, sm->sm_size, sm->sm_shift,
1807 	    &smlock);
1808 
1809 	mutex_enter(&smlock);
1810 
1811 	mutex_enter(&vd->vdev_dtl_lock);
1812 	space_map_walk(sm, space_map_add, &smsync);
1813 	mutex_exit(&vd->vdev_dtl_lock);
1814 
1815 	space_map_truncate(smo, mos, tx);
1816 	space_map_sync(&smsync, SM_ALLOC, smo, mos, tx);
1817 
1818 	space_map_destroy(&smsync);
1819 
1820 	mutex_exit(&smlock);
1821 	mutex_destroy(&smlock);
1822 
1823 	VERIFY(0 == dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
1824 	dmu_buf_will_dirty(db, tx);
1825 	ASSERT3U(db->db_size, >=, sizeof (*smo));
1826 	bcopy(smo, db->db_data, sizeof (*smo));
1827 	dmu_buf_rele(db, FTAG);
1828 
1829 	dmu_tx_commit(tx);
1830 }
1831 
1832 /*
1833  * Determine whether the specified vdev can be offlined/detached/removed
1834  * without losing data.
1835  */
1836 boolean_t
1837 vdev_dtl_required(vdev_t *vd)
1838 {
1839 	spa_t *spa = vd->vdev_spa;
1840 	vdev_t *tvd = vd->vdev_top;
1841 	uint8_t cant_read = vd->vdev_cant_read;
1842 	boolean_t required;
1843 
1844 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1845 
1846 	if (vd == spa->spa_root_vdev || vd == tvd)
1847 		return (B_TRUE);
1848 
1849 	/*
1850 	 * Temporarily mark the device as unreadable, and then determine
1851 	 * whether this results in any DTL outages in the top-level vdev.
1852 	 * If not, we can safely offline/detach/remove the device.
1853 	 */
1854 	vd->vdev_cant_read = B_TRUE;
1855 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1856 	required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
1857 	vd->vdev_cant_read = cant_read;
1858 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1859 
1860 	if (!required && zio_injection_enabled)
1861 		required = !!zio_handle_device_injection(vd, NULL, ECHILD);
1862 
1863 	return (required);
1864 }
1865 
1866 /*
1867  * Determine if resilver is needed, and if so the txg range.
1868  */
1869 boolean_t
1870 vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
1871 {
1872 	boolean_t needed = B_FALSE;
1873 	uint64_t thismin = UINT64_MAX;
1874 	uint64_t thismax = 0;
1875 
1876 	if (vd->vdev_children == 0) {
1877 		mutex_enter(&vd->vdev_dtl_lock);
1878 		if (vd->vdev_dtl[DTL_MISSING].sm_space != 0 &&
1879 		    vdev_writeable(vd)) {
1880 			space_seg_t *ss;
1881 
1882 			ss = avl_first(&vd->vdev_dtl[DTL_MISSING].sm_root);
1883 			thismin = ss->ss_start - 1;
1884 			ss = avl_last(&vd->vdev_dtl[DTL_MISSING].sm_root);
1885 			thismax = ss->ss_end;
1886 			needed = B_TRUE;
1887 		}
1888 		mutex_exit(&vd->vdev_dtl_lock);
1889 	} else {
1890 		for (int c = 0; c < vd->vdev_children; c++) {
1891 			vdev_t *cvd = vd->vdev_child[c];
1892 			uint64_t cmin, cmax;
1893 
1894 			if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
1895 				thismin = MIN(thismin, cmin);
1896 				thismax = MAX(thismax, cmax);
1897 				needed = B_TRUE;
1898 			}
1899 		}
1900 	}
1901 
1902 	if (needed && minp) {
1903 		*minp = thismin;
1904 		*maxp = thismax;
1905 	}
1906 	return (needed);
1907 }
1908 
1909 void
1910 vdev_load(vdev_t *vd)
1911 {
1912 	/*
1913 	 * Recursively load all children.
1914 	 */
1915 	for (int c = 0; c < vd->vdev_children; c++)
1916 		vdev_load(vd->vdev_child[c]);
1917 
1918 	/*
1919 	 * If this is a top-level vdev, initialize its metaslabs.
1920 	 */
1921 	if (vd == vd->vdev_top && !vd->vdev_ishole &&
1922 	    (vd->vdev_ashift == 0 || vd->vdev_asize == 0 ||
1923 	    vdev_metaslab_init(vd, 0) != 0))
1924 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1925 		    VDEV_AUX_CORRUPT_DATA);
1926 
1927 	/*
1928 	 * If this is a leaf vdev, load its DTL.
1929 	 */
1930 	if (vd->vdev_ops->vdev_op_leaf && vdev_dtl_load(vd) != 0)
1931 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1932 		    VDEV_AUX_CORRUPT_DATA);
1933 }
1934 
1935 /*
1936  * The special vdev case is used for hot spares and l2cache devices.  Its
1937  * sole purpose it to set the vdev state for the associated vdev.  To do this,
1938  * we make sure that we can open the underlying device, then try to read the
1939  * label, and make sure that the label is sane and that it hasn't been
1940  * repurposed to another pool.
1941  */
1942 int
1943 vdev_validate_aux(vdev_t *vd)
1944 {
1945 	nvlist_t *label;
1946 	uint64_t guid, version;
1947 	uint64_t state;
1948 
1949 	if (!vdev_readable(vd))
1950 		return (0);
1951 
1952 	if ((label = vdev_label_read_config(vd)) == NULL) {
1953 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1954 		    VDEV_AUX_CORRUPT_DATA);
1955 		return (-1);
1956 	}
1957 
1958 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
1959 	    version > SPA_VERSION ||
1960 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
1961 	    guid != vd->vdev_guid ||
1962 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
1963 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1964 		    VDEV_AUX_CORRUPT_DATA);
1965 		nvlist_free(label);
1966 		return (-1);
1967 	}
1968 
1969 	/*
1970 	 * We don't actually check the pool state here.  If it's in fact in
1971 	 * use by another pool, we update this fact on the fly when requested.
