xref: /freebsd/sys/contrib/openzfs/module/zfs/vdev.c (revision c66ec88fed842fbaad62c30d510644ceb7bd2d71)
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, 2020 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  * Copyright (c) 2017, Intel Corporation.
30  * Copyright (c) 2019, Datto Inc. All rights reserved.
31  */
32 
33 #include <sys/zfs_context.h>
34 #include <sys/fm/fs/zfs.h>
35 #include <sys/spa.h>
36 #include <sys/spa_impl.h>
37 #include <sys/bpobj.h>
38 #include <sys/dmu.h>
39 #include <sys/dmu_tx.h>
40 #include <sys/dsl_dir.h>
41 #include <sys/vdev_impl.h>
42 #include <sys/vdev_rebuild.h>
43 #include <sys/vdev_draid.h>
44 #include <sys/uberblock_impl.h>
45 #include <sys/metaslab.h>
46 #include <sys/metaslab_impl.h>
47 #include <sys/space_map.h>
48 #include <sys/space_reftree.h>
49 #include <sys/zio.h>
50 #include <sys/zap.h>
51 #include <sys/fs/zfs.h>
52 #include <sys/arc.h>
53 #include <sys/zil.h>
54 #include <sys/dsl_scan.h>
55 #include <sys/vdev_raidz.h>
56 #include <sys/abd.h>
57 #include <sys/vdev_initialize.h>
58 #include <sys/vdev_trim.h>
59 #include <sys/zvol.h>
60 #include <sys/zfs_ratelimit.h>
61 
62 /* default target for number of metaslabs per top-level vdev */
63 int zfs_vdev_default_ms_count = 200;
64 
65 /* minimum number of metaslabs per top-level vdev */
66 int zfs_vdev_min_ms_count = 16;
67 
68 /* practical upper limit of total metaslabs per top-level vdev */
69 int zfs_vdev_ms_count_limit = 1ULL << 17;
70 
71 /* lower limit for metaslab size (512M) */
72 int zfs_vdev_default_ms_shift = 29;
73 
74 /* upper limit for metaslab size (16G) */
75 int zfs_vdev_max_ms_shift = 34;
76 
77 int vdev_validate_skip = B_FALSE;
78 
79 /*
80  * Since the DTL space map of a vdev is not expected to have a lot of
81  * entries, we default its block size to 4K.
82  */
83 int zfs_vdev_dtl_sm_blksz = (1 << 12);
84 
85 /*
86  * Rate limit slow IO (delay) events to this many per second.
87  */
88 unsigned int zfs_slow_io_events_per_second = 20;
89 
90 /*
91  * Rate limit checksum events after this many checksum errors per second.
92  */
93 unsigned int zfs_checksum_events_per_second = 20;
94 
95 /*
96  * Ignore errors during scrub/resilver.  Allows to work around resilver
97  * upon import when there are pool errors.
98  */
99 int zfs_scan_ignore_errors = 0;
100 
101 /*
102  * vdev-wide space maps that have lots of entries written to them at
103  * the end of each transaction can benefit from a higher I/O bandwidth
104  * (e.g. vdev_obsolete_sm), thus we default their block size to 128K.
105  */
106 int zfs_vdev_standard_sm_blksz = (1 << 17);
107 
108 /*
109  * Tunable parameter for debugging or performance analysis. Setting this
110  * will cause pool corruption on power loss if a volatile out-of-order
111  * write cache is enabled.
112  */
113 int zfs_nocacheflush = 0;
114 
115 uint64_t zfs_vdev_max_auto_ashift = ASHIFT_MAX;
116 uint64_t zfs_vdev_min_auto_ashift = ASHIFT_MIN;
117 
118 /*PRINTFLIKE2*/
119 void
120 vdev_dbgmsg(vdev_t *vd, const char *fmt, ...)
121 {
122 	va_list adx;
123 	char buf[256];
124 
125 	va_start(adx, fmt);
126 	(void) vsnprintf(buf, sizeof (buf), fmt, adx);
127 	va_end(adx);
128 
129 	if (vd->vdev_path != NULL) {
130 		zfs_dbgmsg("%s vdev '%s': %s", vd->vdev_ops->vdev_op_type,
131 		    vd->vdev_path, buf);
132 	} else {
133 		zfs_dbgmsg("%s-%llu vdev (guid %llu): %s",
134 		    vd->vdev_ops->vdev_op_type,
135 		    (u_longlong_t)vd->vdev_id,
136 		    (u_longlong_t)vd->vdev_guid, buf);
137 	}
138 }
139 
140 void
141 vdev_dbgmsg_print_tree(vdev_t *vd, int indent)
142 {
143 	char state[20];
144 
145 	if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops) {
146 		zfs_dbgmsg("%*svdev %u: %s", indent, "", vd->vdev_id,
147 		    vd->vdev_ops->vdev_op_type);
148 		return;
149 	}
150 
151 	switch (vd->vdev_state) {
152 	case VDEV_STATE_UNKNOWN:
153 		(void) snprintf(state, sizeof (state), "unknown");
154 		break;
155 	case VDEV_STATE_CLOSED:
156 		(void) snprintf(state, sizeof (state), "closed");
157 		break;
158 	case VDEV_STATE_OFFLINE:
159 		(void) snprintf(state, sizeof (state), "offline");
160 		break;
161 	case VDEV_STATE_REMOVED:
162 		(void) snprintf(state, sizeof (state), "removed");
163 		break;
164 	case VDEV_STATE_CANT_OPEN:
165 		(void) snprintf(state, sizeof (state), "can't open");
166 		break;
167 	case VDEV_STATE_FAULTED:
168 		(void) snprintf(state, sizeof (state), "faulted");
169 		break;
170 	case VDEV_STATE_DEGRADED:
171 		(void) snprintf(state, sizeof (state), "degraded");
172 		break;
173 	case VDEV_STATE_HEALTHY:
174 		(void) snprintf(state, sizeof (state), "healthy");
175 		break;
176 	default:
177 		(void) snprintf(state, sizeof (state), "<state %u>",
178 		    (uint_t)vd->vdev_state);
179 	}
180 
181 	zfs_dbgmsg("%*svdev %u: %s%s, guid: %llu, path: %s, %s", indent,
182 	    "", (int)vd->vdev_id, vd->vdev_ops->vdev_op_type,
183 	    vd->vdev_islog ? " (log)" : "",
184 	    (u_longlong_t)vd->vdev_guid,
185 	    vd->vdev_path ? vd->vdev_path : "N/A", state);
186 
187 	for (uint64_t i = 0; i < vd->vdev_children; i++)
188 		vdev_dbgmsg_print_tree(vd->vdev_child[i], indent + 2);
189 }
190 
191 /*
192  * Virtual device management.
193  */
194 
195 static vdev_ops_t *vdev_ops_table[] = {
196 	&vdev_root_ops,
197 	&vdev_raidz_ops,
198 	&vdev_draid_ops,
199 	&vdev_draid_spare_ops,
200 	&vdev_mirror_ops,
201 	&vdev_replacing_ops,
202 	&vdev_spare_ops,
203 	&vdev_disk_ops,
204 	&vdev_file_ops,
205 	&vdev_missing_ops,
206 	&vdev_hole_ops,
207 	&vdev_indirect_ops,
208 	NULL
209 };
210 
211 /*
212  * Given a vdev type, return the appropriate ops vector.
213  */
214 static vdev_ops_t *
215 vdev_getops(const char *type)
216 {
217 	vdev_ops_t *ops, **opspp;
218 
219 	for (opspp = vdev_ops_table; (ops = *opspp) != NULL; opspp++)
220 		if (strcmp(ops->vdev_op_type, type) == 0)
221 			break;
222 
223 	return (ops);
224 }
225 
226 /* ARGSUSED */
227 void
228 vdev_default_xlate(vdev_t *vd, const range_seg64_t *logical_rs,
229     range_seg64_t *physical_rs, range_seg64_t *remain_rs)
230 {
231 	physical_rs->rs_start = logical_rs->rs_start;
232 	physical_rs->rs_end = logical_rs->rs_end;
233 }
234 
235 /*
236  * Derive the enumerated allocation bias from string input.
237  * String origin is either the per-vdev zap or zpool(8).
238  */
239 static vdev_alloc_bias_t
240 vdev_derive_alloc_bias(const char *bias)
241 {
242 	vdev_alloc_bias_t alloc_bias = VDEV_BIAS_NONE;
243 
244 	if (strcmp(bias, VDEV_ALLOC_BIAS_LOG) == 0)
245 		alloc_bias = VDEV_BIAS_LOG;
246 	else if (strcmp(bias, VDEV_ALLOC_BIAS_SPECIAL) == 0)
247 		alloc_bias = VDEV_BIAS_SPECIAL;
248 	else if (strcmp(bias, VDEV_ALLOC_BIAS_DEDUP) == 0)
249 		alloc_bias = VDEV_BIAS_DEDUP;
250 
251 	return (alloc_bias);
252 }
253 
254 /*
255  * Default asize function: return the MAX of psize with the asize of
256  * all children.  This is what's used by anything other than RAID-Z.
257  */
258 uint64_t
259 vdev_default_asize(vdev_t *vd, uint64_t psize)
260 {
261 	uint64_t asize = P2ROUNDUP(psize, 1ULL << vd->vdev_top->vdev_ashift);
262 	uint64_t csize;
263 
264 	for (int c = 0; c < vd->vdev_children; c++) {
265 		csize = vdev_psize_to_asize(vd->vdev_child[c], psize);
266 		asize = MAX(asize, csize);
267 	}
268 
269 	return (asize);
270 }
271 
272 uint64_t
273 vdev_default_min_asize(vdev_t *vd)
274 {
275 	return (vd->vdev_min_asize);
276 }
277 
278 /*
279  * Get the minimum allocatable size. We define the allocatable size as
280  * the vdev's asize rounded to the nearest metaslab. This allows us to
281  * replace or attach devices which don't have the same physical size but
282  * can still satisfy the same number of allocations.
283  */
284 uint64_t
285 vdev_get_min_asize(vdev_t *vd)
286 {
287 	vdev_t *pvd = vd->vdev_parent;
288 
289 	/*
290 	 * If our parent is NULL (inactive spare or cache) or is the root,
291 	 * just return our own asize.
292 	 */
293 	if (pvd == NULL)
294 		return (vd->vdev_asize);
295 
296 	/*
297 	 * The top-level vdev just returns the allocatable size rounded
298 	 * to the nearest metaslab.
299 	 */
300 	if (vd == vd->vdev_top)
301 		return (P2ALIGN(vd->vdev_asize, 1ULL << vd->vdev_ms_shift));
302 
303 	return (pvd->vdev_ops->vdev_op_min_asize(pvd));
304 }
305 
306 void
307 vdev_set_min_asize(vdev_t *vd)
308 {
309 	vd->vdev_min_asize = vdev_get_min_asize(vd);
310 
311 	for (int c = 0; c < vd->vdev_children; c++)
312 		vdev_set_min_asize(vd->vdev_child[c]);
313 }
314 
315 /*
316  * Get the minimal allocation size for the top-level vdev.
317  */
318 uint64_t
319 vdev_get_min_alloc(vdev_t *vd)
320 {
321 	uint64_t min_alloc = 1ULL << vd->vdev_ashift;
322 
323 	if (vd->vdev_ops->vdev_op_min_alloc != NULL)
324 		min_alloc = vd->vdev_ops->vdev_op_min_alloc(vd);
325 
326 	return (min_alloc);
327 }
328 
329 /*
330  * Get the parity level for a top-level vdev.
331  */
332 uint64_t
333 vdev_get_nparity(vdev_t *vd)
334 {
335 	uint64_t nparity = 0;
336 
337 	if (vd->vdev_ops->vdev_op_nparity != NULL)
338 		nparity = vd->vdev_ops->vdev_op_nparity(vd);
339 
340 	return (nparity);
341 }
342 
343 /*
344  * Get the number of data disks for a top-level vdev.
345  */
346 uint64_t
347 vdev_get_ndisks(vdev_t *vd)
348 {
349 	uint64_t ndisks = 1;
350 
351 	if (vd->vdev_ops->vdev_op_ndisks != NULL)
352 		ndisks = vd->vdev_ops->vdev_op_ndisks(vd);
353 
354 	return (ndisks);
355 }
356 
357 vdev_t *
358 vdev_lookup_top(spa_t *spa, uint64_t vdev)
359 {
360 	vdev_t *rvd = spa->spa_root_vdev;
361 
362 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
363 
364 	if (vdev < rvd->vdev_children) {
365 		ASSERT(rvd->vdev_child[vdev] != NULL);
366 		return (rvd->vdev_child[vdev]);
367 	}
368 
369 	return (NULL);
370 }
371 
372 vdev_t *
373 vdev_lookup_by_guid(vdev_t *vd, uint64_t guid)
374 {
375 	vdev_t *mvd;
376 
377 	if (vd->vdev_guid == guid)
378 		return (vd);
379 
380 	for (int c = 0; c < vd->vdev_children; c++)
381 		if ((mvd = vdev_lookup_by_guid(vd->vdev_child[c], guid)) !=
382 		    NULL)
383 			return (mvd);
384 
385 	return (NULL);
386 }
387 
388 static int
389 vdev_count_leaves_impl(vdev_t *vd)
390 {
391 	int n = 0;
392 
393 	if (vd->vdev_ops->vdev_op_leaf)
394 		return (1);
395 
396 	for (int c = 0; c < vd->vdev_children; c++)
397 		n += vdev_count_leaves_impl(vd->vdev_child[c]);
398 
399 	return (n);
400 }
401 
402 int
403 vdev_count_leaves(spa_t *spa)
404 {
405 	int rc;
406 
407 	spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
408 	rc = vdev_count_leaves_impl(spa->spa_root_vdev);
409 	spa_config_exit(spa, SCL_VDEV, FTAG);
410 
411 	return (rc);
412 }
413 
414 void
415 vdev_add_child(vdev_t *pvd, vdev_t *cvd)
416 {
417 	size_t oldsize, newsize;
418 	uint64_t id = cvd->vdev_id;
419 	vdev_t **newchild;
420 
421 	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
422 	ASSERT(cvd->vdev_parent == NULL);
423 
424 	cvd->vdev_parent = pvd;
425 
426 	if (pvd == NULL)
427 		return;
428 
429 	ASSERT(id >= pvd->vdev_children || pvd->vdev_child[id] == NULL);
430 
431 	oldsize = pvd->vdev_children * sizeof (vdev_t *);
432 	pvd->vdev_children = MAX(pvd->vdev_children, id + 1);
433 	newsize = pvd->vdev_children * sizeof (vdev_t *);
434 
435 	newchild = kmem_alloc(newsize, KM_SLEEP);
436 	if (pvd->vdev_child != NULL) {
437 		bcopy(pvd->vdev_child, newchild, oldsize);
438 		kmem_free(pvd->vdev_child, oldsize);
439 	}
440 
441 	pvd->vdev_child = newchild;
442 	pvd->vdev_child[id] = cvd;
443 
444 	cvd->vdev_top = (pvd->vdev_top ? pvd->vdev_top: cvd);
445 	ASSERT(cvd->vdev_top->vdev_parent->vdev_parent == NULL);
446 
447 	/*
448 	 * Walk up all ancestors to update guid sum.
449 	 */
450 	for (; pvd != NULL; pvd = pvd->vdev_parent)
451 		pvd->vdev_guid_sum += cvd->vdev_guid_sum;
452 
453 	if (cvd->vdev_ops->vdev_op_leaf) {
454 		list_insert_head(&cvd->vdev_spa->spa_leaf_list, cvd);
455 		cvd->vdev_spa->spa_leaf_list_gen++;
456 	}
457 }
458 
459 void
460 vdev_remove_child(vdev_t *pvd, vdev_t *cvd)
461 {
462 	int c;
463 	uint_t id = cvd->vdev_id;
464 
465 	ASSERT(cvd->vdev_parent == pvd);
466 
467 	if (pvd == NULL)
468 		return;
469 
470 	ASSERT(id < pvd->vdev_children);
471 	ASSERT(pvd->vdev_child[id] == cvd);
472 
473 	pvd->vdev_child[id] = NULL;
474 	cvd->vdev_parent = NULL;
475 
476 	for (c = 0; c < pvd->vdev_children; c++)
477 		if (pvd->vdev_child[c])
478 			break;
479 
480 	if (c == pvd->vdev_children) {
481 		kmem_free(pvd->vdev_child, c * sizeof (vdev_t *));
482 		pvd->vdev_child = NULL;
483 		pvd->vdev_children = 0;
484 	}
485 
486 	if (cvd->vdev_ops->vdev_op_leaf) {
487 		spa_t *spa = cvd->vdev_spa;
488 		list_remove(&spa->spa_leaf_list, cvd);
489 		spa->spa_leaf_list_gen++;
490 	}
491 
492 	/*
493 	 * Walk up all ancestors to update guid sum.
