xref: /illumos-gate/usr/src/uts/common/fs/zfs/vdev.c (revision 20b5dafb425396adaebd0267d29e1026fc4dc413)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
25  * Copyright 2017 Nexenta Systems, Inc.
26  * Copyright (c) 2014 Integros [integros.com]
27  * Copyright 2016 Toomas Soome <tsoome@me.com>
28  * Copyright 2017 Joyent, Inc.
29  */
30 
31 #include <sys/zfs_context.h>
32 #include <sys/fm/fs/zfs.h>
33 #include <sys/spa.h>
34 #include <sys/spa_impl.h>
35 #include <sys/bpobj.h>
36 #include <sys/dmu.h>
37 #include <sys/dmu_tx.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/vdev_impl.h>
40 #include <sys/uberblock_impl.h>
41 #include <sys/metaslab.h>
42 #include <sys/metaslab_impl.h>
43 #include <sys/space_map.h>
44 #include <sys/space_reftree.h>
45 #include <sys/zio.h>
46 #include <sys/zap.h>
47 #include <sys/fs/zfs.h>
48 #include <sys/arc.h>
49 #include <sys/zil.h>
50 #include <sys/dsl_scan.h>
51 #include <sys/abd.h>
52 #include <sys/vdev_initialize.h>
53 
54 /*
55  * Virtual device management.
56  */
57 
58 static vdev_ops_t *vdev_ops_table[] = {
59 	&vdev_root_ops,
60 	&vdev_raidz_ops,
61 	&vdev_mirror_ops,
62 	&vdev_replacing_ops,
63 	&vdev_spare_ops,
64 	&vdev_disk_ops,
65 	&vdev_file_ops,
66 	&vdev_missing_ops,
67 	&vdev_hole_ops,
68 	&vdev_indirect_ops,
69 	NULL
70 };
71 
72 /* maximum scrub/resilver I/O queue per leaf vdev */
73 int zfs_scrub_limit = 10;
74 
75 /* maximum number of metaslabs per top-level vdev */
76 int vdev_max_ms_count = 200;
77 
78 /* minimum amount of metaslabs per top-level vdev */
79 int vdev_min_ms_count = 16;
80 
81 /* see comment in vdev_metaslab_set_size() */
82 int vdev_default_ms_shift = 29;
83 
84 boolean_t vdev_validate_skip = B_FALSE;
85 
86 /*
87  * Since the DTL space map of a vdev is not expected to have a lot of
88  * entries, we default its block size to 4K.
89  */
90 int vdev_dtl_sm_blksz = (1 << 12);
91 
92 /*
93  * vdev-wide space maps that have lots of entries written to them at
94  * the end of each transaction can benefit from a higher I/O bandwidth
95  * (e.g. vdev_obsolete_sm), thus we default their block size to 128K.
96  */
97 int vdev_standard_sm_blksz = (1 << 17);
98 
99 int zfs_ashift_min;
100 
101 /*PRINTFLIKE2*/
102 void
103 vdev_dbgmsg(vdev_t *vd, const char *fmt, ...)
104 {
105 	va_list adx;
106 	char buf[256];
107 
108 	va_start(adx, fmt);
109 	(void) vsnprintf(buf, sizeof (buf), fmt, adx);
110 	va_end(adx);
111 
112 	if (vd->vdev_path != NULL) {
113 		zfs_dbgmsg("%s vdev '%s': %s", vd->vdev_ops->vdev_op_type,
114 		    vd->vdev_path, buf);
115 	} else {
116 		zfs_dbgmsg("%s-%llu vdev (guid %llu): %s",
117 		    vd->vdev_ops->vdev_op_type,
118 		    (u_longlong_t)vd->vdev_id,
119 		    (u_longlong_t)vd->vdev_guid, buf);
120 	}
121 }
122 
123 void
124 vdev_dbgmsg_print_tree(vdev_t *vd, int indent)
125 {
126 	char state[20];
127 
128 	if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops) {
129 		zfs_dbgmsg("%*svdev %u: %s", indent, "", vd->vdev_id,
130 		    vd->vdev_ops->vdev_op_type);
131 		return;
132 	}
133 
134 	switch (vd->vdev_state) {
135 	case VDEV_STATE_UNKNOWN:
136 		(void) snprintf(state, sizeof (state), "unknown");
137 		break;
138 	case VDEV_STATE_CLOSED:
139 		(void) snprintf(state, sizeof (state), "closed");
140 		break;
141 	case VDEV_STATE_OFFLINE:
142 		(void) snprintf(state, sizeof (state), "offline");
143 		break;
144 	case VDEV_STATE_REMOVED:
145 		(void) snprintf(state, sizeof (state), "removed");
146 		break;
147 	case VDEV_STATE_CANT_OPEN:
148 		(void) snprintf(state, sizeof (state), "can't open");
149 		break;
150 	case VDEV_STATE_FAULTED:
151 		(void) snprintf(state, sizeof (state), "faulted");
152 		break;
153 	case VDEV_STATE_DEGRADED:
154 		(void) snprintf(state, sizeof (state), "degraded");
155 		break;
156 	case VDEV_STATE_HEALTHY:
157 		(void) snprintf(state, sizeof (state), "healthy");
158 		break;
159 	default:
160 		(void) snprintf(state, sizeof (state), "<state %u>",
161 		    (uint_t)vd->vdev_state);
162 	}
163 
164 	zfs_dbgmsg("%*svdev %u: %s%s, guid: %llu, path: %s, %s", indent,
165 	    "", vd->vdev_id, vd->vdev_ops->vdev_op_type,
166 	    vd->vdev_islog ? " (log)" : "",
167 	    (u_longlong_t)vd->vdev_guid,
168 	    vd->vdev_path ? vd->vdev_path : "N/A", state);
169 
170 	for (uint64_t i = 0; i < vd->vdev_children; i++)
171 		vdev_dbgmsg_print_tree(vd->vdev_child[i], indent + 2);
172 }
173 
174 /*
175  * Given a vdev type, return the appropriate ops vector.
176  */
177 static vdev_ops_t *
178 vdev_getops(const char *type)
179 {
180 	vdev_ops_t *ops, **opspp;
181 
182 	for (opspp = vdev_ops_table; (ops = *opspp) != NULL; opspp++)
183 		if (strcmp(ops->vdev_op_type, type) == 0)
184 			break;
185 
186 	return (ops);
187 }
188 
189 /* ARGSUSED */
190 void
191 vdev_default_xlate(vdev_t *vd, const range_seg_t *in, range_seg_t *res)
192 {
193 	res->rs_start = in->rs_start;
194 	res->rs_end = in->rs_end;
195 }
196 
197 /*
198  * Default asize function: return the MAX of psize with the asize of
199  * all children.  This is what's used by anything other than RAID-Z.
200  */
201 uint64_t
202 vdev_default_asize(vdev_t *vd, uint64_t psize)
203 {
204 	uint64_t asize = P2ROUNDUP(psize, 1ULL << vd->vdev_top->vdev_ashift);
205 	uint64_t csize;
206 
207 	for (int c = 0; c < vd->vdev_children; c++) {
208 		csize = vdev_psize_to_asize(vd->vdev_child[c], psize);
209 		asize = MAX(asize, csize);
210 	}
211 
212 	return (asize);
213 }
214 
215 /*
216  * Get the minimum allocatable size. We define the allocatable size as
217  * the vdev's asize rounded to the nearest metaslab. This allows us to
218  * replace or attach devices which don't have the same physical size but
219  * can still satisfy the same number of allocations.
220  */
221 uint64_t
222 vdev_get_min_asize(vdev_t *vd)
223 {
224 	vdev_t *pvd = vd->vdev_parent;
225 
226 	/*
227 	 * If our parent is NULL (inactive spare or cache) or is the root,
228 	 * just return our own asize.
229 	 */
230 	if (pvd == NULL)
231 		return (vd->vdev_asize);
232 
233 	/*
234 	 * The top-level vdev just returns the allocatable size rounded
235 	 * to the nearest metaslab.
236 	 */
237 	if (vd == vd->vdev_top)
238 		return (P2ALIGN(vd->vdev_asize, 1ULL << vd->vdev_ms_shift));
239 
240 	/*
241 	 * The allocatable space for a raidz vdev is N * sizeof(smallest child),
242 	 * so each child must provide at least 1/Nth of its asize.
243 	 */
244 	if (pvd->vdev_ops == &vdev_raidz_ops)
245 		return ((pvd->vdev_min_asize + pvd->vdev_children - 1) /
246 		    pvd->vdev_children);
247 
248 	return (pvd->vdev_min_asize);
249 }
250 
251 void
252 vdev_set_min_asize(vdev_t *vd)
253 {
254 	vd->vdev_min_asize = vdev_get_min_asize(vd);
255 
256 	for (int c = 0; c < vd->vdev_children; c++)
257 		vdev_set_min_asize(vd->vdev_child[c]);
258 }
259 
260 vdev_t *
261 vdev_lookup_top(spa_t *spa, uint64_t vdev)
262 {
263 	vdev_t *rvd = spa->spa_root_vdev;
264 
265 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
266 
267 	if (vdev < rvd->vdev_children) {
268 		ASSERT(rvd->vdev_child[vdev] != NULL);
269 		return (rvd->vdev_child[vdev]);
270 	}
271 
272 	return (NULL);
273 }
274 
275 vdev_t *
276 vdev_lookup_by_guid(vdev_t *vd, uint64_t guid)
277 {
278 	vdev_t *mvd;
279 
280 	if (vd->vdev_guid == guid)
281 		return (vd);
282 
283 	for (int c = 0; c < vd->vdev_children; c++)
284 		if ((mvd = vdev_lookup_by_guid(vd->vdev_child[c], guid)) !=
285 		    NULL)
286 			return (mvd);
287 
288 	return (NULL);
289 }
290 
291 static int
292 vdev_count_leaves_impl(vdev_t *vd)
293 {
294 	int n = 0;
295 
296 	if (vd->vdev_ops->vdev_op_leaf)
297 		return (1);
298 
299 	for (int c = 0; c < vd->vdev_children; c++)
300 		n += vdev_count_leaves_impl(vd->vdev_child[c]);
301 
302 	return (n);
303 }
304 
305 int
306 vdev_count_leaves(spa_t *spa)
307 {
308 	return (vdev_count_leaves_impl(spa->spa_root_vdev));
309 }
310 
311 void
312 vdev_add_child(vdev_t *pvd, vdev_t *cvd)
313 {
314 	size_t oldsize, newsize;
315 	uint64_t id = cvd->vdev_id;
316 	vdev_t **newchild;
317 	spa_t *spa = cvd->vdev_spa;
318 
319 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
320 	ASSERT(cvd->vdev_parent == NULL);
321 
322 	cvd->vdev_parent = pvd;
323 
324 	if (pvd == NULL)
325 		return;
326 
327 	ASSERT(id >= pvd->vdev_children || pvd->vdev_child[id] == NULL);
328 
329 	oldsize = pvd->vdev_children * sizeof (vdev_t *);
330 	pvd->vdev_children = MAX(pvd->vdev_children, id + 1);
331 	newsize = pvd->vdev_children * sizeof (vdev_t *);
332 
333 	newchild = kmem_zalloc(newsize, KM_SLEEP);
334 	if (pvd->vdev_child != NULL) {
335 		bcopy(pvd->vdev_child, newchild, oldsize);
336 		kmem_free(pvd->vdev_child, oldsize);
337 	}
338 
339 	pvd->vdev_child = newchild;
340 	pvd->vdev_child[id] = cvd;
341 
342 	cvd->vdev_top = (pvd->vdev_top ? pvd->vdev_top: cvd);
343 	ASSERT(cvd->vdev_top->vdev_parent->vdev_parent == NULL);
344 
345 	/*
346 	 * Walk up all ancestors to update guid sum.
347 	 */
348 	for (; pvd != NULL; pvd = pvd->vdev_parent)
349 		pvd->vdev_guid_sum += cvd->vdev_guid_sum;
350 }
351 
352 void
353 vdev_remove_child(vdev_t *pvd, vdev_t *cvd)
354 {
355 	int c;
356 	uint_t id = cvd->vdev_id;
357 
358 	ASSERT(cvd->vdev_parent == pvd);
359 
360 	if (pvd == NULL)
361 		return;
362 
363 	ASSERT(id < pvd->vdev_children);
364 	ASSERT(pvd->vdev_child[id] == cvd);
365 
366 	pvd->vdev_child[id] = NULL;
367 	cvd->vdev_parent = NULL;
368 
369 	for (c = 0; c < pvd->vdev_children; c++)
370 		if (pvd->vdev_child[c])
371 			break;
372 
373 	if (c == pvd->vdev_children) {
374 		kmem_free(pvd->vdev_child, c * sizeof (vdev_t *));
375 		pvd->vdev_child = NULL;
376 		pvd->vdev_children = 0;
377 	}
378 
379 	/*
380 	 * Walk up all ancestors to update guid sum.
381 	 */
382 	for (; pvd != NULL; pvd = pvd->vdev_parent)
383 		pvd->vdev_guid_sum -= cvd->vdev_guid_sum;
384 }
385 
386 /*
387  * Remove any holes in the child array.
388  */
389 void
390 vdev_compact_children(vdev_t *pvd)
391 {
392 	vdev_t **newchild, *cvd;
393 	int oldc = pvd->vdev_children;
394 	int newc;
395 
396 	ASSERT(spa_config_held(pvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
397 
398 	for (int c = newc = 0; c < oldc; c++)
399 		if (pvd->vdev_child[c])
400 			newc++;
401 
402 	newchild = kmem_alloc(newc * sizeof (vdev_t *), KM_SLEEP);
403 
404 	for (int c = newc = 0; c < oldc; c++) {
405 		if ((cvd = pvd->vdev_child[c]) != NULL) {
406 			newchild[newc] = cvd;
407 			cvd->vdev_id = newc++;
408 		}
409 	}
410 
411 	kmem_free(pvd->vdev_child, oldc * sizeof (vdev_t *));
412 	pvd->vdev_child = newchild;
413 	pvd->vdev_children = newc;
414 }
415 
416 /*
417  * Allocate and minimally initialize a vdev_t.
418  */
419 vdev_t *
420 vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
421 {
422 	vdev_t *vd;
423 	vdev_indirect_config_t *vic;
424 
425 	vd = kmem_zalloc(sizeof (vdev_t), KM_SLEEP);
426 	vic = &vd->vdev_indirect_config;
427 
428 	if (spa->spa_root_vdev == NULL) {
429 		ASSERT(ops == &vdev_root_ops);
430 		spa->spa_root_vdev = vd;
431 		spa->spa_load_guid = spa_generate_guid(NULL);
432 	}
433 
434 	if (guid == 0 && ops != &vdev_hole_ops) {
435 		if (spa->spa_root_vdev == vd) {
436 			/*
437 			 * The root vdev's guid will also be the pool guid,
438 			 * which must be unique among all pools.
439 			 */
440 			guid = spa_generate_guid(NULL);
441 		} else {
442 			/*
443 			 * Any other vdev's guid must be unique within the pool.
444 			 */
445 			guid = spa_generate_guid(spa);
446 		}
447 		ASSERT(!spa_guid_exists(spa_guid(spa), guid));
448 	}
449 
450 	vd->vdev_spa = spa;
451 	vd->vdev_id = id;
452 	vd->vdev_guid = guid;
453 	vd->vdev_guid_sum = guid;
454 	vd->vdev_ops = ops;
455 	vd->vdev_state = VDEV_STATE_CLOSED;
456 	vd->vdev_ishole = (ops == &vdev_hole_ops);
457 	vic->vic_prev_indirect_vdev = UINT64_MAX;
458 
459 	rw_init(&vd->vdev_indirect_rwlock, NULL, RW_DEFAULT, NULL);
460 	mutex_init(&vd->vdev_obsolete_lock, NULL, MUTEX_DEFAULT, NULL);
461 	vd->vdev_obsolete_segments = range_tree_create(NULL, NULL);
462 
463 	mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_DEFAULT, NULL);
464 	mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL);
465 	mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL);
466 	mutex_init(&vd->vdev_queue_lock, NULL, MUTEX_DEFAULT, NULL);
467 	mutex_init(&vd->vdev_initialize_lock, NULL, MUTEX_DEFAULT, NULL);
468 	mutex_init(&vd->vdev_initialize_io_lock, NULL, MUTEX_DEFAULT, NULL);
469 	cv_init(&vd->vdev_initialize_cv, NULL, CV_DEFAULT, NULL);
470 	cv_init(&vd->vdev_initialize_io_cv, NULL, CV_DEFAULT, NULL);
471 
472 	for (int t = 0; t < DTL_TYPES; t++) {
473 		vd->vdev_dtl[t] = range_tree_create(NULL, NULL);
474 	}
475 	txg_list_create(&vd->vdev_ms_list, spa,
476 	    offsetof(struct metaslab, ms_txg_node));
477 	txg_list_create(&vd->vdev_dtl_list, spa,
478 	    offsetof(struct vdev, vdev_dtl_node));
479 	vd->vdev_stat.vs_timestamp = gethrtime();
480 	vdev_queue_init(vd);
481 	vdev_cache_init(vd);
482 
483 	return (vd);
484 }
485 
486 /*
487  * Allocate a new vdev.  The 'alloctype' is used to control whether we are
488  * creating a new vdev or loading an existing one - the behavior is slightly
489  * different for each case.
490  */
491 int
492 vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id,
493     int alloctype)
494 {
495 	vdev_ops_t *ops;
496 	char *type;
497 	uint64_t guid = 0, islog, nparity;
498 	vdev_t *vd;
499 	vdev_indirect_config_t *vic;
500 
501 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
502 
503 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) != 0)
504 		return (SET_ERROR(EINVAL));
505 
506 	if ((ops = vdev_getops(type)) == NULL)
507 		return (SET_ERROR(EINVAL));
508 
509 	/*
510 	 * If this is a load, get the vdev guid from the nvlist.
511 	 * Otherwise, vdev_alloc_common() will generate one for us.
512 	 */
513 	if (alloctype == VDEV_ALLOC_LOAD) {
514 		uint64_t label_id;
515 
516 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ID, &label_id) ||
517 		    label_id != id)
518 			return (SET_ERROR(EINVAL));
519 
520 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
521 			return (SET_ERROR(EINVAL));
522 	} else if (alloctype == VDEV_ALLOC_SPARE) {
523 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
524 			return (SET_ERROR(EINVAL));
525 	} else if (alloctype == VDEV_ALLOC_L2CACHE) {
526 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
527 			return (SET_ERROR(EINVAL));
528 	} else if (alloctype == VDEV_ALLOC_ROOTPOOL) {
529 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
530 			return (SET_ERROR(EINVAL));
531 	}
532 
533 	/*
534 	 * The first allocated vdev must be of type 'root'.
535 	 */
536 	if (ops != &vdev_root_ops && spa->spa_root_vdev == NULL)
537 		return (SET_ERROR(EINVAL));
538 
539 	/*
540 	 * Determine whether we're a log vdev.
541 	 */
542 	islog = 0;
543 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &islog);
544 	if (islog && spa_version(spa) < SPA_VERSION_SLOGS)
545 		return (SET_ERROR(ENOTSUP));
546 
547 	if (ops == &vdev_hole_ops && spa_version(spa) < SPA_VERSION_HOLES)
548 		return (SET_ERROR(ENOTSUP));
549 
550 	/*
551 	 * Set the nparity property for RAID-Z vdevs.
552 	 */
553 	nparity = -1ULL;
554 	if (ops == &vdev_raidz_ops) {
555 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY,
556 		    &nparity) == 0) {
557 			if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY)
558 				return (SET_ERROR(EINVAL));
559 			/*
560 			 * Previous versions could only support 1 or 2 parity
561 			 * device.
