xref: /illumos-gate/usr/src/uts/common/fs/zfs/vdev.c (revision d616ad8e)
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 2009 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #include <sys/zfs_context.h>
28 #include <sys/fm/fs/zfs.h>
29 #include <sys/spa.h>
30 #include <sys/spa_impl.h>
31 #include <sys/dmu.h>
32 #include <sys/dmu_tx.h>
33 #include <sys/vdev_impl.h>
34 #include <sys/uberblock_impl.h>
35 #include <sys/metaslab.h>
36 #include <sys/metaslab_impl.h>
37 #include <sys/space_map.h>
38 #include <sys/zio.h>
39 #include <sys/zap.h>
40 #include <sys/fs/zfs.h>
41 #include <sys/arc.h>
42 #include <sys/zil.h>
43 
44 /*
45  * Virtual device management.
46  */
47 
48 static vdev_ops_t *vdev_ops_table[] = {
49 	&vdev_root_ops,
50 	&vdev_raidz_ops,
51 	&vdev_mirror_ops,
52 	&vdev_replacing_ops,
53 	&vdev_spare_ops,
54 	&vdev_disk_ops,
55 	&vdev_file_ops,
56 	&vdev_missing_ops,
57 	&vdev_hole_ops,
58 	NULL
59 };
60 
61 /* maximum scrub/resilver I/O queue per leaf vdev */
62 int zfs_scrub_limit = 10;
63 
64 /*
65  * Given a vdev type, return the appropriate ops vector.
66  */
67 static vdev_ops_t *
68 vdev_getops(const char *type)
69 {
70 	vdev_ops_t *ops, **opspp;
71 
72 	for (opspp = vdev_ops_table; (ops = *opspp) != NULL; opspp++)
73 		if (strcmp(ops->vdev_op_type, type) == 0)
74 			break;
75 
76 	return (ops);
77 }
78 
79 /*
80  * Default asize function: return the MAX of psize with the asize of
81  * all children.  This is what's used by anything other than RAID-Z.
82  */
83 uint64_t
84 vdev_default_asize(vdev_t *vd, uint64_t psize)
85 {
86 	uint64_t asize = P2ROUNDUP(psize, 1ULL << vd->vdev_top->vdev_ashift);
87 	uint64_t csize;
88 
89 	for (int c = 0; c < vd->vdev_children; c++) {
90 		csize = vdev_psize_to_asize(vd->vdev_child[c], psize);
91 		asize = MAX(asize, csize);
92 	}
93 
94 	return (asize);
95 }
96 
97 /*
98  * Get the minimum allocatable size. We define the allocatable size as
99  * the vdev's asize rounded to the nearest metaslab. This allows us to
100  * replace or attach devices which don't have the same physical size but
101  * can still satisfy the same number of allocations.
102  */
103 uint64_t
104 vdev_get_min_asize(vdev_t *vd)
105 {
106 	vdev_t *pvd = vd->vdev_parent;
107 
108 	/*
109 	 * The our parent is NULL (inactive spare or cache) or is the root,
110 	 * just return our own asize.
111 	 */
112 	if (pvd == NULL)
113 		return (vd->vdev_asize);
114 
115 	/*
116 	 * The top-level vdev just returns the allocatable size rounded
117 	 * to the nearest metaslab.
118 	 */
119 	if (vd == vd->vdev_top)
120 		return (P2ALIGN(vd->vdev_asize, 1ULL << vd->vdev_ms_shift));
121 
122 	/*
123 	 * The allocatable space for a raidz vdev is N * sizeof(smallest child),
124 	 * so each child must provide at least 1/Nth of its asize.
125 	 */
126 	if (pvd->vdev_ops == &vdev_raidz_ops)
127 		return (pvd->vdev_min_asize / pvd->vdev_children);
128 
129 	return (pvd->vdev_min_asize);
130 }
131 
132 void
133 vdev_set_min_asize(vdev_t *vd)
134 {
135 	vd->vdev_min_asize = vdev_get_min_asize(vd);
136 
137 	for (int c = 0; c < vd->vdev_children; c++)
138 		vdev_set_min_asize(vd->vdev_child[c]);
139 }
140 
141 vdev_t *
142 vdev_lookup_top(spa_t *spa, uint64_t vdev)
143 {
144 	vdev_t *rvd = spa->spa_root_vdev;
145 
146 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
147 
148 	if (vdev < rvd->vdev_children) {
149 		ASSERT(rvd->vdev_child[vdev] != NULL);
150 		return (rvd->vdev_child[vdev]);
151 	}
152 
153 	return (NULL);
154 }
155 
156 vdev_t *
157 vdev_lookup_by_guid(vdev_t *vd, uint64_t guid)
158 {
159 	vdev_t *mvd;
160 
161 	if (vd->vdev_guid == guid)
162 		return (vd);
163 
164 	for (int c = 0; c < vd->vdev_children; c++)
165 		if ((mvd = vdev_lookup_by_guid(vd->vdev_child[c], guid)) !=
166 		    NULL)
167 			return (mvd);
168 
169 	return (NULL);
170 }
171 
172 void
173 vdev_add_child(vdev_t *pvd, vdev_t *cvd)
174 {
175 	size_t oldsize, newsize;
176 	uint64_t id = cvd->vdev_id;
177 	vdev_t **newchild;
178 
179 	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
180 	ASSERT(cvd->vdev_parent == NULL);
181 
182 	cvd->vdev_parent = pvd;
183 
184 	if (pvd == NULL)
185 		return;
186 
187 	ASSERT(id >= pvd->vdev_children || pvd->vdev_child[id] == NULL);
188 
189 	oldsize = pvd->vdev_children * sizeof (vdev_t *);
190 	pvd->vdev_children = MAX(pvd->vdev_children, id + 1);
191 	newsize = pvd->vdev_children * sizeof (vdev_t *);
192 
193 	newchild = kmem_zalloc(newsize, KM_SLEEP);
194 	if (pvd->vdev_child != NULL) {
195 		bcopy(pvd->vdev_child, newchild, oldsize);
196 		kmem_free(pvd->vdev_child, oldsize);
197 	}
198 
199 	pvd->vdev_child = newchild;
200 	pvd->vdev_child[id] = cvd;
201 
202 	cvd->vdev_top = (pvd->vdev_top ? pvd->vdev_top: cvd);
203 	ASSERT(cvd->vdev_top->vdev_parent->vdev_parent == NULL);
204 
205 	/*
206 	 * Walk up all ancestors to update guid sum.
207 	 */
208 	for (; pvd != NULL; pvd = pvd->vdev_parent)
209 		pvd->vdev_guid_sum += cvd->vdev_guid_sum;
210 
211 	if (cvd->vdev_ops->vdev_op_leaf)
212 		cvd->vdev_spa->spa_scrub_maxinflight += zfs_scrub_limit;
213 }
214 
215 void
216 vdev_remove_child(vdev_t *pvd, vdev_t *cvd)
217 {
218 	int c;
219 	uint_t id = cvd->vdev_id;
220 
221 	ASSERT(cvd->vdev_parent == pvd);
222 
223 	if (pvd == NULL)
224 		return;
225 
226 	ASSERT(id < pvd->vdev_children);
227 	ASSERT(pvd->vdev_child[id] == cvd);
228 
229 	pvd->vdev_child[id] = NULL;
230 	cvd->vdev_parent = NULL;
231 
232 	for (c = 0; c < pvd->vdev_children; c++)
233 		if (pvd->vdev_child[c])
234 			break;
235 
236 	if (c == pvd->vdev_children) {
237 		kmem_free(pvd->vdev_child, c * sizeof (vdev_t *));
238 		pvd->vdev_child = NULL;
239 		pvd->vdev_children = 0;
240 	}
241 
242 	/*
243 	 * Walk up all ancestors to update guid sum.
244 	 */
245 	for (; pvd != NULL; pvd = pvd->vdev_parent)
246 		pvd->vdev_guid_sum -= cvd->vdev_guid_sum;
247 
248 	if (cvd->vdev_ops->vdev_op_leaf)
249 		cvd->vdev_spa->spa_scrub_maxinflight -= zfs_scrub_limit;
250 }
251 
252 /*
253  * Remove any holes in the child array.
254  */
255 void
256 vdev_compact_children(vdev_t *pvd)
257 {
258 	vdev_t **newchild, *cvd;
259 	int oldc = pvd->vdev_children;
260 	int newc;
261 
262 	ASSERT(spa_config_held(pvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
263 
264 	for (int c = newc = 0; c < oldc; c++)
265 		if (pvd->vdev_child[c])
266 			newc++;
267 
268 	newchild = kmem_alloc(newc * sizeof (vdev_t *), KM_SLEEP);
269 
270 	for (int c = newc = 0; c < oldc; c++) {
271 		if ((cvd = pvd->vdev_child[c]) != NULL) {
272 			newchild[newc] = cvd;
273 			cvd->vdev_id = newc++;
274 		}
275 	}
276 
277 	kmem_free(pvd->vdev_child, oldc * sizeof (vdev_t *));
278 	pvd->vdev_child = newchild;
279 	pvd->vdev_children = newc;
280 }
281 
282 /*
283  * Allocate and minimally initialize a vdev_t.
284  */
285 vdev_t *
286 vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
287 {
288 	vdev_t *vd;
289 
290 	vd = kmem_zalloc(sizeof (vdev_t), KM_SLEEP);
291 
292 	if (spa->spa_root_vdev == NULL) {
293 		ASSERT(ops == &vdev_root_ops);
294 		spa->spa_root_vdev = vd;
295 	}
296 
297 	if (guid == 0 && ops != &vdev_hole_ops) {
298 		if (spa->spa_root_vdev == vd) {
299 			/*
300 			 * The root vdev's guid will also be the pool guid,
301 			 * which must be unique among all pools.
302 			 */
303 			while (guid == 0 || spa_guid_exists(guid, 0))
304 				guid = spa_get_random(-1ULL);
305 		} else {
306 			/*
307 			 * Any other vdev's guid must be unique within the pool.
308 			 */
309 			while (guid == 0 ||
310 			    spa_guid_exists(spa_guid(spa), guid))
311 				guid = spa_get_random(-1ULL);
312 		}
313 		ASSERT(!spa_guid_exists(spa_guid(spa), guid));
314 	}
315 
316 	vd->vdev_spa = spa;
317 	vd->vdev_id = id;
318 	vd->vdev_guid = guid;
319 	vd->vdev_guid_sum = guid;
320 	vd->vdev_ops = ops;
321 	vd->vdev_state = VDEV_STATE_CLOSED;
322 	vd->vdev_ishole = (ops == &vdev_hole_ops);
323 
324 	mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_DEFAULT, NULL);
325 	mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL);
326 	mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL);
327 	for (int t = 0; t < DTL_TYPES; t++) {
328 		space_map_create(&vd->vdev_dtl[t], 0, -1ULL, 0,
329 		    &vd->vdev_dtl_lock);
330 	}
331 	txg_list_create(&vd->vdev_ms_list,
332 	    offsetof(struct metaslab, ms_txg_node));
333 	txg_list_create(&vd->vdev_dtl_list,
334 	    offsetof(struct vdev, vdev_dtl_node));
335 	vd->vdev_stat.vs_timestamp = gethrtime();
336 	vdev_queue_init(vd);
337 	vdev_cache_init(vd);
338 
339 	return (vd);
340 }
341 
342 /*
343  * Allocate a new vdev.  The 'alloctype' is used to control whether we are
344  * creating a new vdev or loading an existing one - the behavior is slightly
345  * different for each case.
346  */
347 int
348 vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id,
349     int alloctype)
350 {
351 	vdev_ops_t *ops;
352 	char *type;
353 	uint64_t guid = 0, islog, nparity;
354 	vdev_t *vd;
355 
356 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
357 
358 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) != 0)
359 		return (EINVAL);
360 
361 	if ((ops = vdev_getops(type)) == NULL)
362 		return (EINVAL);
363 
364 	/*
365 	 * If this is a load, get the vdev guid from the nvlist.
366 	 * Otherwise, vdev_alloc_common() will generate one for us.
367 	 */
368 	if (alloctype == VDEV_ALLOC_LOAD) {
369 		uint64_t label_id;
370 
371 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ID, &label_id) ||
372 		    label_id != id)
373 			return (EINVAL);
374 
375 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
376 			return (EINVAL);
377 	} else if (alloctype == VDEV_ALLOC_SPARE) {
378 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
379 			return (EINVAL);
380 	} else if (alloctype == VDEV_ALLOC_L2CACHE) {
381 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
382 			return (EINVAL);
383 	} else if (alloctype == VDEV_ALLOC_ROOTPOOL) {
384 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
385 			return (EINVAL);
386 	}
387 
388 	/*
389 	 * The first allocated vdev must be of type 'root'.
390 	 */
391 	if (ops != &vdev_root_ops && spa->spa_root_vdev == NULL)
392 		return (EINVAL);
393 
394 	/*
395 	 * Determine whether we're a log vdev.
396 	 */
397 	islog = 0;
398 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &islog);
399 	if (islog && spa_version(spa) < SPA_VERSION_SLOGS)
400 		return (ENOTSUP);
401 
402 	if (ops == &vdev_hole_ops && spa_version(spa) < SPA_VERSION_HOLES)
403 		return (ENOTSUP);
404 
405 	/*
406 	 * Set the nparity property for RAID-Z vdevs.