1972 	 */
1973 	nvlist_free(label);
1974 	return (0);
1975 }
1976 
1977 void
1978 vdev_remove(vdev_t *vd, uint64_t txg)
1979 {
1980 	spa_t *spa = vd->vdev_spa;
1981 	objset_t *mos = spa->spa_meta_objset;
1982 	dmu_tx_t *tx;
1983 
1984 	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
1985 
1986 	if (vd->vdev_dtl_smo.smo_object) {
1987 		ASSERT3U(vd->vdev_dtl_smo.smo_alloc, ==, 0);
1988 		(void) dmu_object_free(mos, vd->vdev_dtl_smo.smo_object, tx);
1989 		vd->vdev_dtl_smo.smo_object = 0;
1990 	}
1991 
1992 	if (vd->vdev_ms != NULL) {
1993 		for (int m = 0; m < vd->vdev_ms_count; m++) {
1994 			metaslab_t *msp = vd->vdev_ms[m];
1995 
1996 			if (msp == NULL || msp->ms_smo.smo_object == 0)
1997 				continue;
1998 
1999 			ASSERT3U(msp->ms_smo.smo_alloc, ==, 0);
2000 			(void) dmu_object_free(mos, msp->ms_smo.smo_object, tx);
2001 			msp->ms_smo.smo_object = 0;
2002 		}
2003 	}
2004 
2005 	if (vd->vdev_ms_array) {
2006 		(void) dmu_object_free(mos, vd->vdev_ms_array, tx);
2007 		vd->vdev_ms_array = 0;
2008 		vd->vdev_ms_shift = 0;
2009 	}
2010 	dmu_tx_commit(tx);
2011 }
2012 
2013 void
2014 vdev_sync_done(vdev_t *vd, uint64_t txg)
2015 {
2016 	metaslab_t *msp;
2017 	boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg));
2018 
2019 	ASSERT(!vd->vdev_ishole);
2020 
2021 	while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
2022 		metaslab_sync_done(msp, txg);
2023 
2024 	if (reassess)
2025 		metaslab_sync_reassess(vd->vdev_mg);
2026 }
2027 
2028 void
2029 vdev_sync(vdev_t *vd, uint64_t txg)
2030 {
2031 	spa_t *spa = vd->vdev_spa;
2032 	vdev_t *lvd;
2033 	metaslab_t *msp;
2034 	dmu_tx_t *tx;
2035 
2036 	ASSERT(!vd->vdev_ishole);
2037 
2038 	if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0) {
2039 		ASSERT(vd == vd->vdev_top);
2040 		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2041 		vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
2042 		    DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
2043 		ASSERT(vd->vdev_ms_array != 0);
2044 		vdev_config_dirty(vd);
2045 		dmu_tx_commit(tx);
2046 	}
2047 
2048 	/*
2049 	 * Remove the metadata associated with this vdev once it's empty.
2050 	 */
2051 	if (vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing)
2052 		vdev_remove(vd, txg);
2053 
2054 	while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
2055 		metaslab_sync(msp, txg);
2056 		(void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
2057 	}
2058 
2059 	while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
2060 		vdev_dtl_sync(lvd, txg);
2061 
2062 	(void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
2063 }
2064 
2065 uint64_t
2066 vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
2067 {
2068 	return (vd->vdev_ops->vdev_op_asize(vd, psize));
2069 }
2070 
2071 /*
2072  * Mark the given vdev faulted.  A faulted vdev behaves as if the device could
2073  * not be opened, and no I/O is attempted.
2074  */
2075 int
2076 vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2077 {
2078 	vdev_t *vd, *tvd;
2079 
2080 	spa_vdev_state_enter(spa, SCL_NONE);
2081 
2082 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2083 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2084 
2085 	if (!vd->vdev_ops->vdev_op_leaf)
2086 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2087 
2088 	tvd = vd->vdev_top;
2089 
2090 	/*
2091 	 * We don't directly use the aux state here, but if we do a
2092 	 * vdev_reopen(), we need this value to be present to remember why we
2093 	 * were faulted.
2094 	 */
2095 	vd->vdev_label_aux = aux;
2096 
2097 	/*
2098 	 * Faulted state takes precedence over degraded.
2099 	 */
2100 	vd->vdev_delayed_close = B_FALSE;
2101 	vd->vdev_faulted = 1ULL;
2102 	vd->vdev_degraded = 0ULL;
2103 	vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
2104 
2105 	/*
2106 	 * If this device has the only valid copy of the data, then
2107 	 * back off and simply mark the vdev as degraded instead.
2108 	 */
2109 	if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) {
2110 		vd->vdev_degraded = 1ULL;
2111 		vd->vdev_faulted = 0ULL;
2112 
2113 		/*
2114 		 * If we reopen the device and it's not dead, only then do we
2115 		 * mark it degraded.
2116 		 */
2117 		vdev_reopen(tvd);
2118 
2119 		if (vdev_readable(vd))
2120 			vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
2121 	}
2122 
2123 	return (spa_vdev_state_exit(spa, vd, 0));
2124 }
2125 
2126 /*
2127  * Mark the given vdev degraded.  A degraded vdev is purely an indication to the
2128  * user that something is wrong.  The vdev continues to operate as normal as far
2129  * as I/O is concerned.
2130  */
2131 int
2132 vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2133 {
2134 	vdev_t *vd;
2135 
2136 	spa_vdev_state_enter(spa, SCL_NONE);
2137 
2138 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2139 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2140 
2141 	if (!vd->vdev_ops->vdev_op_leaf)
2142 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2143 
2144 	/*
2145 	 * If the vdev is already faulted, then don't do anything.
2146 	 */
2147 	if (vd->vdev_faulted || vd->vdev_degraded)
2148 		return (spa_vdev_state_exit(spa, NULL, 0));
2149 
2150 	vd->vdev_degraded = 1ULL;
2151 	if (!vdev_is_dead(vd))
2152 		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
2153 		    aux);
2154 
2155 	return (spa_vdev_state_exit(spa, vd, 0));
2156 }
2157 
2158 /*
2159  * Online the given vdev.  If 'unspare' is set, it implies two things.  First,
2160  * any attached spare device should be detached when the device finishes
2161  * resilvering.  Second, the online should be treated like a 'test' online case,
2162  * so no FMA events are generated if the device fails to open.
