xref: /illumos-gate/usr/src/uts/common/fs/zfs/vdev.c (revision 99bb17e292915cc12ff98869907cfba546aafdf3)
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 			/*
413 			 * Currently, we can only support 3 parity devices.
414 			 */
415 			if (nparity == 0 || nparity > 3)
416 				return (EINVAL);
417 			/*
418 			 * Previous versions could only support 1 or 2 parity
419 			 * device.
420 			 */
421 			if (nparity > 1 &&
422 			    spa_version(spa) < SPA_VERSION_RAIDZ2)
423 				return (ENOTSUP);
424 			if (nparity > 2 &&
425 			    spa_version(spa) < SPA_VERSION_RAIDZ3)
426 				return (ENOTSUP);
427 		} else {
428 			/*
429 			 * We require the parity to be specified for SPAs that
430 			 * support multiple parity levels.
431 			 */
432 			if (spa_version(spa) >= SPA_VERSION_RAIDZ2)
433 				return (EINVAL);
434 			/*
435 			 * Otherwise, we default to 1 parity device for RAID-Z.
436 			 */
437 			nparity = 1;
438 		}
439 	} else {
440 		nparity = 0;
441 	}
442 	ASSERT(nparity != -1ULL);
443 
444 	vd = vdev_alloc_common(spa, id, guid, ops);
445 
446 	vd->vdev_islog = islog;
447 	vd->vdev_nparity = nparity;
448 
449 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &vd->vdev_path) == 0)
450 		vd->vdev_path = spa_strdup(vd->vdev_path);
451 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &vd->vdev_devid) == 0)
452 		vd->vdev_devid = spa_strdup(vd->vdev_devid);
453 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH,
454 	    &vd->vdev_physpath) == 0)
455 		vd->vdev_physpath = spa_strdup(vd->vdev_physpath);
456 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &vd->vdev_fru) == 0)
457 		vd->vdev_fru = spa_strdup(vd->vdev_fru);
458 
459 	/*
460 	 * Set the whole_disk property.  If it's not specified, leave the value
461 	 * as -1.
462 	 */
463 	if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
464 	    &vd->vdev_wholedisk) != 0)
465 		vd->vdev_wholedisk = -1ULL;
466 
467 	/*
468 	 * Look for the 'not present' flag.  This will only be set if the device
469 	 * was not present at the time of import.
470 	 */
471 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT,
472 	    &vd->vdev_not_present);
473 
474 	/*
475 	 * Get the alignment requirement.
476 	 */
477 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT, &vd->vdev_ashift);
478 
479 	/*
480 	 * Retrieve the vdev creation time.
481 	 */
482 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_CREATE_TXG,
483 	    &vd->vdev_crtxg);
484 
485 	/*
486 	 * If we're a top-level vdev, try to load the allocation parameters.
487 	 */
488 	if (parent && !parent->vdev_parent && alloctype == VDEV_ALLOC_LOAD) {
489 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
490 		    &vd->vdev_ms_array);
491 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
492 		    &vd->vdev_ms_shift);
493 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE,
494 		    &vd->vdev_asize);
495 	}
496 
497 	/*
498 	 * If we're a leaf vdev, try to load the DTL object and other state.
499 	 */
500 	if (vd->vdev_ops->vdev_op_leaf &&
501 	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE ||
502 	    alloctype == VDEV_ALLOC_ROOTPOOL)) {
503 		if (alloctype == VDEV_ALLOC_LOAD) {
504 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL,
505 			    &vd->vdev_dtl_smo.smo_object);
506 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE,
507 			    &vd->vdev_unspare);
508 		}
509 
510 		if (alloctype == VDEV_ALLOC_ROOTPOOL) {
511 			uint64_t spare = 0;
512 
513 			if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE,
514 			    &spare) == 0 && spare)
515 				spa_spare_add(vd);
516 		}
517 
518 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE,
519 		    &vd->vdev_offline);
520 
521 		/*
522 		 * When importing a pool, we want to ignore the persistent fault
523 		 * state, as the diagnosis made on another system may not be
524 		 * valid in the current context.  Local vdevs will
525 		 * remain in the faulted state.
526 		 */
527 		if (spa->spa_load_state == SPA_LOAD_OPEN) {
528 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED,
529 			    &vd->vdev_faulted);
530 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED,
531 			    &vd->vdev_degraded);
532 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED,
533 			    &vd->vdev_removed);
534 
535 			if (vd->vdev_faulted || vd->vdev_degraded) {
536 				char *aux;
537 
538 				vd->vdev_label_aux =
539 				    VDEV_AUX_ERR_EXCEEDED;
540 				if (nvlist_lookup_string(nv,
541 				    ZPOOL_CONFIG_AUX_STATE, &aux) == 0 &&
542 				    strcmp(aux, "external") == 0)
543 					vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
544 			}
545 		}
546 	}
547 
548 	/*
549 	 * Add ourselves to the parent's list of children.
550 	 */
551 	vdev_add_child(parent, vd);
552 
553 	*vdp = vd;
554 
555 	return (0);
556 }
557 
558 void
559 vdev_free(vdev_t *vd)
560 {
561 	spa_t *spa = vd->vdev_spa;
562 
563 	/*
564 	 * vdev_free() implies closing the vdev first.  This is simpler than
565 	 * trying to ensure complicated semantics for all callers.
566 	 */
567 	vdev_close(vd);
568 
569 	ASSERT(!list_link_active(&vd->vdev_config_dirty_node));
570 
571 	/*
572 	 * Free all children.
573 	 */
574 	for (int c = 0; c < vd->vdev_children; c++)
575 		vdev_free(vd->vdev_child[c]);
576 
577 	ASSERT(vd->vdev_child == NULL);
578 	ASSERT(vd->vdev_guid_sum == vd->vdev_guid);
579 
580 	/*
581 	 * Discard allocation state.
582 	 */
583 	if (vd == vd->vdev_top)
584 		vdev_metaslab_fini(vd);
585 
586 	ASSERT3U(vd->vdev_stat.vs_space, ==, 0);
587 	ASSERT3U(vd->vdev_stat.vs_dspace, ==, 0);
588 	ASSERT3U(vd->vdev_stat.vs_alloc, ==, 0);
589 
590 	/*
591 	 * Remove this vdev from its parent's child list.
592 	 */
593 	vdev_remove_child(vd->vdev_parent, vd);
594 
595 	ASSERT(vd->vdev_parent == NULL);
596 
597 	/*
598 	 * Clean up vdev structure.
599 	 */
600 	vdev_queue_fini(vd);
601 	vdev_cache_fini(vd);
602 
603 	if (vd->vdev_path)
604 		spa_strfree(vd->vdev_path);
605 	if (vd->vdev_devid)
606 		spa_strfree(vd->vdev_devid);
607 	if (vd->vdev_physpath)
608 		spa_strfree(vd->vdev_physpath);
609 	if (vd->vdev_fru)
610 		spa_strfree(vd->vdev_fru);
611 
612 	if (vd->vdev_isspare)
613 		spa_spare_remove(vd);
614 	if (vd->vdev_isl2cache)
615 		spa_l2cache_remove(vd);
616 
617 	txg_list_destroy(&vd->vdev_ms_list);
618 	txg_list_destroy(&vd->vdev_dtl_list);
619 
620 	mutex_enter(&vd->vdev_dtl_lock);
621 	for (int t = 0; t < DTL_TYPES; t++) {
622 		space_map_unload(&vd->vdev_dtl[t]);
623 		space_map_destroy(&vd->vdev_dtl[t]);
624 	}
625 	mutex_exit(&vd->vdev_dtl_lock);
626 
627 	mutex_destroy(&vd->vdev_dtl_lock);
628 	mutex_destroy(&vd->vdev_stat_lock);
629 	mutex_destroy(&vd->vdev_probe_lock);
630 
631 	if (vd == spa->spa_root_vdev)
632 		spa->spa_root_vdev = NULL;
633 
634 	kmem_free(vd, sizeof (vdev_t));
635 }
636 
637 /*
638  * Transfer top-level vdev state from svd to tvd.
639  */
640 static void
641 vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
642 {
643 	spa_t *spa = svd->vdev_spa;
644 	metaslab_t *msp;
645 	vdev_t *vd;
646 	int t;
647 
648 	ASSERT(tvd == tvd->vdev_top);
649 
650 	tvd->vdev_ms_array = svd->vdev_ms_array;
651 	tvd->vdev_ms_shift = svd->vdev_ms_shift;
652 	tvd->vdev_ms_count = svd->vdev_ms_count;
653 
654 	svd->vdev_ms_array = 0;
655 	svd->vdev_ms_shift = 0;
656 	svd->vdev_ms_count = 0;
657 
658 	tvd->vdev_mg = svd->vdev_mg;
659 	tvd->vdev_ms = svd->vdev_ms;
660 
661 	svd->vdev_mg = NULL;
662 	svd->vdev_ms = NULL;
663 
664 	if (tvd->vdev_mg != NULL)
665 		tvd->vdev_mg->mg_vd = tvd;
666 
667 	tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
668 	tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
669 	tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
670 
671 	svd->vdev_stat.vs_alloc = 0;
672 	svd->vdev_stat.vs_space = 0;
673 	svd->vdev_stat.vs_dspace = 0;
674 
675 	for (t = 0; t < TXG_SIZE; t++) {
676 		while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
677 			(void) txg_list_add(&tvd->vdev_ms_list, msp, t);
678 		while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
679 			(void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
680 		if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
681 			(void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
682 	}
683 
684 	if (list_link_active(&svd->vdev_config_dirty_node)) {
685 		vdev_config_clean(svd);
686 		vdev_config_dirty(tvd);
687 	}
688 
689 	if (list_link_active(&svd->vdev_state_dirty_node)) {
690 		vdev_state_clean(svd);
691 		vdev_state_dirty(tvd);
692 	}
693 
694 	tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
695 	svd->vdev_deflate_ratio = 0;
696 
697 	tvd->vdev_islog = svd->vdev_islog;
698 	svd->vdev_islog = 0;
699 }
700 
701 static void
702 vdev_top_update(vdev_t *tvd, vdev_t *vd)
703 {
704 	if (vd == NULL)
705 		return;
706 
707 	vd->vdev_top = tvd;
708 
709 	for (int c = 0; c < vd->vdev_children; c++)
710 		vdev_top_update(tvd, vd->vdev_child[c]);
711 }
712 
713 /*
714  * Add a mirror/replacing vdev above an existing vdev.
715  */
716 vdev_t *
717 vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
718 {
719 	spa_t *spa = cvd->vdev_spa;
720 	vdev_t *pvd = cvd->vdev_parent;
721 	vdev_t *mvd;
722 
723 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
724 
725 	mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
726 
727 	mvd->vdev_asize = cvd->vdev_asize;
728 	mvd->vdev_min_asize = cvd->vdev_min_asize;
729 	mvd->vdev_ashift = cvd->vdev_ashift;
730 	mvd->vdev_state = cvd->vdev_state;
731 	mvd->vdev_crtxg = cvd->vdev_crtxg;
732 
733 	vdev_remove_child(pvd, cvd);
734 	vdev_add_child(pvd, mvd);
735 	cvd->vdev_id = mvd->vdev_children;
736 	vdev_add_child(mvd, cvd);
737 	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
738 
739 	if (mvd == mvd->vdev_top)
740 		vdev_top_transfer(cvd, mvd);
741 
742 	return (mvd);
743 }
744 
745 /*
746  * Remove a 1-way mirror/replacing vdev from the tree.
