xref: /titanic_51/usr/src/uts/common/fs/zfs/vdev.c (revision 87d5c1f9b787004b82b01c61d7159df8a96de8e0)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #include <sys/zfs_context.h>
28 #include <sys/fm/fs/zfs.h>
29 #include <sys/spa.h>
30 #include <sys/spa_impl.h>
31 #include <sys/dmu.h>
32 #include <sys/dmu_tx.h>
33 #include <sys/vdev_impl.h>
34 #include <sys/uberblock_impl.h>
35 #include <sys/metaslab.h>
36 #include <sys/metaslab_impl.h>
37 #include <sys/space_map.h>
38 #include <sys/zio.h>
39 #include <sys/zap.h>
40 #include <sys/fs/zfs.h>
41 #include <sys/arc.h>
42 #include <sys/zil.h>
43 
44 /*
45  * Virtual device management.
46  */
47 
48 static vdev_ops_t *vdev_ops_table[] = {
49 	&vdev_root_ops,
50 	&vdev_raidz_ops,
51 	&vdev_mirror_ops,
52 	&vdev_replacing_ops,
53 	&vdev_spare_ops,
54 	&vdev_disk_ops,
55 	&vdev_file_ops,
56 	&vdev_missing_ops,
57 	NULL
58 };
59 
60 /* maximum scrub/resilver I/O queue per leaf vdev */
61 int zfs_scrub_limit = 10;
62 
63 /*
64  * Given a vdev type, return the appropriate ops vector.
65  */
66 static vdev_ops_t *
67 vdev_getops(const char *type)
68 {
69 	vdev_ops_t *ops, **opspp;
70 
71 	for (opspp = vdev_ops_table; (ops = *opspp) != NULL; opspp++)
72 		if (strcmp(ops->vdev_op_type, type) == 0)
73 			break;
74 
75 	return (ops);
76 }
77 
78 /*
79  * Default asize function: return the MAX of psize with the asize of
80  * all children.  This is what's used by anything other than RAID-Z.
81  */
82 uint64_t
83 vdev_default_asize(vdev_t *vd, uint64_t psize)
84 {
85 	uint64_t asize = P2ROUNDUP(psize, 1ULL << vd->vdev_top->vdev_ashift);
86 	uint64_t csize;
87 
88 	for (int c = 0; c < vd->vdev_children; c++) {
89 		csize = vdev_psize_to_asize(vd->vdev_child[c], psize);
90 		asize = MAX(asize, csize);
91 	}
92 
93 	return (asize);
94 }
95 
96 /*
97  * Get the minimum allocatable size. We define the allocatable size as
98  * the vdev's asize rounded to the nearest metaslab. This allows us to
99  * replace or attach devices which don't have the same physical size but
100  * can still satisfy the same number of allocations.
101  */
102 uint64_t
103 vdev_get_min_asize(vdev_t *vd)
104 {
105 	vdev_t *pvd = vd->vdev_parent;
106 
107 	/*
108 	 * The our parent is NULL (inactive spare or cache) or is the root,
109 	 * just return our own asize.
110 	 */
111 	if (pvd == NULL)
112 		return (vd->vdev_asize);
113 
114 	/*
115 	 * The top-level vdev just returns the allocatable size rounded
116 	 * to the nearest metaslab.
117 	 */
118 	if (vd == vd->vdev_top)
119 		return (P2ALIGN(vd->vdev_asize, 1ULL << vd->vdev_ms_shift));
120 
121 	/*
122 	 * The allocatable space for a raidz vdev is N * sizeof(smallest child),
123 	 * so each child must provide at least 1/Nth of its asize.
124 	 */
125 	if (pvd->vdev_ops == &vdev_raidz_ops)
126 		return (pvd->vdev_min_asize / pvd->vdev_children);
127 
128 	return (pvd->vdev_min_asize);
129 }
130 
131 void
132 vdev_set_min_asize(vdev_t *vd)
133 {
134 	vd->vdev_min_asize = vdev_get_min_asize(vd);
135 
136 	for (int c = 0; c < vd->vdev_children; c++)
137 		vdev_set_min_asize(vd->vdev_child[c]);
138 }
139 
140 vdev_t *
141 vdev_lookup_top(spa_t *spa, uint64_t vdev)
142 {
143 	vdev_t *rvd = spa->spa_root_vdev;
144 
145 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
146 
147 	if (vdev < rvd->vdev_children) {
148 		ASSERT(rvd->vdev_child[vdev] != NULL);
149 		return (rvd->vdev_child[vdev]);
150 	}
151 
152 	return (NULL);
153 }
154 
155 vdev_t *
156 vdev_lookup_by_guid(vdev_t *vd, uint64_t guid)
157 {
158 	vdev_t *mvd;
159 
160 	if (vd->vdev_guid == guid)
161 		return (vd);
162 
163 	for (int c = 0; c < vd->vdev_children; c++)
164 		if ((mvd = vdev_lookup_by_guid(vd->vdev_child[c], guid)) !=
165 		    NULL)
166 			return (mvd);
167 
168 	return (NULL);
169 }
170 
171 void
172 vdev_add_child(vdev_t *pvd, vdev_t *cvd)
173 {
174 	size_t oldsize, newsize;
175 	uint64_t id = cvd->vdev_id;
176 	vdev_t **newchild;
177 
178 	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
179 	ASSERT(cvd->vdev_parent == NULL);
180 
181 	cvd->vdev_parent = pvd;
182 
183 	if (pvd == NULL)
184 		return;
185 
186 	ASSERT(id >= pvd->vdev_children || pvd->vdev_child[id] == NULL);
187 
188 	oldsize = pvd->vdev_children * sizeof (vdev_t *);
189 	pvd->vdev_children = MAX(pvd->vdev_children, id + 1);
190 	newsize = pvd->vdev_children * sizeof (vdev_t *);
191 
192 	newchild = kmem_zalloc(newsize, KM_SLEEP);
193 	if (pvd->vdev_child != NULL) {
194 		bcopy(pvd->vdev_child, newchild, oldsize);
195 		kmem_free(pvd->vdev_child, oldsize);
196 	}
197 
198 	pvd->vdev_child = newchild;
199 	pvd->vdev_child[id] = cvd;
200 
201 	cvd->vdev_top = (pvd->vdev_top ? pvd->vdev_top: cvd);
202 	ASSERT(cvd->vdev_top->vdev_parent->vdev_parent == NULL);
203 
204 	/*
205 	 * Walk up all ancestors to update guid sum.
206 	 */
207 	for (; pvd != NULL; pvd = pvd->vdev_parent)
208 		pvd->vdev_guid_sum += cvd->vdev_guid_sum;
209 
210 	if (cvd->vdev_ops->vdev_op_leaf)
211 		cvd->vdev_spa->spa_scrub_maxinflight += zfs_scrub_limit;
212 }
213 
214 void
215 vdev_remove_child(vdev_t *pvd, vdev_t *cvd)
216 {
217 	int c;
218 	uint_t id = cvd->vdev_id;
219 
220 	ASSERT(cvd->vdev_parent == pvd);
221 
222 	if (pvd == NULL)
223 		return;
224 
225 	ASSERT(id < pvd->vdev_children);
226 	ASSERT(pvd->vdev_child[id] == cvd);
227 
228 	pvd->vdev_child[id] = NULL;
229 	cvd->vdev_parent = NULL;
230 
231 	for (c = 0; c < pvd->vdev_children; c++)
232 		if (pvd->vdev_child[c])
233 			break;
234 
235 	if (c == pvd->vdev_children) {
236 		kmem_free(pvd->vdev_child, c * sizeof (vdev_t *));
237 		pvd->vdev_child = NULL;
238 		pvd->vdev_children = 0;
239 	}
240 
241 	/*
242 	 * Walk up all ancestors to update guid sum.
243 	 */
244 	for (; pvd != NULL; pvd = pvd->vdev_parent)
245 		pvd->vdev_guid_sum -= cvd->vdev_guid_sum;
246 
247 	if (cvd->vdev_ops->vdev_op_leaf)
248 		cvd->vdev_spa->spa_scrub_maxinflight -= zfs_scrub_limit;
249 }
250 
251 /*
252  * Remove any holes in the child array.
253  */
254 void
255 vdev_compact_children(vdev_t *pvd)
256 {
257 	vdev_t **newchild, *cvd;
258 	int oldc = pvd->vdev_children;
259 	int newc;
260 
261 	ASSERT(spa_config_held(pvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
262 
263 	for (int c = newc = 0; c < oldc; c++)
264 		if (pvd->vdev_child[c])
265 			newc++;
266 
267 	newchild = kmem_alloc(newc * sizeof (vdev_t *), KM_SLEEP);
268 
269 	for (int c = newc = 0; c < oldc; c++) {
270 		if ((cvd = pvd->vdev_child[c]) != NULL) {
271 			newchild[newc] = cvd;
272 			cvd->vdev_id = newc++;
273 		}
274 	}
275 
276 	kmem_free(pvd->vdev_child, oldc * sizeof (vdev_t *));
277 	pvd->vdev_child = newchild;
278 	pvd->vdev_children = newc;
279 }
280 
281 /*
282  * Allocate and minimally initialize a vdev_t.
283  */
284 static vdev_t *
285 vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
286 {
287 	vdev_t *vd;
288 
289 	vd = kmem_zalloc(sizeof (vdev_t), KM_SLEEP);
290 
291 	if (spa->spa_root_vdev == NULL) {
292 		ASSERT(ops == &vdev_root_ops);
293 		spa->spa_root_vdev = vd;
294 	}
295 
296 	if (guid == 0) {
297 		if (spa->spa_root_vdev == vd) {
298 			/*
299 			 * The root vdev's guid will also be the pool guid,
300 			 * which must be unique among all pools.
301 			 */
302 			while (guid == 0 || spa_guid_exists(guid, 0))
303 				guid = spa_get_random(-1ULL);
304 		} else {
305 			/*
306 			 * Any other vdev's guid must be unique within the pool.
307 			 */
308 			while (guid == 0 ||
309 			    spa_guid_exists(spa_guid(spa), guid))
310 				guid = spa_get_random(-1ULL);
311 		}
312 		ASSERT(!spa_guid_exists(spa_guid(spa), guid));
313 	}
314 
315 	vd->vdev_spa = spa;
316 	vd->vdev_id = id;
317 	vd->vdev_guid = guid;
318 	vd->vdev_guid_sum = guid;
319 	vd->vdev_ops = ops;
320 	vd->vdev_state = VDEV_STATE_CLOSED;
321 
322 	mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_DEFAULT, NULL);
323 	mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL);
324 	mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL);
325 	for (int t = 0; t < DTL_TYPES; t++) {
326 		space_map_create(&vd->vdev_dtl[t], 0, -1ULL, 0,
327 		    &vd->vdev_dtl_lock);
328 	}
329 	txg_list_create(&vd->vdev_ms_list,
330 	    offsetof(struct metaslab, ms_txg_node));
331 	txg_list_create(&vd->vdev_dtl_list,
332 	    offsetof(struct vdev, vdev_dtl_node));
333 	vd->vdev_stat.vs_timestamp = gethrtime();
334 	vdev_queue_init(vd);
335 	vdev_cache_init(vd);
336 
337 	return (vd);
338 }
339 
340 /*
341  * Allocate a new vdev.  The 'alloctype' is used to control whether we are
342  * creating a new vdev or loading an existing one - the behavior is slightly
343  * different for each case.
344  */
345 int
346 vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id,
347     int alloctype)
348 {
349 	vdev_ops_t *ops;
350 	char *type;
351 	uint64_t guid = 0, islog, nparity;
352 	vdev_t *vd;
353 
354 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
355 
356 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) != 0)
357 		return (EINVAL);
358 
359 	if ((ops = vdev_getops(type)) == NULL)
360 		return (EINVAL);
361 
362 	/*
363 	 * If this is a load, get the vdev guid from the nvlist.
364 	 * Otherwise, vdev_alloc_common() will generate one for us.
