xref: /illumos-gate/usr/src/uts/common/fs/zfs/vdev.c (revision 4263d13f00c9691fa14620eff82abef795be0693)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
25  * Copyright (c) 2012 by Delphix. All rights reserved.
26  */
27 
28 #include <sys/zfs_context.h>
29 #include <sys/fm/fs/zfs.h>
30 #include <sys/spa.h>
31 #include <sys/spa_impl.h>
32 #include <sys/dmu.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/vdev_impl.h>
35 #include <sys/uberblock_impl.h>
36 #include <sys/metaslab.h>
37 #include <sys/metaslab_impl.h>
38 #include <sys/space_map.h>
39 #include <sys/zio.h>
40 #include <sys/zap.h>
41 #include <sys/fs/zfs.h>
42 #include <sys/arc.h>
43 #include <sys/zil.h>
44 #include <sys/dsl_scan.h>
45 
46 /*
47  * Virtual device management.
48  */
49 
50 static vdev_ops_t *vdev_ops_table[] = {
51 	&vdev_root_ops,
52 	&vdev_raidz_ops,
53 	&vdev_mirror_ops,
54 	&vdev_replacing_ops,
55 	&vdev_spare_ops,
56 	&vdev_disk_ops,
57 	&vdev_file_ops,
58 	&vdev_missing_ops,
59 	&vdev_hole_ops,
60 	NULL
61 };
62 
63 /* maximum scrub/resilver I/O queue per leaf vdev */
64 int zfs_scrub_limit = 10;
65 
66 /*
67  * Given a vdev type, return the appropriate ops vector.
68  */
69 static vdev_ops_t *
70 vdev_getops(const char *type)
71 {
72 	vdev_ops_t *ops, **opspp;
73 
74 	for (opspp = vdev_ops_table; (ops = *opspp) != NULL; opspp++)
75 		if (strcmp(ops->vdev_op_type, type) == 0)
76 			break;
77 
78 	return (ops);
79 }
80 
81 /*
82  * Default asize function: return the MAX of psize with the asize of
83  * all children.  This is what's used by anything other than RAID-Z.
84  */
85 uint64_t
86 vdev_default_asize(vdev_t *vd, uint64_t psize)
87 {
88 	uint64_t asize = P2ROUNDUP(psize, 1ULL << vd->vdev_top->vdev_ashift);
89 	uint64_t csize;
90 
91 	for (int c = 0; c < vd->vdev_children; c++) {
92 		csize = vdev_psize_to_asize(vd->vdev_child[c], psize);
93 		asize = MAX(asize, csize);
94 	}
95 
96 	return (asize);
97 }
98 
99 /*
100  * Get the minimum allocatable size. We define the allocatable size as
101  * the vdev's asize rounded to the nearest metaslab. This allows us to
102  * replace or attach devices which don't have the same physical size but
103  * can still satisfy the same number of allocations.
104  */
105 uint64_t
106 vdev_get_min_asize(vdev_t *vd)
107 {
108 	vdev_t *pvd = vd->vdev_parent;
109 
110 	/*
111 	 * If our parent is NULL (inactive spare or cache) or is the root,
112 	 * just return our own asize.
113 	 */
114 	if (pvd == NULL)
115 		return (vd->vdev_asize);
116 
117 	/*
118 	 * The top-level vdev just returns the allocatable size rounded
119 	 * to the nearest metaslab.
120 	 */
121 	if (vd == vd->vdev_top)
122 		return (P2ALIGN(vd->vdev_asize, 1ULL << vd->vdev_ms_shift));
123 
124 	/*
125 	 * The allocatable space for a raidz vdev is N * sizeof(smallest child),
126 	 * so each child must provide at least 1/Nth of its asize.
127 	 */
128 	if (pvd->vdev_ops == &vdev_raidz_ops)
129 		return (pvd->vdev_min_asize / pvd->vdev_children);
130 
131 	return (pvd->vdev_min_asize);
132 }
133 
134 void
135 vdev_set_min_asize(vdev_t *vd)
136 {
137 	vd->vdev_min_asize = vdev_get_min_asize(vd);
138 
139 	for (int c = 0; c < vd->vdev_children; c++)
140 		vdev_set_min_asize(vd->vdev_child[c]);
141 }
142 
143 vdev_t *
144 vdev_lookup_top(spa_t *spa, uint64_t vdev)
145 {
146 	vdev_t *rvd = spa->spa_root_vdev;
147 
148 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
149 
150 	if (vdev < rvd->vdev_children) {
151 		ASSERT(rvd->vdev_child[vdev] != NULL);
152 		return (rvd->vdev_child[vdev]);
153 	}
154 
155 	return (NULL);
156 }
157 
158 vdev_t *
159 vdev_lookup_by_guid(vdev_t *vd, uint64_t guid)
160 {
161 	vdev_t *mvd;
162 
163 	if (vd->vdev_guid == guid)
164 		return (vd);
165 
166 	for (int c = 0; c < vd->vdev_children; c++)
167 		if ((mvd = vdev_lookup_by_guid(vd->vdev_child[c], guid)) !=
168 		    NULL)
169 			return (mvd);
170 
171 	return (NULL);
172 }
173 
174 void
175 vdev_add_child(vdev_t *pvd, vdev_t *cvd)
176 {
177 	size_t oldsize, newsize;
178 	uint64_t id = cvd->vdev_id;
179 	vdev_t **newchild;
180 
181 	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
182 	ASSERT(cvd->vdev_parent == NULL);
183 
184 	cvd->vdev_parent = pvd;
185 
186 	if (pvd == NULL)
187 		return;
188 
189 	ASSERT(id >= pvd->vdev_children || pvd->vdev_child[id] == NULL);
190 
191 	oldsize = pvd->vdev_children * sizeof (vdev_t *);
192 	pvd->vdev_children = MAX(pvd->vdev_children, id + 1);
193 	newsize = pvd->vdev_children * sizeof (vdev_t *);
194 
195 	newchild = kmem_zalloc(newsize, KM_SLEEP);
196 	if (pvd->vdev_child != NULL) {
197 		bcopy(pvd->vdev_child, newchild, oldsize);
198 		kmem_free(pvd->vdev_child, oldsize);
199 	}
200 
201 	pvd->vdev_child = newchild;
202 	pvd->vdev_child[id] = cvd;
203 
204 	cvd->vdev_top = (pvd->vdev_top ? pvd->vdev_top: cvd);
205 	ASSERT(cvd->vdev_top->vdev_parent->vdev_parent == NULL);
206 
207 	/*
208 	 * Walk up all ancestors to update guid sum.
209 	 */
210 	for (; pvd != NULL; pvd = pvd->vdev_parent)
211 		pvd->vdev_guid_sum += cvd->vdev_guid_sum;
212 }
213 
214 void
215 vdev_remove_child(vdev_t *pvd, vdev_t *cvd)
216 {
217 	int c;
218 	uint_t id = cvd->vdev_id;
219 
220 	ASSERT(cvd->vdev_parent == pvd);
221 
222 	if (pvd == NULL)
223 		return;
224 
225 	ASSERT(id < pvd->vdev_children);
226 	ASSERT(pvd->vdev_child[id] == cvd);
227 
228 	pvd->vdev_child[id] = NULL;
229 	cvd->vdev_parent = NULL;
230 
231 	for (c = 0; c < pvd->vdev_children; c++)
232 		if (pvd->vdev_child[c])
233 			break;
234 
235 	if (c == pvd->vdev_children) {
236 		kmem_free(pvd->vdev_child, c * sizeof (vdev_t *));
237 		pvd->vdev_child = NULL;
238 		pvd->vdev_children = 0;
239 	}
240 
241 	/*
242 	 * Walk up all ancestors to update guid sum.
243 	 */
244 	for (; pvd != NULL; pvd = pvd->vdev_parent)
245 		pvd->vdev_guid_sum -= cvd->vdev_guid_sum;
246 }
247 
248 /*
249  * Remove any holes in the child array.
250  */
251 void
252 vdev_compact_children(vdev_t *pvd)
253 {
254 	vdev_t **newchild, *cvd;
255 	int oldc = pvd->vdev_children;
256 	int newc;
257 
258 	ASSERT(spa_config_held(pvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
259 
260 	for (int c = newc = 0; c < oldc; c++)
261 		if (pvd->vdev_child[c])
262 			newc++;
263 
264 	newchild = kmem_alloc(newc * sizeof (vdev_t *), KM_SLEEP);
265 
266 	for (int c = newc = 0; c < oldc; c++) {
267 		if ((cvd = pvd->vdev_child[c]) != NULL) {
268 			newchild[newc] = cvd;
269 			cvd->vdev_id = newc++;
270 		}
271 	}
272 
273 	kmem_free(pvd->vdev_child, oldc * sizeof (vdev_t *));
274 	pvd->vdev_child = newchild;
275 	pvd->vdev_children = newc;
276 }
277 
278 /*
279  * Allocate and minimally initialize a vdev_t.
280  */
281 vdev_t *
282 vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
283 {
284 	vdev_t *vd;
285 
286 	vd = kmem_zalloc(sizeof (vdev_t), KM_SLEEP);
287 
288 	if (spa->spa_root_vdev == NULL) {
289 		ASSERT(ops == &vdev_root_ops);
290 		spa->spa_root_vdev = vd;
291 		spa->spa_load_guid = spa_generate_guid(NULL);
292 	}
293 
294 	if (guid == 0 && ops != &vdev_hole_ops) {
295 		if (spa->spa_root_vdev == vd) {
296 			/*
297 			 * The root vdev's guid will also be the pool guid,
298 			 * which must be unique among all pools.
299 			 */
300 			guid = spa_generate_guid(NULL);
301 		} else {
302 			/*
303 			 * Any other vdev's guid must be unique within the pool.
304 			 */
305 			guid = spa_generate_guid(spa);
306 		}
307 		ASSERT(!spa_guid_exists(spa_guid(spa), guid));
308 	}
309 
310 	vd->vdev_spa = spa;
311 	vd->vdev_id = id;
312 	vd->vdev_guid = guid;
313 	vd->vdev_guid_sum = guid;
314 	vd->vdev_ops = ops;
315 	vd->vdev_state = VDEV_STATE_CLOSED;
316 	vd->vdev_ishole = (ops == &vdev_hole_ops);
317 
318 	mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_DEFAULT, NULL);
319 	mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL);
320 	mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL);
321 	for (int t = 0; t < DTL_TYPES; t++) {
322 		space_map_create(&vd->vdev_dtl[t], 0, -1ULL, 0,
323 		    &vd->vdev_dtl_lock);
324 	}
325 	txg_list_create(&vd->vdev_ms_list,
326 	    offsetof(struct metaslab, ms_txg_node));
327 	txg_list_create(&vd->vdev_dtl_list,
328 	    offsetof(struct vdev, vdev_dtl_node));
329 	vd->vdev_stat.vs_timestamp = gethrtime();
330 	vdev_queue_init(vd);
331 	vdev_cache_init(vd);
332 
333 	return (vd);
334 }
335 
336 /*
337  * Allocate a new vdev.  The 'alloctype' is used to control whether we are
338  * creating a new vdev or loading an existing one - the behavior is slightly
339  * different for each case.
340  */
341 int
342 vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id,
343     int alloctype)
344 {
345 	vdev_ops_t *ops;
346 	char *type;
347 	uint64_t guid = 0, islog, nparity;
348 	vdev_t *vd;
349 
350 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
351 
352 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) != 0)
353 		return (EINVAL);
354 
355 	if ((ops = vdev_getops(type)) == NULL)
356 		return (EINVAL);
357 
358 	/*
359 	 * If this is a load, get the vdev guid from the nvlist.
360 	 * Otherwise, vdev_alloc_common() will generate one for us.
361 	 */
362 	if (alloctype == VDEV_ALLOC_LOAD) {
363 		uint64_t label_id;
364 
365 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ID, &label_id) ||
366 		    label_id != id)
367 			return (EINVAL);
368 
369 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
370 			return (EINVAL);
371 	} else if (alloctype == VDEV_ALLOC_SPARE) {
372 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
373 			return (EINVAL);
374 	} else if (alloctype == VDEV_ALLOC_L2CACHE) {
375 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
376 			return (EINVAL);
377 	} else if (alloctype == VDEV_ALLOC_ROOTPOOL) {
378 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
379 			return (EINVAL);
380 	}
381 
382 	/*
383 	 * The first allocated vdev must be of type 'root'.
384 	 */
385 	if (ops != &vdev_root_ops && spa->spa_root_vdev == NULL)
386 		return (EINVAL);
387 
388 	/*
389 	 * Determine whether we're a log vdev.
390 	 */
391 	islog = 0;
392 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &islog);
393 	if (islog && spa_version(spa) < SPA_VERSION_SLOGS)
394 		return (ENOTSUP);
395 
396 	if (ops == &vdev_hole_ops && spa_version(spa) < SPA_VERSION_HOLES)
397 		return (ENOTSUP);
398 
399 	/*
400 	 * Set the nparity property for RAID-Z vdevs.
401 	 */
402 	nparity = -1ULL;
403 	if (ops == &vdev_raidz_ops) {
404 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY,
405 		    &nparity) == 0) {
406 			if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY)
407 				return (EINVAL);
408 			/*
409 			 * Previous versions could only support 1 or 2 parity
410 			 * device.
411 			 */
412 			if (nparity > 1 &&
413 			    spa_version(spa) < SPA_VERSION_RAIDZ2)
414 				return (ENOTSUP);
415 			if (nparity > 2 &&
416 			    spa_version(spa) < SPA_VERSION_RAIDZ3)
417 				return (ENOTSUP);
418 		} else {
419 			/*
420 			 * We require the parity to be specified for SPAs that
421 			 * support multiple parity levels.
422 			 */
423 			if (spa_version(spa) >= SPA_VERSION_RAIDZ2)
424 				return (EINVAL);
425 			/*
426 			 * Otherwise, we default to 1 parity device for RAID-Z.