494 	 */
495 	for (; pvd != NULL; pvd = pvd->vdev_parent)
496 		pvd->vdev_guid_sum -= cvd->vdev_guid_sum;
497 }
498 
499 /*
500  * Remove any holes in the child array.
501  */
502 void
503 vdev_compact_children(vdev_t *pvd)
504 {
505 	vdev_t **newchild, *cvd;
506 	int oldc = pvd->vdev_children;
507 	int newc;
508 
509 	ASSERT(spa_config_held(pvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
510 
511 	if (oldc == 0)
512 		return;
513 
514 	for (int c = newc = 0; c < oldc; c++)
515 		if (pvd->vdev_child[c])
516 			newc++;
517 
518 	if (newc > 0) {
519 		newchild = kmem_zalloc(newc * sizeof (vdev_t *), KM_SLEEP);
520 
521 		for (int c = newc = 0; c < oldc; c++) {
522 			if ((cvd = pvd->vdev_child[c]) != NULL) {
523 				newchild[newc] = cvd;
524 				cvd->vdev_id = newc++;
525 			}
526 		}
527 	} else {
528 		newchild = NULL;
529 	}
530 
531 	kmem_free(pvd->vdev_child, oldc * sizeof (vdev_t *));
532 	pvd->vdev_child = newchild;
533 	pvd->vdev_children = newc;
534 }
535 
536 /*
537  * Allocate and minimally initialize a vdev_t.
538  */
539 vdev_t *
540 vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
541 {
542 	vdev_t *vd;
543 	vdev_indirect_config_t *vic;
544 
545 	vd = kmem_zalloc(sizeof (vdev_t), KM_SLEEP);
546 	vic = &vd->vdev_indirect_config;
547 
548 	if (spa->spa_root_vdev == NULL) {
549 		ASSERT(ops == &vdev_root_ops);
550 		spa->spa_root_vdev = vd;
551 		spa->spa_load_guid = spa_generate_guid(NULL);
552 	}
553 
554 	if (guid == 0 && ops != &vdev_hole_ops) {
555 		if (spa->spa_root_vdev == vd) {
556 			/*
557 			 * The root vdev's guid will also be the pool guid,
558 			 * which must be unique among all pools.
559 			 */
560 			guid = spa_generate_guid(NULL);
561 		} else {
562 			/*
563 			 * Any other vdev's guid must be unique within the pool.
564 			 */
565 			guid = spa_generate_guid(spa);
566 		}
567 		ASSERT(!spa_guid_exists(spa_guid(spa), guid));
568 	}
569 
570 	vd->vdev_spa = spa;
571 	vd->vdev_id = id;
572 	vd->vdev_guid = guid;
573 	vd->vdev_guid_sum = guid;
574 	vd->vdev_ops = ops;
575 	vd->vdev_state = VDEV_STATE_CLOSED;
576 	vd->vdev_ishole = (ops == &vdev_hole_ops);
577 	vic->vic_prev_indirect_vdev = UINT64_MAX;
578 
579 	rw_init(&vd->vdev_indirect_rwlock, NULL, RW_DEFAULT, NULL);
580 	mutex_init(&vd->vdev_obsolete_lock, NULL, MUTEX_DEFAULT, NULL);
581 	vd->vdev_obsolete_segments = range_tree_create(NULL, RANGE_SEG64, NULL,
582 	    0, 0);
583 
584 	/*
585 	 * Initialize rate limit structs for events.  We rate limit ZIO delay
586 	 * and checksum events so that we don't overwhelm ZED with thousands
587 	 * of events when a disk is acting up.
588 	 */
589 	zfs_ratelimit_init(&vd->vdev_delay_rl, &zfs_slow_io_events_per_second,
590 	    1);
591 	zfs_ratelimit_init(&vd->vdev_checksum_rl,
592 	    &zfs_checksum_events_per_second, 1);
593 
594 	list_link_init(&vd->vdev_config_dirty_node);
595 	list_link_init(&vd->vdev_state_dirty_node);
596 	list_link_init(&vd->vdev_initialize_node);
597 	list_link_init(&vd->vdev_leaf_node);
598 	list_link_init(&vd->vdev_trim_node);
599 
600 	mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_NOLOCKDEP, NULL);
601 	mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL);
602 	mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL);
603 	mutex_init(&vd->vdev_scan_io_queue_lock, NULL, MUTEX_DEFAULT, NULL);
604 
605 	mutex_init(&vd->vdev_initialize_lock, NULL, MUTEX_DEFAULT, NULL);
606 	mutex_init(&vd->vdev_initialize_io_lock, NULL, MUTEX_DEFAULT, NULL);
607 	cv_init(&vd->vdev_initialize_cv, NULL, CV_DEFAULT, NULL);
608 	cv_init(&vd->vdev_initialize_io_cv, NULL, CV_DEFAULT, NULL);
609 
610 	mutex_init(&vd->vdev_trim_lock, NULL, MUTEX_DEFAULT, NULL);
611 	mutex_init(&vd->vdev_autotrim_lock, NULL, MUTEX_DEFAULT, NULL);
612 	mutex_init(&vd->vdev_trim_io_lock, NULL, MUTEX_DEFAULT, NULL);
613 	cv_init(&vd->vdev_trim_cv, NULL, CV_DEFAULT, NULL);
614 	cv_init(&vd->vdev_autotrim_cv, NULL, CV_DEFAULT, NULL);
615 	cv_init(&vd->vdev_trim_io_cv, NULL, CV_DEFAULT, NULL);
616 
617 	mutex_init(&vd->vdev_rebuild_lock, NULL, MUTEX_DEFAULT, NULL);
618 	cv_init(&vd->vdev_rebuild_cv, NULL, CV_DEFAULT, NULL);
619 
620 	for (int t = 0; t < DTL_TYPES; t++) {
621 		vd->vdev_dtl[t] = range_tree_create(NULL, RANGE_SEG64, NULL, 0,
622 		    0);
623 	}
624 
625 	txg_list_create(&vd->vdev_ms_list, spa,
626 	    offsetof(struct metaslab, ms_txg_node));
627 	txg_list_create(&vd->vdev_dtl_list, spa,
628 	    offsetof(struct vdev, vdev_dtl_node));
629 	vd->vdev_stat.vs_timestamp = gethrtime();
630 	vdev_queue_init(vd);
631 	vdev_cache_init(vd);
632 
633 	return (vd);
634 }
635 
636 /*
637  * Allocate a new vdev.  The 'alloctype' is used to control whether we are
638  * creating a new vdev or loading an existing one - the behavior is slightly
639  * different for each case.
640  */
641 int
642 vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id,
643     int alloctype)
644 {
645 	vdev_ops_t *ops;
646 	char *type;
647 	uint64_t guid = 0, islog;
648 	vdev_t *vd;
649 	vdev_indirect_config_t *vic;
650 	char *tmp = NULL;
651 	int rc;
652 	vdev_alloc_bias_t alloc_bias = VDEV_BIAS_NONE;
653 	boolean_t top_level = (parent && !parent->vdev_parent);
654 
655 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
656 
657 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) != 0)
658 		return (SET_ERROR(EINVAL));
659 
660 	if ((ops = vdev_getops(type)) == NULL)
661 		return (SET_ERROR(EINVAL));
662 
663 	/*
664 	 * If this is a load, get the vdev guid from the nvlist.
665 	 * Otherwise, vdev_alloc_common() will generate one for us.
666 	 */
667 	if (alloctype == VDEV_ALLOC_LOAD) {
668 		uint64_t label_id;
669 
670 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ID, &label_id) ||
671 		    label_id != id)
672 			return (SET_ERROR(EINVAL));
673 
674 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
675 			return (SET_ERROR(EINVAL));
676 	} else if (alloctype == VDEV_ALLOC_SPARE) {
677 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
678 			return (SET_ERROR(EINVAL));
679 	} else if (alloctype == VDEV_ALLOC_L2CACHE) {
680 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
681 			return (SET_ERROR(EINVAL));
682 	} else if (alloctype == VDEV_ALLOC_ROOTPOOL) {
683 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
684 			return (SET_ERROR(EINVAL));
685 	}
686 
687 	/*
688 	 * The first allocated vdev must be of type 'root'.
689 	 */
690 	if (ops != &vdev_root_ops && spa->spa_root_vdev == NULL)
691 		return (SET_ERROR(EINVAL));
692 
693 	/*
694 	 * Determine whether we're a log vdev.
695 	 */
696 	islog = 0;
697 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &islog);
698 	if (islog && spa_version(spa) < SPA_VERSION_SLOGS)
699 		return (SET_ERROR(ENOTSUP));
700 
701 	if (ops == &vdev_hole_ops && spa_version(spa) < SPA_VERSION_HOLES)
702 		return (SET_ERROR(ENOTSUP));
703 
704 	if (top_level && alloctype == VDEV_ALLOC_ADD) {
705 		char *bias;
706 
707 		/*
708 		 * If creating a top-level vdev, check for allocation
709 		 * classes input.
710 		 */
711 		if (nvlist_lookup_string(nv, ZPOOL_CONFIG_ALLOCATION_BIAS,
712 		    &bias) == 0) {
713 			alloc_bias = vdev_derive_alloc_bias(bias);
714 
715 			/* spa_vdev_add() expects feature to be enabled */
716 			if (spa->spa_load_state != SPA_LOAD_CREATE &&
717 			    !spa_feature_is_enabled(spa,
718 			    SPA_FEATURE_ALLOCATION_CLASSES)) {
719 				return (SET_ERROR(ENOTSUP));
720 			}
721 		}
722 
723 		/* spa_vdev_add() expects feature to be enabled */
724 		if (ops == &vdev_draid_ops &&
725 		    spa->spa_load_state != SPA_LOAD_CREATE &&
726 		    !spa_feature_is_enabled(spa, SPA_FEATURE_DRAID)) {
727 			return (SET_ERROR(ENOTSUP));
728 		}
729 	}
730 
731 	/*
732 	 * Initialize the vdev specific data.  This is done before calling
733 	 * vdev_alloc_common() since it may fail and this simplifies the
734 	 * error reporting and cleanup code paths.
735 	 */
736 	void *tsd = NULL;
737 	if (ops->vdev_op_init != NULL) {
738 		rc = ops->vdev_op_init(spa, nv, &tsd);
739 		if (rc != 0) {
740 			return (rc);
741 		}
742 	}
743 
744 	vd = vdev_alloc_common(spa, id, guid, ops);
745 	vd->vdev_tsd = tsd;
746 	vd->vdev_islog = islog;
747 
748 	if (top_level && alloc_bias != VDEV_BIAS_NONE)
749 		vd->vdev_alloc_bias = alloc_bias;
750 
751 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &vd->vdev_path) == 0)
752 		vd->vdev_path = spa_strdup(vd->vdev_path);
753 
754 	/*
755 	 * ZPOOL_CONFIG_AUX_STATE = "external" means we previously forced a
756 	 * fault on a vdev and want it to persist across imports (like with
757 	 * zpool offline -f).
758 	 */
759 	rc = nvlist_lookup_string(nv, ZPOOL_CONFIG_AUX_STATE, &tmp);
760 	if (rc == 0 && tmp != NULL && strcmp(tmp, "external") == 0) {
761 		vd->vdev_stat.vs_aux = VDEV_AUX_EXTERNAL;
762 		vd->vdev_faulted = 1;
763 		vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
764 	}
765 
766 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &vd->vdev_devid) == 0)
767 		vd->vdev_devid = spa_strdup(vd->vdev_devid);
768 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH,
769 	    &vd->vdev_physpath) == 0)
770 		vd->vdev_physpath = spa_strdup(vd->vdev_physpath);
771 
772 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_VDEV_ENC_SYSFS_PATH,
773 	    &vd->vdev_enc_sysfs_path) == 0)
774 		vd->vdev_enc_sysfs_path = spa_strdup(vd->vdev_enc_sysfs_path);
775 
776 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &vd->vdev_fru) == 0)
777 		vd->vdev_fru = spa_strdup(vd->vdev_fru);
778 
779 	/*
780 	 * Set the whole_disk property.  If it's not specified, leave the value
781 	 * as -1.
782 	 */
783 	if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
784 	    &vd->vdev_wholedisk) != 0)
785 		vd->vdev_wholedisk = -1ULL;
786 
787 	vic = &vd->vdev_indirect_config;
788 
789 	ASSERT0(vic->vic_mapping_object);
790 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_INDIRECT_OBJECT,
791 	    &vic->vic_mapping_object);
792 	ASSERT0(vic->vic_births_object);
793 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_INDIRECT_BIRTHS,
794 	    &vic->vic_births_object);
795 	ASSERT3U(vic->vic_prev_indirect_vdev, ==, UINT64_MAX);
796 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_PREV_INDIRECT_VDEV,
797 	    &vic->vic_prev_indirect_vdev);
798 
799 	/*
800 	 * Look for the 'not present' flag.  This will only be set if the device
801 	 * was not present at the time of import.
802 	 */
803 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT,
804 	    &vd->vdev_not_present);
805 
806 	/*
807 	 * Get the alignment requirement.
808 	 */
809 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT, &vd->vdev_ashift);
810 
811 	/*
812 	 * Retrieve the vdev creation time.
813 	 */
814 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_CREATE_TXG,
815 	    &vd->vdev_crtxg);
816 
817 	/*
818 	 * If we're a top-level vdev, try to load the allocation parameters.
819 	 */
820 	if (top_level &&
821 	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
822 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
823 		    &vd->vdev_ms_array);
824 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
825 		    &vd->vdev_ms_shift);
826 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE,
827 		    &vd->vdev_asize);
828 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVING,
829 		    &vd->vdev_removing);
830 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_VDEV_TOP_ZAP,
831 		    &vd->vdev_top_zap);
832 	} else {
833 		ASSERT0(vd->vdev_top_zap);
834 	}
835 
836 	if (top_level && alloctype != VDEV_ALLOC_ATTACH) {
837 		ASSERT(alloctype == VDEV_ALLOC_LOAD ||
838 		    alloctype == VDEV_ALLOC_ADD ||
839 		    alloctype == VDEV_ALLOC_SPLIT ||
840 		    alloctype == VDEV_ALLOC_ROOTPOOL);
841 		/* Note: metaslab_group_create() is now deferred */
842 	}
843 
844 	if (vd->vdev_ops->vdev_op_leaf &&
845 	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
846 		(void) nvlist_lookup_uint64(nv,
847 		    ZPOOL_CONFIG_VDEV_LEAF_ZAP, &vd->vdev_leaf_zap);
848 	} else {
849 		ASSERT0(vd->vdev_leaf_zap);
850 	}
851 
852 	/*
853 	 * If we're a leaf vdev, try to load the DTL object and other state.
854 	 */
855 
856 	if (vd->vdev_ops->vdev_op_leaf &&
857 	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE ||
858 	    alloctype == VDEV_ALLOC_ROOTPOOL)) {
859 		if (alloctype == VDEV_ALLOC_LOAD) {
860 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL,
861 			    &vd->vdev_dtl_object);
862 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE,
863 			    &vd->vdev_unspare);
864 		}
865 
866 		if (alloctype == VDEV_ALLOC_ROOTPOOL) {
867 			uint64_t spare = 0;
868 
869 			if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE,
870 			    &spare) == 0 && spare)
871 				spa_spare_add(vd);
872 		}
873 
874 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE,
875 		    &vd->vdev_offline);
876 
877 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_RESILVER_TXG,
878 		    &vd->vdev_resilver_txg);
879 
880 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REBUILD_TXG,
881 		    &vd->vdev_rebuild_txg);
882 
883 		if (nvlist_exists(nv, ZPOOL_CONFIG_RESILVER_DEFER))
884 			vdev_defer_resilver(vd);
885 
886 		/*
887 		 * In general, when importing a pool we want to ignore the
888 		 * persistent fault state, as the diagnosis made on another
889 		 * system may not be valid in the current context.  The only
890 		 * exception is if we forced a vdev to a persistently faulted
891 		 * state with 'zpool offline -f'.  The persistent fault will
892 		 * remain across imports until cleared.
893 		 *
894 		 * Local vdevs will remain in the faulted state.
895 		 */
896 		if (spa_load_state(spa) == SPA_LOAD_OPEN ||
897 		    spa_load_state(spa) == SPA_LOAD_IMPORT) {
898 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED,
899 			    &vd->vdev_faulted);
900 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED,
901 			    &vd->vdev_degraded);
902 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED,
903 			    &vd->vdev_removed);
904 
905 			if (vd->vdev_faulted || vd->vdev_degraded) {
906 				char *aux;
907 
908 				vd->vdev_label_aux =
909 				    VDEV_AUX_ERR_EXCEEDED;
910 				if (nvlist_lookup_string(nv,
911 				    ZPOOL_CONFIG_AUX_STATE, &aux) == 0 &&
912 				    strcmp(aux, "external") == 0)
913 					vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
914 				else
915 					vd->vdev_faulted = 0ULL;
916 			}
917 		}
918 	}
919 
920 	/*
921 	 * Add ourselves to the parent's list of children.