365 	 */
366 	if (alloctype == VDEV_ALLOC_LOAD) {
367 		uint64_t label_id;
368 
369 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ID, &label_id) ||
370 		    label_id != id)
371 			return (EINVAL);
372 
373 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
374 			return (EINVAL);
375 	} else if (alloctype == VDEV_ALLOC_SPARE) {
376 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
377 			return (EINVAL);
378 	} else if (alloctype == VDEV_ALLOC_L2CACHE) {
379 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
380 			return (EINVAL);
381 	} else if (alloctype == VDEV_ALLOC_ROOTPOOL) {
382 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
383 			return (EINVAL);
384 	}
385 
386 	/*
387 	 * The first allocated vdev must be of type 'root'.
388 	 */
389 	if (ops != &vdev_root_ops && spa->spa_root_vdev == NULL)
390 		return (EINVAL);
391 
392 	/*
393 	 * Determine whether we're a log vdev.
394 	 */
395 	islog = 0;
396 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &islog);
397 	if (islog && spa_version(spa) < SPA_VERSION_SLOGS)
398 		return (ENOTSUP);
399 
400 	/*
401 	 * Set the nparity property for RAID-Z vdevs.
402 	 */
403 	nparity = -1ULL;
404 	if (ops == &vdev_raidz_ops) {
405 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY,
406 		    &nparity) == 0) {
407 			/*
408 			 * Currently, we can only support 2 parity devices.
409 			 */
410 			if (nparity == 0 || nparity > 2)
411 				return (EINVAL);
412 			/*
413 			 * Older versions can only support 1 parity device.
414 			 */
415 			if (nparity == 2 &&
416 			    spa_version(spa) < SPA_VERSION_RAID6)
417 				return (ENOTSUP);
418 		} else {
419 			/*
420 			 * We require the parity to be specified for SPAs that
421 			 * support multiple parity levels.
422 			 */
423 			if (spa_version(spa) >= SPA_VERSION_RAID6)
424 				return (EINVAL);
425 			/*
426 			 * Otherwise, we default to 1 parity device for RAID-Z.
427 			 */
428 			nparity = 1;
429 		}
430 	} else {
431 		nparity = 0;
432 	}
433 	ASSERT(nparity != -1ULL);
434 
435 	vd = vdev_alloc_common(spa, id, guid, ops);
436 
437 	vd->vdev_islog = islog;
438 	vd->vdev_nparity = nparity;
439 
440 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &vd->vdev_path) == 0)
441 		vd->vdev_path = spa_strdup(vd->vdev_path);
442 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &vd->vdev_devid) == 0)
443 		vd->vdev_devid = spa_strdup(vd->vdev_devid);
444 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH,
445 	    &vd->vdev_physpath) == 0)
446 		vd->vdev_physpath = spa_strdup(vd->vdev_physpath);
447 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &vd->vdev_fru) == 0)
448 		vd->vdev_fru = spa_strdup(vd->vdev_fru);
449 
450 	/*
451 	 * Set the whole_disk property.  If it's not specified, leave the value
452 	 * as -1.
453 	 */
454 	if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
455 	    &vd->vdev_wholedisk) != 0)
456 		vd->vdev_wholedisk = -1ULL;
457 
458 	/*
459 	 * Look for the 'not present' flag.  This will only be set if the device
460 	 * was not present at the time of import.
461 	 */
462 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT,
463 	    &vd->vdev_not_present);
464 
465 	/*
466 	 * Get the alignment requirement.
467 	 */
468 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT, &vd->vdev_ashift);
469 
470 	/*
471 	 * If we're a top-level vdev, try to load the allocation parameters.
472 	 */
473 	if (parent && !parent->vdev_parent && alloctype == VDEV_ALLOC_LOAD) {
474 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
475 		    &vd->vdev_ms_array);
476 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
477 		    &vd->vdev_ms_shift);
478 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE,
479 		    &vd->vdev_asize);
480 	}
481 
482 	/*
483 	 * If we're a leaf vdev, try to load the DTL object and other state.
484 	 */
485 	if (vd->vdev_ops->vdev_op_leaf &&
486 	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE ||
487 	    alloctype == VDEV_ALLOC_ROOTPOOL)) {
488 		if (alloctype == VDEV_ALLOC_LOAD) {
489 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL,
490 			    &vd->vdev_dtl_smo.smo_object);
491 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE,
492 			    &vd->vdev_unspare);
493 		}
494 
495 		if (alloctype == VDEV_ALLOC_ROOTPOOL) {
496 			uint64_t spare = 0;
497 
498 			if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE,
499 			    &spare) == 0 && spare)
500 				spa_spare_add(vd);
501 		}
502 
503 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE,
504 		    &vd->vdev_offline);
505 
506 		/*
507 		 * When importing a pool, we want to ignore the persistent fault
508 		 * state, as the diagnosis made on another system may not be
509 		 * valid in the current context.
510 		 */
511 		if (spa->spa_load_state == SPA_LOAD_OPEN) {
512 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED,
513 			    &vd->vdev_faulted);
514 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED,
515 			    &vd->vdev_degraded);
516 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED,
517 			    &vd->vdev_removed);
518 		}
519 	}
520 
521 	/*
522 	 * Add ourselves to the parent's list of children.
523 	 */
524 	vdev_add_child(parent, vd);
525 
526 	*vdp = vd;
527 
528 	return (0);
529 }
530 
531 void
532 vdev_free(vdev_t *vd)
533 {
534 	spa_t *spa = vd->vdev_spa;
535 
536 	/*
537 	 * vdev_free() implies closing the vdev first.  This is simpler than
538 	 * trying to ensure complicated semantics for all callers.
539 	 */
540 	vdev_close(vd);
541 
542 	ASSERT(!list_link_active(&vd->vdev_config_dirty_node));
543 
544 	/*
545 	 * Free all children.
546 	 */
547 	for (int c = 0; c < vd->vdev_children; c++)
548 		vdev_free(vd->vdev_child[c]);
549 
550 	ASSERT(vd->vdev_child == NULL);
551 	ASSERT(vd->vdev_guid_sum == vd->vdev_guid);
552 
553 	/*
554 	 * Discard allocation state.
555 	 */
556 	if (vd == vd->vdev_top)
557 		vdev_metaslab_fini(vd);
558 
559 	ASSERT3U(vd->vdev_stat.vs_space, ==, 0);
560 	ASSERT3U(vd->vdev_stat.vs_dspace, ==, 0);
561 	ASSERT3U(vd->vdev_stat.vs_alloc, ==, 0);
562 
563 	/*
564 	 * Remove this vdev from its parent's child list.
565 	 */
566 	vdev_remove_child(vd->vdev_parent, vd);
567 
568 	ASSERT(vd->vdev_parent == NULL);
569 
570 	/*
571 	 * Clean up vdev structure.
572 	 */
573 	vdev_queue_fini(vd);
574 	vdev_cache_fini(vd);
575 
576 	if (vd->vdev_path)
577 		spa_strfree(vd->vdev_path);
578 	if (vd->vdev_devid)
579 		spa_strfree(vd->vdev_devid);
580 	if (vd->vdev_physpath)
581 		spa_strfree(vd->vdev_physpath);
582 	if (vd->vdev_fru)
583 		spa_strfree(vd->vdev_fru);
584 
585 	if (vd->vdev_isspare)
586 		spa_spare_remove(vd);
587 	if (vd->vdev_isl2cache)
588 		spa_l2cache_remove(vd);
589 
590 	txg_list_destroy(&vd->vdev_ms_list);
591 	txg_list_destroy(&vd->vdev_dtl_list);
592 
593 	mutex_enter(&vd->vdev_dtl_lock);
594 	for (int t = 0; t < DTL_TYPES; t++) {
595 		space_map_unload(&vd->vdev_dtl[t]);
596 		space_map_destroy(&vd->vdev_dtl[t]);
597 	}
598 	mutex_exit(&vd->vdev_dtl_lock);
599 
600 	mutex_destroy(&vd->vdev_dtl_lock);
601 	mutex_destroy(&vd->vdev_stat_lock);
602 	mutex_destroy(&vd->vdev_probe_lock);
603 
604 	if (vd == spa->spa_root_vdev)
605 		spa->spa_root_vdev = NULL;
606 
607 	kmem_free(vd, sizeof (vdev_t));
608 }
609 
610 /*
611  * Transfer top-level vdev state from svd to tvd.
612  */
613 static void
614 vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
615 {
616 	spa_t *spa = svd->vdev_spa;
617 	metaslab_t *msp;
618 	vdev_t *vd;
619 	int t;
620 
621 	ASSERT(tvd == tvd->vdev_top);
622 
623 	tvd->vdev_ms_array = svd->vdev_ms_array;
624 	tvd->vdev_ms_shift = svd->vdev_ms_shift;
625 	tvd->vdev_ms_count = svd->vdev_ms_count;
626 
627 	svd->vdev_ms_array = 0;
628 	svd->vdev_ms_shift = 0;
629 	svd->vdev_ms_count = 0;
630 
631 	tvd->vdev_mg = svd->vdev_mg;
632 	tvd->vdev_ms = svd->vdev_ms;
633 
634 	svd->vdev_mg = NULL;
635 	svd->vdev_ms = NULL;
636 
637 	if (tvd->vdev_mg != NULL)
638 		tvd->vdev_mg->mg_vd = tvd;
639 
640 	tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
641 	tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
642 	tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
643 
644 	svd->vdev_stat.vs_alloc = 0;
645 	svd->vdev_stat.vs_space = 0;
646 	svd->vdev_stat.vs_dspace = 0;
647 
648 	for (t = 0; t < TXG_SIZE; t++) {
649 		while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
650 			(void) txg_list_add(&tvd->vdev_ms_list, msp, t);
651 		while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
652 			(void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
653 		if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
654 			(void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
655 	}
656 
657 	if (list_link_active(&svd->vdev_config_dirty_node)) {
658 		vdev_config_clean(svd);
659 		vdev_config_dirty(tvd);
660 	}
661 
662 	if (list_link_active(&svd->vdev_state_dirty_node)) {
663 		vdev_state_clean(svd);
664 		vdev_state_dirty(tvd);
665 	}
666 
667 	tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
668 	svd->vdev_deflate_ratio = 0;
669 
670 	tvd->vdev_islog = svd->vdev_islog;
671 	svd->vdev_islog = 0;
672 }
673 
674 static void
675 vdev_top_update(vdev_t *tvd, vdev_t *vd)
676 {
677 	if (vd == NULL)
678 		return;
679 
680 	vd->vdev_top = tvd;
681 
682 	for (int c = 0; c < vd->vdev_children; c++)
683 		vdev_top_update(tvd, vd->vdev_child[c]);
684 }
685 
686 /*
687  * Add a mirror/replacing vdev above an existing vdev.
688  */
689 vdev_t *
690 vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
691 {
692 	spa_t *spa = cvd->vdev_spa;
693 	vdev_t *pvd = cvd->vdev_parent;
694 	vdev_t *mvd;
695 
696 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
697 
698 	mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
699 
700 	mvd->vdev_asize = cvd->vdev_asize;
701 	mvd->vdev_min_asize = cvd->vdev_min_asize;
702 	mvd->vdev_ashift = cvd->vdev_ashift;
703 	mvd->vdev_state = cvd->vdev_state;
704 
705 	vdev_remove_child(pvd, cvd);
706 	vdev_add_child(pvd, mvd);
707 	cvd->vdev_id = mvd->vdev_children;
708 	vdev_add_child(mvd, cvd);
709 	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
710 
711 	if (mvd == mvd->vdev_top)
712 		vdev_top_transfer(cvd, mvd);
713 
714 	return (mvd);
715 }
716 
717 /*
718  * Remove a 1-way mirror/replacing vdev from the tree.
719  */
720 void
721 vdev_remove_parent(vdev_t *cvd)
722 {
723 	vdev_t *mvd = cvd->vdev_parent;
724 	vdev_t *pvd = mvd->vdev_parent;
725 
726 	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
727 
728 	ASSERT(mvd->vdev_children == 1);
729 	ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
730 	    mvd->vdev_ops == &vdev_replacing_ops ||
731 	    mvd->vdev_ops == &vdev_spare_ops);
732 	cvd->vdev_ashift = mvd->vdev_ashift;
733 
734 	vdev_remove_child(mvd, cvd);
735 	vdev_remove_child(pvd, mvd);
736 
737 	/*
738 	 * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
739 	 * Otherwise, we could have detached an offline device, and when we
740 	 * go to import the pool we'll think we have two top-level vdevs,
741 	 * instead of a different version of the same top-level vdev.