407 	 */
408 	nparity = -1ULL;
409 	if (ops == &vdev_raidz_ops) {
410 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY,
411 		    &nparity) == 0) {
412 			if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY)
413 				return (EINVAL);
414 			/*
415 			 * Previous versions could only support 1 or 2 parity
416 			 * device.
417 			 */
418 			if (nparity > 1 &&
419 			    spa_version(spa) < SPA_VERSION_RAIDZ2)
420 				return (ENOTSUP);
421 			if (nparity > 2 &&
422 			    spa_version(spa) < SPA_VERSION_RAIDZ3)
423 				return (ENOTSUP);
424 		} else {
425 			/*
426 			 * We require the parity to be specified for SPAs that
427 			 * support multiple parity levels.
428 			 */
429 			if (spa_version(spa) >= SPA_VERSION_RAIDZ2)
430 				return (EINVAL);
431 			/*
432 			 * Otherwise, we default to 1 parity device for RAID-Z.
433 			 */
434 			nparity = 1;
435 		}
436 	} else {
437 		nparity = 0;
438 	}
439 	ASSERT(nparity != -1ULL);
440 
441 	vd = vdev_alloc_common(spa, id, guid, ops);
442 
443 	vd->vdev_islog = islog;
444 	vd->vdev_nparity = nparity;
445 
446 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &vd->vdev_path) == 0)
447 		vd->vdev_path = spa_strdup(vd->vdev_path);
448 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &vd->vdev_devid) == 0)
449 		vd->vdev_devid = spa_strdup(vd->vdev_devid);
450 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH,
451 	    &vd->vdev_physpath) == 0)
452 		vd->vdev_physpath = spa_strdup(vd->vdev_physpath);
453 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &vd->vdev_fru) == 0)
454 		vd->vdev_fru = spa_strdup(vd->vdev_fru);
455 
456 	/*
457 	 * Set the whole_disk property.  If it's not specified, leave the value
458 	 * as -1.
459 	 */
460 	if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
461 	    &vd->vdev_wholedisk) != 0)
462 		vd->vdev_wholedisk = -1ULL;
463 
464 	/*
465 	 * Look for the 'not present' flag.  This will only be set if the device
466 	 * was not present at the time of import.
467 	 */
468 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT,
469 	    &vd->vdev_not_present);
470 
471 	/*
472 	 * Get the alignment requirement.
473 	 */
474 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT, &vd->vdev_ashift);
475 
476 	/*
477 	 * Retrieve the vdev creation time.
478 	 */
479 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_CREATE_TXG,
480 	    &vd->vdev_crtxg);
481 
482 	/*
483 	 * If we're a top-level vdev, try to load the allocation parameters.
484 	 */
485 	if (parent && !parent->vdev_parent && alloctype == VDEV_ALLOC_LOAD) {
486 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
487 		    &vd->vdev_ms_array);
488 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
489 		    &vd->vdev_ms_shift);
490 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE,
491 		    &vd->vdev_asize);
492 	}
493 
494 	if (parent && !parent->vdev_parent) {
495 		ASSERT(alloctype == VDEV_ALLOC_LOAD ||
496 		    alloctype == VDEV_ALLOC_ADD);
497 		vd->vdev_mg = metaslab_group_create(islog ?
498 		    spa_log_class(spa) : spa_normal_class(spa), vd);
499 	}
500 
501 	/*
502 	 * If we're a leaf vdev, try to load the DTL object and other state.
503 	 */
504 	if (vd->vdev_ops->vdev_op_leaf &&
505 	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE ||
506 	    alloctype == VDEV_ALLOC_ROOTPOOL)) {
507 		if (alloctype == VDEV_ALLOC_LOAD) {
508 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL,
509 			    &vd->vdev_dtl_smo.smo_object);
510 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE,
511 			    &vd->vdev_unspare);
512 		}
513 
514 		if (alloctype == VDEV_ALLOC_ROOTPOOL) {
515 			uint64_t spare = 0;
516 
517 			if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE,
518 			    &spare) == 0 && spare)
519 				spa_spare_add(vd);
520 		}
521 
522 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE,
523 		    &vd->vdev_offline);
524 
525 		/*
526 		 * When importing a pool, we want to ignore the persistent fault
527 		 * state, as the diagnosis made on another system may not be
528 		 * valid in the current context.  Local vdevs will
529 		 * remain in the faulted state.
530 		 */
531 		if (spa->spa_load_state == SPA_LOAD_OPEN) {
532 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED,
533 			    &vd->vdev_faulted);
534 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED,
535 			    &vd->vdev_degraded);
536 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED,
537 			    &vd->vdev_removed);
538 
539 			if (vd->vdev_faulted || vd->vdev_degraded) {
540 				char *aux;
541 
542 				vd->vdev_label_aux =
543 				    VDEV_AUX_ERR_EXCEEDED;
544 				if (nvlist_lookup_string(nv,
545 				    ZPOOL_CONFIG_AUX_STATE, &aux) == 0 &&
546 				    strcmp(aux, "external") == 0)
547 					vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
548 			}
549 		}
550 	}
551 
552 	/*
553 	 * Add ourselves to the parent's list of children.
554 	 */
555 	vdev_add_child(parent, vd);
556 
557 	*vdp = vd;
558 
559 	return (0);
560 }
561 
562 void
563 vdev_free(vdev_t *vd)
564 {
565 	spa_t *spa = vd->vdev_spa;
566 
567 	/*
568 	 * vdev_free() implies closing the vdev first.  This is simpler than
569 	 * trying to ensure complicated semantics for all callers.
570 	 */
571 	vdev_close(vd);
572 
573 	ASSERT(!list_link_active(&vd->vdev_config_dirty_node));
574 	ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
575 
576 	/*
577 	 * Free all children.
578 	 */
579 	for (int c = 0; c < vd->vdev_children; c++)
580 		vdev_free(vd->vdev_child[c]);
581 
582 	ASSERT(vd->vdev_child == NULL);
583 	ASSERT(vd->vdev_guid_sum == vd->vdev_guid);
584 
585 	/*
586 	 * Discard allocation state.
587 	 */
588 	if (vd->vdev_mg != NULL) {
589 		vdev_metaslab_fini(vd);
590 		metaslab_group_destroy(vd->vdev_mg);
591 	}
592 
593 	ASSERT3U(vd->vdev_stat.vs_space, ==, 0);
594 	ASSERT3U(vd->vdev_stat.vs_dspace, ==, 0);
595 	ASSERT3U(vd->vdev_stat.vs_alloc, ==, 0);
596 
597 	/*
598 	 * Remove this vdev from its parent's child list.
599 	 */
600 	vdev_remove_child(vd->vdev_parent, vd);
601 
602 	ASSERT(vd->vdev_parent == NULL);
603 
604 	/*
605 	 * Clean up vdev structure.
606 	 */
607 	vdev_queue_fini(vd);
608 	vdev_cache_fini(vd);
609 
610 	if (vd->vdev_path)
611 		spa_strfree(vd->vdev_path);
612 	if (vd->vdev_devid)
613 		spa_strfree(vd->vdev_devid);
614 	if (vd->vdev_physpath)
615 		spa_strfree(vd->vdev_physpath);
616 	if (vd->vdev_fru)
617 		spa_strfree(vd->vdev_fru);
618 
619 	if (vd->vdev_isspare)
620 		spa_spare_remove(vd);
621 	if (vd->vdev_isl2cache)
622 		spa_l2cache_remove(vd);
623 
624 	txg_list_destroy(&vd->vdev_ms_list);
625 	txg_list_destroy(&vd->vdev_dtl_list);
626 
627 	mutex_enter(&vd->vdev_dtl_lock);
628 	for (int t = 0; t < DTL_TYPES; t++) {
629 		space_map_unload(&vd->vdev_dtl[t]);
630 		space_map_destroy(&vd->vdev_dtl[t]);
631 	}
632 	mutex_exit(&vd->vdev_dtl_lock);
633 
634 	mutex_destroy(&vd->vdev_dtl_lock);
635 	mutex_destroy(&vd->vdev_stat_lock);
636 	mutex_destroy(&vd->vdev_probe_lock);
637 
638 	if (vd == spa->spa_root_vdev)
639 		spa->spa_root_vdev = NULL;
640 
641 	kmem_free(vd, sizeof (vdev_t));
642 }
643 
644 /*
645  * Transfer top-level vdev state from svd to tvd.
646  */
647 static void
648 vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
649 {
650 	spa_t *spa = svd->vdev_spa;
651 	metaslab_t *msp;
652 	vdev_t *vd;
653 	int t;
654 
655 	ASSERT(tvd == tvd->vdev_top);
656 
657 	tvd->vdev_ms_array = svd->vdev_ms_array;
658 	tvd->vdev_ms_shift = svd->vdev_ms_shift;
659 	tvd->vdev_ms_count = svd->vdev_ms_count;
660 
661 	svd->vdev_ms_array = 0;
662 	svd->vdev_ms_shift = 0;
663 	svd->vdev_ms_count = 0;
664 
665 	tvd->vdev_mg = svd->vdev_mg;
666 	tvd->vdev_ms = svd->vdev_ms;
667 
668 	svd->vdev_mg = NULL;
669 	svd->vdev_ms = NULL;
670 
671 	if (tvd->vdev_mg != NULL)
672 		tvd->vdev_mg->mg_vd = tvd;
673 
674 	tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
675 	tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
676 	tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
677 
678 	svd->vdev_stat.vs_alloc = 0;
679 	svd->vdev_stat.vs_space = 0;
680 	svd->vdev_stat.vs_dspace = 0;
681 
682 	for (t = 0; t < TXG_SIZE; t++) {
683 		while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
684 			(void) txg_list_add(&tvd->vdev_ms_list, msp, t);
685 		while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
686 			(void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
687 		if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
688 			(void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
689 	}
690 
691 	if (list_link_active(&svd->vdev_config_dirty_node)) {
692 		vdev_config_clean(svd);
693 		vdev_config_dirty(tvd);
694 	}
695 
696 	if (list_link_active(&svd->vdev_state_dirty_node)) {
697 		vdev_state_clean(svd);
698 		vdev_state_dirty(tvd);
699 	}
700 
701 	tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
702 	svd->vdev_deflate_ratio = 0;
703 
704 	tvd->vdev_islog = svd->vdev_islog;
705 	svd->vdev_islog = 0;
706 }
707 
708 static void
709 vdev_top_update(vdev_t *tvd, vdev_t *vd)
710 {
711 	if (vd == NULL)
712 		return;
713 
714 	vd->vdev_top = tvd;
715 
716 	for (int c = 0; c < vd->vdev_children; c++)
717 		vdev_top_update(tvd, vd->vdev_child[c]);
718 }
719 
720 /*
721  * Add a mirror/replacing vdev above an existing vdev.
722  */
723 vdev_t *
724 vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
725 {
726 	spa_t *spa = cvd->vdev_spa;
727 	vdev_t *pvd = cvd->vdev_parent;
728 	vdev_t *mvd;
729 
730 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
731 
732 	mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
733 
734 	mvd->vdev_asize = cvd->vdev_asize;
735 	mvd->vdev_min_asize = cvd->vdev_min_asize;
736 	mvd->vdev_ashift = cvd->vdev_ashift;
737 	mvd->vdev_state = cvd->vdev_state;
738 	mvd->vdev_crtxg = cvd->vdev_crtxg;
739 
740 	vdev_remove_child(pvd, cvd);
741 	vdev_add_child(pvd, mvd);
742 	cvd->vdev_id = mvd->vdev_children;
743 	vdev_add_child(mvd, cvd);
744 	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
745 
746 	if (mvd == mvd->vdev_top)
747 		vdev_top_transfer(cvd, mvd);
748 
749 	return (mvd);
750 }
751 
752 /*
753  * Remove a 1-way mirror/replacing vdev from the tree.
754  */
755 void
756 vdev_remove_parent(vdev_t *cvd)
757 {
758 	vdev_t *mvd = cvd->vdev_parent;
759 	vdev_t *pvd = mvd->vdev_parent;
760 
761 	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
762 
763 	ASSERT(mvd->vdev_children == 1);
764 	ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
765 	    mvd->vdev_ops == &vdev_replacing_ops ||
766 	    mvd->vdev_ops == &vdev_spare_ops);
767 	cvd->vdev_ashift = mvd->vdev_ashift;
768 
769 	vdev_remove_child(mvd, cvd);
770 	vdev_remove_child(pvd, mvd);
771 
772 	/*
773 	 * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
774 	 * Otherwise, we could have detached an offline device, and when we
775 	 * go to import the pool we'll think we have two top-level vdevs,
776 	 * instead of a different version of the same top-level vdev.
777 	 */
778 	if (mvd->vdev_top == mvd) {
779 		uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
780 		cvd->vdev_guid += guid_delta;
781 		cvd->vdev_guid_sum += guid_delta;
782 	}
783 	cvd->vdev_id = mvd->vdev_id;
784 	vdev_add_child(pvd, cvd);
785 	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
786 
787 	if (cvd == cvd->vdev_top)
788 		vdev_top_transfer(mvd, cvd);
789 
790 	ASSERT(mvd->vdev_children == 0);
791 	vdev_free(mvd);
792 }
793 
794 int
795 vdev_metaslab_init(vdev_t *vd, uint64_t txg)
796 {
797 	spa_t *spa = vd->vdev_spa;
798 	objset_t *mos = spa->spa_meta_objset;
799 	uint64_t m;
800 	uint64_t oldc = vd->vdev_ms_count;
801 	uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
802 	metaslab_t **mspp;
803 	int error;
804 
805 	ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER));
806 
807 	/*
808 	 * This vdev is not being allocated from yet or is a hole.