747  */
748 void
749 vdev_remove_parent(vdev_t *cvd)
750 {
751 	vdev_t *mvd = cvd->vdev_parent;
752 	vdev_t *pvd = mvd->vdev_parent;
753 
754 	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
755 
756 	ASSERT(mvd->vdev_children == 1);
757 	ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
758 	    mvd->vdev_ops == &vdev_replacing_ops ||
759 	    mvd->vdev_ops == &vdev_spare_ops);
760 	cvd->vdev_ashift = mvd->vdev_ashift;
761 
762 	vdev_remove_child(mvd, cvd);
763 	vdev_remove_child(pvd, mvd);
764 
765 	/*
766 	 * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
767 	 * Otherwise, we could have detached an offline device, and when we
768 	 * go to import the pool we'll think we have two top-level vdevs,
769 	 * instead of a different version of the same top-level vdev.
770 	 */
771 	if (mvd->vdev_top == mvd) {
772 		uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
773 		cvd->vdev_guid += guid_delta;
774 		cvd->vdev_guid_sum += guid_delta;
775 	}
776 	cvd->vdev_id = mvd->vdev_id;
777 	vdev_add_child(pvd, cvd);
778 	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
779 
780 	if (cvd == cvd->vdev_top)
781 		vdev_top_transfer(mvd, cvd);
782 
783 	ASSERT(mvd->vdev_children == 0);
784 	vdev_free(mvd);
785 }
786 
787 int
788 vdev_metaslab_init(vdev_t *vd, uint64_t txg)
789 {
790 	spa_t *spa = vd->vdev_spa;
791 	objset_t *mos = spa->spa_meta_objset;
792 	metaslab_class_t *mc;
793 	uint64_t m;
794 	uint64_t oldc = vd->vdev_ms_count;
795 	uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
796 	metaslab_t **mspp;
797 	int error;
798 
799 	/*
800 	 * This vdev is not being allocated from yet or is a hole.
801 	 */
802 	if (vd->vdev_ms_shift == 0)
803 		return (0);
804 
805 	ASSERT(!vd->vdev_ishole);
806 
807 	/*
808 	 * Compute the raidz-deflation ratio.  Note, we hard-code
809 	 * in 128k (1 << 17) because it is the current "typical" blocksize.
810 	 * Even if SPA_MAXBLOCKSIZE changes, this algorithm must never change,
811 	 * or we will inconsistently account for existing bp's.
812 	 */
813 	vd->vdev_deflate_ratio = (1 << 17) /
814 	    (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
815 
816 	ASSERT(oldc <= newc);
817 
818 	if (vd->vdev_islog)
819 		mc = spa->spa_log_class;
820 	else
821 		mc = spa->spa_normal_class;
822 
823 	if (vd->vdev_mg == NULL)
824 		vd->vdev_mg = metaslab_group_create(mc, vd);
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 	return (0);
861 }
862 
863 void
864 vdev_metaslab_fini(vdev_t *vd)
865 {
866 	uint64_t m;
867 	uint64_t count = vd->vdev_ms_count;
868 
869 	if (vd->vdev_ms != NULL) {
870 		for (m = 0; m < count; m++)
871 			if (vd->vdev_ms[m] != NULL)
872 				metaslab_fini(vd->vdev_ms[m]);
873 		kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
874 		vd->vdev_ms = NULL;
875 	}
876 }
877 
878 typedef struct vdev_probe_stats {
879 	boolean_t	vps_readable;
880 	boolean_t	vps_writeable;
881 	int		vps_flags;
882 } vdev_probe_stats_t;
883 
884 static void
885 vdev_probe_done(zio_t *zio)
886 {
887 	spa_t *spa = zio->io_spa;
888 	vdev_t *vd = zio->io_vd;
889 	vdev_probe_stats_t *vps = zio->io_private;
890 
891 	ASSERT(vd->vdev_probe_zio != NULL);
892 
893 	if (zio->io_type == ZIO_TYPE_READ) {
894 		if (zio->io_error == 0)
895 			vps->vps_readable = 1;
896 		if (zio->io_error == 0 && spa_writeable(spa)) {
897 			zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
898 			    zio->io_offset, zio->io_size, zio->io_data,
899 			    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
900 			    ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
901 		} else {
902 			zio_buf_free(zio->io_data, zio->io_size);
903 		}
904 	} else if (zio->io_type == ZIO_TYPE_WRITE) {
905 		if (zio->io_error == 0)
906 			vps->vps_writeable = 1;
907 		zio_buf_free(zio->io_data, zio->io_size);
908 	} else if (zio->io_type == ZIO_TYPE_NULL) {
909 		zio_t *pio;
910 
911 		vd->vdev_cant_read |= !vps->vps_readable;
912 		vd->vdev_cant_write |= !vps->vps_writeable;
913 
914 		if (vdev_readable(vd) &&
915 		    (vdev_writeable(vd) || !spa_writeable(spa))) {
916 			zio->io_error = 0;
917 		} else {
918 			ASSERT(zio->io_error != 0);
919 			zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
920 			    spa, vd, NULL, 0, 0);
921 			zio->io_error = ENXIO;
922 		}
923 
924 		mutex_enter(&vd->vdev_probe_lock);
925 		ASSERT(vd->vdev_probe_zio == zio);
926 		vd->vdev_probe_zio = NULL;
927 		mutex_exit(&vd->vdev_probe_lock);
928 
929 		while ((pio = zio_walk_parents(zio)) != NULL)
930 			if (!vdev_accessible(vd, pio))
931 				pio->io_error = ENXIO;
932 
933 		kmem_free(vps, sizeof (*vps));
934 	}
935 }
936 
937 /*
938  * Determine whether this device is accessible by reading and writing
939  * to several known locations: the pad regions of each vdev label
940  * but the first (which we leave alone in case it contains a VTOC).
941  */
942 zio_t *
943 vdev_probe(vdev_t *vd, zio_t *zio)
944 {
945 	spa_t *spa = vd->vdev_spa;
946 	vdev_probe_stats_t *vps = NULL;
947 	zio_t *pio;
948 
949 	ASSERT(vd->vdev_ops->vdev_op_leaf);
950 
951 	/*
952 	 * Don't probe the probe.
953 	 */
954 	if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
955 		return (NULL);
956 
957 	/*
958 	 * To prevent 'probe storms' when a device fails, we create
959 	 * just one probe i/o at a time.  All zios that want to probe
960 	 * this vdev will become parents of the probe io.
961 	 */
962 	mutex_enter(&vd->vdev_probe_lock);
963 
964 	if ((pio = vd->vdev_probe_zio) == NULL) {
965 		vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
966 
967 		vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
968 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
969 		    ZIO_FLAG_TRYHARD;
970 
971 		if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
972 			/*
973 			 * vdev_cant_read and vdev_cant_write can only
974 			 * transition from TRUE to FALSE when we have the
975 			 * SCL_ZIO lock as writer; otherwise they can only
976 			 * transition from FALSE to TRUE.  This ensures that
977 			 * any zio looking at these values can assume that
978 			 * failures persist for the life of the I/O.  That's
979 			 * important because when a device has intermittent
980 			 * connectivity problems, we want to ensure that
981 			 * they're ascribed to the device (ENXIO) and not
982 			 * the zio (EIO).
983 			 *
984 			 * Since we hold SCL_ZIO as writer here, clear both
985 			 * values so the probe can reevaluate from first
986 			 * principles.
987 			 */
988 			vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
989 			vd->vdev_cant_read = B_FALSE;
990 			vd->vdev_cant_write = B_FALSE;
991 		}
992 
993 		vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
994 		    vdev_probe_done, vps,
995 		    vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
996 
997 		if (zio != NULL) {
998 			vd->vdev_probe_wanted = B_TRUE;
999 			spa_async_request(spa, SPA_ASYNC_PROBE);
1000 		}
1001 	}
1002 
1003 	if (zio != NULL)
1004 		zio_add_child(zio, pio);
1005 
1006 	mutex_exit(&vd->vdev_probe_lock);
1007 
1008 	if (vps == NULL) {
1009 		ASSERT(zio != NULL);
1010 		return (NULL);
1011 	}
1012 
1013 	for (int l = 1; l < VDEV_LABELS; l++) {
1014 		zio_nowait(zio_read_phys(pio, vd,
1015 		    vdev_label_offset(vd->vdev_psize, l,
1016 		    offsetof(vdev_label_t, vl_pad2)),
1017 		    VDEV_PAD_SIZE, zio_buf_alloc(VDEV_PAD_SIZE),
1018 		    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1019 		    ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1020 	}
1021 
1022 	if (zio == NULL)
1023 		return (pio);
1024 
1025 	zio_nowait(pio);
1026 	return (NULL);
1027 }
1028 
1029 static void
1030 vdev_open_child(void *arg)
1031 {
1032 	vdev_t *vd = arg;
1033 
1034 	vd->vdev_open_thread = curthread;
1035 	vd->vdev_open_error = vdev_open(vd);
1036 	vd->vdev_open_thread = NULL;
1037 }
1038 
1039 boolean_t
1040 vdev_uses_zvols(vdev_t *vd)
1041 {
1042 	if (vd->vdev_path && strncmp(vd->vdev_path, ZVOL_DIR,
1043 	    strlen(ZVOL_DIR)) == 0)
1044 		return (B_TRUE);
1045 	for (int c = 0; c < vd->vdev_children; c++)
1046 		if (vdev_uses_zvols(vd->vdev_child[c]))
1047 			return (B_TRUE);
1048 	return (B_FALSE);
1049 }
1050 
1051 void
1052 vdev_open_children(vdev_t *vd)
1053 {
1054 	taskq_t *tq;
1055 	int children = vd->vdev_children;
1056 
1057 	/*
1058 	 * in order to handle pools on top of zvols, do the opens
1059 	 * in a single thread so that the same thread holds the
1060 	 * spa_namespace_lock
1061 	 */
1062 	if (vdev_uses_zvols(vd)) {
1063 		for (int c = 0; c < children; c++)
1064 			vd->vdev_child[c]->vdev_open_error =
1065 			    vdev_open(vd->vdev_child[c]);
1066 		return;
1067 	}
1068 	tq = taskq_create("vdev_open", children, minclsyspri,
1069 	    children, children, TASKQ_PREPOPULATE);
1070 
1071 	for (int c = 0; c < children; c++)
1072 		VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c],
1073 		    TQ_SLEEP) != NULL);
1074 
1075 	taskq_destroy(tq);
1076 }
1077 
1078 /*
1079  * Prepare a virtual device for access.
1080  */
1081 int
1082 vdev_open(vdev_t *vd)
1083 {
1084 	spa_t *spa = vd->vdev_spa;
1085 	int error;
1086 	uint64_t osize = 0;
1087 	uint64_t asize, psize;
1088 	uint64_t ashift = 0;
1089 
1090 	ASSERT(vd->vdev_open_thread == curthread ||
1091 	    spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1092 	ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1093 	    vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1094 	    vd->vdev_state == VDEV_STATE_OFFLINE);
1095 
1096 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1097 	vd->vdev_cant_read = B_FALSE;
1098 	vd->vdev_cant_write = B_FALSE;
1099 	vd->vdev_min_asize = vdev_get_min_asize(vd);
1100 
1101 	/*
1102 	 * If this vdev is not removed, check its fault status.  If it's
1103 	 * faulted, bail out of the open.
