xref: /illumos-gate/usr/src/uts/common/fs/zfs/vdev.c (revision 069f55e237020c4a4907b235fc38fafc6442ce94)
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 	if (zio_injection_enabled && error == 0)
1121 		error = zio_handle_device_injection(vd, NULL, ENXIO);
1122 
1123 	if (error) {
1124 		if (vd->vdev_removed &&
1125 		    vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1126 			vd->vdev_removed = B_FALSE;
1127 
1128 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1129 		    vd->vdev_stat.vs_aux);
1130 		return (error);
1131 	}
1132 
1133 	vd->vdev_removed = B_FALSE;
1134 
1135 	if (vd->vdev_degraded) {
1136 		ASSERT(vd->vdev_children == 0);
1137 		vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1138 		    VDEV_AUX_ERR_EXCEEDED);
1139 	} else {
1140 		vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1141 	}
1142 
1143 	/*
1144 	 * For hole or missing vdevs we just return success.
1145 	 */
1146 	if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1147 		return (0);
1148 
1149 	for (int c = 0; c < vd->vdev_children; c++) {
1150 		if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1151 			vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1152 			    VDEV_AUX_NONE);
1153 			break;
1154 		}
1155 	}
1156 
1157 	osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1158 
1159 	if (vd->vdev_children == 0) {
1160 		if (osize < SPA_MINDEVSIZE) {
1161 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1162 			    VDEV_AUX_TOO_SMALL);
1163 			return (EOVERFLOW);
1164 		}
1165 		psize = osize;
1166 		asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1167 	} else {
1168 		if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1169 		    (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1170 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1171 			    VDEV_AUX_TOO_SMALL);
1172 			return (EOVERFLOW);
1173 		}
1174 		psize = 0;
1175 		asize = osize;
1176 	}
1177 
1178 	vd->vdev_psize = psize;
1179 
1180 	/*
1181 	 * Make sure the allocatable size hasn't shrunk.
1182 	 */
1183 	if (asize < vd->vdev_min_asize) {
1184 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1185 		    VDEV_AUX_BAD_LABEL);
1186 		return (EINVAL);
1187 	}
1188 
1189 	if (vd->vdev_asize == 0) {
1190 		/*
1191 		 * This is the first-ever open, so use the computed values.
1192 		 * For testing purposes, a higher ashift can be requested.
1193 		 */
1194 		vd->vdev_asize = asize;
1195 		vd->vdev_ashift = MAX(ashift, vd->vdev_ashift);
1196 	} else {
1197 		/*
1198 		 * Make sure the alignment requirement hasn't increased.
1199 		 */
1200 		if (ashift > vd->vdev_top->vdev_ashift) {
1201 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1202 			    VDEV_AUX_BAD_LABEL);
1203 			return (EINVAL);
1204 		}
1205 	}
1206 
1207 	/*
1208 	 * If all children are healthy and the asize has increased,
1209 	 * then we've experienced dynamic LUN growth.  If automatic
1210 	 * expansion is enabled then use the additional space.
1211 	 */
1212 	if (vd->vdev_state == VDEV_STATE_HEALTHY && asize > vd->vdev_asize &&
1213 	    (vd->vdev_expanding || spa->spa_autoexpand))
1214 		vd->vdev_asize = asize;
1215 
1216 	vdev_set_min_asize(vd);
1217 
1218 	/*
1219 	 * Ensure we can issue some IO before declaring the
1220 	 * vdev open for business.
1221 	 */
1222 	if (vd->vdev_ops->vdev_op_leaf &&
1223 	    (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1224 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1225 		    VDEV_AUX_IO_FAILURE);
1226 		return (error);
1227 	}
1228 
1229 	/*
1230 	 * If a leaf vdev has a DTL, and seems healthy, then kick off a
1231 	 * resilver.  But don't do this if we are doing a reopen for a scrub,
1232 	 * since this would just restart the scrub we are already doing.
1233 	 */
1234 	if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1235 	    vdev_resilver_needed(vd, NULL, NULL))
1236 		spa_async_request(spa, SPA_ASYNC_RESILVER);
1237 
1238 	return (0);
1239 }
1240 
1241 /*
1242  * Called once the vdevs are all opened, this routine validates the label
1243  * contents.  This needs to be done before vdev_load() so that we don't
1244  * inadvertently do repair I/Os to the wrong device.
1245  *
1246  * This function will only return failure if one of the vdevs indicates that it
1247  * has since been destroyed or exported.  This is only possible if
1248  * /etc/zfs/zpool.cache was readonly at the time.  Otherwise, the vdev state
1249  * will be updated but the function will return 0.
1250  */
1251 int
1252 vdev_validate(vdev_t *vd)
1253 {
1254 	spa_t *spa = vd->vdev_spa;
1255 	nvlist_t *label;
1256 	uint64_t guid, top_guid;
1257 	uint64_t state;
1258 
1259 	for (int c = 0; c < vd->vdev_children; c++)
1260 		if (vdev_validate(vd->vdev_child[c]) != 0)
1261 			return (EBADF);
1262 
1263 	/*
1264 	 * If the device has already failed, or was marked offline, don't do
1265 	 * any further validation.  Otherwise, label I/O will fail and we will
1266 	 * overwrite the previous state.
1267 	 */
1268 	if (vd->vdev_ops->vdev_op_leaf && vdev_readable(vd)) {
1269 
1270 		if ((label = vdev_label_read_config(vd)) == NULL) {
1271 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1272 			    VDEV_AUX_BAD_LABEL);
1273 			return (0);
1274 		}
1275 
1276 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_GUID,
1277 		    &guid) != 0 || guid != spa_guid(spa)) {
1278 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1279 			    VDEV_AUX_CORRUPT_DATA);
1280 			nvlist_free(label);
1281 			return (0);
1282 		}
1283 
1284 		/*
1285 		 * If this vdev just became a top-level vdev because its
1286 		 * sibling was detached, it will have adopted the parent's
1287 		 * vdev guid -- but the label may or may not be on disk yet.
1288 		 * Fortunately, either version of the label will have the
1289 		 * same top guid, so if we're a top-level vdev, we can
1290 		 * safely compare to that instead.
1291 		 */
1292 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID,
1293 		    &guid) != 0 ||
1294 		    nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID,
1295 		    &top_guid) != 0 ||
1296 		    (vd->vdev_guid != guid &&
1297 		    (vd->vdev_guid != top_guid || vd != vd->vdev_top))) {
1298 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1299 			    VDEV_AUX_CORRUPT_DATA);
1300 			nvlist_free(label);
1301 			return (0);
1302 		}
1303 
1304 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1305 		    &state) != 0) {
1306 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1307 			    VDEV_AUX_CORRUPT_DATA);
1308 			nvlist_free(label);
1309 			return (0);
1310 		}
1311 
1312 		nvlist_free(label);
1313 
1314 		/*
1315 		 * If spa->spa_load_verbatim is true, no need to check the
1316 		 * state of the pool.
1317 		 */
1318 		if (!spa->spa_load_verbatim &&
1319 		    spa->spa_load_state == SPA_LOAD_OPEN &&
1320 		    state != POOL_STATE_ACTIVE)
1321 			return (EBADF);
1322 
1323 		/*
1324 		 * If we were able to open and validate a vdev that was
1325 		 * previously marked permanently unavailable, clear that state
1326 		 * now.
1327 		 */
1328 		if (vd->vdev_not_present)
1329 			vd->vdev_not_present = 0;
1330 	}
1331 
1332 	return (0);
1333 }
1334 
1335 /*
1336  * Close a virtual device.
1337  */
1338 void
1339 vdev_close(vdev_t *vd)
1340 {
1341 	spa_t *spa = vd->vdev_spa;
1342 
1343 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1344 
1345 	vd->vdev_ops->vdev_op_close(vd);
1346 
1347 	vdev_cache_purge(vd);
1348 
1349 	/*
1350 	 * We record the previous state before we close it, so that if we are
1351 	 * doing a reopen(), we don't generate FMA ereports if we notice that
1352 	 * it's still faulted.
1353 	 */
1354 	vd->vdev_prevstate = vd->vdev_state;
1355 
1356 	if (vd->vdev_offline)
1357 		vd->vdev_state = VDEV_STATE_OFFLINE;
1358 	else
1359 		vd->vdev_state = VDEV_STATE_CLOSED;
1360 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1361 }
1362 
1363 void
1364 vdev_reopen(vdev_t *vd)
1365 {
1366 	spa_t *spa = vd->vdev_spa;
1367 
1368 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1369 
1370 	vdev_close(vd);
1371 	(void) vdev_open(vd);
1372 
1373 	/*
1374 	 * Call vdev_validate() here to make sure we have the same device.