809 	 */
810 	if (vd->vdev_ms_shift == 0)
811 		return (0);
812 
813 	ASSERT(!vd->vdev_ishole);
814 
815 	/*
816 	 * Compute the raidz-deflation ratio.  Note, we hard-code
817 	 * in 128k (1 << 17) because it is the current "typical" blocksize.
818 	 * Even if SPA_MAXBLOCKSIZE changes, this algorithm must never change,
819 	 * or we will inconsistently account for existing bp's.
820 	 */
821 	vd->vdev_deflate_ratio = (1 << 17) /
822 	    (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
823 
824 	ASSERT(oldc <= newc);
825 
826 	mspp = kmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
827 
828 	if (oldc != 0) {
829 		bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp));
830 		kmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
831 	}
832 
833 	vd->vdev_ms = mspp;
834 	vd->vdev_ms_count = newc;
835 
836 	for (m = oldc; m < newc; m++) {
837 		space_map_obj_t smo = { 0, 0, 0 };
838 		if (txg == 0) {
839 			uint64_t object = 0;
840 			error = dmu_read(mos, vd->vdev_ms_array,
841 			    m * sizeof (uint64_t), sizeof (uint64_t), &object,
842 			    DMU_READ_PREFETCH);
843 			if (error)
844 				return (error);
845 			if (object != 0) {
846 				dmu_buf_t *db;
847 				error = dmu_bonus_hold(mos, object, FTAG, &db);
848 				if (error)
849 					return (error);
850 				ASSERT3U(db->db_size, >=, sizeof (smo));
851 				bcopy(db->db_data, &smo, sizeof (smo));
852 				ASSERT3U(smo.smo_object, ==, object);
853 				dmu_buf_rele(db, FTAG);
854 			}
855 		}
856 		vd->vdev_ms[m] = metaslab_init(vd->vdev_mg, &smo,
857 		    m << vd->vdev_ms_shift, 1ULL << vd->vdev_ms_shift, txg);
858 	}
859 
860 	if (txg == 0)
861 		spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER);
862 
863 	if (oldc == 0)
864 		metaslab_group_activate(vd->vdev_mg);
865 
866 	if (txg == 0)
867 		spa_config_exit(spa, SCL_ALLOC, FTAG);
868 
869 	return (0);
870 }
871 
872 void
873 vdev_metaslab_fini(vdev_t *vd)
874 {
875 	uint64_t m;
876 	uint64_t count = vd->vdev_ms_count;
877 
878 	if (vd->vdev_ms != NULL) {
879 		metaslab_group_passivate(vd->vdev_mg);
880 		for (m = 0; m < count; m++)
881 			if (vd->vdev_ms[m] != NULL)
882 				metaslab_fini(vd->vdev_ms[m]);
883 		kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
884 		vd->vdev_ms = NULL;
885 	}
886 }
887 
888 typedef struct vdev_probe_stats {
889 	boolean_t	vps_readable;
890 	boolean_t	vps_writeable;
891 	int		vps_flags;
892 } vdev_probe_stats_t;
893 
894 static void
895 vdev_probe_done(zio_t *zio)
896 {
897 	spa_t *spa = zio->io_spa;
898 	vdev_t *vd = zio->io_vd;
899 	vdev_probe_stats_t *vps = zio->io_private;
900 
901 	ASSERT(vd->vdev_probe_zio != NULL);
902 
903 	if (zio->io_type == ZIO_TYPE_READ) {
904 		if (zio->io_error == 0)
905 			vps->vps_readable = 1;
906 		if (zio->io_error == 0 && spa_writeable(spa)) {
907 			zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
908 			    zio->io_offset, zio->io_size, zio->io_data,
909 			    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
910 			    ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
911 		} else {
912 			zio_buf_free(zio->io_data, zio->io_size);
913 		}
914 	} else if (zio->io_type == ZIO_TYPE_WRITE) {
915 		if (zio->io_error == 0)
916 			vps->vps_writeable = 1;
917 		zio_buf_free(zio->io_data, zio->io_size);
918 	} else if (zio->io_type == ZIO_TYPE_NULL) {
919 		zio_t *pio;
920 
921 		vd->vdev_cant_read |= !vps->vps_readable;
922 		vd->vdev_cant_write |= !vps->vps_writeable;
923 
924 		if (vdev_readable(vd) &&
925 		    (vdev_writeable(vd) || !spa_writeable(spa))) {
926 			zio->io_error = 0;
927 		} else {
928 			ASSERT(zio->io_error != 0);
929 			zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
930 			    spa, vd, NULL, 0, 0);
931 			zio->io_error = ENXIO;
932 		}
933 
934 		mutex_enter(&vd->vdev_probe_lock);
935 		ASSERT(vd->vdev_probe_zio == zio);
936 		vd->vdev_probe_zio = NULL;
937 		mutex_exit(&vd->vdev_probe_lock);
938 
939 		while ((pio = zio_walk_parents(zio)) != NULL)
940 			if (!vdev_accessible(vd, pio))
941 				pio->io_error = ENXIO;
942 
943 		kmem_free(vps, sizeof (*vps));
944 	}
945 }
946 
947 /*
948  * Determine whether this device is accessible by reading and writing
949  * to several known locations: the pad regions of each vdev label
950  * but the first (which we leave alone in case it contains a VTOC).
951  */
952 zio_t *
953 vdev_probe(vdev_t *vd, zio_t *zio)
954 {
955 	spa_t *spa = vd->vdev_spa;
956 	vdev_probe_stats_t *vps = NULL;
957 	zio_t *pio;
958 
959 	ASSERT(vd->vdev_ops->vdev_op_leaf);
960 
961 	/*
962 	 * Don't probe the probe.
963 	 */
964 	if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
965 		return (NULL);
966 
967 	/*
968 	 * To prevent 'probe storms' when a device fails, we create
969 	 * just one probe i/o at a time.  All zios that want to probe
970 	 * this vdev will become parents of the probe io.
971 	 */
972 	mutex_enter(&vd->vdev_probe_lock);
973 
974 	if ((pio = vd->vdev_probe_zio) == NULL) {
975 		vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
976 
977 		vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
978 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
979 		    ZIO_FLAG_TRYHARD;
980 
981 		if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
982 			/*
983 			 * vdev_cant_read and vdev_cant_write can only
984 			 * transition from TRUE to FALSE when we have the
985 			 * SCL_ZIO lock as writer; otherwise they can only
986 			 * transition from FALSE to TRUE.  This ensures that
987 			 * any zio looking at these values can assume that
988 			 * failures persist for the life of the I/O.  That's
989 			 * important because when a device has intermittent
990 			 * connectivity problems, we want to ensure that
991 			 * they're ascribed to the device (ENXIO) and not
992 			 * the zio (EIO).
993 			 *
994 			 * Since we hold SCL_ZIO as writer here, clear both
995 			 * values so the probe can reevaluate from first
996 			 * principles.
997 			 */
998 			vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
999 			vd->vdev_cant_read = B_FALSE;
1000 			vd->vdev_cant_write = B_FALSE;
1001 		}
1002 
1003 		vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
1004 		    vdev_probe_done, vps,
1005 		    vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
1006 
1007 		if (zio != NULL) {
1008 			vd->vdev_probe_wanted = B_TRUE;
1009 			spa_async_request(spa, SPA_ASYNC_PROBE);
1010 		}
1011 	}
1012 
1013 	if (zio != NULL)
1014 		zio_add_child(zio, pio);
1015 
1016 	mutex_exit(&vd->vdev_probe_lock);
1017 
1018 	if (vps == NULL) {
1019 		ASSERT(zio != NULL);
1020 		return (NULL);
1021 	}
1022 
1023 	for (int l = 1; l < VDEV_LABELS; l++) {
1024 		zio_nowait(zio_read_phys(pio, vd,
1025 		    vdev_label_offset(vd->vdev_psize, l,
1026 		    offsetof(vdev_label_t, vl_pad2)),
1027 		    VDEV_PAD_SIZE, zio_buf_alloc(VDEV_PAD_SIZE),
1028 		    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1029 		    ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1030 	}
1031 
1032 	if (zio == NULL)
1033 		return (pio);
1034 
1035 	zio_nowait(pio);
1036 	return (NULL);
1037 }
1038 
1039 static void
1040 vdev_open_child(void *arg)
1041 {
1042 	vdev_t *vd = arg;
1043 
1044 	vd->vdev_open_thread = curthread;
1045 	vd->vdev_open_error = vdev_open(vd);
1046 	vd->vdev_open_thread = NULL;
1047 }
1048 
1049 boolean_t
1050 vdev_uses_zvols(vdev_t *vd)
1051 {
1052 	if (vd->vdev_path && strncmp(vd->vdev_path, ZVOL_DIR,
1053 	    strlen(ZVOL_DIR)) == 0)
1054 		return (B_TRUE);
1055 	for (int c = 0; c < vd->vdev_children; c++)
1056 		if (vdev_uses_zvols(vd->vdev_child[c]))
1057 			return (B_TRUE);
1058 	return (B_FALSE);
1059 }
1060 
1061 void
1062 vdev_open_children(vdev_t *vd)
1063 {
1064 	taskq_t *tq;
1065 	int children = vd->vdev_children;
1066 
1067 	/*
1068 	 * in order to handle pools on top of zvols, do the opens
1069 	 * in a single thread so that the same thread holds the
1070 	 * spa_namespace_lock
1071 	 */
1072 	if (vdev_uses_zvols(vd)) {
1073 		for (int c = 0; c < children; c++)
1074 			vd->vdev_child[c]->vdev_open_error =
1075 			    vdev_open(vd->vdev_child[c]);
1076 		return;
1077 	}
1078 	tq = taskq_create("vdev_open", children, minclsyspri,
1079 	    children, children, TASKQ_PREPOPULATE);
1080 
1081 	for (int c = 0; c < children; c++)
1082 		VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c],
1083 		    TQ_SLEEP) != NULL);
1084 
1085 	taskq_destroy(tq);
1086 }
1087 
1088 /*
1089  * Prepare a virtual device for access.
1090  */
1091 int
1092 vdev_open(vdev_t *vd)
1093 {
1094 	spa_t *spa = vd->vdev_spa;
1095 	int error;
1096 	uint64_t osize = 0;
1097 	uint64_t asize, psize;
1098 	uint64_t ashift = 0;
1099 
1100 	ASSERT(vd->vdev_open_thread == curthread ||
1101 	    spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1102 	ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1103 	    vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1104 	    vd->vdev_state == VDEV_STATE_OFFLINE);
1105 
1106 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1107 	vd->vdev_cant_read = B_FALSE;
1108 	vd->vdev_cant_write = B_FALSE;
1109 	vd->vdev_min_asize = vdev_get_min_asize(vd);
1110 
1111 	/*
1112 	 * If this vdev is not removed, check its fault status.  If it's
1113 	 * faulted, bail out of the open.
1114 	 */
1115 	if (!vd->vdev_removed && vd->vdev_faulted) {
1116 		ASSERT(vd->vdev_children == 0);
1117 		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1118 		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1119 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1120 		    vd->vdev_label_aux);
1121 		return (ENXIO);
1122 	} else if (vd->vdev_offline) {
1123 		ASSERT(vd->vdev_children == 0);
1124 		vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1125 		return (ENXIO);
1126 	}
1127 
1128 	error = vd->vdev_ops->vdev_op_open(vd, &osize, &ashift);
1129 
1130 	/*
1131 	 * Reset the vdev_reopening flag so that we actually close
1132 	 * the vdev on error.
1133 	 */
1134 	vd->vdev_reopening = B_FALSE;
1135 	if (zio_injection_enabled && error == 0)
1136 		error = zio_handle_device_injection(vd, NULL, ENXIO);
1137 
1138 	if (error) {
1139 		if (vd->vdev_removed &&
1140 		    vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1141 			vd->vdev_removed = B_FALSE;
1142 
1143 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1144 		    vd->vdev_stat.vs_aux);
1145 		return (error);
1146 	}
1147 
1148 	vd->vdev_removed = B_FALSE;
1149 
1150 	/*
1151 	 * Recheck the faulted flag now that we have confirmed that
1152 	 * the vdev is accessible.  If we're faulted, bail.
1153 	 */
1154 	if (vd->vdev_faulted) {
1155 		ASSERT(vd->vdev_children == 0);
1156 		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1157 		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1158 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1159 		    vd->vdev_label_aux);
1160 		return (ENXIO);
1161 	}
1162 
1163 	if (vd->vdev_degraded) {
1164 		ASSERT(vd->vdev_children == 0);
1165 		vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1166 		    VDEV_AUX_ERR_EXCEEDED);
1167 	} else {
1168 		vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1169 	}
1170 
1171 	/*
1172 	 * For hole or missing vdevs we just return success.
1173 	 */
1174 	if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1175 		return (0);
1176 
1177 	for (int c = 0; c < vd->vdev_children; c++) {
1178 		if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1179 			vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1180 			    VDEV_AUX_NONE);
1181 			break;
1182 		}
1183 	}
1184 
1185 	osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1186 
1187 	if (vd->vdev_children == 0) {
1188 		if (osize < SPA_MINDEVSIZE) {
1189 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1190 			    VDEV_AUX_TOO_SMALL);
1191 			return (EOVERFLOW);
1192 		}
1193 		psize = osize;
1194 		asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1195 	} else {
1196 		if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1197 		    (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1198 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1199 			    VDEV_AUX_TOO_SMALL);
1200 			return (EOVERFLOW);
1201 		}
1202 		psize = 0;
1203 		asize = osize;
1204 	}
1205 
1206 	vd->vdev_psize = psize;
1207 
1208 	/*
1209 	 * Make sure the allocatable size hasn't shrunk.