2163  */
2164 int
2165 vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
2166 {
2167 	vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
2168 
2169 	spa_vdev_state_enter(spa, SCL_NONE);
2170 
2171 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2172 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2173 
2174 	if (!vd->vdev_ops->vdev_op_leaf)
2175 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2176 
2177 	tvd = vd->vdev_top;
2178 	vd->vdev_offline = B_FALSE;
2179 	vd->vdev_tmpoffline = B_FALSE;
2180 	vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
2181 	vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
2182 
2183 	/* XXX - L2ARC 1.0 does not support expansion */
2184 	if (!vd->vdev_aux) {
2185 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2186 			pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
2187 	}
2188 
2189 	vdev_reopen(tvd);
2190 	vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
2191 
2192 	if (!vd->vdev_aux) {
2193 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2194 			pvd->vdev_expanding = B_FALSE;
2195 	}
2196 
2197 	if (newstate)
2198 		*newstate = vd->vdev_state;
2199 	if ((flags & ZFS_ONLINE_UNSPARE) &&
2200 	    !vdev_is_dead(vd) && vd->vdev_parent &&
2201 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2202 	    vd->vdev_parent->vdev_child[0] == vd)
2203 		vd->vdev_unspare = B_TRUE;
2204 
2205 	if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
2206 
2207 		/* XXX - L2ARC 1.0 does not support expansion */
2208 		if (vd->vdev_aux)
2209 			return (spa_vdev_state_exit(spa, vd, ENOTSUP));
2210 		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
2211 	}
2212 	return (spa_vdev_state_exit(spa, vd, 0));
2213 }
2214 
2215 static int
2216 vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
2217 {
2218 	vdev_t *vd, *tvd;
2219 	int error = 0;
2220 	uint64_t generation;
2221 	metaslab_group_t *mg;
2222 
2223 top:
2224 	spa_vdev_state_enter(spa, SCL_ALLOC);
2225 
2226 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2227 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2228 
2229 	if (!vd->vdev_ops->vdev_op_leaf)
2230 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2231 
2232 	tvd = vd->vdev_top;
2233 	mg = tvd->vdev_mg;
2234 	generation = spa->spa_config_generation + 1;
2235 
2236 	/*
2237 	 * If the device isn't already offline, try to offline it.
2238 	 */
2239 	if (!vd->vdev_offline) {
2240 		/*
2241 		 * If this device has the only valid copy of some data,
2242 		 * don't allow it to be offlined. Log devices are always
2243 		 * expendable.
2244 		 */
2245 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2246 		    vdev_dtl_required(vd))
2247 			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2248 
2249 		/*
2250 		 * If the top-level is a slog and it has had allocations
2251 		 * then proceed.  We check that the vdev's metaslab group
2252 		 * is not NULL since it's possible that we may have just
2253 		 * added this vdev but not yet initialized its metaslabs.
2254 		 */
2255 		if (tvd->vdev_islog && mg != NULL) {
2256 			/*
2257 			 * Prevent any future allocations.
2258 			 */
2259 			metaslab_group_passivate(mg);
2260 			(void) spa_vdev_state_exit(spa, vd, 0);
2261 
2262 			error = spa_offline_log(spa);
2263 
2264 			spa_vdev_state_enter(spa, SCL_ALLOC);
2265 
2266 			/*
2267 			 * Check to see if the config has changed.
2268 			 */
2269 			if (error || generation != spa->spa_config_generation) {
2270 				metaslab_group_activate(mg);
2271 				if (error)
2272 					return (spa_vdev_state_exit(spa,
2273 					    vd, error));
2274 				(void) spa_vdev_state_exit(spa, vd, 0);
2275 				goto top;
2276 			}
2277 			ASSERT3U(tvd->vdev_stat.vs_alloc, ==, 0);
2278 		}
2279 
2280 		/*
2281 		 * Offline this device and reopen its top-level vdev.
2282 		 * If the top-level vdev is a log device then just offline
2283 		 * it. Otherwise, if this action results in the top-level
2284 		 * vdev becoming unusable, undo it and fail the request.
2285 		 */
2286 		vd->vdev_offline = B_TRUE;
2287 		vdev_reopen(tvd);
2288 
2289 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2290 		    vdev_is_dead(tvd)) {
2291 			vd->vdev_offline = B_FALSE;
2292 			vdev_reopen(tvd);
2293 			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2294 		}
2295 
2296 		/*
2297 		 * Add the device back into the metaslab rotor so that
2298 		 * once we online the device it's open for business.
2299 		 */
2300 		if (tvd->vdev_islog && mg != NULL)
2301 			metaslab_group_activate(mg);
2302 	}
2303 
2304 	vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
2305 
2306 	return (spa_vdev_state_exit(spa, vd, 0));
2307 }
2308 
2309 int
2310 vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
2311 {
2312 	int error;
2313 
2314 	mutex_enter(&spa->spa_vdev_top_lock);
2315 	error = vdev_offline_locked(spa, guid, flags);
2316 	mutex_exit(&spa->spa_vdev_top_lock);
2317 
2318 	return (error);
2319 }
2320 
2321 /*
2322  * Clear the error counts associated with this vdev.  Unlike vdev_online() and
2323  * vdev_offline(), we assume the spa config is locked.  We also clear all
2324  * children.  If 'vd' is NULL, then the user wants to clear all vdevs.
2325  */
2326 void
2327 vdev_clear(spa_t *spa, vdev_t *vd)
2328 {
2329 	vdev_t *rvd = spa->spa_root_vdev;
2330 
2331 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2332 
2333 	if (vd == NULL)
2334 		vd = rvd;
2335 
2336 	vd->vdev_stat.vs_read_errors = 0;
2337 	vd->vdev_stat.vs_write_errors = 0;
2338 	vd->vdev_stat.vs_checksum_errors = 0;
2339 
2340 	for (int c = 0; c < vd->vdev_children; c++)
2341 		vdev_clear(spa, vd->vdev_child[c]);
2342 
2343 	/*
2344 	 * If we're in the FAULTED state or have experienced failed I/O, then
2345 	 * clear the persistent state and attempt to reopen the device.  We
2346 	 * also mark the vdev config dirty, so that the new faulted state is
2347 	 * written out to disk.
