xref: /titanic_51/usr/src/uts/common/fs/zfs/vdev.c (revision dd51520e127b452179a2ce4ea3bd8dee949f9afe)
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  */
25 
26 #include <sys/zfs_context.h>
27 #include <sys/fm/fs/zfs.h>
28 #include <sys/spa.h>
29 #include <sys/spa_impl.h>
30 #include <sys/dmu.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/uberblock_impl.h>
34 #include <sys/metaslab.h>
35 #include <sys/metaslab_impl.h>
36 #include <sys/space_map.h>
37 #include <sys/zio.h>
38 #include <sys/zap.h>
39 #include <sys/fs/zfs.h>
40 #include <sys/arc.h>
41 #include <sys/zil.h>
42 #include <sys/dsl_scan.h>
43 
44 /*
45  * Virtual device management.
46  */
47 
48 static vdev_ops_t *vdev_ops_table[] = {
49 	&vdev_root_ops,
50 	&vdev_raidz_ops,
51 	&vdev_mirror_ops,
52 	&vdev_replacing_ops,
53 	&vdev_spare_ops,
54 	&vdev_disk_ops,
55 	&vdev_file_ops,
56 	&vdev_missing_ops,
57 	&vdev_hole_ops,
58 	NULL
59 };
60 
61 /* maximum scrub/resilver I/O queue per leaf vdev */
62 int zfs_scrub_limit = 10;
63 
64 /*
65  * Given a vdev type, return the appropriate ops vector.
66  */
67 static vdev_ops_t *
68 vdev_getops(const char *type)
69 {
70 	vdev_ops_t *ops, **opspp;
71 
72 	for (opspp = vdev_ops_table; (ops = *opspp) != NULL; opspp++)
73 		if (strcmp(ops->vdev_op_type, type) == 0)
74 			break;
75 
76 	return (ops);
77 }
78 
79 /*
80  * Default asize function: return the MAX of psize with the asize of
81  * all children.  This is what's used by anything other than RAID-Z.
82  */
83 uint64_t
84 vdev_default_asize(vdev_t *vd, uint64_t psize)
85 {
86 	uint64_t asize = P2ROUNDUP(psize, 1ULL << vd->vdev_top->vdev_ashift);
87 	uint64_t csize;
88 
89 	for (int c = 0; c < vd->vdev_children; c++) {
90 		csize = vdev_psize_to_asize(vd->vdev_child[c], psize);
91 		asize = MAX(asize, csize);
92 	}
93 
94 	return (asize);
95 }
96 
97 /*
98  * Get the minimum allocatable size. We define the allocatable size as
99  * the vdev's asize rounded to the nearest metaslab. This allows us to
100  * replace or attach devices which don't have the same physical size but
101  * can still satisfy the same number of allocations.
102  */
103 uint64_t
104 vdev_get_min_asize(vdev_t *vd)
105 {
106 	vdev_t *pvd = vd->vdev_parent;
107 
108 	/*
109 	 * The our parent is NULL (inactive spare or cache) or is the root,
110 	 * just return our own asize.
111 	 */
112 	if (pvd == NULL)
113 		return (vd->vdev_asize);
114 
115 	/*
116 	 * The top-level vdev just returns the allocatable size rounded
117 	 * to the nearest metaslab.
118 	 */
119 	if (vd == vd->vdev_top)
120 		return (P2ALIGN(vd->vdev_asize, 1ULL << vd->vdev_ms_shift));
121 
122 	/*
123 	 * The allocatable space for a raidz vdev is N * sizeof(smallest child),
124 	 * so each child must provide at least 1/Nth of its asize.
125 	 */
126 	if (pvd->vdev_ops == &vdev_raidz_ops)
127 		return (pvd->vdev_min_asize / pvd->vdev_children);
128 
129 	return (pvd->vdev_min_asize);
130 }
131 
132 void
133 vdev_set_min_asize(vdev_t *vd)
134 {
135 	vd->vdev_min_asize = vdev_get_min_asize(vd);
136 
137 	for (int c = 0; c < vd->vdev_children; c++)
138 		vdev_set_min_asize(vd->vdev_child[c]);
139 }
140 
141 vdev_t *
142 vdev_lookup_top(spa_t *spa, uint64_t vdev)
143 {
144 	vdev_t *rvd = spa->spa_root_vdev;
145 
146 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
147 
148 	if (vdev < rvd->vdev_children) {
149 		ASSERT(rvd->vdev_child[vdev] != NULL);
150 		return (rvd->vdev_child[vdev]);
151 	}
152 
153 	return (NULL);
154 }
155 
156 vdev_t *
157 vdev_lookup_by_guid(vdev_t *vd, uint64_t guid)
158 {
159 	vdev_t *mvd;
160 
161 	if (vd->vdev_guid == guid)
162 		return (vd);
163 
164 	for (int c = 0; c < vd->vdev_children; c++)
165 		if ((mvd = vdev_lookup_by_guid(vd->vdev_child[c], guid)) !=
166 		    NULL)
167 			return (mvd);
168 
169 	return (NULL);
170 }
171 
172 void
173 vdev_add_child(vdev_t *pvd, vdev_t *cvd)
174 {
175 	size_t oldsize, newsize;
176 	uint64_t id = cvd->vdev_id;
177 	vdev_t **newchild;
178 
179 	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
180 	ASSERT(cvd->vdev_parent == NULL);
181 
182 	cvd->vdev_parent = pvd;
183 
184 	if (pvd == NULL)
185 		return;
186 
187 	ASSERT(id >= pvd->vdev_children || pvd->vdev_child[id] == NULL);
188 
189 	oldsize = pvd->vdev_children * sizeof (vdev_t *);
190 	pvd->vdev_children = MAX(pvd->vdev_children, id + 1);
191 	newsize = pvd->vdev_children * sizeof (vdev_t *);
192 
193 	newchild = kmem_zalloc(newsize, KM_SLEEP);
194 	if (pvd->vdev_child != NULL) {
195 		bcopy(pvd->vdev_child, newchild, oldsize);
196 		kmem_free(pvd->vdev_child, oldsize);
197 	}
198 
199 	pvd->vdev_child = newchild;
200 	pvd->vdev_child[id] = cvd;
201 
202 	cvd->vdev_top = (pvd->vdev_top ? pvd->vdev_top: cvd);
203 	ASSERT(cvd->vdev_top->vdev_parent->vdev_parent == NULL);
204 
205 	/*
206 	 * Walk up all ancestors to update guid sum.
207 	 */
208 	for (; pvd != NULL; pvd = pvd->vdev_parent)
209 		pvd->vdev_guid_sum += cvd->vdev_guid_sum;
210 }
211 
212 void
213 vdev_remove_child(vdev_t *pvd, vdev_t *cvd)
214 {
215 	int c;
216 	uint_t id = cvd->vdev_id;
217 
218 	ASSERT(cvd->vdev_parent == pvd);
219 
220 	if (pvd == NULL)
221 		return;
222 
223 	ASSERT(id < pvd->vdev_children);
224 	ASSERT(pvd->vdev_child[id] == cvd);
225 
226 	pvd->vdev_child[id] = NULL;
227 	cvd->vdev_parent = NULL;
228 
229 	for (c = 0; c < pvd->vdev_children; c++)
230 		if (pvd->vdev_child[c])
231 			break;
232 
233 	if (c == pvd->vdev_children) {
234 		kmem_free(pvd->vdev_child, c * sizeof (vdev_t *));
235 		pvd->vdev_child = NULL;
236 		pvd->vdev_children = 0;
237 	}
238 
239 	/*
240 	 * Walk up all ancestors to update guid sum.
241 	 */
242 	for (; pvd != NULL; pvd = pvd->vdev_parent)
243 		pvd->vdev_guid_sum -= cvd->vdev_guid_sum;
244 }
245 
246 /*
247  * Remove any holes in the child array.
248  */
249 void
250 vdev_compact_children(vdev_t *pvd)
251 {
252 	vdev_t **newchild, *cvd;
253 	int oldc = pvd->vdev_children;
254 	int newc;
255 
256 	ASSERT(spa_config_held(pvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
257 
258 	for (int c = newc = 0; c < oldc; c++)
259 		if (pvd->vdev_child[c])
260 			newc++;
261 
262 	newchild = kmem_alloc(newc * sizeof (vdev_t *), KM_SLEEP);
263 
264 	for (int c = newc = 0; c < oldc; c++) {
265 		if ((cvd = pvd->vdev_child[c]) != NULL) {
266 			newchild[newc] = cvd;
267 			cvd->vdev_id = newc++;
268 		}
269 	}
270 
271 	kmem_free(pvd->vdev_child, oldc * sizeof (vdev_t *));
272 	pvd->vdev_child = newchild;
273 	pvd->vdev_children = newc;
274 }
275 
276 /*
277  * Allocate and minimally initialize a vdev_t.
278  */
279 vdev_t *
280 vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
281 {
282 	vdev_t *vd;
283 
284 	vd = kmem_zalloc(sizeof (vdev_t), KM_SLEEP);
285 
286 	if (spa->spa_root_vdev == NULL) {
287 		ASSERT(ops == &vdev_root_ops);
288 		spa->spa_root_vdev = vd;
289 	}
290 
291 	if (guid == 0 && ops != &vdev_hole_ops) {
292 		if (spa->spa_root_vdev == vd) {
293 			/*
294 			 * The root vdev's guid will also be the pool guid,
295 			 * which must be unique among all pools.
296 			 */
297 			guid = spa_generate_guid(NULL);
298 		} else {
299 			/*
300 			 * Any other vdev's guid must be unique within the pool.
301 			 */
302 			guid = spa_generate_guid(spa);
303 		}
304 		ASSERT(!spa_guid_exists(spa_guid(spa), guid));
305 	}
306 
307 	vd->vdev_spa = spa;
308 	vd->vdev_id = id;
309 	vd->vdev_guid = guid;
310 	vd->vdev_guid_sum = guid;
311 	vd->vdev_ops = ops;
312 	vd->vdev_state = VDEV_STATE_CLOSED;
313 	vd->vdev_ishole = (ops == &vdev_hole_ops);
314 
315 	mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_DEFAULT, NULL);
316 	mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL);
317 	mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL);
318 	for (int t = 0; t < DTL_TYPES; t++) {
319 		space_map_create(&vd->vdev_dtl[t], 0, -1ULL, 0,
320 		    &vd->vdev_dtl_lock);
321 	}
322 	txg_list_create(&vd->vdev_ms_list,
323 	    offsetof(struct metaslab, ms_txg_node));
324 	txg_list_create(&vd->vdev_dtl_list,
325 	    offsetof(struct vdev, vdev_dtl_node));
326 	vd->vdev_stat.vs_timestamp = gethrtime();
327 	vdev_queue_init(vd);
328 	vdev_cache_init(vd);
329 
330 	return (vd);
331 }
332 
333 /*
334  * Allocate a new vdev.  The 'alloctype' is used to control whether we are
335  * creating a new vdev or loading an existing one - the behavior is slightly
336  * different for each case.
337  */
338 int
339 vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id,
340     int alloctype)
341 {
342 	vdev_ops_t *ops;
343 	char *type;
344 	uint64_t guid = 0, islog, nparity;
345 	vdev_t *vd;
346 
347 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
348 
349 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) != 0)
350 		return (EINVAL);
351 
352 	if ((ops = vdev_getops(type)) == NULL)
353 		return (EINVAL);
354 
355 	/*
356 	 * If this is a load, get the vdev guid from the nvlist.
357 	 * Otherwise, vdev_alloc_common() will generate one for us.
358 	 */
359 	if (alloctype == VDEV_ALLOC_LOAD) {
360 		uint64_t label_id;
361 
362 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ID, &label_id) ||
363 		    label_id != id)
364 			return (EINVAL);
365 
366 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
367 			return (EINVAL);
368 	} else if (alloctype == VDEV_ALLOC_SPARE) {
369 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
370 			return (EINVAL);
371 	} else if (alloctype == VDEV_ALLOC_L2CACHE) {
372 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
373 			return (EINVAL);
374 	} else if (alloctype == VDEV_ALLOC_ROOTPOOL) {
375 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
376 			return (EINVAL);
377 	}
378 
379 	/*
380 	 * The first allocated vdev must be of type 'root'.
381 	 */
382 	if (ops != &vdev_root_ops && spa->spa_root_vdev == NULL)
383 		return (EINVAL);
384 
385 	/*
386 	 * Determine whether we're a log vdev.
387 	 */
388 	islog = 0;
389 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &islog);
390 	if (islog && spa_version(spa) < SPA_VERSION_SLOGS)
391 		return (ENOTSUP);
392 
393 	if (ops == &vdev_hole_ops && spa_version(spa) < SPA_VERSION_HOLES)
394 		return (ENOTSUP);
395 
396 	/*
397 	 * Set the nparity property for RAID-Z vdevs.
398 	 */
399 	nparity = -1ULL;
400 	if (ops == &vdev_raidz_ops) {
401 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY,
402 		    &nparity) == 0) {
403 			if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY)
404 				return (EINVAL);
405 			/*
406 			 * Previous versions could only support 1 or 2 parity
407 			 * device.
408 			 */
409 			if (nparity > 1 &&
410 			    spa_version(spa) < SPA_VERSION_RAIDZ2)
411 				return (ENOTSUP);
412 			if (nparity > 2 &&
413 			    spa_version(spa) < SPA_VERSION_RAIDZ3)
414 				return (ENOTSUP);
415 		} else {
416 			/*
417 			 * We require the parity to be specified for SPAs that
418 			 * support multiple parity levels.
419 			 */
420 			if (spa_version(spa) >= SPA_VERSION_RAIDZ2)
421 				return (EINVAL);
422 			/*
423 			 * Otherwise, we default to 1 parity device for RAID-Z.