562 			 */
563 			if (nparity > 1 &&
564 			    spa_version(spa) < SPA_VERSION_RAIDZ2)
565 				return (SET_ERROR(ENOTSUP));
566 			if (nparity > 2 &&
567 			    spa_version(spa) < SPA_VERSION_RAIDZ3)
568 				return (SET_ERROR(ENOTSUP));
569 		} else {
570 			/*
571 			 * We require the parity to be specified for SPAs that
572 			 * support multiple parity levels.
573 			 */
574 			if (spa_version(spa) >= SPA_VERSION_RAIDZ2)
575 				return (SET_ERROR(EINVAL));
576 			/*
577 			 * Otherwise, we default to 1 parity device for RAID-Z.
578 			 */
579 			nparity = 1;
580 		}
581 	} else {
582 		nparity = 0;
583 	}
584 	ASSERT(nparity != -1ULL);
585 
586 	vd = vdev_alloc_common(spa, id, guid, ops);
587 	vic = &vd->vdev_indirect_config;
588 
589 	vd->vdev_islog = islog;
590 	vd->vdev_nparity = nparity;
591 
592 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &vd->vdev_path) == 0)
593 		vd->vdev_path = spa_strdup(vd->vdev_path);
594 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &vd->vdev_devid) == 0)
595 		vd->vdev_devid = spa_strdup(vd->vdev_devid);
596 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH,
597 	    &vd->vdev_physpath) == 0)
598 		vd->vdev_physpath = spa_strdup(vd->vdev_physpath);
599 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &vd->vdev_fru) == 0)
600 		vd->vdev_fru = spa_strdup(vd->vdev_fru);
601 
602 	/*
603 	 * Set the whole_disk property.  If it's not specified, leave the value
604 	 * as -1.
605 	 */
606 	if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
607 	    &vd->vdev_wholedisk) != 0)
608 		vd->vdev_wholedisk = -1ULL;
609 
610 	ASSERT0(vic->vic_mapping_object);
611 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_INDIRECT_OBJECT,
612 	    &vic->vic_mapping_object);
613 	ASSERT0(vic->vic_births_object);
614 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_INDIRECT_BIRTHS,
615 	    &vic->vic_births_object);
616 	ASSERT3U(vic->vic_prev_indirect_vdev, ==, UINT64_MAX);
617 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_PREV_INDIRECT_VDEV,
618 	    &vic->vic_prev_indirect_vdev);
619 
620 	/*
621 	 * Look for the 'not present' flag.  This will only be set if the device
622 	 * was not present at the time of import.
623 	 */
624 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT,
625 	    &vd->vdev_not_present);
626 
627 	/*
628 	 * Get the alignment requirement.
629 	 */
630 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT, &vd->vdev_ashift);
631 
632 	/*
633 	 * Retrieve the vdev creation time.
634 	 */
635 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_CREATE_TXG,
636 	    &vd->vdev_crtxg);
637 
638 	/*
639 	 * If we're a top-level vdev, try to load the allocation parameters.
640 	 */
641 	if (parent && !parent->vdev_parent &&
642 	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
643 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
644 		    &vd->vdev_ms_array);
645 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
646 		    &vd->vdev_ms_shift);
647 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE,
648 		    &vd->vdev_asize);
649 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVING,
650 		    &vd->vdev_removing);
651 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_VDEV_TOP_ZAP,
652 		    &vd->vdev_top_zap);
653 	} else {
654 		ASSERT0(vd->vdev_top_zap);
655 	}
656 
657 	if (parent && !parent->vdev_parent && alloctype != VDEV_ALLOC_ATTACH) {
658 		ASSERT(alloctype == VDEV_ALLOC_LOAD ||
659 		    alloctype == VDEV_ALLOC_ADD ||
660 		    alloctype == VDEV_ALLOC_SPLIT ||
661 		    alloctype == VDEV_ALLOC_ROOTPOOL);
662 		vd->vdev_mg = metaslab_group_create(islog ?
663 		    spa_log_class(spa) : spa_normal_class(spa), vd,
664 		    spa->spa_alloc_count);
665 	}
666 
667 	if (vd->vdev_ops->vdev_op_leaf &&
668 	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
669 		(void) nvlist_lookup_uint64(nv,
670 		    ZPOOL_CONFIG_VDEV_LEAF_ZAP, &vd->vdev_leaf_zap);
671 	} else {
672 		ASSERT0(vd->vdev_leaf_zap);
673 	}
674 
675 	/*
676 	 * If we're a leaf vdev, try to load the DTL object and other state.
677 	 */
678 
679 	if (vd->vdev_ops->vdev_op_leaf &&
680 	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE ||
681 	    alloctype == VDEV_ALLOC_ROOTPOOL)) {
682 		if (alloctype == VDEV_ALLOC_LOAD) {
683 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL,
684 			    &vd->vdev_dtl_object);
685 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE,
686 			    &vd->vdev_unspare);
687 		}
688 
689 		if (alloctype == VDEV_ALLOC_ROOTPOOL) {
690 			uint64_t spare = 0;
691 
692 			if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE,
693 			    &spare) == 0 && spare)
694 				spa_spare_add(vd);
695 		}
696 
697 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE,
698 		    &vd->vdev_offline);
699 
700 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_RESILVER_TXG,
701 		    &vd->vdev_resilver_txg);
702 
703 		/*
704 		 * When importing a pool, we want to ignore the persistent fault
705 		 * state, as the diagnosis made on another system may not be
706 		 * valid in the current context.  Local vdevs will
707 		 * remain in the faulted state.
708 		 */
709 		if (spa_load_state(spa) == SPA_LOAD_OPEN) {
710 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED,
711 			    &vd->vdev_faulted);
712 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED,
713 			    &vd->vdev_degraded);
714 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED,
715 			    &vd->vdev_removed);
716 
717 			if (vd->vdev_faulted || vd->vdev_degraded) {
718 				char *aux;
719 
720 				vd->vdev_label_aux =
721 				    VDEV_AUX_ERR_EXCEEDED;
722 				if (nvlist_lookup_string(nv,
723 				    ZPOOL_CONFIG_AUX_STATE, &aux) == 0 &&
724 				    strcmp(aux, "external") == 0)
725 					vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
726 			}
727 		}
728 	}
729 
730 	/*
731 	 * Add ourselves to the parent's list of children.
732 	 */
733 	vdev_add_child(parent, vd);
734 
735 	*vdp = vd;
736 
737 	return (0);
738 }
739 
740 void
741 vdev_free(vdev_t *vd)
742 {
743 	spa_t *spa = vd->vdev_spa;
744 	ASSERT3P(vd->vdev_initialize_thread, ==, NULL);
745 
746 	/*
747 	 * vdev_free() implies closing the vdev first.  This is simpler than
748 	 * trying to ensure complicated semantics for all callers.
749 	 */
750 	vdev_close(vd);
751 
752 	ASSERT(!list_link_active(&vd->vdev_config_dirty_node));
753 	ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
754 
755 	/*
756 	 * Free all children.
757 	 */
758 	for (int c = 0; c < vd->vdev_children; c++)
759 		vdev_free(vd->vdev_child[c]);
760 
761 	ASSERT(vd->vdev_child == NULL);
762 	ASSERT(vd->vdev_guid_sum == vd->vdev_guid);
763 	ASSERT(vd->vdev_initialize_thread == NULL);
764 
765 	/*
766 	 * Discard allocation state.
767 	 */
768 	if (vd->vdev_mg != NULL) {
769 		vdev_metaslab_fini(vd);
770 		metaslab_group_destroy(vd->vdev_mg);
771 	}
772 
773 	ASSERT0(vd->vdev_stat.vs_space);
774 	ASSERT0(vd->vdev_stat.vs_dspace);
775 	ASSERT0(vd->vdev_stat.vs_alloc);
776 
777 	/*
778 	 * Remove this vdev from its parent's child list.
779 	 */
780 	vdev_remove_child(vd->vdev_parent, vd);
781 
782 	ASSERT(vd->vdev_parent == NULL);
783 
784 	/*
785 	 * Clean up vdev structure.
786 	 */
787 	vdev_queue_fini(vd);
788 	vdev_cache_fini(vd);
789 
790 	if (vd->vdev_path)
791 		spa_strfree(vd->vdev_path);
792 	if (vd->vdev_devid)
793 		spa_strfree(vd->vdev_devid);
794 	if (vd->vdev_physpath)
795 		spa_strfree(vd->vdev_physpath);
796 	if (vd->vdev_fru)
797 		spa_strfree(vd->vdev_fru);
798 
799 	if (vd->vdev_isspare)
800 		spa_spare_remove(vd);
801 	if (vd->vdev_isl2cache)
802 		spa_l2cache_remove(vd);
803 
804 	txg_list_destroy(&vd->vdev_ms_list);
805 	txg_list_destroy(&vd->vdev_dtl_list);
806 
807 	mutex_enter(&vd->vdev_dtl_lock);
808 	space_map_close(vd->vdev_dtl_sm);
809 	for (int t = 0; t < DTL_TYPES; t++) {
810 		range_tree_vacate(vd->vdev_dtl[t], NULL, NULL);
811 		range_tree_destroy(vd->vdev_dtl[t]);
812 	}
813 	mutex_exit(&vd->vdev_dtl_lock);
814 
815 	EQUIV(vd->vdev_indirect_births != NULL,
816 	    vd->vdev_indirect_mapping != NULL);
817 	if (vd->vdev_indirect_births != NULL) {
818 		vdev_indirect_mapping_close(vd->vdev_indirect_mapping);
819 		vdev_indirect_births_close(vd->vdev_indirect_births);
820 	}
821 
822 	if (vd->vdev_obsolete_sm != NULL) {
823 		ASSERT(vd->vdev_removing ||
824 		    vd->vdev_ops == &vdev_indirect_ops);
825 		space_map_close(vd->vdev_obsolete_sm);
826 		vd->vdev_obsolete_sm = NULL;
827 	}
828 	range_tree_destroy(vd->vdev_obsolete_segments);
829 	rw_destroy(&vd->vdev_indirect_rwlock);
830 	mutex_destroy(&vd->vdev_obsolete_lock);
831 
832 	mutex_destroy(&vd->vdev_queue_lock);
833 	mutex_destroy(&vd->vdev_dtl_lock);
834 	mutex_destroy(&vd->vdev_stat_lock);
835 	mutex_destroy(&vd->vdev_probe_lock);
836 	mutex_destroy(&vd->vdev_initialize_lock);
837 	mutex_destroy(&vd->vdev_initialize_io_lock);
838 	cv_destroy(&vd->vdev_initialize_io_cv);
839 	cv_destroy(&vd->vdev_initialize_cv);
840 
841 	if (vd == spa->spa_root_vdev)
842 		spa->spa_root_vdev = NULL;
843 
844 	kmem_free(vd, sizeof (vdev_t));
845 }
846 
847 /*
848  * Transfer top-level vdev state from svd to tvd.
849  */
850 static void
851 vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
852 {
853 	spa_t *spa = svd->vdev_spa;
854 	metaslab_t *msp;
855 	vdev_t *vd;
856 	int t;
857 
858 	ASSERT(tvd == tvd->vdev_top);
859 
860 	tvd->vdev_ms_array = svd->vdev_ms_array;
861 	tvd->vdev_ms_shift = svd->vdev_ms_shift;
862 	tvd->vdev_ms_count = svd->vdev_ms_count;
863 	tvd->vdev_top_zap = svd->vdev_top_zap;
864 
865 	svd->vdev_ms_array = 0;
866 	svd->vdev_ms_shift = 0;
867 	svd->vdev_ms_count = 0;
868 	svd->vdev_top_zap = 0;
869 
870 	if (tvd->vdev_mg)
871 		ASSERT3P(tvd->vdev_mg, ==, svd->vdev_mg);
872 	tvd->vdev_mg = svd->vdev_mg;
873 	tvd->vdev_ms = svd->vdev_ms;
874 
875 	svd->vdev_mg = NULL;
876 	svd->vdev_ms = NULL;
877 
878 	if (tvd->vdev_mg != NULL)
879 		tvd->vdev_mg->mg_vd = tvd;
880 
881 	tvd->vdev_checkpoint_sm = svd->vdev_checkpoint_sm;
882 	svd->vdev_checkpoint_sm = NULL;
883 
884 	tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
885 	tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
886 	tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
887 
888 	svd->vdev_stat.vs_alloc = 0;
889 	svd->vdev_stat.vs_space = 0;
890 	svd->vdev_stat.vs_dspace = 0;
891 
892 	/*
893 	 * State which may be set on a top-level vdev that's in the
894 	 * process of being removed.
895 	 */
896 	ASSERT0(tvd->vdev_indirect_config.vic_births_object);
897 	ASSERT0(tvd->vdev_indirect_config.vic_mapping_object);
898 	ASSERT3U(tvd->vdev_indirect_config.vic_prev_indirect_vdev, ==, -1ULL);
899 	ASSERT3P(tvd->vdev_indirect_mapping, ==, NULL);
900 	ASSERT3P(tvd->vdev_indirect_births, ==, NULL);
901 	ASSERT3P(tvd->vdev_obsolete_sm, ==, NULL);
902 	ASSERT0(tvd->vdev_removing);
903 	tvd->vdev_removing = svd->vdev_removing;
904 	tvd->vdev_indirect_config = svd->vdev_indirect_config;
905 	tvd->vdev_indirect_mapping = svd->vdev_indirect_mapping;
906 	tvd->vdev_indirect_births = svd->vdev_indirect_births;
907 	range_tree_swap(&svd->vdev_obsolete_segments,
908 	    &tvd->vdev_obsolete_segments);
909 	tvd->vdev_obsolete_sm = svd->vdev_obsolete_sm;
910 	svd->vdev_indirect_config.vic_mapping_object = 0;
911 	svd->vdev_indirect_config.vic_births_object = 0;
912 	svd->vdev_indirect_config.vic_prev_indirect_vdev = -1ULL;
913 	svd->vdev_indirect_mapping = NULL;
914 	svd->vdev_indirect_births = NULL;
915 	svd->vdev_obsolete_sm = NULL;
916 	svd->vdev_removing = 0;
917 
918 	for (t = 0; t < TXG_SIZE; t++) {
919 		while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
920 			(void) txg_list_add(&tvd->vdev_ms_list, msp, t);
921 		while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
922 			(void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
923 		if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
924 			(void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
925 	}
926 
927 	if (list_link_active(&svd->vdev_config_dirty_node)) {
928 		vdev_config_clean(svd);
929 		vdev_config_dirty(tvd);
930 	}
931 
932 	if (list_link_active(&svd->vdev_state_dirty_node)) {
933 		vdev_state_clean(svd);
934 		vdev_state_dirty(tvd);
935 	}
936 
937 	tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
938 	svd->vdev_deflate_ratio = 0;
939 
940 	tvd->vdev_islog = svd->vdev_islog;
941 	svd->vdev_islog = 0;
942 }
943 
944 static void
945 vdev_top_update(vdev_t *tvd, vdev_t *vd)
946 {
947 	if (vd == NULL)
948 		return;
949 
950 	vd->vdev_top = tvd;
951 
952 	for (int c = 0; c < vd->vdev_children; c++)
953 		vdev_top_update(tvd, vd->vdev_child[c]);
954 }
955 
956 /*
957  * Add a mirror/replacing vdev above an existing vdev.
958  */
959 vdev_t *
960 vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
961 {
962 	spa_t *spa = cvd->vdev_spa;
963 	vdev_t *pvd = cvd->vdev_parent;
964 	vdev_t *mvd;
965 
966 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
967 
968 	mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
969 
970 	mvd->vdev_asize = cvd->vdev_asize;
971 	mvd->vdev_min_asize = cvd->vdev_min_asize;
972 	mvd->vdev_max_asize = cvd->vdev_max_asize;
973 	mvd->vdev_psize = cvd->vdev_psize;
974 	mvd->vdev_ashift = cvd->vdev_ashift;
975 	mvd->vdev_state = cvd->vdev_state;
976 	mvd->vdev_crtxg = cvd->vdev_crtxg;
977 
978 	vdev_remove_child(pvd, cvd);
979 	vdev_add_child(pvd, mvd);
980 	cvd->vdev_id = mvd->vdev_children;
981 	vdev_add_child(mvd, cvd);
982 	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
983 
984 	if (mvd == mvd->vdev_top)
985 		vdev_top_transfer(cvd, mvd);
986 
987 	return (mvd);
988 }
989 
990 /*
991  * Remove a 1-way mirror/replacing vdev from the tree.
992  */
993 void
994 vdev_remove_parent(vdev_t *cvd)
995 {
996 	vdev_t *mvd = cvd->vdev_parent;
997 	vdev_t *pvd = mvd->vdev_parent;
998 
999 	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1000 
1001 	ASSERT(mvd->vdev_children == 1);
1002 	ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
1003 	    mvd->vdev_ops == &vdev_replacing_ops ||
1004 	    mvd->vdev_ops == &vdev_spare_ops);
1005 	cvd->vdev_ashift = mvd->vdev_ashift;
1006 
1007 	vdev_remove_child(mvd, cvd);
1008 	vdev_remove_child(pvd, mvd);
1009 
1010 	/*
1011 	 * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
1012 	 * Otherwise, we could have detached an offline device, and when we
1013 	 * go to import the pool we'll think we have two top-level vdevs,
1014 	 * instead of a different version of the same top-level vdev.
1015 	 */
1016 	if (mvd->vdev_top == mvd) {
1017 		uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
1018 		cvd->vdev_orig_guid = cvd->vdev_guid;
1019 		cvd->vdev_guid += guid_delta;
1020 		cvd->vdev_guid_sum += guid_delta;
1021 	}
1022 	cvd->vdev_id = mvd->vdev_id;
1023 	vdev_add_child(pvd, cvd);
1024 	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
1025 
1026 	if (cvd == cvd->vdev_top)
1027 		vdev_top_transfer(mvd, cvd);
1028 
1029 	ASSERT(mvd->vdev_children == 0);
1030 	vdev_free(mvd);
1031 }
1032 
1033 int
1034 vdev_metaslab_init(vdev_t *vd, uint64_t txg)
1035 {
1036 	spa_t *spa = vd->vdev_spa;
1037 	objset_t *mos = spa->spa_meta_objset;
1038 	uint64_t m;
1039 	uint64_t oldc = vd->vdev_ms_count;
1040 	uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
1041 	metaslab_t **mspp;
1042 	int error;
1043 
1044 	ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER));
1045 
1046 	/*
1047 	 * This vdev is not being allocated from yet or is a hole.
1048 	 */
1049 	if (vd->vdev_ms_shift == 0)
1050 		return (0);
1051 
1052 	ASSERT(!vd->vdev_ishole);
1053 
1054 	ASSERT(oldc <= newc);
1055 
1056 	mspp = kmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
1057 
1058 	if (oldc != 0) {
1059 		bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp));
1060 		kmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
1061 	}
1062 
1063 	vd->vdev_ms = mspp;
1064 	vd->vdev_ms_count = newc;
1065 	for (m = oldc; m < newc; m++) {
1066 		uint64_t object = 0;
1067 
1068 		/*
1069 		 * vdev_ms_array may be 0 if we are creating the "fake"
1070 		 * metaslabs for an indirect vdev for zdb's leak detection.
1071 		 * See zdb_leak_init().