922 	 */
923 	vdev_add_child(parent, vd);
924 
925 	*vdp = vd;
926 
927 	return (0);
928 }
929 
930 void
931 vdev_free(vdev_t *vd)
932 {
933 	spa_t *spa = vd->vdev_spa;
934 
935 	ASSERT3P(vd->vdev_initialize_thread, ==, NULL);
936 	ASSERT3P(vd->vdev_trim_thread, ==, NULL);
937 	ASSERT3P(vd->vdev_autotrim_thread, ==, NULL);
938 	ASSERT3P(vd->vdev_rebuild_thread, ==, NULL);
939 
940 	/*
941 	 * Scan queues are normally destroyed at the end of a scan. If the
942 	 * queue exists here, that implies the vdev is being removed while
943 	 * the scan is still running.
944 	 */
945 	if (vd->vdev_scan_io_queue != NULL) {
946 		mutex_enter(&vd->vdev_scan_io_queue_lock);
947 		dsl_scan_io_queue_destroy(vd->vdev_scan_io_queue);
948 		vd->vdev_scan_io_queue = NULL;
949 		mutex_exit(&vd->vdev_scan_io_queue_lock);
950 	}
951 
952 	/*
953 	 * vdev_free() implies closing the vdev first.  This is simpler than
954 	 * trying to ensure complicated semantics for all callers.
955 	 */
956 	vdev_close(vd);
957 
958 	ASSERT(!list_link_active(&vd->vdev_config_dirty_node));
959 	ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
960 
961 	/*
962 	 * Free all children.
963 	 */
964 	for (int c = 0; c < vd->vdev_children; c++)
965 		vdev_free(vd->vdev_child[c]);
966 
967 	ASSERT(vd->vdev_child == NULL);
968 	ASSERT(vd->vdev_guid_sum == vd->vdev_guid);
969 
970 	if (vd->vdev_ops->vdev_op_fini != NULL)
971 		vd->vdev_ops->vdev_op_fini(vd);
972 
973 	/*
974 	 * Discard allocation state.
975 	 */
976 	if (vd->vdev_mg != NULL) {
977 		vdev_metaslab_fini(vd);
978 		metaslab_group_destroy(vd->vdev_mg);
979 		vd->vdev_mg = NULL;
980 	}
981 
982 	ASSERT0(vd->vdev_stat.vs_space);
983 	ASSERT0(vd->vdev_stat.vs_dspace);
984 	ASSERT0(vd->vdev_stat.vs_alloc);
985 
986 	/*
987 	 * Remove this vdev from its parent's child list.
988 	 */
989 	vdev_remove_child(vd->vdev_parent, vd);
990 
991 	ASSERT(vd->vdev_parent == NULL);
992 	ASSERT(!list_link_active(&vd->vdev_leaf_node));
993 
994 	/*
995 	 * Clean up vdev structure.
996 	 */
997 	vdev_queue_fini(vd);
998 	vdev_cache_fini(vd);
999 
1000 	if (vd->vdev_path)
1001 		spa_strfree(vd->vdev_path);
1002 	if (vd->vdev_devid)
1003 		spa_strfree(vd->vdev_devid);
1004 	if (vd->vdev_physpath)
1005 		spa_strfree(vd->vdev_physpath);
1006 
1007 	if (vd->vdev_enc_sysfs_path)
1008 		spa_strfree(vd->vdev_enc_sysfs_path);
1009 
1010 	if (vd->vdev_fru)
1011 		spa_strfree(vd->vdev_fru);
1012 
1013 	if (vd->vdev_isspare)
1014 		spa_spare_remove(vd);
1015 	if (vd->vdev_isl2cache)
1016 		spa_l2cache_remove(vd);
1017 
1018 	txg_list_destroy(&vd->vdev_ms_list);
1019 	txg_list_destroy(&vd->vdev_dtl_list);
1020 
1021 	mutex_enter(&vd->vdev_dtl_lock);
1022 	space_map_close(vd->vdev_dtl_sm);
1023 	for (int t = 0; t < DTL_TYPES; t++) {
1024 		range_tree_vacate(vd->vdev_dtl[t], NULL, NULL);
1025 		range_tree_destroy(vd->vdev_dtl[t]);
1026 	}
1027 	mutex_exit(&vd->vdev_dtl_lock);
1028 
1029 	EQUIV(vd->vdev_indirect_births != NULL,
1030 	    vd->vdev_indirect_mapping != NULL);
1031 	if (vd->vdev_indirect_births != NULL) {
1032 		vdev_indirect_mapping_close(vd->vdev_indirect_mapping);
1033 		vdev_indirect_births_close(vd->vdev_indirect_births);
1034 	}
1035 
1036 	if (vd->vdev_obsolete_sm != NULL) {
1037 		ASSERT(vd->vdev_removing ||
1038 		    vd->vdev_ops == &vdev_indirect_ops);
1039 		space_map_close(vd->vdev_obsolete_sm);
1040 		vd->vdev_obsolete_sm = NULL;
1041 	}
1042 	range_tree_destroy(vd->vdev_obsolete_segments);
1043 	rw_destroy(&vd->vdev_indirect_rwlock);
1044 	mutex_destroy(&vd->vdev_obsolete_lock);
1045 
1046 	mutex_destroy(&vd->vdev_dtl_lock);
1047 	mutex_destroy(&vd->vdev_stat_lock);
1048 	mutex_destroy(&vd->vdev_probe_lock);
1049 	mutex_destroy(&vd->vdev_scan_io_queue_lock);
1050 
1051 	mutex_destroy(&vd->vdev_initialize_lock);
1052 	mutex_destroy(&vd->vdev_initialize_io_lock);
1053 	cv_destroy(&vd->vdev_initialize_io_cv);
1054 	cv_destroy(&vd->vdev_initialize_cv);
1055 
1056 	mutex_destroy(&vd->vdev_trim_lock);
1057 	mutex_destroy(&vd->vdev_autotrim_lock);
1058 	mutex_destroy(&vd->vdev_trim_io_lock);
1059 	cv_destroy(&vd->vdev_trim_cv);
1060 	cv_destroy(&vd->vdev_autotrim_cv);
1061 	cv_destroy(&vd->vdev_trim_io_cv);
1062 
1063 	mutex_destroy(&vd->vdev_rebuild_lock);
1064 	cv_destroy(&vd->vdev_rebuild_cv);
1065 
1066 	zfs_ratelimit_fini(&vd->vdev_delay_rl);
1067 	zfs_ratelimit_fini(&vd->vdev_checksum_rl);
1068 
1069 	if (vd == spa->spa_root_vdev)
1070 		spa->spa_root_vdev = NULL;
1071 
1072 	kmem_free(vd, sizeof (vdev_t));
1073 }
1074 
1075 /*
1076  * Transfer top-level vdev state from svd to tvd.
1077  */
1078 static void
1079 vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
1080 {
1081 	spa_t *spa = svd->vdev_spa;
1082 	metaslab_t *msp;
1083 	vdev_t *vd;
1084 	int t;
1085 
1086 	ASSERT(tvd == tvd->vdev_top);
1087 
1088 	tvd->vdev_pending_fastwrite = svd->vdev_pending_fastwrite;
1089 	tvd->vdev_ms_array = svd->vdev_ms_array;
1090 	tvd->vdev_ms_shift = svd->vdev_ms_shift;
1091 	tvd->vdev_ms_count = svd->vdev_ms_count;
1092 	tvd->vdev_top_zap = svd->vdev_top_zap;
1093 
1094 	svd->vdev_ms_array = 0;
1095 	svd->vdev_ms_shift = 0;
1096 	svd->vdev_ms_count = 0;
1097 	svd->vdev_top_zap = 0;
1098 
1099 	if (tvd->vdev_mg)
1100 		ASSERT3P(tvd->vdev_mg, ==, svd->vdev_mg);
1101 	tvd->vdev_mg = svd->vdev_mg;
1102 	tvd->vdev_ms = svd->vdev_ms;
1103 
1104 	svd->vdev_mg = NULL;
1105 	svd->vdev_ms = NULL;
1106 
1107 	if (tvd->vdev_mg != NULL)
1108 		tvd->vdev_mg->mg_vd = tvd;
1109 
1110 	tvd->vdev_checkpoint_sm = svd->vdev_checkpoint_sm;
1111 	svd->vdev_checkpoint_sm = NULL;
1112 
1113 	tvd->vdev_alloc_bias = svd->vdev_alloc_bias;
1114 	svd->vdev_alloc_bias = VDEV_BIAS_NONE;
1115 
1116 	tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
1117 	tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
1118 	tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
1119 
1120 	svd->vdev_stat.vs_alloc = 0;
1121 	svd->vdev_stat.vs_space = 0;
1122 	svd->vdev_stat.vs_dspace = 0;
1123 
1124 	/*
1125 	 * State which may be set on a top-level vdev that's in the
1126 	 * process of being removed.
1127 	 */
1128 	ASSERT0(tvd->vdev_indirect_config.vic_births_object);
1129 	ASSERT0(tvd->vdev_indirect_config.vic_mapping_object);
1130 	ASSERT3U(tvd->vdev_indirect_config.vic_prev_indirect_vdev, ==, -1ULL);
1131 	ASSERT3P(tvd->vdev_indirect_mapping, ==, NULL);
1132 	ASSERT3P(tvd->vdev_indirect_births, ==, NULL);
1133 	ASSERT3P(tvd->vdev_obsolete_sm, ==, NULL);
1134 	ASSERT0(tvd->vdev_removing);
1135 	ASSERT0(tvd->vdev_rebuilding);
1136 	tvd->vdev_removing = svd->vdev_removing;
1137 	tvd->vdev_rebuilding = svd->vdev_rebuilding;
1138 	tvd->vdev_rebuild_config = svd->vdev_rebuild_config;
1139 	tvd->vdev_indirect_config = svd->vdev_indirect_config;
1140 	tvd->vdev_indirect_mapping = svd->vdev_indirect_mapping;
1141 	tvd->vdev_indirect_births = svd->vdev_indirect_births;
1142 	range_tree_swap(&svd->vdev_obsolete_segments,
1143 	    &tvd->vdev_obsolete_segments);
1144 	tvd->vdev_obsolete_sm = svd->vdev_obsolete_sm;
1145 	svd->vdev_indirect_config.vic_mapping_object = 0;
1146 	svd->vdev_indirect_config.vic_births_object = 0;
1147 	svd->vdev_indirect_config.vic_prev_indirect_vdev = -1ULL;
1148 	svd->vdev_indirect_mapping = NULL;
1149 	svd->vdev_indirect_births = NULL;
1150 	svd->vdev_obsolete_sm = NULL;
1151 	svd->vdev_removing = 0;
1152 	svd->vdev_rebuilding = 0;
1153 
1154 	for (t = 0; t < TXG_SIZE; t++) {
1155 		while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
1156 			(void) txg_list_add(&tvd->vdev_ms_list, msp, t);
1157 		while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
1158 			(void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
1159 		if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
1160 			(void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
1161 	}
1162 
1163 	if (list_link_active(&svd->vdev_config_dirty_node)) {
1164 		vdev_config_clean(svd);
1165 		vdev_config_dirty(tvd);
1166 	}
1167 
1168 	if (list_link_active(&svd->vdev_state_dirty_node)) {
1169 		vdev_state_clean(svd);
1170 		vdev_state_dirty(tvd);
1171 	}
1172 
1173 	tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
1174 	svd->vdev_deflate_ratio = 0;
1175 
1176 	tvd->vdev_islog = svd->vdev_islog;
1177 	svd->vdev_islog = 0;
1178 
1179 	dsl_scan_io_queue_vdev_xfer(svd, tvd);
1180 }
1181 
1182 static void
1183 vdev_top_update(vdev_t *tvd, vdev_t *vd)
1184 {
1185 	if (vd == NULL)
1186 		return;
1187 
1188 	vd->vdev_top = tvd;
1189 
1190 	for (int c = 0; c < vd->vdev_children; c++)
1191 		vdev_top_update(tvd, vd->vdev_child[c]);
1192 }
1193 
1194 /*
1195  * Add a mirror/replacing vdev above an existing vdev.  There is no need to
1196  * call .vdev_op_init() since mirror/replacing vdevs do not have private state.
1197  */
1198 vdev_t *
1199 vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
1200 {
1201 	spa_t *spa = cvd->vdev_spa;
1202 	vdev_t *pvd = cvd->vdev_parent;
1203 	vdev_t *mvd;
1204 
1205 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1206 
1207 	mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
1208 
1209 	mvd->vdev_asize = cvd->vdev_asize;
1210 	mvd->vdev_min_asize = cvd->vdev_min_asize;
1211 	mvd->vdev_max_asize = cvd->vdev_max_asize;
1212 	mvd->vdev_psize = cvd->vdev_psize;
1213 	mvd->vdev_ashift = cvd->vdev_ashift;
1214 	mvd->vdev_logical_ashift = cvd->vdev_logical_ashift;
1215 	mvd->vdev_physical_ashift = cvd->vdev_physical_ashift;
1216 	mvd->vdev_state = cvd->vdev_state;
1217 	mvd->vdev_crtxg = cvd->vdev_crtxg;
1218 
1219 	vdev_remove_child(pvd, cvd);
1220 	vdev_add_child(pvd, mvd);
1221 	cvd->vdev_id = mvd->vdev_children;
1222 	vdev_add_child(mvd, cvd);
1223 	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
1224 
1225 	if (mvd == mvd->vdev_top)
1226 		vdev_top_transfer(cvd, mvd);
1227 
1228 	return (mvd);
1229 }
1230 
1231 /*
1232  * Remove a 1-way mirror/replacing vdev from the tree.
1233  */
1234 void
1235 vdev_remove_parent(vdev_t *cvd)
1236 {
1237 	vdev_t *mvd = cvd->vdev_parent;
1238 	vdev_t *pvd = mvd->vdev_parent;
1239 
1240 	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1241 
1242 	ASSERT(mvd->vdev_children == 1);
1243 	ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
1244 	    mvd->vdev_ops == &vdev_replacing_ops ||
1245 	    mvd->vdev_ops == &vdev_spare_ops);
1246 	cvd->vdev_ashift = mvd->vdev_ashift;
1247 	cvd->vdev_logical_ashift = mvd->vdev_logical_ashift;
1248 	cvd->vdev_physical_ashift = mvd->vdev_physical_ashift;
1249 	vdev_remove_child(mvd, cvd);
1250 	vdev_remove_child(pvd, mvd);
1251 
1252 	/*
1253 	 * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
1254 	 * Otherwise, we could have detached an offline device, and when we
1255 	 * go to import the pool we'll think we have two top-level vdevs,
1256 	 * instead of a different version of the same top-level vdev.
1257 	 */
1258 	if (mvd->vdev_top == mvd) {
1259 		uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
1260 		cvd->vdev_orig_guid = cvd->vdev_guid;
1261 		cvd->vdev_guid += guid_delta;
1262 		cvd->vdev_guid_sum += guid_delta;
1263 
1264 		/*
1265 		 * If pool not set for autoexpand, we need to also preserve
1266 		 * mvd's asize to prevent automatic expansion of cvd.
1267 		 * Otherwise if we are adjusting the mirror by attaching and
1268 		 * detaching children of non-uniform sizes, the mirror could
1269 		 * autoexpand, unexpectedly requiring larger devices to
1270 		 * re-establish the mirror.
1271 		 */
1272 		if (!cvd->vdev_spa->spa_autoexpand)
1273 			cvd->vdev_asize = mvd->vdev_asize;
1274 	}
1275 	cvd->vdev_id = mvd->vdev_id;
1276 	vdev_add_child(pvd, cvd);
1277 	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
1278 
1279 	if (cvd == cvd->vdev_top)
1280 		vdev_top_transfer(mvd, cvd);
1281 
1282 	ASSERT(mvd->vdev_children == 0);
1283 	vdev_free(mvd);
1284 }
1285 
1286 static void
1287 vdev_metaslab_group_create(vdev_t *vd)
1288 {
1289 	spa_t *spa = vd->vdev_spa;
1290 
1291 	/*
1292 	 * metaslab_group_create was delayed until allocation bias was available
1293 	 */
1294 	if (vd->vdev_mg == NULL) {
1295 		metaslab_class_t *mc;
1296 
1297 		if (vd->vdev_islog && vd->vdev_alloc_bias == VDEV_BIAS_NONE)
1298 			vd->vdev_alloc_bias = VDEV_BIAS_LOG;
1299 
1300 		ASSERT3U(vd->vdev_islog, ==,
1301 		    (vd->vdev_alloc_bias == VDEV_BIAS_LOG));
1302 
1303 		switch (vd->vdev_alloc_bias) {
1304 		case VDEV_BIAS_LOG:
1305 			mc = spa_log_class(spa);
1306 			break;
1307 		case VDEV_BIAS_SPECIAL:
1308 			mc = spa_special_class(spa);
1309 			break;
1310 		case VDEV_BIAS_DEDUP:
1311 			mc = spa_dedup_class(spa);
1312 			break;
1313 		default:
1314 			mc = spa_normal_class(spa);
1315 		}
1316 
1317 		vd->vdev_mg = metaslab_group_create(mc, vd,
1318 		    spa->spa_alloc_count);
1319 
1320 		/*
1321 		 * The spa ashift min/max only apply for the normal metaslab
1322 		 * class. Class destination is late binding so ashift boundry
1323 		 * setting had to wait until now.