2348 	 */
2349 	if (vd->vdev_faulted || vd->vdev_degraded ||
2350 	    !vdev_readable(vd) || !vdev_writeable(vd)) {
2351 
2352 		/*
2353 		 * When reopening in reponse to a clear event, it may be due to
2354 		 * a fmadm repair request.  In this case, if the device is
2355 		 * still broken, we want to still post the ereport again.
2356 		 */
2357 		vd->vdev_forcefault = B_TRUE;
2358 
2359 		vd->vdev_faulted = vd->vdev_degraded = 0ULL;
2360 		vd->vdev_cant_read = B_FALSE;
2361 		vd->vdev_cant_write = B_FALSE;
2362 
2363 		vdev_reopen(vd == rvd ? rvd : vd->vdev_top);
2364 
2365 		vd->vdev_forcefault = B_FALSE;
2366 
2367 		if (vd != rvd && vdev_writeable(vd->vdev_top))
2368 			vdev_state_dirty(vd->vdev_top);
2369 
2370 		if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
2371 			spa_async_request(spa, SPA_ASYNC_RESILVER);
2372 
2373 		spa_event_notify(spa, vd, ESC_ZFS_VDEV_CLEAR);
2374 	}
2375 
2376 	/*
2377 	 * When clearing a FMA-diagnosed fault, we always want to
2378 	 * unspare the device, as we assume that the original spare was
2379 	 * done in response to the FMA fault.
2380 	 */
2381 	if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
2382 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2383 	    vd->vdev_parent->vdev_child[0] == vd)
2384 		vd->vdev_unspare = B_TRUE;
2385 }
2386 
2387 boolean_t
2388 vdev_is_dead(vdev_t *vd)
2389 {
2390 	/*
2391 	 * Holes and missing devices are always considered "dead".
2392 	 * This simplifies the code since we don't have to check for
2393 	 * these types of devices in the various code paths.
2394 	 * Instead we rely on the fact that we skip over dead devices
2395 	 * before issuing I/O to them.
2396 	 */
2397 	return (vd->vdev_state < VDEV_STATE_DEGRADED || vd->vdev_ishole ||
2398 	    vd->vdev_ops == &vdev_missing_ops);
2399 }
2400 
2401 boolean_t
2402 vdev_readable(vdev_t *vd)
2403 {
2404 	return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
2405 }
2406 
2407 boolean_t
2408 vdev_writeable(vdev_t *vd)
2409 {
2410 	return (!vdev_is_dead(vd) && !vd->vdev_cant_write);
2411 }
2412 
2413 boolean_t
2414 vdev_allocatable(vdev_t *vd)
2415 {
2416 	uint64_t state = vd->vdev_state;
2417 
2418 	/*
2419 	 * We currently allow allocations from vdevs which may be in the
2420 	 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
2421 	 * fails to reopen then we'll catch it later when we're holding
2422 	 * the proper locks.  Note that we have to get the vdev state
2423 	 * in a local variable because although it changes atomically,
2424 	 * we're asking two separate questions about it.
2425 	 */
2426 	return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
2427 	    !vd->vdev_cant_write && !vd->vdev_ishole);
2428 }
2429 
2430 boolean_t
2431 vdev_accessible(vdev_t *vd, zio_t *zio)
2432 {
2433 	ASSERT(zio->io_vd == vd);
2434 
2435 	if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
2436 		return (B_FALSE);
2437 
2438 	if (zio->io_type == ZIO_TYPE_READ)
2439 		return (!vd->vdev_cant_read);
2440 
2441 	if (zio->io_type == ZIO_TYPE_WRITE)
2442 		return (!vd->vdev_cant_write);
2443 
2444 	return (B_TRUE);
2445 }
2446 
2447 /*
2448  * Get statistics for the given vdev.
2449  */
2450 void
2451 vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
2452 {
2453 	vdev_t *rvd = vd->vdev_spa->spa_root_vdev;
2454 
2455 	mutex_enter(&vd->vdev_stat_lock);
2456 	bcopy(&vd->vdev_stat, vs, sizeof (*vs));
2457 	vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
2458 	vs->vs_state = vd->vdev_state;
2459 	vs->vs_rsize = vdev_get_min_asize(vd);
2460 	if (vd->vdev_ops->vdev_op_leaf)
2461 		vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
2462 	mutex_exit(&vd->vdev_stat_lock);
2463 
2464 	/*
2465 	 * If we're getting stats on the root vdev, aggregate the I/O counts
2466 	 * over all top-level vdevs (i.e. the direct children of the root).
2467 	 */
2468 	if (vd == rvd) {
2469 		for (int c = 0; c < rvd->vdev_children; c++) {
2470 			vdev_t *cvd = rvd->vdev_child[c];
2471 			vdev_stat_t *cvs = &cvd->vdev_stat;
2472 
2473 			mutex_enter(&vd->vdev_stat_lock);
2474 			for (int t = 0; t < ZIO_TYPES; t++) {
2475 				vs->vs_ops[t] += cvs->vs_ops[t];
2476 				vs->vs_bytes[t] += cvs->vs_bytes[t];
2477 			}
2478 			cvs->vs_scan_removing = cvd->vdev_removing;
2479 			mutex_exit(&vd->vdev_stat_lock);
2480 		}
2481 	}
2482 }
2483 
2484 void
2485 vdev_clear_stats(vdev_t *vd)
2486 {
2487 	mutex_enter(&vd->vdev_stat_lock);
2488 	vd->vdev_stat.vs_space = 0;
2489 	vd->vdev_stat.vs_dspace = 0;
2490 	vd->vdev_stat.vs_alloc = 0;
2491 	mutex_exit(&vd->vdev_stat_lock);
2492 }
2493 
2494 void
2495 vdev_scan_stat_init(vdev_t *vd)
2496 {
2497 	vdev_stat_t *vs = &vd->vdev_stat;
2498 
2499 	for (int c = 0; c < vd->vdev_children; c++)
2500 		vdev_scan_stat_init(vd->vdev_child[c]);
2501 
2502 	mutex_enter(&vd->vdev_stat_lock);
2503 	vs->vs_scan_processed = 0;
2504 	mutex_exit(&vd->vdev_stat_lock);
2505 }
2506 
2507 void
2508 vdev_stat_update(zio_t *zio, uint64_t psize)
2509 {
2510 	spa_t *spa = zio->io_spa;
2511 	vdev_t *rvd = spa->spa_root_vdev;
2512 	vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
2513 	vdev_t *pvd;
2514 	uint64_t txg = zio->io_txg;
2515 	vdev_stat_t *vs = &vd->vdev_stat;
2516 	zio_type_t type = zio->io_type;
2517 	int flags = zio->io_flags;
2518 
2519 	/*
2520 	 * If this i/o is a gang leader, it didn't do any actual work.