1072 		 */
1073 		if (txg == 0 && vd->vdev_ms_array != 0) {
1074 			error = dmu_read(mos, vd->vdev_ms_array,
1075 			    m * sizeof (uint64_t), sizeof (uint64_t), &object,
1076 			    DMU_READ_PREFETCH);
1077 			if (error != 0) {
1078 				vdev_dbgmsg(vd, "unable to read the metaslab "
1079 				    "array [error=%d]", error);
1080 				return (error);
1081 			}
1082 		}
1083 
1084 		error = metaslab_init(vd->vdev_mg, m, object, txg,
1085 		    &(vd->vdev_ms[m]));
1086 		if (error != 0) {
1087 			vdev_dbgmsg(vd, "metaslab_init failed [error=%d]",
1088 			    error);
1089 			return (error);
1090 		}
1091 	}
1092 
1093 	if (txg == 0)
1094 		spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER);
1095 
1096 	/*
1097 	 * If the vdev is being removed we don't activate
1098 	 * the metaslabs since we want to ensure that no new
1099 	 * allocations are performed on this device.
1100 	 */
1101 	if (oldc == 0 && !vd->vdev_removing)
1102 		metaslab_group_activate(vd->vdev_mg);
1103 
1104 	if (txg == 0)
1105 		spa_config_exit(spa, SCL_ALLOC, FTAG);
1106 
1107 	return (0);
1108 }
1109 
1110 void
1111 vdev_metaslab_fini(vdev_t *vd)
1112 {
1113 	if (vd->vdev_checkpoint_sm != NULL) {
1114 		ASSERT(spa_feature_is_active(vd->vdev_spa,
1115 		    SPA_FEATURE_POOL_CHECKPOINT));
1116 		space_map_close(vd->vdev_checkpoint_sm);
1117 		/*
1118 		 * Even though we close the space map, we need to set its
1119 		 * pointer to NULL. The reason is that vdev_metaslab_fini()
1120 		 * may be called multiple times for certain operations
1121 		 * (i.e. when destroying a pool) so we need to ensure that
1122 		 * this clause never executes twice. This logic is similar
1123 		 * to the one used for the vdev_ms clause below.
1124 		 */
1125 		vd->vdev_checkpoint_sm = NULL;
1126 	}
1127 
1128 	if (vd->vdev_ms != NULL) {
1129 		uint64_t count = vd->vdev_ms_count;
1130 
1131 		metaslab_group_passivate(vd->vdev_mg);
1132 		for (uint64_t m = 0; m < count; m++) {
1133 			metaslab_t *msp = vd->vdev_ms[m];
1134 
1135 			if (msp != NULL)
1136 				metaslab_fini(msp);
1137 		}
1138 		kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
1139 		vd->vdev_ms = NULL;
1140 
1141 		vd->vdev_ms_count = 0;
1142 	}
1143 	ASSERT0(vd->vdev_ms_count);
1144 }
1145 
1146 typedef struct vdev_probe_stats {
1147 	boolean_t	vps_readable;
1148 	boolean_t	vps_writeable;
1149 	int		vps_flags;
1150 } vdev_probe_stats_t;
1151 
1152 static void
1153 vdev_probe_done(zio_t *zio)
1154 {
1155 	spa_t *spa = zio->io_spa;
1156 	vdev_t *vd = zio->io_vd;
1157 	vdev_probe_stats_t *vps = zio->io_private;
1158 
1159 	ASSERT(vd->vdev_probe_zio != NULL);
1160 
1161 	if (zio->io_type == ZIO_TYPE_READ) {
1162 		if (zio->io_error == 0)
1163 			vps->vps_readable = 1;
1164 		if (zio->io_error == 0 && spa_writeable(spa)) {
1165 			zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
1166 			    zio->io_offset, zio->io_size, zio->io_abd,
1167 			    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1168 			    ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
1169 		} else {
1170 			abd_free(zio->io_abd);
1171 		}
1172 	} else if (zio->io_type == ZIO_TYPE_WRITE) {
1173 		if (zio->io_error == 0)
1174 			vps->vps_writeable = 1;
1175 		abd_free(zio->io_abd);
1176 	} else if (zio->io_type == ZIO_TYPE_NULL) {
1177 		zio_t *pio;
1178 
1179 		vd->vdev_cant_read |= !vps->vps_readable;
1180 		vd->vdev_cant_write |= !vps->vps_writeable;
1181 
1182 		if (vdev_readable(vd) &&
1183 		    (vdev_writeable(vd) || !spa_writeable(spa))) {
1184 			zio->io_error = 0;
1185 		} else {
1186 			ASSERT(zio->io_error != 0);
1187 			vdev_dbgmsg(vd, "failed probe");
1188 			zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
1189 			    spa, vd, NULL, 0, 0);
1190 			zio->io_error = SET_ERROR(ENXIO);
1191 		}
1192 
1193 		mutex_enter(&vd->vdev_probe_lock);
1194 		ASSERT(vd->vdev_probe_zio == zio);
1195 		vd->vdev_probe_zio = NULL;
1196 		mutex_exit(&vd->vdev_probe_lock);
1197 
1198 		zio_link_t *zl = NULL;
1199 		while ((pio = zio_walk_parents(zio, &zl)) != NULL)
1200 			if (!vdev_accessible(vd, pio))
1201 				pio->io_error = SET_ERROR(ENXIO);
1202 
1203 		kmem_free(vps, sizeof (*vps));
1204 	}
1205 }
1206 
1207 /*
1208  * Determine whether this device is accessible.
1209  *
1210  * Read and write to several known locations: the pad regions of each
1211  * vdev label but the first, which we leave alone in case it contains
1212  * a VTOC.
1213  */
1214 zio_t *
1215 vdev_probe(vdev_t *vd, zio_t *zio)
1216 {
1217 	spa_t *spa = vd->vdev_spa;
1218 	vdev_probe_stats_t *vps = NULL;
1219 	zio_t *pio;
1220 
1221 	ASSERT(vd->vdev_ops->vdev_op_leaf);
1222 
1223 	/*
1224 	 * Don't probe the probe.
1225 	 */
1226 	if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
1227 		return (NULL);
1228 
1229 	/*
1230 	 * To prevent 'probe storms' when a device fails, we create
1231 	 * just one probe i/o at a time.  All zios that want to probe
1232 	 * this vdev will become parents of the probe io.
1233 	 */
1234 	mutex_enter(&vd->vdev_probe_lock);
1235 
1236 	if ((pio = vd->vdev_probe_zio) == NULL) {
1237 		vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
1238 
1239 		vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
1240 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
1241 		    ZIO_FLAG_TRYHARD;
1242 
1243 		if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
1244 			/*
1245 			 * vdev_cant_read and vdev_cant_write can only
1246 			 * transition from TRUE to FALSE when we have the
1247 			 * SCL_ZIO lock as writer; otherwise they can only
1248 			 * transition from FALSE to TRUE.  This ensures that
1249 			 * any zio looking at these values can assume that
1250 			 * failures persist for the life of the I/O.  That's
1251 			 * important because when a device has intermittent
1252 			 * connectivity problems, we want to ensure that
1253 			 * they're ascribed to the device (ENXIO) and not
1254 			 * the zio (EIO).
1255 			 *
1256 			 * Since we hold SCL_ZIO as writer here, clear both
1257 			 * values so the probe can reevaluate from first
1258 			 * principles.
1259 			 */
1260 			vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
1261 			vd->vdev_cant_read = B_FALSE;
1262 			vd->vdev_cant_write = B_FALSE;
1263 		}
1264 
1265 		vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
1266 		    vdev_probe_done, vps,
1267 		    vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
1268 
1269 		/*
1270 		 * We can't change the vdev state in this context, so we
1271 		 * kick off an async task to do it on our behalf.
1272 		 */
1273 		if (zio != NULL) {
1274 			vd->vdev_probe_wanted = B_TRUE;
1275 			spa_async_request(spa, SPA_ASYNC_PROBE);
1276 		}
1277 	}
1278 
1279 	if (zio != NULL)
1280 		zio_add_child(zio, pio);
1281 
1282 	mutex_exit(&vd->vdev_probe_lock);
1283 
1284 	if (vps == NULL) {
1285 		ASSERT(zio != NULL);
1286 		return (NULL);
1287 	}
1288 
1289 	for (int l = 1; l < VDEV_LABELS; l++) {
1290 		zio_nowait(zio_read_phys(pio, vd,
1291 		    vdev_label_offset(vd->vdev_psize, l,
1292 		    offsetof(vdev_label_t, vl_pad2)), VDEV_PAD_SIZE,
1293 		    abd_alloc_for_io(VDEV_PAD_SIZE, B_TRUE),
1294 		    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1295 		    ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1296 	}
1297 
1298 	if (zio == NULL)
1299 		return (pio);
1300 
1301 	zio_nowait(pio);
1302 	return (NULL);
1303 }
1304 
1305 static void
1306 vdev_open_child(void *arg)
1307 {
1308 	vdev_t *vd = arg;
1309 
1310 	vd->vdev_open_thread = curthread;
1311 	vd->vdev_open_error = vdev_open(vd);
1312 	vd->vdev_open_thread = NULL;
1313 }
1314 
1315 boolean_t
1316 vdev_uses_zvols(vdev_t *vd)
1317 {
1318 	if (vd->vdev_path && strncmp(vd->vdev_path, ZVOL_DIR,
1319 	    strlen(ZVOL_DIR)) == 0)
1320 		return (B_TRUE);
1321 	for (int c = 0; c < vd->vdev_children; c++)
1322 		if (vdev_uses_zvols(vd->vdev_child[c]))
1323 			return (B_TRUE);
1324 	return (B_FALSE);
1325 }
1326 
1327 void
1328 vdev_open_children(vdev_t *vd)
1329 {
1330 	taskq_t *tq;
1331 	int children = vd->vdev_children;
1332 
1333 	/*
1334 	 * in order to handle pools on top of zvols, do the opens
1335 	 * in a single thread so that the same thread holds the
1336 	 * spa_namespace_lock
1337 	 */
1338 	if (vdev_uses_zvols(vd)) {
1339 		for (int c = 0; c < children; c++)
1340 			vd->vdev_child[c]->vdev_open_error =
1341 			    vdev_open(vd->vdev_child[c]);
1342 		return;
1343 	}
1344 	tq = taskq_create("vdev_open", children, minclsyspri,
1345 	    children, children, TASKQ_PREPOPULATE);
1346 
1347 	for (int c = 0; c < children; c++)
1348 		VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c],
1349 		    TQ_SLEEP) != NULL);
1350 
1351 	taskq_destroy(tq);
1352 }
1353 
1354 /*
1355  * Compute the raidz-deflation ratio.  Note, we hard-code
1356  * in 128k (1 << 17) because it is the "typical" blocksize.
1357  * Even though SPA_MAXBLOCKSIZE changed, this algorithm can not change,
1358  * otherwise it would inconsistently account for existing bp's.
1359  */
1360 static void
1361 vdev_set_deflate_ratio(vdev_t *vd)
1362 {
1363 	if (vd == vd->vdev_top && !vd->vdev_ishole && vd->vdev_ashift != 0) {
1364 		vd->vdev_deflate_ratio = (1 << 17) /
1365 		    (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
1366 	}
1367 }
1368 
1369 /*
1370  * Prepare a virtual device for access.
1371  */
1372 int
1373 vdev_open(vdev_t *vd)
1374 {
1375 	spa_t *spa = vd->vdev_spa;
1376 	int error;
1377 	uint64_t osize = 0;
1378 	uint64_t max_osize = 0;
1379 	uint64_t asize, max_asize, psize;
1380 	uint64_t ashift = 0;
1381 
1382 	ASSERT(vd->vdev_open_thread == curthread ||
1383 	    spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1384 	ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1385 	    vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1386 	    vd->vdev_state == VDEV_STATE_OFFLINE);
1387 
1388 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1389 	vd->vdev_cant_read = B_FALSE;
1390 	vd->vdev_cant_write = B_FALSE;
1391 	vd->vdev_min_asize = vdev_get_min_asize(vd);
1392 
1393 	/*
1394 	 * If this vdev is not removed, check its fault status.  If it's
1395 	 * faulted, bail out of the open.
1396 	 */
1397 	if (!vd->vdev_removed && vd->vdev_faulted) {
1398 		ASSERT(vd->vdev_children == 0);
1399 		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1400 		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1401 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1402 		    vd->vdev_label_aux);
1403 		return (SET_ERROR(ENXIO));
1404 	} else if (vd->vdev_offline) {
1405 		ASSERT(vd->vdev_children == 0);
1406 		vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1407 		return (SET_ERROR(ENXIO));
1408 	}
1409 
1410 	error = vd->vdev_ops->vdev_op_open(vd, &osize, &max_osize, &ashift);
1411 
1412 	/*
1413 	 * Reset the vdev_reopening flag so that we actually close
1414 	 * the vdev on error.
1415 	 */
1416 	vd->vdev_reopening = B_FALSE;
1417 	if (zio_injection_enabled && error == 0)
1418 		error = zio_handle_device_injection(vd, NULL, ENXIO);
1419 
1420 	if (error) {
1421 		if (vd->vdev_removed &&
1422 		    vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1423 			vd->vdev_removed = B_FALSE;
1424 
1425 		if (vd->vdev_stat.vs_aux == VDEV_AUX_CHILDREN_OFFLINE) {
1426 			vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE,
1427 			    vd->vdev_stat.vs_aux);
1428 		} else {
1429 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1430 			    vd->vdev_stat.vs_aux);
1431 		}
1432 		return (error);
1433 	}
1434 
1435 	vd->vdev_removed = B_FALSE;
1436 
1437 	/*
1438 	 * Recheck the faulted flag now that we have confirmed that
1439 	 * the vdev is accessible.  If we're faulted, bail.
1440 	 */
1441 	if (vd->vdev_faulted) {
1442 		ASSERT(vd->vdev_children == 0);
1443 		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1444 		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1445 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1446 		    vd->vdev_label_aux);
1447 		return (SET_ERROR(ENXIO));
1448 	}
1449 
1450 	if (vd->vdev_degraded) {
1451 		ASSERT(vd->vdev_children == 0);
1452 		vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1453 		    VDEV_AUX_ERR_EXCEEDED);
1454 	} else {
1455 		vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1456 	}
1457 
1458 	/*
1459 	 * For hole or missing vdevs we just return success.
1460 	 */
1461 	if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1462 		return (0);
1463 
1464 	for (int c = 0; c < vd->vdev_children; c++) {
1465 		if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1466 			vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1467 			    VDEV_AUX_NONE);
1468 			break;
1469 		}
1470 	}
1471 
1472 	osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1473 	max_osize = P2ALIGN(max_osize, (uint64_t)sizeof (vdev_label_t));
1474 
1475 	if (vd->vdev_children == 0) {
1476 		if (osize < SPA_MINDEVSIZE) {
1477 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1478 			    VDEV_AUX_TOO_SMALL);
1479 			return (SET_ERROR(EOVERFLOW));
1480 		}
1481 		psize = osize;
1482 		asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1483 		max_asize = max_osize - (VDEV_LABEL_START_SIZE +
1484 		    VDEV_LABEL_END_SIZE);
1485 	} else {
1486 		if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1487 		    (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1488 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1489 			    VDEV_AUX_TOO_SMALL);
1490 			return (SET_ERROR(EOVERFLOW));
1491 		}
1492 		psize = 0;
1493 		asize = osize;
1494 		max_asize = max_osize;
1495 	}
1496 
1497 	vd->vdev_psize = psize;
1498 
1499 	/*
1500 	 * Make sure the allocatable size hasn't shrunk too much.
1501 	 */
1502 	if (asize < vd->vdev_min_asize) {
1503 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1504 		    VDEV_AUX_BAD_LABEL);
1505 		return (SET_ERROR(EINVAL));
1506 	}
1507 
1508 	if (vd->vdev_asize == 0) {
1509 		/*
1510 		 * This is the first-ever open, so use the computed values.
1511 		 * For testing purposes, a higher ashift can be requested.
1512 		 */
1513 		vd->vdev_asize = asize;
1514 		vd->vdev_max_asize = max_asize;
1515 		vd->vdev_ashift = MAX(ashift, vd->vdev_ashift);
1516 		vd->vdev_ashift = MAX(zfs_ashift_min, vd->vdev_ashift);
1517 	} else {
1518 		/*
1519 		 * Detect if the alignment requirement has increased.
1520 		 * We don't want to make the pool unavailable, just
1521 		 * issue a warning instead.
1522 		 */
1523 		if (ashift > vd->vdev_top->vdev_ashift &&
1524 		    vd->vdev_ops->vdev_op_leaf) {
1525 			cmn_err(CE_WARN,
1526 			    "Disk, '%s', has a block alignment that is "
1527 			    "larger than the pool's alignment\n",
1528 			    vd->vdev_path);
1529 		}
1530 		vd->vdev_max_asize = max_asize;
1531 	}
1532 
1533 	/*
1534 	 * If all children are healthy we update asize if either:
1535 	 * The asize has increased, due to a device expansion caused by dynamic
1536 	 * LUN growth or vdev replacement, and automatic expansion is enabled;
1537 	 * making the additional space available.
1538 	 *
1539 	 * The asize has decreased, due to a device shrink usually caused by a
1540 	 * vdev replace with a smaller device. This ensures that calculations
1541 	 * based of max_asize and asize e.g. esize are always valid. It's safe
1542 	 * to do this as we've already validated that asize is greater than
1543 	 * vdev_min_asize.
1544 	 */
1545 	if (vd->vdev_state == VDEV_STATE_HEALTHY &&
1546 	    ((asize > vd->vdev_asize &&
1547 	    (vd->vdev_expanding || spa->spa_autoexpand)) ||
1548 	    (asize < vd->vdev_asize)))
1549 		vd->vdev_asize = asize;
1550 
1551 	vdev_set_min_asize(vd);
1552 
1553 	/*
1554 	 * Ensure we can issue some IO before declaring the
1555 	 * vdev open for business.
1556 	 */
1557 	if (vd->vdev_ops->vdev_op_leaf &&
1558 	    (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1559 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1560 		    VDEV_AUX_ERR_EXCEEDED);
1561 		return (error);
1562 	}
1563 
1564 	/*
1565 	 * Track the min and max ashift values for normal data devices.
1566 	 */
1567 	if (vd->vdev_top == vd && vd->vdev_ashift != 0 &&
1568 	    !vd->vdev_islog && vd->vdev_aux == NULL) {
1569 		if (vd->vdev_ashift > spa->spa_max_ashift)
1570 			spa->spa_max_ashift = vd->vdev_ashift;
1571 		if (vd->vdev_ashift < spa->spa_min_ashift)
1572 			spa->spa_min_ashift = vd->vdev_ashift;
1573 	}
1574 
1575 	/*
1576 	 * If a leaf vdev has a DTL, and seems healthy, then kick off a
1577 	 * resilver.  But don't do this if we are doing a reopen for a scrub,
1578 	 * since this would just restart the scrub we are already doing.
1579 	 */
1580 	if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1581 	    vdev_resilver_needed(vd, NULL, NULL))
1582 		spa_async_request(spa, SPA_ASYNC_RESILVER);
1583 
1584 	return (0);
1585 }
1586 
1587 /*
1588  * Called once the vdevs are all opened, this routine validates the label
1589  * contents. This needs to be done before vdev_load() so that we don't
1590  * inadvertently do repair I/Os to the wrong device.
1591  *
1592  * This function will only return failure if one of the vdevs indicates that it
1593  * has since been destroyed or exported.  This is only possible if
1594  * /etc/zfs/zpool.cache was readonly at the time.  Otherwise, the vdev state
1595  * will be updated but the function will return 0.