424 			 */
425 			nparity = 1;
426 		}
427 	} else {
428 		nparity = 0;
429 	}
430 	ASSERT(nparity != -1ULL);
431 
432 	vd = vdev_alloc_common(spa, id, guid, ops);
433 
434 	vd->vdev_islog = islog;
435 	vd->vdev_nparity = nparity;
436 
437 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &vd->vdev_path) == 0)
438 		vd->vdev_path = spa_strdup(vd->vdev_path);
439 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &vd->vdev_devid) == 0)
440 		vd->vdev_devid = spa_strdup(vd->vdev_devid);
441 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH,
442 	    &vd->vdev_physpath) == 0)
443 		vd->vdev_physpath = spa_strdup(vd->vdev_physpath);
444 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &vd->vdev_fru) == 0)
445 		vd->vdev_fru = spa_strdup(vd->vdev_fru);
446 
447 	/*
448 	 * Set the whole_disk property.  If it's not specified, leave the value
449 	 * as -1.
450 	 */
451 	if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
452 	    &vd->vdev_wholedisk) != 0)
453 		vd->vdev_wholedisk = -1ULL;
454 
455 	/*
456 	 * Look for the 'not present' flag.  This will only be set if the device
457 	 * was not present at the time of import.
458 	 */
459 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT,
460 	    &vd->vdev_not_present);
461 
462 	/*
463 	 * Get the alignment requirement.
464 	 */
465 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT, &vd->vdev_ashift);
466 
467 	/*
468 	 * Retrieve the vdev creation time.
469 	 */
470 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_CREATE_TXG,
471 	    &vd->vdev_crtxg);
472 
473 	/*
474 	 * If we're a top-level vdev, try to load the allocation parameters.
475 	 */
476 	if (parent && !parent->vdev_parent &&
477 	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
478 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
479 		    &vd->vdev_ms_array);
480 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
481 		    &vd->vdev_ms_shift);
482 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE,
483 		    &vd->vdev_asize);
484 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVING,
485 		    &vd->vdev_removing);
486 	}
487 
488 	if (parent && !parent->vdev_parent) {
489 		ASSERT(alloctype == VDEV_ALLOC_LOAD ||
490 		    alloctype == VDEV_ALLOC_ADD ||
491 		    alloctype == VDEV_ALLOC_SPLIT ||
492 		    alloctype == VDEV_ALLOC_ROOTPOOL);
493 		vd->vdev_mg = metaslab_group_create(islog ?
494 		    spa_log_class(spa) : spa_normal_class(spa), vd);
495 	}
496 
497 	/*
498 	 * If we're a leaf vdev, try to load the DTL object and other state.
499 	 */
500 	if (vd->vdev_ops->vdev_op_leaf &&
501 	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE ||
502 	    alloctype == VDEV_ALLOC_ROOTPOOL)) {
503 		if (alloctype == VDEV_ALLOC_LOAD) {
504 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL,
505 			    &vd->vdev_dtl_smo.smo_object);
506 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE,
507 			    &vd->vdev_unspare);
508 		}
509 
510 		if (alloctype == VDEV_ALLOC_ROOTPOOL) {
511 			uint64_t spare = 0;
512 
513 			if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE,
514 			    &spare) == 0 && spare)
515 				spa_spare_add(vd);
516 		}
517 
518 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE,
519 		    &vd->vdev_offline);
520 
521 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_RESILVERING,
522 		    &vd->vdev_resilvering);
523 
524 		/*
525 		 * When importing a pool, we want to ignore the persistent fault
526 		 * state, as the diagnosis made on another system may not be
527 		 * valid in the current context.  Local vdevs will
528 		 * remain in the faulted state.
529 		 */
530 		if (spa_load_state(spa) == SPA_LOAD_OPEN) {
531 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED,
532 			    &vd->vdev_faulted);
533 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED,
534 			    &vd->vdev_degraded);
535 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED,
536 			    &vd->vdev_removed);
537 
538 			if (vd->vdev_faulted || vd->vdev_degraded) {
539 				char *aux;
540 
541 				vd->vdev_label_aux =
542 				    VDEV_AUX_ERR_EXCEEDED;
543 				if (nvlist_lookup_string(nv,
544 				    ZPOOL_CONFIG_AUX_STATE, &aux) == 0 &&
545 				    strcmp(aux, "external") == 0)
546 					vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
547 			}
548 		}
549 	}
550 
551 	/*
552 	 * Add ourselves to the parent's list of children.
553 	 */
554 	vdev_add_child(parent, vd);
555 
556 	*vdp = vd;
557 
558 	return (0);
559 }
560 
561 void
562 vdev_free(vdev_t *vd)
563 {
564 	spa_t *spa = vd->vdev_spa;
565 
566 	/*
567 	 * vdev_free() implies closing the vdev first.  This is simpler than
568 	 * trying to ensure complicated semantics for all callers.
569 	 */
570 	vdev_close(vd);
571 
572 	ASSERT(!list_link_active(&vd->vdev_config_dirty_node));
573 	ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
574 
575 	/*
576 	 * Free all children.
577 	 */
578 	for (int c = 0; c < vd->vdev_children; c++)
579 		vdev_free(vd->vdev_child[c]);
580 
581 	ASSERT(vd->vdev_child == NULL);
582 	ASSERT(vd->vdev_guid_sum == vd->vdev_guid);
583 
584 	/*
585 	 * Discard allocation state.
586 	 */
587 	if (vd->vdev_mg != NULL) {
588 		vdev_metaslab_fini(vd);
589 		metaslab_group_destroy(vd->vdev_mg);
590 	}
591 
592 	ASSERT3U(vd->vdev_stat.vs_space, ==, 0);
593 	ASSERT3U(vd->vdev_stat.vs_dspace, ==, 0);
594 	ASSERT3U(vd->vdev_stat.vs_alloc, ==, 0);
595 
596 	/*
597 	 * Remove this vdev from its parent's child list.
598 	 */
599 	vdev_remove_child(vd->vdev_parent, vd);
600 
601 	ASSERT(vd->vdev_parent == NULL);
602 
603 	/*
604 	 * Clean up vdev structure.
605 	 */
606 	vdev_queue_fini(vd);
607 	vdev_cache_fini(vd);
608 
609 	if (vd->vdev_path)
610 		spa_strfree(vd->vdev_path);
611 	if (vd->vdev_devid)
612 		spa_strfree(vd->vdev_devid);
613 	if (vd->vdev_physpath)
614 		spa_strfree(vd->vdev_physpath);
615 	if (vd->vdev_fru)
616 		spa_strfree(vd->vdev_fru);
617 
618 	if (vd->vdev_isspare)
619 		spa_spare_remove(vd);
620 	if (vd->vdev_isl2cache)
621 		spa_l2cache_remove(vd);
622 
623 	txg_list_destroy(&vd->vdev_ms_list);
624 	txg_list_destroy(&vd->vdev_dtl_list);
625 
626 	mutex_enter(&vd->vdev_dtl_lock);
627 	for (int t = 0; t < DTL_TYPES; t++) {
628 		space_map_unload(&vd->vdev_dtl[t]);
629 		space_map_destroy(&vd->vdev_dtl[t]);
630 	}
631 	mutex_exit(&vd->vdev_dtl_lock);
632 
633 	mutex_destroy(&vd->vdev_dtl_lock);
634 	mutex_destroy(&vd->vdev_stat_lock);
635 	mutex_destroy(&vd->vdev_probe_lock);
636 
637 	if (vd == spa->spa_root_vdev)
638 		spa->spa_root_vdev = NULL;
639 
640 	kmem_free(vd, sizeof (vdev_t));
641 }
642 
643 /*
644  * Transfer top-level vdev state from svd to tvd.
645  */
646 static void
647 vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
648 {
649 	spa_t *spa = svd->vdev_spa;
650 	metaslab_t *msp;
651 	vdev_t *vd;
652 	int t;
653 
654 	ASSERT(tvd == tvd->vdev_top);
655 
656 	tvd->vdev_ms_array = svd->vdev_ms_array;
657 	tvd->vdev_ms_shift = svd->vdev_ms_shift;
658 	tvd->vdev_ms_count = svd->vdev_ms_count;
659 
660 	svd->vdev_ms_array = 0;
661 	svd->vdev_ms_shift = 0;
662 	svd->vdev_ms_count = 0;
663 
664 	tvd->vdev_mg = svd->vdev_mg;
665 	tvd->vdev_ms = svd->vdev_ms;
666 
667 	svd->vdev_mg = NULL;
668 	svd->vdev_ms = NULL;
669 
670 	if (tvd->vdev_mg != NULL)
671 		tvd->vdev_mg->mg_vd = tvd;
672 
673 	tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
674 	tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
675 	tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
676 
677 	svd->vdev_stat.vs_alloc = 0;
678 	svd->vdev_stat.vs_space = 0;
679 	svd->vdev_stat.vs_dspace = 0;
680 
681 	for (t = 0; t < TXG_SIZE; t++) {
682 		while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
683 			(void) txg_list_add(&tvd->vdev_ms_list, msp, t);
684 		while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
685 			(void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
686 		if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
687 			(void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
688 	}
689 
690 	if (list_link_active(&svd->vdev_config_dirty_node)) {
691 		vdev_config_clean(svd);
692 		vdev_config_dirty(tvd);
693 	}
694 
695 	if (list_link_active(&svd->vdev_state_dirty_node)) {
696 		vdev_state_clean(svd);
697 		vdev_state_dirty(tvd);
698 	}
699 
700 	tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
701 	svd->vdev_deflate_ratio = 0;
702 
703 	tvd->vdev_islog = svd->vdev_islog;
704 	svd->vdev_islog = 0;
705 }
706 
707 static void
708 vdev_top_update(vdev_t *tvd, vdev_t *vd)
709 {
710 	if (vd == NULL)
711 		return;
712 
713 	vd->vdev_top = tvd;
714 
715 	for (int c = 0; c < vd->vdev_children; c++)
716 		vdev_top_update(tvd, vd->vdev_child[c]);
717 }
718 
719 /*
720  * Add a mirror/replacing vdev above an existing vdev.
721  */
722 vdev_t *
723 vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
724 {
725 	spa_t *spa = cvd->vdev_spa;
726 	vdev_t *pvd = cvd->vdev_parent;
727 	vdev_t *mvd;
728 
729 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
730 
731 	mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
732 
733 	mvd->vdev_asize = cvd->vdev_asize;
734 	mvd->vdev_min_asize = cvd->vdev_min_asize;
735 	mvd->vdev_ashift = cvd->vdev_ashift;
736 	mvd->vdev_state = cvd->vdev_state;
737 	mvd->vdev_crtxg = cvd->vdev_crtxg;
738 
739 	vdev_remove_child(pvd, cvd);
740 	vdev_add_child(pvd, mvd);
741 	cvd->vdev_id = mvd->vdev_children;
742 	vdev_add_child(mvd, cvd);
743 	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
744 
745 	if (mvd == mvd->vdev_top)
746 		vdev_top_transfer(cvd, mvd);
747 
748 	return (mvd);
749 }
750 
751 /*
752  * Remove a 1-way mirror/replacing vdev from the tree.
753  */
754 void
755 vdev_remove_parent(vdev_t *cvd)
756 {
757 	vdev_t *mvd = cvd->vdev_parent;
758 	vdev_t *pvd = mvd->vdev_parent;
759 
760 	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
761 
762 	ASSERT(mvd->vdev_children == 1);
763 	ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
764 	    mvd->vdev_ops == &vdev_replacing_ops ||
765 	    mvd->vdev_ops == &vdev_spare_ops);
766 	cvd->vdev_ashift = mvd->vdev_ashift;
767 
768 	vdev_remove_child(mvd, cvd);
769 	vdev_remove_child(pvd, mvd);
770 
771 	/*
772 	 * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
773 	 * Otherwise, we could have detached an offline device, and when we
774 	 * go to import the pool we'll think we have two top-level vdevs,
775 	 * instead of a different version of the same top-level vdev.
776 	 */
777 	if (mvd->vdev_top == mvd) {
778 		uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
779 		cvd->vdev_orig_guid = cvd->vdev_guid;
780 		cvd->vdev_guid += guid_delta;
781 		cvd->vdev_guid_sum += guid_delta;
782 	}
783 	cvd->vdev_id = mvd->vdev_id;
784 	vdev_add_child(pvd, cvd);
785 	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
786 
787 	if (cvd == cvd->vdev_top)
788 		vdev_top_transfer(mvd, cvd);
789 
790 	ASSERT(mvd->vdev_children == 0);
791 	vdev_free(mvd);
792 }
793 
794 int
795 vdev_metaslab_init(vdev_t *vd, uint64_t txg)
796 {
797 	spa_t *spa = vd->vdev_spa;
798 	objset_t *mos = spa->spa_meta_objset;
799 	uint64_t m;
800 	uint64_t oldc = vd->vdev_ms_count;
801 	uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
802 	metaslab_t **mspp;
803 	int error;
804 
805 	ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER));
806 
807 	/*
808 	 * This vdev is not being allocated from yet or is a hole.
809 	 */
810 	if (vd->vdev_ms_shift == 0)
811 		return (0);
812 
813 	ASSERT(!vd->vdev_ishole);
814 
815 	/*
816 	 * Compute the raidz-deflation ratio.  Note, we hard-code
817 	 * in 128k (1 << 17) because it is the current "typical" blocksize.
818 	 * Even if SPA_MAXBLOCKSIZE changes, this algorithm must never change,
819 	 * or we will inconsistently account for existing bp's.