2521 	 */
2522 	if (zio->io_gang_tree)
2523 		return;
2524 
2525 	if (zio->io_error == 0) {
2526 		/*
2527 		 * If this is a root i/o, don't count it -- we've already
2528 		 * counted the top-level vdevs, and vdev_get_stats() will
2529 		 * aggregate them when asked.  This reduces contention on
2530 		 * the root vdev_stat_lock and implicitly handles blocks
2531 		 * that compress away to holes, for which there is no i/o.
2532 		 * (Holes never create vdev children, so all the counters
2533 		 * remain zero, which is what we want.)
2534 		 *
2535 		 * Note: this only applies to successful i/o (io_error == 0)
2536 		 * because unlike i/o counts, errors are not additive.
2537 		 * When reading a ditto block, for example, failure of
2538 		 * one top-level vdev does not imply a root-level error.
2539 		 */
2540 		if (vd == rvd)
2541 			return;
2542 
2543 		ASSERT(vd == zio->io_vd);
2544 
2545 		if (flags & ZIO_FLAG_IO_BYPASS)
2546 			return;
2547 
2548 		mutex_enter(&vd->vdev_stat_lock);
2549 
2550 		if (flags & ZIO_FLAG_IO_REPAIR) {
2551 			if (flags & ZIO_FLAG_SCAN_THREAD) {
2552 				dsl_scan_phys_t *scn_phys =
2553 				    &spa->spa_dsl_pool->dp_scan->scn_phys;
2554 				uint64_t *processed = &scn_phys->scn_processed;
2555 
2556 				/* XXX cleanup? */
2557 				if (vd->vdev_ops->vdev_op_leaf)
2558 					atomic_add_64(processed, psize);
2559 				vs->vs_scan_processed += psize;
2560 			}
2561 
2562 			if (flags & ZIO_FLAG_SELF_HEAL)
2563 				vs->vs_self_healed += psize;
2564 		}
2565 
2566 		vs->vs_ops[type]++;
2567 		vs->vs_bytes[type] += psize;
2568 
2569 		mutex_exit(&vd->vdev_stat_lock);
2570 		return;
2571 	}
2572 
2573 	if (flags & ZIO_FLAG_SPECULATIVE)
2574 		return;
2575 
2576 	/*
2577 	 * If this is an I/O error that is going to be retried, then ignore the
2578 	 * error.  Otherwise, the user may interpret B_FAILFAST I/O errors as
2579 	 * hard errors, when in reality they can happen for any number of
2580 	 * innocuous reasons (bus resets, MPxIO link failure, etc).
2581 	 */
2582 	if (zio->io_error == EIO &&
2583 	    !(zio->io_flags & ZIO_FLAG_IO_RETRY))
2584 		return;
2585 
2586 	/*
2587 	 * Intent logs writes won't propagate their error to the root
2588 	 * I/O so don't mark these types of failures as pool-level
2589 	 * errors.
2590 	 */
2591 	if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
2592 		return;
2593 
2594 	mutex_enter(&vd->vdev_stat_lock);
2595 	if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
2596 		if (zio->io_error == ECKSUM)
2597 			vs->vs_checksum_errors++;
2598 		else
2599 			vs->vs_read_errors++;
2600 	}
2601 	if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
2602 		vs->vs_write_errors++;
2603 	mutex_exit(&vd->vdev_stat_lock);
2604 
2605 	if (type == ZIO_TYPE_WRITE && txg != 0 &&
2606 	    (!(flags & ZIO_FLAG_IO_REPAIR) ||
2607 	    (flags & ZIO_FLAG_SCAN_THREAD) ||
2608 	    spa->spa_claiming)) {
2609 		/*
2610 		 * This is either a normal write (not a repair), or it's
2611 		 * a repair induced by the scrub thread, or it's a repair
2612 		 * made by zil_claim() during spa_load() in the first txg.
2613 		 * In the normal case, we commit the DTL change in the same
2614 		 * txg as the block was born.  In the scrub-induced repair
2615 		 * case, we know that scrubs run in first-pass syncing context,
2616 		 * so we commit the DTL change in spa_syncing_txg(spa).
2617 		 * In the zil_claim() case, we commit in spa_first_txg(spa).
2618 		 *
2619 		 * We currently do not make DTL entries for failed spontaneous
2620 		 * self-healing writes triggered by normal (non-scrubbing)
2621 		 * reads, because we have no transactional context in which to
2622 		 * do so -- and it's not clear that it'd be desirable anyway.
2623 		 */
2624 		if (vd->vdev_ops->vdev_op_leaf) {
2625 			uint64_t commit_txg = txg;
2626 			if (flags & ZIO_FLAG_SCAN_THREAD) {
2627 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2628 				ASSERT(spa_sync_pass(spa) == 1);
2629 				vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
2630 				commit_txg = spa_syncing_txg(spa);
2631 			} else if (spa->spa_claiming) {
2632 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2633 				commit_txg = spa_first_txg(spa);
2634 			}
2635 			ASSERT(commit_txg >= spa_syncing_txg(spa));
2636 			if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
2637 				return;
2638 			for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2639 				vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
2640 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
2641 		}
2642 		if (vd != rvd)
2643 			vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
2644 	}
2645 }
2646 
2647 /*
2648  * Update the in-core space usage stats for this vdev, its metaslab class,
2649  * and the root vdev.