1596  */
1597 int
1598 vdev_validate(vdev_t *vd)
1599 {
1600 	spa_t *spa = vd->vdev_spa;
1601 	nvlist_t *label;
1602 	uint64_t guid = 0, aux_guid = 0, top_guid;
1603 	uint64_t state;
1604 	nvlist_t *nvl;
1605 	uint64_t txg;
1606 
1607 	if (vdev_validate_skip)
1608 		return (0);
1609 
1610 	for (uint64_t c = 0; c < vd->vdev_children; c++)
1611 		if (vdev_validate(vd->vdev_child[c]) != 0)
1612 			return (SET_ERROR(EBADF));
1613 
1614 	/*
1615 	 * If the device has already failed, or was marked offline, don't do
1616 	 * any further validation.  Otherwise, label I/O will fail and we will
1617 	 * overwrite the previous state.
1618 	 */
1619 	if (!vd->vdev_ops->vdev_op_leaf || !vdev_readable(vd))
1620 		return (0);
1621 
1622 	/*
1623 	 * If we are performing an extreme rewind, we allow for a label that
1624 	 * was modified at a point after the current txg.
1625 	 * If config lock is not held do not check for the txg. spa_sync could
1626 	 * be updating the vdev's label before updating spa_last_synced_txg.
1627 	 */
1628 	if (spa->spa_extreme_rewind || spa_last_synced_txg(spa) == 0 ||
1629 	    spa_config_held(spa, SCL_CONFIG, RW_WRITER) != SCL_CONFIG)
1630 		txg = UINT64_MAX;
1631 	else
1632 		txg = spa_last_synced_txg(spa);
1633 
1634 	if ((label = vdev_label_read_config(vd, txg)) == NULL) {
1635 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1636 		    VDEV_AUX_BAD_LABEL);
1637 		vdev_dbgmsg(vd, "vdev_validate: failed reading config for "
1638 		    "txg %llu", (u_longlong_t)txg);
1639 		return (0);
1640 	}
1641 
1642 	/*
1643 	 * Determine if this vdev has been split off into another
1644 	 * pool.  If so, then refuse to open it.
1645 	 */
1646 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID,
1647 	    &aux_guid) == 0 && aux_guid == spa_guid(spa)) {
1648 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1649 		    VDEV_AUX_SPLIT_POOL);
1650 		nvlist_free(label);
1651 		vdev_dbgmsg(vd, "vdev_validate: vdev split into other pool");
1652 		return (0);
1653 	}
1654 
1655 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_GUID, &guid) != 0) {
1656 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1657 		    VDEV_AUX_CORRUPT_DATA);
1658 		nvlist_free(label);
1659 		vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
1660 		    ZPOOL_CONFIG_POOL_GUID);
1661 		return (0);
1662 	}
1663 
1664 	/*
1665 	 * If config is not trusted then ignore the spa guid check. This is
1666 	 * necessary because if the machine crashed during a re-guid the new
1667 	 * guid might have been written to all of the vdev labels, but not the
1668 	 * cached config. The check will be performed again once we have the
1669 	 * trusted config from the MOS.
1670 	 */
1671 	if (spa->spa_trust_config && guid != spa_guid(spa)) {
1672 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1673 		    VDEV_AUX_CORRUPT_DATA);
1674 		nvlist_free(label);
1675 		vdev_dbgmsg(vd, "vdev_validate: vdev label pool_guid doesn't "
1676 		    "match config (%llu != %llu)", (u_longlong_t)guid,
1677 		    (u_longlong_t)spa_guid(spa));
1678 		return (0);
1679 	}
1680 
1681 	if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl)
1682 	    != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID,
1683 	    &aux_guid) != 0)
1684 		aux_guid = 0;
1685 
1686 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0) {
1687 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1688 		    VDEV_AUX_CORRUPT_DATA);
1689 		nvlist_free(label);
1690 		vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
1691 		    ZPOOL_CONFIG_GUID);
1692 		return (0);
1693 	}
1694 
1695 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID, &top_guid)
1696 	    != 0) {
1697 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1698 		    VDEV_AUX_CORRUPT_DATA);
1699 		nvlist_free(label);
1700 		vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
1701 		    ZPOOL_CONFIG_TOP_GUID);
1702 		return (0);
1703 	}
1704 
1705 	/*
1706 	 * If this vdev just became a top-level vdev because its sibling was
1707 	 * detached, it will have adopted the parent's vdev guid -- but the
1708 	 * label may or may not be on disk yet. Fortunately, either version
1709 	 * of the label will have the same top guid, so if we're a top-level
1710 	 * vdev, we can safely compare to that instead.
1711 	 * However, if the config comes from a cachefile that failed to update
1712 	 * after the detach, a top-level vdev will appear as a non top-level
1713 	 * vdev in the config. Also relax the constraints if we perform an
1714 	 * extreme rewind.
1715 	 *
1716 	 * If we split this vdev off instead, then we also check the
1717 	 * original pool's guid. We don't want to consider the vdev
1718 	 * corrupt if it is partway through a split operation.
1719 	 */
1720 	if (vd->vdev_guid != guid && vd->vdev_guid != aux_guid) {
1721 		boolean_t mismatch = B_FALSE;
1722 		if (spa->spa_trust_config && !spa->spa_extreme_rewind) {
1723 			if (vd != vd->vdev_top || vd->vdev_guid != top_guid)
1724 				mismatch = B_TRUE;
1725 		} else {
1726 			if (vd->vdev_guid != top_guid &&
1727 			    vd->vdev_top->vdev_guid != guid)
1728 				mismatch = B_TRUE;
1729 		}
1730 
1731 		if (mismatch) {
1732 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1733 			    VDEV_AUX_CORRUPT_DATA);
1734 			nvlist_free(label);
1735 			vdev_dbgmsg(vd, "vdev_validate: config guid "
1736 			    "doesn't match label guid");
1737 			vdev_dbgmsg(vd, "CONFIG: guid %llu, top_guid %llu",
1738 			    (u_longlong_t)vd->vdev_guid,
1739 			    (u_longlong_t)vd->vdev_top->vdev_guid);
1740 			vdev_dbgmsg(vd, "LABEL: guid %llu, top_guid %llu, "
1741 			    "aux_guid %llu", (u_longlong_t)guid,
1742 			    (u_longlong_t)top_guid, (u_longlong_t)aux_guid);
1743 			return (0);
1744 		}
1745 	}
1746 
1747 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1748 	    &state) != 0) {
1749 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1750 		    VDEV_AUX_CORRUPT_DATA);
1751 		nvlist_free(label);
1752 		vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
1753 		    ZPOOL_CONFIG_POOL_STATE);
1754 		return (0);
1755 	}
1756 
1757 	nvlist_free(label);
1758 
1759 	/*
1760 	 * If this is a verbatim import, no need to check the
1761 	 * state of the pool.
1762 	 */
1763 	if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) &&
1764 	    spa_load_state(spa) == SPA_LOAD_OPEN &&
1765 	    state != POOL_STATE_ACTIVE) {
1766 		vdev_dbgmsg(vd, "vdev_validate: invalid pool state (%llu) "
1767 		    "for spa %s", (u_longlong_t)state, spa->spa_name);
1768 		return (SET_ERROR(EBADF));
1769 	}
1770 
1771 	/*
1772 	 * If we were able to open and validate a vdev that was
1773 	 * previously marked permanently unavailable, clear that state
1774 	 * now.
1775 	 */
1776 	if (vd->vdev_not_present)
1777 		vd->vdev_not_present = 0;
1778 
1779 	return (0);
1780 }
1781 
1782 static void
1783 vdev_copy_path_impl(vdev_t *svd, vdev_t *dvd)
1784 {
1785 	if (svd->vdev_path != NULL && dvd->vdev_path != NULL) {
1786 		if (strcmp(svd->vdev_path, dvd->vdev_path) != 0) {
1787 			zfs_dbgmsg("vdev_copy_path: vdev %llu: path changed "
1788 			    "from '%s' to '%s'", (u_longlong_t)dvd->vdev_guid,
1789 			    dvd->vdev_path, svd->vdev_path);
1790 			spa_strfree(dvd->vdev_path);
1791 			dvd->vdev_path = spa_strdup(svd->vdev_path);
1792 		}
1793 	} else if (svd->vdev_path != NULL) {
1794 		dvd->vdev_path = spa_strdup(svd->vdev_path);
1795 		zfs_dbgmsg("vdev_copy_path: vdev %llu: path set to '%s'",
1796 		    (u_longlong_t)dvd->vdev_guid, dvd->vdev_path);
1797 	}
1798 }
1799 
1800 /*
1801  * Recursively copy vdev paths from one vdev to another. Source and destination
1802  * vdev trees must have same geometry otherwise return error. Intended to copy
1803  * paths from userland config into MOS config.
1804  */
1805 int
1806 vdev_copy_path_strict(vdev_t *svd, vdev_t *dvd)
1807 {
1808 	if ((svd->vdev_ops == &vdev_missing_ops) ||
1809 	    (svd->vdev_ishole && dvd->vdev_ishole) ||
1810 	    (dvd->vdev_ops == &vdev_indirect_ops))
1811 		return (0);
1812 
1813 	if (svd->vdev_ops != dvd->vdev_ops) {
1814 		vdev_dbgmsg(svd, "vdev_copy_path: vdev type mismatch: %s != %s",
1815 		    svd->vdev_ops->vdev_op_type, dvd->vdev_ops->vdev_op_type);
1816 		return (SET_ERROR(EINVAL));
1817 	}
1818 
1819 	if (svd->vdev_guid != dvd->vdev_guid) {
1820 		vdev_dbgmsg(svd, "vdev_copy_path: guids mismatch (%llu != "
1821 		    "%llu)", (u_longlong_t)svd->vdev_guid,
1822 		    (u_longlong_t)dvd->vdev_guid);
1823 		return (SET_ERROR(EINVAL));
1824 	}
1825 
1826 	if (svd->vdev_children != dvd->vdev_children) {
1827 		vdev_dbgmsg(svd, "vdev_copy_path: children count mismatch: "
1828 		    "%llu != %llu", (u_longlong_t)svd->vdev_children,
1829 		    (u_longlong_t)dvd->vdev_children);
1830 		return (SET_ERROR(EINVAL));
1831 	}
1832 
1833 	for (uint64_t i = 0; i < svd->vdev_children; i++) {
1834 		int error = vdev_copy_path_strict(svd->vdev_child[i],
1835 		    dvd->vdev_child[i]);
1836 		if (error != 0)
1837 			return (error);
1838 	}
1839 
1840 	if (svd->vdev_ops->vdev_op_leaf)
1841 		vdev_copy_path_impl(svd, dvd);
1842 
1843 	return (0);
1844 }
1845 
1846 static void
1847 vdev_copy_path_search(vdev_t *stvd, vdev_t *dvd)
1848 {
1849 	ASSERT(stvd->vdev_top == stvd);
1850 	ASSERT3U(stvd->vdev_id, ==, dvd->vdev_top->vdev_id);
1851 
1852 	for (uint64_t i = 0; i < dvd->vdev_children; i++) {
1853 		vdev_copy_path_search(stvd, dvd->vdev_child[i]);
1854 	}
1855 
1856 	if (!dvd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(dvd))
1857 		return;
1858 
1859 	/*
1860 	 * The idea here is that while a vdev can shift positions within
1861 	 * a top vdev (when replacing, attaching mirror, etc.) it cannot
1862 	 * step outside of it.
1863 	 */
1864 	vdev_t *vd = vdev_lookup_by_guid(stvd, dvd->vdev_guid);
1865 
1866 	if (vd == NULL || vd->vdev_ops != dvd->vdev_ops)
1867 		return;
1868 
1869 	ASSERT(vd->vdev_ops->vdev_op_leaf);
1870 
1871 	vdev_copy_path_impl(vd, dvd);
1872 }
1873 
1874 /*
1875  * Recursively copy vdev paths from one root vdev to another. Source and
1876  * destination vdev trees may differ in geometry. For each destination leaf
1877  * vdev, search a vdev with the same guid and top vdev id in the source.
1878  * Intended to copy paths from userland config into MOS config.
1879  */
1880 void
1881 vdev_copy_path_relaxed(vdev_t *srvd, vdev_t *drvd)
1882 {
1883 	uint64_t children = MIN(srvd->vdev_children, drvd->vdev_children);
1884 	ASSERT(srvd->vdev_ops == &vdev_root_ops);
1885 	ASSERT(drvd->vdev_ops == &vdev_root_ops);
1886 
1887 	for (uint64_t i = 0; i < children; i++) {
1888 		vdev_copy_path_search(srvd->vdev_child[i],
1889 		    drvd->vdev_child[i]);
1890 	}
1891 }
1892 
1893 /*
1894  * Close a virtual device.
1895  */
1896 void
1897 vdev_close(vdev_t *vd)
1898 {
1899 	spa_t *spa = vd->vdev_spa;
1900 	vdev_t *pvd = vd->vdev_parent;
1901 
1902 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1903 
1904 	/*
1905 	 * If our parent is reopening, then we are as well, unless we are
1906 	 * going offline.
1907 	 */
1908 	if (pvd != NULL && pvd->vdev_reopening)
1909 		vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline);
1910 
1911 	vd->vdev_ops->vdev_op_close(vd);
1912 
1913 	vdev_cache_purge(vd);
1914 
1915 	/*
1916 	 * We record the previous state before we close it, so that if we are
1917 	 * doing a reopen(), we don't generate FMA ereports if we notice that
1918 	 * it's still faulted.
1919 	 */
1920 	vd->vdev_prevstate = vd->vdev_state;
1921 
1922 	if (vd->vdev_offline)
1923 		vd->vdev_state = VDEV_STATE_OFFLINE;
1924 	else
1925 		vd->vdev_state = VDEV_STATE_CLOSED;
1926 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1927 }
1928 
1929 void
1930 vdev_hold(vdev_t *vd)
1931 {
1932 	spa_t *spa = vd->vdev_spa;
1933 
1934 	ASSERT(spa_is_root(spa));
1935 	if (spa->spa_state == POOL_STATE_UNINITIALIZED)
1936 		return;
1937 
1938 	for (int c = 0; c < vd->vdev_children; c++)
1939 		vdev_hold(vd->vdev_child[c]);
1940 
1941 	if (vd->vdev_ops->vdev_op_leaf)
1942 		vd->vdev_ops->vdev_op_hold(vd);
1943 }
1944 
1945 void
1946 vdev_rele(vdev_t *vd)
1947 {
1948 	spa_t *spa = vd->vdev_spa;
1949 
1950 	ASSERT(spa_is_root(spa));
1951 	for (int c = 0; c < vd->vdev_children; c++)
1952 		vdev_rele(vd->vdev_child[c]);
1953 
1954 	if (vd->vdev_ops->vdev_op_leaf)
1955 		vd->vdev_ops->vdev_op_rele(vd);
1956 }
1957 
1958 /*
1959  * Reopen all interior vdevs and any unopened leaves.  We don't actually
1960  * reopen leaf vdevs which had previously been opened as they might deadlock
1961  * on the spa_config_lock.  Instead we only obtain the leaf's physical size.
1962  * If the leaf has never been opened then open it, as usual.
1963  */
1964 void
1965 vdev_reopen(vdev_t *vd)
1966 {
1967 	spa_t *spa = vd->vdev_spa;
1968 
1969 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1970 
1971 	/* set the reopening flag unless we're taking the vdev offline */
1972 	vd->vdev_reopening = !vd->vdev_offline;
1973 	vdev_close(vd);
1974 	(void) vdev_open(vd);
1975 
1976 	/*
1977 	 * Call vdev_validate() here to make sure we have the same device.
1978 	 * Otherwise, a device with an invalid label could be successfully
1979 	 * opened in response to vdev_reopen().
1980 	 */
1981 	if (vd->vdev_aux) {
1982 		(void) vdev_validate_aux(vd);
1983 		if (vdev_readable(vd) && vdev_writeable(vd) &&
1984 		    vd->vdev_aux == &spa->spa_l2cache &&
1985 		    !l2arc_vdev_present(vd))
1986 			l2arc_add_vdev(spa, vd);
1987 	} else {
1988 		(void) vdev_validate(vd);
1989 	}
1990 
1991 	/*
1992 	 * Reassess parent vdev's health.
1993 	 */
1994 	vdev_propagate_state(vd);
1995 }
1996 
1997 int
1998 vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1999 {
2000 	int error;
2001 
2002 	/*
2003 	 * Normally, partial opens (e.g. of a mirror) are allowed.
2004 	 * For a create, however, we want to fail the request if
2005 	 * there are any components we can't open.
2006 	 */
2007 	error = vdev_open(vd);
2008 
2009 	if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
2010 		vdev_close(vd);
2011 		return (error ? error : ENXIO);
2012 	}
2013 
2014 	/*
2015 	 * Recursively load DTLs and initialize all labels.
2016 	 */
2017 	if ((error = vdev_dtl_load(vd)) != 0 ||
2018 	    (error = vdev_label_init(vd, txg, isreplacing ?
2019 	    VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
2020 		vdev_close(vd);
2021 		return (error);
2022 	}
2023 
2024 	return (0);
2025 }
2026 
2027 void
2028 vdev_metaslab_set_size(vdev_t *vd)
2029 {
2030 	uint64_t asize = vd->vdev_asize;
2031 	uint64_t ms_shift = 0;
2032 
2033 	/*
2034 	 * For vdevs that are bigger than 8G the metaslab size varies in
2035 	 * a way that the number of metaslabs increases in powers of two,
2036 	 * linearly in terms of vdev_asize, starting from 16 metaslabs.
2037 	 * So for vdev_asize of 8G we get 16 metaslabs, for 16G, we get 32,
2038 	 * and so on, until we hit the maximum metaslab count limit
2039 	 * [vdev_max_ms_count] from which point the metaslab count stays
2040 	 * the same.
2041 	 */
2042 	ms_shift = vdev_default_ms_shift;
2043 
2044 	if ((asize >> ms_shift) < vdev_min_ms_count) {
2045 		/*
2046 		 * For devices that are less than 8G we want to have
2047 		 * exactly 16 metaslabs. We don't want less as integer
2048 		 * division rounds down, so less metaslabs mean more
2049 		 * wasted space. We don't want more as these vdevs are
2050 		 * small and in the likely event that we are running
2051 		 * out of space, the SPA will have a hard time finding
2052 		 * space due to fragmentation.
2053 		 */
2054 		ms_shift = highbit64(asize / vdev_min_ms_count);
2055 		ms_shift = MAX(ms_shift, SPA_MAXBLOCKSHIFT);
2056 
2057 	} else if ((asize >> ms_shift) > vdev_max_ms_count) {
2058 		ms_shift = highbit64(asize / vdev_max_ms_count);
2059 	}
2060 
2061 	vd->vdev_ms_shift = ms_shift;
2062 	ASSERT3U(vd->vdev_ms_shift, >=, SPA_MAXBLOCKSHIFT);
2063 }
2064 
2065 void
2066 vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
2067 {
2068 	ASSERT(vd == vd->vdev_top);
2069 	/* indirect vdevs don't have metaslabs or dtls */
2070 	ASSERT(vdev_is_concrete(vd) || flags == 0);
2071 	ASSERT(ISP2(flags));
2072 	ASSERT(spa_writeable(vd->vdev_spa));
2073 
2074 	if (flags & VDD_METASLAB)
2075 		(void) txg_list_add(&vd->vdev_ms_list, arg, txg);
2076 
2077 	if (flags & VDD_DTL)
2078 		(void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
2079 
2080 	(void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
2081 }
2082 
2083 void
2084 vdev_dirty_leaves(vdev_t *vd, int flags, uint64_t txg)
2085 {
2086 	for (int c = 0; c < vd->vdev_children; c++)
2087 		vdev_dirty_leaves(vd->vdev_child[c], flags, txg);
2088 
2089 	if (vd->vdev_ops->vdev_op_leaf)
2090 		vdev_dirty(vd->vdev_top, flags, vd, txg);
2091 }
2092 
2093 /*
2094  * DTLs.