1210 	 */
1211 	if (asize < vd->vdev_min_asize) {
1212 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1213 		    VDEV_AUX_BAD_LABEL);
1214 		return (EINVAL);
1215 	}
1216 
1217 	if (vd->vdev_asize == 0) {
1218 		/*
1219 		 * This is the first-ever open, so use the computed values.
1220 		 * For testing purposes, a higher ashift can be requested.
1221 		 */
1222 		vd->vdev_asize = asize;
1223 		vd->vdev_ashift = MAX(ashift, vd->vdev_ashift);
1224 	} else {
1225 		/*
1226 		 * Make sure the alignment requirement hasn't increased.
1227 		 */
1228 		if (ashift > vd->vdev_top->vdev_ashift) {
1229 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1230 			    VDEV_AUX_BAD_LABEL);
1231 			return (EINVAL);
1232 		}
1233 	}
1234 
1235 	/*
1236 	 * If all children are healthy and the asize has increased,
1237 	 * then we've experienced dynamic LUN growth.  If automatic
1238 	 * expansion is enabled then use the additional space.
1239 	 */
1240 	if (vd->vdev_state == VDEV_STATE_HEALTHY && asize > vd->vdev_asize &&
1241 	    (vd->vdev_expanding || spa->spa_autoexpand))
1242 		vd->vdev_asize = asize;
1243 
1244 	vdev_set_min_asize(vd);
1245 
1246 	/*
1247 	 * Ensure we can issue some IO before declaring the
1248 	 * vdev open for business.
1249 	 */
1250 	if (vd->vdev_ops->vdev_op_leaf &&
1251 	    (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1252 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1253 		    VDEV_AUX_IO_FAILURE);
1254 		return (error);
1255 	}
1256 
1257 	/*
1258 	 * If a leaf vdev has a DTL, and seems healthy, then kick off a
1259 	 * resilver.  But don't do this if we are doing a reopen for a scrub,
1260 	 * since this would just restart the scrub we are already doing.
1261 	 */
1262 	if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1263 	    vdev_resilver_needed(vd, NULL, NULL))
1264 		spa_async_request(spa, SPA_ASYNC_RESILVER);
1265 
1266 	return (0);
1267 }
1268 
1269 /*
1270  * Called once the vdevs are all opened, this routine validates the label
1271  * contents.  This needs to be done before vdev_load() so that we don't
1272  * inadvertently do repair I/Os to the wrong device.
1273  *
1274  * This function will only return failure if one of the vdevs indicates that it
1275  * has since been destroyed or exported.  This is only possible if
1276  * /etc/zfs/zpool.cache was readonly at the time.  Otherwise, the vdev state
1277  * will be updated but the function will return 0.
1278  */
1279 int
1280 vdev_validate(vdev_t *vd)
1281 {
1282 	spa_t *spa = vd->vdev_spa;
1283 	nvlist_t *label;
1284 	uint64_t guid, top_guid;
1285 	uint64_t state;
1286 
1287 	for (int c = 0; c < vd->vdev_children; c++)
1288 		if (vdev_validate(vd->vdev_child[c]) != 0)
1289 			return (EBADF);
1290 
1291 	/*
1292 	 * If the device has already failed, or was marked offline, don't do
1293 	 * any further validation.  Otherwise, label I/O will fail and we will
1294 	 * overwrite the previous state.
1295 	 */
1296 	if (vd->vdev_ops->vdev_op_leaf && vdev_readable(vd)) {
1297 
1298 		if ((label = vdev_label_read_config(vd)) == NULL) {
1299 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1300 			    VDEV_AUX_BAD_LABEL);
1301 			return (0);
1302 		}
1303 
1304 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_GUID,
1305 		    &guid) != 0 || guid != spa_guid(spa)) {
1306 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1307 			    VDEV_AUX_CORRUPT_DATA);
1308 			nvlist_free(label);
1309 			return (0);
1310 		}
1311 
1312 		/*
1313 		 * If this vdev just became a top-level vdev because its
1314 		 * sibling was detached, it will have adopted the parent's
1315 		 * vdev guid -- but the label may or may not be on disk yet.
1316 		 * Fortunately, either version of the label will have the
1317 		 * same top guid, so if we're a top-level vdev, we can
1318 		 * safely compare to that instead.
1319 		 */
1320 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID,
1321 		    &guid) != 0 ||
1322 		    nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID,
1323 		    &top_guid) != 0 ||
1324 		    (vd->vdev_guid != guid &&
1325 		    (vd->vdev_guid != top_guid || vd != vd->vdev_top))) {
1326 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1327 			    VDEV_AUX_CORRUPT_DATA);
1328 			nvlist_free(label);
1329 			return (0);
1330 		}
1331 
1332 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1333 		    &state) != 0) {
1334 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1335 			    VDEV_AUX_CORRUPT_DATA);
1336 			nvlist_free(label);
1337 			return (0);
1338 		}
1339 
1340 		nvlist_free(label);
1341 
1342 		/*
1343 		 * If spa->spa_load_verbatim is true, no need to check the
1344 		 * state of the pool.
1345 		 */
1346 		if (!spa->spa_load_verbatim &&
1347 		    spa->spa_load_state == SPA_LOAD_OPEN &&
1348 		    state != POOL_STATE_ACTIVE)
1349 			return (EBADF);
1350 
1351 		/*
1352 		 * If we were able to open and validate a vdev that was
1353 		 * previously marked permanently unavailable, clear that state
1354 		 * now.
1355 		 */
1356 		if (vd->vdev_not_present)
1357 			vd->vdev_not_present = 0;
1358 	}
1359 
1360 	return (0);
1361 }
1362 
1363 /*
1364  * Close a virtual device.
1365  */
1366 void
1367 vdev_close(vdev_t *vd)
1368 {
1369 	spa_t *spa = vd->vdev_spa;
1370 	vdev_t *pvd = vd->vdev_parent;
1371 
1372 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1373 
1374 	if (pvd != NULL && pvd->vdev_reopening)
1375 		vd->vdev_reopening = pvd->vdev_reopening;
1376 
1377 	vd->vdev_ops->vdev_op_close(vd);
1378 
1379 	vdev_cache_purge(vd);
1380 
1381 	/*
1382 	 * We record the previous state before we close it, so that if we are
1383 	 * doing a reopen(), we don't generate FMA ereports if we notice that
1384 	 * it's still faulted.
1385 	 */
1386 	vd->vdev_prevstate = vd->vdev_state;
1387 
1388 	if (vd->vdev_offline)
1389 		vd->vdev_state = VDEV_STATE_OFFLINE;
1390 	else
1391 		vd->vdev_state = VDEV_STATE_CLOSED;
1392 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1393 }
1394 
1395 /*
1396  * Reopen all interior vdevs and any unopened leaves.  We don't actually
1397  * reopen leaf vdevs which had previously been opened as they might deadlock
1398  * on the spa_config_lock.  Instead we only obtain the leaf's physical size.
1399  * If the leaf has never been opened then open it, as usual.
1400  */
1401 void
1402 vdev_reopen(vdev_t *vd)
1403 {
1404 	spa_t *spa = vd->vdev_spa;
1405 
1406 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1407 
1408 	vd->vdev_reopening = B_TRUE;
1409 	vdev_close(vd);
1410 	(void) vdev_open(vd);
1411 
1412 	/*
1413 	 * Call vdev_validate() here to make sure we have the same device.
1414 	 * Otherwise, a device with an invalid label could be successfully
1415 	 * opened in response to vdev_reopen().
1416 	 */
1417 	if (vd->vdev_aux) {
1418 		(void) vdev_validate_aux(vd);
1419 		if (vdev_readable(vd) && vdev_writeable(vd) &&
1420 		    vd->vdev_aux == &spa->spa_l2cache &&
1421 		    !l2arc_vdev_present(vd))
1422 			l2arc_add_vdev(spa, vd);
1423 	} else {
1424 		(void) vdev_validate(vd);
1425 	}
1426 
1427 	/*
1428 	 * Reassess parent vdev's health.
1429 	 */
1430 	vdev_propagate_state(vd);
1431 }
1432 
1433 int
1434 vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1435 {
1436 	int error;
1437 
1438 	/*
1439 	 * Normally, partial opens (e.g. of a mirror) are allowed.
1440 	 * For a create, however, we want to fail the request if
1441 	 * there are any components we can't open.
1442 	 */
1443 	error = vdev_open(vd);
1444 
1445 	if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
1446 		vdev_close(vd);
1447 		return (error ? error : ENXIO);
1448 	}
1449 
1450 	/*
1451 	 * Recursively initialize all labels.
1452 	 */
1453 	if ((error = vdev_label_init(vd, txg, isreplacing ?
1454 	    VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
1455 		vdev_close(vd);
1456 		return (error);
1457 	}
1458 
1459 	return (0);
1460 }
1461 
1462 void
1463 vdev_metaslab_set_size(vdev_t *vd)
1464 {
1465 	/*
1466 	 * Aim for roughly 200 metaslabs per vdev.
1467 	 */
1468 	vd->vdev_ms_shift = highbit(vd->vdev_asize / 200);
1469 	vd->vdev_ms_shift = MAX(vd->vdev_ms_shift, SPA_MAXBLOCKSHIFT);
1470 }
1471 
1472 void
1473 vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
1474 {
1475 	ASSERT(vd == vd->vdev_top);
1476 	ASSERT(!vd->vdev_ishole);
1477 	ASSERT(ISP2(flags));
1478 
1479 	if (flags & VDD_METASLAB)
1480 		(void) txg_list_add(&vd->vdev_ms_list, arg, txg);
1481 
1482 	if (flags & VDD_DTL)
1483 		(void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
1484 
1485 	(void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
1486 }
1487 
1488 /*
1489  * DTLs.
1490  *
1491  * A vdev's DTL (dirty time log) is the set of transaction groups for which
1492  * the vdev has less than perfect replication.  There are three kinds of DTL:
1493  *
1494  * DTL_MISSING: txgs for which the vdev has no valid copies of the data
1495  *
1496  * DTL_PARTIAL: txgs for which data is available, but not fully replicated
1497  *
1498  * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
1499  *	scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
1500  *	txgs that was scrubbed.
1501  *
1502  * DTL_OUTAGE: txgs which cannot currently be read, whether due to
1503  *	persistent errors or just some device being offline.
1504  *	Unlike the other three, the DTL_OUTAGE map is not generally
1505  *	maintained; it's only computed when needed, typically to
1506  *	determine whether a device can be detached.
1507  *
1508  * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
1509  * either has the data or it doesn't.
1510  *
1511  * For interior vdevs such as mirror and RAID-Z the picture is more complex.
1512  * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
1513  * if any child is less than fully replicated, then so is its parent.
1514  * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
1515  * comprising only those txgs which appear in 'maxfaults' or more children;
1516  * those are the txgs we don't have enough replication to read.  For example,
1517  * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
1518  * thus, its DTL_MISSING consists of the set of txgs that appear in more than
1519  * two child DTL_MISSING maps.
1520  *
1521  * It should be clear from the above that to compute the DTLs and outage maps
1522  * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
1523  * Therefore, that is all we keep on disk.  When loading the pool, or after
1524  * a configuration change, we generate all other DTLs from first principles.
1525  */
1526 void
1527 vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1528 {
1529 	space_map_t *sm = &vd->vdev_dtl[t];
1530 
1531 	ASSERT(t < DTL_TYPES);
1532 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1533 
1534 	mutex_enter(sm->sm_lock);
1535 	if (!space_map_contains(sm, txg, size))
1536 		space_map_add(sm, txg, size);
1537 	mutex_exit(sm->sm_lock);
1538 }
1539 
1540 boolean_t
1541 vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1542 {
1543 	space_map_t *sm = &vd->vdev_dtl[t];
1544 	boolean_t dirty = B_FALSE;
1545 
1546 	ASSERT(t < DTL_TYPES);
1547 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1548 
1549 	mutex_enter(sm->sm_lock);
1550 	if (sm->sm_space != 0)
1551 		dirty = space_map_contains(sm, txg, size);
1552 	mutex_exit(sm->sm_lock);
1553 
1554 	return (dirty);
1555 }
1556 
1557 boolean_t
1558 vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
1559 {
1560 	space_map_t *sm = &vd->vdev_dtl[t];
1561 	boolean_t empty;
1562 
1563 	mutex_enter(sm->sm_lock);
1564 	empty = (sm->sm_space == 0);
1565 	mutex_exit(sm->sm_lock);
1566 
1567 	return (empty);
1568 }
1569 
1570 /*
1571  * Reassess DTLs after a config change or scrub completion.
1572  */
1573 void
1574 vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
1575 {
1576 	spa_t *spa = vd->vdev_spa;
1577 	avl_tree_t reftree;
1578 	int minref;
1579 
1580 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
1581 
1582 	for (int c = 0; c < vd->vdev_children; c++)
1583 		vdev_dtl_reassess(vd->vdev_child[c], txg,
1584 		    scrub_txg, scrub_done);
1585 
1586 	if (vd == spa->spa_root_vdev || vd->vdev_ishole || vd->vdev_aux)
1587 		return;
1588 
1589 	if (vd->vdev_ops->vdev_op_leaf) {
1590 		mutex_enter(&vd->vdev_dtl_lock);
1591 		if (scrub_txg != 0 &&
1592 		    (spa->spa_scrub_started || spa->spa_scrub_errors == 0)) {
1593 			/* XXX should check scrub_done? */
1594 			/*
1595 			 * We completed a scrub up to scrub_txg.  If we
1596 			 * did it without rebooting, then the scrub dtl
1597 			 * will be valid, so excise the old region and
1598 			 * fold in the scrub dtl.  Otherwise, leave the
1599 			 * dtl as-is if there was an error.