820 	 */
821 	vd->vdev_deflate_ratio = (1 << 17) /
822 	    (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
823 
824 	ASSERT(oldc <= newc);
825 
826 	mspp = kmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
827 
828 	if (oldc != 0) {
829 		bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp));
830 		kmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
831 	}
832 
833 	vd->vdev_ms = mspp;
834 	vd->vdev_ms_count = newc;
835 
836 	for (m = oldc; m < newc; m++) {
837 		space_map_obj_t smo = { 0, 0, 0 };
838 		if (txg == 0) {
839 			uint64_t object = 0;
840 			error = dmu_read(mos, vd->vdev_ms_array,
841 			    m * sizeof (uint64_t), sizeof (uint64_t), &object,
842 			    DMU_READ_PREFETCH);
843 			if (error)
844 				return (error);
845 			if (object != 0) {
846 				dmu_buf_t *db;
847 				error = dmu_bonus_hold(mos, object, FTAG, &db);
848 				if (error)
849 					return (error);
850 				ASSERT3U(db->db_size, >=, sizeof (smo));
851 				bcopy(db->db_data, &smo, sizeof (smo));
852 				ASSERT3U(smo.smo_object, ==, object);
853 				dmu_buf_rele(db, FTAG);
854 			}
855 		}
856 		vd->vdev_ms[m] = metaslab_init(vd->vdev_mg, &smo,
857 		    m << vd->vdev_ms_shift, 1ULL << vd->vdev_ms_shift, txg);
858 	}
859 
860 	if (txg == 0)
861 		spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER);
862 
863 	/*
864 	 * If the vdev is being removed we don't activate
865 	 * the metaslabs since we want to ensure that no new
866 	 * allocations are performed on this device.
867 	 */
868 	if (oldc == 0 && !vd->vdev_removing)
869 		metaslab_group_activate(vd->vdev_mg);
870 
871 	if (txg == 0)
872 		spa_config_exit(spa, SCL_ALLOC, FTAG);
873 
874 	return (0);
875 }
876 
877 void
878 vdev_metaslab_fini(vdev_t *vd)
879 {
880 	uint64_t m;
881 	uint64_t count = vd->vdev_ms_count;
882 
883 	if (vd->vdev_ms != NULL) {
884 		metaslab_group_passivate(vd->vdev_mg);
885 		for (m = 0; m < count; m++)
886 			if (vd->vdev_ms[m] != NULL)
887 				metaslab_fini(vd->vdev_ms[m]);
888 		kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
889 		vd->vdev_ms = NULL;
890 	}
891 }
892 
893 typedef struct vdev_probe_stats {
894 	boolean_t	vps_readable;
895 	boolean_t	vps_writeable;
896 	int		vps_flags;
897 } vdev_probe_stats_t;
898 
899 static void
900 vdev_probe_done(zio_t *zio)
901 {
902 	spa_t *spa = zio->io_spa;
903 	vdev_t *vd = zio->io_vd;
904 	vdev_probe_stats_t *vps = zio->io_private;
905 
906 	ASSERT(vd->vdev_probe_zio != NULL);
907 
908 	if (zio->io_type == ZIO_TYPE_READ) {
909 		if (zio->io_error == 0)
910 			vps->vps_readable = 1;
911 		if (zio->io_error == 0 && spa_writeable(spa)) {
912 			zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
913 			    zio->io_offset, zio->io_size, zio->io_data,
914 			    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
915 			    ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
916 		} else {
917 			zio_buf_free(zio->io_data, zio->io_size);
918 		}
919 	} else if (zio->io_type == ZIO_TYPE_WRITE) {
920 		if (zio->io_error == 0)
921 			vps->vps_writeable = 1;
922 		zio_buf_free(zio->io_data, zio->io_size);
923 	} else if (zio->io_type == ZIO_TYPE_NULL) {
924 		zio_t *pio;
925 
926 		vd->vdev_cant_read |= !vps->vps_readable;
927 		vd->vdev_cant_write |= !vps->vps_writeable;
928 
929 		if (vdev_readable(vd) &&
930 		    (vdev_writeable(vd) || !spa_writeable(spa))) {
931 			zio->io_error = 0;
932 		} else {
933 			ASSERT(zio->io_error != 0);
934 			zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
935 			    spa, vd, NULL, 0, 0);
936 			zio->io_error = ENXIO;
937 		}
938 
939 		mutex_enter(&vd->vdev_probe_lock);
940 		ASSERT(vd->vdev_probe_zio == zio);
941 		vd->vdev_probe_zio = NULL;
942 		mutex_exit(&vd->vdev_probe_lock);
943 
944 		while ((pio = zio_walk_parents(zio)) != NULL)
945 			if (!vdev_accessible(vd, pio))
946 				pio->io_error = ENXIO;
947 
948 		kmem_free(vps, sizeof (*vps));
949 	}
950 }
951 
952 /*
953  * Determine whether this device is accessible by reading and writing
954  * to several known locations: the pad regions of each vdev label
955  * but the first (which we leave alone in case it contains a VTOC).
956  */
957 zio_t *
958 vdev_probe(vdev_t *vd, zio_t *zio)
959 {
960 	spa_t *spa = vd->vdev_spa;
961 	vdev_probe_stats_t *vps = NULL;
962 	zio_t *pio;
963 
964 	ASSERT(vd->vdev_ops->vdev_op_leaf);
965 
966 	/*
967 	 * Don't probe the probe.
968 	 */
969 	if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
970 		return (NULL);
971 
972 	/*
973 	 * To prevent 'probe storms' when a device fails, we create
974 	 * just one probe i/o at a time.  All zios that want to probe
975 	 * this vdev will become parents of the probe io.
976 	 */
977 	mutex_enter(&vd->vdev_probe_lock);
978 
979 	if ((pio = vd->vdev_probe_zio) == NULL) {
980 		vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
981 
982 		vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
983 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
984 		    ZIO_FLAG_TRYHARD;
985 
986 		if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
987 			/*
988 			 * vdev_cant_read and vdev_cant_write can only
989 			 * transition from TRUE to FALSE when we have the
990 			 * SCL_ZIO lock as writer; otherwise they can only
991 			 * transition from FALSE to TRUE.  This ensures that
992 			 * any zio looking at these values can assume that
993 			 * failures persist for the life of the I/O.  That's
994 			 * important because when a device has intermittent
995 			 * connectivity problems, we want to ensure that
996 			 * they're ascribed to the device (ENXIO) and not
997 			 * the zio (EIO).
998 			 *
999 			 * Since we hold SCL_ZIO as writer here, clear both
1000 			 * values so the probe can reevaluate from first
1001 			 * principles.
1002 			 */
1003 			vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
1004 			vd->vdev_cant_read = B_FALSE;
1005 			vd->vdev_cant_write = B_FALSE;
1006 		}
1007 
1008 		vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
1009 		    vdev_probe_done, vps,
1010 		    vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
1011 
1012 		/*
1013 		 * We can't change the vdev state in this context, so we
1014 		 * kick off an async task to do it on our behalf.
1015 		 */
1016 		if (zio != NULL) {
1017 			vd->vdev_probe_wanted = B_TRUE;
1018 			spa_async_request(spa, SPA_ASYNC_PROBE);
1019 		}
1020 	}
1021 
1022 	if (zio != NULL)
1023 		zio_add_child(zio, pio);
1024 
1025 	mutex_exit(&vd->vdev_probe_lock);
1026 
1027 	if (vps == NULL) {
1028 		ASSERT(zio != NULL);
1029 		return (NULL);
1030 	}
1031 
1032 	for (int l = 1; l < VDEV_LABELS; l++) {
1033 		zio_nowait(zio_read_phys(pio, vd,
1034 		    vdev_label_offset(vd->vdev_psize, l,
1035 		    offsetof(vdev_label_t, vl_pad2)),
1036 		    VDEV_PAD_SIZE, zio_buf_alloc(VDEV_PAD_SIZE),
1037 		    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1038 		    ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1039 	}
1040 
1041 	if (zio == NULL)
1042 		return (pio);
1043 
1044 	zio_nowait(pio);
1045 	return (NULL);
1046 }
1047 
1048 static void
1049 vdev_open_child(void *arg)
1050 {
1051 	vdev_t *vd = arg;
1052 
1053 	vd->vdev_open_thread = curthread;
1054 	vd->vdev_open_error = vdev_open(vd);
1055 	vd->vdev_open_thread = NULL;
1056 }
1057 
1058 boolean_t
1059 vdev_uses_zvols(vdev_t *vd)
1060 {
1061 	if (vd->vdev_path && strncmp(vd->vdev_path, ZVOL_DIR,
1062 	    strlen(ZVOL_DIR)) == 0)
1063 		return (B_TRUE);
1064 	for (int c = 0; c < vd->vdev_children; c++)
1065 		if (vdev_uses_zvols(vd->vdev_child[c]))
1066 			return (B_TRUE);
1067 	return (B_FALSE);
1068 }
1069 
1070 void
1071 vdev_open_children(vdev_t *vd)
1072 {
1073 	taskq_t *tq;
1074 	int children = vd->vdev_children;
1075 
1076 	/*
1077 	 * in order to handle pools on top of zvols, do the opens
1078 	 * in a single thread so that the same thread holds the
1079 	 * spa_namespace_lock
1080 	 */
1081 	if (vdev_uses_zvols(vd)) {
1082 		for (int c = 0; c < children; c++)
1083 			vd->vdev_child[c]->vdev_open_error =
1084 			    vdev_open(vd->vdev_child[c]);
1085 		return;
1086 	}
1087 	tq = taskq_create("vdev_open", children, minclsyspri,
1088 	    children, children, TASKQ_PREPOPULATE);
1089 
1090 	for (int c = 0; c < children; c++)
1091 		VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c],
1092 		    TQ_SLEEP) != NULL);
1093 
1094 	taskq_destroy(tq);
1095 }
1096 
1097 /*
1098  * Prepare a virtual device for access.
1099  */
1100 int
1101 vdev_open(vdev_t *vd)
1102 {
1103 	spa_t *spa = vd->vdev_spa;
1104 	int error;
1105 	uint64_t osize = 0;
1106 	uint64_t asize, psize;
1107 	uint64_t ashift = 0;
1108 
1109 	ASSERT(vd->vdev_open_thread == curthread ||
1110 	    spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1111 	ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1112 	    vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1113 	    vd->vdev_state == VDEV_STATE_OFFLINE);
1114 
1115 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1116 	vd->vdev_cant_read = B_FALSE;
1117 	vd->vdev_cant_write = B_FALSE;
1118 	vd->vdev_min_asize = vdev_get_min_asize(vd);
1119 
1120 	/*
1121 	 * If this vdev is not removed, check its fault status.  If it's
1122 	 * faulted, bail out of the open.
1123 	 */
1124 	if (!vd->vdev_removed && vd->vdev_faulted) {
1125 		ASSERT(vd->vdev_children == 0);
1126 		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1127 		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1128 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1129 		    vd->vdev_label_aux);
1130 		return (ENXIO);
1131 	} else if (vd->vdev_offline) {
1132 		ASSERT(vd->vdev_children == 0);
1133 		vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1134 		return (ENXIO);
1135 	}
1136 
1137 	error = vd->vdev_ops->vdev_op_open(vd, &osize, &ashift);
1138 
1139 	/*
1140 	 * Reset the vdev_reopening flag so that we actually close
1141 	 * the vdev on error.
1142 	 */
1143 	vd->vdev_reopening = B_FALSE;
1144 	if (zio_injection_enabled && error == 0)
1145 		error = zio_handle_device_injection(vd, NULL, ENXIO);
1146 
1147 	if (error) {
1148 		if (vd->vdev_removed &&
1149 		    vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1150 			vd->vdev_removed = B_FALSE;
1151 
1152 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1153 		    vd->vdev_stat.vs_aux);
1154 		return (error);
1155 	}
1156 
1157 	vd->vdev_removed = B_FALSE;
1158 
1159 	/*
1160 	 * Recheck the faulted flag now that we have confirmed that
1161 	 * the vdev is accessible.  If we're faulted, bail.
1162 	 */
1163 	if (vd->vdev_faulted) {
1164 		ASSERT(vd->vdev_children == 0);
1165 		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1166 		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1167 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1168 		    vd->vdev_label_aux);
1169 		return (ENXIO);
1170 	}
1171 
1172 	if (vd->vdev_degraded) {
1173 		ASSERT(vd->vdev_children == 0);
1174 		vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1175 		    VDEV_AUX_ERR_EXCEEDED);
1176 	} else {
1177 		vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1178 	}
1179 
1180 	/*
1181 	 * For hole or missing vdevs we just return success.
1182 	 */
1183 	if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1184 		return (0);
1185 
1186 	for (int c = 0; c < vd->vdev_children; c++) {
1187 		if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1188 			vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1189 			    VDEV_AUX_NONE);
1190 			break;
1191 		}
1192 	}
1193 
1194 	osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1195 
1196 	if (vd->vdev_children == 0) {
1197 		if (osize < SPA_MINDEVSIZE) {
1198 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1199 			    VDEV_AUX_TOO_SMALL);
1200 			return (EOVERFLOW);
1201 		}
1202 		psize = osize;
1203 		asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1204 	} else {
1205 		if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1206 		    (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1207 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1208 			    VDEV_AUX_TOO_SMALL);
1209 			return (EOVERFLOW);
1210 		}
1211 		psize = 0;
1212 		asize = osize;
1213 	}
1214 
1215 	vd->vdev_psize = psize;
1216 
1217 	/*
1218 	 * Make sure the allocatable size hasn't shrunk.
1219 	 */
1220 	if (asize < vd->vdev_min_asize) {
1221 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1222 		    VDEV_AUX_BAD_LABEL);
1223 		return (EINVAL);
1224 	}
1225 
1226 	if (vd->vdev_asize == 0) {
1227 		/*
1228 		 * This is the first-ever open, so use the computed values.
1229 		 * For testing purposes, a higher ashift can be requested.
1230 		 */
1231 		vd->vdev_asize = asize;
1232 		vd->vdev_ashift = MAX(ashift, vd->vdev_ashift);
1233 	} else {
1234 		/*
1235 		 * Make sure the alignment requirement hasn't increased.