742 	 */
743 	if (mvd->vdev_top == mvd) {
744 		uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
745 		cvd->vdev_guid += guid_delta;
746 		cvd->vdev_guid_sum += guid_delta;
747 	}
748 	cvd->vdev_id = mvd->vdev_id;
749 	vdev_add_child(pvd, cvd);
750 	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
751 
752 	if (cvd == cvd->vdev_top)
753 		vdev_top_transfer(mvd, cvd);
754 
755 	ASSERT(mvd->vdev_children == 0);
756 	vdev_free(mvd);
757 }
758 
759 int
760 vdev_metaslab_init(vdev_t *vd, uint64_t txg)
761 {
762 	spa_t *spa = vd->vdev_spa;
763 	objset_t *mos = spa->spa_meta_objset;
764 	metaslab_class_t *mc;
765 	uint64_t m;
766 	uint64_t oldc = vd->vdev_ms_count;
767 	uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
768 	metaslab_t **mspp;
769 	int error;
770 
771 	if (vd->vdev_ms_shift == 0)	/* not being allocated from yet */
772 		return (0);
773 
774 	/*
775 	 * Compute the raidz-deflation ratio.  Note, we hard-code
776 	 * in 128k (1 << 17) because it is the current "typical" blocksize.
777 	 * Even if SPA_MAXBLOCKSIZE changes, this algorithm must never change,
778 	 * or we will inconsistently account for existing bp's.
779 	 */
780 	vd->vdev_deflate_ratio = (1 << 17) /
781 	    (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
782 
783 	ASSERT(oldc <= newc);
784 
785 	if (vd->vdev_islog)
786 		mc = spa->spa_log_class;
787 	else
788 		mc = spa->spa_normal_class;
789 
790 	if (vd->vdev_mg == NULL)
791 		vd->vdev_mg = metaslab_group_create(mc, vd);
792 
793 	mspp = kmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
794 
795 	if (oldc != 0) {
796 		bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp));
797 		kmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
798 	}
799 
800 	vd->vdev_ms = mspp;
801 	vd->vdev_ms_count = newc;
802 
803 	for (m = oldc; m < newc; m++) {
804 		space_map_obj_t smo = { 0, 0, 0 };
805 		if (txg == 0) {
806 			uint64_t object = 0;
807 			error = dmu_read(mos, vd->vdev_ms_array,
808 			    m * sizeof (uint64_t), sizeof (uint64_t), &object,
809 			    DMU_READ_PREFETCH);
810 			if (error)
811 				return (error);
812 			if (object != 0) {
813 				dmu_buf_t *db;
814 				error = dmu_bonus_hold(mos, object, FTAG, &db);
815 				if (error)
816 					return (error);
817 				ASSERT3U(db->db_size, >=, sizeof (smo));
818 				bcopy(db->db_data, &smo, sizeof (smo));
819 				ASSERT3U(smo.smo_object, ==, object);
820 				dmu_buf_rele(db, FTAG);
821 			}
822 		}
823 		vd->vdev_ms[m] = metaslab_init(vd->vdev_mg, &smo,
824 		    m << vd->vdev_ms_shift, 1ULL << vd->vdev_ms_shift, txg);
825 	}
826 
827 	return (0);
828 }
829 
830 void
831 vdev_metaslab_fini(vdev_t *vd)
832 {
833 	uint64_t m;
834 	uint64_t count = vd->vdev_ms_count;
835 
836 	if (vd->vdev_ms != NULL) {
837 		for (m = 0; m < count; m++)
838 			if (vd->vdev_ms[m] != NULL)
839 				metaslab_fini(vd->vdev_ms[m]);
840 		kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
841 		vd->vdev_ms = NULL;
842 	}
843 }
844 
845 typedef struct vdev_probe_stats {
846 	boolean_t	vps_readable;
847 	boolean_t	vps_writeable;
848 	int		vps_flags;
849 } vdev_probe_stats_t;
850 
851 static void
852 vdev_probe_done(zio_t *zio)
853 {
854 	spa_t *spa = zio->io_spa;
855 	vdev_t *vd = zio->io_vd;
856 	vdev_probe_stats_t *vps = zio->io_private;
857 
858 	ASSERT(vd->vdev_probe_zio != NULL);
859 
860 	if (zio->io_type == ZIO_TYPE_READ) {
861 		if (zio->io_error == 0)
862 			vps->vps_readable = 1;
863 		if (zio->io_error == 0 && spa_writeable(spa)) {
864 			zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
865 			    zio->io_offset, zio->io_size, zio->io_data,
866 			    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
867 			    ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
868 		} else {
869 			zio_buf_free(zio->io_data, zio->io_size);
870 		}
871 	} else if (zio->io_type == ZIO_TYPE_WRITE) {
872 		if (zio->io_error == 0)
873 			vps->vps_writeable = 1;
874 		zio_buf_free(zio->io_data, zio->io_size);
875 	} else if (zio->io_type == ZIO_TYPE_NULL) {
876 		zio_t *pio;
877 
878 		vd->vdev_cant_read |= !vps->vps_readable;
879 		vd->vdev_cant_write |= !vps->vps_writeable;
880 
881 		if (vdev_readable(vd) &&
882 		    (vdev_writeable(vd) || !spa_writeable(spa))) {
883 			zio->io_error = 0;
884 		} else {
885 			ASSERT(zio->io_error != 0);
886 			zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
887 			    spa, vd, NULL, 0, 0);
888 			zio->io_error = ENXIO;
889 		}
890 
891 		mutex_enter(&vd->vdev_probe_lock);
892 		ASSERT(vd->vdev_probe_zio == zio);
893 		vd->vdev_probe_zio = NULL;
894 		mutex_exit(&vd->vdev_probe_lock);
895 
896 		while ((pio = zio_walk_parents(zio)) != NULL)
897 			if (!vdev_accessible(vd, pio))
898 				pio->io_error = ENXIO;
899 
900 		kmem_free(vps, sizeof (*vps));
901 	}
902 }
903 
904 /*
905  * Determine whether this device is accessible by reading and writing
906  * to several known locations: the pad regions of each vdev label
907  * but the first (which we leave alone in case it contains a VTOC).
908  */
909 zio_t *
910 vdev_probe(vdev_t *vd, zio_t *zio)
911 {
912 	spa_t *spa = vd->vdev_spa;
913 	vdev_probe_stats_t *vps = NULL;
914 	zio_t *pio;
915 
916 	ASSERT(vd->vdev_ops->vdev_op_leaf);
917 
918 	/*
919 	 * Don't probe the probe.
920 	 */
921 	if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
922 		return (NULL);
923 
924 	/*
925 	 * To prevent 'probe storms' when a device fails, we create
926 	 * just one probe i/o at a time.  All zios that want to probe
927 	 * this vdev will become parents of the probe io.
928 	 */
929 	mutex_enter(&vd->vdev_probe_lock);
930 
931 	if ((pio = vd->vdev_probe_zio) == NULL) {
932 		vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
933 
934 		vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
935 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
936 		    ZIO_FLAG_TRYHARD;
937 
938 		if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
939 			/*
940 			 * vdev_cant_read and vdev_cant_write can only
941 			 * transition from TRUE to FALSE when we have the
942 			 * SCL_ZIO lock as writer; otherwise they can only
943 			 * transition from FALSE to TRUE.  This ensures that
944 			 * any zio looking at these values can assume that
945 			 * failures persist for the life of the I/O.  That's
946 			 * important because when a device has intermittent
947 			 * connectivity problems, we want to ensure that
948 			 * they're ascribed to the device (ENXIO) and not
949 			 * the zio (EIO).
950 			 *
951 			 * Since we hold SCL_ZIO as writer here, clear both
952 			 * values so the probe can reevaluate from first
953 			 * principles.
954 			 */
955 			vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
956 			vd->vdev_cant_read = B_FALSE;
957 			vd->vdev_cant_write = B_FALSE;
958 		}
959 
960 		vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
961 		    vdev_probe_done, vps,
962 		    vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
963 
964 		if (zio != NULL) {
965 			vd->vdev_probe_wanted = B_TRUE;
966 			spa_async_request(spa, SPA_ASYNC_PROBE);
967 		}
968 	}
969 
970 	if (zio != NULL)
971 		zio_add_child(zio, pio);
972 
973 	mutex_exit(&vd->vdev_probe_lock);
974 
975 	if (vps == NULL) {
976 		ASSERT(zio != NULL);
977 		return (NULL);
978 	}
979 
980 	for (int l = 1; l < VDEV_LABELS; l++) {
981 		zio_nowait(zio_read_phys(pio, vd,
982 		    vdev_label_offset(vd->vdev_psize, l,
983 		    offsetof(vdev_label_t, vl_pad2)),
984 		    VDEV_PAD_SIZE, zio_buf_alloc(VDEV_PAD_SIZE),
985 		    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
986 		    ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
987 	}
988 
989 	if (zio == NULL)
990 		return (pio);
991 
992 	zio_nowait(pio);
993 	return (NULL);
994 }
995 
996 static void
997 vdev_open_child(void *arg)
998 {
999 	vdev_t *vd = arg;
1000 
1001 	vd->vdev_open_thread = curthread;
1002 	vd->vdev_open_error = vdev_open(vd);
1003 	vd->vdev_open_thread = NULL;
1004 }
1005 
1006 void
1007 vdev_open_children(vdev_t *vd)
1008 {
1009 	taskq_t *tq;
1010 	int children = vd->vdev_children;
1011 
1012 	tq = taskq_create("vdev_open", children, minclsyspri,
1013 	    children, children, TASKQ_PREPOPULATE);
1014 
1015 	for (int c = 0; c < children; c++)
1016 		VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c],
1017 		    TQ_SLEEP) != NULL);
1018 
1019 	taskq_destroy(tq);
1020 }
1021 
1022 /*
1023  * Prepare a virtual device for access.
1024  */
1025 int
1026 vdev_open(vdev_t *vd)
1027 {
1028 	spa_t *spa = vd->vdev_spa;
1029 	int error;
1030 	uint64_t osize = 0;
1031 	uint64_t asize, psize;
1032 	uint64_t ashift = 0;
1033 
1034 	ASSERT(vd->vdev_open_thread == curthread ||
1035 	    spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1036 	ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1037 	    vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1038 	    vd->vdev_state == VDEV_STATE_OFFLINE);
1039 
1040 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1041 	vd->vdev_cant_read = B_FALSE;
1042 	vd->vdev_cant_write = B_FALSE;
1043 	vd->vdev_min_asize = vdev_get_min_asize(vd);
1044 
1045 	if (!vd->vdev_removed && vd->vdev_faulted) {
1046 		ASSERT(vd->vdev_children == 0);
1047 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1048 		    VDEV_AUX_ERR_EXCEEDED);
1049 		return (ENXIO);
1050 	} else if (vd->vdev_offline) {
1051 		ASSERT(vd->vdev_children == 0);
1052 		vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1053 		return (ENXIO);
1054 	}
1055 
1056 	error = vd->vdev_ops->vdev_op_open(vd, &osize, &ashift);
1057 
1058 	if (zio_injection_enabled && error == 0)
1059 		error = zio_handle_device_injection(vd, NULL, ENXIO);
1060 
1061 	if (error) {
1062 		if (vd->vdev_removed &&
1063 		    vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1064 			vd->vdev_removed = B_FALSE;
1065 
1066 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1067 		    vd->vdev_stat.vs_aux);
1068 		return (error);
1069 	}
1070 
1071 	vd->vdev_removed = B_FALSE;
1072 
1073 	if (vd->vdev_degraded) {
1074 		ASSERT(vd->vdev_children == 0);
1075 		vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1076 		    VDEV_AUX_ERR_EXCEEDED);
1077 	} else {
1078 		vd->vdev_state = VDEV_STATE_HEALTHY;
1079 	}
1080 
1081 	for (int c = 0; c < vd->vdev_children; c++) {
1082 		if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1083 			vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1084 			    VDEV_AUX_NONE);
1085 			break;
1086 		}
1087 	}
1088 
1089 	osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1090 
1091 	if (vd->vdev_children == 0) {
1092 		if (osize < SPA_MINDEVSIZE) {
1093 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1094 			    VDEV_AUX_TOO_SMALL);
1095 			return (EOVERFLOW);
1096 		}
1097 		psize = osize;
1098 		asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1099 	} else {
1100 		if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1101 		    (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1102 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1103 			    VDEV_AUX_TOO_SMALL);
1104 			return (EOVERFLOW);
1105 		}
1106 		psize = 0;
1107 		asize = osize;
1108 	}
1109 
1110 	vd->vdev_psize = psize;
1111 
1112 	/*
1113 	 * Make sure the allocatable size hasn't shrunk.