427 			 */
428 			nparity = 1;
429 		}
430 	} else {
431 		nparity = 0;
432 	}
433 	ASSERT(nparity != -1ULL);
434 
435 	vd = vdev_alloc_common(spa, id, guid, ops);
436 
437 	vd->vdev_islog = islog;
438 	vd->vdev_nparity = nparity;
439 
440 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &vd->vdev_path) == 0)
441 		vd->vdev_path = spa_strdup(vd->vdev_path);
442 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &vd->vdev_devid) == 0)
443 		vd->vdev_devid = spa_strdup(vd->vdev_devid);
444 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH,
445 	    &vd->vdev_physpath) == 0)
446 		vd->vdev_physpath = spa_strdup(vd->vdev_physpath);
447 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &vd->vdev_fru) == 0)
448 		vd->vdev_fru = spa_strdup(vd->vdev_fru);
449 
450 	/*
451 	 * Set the whole_disk property.  If it's not specified, leave the value
452 	 * as -1.
453 	 */
454 	if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
455 	    &vd->vdev_wholedisk) != 0)
456 		vd->vdev_wholedisk = -1ULL;
457 
458 	/*
459 	 * Look for the 'not present' flag.  This will only be set if the device
460 	 * was not present at the time of import.
461 	 */
462 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT,
463 	    &vd->vdev_not_present);
464 
465 	/*
466 	 * Get the alignment requirement.
467 	 */
468 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT, &vd->vdev_ashift);
469 
470 	/*
471 	 * Retrieve the vdev creation time.
472 	 */
473 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_CREATE_TXG,
474 	    &vd->vdev_crtxg);
475 
476 	/*
477 	 * If we're a top-level vdev, try to load the allocation parameters.
478 	 */
479 	if (parent && !parent->vdev_parent &&
480 	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
481 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
482 		    &vd->vdev_ms_array);
483 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
484 		    &vd->vdev_ms_shift);
485 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE,
486 		    &vd->vdev_asize);
487 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVING,
488 		    &vd->vdev_removing);
489 	}
490 
491 	if (parent && !parent->vdev_parent) {
492 		ASSERT(alloctype == VDEV_ALLOC_LOAD ||
493 		    alloctype == VDEV_ALLOC_ADD ||
494 		    alloctype == VDEV_ALLOC_SPLIT ||
495 		    alloctype == VDEV_ALLOC_ROOTPOOL);
496 		vd->vdev_mg = metaslab_group_create(islog ?
497 		    spa_log_class(spa) : spa_normal_class(spa), vd);
498 	}
499 
500 	/*
501 	 * If we're a leaf vdev, try to load the DTL object and other state.
502 	 */
503 	if (vd->vdev_ops->vdev_op_leaf &&
504 	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE ||
505 	    alloctype == VDEV_ALLOC_ROOTPOOL)) {
506 		if (alloctype == VDEV_ALLOC_LOAD) {
507 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL,
508 			    &vd->vdev_dtl_smo.smo_object);
509 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE,
510 			    &vd->vdev_unspare);
511 		}
512 
513 		if (alloctype == VDEV_ALLOC_ROOTPOOL) {
514 			uint64_t spare = 0;
515 
516 			if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE,
517 			    &spare) == 0 && spare)
518 				spa_spare_add(vd);
519 		}
520 
521 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE,
522 		    &vd->vdev_offline);
523 
524 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_RESILVERING,
525 		    &vd->vdev_resilvering);
526 
527 		/*
528 		 * When importing a pool, we want to ignore the persistent fault
529 		 * state, as the diagnosis made on another system may not be
530 		 * valid in the current context.  Local vdevs will
531 		 * remain in the faulted state.
532 		 */
533 		if (spa_load_state(spa) == SPA_LOAD_OPEN) {
534 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED,
535 			    &vd->vdev_faulted);
536 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED,
537 			    &vd->vdev_degraded);
538 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED,
539 			    &vd->vdev_removed);
540 
541 			if (vd->vdev_faulted || vd->vdev_degraded) {
542 				char *aux;
543 
544 				vd->vdev_label_aux =
545 				    VDEV_AUX_ERR_EXCEEDED;
546 				if (nvlist_lookup_string(nv,
547 				    ZPOOL_CONFIG_AUX_STATE, &aux) == 0 &&
548 				    strcmp(aux, "external") == 0)
549 					vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
550 			}
551 		}
552 	}
553 
554 	/*
555 	 * Add ourselves to the parent's list of children.
556 	 */
557 	vdev_add_child(parent, vd);
558 
559 	*vdp = vd;
560 
561 	return (0);
562 }
563 
564 void
565 vdev_free(vdev_t *vd)
566 {
567 	spa_t *spa = vd->vdev_spa;
568 
569 	/*
570 	 * vdev_free() implies closing the vdev first.  This is simpler than
571 	 * trying to ensure complicated semantics for all callers.
572 	 */
573 	vdev_close(vd);
574 
575 	ASSERT(!list_link_active(&vd->vdev_config_dirty_node));
576 	ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
577 
578 	/*
579 	 * Free all children.
580 	 */
581 	for (int c = 0; c < vd->vdev_children; c++)
582 		vdev_free(vd->vdev_child[c]);
583 
584 	ASSERT(vd->vdev_child == NULL);
585 	ASSERT(vd->vdev_guid_sum == vd->vdev_guid);
586 
587 	/*
588 	 * Discard allocation state.
589 	 */
590 	if (vd->vdev_mg != NULL) {
591 		vdev_metaslab_fini(vd);
592 		metaslab_group_destroy(vd->vdev_mg);
593 	}
594 
595 	ASSERT3U(vd->vdev_stat.vs_space, ==, 0);
596 	ASSERT3U(vd->vdev_stat.vs_dspace, ==, 0);
597 	ASSERT3U(vd->vdev_stat.vs_alloc, ==, 0);
598 
599 	/*
600 	 * Remove this vdev from its parent's child list.
601 	 */
602 	vdev_remove_child(vd->vdev_parent, vd);
603 
604 	ASSERT(vd->vdev_parent == NULL);
605 
606 	/*
607 	 * Clean up vdev structure.
608 	 */
609 	vdev_queue_fini(vd);
610 	vdev_cache_fini(vd);
611 
612 	if (vd->vdev_path)
613 		spa_strfree(vd->vdev_path);
614 	if (vd->vdev_devid)
615 		spa_strfree(vd->vdev_devid);
616 	if (vd->vdev_physpath)
617 		spa_strfree(vd->vdev_physpath);
618 	if (vd->vdev_fru)
619 		spa_strfree(vd->vdev_fru);
620 
621 	if (vd->vdev_isspare)
622 		spa_spare_remove(vd);
623 	if (vd->vdev_isl2cache)
624 		spa_l2cache_remove(vd);
625 
626 	txg_list_destroy(&vd->vdev_ms_list);
627 	txg_list_destroy(&vd->vdev_dtl_list);
628 
629 	mutex_enter(&vd->vdev_dtl_lock);
630 	for (int t = 0; t < DTL_TYPES; t++) {
631 		space_map_unload(&vd->vdev_dtl[t]);
632 		space_map_destroy(&vd->vdev_dtl[t]);
633 	}
634 	mutex_exit(&vd->vdev_dtl_lock);
635 
636 	mutex_destroy(&vd->vdev_dtl_lock);
637 	mutex_destroy(&vd->vdev_stat_lock);
638 	mutex_destroy(&vd->vdev_probe_lock);
639 
640 	if (vd == spa->spa_root_vdev)
641 		spa->spa_root_vdev = NULL;
642 
643 	kmem_free(vd, sizeof (vdev_t));
644 }
645 
646 /*
647  * Transfer top-level vdev state from svd to tvd.
648  */
649 static void
650 vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
651 {
652 	spa_t *spa = svd->vdev_spa;
653 	metaslab_t *msp;
654 	vdev_t *vd;
655 	int t;
656 
657 	ASSERT(tvd == tvd->vdev_top);
658 
659 	tvd->vdev_ms_array = svd->vdev_ms_array;
660 	tvd->vdev_ms_shift = svd->vdev_ms_shift;
661 	tvd->vdev_ms_count = svd->vdev_ms_count;
662 
663 	svd->vdev_ms_array = 0;
664 	svd->vdev_ms_shift = 0;
665 	svd->vdev_ms_count = 0;
666 
667 	tvd->vdev_mg = svd->vdev_mg;
668 	tvd->vdev_ms = svd->vdev_ms;
669 
670 	svd->vdev_mg = NULL;
671 	svd->vdev_ms = NULL;
672 
673 	if (tvd->vdev_mg != NULL)
674 		tvd->vdev_mg->mg_vd = tvd;
675 
676 	tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
677 	tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
678 	tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
679 
680 	svd->vdev_stat.vs_alloc = 0;
681 	svd->vdev_stat.vs_space = 0;
682 	svd->vdev_stat.vs_dspace = 0;
683 
684 	for (t = 0; t < TXG_SIZE; t++) {
685 		while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
686 			(void) txg_list_add(&tvd->vdev_ms_list, msp, t);
687 		while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
688 			(void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
689 		if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
690 			(void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
691 	}
692 
693 	if (list_link_active(&svd->vdev_config_dirty_node)) {
694 		vdev_config_clean(svd);
695 		vdev_config_dirty(tvd);
696 	}
697 
698 	if (list_link_active(&svd->vdev_state_dirty_node)) {
699 		vdev_state_clean(svd);
700 		vdev_state_dirty(tvd);
701 	}
702 
703 	tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
704 	svd->vdev_deflate_ratio = 0;
705 
706 	tvd->vdev_islog = svd->vdev_islog;
707 	svd->vdev_islog = 0;
708 }
709 
710 static void
711 vdev_top_update(vdev_t *tvd, vdev_t *vd)
712 {
713 	if (vd == NULL)
714 		return;
715 
716 	vd->vdev_top = tvd;
717 
718 	for (int c = 0; c < vd->vdev_children; c++)
719 		vdev_top_update(tvd, vd->vdev_child[c]);
720 }
721 
722 /*
723  * Add a mirror/replacing vdev above an existing vdev.
724  */
725 vdev_t *
726 vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
727 {
728 	spa_t *spa = cvd->vdev_spa;
729 	vdev_t *pvd = cvd->vdev_parent;
730 	vdev_t *mvd;
731 
732 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
733 
734 	mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
735 
736 	mvd->vdev_asize = cvd->vdev_asize;
737 	mvd->vdev_min_asize = cvd->vdev_min_asize;
738 	mvd->vdev_max_asize = cvd->vdev_max_asize;
739 	mvd->vdev_ashift = cvd->vdev_ashift;
740 	mvd->vdev_state = cvd->vdev_state;
741 	mvd->vdev_crtxg = cvd->vdev_crtxg;
742 
743 	vdev_remove_child(pvd, cvd);
744 	vdev_add_child(pvd, mvd);
745 	cvd->vdev_id = mvd->vdev_children;
746 	vdev_add_child(mvd, cvd);
747 	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
748 
749 	if (mvd == mvd->vdev_top)
750 		vdev_top_transfer(cvd, mvd);
751 
752 	return (mvd);
753 }
754 
755 /*
756  * Remove a 1-way mirror/replacing vdev from the tree.
757  */
758 void
759 vdev_remove_parent(vdev_t *cvd)
760 {
761 	vdev_t *mvd = cvd->vdev_parent;
762 	vdev_t *pvd = mvd->vdev_parent;
763 
764 	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
765 
766 	ASSERT(mvd->vdev_children == 1);
767 	ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
768 	    mvd->vdev_ops == &vdev_replacing_ops ||
769 	    mvd->vdev_ops == &vdev_spare_ops);
770 	cvd->vdev_ashift = mvd->vdev_ashift;
771 
772 	vdev_remove_child(mvd, cvd);
773 	vdev_remove_child(pvd, mvd);
774 
775 	/*
776 	 * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
777 	 * Otherwise, we could have detached an offline device, and when we
778 	 * go to import the pool we'll think we have two top-level vdevs,
779 	 * instead of a different version of the same top-level vdev.
780 	 */
781 	if (mvd->vdev_top == mvd) {
782 		uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
783 		cvd->vdev_orig_guid = cvd->vdev_guid;
784 		cvd->vdev_guid += guid_delta;
785 		cvd->vdev_guid_sum += guid_delta;
786 	}
787 	cvd->vdev_id = mvd->vdev_id;
788 	vdev_add_child(pvd, cvd);
789 	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
790 
791 	if (cvd == cvd->vdev_top)
792 		vdev_top_transfer(mvd, cvd);
793 
794 	ASSERT(mvd->vdev_children == 0);
795 	vdev_free(mvd);
796 }
797 
798 int
799 vdev_metaslab_init(vdev_t *vd, uint64_t txg)
800 {
801 	spa_t *spa = vd->vdev_spa;
802 	objset_t *mos = spa->spa_meta_objset;
803 	uint64_t m;
804 	uint64_t oldc = vd->vdev_ms_count;
805 	uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
806 	metaslab_t **mspp;
807 	int error;
808 
809 	ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER));
810 
811 	/*
812 	 * This vdev is not being allocated from yet or is a hole.
813 	 */
814 	if (vd->vdev_ms_shift == 0)
815 		return (0);
816 
817 	ASSERT(!vd->vdev_ishole);
818 
819 	/*
820 	 * Compute the raidz-deflation ratio.  Note, we hard-code
821 	 * in 128k (1 << 17) because it is the current "typical" blocksize.
822 	 * Even if SPA_MAXBLOCKSIZE changes, this algorithm must never change,
823 	 * or we will inconsistently account for existing bp's.