1324 		 */
1325 		if (vd->vdev_top == vd && vd->vdev_ashift != 0 &&
1326 		    mc == spa_normal_class(spa) && vd->vdev_aux == NULL) {
1327 			if (vd->vdev_ashift > spa->spa_max_ashift)
1328 				spa->spa_max_ashift = vd->vdev_ashift;
1329 			if (vd->vdev_ashift < spa->spa_min_ashift)
1330 				spa->spa_min_ashift = vd->vdev_ashift;
1331 
1332 			uint64_t min_alloc = vdev_get_min_alloc(vd);
1333 			if (min_alloc < spa->spa_min_alloc)
1334 				spa->spa_min_alloc = min_alloc;
1335 		}
1336 	}
1337 }
1338 
1339 int
1340 vdev_metaslab_init(vdev_t *vd, uint64_t txg)
1341 {
1342 	spa_t *spa = vd->vdev_spa;
1343 	objset_t *mos = spa->spa_meta_objset;
1344 	uint64_t m;
1345 	uint64_t oldc = vd->vdev_ms_count;
1346 	uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
1347 	metaslab_t **mspp;
1348 	int error;
1349 	boolean_t expanding = (oldc != 0);
1350 
1351 	ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER));
1352 
1353 	/*
1354 	 * This vdev is not being allocated from yet or is a hole.
1355 	 */
1356 	if (vd->vdev_ms_shift == 0)
1357 		return (0);
1358 
1359 	ASSERT(!vd->vdev_ishole);
1360 
1361 	ASSERT(oldc <= newc);
1362 
1363 	mspp = vmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
1364 
1365 	if (expanding) {
1366 		bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp));
1367 		vmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
1368 	}
1369 
1370 	vd->vdev_ms = mspp;
1371 	vd->vdev_ms_count = newc;
1372 	for (m = oldc; m < newc; m++) {
1373 		uint64_t object = 0;
1374 
1375 		/*
1376 		 * vdev_ms_array may be 0 if we are creating the "fake"
1377 		 * metaslabs for an indirect vdev for zdb's leak detection.
1378 		 * See zdb_leak_init().
1379 		 */
1380 		if (txg == 0 && vd->vdev_ms_array != 0) {
1381 			error = dmu_read(mos, vd->vdev_ms_array,
1382 			    m * sizeof (uint64_t), sizeof (uint64_t), &object,
1383 			    DMU_READ_PREFETCH);
1384 			if (error != 0) {
1385 				vdev_dbgmsg(vd, "unable to read the metaslab "
1386 				    "array [error=%d]", error);
1387 				return (error);
1388 			}
1389 		}
1390 
1391 #ifndef _KERNEL
1392 		/*
1393 		 * To accommodate zdb_leak_init() fake indirect
1394 		 * metaslabs, we allocate a metaslab group for
1395 		 * indirect vdevs which normally don't have one.
1396 		 */
1397 		if (vd->vdev_mg == NULL) {
1398 			ASSERT0(vdev_is_concrete(vd));
1399 			vdev_metaslab_group_create(vd);
1400 		}
1401 #endif
1402 		error = metaslab_init(vd->vdev_mg, m, object, txg,
1403 		    &(vd->vdev_ms[m]));
1404 		if (error != 0) {
1405 			vdev_dbgmsg(vd, "metaslab_init failed [error=%d]",
1406 			    error);
1407 			return (error);
1408 		}
1409 	}
1410 
1411 	if (txg == 0)
1412 		spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER);
1413 
1414 	/*
1415 	 * If the vdev is being removed we don't activate
1416 	 * the metaslabs since we want to ensure that no new
1417 	 * allocations are performed on this device.
1418 	 */
1419 	if (!expanding && !vd->vdev_removing) {
1420 		metaslab_group_activate(vd->vdev_mg);
1421 	}
1422 
1423 	if (txg == 0)
1424 		spa_config_exit(spa, SCL_ALLOC, FTAG);
1425 
1426 	/*
1427 	 * Regardless whether this vdev was just added or it is being
1428 	 * expanded, the metaslab count has changed. Recalculate the
1429 	 * block limit.
1430 	 */
1431 	spa_log_sm_set_blocklimit(spa);
1432 
1433 	return (0);
1434 }
1435 
1436 void
1437 vdev_metaslab_fini(vdev_t *vd)
1438 {
1439 	if (vd->vdev_checkpoint_sm != NULL) {
1440 		ASSERT(spa_feature_is_active(vd->vdev_spa,
1441 		    SPA_FEATURE_POOL_CHECKPOINT));
1442 		space_map_close(vd->vdev_checkpoint_sm);
1443 		/*
1444 		 * Even though we close the space map, we need to set its
1445 		 * pointer to NULL. The reason is that vdev_metaslab_fini()
1446 		 * may be called multiple times for certain operations
1447 		 * (i.e. when destroying a pool) so we need to ensure that
1448 		 * this clause never executes twice. This logic is similar
1449 		 * to the one used for the vdev_ms clause below.
1450 		 */
1451 		vd->vdev_checkpoint_sm = NULL;
1452 	}
1453 
1454 	if (vd->vdev_ms != NULL) {
1455 		metaslab_group_t *mg = vd->vdev_mg;
1456 		metaslab_group_passivate(mg);
1457 
1458 		uint64_t count = vd->vdev_ms_count;
1459 		for (uint64_t m = 0; m < count; m++) {
1460 			metaslab_t *msp = vd->vdev_ms[m];
1461 			if (msp != NULL)
1462 				metaslab_fini(msp);
1463 		}
1464 		vmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
1465 		vd->vdev_ms = NULL;
1466 
1467 		vd->vdev_ms_count = 0;
1468 
1469 		for (int i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
1470 			ASSERT0(mg->mg_histogram[i]);
1471 	}
1472 	ASSERT0(vd->vdev_ms_count);
1473 	ASSERT3U(vd->vdev_pending_fastwrite, ==, 0);
1474 }
1475 
1476 typedef struct vdev_probe_stats {
1477 	boolean_t	vps_readable;
1478 	boolean_t	vps_writeable;
1479 	int		vps_flags;
1480 } vdev_probe_stats_t;
1481 
1482 static void
1483 vdev_probe_done(zio_t *zio)
1484 {
1485 	spa_t *spa = zio->io_spa;
1486 	vdev_t *vd = zio->io_vd;
1487 	vdev_probe_stats_t *vps = zio->io_private;
1488 
1489 	ASSERT(vd->vdev_probe_zio != NULL);
1490 
1491 	if (zio->io_type == ZIO_TYPE_READ) {
1492 		if (zio->io_error == 0)
1493 			vps->vps_readable = 1;
1494 		if (zio->io_error == 0 && spa_writeable(spa)) {
1495 			zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
1496 			    zio->io_offset, zio->io_size, zio->io_abd,
1497 			    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1498 			    ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
1499 		} else {
1500 			abd_free(zio->io_abd);
1501 		}
1502 	} else if (zio->io_type == ZIO_TYPE_WRITE) {
1503 		if (zio->io_error == 0)
1504 			vps->vps_writeable = 1;
1505 		abd_free(zio->io_abd);
1506 	} else if (zio->io_type == ZIO_TYPE_NULL) {
1507 		zio_t *pio;
1508 		zio_link_t *zl;
1509 
1510 		vd->vdev_cant_read |= !vps->vps_readable;
1511 		vd->vdev_cant_write |= !vps->vps_writeable;
1512 
1513 		if (vdev_readable(vd) &&
1514 		    (vdev_writeable(vd) || !spa_writeable(spa))) {
1515 			zio->io_error = 0;
1516 		} else {
1517 			ASSERT(zio->io_error != 0);
1518 			vdev_dbgmsg(vd, "failed probe");
1519 			(void) zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
1520 			    spa, vd, NULL, NULL, 0);
1521 			zio->io_error = SET_ERROR(ENXIO);
1522 		}
1523 
1524 		mutex_enter(&vd->vdev_probe_lock);
1525 		ASSERT(vd->vdev_probe_zio == zio);
1526 		vd->vdev_probe_zio = NULL;
1527 		mutex_exit(&vd->vdev_probe_lock);
1528 
1529 		zl = NULL;
1530 		while ((pio = zio_walk_parents(zio, &zl)) != NULL)
1531 			if (!vdev_accessible(vd, pio))
1532 				pio->io_error = SET_ERROR(ENXIO);
1533 
1534 		kmem_free(vps, sizeof (*vps));
1535 	}
1536 }
1537 
1538 /*
1539  * Determine whether this device is accessible.
1540  *
1541  * Read and write to several known locations: the pad regions of each
1542  * vdev label but the first, which we leave alone in case it contains
1543  * a VTOC.
1544  */
1545 zio_t *
1546 vdev_probe(vdev_t *vd, zio_t *zio)
1547 {
1548 	spa_t *spa = vd->vdev_spa;
1549 	vdev_probe_stats_t *vps = NULL;
1550 	zio_t *pio;
1551 
1552 	ASSERT(vd->vdev_ops->vdev_op_leaf);
1553 
1554 	/*
1555 	 * Don't probe the probe.
1556 	 */
1557 	if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
1558 		return (NULL);
1559 
1560 	/*
1561 	 * To prevent 'probe storms' when a device fails, we create
1562 	 * just one probe i/o at a time.  All zios that want to probe
1563 	 * this vdev will become parents of the probe io.
1564 	 */
1565 	mutex_enter(&vd->vdev_probe_lock);
1566 
1567 	if ((pio = vd->vdev_probe_zio) == NULL) {
1568 		vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
1569 
1570 		vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
1571 		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
1572 		    ZIO_FLAG_TRYHARD;
1573 
1574 		if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
1575 			/*
1576 			 * vdev_cant_read and vdev_cant_write can only
1577 			 * transition from TRUE to FALSE when we have the
1578 			 * SCL_ZIO lock as writer; otherwise they can only
1579 			 * transition from FALSE to TRUE.  This ensures that
1580 			 * any zio looking at these values can assume that
1581 			 * failures persist for the life of the I/O.  That's
1582 			 * important because when a device has intermittent
1583 			 * connectivity problems, we want to ensure that
1584 			 * they're ascribed to the device (ENXIO) and not
1585 			 * the zio (EIO).
1586 			 *
1587 			 * Since we hold SCL_ZIO as writer here, clear both
1588 			 * values so the probe can reevaluate from first
1589 			 * principles.
1590 			 */
1591 			vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
1592 			vd->vdev_cant_read = B_FALSE;
1593 			vd->vdev_cant_write = B_FALSE;
1594 		}
1595 
1596 		vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
1597 		    vdev_probe_done, vps,
1598 		    vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
1599 
1600 		/*
1601 		 * We can't change the vdev state in this context, so we
1602 		 * kick off an async task to do it on our behalf.
1603 		 */
1604 		if (zio != NULL) {
1605 			vd->vdev_probe_wanted = B_TRUE;
1606 			spa_async_request(spa, SPA_ASYNC_PROBE);
1607 		}
1608 	}
1609 
1610 	if (zio != NULL)
1611 		zio_add_child(zio, pio);
1612 
1613 	mutex_exit(&vd->vdev_probe_lock);
1614 
1615 	if (vps == NULL) {
1616 		ASSERT(zio != NULL);
1617 		return (NULL);
1618 	}
1619 
1620 	for (int l = 1; l < VDEV_LABELS; l++) {
1621 		zio_nowait(zio_read_phys(pio, vd,
1622 		    vdev_label_offset(vd->vdev_psize, l,
1623 		    offsetof(vdev_label_t, vl_be)), VDEV_PAD_SIZE,
1624 		    abd_alloc_for_io(VDEV_PAD_SIZE, B_TRUE),
1625 		    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1626 		    ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1627 	}
1628 
1629 	if (zio == NULL)
1630 		return (pio);
1631 
1632 	zio_nowait(pio);
1633 	return (NULL);
1634 }
1635 
1636 static void
1637 vdev_open_child(void *arg)
1638 {
1639 	vdev_t *vd = arg;
1640 
1641 	vd->vdev_open_thread = curthread;
1642 	vd->vdev_open_error = vdev_open(vd);
1643 	vd->vdev_open_thread = NULL;
1644 }
1645 
1646 static boolean_t
1647 vdev_uses_zvols(vdev_t *vd)
1648 {
1649 #ifdef _KERNEL
1650 	if (zvol_is_zvol(vd->vdev_path))
1651 		return (B_TRUE);
1652 #endif
1653 
1654 	for (int c = 0; c < vd->vdev_children; c++)
1655 		if (vdev_uses_zvols(vd->vdev_child[c]))
1656 			return (B_TRUE);
1657 
1658 	return (B_FALSE);
1659 }
1660 
1661 /*
1662  * Returns B_TRUE if the passed child should be opened.
1663  */
1664 static boolean_t
1665 vdev_default_open_children_func(vdev_t *vd)
1666 {
1667 	return (B_TRUE);
1668 }
1669 
1670 /*
1671  * Open the requested child vdevs.  If any of the leaf vdevs are using
1672  * a ZFS volume then do the opens in a single thread.  This avoids a
1673  * deadlock when the current thread is holding the spa_namespace_lock.
1674  */
1675 static void
1676 vdev_open_children_impl(vdev_t *vd, vdev_open_children_func_t *open_func)
1677 {
1678 	int children = vd->vdev_children;
1679 
1680 	taskq_t *tq = taskq_create("vdev_open", children, minclsyspri,
1681 	    children, children, TASKQ_PREPOPULATE);
1682 	vd->vdev_nonrot = B_TRUE;
1683 
1684 	for (int c = 0; c < children; c++) {
1685 		vdev_t *cvd = vd->vdev_child[c];
1686 
1687 		if (open_func(cvd) == B_FALSE)
1688 			continue;
1689 
1690 		if (tq == NULL || vdev_uses_zvols(vd)) {
1691 			cvd->vdev_open_error = vdev_open(cvd);
1692 		} else {
1693 			VERIFY(taskq_dispatch(tq, vdev_open_child,
1694 			    cvd, TQ_SLEEP) != TASKQID_INVALID);
1695 		}
1696 
1697 		vd->vdev_nonrot &= cvd->vdev_nonrot;
1698 	}
1699 
1700 	if (tq != NULL) {
1701 		taskq_wait(tq);
1702 		taskq_destroy(tq);
1703 	}
1704 }
1705 
1706 /*
1707  * Open all child vdevs.
1708  */
1709 void
1710 vdev_open_children(vdev_t *vd)
1711 {
1712 	vdev_open_children_impl(vd, vdev_default_open_children_func);
1713 }
1714 
1715 /*
1716  * Conditionally open a subset of child vdevs.
1717  */
1718 void
1719 vdev_open_children_subset(vdev_t *vd, vdev_open_children_func_t *open_func)
1720 {
1721 	vdev_open_children_impl(vd, open_func);
1722 }
1723 
1724 /*
1725  * Compute the raidz-deflation ratio.  Note, we hard-code
1726  * in 128k (1 << 17) because it is the "typical" blocksize.
1727  * Even though SPA_MAXBLOCKSIZE changed, this algorithm can not change,
1728  * otherwise it would inconsistently account for existing bp's.
1729  */
1730 static void
1731 vdev_set_deflate_ratio(vdev_t *vd)
1732 {
1733 	if (vd == vd->vdev_top && !vd->vdev_ishole && vd->vdev_ashift != 0) {
1734 		vd->vdev_deflate_ratio = (1 << 17) /
1735 		    (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
1736 	}
1737 }
1738 
1739 /*
1740  * Maximize performance by inflating the configured ashift for top level
1741  * vdevs to be as close to the physical ashift as possible while maintaining
1742  * administrator defined limits and ensuring it doesn't go below the
1743  * logical ashift.
1744  */
1745 static void
1746 vdev_ashift_optimize(vdev_t *vd)
1747 {
1748 	ASSERT(vd == vd->vdev_top);
1749 
1750 	if (vd->vdev_ashift < vd->vdev_physical_ashift) {
1751 		vd->vdev_ashift = MIN(
1752 		    MAX(zfs_vdev_max_auto_ashift, vd->vdev_ashift),
1753 		    MAX(zfs_vdev_min_auto_ashift,
1754 		    vd->vdev_physical_ashift));
1755 	} else {
1756 		/*
1757 		 * If the logical and physical ashifts are the same, then
1758 		 * we ensure that the top-level vdev's ashift is not smaller
1759 		 * than our minimum ashift value. For the unusual case
1760 		 * where logical ashift > physical ashift, we can't cap
1761 		 * the calculated ashift based on max ashift as that
1762 		 * would cause failures.