1600 			 *
1601 			 * There's little trick here: to excise the beginning
1602 			 * of the DTL_MISSING map, we put it into a reference
1603 			 * tree and then add a segment with refcnt -1 that
1604 			 * covers the range [0, scrub_txg).  This means
1605 			 * that each txg in that range has refcnt -1 or 0.
1606 			 * We then add DTL_SCRUB with a refcnt of 2, so that
1607 			 * entries in the range [0, scrub_txg) will have a
1608 			 * positive refcnt -- either 1 or 2.  We then convert
1609 			 * the reference tree into the new DTL_MISSING map.
1610 			 */
1611 			space_map_ref_create(&reftree);
1612 			space_map_ref_add_map(&reftree,
1613 			    &vd->vdev_dtl[DTL_MISSING], 1);
1614 			space_map_ref_add_seg(&reftree, 0, scrub_txg, -1);
1615 			space_map_ref_add_map(&reftree,
1616 			    &vd->vdev_dtl[DTL_SCRUB], 2);
1617 			space_map_ref_generate_map(&reftree,
1618 			    &vd->vdev_dtl[DTL_MISSING], 1);
1619 			space_map_ref_destroy(&reftree);
1620 		}
1621 		space_map_vacate(&vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
1622 		space_map_walk(&vd->vdev_dtl[DTL_MISSING],
1623 		    space_map_add, &vd->vdev_dtl[DTL_PARTIAL]);
1624 		if (scrub_done)
1625 			space_map_vacate(&vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
1626 		space_map_vacate(&vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
1627 		if (!vdev_readable(vd))
1628 			space_map_add(&vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
1629 		else
1630 			space_map_walk(&vd->vdev_dtl[DTL_MISSING],
1631 			    space_map_add, &vd->vdev_dtl[DTL_OUTAGE]);
1632 		mutex_exit(&vd->vdev_dtl_lock);
1633 
1634 		if (txg != 0)
1635 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
1636 		return;
1637 	}
1638 
1639 	mutex_enter(&vd->vdev_dtl_lock);
1640 	for (int t = 0; t < DTL_TYPES; t++) {
1641 		/* account for child's outage in parent's missing map */
1642 		int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
1643 		if (t == DTL_SCRUB)
1644 			continue;			/* leaf vdevs only */
1645 		if (t == DTL_PARTIAL)
1646 			minref = 1;			/* i.e. non-zero */
1647 		else if (vd->vdev_nparity != 0)
1648 			minref = vd->vdev_nparity + 1;	/* RAID-Z */
1649 		else
1650 			minref = vd->vdev_children;	/* any kind of mirror */
1651 		space_map_ref_create(&reftree);
1652 		for (int c = 0; c < vd->vdev_children; c++) {
1653 			vdev_t *cvd = vd->vdev_child[c];
1654 			mutex_enter(&cvd->vdev_dtl_lock);
1655 			space_map_ref_add_map(&reftree, &cvd->vdev_dtl[s], 1);
1656 			mutex_exit(&cvd->vdev_dtl_lock);
1657 		}
1658 		space_map_ref_generate_map(&reftree, &vd->vdev_dtl[t], minref);
1659 		space_map_ref_destroy(&reftree);
1660 	}
1661 	mutex_exit(&vd->vdev_dtl_lock);
1662 }
1663 
1664 static int
1665 vdev_dtl_load(vdev_t *vd)
1666 {
1667 	spa_t *spa = vd->vdev_spa;
1668 	space_map_obj_t *smo = &vd->vdev_dtl_smo;
1669 	objset_t *mos = spa->spa_meta_objset;
1670 	dmu_buf_t *db;
1671 	int error;
1672 
1673 	ASSERT(vd->vdev_children == 0);
1674 
1675 	if (smo->smo_object == 0)
1676 		return (0);
1677 
1678 	ASSERT(!vd->vdev_ishole);
1679 
1680 	if ((error = dmu_bonus_hold(mos, smo->smo_object, FTAG, &db)) != 0)
1681 		return (error);
1682 
1683 	ASSERT3U(db->db_size, >=, sizeof (*smo));
1684 	bcopy(db->db_data, smo, sizeof (*smo));
1685 	dmu_buf_rele(db, FTAG);
1686 
1687 	mutex_enter(&vd->vdev_dtl_lock);
1688 	error = space_map_load(&vd->vdev_dtl[DTL_MISSING],
1689 	    NULL, SM_ALLOC, smo, mos);
1690 	mutex_exit(&vd->vdev_dtl_lock);
1691 
1692 	return (error);
1693 }
1694 
1695 void
1696 vdev_dtl_sync(vdev_t *vd, uint64_t txg)
1697 {
1698 	spa_t *spa = vd->vdev_spa;
1699 	space_map_obj_t *smo = &vd->vdev_dtl_smo;
1700 	space_map_t *sm = &vd->vdev_dtl[DTL_MISSING];
1701 	objset_t *mos = spa->spa_meta_objset;
1702 	space_map_t smsync;
1703 	kmutex_t smlock;
1704 	dmu_buf_t *db;
1705 	dmu_tx_t *tx;
1706 
1707 	ASSERT(!vd->vdev_ishole);
1708 
1709 	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1710 
1711 	if (vd->vdev_detached) {
1712 		if (smo->smo_object != 0) {
1713 			int err = dmu_object_free(mos, smo->smo_object, tx);
1714 			ASSERT3U(err, ==, 0);
1715 			smo->smo_object = 0;
1716 		}
1717 		dmu_tx_commit(tx);
1718 		return;
1719 	}
1720 
1721 	if (smo->smo_object == 0) {
1722 		ASSERT(smo->smo_objsize == 0);
1723 		ASSERT(smo->smo_alloc == 0);
1724 		smo->smo_object = dmu_object_alloc(mos,
1725 		    DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
1726 		    DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
1727 		ASSERT(smo->smo_object != 0);
1728 		vdev_config_dirty(vd->vdev_top);
1729 	}
1730 
1731 	mutex_init(&smlock, NULL, MUTEX_DEFAULT, NULL);
1732 
1733 	space_map_create(&smsync, sm->sm_start, sm->sm_size, sm->sm_shift,
1734 	    &smlock);
1735 
1736 	mutex_enter(&smlock);
1737 
1738 	mutex_enter(&vd->vdev_dtl_lock);
1739 	space_map_walk(sm, space_map_add, &smsync);
1740 	mutex_exit(&vd->vdev_dtl_lock);
1741 
1742 	space_map_truncate(smo, mos, tx);
1743 	space_map_sync(&smsync, SM_ALLOC, smo, mos, tx);
1744 
1745 	space_map_destroy(&smsync);
1746 
1747 	mutex_exit(&smlock);
1748 	mutex_destroy(&smlock);
1749 
1750 	VERIFY(0 == dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
1751 	dmu_buf_will_dirty(db, tx);
1752 	ASSERT3U(db->db_size, >=, sizeof (*smo));
1753 	bcopy(smo, db->db_data, sizeof (*smo));
1754 	dmu_buf_rele(db, FTAG);
1755 
1756 	dmu_tx_commit(tx);
1757 }
1758 
1759 /*
1760  * Determine whether the specified vdev can be offlined/detached/removed
1761  * without losing data.
1762  */
1763 boolean_t
1764 vdev_dtl_required(vdev_t *vd)
1765 {
1766 	spa_t *spa = vd->vdev_spa;
1767 	vdev_t *tvd = vd->vdev_top;
1768 	uint8_t cant_read = vd->vdev_cant_read;
1769 	boolean_t required;
1770 
1771 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1772 
1773 	if (vd == spa->spa_root_vdev || vd == tvd)
1774 		return (B_TRUE);
1775 
1776 	/*
1777 	 * Temporarily mark the device as unreadable, and then determine
1778 	 * whether this results in any DTL outages in the top-level vdev.
1779 	 * If not, we can safely offline/detach/remove the device.
1780 	 */
1781 	vd->vdev_cant_read = B_TRUE;
1782 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1783 	required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
1784 	vd->vdev_cant_read = cant_read;
1785 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1786 
1787 	return (required);
1788 }
1789 
1790 /*
1791  * Determine if resilver is needed, and if so the txg range.
1792  */
1793 boolean_t
1794 vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
1795 {
1796 	boolean_t needed = B_FALSE;
1797 	uint64_t thismin = UINT64_MAX;
1798 	uint64_t thismax = 0;
1799 
1800 	if (vd->vdev_children == 0) {
1801 		mutex_enter(&vd->vdev_dtl_lock);
1802 		if (vd->vdev_dtl[DTL_MISSING].sm_space != 0 &&
1803 		    vdev_writeable(vd)) {
1804 			space_seg_t *ss;
1805 
1806 			ss = avl_first(&vd->vdev_dtl[DTL_MISSING].sm_root);
1807 			thismin = ss->ss_start - 1;
1808 			ss = avl_last(&vd->vdev_dtl[DTL_MISSING].sm_root);
1809 			thismax = ss->ss_end;
1810 			needed = B_TRUE;
1811 		}
1812 		mutex_exit(&vd->vdev_dtl_lock);
1813 	} else {
1814 		for (int c = 0; c < vd->vdev_children; c++) {
1815 			vdev_t *cvd = vd->vdev_child[c];
1816 			uint64_t cmin, cmax;
1817 
1818 			if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
1819 				thismin = MIN(thismin, cmin);
1820 				thismax = MAX(thismax, cmax);
1821 				needed = B_TRUE;
1822 			}
1823 		}
1824 	}
1825 
1826 	if (needed && minp) {
1827 		*minp = thismin;
1828 		*maxp = thismax;
1829 	}
1830 	return (needed);
1831 }
1832 
1833 void
1834 vdev_load(vdev_t *vd)
1835 {
1836 	/*
1837 	 * Recursively load all children.
1838 	 */
1839 	for (int c = 0; c < vd->vdev_children; c++)
1840 		vdev_load(vd->vdev_child[c]);
1841 
1842 	/*
1843 	 * If this is a top-level vdev, initialize its metaslabs.
1844 	 */
1845 	if (vd == vd->vdev_top && !vd->vdev_ishole &&
1846 	    (vd->vdev_ashift == 0 || vd->vdev_asize == 0 ||
1847 	    vdev_metaslab_init(vd, 0) != 0))
1848 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1849 		    VDEV_AUX_CORRUPT_DATA);
1850 
1851 	/*
1852 	 * If this is a leaf vdev, load its DTL.
1853 	 */
1854 	if (vd->vdev_ops->vdev_op_leaf && vdev_dtl_load(vd) != 0)
1855 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1856 		    VDEV_AUX_CORRUPT_DATA);
1857 }
1858 
1859 /*
1860  * The special vdev case is used for hot spares and l2cache devices.  Its
1861  * sole purpose it to set the vdev state for the associated vdev.  To do this,
1862  * we make sure that we can open the underlying device, then try to read the
1863  * label, and make sure that the label is sane and that it hasn't been
1864  * repurposed to another pool.
1865  */
1866 int
1867 vdev_validate_aux(vdev_t *vd)
1868 {
1869 	nvlist_t *label;
1870 	uint64_t guid, version;
1871 	uint64_t state;
1872 
1873 	if (!vdev_readable(vd))
1874 		return (0);
1875 
1876 	if ((label = vdev_label_read_config(vd)) == NULL) {
1877 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1878 		    VDEV_AUX_CORRUPT_DATA);
1879 		return (-1);
1880 	}
1881 
1882 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
1883 	    version > SPA_VERSION ||
1884 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
1885 	    guid != vd->vdev_guid ||
1886 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
1887 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1888 		    VDEV_AUX_CORRUPT_DATA);
1889 		nvlist_free(label);
1890 		return (-1);
1891 	}
1892 
1893 	/*
1894 	 * We don't actually check the pool state here.  If it's in fact in
1895 	 * use by another pool, we update this fact on the fly when requested.