1236 		 */
1237 		if (ashift > vd->vdev_top->vdev_ashift) {
1238 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1239 			    VDEV_AUX_BAD_LABEL);
1240 			return (EINVAL);
1241 		}
1242 	}
1243 
1244 	/*
1245 	 * If all children are healthy and the asize has increased,
1246 	 * then we've experienced dynamic LUN growth.  If automatic
1247 	 * expansion is enabled then use the additional space.
1248 	 */
1249 	if (vd->vdev_state == VDEV_STATE_HEALTHY && asize > vd->vdev_asize &&
1250 	    (vd->vdev_expanding || spa->spa_autoexpand))
1251 		vd->vdev_asize = asize;
1252 
1253 	vdev_set_min_asize(vd);
1254 
1255 	/*
1256 	 * Ensure we can issue some IO before declaring the
1257 	 * vdev open for business.
1258 	 */
1259 	if (vd->vdev_ops->vdev_op_leaf &&
1260 	    (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1261 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1262 		    VDEV_AUX_ERR_EXCEEDED);
1263 		return (error);
1264 	}
1265 
1266 	/*
1267 	 * If a leaf vdev has a DTL, and seems healthy, then kick off a
1268 	 * resilver.  But don't do this if we are doing a reopen for a scrub,
1269 	 * since this would just restart the scrub we are already doing.
1270 	 */
1271 	if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1272 	    vdev_resilver_needed(vd, NULL, NULL))
1273 		spa_async_request(spa, SPA_ASYNC_RESILVER);
1274 
1275 	return (0);
1276 }
1277 
1278 /*
1279  * Called once the vdevs are all opened, this routine validates the label
1280  * contents.  This needs to be done before vdev_load() so that we don't
1281  * inadvertently do repair I/Os to the wrong device.
1282  *
1283  * This function will only return failure if one of the vdevs indicates that it
1284  * has since been destroyed or exported.  This is only possible if
1285  * /etc/zfs/zpool.cache was readonly at the time.  Otherwise, the vdev state
1286  * will be updated but the function will return 0.
1287  */
1288 int
1289 vdev_validate(vdev_t *vd)
1290 {
1291 	spa_t *spa = vd->vdev_spa;
1292 	nvlist_t *label;
1293 	uint64_t guid = 0, top_guid;
1294 	uint64_t state;
1295 
1296 	for (int c = 0; c < vd->vdev_children; c++)
1297 		if (vdev_validate(vd->vdev_child[c]) != 0)
1298 			return (EBADF);
1299 
1300 	/*
1301 	 * If the device has already failed, or was marked offline, don't do
1302 	 * any further validation.  Otherwise, label I/O will fail and we will
1303 	 * overwrite the previous state.
1304 	 */
1305 	if (vd->vdev_ops->vdev_op_leaf && vdev_readable(vd)) {
1306 		uint64_t aux_guid = 0;
1307 		nvlist_t *nvl;
1308 
1309 		if ((label = vdev_label_read_config(vd)) == NULL) {
1310 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1311 			    VDEV_AUX_BAD_LABEL);
1312 			return (0);
1313 		}
1314 
1315 		/*
1316 		 * Determine if this vdev has been split off into another
1317 		 * pool.  If so, then refuse to open it.
1318 		 */
1319 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID,
1320 		    &aux_guid) == 0 && aux_guid == spa_guid(spa)) {
1321 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1322 			    VDEV_AUX_SPLIT_POOL);
1323 			nvlist_free(label);
1324 			return (0);
1325 		}
1326 
1327 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_GUID,
1328 		    &guid) != 0 || guid != spa_guid(spa)) {
1329 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1330 			    VDEV_AUX_CORRUPT_DATA);
1331 			nvlist_free(label);
1332 			return (0);
1333 		}
1334 
1335 		if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl)
1336 		    != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID,
1337 		    &aux_guid) != 0)
1338 			aux_guid = 0;
1339 
1340 		/*
1341 		 * If this vdev just became a top-level vdev because its
1342 		 * sibling was detached, it will have adopted the parent's
1343 		 * vdev guid -- but the label may or may not be on disk yet.
1344 		 * Fortunately, either version of the label will have the
1345 		 * same top guid, so if we're a top-level vdev, we can
1346 		 * safely compare to that instead.
1347 		 *
1348 		 * If we split this vdev off instead, then we also check the
1349 		 * original pool's guid.  We don't want to consider the vdev
1350 		 * corrupt if it is partway through a split operation.
1351 		 */
1352 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID,
1353 		    &guid) != 0 ||
1354 		    nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID,
1355 		    &top_guid) != 0 ||
1356 		    ((vd->vdev_guid != guid && vd->vdev_guid != aux_guid) &&
1357 		    (vd->vdev_guid != top_guid || vd != vd->vdev_top))) {
1358 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1359 			    VDEV_AUX_CORRUPT_DATA);
1360 			nvlist_free(label);
1361 			return (0);
1362 		}
1363 
1364 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1365 		    &state) != 0) {
1366 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1367 			    VDEV_AUX_CORRUPT_DATA);
1368 			nvlist_free(label);
1369 			return (0);
1370 		}
1371 
1372 		nvlist_free(label);
1373 
1374 		/*
1375 		 * If this is a verbatim import, no need to check the
1376 		 * state of the pool.
1377 		 */
1378 		if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) &&
1379 		    spa_load_state(spa) == SPA_LOAD_OPEN &&
1380 		    state != POOL_STATE_ACTIVE)
1381 			return (EBADF);
1382 
1383 		/*
1384 		 * If we were able to open and validate a vdev that was
1385 		 * previously marked permanently unavailable, clear that state
1386 		 * now.
1387 		 */
1388 		if (vd->vdev_not_present)
1389 			vd->vdev_not_present = 0;
1390 	}
1391 
1392 	return (0);
1393 }
1394 
1395 /*
1396  * Close a virtual device.
1397  */
1398 void
1399 vdev_close(vdev_t *vd)
1400 {
1401 	spa_t *spa = vd->vdev_spa;
1402 	vdev_t *pvd = vd->vdev_parent;
1403 
1404 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1405 
1406 	/*
1407 	 * If our parent is reopening, then we are as well, unless we are
1408 	 * going offline.
1409 	 */
1410 	if (pvd != NULL && pvd->vdev_reopening)
1411 		vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline);
1412 
1413 	vd->vdev_ops->vdev_op_close(vd);
1414 
1415 	vdev_cache_purge(vd);
1416 
1417 	/*
1418 	 * We record the previous state before we close it, so that if we are
1419 	 * doing a reopen(), we don't generate FMA ereports if we notice that
1420 	 * it's still faulted.
1421 	 */
1422 	vd->vdev_prevstate = vd->vdev_state;
1423 
1424 	if (vd->vdev_offline)
1425 		vd->vdev_state = VDEV_STATE_OFFLINE;
1426 	else
1427 		vd->vdev_state = VDEV_STATE_CLOSED;
1428 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1429 }
1430 
1431 void
1432 vdev_hold(vdev_t *vd)
1433 {
1434 	spa_t *spa = vd->vdev_spa;
1435 
1436 	ASSERT(spa_is_root(spa));
1437 	if (spa->spa_state == POOL_STATE_UNINITIALIZED)
1438 		return;
1439 
1440 	for (int c = 0; c < vd->vdev_children; c++)
1441 		vdev_hold(vd->vdev_child[c]);
1442 
1443 	if (vd->vdev_ops->vdev_op_leaf)
1444 		vd->vdev_ops->vdev_op_hold(vd);
1445 }
1446 
1447 void
1448 vdev_rele(vdev_t *vd)
1449 {
1450 	spa_t *spa = vd->vdev_spa;
1451 
1452 	ASSERT(spa_is_root(spa));
1453 	for (int c = 0; c < vd->vdev_children; c++)
1454 		vdev_rele(vd->vdev_child[c]);
1455 
1456 	if (vd->vdev_ops->vdev_op_leaf)
1457 		vd->vdev_ops->vdev_op_rele(vd);
1458 }
1459 
1460 /*
1461  * Reopen all interior vdevs and any unopened leaves.  We don't actually
1462  * reopen leaf vdevs which had previously been opened as they might deadlock
1463  * on the spa_config_lock.  Instead we only obtain the leaf's physical size.
1464  * If the leaf has never been opened then open it, as usual.
1465  */
1466 void
1467 vdev_reopen(vdev_t *vd)
1468 {
1469 	spa_t *spa = vd->vdev_spa;
1470 
1471 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1472 
1473 	/* set the reopening flag unless we're taking the vdev offline */
1474 	vd->vdev_reopening = !vd->vdev_offline;
1475 	vdev_close(vd);
1476 	(void) vdev_open(vd);
1477 
1478 	/*
1479 	 * Call vdev_validate() here to make sure we have the same device.
1480 	 * Otherwise, a device with an invalid label could be successfully
1481 	 * opened in response to vdev_reopen().
1482 	 */
1483 	if (vd->vdev_aux) {
1484 		(void) vdev_validate_aux(vd);
1485 		if (vdev_readable(vd) && vdev_writeable(vd) &&
1486 		    vd->vdev_aux == &spa->spa_l2cache &&
1487 		    !l2arc_vdev_present(vd))
1488 			l2arc_add_vdev(spa, vd);
1489 	} else {
1490 		(void) vdev_validate(vd);
1491 	}
1492 
1493 	/*
1494 	 * Reassess parent vdev's health.
1495 	 */
1496 	vdev_propagate_state(vd);
1497 }
1498 
1499 int
1500 vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1501 {
1502 	int error;
1503 
1504 	/*
1505 	 * Normally, partial opens (e.g. of a mirror) are allowed.
1506 	 * For a create, however, we want to fail the request if
1507 	 * there are any components we can't open.
1508 	 */
1509 	error = vdev_open(vd);
1510 
1511 	if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
1512 		vdev_close(vd);
1513 		return (error ? error : ENXIO);
1514 	}
1515 
1516 	/*
1517 	 * Recursively initialize all labels.
1518 	 */
1519 	if ((error = vdev_label_init(vd, txg, isreplacing ?
1520 	    VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
1521 		vdev_close(vd);
1522 		return (error);
1523 	}
1524 
1525 	return (0);
1526 }
1527 
1528 void
1529 vdev_metaslab_set_size(vdev_t *vd)
1530 {
1531 	/*
1532 	 * Aim for roughly 200 metaslabs per vdev.
1533 	 */
1534 	vd->vdev_ms_shift = highbit(vd->vdev_asize / 200);
1535 	vd->vdev_ms_shift = MAX(vd->vdev_ms_shift, SPA_MAXBLOCKSHIFT);
1536 }
1537 
1538 void
1539 vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
1540 {
1541 	ASSERT(vd == vd->vdev_top);
1542 	ASSERT(!vd->vdev_ishole);
1543 	ASSERT(ISP2(flags));
1544 	ASSERT(spa_writeable(vd->vdev_spa));
1545 
1546 	if (flags & VDD_METASLAB)
1547 		(void) txg_list_add(&vd->vdev_ms_list, arg, txg);
1548 
1549 	if (flags & VDD_DTL)
1550 		(void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
1551 
1552 	(void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
1553 }
1554 
1555 /*
1556  * DTLs.
1557  *
1558  * A vdev's DTL (dirty time log) is the set of transaction groups for which
1559  * the vdev has less than perfect replication.  There are four kinds of DTL:
1560  *
1561  * DTL_MISSING: txgs for which the vdev has no valid copies of the data
1562  *
1563  * DTL_PARTIAL: txgs for which data is available, but not fully replicated
1564  *
1565  * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
1566  *	scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
1567  *	txgs that was scrubbed.
1568  *
1569  * DTL_OUTAGE: txgs which cannot currently be read, whether due to
1570  *	persistent errors or just some device being offline.
1571  *	Unlike the other three, the DTL_OUTAGE map is not generally
1572  *	maintained; it's only computed when needed, typically to
1573  *	determine whether a device can be detached.
1574  *
1575  * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
1576  * either has the data or it doesn't.
1577  *
1578  * For interior vdevs such as mirror and RAID-Z the picture is more complex.
1579  * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
1580  * if any child is less than fully replicated, then so is its parent.
1581  * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
1582  * comprising only those txgs which appear in 'maxfaults' or more children;
1583  * those are the txgs we don't have enough replication to read.  For example,
1584  * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
1585  * thus, its DTL_MISSING consists of the set of txgs that appear in more than
1586  * two child DTL_MISSING maps.
1587  *
1588  * It should be clear from the above that to compute the DTLs and outage maps
1589  * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
1590  * Therefore, that is all we keep on disk.  When loading the pool, or after
1591  * a configuration change, we generate all other DTLs from first principles.
1592  */
1593 void
1594 vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1595 {
1596 	space_map_t *sm = &vd->vdev_dtl[t];
1597 
1598 	ASSERT(t < DTL_TYPES);
1599 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1600 	ASSERT(spa_writeable(vd->vdev_spa));
1601 
1602 	mutex_enter(sm->sm_lock);
1603 	if (!space_map_contains(sm, txg, size))
1604 		space_map_add(sm, txg, size);
1605 	mutex_exit(sm->sm_lock);
1606 }
1607 
1608 boolean_t
1609 vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1610 {
1611 	space_map_t *sm = &vd->vdev_dtl[t];
1612 	boolean_t dirty = B_FALSE;
1613 
1614 	ASSERT(t < DTL_TYPES);
1615 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1616 
1617 	mutex_enter(sm->sm_lock);
1618 	if (sm->sm_space != 0)
1619 		dirty = space_map_contains(sm, txg, size);
1620 	mutex_exit(sm->sm_lock);
1621 
1622 	return (dirty);
1623 }
1624 
1625 boolean_t
1626 vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
1627 {
1628 	space_map_t *sm = &vd->vdev_dtl[t];
1629 	boolean_t empty;
1630 
1631 	mutex_enter(sm->sm_lock);
1632 	empty = (sm->sm_space == 0);
1633 	mutex_exit(sm->sm_lock);
1634 
1635 	return (empty);
1636 }
1637 
1638 /*
1639  * Reassess DTLs after a config change or scrub completion.