2095  *
2096  * A vdev's DTL (dirty time log) is the set of transaction groups for which
2097  * the vdev has less than perfect replication.  There are four kinds of DTL:
2098  *
2099  * DTL_MISSING: txgs for which the vdev has no valid copies of the data
2100  *
2101  * DTL_PARTIAL: txgs for which data is available, but not fully replicated
2102  *
2103  * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
2104  *	scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
2105  *	txgs that was scrubbed.
2106  *
2107  * DTL_OUTAGE: txgs which cannot currently be read, whether due to
2108  *	persistent errors or just some device being offline.
2109  *	Unlike the other three, the DTL_OUTAGE map is not generally
2110  *	maintained; it's only computed when needed, typically to
2111  *	determine whether a device can be detached.
2112  *
2113  * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
2114  * either has the data or it doesn't.
2115  *
2116  * For interior vdevs such as mirror and RAID-Z the picture is more complex.
2117  * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
2118  * if any child is less than fully replicated, then so is its parent.
2119  * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
2120  * comprising only those txgs which appear in 'maxfaults' or more children;
2121  * those are the txgs we don't have enough replication to read.  For example,
2122  * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
2123  * thus, its DTL_MISSING consists of the set of txgs that appear in more than
2124  * two child DTL_MISSING maps.
2125  *
2126  * It should be clear from the above that to compute the DTLs and outage maps
2127  * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
2128  * Therefore, that is all we keep on disk.  When loading the pool, or after
2129  * a configuration change, we generate all other DTLs from first principles.
2130  */
2131 void
2132 vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
2133 {
2134 	range_tree_t *rt = vd->vdev_dtl[t];
2135 
2136 	ASSERT(t < DTL_TYPES);
2137 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
2138 	ASSERT(spa_writeable(vd->vdev_spa));
2139 
2140 	mutex_enter(&vd->vdev_dtl_lock);
2141 	if (!range_tree_contains(rt, txg, size))
2142 		range_tree_add(rt, txg, size);
2143 	mutex_exit(&vd->vdev_dtl_lock);
2144 }
2145 
2146 boolean_t
2147 vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
2148 {
2149 	range_tree_t *rt = vd->vdev_dtl[t];
2150 	boolean_t dirty = B_FALSE;
2151 
2152 	ASSERT(t < DTL_TYPES);
2153 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
2154 
2155 	/*
2156 	 * While we are loading the pool, the DTLs have not been loaded yet.
2157 	 * Ignore the DTLs and try all devices.  This avoids a recursive
2158 	 * mutex enter on the vdev_dtl_lock, and also makes us try hard
2159 	 * when loading the pool (relying on the checksum to ensure that
2160 	 * we get the right data -- note that we while loading, we are
2161 	 * only reading the MOS, which is always checksummed).
2162 	 */
2163 	if (vd->vdev_spa->spa_load_state != SPA_LOAD_NONE)
2164 		return (B_FALSE);
2165 
2166 	mutex_enter(&vd->vdev_dtl_lock);
2167 	if (!range_tree_is_empty(rt))
2168 		dirty = range_tree_contains(rt, txg, size);
2169 	mutex_exit(&vd->vdev_dtl_lock);
2170 
2171 	return (dirty);
2172 }
2173 
2174 boolean_t
2175 vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
2176 {
2177 	range_tree_t *rt = vd->vdev_dtl[t];
2178 	boolean_t empty;
2179 
2180 	mutex_enter(&vd->vdev_dtl_lock);
2181 	empty = range_tree_is_empty(rt);
2182 	mutex_exit(&vd->vdev_dtl_lock);
2183 
2184 	return (empty);
2185 }
2186 
2187 /*
2188  * Returns the lowest txg in the DTL range.
2189  */
2190 static uint64_t
2191 vdev_dtl_min(vdev_t *vd)
2192 {
2193 	range_seg_t *rs;
2194 
2195 	ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
2196 	ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
2197 	ASSERT0(vd->vdev_children);
2198 
2199 	rs = avl_first(&vd->vdev_dtl[DTL_MISSING]->rt_root);
2200 	return (rs->rs_start - 1);
2201 }
2202 
2203 /*
2204  * Returns the highest txg in the DTL.
2205  */
2206 static uint64_t
2207 vdev_dtl_max(vdev_t *vd)
2208 {
2209 	range_seg_t *rs;
2210 
2211 	ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
2212 	ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
2213 	ASSERT0(vd->vdev_children);
2214 
2215 	rs = avl_last(&vd->vdev_dtl[DTL_MISSING]->rt_root);
2216 	return (rs->rs_end);
2217 }
2218 
2219 /*
2220  * Determine if a resilvering vdev should remove any DTL entries from
2221  * its range. If the vdev was resilvering for the entire duration of the
2222  * scan then it should excise that range from its DTLs. Otherwise, this
2223  * vdev is considered partially resilvered and should leave its DTL
2224  * entries intact. The comment in vdev_dtl_reassess() describes how we
2225  * excise the DTLs.
2226  */
2227 static boolean_t
2228 vdev_dtl_should_excise(vdev_t *vd)
2229 {
2230 	spa_t *spa = vd->vdev_spa;
2231 	dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
2232 
2233 	ASSERT0(scn->scn_phys.scn_errors);
2234 	ASSERT0(vd->vdev_children);
2235 
2236 	if (vd->vdev_state < VDEV_STATE_DEGRADED)
2237 		return (B_FALSE);
2238 
2239 	if (vd->vdev_resilver_txg == 0 ||
2240 	    range_tree_is_empty(vd->vdev_dtl[DTL_MISSING]))
2241 		return (B_TRUE);
2242 
2243 	/*
2244 	 * When a resilver is initiated the scan will assign the scn_max_txg
2245 	 * value to the highest txg value that exists in all DTLs. If this
2246 	 * device's max DTL is not part of this scan (i.e. it is not in
2247 	 * the range (scn_min_txg, scn_max_txg] then it is not eligible
2248 	 * for excision.
2249 	 */
2250 	if (vdev_dtl_max(vd) <= scn->scn_phys.scn_max_txg) {
2251 		ASSERT3U(scn->scn_phys.scn_min_txg, <=, vdev_dtl_min(vd));
2252 		ASSERT3U(scn->scn_phys.scn_min_txg, <, vd->vdev_resilver_txg);
2253 		ASSERT3U(vd->vdev_resilver_txg, <=, scn->scn_phys.scn_max_txg);
2254 		return (B_TRUE);
2255 	}
2256 	return (B_FALSE);
2257 }
2258 
2259 /*
2260  * Reassess DTLs after a config change or scrub completion.
2261  */
2262 void
2263 vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
2264 {
2265 	spa_t *spa = vd->vdev_spa;
2266 	avl_tree_t reftree;
2267 	int minref;
2268 
2269 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
2270 
2271 	for (int c = 0; c < vd->vdev_children; c++)
2272 		vdev_dtl_reassess(vd->vdev_child[c], txg,
2273 		    scrub_txg, scrub_done);
2274 
2275 	if (vd == spa->spa_root_vdev || !vdev_is_concrete(vd) || vd->vdev_aux)
2276 		return;
2277 
2278 	if (vd->vdev_ops->vdev_op_leaf) {
2279 		dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
2280 
2281 		mutex_enter(&vd->vdev_dtl_lock);
2282 
2283 		/*
2284 		 * If we've completed a scan cleanly then determine
2285 		 * if this vdev should remove any DTLs. We only want to
2286 		 * excise regions on vdevs that were available during
2287 		 * the entire duration of this scan.
2288 		 */
2289 		if (scrub_txg != 0 &&
2290 		    (spa->spa_scrub_started ||
2291 		    (scn != NULL && scn->scn_phys.scn_errors == 0)) &&
2292 		    vdev_dtl_should_excise(vd)) {
2293 			/*
2294 			 * We completed a scrub up to scrub_txg.  If we
2295 			 * did it without rebooting, then the scrub dtl
2296 			 * will be valid, so excise the old region and
2297 			 * fold in the scrub dtl.  Otherwise, leave the
2298 			 * dtl as-is if there was an error.
2299 			 *
2300 			 * There's little trick here: to excise the beginning
2301 			 * of the DTL_MISSING map, we put it into a reference
2302 			 * tree and then add a segment with refcnt -1 that
2303 			 * covers the range [0, scrub_txg).  This means
2304 			 * that each txg in that range has refcnt -1 or 0.
2305 			 * We then add DTL_SCRUB with a refcnt of 2, so that
2306 			 * entries in the range [0, scrub_txg) will have a
2307 			 * positive refcnt -- either 1 or 2.  We then convert
2308 			 * the reference tree into the new DTL_MISSING map.
2309 			 */
2310 			space_reftree_create(&reftree);
2311 			space_reftree_add_map(&reftree,
2312 			    vd->vdev_dtl[DTL_MISSING], 1);
2313 			space_reftree_add_seg(&reftree, 0, scrub_txg, -1);
2314 			space_reftree_add_map(&reftree,
2315 			    vd->vdev_dtl[DTL_SCRUB], 2);
2316 			space_reftree_generate_map(&reftree,
2317 			    vd->vdev_dtl[DTL_MISSING], 1);
2318 			space_reftree_destroy(&reftree);
2319 		}
2320 		range_tree_vacate(vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
2321 		range_tree_walk(vd->vdev_dtl[DTL_MISSING],
2322 		    range_tree_add, vd->vdev_dtl[DTL_PARTIAL]);
2323 		if (scrub_done)
2324 			range_tree_vacate(vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
2325 		range_tree_vacate(vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
2326 		if (!vdev_readable(vd))
2327 			range_tree_add(vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
2328 		else
2329 			range_tree_walk(vd->vdev_dtl[DTL_MISSING],
2330 			    range_tree_add, vd->vdev_dtl[DTL_OUTAGE]);
2331 
2332 		/*
2333 		 * If the vdev was resilvering and no longer has any
2334 		 * DTLs then reset its resilvering flag.
2335 		 */
2336 		if (vd->vdev_resilver_txg != 0 &&
2337 		    range_tree_is_empty(vd->vdev_dtl[DTL_MISSING]) &&
2338 		    range_tree_is_empty(vd->vdev_dtl[DTL_OUTAGE]))
2339 			vd->vdev_resilver_txg = 0;
2340 
2341 		mutex_exit(&vd->vdev_dtl_lock);
2342 
2343 		if (txg != 0)
2344 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
2345 		return;
2346 	}
2347 
2348 	mutex_enter(&vd->vdev_dtl_lock);
2349 	for (int t = 0; t < DTL_TYPES; t++) {
2350 		/* account for child's outage in parent's missing map */
2351 		int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
2352 		if (t == DTL_SCRUB)
2353 			continue;			/* leaf vdevs only */
2354 		if (t == DTL_PARTIAL)
2355 			minref = 1;			/* i.e. non-zero */
2356 		else if (vd->vdev_nparity != 0)
2357 			minref = vd->vdev_nparity + 1;	/* RAID-Z */
2358 		else
2359 			minref = vd->vdev_children;	/* any kind of mirror */
2360 		space_reftree_create(&reftree);
2361 		for (int c = 0; c < vd->vdev_children; c++) {
2362 			vdev_t *cvd = vd->vdev_child[c];
2363 			mutex_enter(&cvd->vdev_dtl_lock);
2364 			space_reftree_add_map(&reftree, cvd->vdev_dtl[s], 1);
2365 			mutex_exit(&cvd->vdev_dtl_lock);
2366 		}
2367 		space_reftree_generate_map(&reftree, vd->vdev_dtl[t], minref);
2368 		space_reftree_destroy(&reftree);
2369 	}
2370 	mutex_exit(&vd->vdev_dtl_lock);
2371 }
2372 
2373 int
2374 vdev_dtl_load(vdev_t *vd)
2375 {
2376 	spa_t *spa = vd->vdev_spa;
2377 	objset_t *mos = spa->spa_meta_objset;
2378 	int error = 0;
2379 
2380 	if (vd->vdev_ops->vdev_op_leaf && vd->vdev_dtl_object != 0) {
2381 		ASSERT(vdev_is_concrete(vd));
2382 
2383 		error = space_map_open(&vd->vdev_dtl_sm, mos,
2384 		    vd->vdev_dtl_object, 0, -1ULL, 0);
2385 		if (error)
2386 			return (error);
2387 		ASSERT(vd->vdev_dtl_sm != NULL);
2388 
2389 		mutex_enter(&vd->vdev_dtl_lock);
2390 
2391 		/*
2392 		 * Now that we've opened the space_map we need to update
2393 		 * the in-core DTL.
2394 		 */
2395 		space_map_update(vd->vdev_dtl_sm);
2396 
2397 		error = space_map_load(vd->vdev_dtl_sm,
2398 		    vd->vdev_dtl[DTL_MISSING], SM_ALLOC);
2399 		mutex_exit(&vd->vdev_dtl_lock);
2400 
2401 		return (error);
2402 	}
2403 
2404 	for (int c = 0; c < vd->vdev_children; c++) {
2405 		error = vdev_dtl_load(vd->vdev_child[c]);
2406 		if (error != 0)
2407 			break;
2408 	}
2409 
2410 	return (error);
2411 }
2412 
2413 void
2414 vdev_destroy_unlink_zap(vdev_t *vd, uint64_t zapobj, dmu_tx_t *tx)
2415 {
2416 	spa_t *spa = vd->vdev_spa;
2417 
2418 	VERIFY0(zap_destroy(spa->spa_meta_objset, zapobj, tx));
2419 	VERIFY0(zap_remove_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
2420 	    zapobj, tx));
2421 }
2422 
2423 uint64_t
2424 vdev_create_link_zap(vdev_t *vd, dmu_tx_t *tx)
2425 {
2426 	spa_t *spa = vd->vdev_spa;
2427 	uint64_t zap = zap_create(spa->spa_meta_objset, DMU_OTN_ZAP_METADATA,
2428 	    DMU_OT_NONE, 0, tx);
2429 
2430 	ASSERT(zap != 0);
2431 	VERIFY0(zap_add_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
2432 	    zap, tx));
2433 
2434 	return (zap);
2435 }
2436 
2437 void
2438 vdev_construct_zaps(vdev_t *vd, dmu_tx_t *tx)
2439 {
2440 	if (vd->vdev_ops != &vdev_hole_ops &&
2441 	    vd->vdev_ops != &vdev_missing_ops &&
2442 	    vd->vdev_ops != &vdev_root_ops &&
2443 	    !vd->vdev_top->vdev_removing) {
2444 		if (vd->vdev_ops->vdev_op_leaf && vd->vdev_leaf_zap == 0) {
2445 			vd->vdev_leaf_zap = vdev_create_link_zap(vd, tx);
2446 		}
2447 		if (vd == vd->vdev_top && vd->vdev_top_zap == 0) {
2448 			vd->vdev_top_zap = vdev_create_link_zap(vd, tx);
2449 		}
2450 	}
2451 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
2452 		vdev_construct_zaps(vd->vdev_child[i], tx);
2453 	}
2454 }
2455 
2456 void
2457 vdev_dtl_sync(vdev_t *vd, uint64_t txg)
2458 {
2459 	spa_t *spa = vd->vdev_spa;
2460 	range_tree_t *rt = vd->vdev_dtl[DTL_MISSING];
2461 	objset_t *mos = spa->spa_meta_objset;
2462 	range_tree_t *rtsync;
2463 	dmu_tx_t *tx;
2464 	uint64_t object = space_map_object(vd->vdev_dtl_sm);
2465 
2466 	ASSERT(vdev_is_concrete(vd));
2467 	ASSERT(vd->vdev_ops->vdev_op_leaf);
2468 
2469 	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2470 
2471 	if (vd->vdev_detached || vd->vdev_top->vdev_removing) {
2472 		mutex_enter(&vd->vdev_dtl_lock);
2473 		space_map_free(vd->vdev_dtl_sm, tx);
2474 		space_map_close(vd->vdev_dtl_sm);
2475 		vd->vdev_dtl_sm = NULL;
2476 		mutex_exit(&vd->vdev_dtl_lock);
2477 
2478 		/*
2479 		 * We only destroy the leaf ZAP for detached leaves or for
2480 		 * removed log devices. Removed data devices handle leaf ZAP
2481 		 * cleanup later, once cancellation is no longer possible.
2482 		 */
2483 		if (vd->vdev_leaf_zap != 0 && (vd->vdev_detached ||
2484 		    vd->vdev_top->vdev_islog)) {
2485 			vdev_destroy_unlink_zap(vd, vd->vdev_leaf_zap, tx);
2486 			vd->vdev_leaf_zap = 0;
2487 		}
2488 
2489 		dmu_tx_commit(tx);
2490 		return;
2491 	}
2492 
2493 	if (vd->vdev_dtl_sm == NULL) {
2494 		uint64_t new_object;
2495 
2496 		new_object = space_map_alloc(mos, vdev_dtl_sm_blksz, tx);
2497 		VERIFY3U(new_object, !=, 0);
2498 
2499 		VERIFY0(space_map_open(&vd->vdev_dtl_sm, mos, new_object,
2500 		    0, -1ULL, 0));
2501 		ASSERT(vd->vdev_dtl_sm != NULL);
2502 	}
2503 
2504 	rtsync = range_tree_create(NULL, NULL);
2505 
2506 	mutex_enter(&vd->vdev_dtl_lock);
2507 	range_tree_walk(rt, range_tree_add, rtsync);
2508 	mutex_exit(&vd->vdev_dtl_lock);
2509 
2510 	space_map_truncate(vd->vdev_dtl_sm, vdev_dtl_sm_blksz, tx);
2511 	space_map_write(vd->vdev_dtl_sm, rtsync, SM_ALLOC, SM_NO_VDEVID, tx);
2512 	range_tree_vacate(rtsync, NULL, NULL);
2513 
2514 	range_tree_destroy(rtsync);
2515 
2516 	/*
2517 	 * If the object for the space map has changed then dirty
2518 	 * the top level so that we update the config.
2519 	 */
2520 	if (object != space_map_object(vd->vdev_dtl_sm)) {
2521 		vdev_dbgmsg(vd, "txg %llu, spa %s, DTL old object %llu, "
2522 		    "new object %llu", (u_longlong_t)txg, spa_name(spa),
2523 		    (u_longlong_t)object,
2524 		    (u_longlong_t)space_map_object(vd->vdev_dtl_sm));
2525 		vdev_config_dirty(vd->vdev_top);
2526 	}
2527 
2528 	dmu_tx_commit(tx);
2529 
2530 	mutex_enter(&vd->vdev_dtl_lock);
2531 	space_map_update(vd->vdev_dtl_sm);
2532 	mutex_exit(&vd->vdev_dtl_lock);
2533 }
2534 
2535 /*
2536  * Determine whether the specified vdev can be offlined/detached/removed
2537  * without losing data.
2538  */
2539 boolean_t
2540 vdev_dtl_required(vdev_t *vd)
2541 {
2542 	spa_t *spa = vd->vdev_spa;
2543 	vdev_t *tvd = vd->vdev_top;
2544 	uint8_t cant_read = vd->vdev_cant_read;
2545 	boolean_t required;
2546 
2547 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2548 
2549 	if (vd == spa->spa_root_vdev || vd == tvd)
2550 		return (B_TRUE);
2551 
2552 	/*
2553 	 * Temporarily mark the device as unreadable, and then determine
2554 	 * whether this results in any DTL outages in the top-level vdev.
2555 	 * If not, we can safely offline/detach/remove the device.