1763 		 * We still check if we need to increase it to match
1764 		 * the min ashift.
1765 		 */
1766 		vd->vdev_ashift = MAX(zfs_vdev_min_auto_ashift,
1767 		    vd->vdev_ashift);
1768 	}
1769 }
1770 
1771 /*
1772  * Prepare a virtual device for access.
1773  */
1774 int
1775 vdev_open(vdev_t *vd)
1776 {
1777 	spa_t *spa = vd->vdev_spa;
1778 	int error;
1779 	uint64_t osize = 0;
1780 	uint64_t max_osize = 0;
1781 	uint64_t asize, max_asize, psize;
1782 	uint64_t logical_ashift = 0;
1783 	uint64_t physical_ashift = 0;
1784 
1785 	ASSERT(vd->vdev_open_thread == curthread ||
1786 	    spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1787 	ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1788 	    vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1789 	    vd->vdev_state == VDEV_STATE_OFFLINE);
1790 
1791 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1792 	vd->vdev_cant_read = B_FALSE;
1793 	vd->vdev_cant_write = B_FALSE;
1794 	vd->vdev_min_asize = vdev_get_min_asize(vd);
1795 
1796 	/*
1797 	 * If this vdev is not removed, check its fault status.  If it's
1798 	 * faulted, bail out of the open.
1799 	 */
1800 	if (!vd->vdev_removed && vd->vdev_faulted) {
1801 		ASSERT(vd->vdev_children == 0);
1802 		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1803 		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1804 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1805 		    vd->vdev_label_aux);
1806 		return (SET_ERROR(ENXIO));
1807 	} else if (vd->vdev_offline) {
1808 		ASSERT(vd->vdev_children == 0);
1809 		vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1810 		return (SET_ERROR(ENXIO));
1811 	}
1812 
1813 	error = vd->vdev_ops->vdev_op_open(vd, &osize, &max_osize,
1814 	    &logical_ashift, &physical_ashift);
1815 	/*
1816 	 * Physical volume size should never be larger than its max size, unless
1817 	 * the disk has shrunk while we were reading it or the device is buggy
1818 	 * or damaged: either way it's not safe for use, bail out of the open.
1819 	 */
1820 	if (osize > max_osize) {
1821 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1822 		    VDEV_AUX_OPEN_FAILED);
1823 		return (SET_ERROR(ENXIO));
1824 	}
1825 
1826 	/*
1827 	 * Reset the vdev_reopening flag so that we actually close
1828 	 * the vdev on error.
1829 	 */
1830 	vd->vdev_reopening = B_FALSE;
1831 	if (zio_injection_enabled && error == 0)
1832 		error = zio_handle_device_injection(vd, NULL, SET_ERROR(ENXIO));
1833 
1834 	if (error) {
1835 		if (vd->vdev_removed &&
1836 		    vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1837 			vd->vdev_removed = B_FALSE;
1838 
1839 		if (vd->vdev_stat.vs_aux == VDEV_AUX_CHILDREN_OFFLINE) {
1840 			vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE,
1841 			    vd->vdev_stat.vs_aux);
1842 		} else {
1843 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1844 			    vd->vdev_stat.vs_aux);
1845 		}
1846 		return (error);
1847 	}
1848 
1849 	vd->vdev_removed = B_FALSE;
1850 
1851 	/*
1852 	 * Recheck the faulted flag now that we have confirmed that
1853 	 * the vdev is accessible.  If we're faulted, bail.
1854 	 */
1855 	if (vd->vdev_faulted) {
1856 		ASSERT(vd->vdev_children == 0);
1857 		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1858 		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1859 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1860 		    vd->vdev_label_aux);
1861 		return (SET_ERROR(ENXIO));
1862 	}
1863 
1864 	if (vd->vdev_degraded) {
1865 		ASSERT(vd->vdev_children == 0);
1866 		vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1867 		    VDEV_AUX_ERR_EXCEEDED);
1868 	} else {
1869 		vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1870 	}
1871 
1872 	/*
1873 	 * For hole or missing vdevs we just return success.
1874 	 */
1875 	if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1876 		return (0);
1877 
1878 	for (int c = 0; c < vd->vdev_children; c++) {
1879 		if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1880 			vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1881 			    VDEV_AUX_NONE);
1882 			break;
1883 		}
1884 	}
1885 
1886 	osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1887 	max_osize = P2ALIGN(max_osize, (uint64_t)sizeof (vdev_label_t));
1888 
1889 	if (vd->vdev_children == 0) {
1890 		if (osize < SPA_MINDEVSIZE) {
1891 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1892 			    VDEV_AUX_TOO_SMALL);
1893 			return (SET_ERROR(EOVERFLOW));
1894 		}
1895 		psize = osize;
1896 		asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1897 		max_asize = max_osize - (VDEV_LABEL_START_SIZE +
1898 		    VDEV_LABEL_END_SIZE);
1899 	} else {
1900 		if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1901 		    (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1902 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1903 			    VDEV_AUX_TOO_SMALL);
1904 			return (SET_ERROR(EOVERFLOW));
1905 		}
1906 		psize = 0;
1907 		asize = osize;
1908 		max_asize = max_osize;
1909 	}
1910 
1911 	/*
1912 	 * If the vdev was expanded, record this so that we can re-create the
1913 	 * uberblock rings in labels {2,3}, during the next sync.
1914 	 */
1915 	if ((psize > vd->vdev_psize) && (vd->vdev_psize != 0))
1916 		vd->vdev_copy_uberblocks = B_TRUE;
1917 
1918 	vd->vdev_psize = psize;
1919 
1920 	/*
1921 	 * Make sure the allocatable size hasn't shrunk too much.
1922 	 */
1923 	if (asize < vd->vdev_min_asize) {
1924 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1925 		    VDEV_AUX_BAD_LABEL);
1926 		return (SET_ERROR(EINVAL));
1927 	}
1928 
1929 	/*
1930 	 * We can always set the logical/physical ashift members since
1931 	 * their values are only used to calculate the vdev_ashift when
1932 	 * the device is first added to the config. These values should
1933 	 * not be used for anything else since they may change whenever
1934 	 * the device is reopened and we don't store them in the label.
1935 	 */
1936 	vd->vdev_physical_ashift =
1937 	    MAX(physical_ashift, vd->vdev_physical_ashift);
1938 	vd->vdev_logical_ashift = MAX(logical_ashift,
1939 	    vd->vdev_logical_ashift);
1940 
1941 	if (vd->vdev_asize == 0) {
1942 		/*
1943 		 * This is the first-ever open, so use the computed values.
1944 		 * For compatibility, a different ashift can be requested.
1945 		 */
1946 		vd->vdev_asize = asize;
1947 		vd->vdev_max_asize = max_asize;
1948 
1949 		/*
1950 		 * If the vdev_ashift was not overriden at creation time,
1951 		 * then set it the logical ashift and optimize the ashift.
1952 		 */
1953 		if (vd->vdev_ashift == 0) {
1954 			vd->vdev_ashift = vd->vdev_logical_ashift;
1955 
1956 			if (vd->vdev_logical_ashift > ASHIFT_MAX) {
1957 				vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1958 				    VDEV_AUX_ASHIFT_TOO_BIG);
1959 				return (SET_ERROR(EDOM));
1960 			}
1961 
1962 			if (vd->vdev_top == vd) {
1963 				vdev_ashift_optimize(vd);
1964 			}
1965 		}
1966 		if (vd->vdev_ashift != 0 && (vd->vdev_ashift < ASHIFT_MIN ||
1967 		    vd->vdev_ashift > ASHIFT_MAX)) {
1968 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1969 			    VDEV_AUX_BAD_ASHIFT);
1970 			return (SET_ERROR(EDOM));
1971 		}
1972 	} else {
1973 		/*
1974 		 * Make sure the alignment required hasn't increased.
1975 		 */
1976 		if (vd->vdev_ashift > vd->vdev_top->vdev_ashift &&
1977 		    vd->vdev_ops->vdev_op_leaf) {
1978 			(void) zfs_ereport_post(
1979 			    FM_EREPORT_ZFS_DEVICE_BAD_ASHIFT,
1980 			    spa, vd, NULL, NULL, 0);
1981 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1982 			    VDEV_AUX_BAD_LABEL);
1983 			return (SET_ERROR(EDOM));
1984 		}
1985 		vd->vdev_max_asize = max_asize;
1986 	}
1987 
1988 	/*
1989 	 * If all children are healthy we update asize if either:
1990 	 * The asize has increased, due to a device expansion caused by dynamic
1991 	 * LUN growth or vdev replacement, and automatic expansion is enabled;
1992 	 * making the additional space available.
1993 	 *
1994 	 * The asize has decreased, due to a device shrink usually caused by a
1995 	 * vdev replace with a smaller device. This ensures that calculations
1996 	 * based of max_asize and asize e.g. esize are always valid. It's safe
1997 	 * to do this as we've already validated that asize is greater than
1998 	 * vdev_min_asize.
1999 	 */
2000 	if (vd->vdev_state == VDEV_STATE_HEALTHY &&
2001 	    ((asize > vd->vdev_asize &&
2002 	    (vd->vdev_expanding || spa->spa_autoexpand)) ||
2003 	    (asize < vd->vdev_asize)))
2004 		vd->vdev_asize = asize;
2005 
2006 	vdev_set_min_asize(vd);
2007 
2008 	/*
2009 	 * Ensure we can issue some IO before declaring the
2010 	 * vdev open for business.
2011 	 */
2012 	if (vd->vdev_ops->vdev_op_leaf &&
2013 	    (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
2014 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
2015 		    VDEV_AUX_ERR_EXCEEDED);
2016 		return (error);
2017 	}
2018 
2019 	/*
2020 	 * Track the the minimum allocation size.
2021 	 */
2022 	if (vd->vdev_top == vd && vd->vdev_ashift != 0 &&
2023 	    vd->vdev_islog == 0 && vd->vdev_aux == NULL) {
2024 		uint64_t min_alloc = vdev_get_min_alloc(vd);
2025 		if (min_alloc < spa->spa_min_alloc)
2026 			spa->spa_min_alloc = min_alloc;
2027 	}
2028 
2029 	/*
2030 	 * If this is a leaf vdev, assess whether a resilver is needed.
2031 	 * But don't do this if we are doing a reopen for a scrub, since
2032 	 * this would just restart the scrub we are already doing.
2033 	 */
2034 	if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen)
2035 		dsl_scan_assess_vdev(spa->spa_dsl_pool, vd);
2036 
2037 	return (0);
2038 }
2039 
2040 /*
2041  * Called once the vdevs are all opened, this routine validates the label
2042  * contents. This needs to be done before vdev_load() so that we don't
2043  * inadvertently do repair I/Os to the wrong device.
2044  *
2045  * This function will only return failure if one of the vdevs indicates that it
2046  * has since been destroyed or exported.  This is only possible if
2047  * /etc/zfs/zpool.cache was readonly at the time.  Otherwise, the vdev state
2048  * will be updated but the function will return 0.
2049  */
2050 int
2051 vdev_validate(vdev_t *vd)
2052 {
2053 	spa_t *spa = vd->vdev_spa;
2054 	nvlist_t *label;
2055 	uint64_t guid = 0, aux_guid = 0, top_guid;
2056 	uint64_t state;
2057 	nvlist_t *nvl;
2058 	uint64_t txg;
2059 
2060 	if (vdev_validate_skip)
2061 		return (0);
2062 
2063 	for (uint64_t c = 0; c < vd->vdev_children; c++)
2064 		if (vdev_validate(vd->vdev_child[c]) != 0)
2065 			return (SET_ERROR(EBADF));
2066 
2067 	/*
2068 	 * If the device has already failed, or was marked offline, don't do
2069 	 * any further validation.  Otherwise, label I/O will fail and we will
2070 	 * overwrite the previous state.
2071 	 */
2072 	if (!vd->vdev_ops->vdev_op_leaf || !vdev_readable(vd))
2073 		return (0);
2074 
2075 	/*
2076 	 * If we are performing an extreme rewind, we allow for a label that
2077 	 * was modified at a point after the current txg.
2078 	 * If config lock is not held do not check for the txg. spa_sync could
2079 	 * be updating the vdev's label before updating spa_last_synced_txg.
2080 	 */
2081 	if (spa->spa_extreme_rewind || spa_last_synced_txg(spa) == 0 ||
2082 	    spa_config_held(spa, SCL_CONFIG, RW_WRITER) != SCL_CONFIG)
2083 		txg = UINT64_MAX;
2084 	else
2085 		txg = spa_last_synced_txg(spa);
2086 
2087 	if ((label = vdev_label_read_config(vd, txg)) == NULL) {
2088 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2089 		    VDEV_AUX_BAD_LABEL);
2090 		vdev_dbgmsg(vd, "vdev_validate: failed reading config for "
2091 		    "txg %llu", (u_longlong_t)txg);
2092 		return (0);
2093 	}
2094 
2095 	/*
2096 	 * Determine if this vdev has been split off into another
2097 	 * pool.  If so, then refuse to open it.
2098 	 */
2099 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID,
2100 	    &aux_guid) == 0 && aux_guid == spa_guid(spa)) {
2101 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2102 		    VDEV_AUX_SPLIT_POOL);
2103 		nvlist_free(label);
2104 		vdev_dbgmsg(vd, "vdev_validate: vdev split into other pool");
2105 		return (0);
2106 	}
2107 
2108 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_GUID, &guid) != 0) {
2109 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2110 		    VDEV_AUX_CORRUPT_DATA);
2111 		nvlist_free(label);
2112 		vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
2113 		    ZPOOL_CONFIG_POOL_GUID);
2114 		return (0);
2115 	}
2116 
2117 	/*
2118 	 * If config is not trusted then ignore the spa guid check. This is
2119 	 * necessary because if the machine crashed during a re-guid the new
2120 	 * guid might have been written to all of the vdev labels, but not the
2121 	 * cached config. The check will be performed again once we have the
2122 	 * trusted config from the MOS.
2123 	 */
2124 	if (spa->spa_trust_config && guid != spa_guid(spa)) {
2125 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2126 		    VDEV_AUX_CORRUPT_DATA);
2127 		nvlist_free(label);
2128 		vdev_dbgmsg(vd, "vdev_validate: vdev label pool_guid doesn't "
2129 		    "match config (%llu != %llu)", (u_longlong_t)guid,
2130 		    (u_longlong_t)spa_guid(spa));
2131 		return (0);
2132 	}
2133 
2134 	if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl)
2135 	    != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID,
2136 	    &aux_guid) != 0)
2137 		aux_guid = 0;
2138 
2139 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0) {
2140 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2141 		    VDEV_AUX_CORRUPT_DATA);
2142 		nvlist_free(label);
2143 		vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
2144 		    ZPOOL_CONFIG_GUID);
2145 		return (0);
2146 	}
2147 
2148 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID, &top_guid)
2149 	    != 0) {
2150 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2151 		    VDEV_AUX_CORRUPT_DATA);
2152 		nvlist_free(label);
2153 		vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
2154 		    ZPOOL_CONFIG_TOP_GUID);
2155 		return (0);
2156 	}
2157 
2158 	/*
2159 	 * If this vdev just became a top-level vdev because its sibling was
2160 	 * detached, it will have adopted the parent's vdev guid -- but the
2161 	 * label may or may not be on disk yet. Fortunately, either version
2162 	 * of the label will have the same top guid, so if we're a top-level
2163 	 * vdev, we can safely compare to that instead.
2164 	 * However, if the config comes from a cachefile that failed to update
2165 	 * after the detach, a top-level vdev will appear as a non top-level
2166 	 * vdev in the config. Also relax the constraints if we perform an
2167 	 * extreme rewind.
2168 	 *
2169 	 * If we split this vdev off instead, then we also check the
2170 	 * original pool's guid. We don't want to consider the vdev
2171 	 * corrupt if it is partway through a split operation.