824 	 */
825 	vd->vdev_deflate_ratio = (1 << 17) /
826 	    (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
827 
828 	ASSERT(oldc <= newc);
829 
830 	mspp = kmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
831 
832 	if (oldc != 0) {
833 		bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp));
834 		kmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
835 	}
836 
837 	vd->vdev_ms = mspp;
838 	vd->vdev_ms_count = newc;
839 
840 	for (m = oldc; m < newc; m++) {
841 		space_map_obj_t smo = { 0, 0, 0 };
842 		if (txg == 0) {
843 			uint64_t object = 0;
844 			error = dmu_read(mos, vd->vdev_ms_array,
845 			    m * sizeof (uint64_t), sizeof (uint64_t), &object,
846 			    DMU_READ_PREFETCH);
847 			if (error)
848 				return (error);
849 			if (object != 0) {
850 				dmu_buf_t *db;
851 				error = dmu_bonus_hold(mos, object, FTAG, &db);
852 				if (error)
853 					return (error);
854 				ASSERT3U(db->db_size, >=, sizeof (smo));
855 				bcopy(db->db_data, &smo, sizeof (smo));
856 				ASSERT3U(smo.smo_object, ==, object);
857 				dmu_buf_rele(db, FTAG);
858 			}
859 		}
860 		vd->vdev_ms[m] = metaslab_init(vd->vdev_mg, &smo,
861 		    m << vd->vdev_ms_shift, 1ULL << vd->vdev_ms_shift, txg);
862 	}
863 
864 	if (txg == 0)
865 		spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER);
866 
867 	/*
868 	 * If the vdev is being removed we don't activate
869 	 * the metaslabs since we want to ensure that no new
870 	 * allocations are performed on this device.
871 	 */
872 	if (oldc == 0 && !vd->vdev_removing)
873 		metaslab_group_activate(vd->vdev_mg);
874 
875 	if (txg == 0)
876 		spa_config_exit(spa, SCL_ALLOC, FTAG);
877 
878 	return (0);
879 }
880 
881 void
882 vdev_metaslab_fini(vdev_t *vd)
883 {
884 	uint64_t m;
885 	uint64_t count = vd->vdev_ms_count;
886 
887 	if (vd->vdev_ms != NULL) {
888 		metaslab_group_passivate(vd->vdev_mg);
889 		for (m = 0; m < count; m++)
890 			if (vd->vdev_ms[m] != NULL)
891 				metaslab_fini(vd->vdev_ms[m]);
892 		kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
893 		vd->vdev_ms = NULL;
894 	}
895 }
896 
897 typedef struct vdev_probe_stats {
898 	boolean_t	vps_readable;
899 	boolean_t	vps_writeable;
900 	int		vps_flags;
901 } vdev_probe_stats_t;
902 
903 static void
904 vdev_probe_done(zio_t *zio)
905 {
906 	spa_t *spa = zio->io_spa;
907 	vdev_t *vd = zio->io_vd;
908 	vdev_probe_stats_t *vps = zio->io_private;
909 
910 	ASSERT(vd->vdev_probe_zio != NULL);
911 
912 	if (zio->io_type == ZIO_TYPE_READ) {
913 		if (zio->io_error == 0)
914 			vps->vps_readable = 1;
915 		if (zio->io_error == 0 && spa_writeable(spa)) {
916 			zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
917 			    zio->io_offset, zio->io_size, zio->io_data,
918 			    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
919 			    ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
920 		} else {
921 			zio_buf_free(zio->io_data, zio->io_size);
922 		}
923 	} else if (zio->io_type == ZIO_TYPE_WRITE) {
924 		if (zio->io_error == 0)
925 			vps->vps_writeable = 1;
926 		zio_buf_free(zio->io_data, zio->io_size);
927 	} else if (zio->io_type == ZIO_TYPE_NULL) {
928 		zio_t *pio;
929 
930 		vd->vdev_cant_read |= !vps->vps_readable;
931 		vd->vdev_cant_write |= !vps->vps_writeable;
932 
933 		if (vdev_readable(vd) &&
934 		    (vdev_writeable(vd) || !spa_writeable(spa))) {
935 			zio->io_error = 0;
936 		} else {
937 			ASSERT(zio->io_error != 0);
938 			zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
939 			    spa, vd, NULL, 0, 0);
940 			zio->io_error = ENXIO;
941 		}
942 
943 		mutex_enter(&vd->vdev_probe_lock);
944 		ASSERT(vd->vdev_probe_zio == zio);
945 		vd->vdev_probe_zio = NULL;
946 		mutex_exit(&vd->vdev_probe_lock);
947 
948 		while ((pio = zio_walk_parents(zio)) != NULL)
949 			if (!vdev_accessible(vd, pio))
950 				pio->io_error = ENXIO;
951 
952 		kmem_free(vps, sizeof (*vps));
953 	}
954 }
955 
956 /*
957  * Determine whether this device is accessible by reading and writing
958  * to several known locations: the pad regions of each vdev label
959  * but the first (which we leave alone in case it contains a VTOC).
960  */
961 zio_t *
962 vdev_probe(vdev_t *vd, zio_t *zio)
963 {
964 	spa_t *spa = vd->vdev_spa;
965 	vdev_probe_stats_t *vps = NULL;
966 	zio_t *pio;
967 
968 	ASSERT(vd->vdev_ops->vdev_op_leaf);
969 
970 	/*
971 	 * Don't probe the probe.
972 	 */
973 	if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
974 		return (NULL);
975 
976 	/*
977 	 * To prevent 'probe storms' when a device fails, we create
978 	 * just one probe i/o at a time.  All zios that want to probe
979 	 * this vdev will become parents of the probe io.
980 	 */
981 	mutex_enter(&vd->vdev_probe_lock);
982 
983 	if ((pio = vd->vdev_probe_zio) == NULL) {
984 		vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
985 
986 		vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
987 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
988 		    ZIO_FLAG_TRYHARD;
989 
990 		if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
991 			/*
992 			 * vdev_cant_read and vdev_cant_write can only
993 			 * transition from TRUE to FALSE when we have the
994 			 * SCL_ZIO lock as writer; otherwise they can only
995 			 * transition from FALSE to TRUE.  This ensures that
996 			 * any zio looking at these values can assume that
997 			 * failures persist for the life of the I/O.  That's
998 			 * important because when a device has intermittent
999 			 * connectivity problems, we want to ensure that
1000 			 * they're ascribed to the device (ENXIO) and not
1001 			 * the zio (EIO).
1002 			 *
1003 			 * Since we hold SCL_ZIO as writer here, clear both
1004 			 * values so the probe can reevaluate from first
1005 			 * principles.
1006 			 */
1007 			vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
1008 			vd->vdev_cant_read = B_FALSE;
1009 			vd->vdev_cant_write = B_FALSE;
1010 		}
1011 
1012 		vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
1013 		    vdev_probe_done, vps,
1014 		    vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
1015 
1016 		/*
1017 		 * We can't change the vdev state in this context, so we
1018 		 * kick off an async task to do it on our behalf.
1019 		 */
1020 		if (zio != NULL) {
1021 			vd->vdev_probe_wanted = B_TRUE;
1022 			spa_async_request(spa, SPA_ASYNC_PROBE);
1023 		}
1024 	}
1025 
1026 	if (zio != NULL)
1027 		zio_add_child(zio, pio);
1028 
1029 	mutex_exit(&vd->vdev_probe_lock);
1030 
1031 	if (vps == NULL) {
1032 		ASSERT(zio != NULL);
1033 		return (NULL);
1034 	}
1035 
1036 	for (int l = 1; l < VDEV_LABELS; l++) {
1037 		zio_nowait(zio_read_phys(pio, vd,
1038 		    vdev_label_offset(vd->vdev_psize, l,
1039 		    offsetof(vdev_label_t, vl_pad2)),
1040 		    VDEV_PAD_SIZE, zio_buf_alloc(VDEV_PAD_SIZE),
1041 		    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1042 		    ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1043 	}
1044 
1045 	if (zio == NULL)
1046 		return (pio);
1047 
1048 	zio_nowait(pio);
1049 	return (NULL);
1050 }
1051 
1052 static void
1053 vdev_open_child(void *arg)
1054 {
1055 	vdev_t *vd = arg;
1056 
1057 	vd->vdev_open_thread = curthread;
1058 	vd->vdev_open_error = vdev_open(vd);
1059 	vd->vdev_open_thread = NULL;
1060 }
1061 
1062 boolean_t
1063 vdev_uses_zvols(vdev_t *vd)
1064 {
1065 	if (vd->vdev_path && strncmp(vd->vdev_path, ZVOL_DIR,
1066 	    strlen(ZVOL_DIR)) == 0)
1067 		return (B_TRUE);
1068 	for (int c = 0; c < vd->vdev_children; c++)
1069 		if (vdev_uses_zvols(vd->vdev_child[c]))
1070 			return (B_TRUE);
1071 	return (B_FALSE);
1072 }
1073 
1074 void
1075 vdev_open_children(vdev_t *vd)
1076 {
1077 	taskq_t *tq;
1078 	int children = vd->vdev_children;
1079 
1080 	/*
1081 	 * in order to handle pools on top of zvols, do the opens
1082 	 * in a single thread so that the same thread holds the
1083 	 * spa_namespace_lock
1084 	 */
1085 	if (vdev_uses_zvols(vd)) {
1086 		for (int c = 0; c < children; c++)
1087 			vd->vdev_child[c]->vdev_open_error =
1088 			    vdev_open(vd->vdev_child[c]);
1089 		return;
1090 	}
1091 	tq = taskq_create("vdev_open", children, minclsyspri,
1092 	    children, children, TASKQ_PREPOPULATE);
1093 
1094 	for (int c = 0; c < children; c++)
1095 		VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c],
1096 		    TQ_SLEEP) != NULL);
1097 
1098 	taskq_destroy(tq);
1099 }
1100 
1101 /*
1102  * Prepare a virtual device for access.
1103  */
1104 int
1105 vdev_open(vdev_t *vd)
1106 {
1107 	spa_t *spa = vd->vdev_spa;
1108 	int error;
1109 	uint64_t osize = 0;
1110 	uint64_t max_osize = 0;
1111 	uint64_t asize, max_asize, psize;
1112 	uint64_t ashift = 0;
1113 
1114 	ASSERT(vd->vdev_open_thread == curthread ||
1115 	    spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1116 	ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1117 	    vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1118 	    vd->vdev_state == VDEV_STATE_OFFLINE);
1119 
1120 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1121 	vd->vdev_cant_read = B_FALSE;
1122 	vd->vdev_cant_write = B_FALSE;
1123 	vd->vdev_min_asize = vdev_get_min_asize(vd);
1124 
1125 	/*
1126 	 * If this vdev is not removed, check its fault status.  If it's
1127 	 * faulted, bail out of the open.
1128 	 */
1129 	if (!vd->vdev_removed && vd->vdev_faulted) {
1130 		ASSERT(vd->vdev_children == 0);
1131 		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1132 		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1133 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1134 		    vd->vdev_label_aux);
1135 		return (ENXIO);
1136 	} else if (vd->vdev_offline) {
1137 		ASSERT(vd->vdev_children == 0);
1138 		vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1139 		return (ENXIO);
1140 	}
1141 
1142 	error = vd->vdev_ops->vdev_op_open(vd, &osize, &max_osize, &ashift);
1143 
1144 	/*
1145 	 * Reset the vdev_reopening flag so that we actually close
1146 	 * the vdev on error.
1147 	 */
1148 	vd->vdev_reopening = B_FALSE;
1149 	if (zio_injection_enabled && error == 0)
1150 		error = zio_handle_device_injection(vd, NULL, ENXIO);
1151 
1152 	if (error) {
1153 		if (vd->vdev_removed &&
1154 		    vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1155 			vd->vdev_removed = B_FALSE;
1156 
1157 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1158 		    vd->vdev_stat.vs_aux);
1159 		return (error);
1160 	}
1161 
1162 	vd->vdev_removed = B_FALSE;
1163 
1164 	/*
1165 	 * Recheck the faulted flag now that we have confirmed that
1166 	 * the vdev is accessible.  If we're faulted, bail.
1167 	 */
1168 	if (vd->vdev_faulted) {
1169 		ASSERT(vd->vdev_children == 0);
1170 		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1171 		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1172 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1173 		    vd->vdev_label_aux);
1174 		return (ENXIO);
1175 	}
1176 
1177 	if (vd->vdev_degraded) {
1178 		ASSERT(vd->vdev_children == 0);
1179 		vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1180 		    VDEV_AUX_ERR_EXCEEDED);
1181 	} else {
1182 		vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1183 	}
1184 
1185 	/*
1186 	 * For hole or missing vdevs we just return success.
1187 	 */
1188 	if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1189 		return (0);
1190 
1191 	for (int c = 0; c < vd->vdev_children; c++) {
1192 		if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1193 			vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1194 			    VDEV_AUX_NONE);
1195 			break;
1196 		}
1197 	}
1198 
1199 	osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1200 	max_osize = P2ALIGN(max_osize, (uint64_t)sizeof (vdev_label_t));
1201 
1202 	if (vd->vdev_children == 0) {
1203 		if (osize < SPA_MINDEVSIZE) {
1204 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1205 			    VDEV_AUX_TOO_SMALL);
1206 			return (EOVERFLOW);
1207 		}
1208 		psize = osize;
1209 		asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1210 		max_asize = max_osize - (VDEV_LABEL_START_SIZE +
1211 		    VDEV_LABEL_END_SIZE);
1212 	} else {
1213 		if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1214 		    (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1215 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1216 			    VDEV_AUX_TOO_SMALL);
1217 			return (EOVERFLOW);
1218 		}
1219 		psize = 0;
1220 		asize = osize;
1221 		max_asize = max_osize;
1222 	}
1223 
1224 	vd->vdev_psize = psize;
1225 
1226 	/*
1227 	 * Make sure the allocatable size hasn't shrunk.
1228 	 */
1229 	if (asize < vd->vdev_min_asize) {
1230 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1231 		    VDEV_AUX_BAD_LABEL);
1232 		return (EINVAL);
1233 	}
1234 
1235 	if (vd->vdev_asize == 0) {
1236 		/*
1237 		 * This is the first-ever open, so use the computed values.
1238 		 * For testing purposes, a higher ashift can be requested.