2650  */
2651 void
2652 vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
2653     int64_t space_delta)
2654 {
2655 	int64_t dspace_delta = space_delta;
2656 	spa_t *spa = vd->vdev_spa;
2657 	vdev_t *rvd = spa->spa_root_vdev;
2658 	metaslab_group_t *mg = vd->vdev_mg;
2659 	metaslab_class_t *mc = mg ? mg->mg_class : NULL;
2660 
2661 	ASSERT(vd == vd->vdev_top);
2662 
2663 	/*
2664 	 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
2665 	 * factor.  We must calculate this here and not at the root vdev
2666 	 * because the root vdev's psize-to-asize is simply the max of its
2667 	 * childrens', thus not accurate enough for us.
2668 	 */
2669 	ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
2670 	ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
2671 	dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
2672 	    vd->vdev_deflate_ratio;
2673 
2674 	mutex_enter(&vd->vdev_stat_lock);
2675 	vd->vdev_stat.vs_alloc += alloc_delta;
2676 	vd->vdev_stat.vs_space += space_delta;
2677 	vd->vdev_stat.vs_dspace += dspace_delta;
2678 	mutex_exit(&vd->vdev_stat_lock);
2679 
2680 	if (mc == spa_normal_class(spa)) {
2681 		mutex_enter(&rvd->vdev_stat_lock);
2682 		rvd->vdev_stat.vs_alloc += alloc_delta;
2683 		rvd->vdev_stat.vs_space += space_delta;
2684 		rvd->vdev_stat.vs_dspace += dspace_delta;
2685 		mutex_exit(&rvd->vdev_stat_lock);
2686 	}
2687 
2688 	if (mc != NULL) {
2689 		ASSERT(rvd == vd->vdev_parent);
2690 		ASSERT(vd->vdev_ms_count != 0);
2691 
2692 		metaslab_class_space_update(mc,
2693 		    alloc_delta, defer_delta, space_delta, dspace_delta);
2694 	}
2695 }
2696 
2697 /*
2698  * Mark a top-level vdev's config as dirty, placing it on the dirty list
2699  * so that it will be written out next time the vdev configuration is synced.
2700  * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
2701  */
2702 void
2703 vdev_config_dirty(vdev_t *vd)
2704 {
2705 	spa_t *spa = vd->vdev_spa;
2706 	vdev_t *rvd = spa->spa_root_vdev;
2707 	int c;
2708 
2709 	ASSERT(spa_writeable(spa));
2710 
2711 	/*
2712 	 * If this is an aux vdev (as with l2cache and spare devices), then we
2713 	 * update the vdev config manually and set the sync flag.
2714 	 */
2715 	if (vd->vdev_aux != NULL) {
2716 		spa_aux_vdev_t *sav = vd->vdev_aux;
2717 		nvlist_t **aux;
2718 		uint_t naux;
2719 
2720 		for (c = 0; c < sav->sav_count; c++) {
2721 			if (sav->sav_vdevs[c] == vd)
2722 				break;
2723 		}
2724 
2725 		if (c == sav->sav_count) {
2726 			/*
2727 			 * We're being removed.  There's nothing more to do.
2728 			 */
2729 			ASSERT(sav->sav_sync == B_TRUE);
2730 			return;
2731 		}
2732 
2733 		sav->sav_sync = B_TRUE;
2734 
2735 		if (nvlist_lookup_nvlist_array(sav->sav_config,
2736 		    ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
2737 			VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
2738 			    ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
2739 		}
2740 
2741 		ASSERT(c < naux);
2742 
2743 		/*
2744 		 * Setting the nvlist in the middle if the array is a little
2745 		 * sketchy, but it will work.
2746 		 */
2747 		nvlist_free(aux[c]);
2748 		aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0);
2749 
2750 		return;
2751 	}
2752 
2753 	/*
2754 	 * The dirty list is protected by the SCL_CONFIG lock.  The caller
2755 	 * must either hold SCL_CONFIG as writer, or must be the sync thread
2756 	 * (which holds SCL_CONFIG as reader).  There's only one sync thread,
2757 	 * so this is sufficient to ensure mutual exclusion.
2758 	 */
2759 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2760 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2761 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
2762 
2763 	if (vd == rvd) {
2764 		for (c = 0; c < rvd->vdev_children; c++)
2765 			vdev_config_dirty(rvd->vdev_child[c]);
2766 	} else {
2767 		ASSERT(vd == vd->vdev_top);
2768 
2769 		if (!list_link_active(&vd->vdev_config_dirty_node) &&
2770 		    !vd->vdev_ishole)
2771 			list_insert_head(&spa->spa_config_dirty_list, vd);
2772 	}
2773 }
2774 
2775 void
2776 vdev_config_clean(vdev_t *vd)
2777 {
2778 	spa_t *spa = vd->vdev_spa;
2779 
2780 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2781 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2782 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
2783 
2784 	ASSERT(list_link_active(&vd->vdev_config_dirty_node));
2785 	list_remove(&spa->spa_config_dirty_list, vd);
2786 }
2787 
2788 /*
2789  * Mark a top-level vdev's state as dirty, so that the next pass of
2790  * spa_sync() can convert this into vdev_config_dirty().  We distinguish
2791  * the state changes from larger config changes because they require
2792  * much less locking, and are often needed for administrative actions.
2793  */
2794 void
2795 vdev_state_dirty(vdev_t *vd)
2796 {
2797 	spa_t *spa = vd->vdev_spa;
2798 
2799 	ASSERT(spa_writeable(spa));
2800 	ASSERT(vd == vd->vdev_top);
2801 
2802 	/*
2803 	 * The state list is protected by the SCL_STATE lock.  The caller
2804 	 * must either hold SCL_STATE as writer, or must be the sync thread
2805 	 * (which holds SCL_STATE as reader).  There's only one sync thread,
2806 	 * so this is sufficient to ensure mutual exclusion.