2172 	 */
2173 	if (vd->vdev_guid != guid && vd->vdev_guid != aux_guid) {
2174 		boolean_t mismatch = B_FALSE;
2175 		if (spa->spa_trust_config && !spa->spa_extreme_rewind) {
2176 			if (vd != vd->vdev_top || vd->vdev_guid != top_guid)
2177 				mismatch = B_TRUE;
2178 		} else {
2179 			if (vd->vdev_guid != top_guid &&
2180 			    vd->vdev_top->vdev_guid != guid)
2181 				mismatch = B_TRUE;
2182 		}
2183 
2184 		if (mismatch) {
2185 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2186 			    VDEV_AUX_CORRUPT_DATA);
2187 			nvlist_free(label);
2188 			vdev_dbgmsg(vd, "vdev_validate: config guid "
2189 			    "doesn't match label guid");
2190 			vdev_dbgmsg(vd, "CONFIG: guid %llu, top_guid %llu",
2191 			    (u_longlong_t)vd->vdev_guid,
2192 			    (u_longlong_t)vd->vdev_top->vdev_guid);
2193 			vdev_dbgmsg(vd, "LABEL: guid %llu, top_guid %llu, "
2194 			    "aux_guid %llu", (u_longlong_t)guid,
2195 			    (u_longlong_t)top_guid, (u_longlong_t)aux_guid);
2196 			return (0);
2197 		}
2198 	}
2199 
2200 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
2201 	    &state) != 0) {
2202 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2203 		    VDEV_AUX_CORRUPT_DATA);
2204 		nvlist_free(label);
2205 		vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
2206 		    ZPOOL_CONFIG_POOL_STATE);
2207 		return (0);
2208 	}
2209 
2210 	nvlist_free(label);
2211 
2212 	/*
2213 	 * If this is a verbatim import, no need to check the
2214 	 * state of the pool.
2215 	 */
2216 	if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) &&
2217 	    spa_load_state(spa) == SPA_LOAD_OPEN &&
2218 	    state != POOL_STATE_ACTIVE) {
2219 		vdev_dbgmsg(vd, "vdev_validate: invalid pool state (%llu) "
2220 		    "for spa %s", (u_longlong_t)state, spa->spa_name);
2221 		return (SET_ERROR(EBADF));
2222 	}
2223 
2224 	/*
2225 	 * If we were able to open and validate a vdev that was
2226 	 * previously marked permanently unavailable, clear that state
2227 	 * now.
2228 	 */
2229 	if (vd->vdev_not_present)
2230 		vd->vdev_not_present = 0;
2231 
2232 	return (0);
2233 }
2234 
2235 static void
2236 vdev_copy_path_impl(vdev_t *svd, vdev_t *dvd)
2237 {
2238 	if (svd->vdev_path != NULL && dvd->vdev_path != NULL) {
2239 		if (strcmp(svd->vdev_path, dvd->vdev_path) != 0) {
2240 			zfs_dbgmsg("vdev_copy_path: vdev %llu: path changed "
2241 			    "from '%s' to '%s'", (u_longlong_t)dvd->vdev_guid,
2242 			    dvd->vdev_path, svd->vdev_path);
2243 			spa_strfree(dvd->vdev_path);
2244 			dvd->vdev_path = spa_strdup(svd->vdev_path);
2245 		}
2246 	} else if (svd->vdev_path != NULL) {
2247 		dvd->vdev_path = spa_strdup(svd->vdev_path);
2248 		zfs_dbgmsg("vdev_copy_path: vdev %llu: path set to '%s'",
2249 		    (u_longlong_t)dvd->vdev_guid, dvd->vdev_path);
2250 	}
2251 }
2252 
2253 /*
2254  * Recursively copy vdev paths from one vdev to another. Source and destination
2255  * vdev trees must have same geometry otherwise return error. Intended to copy
2256  * paths from userland config into MOS config.
2257  */
2258 int
2259 vdev_copy_path_strict(vdev_t *svd, vdev_t *dvd)
2260 {
2261 	if ((svd->vdev_ops == &vdev_missing_ops) ||
2262 	    (svd->vdev_ishole && dvd->vdev_ishole) ||
2263 	    (dvd->vdev_ops == &vdev_indirect_ops))
2264 		return (0);
2265 
2266 	if (svd->vdev_ops != dvd->vdev_ops) {
2267 		vdev_dbgmsg(svd, "vdev_copy_path: vdev type mismatch: %s != %s",
2268 		    svd->vdev_ops->vdev_op_type, dvd->vdev_ops->vdev_op_type);
2269 		return (SET_ERROR(EINVAL));
2270 	}
2271 
2272 	if (svd->vdev_guid != dvd->vdev_guid) {
2273 		vdev_dbgmsg(svd, "vdev_copy_path: guids mismatch (%llu != "
2274 		    "%llu)", (u_longlong_t)svd->vdev_guid,
2275 		    (u_longlong_t)dvd->vdev_guid);
2276 		return (SET_ERROR(EINVAL));
2277 	}
2278 
2279 	if (svd->vdev_children != dvd->vdev_children) {
2280 		vdev_dbgmsg(svd, "vdev_copy_path: children count mismatch: "
2281 		    "%llu != %llu", (u_longlong_t)svd->vdev_children,
2282 		    (u_longlong_t)dvd->vdev_children);
2283 		return (SET_ERROR(EINVAL));
2284 	}
2285 
2286 	for (uint64_t i = 0; i < svd->vdev_children; i++) {
2287 		int error = vdev_copy_path_strict(svd->vdev_child[i],
2288 		    dvd->vdev_child[i]);
2289 		if (error != 0)
2290 			return (error);
2291 	}
2292 
2293 	if (svd->vdev_ops->vdev_op_leaf)
2294 		vdev_copy_path_impl(svd, dvd);
2295 
2296 	return (0);
2297 }
2298 
2299 static void
2300 vdev_copy_path_search(vdev_t *stvd, vdev_t *dvd)
2301 {
2302 	ASSERT(stvd->vdev_top == stvd);
2303 	ASSERT3U(stvd->vdev_id, ==, dvd->vdev_top->vdev_id);
2304 
2305 	for (uint64_t i = 0; i < dvd->vdev_children; i++) {
2306 		vdev_copy_path_search(stvd, dvd->vdev_child[i]);
2307 	}
2308 
2309 	if (!dvd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(dvd))
2310 		return;
2311 
2312 	/*
2313 	 * The idea here is that while a vdev can shift positions within
2314 	 * a top vdev (when replacing, attaching mirror, etc.) it cannot
2315 	 * step outside of it.
2316 	 */
2317 	vdev_t *vd = vdev_lookup_by_guid(stvd, dvd->vdev_guid);
2318 
2319 	if (vd == NULL || vd->vdev_ops != dvd->vdev_ops)
2320 		return;
2321 
2322 	ASSERT(vd->vdev_ops->vdev_op_leaf);
2323 
2324 	vdev_copy_path_impl(vd, dvd);
2325 }
2326 
2327 /*
2328  * Recursively copy vdev paths from one root vdev to another. Source and
2329  * destination vdev trees may differ in geometry. For each destination leaf
2330  * vdev, search a vdev with the same guid and top vdev id in the source.
2331  * Intended to copy paths from userland config into MOS config.
2332  */
2333 void
2334 vdev_copy_path_relaxed(vdev_t *srvd, vdev_t *drvd)
2335 {
2336 	uint64_t children = MIN(srvd->vdev_children, drvd->vdev_children);
2337 	ASSERT(srvd->vdev_ops == &vdev_root_ops);
2338 	ASSERT(drvd->vdev_ops == &vdev_root_ops);
2339 
2340 	for (uint64_t i = 0; i < children; i++) {
2341 		vdev_copy_path_search(srvd->vdev_child[i],
2342 		    drvd->vdev_child[i]);
2343 	}
2344 }
2345 
2346 /*
2347  * Close a virtual device.
2348  */
2349 void
2350 vdev_close(vdev_t *vd)
2351 {
2352 	vdev_t *pvd = vd->vdev_parent;
2353 	spa_t *spa __maybe_unused = vd->vdev_spa;
2354 
2355 	ASSERT(vd != NULL);
2356 	ASSERT(vd->vdev_open_thread == curthread ||
2357 	    spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2358 
2359 	/*
2360 	 * If our parent is reopening, then we are as well, unless we are
2361 	 * going offline.
2362 	 */
2363 	if (pvd != NULL && pvd->vdev_reopening)
2364 		vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline);
2365 
2366 	vd->vdev_ops->vdev_op_close(vd);
2367 
2368 	vdev_cache_purge(vd);
2369 
2370 	/*
2371 	 * We record the previous state before we close it, so that if we are
2372 	 * doing a reopen(), we don't generate FMA ereports if we notice that
2373 	 * it's still faulted.
2374 	 */
2375 	vd->vdev_prevstate = vd->vdev_state;
2376 
2377 	if (vd->vdev_offline)
2378 		vd->vdev_state = VDEV_STATE_OFFLINE;
2379 	else
2380 		vd->vdev_state = VDEV_STATE_CLOSED;
2381 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
2382 }
2383 
2384 void
2385 vdev_hold(vdev_t *vd)
2386 {
2387 	spa_t *spa = vd->vdev_spa;
2388 
2389 	ASSERT(spa_is_root(spa));
2390 	if (spa->spa_state == POOL_STATE_UNINITIALIZED)
2391 		return;
2392 
2393 	for (int c = 0; c < vd->vdev_children; c++)
2394 		vdev_hold(vd->vdev_child[c]);
2395 
2396 	if (vd->vdev_ops->vdev_op_leaf)
2397 		vd->vdev_ops->vdev_op_hold(vd);
2398 }
2399 
2400 void
2401 vdev_rele(vdev_t *vd)
2402 {
2403 	ASSERT(spa_is_root(vd->vdev_spa));
2404 	for (int c = 0; c < vd->vdev_children; c++)
2405 		vdev_rele(vd->vdev_child[c]);
2406 
2407 	if (vd->vdev_ops->vdev_op_leaf)
2408 		vd->vdev_ops->vdev_op_rele(vd);
2409 }
2410 
2411 /*
2412  * Reopen all interior vdevs and any unopened leaves.  We don't actually
2413  * reopen leaf vdevs which had previously been opened as they might deadlock
2414  * on the spa_config_lock.  Instead we only obtain the leaf's physical size.
2415  * If the leaf has never been opened then open it, as usual.
2416  */
2417 void
2418 vdev_reopen(vdev_t *vd)
2419 {
2420 	spa_t *spa = vd->vdev_spa;
2421 
2422 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2423 
2424 	/* set the reopening flag unless we're taking the vdev offline */
2425 	vd->vdev_reopening = !vd->vdev_offline;
2426 	vdev_close(vd);
2427 	(void) vdev_open(vd);
2428 
2429 	/*
2430 	 * Call vdev_validate() here to make sure we have the same device.
2431 	 * Otherwise, a device with an invalid label could be successfully
2432 	 * opened in response to vdev_reopen().
2433 	 */
2434 	if (vd->vdev_aux) {
2435 		(void) vdev_validate_aux(vd);
2436 		if (vdev_readable(vd) && vdev_writeable(vd) &&
2437 		    vd->vdev_aux == &spa->spa_l2cache) {
2438 			/*
2439 			 * In case the vdev is present we should evict all ARC
2440 			 * buffers and pointers to log blocks and reclaim their
2441 			 * space before restoring its contents to L2ARC.
2442 			 */
2443 			if (l2arc_vdev_present(vd)) {
2444 				l2arc_rebuild_vdev(vd, B_TRUE);
2445 			} else {
2446 				l2arc_add_vdev(spa, vd);
2447 			}
2448 			spa_async_request(spa, SPA_ASYNC_L2CACHE_REBUILD);
2449 			spa_async_request(spa, SPA_ASYNC_L2CACHE_TRIM);
2450 		}
2451 	} else {
2452 		(void) vdev_validate(vd);
2453 	}
2454 
2455 	/*
2456 	 * Reassess parent vdev's health.
2457 	 */
2458 	vdev_propagate_state(vd);
2459 }
2460 
2461 int
2462 vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
2463 {
2464 	int error;
2465 
2466 	/*
2467 	 * Normally, partial opens (e.g. of a mirror) are allowed.
2468 	 * For a create, however, we want to fail the request if
2469 	 * there are any components we can't open.
2470 	 */
2471 	error = vdev_open(vd);
2472 
2473 	if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
2474 		vdev_close(vd);
2475 		return (error ? error : SET_ERROR(ENXIO));
2476 	}
2477 
2478 	/*
2479 	 * Recursively load DTLs and initialize all labels.
2480 	 */
2481 	if ((error = vdev_dtl_load(vd)) != 0 ||
2482 	    (error = vdev_label_init(vd, txg, isreplacing ?
2483 	    VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
2484 		vdev_close(vd);
2485 		return (error);
2486 	}
2487 
2488 	return (0);
2489 }
2490 
2491 void
2492 vdev_metaslab_set_size(vdev_t *vd)
2493 {
2494 	uint64_t asize = vd->vdev_asize;
2495 	uint64_t ms_count = asize >> zfs_vdev_default_ms_shift;
2496 	uint64_t ms_shift;
2497 
2498 	/*
2499 	 * There are two dimensions to the metaslab sizing calculation:
2500 	 * the size of the metaslab and the count of metaslabs per vdev.
2501 	 *
2502 	 * The default values used below are a good balance between memory
2503 	 * usage (larger metaslab size means more memory needed for loaded
2504 	 * metaslabs; more metaslabs means more memory needed for the
2505 	 * metaslab_t structs), metaslab load time (larger metaslabs take
2506 	 * longer to load), and metaslab sync time (more metaslabs means
2507 	 * more time spent syncing all of them).
2508 	 *
2509 	 * In general, we aim for zfs_vdev_default_ms_count (200) metaslabs.
2510 	 * The range of the dimensions are as follows:
2511 	 *
2512 	 *	2^29 <= ms_size  <= 2^34
2513 	 *	  16 <= ms_count <= 131,072
2514 	 *
2515 	 * On the lower end of vdev sizes, we aim for metaslabs sizes of
2516 	 * at least 512MB (2^29) to minimize fragmentation effects when
2517 	 * testing with smaller devices.  However, the count constraint
2518 	 * of at least 16 metaslabs will override this minimum size goal.
2519 	 *
2520 	 * On the upper end of vdev sizes, we aim for a maximum metaslab
2521 	 * size of 16GB.  However, we will cap the total count to 2^17
2522 	 * metaslabs to keep our memory footprint in check and let the
2523 	 * metaslab size grow from there if that limit is hit.
2524 	 *
2525 	 * The net effect of applying above constrains is summarized below.
2526 	 *
2527 	 *   vdev size       metaslab count
2528 	 *  --------------|-----------------
2529 	 *      < 8GB        ~16
2530 	 *  8GB   - 100GB   one per 512MB
2531 	 *  100GB - 3TB     ~200
2532 	 *  3TB   - 2PB     one per 16GB
2533 	 *      > 2PB       ~131,072
2534 	 *  --------------------------------
2535 	 *
2536 	 *  Finally, note that all of the above calculate the initial
2537 	 *  number of metaslabs. Expanding a top-level vdev will result
2538 	 *  in additional metaslabs being allocated making it possible
2539 	 *  to exceed the zfs_vdev_ms_count_limit.
2540 	 */
2541 
2542 	if (ms_count < zfs_vdev_min_ms_count)
2543 		ms_shift = highbit64(asize / zfs_vdev_min_ms_count);
2544 	else if (ms_count > zfs_vdev_default_ms_count)
2545 		ms_shift = highbit64(asize / zfs_vdev_default_ms_count);
2546 	else
2547 		ms_shift = zfs_vdev_default_ms_shift;
2548 
2549 	if (ms_shift < SPA_MAXBLOCKSHIFT) {
2550 		ms_shift = SPA_MAXBLOCKSHIFT;
2551 	} else if (ms_shift > zfs_vdev_max_ms_shift) {
2552 		ms_shift = zfs_vdev_max_ms_shift;
2553 		/* cap the total count to constrain memory footprint */
2554 		if ((asize >> ms_shift) > zfs_vdev_ms_count_limit)
2555 			ms_shift = highbit64(asize / zfs_vdev_ms_count_limit);
2556 	}
2557 
2558 	vd->vdev_ms_shift = ms_shift;
2559 	ASSERT3U(vd->vdev_ms_shift, >=, SPA_MAXBLOCKSHIFT);
2560 }
2561 
2562 void
2563 vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
2564 {
2565 	ASSERT(vd == vd->vdev_top);
2566 	/* indirect vdevs don't have metaslabs or dtls */
2567 	ASSERT(vdev_is_concrete(vd) || flags == 0);
2568 	ASSERT(ISP2(flags));
2569 	ASSERT(spa_writeable(vd->vdev_spa));
2570 
2571 	if (flags & VDD_METASLAB)
2572 		(void) txg_list_add(&vd->vdev_ms_list, arg, txg);
2573 
2574 	if (flags & VDD_DTL)
2575 		(void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
2576 
2577 	(void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
2578 }
2579 
2580 void
2581 vdev_dirty_leaves(vdev_t *vd, int flags, uint64_t txg)
2582 {
2583 	for (int c = 0; c < vd->vdev_children; c++)
2584 		vdev_dirty_leaves(vd->vdev_child[c], flags, txg);
2585 
2586 	if (vd->vdev_ops->vdev_op_leaf)
2587 		vdev_dirty(vd->vdev_top, flags, vd, txg);
2588 }
2589 
2590 /*
2591  * DTLs.