1375 	 * Otherwise, a device with an invalid label could be successfully
1376 	 * opened in response to vdev_reopen().
1377 	 */
1378 	if (vd->vdev_aux) {
1379 		(void) vdev_validate_aux(vd);
1380 		if (vdev_readable(vd) && vdev_writeable(vd) &&
1381 		    vd->vdev_aux == &spa->spa_l2cache &&
1382 		    !l2arc_vdev_present(vd))
1383 			l2arc_add_vdev(spa, vd);
1384 	} else {
1385 		(void) vdev_validate(vd);
1386 	}
1387 
1388 	/*
1389 	 * Reassess parent vdev's health.
1390 	 */
1391 	vdev_propagate_state(vd);
1392 }
1393 
1394 int
1395 vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1396 {
1397 	int error;
1398 
1399 	/*
1400 	 * Normally, partial opens (e.g. of a mirror) are allowed.
1401 	 * For a create, however, we want to fail the request if
1402 	 * there are any components we can't open.
1403 	 */
1404 	error = vdev_open(vd);
1405 
1406 	if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
1407 		vdev_close(vd);
1408 		return (error ? error : ENXIO);
1409 	}
1410 
1411 	/*
1412 	 * Recursively initialize all labels.
1413 	 */
1414 	if ((error = vdev_label_init(vd, txg, isreplacing ?
1415 	    VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
1416 		vdev_close(vd);
1417 		return (error);
1418 	}
1419 
1420 	return (0);
1421 }
1422 
1423 void
1424 vdev_metaslab_set_size(vdev_t *vd)
1425 {
1426 	/*
1427 	 * Aim for roughly 200 metaslabs per vdev.
1428 	 */
1429 	vd->vdev_ms_shift = highbit(vd->vdev_asize / 200);
1430 	vd->vdev_ms_shift = MAX(vd->vdev_ms_shift, SPA_MAXBLOCKSHIFT);
1431 }
1432 
1433 void
1434 vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
1435 {
1436 	ASSERT(vd == vd->vdev_top);
1437 	ASSERT(!vd->vdev_ishole);
1438 	ASSERT(ISP2(flags));
1439 
1440 	if (flags & VDD_METASLAB)
1441 		(void) txg_list_add(&vd->vdev_ms_list, arg, txg);
1442 
1443 	if (flags & VDD_DTL)
1444 		(void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
1445 
1446 	(void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
1447 }
1448 
1449 /*
1450  * DTLs.
1451  *
1452  * A vdev's DTL (dirty time log) is the set of transaction groups for which
1453  * the vdev has less than perfect replication.  There are three kinds of DTL:
1454  *
1455  * DTL_MISSING: txgs for which the vdev has no valid copies of the data
1456  *
1457  * DTL_PARTIAL: txgs for which data is available, but not fully replicated
1458  *
1459  * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
1460  *	scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
1461  *	txgs that was scrubbed.
1462  *
1463  * DTL_OUTAGE: txgs which cannot currently be read, whether due to
1464  *	persistent errors or just some device being offline.
1465  *	Unlike the other three, the DTL_OUTAGE map is not generally
1466  *	maintained; it's only computed when needed, typically to
1467  *	determine whether a device can be detached.
1468  *
1469  * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
1470  * either has the data or it doesn't.
1471  *
1472  * For interior vdevs such as mirror and RAID-Z the picture is more complex.
1473  * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
1474  * if any child is less than fully replicated, then so is its parent.
1475  * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
1476  * comprising only those txgs which appear in 'maxfaults' or more children;
1477  * those are the txgs we don't have enough replication to read.  For example,
1478  * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
1479  * thus, its DTL_MISSING consists of the set of txgs that appear in more than
1480  * two child DTL_MISSING maps.
1481  *
1482  * It should be clear from the above that to compute the DTLs and outage maps
1483  * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
1484  * Therefore, that is all we keep on disk.  When loading the pool, or after
1485  * a configuration change, we generate all other DTLs from first principles.
1486  */
1487 void
1488 vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1489 {
1490 	space_map_t *sm = &vd->vdev_dtl[t];
1491 
1492 	ASSERT(t < DTL_TYPES);
1493 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1494 
1495 	mutex_enter(sm->sm_lock);
1496 	if (!space_map_contains(sm, txg, size))
1497 		space_map_add(sm, txg, size);
1498 	mutex_exit(sm->sm_lock);
1499 }
1500 
1501 boolean_t
1502 vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1503 {
1504 	space_map_t *sm = &vd->vdev_dtl[t];
1505 	boolean_t dirty = B_FALSE;
1506 
1507 	ASSERT(t < DTL_TYPES);
1508 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1509 
1510 	mutex_enter(sm->sm_lock);
1511 	if (sm->sm_space != 0)
1512 		dirty = space_map_contains(sm, txg, size);
1513 	mutex_exit(sm->sm_lock);
1514 
1515 	return (dirty);
1516 }
1517 
1518 boolean_t
1519 vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
1520 {
1521 	space_map_t *sm = &vd->vdev_dtl[t];
1522 	boolean_t empty;
1523 
1524 	mutex_enter(sm->sm_lock);
1525 	empty = (sm->sm_space == 0);
1526 	mutex_exit(sm->sm_lock);
1527 
1528 	return (empty);
1529 }
1530 
1531 /*
1532  * Reassess DTLs after a config change or scrub completion.
1533  */
1534 void
1535 vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
1536 {
1537 	spa_t *spa = vd->vdev_spa;
1538 	avl_tree_t reftree;
1539 	int minref;
1540 
1541 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
1542 
1543 	for (int c = 0; c < vd->vdev_children; c++)
1544 		vdev_dtl_reassess(vd->vdev_child[c], txg,
1545 		    scrub_txg, scrub_done);
1546 
1547 	if (vd == spa->spa_root_vdev || vd->vdev_ishole)
1548 		return;
1549 
1550 	if (vd->vdev_ops->vdev_op_leaf) {
1551 		mutex_enter(&vd->vdev_dtl_lock);
1552 		if (scrub_txg != 0 &&
1553 		    (spa->spa_scrub_started || spa->spa_scrub_errors == 0)) {
1554 			/* XXX should check scrub_done? */
1555 			/*
1556 			 * We completed a scrub up to scrub_txg.  If we
1557 			 * did it without rebooting, then the scrub dtl
1558 			 * will be valid, so excise the old region and
1559 			 * fold in the scrub dtl.  Otherwise, leave the
1560 			 * dtl as-is if there was an error.
1561 			 *
1562 			 * There's little trick here: to excise the beginning
1563 			 * of the DTL_MISSING map, we put it into a reference
1564 			 * tree and then add a segment with refcnt -1 that
1565 			 * covers the range [0, scrub_txg).  This means
1566 			 * that each txg in that range has refcnt -1 or 0.
1567 			 * We then add DTL_SCRUB with a refcnt of 2, so that
1568 			 * entries in the range [0, scrub_txg) will have a
1569 			 * positive refcnt -- either 1 or 2.  We then convert
1570 			 * the reference tree into the new DTL_MISSING map.