1896 	 */
1897 	nvlist_free(label);
1898 	return (0);
1899 }
1900 
1901 void
1902 vdev_remove(vdev_t *vd, uint64_t txg)
1903 {
1904 	spa_t *spa = vd->vdev_spa;
1905 	objset_t *mos = spa->spa_meta_objset;
1906 	dmu_tx_t *tx;
1907 
1908 	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
1909 
1910 	if (vd->vdev_dtl_smo.smo_object) {
1911 		ASSERT3U(vd->vdev_dtl_smo.smo_alloc, ==, 0);
1912 		(void) dmu_object_free(mos, vd->vdev_dtl_smo.smo_object, tx);
1913 		vd->vdev_dtl_smo.smo_object = 0;
1914 	}
1915 
1916 	if (vd->vdev_ms != NULL) {
1917 		for (int m = 0; m < vd->vdev_ms_count; m++) {
1918 			metaslab_t *msp = vd->vdev_ms[m];
1919 
1920 			if (msp == NULL || msp->ms_smo.smo_object == 0)
1921 				continue;
1922 
1923 			ASSERT3U(msp->ms_smo.smo_alloc, ==, 0);
1924 			(void) dmu_object_free(mos, msp->ms_smo.smo_object, tx);
1925 			msp->ms_smo.smo_object = 0;
1926 		}
1927 	}
1928 
1929 	if (vd->vdev_ms_array) {
1930 		(void) dmu_object_free(mos, vd->vdev_ms_array, tx);
1931 		vd->vdev_ms_array = 0;
1932 		vd->vdev_ms_shift = 0;
1933 	}
1934 	dmu_tx_commit(tx);
1935 }
1936 
1937 void
1938 vdev_sync_done(vdev_t *vd, uint64_t txg)
1939 {
1940 	metaslab_t *msp;
1941 
1942 	ASSERT(!vd->vdev_ishole);
1943 
1944 	while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
1945 		metaslab_sync_done(msp, txg);
1946 }
1947 
1948 void
1949 vdev_sync(vdev_t *vd, uint64_t txg)
1950 {
1951 	spa_t *spa = vd->vdev_spa;
1952 	vdev_t *lvd;
1953 	metaslab_t *msp;
1954 	dmu_tx_t *tx;
1955 
1956 	ASSERT(!vd->vdev_ishole);
1957 
1958 	if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0) {
1959 		ASSERT(vd == vd->vdev_top);
1960 		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1961 		vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
1962 		    DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
1963 		ASSERT(vd->vdev_ms_array != 0);
1964 		vdev_config_dirty(vd);
1965 		dmu_tx_commit(tx);
1966 	}
1967 
1968 	if (vd->vdev_removing)
1969 		vdev_remove(vd, txg);
1970 
1971 	while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
1972 		metaslab_sync(msp, txg);
1973 		(void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
1974 	}
1975 
1976 	while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
1977 		vdev_dtl_sync(lvd, txg);
1978 
1979 	(void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
1980 }
1981 
1982 uint64_t
1983 vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
1984 {
1985 	return (vd->vdev_ops->vdev_op_asize(vd, psize));
1986 }
1987 
1988 /*
1989  * Mark the given vdev faulted.  A faulted vdev behaves as if the device could
1990  * not be opened, and no I/O is attempted.
1991  */
1992 int
1993 vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
1994 {
1995 	vdev_t *vd;
1996 
1997 	spa_vdev_state_enter(spa, SCL_NONE);
1998 
1999 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2000 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2001 
2002 	if (!vd->vdev_ops->vdev_op_leaf)
2003 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2004 
2005 	/*
2006 	 * We don't directly use the aux state here, but if we do a
2007 	 * vdev_reopen(), we need this value to be present to remember why we
2008 	 * were faulted.
2009 	 */
2010 	vd->vdev_label_aux = aux;
2011 
2012 	/*
2013 	 * Faulted state takes precedence over degraded.
2014 	 */
2015 	vd->vdev_faulted = 1ULL;
2016 	vd->vdev_degraded = 0ULL;
2017 	vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
2018 
2019 	/*
2020 	 * If marking the vdev as faulted cause the top-level vdev to become
2021 	 * unavailable, then back off and simply mark the vdev as degraded
2022 	 * instead.
2023 	 */
2024 	if (vdev_is_dead(vd->vdev_top) && !vd->vdev_islog &&
2025 	    vd->vdev_aux == NULL) {
2026 		vd->vdev_degraded = 1ULL;
2027 		vd->vdev_faulted = 0ULL;
2028 
2029 		/*
2030 		 * If we reopen the device and it's not dead, only then do we
2031 		 * mark it degraded.
2032 		 */
2033 		vdev_reopen(vd);
2034 
2035 		if (vdev_readable(vd))
2036 			vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
2037 	}
2038 
2039 	return (spa_vdev_state_exit(spa, vd, 0));
2040 }
2041 
2042 /*
2043  * Mark the given vdev degraded.  A degraded vdev is purely an indication to the
2044  * user that something is wrong.  The vdev continues to operate as normal as far
2045  * as I/O is concerned.
2046  */
2047 int
2048 vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2049 {
2050 	vdev_t *vd;
2051 
2052 	spa_vdev_state_enter(spa, SCL_NONE);
2053 
2054 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2055 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2056 
2057 	if (!vd->vdev_ops->vdev_op_leaf)
2058 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2059 
2060 	/*
2061 	 * If the vdev is already faulted, then don't do anything.
2062 	 */
2063 	if (vd->vdev_faulted || vd->vdev_degraded)
2064 		return (spa_vdev_state_exit(spa, NULL, 0));
2065 
2066 	vd->vdev_degraded = 1ULL;
2067 	if (!vdev_is_dead(vd))
2068 		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
2069 		    aux);
2070 
2071 	return (spa_vdev_state_exit(spa, vd, 0));
2072 }
2073 
2074 /*
2075  * Online the given vdev.  If 'unspare' is set, it implies two things.  First,
2076  * any attached spare device should be detached when the device finishes
2077  * resilvering.  Second, the online should be treated like a 'test' online case,
2078  * so no FMA events are generated if the device fails to open.
2079  */
2080 int
2081 vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
2082 {
2083 	vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
2084 
2085 	spa_vdev_state_enter(spa, SCL_NONE);
2086 
2087 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2088 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2089 
2090 	if (!vd->vdev_ops->vdev_op_leaf)
2091 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2092 
2093 	tvd = vd->vdev_top;
2094 	vd->vdev_offline = B_FALSE;
2095 	vd->vdev_tmpoffline = B_FALSE;
2096 	vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
2097 	vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
2098 
2099 	/* XXX - L2ARC 1.0 does not support expansion */
2100 	if (!vd->vdev_aux) {
2101 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2102 			pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
2103 	}
2104 
2105 	vdev_reopen(tvd);
2106 	vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
2107 
2108 	if (!vd->vdev_aux) {
2109 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2110 			pvd->vdev_expanding = B_FALSE;
2111 	}
2112 
2113 	if (newstate)
2114 		*newstate = vd->vdev_state;
2115 	if ((flags & ZFS_ONLINE_UNSPARE) &&
2116 	    !vdev_is_dead(vd) && vd->vdev_parent &&
2117 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2118 	    vd->vdev_parent->vdev_child[0] == vd)
2119 		vd->vdev_unspare = B_TRUE;
2120 
2121 	if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
2122 
2123 		/* XXX - L2ARC 1.0 does not support expansion */
2124 		if (vd->vdev_aux)
2125 			return (spa_vdev_state_exit(spa, vd, ENOTSUP));
2126 		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
2127 	}
2128 	return (spa_vdev_state_exit(spa, vd, 0));
2129 }
2130 
2131 int
2132 vdev_offline_log(spa_t *spa)
2133 {
2134 	int error = 0;
2135 
2136 	if ((error = dmu_objset_find(spa_name(spa), zil_vdev_offline,
2137 	    NULL, DS_FIND_CHILDREN)) == 0) {
2138 
2139 		/*
2140 		 * We successfully offlined the log device, sync out the
2141 		 * current txg so that the "stubby" block can be removed
2142 		 * by zil_sync().
2143 		 */
2144 		txg_wait_synced(spa->spa_dsl_pool, 0);
2145 	}
2146 	return (error);
2147 }
2148 
2149 static int
2150 vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
2151 {
2152 	vdev_t *vd, *tvd;
2153 	int error = 0;
2154 	uint64_t generation;
2155 	metaslab_group_t *mg;
2156 
2157 top:
2158 	spa_vdev_state_enter(spa, SCL_ALLOC);
2159 
2160 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2161 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2162 
2163 	if (!vd->vdev_ops->vdev_op_leaf)
2164 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2165 
2166 	tvd = vd->vdev_top;
2167 	mg = tvd->vdev_mg;
2168 	generation = spa->spa_config_generation + 1;
2169 
2170 	/*
2171 	 * If the device isn't already offline, try to offline it.
2172 	 */
2173 	if (!vd->vdev_offline) {
2174 		/*
2175 		 * If this device has the only valid copy of some data,
2176 		 * don't allow it to be offlined. Log devices are always
2177 		 * expendable.
2178 		 */
2179 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2180 		    vdev_dtl_required(vd))
2181 			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2182 
2183 		/*
2184 		 * If the top-level is a slog and it has had allocations
2185 		 * then proceed.  We check that the vdev's metaslab group
2186 		 * is not NULL since it's possible that we may have just
2187 		 * added this vdev but not yet initialized its metaslabs.
2188 		 */
2189 		if (tvd->vdev_islog && mg != NULL) {
2190 			/*
2191 			 * Prevent any future allocations.
2192 			 */
2193 			metaslab_group_passivate(mg);
2194 			(void) spa_vdev_state_exit(spa, vd, 0);
2195 
2196 			error = vdev_offline_log(spa);
2197 
2198 			spa_vdev_state_enter(spa, SCL_ALLOC);
2199 
2200 			/*
2201 			 * Check to see if the config has changed.
2202 			 */
2203 			if (error || generation != spa->spa_config_generation) {
2204 				metaslab_group_activate(mg);
2205 				if (error)
2206 					return (spa_vdev_state_exit(spa,
2207 					    vd, error));
2208 				(void) spa_vdev_state_exit(spa, vd, 0);
2209 				goto top;
2210 			}
2211 			ASSERT3U(tvd->vdev_stat.vs_alloc, ==, 0);
2212 		}
2213 
2214 		/*
2215 		 * Offline this device and reopen its top-level vdev.
2216 		 * If the top-level vdev is a log device then just offline
2217 		 * it. Otherwise, if this action results in the top-level
2218 		 * vdev becoming unusable, undo it and fail the request.
2219 		 */
2220 		vd->vdev_offline = B_TRUE;
2221 		vdev_reopen(tvd);
2222 
2223 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2224 		    vdev_is_dead(tvd)) {
2225 			vd->vdev_offline = B_FALSE;
2226 			vdev_reopen(tvd);
2227 			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2228 		}
2229 
2230 		/*
2231 		 * Add the device back into the metaslab rotor so that
2232 		 * once we online the device it's open for business.
2233 		 */
2234 		if (tvd->vdev_islog && mg != NULL)
2235 			metaslab_group_activate(mg);
2236 	}
2237 
2238 	vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
2239 
2240 	return (spa_vdev_state_exit(spa, vd, 0));
2241 }
2242 
2243 int
2244 vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
2245 {
2246 	int error;
2247 
2248 	mutex_enter(&spa->spa_vdev_top_lock);
2249 	error = vdev_offline_locked(spa, guid, flags);
2250 	mutex_exit(&spa->spa_vdev_top_lock);
2251 
2252 	return (error);
2253 }
2254 
2255 /*
2256  * Clear the error counts associated with this vdev.  Unlike vdev_online() and
2257  * vdev_offline(), we assume the spa config is locked.  We also clear all
2258  * children.  If 'vd' is NULL, then the user wants to clear all vdevs.
2259  */
2260 void
2261 vdev_clear(spa_t *spa, vdev_t *vd)
2262 {
2263 	vdev_t *rvd = spa->spa_root_vdev;
2264 
2265 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2266 
2267 	if (vd == NULL)
2268 		vd = rvd;
2269 
2270 	vd->vdev_stat.vs_read_errors = 0;
2271 	vd->vdev_stat.vs_write_errors = 0;
2272 	vd->vdev_stat.vs_checksum_errors = 0;
2273 
2274 	for (int c = 0; c < vd->vdev_children; c++)
2275 		vdev_clear(spa, vd->vdev_child[c]);
2276 
2277 	/*
2278 	 * If we're in the FAULTED state or have experienced failed I/O, then
2279 	 * clear the persistent state and attempt to reopen the device.  We
2280 	 * also mark the vdev config dirty, so that the new faulted state is
2281 	 * written out to disk.
2282 	 */
2283 	if (vd->vdev_faulted || vd->vdev_degraded ||
2284 	    !vdev_readable(vd) || !vdev_writeable(vd)) {
2285 
2286 		/*
2287 		 * When reopening in reponse to a clear event, it may be due to
2288 		 * a fmadm repair request.  In this case, if the device is
2289 		 * still broken, we want to still post the ereport again.
2290 		 */
2291 		vd->vdev_forcefault = B_TRUE;
2292 
2293 		vd->vdev_faulted = vd->vdev_degraded = 0;
2294 		vd->vdev_cant_read = B_FALSE;
2295 		vd->vdev_cant_write = B_FALSE;
2296 
2297 		vdev_reopen(vd);
2298 
2299 		vd->vdev_forcefault = B_FALSE;
2300 
2301 		if (vd != rvd)
2302 			vdev_state_dirty(vd->vdev_top);
2303 
2304 		if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
2305 			spa_async_request(spa, SPA_ASYNC_RESILVER);
2306 
2307 		spa_event_notify(spa, vd, ESC_ZFS_VDEV_CLEAR);
2308 	}
2309 
2310 	/*
2311 	 * When clearing a FMA-diagnosed fault, we always want to
2312 	 * unspare the device, as we assume that the original spare was
2313 	 * done in response to the FMA fault.
2314 	 */
2315 	if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
2316 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2317 	    vd->vdev_parent->vdev_child[0] == vd)
2318 		vd->vdev_unspare = B_TRUE;
2319 }
2320 
2321 boolean_t
2322 vdev_is_dead(vdev_t *vd)
2323 {
2324 	/*
2325 	 * Holes and missing devices are always considered "dead".
2326 	 * This simplifies the code since we don't have to check for
2327 	 * these types of devices in the various code paths.
2328 	 * Instead we rely on the fact that we skip over dead devices
2329 	 * before issuing I/O to them.