1104 	 */
1105 	if (!vd->vdev_removed && vd->vdev_faulted) {
1106 		ASSERT(vd->vdev_children == 0);
1107 		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1108 		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1109 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1110 		    vd->vdev_label_aux);
1111 		return (ENXIO);
1112 	} else if (vd->vdev_offline) {
1113 		ASSERT(vd->vdev_children == 0);
1114 		vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1115 		return (ENXIO);
1116 	}
1117 
1118 	error = vd->vdev_ops->vdev_op_open(vd, &osize, &ashift);
1119 
1120 	/*
1121 	 * Reset the vdev_reopening flag so that we actually close
1122 	 * the vdev on error.
1123 	 */
1124 	vd->vdev_reopening = B_FALSE;
1125 	if (zio_injection_enabled && error == 0)
1126 		error = zio_handle_device_injection(vd, NULL, ENXIO);
1127 
1128 	if (error) {
1129 		if (vd->vdev_removed &&
1130 		    vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1131 			vd->vdev_removed = B_FALSE;
1132 
1133 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1134 		    vd->vdev_stat.vs_aux);
1135 		return (error);
1136 	}
1137 
1138 	vd->vdev_removed = B_FALSE;
1139 
1140 	/*
1141 	 * Recheck the faulted flag now that we have confirmed that
1142 	 * the vdev is accessible.  If we're faulted, bail.
1143 	 */
1144 	if (vd->vdev_faulted) {
1145 		ASSERT(vd->vdev_children == 0);
1146 		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1147 		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1148 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1149 		    vd->vdev_label_aux);
1150 		return (ENXIO);
1151 	}
1152 
1153 	if (vd->vdev_degraded) {
1154 		ASSERT(vd->vdev_children == 0);
1155 		vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1156 		    VDEV_AUX_ERR_EXCEEDED);
1157 	} else {
1158 		vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1159 	}
1160 
1161 	/*
1162 	 * For hole or missing vdevs we just return success.
1163 	 */
1164 	if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1165 		return (0);
1166 
1167 	for (int c = 0; c < vd->vdev_children; c++) {
1168 		if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1169 			vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1170 			    VDEV_AUX_NONE);
1171 			break;
1172 		}
1173 	}
1174 
1175 	osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1176 
1177 	if (vd->vdev_children == 0) {
1178 		if (osize < SPA_MINDEVSIZE) {
1179 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1180 			    VDEV_AUX_TOO_SMALL);
1181 			return (EOVERFLOW);
1182 		}
1183 		psize = osize;
1184 		asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1185 	} else {
1186 		if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1187 		    (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1188 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1189 			    VDEV_AUX_TOO_SMALL);
1190 			return (EOVERFLOW);
1191 		}
1192 		psize = 0;
1193 		asize = osize;
1194 	}
1195 
1196 	vd->vdev_psize = psize;
1197 
1198 	/*
1199 	 * Make sure the allocatable size hasn't shrunk.
1200 	 */
1201 	if (asize < vd->vdev_min_asize) {
1202 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1203 		    VDEV_AUX_BAD_LABEL);
1204 		return (EINVAL);
1205 	}
1206 
1207 	if (vd->vdev_asize == 0) {
1208 		/*
1209 		 * This is the first-ever open, so use the computed values.
1210 		 * For testing purposes, a higher ashift can be requested.
1211 		 */
1212 		vd->vdev_asize = asize;
1213 		vd->vdev_ashift = MAX(ashift, vd->vdev_ashift);
1214 	} else {
1215 		/*
1216 		 * Make sure the alignment requirement hasn't increased.
1217 		 */
1218 		if (ashift > vd->vdev_top->vdev_ashift) {
1219 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1220 			    VDEV_AUX_BAD_LABEL);
1221 			return (EINVAL);
1222 		}
1223 	}
1224 
1225 	/*
1226 	 * If all children are healthy and the asize has increased,
1227 	 * then we've experienced dynamic LUN growth.  If automatic
1228 	 * expansion is enabled then use the additional space.
1229 	 */
1230 	if (vd->vdev_state == VDEV_STATE_HEALTHY && asize > vd->vdev_asize &&
1231 	    (vd->vdev_expanding || spa->spa_autoexpand))
1232 		vd->vdev_asize = asize;
1233 
1234 	vdev_set_min_asize(vd);
1235 
1236 	/*
1237 	 * Ensure we can issue some IO before declaring the
1238 	 * vdev open for business.
1239 	 */
1240 	if (vd->vdev_ops->vdev_op_leaf &&
1241 	    (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1242 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1243 		    VDEV_AUX_IO_FAILURE);
1244 		return (error);
1245 	}
1246 
1247 	/*
1248 	 * If a leaf vdev has a DTL, and seems healthy, then kick off a
1249 	 * resilver.  But don't do this if we are doing a reopen for a scrub,
1250 	 * since this would just restart the scrub we are already doing.
1251 	 */
1252 	if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1253 	    vdev_resilver_needed(vd, NULL, NULL))
1254 		spa_async_request(spa, SPA_ASYNC_RESILVER);
1255 
1256 	return (0);
1257 }
1258 
1259 /*
1260  * Called once the vdevs are all opened, this routine validates the label
1261  * contents.  This needs to be done before vdev_load() so that we don't
1262  * inadvertently do repair I/Os to the wrong device.
1263  *
1264  * This function will only return failure if one of the vdevs indicates that it
1265  * has since been destroyed or exported.  This is only possible if
1266  * /etc/zfs/zpool.cache was readonly at the time.  Otherwise, the vdev state
1267  * will be updated but the function will return 0.
1268  */
1269 int
1270 vdev_validate(vdev_t *vd)
1271 {
1272 	spa_t *spa = vd->vdev_spa;
1273 	nvlist_t *label;
1274 	uint64_t guid, top_guid;
1275 	uint64_t state;
1276 
1277 	for (int c = 0; c < vd->vdev_children; c++)
1278 		if (vdev_validate(vd->vdev_child[c]) != 0)
1279 			return (EBADF);
1280 
1281 	/*
1282 	 * If the device has already failed, or was marked offline, don't do
1283 	 * any further validation.  Otherwise, label I/O will fail and we will
1284 	 * overwrite the previous state.
1285 	 */
1286 	if (vd->vdev_ops->vdev_op_leaf && vdev_readable(vd)) {
1287 
1288 		if ((label = vdev_label_read_config(vd)) == NULL) {
1289 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1290 			    VDEV_AUX_BAD_LABEL);
1291 			return (0);
1292 		}
1293 
1294 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_GUID,
1295 		    &guid) != 0 || guid != spa_guid(spa)) {
1296 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1297 			    VDEV_AUX_CORRUPT_DATA);
1298 			nvlist_free(label);
1299 			return (0);
1300 		}
1301 
1302 		/*
1303 		 * If this vdev just became a top-level vdev because its
1304 		 * sibling was detached, it will have adopted the parent's
1305 		 * vdev guid -- but the label may or may not be on disk yet.
1306 		 * Fortunately, either version of the label will have the
1307 		 * same top guid, so if we're a top-level vdev, we can
1308 		 * safely compare to that instead.
1309 		 */
1310 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID,
1311 		    &guid) != 0 ||
1312 		    nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID,
1313 		    &top_guid) != 0 ||
1314 		    (vd->vdev_guid != guid &&
1315 		    (vd->vdev_guid != top_guid || vd != vd->vdev_top))) {
1316 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1317 			    VDEV_AUX_CORRUPT_DATA);
1318 			nvlist_free(label);
1319 			return (0);
1320 		}
1321 
1322 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1323 		    &state) != 0) {
1324 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1325 			    VDEV_AUX_CORRUPT_DATA);
1326 			nvlist_free(label);
1327 			return (0);
1328 		}
1329 
1330 		nvlist_free(label);
1331 
1332 		/*
1333 		 * If spa->spa_load_verbatim is true, no need to check the
1334 		 * state of the pool.
1335 		 */
1336 		if (!spa->spa_load_verbatim &&
1337 		    spa->spa_load_state == SPA_LOAD_OPEN &&
1338 		    state != POOL_STATE_ACTIVE)
1339 			return (EBADF);
1340 
1341 		/*
1342 		 * If we were able to open and validate a vdev that was
1343 		 * previously marked permanently unavailable, clear that state
1344 		 * now.
1345 		 */
1346 		if (vd->vdev_not_present)
1347 			vd->vdev_not_present = 0;
1348 	}
1349 
1350 	return (0);
1351 }
1352 
1353 /*
1354  * Close a virtual device.
1355  */
1356 void
1357 vdev_close(vdev_t *vd)
1358 {
1359 	spa_t *spa = vd->vdev_spa;
1360 	vdev_t *pvd = vd->vdev_parent;
1361 
1362 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1363 
1364 	if (pvd != NULL && pvd->vdev_reopening)
1365 		vd->vdev_reopening = pvd->vdev_reopening;
1366 
1367 	vd->vdev_ops->vdev_op_close(vd);
1368 
1369 	vdev_cache_purge(vd);
1370 
1371 	/*
1372 	 * We record the previous state before we close it, so that if we are
1373 	 * doing a reopen(), we don't generate FMA ereports if we notice that
1374 	 * it's still faulted.
1375 	 */
1376 	vd->vdev_prevstate = vd->vdev_state;
1377 
1378 	if (vd->vdev_offline)
1379 		vd->vdev_state = VDEV_STATE_OFFLINE;
1380 	else
1381 		vd->vdev_state = VDEV_STATE_CLOSED;
1382 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1383 }
1384 
1385 /*
1386  * Reopen all interior vdevs and any unopened leaves.  We don't actually
1387  * reopen leaf vdevs which had previously been opened as they might deadlock
1388  * on the spa_config_lock.  Instead we only obtain the leaf's physical size.
1389  * If the leaf has never been opened then open it, as usual.
1390  */
1391 void
1392 vdev_reopen(vdev_t *vd)
1393 {
1394 	spa_t *spa = vd->vdev_spa;
1395 
1396 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1397 
1398 	vd->vdev_reopening = B_TRUE;
1399 	vdev_close(vd);
1400 	(void) vdev_open(vd);
1401 
1402 	/*
1403 	 * Call vdev_validate() here to make sure we have the same device.
1404 	 * Otherwise, a device with an invalid label could be successfully
1405 	 * opened in response to vdev_reopen().
1406 	 */
1407 	if (vd->vdev_aux) {
1408 		(void) vdev_validate_aux(vd);
1409 		if (vdev_readable(vd) && vdev_writeable(vd) &&
1410 		    vd->vdev_aux == &spa->spa_l2cache &&
1411 		    !l2arc_vdev_present(vd))
1412 			l2arc_add_vdev(spa, vd);
1413 	} else {
1414 		(void) vdev_validate(vd);
1415 	}
1416 
1417 	/*
1418 	 * Reassess parent vdev's health.
1419 	 */
1420 	vdev_propagate_state(vd);
1421 }
1422 
1423 int
1424 vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1425 {
1426 	int error;
1427 
1428 	/*
1429 	 * Normally, partial opens (e.g. of a mirror) are allowed.
1430 	 * For a create, however, we want to fail the request if
1431 	 * there are any components we can't open.
1432 	 */
1433 	error = vdev_open(vd);
1434 
1435 	if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
1436 		vdev_close(vd);
1437 		return (error ? error : ENXIO);
1438 	}
1439 
1440 	/*
1441 	 * Recursively initialize all labels.
1442 	 */
1443 	if ((error = vdev_label_init(vd, txg, isreplacing ?
1444 	    VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
1445 		vdev_close(vd);
1446 		return (error);
1447 	}
1448 
1449 	return (0);
1450 }
1451 
1452 void
1453 vdev_metaslab_set_size(vdev_t *vd)
1454 {
1455 	/*
1456 	 * Aim for roughly 200 metaslabs per vdev.