2556 	 */
2557 	vd->vdev_cant_read = B_TRUE;
2558 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
2559 	required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
2560 	vd->vdev_cant_read = cant_read;
2561 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
2562 
2563 	if (!required && zio_injection_enabled)
2564 		required = !!zio_handle_device_injection(vd, NULL, ECHILD);
2565 
2566 	return (required);
2567 }
2568 
2569 /*
2570  * Determine if resilver is needed, and if so the txg range.
2571  */
2572 boolean_t
2573 vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
2574 {
2575 	boolean_t needed = B_FALSE;
2576 	uint64_t thismin = UINT64_MAX;
2577 	uint64_t thismax = 0;
2578 
2579 	if (vd->vdev_children == 0) {
2580 		mutex_enter(&vd->vdev_dtl_lock);
2581 		if (!range_tree_is_empty(vd->vdev_dtl[DTL_MISSING]) &&
2582 		    vdev_writeable(vd)) {
2583 
2584 			thismin = vdev_dtl_min(vd);
2585 			thismax = vdev_dtl_max(vd);
2586 			needed = B_TRUE;
2587 		}
2588 		mutex_exit(&vd->vdev_dtl_lock);
2589 	} else {
2590 		for (int c = 0; c < vd->vdev_children; c++) {
2591 			vdev_t *cvd = vd->vdev_child[c];
2592 			uint64_t cmin, cmax;
2593 
2594 			if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
2595 				thismin = MIN(thismin, cmin);
2596 				thismax = MAX(thismax, cmax);
2597 				needed = B_TRUE;
2598 			}
2599 		}
2600 	}
2601 
2602 	if (needed && minp) {
2603 		*minp = thismin;
2604 		*maxp = thismax;
2605 	}
2606 	return (needed);
2607 }
2608 
2609 /*
2610  * Gets the checkpoint space map object from the vdev's ZAP.
2611  * Returns the spacemap object, or 0 if it wasn't in the ZAP
2612  * or the ZAP doesn't exist yet.
2613  */
2614 int
2615 vdev_checkpoint_sm_object(vdev_t *vd)
2616 {
2617 	ASSERT0(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER));
2618 	if (vd->vdev_top_zap == 0) {
2619 		return (0);
2620 	}
2621 
2622 	uint64_t sm_obj = 0;
2623 	int err = zap_lookup(spa_meta_objset(vd->vdev_spa), vd->vdev_top_zap,
2624 	    VDEV_TOP_ZAP_POOL_CHECKPOINT_SM, sizeof (uint64_t), 1, &sm_obj);
2625 
2626 	ASSERT(err == 0 || err == ENOENT);
2627 
2628 	return (sm_obj);
2629 }
2630 
2631 int
2632 vdev_load(vdev_t *vd)
2633 {
2634 	int error = 0;
2635 	/*
2636 	 * Recursively load all children.
2637 	 */
2638 	for (int c = 0; c < vd->vdev_children; c++) {
2639 		error = vdev_load(vd->vdev_child[c]);
2640 		if (error != 0) {
2641 			return (error);
2642 		}
2643 	}
2644 
2645 	vdev_set_deflate_ratio(vd);
2646 
2647 	/*
2648 	 * If this is a top-level vdev, initialize its metaslabs.
2649 	 */
2650 	if (vd == vd->vdev_top && vdev_is_concrete(vd)) {
2651 		if (vd->vdev_ashift == 0 || vd->vdev_asize == 0) {
2652 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2653 			    VDEV_AUX_CORRUPT_DATA);
2654 			vdev_dbgmsg(vd, "vdev_load: invalid size. ashift=%llu, "
2655 			    "asize=%llu", (u_longlong_t)vd->vdev_ashift,
2656 			    (u_longlong_t)vd->vdev_asize);
2657 			return (SET_ERROR(ENXIO));
2658 		} else if ((error = vdev_metaslab_init(vd, 0)) != 0) {
2659 			vdev_dbgmsg(vd, "vdev_load: metaslab_init failed "
2660 			    "[error=%d]", error);
2661 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2662 			    VDEV_AUX_CORRUPT_DATA);
2663 			return (error);
2664 		}
2665 
2666 		uint64_t checkpoint_sm_obj = vdev_checkpoint_sm_object(vd);
2667 		if (checkpoint_sm_obj != 0) {
2668 			objset_t *mos = spa_meta_objset(vd->vdev_spa);
2669 			ASSERT(vd->vdev_asize != 0);
2670 			ASSERT3P(vd->vdev_checkpoint_sm, ==, NULL);
2671 
2672 			if ((error = space_map_open(&vd->vdev_checkpoint_sm,
2673 			    mos, checkpoint_sm_obj, 0, vd->vdev_asize,
2674 			    vd->vdev_ashift))) {
2675 				vdev_dbgmsg(vd, "vdev_load: space_map_open "
2676 				    "failed for checkpoint spacemap (obj %llu) "
2677 				    "[error=%d]",
2678 				    (u_longlong_t)checkpoint_sm_obj, error);
2679 				return (error);
2680 			}
2681 			ASSERT3P(vd->vdev_checkpoint_sm, !=, NULL);
2682 			space_map_update(vd->vdev_checkpoint_sm);
2683 
2684 			/*
2685 			 * Since the checkpoint_sm contains free entries
2686 			 * exclusively we can use sm_alloc to indicate the
2687 			 * culmulative checkpointed space that has been freed.
2688 			 */
2689 			vd->vdev_stat.vs_checkpoint_space =
2690 			    -vd->vdev_checkpoint_sm->sm_alloc;
2691 			vd->vdev_spa->spa_checkpoint_info.sci_dspace +=
2692 			    vd->vdev_stat.vs_checkpoint_space;
2693 		}
2694 	}
2695 
2696 	/*
2697 	 * If this is a leaf vdev, load its DTL.
2698 	 */
2699 	if (vd->vdev_ops->vdev_op_leaf && (error = vdev_dtl_load(vd)) != 0) {
2700 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2701 		    VDEV_AUX_CORRUPT_DATA);
2702 		vdev_dbgmsg(vd, "vdev_load: vdev_dtl_load failed "
2703 		    "[error=%d]", error);
2704 		return (error);
2705 	}
2706 
2707 	uint64_t obsolete_sm_object = vdev_obsolete_sm_object(vd);
2708 	if (obsolete_sm_object != 0) {
2709 		objset_t *mos = vd->vdev_spa->spa_meta_objset;
2710 		ASSERT(vd->vdev_asize != 0);
2711 		ASSERT3P(vd->vdev_obsolete_sm, ==, NULL);
2712 
2713 		if ((error = space_map_open(&vd->vdev_obsolete_sm, mos,
2714 		    obsolete_sm_object, 0, vd->vdev_asize, 0))) {
2715 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2716 			    VDEV_AUX_CORRUPT_DATA);
2717 			vdev_dbgmsg(vd, "vdev_load: space_map_open failed for "
2718 			    "obsolete spacemap (obj %llu) [error=%d]",
2719 			    (u_longlong_t)obsolete_sm_object, error);
2720 			return (error);
2721 		}
2722 		space_map_update(vd->vdev_obsolete_sm);
2723 	}
2724 
2725 	return (0);
2726 }
2727 
2728 /*
2729  * The special vdev case is used for hot spares and l2cache devices.  Its
2730  * sole purpose it to set the vdev state for the associated vdev.  To do this,
2731  * we make sure that we can open the underlying device, then try to read the
2732  * label, and make sure that the label is sane and that it hasn't been
2733  * repurposed to another pool.
2734  */
2735 int
2736 vdev_validate_aux(vdev_t *vd)
2737 {
2738 	nvlist_t *label;
2739 	uint64_t guid, version;
2740 	uint64_t state;
2741 
2742 	if (!vdev_readable(vd))
2743 		return (0);
2744 
2745 	if ((label = vdev_label_read_config(vd, -1ULL)) == NULL) {
2746 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2747 		    VDEV_AUX_CORRUPT_DATA);
2748 		return (-1);
2749 	}
2750 
2751 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
2752 	    !SPA_VERSION_IS_SUPPORTED(version) ||
2753 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
2754 	    guid != vd->vdev_guid ||
2755 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
2756 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2757 		    VDEV_AUX_CORRUPT_DATA);
2758 		nvlist_free(label);
2759 		return (-1);
2760 	}
2761 
2762 	/*
2763 	 * We don't actually check the pool state here.  If it's in fact in
2764 	 * use by another pool, we update this fact on the fly when requested.
2765 	 */
2766 	nvlist_free(label);
2767 	return (0);
2768 }
2769 
2770 /*
2771  * Free the objects used to store this vdev's spacemaps, and the array
2772  * that points to them.
2773  */
2774 void
2775 vdev_destroy_spacemaps(vdev_t *vd, dmu_tx_t *tx)
2776 {
2777 	if (vd->vdev_ms_array == 0)
2778 		return;
2779 
2780 	objset_t *mos = vd->vdev_spa->spa_meta_objset;
2781 	uint64_t array_count = vd->vdev_asize >> vd->vdev_ms_shift;
2782 	size_t array_bytes = array_count * sizeof (uint64_t);
2783 	uint64_t *smobj_array = kmem_alloc(array_bytes, KM_SLEEP);
2784 	VERIFY0(dmu_read(mos, vd->vdev_ms_array, 0,
2785 	    array_bytes, smobj_array, 0));
2786 
2787 	for (uint64_t i = 0; i < array_count; i++) {
2788 		uint64_t smobj = smobj_array[i];
2789 		if (smobj == 0)
2790 			continue;
2791 
2792 		space_map_free_obj(mos, smobj, tx);
2793 	}
2794 
2795 	kmem_free(smobj_array, array_bytes);
2796 	VERIFY0(dmu_object_free(mos, vd->vdev_ms_array, tx));
2797 	vd->vdev_ms_array = 0;
2798 }
2799 
2800 static void
2801 vdev_remove_empty(vdev_t *vd, uint64_t txg)
2802 {
2803 	spa_t *spa = vd->vdev_spa;
2804 	dmu_tx_t *tx;
2805 
2806 	ASSERT(vd == vd->vdev_top);
2807 	ASSERT3U(txg, ==, spa_syncing_txg(spa));
2808 
2809 	if (vd->vdev_ms != NULL) {
2810 		metaslab_group_t *mg = vd->vdev_mg;
2811 
2812 		metaslab_group_histogram_verify(mg);
2813 		metaslab_class_histogram_verify(mg->mg_class);
2814 
2815 		for (int m = 0; m < vd->vdev_ms_count; m++) {
2816 			metaslab_t *msp = vd->vdev_ms[m];
2817 
2818 			if (msp == NULL || msp->ms_sm == NULL)
2819 				continue;
2820 
2821 			mutex_enter(&msp->ms_lock);
2822 			/*
2823 			 * If the metaslab was not loaded when the vdev
2824 			 * was removed then the histogram accounting may
2825 			 * not be accurate. Update the histogram information
2826 			 * here so that we ensure that the metaslab group
2827 			 * and metaslab class are up-to-date.
2828 			 */
2829 			metaslab_group_histogram_remove(mg, msp);
2830 
2831 			VERIFY0(space_map_allocated(msp->ms_sm));
2832 			space_map_close(msp->ms_sm);
2833 			msp->ms_sm = NULL;
2834 			mutex_exit(&msp->ms_lock);
2835 		}
2836 
2837 		if (vd->vdev_checkpoint_sm != NULL) {
2838 			ASSERT(spa_has_checkpoint(spa));
2839 			space_map_close(vd->vdev_checkpoint_sm);
2840 			vd->vdev_checkpoint_sm = NULL;
2841 		}
2842 
2843 		metaslab_group_histogram_verify(mg);
2844 		metaslab_class_histogram_verify(mg->mg_class);
2845 		for (int i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
2846 			ASSERT0(mg->mg_histogram[i]);
2847 	}
2848 
2849 	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2850 	vdev_destroy_spacemaps(vd, tx);
2851 
2852 	if (vd->vdev_islog && vd->vdev_top_zap != 0) {
2853 		vdev_destroy_unlink_zap(vd, vd->vdev_top_zap, tx);
2854 		vd->vdev_top_zap = 0;
2855 	}
2856 	dmu_tx_commit(tx);
2857 }
2858 
2859 void
2860 vdev_sync_done(vdev_t *vd, uint64_t txg)
2861 {
2862 	metaslab_t *msp;
2863 	boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg));
2864 
2865 	ASSERT(vdev_is_concrete(vd));
2866 
2867 	while ((msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
2868 	    != NULL)
2869 		metaslab_sync_done(msp, txg);
2870 
2871 	if (reassess)
2872 		metaslab_sync_reassess(vd->vdev_mg);
2873 }
2874 
2875 void
2876 vdev_sync(vdev_t *vd, uint64_t txg)
2877 {
2878 	spa_t *spa = vd->vdev_spa;
2879 	vdev_t *lvd;
2880 	metaslab_t *msp;
2881 	dmu_tx_t *tx;
2882 
2883 	if (range_tree_space(vd->vdev_obsolete_segments) > 0) {
2884 		dmu_tx_t *tx;
2885 
2886 		ASSERT(vd->vdev_removing ||
2887 		    vd->vdev_ops == &vdev_indirect_ops);
2888 
2889 		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2890 		vdev_indirect_sync_obsolete(vd, tx);
2891 		dmu_tx_commit(tx);
2892 
2893 		/*
2894 		 * If the vdev is indirect, it can't have dirty
2895 		 * metaslabs or DTLs.
2896 		 */
2897 		if (vd->vdev_ops == &vdev_indirect_ops) {
2898 			ASSERT(txg_list_empty(&vd->vdev_ms_list, txg));
2899 			ASSERT(txg_list_empty(&vd->vdev_dtl_list, txg));
2900 			return;
2901 		}
2902 	}
2903 
2904 	ASSERT(vdev_is_concrete(vd));
2905 
2906 	if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0 &&
2907 	    !vd->vdev_removing) {
2908 		ASSERT(vd == vd->vdev_top);
2909 		ASSERT0(vd->vdev_indirect_config.vic_mapping_object);
2910 		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2911 		vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
2912 		    DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
2913 		ASSERT(vd->vdev_ms_array != 0);
2914 		vdev_config_dirty(vd);
2915 		dmu_tx_commit(tx);
2916 	}
2917 
2918 	while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
2919 		metaslab_sync(msp, txg);
2920 		(void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
2921 	}
2922 
2923 	while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
2924 		vdev_dtl_sync(lvd, txg);
2925 
2926 	/*
2927 	 * Remove the metadata associated with this vdev once it's empty.
2928 	 * Note that this is typically used for log/cache device removal;
2929 	 * we don't empty toplevel vdevs when removing them.  But if
2930 	 * a toplevel happens to be emptied, this is not harmful.
2931 	 */
2932 	if (vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing) {
2933 		vdev_remove_empty(vd, txg);
2934 	}
2935 
2936 	(void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
2937 }
2938 
2939 uint64_t
2940 vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
2941 {
2942 	return (vd->vdev_ops->vdev_op_asize(vd, psize));
2943 }
2944 
2945 /*
2946  * Mark the given vdev faulted.  A faulted vdev behaves as if the device could
2947  * not be opened, and no I/O is attempted.
2948  */
2949 int
2950 vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2951 {
2952 	vdev_t *vd, *tvd;
2953 
2954 	spa_vdev_state_enter(spa, SCL_NONE);
2955 
2956 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2957 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2958 
2959 	if (!vd->vdev_ops->vdev_op_leaf)
2960 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2961 
2962 	tvd = vd->vdev_top;
2963 
2964 	/*
2965 	 * We don't directly use the aux state here, but if we do a
2966 	 * vdev_reopen(), we need this value to be present to remember why we
2967 	 * were faulted.
2968 	 */
2969 	vd->vdev_label_aux = aux;
2970 
2971 	/*
2972 	 * Faulted state takes precedence over degraded.
2973 	 */
2974 	vd->vdev_delayed_close = B_FALSE;
2975 	vd->vdev_faulted = 1ULL;
2976 	vd->vdev_degraded = 0ULL;
2977 	vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
2978 
2979 	/*
2980 	 * If this device has the only valid copy of the data, then
2981 	 * back off and simply mark the vdev as degraded instead.
2982 	 */
2983 	if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) {
2984 		vd->vdev_degraded = 1ULL;
2985 		vd->vdev_faulted = 0ULL;
2986 
2987 		/*
2988 		 * If we reopen the device and it's not dead, only then do we
2989 		 * mark it degraded.
2990 		 */
2991 		vdev_reopen(tvd);
2992 
2993 		if (vdev_readable(vd))
2994 			vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
2995 	}
2996 
2997 	return (spa_vdev_state_exit(spa, vd, 0));
2998 }
2999 
3000 /*
3001  * Mark the given vdev degraded.  A degraded vdev is purely an indication to the
3002  * user that something is wrong.  The vdev continues to operate as normal as far
3003  * as I/O is concerned.
3004  */
3005 int
3006 vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
3007 {
3008 	vdev_t *vd;
3009 
3010 	spa_vdev_state_enter(spa, SCL_NONE);
3011 
3012 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
3013 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
3014 
3015 	if (!vd->vdev_ops->vdev_op_leaf)
3016 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
3017 
3018 	/*
3019 	 * If the vdev is already faulted, then don't do anything.
3020 	 */
3021 	if (vd->vdev_faulted || vd->vdev_degraded)
3022 		return (spa_vdev_state_exit(spa, NULL, 0));
3023 
3024 	vd->vdev_degraded = 1ULL;
3025 	if (!vdev_is_dead(vd))
3026 		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
3027 		    aux);
3028 
3029 	return (spa_vdev_state_exit(spa, vd, 0));
3030 }
3031 
3032 /*
3033  * Online the given vdev.
3034  *
3035  * If 'ZFS_ONLINE_UNSPARE' is set, it implies two things.  First, any attached
3036  * spare device should be detached when the device finishes resilvering.
3037  * Second, the online should be treated like a 'test' online case, so no FMA
3038  * events are generated if the device fails to open.
3039  */
3040 int
3041 vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
3042 {
3043 	vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
3044 	boolean_t wasoffline;
3045 	vdev_state_t oldstate;
3046 
3047 	spa_vdev_state_enter(spa, SCL_NONE);
3048 
3049 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
3050 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
3051 
3052 	if (!vd->vdev_ops->vdev_op_leaf)
3053 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
3054 
3055 	wasoffline = (vd->vdev_offline || vd->vdev_tmpoffline);
3056 	oldstate = vd->vdev_state;
3057 
3058 	tvd = vd->vdev_top;
3059 	vd->vdev_offline = B_FALSE;
3060 	vd->vdev_tmpoffline = B_FALSE;
3061 	vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
3062 	vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
3063 
3064 	/* XXX - L2ARC 1.0 does not support expansion */
3065 	if (!vd->vdev_aux) {
3066 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
3067 			pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
3068 	}
3069 
3070 	vdev_reopen(tvd);
3071 	vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
3072 
3073 	if (!vd->vdev_aux) {
3074 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
3075 			pvd->vdev_expanding = B_FALSE;
3076 	}
3077 
3078 	if (newstate)
3079 		*newstate = vd->vdev_state;
3080 	if ((flags & ZFS_ONLINE_UNSPARE) &&
3081 	    !vdev_is_dead(vd) && vd->vdev_parent &&
3082 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
3083 	    vd->vdev_parent->vdev_child[0] == vd)
3084 		vd->vdev_unspare = B_TRUE;
3085 
3086 	if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
3087 
3088 		/* XXX - L2ARC 1.0 does not support expansion */
3089 		if (vd->vdev_aux)
3090 			return (spa_vdev_state_exit(spa, vd, ENOTSUP));
3091 		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
3092 	}
3093 
3094 	/* Restart initializing if necessary */
3095 	mutex_enter(&vd->vdev_initialize_lock);
3096 	if (vdev_writeable(vd) &&
3097 	    vd->vdev_initialize_thread == NULL &&
3098 	    vd->vdev_initialize_state == VDEV_INITIALIZE_ACTIVE) {
3099 		(void) vdev_initialize(vd);
3100 	}
3101 	mutex_exit(&vd->vdev_initialize_lock);
3102 
3103 	if (wasoffline ||
3104 	    (oldstate < VDEV_STATE_DEGRADED &&
3105 	    vd->vdev_state >= VDEV_STATE_DEGRADED))
3106 		spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_ONLINE);
3107 
3108 	return (spa_vdev_state_exit(spa, vd, 0));
3109 }
3110 
3111 static int
3112 vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
3113 {
3114 	vdev_t *vd, *tvd;
3115 	int error = 0;
3116 	uint64_t generation;
3117 	metaslab_group_t *mg;
3118 
3119 top:
3120 	spa_vdev_state_enter(spa, SCL_ALLOC);
3121 
3122 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
3123 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
3124 
3125 	if (!vd->vdev_ops->vdev_op_leaf)
3126 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
3127 
3128 	tvd = vd->vdev_top;
3129 	mg = tvd->vdev_mg;
3130 	generation = spa->spa_config_generation + 1;
3131 
3132 	/*
3133 	 * If the device isn't already offline, try to offline it.