1114 	 */
1115 	if (asize < vd->vdev_min_asize) {
1116 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1117 		    VDEV_AUX_BAD_LABEL);
1118 		return (EINVAL);
1119 	}
1120 
1121 	if (vd->vdev_asize == 0) {
1122 		/*
1123 		 * This is the first-ever open, so use the computed values.
1124 		 * For testing purposes, a higher ashift can be requested.
1125 		 */
1126 		vd->vdev_asize = asize;
1127 		vd->vdev_ashift = MAX(ashift, vd->vdev_ashift);
1128 	} else {
1129 		/*
1130 		 * Make sure the alignment requirement hasn't increased.
1131 		 */
1132 		if (ashift > vd->vdev_top->vdev_ashift) {
1133 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1134 			    VDEV_AUX_BAD_LABEL);
1135 			return (EINVAL);
1136 		}
1137 	}
1138 
1139 	/*
1140 	 * If all children are healthy and the asize has increased,
1141 	 * then we've experienced dynamic LUN growth.  If automatic
1142 	 * expansion is enabled then use the additional space.
1143 	 */
1144 	if (vd->vdev_state == VDEV_STATE_HEALTHY && asize > vd->vdev_asize &&
1145 	    (vd->vdev_expanding || spa->spa_autoexpand))
1146 		vd->vdev_asize = asize;
1147 
1148 	vdev_set_min_asize(vd);
1149 
1150 	/*
1151 	 * Ensure we can issue some IO before declaring the
1152 	 * vdev open for business.
1153 	 */
1154 	if (vd->vdev_ops->vdev_op_leaf &&
1155 	    (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1156 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1157 		    VDEV_AUX_IO_FAILURE);
1158 		return (error);
1159 	}
1160 
1161 	/*
1162 	 * If a leaf vdev has a DTL, and seems healthy, then kick off a
1163 	 * resilver.  But don't do this if we are doing a reopen for a scrub,
1164 	 * since this would just restart the scrub we are already doing.
1165 	 */
1166 	if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1167 	    vdev_resilver_needed(vd, NULL, NULL))
1168 		spa_async_request(spa, SPA_ASYNC_RESILVER);
1169 
1170 	return (0);
1171 }
1172 
1173 /*
1174  * Called once the vdevs are all opened, this routine validates the label
1175  * contents.  This needs to be done before vdev_load() so that we don't
1176  * inadvertently do repair I/Os to the wrong device.
1177  *
1178  * This function will only return failure if one of the vdevs indicates that it
1179  * has since been destroyed or exported.  This is only possible if
1180  * /etc/zfs/zpool.cache was readonly at the time.  Otherwise, the vdev state
1181  * will be updated but the function will return 0.
1182  */
1183 int
1184 vdev_validate(vdev_t *vd)
1185 {
1186 	spa_t *spa = vd->vdev_spa;
1187 	nvlist_t *label;
1188 	uint64_t guid, top_guid;
1189 	uint64_t state;
1190 
1191 	for (int c = 0; c < vd->vdev_children; c++)
1192 		if (vdev_validate(vd->vdev_child[c]) != 0)
1193 			return (EBADF);
1194 
1195 	/*
1196 	 * If the device has already failed, or was marked offline, don't do
1197 	 * any further validation.  Otherwise, label I/O will fail and we will
1198 	 * overwrite the previous state.
1199 	 */
1200 	if (vd->vdev_ops->vdev_op_leaf && vdev_readable(vd)) {
1201 
1202 		if ((label = vdev_label_read_config(vd)) == NULL) {
1203 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1204 			    VDEV_AUX_BAD_LABEL);
1205 			return (0);
1206 		}
1207 
1208 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_GUID,
1209 		    &guid) != 0 || guid != spa_guid(spa)) {
1210 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1211 			    VDEV_AUX_CORRUPT_DATA);
1212 			nvlist_free(label);
1213 			return (0);
1214 		}
1215 
1216 		/*
1217 		 * If this vdev just became a top-level vdev because its
1218 		 * sibling was detached, it will have adopted the parent's
1219 		 * vdev guid -- but the label may or may not be on disk yet.
1220 		 * Fortunately, either version of the label will have the
1221 		 * same top guid, so if we're a top-level vdev, we can
1222 		 * safely compare to that instead.
1223 		 */
1224 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID,
1225 		    &guid) != 0 ||
1226 		    nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID,
1227 		    &top_guid) != 0 ||
1228 		    (vd->vdev_guid != guid &&
1229 		    (vd->vdev_guid != top_guid || vd != vd->vdev_top))) {
1230 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1231 			    VDEV_AUX_CORRUPT_DATA);
1232 			nvlist_free(label);
1233 			return (0);
1234 		}
1235 
1236 		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1237 		    &state) != 0) {
1238 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1239 			    VDEV_AUX_CORRUPT_DATA);
1240 			nvlist_free(label);
1241 			return (0);
1242 		}
1243 
1244 		nvlist_free(label);
1245 
1246 		if (spa->spa_load_state == SPA_LOAD_OPEN &&
1247 		    state != POOL_STATE_ACTIVE)
1248 			return (EBADF);
1249 
1250 		/*
1251 		 * If we were able to open and validate a vdev that was
1252 		 * previously marked permanently unavailable, clear that state
1253 		 * now.
1254 		 */
1255 		if (vd->vdev_not_present)
1256 			vd->vdev_not_present = 0;
1257 	}
1258 
1259 	return (0);
1260 }
1261 
1262 /*
1263  * Close a virtual device.
1264  */
1265 void
1266 vdev_close(vdev_t *vd)
1267 {
1268 	spa_t *spa = vd->vdev_spa;
1269 
1270 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1271 
1272 	vd->vdev_ops->vdev_op_close(vd);
1273 
1274 	vdev_cache_purge(vd);
1275 
1276 	/*
1277 	 * We record the previous state before we close it, so that if we are
1278 	 * doing a reopen(), we don't generate FMA ereports if we notice that
1279 	 * it's still faulted.
1280 	 */
1281 	vd->vdev_prevstate = vd->vdev_state;
1282 
1283 	if (vd->vdev_offline)
1284 		vd->vdev_state = VDEV_STATE_OFFLINE;
1285 	else
1286 		vd->vdev_state = VDEV_STATE_CLOSED;
1287 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1288 }
1289 
1290 void
1291 vdev_reopen(vdev_t *vd)
1292 {
1293 	spa_t *spa = vd->vdev_spa;
1294 
1295 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1296 
1297 	vdev_close(vd);
1298 	(void) vdev_open(vd);
1299 
1300 	/*
1301 	 * Call vdev_validate() here to make sure we have the same device.
1302 	 * Otherwise, a device with an invalid label could be successfully
1303 	 * opened in response to vdev_reopen().
1304 	 */
1305 	if (vd->vdev_aux) {
1306 		(void) vdev_validate_aux(vd);
1307 		if (vdev_readable(vd) && vdev_writeable(vd) &&
1308 		    vd->vdev_aux == &spa->spa_l2cache &&
1309 		    !l2arc_vdev_present(vd))
1310 			l2arc_add_vdev(spa, vd);
1311 	} else {
1312 		(void) vdev_validate(vd);
1313 	}
1314 
1315 	/*
1316 	 * Reassess parent vdev's health.
1317 	 */
1318 	vdev_propagate_state(vd);
1319 }
1320 
1321 int
1322 vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1323 {
1324 	int error;
1325 
1326 	/*
1327 	 * Normally, partial opens (e.g. of a mirror) are allowed.
1328 	 * For a create, however, we want to fail the request if
1329 	 * there are any components we can't open.
1330 	 */
1331 	error = vdev_open(vd);
1332 
1333 	if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
1334 		vdev_close(vd);
1335 		return (error ? error : ENXIO);
1336 	}
1337 
1338 	/*
1339 	 * Recursively initialize all labels.
1340 	 */
1341 	if ((error = vdev_label_init(vd, txg, isreplacing ?
1342 	    VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
1343 		vdev_close(vd);
1344 		return (error);
1345 	}
1346 
1347 	return (0);
1348 }
1349 
1350 void
1351 vdev_metaslab_set_size(vdev_t *vd)
1352 {
1353 	/*
1354 	 * Aim for roughly 200 metaslabs per vdev.
1355 	 */
1356 	vd->vdev_ms_shift = highbit(vd->vdev_asize / 200);
1357 	vd->vdev_ms_shift = MAX(vd->vdev_ms_shift, SPA_MAXBLOCKSHIFT);
1358 }
1359 
1360 void
1361 vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
1362 {
1363 	ASSERT(vd == vd->vdev_top);
1364 	ASSERT(ISP2(flags));
1365 
1366 	if (flags & VDD_METASLAB)
1367 		(void) txg_list_add(&vd->vdev_ms_list, arg, txg);
1368 
1369 	if (flags & VDD_DTL)
1370 		(void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
1371 
1372 	(void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
1373 }
1374 
1375 /*
1376  * DTLs.
1377  *
1378  * A vdev's DTL (dirty time log) is the set of transaction groups for which
1379  * the vdev has less than perfect replication.  There are three kinds of DTL:
1380  *
1381  * DTL_MISSING: txgs for which the vdev has no valid copies of the data
1382  *
1383  * DTL_PARTIAL: txgs for which data is available, but not fully replicated
1384  *
1385  * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
1386  *	scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
1387  *	txgs that was scrubbed.
1388  *
1389  * DTL_OUTAGE: txgs which cannot currently be read, whether due to
1390  *	persistent errors or just some device being offline.
1391  *	Unlike the other three, the DTL_OUTAGE map is not generally
1392  *	maintained; it's only computed when needed, typically to
1393  *	determine whether a device can be detached.
1394  *
1395  * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
1396  * either has the data or it doesn't.
1397  *
1398  * For interior vdevs such as mirror and RAID-Z the picture is more complex.
1399  * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
1400  * if any child is less than fully replicated, then so is its parent.
1401  * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
1402  * comprising only those txgs which appear in 'maxfaults' or more children;
1403  * those are the txgs we don't have enough replication to read.  For example,
1404  * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
1405  * thus, its DTL_MISSING consists of the set of txgs that appear in more than
1406  * two child DTL_MISSING maps.
1407  *
1408  * It should be clear from the above that to compute the DTLs and outage maps
1409  * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
1410  * Therefore, that is all we keep on disk.  When loading the pool, or after
1411  * a configuration change, we generate all other DTLs from first principles.
1412  */
1413 void
1414 vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1415 {
1416 	space_map_t *sm = &vd->vdev_dtl[t];
1417 
1418 	ASSERT(t < DTL_TYPES);
1419 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1420 
1421 	mutex_enter(sm->sm_lock);
1422 	if (!space_map_contains(sm, txg, size))
1423 		space_map_add(sm, txg, size);
1424 	mutex_exit(sm->sm_lock);
1425 }
1426 
1427 boolean_t
1428 vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1429 {
1430 	space_map_t *sm = &vd->vdev_dtl[t];
1431 	boolean_t dirty = B_FALSE;
1432 
1433 	ASSERT(t < DTL_TYPES);
1434 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1435 
1436 	mutex_enter(sm->sm_lock);
1437 	if (sm->sm_space != 0)
1438 		dirty = space_map_contains(sm, txg, size);
1439 	mutex_exit(sm->sm_lock);
1440 
1441 	return (dirty);
1442 }
1443 
1444 boolean_t
1445 vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
1446 {
1447 	space_map_t *sm = &vd->vdev_dtl[t];
1448 	boolean_t empty;
1449 
1450 	mutex_enter(sm->sm_lock);
1451 	empty = (sm->sm_space == 0);
1452 	mutex_exit(sm->sm_lock);
1453 
1454 	return (empty);
1455 }
1456 
1457 /*
1458  * Reassess DTLs after a config change or scrub completion.