1457 	 */
1458 	vd->vdev_ms_shift = highbit(vd->vdev_asize / 200);
1459 	vd->vdev_ms_shift = MAX(vd->vdev_ms_shift, SPA_MAXBLOCKSHIFT);
1460 }
1461 
1462 void
1463 vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
1464 {
1465 	ASSERT(vd == vd->vdev_top);
1466 	ASSERT(!vd->vdev_ishole);
1467 	ASSERT(ISP2(flags));
1468 
1469 	if (flags & VDD_METASLAB)
1470 		(void) txg_list_add(&vd->vdev_ms_list, arg, txg);
1471 
1472 	if (flags & VDD_DTL)
1473 		(void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
1474 
1475 	(void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
1476 }
1477 
1478 /*
1479  * DTLs.
1480  *
1481  * A vdev's DTL (dirty time log) is the set of transaction groups for which
1482  * the vdev has less than perfect replication.  There are three kinds of DTL:
1483  *
1484  * DTL_MISSING: txgs for which the vdev has no valid copies of the data
1485  *
1486  * DTL_PARTIAL: txgs for which data is available, but not fully replicated
1487  *
1488  * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
1489  *	scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
1490  *	txgs that was scrubbed.
1491  *
1492  * DTL_OUTAGE: txgs which cannot currently be read, whether due to
1493  *	persistent errors or just some device being offline.
1494  *	Unlike the other three, the DTL_OUTAGE map is not generally
1495  *	maintained; it's only computed when needed, typically to
1496  *	determine whether a device can be detached.
1497  *
1498  * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
1499  * either has the data or it doesn't.
1500  *
1501  * For interior vdevs such as mirror and RAID-Z the picture is more complex.
1502  * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
1503  * if any child is less than fully replicated, then so is its parent.
1504  * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
1505  * comprising only those txgs which appear in 'maxfaults' or more children;
1506  * those are the txgs we don't have enough replication to read.  For example,
1507  * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
1508  * thus, its DTL_MISSING consists of the set of txgs that appear in more than
1509  * two child DTL_MISSING maps.
1510  *
1511  * It should be clear from the above that to compute the DTLs and outage maps
1512  * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
1513  * Therefore, that is all we keep on disk.  When loading the pool, or after
1514  * a configuration change, we generate all other DTLs from first principles.
1515  */
1516 void
1517 vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1518 {
1519 	space_map_t *sm = &vd->vdev_dtl[t];
1520 
1521 	ASSERT(t < DTL_TYPES);
1522 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1523 
1524 	mutex_enter(sm->sm_lock);
1525 	if (!space_map_contains(sm, txg, size))
1526 		space_map_add(sm, txg, size);
1527 	mutex_exit(sm->sm_lock);
1528 }
1529 
1530 boolean_t
1531 vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1532 {
1533 	space_map_t *sm = &vd->vdev_dtl[t];
1534 	boolean_t dirty = B_FALSE;
1535 
1536 	ASSERT(t < DTL_TYPES);
1537 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1538 
1539 	mutex_enter(sm->sm_lock);
1540 	if (sm->sm_space != 0)
1541 		dirty = space_map_contains(sm, txg, size);
1542 	mutex_exit(sm->sm_lock);
1543 
1544 	return (dirty);
1545 }
1546 
1547 boolean_t
1548 vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
1549 {
1550 	space_map_t *sm = &vd->vdev_dtl[t];
1551 	boolean_t empty;
1552 
1553 	mutex_enter(sm->sm_lock);
1554 	empty = (sm->sm_space == 0);
1555 	mutex_exit(sm->sm_lock);
1556 
1557 	return (empty);
1558 }
1559 
1560 /*
1561  * Reassess DTLs after a config change or scrub completion.
1562  */
1563 void
1564 vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
1565 {
1566 	spa_t *spa = vd->vdev_spa;
1567 	avl_tree_t reftree;
1568 	int minref;
1569 
1570 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
1571 
1572 	for (int c = 0; c < vd->vdev_children; c++)
1573 		vdev_dtl_reassess(vd->vdev_child[c], txg,
1574 		    scrub_txg, scrub_done);
1575 
1576 	if (vd == spa->spa_root_vdev || vd->vdev_ishole)
1577 		return;
1578 
1579 	if (vd->vdev_ops->vdev_op_leaf) {
1580 		mutex_enter(&vd->vdev_dtl_lock);
1581 		if (scrub_txg != 0 &&
1582 		    (spa->spa_scrub_started || spa->spa_scrub_errors == 0)) {
1583 			/* XXX should check scrub_done? */
1584 			/*
1585 			 * We completed a scrub up to scrub_txg.  If we
1586 			 * did it without rebooting, then the scrub dtl
1587 			 * will be valid, so excise the old region and
1588 			 * fold in the scrub dtl.  Otherwise, leave the
1589 			 * dtl as-is if there was an error.
1590 			 *
1591 			 * There's little trick here: to excise the beginning
1592 			 * of the DTL_MISSING map, we put it into a reference
1593 			 * tree and then add a segment with refcnt -1 that
1594 			 * covers the range [0, scrub_txg).  This means
1595 			 * that each txg in that range has refcnt -1 or 0.
1596 			 * We then add DTL_SCRUB with a refcnt of 2, so that
1597 			 * entries in the range [0, scrub_txg) will have a
1598 			 * positive refcnt -- either 1 or 2.  We then convert
1599 			 * the reference tree into the new DTL_MISSING map.
1600 			 */
1601 			space_map_ref_create(&reftree);
1602 			space_map_ref_add_map(&reftree,
1603 			    &vd->vdev_dtl[DTL_MISSING], 1);
1604 			space_map_ref_add_seg(&reftree, 0, scrub_txg, -1);
1605 			space_map_ref_add_map(&reftree,
1606 			    &vd->vdev_dtl[DTL_SCRUB], 2);
1607 			space_map_ref_generate_map(&reftree,
1608 			    &vd->vdev_dtl[DTL_MISSING], 1);
1609 			space_map_ref_destroy(&reftree);
1610 		}
1611 		space_map_vacate(&vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
1612 		space_map_walk(&vd->vdev_dtl[DTL_MISSING],
1613 		    space_map_add, &vd->vdev_dtl[DTL_PARTIAL]);
1614 		if (scrub_done)
1615 			space_map_vacate(&vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
1616 		space_map_vacate(&vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
1617 		if (!vdev_readable(vd))
1618 			space_map_add(&vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
1619 		else
1620 			space_map_walk(&vd->vdev_dtl[DTL_MISSING],
1621 			    space_map_add, &vd->vdev_dtl[DTL_OUTAGE]);
1622 		mutex_exit(&vd->vdev_dtl_lock);
1623 
1624 		if (txg != 0)
1625 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
1626 		return;
1627 	}
1628 
1629 	mutex_enter(&vd->vdev_dtl_lock);
1630 	for (int t = 0; t < DTL_TYPES; t++) {
1631 		/* account for child's outage in parent's missing map */
1632 		int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
1633 		if (t == DTL_SCRUB)
1634 			continue;			/* leaf vdevs only */
1635 		if (t == DTL_PARTIAL)
1636 			minref = 1;			/* i.e. non-zero */
1637 		else if (vd->vdev_nparity != 0)
1638 			minref = vd->vdev_nparity + 1;	/* RAID-Z */
1639 		else
1640 			minref = vd->vdev_children;	/* any kind of mirror */
1641 		space_map_ref_create(&reftree);
1642 		for (int c = 0; c < vd->vdev_children; c++) {
1643 			vdev_t *cvd = vd->vdev_child[c];
1644 			mutex_enter(&cvd->vdev_dtl_lock);
1645 			space_map_ref_add_map(&reftree, &cvd->vdev_dtl[s], 1);
1646 			mutex_exit(&cvd->vdev_dtl_lock);
1647 		}
1648 		space_map_ref_generate_map(&reftree, &vd->vdev_dtl[t], minref);
1649 		space_map_ref_destroy(&reftree);
1650 	}
1651 	mutex_exit(&vd->vdev_dtl_lock);
1652 }
1653 
1654 static int
1655 vdev_dtl_load(vdev_t *vd)
1656 {
1657 	spa_t *spa = vd->vdev_spa;
1658 	space_map_obj_t *smo = &vd->vdev_dtl_smo;
1659 	objset_t *mos = spa->spa_meta_objset;
1660 	dmu_buf_t *db;
1661 	int error;
1662 
1663 	ASSERT(vd->vdev_children == 0);
1664 
1665 	if (smo->smo_object == 0)
1666 		return (0);
1667 
1668 	ASSERT(!vd->vdev_ishole);
1669 
1670 	if ((error = dmu_bonus_hold(mos, smo->smo_object, FTAG, &db)) != 0)
1671 		return (error);
1672 
1673 	ASSERT3U(db->db_size, >=, sizeof (*smo));
1674 	bcopy(db->db_data, smo, sizeof (*smo));
1675 	dmu_buf_rele(db, FTAG);
1676 
1677 	mutex_enter(&vd->vdev_dtl_lock);
1678 	error = space_map_load(&vd->vdev_dtl[DTL_MISSING],
1679 	    NULL, SM_ALLOC, smo, mos);
1680 	mutex_exit(&vd->vdev_dtl_lock);
1681 
1682 	return (error);
1683 }
1684 
1685 void
1686 vdev_dtl_sync(vdev_t *vd, uint64_t txg)
1687 {
1688 	spa_t *spa = vd->vdev_spa;
1689 	space_map_obj_t *smo = &vd->vdev_dtl_smo;
1690 	space_map_t *sm = &vd->vdev_dtl[DTL_MISSING];
1691 	objset_t *mos = spa->spa_meta_objset;
1692 	space_map_t smsync;
1693 	kmutex_t smlock;
1694 	dmu_buf_t *db;
1695 	dmu_tx_t *tx;
1696 
1697 	ASSERT(!vd->vdev_ishole);
1698 
1699 	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1700 
1701 	if (vd->vdev_detached) {
1702 		if (smo->smo_object != 0) {
1703 			int err = dmu_object_free(mos, smo->smo_object, tx);
1704 			ASSERT3U(err, ==, 0);
1705 			smo->smo_object = 0;
1706 		}
1707 		dmu_tx_commit(tx);
1708 		return;
1709 	}
1710 
1711 	if (smo->smo_object == 0) {
1712 		ASSERT(smo->smo_objsize == 0);
1713 		ASSERT(smo->smo_alloc == 0);
1714 		smo->smo_object = dmu_object_alloc(mos,
1715 		    DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
1716 		    DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
1717 		ASSERT(smo->smo_object != 0);
1718 		vdev_config_dirty(vd->vdev_top);
1719 	}
1720 
1721 	mutex_init(&smlock, NULL, MUTEX_DEFAULT, NULL);
1722 
1723 	space_map_create(&smsync, sm->sm_start, sm->sm_size, sm->sm_shift,
1724 	    &smlock);
1725 
1726 	mutex_enter(&smlock);
1727 
1728 	mutex_enter(&vd->vdev_dtl_lock);
1729 	space_map_walk(sm, space_map_add, &smsync);
1730 	mutex_exit(&vd->vdev_dtl_lock);
1731 
1732 	space_map_truncate(smo, mos, tx);
1733 	space_map_sync(&smsync, SM_ALLOC, smo, mos, tx);
1734 
1735 	space_map_destroy(&smsync);
1736 
1737 	mutex_exit(&smlock);
1738 	mutex_destroy(&smlock);
1739 
1740 	VERIFY(0 == dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
1741 	dmu_buf_will_dirty(db, tx);
1742 	ASSERT3U(db->db_size, >=, sizeof (*smo));
1743 	bcopy(smo, db->db_data, sizeof (*smo));
1744 	dmu_buf_rele(db, FTAG);
1745 
1746 	dmu_tx_commit(tx);
1747 }
1748 
1749 /*
1750  * Determine whether the specified vdev can be offlined/detached/removed
1751  * without losing data.