2807 	 */
2808 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2809 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2810 	    spa_config_held(spa, SCL_STATE, RW_READER)));
2811 
2812 	if (!list_link_active(&vd->vdev_state_dirty_node) && !vd->vdev_ishole)
2813 		list_insert_head(&spa->spa_state_dirty_list, vd);
2814 }
2815 
2816 void
2817 vdev_state_clean(vdev_t *vd)
2818 {
2819 	spa_t *spa = vd->vdev_spa;
2820 
2821 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2822 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2823 	    spa_config_held(spa, SCL_STATE, RW_READER)));
2824 
2825 	ASSERT(list_link_active(&vd->vdev_state_dirty_node));
2826 	list_remove(&spa->spa_state_dirty_list, vd);
2827 }
2828 
2829 /*
2830  * Propagate vdev state up from children to parent.
2831  */
2832 void
2833 vdev_propagate_state(vdev_t *vd)
2834 {
2835 	spa_t *spa = vd->vdev_spa;
2836 	vdev_t *rvd = spa->spa_root_vdev;
2837 	int degraded = 0, faulted = 0;
2838 	int corrupted = 0;
2839 	vdev_t *child;
2840 
2841 	if (vd->vdev_children > 0) {
2842 		for (int c = 0; c < vd->vdev_children; c++) {
2843 			child = vd->vdev_child[c];
2844 
2845 			/*
2846 			 * Don't factor holes into the decision.
2847 			 */
2848 			if (child->vdev_ishole)
2849 				continue;
2850 
2851 			if (!vdev_readable(child) ||
2852 			    (!vdev_writeable(child) && spa_writeable(spa))) {
2853 				/*
2854 				 * Root special: if there is a top-level log
2855 				 * device, treat the root vdev as if it were
2856 				 * degraded.
2857 				 */
2858 				if (child->vdev_islog && vd == rvd)
2859 					degraded++;
2860 				else
2861 					faulted++;
2862 			} else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
2863 				degraded++;
2864 			}
2865 
2866 			if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
2867 				corrupted++;
2868 		}
2869 
2870 		vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
2871 
2872 		/*
2873 		 * Root special: if there is a top-level vdev that cannot be
2874 		 * opened due to corrupted metadata, then propagate the root
2875 		 * vdev's aux state as 'corrupt' rather than 'insufficient
2876 		 * replicas'.
2877 		 */
2878 		if (corrupted && vd == rvd &&
2879 		    rvd->vdev_state == VDEV_STATE_CANT_OPEN)
2880 			vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
2881 			    VDEV_AUX_CORRUPT_DATA);
2882 	}
2883 
2884 	if (vd->vdev_parent)
2885 		vdev_propagate_state(vd->vdev_parent);
2886 }
2887 
2888 /*
2889  * Set a vdev's state.  If this is during an open, we don't update the parent
2890  * state, because we're in the process of opening children depth-first.
2891  * Otherwise, we propagate the change to the parent.
2892  *
2893  * If this routine places a device in a faulted state, an appropriate ereport is
2894  * generated.
2895  */
2896 void
2897 vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
2898 {
2899 	uint64_t save_state;
2900 	spa_t *spa = vd->vdev_spa;
2901 
2902 	if (state == vd->vdev_state) {
2903 		vd->vdev_stat.vs_aux = aux;
2904 		return;
2905 	}
2906 
2907 	save_state = vd->vdev_state;
2908 
2909 	vd->vdev_state = state;
2910 	vd->vdev_stat.vs_aux = aux;
2911 
2912 	/*
2913 	 * If we are setting the vdev state to anything but an open state, then
2914 	 * always close the underlying device unless the device has requested
2915 	 * a delayed close (i.e. we're about to remove or fault the device).
2916 	 * Otherwise, we keep accessible but invalid devices open forever.
2917 	 * We don't call vdev_close() itself, because that implies some extra
2918 	 * checks (offline, etc) that we don't want here.  This is limited to
2919 	 * leaf devices, because otherwise closing the device will affect other
2920 	 * children.
2921 	 */
2922 	if (!vd->vdev_delayed_close && vdev_is_dead(vd) &&
2923 	    vd->vdev_ops->vdev_op_leaf)
2924 		vd->vdev_ops->vdev_op_close(vd);
2925 
2926 	/*
2927 	 * If we have brought this vdev back into service, we need
2928 	 * to notify fmd so that it can gracefully repair any outstanding
2929 	 * cases due to a missing device.  We do this in all cases, even those
2930 	 * that probably don't correlate to a repaired fault.  This is sure to
2931 	 * catch all cases, and we let the zfs-retire agent sort it out.  If
2932 	 * this is a transient state it's OK, as the retire agent will
2933 	 * double-check the state of the vdev before repairing it.
2934 	 */
2935 	if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf &&
2936 	    vd->vdev_prevstate != state)
2937 		zfs_post_state_change(spa, vd);
2938 
2939 	if (vd->vdev_removed &&
2940 	    state == VDEV_STATE_CANT_OPEN &&
2941 	    (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
2942 		/*
2943 		 * If the previous state is set to VDEV_STATE_REMOVED, then this
2944 		 * device was previously marked removed and someone attempted to
2945 		 * reopen it.  If this failed due to a nonexistent device, then
2946 		 * keep the device in the REMOVED state.  We also let this be if
2947 		 * it is one of our special test online cases, which is only
2948 		 * attempting to online the device and shouldn't generate an FMA
2949 		 * fault.
2950 		 */
2951 		vd->vdev_state = VDEV_STATE_REMOVED;
2952 		vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
2953 	} else if (state == VDEV_STATE_REMOVED) {
2954 		vd->vdev_removed = B_TRUE;
2955 	} else if (state == VDEV_STATE_CANT_OPEN) {
2956 		/*
2957 		 * If we fail to open a vdev during an import or recovery, we
2958 		 * mark it as "not available", which signifies that it was
2959 		 * never there to begin with.  Failure to open such a device
2960 		 * is not considered an error.
2961 		 */
2962 		if ((spa_load_state(spa) == SPA_LOAD_IMPORT ||
2963 		    spa_load_state(spa) == SPA_LOAD_RECOVER) &&
2964 		    vd->vdev_ops->vdev_op_leaf)
2965 			vd->vdev_not_present = 1;
2966 
2967 		/*
2968 		 * Post the appropriate ereport.  If the 'prevstate' field is
2969 		 * set to something other than VDEV_STATE_UNKNOWN, it indicates
2970 		 * that this is part of a vdev_reopen().  In this case, we don't
2971 		 * want to post the ereport if the device was already in the
2972 		 * CANT_OPEN state beforehand.