2592  *
2593  * A vdev's DTL (dirty time log) is the set of transaction groups for which
2594  * the vdev has less than perfect replication.  There are four kinds of DTL:
2595  *
2596  * DTL_MISSING: txgs for which the vdev has no valid copies of the data
2597  *
2598  * DTL_PARTIAL: txgs for which data is available, but not fully replicated
2599  *
2600  * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
2601  *	scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
2602  *	txgs that was scrubbed.
2603  *
2604  * DTL_OUTAGE: txgs which cannot currently be read, whether due to
2605  *	persistent errors or just some device being offline.
2606  *	Unlike the other three, the DTL_OUTAGE map is not generally
2607  *	maintained; it's only computed when needed, typically to
2608  *	determine whether a device can be detached.
2609  *
2610  * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
2611  * either has the data or it doesn't.
2612  *
2613  * For interior vdevs such as mirror and RAID-Z the picture is more complex.
2614  * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
2615  * if any child is less than fully replicated, then so is its parent.
2616  * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
2617  * comprising only those txgs which appear in 'maxfaults' or more children;
2618  * those are the txgs we don't have enough replication to read.  For example,
2619  * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
2620  * thus, its DTL_MISSING consists of the set of txgs that appear in more than
2621  * two child DTL_MISSING maps.
2622  *
2623  * It should be clear from the above that to compute the DTLs and outage maps
2624  * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
2625  * Therefore, that is all we keep on disk.  When loading the pool, or after
2626  * a configuration change, we generate all other DTLs from first principles.
2627  */
2628 void
2629 vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
2630 {
2631 	range_tree_t *rt = vd->vdev_dtl[t];
2632 
2633 	ASSERT(t < DTL_TYPES);
2634 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
2635 	ASSERT(spa_writeable(vd->vdev_spa));
2636 
2637 	mutex_enter(&vd->vdev_dtl_lock);
2638 	if (!range_tree_contains(rt, txg, size))
2639 		range_tree_add(rt, txg, size);
2640 	mutex_exit(&vd->vdev_dtl_lock);
2641 }
2642 
2643 boolean_t
2644 vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
2645 {
2646 	range_tree_t *rt = vd->vdev_dtl[t];
2647 	boolean_t dirty = B_FALSE;
2648 
2649 	ASSERT(t < DTL_TYPES);
2650 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
2651 
2652 	/*
2653 	 * While we are loading the pool, the DTLs have not been loaded yet.
2654 	 * This isn't a problem but it can result in devices being tried
2655 	 * which are known to not have the data.  In which case, the import
2656 	 * is relying on the checksum to ensure that we get the right data.
2657 	 * Note that while importing we are only reading the MOS, which is
2658 	 * always checksummed.
2659 	 */
2660 	mutex_enter(&vd->vdev_dtl_lock);
2661 	if (!range_tree_is_empty(rt))
2662 		dirty = range_tree_contains(rt, txg, size);
2663 	mutex_exit(&vd->vdev_dtl_lock);
2664 
2665 	return (dirty);
2666 }
2667 
2668 boolean_t
2669 vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
2670 {
2671 	range_tree_t *rt = vd->vdev_dtl[t];
2672 	boolean_t empty;
2673 
2674 	mutex_enter(&vd->vdev_dtl_lock);
2675 	empty = range_tree_is_empty(rt);
2676 	mutex_exit(&vd->vdev_dtl_lock);
2677 
2678 	return (empty);
2679 }
2680 
2681 /*
2682  * Check if the txg falls within the range which must be
2683  * resilvered.  DVAs outside this range can always be skipped.
2684  */
2685 boolean_t
2686 vdev_default_need_resilver(vdev_t *vd, const dva_t *dva, size_t psize,
2687     uint64_t phys_birth)
2688 {
2689 	/* Set by sequential resilver. */
2690 	if (phys_birth == TXG_UNKNOWN)
2691 		return (B_TRUE);
2692 
2693 	return (vdev_dtl_contains(vd, DTL_PARTIAL, phys_birth, 1));
2694 }
2695 
2696 /*
2697  * Returns B_TRUE if the vdev determines the DVA needs to be resilvered.
2698  */
2699 boolean_t
2700 vdev_dtl_need_resilver(vdev_t *vd, const dva_t *dva, size_t psize,
2701     uint64_t phys_birth)
2702 {
2703 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
2704 
2705 	if (vd->vdev_ops->vdev_op_need_resilver == NULL ||
2706 	    vd->vdev_ops->vdev_op_leaf)
2707 		return (B_TRUE);
2708 
2709 	return (vd->vdev_ops->vdev_op_need_resilver(vd, dva, psize,
2710 	    phys_birth));
2711 }
2712 
2713 /*
2714  * Returns the lowest txg in the DTL range.
2715  */
2716 static uint64_t
2717 vdev_dtl_min(vdev_t *vd)
2718 {
2719 	ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
2720 	ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
2721 	ASSERT0(vd->vdev_children);
2722 
2723 	return (range_tree_min(vd->vdev_dtl[DTL_MISSING]) - 1);
2724 }
2725 
2726 /*
2727  * Returns the highest txg in the DTL.
2728  */
2729 static uint64_t
2730 vdev_dtl_max(vdev_t *vd)
2731 {
2732 	ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
2733 	ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
2734 	ASSERT0(vd->vdev_children);
2735 
2736 	return (range_tree_max(vd->vdev_dtl[DTL_MISSING]));
2737 }
2738 
2739 /*
2740  * Determine if a resilvering vdev should remove any DTL entries from
2741  * its range. If the vdev was resilvering for the entire duration of the
2742  * scan then it should excise that range from its DTLs. Otherwise, this
2743  * vdev is considered partially resilvered and should leave its DTL
2744  * entries intact. The comment in vdev_dtl_reassess() describes how we
2745  * excise the DTLs.
2746  */
2747 static boolean_t
2748 vdev_dtl_should_excise(vdev_t *vd, boolean_t rebuild_done)
2749 {
2750 	ASSERT0(vd->vdev_children);
2751 
2752 	if (vd->vdev_state < VDEV_STATE_DEGRADED)
2753 		return (B_FALSE);
2754 
2755 	if (vd->vdev_resilver_deferred)
2756 		return (B_FALSE);
2757 
2758 	if (range_tree_is_empty(vd->vdev_dtl[DTL_MISSING]))
2759 		return (B_TRUE);
2760 
2761 	if (rebuild_done) {
2762 		vdev_rebuild_t *vr = &vd->vdev_top->vdev_rebuild_config;
2763 		vdev_rebuild_phys_t *vrp = &vr->vr_rebuild_phys;
2764 
2765 		/* Rebuild not initiated by attach */
2766 		if (vd->vdev_rebuild_txg == 0)
2767 			return (B_TRUE);
2768 
2769 		/*
2770 		 * When a rebuild completes without error then all missing data
2771 		 * up to the rebuild max txg has been reconstructed and the DTL
2772 		 * is eligible for excision.
2773 		 */
2774 		if (vrp->vrp_rebuild_state == VDEV_REBUILD_COMPLETE &&
2775 		    vdev_dtl_max(vd) <= vrp->vrp_max_txg) {
2776 			ASSERT3U(vrp->vrp_min_txg, <=, vdev_dtl_min(vd));
2777 			ASSERT3U(vrp->vrp_min_txg, <, vd->vdev_rebuild_txg);
2778 			ASSERT3U(vd->vdev_rebuild_txg, <=, vrp->vrp_max_txg);
2779 			return (B_TRUE);
2780 		}
2781 	} else {
2782 		dsl_scan_t *scn = vd->vdev_spa->spa_dsl_pool->dp_scan;
2783 		dsl_scan_phys_t *scnp __maybe_unused = &scn->scn_phys;
2784 
2785 		/* Resilver not initiated by attach */
2786 		if (vd->vdev_resilver_txg == 0)
2787 			return (B_TRUE);
2788 
2789 		/*
2790 		 * When a resilver is initiated the scan will assign the
2791 		 * scn_max_txg value to the highest txg value that exists
2792 		 * in all DTLs. If this device's max DTL is not part of this
2793 		 * scan (i.e. it is not in the range (scn_min_txg, scn_max_txg]
2794 		 * then it is not eligible for excision.
2795 		 */
2796 		if (vdev_dtl_max(vd) <= scn->scn_phys.scn_max_txg) {
2797 			ASSERT3U(scnp->scn_min_txg, <=, vdev_dtl_min(vd));
2798 			ASSERT3U(scnp->scn_min_txg, <, vd->vdev_resilver_txg);
2799 			ASSERT3U(vd->vdev_resilver_txg, <=, scnp->scn_max_txg);
2800 			return (B_TRUE);
2801 		}
2802 	}
2803 
2804 	return (B_FALSE);
2805 }
2806 
2807 /*
2808  * Reassess DTLs after a config change or scrub completion. If txg == 0 no
2809  * write operations will be issued to the pool.
2810  */
2811 void
2812 vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg,
2813     boolean_t scrub_done, boolean_t rebuild_done)
2814 {
2815 	spa_t *spa = vd->vdev_spa;
2816 	avl_tree_t reftree;
2817 	int minref;
2818 
2819 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
2820 
2821 	for (int c = 0; c < vd->vdev_children; c++)
2822 		vdev_dtl_reassess(vd->vdev_child[c], txg,
2823 		    scrub_txg, scrub_done, rebuild_done);
2824 
2825 	if (vd == spa->spa_root_vdev || !vdev_is_concrete(vd) || vd->vdev_aux)
2826 		return;
2827 
2828 	if (vd->vdev_ops->vdev_op_leaf) {
2829 		dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
2830 		vdev_rebuild_t *vr = &vd->vdev_top->vdev_rebuild_config;
2831 		boolean_t check_excise = B_FALSE;
2832 		boolean_t wasempty = B_TRUE;
2833 
2834 		mutex_enter(&vd->vdev_dtl_lock);
2835 
2836 		/*
2837 		 * If requested, pretend the scan or rebuild completed cleanly.
2838 		 */
2839 		if (zfs_scan_ignore_errors) {
2840 			if (scn != NULL)
2841 				scn->scn_phys.scn_errors = 0;
2842 			if (vr != NULL)
2843 				vr->vr_rebuild_phys.vrp_errors = 0;
2844 		}
2845 
2846 		if (scrub_txg != 0 &&
2847 		    !range_tree_is_empty(vd->vdev_dtl[DTL_MISSING])) {
2848 			wasempty = B_FALSE;
2849 			zfs_dbgmsg("guid:%llu txg:%llu scrub:%llu started:%d "
2850 			    "dtl:%llu/%llu errors:%llu",
2851 			    (u_longlong_t)vd->vdev_guid, (u_longlong_t)txg,
2852 			    (u_longlong_t)scrub_txg, spa->spa_scrub_started,
2853 			    (u_longlong_t)vdev_dtl_min(vd),
2854 			    (u_longlong_t)vdev_dtl_max(vd),
2855 			    (u_longlong_t)(scn ? scn->scn_phys.scn_errors : 0));
2856 		}
2857 
2858 		/*
2859 		 * If we've completed a scrub/resilver or a rebuild cleanly
2860 		 * then determine if this vdev should remove any DTLs. We
2861 		 * only want to excise regions on vdevs that were available
2862 		 * during the entire duration of this scan.
2863 		 */
2864 		if (rebuild_done &&
2865 		    vr != NULL && vr->vr_rebuild_phys.vrp_errors == 0) {
2866 			check_excise = B_TRUE;
2867 		} else {
2868 			if (spa->spa_scrub_started ||
2869 			    (scn != NULL && scn->scn_phys.scn_errors == 0)) {
2870 				check_excise = B_TRUE;
2871 			}
2872 		}
2873 
2874 		if (scrub_txg && check_excise &&
2875 		    vdev_dtl_should_excise(vd, rebuild_done)) {
2876 			/*
2877 			 * We completed a scrub, resilver or rebuild up to
2878 			 * scrub_txg.  If we did it without rebooting, then
2879 			 * the scrub dtl will be valid, so excise the old
2880 			 * region and fold in the scrub dtl.  Otherwise,
2881 			 * leave the dtl as-is if there was an error.
2882 			 *
2883 			 * There's little trick here: to excise the beginning
2884 			 * of the DTL_MISSING map, we put it into a reference
2885 			 * tree and then add a segment with refcnt -1 that
2886 			 * covers the range [0, scrub_txg).  This means
2887 			 * that each txg in that range has refcnt -1 or 0.
2888 			 * We then add DTL_SCRUB with a refcnt of 2, so that
2889 			 * entries in the range [0, scrub_txg) will have a
2890 			 * positive refcnt -- either 1 or 2.  We then convert
2891 			 * the reference tree into the new DTL_MISSING map.
2892 			 */
2893 			space_reftree_create(&reftree);
2894 			space_reftree_add_map(&reftree,
2895 			    vd->vdev_dtl[DTL_MISSING], 1);
2896 			space_reftree_add_seg(&reftree, 0, scrub_txg, -1);
2897 			space_reftree_add_map(&reftree,
2898 			    vd->vdev_dtl[DTL_SCRUB], 2);
2899 			space_reftree_generate_map(&reftree,
2900 			    vd->vdev_dtl[DTL_MISSING], 1);
2901 			space_reftree_destroy(&reftree);
2902 
2903 			if (!range_tree_is_empty(vd->vdev_dtl[DTL_MISSING])) {
2904 				zfs_dbgmsg("update DTL_MISSING:%llu/%llu",
2905 				    (u_longlong_t)vdev_dtl_min(vd),
2906 				    (u_longlong_t)vdev_dtl_max(vd));
2907 			} else if (!wasempty) {
2908 				zfs_dbgmsg("DTL_MISSING is now empty");
2909 			}
2910 		}
2911 		range_tree_vacate(vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
2912 		range_tree_walk(vd->vdev_dtl[DTL_MISSING],
2913 		    range_tree_add, vd->vdev_dtl[DTL_PARTIAL]);
2914 		if (scrub_done)
2915 			range_tree_vacate(vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
2916 		range_tree_vacate(vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
2917 		if (!vdev_readable(vd))
2918 			range_tree_add(vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
2919 		else
2920 			range_tree_walk(vd->vdev_dtl[DTL_MISSING],
2921 			    range_tree_add, vd->vdev_dtl[DTL_OUTAGE]);
2922 
2923 		/*
2924 		 * If the vdev was resilvering or rebuilding and no longer
2925 		 * has any DTLs then reset the appropriate flag and dirty
2926 		 * the top level so that we persist the change.