1239 		 */
1240 		vd->vdev_asize = asize;
1241 		vd->vdev_max_asize = max_asize;
1242 		vd->vdev_ashift = MAX(ashift, vd->vdev_ashift);
1243 	} else {
1244 		/*
1245 		 * Make sure the alignment requirement hasn't increased.
1246 		 */
1247 		if (ashift > vd->vdev_top->vdev_ashift) {
1248 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1249 			    VDEV_AUX_BAD_LABEL);
1250 			return (EINVAL);
1251 		}
1252 		vd->vdev_max_asize = max_asize;
1253 	}
1254 
1255 	/*
1256 	 * If all children are healthy and the asize has increased,
1257 	 * then we've experienced dynamic LUN growth.  If automatic
1258 	 * expansion is enabled then use the additional space.
1259 	 */
1260 	if (vd->vdev_state == VDEV_STATE_HEALTHY && asize > vd->vdev_asize &&
1261 	    (vd->vdev_expanding || spa->spa_autoexpand))
1262 		vd->vdev_asize = asize;
1263 
1264 	vdev_set_min_asize(vd);
1265 
1266 	/*
1267 	 * Ensure we can issue some IO before declaring the
1268 	 * vdev open for business.
1269 	 */
1270 	if (vd->vdev_ops->vdev_op_leaf &&
1271 	    (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1272 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1273 		    VDEV_AUX_ERR_EXCEEDED);
1274 		return (error);
1275 	}
1276 
1277 	/*
1278 	 * If a leaf vdev has a DTL, and seems healthy, then kick off a
1279 	 * resilver.  But don't do this if we are doing a reopen for a scrub,
1280 	 * since this would just restart the scrub we are already doing.
1281 	 */
1282 	if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1283 	    vdev_resilver_needed(vd, NULL, NULL))
1284 		spa_async_request(spa, SPA_ASYNC_RESILVER);
1285 
1286 	return (0);
1287 }
1288 
1289 /*
1290  * Called once the vdevs are all opened, this routine validates the label
1291  * contents.  This needs to be done before vdev_load() so that we don't
1292  * inadvertently do repair I/Os to the wrong device.
1293  *
1294  * This function will only return failure if one of the vdevs indicates that it
1295  * has since been destroyed or exported.  This is only possible if
1296  * /etc/zfs/zpool.cache was readonly at the time.  Otherwise, the vdev state
1297  * will be updated but the function will return 0.
1298  */
1299 int
1300 vdev_validate(vdev_t *vd)
1301 {
1302 	spa_t *spa = vd->vdev_spa;
1303 	nvlist_t *label;
1304 	uint64_t guid = 0, top_guid;
1305 	uint64_t state;
1306 
1307 	for (int c = 0; c < vd->vdev_children; c++)
1308 		if (vdev_validate(vd->vdev_child[c]) != 0)
1309 			return (EBADF);
1310 
1311 	/*
1312 	 * If the device has already failed, or was marked offline, don't do
1313 	 * any further validation.  Otherwise, label I/O will fail and we will
1314 	 * overwrite the previous state.
1315 	 */
1316 	if (vd->vdev_ops->vdev_op_leaf && vdev_readable(vd)) {
1317 		uint64_t aux_guid = 0;
1318 		nvlist_t *nvl;
1319 
1320 		if ((label = vdev_label_read_config(vd)) == NULL) {
1321 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1322 			    VDEV_AUX_BAD_LABEL);
1323 			return (0);
1324 		}
1325 
1326 		/*
1327 		 * Determine if this vdev has been split off into another
1328 		 * pool.  If so, then refuse to open it.
1329 		 */
1330 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID,
1331 		    &aux_guid) == 0 && aux_guid == spa_guid(spa)) {
1332 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1333 			    VDEV_AUX_SPLIT_POOL);
1334 			nvlist_free(label);
1335 			return (0);
1336 		}
1337 
1338 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_GUID,
1339 		    &guid) != 0 || guid != spa_guid(spa)) {
1340 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1341 			    VDEV_AUX_CORRUPT_DATA);
1342 			nvlist_free(label);
1343 			return (0);
1344 		}
1345 
1346 		if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl)
1347 		    != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID,
1348 		    &aux_guid) != 0)
1349 			aux_guid = 0;
1350 
1351 		/*
1352 		 * If this vdev just became a top-level vdev because its
1353 		 * sibling was detached, it will have adopted the parent's
1354 		 * vdev guid -- but the label may or may not be on disk yet.
1355 		 * Fortunately, either version of the label will have the
1356 		 * same top guid, so if we're a top-level vdev, we can
1357 		 * safely compare to that instead.
1358 		 *
1359 		 * If we split this vdev off instead, then we also check the
1360 		 * original pool's guid.  We don't want to consider the vdev
1361 		 * corrupt if it is partway through a split operation.
1362 		 */
1363 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID,
1364 		    &guid) != 0 ||
1365 		    nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID,
1366 		    &top_guid) != 0 ||
1367 		    ((vd->vdev_guid != guid && vd->vdev_guid != aux_guid) &&
1368 		    (vd->vdev_guid != top_guid || vd != vd->vdev_top))) {
1369 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1370 			    VDEV_AUX_CORRUPT_DATA);
1371 			nvlist_free(label);
1372 			return (0);
1373 		}
1374 
1375 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1376 		    &state) != 0) {
1377 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1378 			    VDEV_AUX_CORRUPT_DATA);
1379 			nvlist_free(label);
1380 			return (0);
1381 		}
1382 
1383 		nvlist_free(label);
1384 
1385 		/*
1386 		 * If this is a verbatim import, no need to check the
1387 		 * state of the pool.
1388 		 */
1389 		if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) &&
1390 		    spa_load_state(spa) == SPA_LOAD_OPEN &&
1391 		    state != POOL_STATE_ACTIVE)
1392 			return (EBADF);
1393 
1394 		/*
1395 		 * If we were able to open and validate a vdev that was
1396 		 * previously marked permanently unavailable, clear that state
1397 		 * now.
1398 		 */
1399 		if (vd->vdev_not_present)
1400 			vd->vdev_not_present = 0;
1401 	}
1402 
1403 	return (0);
1404 }
1405 
1406 /*
1407  * Close a virtual device.
1408  */
1409 void
1410 vdev_close(vdev_t *vd)
1411 {
1412 	spa_t *spa = vd->vdev_spa;
1413 	vdev_t *pvd = vd->vdev_parent;
1414 
1415 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1416 
1417 	/*
1418 	 * If our parent is reopening, then we are as well, unless we are
1419 	 * going offline.
1420 	 */
1421 	if (pvd != NULL && pvd->vdev_reopening)
1422 		vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline);
1423 
1424 	vd->vdev_ops->vdev_op_close(vd);
1425 
1426 	vdev_cache_purge(vd);
1427 
1428 	/*
1429 	 * We record the previous state before we close it, so that if we are
1430 	 * doing a reopen(), we don't generate FMA ereports if we notice that
1431 	 * it's still faulted.
1432 	 */
1433 	vd->vdev_prevstate = vd->vdev_state;
1434 
1435 	if (vd->vdev_offline)
1436 		vd->vdev_state = VDEV_STATE_OFFLINE;
1437 	else
1438 		vd->vdev_state = VDEV_STATE_CLOSED;
1439 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1440 }
1441 
1442 void
1443 vdev_hold(vdev_t *vd)
1444 {
1445 	spa_t *spa = vd->vdev_spa;
1446 
1447 	ASSERT(spa_is_root(spa));
1448 	if (spa->spa_state == POOL_STATE_UNINITIALIZED)
1449 		return;
1450 
1451 	for (int c = 0; c < vd->vdev_children; c++)
1452 		vdev_hold(vd->vdev_child[c]);
1453 
1454 	if (vd->vdev_ops->vdev_op_leaf)
1455 		vd->vdev_ops->vdev_op_hold(vd);
1456 }
1457 
1458 void
1459 vdev_rele(vdev_t *vd)
1460 {
1461 	spa_t *spa = vd->vdev_spa;
1462 
1463 	ASSERT(spa_is_root(spa));
1464 	for (int c = 0; c < vd->vdev_children; c++)
1465 		vdev_rele(vd->vdev_child[c]);
1466 
1467 	if (vd->vdev_ops->vdev_op_leaf)
1468 		vd->vdev_ops->vdev_op_rele(vd);
1469 }
1470 
1471 /*
1472  * Reopen all interior vdevs and any unopened leaves.  We don't actually
1473  * reopen leaf vdevs which had previously been opened as they might deadlock
1474  * on the spa_config_lock.  Instead we only obtain the leaf's physical size.
1475  * If the leaf has never been opened then open it, as usual.
1476  */
1477 void
1478 vdev_reopen(vdev_t *vd)
1479 {
1480 	spa_t *spa = vd->vdev_spa;
1481 
1482 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1483 
1484 	/* set the reopening flag unless we're taking the vdev offline */
1485 	vd->vdev_reopening = !vd->vdev_offline;
1486 	vdev_close(vd);
1487 	(void) vdev_open(vd);
1488 
1489 	/*
1490 	 * Call vdev_validate() here to make sure we have the same device.
1491 	 * Otherwise, a device with an invalid label could be successfully
1492 	 * opened in response to vdev_reopen().
1493 	 */
1494 	if (vd->vdev_aux) {
1495 		(void) vdev_validate_aux(vd);
1496 		if (vdev_readable(vd) && vdev_writeable(vd) &&
1497 		    vd->vdev_aux == &spa->spa_l2cache &&
1498 		    !l2arc_vdev_present(vd))
1499 			l2arc_add_vdev(spa, vd);
1500 	} else {
1501 		(void) vdev_validate(vd);
1502 	}
1503 
1504 	/*
1505 	 * Reassess parent vdev's health.
1506 	 */
1507 	vdev_propagate_state(vd);
1508 }
1509 
1510 int
1511 vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1512 {
1513 	int error;
1514 
1515 	/*
1516 	 * Normally, partial opens (e.g. of a mirror) are allowed.
1517 	 * For a create, however, we want to fail the request if
1518 	 * there are any components we can't open.
1519 	 */
1520 	error = vdev_open(vd);
1521 
1522 	if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
1523 		vdev_close(vd);
1524 		return (error ? error : ENXIO);
1525 	}
1526 
1527 	/*
1528 	 * Recursively initialize all labels.
1529 	 */
1530 	if ((error = vdev_label_init(vd, txg, isreplacing ?
1531 	    VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
1532 		vdev_close(vd);
1533 		return (error);
1534 	}
1535 
1536 	return (0);
1537 }
1538 
1539 void
1540 vdev_metaslab_set_size(vdev_t *vd)
1541 {
1542 	/*
1543 	 * Aim for roughly 200 metaslabs per vdev.
1544 	 */
1545 	vd->vdev_ms_shift = highbit(vd->vdev_asize / 200);
1546 	vd->vdev_ms_shift = MAX(vd->vdev_ms_shift, SPA_MAXBLOCKSHIFT);
1547 }
1548 
1549 void
1550 vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
1551 {
1552 	ASSERT(vd == vd->vdev_top);
1553 	ASSERT(!vd->vdev_ishole);
1554 	ASSERT(ISP2(flags));
1555 	ASSERT(spa_writeable(vd->vdev_spa));
1556 
1557 	if (flags & VDD_METASLAB)
1558 		(void) txg_list_add(&vd->vdev_ms_list, arg, txg);
1559 
1560 	if (flags & VDD_DTL)
1561 		(void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
1562 
1563 	(void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
1564 }
1565 
1566 /*
1567  * DTLs.
1568  *
1569  * A vdev's DTL (dirty time log) is the set of transaction groups for which
1570  * the vdev has less than perfect replication.  There are four kinds of DTL:
1571  *
1572  * DTL_MISSING: txgs for which the vdev has no valid copies of the data
1573  *
1574  * DTL_PARTIAL: txgs for which data is available, but not fully replicated
1575  *
1576  * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
1577  *	scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
1578  *	txgs that was scrubbed.
1579  *
1580  * DTL_OUTAGE: txgs which cannot currently be read, whether due to
1581  *	persistent errors or just some device being offline.
1582  *	Unlike the other three, the DTL_OUTAGE map is not generally
1583  *	maintained; it's only computed when needed, typically to
1584  *	determine whether a device can be detached.
1585  *
1586  * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
1587  * either has the data or it doesn't.
1588  *
1589  * For interior vdevs such as mirror and RAID-Z the picture is more complex.
1590  * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
1591  * if any child is less than fully replicated, then so is its parent.
1592  * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
1593  * comprising only those txgs which appear in 'maxfaults' or more children;
1594  * those are the txgs we don't have enough replication to read.  For example,
1595  * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
1596  * thus, its DTL_MISSING consists of the set of txgs that appear in more than
1597  * two child DTL_MISSING maps.
1598  *
1599  * It should be clear from the above that to compute the DTLs and outage maps
1600  * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
1601  * Therefore, that is all we keep on disk.  When loading the pool, or after
1602  * a configuration change, we generate all other DTLs from first principles.
1603  */
1604 void
1605 vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1606 {
1607 	space_map_t *sm = &vd->vdev_dtl[t];
1608 
1609 	ASSERT(t < DTL_TYPES);
1610 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1611 	ASSERT(spa_writeable(vd->vdev_spa));
1612 
1613 	mutex_enter(sm->sm_lock);
1614 	if (!space_map_contains(sm, txg, size))
1615 		space_map_add(sm, txg, size);
1616 	mutex_exit(sm->sm_lock);
1617 }
1618 
1619 boolean_t
1620 vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1621 {
1622 	space_map_t *sm = &vd->vdev_dtl[t];
1623 	boolean_t dirty = B_FALSE;
1624 
1625 	ASSERT(t < DTL_TYPES);
1626 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1627 
1628 	mutex_enter(sm->sm_lock);
1629 	if (sm->sm_space != 0)
1630 		dirty = space_map_contains(sm, txg, size);
1631 	mutex_exit(sm->sm_lock);
1632 
1633 	return (dirty);
1634 }
1635 
1636 boolean_t
1637 vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
1638 {
1639 	space_map_t *sm = &vd->vdev_dtl[t];
1640 	boolean_t empty;
1641 
1642 	mutex_enter(sm->sm_lock);
1643 	empty = (sm->sm_space == 0);
1644 	mutex_exit(sm->sm_lock);
1645 
1646 	return (empty);
1647 }
1648 
1649 /*
1650  * Reassess DTLs after a config change or scrub completion.