1571 			 */
1572 			space_map_ref_create(&reftree);
1573 			space_map_ref_add_map(&reftree,
1574 			    &vd->vdev_dtl[DTL_MISSING], 1);
1575 			space_map_ref_add_seg(&reftree, 0, scrub_txg, -1);
1576 			space_map_ref_add_map(&reftree,
1577 			    &vd->vdev_dtl[DTL_SCRUB], 2);
1578 			space_map_ref_generate_map(&reftree,
1579 			    &vd->vdev_dtl[DTL_MISSING], 1);
1580 			space_map_ref_destroy(&reftree);
1581 		}
1582 		space_map_vacate(&vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
1583 		space_map_walk(&vd->vdev_dtl[DTL_MISSING],
1584 		    space_map_add, &vd->vdev_dtl[DTL_PARTIAL]);
1585 		if (scrub_done)
1586 			space_map_vacate(&vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
1587 		space_map_vacate(&vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
1588 		if (!vdev_readable(vd))
1589 			space_map_add(&vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
1590 		else
1591 			space_map_walk(&vd->vdev_dtl[DTL_MISSING],
1592 			    space_map_add, &vd->vdev_dtl[DTL_OUTAGE]);
1593 		mutex_exit(&vd->vdev_dtl_lock);
1594 
1595 		if (txg != 0)
1596 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
1597 		return;
1598 	}
1599 
1600 	mutex_enter(&vd->vdev_dtl_lock);
1601 	for (int t = 0; t < DTL_TYPES; t++) {
1602 		if (t == DTL_SCRUB)
1603 			continue;			/* leaf vdevs only */
1604 		if (t == DTL_PARTIAL)
1605 			minref = 1;			/* i.e. non-zero */
1606 		else if (vd->vdev_nparity != 0)
1607 			minref = vd->vdev_nparity + 1;	/* RAID-Z */
1608 		else
1609 			minref = vd->vdev_children;	/* any kind of mirror */
1610 		space_map_ref_create(&reftree);
1611 		for (int c = 0; c < vd->vdev_children; c++) {
1612 			vdev_t *cvd = vd->vdev_child[c];
1613 			mutex_enter(&cvd->vdev_dtl_lock);
1614 			space_map_ref_add_map(&reftree, &cvd->vdev_dtl[t], 1);
1615 			mutex_exit(&cvd->vdev_dtl_lock);
1616 		}
1617 		space_map_ref_generate_map(&reftree, &vd->vdev_dtl[t], minref);
1618 		space_map_ref_destroy(&reftree);
1619 	}
1620 	mutex_exit(&vd->vdev_dtl_lock);
1621 }
1622 
1623 static int
1624 vdev_dtl_load(vdev_t *vd)
1625 {
1626 	spa_t *spa = vd->vdev_spa;
1627 	space_map_obj_t *smo = &vd->vdev_dtl_smo;
1628 	objset_t *mos = spa->spa_meta_objset;
1629 	dmu_buf_t *db;
1630 	int error;
1631 
1632 	ASSERT(vd->vdev_children == 0);
1633 
1634 	if (smo->smo_object == 0)
1635 		return (0);
1636 
1637 	ASSERT(!vd->vdev_ishole);
1638 
1639 	if ((error = dmu_bonus_hold(mos, smo->smo_object, FTAG, &db)) != 0)
1640 		return (error);
1641 
1642 	ASSERT3U(db->db_size, >=, sizeof (*smo));
1643 	bcopy(db->db_data, smo, sizeof (*smo));
1644 	dmu_buf_rele(db, FTAG);
1645 
1646 	mutex_enter(&vd->vdev_dtl_lock);
1647 	error = space_map_load(&vd->vdev_dtl[DTL_MISSING],
1648 	    NULL, SM_ALLOC, smo, mos);
1649 	mutex_exit(&vd->vdev_dtl_lock);
1650 
1651 	return (error);
1652 }
1653 
1654 void
1655 vdev_dtl_sync(vdev_t *vd, uint64_t txg)
1656 {
1657 	spa_t *spa = vd->vdev_spa;
1658 	space_map_obj_t *smo = &vd->vdev_dtl_smo;
1659 	space_map_t *sm = &vd->vdev_dtl[DTL_MISSING];
1660 	objset_t *mos = spa->spa_meta_objset;
1661 	space_map_t smsync;
1662 	kmutex_t smlock;
1663 	dmu_buf_t *db;
1664 	dmu_tx_t *tx;
1665 
1666 	ASSERT(!vd->vdev_ishole);
1667 
1668 	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1669 
1670 	if (vd->vdev_detached) {
1671 		if (smo->smo_object != 0) {
1672 			int err = dmu_object_free(mos, smo->smo_object, tx);
1673 			ASSERT3U(err, ==, 0);
1674 			smo->smo_object = 0;
1675 		}
1676 		dmu_tx_commit(tx);
1677 		return;
1678 	}
1679 
1680 	if (smo->smo_object == 0) {
1681 		ASSERT(smo->smo_objsize == 0);
1682 		ASSERT(smo->smo_alloc == 0);
1683 		smo->smo_object = dmu_object_alloc(mos,
1684 		    DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
1685 		    DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
1686 		ASSERT(smo->smo_object != 0);
1687 		vdev_config_dirty(vd->vdev_top);
1688 	}
1689 
1690 	mutex_init(&smlock, NULL, MUTEX_DEFAULT, NULL);
1691 
1692 	space_map_create(&smsync, sm->sm_start, sm->sm_size, sm->sm_shift,
1693 	    &smlock);
1694 
1695 	mutex_enter(&smlock);
1696 
1697 	mutex_enter(&vd->vdev_dtl_lock);
1698 	space_map_walk(sm, space_map_add, &smsync);
1699 	mutex_exit(&vd->vdev_dtl_lock);
1700 
1701 	space_map_truncate(smo, mos, tx);
1702 	space_map_sync(&smsync, SM_ALLOC, smo, mos, tx);
1703 
1704 	space_map_destroy(&smsync);
1705 
1706 	mutex_exit(&smlock);
1707 	mutex_destroy(&smlock);
1708 
1709 	VERIFY(0 == dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
1710 	dmu_buf_will_dirty(db, tx);
1711 	ASSERT3U(db->db_size, >=, sizeof (*smo));
1712 	bcopy(smo, db->db_data, sizeof (*smo));
1713 	dmu_buf_rele(db, FTAG);
1714 
1715 	dmu_tx_commit(tx);
1716 }
1717 
1718 /*
1719  * Determine whether the specified vdev can be offlined/detached/removed
1720  * without losing data.
1721  */
1722 boolean_t
1723 vdev_dtl_required(vdev_t *vd)
1724 {
1725 	spa_t *spa = vd->vdev_spa;
1726 	vdev_t *tvd = vd->vdev_top;
1727 	uint8_t cant_read = vd->vdev_cant_read;
1728 	boolean_t required;
1729 
1730 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1731 
1732 	if (vd == spa->spa_root_vdev || vd == tvd)
1733 		return (B_TRUE);
1734 
1735 	/*
1736 	 * Temporarily mark the device as unreadable, and then determine
1737 	 * whether this results in any DTL outages in the top-level vdev.
1738 	 * If not, we can safely offline/detach/remove the device.
1739 	 */
1740 	vd->vdev_cant_read = B_TRUE;
1741 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1742 	required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
1743 	vd->vdev_cant_read = cant_read;
1744 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1745 
1746 	return (required);
1747 }
1748 
1749 /*
1750  * Determine if resilver is needed, and if so the txg range.
1751  */
1752 boolean_t
1753 vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
1754 {
1755 	boolean_t needed = B_FALSE;
1756 	uint64_t thismin = UINT64_MAX;
1757 	uint64_t thismax = 0;
1758 
1759 	if (vd->vdev_children == 0) {
1760 		mutex_enter(&vd->vdev_dtl_lock);
1761 		if (vd->vdev_dtl[DTL_MISSING].sm_space != 0 &&
1762 		    vdev_writeable(vd)) {
1763 			space_seg_t *ss;
1764 
1765 			ss = avl_first(&vd->vdev_dtl[DTL_MISSING].sm_root);
1766 			thismin = ss->ss_start - 1;
1767 			ss = avl_last(&vd->vdev_dtl[DTL_MISSING].sm_root);
1768 			thismax = ss->ss_end;
1769 			needed = B_TRUE;
1770 		}
1771 		mutex_exit(&vd->vdev_dtl_lock);
1772 	} else {
1773 		for (int c = 0; c < vd->vdev_children; c++) {
1774 			vdev_t *cvd = vd->vdev_child[c];
1775 			uint64_t cmin, cmax;
1776 
1777 			if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
1778 				thismin = MIN(thismin, cmin);
1779 				thismax = MAX(thismax, cmax);
1780 				needed = B_TRUE;
1781 			}
1782 		}
1783 	}
1784 
1785 	if (needed && minp) {
1786 		*minp = thismin;
1787 		*maxp = thismax;
1788 	}
1789 	return (needed);
1790 }
1791 
1792 void
1793 vdev_load(vdev_t *vd)
1794 {
1795 	/*
1796 	 * Recursively load all children.
1797 	 */
1798 	for (int c = 0; c < vd->vdev_children; c++)
1799 		vdev_load(vd->vdev_child[c]);
1800 
1801 	/*
1802 	 * If this is a top-level vdev, initialize its metaslabs.
1803 	 */
1804 	if (vd == vd->vdev_top && !vd->vdev_ishole &&
1805 	    (vd->vdev_ashift == 0 || vd->vdev_asize == 0 ||
1806 	    vdev_metaslab_init(vd, 0) != 0))
1807 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1808 		    VDEV_AUX_CORRUPT_DATA);
1809 
1810 	/*
1811 	 * If this is a leaf vdev, load its DTL.