1459  */
1460 void
1461 vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
1462 {
1463 	spa_t *spa = vd->vdev_spa;
1464 	avl_tree_t reftree;
1465 	int minref;
1466 
1467 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
1468 
1469 	for (int c = 0; c < vd->vdev_children; c++)
1470 		vdev_dtl_reassess(vd->vdev_child[c], txg,
1471 		    scrub_txg, scrub_done);
1472 
1473 	if (vd == spa->spa_root_vdev)
1474 		return;
1475 
1476 	if (vd->vdev_ops->vdev_op_leaf) {
1477 		mutex_enter(&vd->vdev_dtl_lock);
1478 		if (scrub_txg != 0 &&
1479 		    (spa->spa_scrub_started || spa->spa_scrub_errors == 0)) {
1480 			/* XXX should check scrub_done? */
1481 			/*
1482 			 * We completed a scrub up to scrub_txg.  If we
1483 			 * did it without rebooting, then the scrub dtl
1484 			 * will be valid, so excise the old region and
1485 			 * fold in the scrub dtl.  Otherwise, leave the
1486 			 * dtl as-is if there was an error.
1487 			 *
1488 			 * There's little trick here: to excise the beginning
1489 			 * of the DTL_MISSING map, we put it into a reference
1490 			 * tree and then add a segment with refcnt -1 that
1491 			 * covers the range [0, scrub_txg).  This means
1492 			 * that each txg in that range has refcnt -1 or 0.
1493 			 * We then add DTL_SCRUB with a refcnt of 2, so that
1494 			 * entries in the range [0, scrub_txg) will have a
1495 			 * positive refcnt -- either 1 or 2.  We then convert
1496 			 * the reference tree into the new DTL_MISSING map.
1497 			 */
1498 			space_map_ref_create(&reftree);
1499 			space_map_ref_add_map(&reftree,
1500 			    &vd->vdev_dtl[DTL_MISSING], 1);
1501 			space_map_ref_add_seg(&reftree, 0, scrub_txg, -1);
1502 			space_map_ref_add_map(&reftree,
1503 			    &vd->vdev_dtl[DTL_SCRUB], 2);
1504 			space_map_ref_generate_map(&reftree,
1505 			    &vd->vdev_dtl[DTL_MISSING], 1);
1506 			space_map_ref_destroy(&reftree);
1507 		}
1508 		space_map_vacate(&vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
1509 		space_map_walk(&vd->vdev_dtl[DTL_MISSING],
1510 		    space_map_add, &vd->vdev_dtl[DTL_PARTIAL]);
1511 		if (scrub_done)
1512 			space_map_vacate(&vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
1513 		space_map_vacate(&vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
1514 		if (!vdev_readable(vd))
1515 			space_map_add(&vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
1516 		else
1517 			space_map_walk(&vd->vdev_dtl[DTL_MISSING],
1518 			    space_map_add, &vd->vdev_dtl[DTL_OUTAGE]);
1519 		mutex_exit(&vd->vdev_dtl_lock);
1520 
1521 		if (txg != 0)
1522 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
1523 		return;
1524 	}
1525 
1526 	mutex_enter(&vd->vdev_dtl_lock);
1527 	for (int t = 0; t < DTL_TYPES; t++) {
1528 		if (t == DTL_SCRUB)
1529 			continue;			/* leaf vdevs only */
1530 		if (t == DTL_PARTIAL)
1531 			minref = 1;			/* i.e. non-zero */
1532 		else if (vd->vdev_nparity != 0)
1533 			minref = vd->vdev_nparity + 1;	/* RAID-Z */
1534 		else
1535 			minref = vd->vdev_children;	/* any kind of mirror */
1536 		space_map_ref_create(&reftree);
1537 		for (int c = 0; c < vd->vdev_children; c++) {
1538 			vdev_t *cvd = vd->vdev_child[c];
1539 			mutex_enter(&cvd->vdev_dtl_lock);
1540 			space_map_ref_add_map(&reftree, &cvd->vdev_dtl[t], 1);
1541 			mutex_exit(&cvd->vdev_dtl_lock);
1542 		}
1543 		space_map_ref_generate_map(&reftree, &vd->vdev_dtl[t], minref);
1544 		space_map_ref_destroy(&reftree);
1545 	}
1546 	mutex_exit(&vd->vdev_dtl_lock);
1547 }
1548 
1549 static int
1550 vdev_dtl_load(vdev_t *vd)
1551 {
1552 	spa_t *spa = vd->vdev_spa;
1553 	space_map_obj_t *smo = &vd->vdev_dtl_smo;
1554 	objset_t *mos = spa->spa_meta_objset;
1555 	dmu_buf_t *db;
1556 	int error;
1557 
1558 	ASSERT(vd->vdev_children == 0);
1559 
1560 	if (smo->smo_object == 0)
1561 		return (0);
1562 
1563 	if ((error = dmu_bonus_hold(mos, smo->smo_object, FTAG, &db)) != 0)
1564 		return (error);
1565 
1566 	ASSERT3U(db->db_size, >=, sizeof (*smo));
1567 	bcopy(db->db_data, smo, sizeof (*smo));
1568 	dmu_buf_rele(db, FTAG);
1569 
1570 	mutex_enter(&vd->vdev_dtl_lock);
1571 	error = space_map_load(&vd->vdev_dtl[DTL_MISSING],
1572 	    NULL, SM_ALLOC, smo, mos);
1573 	mutex_exit(&vd->vdev_dtl_lock);
1574 
1575 	return (error);
1576 }
1577 
1578 void
1579 vdev_dtl_sync(vdev_t *vd, uint64_t txg)
1580 {
1581 	spa_t *spa = vd->vdev_spa;
1582 	space_map_obj_t *smo = &vd->vdev_dtl_smo;
1583 	space_map_t *sm = &vd->vdev_dtl[DTL_MISSING];
1584 	objset_t *mos = spa->spa_meta_objset;
1585 	space_map_t smsync;
1586 	kmutex_t smlock;
1587 	dmu_buf_t *db;
1588 	dmu_tx_t *tx;
1589 
1590 	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1591 
1592 	if (vd->vdev_detached) {
1593 		if (smo->smo_object != 0) {
1594 			int err = dmu_object_free(mos, smo->smo_object, tx);
1595 			ASSERT3U(err, ==, 0);
1596 			smo->smo_object = 0;
1597 		}
1598 		dmu_tx_commit(tx);
1599 		return;
1600 	}
1601 
1602 	if (smo->smo_object == 0) {
1603 		ASSERT(smo->smo_objsize == 0);
1604 		ASSERT(smo->smo_alloc == 0);
1605 		smo->smo_object = dmu_object_alloc(mos,
1606 		    DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
1607 		    DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
1608 		ASSERT(smo->smo_object != 0);
1609 		vdev_config_dirty(vd->vdev_top);
1610 	}
1611 
1612 	mutex_init(&smlock, NULL, MUTEX_DEFAULT, NULL);
1613 
1614 	space_map_create(&smsync, sm->sm_start, sm->sm_size, sm->sm_shift,
1615 	    &smlock);
1616 
1617 	mutex_enter(&smlock);
1618 
1619 	mutex_enter(&vd->vdev_dtl_lock);
1620 	space_map_walk(sm, space_map_add, &smsync);
1621 	mutex_exit(&vd->vdev_dtl_lock);
1622 
1623 	space_map_truncate(smo, mos, tx);
1624 	space_map_sync(&smsync, SM_ALLOC, smo, mos, tx);
1625 
1626 	space_map_destroy(&smsync);
1627 
1628 	mutex_exit(&smlock);
1629 	mutex_destroy(&smlock);
1630 
1631 	VERIFY(0 == dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
1632 	dmu_buf_will_dirty(db, tx);
1633 	ASSERT3U(db->db_size, >=, sizeof (*smo));
1634 	bcopy(smo, db->db_data, sizeof (*smo));
1635 	dmu_buf_rele(db, FTAG);
1636 
1637 	dmu_tx_commit(tx);
1638 }
1639 
1640 /*
1641  * Determine whether the specified vdev can be offlined/detached/removed
1642  * without losing data.
1643  */
1644 boolean_t
1645 vdev_dtl_required(vdev_t *vd)
1646 {
1647 	spa_t *spa = vd->vdev_spa;
1648 	vdev_t *tvd = vd->vdev_top;
1649 	uint8_t cant_read = vd->vdev_cant_read;
1650 	boolean_t required;
1651 
1652 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1653 
1654 	if (vd == spa->spa_root_vdev || vd == tvd)
1655 		return (B_TRUE);
1656 
1657 	/*
1658 	 * Temporarily mark the device as unreadable, and then determine
1659 	 * whether this results in any DTL outages in the top-level vdev.
1660 	 * If not, we can safely offline/detach/remove the device.
1661 	 */
1662 	vd->vdev_cant_read = B_TRUE;
1663 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1664 	required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
1665 	vd->vdev_cant_read = cant_read;
1666 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1667 
1668 	return (required);
1669 }
1670 
1671 /*
1672  * Determine if resilver is needed, and if so the txg range.
1673  */
1674 boolean_t
1675 vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
1676 {
1677 	boolean_t needed = B_FALSE;
1678 	uint64_t thismin = UINT64_MAX;
1679 	uint64_t thismax = 0;
1680 
1681 	if (vd->vdev_children == 0) {
1682 		mutex_enter(&vd->vdev_dtl_lock);
1683 		if (vd->vdev_dtl[DTL_MISSING].sm_space != 0 &&
1684 		    vdev_writeable(vd)) {
1685 			space_seg_t *ss;
1686 
1687 			ss = avl_first(&vd->vdev_dtl[DTL_MISSING].sm_root);
1688 			thismin = ss->ss_start - 1;
1689 			ss = avl_last(&vd->vdev_dtl[DTL_MISSING].sm_root);
1690 			thismax = ss->ss_end;
1691 			needed = B_TRUE;
1692 		}
1693 		mutex_exit(&vd->vdev_dtl_lock);
1694 	} else {
1695 		for (int c = 0; c < vd->vdev_children; c++) {
1696 			vdev_t *cvd = vd->vdev_child[c];
1697 			uint64_t cmin, cmax;
1698 
1699 			if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
1700 				thismin = MIN(thismin, cmin);
1701 				thismax = MAX(thismax, cmax);
1702 				needed = B_TRUE;
1703 			}
1704 		}
1705 	}
1706 
1707 	if (needed && minp) {
1708 		*minp = thismin;
1709 		*maxp = thismax;
1710 	}
1711 	return (needed);
1712 }
1713 
1714 void
1715 vdev_load(vdev_t *vd)
1716 {
1717 	/*
1718 	 * Recursively load all children.
1719 	 */
1720 	for (int c = 0; c < vd->vdev_children; c++)
1721 		vdev_load(vd->vdev_child[c]);
1722 
1723 	/*
1724 	 * If this is a top-level vdev, initialize its metaslabs.
1725 	 */
1726 	if (vd == vd->vdev_top &&
1727 	    (vd->vdev_ashift == 0 || vd->vdev_asize == 0 ||
1728 	    vdev_metaslab_init(vd, 0) != 0))
1729 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1730 		    VDEV_AUX_CORRUPT_DATA);
1731 
1732 	/*
1733 	 * If this is a leaf vdev, load its DTL.
1734 	 */
1735 	if (vd->vdev_ops->vdev_op_leaf && vdev_dtl_load(vd) != 0)
1736 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1737 		    VDEV_AUX_CORRUPT_DATA);
1738 }
1739 
1740 /*
1741  * The special vdev case is used for hot spares and l2cache devices.  Its
1742  * sole purpose it to set the vdev state for the associated vdev.  To do this,
1743  * we make sure that we can open the underlying device, then try to read the
1744  * label, and make sure that the label is sane and that it hasn't been
1745  * repurposed to another pool.