1752  */
1753 boolean_t
1754 vdev_dtl_required(vdev_t *vd)
1755 {
1756 	spa_t *spa = vd->vdev_spa;
1757 	vdev_t *tvd = vd->vdev_top;
1758 	uint8_t cant_read = vd->vdev_cant_read;
1759 	boolean_t required;
1760 
1761 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1762 
1763 	if (vd == spa->spa_root_vdev || vd == tvd)
1764 		return (B_TRUE);
1765 
1766 	/*
1767 	 * Temporarily mark the device as unreadable, and then determine
1768 	 * whether this results in any DTL outages in the top-level vdev.
1769 	 * If not, we can safely offline/detach/remove the device.
1770 	 */
1771 	vd->vdev_cant_read = B_TRUE;
1772 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1773 	required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
1774 	vd->vdev_cant_read = cant_read;
1775 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1776 
1777 	return (required);
1778 }
1779 
1780 /*
1781  * Determine if resilver is needed, and if so the txg range.
1782  */
1783 boolean_t
1784 vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
1785 {
1786 	boolean_t needed = B_FALSE;
1787 	uint64_t thismin = UINT64_MAX;
1788 	uint64_t thismax = 0;
1789 
1790 	if (vd->vdev_children == 0) {
1791 		mutex_enter(&vd->vdev_dtl_lock);
1792 		if (vd->vdev_dtl[DTL_MISSING].sm_space != 0 &&
1793 		    vdev_writeable(vd)) {
1794 			space_seg_t *ss;
1795 
1796 			ss = avl_first(&vd->vdev_dtl[DTL_MISSING].sm_root);
1797 			thismin = ss->ss_start - 1;
1798 			ss = avl_last(&vd->vdev_dtl[DTL_MISSING].sm_root);
1799 			thismax = ss->ss_end;
1800 			needed = B_TRUE;
1801 		}
1802 		mutex_exit(&vd->vdev_dtl_lock);
1803 	} else {
1804 		for (int c = 0; c < vd->vdev_children; c++) {
1805 			vdev_t *cvd = vd->vdev_child[c];
1806 			uint64_t cmin, cmax;
1807 
1808 			if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
1809 				thismin = MIN(thismin, cmin);
1810 				thismax = MAX(thismax, cmax);
1811 				needed = B_TRUE;
1812 			}
1813 		}
1814 	}
1815 
1816 	if (needed && minp) {
1817 		*minp = thismin;
1818 		*maxp = thismax;
1819 	}
1820 	return (needed);
1821 }
1822 
1823 void
1824 vdev_load(vdev_t *vd)
1825 {
1826 	/*
1827 	 * Recursively load all children.
1828 	 */
1829 	for (int c = 0; c < vd->vdev_children; c++)
1830 		vdev_load(vd->vdev_child[c]);
1831 
1832 	/*
1833 	 * If this is a top-level vdev, initialize its metaslabs.
1834 	 */
1835 	if (vd == vd->vdev_top && !vd->vdev_ishole &&
1836 	    (vd->vdev_ashift == 0 || vd->vdev_asize == 0 ||
1837 	    vdev_metaslab_init(vd, 0) != 0))
1838 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1839 		    VDEV_AUX_CORRUPT_DATA);
1840 
1841 	/*
1842 	 * If this is a leaf vdev, load its DTL.
1843 	 */
1844 	if (vd->vdev_ops->vdev_op_leaf && vdev_dtl_load(vd) != 0)
1845 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1846 		    VDEV_AUX_CORRUPT_DATA);
1847 }
1848 
1849 /*
1850  * The special vdev case is used for hot spares and l2cache devices.  Its
1851  * sole purpose it to set the vdev state for the associated vdev.  To do this,
1852  * we make sure that we can open the underlying device, then try to read the
1853  * label, and make sure that the label is sane and that it hasn't been
1854  * repurposed to another pool.
1855  */
1856 int
1857 vdev_validate_aux(vdev_t *vd)
1858 {
1859 	nvlist_t *label;
1860 	uint64_t guid, version;
1861 	uint64_t state;
1862 
1863 	if (!vdev_readable(vd))
1864 		return (0);
1865 
1866 	if ((label = vdev_label_read_config(vd)) == NULL) {
1867 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1868 		    VDEV_AUX_CORRUPT_DATA);
1869 		return (-1);
1870 	}
1871 
1872 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
1873 	    version > SPA_VERSION ||
1874 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
1875 	    guid != vd->vdev_guid ||
1876 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
1877 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1878 		    VDEV_AUX_CORRUPT_DATA);
1879 		nvlist_free(label);
1880 		return (-1);
1881 	}
1882 
1883 	/*
1884 	 * We don't actually check the pool state here.  If it's in fact in
1885 	 * use by another pool, we update this fact on the fly when requested.
1886 	 */
1887 	nvlist_free(label);
1888 	return (0);
1889 }
1890 
1891 void
1892 vdev_remove(vdev_t *vd, uint64_t txg)
1893 {
1894 	spa_t *spa = vd->vdev_spa;
1895 	objset_t *mos = spa->spa_meta_objset;
1896 	dmu_tx_t *tx;
1897 
1898 	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
1899 
1900 	if (vd->vdev_dtl_smo.smo_object) {
1901 		ASSERT3U(vd->vdev_dtl_smo.smo_alloc, ==, 0);
1902 		(void) dmu_object_free(mos, vd->vdev_dtl_smo.smo_object, tx);
1903 		vd->vdev_dtl_smo.smo_object = 0;
1904 	}
1905 
1906 	if (vd->vdev_ms != NULL) {
1907 		for (int m = 0; m < vd->vdev_ms_count; m++) {
1908 			metaslab_t *msp = vd->vdev_ms[m];
1909 
1910 			if (msp == NULL || msp->ms_smo.smo_object == 0)
1911 				continue;
1912 
1913 			ASSERT3U(msp->ms_smo.smo_alloc, ==, 0);
1914 			(void) dmu_object_free(mos, msp->ms_smo.smo_object, tx);
1915 			msp->ms_smo.smo_object = 0;
1916 		}
1917 	}
1918 
1919 	if (vd->vdev_ms_array) {
1920 		(void) dmu_object_free(mos, vd->vdev_ms_array, tx);
1921 		vd->vdev_ms_array = 0;
1922 		vd->vdev_ms_shift = 0;
1923 	}
1924 	dmu_tx_commit(tx);
1925 }
1926 
1927 void
1928 vdev_sync_done(vdev_t *vd, uint64_t txg)
1929 {
1930 	metaslab_t *msp;
1931 
1932 	ASSERT(!vd->vdev_ishole);
1933 
1934 	while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
1935 		metaslab_sync_done(msp, txg);
1936 }
1937 
1938 void
1939 vdev_sync(vdev_t *vd, uint64_t txg)
1940 {
1941 	spa_t *spa = vd->vdev_spa;
1942 	vdev_t *lvd;
1943 	metaslab_t *msp;
1944 	dmu_tx_t *tx;
1945 
1946 	ASSERT(!vd->vdev_ishole);
1947 
1948 	if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0) {
1949 		ASSERT(vd == vd->vdev_top);
1950 		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1951 		vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
1952 		    DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
1953 		ASSERT(vd->vdev_ms_array != 0);
1954 		vdev_config_dirty(vd);
1955 		dmu_tx_commit(tx);
1956 	}
1957 
1958 	if (vd->vdev_removing)
1959 		vdev_remove(vd, txg);
1960 
1961 	while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
1962 		metaslab_sync(msp, txg);
1963 		(void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
1964 	}
1965 
1966 	while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
1967 		vdev_dtl_sync(lvd, txg);
1968 
1969 	(void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
1970 }
1971 
1972 uint64_t
1973 vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
1974 {
1975 	return (vd->vdev_ops->vdev_op_asize(vd, psize));
1976 }
1977 
1978 /*
1979  * Mark the given vdev faulted.  A faulted vdev behaves as if the device could
1980  * not be opened, and no I/O is attempted.
1981  */
1982 int
1983 vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
1984 {
1985 	vdev_t *vd;
1986 
1987 	spa_vdev_state_enter(spa, SCL_NONE);
1988 
1989 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
1990 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
1991 
1992 	if (!vd->vdev_ops->vdev_op_leaf)
1993 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
1994 
1995 	/*
1996 	 * We don't directly use the aux state here, but if we do a
1997 	 * vdev_reopen(), we need this value to be present to remember why we
1998 	 * were faulted.
1999 	 */
2000 	vd->vdev_label_aux = aux;
2001 
2002 	/*
2003 	 * Faulted state takes precedence over degraded.
2004 	 */
2005 	vd->vdev_faulted = 1ULL;
2006 	vd->vdev_degraded = 0ULL;
2007 	vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
2008 
2009 	/*
2010 	 * If marking the vdev as faulted cause the top-level vdev to become
2011 	 * unavailable, then back off and simply mark the vdev as degraded
2012 	 * instead.
2013 	 */
2014 	if (vdev_is_dead(vd->vdev_top) && !vd->vdev_islog &&
2015 	    vd->vdev_aux == NULL) {
2016 		vd->vdev_degraded = 1ULL;
2017 		vd->vdev_faulted = 0ULL;
2018 
2019 		/*
2020 		 * If we reopen the device and it's not dead, only then do we
2021 		 * mark it degraded.
2022 		 */
2023 		vdev_reopen(vd);
2024 
2025 		if (vdev_readable(vd))
2026 			vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
2027 	}
2028 
2029 	return (spa_vdev_state_exit(spa, vd, 0));
2030 }
2031 
2032 /*
2033  * Mark the given vdev degraded.  A degraded vdev is purely an indication to the
2034  * user that something is wrong.  The vdev continues to operate as normal as far
2035  * as I/O is concerned.
2036  */
2037 int
2038 vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2039 {
2040 	vdev_t *vd;
2041 
2042 	spa_vdev_state_enter(spa, SCL_NONE);
2043 
2044 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2045 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2046 
2047 	if (!vd->vdev_ops->vdev_op_leaf)
2048 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2049 
2050 	/*
2051 	 * If the vdev is already faulted, then don't do anything.
2052 	 */
2053 	if (vd->vdev_faulted || vd->vdev_degraded)
2054 		return (spa_vdev_state_exit(spa, NULL, 0));
2055 
2056 	vd->vdev_degraded = 1ULL;
2057 	if (!vdev_is_dead(vd))
2058 		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
2059 		    aux);
2060 
2061 	return (spa_vdev_state_exit(spa, vd, 0));
2062 }
2063 
2064 /*
2065  * Online the given vdev.  If 'unspare' is set, it implies two things.  First,
2066  * any attached spare device should be detached when the device finishes
2067  * resilvering.  Second, the online should be treated like a 'test' online case,
2068  * so no FMA events are generated if the device fails to open.