1651  */
1652 void
1653 vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
1654 {
1655 	spa_t *spa = vd->vdev_spa;
1656 	avl_tree_t reftree;
1657 	int minref;
1658 
1659 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
1660 
1661 	for (int c = 0; c < vd->vdev_children; c++)
1662 		vdev_dtl_reassess(vd->vdev_child[c], txg,
1663 		    scrub_txg, scrub_done);
1664 
1665 	if (vd == spa->spa_root_vdev || vd->vdev_ishole || vd->vdev_aux)
1666 		return;
1667 
1668 	if (vd->vdev_ops->vdev_op_leaf) {
1669 		dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1670 
1671 		mutex_enter(&vd->vdev_dtl_lock);
1672 		if (scrub_txg != 0 &&
1673 		    (spa->spa_scrub_started ||
1674 		    (scn && scn->scn_phys.scn_errors == 0))) {
1675 			/*
1676 			 * We completed a scrub up to scrub_txg.  If we
1677 			 * did it without rebooting, then the scrub dtl
1678 			 * will be valid, so excise the old region and
1679 			 * fold in the scrub dtl.  Otherwise, leave the
1680 			 * dtl as-is if there was an error.
1681 			 *
1682 			 * There's little trick here: to excise the beginning
1683 			 * of the DTL_MISSING map, we put it into a reference
1684 			 * tree and then add a segment with refcnt -1 that
1685 			 * covers the range [0, scrub_txg).  This means
1686 			 * that each txg in that range has refcnt -1 or 0.
1687 			 * We then add DTL_SCRUB with a refcnt of 2, so that
1688 			 * entries in the range [0, scrub_txg) will have a
1689 			 * positive refcnt -- either 1 or 2.  We then convert
1690 			 * the reference tree into the new DTL_MISSING map.
1691 			 */
1692 			space_map_ref_create(&reftree);
1693 			space_map_ref_add_map(&reftree,
1694 			    &vd->vdev_dtl[DTL_MISSING], 1);
1695 			space_map_ref_add_seg(&reftree, 0, scrub_txg, -1);
1696 			space_map_ref_add_map(&reftree,
1697 			    &vd->vdev_dtl[DTL_SCRUB], 2);
1698 			space_map_ref_generate_map(&reftree,
1699 			    &vd->vdev_dtl[DTL_MISSING], 1);
1700 			space_map_ref_destroy(&reftree);
1701 		}
1702 		space_map_vacate(&vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
1703 		space_map_walk(&vd->vdev_dtl[DTL_MISSING],
1704 		    space_map_add, &vd->vdev_dtl[DTL_PARTIAL]);
1705 		if (scrub_done)
1706 			space_map_vacate(&vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
1707 		space_map_vacate(&vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
1708 		if (!vdev_readable(vd))
1709 			space_map_add(&vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
1710 		else
1711 			space_map_walk(&vd->vdev_dtl[DTL_MISSING],
1712 			    space_map_add, &vd->vdev_dtl[DTL_OUTAGE]);
1713 		mutex_exit(&vd->vdev_dtl_lock);
1714 
1715 		if (txg != 0)
1716 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
1717 		return;
1718 	}
1719 
1720 	mutex_enter(&vd->vdev_dtl_lock);
1721 	for (int t = 0; t < DTL_TYPES; t++) {
1722 		/* account for child's outage in parent's missing map */
1723 		int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
1724 		if (t == DTL_SCRUB)
1725 			continue;			/* leaf vdevs only */
1726 		if (t == DTL_PARTIAL)
1727 			minref = 1;			/* i.e. non-zero */
1728 		else if (vd->vdev_nparity != 0)
1729 			minref = vd->vdev_nparity + 1;	/* RAID-Z */
1730 		else
1731 			minref = vd->vdev_children;	/* any kind of mirror */
1732 		space_map_ref_create(&reftree);
1733 		for (int c = 0; c < vd->vdev_children; c++) {
1734 			vdev_t *cvd = vd->vdev_child[c];
1735 			mutex_enter(&cvd->vdev_dtl_lock);
1736 			space_map_ref_add_map(&reftree, &cvd->vdev_dtl[s], 1);
1737 			mutex_exit(&cvd->vdev_dtl_lock);
1738 		}
1739 		space_map_ref_generate_map(&reftree, &vd->vdev_dtl[t], minref);
1740 		space_map_ref_destroy(&reftree);
1741 	}
1742 	mutex_exit(&vd->vdev_dtl_lock);
1743 }
1744 
1745 static int
1746 vdev_dtl_load(vdev_t *vd)
1747 {
1748 	spa_t *spa = vd->vdev_spa;
1749 	space_map_obj_t *smo = &vd->vdev_dtl_smo;
1750 	objset_t *mos = spa->spa_meta_objset;
1751 	dmu_buf_t *db;
1752 	int error;
1753 
1754 	ASSERT(vd->vdev_children == 0);
1755 
1756 	if (smo->smo_object == 0)
1757 		return (0);
1758 
1759 	ASSERT(!vd->vdev_ishole);
1760 
1761 	if ((error = dmu_bonus_hold(mos, smo->smo_object, FTAG, &db)) != 0)
1762 		return (error);
1763 
1764 	ASSERT3U(db->db_size, >=, sizeof (*smo));
1765 	bcopy(db->db_data, smo, sizeof (*smo));
1766 	dmu_buf_rele(db, FTAG);
1767 
1768 	mutex_enter(&vd->vdev_dtl_lock);
1769 	error = space_map_load(&vd->vdev_dtl[DTL_MISSING],
1770 	    NULL, SM_ALLOC, smo, mos);
1771 	mutex_exit(&vd->vdev_dtl_lock);
1772 
1773 	return (error);
1774 }
1775 
1776 void
1777 vdev_dtl_sync(vdev_t *vd, uint64_t txg)
1778 {
1779 	spa_t *spa = vd->vdev_spa;
1780 	space_map_obj_t *smo = &vd->vdev_dtl_smo;
1781 	space_map_t *sm = &vd->vdev_dtl[DTL_MISSING];
1782 	objset_t *mos = spa->spa_meta_objset;
1783 	space_map_t smsync;
1784 	kmutex_t smlock;
1785 	dmu_buf_t *db;
1786 	dmu_tx_t *tx;
1787 
1788 	ASSERT(!vd->vdev_ishole);
1789 
1790 	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1791 
1792 	if (vd->vdev_detached) {
1793 		if (smo->smo_object != 0) {
1794 			int err = dmu_object_free(mos, smo->smo_object, tx);
1795 			ASSERT3U(err, ==, 0);
1796 			smo->smo_object = 0;
1797 		}
1798 		dmu_tx_commit(tx);
1799 		return;
1800 	}
1801 
1802 	if (smo->smo_object == 0) {
1803 		ASSERT(smo->smo_objsize == 0);
1804 		ASSERT(smo->smo_alloc == 0);
1805 		smo->smo_object = dmu_object_alloc(mos,
1806 		    DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
1807 		    DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
1808 		ASSERT(smo->smo_object != 0);
1809 		vdev_config_dirty(vd->vdev_top);
1810 	}
1811 
1812 	mutex_init(&smlock, NULL, MUTEX_DEFAULT, NULL);
1813 
1814 	space_map_create(&smsync, sm->sm_start, sm->sm_size, sm->sm_shift,
1815 	    &smlock);
1816 
1817 	mutex_enter(&smlock);
1818 
1819 	mutex_enter(&vd->vdev_dtl_lock);
1820 	space_map_walk(sm, space_map_add, &smsync);
1821 	mutex_exit(&vd->vdev_dtl_lock);
1822 
1823 	space_map_truncate(smo, mos, tx);
1824 	space_map_sync(&smsync, SM_ALLOC, smo, mos, tx);
1825 
1826 	space_map_destroy(&smsync);
1827 
1828 	mutex_exit(&smlock);
1829 	mutex_destroy(&smlock);
1830 
1831 	VERIFY(0 == dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
1832 	dmu_buf_will_dirty(db, tx);
1833 	ASSERT3U(db->db_size, >=, sizeof (*smo));
1834 	bcopy(smo, db->db_data, sizeof (*smo));
1835 	dmu_buf_rele(db, FTAG);
1836 
1837 	dmu_tx_commit(tx);
1838 }
1839 
1840 /*
1841  * Determine whether the specified vdev can be offlined/detached/removed
1842  * without losing data.
1843  */
1844 boolean_t
1845 vdev_dtl_required(vdev_t *vd)
1846 {
1847 	spa_t *spa = vd->vdev_spa;
1848 	vdev_t *tvd = vd->vdev_top;
1849 	uint8_t cant_read = vd->vdev_cant_read;
1850 	boolean_t required;
1851 
1852 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1853 
1854 	if (vd == spa->spa_root_vdev || vd == tvd)
1855 		return (B_TRUE);
1856 
1857 	/*
1858 	 * Temporarily mark the device as unreadable, and then determine
1859 	 * whether this results in any DTL outages in the top-level vdev.
1860 	 * If not, we can safely offline/detach/remove the device.
1861 	 */
1862 	vd->vdev_cant_read = B_TRUE;
1863 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1864 	required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
1865 	vd->vdev_cant_read = cant_read;
1866 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1867 
1868 	if (!required && zio_injection_enabled)
1869 		required = !!zio_handle_device_injection(vd, NULL, ECHILD);
1870 
1871 	return (required);
1872 }
1873 
1874 /*
1875  * Determine if resilver is needed, and if so the txg range.
1876  */
1877 boolean_t
1878 vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
1879 {
1880 	boolean_t needed = B_FALSE;
1881 	uint64_t thismin = UINT64_MAX;
1882 	uint64_t thismax = 0;
1883 
1884 	if (vd->vdev_children == 0) {
1885 		mutex_enter(&vd->vdev_dtl_lock);
1886 		if (vd->vdev_dtl[DTL_MISSING].sm_space != 0 &&
1887 		    vdev_writeable(vd)) {
1888 			space_seg_t *ss;
1889 
1890 			ss = avl_first(&vd->vdev_dtl[DTL_MISSING].sm_root);
1891 			thismin = ss->ss_start - 1;
1892 			ss = avl_last(&vd->vdev_dtl[DTL_MISSING].sm_root);
1893 			thismax = ss->ss_end;
1894 			needed = B_TRUE;
1895 		}
1896 		mutex_exit(&vd->vdev_dtl_lock);
1897 	} else {
1898 		for (int c = 0; c < vd->vdev_children; c++) {
1899 			vdev_t *cvd = vd->vdev_child[c];
1900 			uint64_t cmin, cmax;
1901 
1902 			if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
1903 				thismin = MIN(thismin, cmin);
1904 				thismax = MAX(thismax, cmax);
1905 				needed = B_TRUE;
1906 			}
1907 		}
1908 	}
1909 
1910 	if (needed && minp) {
1911 		*minp = thismin;
1912 		*maxp = thismax;
1913 	}
1914 	return (needed);
1915 }
1916 
1917 void
1918 vdev_load(vdev_t *vd)
1919 {
1920 	/*
1921 	 * Recursively load all children.
1922 	 */
1923 	for (int c = 0; c < vd->vdev_children; c++)
1924 		vdev_load(vd->vdev_child[c]);
1925 
1926 	/*
1927 	 * If this is a top-level vdev, initialize its metaslabs.
1928 	 */
1929 	if (vd == vd->vdev_top && !vd->vdev_ishole &&
1930 	    (vd->vdev_ashift == 0 || vd->vdev_asize == 0 ||
1931 	    vdev_metaslab_init(vd, 0) != 0))
1932 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1933 		    VDEV_AUX_CORRUPT_DATA);
1934 
1935 	/*
1936 	 * If this is a leaf vdev, load its DTL.
1937 	 */
1938 	if (vd->vdev_ops->vdev_op_leaf && vdev_dtl_load(vd) != 0)
1939 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1940 		    VDEV_AUX_CORRUPT_DATA);
1941 }
1942 
1943 /*
1944  * The special vdev case is used for hot spares and l2cache devices.  Its
1945  * sole purpose it to set the vdev state for the associated vdev.  To do this,
1946  * we make sure that we can open the underlying device, then try to read the
1947  * label, and make sure that the label is sane and that it hasn't been
1948  * repurposed to another pool.
1949  */
1950 int
1951 vdev_validate_aux(vdev_t *vd)
1952 {
1953 	nvlist_t *label;
1954 	uint64_t guid, version;
1955 	uint64_t state;
1956 
1957 	if (!vdev_readable(vd))
1958 		return (0);
1959 
1960 	if ((label = vdev_label_read_config(vd)) == NULL) {
1961 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1962 		    VDEV_AUX_CORRUPT_DATA);
1963 		return (-1);
1964 	}
1965 
1966 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
1967 	    version > SPA_VERSION ||
1968 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
1969 	    guid != vd->vdev_guid ||
1970 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
1971 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1972 		    VDEV_AUX_CORRUPT_DATA);
1973 		nvlist_free(label);
1974 		return (-1);
1975 	}
1976 
1977 	/*
1978 	 * We don't actually check the pool state here.  If it's in fact in
1979 	 * use by another pool, we update this fact on the fly when requested.