1640  */
1641 void
1642 vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
1643 {
1644 	spa_t *spa = vd->vdev_spa;
1645 	avl_tree_t reftree;
1646 	int minref;
1647 
1648 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
1649 
1650 	for (int c = 0; c < vd->vdev_children; c++)
1651 		vdev_dtl_reassess(vd->vdev_child[c], txg,
1652 		    scrub_txg, scrub_done);
1653 
1654 	if (vd == spa->spa_root_vdev || vd->vdev_ishole || vd->vdev_aux)
1655 		return;
1656 
1657 	if (vd->vdev_ops->vdev_op_leaf) {
1658 		dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1659 
1660 		mutex_enter(&vd->vdev_dtl_lock);
1661 		if (scrub_txg != 0 &&
1662 		    (spa->spa_scrub_started ||
1663 		    (scn && scn->scn_phys.scn_errors == 0))) {
1664 			/*
1665 			 * We completed a scrub up to scrub_txg.  If we
1666 			 * did it without rebooting, then the scrub dtl
1667 			 * will be valid, so excise the old region and
1668 			 * fold in the scrub dtl.  Otherwise, leave the
1669 			 * dtl as-is if there was an error.
1670 			 *
1671 			 * There's little trick here: to excise the beginning
1672 			 * of the DTL_MISSING map, we put it into a reference
1673 			 * tree and then add a segment with refcnt -1 that
1674 			 * covers the range [0, scrub_txg).  This means
1675 			 * that each txg in that range has refcnt -1 or 0.
1676 			 * We then add DTL_SCRUB with a refcnt of 2, so that
1677 			 * entries in the range [0, scrub_txg) will have a
1678 			 * positive refcnt -- either 1 or 2.  We then convert
1679 			 * the reference tree into the new DTL_MISSING map.
1680 			 */
1681 			space_map_ref_create(&reftree);
1682 			space_map_ref_add_map(&reftree,
1683 			    &vd->vdev_dtl[DTL_MISSING], 1);
1684 			space_map_ref_add_seg(&reftree, 0, scrub_txg, -1);
1685 			space_map_ref_add_map(&reftree,
1686 			    &vd->vdev_dtl[DTL_SCRUB], 2);
1687 			space_map_ref_generate_map(&reftree,
1688 			    &vd->vdev_dtl[DTL_MISSING], 1);
1689 			space_map_ref_destroy(&reftree);
1690 		}
1691 		space_map_vacate(&vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
1692 		space_map_walk(&vd->vdev_dtl[DTL_MISSING],
1693 		    space_map_add, &vd->vdev_dtl[DTL_PARTIAL]);
1694 		if (scrub_done)
1695 			space_map_vacate(&vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
1696 		space_map_vacate(&vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
1697 		if (!vdev_readable(vd))
1698 			space_map_add(&vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
1699 		else
1700 			space_map_walk(&vd->vdev_dtl[DTL_MISSING],
1701 			    space_map_add, &vd->vdev_dtl[DTL_OUTAGE]);
1702 		mutex_exit(&vd->vdev_dtl_lock);
1703 
1704 		if (txg != 0)
1705 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
1706 		return;
1707 	}
1708 
1709 	mutex_enter(&vd->vdev_dtl_lock);
1710 	for (int t = 0; t < DTL_TYPES; t++) {
1711 		/* account for child's outage in parent's missing map */
1712 		int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
1713 		if (t == DTL_SCRUB)
1714 			continue;			/* leaf vdevs only */
1715 		if (t == DTL_PARTIAL)
1716 			minref = 1;			/* i.e. non-zero */
1717 		else if (vd->vdev_nparity != 0)
1718 			minref = vd->vdev_nparity + 1;	/* RAID-Z */
1719 		else
1720 			minref = vd->vdev_children;	/* any kind of mirror */
1721 		space_map_ref_create(&reftree);
1722 		for (int c = 0; c < vd->vdev_children; c++) {
1723 			vdev_t *cvd = vd->vdev_child[c];
1724 			mutex_enter(&cvd->vdev_dtl_lock);
1725 			space_map_ref_add_map(&reftree, &cvd->vdev_dtl[s], 1);
1726 			mutex_exit(&cvd->vdev_dtl_lock);
1727 		}
1728 		space_map_ref_generate_map(&reftree, &vd->vdev_dtl[t], minref);
1729 		space_map_ref_destroy(&reftree);
1730 	}
1731 	mutex_exit(&vd->vdev_dtl_lock);
1732 }
1733 
1734 static int
1735 vdev_dtl_load(vdev_t *vd)
1736 {
1737 	spa_t *spa = vd->vdev_spa;
1738 	space_map_obj_t *smo = &vd->vdev_dtl_smo;
1739 	objset_t *mos = spa->spa_meta_objset;
1740 	dmu_buf_t *db;
1741 	int error;
1742 
1743 	ASSERT(vd->vdev_children == 0);
1744 
1745 	if (smo->smo_object == 0)
1746 		return (0);
1747 
1748 	ASSERT(!vd->vdev_ishole);
1749 
1750 	if ((error = dmu_bonus_hold(mos, smo->smo_object, FTAG, &db)) != 0)
1751 		return (error);
1752 
1753 	ASSERT3U(db->db_size, >=, sizeof (*smo));
1754 	bcopy(db->db_data, smo, sizeof (*smo));
1755 	dmu_buf_rele(db, FTAG);
1756 
1757 	mutex_enter(&vd->vdev_dtl_lock);
1758 	error = space_map_load(&vd->vdev_dtl[DTL_MISSING],
1759 	    NULL, SM_ALLOC, smo, mos);
1760 	mutex_exit(&vd->vdev_dtl_lock);
1761 
1762 	return (error);
1763 }
1764 
1765 void
1766 vdev_dtl_sync(vdev_t *vd, uint64_t txg)
1767 {
1768 	spa_t *spa = vd->vdev_spa;
1769 	space_map_obj_t *smo = &vd->vdev_dtl_smo;
1770 	space_map_t *sm = &vd->vdev_dtl[DTL_MISSING];
1771 	objset_t *mos = spa->spa_meta_objset;
1772 	space_map_t smsync;
1773 	kmutex_t smlock;
1774 	dmu_buf_t *db;
1775 	dmu_tx_t *tx;
1776 
1777 	ASSERT(!vd->vdev_ishole);
1778 
1779 	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1780 
1781 	if (vd->vdev_detached) {
1782 		if (smo->smo_object != 0) {
1783 			int err = dmu_object_free(mos, smo->smo_object, tx);
1784 			ASSERT3U(err, ==, 0);
1785 			smo->smo_object = 0;
1786 		}
1787 		dmu_tx_commit(tx);
1788 		return;
1789 	}
1790 
1791 	if (smo->smo_object == 0) {
1792 		ASSERT(smo->smo_objsize == 0);
1793 		ASSERT(smo->smo_alloc == 0);
1794 		smo->smo_object = dmu_object_alloc(mos,
1795 		    DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
1796 		    DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
1797 		ASSERT(smo->smo_object != 0);
1798 		vdev_config_dirty(vd->vdev_top);
1799 	}
1800 
1801 	mutex_init(&smlock, NULL, MUTEX_DEFAULT, NULL);
1802 
1803 	space_map_create(&smsync, sm->sm_start, sm->sm_size, sm->sm_shift,
1804 	    &smlock);
1805 
1806 	mutex_enter(&smlock);
1807 
1808 	mutex_enter(&vd->vdev_dtl_lock);
1809 	space_map_walk(sm, space_map_add, &smsync);
1810 	mutex_exit(&vd->vdev_dtl_lock);
1811 
1812 	space_map_truncate(smo, mos, tx);
1813 	space_map_sync(&smsync, SM_ALLOC, smo, mos, tx);
1814 
1815 	space_map_destroy(&smsync);
1816 
1817 	mutex_exit(&smlock);
1818 	mutex_destroy(&smlock);
1819 
1820 	VERIFY(0 == dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
1821 	dmu_buf_will_dirty(db, tx);
1822 	ASSERT3U(db->db_size, >=, sizeof (*smo));
1823 	bcopy(smo, db->db_data, sizeof (*smo));
1824 	dmu_buf_rele(db, FTAG);
1825 
1826 	dmu_tx_commit(tx);
1827 }
1828 
1829 /*
1830  * Determine whether the specified vdev can be offlined/detached/removed
1831  * without losing data.
1832  */
1833 boolean_t
1834 vdev_dtl_required(vdev_t *vd)
1835 {
1836 	spa_t *spa = vd->vdev_spa;
1837 	vdev_t *tvd = vd->vdev_top;
1838 	uint8_t cant_read = vd->vdev_cant_read;
1839 	boolean_t required;
1840 
1841 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1842 
1843 	if (vd == spa->spa_root_vdev || vd == tvd)
1844 		return (B_TRUE);
1845 
1846 	/*
1847 	 * Temporarily mark the device as unreadable, and then determine
1848 	 * whether this results in any DTL outages in the top-level vdev.
1849 	 * If not, we can safely offline/detach/remove the device.
1850 	 */
1851 	vd->vdev_cant_read = B_TRUE;
1852 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1853 	required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
1854 	vd->vdev_cant_read = cant_read;
1855 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1856 
1857 	if (!required && zio_injection_enabled)
1858 		required = !!zio_handle_device_injection(vd, NULL, ECHILD);
1859 
1860 	return (required);
1861 }
1862 
1863 /*
1864  * Determine if resilver is needed, and if so the txg range.
1865  */
1866 boolean_t
1867 vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
1868 {
1869 	boolean_t needed = B_FALSE;
1870 	uint64_t thismin = UINT64_MAX;
1871 	uint64_t thismax = 0;
1872 
1873 	if (vd->vdev_children == 0) {
1874 		mutex_enter(&vd->vdev_dtl_lock);
1875 		if (vd->vdev_dtl[DTL_MISSING].sm_space != 0 &&
1876 		    vdev_writeable(vd)) {
1877 			space_seg_t *ss;
1878 
1879 			ss = avl_first(&vd->vdev_dtl[DTL_MISSING].sm_root);
1880 			thismin = ss->ss_start - 1;
1881 			ss = avl_last(&vd->vdev_dtl[DTL_MISSING].sm_root);
1882 			thismax = ss->ss_end;
1883 			needed = B_TRUE;
1884 		}
1885 		mutex_exit(&vd->vdev_dtl_lock);
1886 	} else {
1887 		for (int c = 0; c < vd->vdev_children; c++) {
1888 			vdev_t *cvd = vd->vdev_child[c];
1889 			uint64_t cmin, cmax;
1890 
1891 			if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
1892 				thismin = MIN(thismin, cmin);
1893 				thismax = MAX(thismax, cmax);
1894 				needed = B_TRUE;
1895 			}
1896 		}
1897 	}
1898 
1899 	if (needed && minp) {
1900 		*minp = thismin;
1901 		*maxp = thismax;
1902 	}
1903 	return (needed);
1904 }
1905 
1906 void
1907 vdev_load(vdev_t *vd)
1908 {
1909 	/*
1910 	 * Recursively load all children.
1911 	 */
1912 	for (int c = 0; c < vd->vdev_children; c++)
1913 		vdev_load(vd->vdev_child[c]);
1914 
1915 	/*
1916 	 * If this is a top-level vdev, initialize its metaslabs.
1917 	 */
1918 	if (vd == vd->vdev_top && !vd->vdev_ishole &&
1919 	    (vd->vdev_ashift == 0 || vd->vdev_asize == 0 ||
1920 	    vdev_metaslab_init(vd, 0) != 0))
1921 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1922 		    VDEV_AUX_CORRUPT_DATA);
1923 
1924 	/*
1925 	 * If this is a leaf vdev, load its DTL.
1926 	 */
1927 	if (vd->vdev_ops->vdev_op_leaf && vdev_dtl_load(vd) != 0)
1928 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1929 		    VDEV_AUX_CORRUPT_DATA);
1930 }
1931 
1932 /*
1933  * The special vdev case is used for hot spares and l2cache devices.  Its
1934  * sole purpose it to set the vdev state for the associated vdev.  To do this,
1935  * we make sure that we can open the underlying device, then try to read the
1936  * label, and make sure that the label is sane and that it hasn't been
1937  * repurposed to another pool.
1938  */
1939 int
1940 vdev_validate_aux(vdev_t *vd)
1941 {
1942 	nvlist_t *label;
1943 	uint64_t guid, version;
1944 	uint64_t state;
1945 
1946 	if (!vdev_readable(vd))
1947 		return (0);
1948 
1949 	if ((label = vdev_label_read_config(vd)) == NULL) {
1950 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1951 		    VDEV_AUX_CORRUPT_DATA);
1952 		return (-1);
1953 	}
1954 
1955 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
1956 	    version > SPA_VERSION ||
1957 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
1958 	    guid != vd->vdev_guid ||
1959 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
1960 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1961 		    VDEV_AUX_CORRUPT_DATA);
1962 		nvlist_free(label);
1963 		return (-1);
1964 	}
1965 
1966 	/*
1967 	 * We don't actually check the pool state here.  If it's in fact in
1968 	 * use by another pool, we update this fact on the fly when requested.