2069  */
2070 int
2071 vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
2072 {
2073 	vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
2074 
2075 	spa_vdev_state_enter(spa, SCL_NONE);
2076 
2077 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2078 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2079 
2080 	if (!vd->vdev_ops->vdev_op_leaf)
2081 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2082 
2083 	tvd = vd->vdev_top;
2084 	vd->vdev_offline = B_FALSE;
2085 	vd->vdev_tmpoffline = B_FALSE;
2086 	vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
2087 	vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
2088 
2089 	/* XXX - L2ARC 1.0 does not support expansion */
2090 	if (!vd->vdev_aux) {
2091 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2092 			pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
2093 	}
2094 
2095 	vdev_reopen(tvd);
2096 	vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
2097 
2098 	if (!vd->vdev_aux) {
2099 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2100 			pvd->vdev_expanding = B_FALSE;
2101 	}
2102 
2103 	if (newstate)
2104 		*newstate = vd->vdev_state;
2105 	if ((flags & ZFS_ONLINE_UNSPARE) &&
2106 	    !vdev_is_dead(vd) && vd->vdev_parent &&
2107 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2108 	    vd->vdev_parent->vdev_child[0] == vd)
2109 		vd->vdev_unspare = B_TRUE;
2110 
2111 	if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
2112 
2113 		/* XXX - L2ARC 1.0 does not support expansion */
2114 		if (vd->vdev_aux)
2115 			return (spa_vdev_state_exit(spa, vd, ENOTSUP));
2116 		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
2117 	}
2118 	return (spa_vdev_state_exit(spa, vd, 0));
2119 }
2120 
2121 int
2122 vdev_offline_log(spa_t *spa)
2123 {
2124 	int error = 0;
2125 
2126 	if ((error = dmu_objset_find(spa_name(spa), zil_vdev_offline,
2127 	    NULL, DS_FIND_CHILDREN)) == 0) {
2128 
2129 		/*
2130 		 * We successfully offlined the log device, sync out the
2131 		 * current txg so that the "stubby" block can be removed
2132 		 * by zil_sync().
2133 		 */
2134 		txg_wait_synced(spa->spa_dsl_pool, 0);
2135 	}
2136 	return (error);
2137 }
2138 
2139 int
2140 vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
2141 {
2142 	vdev_t *vd, *tvd;
2143 	int error = 0;
2144 	uint64_t generation;
2145 	metaslab_group_t *mg;
2146 
2147 top:
2148 	spa_vdev_state_enter(spa, SCL_ALLOC);
2149 
2150 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2151 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2152 
2153 	if (!vd->vdev_ops->vdev_op_leaf)
2154 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2155 
2156 	tvd = vd->vdev_top;
2157 	mg = tvd->vdev_mg;
2158 	generation = spa->spa_config_generation + 1;
2159 
2160 	/*
2161 	 * If the device isn't already offline, try to offline it.
2162 	 */
2163 	if (!vd->vdev_offline) {
2164 		/*
2165 		 * If this device has the only valid copy of some data,
2166 		 * don't allow it to be offlined. Log devices are always
2167 		 * expendable.
2168 		 */
2169 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2170 		    vdev_dtl_required(vd))
2171 			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2172 
2173 		/*
2174 		 * If the top-level is a slog and it's had allocations
2175 		 * then proceed. We check that the vdev's metaslab
2176 		 * grop is not NULL since it's possible that we may
2177 		 * have just added this vdev and have not yet initialized
2178 		 * it's metaslabs.
2179 		 */
2180 		if (tvd->vdev_islog && mg != NULL) {
2181 			/*
2182 			 * Prevent any future allocations.
2183 			 */
2184 			metaslab_class_remove(spa->spa_log_class, mg);
2185 			(void) spa_vdev_state_exit(spa, vd, 0);
2186 
2187 			error = vdev_offline_log(spa);
2188 
2189 			spa_vdev_state_enter(spa, SCL_ALLOC);
2190 
2191 			/*
2192 			 * Check to see if the config has changed.
2193 			 */
2194 			if (error || generation != spa->spa_config_generation) {
2195 				metaslab_class_add(spa->spa_log_class, mg);
2196 				if (error)
2197 					return (spa_vdev_state_exit(spa,
2198 					    vd, error));
2199 				(void) spa_vdev_state_exit(spa, vd, 0);
2200 				goto top;
2201 			}
2202 			ASSERT3U(tvd->vdev_stat.vs_alloc, ==, 0);
2203 		}
2204 
2205 		/*
2206 		 * Offline this device and reopen its top-level vdev.
2207 		 * If the top-level vdev is a log device then just offline
2208 		 * it. Otherwise, if this action results in the top-level
2209 		 * vdev becoming unusable, undo it and fail the request.
2210 		 */
2211 		vd->vdev_offline = B_TRUE;
2212 		vdev_reopen(tvd);
2213 
2214 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2215 		    vdev_is_dead(tvd)) {
2216 			vd->vdev_offline = B_FALSE;
2217 			vdev_reopen(tvd);
2218 			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2219 		}
2220 
2221 		/*
2222 		 * Add the device back into the metaslab rotor so that
2223 		 * once we online the device it's open for business.
2224 		 */
2225 		if (tvd->vdev_islog && mg != NULL)
2226 			metaslab_class_add(spa->spa_log_class, mg);
2227 	}
2228 
2229 	vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
2230 
2231 	return (spa_vdev_state_exit(spa, vd, 0));
2232 }
2233 
2234 /*
2235  * Clear the error counts associated with this vdev.  Unlike vdev_online() and
2236  * vdev_offline(), we assume the spa config is locked.  We also clear all
2237  * children.  If 'vd' is NULL, then the user wants to clear all vdevs.
2238  */
2239 void
2240 vdev_clear(spa_t *spa, vdev_t *vd)
2241 {
2242 	vdev_t *rvd = spa->spa_root_vdev;
2243 
2244 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2245 
2246 	if (vd == NULL)
2247 		vd = rvd;
2248 
2249 	vd->vdev_stat.vs_read_errors = 0;
2250 	vd->vdev_stat.vs_write_errors = 0;
2251 	vd->vdev_stat.vs_checksum_errors = 0;
2252 
2253 	for (int c = 0; c < vd->vdev_children; c++)
2254 		vdev_clear(spa, vd->vdev_child[c]);
2255 
2256 	/*
2257 	 * If we're in the FAULTED state or have experienced failed I/O, then
2258 	 * clear the persistent state and attempt to reopen the device.  We
2259 	 * also mark the vdev config dirty, so that the new faulted state is
2260 	 * written out to disk.
2261 	 */
2262 	if (vd->vdev_faulted || vd->vdev_degraded ||
2263 	    !vdev_readable(vd) || !vdev_writeable(vd)) {
2264 
2265 		/*
2266 		 * When reopening in reponse to a clear event, it may be due to
2267 		 * a fmadm repair request.  In this case, if the device is
2268 		 * still broken, we want to still post the ereport again.
2269 		 */
2270 		vd->vdev_forcefault = B_TRUE;
2271 
2272 		vd->vdev_faulted = vd->vdev_degraded = 0;
2273 		vd->vdev_cant_read = B_FALSE;
2274 		vd->vdev_cant_write = B_FALSE;
2275 
2276 		vdev_reopen(vd);
2277 
2278 		vd->vdev_forcefault = B_FALSE;
2279 
2280 		if (vd != rvd)
2281 			vdev_state_dirty(vd->vdev_top);
2282 
2283 		if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
2284 			spa_async_request(spa, SPA_ASYNC_RESILVER);
2285 
2286 		spa_event_notify(spa, vd, ESC_ZFS_VDEV_CLEAR);
2287 	}
2288 
2289 	/*
2290 	 * When clearing a FMA-diagnosed fault, we always want to
2291 	 * unspare the device, as we assume that the original spare was
2292 	 * done in response to the FMA fault.
2293 	 */
2294 	if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
2295 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2296 	    vd->vdev_parent->vdev_child[0] == vd)
2297 		vd->vdev_unspare = B_TRUE;
2298 }
2299 
2300 boolean_t
2301 vdev_is_dead(vdev_t *vd)
2302 {
2303 	/*
2304 	 * Holes and missing devices are always considered "dead".
2305 	 * This simplifies the code since we don't have to check for
2306 	 * these types of devices in the various code paths.
2307 	 * Instead we rely on the fact that we skip over dead devices
2308 	 * before issuing I/O to them.
2309 	 */
2310 	return (vd->vdev_state < VDEV_STATE_DEGRADED || vd->vdev_ishole ||
2311 	    vd->vdev_ops == &vdev_missing_ops);
2312 }
2313 
2314 boolean_t
2315 vdev_readable(vdev_t *vd)
2316 {
2317 	return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
2318 }
2319 
2320 boolean_t
2321 vdev_writeable(vdev_t *vd)
2322 {
2323 	return (!vdev_is_dead(vd) && !vd->vdev_cant_write);
2324 }
2325 
2326 boolean_t
2327 vdev_allocatable(vdev_t *vd)
2328 {
2329 	uint64_t state = vd->vdev_state;
2330 
2331 	/*
2332 	 * We currently allow allocations from vdevs which may be in the
2333 	 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
2334 	 * fails to reopen then we'll catch it later when we're holding
2335 	 * the proper locks.  Note that we have to get the vdev state
2336 	 * in a local variable because although it changes atomically,
2337 	 * we're asking two separate questions about it.
2338 	 */
2339 	return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
2340 	    !vd->vdev_cant_write && !vd->vdev_ishole && !vd->vdev_removing);
2341 }
2342 
2343 boolean_t
2344 vdev_accessible(vdev_t *vd, zio_t *zio)
2345 {
2346 	ASSERT(zio->io_vd == vd);
2347 
2348 	if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
2349 		return (B_FALSE);
2350 
2351 	if (zio->io_type == ZIO_TYPE_READ)
2352 		return (!vd->vdev_cant_read);
2353 
2354 	if (zio->io_type == ZIO_TYPE_WRITE)
2355 		return (!vd->vdev_cant_write);
2356 
2357 	return (B_TRUE);
2358 }
2359 
2360 /*
2361  * Get statistics for the given vdev.
2362  */
2363 void
2364 vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
2365 {
2366 	vdev_t *rvd = vd->vdev_spa->spa_root_vdev;
2367 
2368 	mutex_enter(&vd->vdev_stat_lock);
2369 	bcopy(&vd->vdev_stat, vs, sizeof (*vs));
2370 	vs->vs_scrub_errors = vd->vdev_spa->spa_scrub_errors;
2371 	vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
2372 	vs->vs_state = vd->vdev_state;
2373 	vs->vs_rsize = vdev_get_min_asize(vd);
2374 	if (vd->vdev_ops->vdev_op_leaf)
2375 		vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
2376 	mutex_exit(&vd->vdev_stat_lock);
2377 
2378 	/*
2379 	 * If we're getting stats on the root vdev, aggregate the I/O counts
2380 	 * over all top-level vdevs (i.e. the direct children of the root).
2381 	 */
2382 	if (vd == rvd) {
2383 		for (int c = 0; c < rvd->vdev_children; c++) {
2384 			vdev_t *cvd = rvd->vdev_child[c];
2385 			vdev_stat_t *cvs = &cvd->vdev_stat;
2386 
2387 			mutex_enter(&vd->vdev_stat_lock);
2388 			for (int t = 0; t < ZIO_TYPES; t++) {
2389 				vs->vs_ops[t] += cvs->vs_ops[t];
2390 				vs->vs_bytes[t] += cvs->vs_bytes[t];
2391 			}
2392 			vs->vs_scrub_examined += cvs->vs_scrub_examined;
2393 			mutex_exit(&vd->vdev_stat_lock);
2394 		}
2395 	}
2396 }
2397 
2398 void
2399 vdev_clear_stats(vdev_t *vd)
2400 {
2401 	mutex_enter(&vd->vdev_stat_lock);
2402 	vd->vdev_stat.vs_space = 0;
2403 	vd->vdev_stat.vs_dspace = 0;
2404 	vd->vdev_stat.vs_alloc = 0;
2405 	mutex_exit(&vd->vdev_stat_lock);
2406 }
2407 
2408 void
2409 vdev_stat_update(zio_t *zio, uint64_t psize)
2410 {
2411 	spa_t *spa = zio->io_spa;
2412 	vdev_t *rvd = spa->spa_root_vdev;
2413 	vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
2414 	vdev_t *pvd;
2415 	uint64_t txg = zio->io_txg;
2416 	vdev_stat_t *vs = &vd->vdev_stat;
2417 	zio_type_t type = zio->io_type;
2418 	int flags = zio->io_flags;
2419 
2420 	/*
2421 	 * If this i/o is a gang leader, it didn't do any actual work.