1980 	 */
1981 	nvlist_free(label);
1982 	return (0);
1983 }
1984 
1985 void
1986 vdev_remove(vdev_t *vd, uint64_t txg)
1987 {
1988 	spa_t *spa = vd->vdev_spa;
1989 	objset_t *mos = spa->spa_meta_objset;
1990 	dmu_tx_t *tx;
1991 
1992 	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
1993 
1994 	if (vd->vdev_dtl_smo.smo_object) {
1995 		ASSERT3U(vd->vdev_dtl_smo.smo_alloc, ==, 0);
1996 		(void) dmu_object_free(mos, vd->vdev_dtl_smo.smo_object, tx);
1997 		vd->vdev_dtl_smo.smo_object = 0;
1998 	}
1999 
2000 	if (vd->vdev_ms != NULL) {
2001 		for (int m = 0; m < vd->vdev_ms_count; m++) {
2002 			metaslab_t *msp = vd->vdev_ms[m];
2003 
2004 			if (msp == NULL || msp->ms_smo.smo_object == 0)
2005 				continue;
2006 
2007 			ASSERT3U(msp->ms_smo.smo_alloc, ==, 0);
2008 			(void) dmu_object_free(mos, msp->ms_smo.smo_object, tx);
2009 			msp->ms_smo.smo_object = 0;
2010 		}
2011 	}
2012 
2013 	if (vd->vdev_ms_array) {
2014 		(void) dmu_object_free(mos, vd->vdev_ms_array, tx);
2015 		vd->vdev_ms_array = 0;
2016 		vd->vdev_ms_shift = 0;
2017 	}
2018 	dmu_tx_commit(tx);
2019 }
2020 
2021 void
2022 vdev_sync_done(vdev_t *vd, uint64_t txg)
2023 {
2024 	metaslab_t *msp;
2025 	boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg));
2026 
2027 	ASSERT(!vd->vdev_ishole);
2028 
2029 	while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
2030 		metaslab_sync_done(msp, txg);
2031 
2032 	if (reassess)
2033 		metaslab_sync_reassess(vd->vdev_mg);
2034 }
2035 
2036 void
2037 vdev_sync(vdev_t *vd, uint64_t txg)
2038 {
2039 	spa_t *spa = vd->vdev_spa;
2040 	vdev_t *lvd;
2041 	metaslab_t *msp;
2042 	dmu_tx_t *tx;
2043 
2044 	ASSERT(!vd->vdev_ishole);
2045 
2046 	if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0) {
2047 		ASSERT(vd == vd->vdev_top);
2048 		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2049 		vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
2050 		    DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
2051 		ASSERT(vd->vdev_ms_array != 0);
2052 		vdev_config_dirty(vd);
2053 		dmu_tx_commit(tx);
2054 	}
2055 
2056 	/*
2057 	 * Remove the metadata associated with this vdev once it's empty.
2058 	 */
2059 	if (vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing)
2060 		vdev_remove(vd, txg);
2061 
2062 	while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
2063 		metaslab_sync(msp, txg);
2064 		(void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
2065 	}
2066 
2067 	while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
2068 		vdev_dtl_sync(lvd, txg);
2069 
2070 	(void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
2071 }
2072 
2073 uint64_t
2074 vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
2075 {
2076 	return (vd->vdev_ops->vdev_op_asize(vd, psize));
2077 }
2078 
2079 /*
2080  * Mark the given vdev faulted.  A faulted vdev behaves as if the device could
2081  * not be opened, and no I/O is attempted.
2082  */
2083 int
2084 vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2085 {
2086 	vdev_t *vd, *tvd;
2087 
2088 	spa_vdev_state_enter(spa, SCL_NONE);
2089 
2090 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2091 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2092 
2093 	if (!vd->vdev_ops->vdev_op_leaf)
2094 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2095 
2096 	tvd = vd->vdev_top;
2097 
2098 	/*
2099 	 * We don't directly use the aux state here, but if we do a
2100 	 * vdev_reopen(), we need this value to be present to remember why we
2101 	 * were faulted.
2102 	 */
2103 	vd->vdev_label_aux = aux;
2104 
2105 	/*
2106 	 * Faulted state takes precedence over degraded.
2107 	 */
2108 	vd->vdev_delayed_close = B_FALSE;
2109 	vd->vdev_faulted = 1ULL;
2110 	vd->vdev_degraded = 0ULL;
2111 	vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
2112 
2113 	/*
2114 	 * If this device has the only valid copy of the data, then
2115 	 * back off and simply mark the vdev as degraded instead.
2116 	 */
2117 	if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) {
2118 		vd->vdev_degraded = 1ULL;
2119 		vd->vdev_faulted = 0ULL;
2120 
2121 		/*
2122 		 * If we reopen the device and it's not dead, only then do we
2123 		 * mark it degraded.
2124 		 */
2125 		vdev_reopen(tvd);
2126 
2127 		if (vdev_readable(vd))
2128 			vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
2129 	}
2130 
2131 	return (spa_vdev_state_exit(spa, vd, 0));
2132 }
2133 
2134 /*
2135  * Mark the given vdev degraded.  A degraded vdev is purely an indication to the
2136  * user that something is wrong.  The vdev continues to operate as normal as far
2137  * as I/O is concerned.
2138  */
2139 int
2140 vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2141 {
2142 	vdev_t *vd;
2143 
2144 	spa_vdev_state_enter(spa, SCL_NONE);
2145 
2146 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2147 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2148 
2149 	if (!vd->vdev_ops->vdev_op_leaf)
2150 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2151 
2152 	/*
2153 	 * If the vdev is already faulted, then don't do anything.
2154 	 */
2155 	if (vd->vdev_faulted || vd->vdev_degraded)
2156 		return (spa_vdev_state_exit(spa, NULL, 0));
2157 
2158 	vd->vdev_degraded = 1ULL;
2159 	if (!vdev_is_dead(vd))
2160 		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
2161 		    aux);
2162 
2163 	return (spa_vdev_state_exit(spa, vd, 0));
2164 }
2165 
2166 /*
2167  * Online the given vdev.  If 'unspare' is set, it implies two things.  First,
2168  * any attached spare device should be detached when the device finishes
2169  * resilvering.  Second, the online should be treated like a 'test' online case,
2170  * so no FMA events are generated if the device fails to open.
2171  */
2172 int
2173 vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
2174 {
2175 	vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
2176 
2177 	spa_vdev_state_enter(spa, SCL_NONE);
2178 
2179 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2180 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2181 
2182 	if (!vd->vdev_ops->vdev_op_leaf)
2183 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2184 
2185 	tvd = vd->vdev_top;
2186 	vd->vdev_offline = B_FALSE;
2187 	vd->vdev_tmpoffline = B_FALSE;
2188 	vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
2189 	vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
2190 
2191 	/* XXX - L2ARC 1.0 does not support expansion */
2192 	if (!vd->vdev_aux) {
2193 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2194 			pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
2195 	}
2196 
2197 	vdev_reopen(tvd);
2198 	vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
2199 
2200 	if (!vd->vdev_aux) {
2201 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2202 			pvd->vdev_expanding = B_FALSE;
2203 	}
2204 
2205 	if (newstate)
2206 		*newstate = vd->vdev_state;
2207 	if ((flags & ZFS_ONLINE_UNSPARE) &&
2208 	    !vdev_is_dead(vd) && vd->vdev_parent &&
2209 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2210 	    vd->vdev_parent->vdev_child[0] == vd)
2211 		vd->vdev_unspare = B_TRUE;
2212 
2213 	if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
2214 
2215 		/* XXX - L2ARC 1.0 does not support expansion */
2216 		if (vd->vdev_aux)
2217 			return (spa_vdev_state_exit(spa, vd, ENOTSUP));
2218 		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
2219 	}
2220 	return (spa_vdev_state_exit(spa, vd, 0));
2221 }
2222 
2223 static int
2224 vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
2225 {
2226 	vdev_t *vd, *tvd;
2227 	int error = 0;
2228 	uint64_t generation;
2229 	metaslab_group_t *mg;
2230 
2231 top:
2232 	spa_vdev_state_enter(spa, SCL_ALLOC);
2233 
2234 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2235 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2236 
2237 	if (!vd->vdev_ops->vdev_op_leaf)
2238 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2239 
2240 	tvd = vd->vdev_top;
2241 	mg = tvd->vdev_mg;
2242 	generation = spa->spa_config_generation + 1;
2243 
2244 	/*
2245 	 * If the device isn't already offline, try to offline it.
2246 	 */
2247 	if (!vd->vdev_offline) {
2248 		/*
2249 		 * If this device has the only valid copy of some data,
2250 		 * don't allow it to be offlined. Log devices are always
2251 		 * expendable.
2252 		 */
2253 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2254 		    vdev_dtl_required(vd))
2255 			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2256 
2257 		/*
2258 		 * If the top-level is a slog and it has had allocations
2259 		 * then proceed.  We check that the vdev's metaslab group
2260 		 * is not NULL since it's possible that we may have just
2261 		 * added this vdev but not yet initialized its metaslabs.
2262 		 */
2263 		if (tvd->vdev_islog && mg != NULL) {
2264 			/*
2265 			 * Prevent any future allocations.
2266 			 */
2267 			metaslab_group_passivate(mg);
2268 			(void) spa_vdev_state_exit(spa, vd, 0);
2269 
2270 			error = spa_offline_log(spa);
2271 
2272 			spa_vdev_state_enter(spa, SCL_ALLOC);
2273 
2274 			/*
2275 			 * Check to see if the config has changed.
2276 			 */
2277 			if (error || generation != spa->spa_config_generation) {
2278 				metaslab_group_activate(mg);
2279 				if (error)
2280 					return (spa_vdev_state_exit(spa,
2281 					    vd, error));
2282 				(void) spa_vdev_state_exit(spa, vd, 0);
2283 				goto top;
2284 			}
2285 			ASSERT3U(tvd->vdev_stat.vs_alloc, ==, 0);
2286 		}
2287 
2288 		/*
2289 		 * Offline this device and reopen its top-level vdev.
2290 		 * If the top-level vdev is a log device then just offline
2291 		 * it. Otherwise, if this action results in the top-level
2292 		 * vdev becoming unusable, undo it and fail the request.
2293 		 */
2294 		vd->vdev_offline = B_TRUE;
2295 		vdev_reopen(tvd);
2296 
2297 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2298 		    vdev_is_dead(tvd)) {
2299 			vd->vdev_offline = B_FALSE;
2300 			vdev_reopen(tvd);
2301 			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2302 		}
2303 
2304 		/*
2305 		 * Add the device back into the metaslab rotor so that
2306 		 * once we online the device it's open for business.
2307 		 */
2308 		if (tvd->vdev_islog && mg != NULL)
2309 			metaslab_group_activate(mg);
2310 	}
2311 
2312 	vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
2313 
2314 	return (spa_vdev_state_exit(spa, vd, 0));
2315 }
2316 
2317 int
2318 vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
2319 {
2320 	int error;
2321 
2322 	mutex_enter(&spa->spa_vdev_top_lock);
2323 	error = vdev_offline_locked(spa, guid, flags);
2324 	mutex_exit(&spa->spa_vdev_top_lock);
2325 
2326 	return (error);
2327 }
2328 
2329 /*
2330  * Clear the error counts associated with this vdev.  Unlike vdev_online() and
2331  * vdev_offline(), we assume the spa config is locked.  We also clear all
2332  * children.  If 'vd' is NULL, then the user wants to clear all vdevs.
2333  */
2334 void
2335 vdev_clear(spa_t *spa, vdev_t *vd)
2336 {
2337 	vdev_t *rvd = spa->spa_root_vdev;
2338 
2339 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2340 
2341 	if (vd == NULL)
2342 		vd = rvd;
2343 
2344 	vd->vdev_stat.vs_read_errors = 0;
2345 	vd->vdev_stat.vs_write_errors = 0;
2346 	vd->vdev_stat.vs_checksum_errors = 0;
2347 
2348 	for (int c = 0; c < vd->vdev_children; c++)
2349 		vdev_clear(spa, vd->vdev_child[c]);
2350 
2351 	/*
2352 	 * If we're in the FAULTED state or have experienced failed I/O, then
2353 	 * clear the persistent state and attempt to reopen the device.  We
2354 	 * also mark the vdev config dirty, so that the new faulted state is
2355 	 * written out to disk.
2356 	 */
2357 	if (vd->vdev_faulted || vd->vdev_degraded ||
2358 	    !vdev_readable(vd) || !vdev_writeable(vd)) {
2359 
2360 		/*
2361 		 * When reopening in reponse to a clear event, it may be due to
2362 		 * a fmadm repair request.  In this case, if the device is
2363 		 * still broken, we want to still post the ereport again.
2364 		 */
2365 		vd->vdev_forcefault = B_TRUE;
2366 
2367 		vd->vdev_faulted = vd->vdev_degraded = 0ULL;
2368 		vd->vdev_cant_read = B_FALSE;
2369 		vd->vdev_cant_write = B_FALSE;
2370 
2371 		vdev_reopen(vd == rvd ? rvd : vd->vdev_top);
2372 
2373 		vd->vdev_forcefault = B_FALSE;
2374 
2375 		if (vd != rvd && vdev_writeable(vd->vdev_top))
2376 			vdev_state_dirty(vd->vdev_top);
2377 
2378 		if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
2379 			spa_async_request(spa, SPA_ASYNC_RESILVER);
2380 
2381 		spa_event_notify(spa, vd, ESC_ZFS_VDEV_CLEAR);
2382 	}
2383 
2384 	/*
2385 	 * When clearing a FMA-diagnosed fault, we always want to
2386 	 * unspare the device, as we assume that the original spare was
2387 	 * done in response to the FMA fault.
2388 	 */
2389 	if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
2390 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2391 	    vd->vdev_parent->vdev_child[0] == vd)
2392 		vd->vdev_unspare = B_TRUE;
2393 }
2394 
2395 boolean_t
2396 vdev_is_dead(vdev_t *vd)
2397 {
2398 	/*
2399 	 * Holes and missing devices are always considered "dead".
2400 	 * This simplifies the code since we don't have to check for
2401 	 * these types of devices in the various code paths.
2402 	 * Instead we rely on the fact that we skip over dead devices
2403 	 * before issuing I/O to them.