1969 	 */
1970 	nvlist_free(label);
1971 	return (0);
1972 }
1973 
1974 void
1975 vdev_remove(vdev_t *vd, uint64_t txg)
1976 {
1977 	spa_t *spa = vd->vdev_spa;
1978 	objset_t *mos = spa->spa_meta_objset;
1979 	dmu_tx_t *tx;
1980 
1981 	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
1982 
1983 	if (vd->vdev_dtl_smo.smo_object) {
1984 		ASSERT3U(vd->vdev_dtl_smo.smo_alloc, ==, 0);
1985 		(void) dmu_object_free(mos, vd->vdev_dtl_smo.smo_object, tx);
1986 		vd->vdev_dtl_smo.smo_object = 0;
1987 	}
1988 
1989 	if (vd->vdev_ms != NULL) {
1990 		for (int m = 0; m < vd->vdev_ms_count; m++) {
1991 			metaslab_t *msp = vd->vdev_ms[m];
1992 
1993 			if (msp == NULL || msp->ms_smo.smo_object == 0)
1994 				continue;
1995 
1996 			ASSERT3U(msp->ms_smo.smo_alloc, ==, 0);
1997 			(void) dmu_object_free(mos, msp->ms_smo.smo_object, tx);
1998 			msp->ms_smo.smo_object = 0;
1999 		}
2000 	}
2001 
2002 	if (vd->vdev_ms_array) {
2003 		(void) dmu_object_free(mos, vd->vdev_ms_array, tx);
2004 		vd->vdev_ms_array = 0;
2005 		vd->vdev_ms_shift = 0;
2006 	}
2007 	dmu_tx_commit(tx);
2008 }
2009 
2010 void
2011 vdev_sync_done(vdev_t *vd, uint64_t txg)
2012 {
2013 	metaslab_t *msp;
2014 	boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg));
2015 
2016 	ASSERT(!vd->vdev_ishole);
2017 
2018 	while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
2019 		metaslab_sync_done(msp, txg);
2020 
2021 	if (reassess)
2022 		metaslab_sync_reassess(vd->vdev_mg);
2023 }
2024 
2025 void
2026 vdev_sync(vdev_t *vd, uint64_t txg)
2027 {
2028 	spa_t *spa = vd->vdev_spa;
2029 	vdev_t *lvd;
2030 	metaslab_t *msp;
2031 	dmu_tx_t *tx;
2032 
2033 	ASSERT(!vd->vdev_ishole);
2034 
2035 	if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0) {
2036 		ASSERT(vd == vd->vdev_top);
2037 		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2038 		vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
2039 		    DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
2040 		ASSERT(vd->vdev_ms_array != 0);
2041 		vdev_config_dirty(vd);
2042 		dmu_tx_commit(tx);
2043 	}
2044 
2045 	/*
2046 	 * Remove the metadata associated with this vdev once it's empty.
2047 	 */
2048 	if (vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing)
2049 		vdev_remove(vd, txg);
2050 
2051 	while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
2052 		metaslab_sync(msp, txg);
2053 		(void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
2054 	}
2055 
2056 	while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
2057 		vdev_dtl_sync(lvd, txg);
2058 
2059 	(void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
2060 }
2061 
2062 uint64_t
2063 vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
2064 {
2065 	return (vd->vdev_ops->vdev_op_asize(vd, psize));
2066 }
2067 
2068 /*
2069  * Mark the given vdev faulted.  A faulted vdev behaves as if the device could
2070  * not be opened, and no I/O is attempted.
2071  */
2072 int
2073 vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2074 {
2075 	vdev_t *vd, *tvd;
2076 
2077 	spa_vdev_state_enter(spa, SCL_NONE);
2078 
2079 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2080 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2081 
2082 	if (!vd->vdev_ops->vdev_op_leaf)
2083 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2084 
2085 	tvd = vd->vdev_top;
2086 
2087 	/*
2088 	 * We don't directly use the aux state here, but if we do a
2089 	 * vdev_reopen(), we need this value to be present to remember why we
2090 	 * were faulted.
2091 	 */
2092 	vd->vdev_label_aux = aux;
2093 
2094 	/*
2095 	 * Faulted state takes precedence over degraded.
2096 	 */
2097 	vd->vdev_delayed_close = B_FALSE;
2098 	vd->vdev_faulted = 1ULL;
2099 	vd->vdev_degraded = 0ULL;
2100 	vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
2101 
2102 	/*
2103 	 * If this device has the only valid copy of the data, then
2104 	 * back off and simply mark the vdev as degraded instead.
2105 	 */
2106 	if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) {
2107 		vd->vdev_degraded = 1ULL;
2108 		vd->vdev_faulted = 0ULL;
2109 
2110 		/*
2111 		 * If we reopen the device and it's not dead, only then do we
2112 		 * mark it degraded.
2113 		 */
2114 		vdev_reopen(tvd);
2115 
2116 		if (vdev_readable(vd))
2117 			vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
2118 	}
2119 
2120 	return (spa_vdev_state_exit(spa, vd, 0));
2121 }
2122 
2123 /*
2124  * Mark the given vdev degraded.  A degraded vdev is purely an indication to the
2125  * user that something is wrong.  The vdev continues to operate as normal as far
2126  * as I/O is concerned.
2127  */
2128 int
2129 vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2130 {
2131 	vdev_t *vd;
2132 
2133 	spa_vdev_state_enter(spa, SCL_NONE);
2134 
2135 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2136 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2137 
2138 	if (!vd->vdev_ops->vdev_op_leaf)
2139 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2140 
2141 	/*
2142 	 * If the vdev is already faulted, then don't do anything.
2143 	 */
2144 	if (vd->vdev_faulted || vd->vdev_degraded)
2145 		return (spa_vdev_state_exit(spa, NULL, 0));
2146 
2147 	vd->vdev_degraded = 1ULL;
2148 	if (!vdev_is_dead(vd))
2149 		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
2150 		    aux);
2151 
2152 	return (spa_vdev_state_exit(spa, vd, 0));
2153 }
2154 
2155 /*
2156  * Online the given vdev.  If 'unspare' is set, it implies two things.  First,
2157  * any attached spare device should be detached when the device finishes
2158  * resilvering.  Second, the online should be treated like a 'test' online case,
2159  * so no FMA events are generated if the device fails to open.
2160  */
2161 int
2162 vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
2163 {
2164 	vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
2165 
2166 	spa_vdev_state_enter(spa, SCL_NONE);
2167 
2168 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2169 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2170 
2171 	if (!vd->vdev_ops->vdev_op_leaf)
2172 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2173 
2174 	tvd = vd->vdev_top;
2175 	vd->vdev_offline = B_FALSE;
2176 	vd->vdev_tmpoffline = B_FALSE;
2177 	vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
2178 	vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
2179 
2180 	/* XXX - L2ARC 1.0 does not support expansion */
2181 	if (!vd->vdev_aux) {
2182 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2183 			pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
2184 	}
2185 
2186 	vdev_reopen(tvd);
2187 	vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
2188 
2189 	if (!vd->vdev_aux) {
2190 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2191 			pvd->vdev_expanding = B_FALSE;
2192 	}
2193 
2194 	if (newstate)
2195 		*newstate = vd->vdev_state;
2196 	if ((flags & ZFS_ONLINE_UNSPARE) &&
2197 	    !vdev_is_dead(vd) && vd->vdev_parent &&
2198 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2199 	    vd->vdev_parent->vdev_child[0] == vd)
2200 		vd->vdev_unspare = B_TRUE;
2201 
2202 	if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
2203 
2204 		/* XXX - L2ARC 1.0 does not support expansion */
2205 		if (vd->vdev_aux)
2206 			return (spa_vdev_state_exit(spa, vd, ENOTSUP));
2207 		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
2208 	}
2209 	return (spa_vdev_state_exit(spa, vd, 0));
2210 }
2211 
2212 static int
2213 vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
2214 {
2215 	vdev_t *vd, *tvd;
2216 	int error = 0;
2217 	uint64_t generation;
2218 	metaslab_group_t *mg;
2219 
2220 top:
2221 	spa_vdev_state_enter(spa, SCL_ALLOC);
2222 
2223 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2224 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2225 
2226 	if (!vd->vdev_ops->vdev_op_leaf)
2227 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2228 
2229 	tvd = vd->vdev_top;
2230 	mg = tvd->vdev_mg;
2231 	generation = spa->spa_config_generation + 1;
2232 
2233 	/*
2234 	 * If the device isn't already offline, try to offline it.
2235 	 */
2236 	if (!vd->vdev_offline) {
2237 		/*
2238 		 * If this device has the only valid copy of some data,
2239 		 * don't allow it to be offlined. Log devices are always
2240 		 * expendable.
2241 		 */
2242 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2243 		    vdev_dtl_required(vd))
2244 			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2245 
2246 		/*
2247 		 * If the top-level is a slog and it has had allocations
2248 		 * then proceed.  We check that the vdev's metaslab group
2249 		 * is not NULL since it's possible that we may have just
2250 		 * added this vdev but not yet initialized its metaslabs.
2251 		 */
2252 		if (tvd->vdev_islog && mg != NULL) {
2253 			/*
2254 			 * Prevent any future allocations.
2255 			 */
2256 			metaslab_group_passivate(mg);
2257 			(void) spa_vdev_state_exit(spa, vd, 0);
2258 
2259 			error = spa_offline_log(spa);
2260 
2261 			spa_vdev_state_enter(spa, SCL_ALLOC);
2262 
2263 			/*
2264 			 * Check to see if the config has changed.
2265 			 */
2266 			if (error || generation != spa->spa_config_generation) {
2267 				metaslab_group_activate(mg);
2268 				if (error)
2269 					return (spa_vdev_state_exit(spa,
2270 					    vd, error));
2271 				(void) spa_vdev_state_exit(spa, vd, 0);
2272 				goto top;
2273 			}
2274 			ASSERT3U(tvd->vdev_stat.vs_alloc, ==, 0);
2275 		}
2276 
2277 		/*
2278 		 * Offline this device and reopen its top-level vdev.
2279 		 * If the top-level vdev is a log device then just offline
2280 		 * it. Otherwise, if this action results in the top-level
2281 		 * vdev becoming unusable, undo it and fail the request.
2282 		 */
2283 		vd->vdev_offline = B_TRUE;
2284 		vdev_reopen(tvd);
2285 
2286 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2287 		    vdev_is_dead(tvd)) {
2288 			vd->vdev_offline = B_FALSE;
2289 			vdev_reopen(tvd);
2290 			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2291 		}
2292 
2293 		/*
2294 		 * Add the device back into the metaslab rotor so that
2295 		 * once we online the device it's open for business.
2296 		 */
2297 		if (tvd->vdev_islog && mg != NULL)
2298 			metaslab_group_activate(mg);
2299 	}
2300 
2301 	vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
2302 
2303 	return (spa_vdev_state_exit(spa, vd, 0));
2304 }
2305 
2306 int
2307 vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
2308 {
2309 	int error;
2310 
2311 	mutex_enter(&spa->spa_vdev_top_lock);
2312 	error = vdev_offline_locked(spa, guid, flags);
2313 	mutex_exit(&spa->spa_vdev_top_lock);
2314 
2315 	return (error);
2316 }
2317 
2318 /*
2319  * Clear the error counts associated with this vdev.  Unlike vdev_online() and
2320  * vdev_offline(), we assume the spa config is locked.  We also clear all
2321  * children.  If 'vd' is NULL, then the user wants to clear all vdevs.
2322  */
2323 void
2324 vdev_clear(spa_t *spa, vdev_t *vd)
2325 {
2326 	vdev_t *rvd = spa->spa_root_vdev;
2327 
2328 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2329 
2330 	if (vd == NULL)
2331 		vd = rvd;
2332 
2333 	vd->vdev_stat.vs_read_errors = 0;
2334 	vd->vdev_stat.vs_write_errors = 0;
2335 	vd->vdev_stat.vs_checksum_errors = 0;
2336 
2337 	for (int c = 0; c < vd->vdev_children; c++)
2338 		vdev_clear(spa, vd->vdev_child[c]);
2339 
2340 	/*
2341 	 * If we're in the FAULTED state or have experienced failed I/O, then
2342 	 * clear the persistent state and attempt to reopen the device.  We
2343 	 * also mark the vdev config dirty, so that the new faulted state is
2344 	 * written out to disk.
2345 	 */
2346 	if (vd->vdev_faulted || vd->vdev_degraded ||
2347 	    !vdev_readable(vd) || !vdev_writeable(vd)) {
2348 
2349 		/*
2350 		 * When reopening in reponse to a clear event, it may be due to
2351 		 * a fmadm repair request.  In this case, if the device is
2352 		 * still broken, we want to still post the ereport again.
2353 		 */
2354 		vd->vdev_forcefault = B_TRUE;
2355 
2356 		vd->vdev_faulted = vd->vdev_degraded = 0ULL;
2357 		vd->vdev_cant_read = B_FALSE;
2358 		vd->vdev_cant_write = B_FALSE;
2359 
2360 		vdev_reopen(vd == rvd ? rvd : vd->vdev_top);
2361 
2362 		vd->vdev_forcefault = B_FALSE;
2363 
2364 		if (vd != rvd && vdev_writeable(vd->vdev_top))
2365 			vdev_state_dirty(vd->vdev_top);
2366 
2367 		if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
2368 			spa_async_request(spa, SPA_ASYNC_RESILVER);
2369 
2370 		spa_event_notify(spa, vd, ESC_ZFS_VDEV_CLEAR);
2371 	}
2372 
2373 	/*
2374 	 * When clearing a FMA-diagnosed fault, we always want to
2375 	 * unspare the device, as we assume that the original spare was
2376 	 * done in response to the FMA fault.
2377 	 */
2378 	if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
2379 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2380 	    vd->vdev_parent->vdev_child[0] == vd)
2381 		vd->vdev_unspare = B_TRUE;
2382 }
2383 
2384 boolean_t
2385 vdev_is_dead(vdev_t *vd)
2386 {
2387 	/*
2388 	 * Holes and missing devices are always considered "dead".
2389 	 * This simplifies the code since we don't have to check for
2390 	 * these types of devices in the various code paths.
2391 	 * Instead we rely on the fact that we skip over dead devices
2392 	 * before issuing I/O to them.