2330 	 */
2331 	return (vd->vdev_state < VDEV_STATE_DEGRADED || vd->vdev_ishole ||
2332 	    vd->vdev_ops == &vdev_missing_ops);
2333 }
2334 
2335 boolean_t
2336 vdev_readable(vdev_t *vd)
2337 {
2338 	return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
2339 }
2340 
2341 boolean_t
2342 vdev_writeable(vdev_t *vd)
2343 {
2344 	return (!vdev_is_dead(vd) && !vd->vdev_cant_write);
2345 }
2346 
2347 boolean_t
2348 vdev_allocatable(vdev_t *vd)
2349 {
2350 	uint64_t state = vd->vdev_state;
2351 
2352 	/*
2353 	 * We currently allow allocations from vdevs which may be in the
2354 	 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
2355 	 * fails to reopen then we'll catch it later when we're holding
2356 	 * the proper locks.  Note that we have to get the vdev state
2357 	 * in a local variable because although it changes atomically,
2358 	 * we're asking two separate questions about it.
2359 	 */
2360 	return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
2361 	    !vd->vdev_cant_write && !vd->vdev_ishole && !vd->vdev_removing);
2362 }
2363 
2364 boolean_t
2365 vdev_accessible(vdev_t *vd, zio_t *zio)
2366 {
2367 	ASSERT(zio->io_vd == vd);
2368 
2369 	if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
2370 		return (B_FALSE);
2371 
2372 	if (zio->io_type == ZIO_TYPE_READ)
2373 		return (!vd->vdev_cant_read);
2374 
2375 	if (zio->io_type == ZIO_TYPE_WRITE)
2376 		return (!vd->vdev_cant_write);
2377 
2378 	return (B_TRUE);
2379 }
2380 
2381 /*
2382  * Get statistics for the given vdev.
2383  */
2384 void
2385 vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
2386 {
2387 	vdev_t *rvd = vd->vdev_spa->spa_root_vdev;
2388 
2389 	mutex_enter(&vd->vdev_stat_lock);
2390 	bcopy(&vd->vdev_stat, vs, sizeof (*vs));
2391 	vs->vs_scrub_errors = vd->vdev_spa->spa_scrub_errors;
2392 	vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
2393 	vs->vs_state = vd->vdev_state;
2394 	vs->vs_rsize = vdev_get_min_asize(vd);
2395 	if (vd->vdev_ops->vdev_op_leaf)
2396 		vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
2397 	mutex_exit(&vd->vdev_stat_lock);
2398 
2399 	/*
2400 	 * If we're getting stats on the root vdev, aggregate the I/O counts
2401 	 * over all top-level vdevs (i.e. the direct children of the root).
2402 	 */
2403 	if (vd == rvd) {
2404 		for (int c = 0; c < rvd->vdev_children; c++) {
2405 			vdev_t *cvd = rvd->vdev_child[c];
2406 			vdev_stat_t *cvs = &cvd->vdev_stat;
2407 
2408 			mutex_enter(&vd->vdev_stat_lock);
2409 			for (int t = 0; t < ZIO_TYPES; t++) {
2410 				vs->vs_ops[t] += cvs->vs_ops[t];
2411 				vs->vs_bytes[t] += cvs->vs_bytes[t];
2412 			}
2413 			vs->vs_scrub_examined += cvs->vs_scrub_examined;
2414 			mutex_exit(&vd->vdev_stat_lock);
2415 		}
2416 	}
2417 }
2418 
2419 void
2420 vdev_clear_stats(vdev_t *vd)
2421 {
2422 	mutex_enter(&vd->vdev_stat_lock);
2423 	vd->vdev_stat.vs_space = 0;
2424 	vd->vdev_stat.vs_dspace = 0;
2425 	vd->vdev_stat.vs_alloc = 0;
2426 	mutex_exit(&vd->vdev_stat_lock);
2427 }
2428 
2429 void
2430 vdev_stat_update(zio_t *zio, uint64_t psize)
2431 {
2432 	spa_t *spa = zio->io_spa;
2433 	vdev_t *rvd = spa->spa_root_vdev;
2434 	vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
2435 	vdev_t *pvd;
2436 	uint64_t txg = zio->io_txg;
2437 	vdev_stat_t *vs = &vd->vdev_stat;
2438 	zio_type_t type = zio->io_type;
2439 	int flags = zio->io_flags;
2440 
2441 	/*
2442 	 * If this i/o is a gang leader, it didn't do any actual work.
2443 	 */
2444 	if (zio->io_gang_tree)
2445 		return;
2446 
2447 	if (zio->io_error == 0) {
2448 		/*
2449 		 * If this is a root i/o, don't count it -- we've already
2450 		 * counted the top-level vdevs, and vdev_get_stats() will
2451 		 * aggregate them when asked.  This reduces contention on
2452 		 * the root vdev_stat_lock and implicitly handles blocks
2453 		 * that compress away to holes, for which there is no i/o.
2454 		 * (Holes never create vdev children, so all the counters
2455 		 * remain zero, which is what we want.)
2456 		 *
2457 		 * Note: this only applies to successful i/o (io_error == 0)
2458 		 * because unlike i/o counts, errors are not additive.
2459 		 * When reading a ditto block, for example, failure of
2460 		 * one top-level vdev does not imply a root-level error.
2461 		 */
2462 		if (vd == rvd)
2463 			return;
2464 
2465 		ASSERT(vd == zio->io_vd);
2466 
2467 		if (flags & ZIO_FLAG_IO_BYPASS)
2468 			return;
2469 
2470 		mutex_enter(&vd->vdev_stat_lock);
2471 
2472 		if (flags & ZIO_FLAG_IO_REPAIR) {
2473 			if (flags & ZIO_FLAG_SCRUB_THREAD)
2474 				vs->vs_scrub_repaired += psize;
2475 			if (flags & ZIO_FLAG_SELF_HEAL)
2476 				vs->vs_self_healed += psize;
2477 		}
2478 
2479 		vs->vs_ops[type]++;
2480 		vs->vs_bytes[type] += psize;
2481 
2482 		mutex_exit(&vd->vdev_stat_lock);
2483 		return;
2484 	}
2485 
2486 	if (flags & ZIO_FLAG_SPECULATIVE)
2487 		return;
2488 
2489 	/*
2490 	 * If this is an I/O error that is going to be retried, then ignore the
2491 	 * error.  Otherwise, the user may interpret B_FAILFAST I/O errors as
2492 	 * hard errors, when in reality they can happen for any number of
2493 	 * innocuous reasons (bus resets, MPxIO link failure, etc).
2494 	 */
2495 	if (zio->io_error == EIO &&
2496 	    !(zio->io_flags & ZIO_FLAG_IO_RETRY))
2497 		return;
2498 
2499 	/*
2500 	 * Intent logs writes won't propagate their error to the root
2501 	 * I/O so don't mark these types of failures as pool-level
2502 	 * errors.
2503 	 */
2504 	if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
2505 		return;
2506 
2507 	mutex_enter(&vd->vdev_stat_lock);
2508 	if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
2509 		if (zio->io_error == ECKSUM)
2510 			vs->vs_checksum_errors++;
2511 		else
2512 			vs->vs_read_errors++;
2513 	}
2514 	if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
2515 		vs->vs_write_errors++;
2516 	mutex_exit(&vd->vdev_stat_lock);
2517 
2518 	if (type == ZIO_TYPE_WRITE && txg != 0 &&
2519 	    (!(flags & ZIO_FLAG_IO_REPAIR) ||
2520 	    (flags & ZIO_FLAG_SCRUB_THREAD) ||
2521 	    spa->spa_claiming)) {
2522 		/*
2523 		 * This is either a normal write (not a repair), or it's
2524 		 * a repair induced by the scrub thread, or it's a repair
2525 		 * made by zil_claim() during spa_load() in the first txg.
2526 		 * In the normal case, we commit the DTL change in the same
2527 		 * txg as the block was born.  In the scrub-induced repair
2528 		 * case, we know that scrubs run in first-pass syncing context,
2529 		 * so we commit the DTL change in spa_syncing_txg(spa).
2530 		 * In the zil_claim() case, we commit in spa_first_txg(spa).
2531 		 *
2532 		 * We currently do not make DTL entries for failed spontaneous
2533 		 * self-healing writes triggered by normal (non-scrubbing)
2534 		 * reads, because we have no transactional context in which to
2535 		 * do so -- and it's not clear that it'd be desirable anyway.
2536 		 */
2537 		if (vd->vdev_ops->vdev_op_leaf) {
2538 			uint64_t commit_txg = txg;
2539 			if (flags & ZIO_FLAG_SCRUB_THREAD) {
2540 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2541 				ASSERT(spa_sync_pass(spa) == 1);
2542 				vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
2543 				commit_txg = spa_syncing_txg(spa);
2544 			} else if (spa->spa_claiming) {
2545 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2546 				commit_txg = spa_first_txg(spa);
2547 			}
2548 			ASSERT(commit_txg >= spa_syncing_txg(spa));
2549 			if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
2550 				return;
2551 			for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2552 				vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
2553 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
2554 		}
2555 		if (vd != rvd)
2556 			vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
2557 	}
2558 }
2559 
2560 void
2561 vdev_scrub_stat_update(vdev_t *vd, pool_scrub_type_t type, boolean_t complete)
2562 {
2563 	vdev_stat_t *vs = &vd->vdev_stat;
2564 
2565 	for (int c = 0; c < vd->vdev_children; c++)
2566 		vdev_scrub_stat_update(vd->vdev_child[c], type, complete);
2567 
2568 	mutex_enter(&vd->vdev_stat_lock);
2569 
2570 	if (type == POOL_SCRUB_NONE) {
2571 		/*
2572 		 * Update completion and end time.  Leave everything else alone
2573 		 * so we can report what happened during the previous scrub.
2574 		 */
2575 		vs->vs_scrub_complete = complete;
2576 		vs->vs_scrub_end = gethrestime_sec();
2577 	} else {
2578 		vs->vs_scrub_type = type;
2579 		vs->vs_scrub_complete = 0;
2580 		vs->vs_scrub_examined = 0;
2581 		vs->vs_scrub_repaired = 0;
2582 		vs->vs_scrub_start = gethrestime_sec();
2583 		vs->vs_scrub_end = 0;
2584 	}
2585 
2586 	mutex_exit(&vd->vdev_stat_lock);
2587 }
2588 
2589 /*
2590  * Update the in-core space usage stats for this vdev, its metaslab class,
2591  * and the root vdev.
2592  */
2593 void
2594 vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
2595     int64_t space_delta)
2596 {
2597 	int64_t dspace_delta = space_delta;
2598 	spa_t *spa = vd->vdev_spa;
2599 	vdev_t *rvd = spa->spa_root_vdev;
2600 	metaslab_group_t *mg = vd->vdev_mg;
2601 	metaslab_class_t *mc = mg ? mg->mg_class : NULL;
2602 
2603 	ASSERT(vd == vd->vdev_top);
2604 
2605 	/*
2606 	 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
2607 	 * factor.  We must calculate this here and not at the root vdev
2608 	 * because the root vdev's psize-to-asize is simply the max of its
2609 	 * childrens', thus not accurate enough for us.
2610 	 */
2611 	ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
2612 	ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
2613 	dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
2614 	    vd->vdev_deflate_ratio;
2615 
2616 	mutex_enter(&vd->vdev_stat_lock);
2617 	vd->vdev_stat.vs_alloc += alloc_delta;
2618 	vd->vdev_stat.vs_space += space_delta;
2619 	vd->vdev_stat.vs_dspace += dspace_delta;
2620 	mutex_exit(&vd->vdev_stat_lock);
2621 
2622 	if (mc == spa_normal_class(spa)) {
2623 		mutex_enter(&rvd->vdev_stat_lock);
2624 		rvd->vdev_stat.vs_alloc += alloc_delta;
2625 		rvd->vdev_stat.vs_space += space_delta;
2626 		rvd->vdev_stat.vs_dspace += dspace_delta;
2627 		mutex_exit(&rvd->vdev_stat_lock);
2628 	}
2629 
2630 	if (mc != NULL) {
2631 		ASSERT(rvd == vd->vdev_parent);
2632 		ASSERT(vd->vdev_ms_count != 0);
2633 
2634 		metaslab_class_space_update(mc,
2635 		    alloc_delta, defer_delta, space_delta, dspace_delta);
2636 	}
2637 }
2638 
2639 /*
2640  * Mark a top-level vdev's config as dirty, placing it on the dirty list
2641  * so that it will be written out next time the vdev configuration is synced.
2642  * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
2643  */
2644 void
2645 vdev_config_dirty(vdev_t *vd)
2646 {
2647 	spa_t *spa = vd->vdev_spa;
2648 	vdev_t *rvd = spa->spa_root_vdev;
2649 	int c;
2650 
2651 	/*
2652 	 * If this is an aux vdev (as with l2cache and spare devices), then we
2653 	 * update the vdev config manually and set the sync flag.
2654 	 */
2655 	if (vd->vdev_aux != NULL) {
2656 		spa_aux_vdev_t *sav = vd->vdev_aux;
2657 		nvlist_t **aux;
2658 		uint_t naux;
2659 
2660 		for (c = 0; c < sav->sav_count; c++) {
2661 			if (sav->sav_vdevs[c] == vd)
2662 				break;
2663 		}
2664 
2665 		if (c == sav->sav_count) {
2666 			/*
2667 			 * We're being removed.  There's nothing more to do.