2422 	 */
2423 	if (zio->io_gang_tree)
2424 		return;
2425 
2426 	if (zio->io_error == 0) {
2427 		/*
2428 		 * If this is a root i/o, don't count it -- we've already
2429 		 * counted the top-level vdevs, and vdev_get_stats() will
2430 		 * aggregate them when asked.  This reduces contention on
2431 		 * the root vdev_stat_lock and implicitly handles blocks
2432 		 * that compress away to holes, for which there is no i/o.
2433 		 * (Holes never create vdev children, so all the counters
2434 		 * remain zero, which is what we want.)
2435 		 *
2436 		 * Note: this only applies to successful i/o (io_error == 0)
2437 		 * because unlike i/o counts, errors are not additive.
2438 		 * When reading a ditto block, for example, failure of
2439 		 * one top-level vdev does not imply a root-level error.
2440 		 */
2441 		if (vd == rvd)
2442 			return;
2443 
2444 		ASSERT(vd == zio->io_vd);
2445 
2446 		if (flags & ZIO_FLAG_IO_BYPASS)
2447 			return;
2448 
2449 		mutex_enter(&vd->vdev_stat_lock);
2450 
2451 		if (flags & ZIO_FLAG_IO_REPAIR) {
2452 			if (flags & ZIO_FLAG_SCRUB_THREAD)
2453 				vs->vs_scrub_repaired += psize;
2454 			if (flags & ZIO_FLAG_SELF_HEAL)
2455 				vs->vs_self_healed += psize;
2456 		}
2457 
2458 		vs->vs_ops[type]++;
2459 		vs->vs_bytes[type] += psize;
2460 
2461 		mutex_exit(&vd->vdev_stat_lock);
2462 		return;
2463 	}
2464 
2465 	if (flags & ZIO_FLAG_SPECULATIVE)
2466 		return;
2467 
2468 	/*
2469 	 * If this is an I/O error that is going to be retried, then ignore the
2470 	 * error.  Otherwise, the user may interpret B_FAILFAST I/O errors as
2471 	 * hard errors, when in reality they can happen for any number of
2472 	 * innocuous reasons (bus resets, MPxIO link failure, etc).
2473 	 */
2474 	if (zio->io_error == EIO &&
2475 	    !(zio->io_flags & ZIO_FLAG_IO_RETRY))
2476 		return;
2477 
2478 	/*
2479 	 * Intent logs writes won't propagate their error to the root
2480 	 * I/O so don't mark these types of failures as pool-level
2481 	 * errors.
2482 	 */
2483 	if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
2484 		return;
2485 
2486 	mutex_enter(&vd->vdev_stat_lock);
2487 	if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
2488 		if (zio->io_error == ECKSUM)
2489 			vs->vs_checksum_errors++;
2490 		else
2491 			vs->vs_read_errors++;
2492 	}
2493 	if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
2494 		vs->vs_write_errors++;
2495 	mutex_exit(&vd->vdev_stat_lock);
2496 
2497 	if (type == ZIO_TYPE_WRITE && txg != 0 &&
2498 	    (!(flags & ZIO_FLAG_IO_REPAIR) ||
2499 	    (flags & ZIO_FLAG_SCRUB_THREAD))) {
2500 		/*
2501 		 * This is either a normal write (not a repair), or it's a
2502 		 * repair induced by the scrub thread.  In the normal case,
2503 		 * we commit the DTL change in the same txg as the block
2504 		 * was born.  In the scrub-induced repair case, we know that
2505 		 * scrubs run in first-pass syncing context, so we commit
2506 		 * the DTL change in spa->spa_syncing_txg.
2507 		 *
2508 		 * We currently do not make DTL entries for failed spontaneous
2509 		 * self-healing writes triggered by normal (non-scrubbing)
2510 		 * reads, because we have no transactional context in which to
2511 		 * do so -- and it's not clear that it'd be desirable anyway.
2512 		 */
2513 		if (vd->vdev_ops->vdev_op_leaf) {
2514 			uint64_t commit_txg = txg;
2515 			if (flags & ZIO_FLAG_SCRUB_THREAD) {
2516 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2517 				ASSERT(spa_sync_pass(spa) == 1);
2518 				vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
2519 				commit_txg = spa->spa_syncing_txg;
2520 			}
2521 			ASSERT(commit_txg >= spa->spa_syncing_txg);
2522 			if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
2523 				return;
2524 			for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2525 				vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
2526 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
2527 		}
2528 		if (vd != rvd)
2529 			vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
2530 	}
2531 }
2532 
2533 void
2534 vdev_scrub_stat_update(vdev_t *vd, pool_scrub_type_t type, boolean_t complete)
2535 {
2536 	vdev_stat_t *vs = &vd->vdev_stat;
2537 
2538 	for (int c = 0; c < vd->vdev_children; c++)
2539 		vdev_scrub_stat_update(vd->vdev_child[c], type, complete);
2540 
2541 	mutex_enter(&vd->vdev_stat_lock);
2542 
2543 	if (type == POOL_SCRUB_NONE) {
2544 		/*
2545 		 * Update completion and end time.  Leave everything else alone
2546 		 * so we can report what happened during the previous scrub.
2547 		 */
2548 		vs->vs_scrub_complete = complete;
2549 		vs->vs_scrub_end = gethrestime_sec();
2550 	} else {
2551 		vs->vs_scrub_type = type;
2552 		vs->vs_scrub_complete = 0;
2553 		vs->vs_scrub_examined = 0;
2554 		vs->vs_scrub_repaired = 0;
2555 		vs->vs_scrub_start = gethrestime_sec();
2556 		vs->vs_scrub_end = 0;
2557 	}
2558 
2559 	mutex_exit(&vd->vdev_stat_lock);
2560 }
2561 
2562 /*
2563  * Update the in-core space usage stats for this vdev and the root vdev.
2564  */
2565 void
2566 vdev_space_update(vdev_t *vd, int64_t space_delta, int64_t alloc_delta,
2567     boolean_t update_root)
2568 {
2569 	int64_t dspace_delta = space_delta;
2570 	spa_t *spa = vd->vdev_spa;
2571 	vdev_t *rvd = spa->spa_root_vdev;
2572 
2573 	ASSERT(vd == vd->vdev_top);
2574 
2575 	/*
2576 	 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
2577 	 * factor.  We must calculate this here and not at the root vdev
2578 	 * because the root vdev's psize-to-asize is simply the max of its
2579 	 * childrens', thus not accurate enough for us.
2580 	 */
2581 	ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
2582 	ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
2583 	dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
2584 	    vd->vdev_deflate_ratio;
2585 
2586 	mutex_enter(&vd->vdev_stat_lock);
2587 	vd->vdev_stat.vs_space += space_delta;
2588 	vd->vdev_stat.vs_alloc += alloc_delta;
2589 	vd->vdev_stat.vs_dspace += dspace_delta;
2590 	mutex_exit(&vd->vdev_stat_lock);
2591 
2592 	if (update_root) {
2593 		ASSERT(rvd == vd->vdev_parent);
2594 		ASSERT(vd->vdev_ms_count != 0);
2595 
2596 		/*
2597 		 * Don't count non-normal (e.g. intent log) space as part of
2598 		 * the pool's capacity.
2599 		 */
2600 		if (vd->vdev_islog)
2601 			return;
2602 
2603 		mutex_enter(&rvd->vdev_stat_lock);
2604 		rvd->vdev_stat.vs_space += space_delta;
2605 		rvd->vdev_stat.vs_alloc += alloc_delta;
2606 		rvd->vdev_stat.vs_dspace += dspace_delta;
2607 		mutex_exit(&rvd->vdev_stat_lock);
2608 	}
2609 }
2610 
2611 /*
2612  * Mark a top-level vdev's config as dirty, placing it on the dirty list
2613  * so that it will be written out next time the vdev configuration is synced.
2614  * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
2615  */
2616 void
2617 vdev_config_dirty(vdev_t *vd)
2618 {
2619 	spa_t *spa = vd->vdev_spa;
2620 	vdev_t *rvd = spa->spa_root_vdev;
2621 	int c;
2622 
2623 	/*
2624 	 * If this is an aux vdev (as with l2cache and spare devices), then we
2625 	 * update the vdev config manually and set the sync flag.
2626 	 */
2627 	if (vd->vdev_aux != NULL) {
2628 		spa_aux_vdev_t *sav = vd->vdev_aux;
2629 		nvlist_t **aux;
2630 		uint_t naux;
2631 
2632 		for (c = 0; c < sav->sav_count; c++) {
2633 			if (sav->sav_vdevs[c] == vd)
2634 				break;
2635 		}
2636 
2637 		if (c == sav->sav_count) {
2638 			/*
2639 			 * We're being removed.  There's nothing more to do.
2640 			 */
2641 			ASSERT(sav->sav_sync == B_TRUE);
2642 			return;
2643 		}
2644 
2645 		sav->sav_sync = B_TRUE;
2646 
2647 		if (nvlist_lookup_nvlist_array(sav->sav_config,
2648 		    ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
2649 			VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
2650 			    ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
2651 		}
2652 
2653 		ASSERT(c < naux);
2654 
2655 		/*
2656 		 * Setting the nvlist in the middle if the array is a little
2657 		 * sketchy, but it will work.
2658 		 */
2659 		nvlist_free(aux[c]);
2660 		aux[c] = vdev_config_generate(spa, vd, B_TRUE, B_FALSE, B_TRUE);
2661 
2662 		return;
2663 	}
2664 
2665 	/*
2666 	 * The dirty list is protected by the SCL_CONFIG lock.  The caller
2667 	 * must either hold SCL_CONFIG as writer, or must be the sync thread
2668 	 * (which holds SCL_CONFIG as reader).  There's only one sync thread,
2669 	 * so this is sufficient to ensure mutual exclusion.
2670 	 */
2671 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2672 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2673 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
2674 
2675 	if (vd == rvd) {
2676 		for (c = 0; c < rvd->vdev_children; c++)
2677 			vdev_config_dirty(rvd->vdev_child[c]);
2678 	} else {
2679 		ASSERT(vd == vd->vdev_top);
2680 
2681 		if (!list_link_active(&vd->vdev_config_dirty_node) &&
2682 		    !vd->vdev_ishole)
2683 			list_insert_head(&spa->spa_config_dirty_list, vd);
2684 	}
2685 }
2686 
2687 void
2688 vdev_config_clean(vdev_t *vd)
2689 {
2690 	spa_t *spa = vd->vdev_spa;
2691 
2692 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2693 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2694 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
2695 
2696 	ASSERT(list_link_active(&vd->vdev_config_dirty_node));
2697 	list_remove(&spa->spa_config_dirty_list, vd);
2698 }
2699 
2700 /*
2701  * Mark a top-level vdev's state as dirty, so that the next pass of
2702  * spa_sync() can convert this into vdev_config_dirty().  We distinguish
2703  * the state changes from larger config changes because they require
2704  * much less locking, and are often needed for administrative actions.