1812 	 */
1813 	if (vd->vdev_ops->vdev_op_leaf && vdev_dtl_load(vd) != 0)
1814 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1815 		    VDEV_AUX_CORRUPT_DATA);
1816 }
1817 
1818 /*
1819  * The special vdev case is used for hot spares and l2cache devices.  Its
1820  * sole purpose it to set the vdev state for the associated vdev.  To do this,
1821  * we make sure that we can open the underlying device, then try to read the
1822  * label, and make sure that the label is sane and that it hasn't been
1823  * repurposed to another pool.
1824  */
1825 int
1826 vdev_validate_aux(vdev_t *vd)
1827 {
1828 	nvlist_t *label;
1829 	uint64_t guid, version;
1830 	uint64_t state;
1831 
1832 	if (!vdev_readable(vd))
1833 		return (0);
1834 
1835 	if ((label = vdev_label_read_config(vd)) == NULL) {
1836 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1837 		    VDEV_AUX_CORRUPT_DATA);
1838 		return (-1);
1839 	}
1840 
1841 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
1842 	    version > SPA_VERSION ||
1843 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
1844 	    guid != vd->vdev_guid ||
1845 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
1846 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1847 		    VDEV_AUX_CORRUPT_DATA);
1848 		nvlist_free(label);
1849 		return (-1);
1850 	}
1851 
1852 	/*
1853 	 * We don't actually check the pool state here.  If it's in fact in
1854 	 * use by another pool, we update this fact on the fly when requested.
1855 	 */
1856 	nvlist_free(label);
1857 	return (0);
1858 }
1859 
1860 void
1861 vdev_remove(vdev_t *vd, uint64_t txg)
1862 {
1863 	spa_t *spa = vd->vdev_spa;
1864 	objset_t *mos = spa->spa_meta_objset;
1865 	dmu_tx_t *tx;
1866 
1867 	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
1868 
1869 	if (vd->vdev_dtl_smo.smo_object) {
1870 		ASSERT3U(vd->vdev_dtl_smo.smo_alloc, ==, 0);
1871 		(void) dmu_object_free(mos, vd->vdev_dtl_smo.smo_object, tx);
1872 		vd->vdev_dtl_smo.smo_object = 0;
1873 	}
1874 
1875 	if (vd->vdev_ms != NULL) {
1876 		for (int m = 0; m < vd->vdev_ms_count; m++) {
1877 			metaslab_t *msp = vd->vdev_ms[m];
1878 
1879 			if (msp == NULL || msp->ms_smo.smo_object == 0)
1880 				continue;
1881 
1882 			ASSERT3U(msp->ms_smo.smo_alloc, ==, 0);
1883 			(void) dmu_object_free(mos, msp->ms_smo.smo_object, tx);
1884 			msp->ms_smo.smo_object = 0;
1885 		}
1886 	}
1887 
1888 	if (vd->vdev_ms_array) {
1889 		(void) dmu_object_free(mos, vd->vdev_ms_array, tx);
1890 		vd->vdev_ms_array = 0;
1891 		vd->vdev_ms_shift = 0;
1892 	}
1893 	dmu_tx_commit(tx);
1894 }
1895 
1896 void
1897 vdev_sync_done(vdev_t *vd, uint64_t txg)
1898 {
1899 	metaslab_t *msp;
1900 
1901 	ASSERT(!vd->vdev_ishole);
1902 
1903 	while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
1904 		metaslab_sync_done(msp, txg);
1905 }
1906 
1907 void
1908 vdev_sync(vdev_t *vd, uint64_t txg)
1909 {
1910 	spa_t *spa = vd->vdev_spa;
1911 	vdev_t *lvd;
1912 	metaslab_t *msp;
1913 	dmu_tx_t *tx;
1914 
1915 	ASSERT(!vd->vdev_ishole);
1916 
1917 	if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0) {
1918 		ASSERT(vd == vd->vdev_top);
1919 		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1920 		vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
1921 		    DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
1922 		ASSERT(vd->vdev_ms_array != 0);
1923 		vdev_config_dirty(vd);
1924 		dmu_tx_commit(tx);
1925 	}
1926 
1927 	if (vd->vdev_removing)
1928 		vdev_remove(vd, txg);
1929 
1930 	while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
1931 		metaslab_sync(msp, txg);
1932 		(void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
1933 	}
1934 
1935 	while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
1936 		vdev_dtl_sync(lvd, txg);
1937 
1938 	(void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
1939 }
1940 
1941 uint64_t
1942 vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
1943 {
1944 	return (vd->vdev_ops->vdev_op_asize(vd, psize));
1945 }
1946 
1947 /*
1948  * Mark the given vdev faulted.  A faulted vdev behaves as if the device could
1949  * not be opened, and no I/O is attempted.
1950  */
1951 int
1952 vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
1953 {
1954 	vdev_t *vd;
1955 
1956 	spa_vdev_state_enter(spa, SCL_NONE);
1957 
1958 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
1959 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
1960 
1961 	if (!vd->vdev_ops->vdev_op_leaf)
1962 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
1963 
1964 	/*
1965 	 * We don't directly use the aux state here, but if we do a
1966 	 * vdev_reopen(), we need this value to be present to remember why we
1967 	 * were faulted.
1968 	 */
1969 	vd->vdev_label_aux = aux;
1970 
1971 	/*
1972 	 * Faulted state takes precedence over degraded.
1973 	 */
1974 	vd->vdev_faulted = 1ULL;
1975 	vd->vdev_degraded = 0ULL;
1976 	vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
1977 
1978 	/*
1979 	 * If marking the vdev as faulted cause the top-level vdev to become
1980 	 * unavailable, then back off and simply mark the vdev as degraded
1981 	 * instead.
1982 	 */
1983 	if (vdev_is_dead(vd->vdev_top) && !vd->vdev_islog &&
1984 	    vd->vdev_aux == NULL) {
1985 		vd->vdev_degraded = 1ULL;
1986 		vd->vdev_faulted = 0ULL;
1987 
1988 		/*
1989 		 * If we reopen the device and it's not dead, only then do we
1990 		 * mark it degraded.
1991 		 */
1992 		vdev_reopen(vd);
1993 
1994 		if (vdev_readable(vd))
1995 			vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
1996 	}
1997 
1998 	return (spa_vdev_state_exit(spa, vd, 0));
1999 }
2000 
2001 /*
2002  * Mark the given vdev degraded.  A degraded vdev is purely an indication to the
2003  * user that something is wrong.  The vdev continues to operate as normal as far
2004  * as I/O is concerned.
2005  */
2006 int
2007 vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2008 {
2009 	vdev_t *vd;
2010 
2011 	spa_vdev_state_enter(spa, SCL_NONE);
2012 
2013 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2014 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2015 
2016 	if (!vd->vdev_ops->vdev_op_leaf)
2017 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2018 
2019 	/*
2020 	 * If the vdev is already faulted, then don't do anything.
2021 	 */
2022 	if (vd->vdev_faulted || vd->vdev_degraded)
2023 		return (spa_vdev_state_exit(spa, NULL, 0));
2024 
2025 	vd->vdev_degraded = 1ULL;
2026 	if (!vdev_is_dead(vd))
2027 		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
2028 		    aux);
2029 
2030 	return (spa_vdev_state_exit(spa, vd, 0));
2031 }
2032 
2033 /*
2034  * Online the given vdev.  If 'unspare' is set, it implies two things.  First,
2035  * any attached spare device should be detached when the device finishes
2036  * resilvering.  Second, the online should be treated like a 'test' online case,
2037  * so no FMA events are generated if the device fails to open.