2668 			 */
2669 			ASSERT(sav->sav_sync == B_TRUE);
2670 			return;
2671 		}
2672 
2673 		sav->sav_sync = B_TRUE;
2674 
2675 		if (nvlist_lookup_nvlist_array(sav->sav_config,
2676 		    ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
2677 			VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
2678 			    ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
2679 		}
2680 
2681 		ASSERT(c < naux);
2682 
2683 		/*
2684 		 * Setting the nvlist in the middle if the array is a little
2685 		 * sketchy, but it will work.
2686 		 */
2687 		nvlist_free(aux[c]);
2688 		aux[c] = vdev_config_generate(spa, vd, B_TRUE, B_FALSE, B_TRUE);
2689 
2690 		return;
2691 	}
2692 
2693 	/*
2694 	 * The dirty list is protected by the SCL_CONFIG lock.  The caller
2695 	 * must either hold SCL_CONFIG as writer, or must be the sync thread
2696 	 * (which holds SCL_CONFIG as reader).  There's only one sync thread,
2697 	 * so this is sufficient to ensure mutual exclusion.
2698 	 */
2699 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2700 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2701 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
2702 
2703 	if (vd == rvd) {
2704 		for (c = 0; c < rvd->vdev_children; c++)
2705 			vdev_config_dirty(rvd->vdev_child[c]);
2706 	} else {
2707 		ASSERT(vd == vd->vdev_top);
2708 
2709 		if (!list_link_active(&vd->vdev_config_dirty_node) &&
2710 		    !vd->vdev_ishole)
2711 			list_insert_head(&spa->spa_config_dirty_list, vd);
2712 	}
2713 }
2714 
2715 void
2716 vdev_config_clean(vdev_t *vd)
2717 {
2718 	spa_t *spa = vd->vdev_spa;
2719 
2720 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2721 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2722 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
2723 
2724 	ASSERT(list_link_active(&vd->vdev_config_dirty_node));
2725 	list_remove(&spa->spa_config_dirty_list, vd);
2726 }
2727 
2728 /*
2729  * Mark a top-level vdev's state as dirty, so that the next pass of
2730  * spa_sync() can convert this into vdev_config_dirty().  We distinguish
2731  * the state changes from larger config changes because they require
2732  * much less locking, and are often needed for administrative actions.
2733  */
2734 void
2735 vdev_state_dirty(vdev_t *vd)
2736 {
2737 	spa_t *spa = vd->vdev_spa;
2738 
2739 	ASSERT(vd == vd->vdev_top);
2740 
2741 	/*
2742 	 * The state list is protected by the SCL_STATE lock.  The caller
2743 	 * must either hold SCL_STATE as writer, or must be the sync thread
2744 	 * (which holds SCL_STATE as reader).  There's only one sync thread,
2745 	 * so this is sufficient to ensure mutual exclusion.
2746 	 */
2747 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2748 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2749 	    spa_config_held(spa, SCL_STATE, RW_READER)));
2750 
2751 	if (!list_link_active(&vd->vdev_state_dirty_node) && !vd->vdev_ishole)
2752 		list_insert_head(&spa->spa_state_dirty_list, vd);
2753 }
2754 
2755 void
2756 vdev_state_clean(vdev_t *vd)
2757 {
2758 	spa_t *spa = vd->vdev_spa;
2759 
2760 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2761 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2762 	    spa_config_held(spa, SCL_STATE, RW_READER)));
2763 
2764 	ASSERT(list_link_active(&vd->vdev_state_dirty_node));
2765 	list_remove(&spa->spa_state_dirty_list, vd);
2766 }
2767 
2768 /*
2769  * Propagate vdev state up from children to parent.
2770  */
2771 void
2772 vdev_propagate_state(vdev_t *vd)
2773 {
2774 	spa_t *spa = vd->vdev_spa;
2775 	vdev_t *rvd = spa->spa_root_vdev;
2776 	int degraded = 0, faulted = 0;
2777 	int corrupted = 0;
2778 	vdev_t *child;
2779 
2780 	if (vd->vdev_children > 0) {
2781 		for (int c = 0; c < vd->vdev_children; c++) {
2782 			child = vd->vdev_child[c];
2783 
2784 			/*
2785 			 * Don't factor holes into the decision.
2786 			 */
2787 			if (child->vdev_ishole)
2788 				continue;
2789 
2790 			if (!vdev_readable(child) ||
2791 			    (!vdev_writeable(child) && spa_writeable(spa))) {
2792 				/*
2793 				 * Root special: if there is a top-level log
2794 				 * device, treat the root vdev as if it were
2795 				 * degraded.
2796 				 */
2797 				if (child->vdev_islog && vd == rvd)
2798 					degraded++;
2799 				else
2800 					faulted++;
2801 			} else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
2802 				degraded++;
2803 			}
2804 
2805 			if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
2806 				corrupted++;
2807 		}
2808 
2809 		vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
2810 
2811 		/*
2812 		 * Root special: if there is a top-level vdev that cannot be
2813 		 * opened due to corrupted metadata, then propagate the root
2814 		 * vdev's aux state as 'corrupt' rather than 'insufficient
2815 		 * replicas'.
2816 		 */
2817 		if (corrupted && vd == rvd &&
2818 		    rvd->vdev_state == VDEV_STATE_CANT_OPEN)
2819 			vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
2820 			    VDEV_AUX_CORRUPT_DATA);
2821 	}
2822 
2823 	if (vd->vdev_parent)
2824 		vdev_propagate_state(vd->vdev_parent);
2825 }
2826 
2827 /*
2828  * Set a vdev's state.  If this is during an open, we don't update the parent
2829  * state, because we're in the process of opening children depth-first.
2830  * Otherwise, we propagate the change to the parent.
2831  *
2832  * If this routine places a device in a faulted state, an appropriate ereport is
2833  * generated.
2834  */
2835 void
2836 vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
2837 {
2838 	uint64_t save_state;
2839 	spa_t *spa = vd->vdev_spa;
2840 
2841 	if (state == vd->vdev_state) {
2842 		vd->vdev_stat.vs_aux = aux;
2843 		return;
2844 	}
2845 
2846 	save_state = vd->vdev_state;
2847 
2848 	vd->vdev_state = state;
2849 	vd->vdev_stat.vs_aux = aux;
2850 
2851 	/*
2852 	 * If we are setting the vdev state to anything but an open state, then
2853 	 * always close the underlying device.  Otherwise, we keep accessible
2854 	 * but invalid devices open forever.  We don't call vdev_close() itself,
2855 	 * because that implies some extra checks (offline, etc) that we don't
2856 	 * want here.  This is limited to leaf devices, because otherwise
2857 	 * closing the device will affect other children.
2858 	 */
2859 	if (vdev_is_dead(vd) && vd->vdev_ops->vdev_op_leaf)
2860 		vd->vdev_ops->vdev_op_close(vd);
2861 
2862 	/*
2863 	 * If we have brought this vdev back into service, we need
2864 	 * to notify fmd so that it can gracefully repair any outstanding
2865 	 * cases due to a missing device.  We do this in all cases, even those
2866 	 * that probably don't correlate to a repaired fault.  This is sure to
2867 	 * catch all cases, and we let the zfs-retire agent sort it out.  If
2868 	 * this is a transient state it's OK, as the retire agent will
2869 	 * double-check the state of the vdev before repairing it.
2870 	 */
2871 	if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf &&
2872 	    vd->vdev_prevstate != state)
2873 		zfs_post_state_change(spa, vd);
2874 
2875 	if (vd->vdev_removed &&
2876 	    state == VDEV_STATE_CANT_OPEN &&
2877 	    (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
2878 		/*
2879 		 * If the previous state is set to VDEV_STATE_REMOVED, then this
2880 		 * device was previously marked removed and someone attempted to
2881 		 * reopen it.  If this failed due to a nonexistent device, then
2882 		 * keep the device in the REMOVED state.  We also let this be if
2883 		 * it is one of our special test online cases, which is only
2884 		 * attempting to online the device and shouldn't generate an FMA
2885 		 * fault.
2886 		 */
2887 		vd->vdev_state = VDEV_STATE_REMOVED;
2888 		vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
2889 	} else if (state == VDEV_STATE_REMOVED) {
2890 		vd->vdev_removed = B_TRUE;
2891 	} else if (state == VDEV_STATE_CANT_OPEN) {
2892 		/*
2893 		 * If we fail to open a vdev during an import, we mark it as
2894 		 * "not available", which signifies that it was never there to
2895 		 * begin with.  Failure to open such a device is not considered
2896 		 * an error.
2897 		 */
2898 		if (spa->spa_load_state == SPA_LOAD_IMPORT &&
2899 		    vd->vdev_ops->vdev_op_leaf)
2900 			vd->vdev_not_present = 1;
2901 
2902 		/*
2903 		 * Post the appropriate ereport.  If the 'prevstate' field is
2904 		 * set to something other than VDEV_STATE_UNKNOWN, it indicates
2905 		 * that this is part of a vdev_reopen().  In this case, we don't
2906 		 * want to post the ereport if the device was already in the
2907 		 * CANT_OPEN state beforehand.
2908 		 *
2909 		 * If the 'checkremove' flag is set, then this is an attempt to
2910 		 * online the device in response to an insertion event.  If we
2911 		 * hit this case, then we have detected an insertion event for a
2912 		 * faulted or offline device that wasn't in the removed state.
2913 		 * In this scenario, we don't post an ereport because we are
2914 		 * about to replace the device, or attempt an online with
2915 		 * vdev_forcefault, which will generate the fault for us.
2916 		 */
2917 		if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
2918 		    !vd->vdev_not_present && !vd->vdev_checkremove &&
2919 		    vd != spa->spa_root_vdev) {
2920 			const char *class;
2921 
2922 			switch (aux) {
2923 			case VDEV_AUX_OPEN_FAILED:
2924 				class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
2925 				break;
2926 			case VDEV_AUX_CORRUPT_DATA:
2927 				class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
2928 				break;
2929 			case VDEV_AUX_NO_REPLICAS:
2930 				class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
2931 				break;
2932 			case VDEV_AUX_BAD_GUID_SUM:
2933 				class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
2934 				break;
2935 			case VDEV_AUX_TOO_SMALL:
2936 				class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
2937 				break;
2938 			case VDEV_AUX_BAD_LABEL:
2939 				class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
2940 				break;
2941 			case VDEV_AUX_IO_FAILURE:
2942 				class = FM_EREPORT_ZFS_IO_FAILURE;
2943 				break;
2944 			default:
2945 				class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
2946 			}
2947 
2948 			zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
2949 		}
2950 
2951 		/* Erase any notion of persistent removed state */
2952 		vd->vdev_removed = B_FALSE;
2953 	} else {
2954 		vd->vdev_removed = B_FALSE;
2955 	}
2956 
2957 	if (!isopen && vd->vdev_parent)
2958 		vdev_propagate_state(vd->vdev_parent);
2959 }
2960 
2961 /*
2962  * Check the vdev configuration to ensure that it's capable of supporting
2963  * a root pool. Currently, we do not support RAID-Z or partial configuration.
2964  * In addition, only a single top-level vdev is allowed and none of the leaves
2965  * can be wholedisks.
2966  */
2967 boolean_t
2968 vdev_is_bootable(vdev_t *vd)
2969 {
2970 	if (!vd->vdev_ops->vdev_op_leaf) {
2971 		char *vdev_type = vd->vdev_ops->vdev_op_type;
2972 
2973 		if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
2974 		    vd->vdev_children > 1) {
2975 			return (B_FALSE);
2976 		} else if (strcmp(vdev_type, VDEV_TYPE_RAIDZ) == 0 ||
2977 		    strcmp(vdev_type, VDEV_TYPE_MISSING) == 0) {
2978 			return (B_FALSE);
2979 		}
2980 	} else if (vd->vdev_wholedisk == 1) {
2981 		return (B_FALSE);
2982 	}
2983 
2984 	for (int c = 0; c < vd->vdev_children; c++) {
2985 		if (!vdev_is_bootable(vd->vdev_child[c]))
2986 			return (B_FALSE);
2987 	}
2988 	return (B_TRUE);
2989 }
2990 
2991 /*
2992  * Load the state from the original vdev tree (ovd) which
2993  * we've retrieved from the MOS config object. If the original
2994  * vdev was offline then we transfer that state to the device
2995  * in the current vdev tree (nvd).
2996  */
2997 void
2998 vdev_load_log_state(vdev_t *nvd, vdev_t *ovd)
2999 {
3000 	spa_t *spa = nvd->vdev_spa;
3001 
3002 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3003 	ASSERT3U(nvd->vdev_guid, ==, ovd->vdev_guid);
3004 
3005 	for (int c = 0; c < nvd->vdev_children; c++)
3006 		vdev_load_log_state(nvd->vdev_child[c], ovd->vdev_child[c]);
3007 
3008 	if (nvd->vdev_ops->vdev_op_leaf && ovd->vdev_offline) {
3009 		/*
3010 		 * It would be nice to call vdev_offline()
3011 		 * directly but the pool isn't fully loaded and
3012 		 * the txg threads have not been started yet.
3013 		 */
3014 		nvd->vdev_offline = ovd->vdev_offline;
3015 		vdev_reopen(nvd->vdev_top);
3016 	}
3017 }
3018 
3019 /*
3020  * Expand a vdev if possible.
3021  */
3022 void
3023 vdev_expand(vdev_t *vd, uint64_t txg)
3024 {
3025 	ASSERT(vd->vdev_top == vd);
3026 	ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
3027 
3028 	if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count) {
3029 		VERIFY(vdev_metaslab_init(vd, txg) == 0);
3030 		vdev_config_dirty(vd);
3031 	}
3032 }
3033