2973 		 *
2974 		 * If the 'checkremove' flag is set, then this is an attempt to
2975 		 * online the device in response to an insertion event.  If we
2976 		 * hit this case, then we have detected an insertion event for a
2977 		 * faulted or offline device that wasn't in the removed state.
2978 		 * In this scenario, we don't post an ereport because we are
2979 		 * about to replace the device, or attempt an online with
2980 		 * vdev_forcefault, which will generate the fault for us.
2981 		 */
2982 		if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
2983 		    !vd->vdev_not_present && !vd->vdev_checkremove &&
2984 		    vd != spa->spa_root_vdev) {
2985 			const char *class;
2986 
2987 			switch (aux) {
2988 			case VDEV_AUX_OPEN_FAILED:
2989 				class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
2990 				break;
2991 			case VDEV_AUX_CORRUPT_DATA:
2992 				class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
2993 				break;
2994 			case VDEV_AUX_NO_REPLICAS:
2995 				class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
2996 				break;
2997 			case VDEV_AUX_BAD_GUID_SUM:
2998 				class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
2999 				break;
3000 			case VDEV_AUX_TOO_SMALL:
3001 				class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
3002 				break;
3003 			case VDEV_AUX_BAD_LABEL:
3004 				class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
3005 				break;
3006 			default:
3007 				class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
3008 			}
3009 
3010 			zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
3011 		}
3012 
3013 		/* Erase any notion of persistent removed state */
3014 		vd->vdev_removed = B_FALSE;
3015 	} else {
3016 		vd->vdev_removed = B_FALSE;
3017 	}
3018 
3019 	if (!isopen && vd->vdev_parent)
3020 		vdev_propagate_state(vd->vdev_parent);
3021 }
3022 
3023 /*
3024  * Check the vdev configuration to ensure that it's capable of supporting
3025  * a root pool. Currently, we do not support RAID-Z or partial configuration.
3026  * In addition, only a single top-level vdev is allowed and none of the leaves
3027  * can be wholedisks.
3028  */
3029 boolean_t
3030 vdev_is_bootable(vdev_t *vd)
3031 {
3032 	if (!vd->vdev_ops->vdev_op_leaf) {
3033 		char *vdev_type = vd->vdev_ops->vdev_op_type;
3034 
3035 		if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
3036 		    vd->vdev_children > 1) {
3037 			return (B_FALSE);
3038 		} else if (strcmp(vdev_type, VDEV_TYPE_RAIDZ) == 0 ||
3039 		    strcmp(vdev_type, VDEV_TYPE_MISSING) == 0) {
3040 			return (B_FALSE);
3041 		}
3042 	} else if (vd->vdev_wholedisk == 1) {
3043 		return (B_FALSE);
3044 	}
3045 
3046 	for (int c = 0; c < vd->vdev_children; c++) {
3047 		if (!vdev_is_bootable(vd->vdev_child[c]))
3048 			return (B_FALSE);
3049 	}
3050 	return (B_TRUE);
3051 }
3052 
3053 /*
3054  * Load the state from the original vdev tree (ovd) which
3055  * we've retrieved from the MOS config object. If the original
3056  * vdev was offline or faulted then we transfer that state to the
3057  * device in the current vdev tree (nvd).
3058  */
3059 void
3060 vdev_load_log_state(vdev_t *nvd, vdev_t *ovd)
3061 {
3062 	spa_t *spa = nvd->vdev_spa;
3063 
3064 	ASSERT(nvd->vdev_top->vdev_islog);
3065 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3066 	ASSERT3U(nvd->vdev_guid, ==, ovd->vdev_guid);
3067 
3068 	for (int c = 0; c < nvd->vdev_children; c++)
3069 		vdev_load_log_state(nvd->vdev_child[c], ovd->vdev_child[c]);
3070 
3071 	if (nvd->vdev_ops->vdev_op_leaf) {
3072 		/*
3073 		 * Restore the persistent vdev state
3074 		 */
3075 		nvd->vdev_offline = ovd->vdev_offline;
3076 		nvd->vdev_faulted = ovd->vdev_faulted;
3077 		nvd->vdev_degraded = ovd->vdev_degraded;
3078 		nvd->vdev_removed = ovd->vdev_removed;
3079 	}
3080 }
3081 
3082 /*
3083  * Determine if a log device has valid content.  If the vdev was
3084  * removed or faulted in the MOS config then we know that
3085  * the content on the log device has already been written to the pool.
3086  */
3087 boolean_t
3088 vdev_log_state_valid(vdev_t *vd)
3089 {
3090 	if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted &&
3091 	    !vd->vdev_removed)
3092 		return (B_TRUE);
3093 
3094 	for (int c = 0; c < vd->vdev_children; c++)
3095 		if (vdev_log_state_valid(vd->vdev_child[c]))
3096 			return (B_TRUE);
3097 
3098 	return (B_FALSE);
3099 }
3100 
3101 /*
3102  * Expand a vdev if possible.
3103  */
3104 void
3105 vdev_expand(vdev_t *vd, uint64_t txg)
3106 {
3107 	ASSERT(vd->vdev_top == vd);
3108 	ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
3109 
3110 	if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count) {
3111 		VERIFY(vdev_metaslab_init(vd, txg) == 0);
3112 		vdev_config_dirty(vd);
3113 	}
3114 }
3115 
3116 /*
3117  * Split a vdev.
3118  */
3119 void
3120 vdev_split(vdev_t *vd)
3121 {
3122 	vdev_t *cvd, *pvd = vd->vdev_parent;
3123 
3124 	vdev_remove_child(pvd, vd);
3125 	vdev_compact_children(pvd);
3126 
3127 	cvd = pvd->vdev_child[0];
3128 	if (pvd->vdev_children == 1) {
3129 		vdev_remove_parent(cvd);
3130 		cvd->vdev_splitting = B_TRUE;
3131 	}
3132 	vdev_propagate_state(cvd);
3133 }
3134