3134 	 */
3135 	if (!vd->vdev_offline) {
3136 		/*
3137 		 * If this device has the only valid copy of some data,
3138 		 * don't allow it to be offlined. Log devices are always
3139 		 * expendable.
3140 		 */
3141 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
3142 		    vdev_dtl_required(vd))
3143 			return (spa_vdev_state_exit(spa, NULL, EBUSY));
3144 
3145 		/*
3146 		 * If the top-level is a slog and it has had allocations
3147 		 * then proceed.  We check that the vdev's metaslab group
3148 		 * is not NULL since it's possible that we may have just
3149 		 * added this vdev but not yet initialized its metaslabs.
3150 		 */
3151 		if (tvd->vdev_islog && mg != NULL) {
3152 			/*
3153 			 * Prevent any future allocations.
3154 			 */
3155 			metaslab_group_passivate(mg);
3156 			(void) spa_vdev_state_exit(spa, vd, 0);
3157 
3158 			error = spa_reset_logs(spa);
3159 
3160 			/*
3161 			 * If the log device was successfully reset but has
3162 			 * checkpointed data, do not offline it.
3163 			 */
3164 			if (error == 0 &&
3165 			    tvd->vdev_checkpoint_sm != NULL) {
3166 				ASSERT3U(tvd->vdev_checkpoint_sm->sm_alloc,
3167 				    !=, 0);
3168 				error = ZFS_ERR_CHECKPOINT_EXISTS;
3169 			}
3170 
3171 			spa_vdev_state_enter(spa, SCL_ALLOC);
3172 
3173 			/*
3174 			 * Check to see if the config has changed.
3175 			 */
3176 			if (error || generation != spa->spa_config_generation) {
3177 				metaslab_group_activate(mg);
3178 				if (error)
3179 					return (spa_vdev_state_exit(spa,
3180 					    vd, error));
3181 				(void) spa_vdev_state_exit(spa, vd, 0);
3182 				goto top;
3183 			}
3184 			ASSERT0(tvd->vdev_stat.vs_alloc);
3185 		}
3186 
3187 		/*
3188 		 * Offline this device and reopen its top-level vdev.
3189 		 * If the top-level vdev is a log device then just offline
3190 		 * it. Otherwise, if this action results in the top-level
3191 		 * vdev becoming unusable, undo it and fail the request.
3192 		 */
3193 		vd->vdev_offline = B_TRUE;
3194 		vdev_reopen(tvd);
3195 
3196 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
3197 		    vdev_is_dead(tvd)) {
3198 			vd->vdev_offline = B_FALSE;
3199 			vdev_reopen(tvd);
3200 			return (spa_vdev_state_exit(spa, NULL, EBUSY));
3201 		}
3202 
3203 		/*
3204 		 * Add the device back into the metaslab rotor so that
3205 		 * once we online the device it's open for business.
3206 		 */
3207 		if (tvd->vdev_islog && mg != NULL)
3208 			metaslab_group_activate(mg);
3209 	}
3210 
3211 	vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
3212 
3213 	return (spa_vdev_state_exit(spa, vd, 0));
3214 }
3215 
3216 int
3217 vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
3218 {
3219 	int error;
3220 
3221 	mutex_enter(&spa->spa_vdev_top_lock);
3222 	error = vdev_offline_locked(spa, guid, flags);
3223 	mutex_exit(&spa->spa_vdev_top_lock);
3224 
3225 	return (error);
3226 }
3227 
3228 /*
3229  * Clear the error counts associated with this vdev.  Unlike vdev_online() and
3230  * vdev_offline(), we assume the spa config is locked.  We also clear all
3231  * children.  If 'vd' is NULL, then the user wants to clear all vdevs.
3232  */
3233 void
3234 vdev_clear(spa_t *spa, vdev_t *vd)
3235 {
3236 	vdev_t *rvd = spa->spa_root_vdev;
3237 
3238 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3239 
3240 	if (vd == NULL)
3241 		vd = rvd;
3242 
3243 	vd->vdev_stat.vs_read_errors = 0;
3244 	vd->vdev_stat.vs_write_errors = 0;
3245 	vd->vdev_stat.vs_checksum_errors = 0;
3246 
3247 	for (int c = 0; c < vd->vdev_children; c++)
3248 		vdev_clear(spa, vd->vdev_child[c]);
3249 
3250 	/*
3251 	 * It makes no sense to "clear" an indirect vdev.
3252 	 */
3253 	if (!vdev_is_concrete(vd))
3254 		return;
3255 
3256 	/*
3257 	 * If we're in the FAULTED state or have experienced failed I/O, then
3258 	 * clear the persistent state and attempt to reopen the device.  We
3259 	 * also mark the vdev config dirty, so that the new faulted state is
3260 	 * written out to disk.
3261 	 */
3262 	if (vd->vdev_faulted || vd->vdev_degraded ||
3263 	    !vdev_readable(vd) || !vdev_writeable(vd)) {
3264 
3265 		/*
3266 		 * When reopening in reponse to a clear event, it may be due to
3267 		 * a fmadm repair request.  In this case, if the device is
3268 		 * still broken, we want to still post the ereport again.
3269 		 */
3270 		vd->vdev_forcefault = B_TRUE;
3271 
3272 		vd->vdev_faulted = vd->vdev_degraded = 0ULL;
3273 		vd->vdev_cant_read = B_FALSE;
3274 		vd->vdev_cant_write = B_FALSE;
3275 
3276 		vdev_reopen(vd == rvd ? rvd : vd->vdev_top);
3277 
3278 		vd->vdev_forcefault = B_FALSE;
3279 
3280 		if (vd != rvd && vdev_writeable(vd->vdev_top))
3281 			vdev_state_dirty(vd->vdev_top);
3282 
3283 		if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
3284 			spa_async_request(spa, SPA_ASYNC_RESILVER);
3285 
3286 		spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_CLEAR);
3287 	}
3288 
3289 	/*
3290 	 * When clearing a FMA-diagnosed fault, we always want to
3291 	 * unspare the device, as we assume that the original spare was
3292 	 * done in response to the FMA fault.
3293 	 */
3294 	if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
3295 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
3296 	    vd->vdev_parent->vdev_child[0] == vd)
3297 		vd->vdev_unspare = B_TRUE;
3298 }
3299 
3300 boolean_t
3301 vdev_is_dead(vdev_t *vd)
3302 {
3303 	/*
3304 	 * Holes and missing devices are always considered "dead".
3305 	 * This simplifies the code since we don't have to check for
3306 	 * these types of devices in the various code paths.
3307 	 * Instead we rely on the fact that we skip over dead devices
3308 	 * before issuing I/O to them.
3309 	 */
3310 	return (vd->vdev_state < VDEV_STATE_DEGRADED ||
3311 	    vd->vdev_ops == &vdev_hole_ops ||
3312 	    vd->vdev_ops == &vdev_missing_ops);
3313 }
3314 
3315 boolean_t
3316 vdev_readable(vdev_t *vd)
3317 {
3318 	return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
3319 }
3320 
3321 boolean_t
3322 vdev_writeable(vdev_t *vd)
3323 {
3324 	return (!vdev_is_dead(vd) && !vd->vdev_cant_write &&
3325 	    vdev_is_concrete(vd));
3326 }
3327 
3328 boolean_t
3329 vdev_allocatable(vdev_t *vd)
3330 {
3331 	uint64_t state = vd->vdev_state;
3332 
3333 	/*
3334 	 * We currently allow allocations from vdevs which may be in the
3335 	 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
3336 	 * fails to reopen then we'll catch it later when we're holding
3337 	 * the proper locks.  Note that we have to get the vdev state
3338 	 * in a local variable because although it changes atomically,
3339 	 * we're asking two separate questions about it.
3340 	 */
3341 	return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
3342 	    !vd->vdev_cant_write && vdev_is_concrete(vd) &&
3343 	    vd->vdev_mg->mg_initialized);
3344 }
3345 
3346 boolean_t
3347 vdev_accessible(vdev_t *vd, zio_t *zio)
3348 {
3349 	ASSERT(zio->io_vd == vd);
3350 
3351 	if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
3352 		return (B_FALSE);
3353 
3354 	if (zio->io_type == ZIO_TYPE_READ)
3355 		return (!vd->vdev_cant_read);
3356 
3357 	if (zio->io_type == ZIO_TYPE_WRITE)
3358 		return (!vd->vdev_cant_write);
3359 
3360 	return (B_TRUE);
3361 }
3362 
3363 boolean_t
3364 vdev_is_spacemap_addressable(vdev_t *vd)
3365 {
3366 	/*
3367 	 * Assuming 47 bits of the space map entry dedicated for the entry's
3368 	 * offset (see description in space_map.h), we calculate the maximum
3369 	 * address that can be described by a space map entry for the given
3370 	 * device.
3371 	 */
3372 	uint64_t shift = vd->vdev_ashift + 47;
3373 
3374 	if (shift >= 63) /* detect potential overflow */
3375 		return (B_TRUE);
3376 
3377 	return (vd->vdev_asize < (1ULL << shift));
3378 }
3379 
3380 /*
3381  * Get statistics for the given vdev.
3382  */
3383 void
3384 vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
3385 {
3386 	spa_t *spa = vd->vdev_spa;
3387 	vdev_t *rvd = spa->spa_root_vdev;
3388 	vdev_t *tvd = vd->vdev_top;
3389 
3390 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
3391 
3392 	mutex_enter(&vd->vdev_stat_lock);
3393 	bcopy(&vd->vdev_stat, vs, sizeof (*vs));
3394 	vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
3395 	vs->vs_state = vd->vdev_state;
3396 	vs->vs_rsize = vdev_get_min_asize(vd);
3397 	if (vd->vdev_ops->vdev_op_leaf) {
3398 		vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
3399 		/*
3400 		 * Report intializing progress. Since we don't have the
3401 		 * initializing locks held, this is only an estimate (although a
3402 		 * fairly accurate one).
3403 		 */
3404 		vs->vs_initialize_bytes_done = vd->vdev_initialize_bytes_done;
3405 		vs->vs_initialize_bytes_est = vd->vdev_initialize_bytes_est;
3406 		vs->vs_initialize_state = vd->vdev_initialize_state;
3407 		vs->vs_initialize_action_time = vd->vdev_initialize_action_time;
3408 	}
3409 	/*
3410 	 * Report expandable space on top-level, non-auxillary devices only.
3411 	 * The expandable space is reported in terms of metaslab sized units
3412 	 * since that determines how much space the pool can expand.
3413 	 */
3414 	if (vd->vdev_aux == NULL && tvd != NULL) {
3415 		vs->vs_esize = P2ALIGN(vd->vdev_max_asize - vd->vdev_asize -
3416 		    spa->spa_bootsize, 1ULL << tvd->vdev_ms_shift);
3417 	}
3418 	if (vd->vdev_aux == NULL && vd == vd->vdev_top &&
3419 	    vdev_is_concrete(vd)) {
3420 		vs->vs_fragmentation = vd->vdev_mg->mg_fragmentation;
3421 	}
3422 
3423 	/*
3424 	 * If we're getting stats on the root vdev, aggregate the I/O counts
3425 	 * over all top-level vdevs (i.e. the direct children of the root).
3426 	 */
3427 	if (vd == rvd) {
3428 		for (int c = 0; c < rvd->vdev_children; c++) {
3429 			vdev_t *cvd = rvd->vdev_child[c];
3430 			vdev_stat_t *cvs = &cvd->vdev_stat;
3431 
3432 			for (int t = 0; t < ZIO_TYPES; t++) {
3433 				vs->vs_ops[t] += cvs->vs_ops[t];
3434 				vs->vs_bytes[t] += cvs->vs_bytes[t];
3435 			}
3436 			cvs->vs_scan_removing = cvd->vdev_removing;
3437 		}
3438 	}
3439 	mutex_exit(&vd->vdev_stat_lock);
3440 }
3441 
3442 void
3443 vdev_clear_stats(vdev_t *vd)
3444 {
3445 	mutex_enter(&vd->vdev_stat_lock);
3446 	vd->vdev_stat.vs_space = 0;
3447 	vd->vdev_stat.vs_dspace = 0;
3448 	vd->vdev_stat.vs_alloc = 0;
3449 	mutex_exit(&vd->vdev_stat_lock);
3450 }
3451 
3452 void
3453 vdev_scan_stat_init(vdev_t *vd)
3454 {
3455 	vdev_stat_t *vs = &vd->vdev_stat;
3456 
3457 	for (int c = 0; c < vd->vdev_children; c++)
3458 		vdev_scan_stat_init(vd->vdev_child[c]);
3459 
3460 	mutex_enter(&vd->vdev_stat_lock);
3461 	vs->vs_scan_processed = 0;
3462 	mutex_exit(&vd->vdev_stat_lock);
3463 }
3464 
3465 void
3466 vdev_stat_update(zio_t *zio, uint64_t psize)
3467 {
3468 	spa_t *spa = zio->io_spa;
3469 	vdev_t *rvd = spa->spa_root_vdev;
3470 	vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
3471 	vdev_t *pvd;
3472 	uint64_t txg = zio->io_txg;
3473 	vdev_stat_t *vs = &vd->vdev_stat;
3474 	zio_type_t type = zio->io_type;
3475 	int flags = zio->io_flags;
3476 
3477 	/*
3478 	 * If this i/o is a gang leader, it didn't do any actual work.
3479 	 */
3480 	if (zio->io_gang_tree)
3481 		return;
3482 
3483 	if (zio->io_error == 0) {
3484 		/*
3485 		 * If this is a root i/o, don't count it -- we've already
3486 		 * counted the top-level vdevs, and vdev_get_stats() will
3487 		 * aggregate them when asked.  This reduces contention on
3488 		 * the root vdev_stat_lock and implicitly handles blocks
3489 		 * that compress away to holes, for which there is no i/o.
3490 		 * (Holes never create vdev children, so all the counters
3491 		 * remain zero, which is what we want.)
3492 		 *
3493 		 * Note: this only applies to successful i/o (io_error == 0)
3494 		 * because unlike i/o counts, errors are not additive.
3495 		 * When reading a ditto block, for example, failure of
3496 		 * one top-level vdev does not imply a root-level error.
3497 		 */
3498 		if (vd == rvd)
3499 			return;
3500 
3501 		ASSERT(vd == zio->io_vd);
3502 
3503 		if (flags & ZIO_FLAG_IO_BYPASS)
3504 			return;
3505 
3506 		mutex_enter(&vd->vdev_stat_lock);
3507 
3508 		if (flags & ZIO_FLAG_IO_REPAIR) {
3509 			if (flags & ZIO_FLAG_SCAN_THREAD) {
3510 				dsl_scan_phys_t *scn_phys =
3511 				    &spa->spa_dsl_pool->dp_scan->scn_phys;
3512 				uint64_t *processed = &scn_phys->scn_processed;
3513 
3514 				/* XXX cleanup? */
3515 				if (vd->vdev_ops->vdev_op_leaf)
3516 					atomic_add_64(processed, psize);
3517 				vs->vs_scan_processed += psize;
3518 			}
3519 
3520 			if (flags & ZIO_FLAG_SELF_HEAL)
3521 				vs->vs_self_healed += psize;
3522 		}
3523 
3524 		vs->vs_ops[type]++;
3525 		vs->vs_bytes[type] += psize;
3526 
3527 		mutex_exit(&vd->vdev_stat_lock);
3528 		return;
3529 	}
3530 
3531 	if (flags & ZIO_FLAG_SPECULATIVE)
3532 		return;
3533 
3534 	/*
3535 	 * If this is an I/O error that is going to be retried, then ignore the
3536 	 * error.  Otherwise, the user may interpret B_FAILFAST I/O errors as
3537 	 * hard errors, when in reality they can happen for any number of
3538 	 * innocuous reasons (bus resets, MPxIO link failure, etc).
3539 	 */
3540 	if (zio->io_error == EIO &&
3541 	    !(zio->io_flags & ZIO_FLAG_IO_RETRY))
3542 		return;
3543 
3544 	/*
3545 	 * Intent logs writes won't propagate their error to the root
3546 	 * I/O so don't mark these types of failures as pool-level
3547 	 * errors.
3548 	 */
3549 	if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
3550 		return;
3551 
3552 	mutex_enter(&vd->vdev_stat_lock);
3553 	if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
3554 		if (zio->io_error == ECKSUM)
3555 			vs->vs_checksum_errors++;
3556 		else
3557 			vs->vs_read_errors++;
3558 	}
3559 	if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
3560 		vs->vs_write_errors++;
3561 	mutex_exit(&vd->vdev_stat_lock);
3562 
3563 	if (spa->spa_load_state == SPA_LOAD_NONE &&
3564 	    type == ZIO_TYPE_WRITE && txg != 0 &&
3565 	    (!(flags & ZIO_FLAG_IO_REPAIR) ||
3566 	    (flags & ZIO_FLAG_SCAN_THREAD) ||
3567 	    spa->spa_claiming)) {
3568 		/*
3569 		 * This is either a normal write (not a repair), or it's
3570 		 * a repair induced by the scrub thread, or it's a repair
3571 		 * made by zil_claim() during spa_load() in the first txg.
3572 		 * In the normal case, we commit the DTL change in the same
3573 		 * txg as the block was born.  In the scrub-induced repair
3574 		 * case, we know that scrubs run in first-pass syncing context,
3575 		 * so we commit the DTL change in spa_syncing_txg(spa).
3576 		 * In the zil_claim() case, we commit in spa_first_txg(spa).
3577 		 *
3578 		 * We currently do not make DTL entries for failed spontaneous
3579 		 * self-healing writes triggered by normal (non-scrubbing)
3580 		 * reads, because we have no transactional context in which to
3581 		 * do so -- and it's not clear that it'd be desirable anyway.
3582 		 */
3583 		if (vd->vdev_ops->vdev_op_leaf) {
3584 			uint64_t commit_txg = txg;
3585 			if (flags & ZIO_FLAG_SCAN_THREAD) {
3586 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
3587 				ASSERT(spa_sync_pass(spa) == 1);
3588 				vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
3589 				commit_txg = spa_syncing_txg(spa);
3590 			} else if (spa->spa_claiming) {
3591 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
3592 				commit_txg = spa_first_txg(spa);
3593 			}
3594 			ASSERT(commit_txg >= spa_syncing_txg(spa));
3595 			if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
3596 				return;
3597 			for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
3598 				vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
3599 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
3600 		}
3601 		if (vd != rvd)
3602 			vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
3603 	}
3604 }
3605 
3606 /*
3607  * Update the in-core space usage stats for this vdev, its metaslab class,
3608  * and the root vdev.