2038  */
2039 int
2040 vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
2041 {
2042 	vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
2043 
2044 	spa_vdev_state_enter(spa, SCL_NONE);
2045 
2046 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2047 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2048 
2049 	if (!vd->vdev_ops->vdev_op_leaf)
2050 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2051 
2052 	tvd = vd->vdev_top;
2053 	vd->vdev_offline = B_FALSE;
2054 	vd->vdev_tmpoffline = B_FALSE;
2055 	vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
2056 	vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
2057 
2058 	/* XXX - L2ARC 1.0 does not support expansion */
2059 	if (!vd->vdev_aux) {
2060 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2061 			pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
2062 	}
2063 
2064 	vdev_reopen(tvd);
2065 	vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
2066 
2067 	if (!vd->vdev_aux) {
2068 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2069 			pvd->vdev_expanding = B_FALSE;
2070 	}
2071 
2072 	if (newstate)
2073 		*newstate = vd->vdev_state;
2074 	if ((flags & ZFS_ONLINE_UNSPARE) &&
2075 	    !vdev_is_dead(vd) && vd->vdev_parent &&
2076 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2077 	    vd->vdev_parent->vdev_child[0] == vd)
2078 		vd->vdev_unspare = B_TRUE;
2079 
2080 	if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
2081 
2082 		/* XXX - L2ARC 1.0 does not support expansion */
2083 		if (vd->vdev_aux)
2084 			return (spa_vdev_state_exit(spa, vd, ENOTSUP));
2085 		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
2086 	}
2087 	return (spa_vdev_state_exit(spa, vd, 0));
2088 }
2089 
2090 int
2091 vdev_offline_log(spa_t *spa)
2092 {
2093 	int error = 0;
2094 
2095 	if ((error = dmu_objset_find(spa_name(spa), zil_vdev_offline,
2096 	    NULL, DS_FIND_CHILDREN)) == 0) {
2097 
2098 		/*
2099 		 * We successfully offlined the log device, sync out the
2100 		 * current txg so that the "stubby" block can be removed
2101 		 * by zil_sync().
2102 		 */
2103 		txg_wait_synced(spa->spa_dsl_pool, 0);
2104 	}
2105 	return (error);
2106 }
2107 
2108 int
2109 vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
2110 {
2111 	vdev_t *vd, *tvd;
2112 	int error = 0;
2113 	uint64_t generation;
2114 	metaslab_group_t *mg;
2115 
2116 top:
2117 	spa_vdev_state_enter(spa, SCL_ALLOC);
2118 
2119 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2120 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2121 
2122 	if (!vd->vdev_ops->vdev_op_leaf)
2123 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2124 
2125 	tvd = vd->vdev_top;
2126 	mg = tvd->vdev_mg;
2127 	generation = spa->spa_config_generation + 1;
2128 
2129 	/*
2130 	 * If the device isn't already offline, try to offline it.
2131 	 */
2132 	if (!vd->vdev_offline) {
2133 		/*
2134 		 * If this device has the only valid copy of some data,
2135 		 * don't allow it to be offlined. Log devices are always
2136 		 * expendable.
2137 		 */
2138 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2139 		    vdev_dtl_required(vd))
2140 			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2141 
2142 		/*
2143 		 * If the top-level is a slog and it's had allocations
2144 		 * then proceed. We check that the vdev's metaslab
2145 		 * grop is not NULL since it's possible that we may
2146 		 * have just added this vdev and have not yet initialized
2147 		 * it's metaslabs.
2148 		 */
2149 		if (tvd->vdev_islog && mg != NULL) {
2150 			/*
2151 			 * Prevent any future allocations.
2152 			 */
2153 			metaslab_class_remove(spa->spa_log_class, mg);
2154 			(void) spa_vdev_state_exit(spa, vd, 0);
2155 
2156 			error = vdev_offline_log(spa);
2157 
2158 			spa_vdev_state_enter(spa, SCL_ALLOC);
2159 
2160 			/*
2161 			 * Check to see if the config has changed.
2162 			 */
2163 			if (error || generation != spa->spa_config_generation) {
2164 				metaslab_class_add(spa->spa_log_class, mg);
2165 				if (error)
2166 					return (spa_vdev_state_exit(spa,
2167 					    vd, error));
2168 				(void) spa_vdev_state_exit(spa, vd, 0);
2169 				goto top;
2170 			}
2171 			ASSERT3U(tvd->vdev_stat.vs_alloc, ==, 0);
2172 		}
2173 
2174 		/*
2175 		 * Offline this device and reopen its top-level vdev.
2176 		 * If the top-level vdev is a log device then just offline
2177 		 * it. Otherwise, if this action results in the top-level
2178 		 * vdev becoming unusable, undo it and fail the request.
2179 		 */
2180 		vd->vdev_offline = B_TRUE;
2181 		vdev_reopen(tvd);
2182 
2183 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2184 		    vdev_is_dead(tvd)) {
2185 			vd->vdev_offline = B_FALSE;
2186 			vdev_reopen(tvd);
2187 			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2188 		}
2189 
2190 		/*
2191 		 * Add the device back into the metaslab rotor so that
2192 		 * once we online the device it's open for business.
2193 		 */
2194 		if (tvd->vdev_islog && mg != NULL)
2195 			metaslab_class_add(spa->spa_log_class, mg);
2196 	}
2197 
2198 	vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
2199 
2200 	return (spa_vdev_state_exit(spa, vd, 0));
2201 }
2202 
2203 /*
2204  * Clear the error counts associated with this vdev.  Unlike vdev_online() and
2205  * vdev_offline(), we assume the spa config is locked.  We also clear all
2206  * children.  If 'vd' is NULL, then the user wants to clear all vdevs.
2207  */
2208 void
2209 vdev_clear(spa_t *spa, vdev_t *vd)
2210 {
2211 	vdev_t *rvd = spa->spa_root_vdev;
2212 
2213 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2214 
2215 	if (vd == NULL)
2216 		vd = rvd;
2217 
2218 	vd->vdev_stat.vs_read_errors = 0;
2219 	vd->vdev_stat.vs_write_errors = 0;
2220 	vd->vdev_stat.vs_checksum_errors = 0;
2221 
2222 	for (int c = 0; c < vd->vdev_children; c++)
2223 		vdev_clear(spa, vd->vdev_child[c]);
2224 
2225 	/*
2226 	 * If we're in the FAULTED state or have experienced failed I/O, then
2227 	 * clear the persistent state and attempt to reopen the device.  We
2228 	 * also mark the vdev config dirty, so that the new faulted state is
2229 	 * written out to disk.
2230 	 */
2231 	if (vd->vdev_faulted || vd->vdev_degraded ||
2232 	    !vdev_readable(vd) || !vdev_writeable(vd)) {
2233 
2234 		vd->vdev_faulted = vd->vdev_degraded = 0;
2235 		vd->vdev_cant_read = B_FALSE;
2236 		vd->vdev_cant_write = B_FALSE;
2237 
2238 		vdev_reopen(vd);
2239 
2240 		if (vd != rvd)
2241 			vdev_state_dirty(vd->vdev_top);
2242 
2243 		if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
2244 			spa_async_request(spa, SPA_ASYNC_RESILVER);
2245 
2246 		spa_event_notify(spa, vd, ESC_ZFS_VDEV_CLEAR);
2247 	}
2248 }
2249 
2250 boolean_t
2251 vdev_is_dead(vdev_t *vd)
2252 {
2253 	/*
2254 	 * Holes and missing devices are always considered "dead".
2255 	 * This simplifies the code since we don't have to check for
2256 	 * these types of devices in the various code paths.
2257 	 * Instead we rely on the fact that we skip over dead devices
2258 	 * before issuing I/O to them.
2259 	 */
2260 	return (vd->vdev_state < VDEV_STATE_DEGRADED || vd->vdev_ishole ||
2261 	    vd->vdev_ops == &vdev_missing_ops);
2262 }
2263 
2264 boolean_t
2265 vdev_readable(vdev_t *vd)
2266 {
2267 	return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
2268 }
2269 
2270 boolean_t
2271 vdev_writeable(vdev_t *vd)
2272 {
2273 	return (!vdev_is_dead(vd) && !vd->vdev_cant_write);
2274 }
2275 
2276 boolean_t
2277 vdev_allocatable(vdev_t *vd)
2278 {
2279 	uint64_t state = vd->vdev_state;
2280 
2281 	/*
2282 	 * We currently allow allocations from vdevs which may be in the
2283 	 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
2284 	 * fails to reopen then we'll catch it later when we're holding
2285 	 * the proper locks.  Note that we have to get the vdev state
2286 	 * in a local variable because although it changes atomically,
2287 	 * we're asking two separate questions about it.
2288 	 */
2289 	return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
2290 	    !vd->vdev_cant_write && !vd->vdev_ishole && !vd->vdev_removing);
2291 }
2292 
2293 boolean_t
2294 vdev_accessible(vdev_t *vd, zio_t *zio)
2295 {
2296 	ASSERT(zio->io_vd == vd);
2297 
2298 	if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
2299 		return (B_FALSE);
2300 
2301 	if (zio->io_type == ZIO_TYPE_READ)
2302 		return (!vd->vdev_cant_read);
2303 
2304 	if (zio->io_type == ZIO_TYPE_WRITE)
2305 		return (!vd->vdev_cant_write);
2306 
2307 	return (B_TRUE);
2308 }
2309 
2310 /*
2311  * Get statistics for the given vdev.