1746  */
1747 int
1748 vdev_validate_aux(vdev_t *vd)
1749 {
1750 	nvlist_t *label;
1751 	uint64_t guid, version;
1752 	uint64_t state;
1753 
1754 	if (!vdev_readable(vd))
1755 		return (0);
1756 
1757 	if ((label = vdev_label_read_config(vd)) == NULL) {
1758 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1759 		    VDEV_AUX_CORRUPT_DATA);
1760 		return (-1);
1761 	}
1762 
1763 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
1764 	    version > SPA_VERSION ||
1765 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
1766 	    guid != vd->vdev_guid ||
1767 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
1768 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1769 		    VDEV_AUX_CORRUPT_DATA);
1770 		nvlist_free(label);
1771 		return (-1);
1772 	}
1773 
1774 	/*
1775 	 * We don't actually check the pool state here.  If it's in fact in
1776 	 * use by another pool, we update this fact on the fly when requested.
1777 	 */
1778 	nvlist_free(label);
1779 	return (0);
1780 }
1781 
1782 void
1783 vdev_sync_done(vdev_t *vd, uint64_t txg)
1784 {
1785 	metaslab_t *msp;
1786 
1787 	while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
1788 		metaslab_sync_done(msp, txg);
1789 }
1790 
1791 void
1792 vdev_sync(vdev_t *vd, uint64_t txg)
1793 {
1794 	spa_t *spa = vd->vdev_spa;
1795 	vdev_t *lvd;
1796 	metaslab_t *msp;
1797 	dmu_tx_t *tx;
1798 
1799 	if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0) {
1800 		ASSERT(vd == vd->vdev_top);
1801 		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1802 		vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
1803 		    DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
1804 		ASSERT(vd->vdev_ms_array != 0);
1805 		vdev_config_dirty(vd);
1806 		dmu_tx_commit(tx);
1807 	}
1808 
1809 	while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
1810 		metaslab_sync(msp, txg);
1811 		(void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
1812 	}
1813 
1814 	while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
1815 		vdev_dtl_sync(lvd, txg);
1816 
1817 	(void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
1818 }
1819 
1820 uint64_t
1821 vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
1822 {
1823 	return (vd->vdev_ops->vdev_op_asize(vd, psize));
1824 }
1825 
1826 /*
1827  * Mark the given vdev faulted.  A faulted vdev behaves as if the device could
1828  * not be opened, and no I/O is attempted.
1829  */
1830 int
1831 vdev_fault(spa_t *spa, uint64_t guid)
1832 {
1833 	vdev_t *vd;
1834 
1835 	spa_vdev_state_enter(spa);
1836 
1837 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
1838 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
1839 
1840 	if (!vd->vdev_ops->vdev_op_leaf)
1841 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
1842 
1843 	/*
1844 	 * Faulted state takes precedence over degraded.
1845 	 */
1846 	vd->vdev_faulted = 1ULL;
1847 	vd->vdev_degraded = 0ULL;
1848 	vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, VDEV_AUX_ERR_EXCEEDED);
1849 
1850 	/*
1851 	 * If marking the vdev as faulted cause the top-level vdev to become
1852 	 * unavailable, then back off and simply mark the vdev as degraded
1853 	 * instead.
1854 	 */
1855 	if (vdev_is_dead(vd->vdev_top) && vd->vdev_aux == NULL) {
1856 		vd->vdev_degraded = 1ULL;
1857 		vd->vdev_faulted = 0ULL;
1858 
1859 		/*
1860 		 * If we reopen the device and it's not dead, only then do we
1861 		 * mark it degraded.
1862 		 */
1863 		vdev_reopen(vd);
1864 
1865 		if (vdev_readable(vd)) {
1866 			vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
1867 			    VDEV_AUX_ERR_EXCEEDED);
1868 		}
1869 	}
1870 
1871 	return (spa_vdev_state_exit(spa, vd, 0));
1872 }
1873 
1874 /*
1875  * Mark the given vdev degraded.  A degraded vdev is purely an indication to the
1876  * user that something is wrong.  The vdev continues to operate as normal as far
1877  * as I/O is concerned.
1878  */
1879 int
1880 vdev_degrade(spa_t *spa, uint64_t guid)
1881 {
1882 	vdev_t *vd;
1883 
1884 	spa_vdev_state_enter(spa);
1885 
1886 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
1887 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
1888 
1889 	if (!vd->vdev_ops->vdev_op_leaf)
1890 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
1891 
1892 	/*
1893 	 * If the vdev is already faulted, then don't do anything.
1894 	 */
1895 	if (vd->vdev_faulted || vd->vdev_degraded)
1896 		return (spa_vdev_state_exit(spa, NULL, 0));
1897 
1898 	vd->vdev_degraded = 1ULL;
1899 	if (!vdev_is_dead(vd))
1900 		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
1901 		    VDEV_AUX_ERR_EXCEEDED);
1902 
1903 	return (spa_vdev_state_exit(spa, vd, 0));
1904 }
1905 
1906 /*
1907  * Online the given vdev.  If 'unspare' is set, it implies two things.  First,
1908  * any attached spare device should be detached when the device finishes
1909  * resilvering.  Second, the online should be treated like a 'test' online case,
1910  * so no FMA events are generated if the device fails to open.
1911  */
1912 int
1913 vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
1914 {
1915 	vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
1916 
1917 	spa_vdev_state_enter(spa);
1918 
1919 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
1920 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
1921 
1922 	if (!vd->vdev_ops->vdev_op_leaf)
1923 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
1924 
1925 	tvd = vd->vdev_top;
1926 	vd->vdev_offline = B_FALSE;
1927 	vd->vdev_tmpoffline = B_FALSE;
1928 	vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
1929 	vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
1930 
1931 	/* XXX - L2ARC 1.0 does not support expansion */
1932 	if (!vd->vdev_aux) {
1933 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
1934 			pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
1935 	}
1936 
1937 	vdev_reopen(tvd);
1938 	vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
1939 
1940 	if (!vd->vdev_aux) {
1941 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
1942 			pvd->vdev_expanding = B_FALSE;
1943 	}
1944 
1945 	if (newstate)
1946 		*newstate = vd->vdev_state;
1947 	if ((flags & ZFS_ONLINE_UNSPARE) &&
1948 	    !vdev_is_dead(vd) && vd->vdev_parent &&
1949 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
1950 	    vd->vdev_parent->vdev_child[0] == vd)
1951 		vd->vdev_unspare = B_TRUE;
1952 
1953 	if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
1954 
1955 		/* XXX - L2ARC 1.0 does not support expansion */
1956 		if (vd->vdev_aux)
1957 			return (spa_vdev_state_exit(spa, vd, ENOTSUP));
1958 		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
1959 	}
1960 	return (spa_vdev_state_exit(spa, vd, 0));
1961 }
1962 
1963 int
1964 vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
1965 {
1966 	vdev_t *vd, *tvd;
1967 	int error;
1968 
1969 	spa_vdev_state_enter(spa);
1970 
1971 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
1972 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
1973 
1974 	if (!vd->vdev_ops->vdev_op_leaf)
1975 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
1976 
1977 	tvd = vd->vdev_top;
1978 
1979 	/*
1980 	 * If the device isn't already offline, try to offline it.
1981 	 */
1982 	if (!vd->vdev_offline) {
1983 		/*
1984 		 * If this device has the only valid copy of some data,
1985 		 * don't allow it to be offlined. Log devices are always
1986 		 * expendable.
1987 		 */
1988 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
1989 		    vdev_dtl_required(vd))
1990 			return (spa_vdev_state_exit(spa, NULL, EBUSY));
1991 
1992 		/*
1993 		 * Offline this device and reopen its top-level vdev.
1994 		 * If the top-level vdev is a log device then just offline
1995 		 * it. Otherwise, if this action results in the top-level
1996 		 * vdev becoming unusable, undo it and fail the request.
1997 		 */
1998 		vd->vdev_offline = B_TRUE;
1999 		vdev_reopen(tvd);
2000 
2001 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2002 		    vdev_is_dead(tvd)) {
2003 			vd->vdev_offline = B_FALSE;
2004 			vdev_reopen(tvd);
2005 			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2006 		}
2007 	}
2008 
2009 	vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
2010 
2011 	if (!tvd->vdev_islog || !vdev_is_dead(tvd))
2012 		return (spa_vdev_state_exit(spa, vd, 0));
2013 
2014 	(void) spa_vdev_state_exit(spa, vd, 0);
2015 
2016 	error = dmu_objset_find(spa_name(spa), zil_vdev_offline,
2017 	    NULL, DS_FIND_CHILDREN);
2018 	if (error) {
2019 		(void) vdev_online(spa, guid, 0, NULL);
2020 		return (error);
2021 	}
2022 	/*
2023 	 * If we successfully offlined the log device then we need to
2024 	 * sync out the current txg so that the "stubby" block can be
2025 	 * removed by zil_sync().
2026 	 */
2027 	txg_wait_synced(spa->spa_dsl_pool, 0);
2028 	return (0);
2029 }
2030 
2031 /*
2032  * Clear the error counts associated with this vdev.  Unlike vdev_online() and
2033  * vdev_offline(), we assume the spa config is locked.  We also clear all
2034  * children.  If 'vd' is NULL, then the user wants to clear all vdevs.
2035  */
2036 void
2037 vdev_clear(spa_t *spa, vdev_t *vd)
2038 {
2039 	vdev_t *rvd = spa->spa_root_vdev;
2040 
2041 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2042 
2043 	if (vd == NULL)
2044 		vd = rvd;
2045 
2046 	vd->vdev_stat.vs_read_errors = 0;
2047 	vd->vdev_stat.vs_write_errors = 0;
2048 	vd->vdev_stat.vs_checksum_errors = 0;
2049 
2050 	for (int c = 0; c < vd->vdev_children; c++)
2051 		vdev_clear(spa, vd->vdev_child[c]);
2052 
2053 	/*
2054 	 * If we're in the FAULTED state or have experienced failed I/O, then
2055 	 * clear the persistent state and attempt to reopen the device.  We
2056 	 * also mark the vdev config dirty, so that the new faulted state is
2057 	 * written out to disk.
2058 	 */
2059 	if (vd->vdev_faulted || vd->vdev_degraded ||
2060 	    !vdev_readable(vd) || !vdev_writeable(vd)) {
2061 
2062 		vd->vdev_faulted = vd->vdev_degraded = 0;
2063 		vd->vdev_cant_read = B_FALSE;
2064 		vd->vdev_cant_write = B_FALSE;
2065 
2066 		vdev_reopen(vd);
2067 
2068 		if (vd != rvd)
2069 			vdev_state_dirty(vd->vdev_top);
2070 
2071 		if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
2072 			spa_async_request(spa, SPA_ASYNC_RESILVER);
2073 
2074 		spa_event_notify(spa, vd, ESC_ZFS_VDEV_CLEAR);
2075 	}
2076 }
2077 
2078 boolean_t
2079 vdev_is_dead(vdev_t *vd)
2080 {
2081 	return (vd->vdev_state < VDEV_STATE_DEGRADED);
2082 }
2083 
2084 boolean_t
2085 vdev_readable(vdev_t *vd)
2086 {
2087 	return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
2088 }
2089 
2090 boolean_t
2091 vdev_writeable(vdev_t *vd)
2092 {
2093 	return (!vdev_is_dead(vd) && !vd->vdev_cant_write);
2094 }
2095 
2096 boolean_t
2097 vdev_allocatable(vdev_t *vd)
2098 {
2099 	uint64_t state = vd->vdev_state;
2100 
2101 	/*
2102 	 * We currently allow allocations from vdevs which may be in the
2103 	 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
2104 	 * fails to reopen then we'll catch it later when we're holding
2105 	 * the proper locks.  Note that we have to get the vdev state
2106 	 * in a local variable because although it changes atomically,
2107 	 * we're asking two separate questions about it.
2108 	 */
2109 	return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
2110 	    !vd->vdev_cant_write);
2111 }
2112 
2113 boolean_t
2114 vdev_accessible(vdev_t *vd, zio_t *zio)
2115 {
2116 	ASSERT(zio->io_vd == vd);
2117 
2118 	if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
2119 		return (B_FALSE);
2120 
2121 	if (zio->io_type == ZIO_TYPE_READ)
2122 		return (!vd->vdev_cant_read);
2123 
2124 	if (zio->io_type == ZIO_TYPE_WRITE)
2125 		return (!vd->vdev_cant_write);
2126 
2127 	return (B_TRUE);
2128 }
2129 
2130 /*
2131  * Get statistics for the given vdev.