2404 	 */
2405 	return (vd->vdev_state < VDEV_STATE_DEGRADED || vd->vdev_ishole ||
2406 	    vd->vdev_ops == &vdev_missing_ops);
2407 }
2408 
2409 boolean_t
2410 vdev_readable(vdev_t *vd)
2411 {
2412 	return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
2413 }
2414 
2415 boolean_t
2416 vdev_writeable(vdev_t *vd)
2417 {
2418 	return (!vdev_is_dead(vd) && !vd->vdev_cant_write);
2419 }
2420 
2421 boolean_t
2422 vdev_allocatable(vdev_t *vd)
2423 {
2424 	uint64_t state = vd->vdev_state;
2425 
2426 	/*
2427 	 * We currently allow allocations from vdevs which may be in the
2428 	 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
2429 	 * fails to reopen then we'll catch it later when we're holding
2430 	 * the proper locks.  Note that we have to get the vdev state
2431 	 * in a local variable because although it changes atomically,
2432 	 * we're asking two separate questions about it.
2433 	 */
2434 	return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
2435 	    !vd->vdev_cant_write && !vd->vdev_ishole);
2436 }
2437 
2438 boolean_t
2439 vdev_accessible(vdev_t *vd, zio_t *zio)
2440 {
2441 	ASSERT(zio->io_vd == vd);
2442 
2443 	if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
2444 		return (B_FALSE);
2445 
2446 	if (zio->io_type == ZIO_TYPE_READ)
2447 		return (!vd->vdev_cant_read);
2448 
2449 	if (zio->io_type == ZIO_TYPE_WRITE)
2450 		return (!vd->vdev_cant_write);
2451 
2452 	return (B_TRUE);
2453 }
2454 
2455 /*
2456  * Get statistics for the given vdev.
2457  */
2458 void
2459 vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
2460 {
2461 	vdev_t *rvd = vd->vdev_spa->spa_root_vdev;
2462 
2463 	mutex_enter(&vd->vdev_stat_lock);
2464 	bcopy(&vd->vdev_stat, vs, sizeof (*vs));
2465 	vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
2466 	vs->vs_state = vd->vdev_state;
2467 	vs->vs_rsize = vdev_get_min_asize(vd);
2468 	if (vd->vdev_ops->vdev_op_leaf)
2469 		vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
2470 	vs->vs_esize = vd->vdev_max_asize - vd->vdev_asize;
2471 	mutex_exit(&vd->vdev_stat_lock);
2472 
2473 	/*
2474 	 * If we're getting stats on the root vdev, aggregate the I/O counts
2475 	 * over all top-level vdevs (i.e. the direct children of the root).
2476 	 */
2477 	if (vd == rvd) {
2478 		for (int c = 0; c < rvd->vdev_children; c++) {
2479 			vdev_t *cvd = rvd->vdev_child[c];
2480 			vdev_stat_t *cvs = &cvd->vdev_stat;
2481 
2482 			mutex_enter(&vd->vdev_stat_lock);
2483 			for (int t = 0; t < ZIO_TYPES; t++) {
2484 				vs->vs_ops[t] += cvs->vs_ops[t];
2485 				vs->vs_bytes[t] += cvs->vs_bytes[t];
2486 			}
2487 			cvs->vs_scan_removing = cvd->vdev_removing;
2488 			mutex_exit(&vd->vdev_stat_lock);
2489 		}
2490 	}
2491 }
2492 
2493 void
2494 vdev_clear_stats(vdev_t *vd)
2495 {
2496 	mutex_enter(&vd->vdev_stat_lock);
2497 	vd->vdev_stat.vs_space = 0;
2498 	vd->vdev_stat.vs_dspace = 0;
2499 	vd->vdev_stat.vs_alloc = 0;
2500 	mutex_exit(&vd->vdev_stat_lock);
2501 }
2502 
2503 void
2504 vdev_scan_stat_init(vdev_t *vd)
2505 {
2506 	vdev_stat_t *vs = &vd->vdev_stat;
2507 
2508 	for (int c = 0; c < vd->vdev_children; c++)
2509 		vdev_scan_stat_init(vd->vdev_child[c]);
2510 
2511 	mutex_enter(&vd->vdev_stat_lock);
2512 	vs->vs_scan_processed = 0;
2513 	mutex_exit(&vd->vdev_stat_lock);
2514 }
2515 
2516 void
2517 vdev_stat_update(zio_t *zio, uint64_t psize)
2518 {
2519 	spa_t *spa = zio->io_spa;
2520 	vdev_t *rvd = spa->spa_root_vdev;
2521 	vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
2522 	vdev_t *pvd;
2523 	uint64_t txg = zio->io_txg;
2524 	vdev_stat_t *vs = &vd->vdev_stat;
2525 	zio_type_t type = zio->io_type;
2526 	int flags = zio->io_flags;
2527 
2528 	/*
2529 	 * If this i/o is a gang leader, it didn't do any actual work.
2530 	 */
2531 	if (zio->io_gang_tree)
2532 		return;
2533 
2534 	if (zio->io_error == 0) {
2535 		/*
2536 		 * If this is a root i/o, don't count it -- we've already
2537 		 * counted the top-level vdevs, and vdev_get_stats() will
2538 		 * aggregate them when asked.  This reduces contention on
2539 		 * the root vdev_stat_lock and implicitly handles blocks
2540 		 * that compress away to holes, for which there is no i/o.
2541 		 * (Holes never create vdev children, so all the counters
2542 		 * remain zero, which is what we want.)
2543 		 *
2544 		 * Note: this only applies to successful i/o (io_error == 0)
2545 		 * because unlike i/o counts, errors are not additive.
2546 		 * When reading a ditto block, for example, failure of
2547 		 * one top-level vdev does not imply a root-level error.
2548 		 */
2549 		if (vd == rvd)
2550 			return;
2551 
2552 		ASSERT(vd == zio->io_vd);
2553 
2554 		if (flags & ZIO_FLAG_IO_BYPASS)
2555 			return;
2556 
2557 		mutex_enter(&vd->vdev_stat_lock);
2558 
2559 		if (flags & ZIO_FLAG_IO_REPAIR) {
2560 			if (flags & ZIO_FLAG_SCAN_THREAD) {
2561 				dsl_scan_phys_t *scn_phys =
2562 				    &spa->spa_dsl_pool->dp_scan->scn_phys;
2563 				uint64_t *processed = &scn_phys->scn_processed;
2564 
2565 				/* XXX cleanup? */
2566 				if (vd->vdev_ops->vdev_op_leaf)
2567 					atomic_add_64(processed, psize);
2568 				vs->vs_scan_processed += psize;
2569 			}
2570 
2571 			if (flags & ZIO_FLAG_SELF_HEAL)
2572 				vs->vs_self_healed += psize;
2573 		}
2574 
2575 		vs->vs_ops[type]++;
2576 		vs->vs_bytes[type] += psize;
2577 
2578 		mutex_exit(&vd->vdev_stat_lock);
2579 		return;
2580 	}
2581 
2582 	if (flags & ZIO_FLAG_SPECULATIVE)
2583 		return;
2584 
2585 	/*
2586 	 * If this is an I/O error that is going to be retried, then ignore the
2587 	 * error.  Otherwise, the user may interpret B_FAILFAST I/O errors as
2588 	 * hard errors, when in reality they can happen for any number of
2589 	 * innocuous reasons (bus resets, MPxIO link failure, etc).
2590 	 */
2591 	if (zio->io_error == EIO &&
2592 	    !(zio->io_flags & ZIO_FLAG_IO_RETRY))
2593 		return;
2594 
2595 	/*
2596 	 * Intent logs writes won't propagate their error to the root
2597 	 * I/O so don't mark these types of failures as pool-level
2598 	 * errors.
2599 	 */
2600 	if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
2601 		return;
2602 
2603 	mutex_enter(&vd->vdev_stat_lock);
2604 	if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
2605 		if (zio->io_error == ECKSUM)
2606 			vs->vs_checksum_errors++;
2607 		else
2608 			vs->vs_read_errors++;
2609 	}
2610 	if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
2611 		vs->vs_write_errors++;
2612 	mutex_exit(&vd->vdev_stat_lock);
2613 
2614 	if (type == ZIO_TYPE_WRITE && txg != 0 &&
2615 	    (!(flags & ZIO_FLAG_IO_REPAIR) ||
2616 	    (flags & ZIO_FLAG_SCAN_THREAD) ||
2617 	    spa->spa_claiming)) {
2618 		/*
2619 		 * This is either a normal write (not a repair), or it's
2620 		 * a repair induced by the scrub thread, or it's a repair
2621 		 * made by zil_claim() during spa_load() in the first txg.
2622 		 * In the normal case, we commit the DTL change in the same
2623 		 * txg as the block was born.  In the scrub-induced repair
2624 		 * case, we know that scrubs run in first-pass syncing context,
2625 		 * so we commit the DTL change in spa_syncing_txg(spa).
2626 		 * In the zil_claim() case, we commit in spa_first_txg(spa).
2627 		 *
2628 		 * We currently do not make DTL entries for failed spontaneous
2629 		 * self-healing writes triggered by normal (non-scrubbing)
2630 		 * reads, because we have no transactional context in which to
2631 		 * do so -- and it's not clear that it'd be desirable anyway.
2632 		 */
2633 		if (vd->vdev_ops->vdev_op_leaf) {
2634 			uint64_t commit_txg = txg;
2635 			if (flags & ZIO_FLAG_SCAN_THREAD) {
2636 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2637 				ASSERT(spa_sync_pass(spa) == 1);
2638 				vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
2639 				commit_txg = spa_syncing_txg(spa);
2640 			} else if (spa->spa_claiming) {
2641 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2642 				commit_txg = spa_first_txg(spa);
2643 			}
2644 			ASSERT(commit_txg >= spa_syncing_txg(spa));
2645 			if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
2646 				return;
2647 			for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2648 				vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
2649 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
2650 		}
2651 		if (vd != rvd)
2652 			vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
2653 	}
2654 }
2655 
2656 /*
2657  * Update the in-core space usage stats for this vdev, its metaslab class,
2658  * and the root vdev.
2659  */
2660 void
2661 vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
2662     int64_t space_delta)
2663 {
2664 	int64_t dspace_delta = space_delta;
2665 	spa_t *spa = vd->vdev_spa;
2666 	vdev_t *rvd = spa->spa_root_vdev;
2667 	metaslab_group_t *mg = vd->vdev_mg;
2668 	metaslab_class_t *mc = mg ? mg->mg_class : NULL;
2669 
2670 	ASSERT(vd == vd->vdev_top);
2671 
2672 	/*
2673 	 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
2674 	 * factor.  We must calculate this here and not at the root vdev
2675 	 * because the root vdev's psize-to-asize is simply the max of its
2676 	 * childrens', thus not accurate enough for us.
2677 	 */
2678 	ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
2679 	ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
2680 	dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
2681 	    vd->vdev_deflate_ratio;
2682 
2683 	mutex_enter(&vd->vdev_stat_lock);
2684 	vd->vdev_stat.vs_alloc += alloc_delta;
2685 	vd->vdev_stat.vs_space += space_delta;
2686 	vd->vdev_stat.vs_dspace += dspace_delta;
2687 	mutex_exit(&vd->vdev_stat_lock);
2688 
2689 	if (mc == spa_normal_class(spa)) {
2690 		mutex_enter(&rvd->vdev_stat_lock);
2691 		rvd->vdev_stat.vs_alloc += alloc_delta;
2692 		rvd->vdev_stat.vs_space += space_delta;
2693 		rvd->vdev_stat.vs_dspace += dspace_delta;
2694 		mutex_exit(&rvd->vdev_stat_lock);
2695 	}
2696 
2697 	if (mc != NULL) {
2698 		ASSERT(rvd == vd->vdev_parent);
2699 		ASSERT(vd->vdev_ms_count != 0);
2700 
2701 		metaslab_class_space_update(mc,
2702 		    alloc_delta, defer_delta, space_delta, dspace_delta);
2703 	}
2704 }
2705 
2706 /*
2707  * Mark a top-level vdev's config as dirty, placing it on the dirty list
2708  * so that it will be written out next time the vdev configuration is synced.
2709  * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
2710  */
2711 void
2712 vdev_config_dirty(vdev_t *vd)
2713 {
2714 	spa_t *spa = vd->vdev_spa;
2715 	vdev_t *rvd = spa->spa_root_vdev;
2716 	int c;
2717 
2718 	ASSERT(spa_writeable(spa));
2719 
2720 	/*
2721 	 * If this is an aux vdev (as with l2cache and spare devices), then we
2722 	 * update the vdev config manually and set the sync flag.
2723 	 */
2724 	if (vd->vdev_aux != NULL) {
2725 		spa_aux_vdev_t *sav = vd->vdev_aux;
2726 		nvlist_t **aux;
2727 		uint_t naux;
2728 
2729 		for (c = 0; c < sav->sav_count; c++) {
2730 			if (sav->sav_vdevs[c] == vd)
2731 				break;
2732 		}
2733 
2734 		if (c == sav->sav_count) {
2735 			/*
2736 			 * We're being removed.  There's nothing more to do.
2737 			 */
2738 			ASSERT(sav->sav_sync == B_TRUE);
2739 			return;
2740 		}
2741 
2742 		sav->sav_sync = B_TRUE;
2743 
2744 		if (nvlist_lookup_nvlist_array(sav->sav_config,
2745 		    ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
2746 			VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
2747 			    ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
2748 		}
2749 
2750 		ASSERT(c < naux);
2751 
2752 		/*
2753 		 * Setting the nvlist in the middle if the array is a little
2754 		 * sketchy, but it will work.
2755 		 */
2756 		nvlist_free(aux[c]);
2757 		aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0);
2758 
2759 		return;
2760 	}
2761 
2762 	/*
2763 	 * The dirty list is protected by the SCL_CONFIG lock.  The caller
2764 	 * must either hold SCL_CONFIG as writer, or must be the sync thread
2765 	 * (which holds SCL_CONFIG as reader).  There's only one sync thread,
2766 	 * so this is sufficient to ensure mutual exclusion.