2312  */
2313 void
2314 vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
2315 {
2316 	vdev_t *rvd = vd->vdev_spa->spa_root_vdev;
2317 
2318 	mutex_enter(&vd->vdev_stat_lock);
2319 	bcopy(&vd->vdev_stat, vs, sizeof (*vs));
2320 	vs->vs_scrub_errors = vd->vdev_spa->spa_scrub_errors;
2321 	vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
2322 	vs->vs_state = vd->vdev_state;
2323 	vs->vs_rsize = vdev_get_min_asize(vd);
2324 	if (vd->vdev_ops->vdev_op_leaf)
2325 		vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
2326 	mutex_exit(&vd->vdev_stat_lock);
2327 
2328 	/*
2329 	 * If we're getting stats on the root vdev, aggregate the I/O counts
2330 	 * over all top-level vdevs (i.e. the direct children of the root).
2331 	 */
2332 	if (vd == rvd) {
2333 		for (int c = 0; c < rvd->vdev_children; c++) {
2334 			vdev_t *cvd = rvd->vdev_child[c];
2335 			vdev_stat_t *cvs = &cvd->vdev_stat;
2336 
2337 			mutex_enter(&vd->vdev_stat_lock);
2338 			for (int t = 0; t < ZIO_TYPES; t++) {
2339 				vs->vs_ops[t] += cvs->vs_ops[t];
2340 				vs->vs_bytes[t] += cvs->vs_bytes[t];
2341 			}
2342 			vs->vs_scrub_examined += cvs->vs_scrub_examined;
2343 			mutex_exit(&vd->vdev_stat_lock);
2344 		}
2345 	}
2346 }
2347 
2348 void
2349 vdev_clear_stats(vdev_t *vd)
2350 {
2351 	mutex_enter(&vd->vdev_stat_lock);
2352 	vd->vdev_stat.vs_space = 0;
2353 	vd->vdev_stat.vs_dspace = 0;
2354 	vd->vdev_stat.vs_alloc = 0;
2355 	mutex_exit(&vd->vdev_stat_lock);
2356 }
2357 
2358 void
2359 vdev_stat_update(zio_t *zio, uint64_t psize)
2360 {
2361 	spa_t *spa = zio->io_spa;
2362 	vdev_t *rvd = spa->spa_root_vdev;
2363 	vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
2364 	vdev_t *pvd;
2365 	uint64_t txg = zio->io_txg;
2366 	vdev_stat_t *vs = &vd->vdev_stat;
2367 	zio_type_t type = zio->io_type;
2368 	int flags = zio->io_flags;
2369 
2370 	/*
2371 	 * If this i/o is a gang leader, it didn't do any actual work.
2372 	 */
2373 	if (zio->io_gang_tree)
2374 		return;
2375 
2376 	if (zio->io_error == 0) {
2377 		/*
2378 		 * If this is a root i/o, don't count it -- we've already
2379 		 * counted the top-level vdevs, and vdev_get_stats() will
2380 		 * aggregate them when asked.  This reduces contention on
2381 		 * the root vdev_stat_lock and implicitly handles blocks
2382 		 * that compress away to holes, for which there is no i/o.
2383 		 * (Holes never create vdev children, so all the counters
2384 		 * remain zero, which is what we want.)
2385 		 *
2386 		 * Note: this only applies to successful i/o (io_error == 0)
2387 		 * because unlike i/o counts, errors are not additive.
2388 		 * When reading a ditto block, for example, failure of
2389 		 * one top-level vdev does not imply a root-level error.
2390 		 */
2391 		if (vd == rvd)
2392 			return;
2393 
2394 		ASSERT(vd == zio->io_vd);
2395 
2396 		if (flags & ZIO_FLAG_IO_BYPASS)
2397 			return;
2398 
2399 		mutex_enter(&vd->vdev_stat_lock);
2400 
2401 		if (flags & ZIO_FLAG_IO_REPAIR) {
2402 			if (flags & ZIO_FLAG_SCRUB_THREAD)
2403 				vs->vs_scrub_repaired += psize;
2404 			if (flags & ZIO_FLAG_SELF_HEAL)
2405 				vs->vs_self_healed += psize;
2406 		}
2407 
2408 		vs->vs_ops[type]++;
2409 		vs->vs_bytes[type] += psize;
2410 
2411 		mutex_exit(&vd->vdev_stat_lock);
2412 		return;
2413 	}
2414 
2415 	if (flags & ZIO_FLAG_SPECULATIVE)
2416 		return;
2417 
2418 	/*
2419 	 * If this is an I/O error that is going to be retried, then ignore the
2420 	 * error.  Otherwise, the user may interpret B_FAILFAST I/O errors as
2421 	 * hard errors, when in reality they can happen for any number of
2422 	 * innocuous reasons (bus resets, MPxIO link failure, etc).
2423 	 */
2424 	if (zio->io_error == EIO &&
2425 	    !(zio->io_flags & ZIO_FLAG_IO_RETRY))
2426 		return;
2427 
2428 	/*
2429 	 * Intent logs writes won't propagate their error to the root
2430 	 * I/O so don't mark these types of failures as pool-level
2431 	 * errors.
2432 	 */
2433 	if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
2434 		return;
2435 
2436 	mutex_enter(&vd->vdev_stat_lock);
2437 	if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
2438 		if (zio->io_error == ECKSUM)
2439 			vs->vs_checksum_errors++;
2440 		else
2441 			vs->vs_read_errors++;
2442 	}
2443 	if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
2444 		vs->vs_write_errors++;
2445 	mutex_exit(&vd->vdev_stat_lock);
2446 
2447 	if (type == ZIO_TYPE_WRITE && txg != 0 &&
2448 	    (!(flags & ZIO_FLAG_IO_REPAIR) ||
2449 	    (flags & ZIO_FLAG_SCRUB_THREAD))) {
2450 		/*
2451 		 * This is either a normal write (not a repair), or it's a
2452 		 * repair induced by the scrub thread.  In the normal case,
2453 		 * we commit the DTL change in the same txg as the block
2454 		 * was born.  In the scrub-induced repair case, we know that
2455 		 * scrubs run in first-pass syncing context, so we commit
2456 		 * the DTL change in spa->spa_syncing_txg.
2457 		 *
2458 		 * We currently do not make DTL entries for failed spontaneous
2459 		 * self-healing writes triggered by normal (non-scrubbing)
2460 		 * reads, because we have no transactional context in which to
2461 		 * do so -- and it's not clear that it'd be desirable anyway.
2462 		 */
2463 		if (vd->vdev_ops->vdev_op_leaf) {
2464 			uint64_t commit_txg = txg;
2465 			if (flags & ZIO_FLAG_SCRUB_THREAD) {
2466 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2467 				ASSERT(spa_sync_pass(spa) == 1);
2468 				vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
2469 				commit_txg = spa->spa_syncing_txg;
2470 			}
2471 			ASSERT(commit_txg >= spa->spa_syncing_txg);
2472 			if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
2473 				return;
2474 			for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2475 				vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
2476 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
2477 		}
2478 		if (vd != rvd)
2479 			vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
2480 	}
2481 }
2482 
2483 void
2484 vdev_scrub_stat_update(vdev_t *vd, pool_scrub_type_t type, boolean_t complete)
2485 {
2486 	vdev_stat_t *vs = &vd->vdev_stat;
2487 
2488 	for (int c = 0; c < vd->vdev_children; c++)
2489 		vdev_scrub_stat_update(vd->vdev_child[c], type, complete);
2490 
2491 	mutex_enter(&vd->vdev_stat_lock);
2492 
2493 	if (type == POOL_SCRUB_NONE) {
2494 		/*
2495 		 * Update completion and end time.  Leave everything else alone
2496 		 * so we can report what happened during the previous scrub.
2497 		 */
2498 		vs->vs_scrub_complete = complete;
2499 		vs->vs_scrub_end = gethrestime_sec();
2500 	} else {
2501 		vs->vs_scrub_type = type;
2502 		vs->vs_scrub_complete = 0;
2503 		vs->vs_scrub_examined = 0;
2504 		vs->vs_scrub_repaired = 0;
2505 		vs->vs_scrub_start = gethrestime_sec();
2506 		vs->vs_scrub_end = 0;
2507 	}
2508 
2509 	mutex_exit(&vd->vdev_stat_lock);
2510 }
2511 
2512 /*
2513  * Update the in-core space usage stats for this vdev and the root vdev.