3609  */
3610 void
3611 vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
3612     int64_t space_delta)
3613 {
3614 	int64_t dspace_delta = space_delta;
3615 	spa_t *spa = vd->vdev_spa;
3616 	vdev_t *rvd = spa->spa_root_vdev;
3617 	metaslab_group_t *mg = vd->vdev_mg;
3618 	metaslab_class_t *mc = mg ? mg->mg_class : NULL;
3619 
3620 	ASSERT(vd == vd->vdev_top);
3621 
3622 	/*
3623 	 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
3624 	 * factor.  We must calculate this here and not at the root vdev
3625 	 * because the root vdev's psize-to-asize is simply the max of its
3626 	 * childrens', thus not accurate enough for us.
3627 	 */
3628 	ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
3629 	ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
3630 	dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
3631 	    vd->vdev_deflate_ratio;
3632 
3633 	mutex_enter(&vd->vdev_stat_lock);
3634 	vd->vdev_stat.vs_alloc += alloc_delta;
3635 	vd->vdev_stat.vs_space += space_delta;
3636 	vd->vdev_stat.vs_dspace += dspace_delta;
3637 	mutex_exit(&vd->vdev_stat_lock);
3638 
3639 	if (mc == spa_normal_class(spa)) {
3640 		mutex_enter(&rvd->vdev_stat_lock);
3641 		rvd->vdev_stat.vs_alloc += alloc_delta;
3642 		rvd->vdev_stat.vs_space += space_delta;
3643 		rvd->vdev_stat.vs_dspace += dspace_delta;
3644 		mutex_exit(&rvd->vdev_stat_lock);
3645 	}
3646 
3647 	if (mc != NULL) {
3648 		ASSERT(rvd == vd->vdev_parent);
3649 		ASSERT(vd->vdev_ms_count != 0);
3650 
3651 		metaslab_class_space_update(mc,
3652 		    alloc_delta, defer_delta, space_delta, dspace_delta);
3653 	}
3654 }
3655 
3656 /*
3657  * Mark a top-level vdev's config as dirty, placing it on the dirty list
3658  * so that it will be written out next time the vdev configuration is synced.
3659  * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
3660  */
3661 void
3662 vdev_config_dirty(vdev_t *vd)
3663 {
3664 	spa_t *spa = vd->vdev_spa;
3665 	vdev_t *rvd = spa->spa_root_vdev;
3666 	int c;
3667 
3668 	ASSERT(spa_writeable(spa));
3669 
3670 	/*
3671 	 * If this is an aux vdev (as with l2cache and spare devices), then we
3672 	 * update the vdev config manually and set the sync flag.
3673 	 */
3674 	if (vd->vdev_aux != NULL) {
3675 		spa_aux_vdev_t *sav = vd->vdev_aux;
3676 		nvlist_t **aux;
3677 		uint_t naux;
3678 
3679 		for (c = 0; c < sav->sav_count; c++) {
3680 			if (sav->sav_vdevs[c] == vd)
3681 				break;
3682 		}
3683 
3684 		if (c == sav->sav_count) {
3685 			/*
3686 			 * We're being removed.  There's nothing more to do.
3687 			 */
3688 			ASSERT(sav->sav_sync == B_TRUE);
3689 			return;
3690 		}
3691 
3692 		sav->sav_sync = B_TRUE;
3693 
3694 		if (nvlist_lookup_nvlist_array(sav->sav_config,
3695 		    ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
3696 			VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
3697 			    ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
3698 		}
3699 
3700 		ASSERT(c < naux);
3701 
3702 		/*
3703 		 * Setting the nvlist in the middle if the array is a little
3704 		 * sketchy, but it will work.
3705 		 */
3706 		nvlist_free(aux[c]);
3707 		aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0);
3708 
3709 		return;
3710 	}
3711 
3712 	/*
3713 	 * The dirty list is protected by the SCL_CONFIG lock.  The caller
3714 	 * must either hold SCL_CONFIG as writer, or must be the sync thread
3715 	 * (which holds SCL_CONFIG as reader).  There's only one sync thread,
3716 	 * so this is sufficient to ensure mutual exclusion.
3717 	 */
3718 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
3719 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3720 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
3721 
3722 	if (vd == rvd) {
3723 		for (c = 0; c < rvd->vdev_children; c++)
3724 			vdev_config_dirty(rvd->vdev_child[c]);
3725 	} else {
3726 		ASSERT(vd == vd->vdev_top);
3727 
3728 		if (!list_link_active(&vd->vdev_config_dirty_node) &&
3729 		    vdev_is_concrete(vd)) {
3730 			list_insert_head(&spa->spa_config_dirty_list, vd);
3731 		}
3732 	}
3733 }
3734 
3735 void
3736 vdev_config_clean(vdev_t *vd)
3737 {
3738 	spa_t *spa = vd->vdev_spa;
3739 
3740 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
3741 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3742 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
3743 
3744 	ASSERT(list_link_active(&vd->vdev_config_dirty_node));
3745 	list_remove(&spa->spa_config_dirty_list, vd);
3746 }
3747 
3748 /*
3749  * Mark a top-level vdev's state as dirty, so that the next pass of
3750  * spa_sync() can convert this into vdev_config_dirty().  We distinguish
3751  * the state changes from larger config changes because they require
3752  * much less locking, and are often needed for administrative actions.
3753  */
3754 void
3755 vdev_state_dirty(vdev_t *vd)
3756 {
3757 	spa_t *spa = vd->vdev_spa;
3758 
3759 	ASSERT(spa_writeable(spa));
3760 	ASSERT(vd == vd->vdev_top);
3761 
3762 	/*
3763 	 * The state list is protected by the SCL_STATE lock.  The caller
3764 	 * must either hold SCL_STATE as writer, or must be the sync thread
3765 	 * (which holds SCL_STATE as reader).  There's only one sync thread,
3766 	 * so this is sufficient to ensure mutual exclusion.
3767 	 */
3768 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
3769 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3770 	    spa_config_held(spa, SCL_STATE, RW_READER)));
3771 
3772 	if (!list_link_active(&vd->vdev_state_dirty_node) &&
3773 	    vdev_is_concrete(vd))
3774 		list_insert_head(&spa->spa_state_dirty_list, vd);
3775 }
3776 
3777 void
3778 vdev_state_clean(vdev_t *vd)
3779 {
3780 	spa_t *spa = vd->vdev_spa;
3781 
3782 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
3783 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3784 	    spa_config_held(spa, SCL_STATE, RW_READER)));
3785 
3786 	ASSERT(list_link_active(&vd->vdev_state_dirty_node));
3787 	list_remove(&spa->spa_state_dirty_list, vd);
3788 }
3789 
3790 /*
3791  * Propagate vdev state up from children to parent.
3792  */
3793 void
3794 vdev_propagate_state(vdev_t *vd)
3795 {
3796 	spa_t *spa = vd->vdev_spa;
3797 	vdev_t *rvd = spa->spa_root_vdev;
3798 	int degraded = 0, faulted = 0;
3799 	int corrupted = 0;
3800 	vdev_t *child;
3801 
3802 	if (vd->vdev_children > 0) {
3803 		for (int c = 0; c < vd->vdev_children; c++) {
3804 			child = vd->vdev_child[c];
3805 
3806 			/*
3807 			 * Don't factor holes or indirect vdevs into the
3808 			 * decision.
3809 			 */
3810 			if (!vdev_is_concrete(child))
3811 				continue;
3812 
3813 			if (!vdev_readable(child) ||
3814 			    (!vdev_writeable(child) && spa_writeable(spa))) {
3815 				/*
3816 				 * Root special: if there is a top-level log
3817 				 * device, treat the root vdev as if it were
3818 				 * degraded.
3819 				 */
3820 				if (child->vdev_islog && vd == rvd)
3821 					degraded++;
3822 				else
3823 					faulted++;
3824 			} else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
3825 				degraded++;
3826 			}
3827 
3828 			if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
3829 				corrupted++;
3830 		}
3831 
3832 		vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
3833 
3834 		/*
3835 		 * Root special: if there is a top-level vdev that cannot be
3836 		 * opened due to corrupted metadata, then propagate the root
3837 		 * vdev's aux state as 'corrupt' rather than 'insufficient
3838 		 * replicas'.
3839 		 */
3840 		if (corrupted && vd == rvd &&
3841 		    rvd->vdev_state == VDEV_STATE_CANT_OPEN)
3842 			vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
3843 			    VDEV_AUX_CORRUPT_DATA);
3844 	}
3845 
3846 	if (vd->vdev_parent)
3847 		vdev_propagate_state(vd->vdev_parent);
3848 }
3849 
3850 /*
3851  * Set a vdev's state.  If this is during an open, we don't update the parent
3852  * state, because we're in the process of opening children depth-first.
3853  * Otherwise, we propagate the change to the parent.
3854  *
3855  * If this routine places a device in a faulted state, an appropriate ereport is
3856  * generated.
3857  */
3858 void
3859 vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
3860 {
3861 	uint64_t save_state;
3862 	spa_t *spa = vd->vdev_spa;
3863 
3864 	if (state == vd->vdev_state) {
3865 		vd->vdev_stat.vs_aux = aux;
3866 		return;
3867 	}
3868 
3869 	save_state = vd->vdev_state;
3870 
3871 	vd->vdev_state = state;
3872 	vd->vdev_stat.vs_aux = aux;
3873 
3874 	/*
3875 	 * If we are setting the vdev state to anything but an open state, then
3876 	 * always close the underlying device unless the device has requested
3877 	 * a delayed close (i.e. we're about to remove or fault the device).
3878 	 * Otherwise, we keep accessible but invalid devices open forever.
3879 	 * We don't call vdev_close() itself, because that implies some extra
3880 	 * checks (offline, etc) that we don't want here.  This is limited to
3881 	 * leaf devices, because otherwise closing the device will affect other
3882 	 * children.
3883 	 */
3884 	if (!vd->vdev_delayed_close && vdev_is_dead(vd) &&
3885 	    vd->vdev_ops->vdev_op_leaf)
3886 		vd->vdev_ops->vdev_op_close(vd);
3887 
3888 	/*
3889 	 * If we have brought this vdev back into service, we need
3890 	 * to notify fmd so that it can gracefully repair any outstanding
3891 	 * cases due to a missing device.  We do this in all cases, even those
3892 	 * that probably don't correlate to a repaired fault.  This is sure to
3893 	 * catch all cases, and we let the zfs-retire agent sort it out.  If
3894 	 * this is a transient state it's OK, as the retire agent will
3895 	 * double-check the state of the vdev before repairing it.
3896 	 */
3897 	if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf &&
3898 	    vd->vdev_prevstate != state)
3899 		zfs_post_state_change(spa, vd);
3900 
3901 	if (vd->vdev_removed &&
3902 	    state == VDEV_STATE_CANT_OPEN &&
3903 	    (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
3904 		/*
3905 		 * If the previous state is set to VDEV_STATE_REMOVED, then this
3906 		 * device was previously marked removed and someone attempted to
3907 		 * reopen it.  If this failed due to a nonexistent device, then
3908 		 * keep the device in the REMOVED state.  We also let this be if
3909 		 * it is one of our special test online cases, which is only
3910 		 * attempting to online the device and shouldn't generate an FMA
3911 		 * fault.
3912 		 */
3913 		vd->vdev_state = VDEV_STATE_REMOVED;
3914 		vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
3915 	} else if (state == VDEV_STATE_REMOVED) {
3916 		vd->vdev_removed = B_TRUE;
3917 	} else if (state == VDEV_STATE_CANT_OPEN) {
3918 		/*
3919 		 * If we fail to open a vdev during an import or recovery, we
3920 		 * mark it as "not available", which signifies that it was
3921 		 * never there to begin with.  Failure to open such a device
3922 		 * is not considered an error.
3923 		 */
3924 		if ((spa_load_state(spa) == SPA_LOAD_IMPORT ||
3925 		    spa_load_state(spa) == SPA_LOAD_RECOVER) &&
3926 		    vd->vdev_ops->vdev_op_leaf)
3927 			vd->vdev_not_present = 1;
3928 
3929 		/*
3930 		 * Post the appropriate ereport.  If the 'prevstate' field is
3931 		 * set to something other than VDEV_STATE_UNKNOWN, it indicates
3932 		 * that this is part of a vdev_reopen().  In this case, we don't
3933 		 * want to post the ereport if the device was already in the
3934 		 * CANT_OPEN state beforehand.
3935 		 *
3936 		 * If the 'checkremove' flag is set, then this is an attempt to
3937 		 * online the device in response to an insertion event.  If we
3938 		 * hit this case, then we have detected an insertion event for a
3939 		 * faulted or offline device that wasn't in the removed state.
3940 		 * In this scenario, we don't post an ereport because we are
3941 		 * about to replace the device, or attempt an online with
3942 		 * vdev_forcefault, which will generate the fault for us.
3943 		 */
3944 		if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
3945 		    !vd->vdev_not_present && !vd->vdev_checkremove &&
3946 		    vd != spa->spa_root_vdev) {
3947 			const char *class;
3948 
3949 			switch (aux) {
3950 			case VDEV_AUX_OPEN_FAILED:
3951 				class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
3952 				break;
3953 			case VDEV_AUX_CORRUPT_DATA:
3954 				class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
3955 				break;
3956 			case VDEV_AUX_NO_REPLICAS:
3957 				class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
3958 				break;
3959 			case VDEV_AUX_BAD_GUID_SUM:
3960 				class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
3961 				break;
3962 			case VDEV_AUX_TOO_SMALL:
3963 				class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
3964 				break;
3965 			case VDEV_AUX_BAD_LABEL:
3966 				class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
3967 				break;
3968 			default:
3969 				class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
3970 			}
3971 
3972 			zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
3973 		}
3974 
3975 		/* Erase any notion of persistent removed state */
3976 		vd->vdev_removed = B_FALSE;
3977 	} else {
3978 		vd->vdev_removed = B_FALSE;
3979 	}
3980 
3981 	if (!isopen && vd->vdev_parent)
3982 		vdev_propagate_state(vd->vdev_parent);
3983 }
3984 
3985 boolean_t
3986 vdev_children_are_offline(vdev_t *vd)
3987 {
3988 	ASSERT(!vd->vdev_ops->vdev_op_leaf);
3989 
3990 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
3991 		if (vd->vdev_child[i]->vdev_state != VDEV_STATE_OFFLINE)
3992 			return (B_FALSE);
3993 	}
3994 
3995 	return (B_TRUE);
3996 }
3997 
3998 /*
3999  * Check the vdev configuration to ensure that it's capable of supporting
4000  * a root pool. We do not support partial configuration.
4001  * In addition, only a single top-level vdev is allowed.
4002  */
4003 boolean_t
4004 vdev_is_bootable(vdev_t *vd)
4005 {
4006 	if (!vd->vdev_ops->vdev_op_leaf) {
4007 		char *vdev_type = vd->vdev_ops->vdev_op_type;
4008 
4009 		if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
4010 		    vd->vdev_children > 1) {
4011 			return (B_FALSE);
4012 		} else if (strcmp(vdev_type, VDEV_TYPE_MISSING) == 0 ||
4013 		    strcmp(vdev_type, VDEV_TYPE_INDIRECT) == 0) {
4014 			return (B_FALSE);
4015 		}
4016 	}
4017 
4018 	for (int c = 0; c < vd->vdev_children; c++) {
4019 		if (!vdev_is_bootable(vd->vdev_child[c]))
4020 			return (B_FALSE);
4021 	}
4022 	return (B_TRUE);
4023 }
4024 
4025 boolean_t
4026 vdev_is_concrete(vdev_t *vd)
4027 {
4028 	vdev_ops_t *ops = vd->vdev_ops;
4029 	if (ops == &vdev_indirect_ops || ops == &vdev_hole_ops ||
4030 	    ops == &vdev_missing_ops || ops == &vdev_root_ops) {
4031 		return (B_FALSE);
4032 	} else {
4033 		return (B_TRUE);
4034 	}
4035 }
4036 
4037 /*
4038  * Determine if a log device has valid content.  If the vdev was
4039  * removed or faulted in the MOS config then we know that
4040  * the content on the log device has already been written to the pool.
4041  */
4042 boolean_t
4043 vdev_log_state_valid(vdev_t *vd)
4044 {
4045 	if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted &&
4046 	    !vd->vdev_removed)
4047 		return (B_TRUE);
4048 
4049 	for (int c = 0; c < vd->vdev_children; c++)
4050 		if (vdev_log_state_valid(vd->vdev_child[c]))
4051 			return (B_TRUE);
4052 
4053 	return (B_FALSE);
4054 }
4055 
4056 /*
4057  * Expand a vdev if possible.
4058  */
4059 void
4060 vdev_expand(vdev_t *vd, uint64_t txg)
4061 {
4062 	ASSERT(vd->vdev_top == vd);
4063 	ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
4064 
4065 	vdev_set_deflate_ratio(vd);
4066 
4067 	if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count &&
4068 	    vdev_is_concrete(vd)) {
4069 		VERIFY(vdev_metaslab_init(vd, txg) == 0);
4070 		vdev_config_dirty(vd);
4071 	}
4072 }
4073 
4074 /*
4075  * Split a vdev.
4076  */
4077 void
4078 vdev_split(vdev_t *vd)
4079 {
4080 	vdev_t *cvd, *pvd = vd->vdev_parent;
4081 
4082 	vdev_remove_child(pvd, vd);
4083 	vdev_compact_children(pvd);
4084 
4085 	cvd = pvd->vdev_child[0];
4086 	if (pvd->vdev_children == 1) {
4087 		vdev_remove_parent(cvd);
4088 		cvd->vdev_splitting = B_TRUE;
4089 	}
4090 	vdev_propagate_state(cvd);
4091 }
4092 
4093 void
4094 vdev_deadman(vdev_t *vd)
4095 {
4096 	for (int c = 0; c < vd->vdev_children; c++) {
4097 		vdev_t *cvd = vd->vdev_child[c];
4098 
4099 		vdev_deadman(cvd);
4100 	}
4101 
4102 	if (vd->vdev_ops->vdev_op_leaf) {
4103 		vdev_queue_t *vq = &vd->vdev_queue;
4104 
4105 		mutex_enter(&vq->vq_lock);
4106 		if (avl_numnodes(&vq->vq_active_tree) > 0) {
4107 			spa_t *spa = vd->vdev_spa;
4108 			zio_t *fio;
4109 			uint64_t delta;
4110 
4111 			/*
4112 			 * Look at the head of all the pending queues,
4113 			 * if any I/O has been outstanding for longer than
4114 			 * the spa_deadman_synctime we panic the system.
4115 			 */
4116 			fio = avl_first(&vq->vq_active_tree);
4117 			delta = gethrtime() - fio->io_timestamp;
4118 			if (delta > spa_deadman_synctime(spa)) {
4119 				vdev_dbgmsg(vd, "SLOW IO: zio timestamp "
4120 				    "%lluns, delta %lluns, last io %lluns",
4121 				    fio->io_timestamp, (u_longlong_t)delta,
4122 				    vq->vq_io_complete_ts);
4123 				fm_panic("I/O to pool '%s' appears to be "
4124 				    "hung.", spa_name(spa));
4125 			}
4126 		}
4127 		mutex_exit(&vq->vq_lock);
4128 	}
4129 }
4130