2393 	 */
2394 	return (vd->vdev_state < VDEV_STATE_DEGRADED || vd->vdev_ishole ||
2395 	    vd->vdev_ops == &vdev_missing_ops);
2396 }
2397 
2398 boolean_t
2399 vdev_readable(vdev_t *vd)
2400 {
2401 	return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
2402 }
2403 
2404 boolean_t
2405 vdev_writeable(vdev_t *vd)
2406 {
2407 	return (!vdev_is_dead(vd) && !vd->vdev_cant_write);
2408 }
2409 
2410 boolean_t
2411 vdev_allocatable(vdev_t *vd)
2412 {
2413 	uint64_t state = vd->vdev_state;
2414 
2415 	/*
2416 	 * We currently allow allocations from vdevs which may be in the
2417 	 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
2418 	 * fails to reopen then we'll catch it later when we're holding
2419 	 * the proper locks.  Note that we have to get the vdev state
2420 	 * in a local variable because although it changes atomically,
2421 	 * we're asking two separate questions about it.
2422 	 */
2423 	return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
2424 	    !vd->vdev_cant_write && !vd->vdev_ishole);
2425 }
2426 
2427 boolean_t
2428 vdev_accessible(vdev_t *vd, zio_t *zio)
2429 {
2430 	ASSERT(zio->io_vd == vd);
2431 
2432 	if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
2433 		return (B_FALSE);
2434 
2435 	if (zio->io_type == ZIO_TYPE_READ)
2436 		return (!vd->vdev_cant_read);
2437 
2438 	if (zio->io_type == ZIO_TYPE_WRITE)
2439 		return (!vd->vdev_cant_write);
2440 
2441 	return (B_TRUE);
2442 }
2443 
2444 /*
2445  * Get statistics for the given vdev.
2446  */
2447 void
2448 vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
2449 {
2450 	vdev_t *rvd = vd->vdev_spa->spa_root_vdev;
2451 
2452 	mutex_enter(&vd->vdev_stat_lock);
2453 	bcopy(&vd->vdev_stat, vs, sizeof (*vs));
2454 	vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
2455 	vs->vs_state = vd->vdev_state;
2456 	vs->vs_rsize = vdev_get_min_asize(vd);
2457 	if (vd->vdev_ops->vdev_op_leaf)
2458 		vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
2459 	mutex_exit(&vd->vdev_stat_lock);
2460 
2461 	/*
2462 	 * If we're getting stats on the root vdev, aggregate the I/O counts
2463 	 * over all top-level vdevs (i.e. the direct children of the root).
2464 	 */
2465 	if (vd == rvd) {
2466 		for (int c = 0; c < rvd->vdev_children; c++) {
2467 			vdev_t *cvd = rvd->vdev_child[c];
2468 			vdev_stat_t *cvs = &cvd->vdev_stat;
2469 
2470 			mutex_enter(&vd->vdev_stat_lock);
2471 			for (int t = 0; t < ZIO_TYPES; t++) {
2472 				vs->vs_ops[t] += cvs->vs_ops[t];
2473 				vs->vs_bytes[t] += cvs->vs_bytes[t];
2474 			}
2475 			cvs->vs_scan_removing = cvd->vdev_removing;
2476 			mutex_exit(&vd->vdev_stat_lock);
2477 		}
2478 	}
2479 }
2480 
2481 void
2482 vdev_clear_stats(vdev_t *vd)
2483 {
2484 	mutex_enter(&vd->vdev_stat_lock);
2485 	vd->vdev_stat.vs_space = 0;
2486 	vd->vdev_stat.vs_dspace = 0;
2487 	vd->vdev_stat.vs_alloc = 0;
2488 	mutex_exit(&vd->vdev_stat_lock);
2489 }
2490 
2491 void
2492 vdev_scan_stat_init(vdev_t *vd)
2493 {
2494 	vdev_stat_t *vs = &vd->vdev_stat;
2495 
2496 	for (int c = 0; c < vd->vdev_children; c++)
2497 		vdev_scan_stat_init(vd->vdev_child[c]);
2498 
2499 	mutex_enter(&vd->vdev_stat_lock);
2500 	vs->vs_scan_processed = 0;
2501 	mutex_exit(&vd->vdev_stat_lock);
2502 }
2503 
2504 void
2505 vdev_stat_update(zio_t *zio, uint64_t psize)
2506 {
2507 	spa_t *spa = zio->io_spa;
2508 	vdev_t *rvd = spa->spa_root_vdev;
2509 	vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
2510 	vdev_t *pvd;
2511 	uint64_t txg = zio->io_txg;
2512 	vdev_stat_t *vs = &vd->vdev_stat;
2513 	zio_type_t type = zio->io_type;
2514 	int flags = zio->io_flags;
2515 
2516 	/*
2517 	 * If this i/o is a gang leader, it didn't do any actual work.
2518 	 */
2519 	if (zio->io_gang_tree)
2520 		return;
2521 
2522 	if (zio->io_error == 0) {
2523 		/*
2524 		 * If this is a root i/o, don't count it -- we've already
2525 		 * counted the top-level vdevs, and vdev_get_stats() will
2526 		 * aggregate them when asked.  This reduces contention on
2527 		 * the root vdev_stat_lock and implicitly handles blocks
2528 		 * that compress away to holes, for which there is no i/o.
2529 		 * (Holes never create vdev children, so all the counters
2530 		 * remain zero, which is what we want.)
2531 		 *
2532 		 * Note: this only applies to successful i/o (io_error == 0)
2533 		 * because unlike i/o counts, errors are not additive.
2534 		 * When reading a ditto block, for example, failure of
2535 		 * one top-level vdev does not imply a root-level error.
2536 		 */
2537 		if (vd == rvd)
2538 			return;
2539 
2540 		ASSERT(vd == zio->io_vd);
2541 
2542 		if (flags & ZIO_FLAG_IO_BYPASS)
2543 			return;
2544 
2545 		mutex_enter(&vd->vdev_stat_lock);
2546 
2547 		if (flags & ZIO_FLAG_IO_REPAIR) {
2548 			if (flags & ZIO_FLAG_SCAN_THREAD) {
2549 				dsl_scan_phys_t *scn_phys =
2550 				    &spa->spa_dsl_pool->dp_scan->scn_phys;
2551 				uint64_t *processed = &scn_phys->scn_processed;
2552 
2553 				/* XXX cleanup? */
2554 				if (vd->vdev_ops->vdev_op_leaf)
2555 					atomic_add_64(processed, psize);
2556 				vs->vs_scan_processed += psize;
2557 			}
2558 
2559 			if (flags & ZIO_FLAG_SELF_HEAL)
2560 				vs->vs_self_healed += psize;
2561 		}
2562 
2563 		vs->vs_ops[type]++;
2564 		vs->vs_bytes[type] += psize;
2565 
2566 		mutex_exit(&vd->vdev_stat_lock);
2567 		return;
2568 	}
2569 
2570 	if (flags & ZIO_FLAG_SPECULATIVE)
2571 		return;
2572 
2573 	/*
2574 	 * If this is an I/O error that is going to be retried, then ignore the
2575 	 * error.  Otherwise, the user may interpret B_FAILFAST I/O errors as
2576 	 * hard errors, when in reality they can happen for any number of
2577 	 * innocuous reasons (bus resets, MPxIO link failure, etc).
2578 	 */
2579 	if (zio->io_error == EIO &&
2580 	    !(zio->io_flags & ZIO_FLAG_IO_RETRY))
2581 		return;
2582 
2583 	/*
2584 	 * Intent logs writes won't propagate their error to the root
2585 	 * I/O so don't mark these types of failures as pool-level
2586 	 * errors.
2587 	 */
2588 	if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
2589 		return;
2590 
2591 	mutex_enter(&vd->vdev_stat_lock);
2592 	if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
2593 		if (zio->io_error == ECKSUM)
2594 			vs->vs_checksum_errors++;
2595 		else
2596 			vs->vs_read_errors++;
2597 	}
2598 	if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
2599 		vs->vs_write_errors++;
2600 	mutex_exit(&vd->vdev_stat_lock);
2601 
2602 	if (type == ZIO_TYPE_WRITE && txg != 0 &&
2603 	    (!(flags & ZIO_FLAG_IO_REPAIR) ||
2604 	    (flags & ZIO_FLAG_SCAN_THREAD) ||
2605 	    spa->spa_claiming)) {
2606 		/*
2607 		 * This is either a normal write (not a repair), or it's
2608 		 * a repair induced by the scrub thread, or it's a repair
2609 		 * made by zil_claim() during spa_load() in the first txg.
2610 		 * In the normal case, we commit the DTL change in the same
2611 		 * txg as the block was born.  In the scrub-induced repair
2612 		 * case, we know that scrubs run in first-pass syncing context,
2613 		 * so we commit the DTL change in spa_syncing_txg(spa).
2614 		 * In the zil_claim() case, we commit in spa_first_txg(spa).
2615 		 *
2616 		 * We currently do not make DTL entries for failed spontaneous
2617 		 * self-healing writes triggered by normal (non-scrubbing)
2618 		 * reads, because we have no transactional context in which to
2619 		 * do so -- and it's not clear that it'd be desirable anyway.
2620 		 */
2621 		if (vd->vdev_ops->vdev_op_leaf) {
2622 			uint64_t commit_txg = txg;
2623 			if (flags & ZIO_FLAG_SCAN_THREAD) {
2624 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2625 				ASSERT(spa_sync_pass(spa) == 1);
2626 				vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
2627 				commit_txg = spa_syncing_txg(spa);
2628 			} else if (spa->spa_claiming) {
2629 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2630 				commit_txg = spa_first_txg(spa);
2631 			}
2632 			ASSERT(commit_txg >= spa_syncing_txg(spa));
2633 			if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
2634 				return;
2635 			for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2636 				vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
2637 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
2638 		}
2639 		if (vd != rvd)
2640 			vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
2641 	}
2642 }
2643 
2644 /*
2645  * Update the in-core space usage stats for this vdev, its metaslab class,
2646  * and the root vdev.
2647  */
2648 void
2649 vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
2650     int64_t space_delta)
2651 {
2652 	int64_t dspace_delta = space_delta;
2653 	spa_t *spa = vd->vdev_spa;
2654 	vdev_t *rvd = spa->spa_root_vdev;
2655 	metaslab_group_t *mg = vd->vdev_mg;
2656 	metaslab_class_t *mc = mg ? mg->mg_class : NULL;
2657 
2658 	ASSERT(vd == vd->vdev_top);
2659 
2660 	/*
2661 	 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
2662 	 * factor.  We must calculate this here and not at the root vdev
2663 	 * because the root vdev's psize-to-asize is simply the max of its
2664 	 * childrens', thus not accurate enough for us.
2665 	 */
2666 	ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
2667 	ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
2668 	dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
2669 	    vd->vdev_deflate_ratio;
2670 
2671 	mutex_enter(&vd->vdev_stat_lock);
2672 	vd->vdev_stat.vs_alloc += alloc_delta;
2673 	vd->vdev_stat.vs_space += space_delta;
2674 	vd->vdev_stat.vs_dspace += dspace_delta;
2675 	mutex_exit(&vd->vdev_stat_lock);
2676 
2677 	if (mc == spa_normal_class(spa)) {
2678 		mutex_enter(&rvd->vdev_stat_lock);
2679 		rvd->vdev_stat.vs_alloc += alloc_delta;
2680 		rvd->vdev_stat.vs_space += space_delta;
2681 		rvd->vdev_stat.vs_dspace += dspace_delta;
2682 		mutex_exit(&rvd->vdev_stat_lock);
2683 	}
2684 
2685 	if (mc != NULL) {
2686 		ASSERT(rvd == vd->vdev_parent);
2687 		ASSERT(vd->vdev_ms_count != 0);
2688 
2689 		metaslab_class_space_update(mc,
2690 		    alloc_delta, defer_delta, space_delta, dspace_delta);
2691 	}
2692 }
2693 
2694 /*
2695  * Mark a top-level vdev's config as dirty, placing it on the dirty list
2696  * so that it will be written out next time the vdev configuration is synced.
2697  * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
2698  */
2699 void
2700 vdev_config_dirty(vdev_t *vd)
2701 {
2702 	spa_t *spa = vd->vdev_spa;
2703 	vdev_t *rvd = spa->spa_root_vdev;
2704 	int c;
2705 
2706 	ASSERT(spa_writeable(spa));
2707 
2708 	/*
2709 	 * If this is an aux vdev (as with l2cache and spare devices), then we
2710 	 * update the vdev config manually and set the sync flag.
2711 	 */
2712 	if (vd->vdev_aux != NULL) {
2713 		spa_aux_vdev_t *sav = vd->vdev_aux;
2714 		nvlist_t **aux;
2715 		uint_t naux;
2716 
2717 		for (c = 0; c < sav->sav_count; c++) {
2718 			if (sav->sav_vdevs[c] == vd)
2719 				break;
2720 		}
2721 
2722 		if (c == sav->sav_count) {
2723 			/*
2724 			 * We're being removed.  There's nothing more to do.
2725 			 */
2726 			ASSERT(sav->sav_sync == B_TRUE);
2727 			return;
2728 		}
2729 
2730 		sav->sav_sync = B_TRUE;
2731 
2732 		if (nvlist_lookup_nvlist_array(sav->sav_config,
2733 		    ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
2734 			VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
2735 			    ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
2736 		}
2737 
2738 		ASSERT(c < naux);
2739 
2740 		/*
2741 		 * Setting the nvlist in the middle if the array is a little
2742 		 * sketchy, but it will work.
2743 		 */
2744 		nvlist_free(aux[c]);
2745 		aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0);
2746 
2747 		return;
2748 	}
2749 
2750 	/*
2751 	 * The dirty list is protected by the SCL_CONFIG lock.  The caller
2752 	 * must either hold SCL_CONFIG as writer, or must be the sync thread
2753 	 * (which holds SCL_CONFIG as reader).  There's only one sync thread,
2754 	 * so this is sufficient to ensure mutual exclusion.