2705  */
2706 void
2707 vdev_state_dirty(vdev_t *vd)
2708 {
2709 	spa_t *spa = vd->vdev_spa;
2710 
2711 	ASSERT(vd == vd->vdev_top);
2712 
2713 	/*
2714 	 * The state list is protected by the SCL_STATE lock.  The caller
2715 	 * must either hold SCL_STATE as writer, or must be the sync thread
2716 	 * (which holds SCL_STATE as reader).  There's only one sync thread,
2717 	 * so this is sufficient to ensure mutual exclusion.
2718 	 */
2719 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2720 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2721 	    spa_config_held(spa, SCL_STATE, RW_READER)));
2722 
2723 	if (!list_link_active(&vd->vdev_state_dirty_node))
2724 		list_insert_head(&spa->spa_state_dirty_list, vd);
2725 }
2726 
2727 void
2728 vdev_state_clean(vdev_t *vd)
2729 {
2730 	spa_t *spa = vd->vdev_spa;
2731 
2732 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2733 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2734 	    spa_config_held(spa, SCL_STATE, RW_READER)));
2735 
2736 	ASSERT(list_link_active(&vd->vdev_state_dirty_node));
2737 	list_remove(&spa->spa_state_dirty_list, vd);
2738 }
2739 
2740 /*
2741  * Propagate vdev state up from children to parent.
2742  */
2743 void
2744 vdev_propagate_state(vdev_t *vd)
2745 {
2746 	spa_t *spa = vd->vdev_spa;
2747 	vdev_t *rvd = spa->spa_root_vdev;
2748 	int degraded = 0, faulted = 0;
2749 	int corrupted = 0;
2750 	vdev_t *child;
2751 
2752 	if (vd->vdev_children > 0) {
2753 		for (int c = 0; c < vd->vdev_children; c++) {
2754 			child = vd->vdev_child[c];
2755 
2756 			/*
2757 			 * Don't factor holes into the decision.
2758 			 */
2759 			if (child->vdev_ishole)
2760 				continue;
2761 
2762 			if (!vdev_readable(child) ||
2763 			    (!vdev_writeable(child) && spa_writeable(spa))) {
2764 				/*
2765 				 * Root special: if there is a top-level log
2766 				 * device, treat the root vdev as if it were
2767 				 * degraded.
2768 				 */
2769 				if (child->vdev_islog && vd == rvd)
2770 					degraded++;
2771 				else
2772 					faulted++;
2773 			} else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
2774 				degraded++;
2775 			}
2776 
2777 			if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
2778 				corrupted++;
2779 		}
2780 
2781 		vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
2782 
2783 		/*
2784 		 * Root special: if there is a top-level vdev that cannot be
2785 		 * opened due to corrupted metadata, then propagate the root
2786 		 * vdev's aux state as 'corrupt' rather than 'insufficient
2787 		 * replicas'.
2788 		 */
2789 		if (corrupted && vd == rvd &&
2790 		    rvd->vdev_state == VDEV_STATE_CANT_OPEN)
2791 			vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
2792 			    VDEV_AUX_CORRUPT_DATA);
2793 	}
2794 
2795 	if (vd->vdev_parent)
2796 		vdev_propagate_state(vd->vdev_parent);
2797 }
2798 
2799 /*
2800  * Set a vdev's state.  If this is during an open, we don't update the parent
2801  * state, because we're in the process of opening children depth-first.
2802  * Otherwise, we propagate the change to the parent.
2803  *
2804  * If this routine places a device in a faulted state, an appropriate ereport is
2805  * generated.
2806  */
2807 void
2808 vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
2809 {
2810 	uint64_t save_state;
2811 	spa_t *spa = vd->vdev_spa;
2812 
2813 	if (state == vd->vdev_state) {
2814 		vd->vdev_stat.vs_aux = aux;
2815 		return;
2816 	}
2817 
2818 	save_state = vd->vdev_state;
2819 
2820 	vd->vdev_state = state;
2821 	vd->vdev_stat.vs_aux = aux;
2822 
2823 	/*
2824 	 * If we are setting the vdev state to anything but an open state, then
2825 	 * always close the underlying device.  Otherwise, we keep accessible
2826 	 * but invalid devices open forever.  We don't call vdev_close() itself,
2827 	 * because that implies some extra checks (offline, etc) that we don't
2828 	 * want here.  This is limited to leaf devices, because otherwise
2829 	 * closing the device will affect other children.
2830 	 */
2831 	if (vdev_is_dead(vd) && vd->vdev_ops->vdev_op_leaf)
2832 		vd->vdev_ops->vdev_op_close(vd);
2833 
2834 	/*
2835 	 * If we have brought this vdev back into service, we need
2836 	 * to notify fmd so that it can gracefully repair any outstanding
2837 	 * cases due to a missing device.  We do this in all cases, even those
2838 	 * that probably don't correlate to a repaired fault.  This is sure to
2839 	 * catch all cases, and we let the zfs-retire agent sort it out.  If
2840 	 * this is a transient state it's OK, as the retire agent will
2841 	 * double-check the state of the vdev before repairing it.
2842 	 */
2843 	if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf &&
2844 	    vd->vdev_prevstate != state)
2845 		zfs_post_state_change(spa, vd);
2846 
2847 	if (vd->vdev_removed &&
2848 	    state == VDEV_STATE_CANT_OPEN &&
2849 	    (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
2850 		/*
2851 		 * If the previous state is set to VDEV_STATE_REMOVED, then this
2852 		 * device was previously marked removed and someone attempted to
2853 		 * reopen it.  If this failed due to a nonexistent device, then
2854 		 * keep the device in the REMOVED state.  We also let this be if
2855 		 * it is one of our special test online cases, which is only
2856 		 * attempting to online the device and shouldn't generate an FMA
2857 		 * fault.
2858 		 */
2859 		vd->vdev_state = VDEV_STATE_REMOVED;
2860 		vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
2861 	} else if (state == VDEV_STATE_REMOVED) {
2862 		vd->vdev_removed = B_TRUE;
2863 	} else if (state == VDEV_STATE_CANT_OPEN) {
2864 		/*
2865 		 * If we fail to open a vdev during an import, we mark it as
2866 		 * "not available", which signifies that it was never there to
2867 		 * begin with.  Failure to open such a device is not considered
2868 		 * an error.
2869 		 */
2870 		if (spa->spa_load_state == SPA_LOAD_IMPORT &&
2871 		    vd->vdev_ops->vdev_op_leaf)
2872 			vd->vdev_not_present = 1;
2873 
2874 		/*
2875 		 * Post the appropriate ereport.  If the 'prevstate' field is
2876 		 * set to something other than VDEV_STATE_UNKNOWN, it indicates
2877 		 * that this is part of a vdev_reopen().  In this case, we don't
2878 		 * want to post the ereport if the device was already in the
2879 		 * CANT_OPEN state beforehand.
2880 		 *
2881 		 * If the 'checkremove' flag is set, then this is an attempt to
2882 		 * online the device in response to an insertion event.  If we
2883 		 * hit this case, then we have detected an insertion event for a
2884 		 * faulted or offline device that wasn't in the removed state.
2885 		 * In this scenario, we don't post an ereport because we are
2886 		 * about to replace the device, or attempt an online with
2887 		 * vdev_forcefault, which will generate the fault for us.
2888 		 */
2889 		if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
2890 		    !vd->vdev_not_present && !vd->vdev_checkremove &&
2891 		    vd != spa->spa_root_vdev) {
2892 			const char *class;
2893 
2894 			switch (aux) {
2895 			case VDEV_AUX_OPEN_FAILED:
2896 				class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
2897 				break;
2898 			case VDEV_AUX_CORRUPT_DATA:
2899 				class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
2900 				break;
2901 			case VDEV_AUX_NO_REPLICAS:
2902 				class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
2903 				break;
2904 			case VDEV_AUX_BAD_GUID_SUM:
2905 				class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
2906 				break;
2907 			case VDEV_AUX_TOO_SMALL:
2908 				class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
2909 				break;
2910 			case VDEV_AUX_BAD_LABEL:
2911 				class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
2912 				break;
2913 			case VDEV_AUX_IO_FAILURE:
2914 				class = FM_EREPORT_ZFS_IO_FAILURE;
2915 				break;
2916 			default:
2917 				class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
2918 			}
2919 
2920 			zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
2921 		}
2922 
2923 		/* Erase any notion of persistent removed state */
2924 		vd->vdev_removed = B_FALSE;
2925 	} else {
2926 		vd->vdev_removed = B_FALSE;
2927 	}
2928 
2929 	if (!isopen && vd->vdev_parent)
2930 		vdev_propagate_state(vd->vdev_parent);
2931 }
2932 
2933 /*
2934  * Check the vdev configuration to ensure that it's capable of supporting
2935  * a root pool. Currently, we do not support RAID-Z or partial configuration.
2936  * In addition, only a single top-level vdev is allowed and none of the leaves
2937  * can be wholedisks.
2938  */
2939 boolean_t
2940 vdev_is_bootable(vdev_t *vd)
2941 {
2942 	if (!vd->vdev_ops->vdev_op_leaf) {
2943 		char *vdev_type = vd->vdev_ops->vdev_op_type;
2944 
2945 		if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
2946 		    vd->vdev_children > 1) {
2947 			return (B_FALSE);
2948 		} else if (strcmp(vdev_type, VDEV_TYPE_RAIDZ) == 0 ||
2949 		    strcmp(vdev_type, VDEV_TYPE_MISSING) == 0) {
2950 			return (B_FALSE);
2951 		}
2952 	} else if (vd->vdev_wholedisk == 1) {
2953 		return (B_FALSE);
2954 	}
2955 
2956 	for (int c = 0; c < vd->vdev_children; c++) {
2957 		if (!vdev_is_bootable(vd->vdev_child[c]))
2958 			return (B_FALSE);
2959 	}
2960 	return (B_TRUE);
2961 }
2962 
2963 /*
2964  * Load the state from the original vdev tree (ovd) which
2965  * we've retrieved from the MOS config object. If the original
2966  * vdev was offline then we transfer that state to the device
2967  * in the current vdev tree (nvd).
2968  */
2969 void
2970 vdev_load_log_state(vdev_t *nvd, vdev_t *ovd)
2971 {
2972 	spa_t *spa = nvd->vdev_spa;
2973 
2974 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2975 	ASSERT3U(nvd->vdev_guid, ==, ovd->vdev_guid);
2976 
2977 	for (int c = 0; c < nvd->vdev_children; c++)
2978 		vdev_load_log_state(nvd->vdev_child[c], ovd->vdev_child[c]);
2979 
2980 	if (nvd->vdev_ops->vdev_op_leaf && ovd->vdev_offline) {
2981 		/*
2982 		 * It would be nice to call vdev_offline()
2983 		 * directly but the pool isn't fully loaded and
2984 		 * the txg threads have not been started yet.
2985 		 */
2986 		nvd->vdev_offline = ovd->vdev_offline;
2987 		vdev_reopen(nvd->vdev_top);
2988 	}
2989 }
2990 
2991 /*
2992  * Expand a vdev if possible.
2993  */
2994 void
2995 vdev_expand(vdev_t *vd, uint64_t txg)
2996 {
2997 	ASSERT(vd->vdev_top == vd);
2998 	ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
2999 
3000 	if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count) {
3001 		VERIFY(vdev_metaslab_init(vd, txg) == 0);
3002 		vdev_config_dirty(vd);
3003 	}
3004 }
3005