2767 	 */
2768 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2769 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2770 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
2771 
2772 	if (vd == rvd) {
2773 		for (c = 0; c < rvd->vdev_children; c++)
2774 			vdev_config_dirty(rvd->vdev_child[c]);
2775 	} else {
2776 		ASSERT(vd == vd->vdev_top);
2777 
2778 		if (!list_link_active(&vd->vdev_config_dirty_node) &&
2779 		    !vd->vdev_ishole)
2780 			list_insert_head(&spa->spa_config_dirty_list, vd);
2781 	}
2782 }
2783 
2784 void
2785 vdev_config_clean(vdev_t *vd)
2786 {
2787 	spa_t *spa = vd->vdev_spa;
2788 
2789 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2790 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2791 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
2792 
2793 	ASSERT(list_link_active(&vd->vdev_config_dirty_node));
2794 	list_remove(&spa->spa_config_dirty_list, vd);
2795 }
2796 
2797 /*
2798  * Mark a top-level vdev's state as dirty, so that the next pass of
2799  * spa_sync() can convert this into vdev_config_dirty().  We distinguish
2800  * the state changes from larger config changes because they require
2801  * much less locking, and are often needed for administrative actions.
2802  */
2803 void
2804 vdev_state_dirty(vdev_t *vd)
2805 {
2806 	spa_t *spa = vd->vdev_spa;
2807 
2808 	ASSERT(spa_writeable(spa));
2809 	ASSERT(vd == vd->vdev_top);
2810 
2811 	/*
2812 	 * The state list is protected by the SCL_STATE lock.  The caller
2813 	 * must either hold SCL_STATE as writer, or must be the sync thread
2814 	 * (which holds SCL_STATE as reader).  There's only one sync thread,
2815 	 * so this is sufficient to ensure mutual exclusion.
2816 	 */
2817 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2818 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2819 	    spa_config_held(spa, SCL_STATE, RW_READER)));
2820 
2821 	if (!list_link_active(&vd->vdev_state_dirty_node) && !vd->vdev_ishole)
2822 		list_insert_head(&spa->spa_state_dirty_list, vd);
2823 }
2824 
2825 void
2826 vdev_state_clean(vdev_t *vd)
2827 {
2828 	spa_t *spa = vd->vdev_spa;
2829 
2830 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2831 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2832 	    spa_config_held(spa, SCL_STATE, RW_READER)));
2833 
2834 	ASSERT(list_link_active(&vd->vdev_state_dirty_node));
2835 	list_remove(&spa->spa_state_dirty_list, vd);
2836 }
2837 
2838 /*
2839  * Propagate vdev state up from children to parent.
2840  */
2841 void
2842 vdev_propagate_state(vdev_t *vd)
2843 {
2844 	spa_t *spa = vd->vdev_spa;
2845 	vdev_t *rvd = spa->spa_root_vdev;
2846 	int degraded = 0, faulted = 0;
2847 	int corrupted = 0;
2848 	vdev_t *child;
2849 
2850 	if (vd->vdev_children > 0) {
2851 		for (int c = 0; c < vd->vdev_children; c++) {
2852 			child = vd->vdev_child[c];
2853 
2854 			/*
2855 			 * Don't factor holes into the decision.
2856 			 */
2857 			if (child->vdev_ishole)
2858 				continue;
2859 
2860 			if (!vdev_readable(child) ||
2861 			    (!vdev_writeable(child) && spa_writeable(spa))) {
2862 				/*
2863 				 * Root special: if there is a top-level log
2864 				 * device, treat the root vdev as if it were
2865 				 * degraded.
2866 				 */
2867 				if (child->vdev_islog && vd == rvd)
2868 					degraded++;
2869 				else
2870 					faulted++;
2871 			} else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
2872 				degraded++;
2873 			}
2874 
2875 			if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
2876 				corrupted++;
2877 		}
2878 
2879 		vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
2880 
2881 		/*
2882 		 * Root special: if there is a top-level vdev that cannot be
2883 		 * opened due to corrupted metadata, then propagate the root
2884 		 * vdev's aux state as 'corrupt' rather than 'insufficient
2885 		 * replicas'.
2886 		 */
2887 		if (corrupted && vd == rvd &&
2888 		    rvd->vdev_state == VDEV_STATE_CANT_OPEN)
2889 			vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
2890 			    VDEV_AUX_CORRUPT_DATA);
2891 	}
2892 
2893 	if (vd->vdev_parent)
2894 		vdev_propagate_state(vd->vdev_parent);
2895 }
2896 
2897 /*
2898  * Set a vdev's state.  If this is during an open, we don't update the parent
2899  * state, because we're in the process of opening children depth-first.
2900  * Otherwise, we propagate the change to the parent.
2901  *
2902  * If this routine places a device in a faulted state, an appropriate ereport is
2903  * generated.
2904  */
2905 void
2906 vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
2907 {
2908 	uint64_t save_state;
2909 	spa_t *spa = vd->vdev_spa;
2910 
2911 	if (state == vd->vdev_state) {
2912 		vd->vdev_stat.vs_aux = aux;
2913 		return;
2914 	}
2915 
2916 	save_state = vd->vdev_state;
2917 
2918 	vd->vdev_state = state;
2919 	vd->vdev_stat.vs_aux = aux;
2920 
2921 	/*
2922 	 * If we are setting the vdev state to anything but an open state, then
2923 	 * always close the underlying device unless the device has requested
2924 	 * a delayed close (i.e. we're about to remove or fault the device).
2925 	 * Otherwise, we keep accessible but invalid devices open forever.
2926 	 * We don't call vdev_close() itself, because that implies some extra
2927 	 * checks (offline, etc) that we don't want here.  This is limited to
2928 	 * leaf devices, because otherwise closing the device will affect other
2929 	 * children.
2930 	 */
2931 	if (!vd->vdev_delayed_close && vdev_is_dead(vd) &&
2932 	    vd->vdev_ops->vdev_op_leaf)
2933 		vd->vdev_ops->vdev_op_close(vd);
2934 
2935 	/*
2936 	 * If we have brought this vdev back into service, we need
2937 	 * to notify fmd so that it can gracefully repair any outstanding
2938 	 * cases due to a missing device.  We do this in all cases, even those
2939 	 * that probably don't correlate to a repaired fault.  This is sure to
2940 	 * catch all cases, and we let the zfs-retire agent sort it out.  If
2941 	 * this is a transient state it's OK, as the retire agent will
2942 	 * double-check the state of the vdev before repairing it.
2943 	 */
2944 	if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf &&
2945 	    vd->vdev_prevstate != state)
2946 		zfs_post_state_change(spa, vd);
2947 
2948 	if (vd->vdev_removed &&
2949 	    state == VDEV_STATE_CANT_OPEN &&
2950 	    (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
2951 		/*
2952 		 * If the previous state is set to VDEV_STATE_REMOVED, then this
2953 		 * device was previously marked removed and someone attempted to
2954 		 * reopen it.  If this failed due to a nonexistent device, then
2955 		 * keep the device in the REMOVED state.  We also let this be if
2956 		 * it is one of our special test online cases, which is only
2957 		 * attempting to online the device and shouldn't generate an FMA
2958 		 * fault.
2959 		 */
2960 		vd->vdev_state = VDEV_STATE_REMOVED;
2961 		vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
2962 	} else if (state == VDEV_STATE_REMOVED) {
2963 		vd->vdev_removed = B_TRUE;
2964 	} else if (state == VDEV_STATE_CANT_OPEN) {
2965 		/*
2966 		 * If we fail to open a vdev during an import or recovery, we
2967 		 * mark it as "not available", which signifies that it was
2968 		 * never there to begin with.  Failure to open such a device
2969 		 * is not considered an error.
2970 		 */
2971 		if ((spa_load_state(spa) == SPA_LOAD_IMPORT ||
2972 		    spa_load_state(spa) == SPA_LOAD_RECOVER) &&
2973 		    vd->vdev_ops->vdev_op_leaf)
2974 			vd->vdev_not_present = 1;
2975 
2976 		/*
2977 		 * Post the appropriate ereport.  If the 'prevstate' field is
2978 		 * set to something other than VDEV_STATE_UNKNOWN, it indicates
2979 		 * that this is part of a vdev_reopen().  In this case, we don't
2980 		 * want to post the ereport if the device was already in the
2981 		 * CANT_OPEN state beforehand.
2982 		 *
2983 		 * If the 'checkremove' flag is set, then this is an attempt to
2984 		 * online the device in response to an insertion event.  If we
2985 		 * hit this case, then we have detected an insertion event for a
2986 		 * faulted or offline device that wasn't in the removed state.
2987 		 * In this scenario, we don't post an ereport because we are
2988 		 * about to replace the device, or attempt an online with
2989 		 * vdev_forcefault, which will generate the fault for us.
2990 		 */
2991 		if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
2992 		    !vd->vdev_not_present && !vd->vdev_checkremove &&
2993 		    vd != spa->spa_root_vdev) {
2994 			const char *class;
2995 
2996 			switch (aux) {
2997 			case VDEV_AUX_OPEN_FAILED:
2998 				class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
2999 				break;
3000 			case VDEV_AUX_CORRUPT_DATA:
3001 				class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
3002 				break;
3003 			case VDEV_AUX_NO_REPLICAS:
3004 				class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
3005 				break;
3006 			case VDEV_AUX_BAD_GUID_SUM:
3007 				class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
3008 				break;
3009 			case VDEV_AUX_TOO_SMALL:
3010 				class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
3011 				break;
3012 			case VDEV_AUX_BAD_LABEL:
3013 				class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
3014 				break;
3015 			default:
3016 				class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
3017 			}
3018 
3019 			zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
3020 		}
3021 
3022 		/* Erase any notion of persistent removed state */
3023 		vd->vdev_removed = B_FALSE;
3024 	} else {
3025 		vd->vdev_removed = B_FALSE;
3026 	}
3027 
3028 	if (!isopen && vd->vdev_parent)
3029 		vdev_propagate_state(vd->vdev_parent);
3030 }
3031 
3032 /*
3033  * Check the vdev configuration to ensure that it's capable of supporting
3034  * a root pool. Currently, we do not support RAID-Z or partial configuration.
3035  * In addition, only a single top-level vdev is allowed and none of the leaves
3036  * can be wholedisks.
3037  */
3038 boolean_t
3039 vdev_is_bootable(vdev_t *vd)
3040 {
3041 	if (!vd->vdev_ops->vdev_op_leaf) {
3042 		char *vdev_type = vd->vdev_ops->vdev_op_type;
3043 
3044 		if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
3045 		    vd->vdev_children > 1) {
3046 			return (B_FALSE);
3047 		} else if (strcmp(vdev_type, VDEV_TYPE_RAIDZ) == 0 ||
3048 		    strcmp(vdev_type, VDEV_TYPE_MISSING) == 0) {
3049 			return (B_FALSE);
3050 		}
3051 	} else if (vd->vdev_wholedisk == 1) {
3052 		return (B_FALSE);
3053 	}
3054 
3055 	for (int c = 0; c < vd->vdev_children; c++) {
3056 		if (!vdev_is_bootable(vd->vdev_child[c]))
3057 			return (B_FALSE);
3058 	}
3059 	return (B_TRUE);
3060 }
3061 
3062 /*
3063  * Load the state from the original vdev tree (ovd) which
3064  * we've retrieved from the MOS config object. If the original
3065  * vdev was offline or faulted then we transfer that state to the
3066  * device in the current vdev tree (nvd).
3067  */
3068 void
3069 vdev_load_log_state(vdev_t *nvd, vdev_t *ovd)
3070 {
3071 	spa_t *spa = nvd->vdev_spa;
3072 
3073 	ASSERT(nvd->vdev_top->vdev_islog);
3074 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3075 	ASSERT3U(nvd->vdev_guid, ==, ovd->vdev_guid);
3076 
3077 	for (int c = 0; c < nvd->vdev_children; c++)
3078 		vdev_load_log_state(nvd->vdev_child[c], ovd->vdev_child[c]);
3079 
3080 	if (nvd->vdev_ops->vdev_op_leaf) {
3081 		/*
3082 		 * Restore the persistent vdev state
3083 		 */
3084 		nvd->vdev_offline = ovd->vdev_offline;
3085 		nvd->vdev_faulted = ovd->vdev_faulted;
3086 		nvd->vdev_degraded = ovd->vdev_degraded;
3087 		nvd->vdev_removed = ovd->vdev_removed;
3088 	}
3089 }
3090 
3091 /*
3092  * Determine if a log device has valid content.  If the vdev was
3093  * removed or faulted in the MOS config then we know that
3094  * the content on the log device has already been written to the pool.
3095  */
3096 boolean_t
3097 vdev_log_state_valid(vdev_t *vd)
3098 {
3099 	if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted &&
3100 	    !vd->vdev_removed)
3101 		return (B_TRUE);
3102 
3103 	for (int c = 0; c < vd->vdev_children; c++)
3104 		if (vdev_log_state_valid(vd->vdev_child[c]))
3105 			return (B_TRUE);
3106 
3107 	return (B_FALSE);
3108 }
3109 
3110 /*
3111  * Expand a vdev if possible.
3112  */
3113 void
3114 vdev_expand(vdev_t *vd, uint64_t txg)
3115 {
3116 	ASSERT(vd->vdev_top == vd);
3117 	ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
3118 
3119 	if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count) {
3120 		VERIFY(vdev_metaslab_init(vd, txg) == 0);
3121 		vdev_config_dirty(vd);
3122 	}
3123 }
3124 
3125 /*
3126  * Split a vdev.
3127  */
3128 void
3129 vdev_split(vdev_t *vd)
3130 {
3131 	vdev_t *cvd, *pvd = vd->vdev_parent;
3132 
3133 	vdev_remove_child(pvd, vd);
3134 	vdev_compact_children(pvd);
3135 
3136 	cvd = pvd->vdev_child[0];
3137 	if (pvd->vdev_children == 1) {
3138 		vdev_remove_parent(cvd);
3139 		cvd->vdev_splitting = B_TRUE;
3140 	}
3141 	vdev_propagate_state(cvd);
3142 }
3143