2514  */
2515 void
2516 vdev_space_update(vdev_t *vd, int64_t space_delta, int64_t alloc_delta,
2517     boolean_t update_root)
2518 {
2519 	int64_t dspace_delta = space_delta;
2520 	spa_t *spa = vd->vdev_spa;
2521 	vdev_t *rvd = spa->spa_root_vdev;
2522 
2523 	ASSERT(vd == vd->vdev_top);
2524 
2525 	/*
2526 	 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
2527 	 * factor.  We must calculate this here and not at the root vdev
2528 	 * because the root vdev's psize-to-asize is simply the max of its
2529 	 * childrens', thus not accurate enough for us.
2530 	 */
2531 	ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
2532 	ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
2533 	dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
2534 	    vd->vdev_deflate_ratio;
2535 
2536 	mutex_enter(&vd->vdev_stat_lock);
2537 	vd->vdev_stat.vs_space += space_delta;
2538 	vd->vdev_stat.vs_alloc += alloc_delta;
2539 	vd->vdev_stat.vs_dspace += dspace_delta;
2540 	mutex_exit(&vd->vdev_stat_lock);
2541 
2542 	if (update_root) {
2543 		ASSERT(rvd == vd->vdev_parent);
2544 		ASSERT(vd->vdev_ms_count != 0);
2545 
2546 		/*
2547 		 * Don't count non-normal (e.g. intent log) space as part of
2548 		 * the pool's capacity.
2549 		 */
2550 		if (vd->vdev_islog)
2551 			return;
2552 
2553 		mutex_enter(&rvd->vdev_stat_lock);
2554 		rvd->vdev_stat.vs_space += space_delta;
2555 		rvd->vdev_stat.vs_alloc += alloc_delta;
2556 		rvd->vdev_stat.vs_dspace += dspace_delta;
2557 		mutex_exit(&rvd->vdev_stat_lock);
2558 	}
2559 }
2560 
2561 /*
2562  * Mark a top-level vdev's config as dirty, placing it on the dirty list
2563  * so that it will be written out next time the vdev configuration is synced.
2564  * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
2565  */
2566 void
2567 vdev_config_dirty(vdev_t *vd)
2568 {
2569 	spa_t *spa = vd->vdev_spa;
2570 	vdev_t *rvd = spa->spa_root_vdev;
2571 	int c;
2572 
2573 	/*
2574 	 * If this is an aux vdev (as with l2cache and spare devices), then we
2575 	 * update the vdev config manually and set the sync flag.
2576 	 */
2577 	if (vd->vdev_aux != NULL) {
2578 		spa_aux_vdev_t *sav = vd->vdev_aux;
2579 		nvlist_t **aux;
2580 		uint_t naux;
2581 
2582 		for (c = 0; c < sav->sav_count; c++) {
2583 			if (sav->sav_vdevs[c] == vd)
2584 				break;
2585 		}
2586 
2587 		if (c == sav->sav_count) {
2588 			/*
2589 			 * We're being removed.  There's nothing more to do.
2590 			 */
2591 			ASSERT(sav->sav_sync == B_TRUE);
2592 			return;
2593 		}
2594 
2595 		sav->sav_sync = B_TRUE;
2596 
2597 		if (nvlist_lookup_nvlist_array(sav->sav_config,
2598 		    ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
2599 			VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
2600 			    ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
2601 		}
2602 
2603 		ASSERT(c < naux);
2604 
2605 		/*
2606 		 * Setting the nvlist in the middle if the array is a little
2607 		 * sketchy, but it will work.
2608 		 */
2609 		nvlist_free(aux[c]);
2610 		aux[c] = vdev_config_generate(spa, vd, B_TRUE, B_FALSE, B_TRUE);
2611 
2612 		return;
2613 	}
2614 
2615 	/*
2616 	 * The dirty list is protected by the SCL_CONFIG lock.  The caller
2617 	 * must either hold SCL_CONFIG as writer, or must be the sync thread
2618 	 * (which holds SCL_CONFIG as reader).  There's only one sync thread,
2619 	 * so this is sufficient to ensure mutual exclusion.
2620 	 */
2621 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2622 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2623 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
2624 
2625 	if (vd == rvd) {
2626 		for (c = 0; c < rvd->vdev_children; c++)
2627 			vdev_config_dirty(rvd->vdev_child[c]);
2628 	} else {
2629 		ASSERT(vd == vd->vdev_top);
2630 
2631 		if (!list_link_active(&vd->vdev_config_dirty_node) &&
2632 		    !vd->vdev_ishole)
2633 			list_insert_head(&spa->spa_config_dirty_list, vd);
2634 	}
2635 }
2636 
2637 void
2638 vdev_config_clean(vdev_t *vd)
2639 {
2640 	spa_t *spa = vd->vdev_spa;
2641 
2642 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2643 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2644 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
2645 
2646 	ASSERT(list_link_active(&vd->vdev_config_dirty_node));
2647 	list_remove(&spa->spa_config_dirty_list, vd);
2648 }
2649 
2650 /*
2651  * Mark a top-level vdev's state as dirty, so that the next pass of
2652  * spa_sync() can convert this into vdev_config_dirty().  We distinguish
2653  * the state changes from larger config changes because they require
2654  * much less locking, and are often needed for administrative actions.
2655  */
2656 void
2657 vdev_state_dirty(vdev_t *vd)
2658 {
2659 	spa_t *spa = vd->vdev_spa;
2660 
2661 	ASSERT(vd == vd->vdev_top);
2662 
2663 	/*
2664 	 * The state list is protected by the SCL_STATE lock.  The caller
2665 	 * must either hold SCL_STATE as writer, or must be the sync thread
2666 	 * (which holds SCL_STATE as reader).  There's only one sync thread,
2667 	 * so this is sufficient to ensure mutual exclusion.
2668 	 */
2669 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2670 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2671 	    spa_config_held(spa, SCL_STATE, RW_READER)));
2672 
2673 	if (!list_link_active(&vd->vdev_state_dirty_node))
2674 		list_insert_head(&spa->spa_state_dirty_list, vd);
2675 }
2676 
2677 void
2678 vdev_state_clean(vdev_t *vd)
2679 {
2680 	spa_t *spa = vd->vdev_spa;
2681 
2682 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2683 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2684 	    spa_config_held(spa, SCL_STATE, RW_READER)));
2685 
2686 	ASSERT(list_link_active(&vd->vdev_state_dirty_node));
2687 	list_remove(&spa->spa_state_dirty_list, vd);
2688 }
2689 
2690 /*
2691  * Propagate vdev state up from children to parent.
2692  */
2693 void
2694 vdev_propagate_state(vdev_t *vd)
2695 {
2696 	spa_t *spa = vd->vdev_spa;
2697 	vdev_t *rvd = spa->spa_root_vdev;
2698 	int degraded = 0, faulted = 0;
2699 	int corrupted = 0;
2700 	vdev_t *child;
2701 
2702 	if (vd->vdev_children > 0) {
2703 		for (int c = 0; c < vd->vdev_children; c++) {
2704 			child = vd->vdev_child[c];
2705 
2706 			/*
2707 			 * Don't factor holes into the decision.
2708 			 */
2709 			if (child->vdev_ishole)
2710 				continue;
2711 
2712 			if (!vdev_readable(child) ||
2713 			    (!vdev_writeable(child) && spa_writeable(spa))) {
2714 				/*
2715 				 * Root special: if there is a top-level log
2716 				 * device, treat the root vdev as if it were
2717 				 * degraded.
2718 				 */
2719 				if (child->vdev_islog && vd == rvd)
2720 					degraded++;
2721 				else
2722 					faulted++;
2723 			} else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
2724 				degraded++;
2725 			}
2726 
2727 			if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
2728 				corrupted++;
2729 		}
2730 
2731 		vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
2732 
2733 		/*
2734 		 * Root special: if there is a top-level vdev that cannot be
2735 		 * opened due to corrupted metadata, then propagate the root
2736 		 * vdev's aux state as 'corrupt' rather than 'insufficient
2737 		 * replicas'.