2132  */
2133 void
2134 vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
2135 {
2136 	vdev_t *rvd = vd->vdev_spa->spa_root_vdev;
2137 
2138 	mutex_enter(&vd->vdev_stat_lock);
2139 	bcopy(&vd->vdev_stat, vs, sizeof (*vs));
2140 	vs->vs_scrub_errors = vd->vdev_spa->spa_scrub_errors;
2141 	vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
2142 	vs->vs_state = vd->vdev_state;
2143 	vs->vs_rsize = vdev_get_min_asize(vd);
2144 	if (vd->vdev_ops->vdev_op_leaf)
2145 		vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
2146 	mutex_exit(&vd->vdev_stat_lock);
2147 
2148 	/*
2149 	 * If we're getting stats on the root vdev, aggregate the I/O counts
2150 	 * over all top-level vdevs (i.e. the direct children of the root).
2151 	 */
2152 	if (vd == rvd) {
2153 		for (int c = 0; c < rvd->vdev_children; c++) {
2154 			vdev_t *cvd = rvd->vdev_child[c];
2155 			vdev_stat_t *cvs = &cvd->vdev_stat;
2156 
2157 			mutex_enter(&vd->vdev_stat_lock);
2158 			for (int t = 0; t < ZIO_TYPES; t++) {
2159 				vs->vs_ops[t] += cvs->vs_ops[t];
2160 				vs->vs_bytes[t] += cvs->vs_bytes[t];
2161 			}
2162 			vs->vs_scrub_examined += cvs->vs_scrub_examined;
2163 			mutex_exit(&vd->vdev_stat_lock);
2164 		}
2165 	}
2166 }
2167 
2168 void
2169 vdev_clear_stats(vdev_t *vd)
2170 {
2171 	mutex_enter(&vd->vdev_stat_lock);
2172 	vd->vdev_stat.vs_space = 0;
2173 	vd->vdev_stat.vs_dspace = 0;
2174 	vd->vdev_stat.vs_alloc = 0;
2175 	mutex_exit(&vd->vdev_stat_lock);
2176 }
2177 
2178 void
2179 vdev_stat_update(zio_t *zio, uint64_t psize)
2180 {
2181 	spa_t *spa = zio->io_spa;
2182 	vdev_t *rvd = spa->spa_root_vdev;
2183 	vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
2184 	vdev_t *pvd;
2185 	uint64_t txg = zio->io_txg;
2186 	vdev_stat_t *vs = &vd->vdev_stat;
2187 	zio_type_t type = zio->io_type;
2188 	int flags = zio->io_flags;
2189 
2190 	/*
2191 	 * If this i/o is a gang leader, it didn't do any actual work.
2192 	 */
2193 	if (zio->io_gang_tree)
2194 		return;
2195 
2196 	if (zio->io_error == 0) {
2197 		/*
2198 		 * If this is a root i/o, don't count it -- we've already
2199 		 * counted the top-level vdevs, and vdev_get_stats() will
2200 		 * aggregate them when asked.  This reduces contention on
2201 		 * the root vdev_stat_lock and implicitly handles blocks
2202 		 * that compress away to holes, for which there is no i/o.
2203 		 * (Holes never create vdev children, so all the counters
2204 		 * remain zero, which is what we want.)
2205 		 *
2206 		 * Note: this only applies to successful i/o (io_error == 0)
2207 		 * because unlike i/o counts, errors are not additive.
2208 		 * When reading a ditto block, for example, failure of
2209 		 * one top-level vdev does not imply a root-level error.
2210 		 */
2211 		if (vd == rvd)
2212 			return;
2213 
2214 		ASSERT(vd == zio->io_vd);
2215 
2216 		if (flags & ZIO_FLAG_IO_BYPASS)
2217 			return;
2218 
2219 		mutex_enter(&vd->vdev_stat_lock);
2220 
2221 		if (flags & ZIO_FLAG_IO_REPAIR) {
2222 			if (flags & ZIO_FLAG_SCRUB_THREAD)
2223 				vs->vs_scrub_repaired += psize;
2224 			if (flags & ZIO_FLAG_SELF_HEAL)
2225 				vs->vs_self_healed += psize;
2226 		}
2227 
2228 		vs->vs_ops[type]++;
2229 		vs->vs_bytes[type] += psize;
2230 
2231 		mutex_exit(&vd->vdev_stat_lock);
2232 		return;
2233 	}
2234 
2235 	if (flags & ZIO_FLAG_SPECULATIVE)
2236 		return;
2237 
2238 	/*
2239 	 * If this is an I/O error that is going to be retried, then ignore the
2240 	 * error.  Otherwise, the user may interpret B_FAILFAST I/O errors as
2241 	 * hard errors, when in reality they can happen for any number of
2242 	 * innocuous reasons (bus resets, MPxIO link failure, etc).
2243 	 */
2244 	if (zio->io_error == EIO &&
2245 	    !(zio->io_flags & ZIO_FLAG_IO_RETRY))
2246 		return;
2247 
2248 	mutex_enter(&vd->vdev_stat_lock);
2249 	if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
2250 		if (zio->io_error == ECKSUM)
2251 			vs->vs_checksum_errors++;
2252 		else
2253 			vs->vs_read_errors++;
2254 	}
2255 	if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
2256 		vs->vs_write_errors++;
2257 	mutex_exit(&vd->vdev_stat_lock);
2258 
2259 	if (type == ZIO_TYPE_WRITE && txg != 0 &&
2260 	    (!(flags & ZIO_FLAG_IO_REPAIR) ||
2261 	    (flags & ZIO_FLAG_SCRUB_THREAD))) {
2262 		/*
2263 		 * This is either a normal write (not a repair), or it's a
2264 		 * repair induced by the scrub thread.  In the normal case,
2265 		 * we commit the DTL change in the same txg as the block
2266 		 * was born.  In the scrub-induced repair case, we know that
2267 		 * scrubs run in first-pass syncing context, so we commit
2268 		 * the DTL change in spa->spa_syncing_txg.
2269 		 *
2270 		 * We currently do not make DTL entries for failed spontaneous
2271 		 * self-healing writes triggered by normal (non-scrubbing)
2272 		 * reads, because we have no transactional context in which to
2273 		 * do so -- and it's not clear that it'd be desirable anyway.
2274 		 */
2275 		if (vd->vdev_ops->vdev_op_leaf) {
2276 			uint64_t commit_txg = txg;
2277 			if (flags & ZIO_FLAG_SCRUB_THREAD) {
2278 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2279 				ASSERT(spa_sync_pass(spa) == 1);
2280 				vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
2281 				commit_txg = spa->spa_syncing_txg;
2282 			}
2283 			ASSERT(commit_txg >= spa->spa_syncing_txg);
2284 			if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
2285 				return;
2286 			for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2287 				vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
2288 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
2289 		}
2290 		if (vd != rvd)
2291 			vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
2292 	}
2293 }
2294 
2295 void
2296 vdev_scrub_stat_update(vdev_t *vd, pool_scrub_type_t type, boolean_t complete)
2297 {
2298 	vdev_stat_t *vs = &vd->vdev_stat;
2299 
2300 	for (int c = 0; c < vd->vdev_children; c++)
2301 		vdev_scrub_stat_update(vd->vdev_child[c], type, complete);
2302 
2303 	mutex_enter(&vd->vdev_stat_lock);
2304 
2305 	if (type == POOL_SCRUB_NONE) {
2306 		/*
2307 		 * Update completion and end time.  Leave everything else alone
2308 		 * so we can report what happened during the previous scrub.
2309 		 */
2310 		vs->vs_scrub_complete = complete;
2311 		vs->vs_scrub_end = gethrestime_sec();
2312 	} else {
2313 		vs->vs_scrub_type = type;
2314 		vs->vs_scrub_complete = 0;
2315 		vs->vs_scrub_examined = 0;
2316 		vs->vs_scrub_repaired = 0;
2317 		vs->vs_scrub_start = gethrestime_sec();
2318 		vs->vs_scrub_end = 0;
2319 	}
2320 
2321 	mutex_exit(&vd->vdev_stat_lock);
2322 }
2323 
2324 /*
2325  * Update the in-core space usage stats for this vdev and the root vdev.
2326  */
2327 void
2328 vdev_space_update(vdev_t *vd, int64_t space_delta, int64_t alloc_delta,
2329     boolean_t update_root)
2330 {
2331 	int64_t dspace_delta = space_delta;
2332 	spa_t *spa = vd->vdev_spa;
2333 	vdev_t *rvd = spa->spa_root_vdev;
2334 
2335 	ASSERT(vd == vd->vdev_top);
2336 
2337 	/*
2338 	 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
2339 	 * factor.  We must calculate this here and not at the root vdev
2340 	 * because the root vdev's psize-to-asize is simply the max of its
2341 	 * childrens', thus not accurate enough for us.
2342 	 */
2343 	ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
2344 	ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
2345 	dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
2346 	    vd->vdev_deflate_ratio;
2347 
2348 	mutex_enter(&vd->vdev_stat_lock);
2349 	vd->vdev_stat.vs_space += space_delta;
2350 	vd->vdev_stat.vs_alloc += alloc_delta;
2351 	vd->vdev_stat.vs_dspace += dspace_delta;
2352 	mutex_exit(&vd->vdev_stat_lock);
2353 
2354 	if (update_root) {
2355 		ASSERT(rvd == vd->vdev_parent);
2356 		ASSERT(vd->vdev_ms_count != 0);
2357 
2358 		/*
2359 		 * Don't count non-normal (e.g. intent log) space as part of
2360 		 * the pool's capacity.
2361 		 */
2362 		if (vd->vdev_mg->mg_class != spa->spa_normal_class)
2363 			return;
2364 
2365 		mutex_enter(&rvd->vdev_stat_lock);
2366 		rvd->vdev_stat.vs_space += space_delta;
2367 		rvd->vdev_stat.vs_alloc += alloc_delta;
2368 		rvd->vdev_stat.vs_dspace += dspace_delta;
2369 		mutex_exit(&rvd->vdev_stat_lock);
2370 	}
2371 }
2372 
2373 /*
2374  * Mark a top-level vdev's config as dirty, placing it on the dirty list
2375  * so that it will be written out next time the vdev configuration is synced.
2376  * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
2377  */
2378 void
2379 vdev_config_dirty(vdev_t *vd)
2380 {
2381 	spa_t *spa = vd->vdev_spa;
2382 	vdev_t *rvd = spa->spa_root_vdev;
2383 	int c;
2384 
2385 	/*
2386 	 * If this is an aux vdev (as with l2cache and spare devices), then we
2387 	 * update the vdev config manually and set the sync flag.
2388 	 */
2389 	if (vd->vdev_aux != NULL) {
2390 		spa_aux_vdev_t *sav = vd->vdev_aux;
2391 		nvlist_t **aux;
2392 		uint_t naux;
2393 
2394 		for (c = 0; c < sav->sav_count; c++) {
2395 			if (sav->sav_vdevs[c] == vd)
2396 				break;
2397 		}
2398 
2399 		if (c == sav->sav_count) {
2400 			/*
2401 			 * We're being removed.  There's nothing more to do.
2402 			 */
2403 			ASSERT(sav->sav_sync == B_TRUE);
2404 			return;
2405 		}
2406 
2407 		sav->sav_sync = B_TRUE;
2408 
2409 		if (nvlist_lookup_nvlist_array(sav->sav_config,
2410 		    ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
2411 			VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
2412 			    ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
2413 		}
2414 
2415 		ASSERT(c < naux);
2416 
2417 		/*
2418 		 * Setting the nvlist in the middle if the array is a little
2419 		 * sketchy, but it will work.
2420 		 */
2421 		nvlist_free(aux[c]);
2422 		aux[c] = vdev_config_generate(spa, vd, B_TRUE, B_FALSE, B_TRUE);
2423 
2424 		return;
2425 	}
2426 
2427 	/*
2428 	 * The dirty list is protected by the SCL_CONFIG lock.  The caller
2429 	 * must either hold SCL_CONFIG as writer, or must be the sync thread
2430 	 * (which holds SCL_CONFIG as reader).  There's only one sync thread,
2431 	 * so this is sufficient to ensure mutual exclusion.