2755 	 */
2756 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2757 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2758 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
2759 
2760 	if (vd == rvd) {
2761 		for (c = 0; c < rvd->vdev_children; c++)
2762 			vdev_config_dirty(rvd->vdev_child[c]);
2763 	} else {
2764 		ASSERT(vd == vd->vdev_top);
2765 
2766 		if (!list_link_active(&vd->vdev_config_dirty_node) &&
2767 		    !vd->vdev_ishole)
2768 			list_insert_head(&spa->spa_config_dirty_list, vd);
2769 	}
2770 }
2771 
2772 void
2773 vdev_config_clean(vdev_t *vd)
2774 {
2775 	spa_t *spa = vd->vdev_spa;
2776 
2777 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2778 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2779 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
2780 
2781 	ASSERT(list_link_active(&vd->vdev_config_dirty_node));
2782 	list_remove(&spa->spa_config_dirty_list, vd);
2783 }
2784 
2785 /*
2786  * Mark a top-level vdev's state as dirty, so that the next pass of
2787  * spa_sync() can convert this into vdev_config_dirty().  We distinguish
2788  * the state changes from larger config changes because they require
2789  * much less locking, and are often needed for administrative actions.
2790  */
2791 void
2792 vdev_state_dirty(vdev_t *vd)
2793 {
2794 	spa_t *spa = vd->vdev_spa;
2795 
2796 	ASSERT(spa_writeable(spa));
2797 	ASSERT(vd == vd->vdev_top);
2798 
2799 	/*
2800 	 * The state list is protected by the SCL_STATE lock.  The caller
2801 	 * must either hold SCL_STATE as writer, or must be the sync thread
2802 	 * (which holds SCL_STATE as reader).  There's only one sync thread,
2803 	 * so this is sufficient to ensure mutual exclusion.
2804 	 */
2805 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2806 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2807 	    spa_config_held(spa, SCL_STATE, RW_READER)));
2808 
2809 	if (!list_link_active(&vd->vdev_state_dirty_node) && !vd->vdev_ishole)
2810 		list_insert_head(&spa->spa_state_dirty_list, vd);
2811 }
2812 
2813 void
2814 vdev_state_clean(vdev_t *vd)
2815 {
2816 	spa_t *spa = vd->vdev_spa;
2817 
2818 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2819 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2820 	    spa_config_held(spa, SCL_STATE, RW_READER)));
2821 
2822 	ASSERT(list_link_active(&vd->vdev_state_dirty_node));
2823 	list_remove(&spa->spa_state_dirty_list, vd);
2824 }
2825 
2826 /*
2827  * Propagate vdev state up from children to parent.
2828  */
2829 void
2830 vdev_propagate_state(vdev_t *vd)
2831 {
2832 	spa_t *spa = vd->vdev_spa;
2833 	vdev_t *rvd = spa->spa_root_vdev;
2834 	int degraded = 0, faulted = 0;
2835 	int corrupted = 0;
2836 	vdev_t *child;
2837 
2838 	if (vd->vdev_children > 0) {
2839 		for (int c = 0; c < vd->vdev_children; c++) {
2840 			child = vd->vdev_child[c];
2841 
2842 			/*
2843 			 * Don't factor holes into the decision.
2844 			 */
2845 			if (child->vdev_ishole)
2846 				continue;
2847 
2848 			if (!vdev_readable(child) ||
2849 			    (!vdev_writeable(child) && spa_writeable(spa))) {
2850 				/*
2851 				 * Root special: if there is a top-level log
2852 				 * device, treat the root vdev as if it were
2853 				 * degraded.
2854 				 */
2855 				if (child->vdev_islog && vd == rvd)
2856 					degraded++;
2857 				else
2858 					faulted++;
2859 			} else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
2860 				degraded++;
2861 			}
2862 
2863 			if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
2864 				corrupted++;
2865 		}
2866 
2867 		vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
2868 
2869 		/*
2870 		 * Root special: if there is a top-level vdev that cannot be
2871 		 * opened due to corrupted metadata, then propagate the root
2872 		 * vdev's aux state as 'corrupt' rather than 'insufficient
2873 		 * replicas'.
2874 		 */
2875 		if (corrupted && vd == rvd &&
2876 		    rvd->vdev_state == VDEV_STATE_CANT_OPEN)
2877 			vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
2878 			    VDEV_AUX_CORRUPT_DATA);
2879 	}
2880 
2881 	if (vd->vdev_parent)
2882 		vdev_propagate_state(vd->vdev_parent);
2883 }
2884 
2885 /*
2886  * Set a vdev's state.  If this is during an open, we don't update the parent
2887  * state, because we're in the process of opening children depth-first.
2888  * Otherwise, we propagate the change to the parent.
2889  *
2890  * If this routine places a device in a faulted state, an appropriate ereport is
2891  * generated.
2892  */
2893 void
2894 vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
2895 {
2896 	uint64_t save_state;
2897 	spa_t *spa = vd->vdev_spa;
2898 
2899 	if (state == vd->vdev_state) {
2900 		vd->vdev_stat.vs_aux = aux;
2901 		return;
2902 	}
2903 
2904 	save_state = vd->vdev_state;
2905 
2906 	vd->vdev_state = state;
2907 	vd->vdev_stat.vs_aux = aux;
2908 
2909 	/*
2910 	 * If we are setting the vdev state to anything but an open state, then
2911 	 * always close the underlying device unless the device has requested
2912 	 * a delayed close (i.e. we're about to remove or fault the device).
2913 	 * Otherwise, we keep accessible but invalid devices open forever.
2914 	 * We don't call vdev_close() itself, because that implies some extra
2915 	 * checks (offline, etc) that we don't want here.  This is limited to
2916 	 * leaf devices, because otherwise closing the device will affect other
2917 	 * children.
2918 	 */
2919 	if (!vd->vdev_delayed_close && vdev_is_dead(vd) &&
2920 	    vd->vdev_ops->vdev_op_leaf)
2921 		vd->vdev_ops->vdev_op_close(vd);
2922 
2923 	/*
2924 	 * If we have brought this vdev back into service, we need
2925 	 * to notify fmd so that it can gracefully repair any outstanding
2926 	 * cases due to a missing device.  We do this in all cases, even those
2927 	 * that probably don't correlate to a repaired fault.  This is sure to
2928 	 * catch all cases, and we let the zfs-retire agent sort it out.  If
2929 	 * this is a transient state it's OK, as the retire agent will
2930 	 * double-check the state of the vdev before repairing it.
2931 	 */
2932 	if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf &&
2933 	    vd->vdev_prevstate != state)
2934 		zfs_post_state_change(spa, vd);
2935 
2936 	if (vd->vdev_removed &&
2937 	    state == VDEV_STATE_CANT_OPEN &&
2938 	    (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
2939 		/*
2940 		 * If the previous state is set to VDEV_STATE_REMOVED, then this
2941 		 * device was previously marked removed and someone attempted to
2942 		 * reopen it.  If this failed due to a nonexistent device, then
2943 		 * keep the device in the REMOVED state.  We also let this be if
2944 		 * it is one of our special test online cases, which is only
2945 		 * attempting to online the device and shouldn't generate an FMA
2946 		 * fault.
2947 		 */
2948 		vd->vdev_state = VDEV_STATE_REMOVED;
2949 		vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
2950 	} else if (state == VDEV_STATE_REMOVED) {
2951 		vd->vdev_removed = B_TRUE;
2952 	} else if (state == VDEV_STATE_CANT_OPEN) {
2953 		/*
2954 		 * If we fail to open a vdev during an import or recovery, we
2955 		 * mark it as "not available", which signifies that it was
2956 		 * never there to begin with.  Failure to open such a device
2957 		 * is not considered an error.
2958 		 */
2959 		if ((spa_load_state(spa) == SPA_LOAD_IMPORT ||
2960 		    spa_load_state(spa) == SPA_LOAD_RECOVER) &&
2961 		    vd->vdev_ops->vdev_op_leaf)
2962 			vd->vdev_not_present = 1;
2963 
2964 		/*
2965 		 * Post the appropriate ereport.  If the 'prevstate' field is
2966 		 * set to something other than VDEV_STATE_UNKNOWN, it indicates
2967 		 * that this is part of a vdev_reopen().  In this case, we don't
2968 		 * want to post the ereport if the device was already in the
2969 		 * CANT_OPEN state beforehand.
2970 		 *
2971 		 * If the 'checkremove' flag is set, then this is an attempt to
2972 		 * online the device in response to an insertion event.  If we
2973 		 * hit this case, then we have detected an insertion event for a
2974 		 * faulted or offline device that wasn't in the removed state.
2975 		 * In this scenario, we don't post an ereport because we are
2976 		 * about to replace the device, or attempt an online with
2977 		 * vdev_forcefault, which will generate the fault for us.
2978 		 */
2979 		if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
2980 		    !vd->vdev_not_present && !vd->vdev_checkremove &&
2981 		    vd != spa->spa_root_vdev) {
2982 			const char *class;
2983 
2984 			switch (aux) {
2985 			case VDEV_AUX_OPEN_FAILED:
2986 				class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
2987 				break;
2988 			case VDEV_AUX_CORRUPT_DATA:
2989 				class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
2990 				break;
2991 			case VDEV_AUX_NO_REPLICAS:
2992 				class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
2993 				break;
2994 			case VDEV_AUX_BAD_GUID_SUM:
2995 				class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
2996 				break;
2997 			case VDEV_AUX_TOO_SMALL:
2998 				class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
2999 				break;
3000 			case VDEV_AUX_BAD_LABEL:
3001 				class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
3002 				break;
3003 			default:
3004 				class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
3005 			}
3006 
3007 			zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
3008 		}
3009 
3010 		/* Erase any notion of persistent removed state */
3011 		vd->vdev_removed = B_FALSE;
3012 	} else {
3013 		vd->vdev_removed = B_FALSE;
3014 	}
3015 
3016 	if (!isopen && vd->vdev_parent)
3017 		vdev_propagate_state(vd->vdev_parent);
3018 }
3019 
3020 /*
3021  * Check the vdev configuration to ensure that it's capable of supporting
3022  * a root pool. Currently, we do not support RAID-Z or partial configuration.
3023  * In addition, only a single top-level vdev is allowed and none of the leaves
3024  * can be wholedisks.
3025  */
3026 boolean_t
3027 vdev_is_bootable(vdev_t *vd)
3028 {
3029 	if (!vd->vdev_ops->vdev_op_leaf) {
3030 		char *vdev_type = vd->vdev_ops->vdev_op_type;
3031 
3032 		if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
3033 		    vd->vdev_children > 1) {
3034 			return (B_FALSE);
3035 		} else if (strcmp(vdev_type, VDEV_TYPE_RAIDZ) == 0 ||
3036 		    strcmp(vdev_type, VDEV_TYPE_MISSING) == 0) {
3037 			return (B_FALSE);
3038 		}
3039 	} else if (vd->vdev_wholedisk == 1) {
3040 		return (B_FALSE);
3041 	}
3042 
3043 	for (int c = 0; c < vd->vdev_children; c++) {
3044 		if (!vdev_is_bootable(vd->vdev_child[c]))
3045 			return (B_FALSE);
3046 	}
3047 	return (B_TRUE);
3048 }
3049 
3050 /*
3051  * Load the state from the original vdev tree (ovd) which
3052  * we've retrieved from the MOS config object. If the original
3053  * vdev was offline or faulted then we transfer that state to the
3054  * device in the current vdev tree (nvd).
3055  */
3056 void
3057 vdev_load_log_state(vdev_t *nvd, vdev_t *ovd)
3058 {
3059 	spa_t *spa = nvd->vdev_spa;
3060 
3061 	ASSERT(nvd->vdev_top->vdev_islog);
3062 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3063 	ASSERT3U(nvd->vdev_guid, ==, ovd->vdev_guid);
3064 
3065 	for (int c = 0; c < nvd->vdev_children; c++)
3066 		vdev_load_log_state(nvd->vdev_child[c], ovd->vdev_child[c]);
3067 
3068 	if (nvd->vdev_ops->vdev_op_leaf) {
3069 		/*
3070 		 * Restore the persistent vdev state
3071 		 */
3072 		nvd->vdev_offline = ovd->vdev_offline;
3073 		nvd->vdev_faulted = ovd->vdev_faulted;
3074 		nvd->vdev_degraded = ovd->vdev_degraded;
3075 		nvd->vdev_removed = ovd->vdev_removed;
3076 	}
3077 }
3078 
3079 /*
3080  * Determine if a log device has valid content.  If the vdev was
3081  * removed or faulted in the MOS config then we know that
3082  * the content on the log device has already been written to the pool.
3083  */
3084 boolean_t
3085 vdev_log_state_valid(vdev_t *vd)
3086 {
3087 	if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted &&
3088 	    !vd->vdev_removed)
3089 		return (B_TRUE);
3090 
3091 	for (int c = 0; c < vd->vdev_children; c++)
3092 		if (vdev_log_state_valid(vd->vdev_child[c]))
3093 			return (B_TRUE);
3094 
3095 	return (B_FALSE);
3096 }
3097 
3098 /*
3099  * Expand a vdev if possible.
3100  */
3101 void
3102 vdev_expand(vdev_t *vd, uint64_t txg)
3103 {
3104 	ASSERT(vd->vdev_top == vd);
3105 	ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
3106 
3107 	if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count) {
3108 		VERIFY(vdev_metaslab_init(vd, txg) == 0);
3109 		vdev_config_dirty(vd);
3110 	}
3111 }
3112 
3113 /*
3114  * Split a vdev.
3115  */
3116 void
3117 vdev_split(vdev_t *vd)
3118 {
3119 	vdev_t *cvd, *pvd = vd->vdev_parent;
3120 
3121 	vdev_remove_child(pvd, vd);
3122 	vdev_compact_children(pvd);
3123 
3124 	cvd = pvd->vdev_child[0];
3125 	if (pvd->vdev_children == 1) {
3126 		vdev_remove_parent(cvd);
3127 		cvd->vdev_splitting = B_TRUE;
3128 	}
3129 	vdev_propagate_state(cvd);
3130 }
3131