2738 		 */
2739 		if (corrupted && vd == rvd &&
2740 		    rvd->vdev_state == VDEV_STATE_CANT_OPEN)
2741 			vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
2742 			    VDEV_AUX_CORRUPT_DATA);
2743 	}
2744 
2745 	if (vd->vdev_parent)
2746 		vdev_propagate_state(vd->vdev_parent);
2747 }
2748 
2749 /*
2750  * Set a vdev's state.  If this is during an open, we don't update the parent
2751  * state, because we're in the process of opening children depth-first.
2752  * Otherwise, we propagate the change to the parent.
2753  *
2754  * If this routine places a device in a faulted state, an appropriate ereport is
2755  * generated.
2756  */
2757 void
2758 vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
2759 {
2760 	uint64_t save_state;
2761 	spa_t *spa = vd->vdev_spa;
2762 
2763 	if (state == vd->vdev_state) {
2764 		vd->vdev_stat.vs_aux = aux;
2765 		return;
2766 	}
2767 
2768 	save_state = vd->vdev_state;
2769 
2770 	vd->vdev_state = state;
2771 	vd->vdev_stat.vs_aux = aux;
2772 
2773 	/*
2774 	 * If we are setting the vdev state to anything but an open state, then
2775 	 * always close the underlying device.  Otherwise, we keep accessible
2776 	 * but invalid devices open forever.  We don't call vdev_close() itself,
2777 	 * because that implies some extra checks (offline, etc) that we don't
2778 	 * want here.  This is limited to leaf devices, because otherwise
2779 	 * closing the device will affect other children.
2780 	 */
2781 	if (vdev_is_dead(vd) && vd->vdev_ops->vdev_op_leaf)
2782 		vd->vdev_ops->vdev_op_close(vd);
2783 
2784 	/*
2785 	 * If we have brought this vdev back into service, we need
2786 	 * to notify fmd so that it can gracefully repair any outstanding
2787 	 * cases due to a missing device.  We do this in all cases, even those
2788 	 * that probably don't correlate to a repaired fault.  This is sure to
2789 	 * catch all cases, and we let the zfs-retire agent sort it out.  If
2790 	 * this is a transient state it's OK, as the retire agent will
2791 	 * double-check the state of the vdev before repairing it.
2792 	 */
2793 	if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf &&
2794 	    vd->vdev_prevstate != state)
2795 		zfs_post_state_change(spa, vd);
2796 
2797 	if (vd->vdev_removed &&
2798 	    state == VDEV_STATE_CANT_OPEN &&
2799 	    (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
2800 		/*
2801 		 * If the previous state is set to VDEV_STATE_REMOVED, then this
2802 		 * device was previously marked removed and someone attempted to
2803 		 * reopen it.  If this failed due to a nonexistent device, then
2804 		 * keep the device in the REMOVED state.  We also let this be if
2805 		 * it is one of our special test online cases, which is only
2806 		 * attempting to online the device and shouldn't generate an FMA
2807 		 * fault.
2808 		 */
2809 		vd->vdev_state = VDEV_STATE_REMOVED;
2810 		vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
2811 	} else if (state == VDEV_STATE_REMOVED) {
2812 		vd->vdev_removed = B_TRUE;
2813 	} else if (state == VDEV_STATE_CANT_OPEN) {
2814 		/*
2815 		 * If we fail to open a vdev during an import, we mark it as
2816 		 * "not available", which signifies that it was never there to
2817 		 * begin with.  Failure to open such a device is not considered
2818 		 * an error.
2819 		 */
2820 		if (spa->spa_load_state == SPA_LOAD_IMPORT &&
2821 		    vd->vdev_ops->vdev_op_leaf)
2822 			vd->vdev_not_present = 1;
2823 
2824 		/*
2825 		 * Post the appropriate ereport.  If the 'prevstate' field is
2826 		 * set to something other than VDEV_STATE_UNKNOWN, it indicates
2827 		 * that this is part of a vdev_reopen().  In this case, we don't
2828 		 * want to post the ereport if the device was already in the
2829 		 * CANT_OPEN state beforehand.
2830 		 *
2831 		 * If the 'checkremove' flag is set, then this is an attempt to
2832 		 * online the device in response to an insertion event.  If we
2833 		 * hit this case, then we have detected an insertion event for a
2834 		 * faulted or offline device that wasn't in the removed state.
2835 		 * In this scenario, we don't post an ereport because we are
2836 		 * about to replace the device, or attempt an online with
2837 		 * vdev_forcefault, which will generate the fault for us.
2838 		 */
2839 		if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
2840 		    !vd->vdev_not_present && !vd->vdev_checkremove &&
2841 		    vd != spa->spa_root_vdev) {
2842 			const char *class;
2843 
2844 			switch (aux) {
2845 			case VDEV_AUX_OPEN_FAILED:
2846 				class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
2847 				break;
2848 			case VDEV_AUX_CORRUPT_DATA:
2849 				class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
2850 				break;
2851 			case VDEV_AUX_NO_REPLICAS:
2852 				class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
2853 				break;
2854 			case VDEV_AUX_BAD_GUID_SUM:
2855 				class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
2856 				break;
2857 			case VDEV_AUX_TOO_SMALL:
2858 				class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
2859 				break;
2860 			case VDEV_AUX_BAD_LABEL:
2861 				class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
2862 				break;
2863 			case VDEV_AUX_IO_FAILURE:
2864 				class = FM_EREPORT_ZFS_IO_FAILURE;
2865 				break;
2866 			default:
2867 				class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
2868 			}
2869 
2870 			zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
2871 		}
2872 
2873 		/* Erase any notion of persistent removed state */
2874 		vd->vdev_removed = B_FALSE;
2875 	} else {
2876 		vd->vdev_removed = B_FALSE;
2877 	}
2878 
2879 	if (!isopen && vd->vdev_parent)
2880 		vdev_propagate_state(vd->vdev_parent);
2881 }
2882 
2883 /*
2884  * Check the vdev configuration to ensure that it's capable of supporting
2885  * a root pool. Currently, we do not support RAID-Z or partial configuration.
2886  * In addition, only a single top-level vdev is allowed and none of the leaves
2887  * can be wholedisks.
2888  */
2889 boolean_t
2890 vdev_is_bootable(vdev_t *vd)
2891 {
2892 	if (!vd->vdev_ops->vdev_op_leaf) {
2893 		char *vdev_type = vd->vdev_ops->vdev_op_type;
2894 
2895 		if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
2896 		    vd->vdev_children > 1) {
2897 			return (B_FALSE);
2898 		} else if (strcmp(vdev_type, VDEV_TYPE_RAIDZ) == 0 ||
2899 		    strcmp(vdev_type, VDEV_TYPE_MISSING) == 0) {
2900 			return (B_FALSE);
2901 		}
2902 	} else if (vd->vdev_wholedisk == 1) {
2903 		return (B_FALSE);
2904 	}
2905 
2906 	for (int c = 0; c < vd->vdev_children; c++) {
2907 		if (!vdev_is_bootable(vd->vdev_child[c]))
2908 			return (B_FALSE);
2909 	}
2910 	return (B_TRUE);
2911 }
2912 
2913 /*
2914  * Load the state from the original vdev tree (ovd) which
2915  * we've retrieved from the MOS config object. If the original
2916  * vdev was offline then we transfer that state to the device
2917  * in the current vdev tree (nvd).
2918  */
2919 void
2920 vdev_load_log_state(vdev_t *nvd, vdev_t *ovd)
2921 {
2922 	spa_t *spa = nvd->vdev_spa;
2923 
2924 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2925 	ASSERT3U(nvd->vdev_guid, ==, ovd->vdev_guid);
2926 
2927 	for (int c = 0; c < nvd->vdev_children; c++)
2928 		vdev_load_log_state(nvd->vdev_child[c], ovd->vdev_child[c]);
2929 
2930 	if (nvd->vdev_ops->vdev_op_leaf && ovd->vdev_offline) {
2931 		/*
2932 		 * It would be nice to call vdev_offline()
2933 		 * directly but the pool isn't fully loaded and
2934 		 * the txg threads have not been started yet.
2935 		 */
2936 		nvd->vdev_offline = ovd->vdev_offline;
2937 		vdev_reopen(nvd->vdev_top);
2938 	}
2939 }
2940 
2941 /*
2942  * Expand a vdev if possible.
2943  */
2944 void
2945 vdev_expand(vdev_t *vd, uint64_t txg)
2946 {
2947 	ASSERT(vd->vdev_top == vd);
2948 	ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
2949 
2950 	if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count) {
2951 		VERIFY(vdev_metaslab_init(vd, txg) == 0);
2952 		vdev_config_dirty(vd);
2953 	}
2954 }
2955