2432 	 */
2433 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2434 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2435 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
2436 
2437 	if (vd == rvd) {
2438 		for (c = 0; c < rvd->vdev_children; c++)
2439 			vdev_config_dirty(rvd->vdev_child[c]);
2440 	} else {
2441 		ASSERT(vd == vd->vdev_top);
2442 
2443 		if (!list_link_active(&vd->vdev_config_dirty_node))
2444 			list_insert_head(&spa->spa_config_dirty_list, vd);
2445 	}
2446 }
2447 
2448 void
2449 vdev_config_clean(vdev_t *vd)
2450 {
2451 	spa_t *spa = vd->vdev_spa;
2452 
2453 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2454 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2455 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
2456 
2457 	ASSERT(list_link_active(&vd->vdev_config_dirty_node));
2458 	list_remove(&spa->spa_config_dirty_list, vd);
2459 }
2460 
2461 /*
2462  * Mark a top-level vdev's state as dirty, so that the next pass of
2463  * spa_sync() can convert this into vdev_config_dirty().  We distinguish
2464  * the state changes from larger config changes because they require
2465  * much less locking, and are often needed for administrative actions.
2466  */
2467 void
2468 vdev_state_dirty(vdev_t *vd)
2469 {
2470 	spa_t *spa = vd->vdev_spa;
2471 
2472 	ASSERT(vd == vd->vdev_top);
2473 
2474 	/*
2475 	 * The state list is protected by the SCL_STATE lock.  The caller
2476 	 * must either hold SCL_STATE as writer, or must be the sync thread
2477 	 * (which holds SCL_STATE as reader).  There's only one sync thread,
2478 	 * so this is sufficient to ensure mutual exclusion.
2479 	 */
2480 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2481 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2482 	    spa_config_held(spa, SCL_STATE, RW_READER)));
2483 
2484 	if (!list_link_active(&vd->vdev_state_dirty_node))
2485 		list_insert_head(&spa->spa_state_dirty_list, vd);
2486 }
2487 
2488 void
2489 vdev_state_clean(vdev_t *vd)
2490 {
2491 	spa_t *spa = vd->vdev_spa;
2492 
2493 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2494 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2495 	    spa_config_held(spa, SCL_STATE, RW_READER)));
2496 
2497 	ASSERT(list_link_active(&vd->vdev_state_dirty_node));
2498 	list_remove(&spa->spa_state_dirty_list, vd);
2499 }
2500 
2501 /*
2502  * Propagate vdev state up from children to parent.
2503  */
2504 void
2505 vdev_propagate_state(vdev_t *vd)
2506 {
2507 	spa_t *spa = vd->vdev_spa;
2508 	vdev_t *rvd = spa->spa_root_vdev;
2509 	int degraded = 0, faulted = 0;
2510 	int corrupted = 0;
2511 	vdev_t *child;
2512 
2513 	if (vd->vdev_children > 0) {
2514 		for (int c = 0; c < vd->vdev_children; c++) {
2515 			child = vd->vdev_child[c];
2516 
2517 			if (!vdev_readable(child) ||
2518 			    (!vdev_writeable(child) && spa_writeable(spa))) {
2519 				/*
2520 				 * Root special: if there is a top-level log
2521 				 * device, treat the root vdev as if it were
2522 				 * degraded.
2523 				 */
2524 				if (child->vdev_islog && vd == rvd)
2525 					degraded++;
2526 				else
2527 					faulted++;
2528 			} else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
2529 				degraded++;
2530 			}
2531 
2532 			if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
2533 				corrupted++;
2534 		}
2535 
2536 		vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
2537 
2538 		/*
2539 		 * Root special: if there is a top-level vdev that cannot be
2540 		 * opened due to corrupted metadata, then propagate the root
2541 		 * vdev's aux state as 'corrupt' rather than 'insufficient
2542 		 * replicas'.
2543 		 */
2544 		if (corrupted && vd == rvd &&
2545 		    rvd->vdev_state == VDEV_STATE_CANT_OPEN)
2546 			vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
2547 			    VDEV_AUX_CORRUPT_DATA);
2548 	}
2549 
2550 	if (vd->vdev_parent)
2551 		vdev_propagate_state(vd->vdev_parent);
2552 }
2553 
2554 /*
2555  * Set a vdev's state.  If this is during an open, we don't update the parent
2556  * state, because we're in the process of opening children depth-first.
2557  * Otherwise, we propagate the change to the parent.
2558  *
2559  * If this routine places a device in a faulted state, an appropriate ereport is
2560  * generated.
2561  */
2562 void
2563 vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
2564 {
2565 	uint64_t save_state;
2566 	spa_t *spa = vd->vdev_spa;
2567 
2568 	if (state == vd->vdev_state) {
2569 		vd->vdev_stat.vs_aux = aux;
2570 		return;
2571 	}
2572 
2573 	save_state = vd->vdev_state;
2574 
2575 	vd->vdev_state = state;
2576 	vd->vdev_stat.vs_aux = aux;
2577 
2578 	/*
2579 	 * If we are setting the vdev state to anything but an open state, then
2580 	 * always close the underlying device.  Otherwise, we keep accessible
2581 	 * but invalid devices open forever.  We don't call vdev_close() itself,
2582 	 * because that implies some extra checks (offline, etc) that we don't
2583 	 * want here.  This is limited to leaf devices, because otherwise
2584 	 * closing the device will affect other children.
2585 	 */
2586 	if (vdev_is_dead(vd) && vd->vdev_ops->vdev_op_leaf)
2587 		vd->vdev_ops->vdev_op_close(vd);
2588 
2589 	if (vd->vdev_removed &&
2590 	    state == VDEV_STATE_CANT_OPEN &&
2591 	    (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
2592 		/*
2593 		 * If the previous state is set to VDEV_STATE_REMOVED, then this
2594 		 * device was previously marked removed and someone attempted to
2595 		 * reopen it.  If this failed due to a nonexistent device, then
2596 		 * keep the device in the REMOVED state.  We also let this be if
2597 		 * it is one of our special test online cases, which is only
2598 		 * attempting to online the device and shouldn't generate an FMA
2599 		 * fault.
2600 		 */
2601 		vd->vdev_state = VDEV_STATE_REMOVED;
2602 		vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
2603 	} else if (state == VDEV_STATE_REMOVED) {
2604 		/*
2605 		 * Indicate to the ZFS DE that this device has been removed, and
2606 		 * any recent errors should be ignored.
2607 		 */
2608 		zfs_post_remove(spa, vd);
2609 		vd->vdev_removed = B_TRUE;
2610 	} else if (state == VDEV_STATE_CANT_OPEN) {
2611 		/*
2612 		 * If we fail to open a vdev during an import, we mark it as
2613 		 * "not available", which signifies that it was never there to
2614 		 * begin with.  Failure to open such a device is not considered
2615 		 * an error.
2616 		 */
2617 		if (spa->spa_load_state == SPA_LOAD_IMPORT &&
2618 		    vd->vdev_ops->vdev_op_leaf)
2619 			vd->vdev_not_present = 1;
2620 
2621 		/*
2622 		 * Post the appropriate ereport.  If the 'prevstate' field is
2623 		 * set to something other than VDEV_STATE_UNKNOWN, it indicates
2624 		 * that this is part of a vdev_reopen().  In this case, we don't
2625 		 * want to post the ereport if the device was already in the
2626 		 * CANT_OPEN state beforehand.
2627 		 *
2628 		 * If the 'checkremove' flag is set, then this is an attempt to
2629 		 * online the device in response to an insertion event.  If we
2630 		 * hit this case, then we have detected an insertion event for a
2631 		 * faulted or offline device that wasn't in the removed state.
2632 		 * In this scenario, we don't post an ereport because we are
2633 		 * about to replace the device, or attempt an online with
2634 		 * vdev_forcefault, which will generate the fault for us.
2635 		 */
2636 		if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
2637 		    !vd->vdev_not_present && !vd->vdev_checkremove &&
2638 		    vd != spa->spa_root_vdev) {
2639 			const char *class;
2640 
2641 			switch (aux) {
2642 			case VDEV_AUX_OPEN_FAILED:
2643 				class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
2644 				break;
2645 			case VDEV_AUX_CORRUPT_DATA:
2646 				class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
2647 				break;
2648 			case VDEV_AUX_NO_REPLICAS:
2649 				class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
2650 				break;
2651 			case VDEV_AUX_BAD_GUID_SUM:
2652 				class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
2653 				break;
2654 			case VDEV_AUX_TOO_SMALL:
2655 				class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
2656 				break;
2657 			case VDEV_AUX_BAD_LABEL:
2658 				class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
2659 				break;
2660 			case VDEV_AUX_IO_FAILURE:
2661 				class = FM_EREPORT_ZFS_IO_FAILURE;
2662 				break;
2663 			default:
2664 				class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
2665 			}
2666 
2667 			zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
2668 		}
2669 
2670 		/* Erase any notion of persistent removed state */
2671 		vd->vdev_removed = B_FALSE;
2672 	} else {
2673 		vd->vdev_removed = B_FALSE;
2674 	}
2675 
2676 	if (!isopen && vd->vdev_parent)
2677 		vdev_propagate_state(vd->vdev_parent);
2678 }
2679 
2680 /*
2681  * Check the vdev configuration to ensure that it's capable of supporting
2682  * a root pool. Currently, we do not support RAID-Z or partial configuration.
2683  * In addition, only a single top-level vdev is allowed and none of the leaves
2684  * can be wholedisks.
2685  */
2686 boolean_t
2687 vdev_is_bootable(vdev_t *vd)
2688 {
2689 	if (!vd->vdev_ops->vdev_op_leaf) {
2690 		char *vdev_type = vd->vdev_ops->vdev_op_type;
2691 
2692 		if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
2693 		    vd->vdev_children > 1) {
2694 			return (B_FALSE);
2695 		} else if (strcmp(vdev_type, VDEV_TYPE_RAIDZ) == 0 ||
2696 		    strcmp(vdev_type, VDEV_TYPE_MISSING) == 0) {
2697 			return (B_FALSE);
2698 		}
2699 	} else if (vd->vdev_wholedisk == 1) {
2700 		return (B_FALSE);
2701 	}
2702 
2703 	for (int c = 0; c < vd->vdev_children; c++) {
2704 		if (!vdev_is_bootable(vd->vdev_child[c]))
2705 			return (B_FALSE);
2706 	}
2707 	return (B_TRUE);
2708 }
2709 
2710 void
2711 vdev_load_log_state(vdev_t *vd, nvlist_t *nv)
2712 {
2713 	uint_t children;
2714 	nvlist_t **child;
2715 	uint64_t val;
2716 	spa_t *spa = vd->vdev_spa;
2717 
2718 	if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
2719 	    &child, &children) == 0) {
2720 		for (int c = 0; c < children; c++)
2721 			vdev_load_log_state(vd->vdev_child[c], child[c]);
2722 	}
2723 
2724 	if (vd->vdev_ops->vdev_op_leaf && nvlist_lookup_uint64(nv,
2725 	    ZPOOL_CONFIG_OFFLINE, &val) == 0 && val) {
2726 
2727 		/*
2728 		 * It would be nice to call vdev_offline()
2729 		 * directly but the pool isn't fully loaded and
2730 		 * the txg threads have not been started yet.
2731 		 */
2732 		spa_config_enter(spa, SCL_STATE_ALL, FTAG, RW_WRITER);
2733 		vd->vdev_offline = val;
2734 		vdev_reopen(vd->vdev_top);
2735 		spa_config_exit(spa, SCL_STATE_ALL, FTAG);
2736 	}
2737 }
2738 
2739 /*
2740  * Expand a vdev if possible.
2741  */
2742 void
2743 vdev_expand(vdev_t *vd, uint64_t txg)
2744 {
2745 	ASSERT(vd->vdev_top == vd);
2746 	ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
2747 
2748 	if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count) {
2749 		VERIFY(vdev_metaslab_init(vd, txg) == 0);
2750 		vdev_config_dirty(vd);
2751 	}
2752 }
2753