2927 		 */
2928 		if (txg != 0 &&
2929 		    range_tree_is_empty(vd->vdev_dtl[DTL_MISSING]) &&
2930 		    range_tree_is_empty(vd->vdev_dtl[DTL_OUTAGE])) {
2931 			if (vd->vdev_rebuild_txg != 0) {
2932 				vd->vdev_rebuild_txg = 0;
2933 				vdev_config_dirty(vd->vdev_top);
2934 			} else if (vd->vdev_resilver_txg != 0) {
2935 				vd->vdev_resilver_txg = 0;
2936 				vdev_config_dirty(vd->vdev_top);
2937 			}
2938 		}
2939 
2940 		mutex_exit(&vd->vdev_dtl_lock);
2941 
2942 		if (txg != 0)
2943 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
2944 		return;
2945 	}
2946 
2947 	mutex_enter(&vd->vdev_dtl_lock);
2948 	for (int t = 0; t < DTL_TYPES; t++) {
2949 		/* account for child's outage in parent's missing map */
2950 		int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
2951 		if (t == DTL_SCRUB)
2952 			continue;			/* leaf vdevs only */
2953 		if (t == DTL_PARTIAL)
2954 			minref = 1;			/* i.e. non-zero */
2955 		else if (vdev_get_nparity(vd) != 0)
2956 			minref = vdev_get_nparity(vd) + 1; /* RAID-Z, dRAID */
2957 		else
2958 			minref = vd->vdev_children;	/* any kind of mirror */
2959 		space_reftree_create(&reftree);
2960 		for (int c = 0; c < vd->vdev_children; c++) {
2961 			vdev_t *cvd = vd->vdev_child[c];
2962 			mutex_enter(&cvd->vdev_dtl_lock);
2963 			space_reftree_add_map(&reftree, cvd->vdev_dtl[s], 1);
2964 			mutex_exit(&cvd->vdev_dtl_lock);
2965 		}
2966 		space_reftree_generate_map(&reftree, vd->vdev_dtl[t], minref);
2967 		space_reftree_destroy(&reftree);
2968 	}
2969 	mutex_exit(&vd->vdev_dtl_lock);
2970 }
2971 
2972 int
2973 vdev_dtl_load(vdev_t *vd)
2974 {
2975 	spa_t *spa = vd->vdev_spa;
2976 	objset_t *mos = spa->spa_meta_objset;
2977 	range_tree_t *rt;
2978 	int error = 0;
2979 
2980 	if (vd->vdev_ops->vdev_op_leaf && vd->vdev_dtl_object != 0) {
2981 		ASSERT(vdev_is_concrete(vd));
2982 
2983 		error = space_map_open(&vd->vdev_dtl_sm, mos,
2984 		    vd->vdev_dtl_object, 0, -1ULL, 0);
2985 		if (error)
2986 			return (error);
2987 		ASSERT(vd->vdev_dtl_sm != NULL);
2988 
2989 		rt = range_tree_create(NULL, RANGE_SEG64, NULL, 0, 0);
2990 		error = space_map_load(vd->vdev_dtl_sm, rt, SM_ALLOC);
2991 		if (error == 0) {
2992 			mutex_enter(&vd->vdev_dtl_lock);
2993 			range_tree_walk(rt, range_tree_add,
2994 			    vd->vdev_dtl[DTL_MISSING]);
2995 			mutex_exit(&vd->vdev_dtl_lock);
2996 		}
2997 
2998 		range_tree_vacate(rt, NULL, NULL);
2999 		range_tree_destroy(rt);
3000 
3001 		return (error);
3002 	}
3003 
3004 	for (int c = 0; c < vd->vdev_children; c++) {
3005 		error = vdev_dtl_load(vd->vdev_child[c]);
3006 		if (error != 0)
3007 			break;
3008 	}
3009 
3010 	return (error);
3011 }
3012 
3013 static void
3014 vdev_zap_allocation_data(vdev_t *vd, dmu_tx_t *tx)
3015 {
3016 	spa_t *spa = vd->vdev_spa;
3017 	objset_t *mos = spa->spa_meta_objset;
3018 	vdev_alloc_bias_t alloc_bias = vd->vdev_alloc_bias;
3019 	const char *string;
3020 
3021 	ASSERT(alloc_bias != VDEV_BIAS_NONE);
3022 
3023 	string =
3024 	    (alloc_bias == VDEV_BIAS_LOG) ? VDEV_ALLOC_BIAS_LOG :
3025 	    (alloc_bias == VDEV_BIAS_SPECIAL) ? VDEV_ALLOC_BIAS_SPECIAL :
3026 	    (alloc_bias == VDEV_BIAS_DEDUP) ? VDEV_ALLOC_BIAS_DEDUP : NULL;
3027 
3028 	ASSERT(string != NULL);
3029 	VERIFY0(zap_add(mos, vd->vdev_top_zap, VDEV_TOP_ZAP_ALLOCATION_BIAS,
3030 	    1, strlen(string) + 1, string, tx));
3031 
3032 	if (alloc_bias == VDEV_BIAS_SPECIAL || alloc_bias == VDEV_BIAS_DEDUP) {
3033 		spa_activate_allocation_classes(spa, tx);
3034 	}
3035 }
3036 
3037 void
3038 vdev_destroy_unlink_zap(vdev_t *vd, uint64_t zapobj, dmu_tx_t *tx)
3039 {
3040 	spa_t *spa = vd->vdev_spa;
3041 
3042 	VERIFY0(zap_destroy(spa->spa_meta_objset, zapobj, tx));
3043 	VERIFY0(zap_remove_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
3044 	    zapobj, tx));
3045 }
3046 
3047 uint64_t
3048 vdev_create_link_zap(vdev_t *vd, dmu_tx_t *tx)
3049 {
3050 	spa_t *spa = vd->vdev_spa;
3051 	uint64_t zap = zap_create(spa->spa_meta_objset, DMU_OTN_ZAP_METADATA,
3052 	    DMU_OT_NONE, 0, tx);
3053 
3054 	ASSERT(zap != 0);
3055 	VERIFY0(zap_add_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
3056 	    zap, tx));
3057 
3058 	return (zap);
3059 }
3060 
3061 void
3062 vdev_construct_zaps(vdev_t *vd, dmu_tx_t *tx)
3063 {
3064 	if (vd->vdev_ops != &vdev_hole_ops &&
3065 	    vd->vdev_ops != &vdev_missing_ops &&
3066 	    vd->vdev_ops != &vdev_root_ops &&
3067 	    !vd->vdev_top->vdev_removing) {
3068 		if (vd->vdev_ops->vdev_op_leaf && vd->vdev_leaf_zap == 0) {
3069 			vd->vdev_leaf_zap = vdev_create_link_zap(vd, tx);
3070 		}
3071 		if (vd == vd->vdev_top && vd->vdev_top_zap == 0) {
3072 			vd->vdev_top_zap = vdev_create_link_zap(vd, tx);
3073 			if (vd->vdev_alloc_bias != VDEV_BIAS_NONE)
3074 				vdev_zap_allocation_data(vd, tx);
3075 		}
3076 	}
3077 
3078 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
3079 		vdev_construct_zaps(vd->vdev_child[i], tx);
3080 	}
3081 }
3082 
3083 static void
3084 vdev_dtl_sync(vdev_t *vd, uint64_t txg)
3085 {
3086 	spa_t *spa = vd->vdev_spa;
3087 	range_tree_t *rt = vd->vdev_dtl[DTL_MISSING];
3088 	objset_t *mos = spa->spa_meta_objset;
3089 	range_tree_t *rtsync;
3090 	dmu_tx_t *tx;
3091 	uint64_t object = space_map_object(vd->vdev_dtl_sm);
3092 
3093 	ASSERT(vdev_is_concrete(vd));
3094 	ASSERT(vd->vdev_ops->vdev_op_leaf);
3095 
3096 	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
3097 
3098 	if (vd->vdev_detached || vd->vdev_top->vdev_removing) {
3099 		mutex_enter(&vd->vdev_dtl_lock);
3100 		space_map_free(vd->vdev_dtl_sm, tx);
3101 		space_map_close(vd->vdev_dtl_sm);
3102 		vd->vdev_dtl_sm = NULL;
3103 		mutex_exit(&vd->vdev_dtl_lock);
3104 
3105 		/*
3106 		 * We only destroy the leaf ZAP for detached leaves or for
3107 		 * removed log devices. Removed data devices handle leaf ZAP
3108 		 * cleanup later, once cancellation is no longer possible.
3109 		 */
3110 		if (vd->vdev_leaf_zap != 0 && (vd->vdev_detached ||
3111 		    vd->vdev_top->vdev_islog)) {
3112 			vdev_destroy_unlink_zap(vd, vd->vdev_leaf_zap, tx);
3113 			vd->vdev_leaf_zap = 0;
3114 		}
3115 
3116 		dmu_tx_commit(tx);
3117 		return;
3118 	}
3119 
3120 	if (vd->vdev_dtl_sm == NULL) {
3121 		uint64_t new_object;
3122 
3123 		new_object = space_map_alloc(mos, zfs_vdev_dtl_sm_blksz, tx);
3124 		VERIFY3U(new_object, !=, 0);
3125 
3126 		VERIFY0(space_map_open(&vd->vdev_dtl_sm, mos, new_object,
3127 		    0, -1ULL, 0));
3128 		ASSERT(vd->vdev_dtl_sm != NULL);
3129 	}
3130 
3131 	rtsync = range_tree_create(NULL, RANGE_SEG64, NULL, 0, 0);
3132 
3133 	mutex_enter(&vd->vdev_dtl_lock);
3134 	range_tree_walk(rt, range_tree_add, rtsync);
3135 	mutex_exit(&vd->vdev_dtl_lock);
3136 
3137 	space_map_truncate(vd->vdev_dtl_sm, zfs_vdev_dtl_sm_blksz, tx);
3138 	space_map_write(vd->vdev_dtl_sm, rtsync, SM_ALLOC, SM_NO_VDEVID, tx);
3139 	range_tree_vacate(rtsync, NULL, NULL);
3140 
3141 	range_tree_destroy(rtsync);
3142 
3143 	/*
3144 	 * If the object for the space map has changed then dirty
3145 	 * the top level so that we update the config.
3146 	 */
3147 	if (object != space_map_object(vd->vdev_dtl_sm)) {
3148 		vdev_dbgmsg(vd, "txg %llu, spa %s, DTL old object %llu, "
3149 		    "new object %llu", (u_longlong_t)txg, spa_name(spa),
3150 		    (u_longlong_t)object,
3151 		    (u_longlong_t)space_map_object(vd->vdev_dtl_sm));
3152 		vdev_config_dirty(vd->vdev_top);
3153 	}
3154 
3155 	dmu_tx_commit(tx);
3156 }
3157 
3158 /*
3159  * Determine whether the specified vdev can be offlined/detached/removed
3160  * without losing data.
3161  */
3162 boolean_t
3163 vdev_dtl_required(vdev_t *vd)
3164 {
3165 	spa_t *spa = vd->vdev_spa;
3166 	vdev_t *tvd = vd->vdev_top;
3167 	uint8_t cant_read = vd->vdev_cant_read;
3168 	boolean_t required;
3169 
3170 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3171 
3172 	if (vd == spa->spa_root_vdev || vd == tvd)
3173 		return (B_TRUE);
3174 
3175 	/*
3176 	 * Temporarily mark the device as unreadable, and then determine
3177 	 * whether this results in any DTL outages in the top-level vdev.
3178 	 * If not, we can safely offline/detach/remove the device.
3179 	 */
3180 	vd->vdev_cant_read = B_TRUE;
3181 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE, B_FALSE);
3182 	required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
3183 	vd->vdev_cant_read = cant_read;
3184 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE, B_FALSE);
3185 
3186 	if (!required && zio_injection_enabled) {
3187 		required = !!zio_handle_device_injection(vd, NULL,
3188 		    SET_ERROR(ECHILD));
3189 	}
3190 
3191 	return (required);
3192 }
3193 
3194 /*
3195  * Determine if resilver is needed, and if so the txg range.
3196  */
3197 boolean_t
3198 vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
3199 {
3200 	boolean_t needed = B_FALSE;
3201 	uint64_t thismin = UINT64_MAX;
3202 	uint64_t thismax = 0;
3203 
3204 	if (vd->vdev_children == 0) {
3205 		mutex_enter(&vd->vdev_dtl_lock);
3206 		if (!range_tree_is_empty(vd->vdev_dtl[DTL_MISSING]) &&
3207 		    vdev_writeable(vd)) {
3208 
3209 			thismin = vdev_dtl_min(vd);
3210 			thismax = vdev_dtl_max(vd);
3211 			needed = B_TRUE;
3212 		}
3213 		mutex_exit(&vd->vdev_dtl_lock);
3214 	} else {
3215 		for (int c = 0; c < vd->vdev_children; c++) {
3216 			vdev_t *cvd = vd->vdev_child[c];
3217 			uint64_t cmin, cmax;
3218 
3219 			if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
3220 				thismin = MIN(thismin, cmin);
3221 				thismax = MAX(thismax, cmax);
3222 				needed = B_TRUE;
3223 			}
3224 		}
3225 	}
3226 
3227 	if (needed && minp) {
3228 		*minp = thismin;
3229 		*maxp = thismax;
3230 	}
3231 	return (needed);
3232 }
3233 
3234 /*
3235  * Gets the checkpoint space map object from the vdev's ZAP.  On success sm_obj
3236  * will contain either the checkpoint spacemap object or zero if none exists.
3237  * All other errors are returned to the caller.
3238  */
3239 int
3240 vdev_checkpoint_sm_object(vdev_t *vd, uint64_t *sm_obj)
3241 {
3242 	ASSERT0(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER));
3243 
3244 	if (vd->vdev_top_zap == 0) {
3245 		*sm_obj = 0;
3246 		return (0);
3247 	}
3248 
3249 	int error = zap_lookup(spa_meta_objset(vd->vdev_spa), vd->vdev_top_zap,
3250 	    VDEV_TOP_ZAP_POOL_CHECKPOINT_SM, sizeof (uint64_t), 1, sm_obj);
3251 	if (error == ENOENT) {
3252 		*sm_obj = 0;
3253 		error = 0;
3254 	}
3255 
3256 	return (error);
3257 }
3258 
3259 int
3260 vdev_load(vdev_t *vd)
3261 {
3262 	int error = 0;
3263 
3264 	/*
3265 	 * Recursively load all children.
3266 	 */
3267 	for (int c = 0; c < vd->vdev_children; c++) {
3268 		error = vdev_load(vd->vdev_child[c]);
3269 		if (error != 0) {
3270 			return (error);
3271 		}
3272 	}
3273 
3274 	vdev_set_deflate_ratio(vd);
3275 
3276 	/*
3277 	 * On spa_load path, grab the allocation bias from our zap
3278 	 */
3279 	if (vd == vd->vdev_top && vd->vdev_top_zap != 0) {
3280 		spa_t *spa = vd->vdev_spa;
3281 		char bias_str[64];
3282 
3283 		error = zap_lookup(spa->spa_meta_objset, vd->vdev_top_zap,
3284 		    VDEV_TOP_ZAP_ALLOCATION_BIAS, 1, sizeof (bias_str),
3285 		    bias_str);
3286 		if (error == 0) {
3287 			ASSERT(vd->vdev_alloc_bias == VDEV_BIAS_NONE);
3288 			vd->vdev_alloc_bias = vdev_derive_alloc_bias(bias_str);
3289 		} else if (error != ENOENT) {
3290 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
3291 			    VDEV_AUX_CORRUPT_DATA);
3292 			vdev_dbgmsg(vd, "vdev_load: zap_lookup(top_zap=%llu) "
3293 			    "failed [error=%d]", vd->vdev_top_zap, error);
3294 			return (error);
3295 		}
3296 	}
3297 
3298 	/*
3299 	 * Load any rebuild state from the top-level vdev zap.
3300 	 */
3301 	if (vd == vd->vdev_top && vd->vdev_top_zap != 0) {
3302 		error = vdev_rebuild_load(vd);
3303 		if (error && error != ENOTSUP) {
3304 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
3305 			    VDEV_AUX_CORRUPT_DATA);
3306 			vdev_dbgmsg(vd, "vdev_load: vdev_rebuild_load "
3307 			    "failed [error=%d]", error);
3308 			return (error);
3309 		}
3310 	}
3311 
3312 	/*
3313 	 * If this is a top-level vdev, initialize its metaslabs.
3314 	 */
3315 	if (vd == vd->vdev_top && vdev_is_concrete(vd)) {
3316 		vdev_metaslab_group_create(vd);
3317 
3318 		if (vd->vdev_ashift == 0 || vd->vdev_asize == 0) {
3319 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
3320 			    VDEV_AUX_CORRUPT_DATA);
3321 			vdev_dbgmsg(vd, "vdev_load: invalid size. ashift=%llu, "
3322 			    "asize=%llu", (u_longlong_t)vd->vdev_ashift,
3323 			    (u_longlong_t)vd->vdev_asize);
3324 			return (SET_ERROR(ENXIO));
3325 		}
3326 
3327 		error = vdev_metaslab_init(vd, 0);
3328 		if (error != 0) {
3329 			vdev_dbgmsg(vd, "vdev_load: metaslab_init failed "
3330 			    "[error=%d]", error);
3331 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
3332 			    VDEV_AUX_CORRUPT_DATA);
3333 			return (error);
3334 		}
3335 
3336 		uint64_t checkpoint_sm_obj;
3337 		error = vdev_checkpoint_sm_object(vd, &checkpoint_sm_obj);
3338 		if (error == 0 && checkpoint_sm_obj != 0) {
3339 			objset_t *mos = spa_meta_objset(vd->vdev_spa);
3340 			ASSERT(vd->vdev_asize != 0);
3341 			ASSERT3P(vd->vdev_checkpoint_sm, ==, NULL);
3342 
3343 			error = space_map_open(&vd->vdev_checkpoint_sm,
3344 			    mos, checkpoint_sm_obj, 0, vd->vdev_asize,
3345 			    vd->vdev_ashift);
3346 			if (error != 0) {
3347 				vdev_dbgmsg(vd, "vdev_load: space_map_open "
3348 				    "failed for checkpoint spacemap (obj %llu) "
3349 				    "[error=%d]",
3350 				    (u_longlong_t)checkpoint_sm_obj, error);
3351 				return (error);
3352 			}
3353 			ASSERT3P(vd->vdev_checkpoint_sm, !=, NULL);
3354 
3355 			/*
3356 			 * Since the checkpoint_sm contains free entries
3357 			 * exclusively we can use space_map_allocated() to
3358 			 * indicate the cumulative checkpointed space that
3359 			 * has been freed.
3360 			 */
3361 			vd->vdev_stat.vs_checkpoint_space =
3362 			    -space_map_allocated(vd->vdev_checkpoint_sm);
3363 			vd->vdev_spa->spa_checkpoint_info.sci_dspace +=
3364 			    vd->vdev_stat.vs_checkpoint_space;
3365 		} else if (error != 0) {
3366 			vdev_dbgmsg(vd, "vdev_load: failed to retrieve "
3367 			    "checkpoint space map object from vdev ZAP "
3368 			    "[error=%d]", error);
3369 			return (error);
3370 		}
3371 	}
3372 
3373 	/*
3374 	 * If this is a leaf vdev, load its DTL.
3375 	 */
3376 	if (vd->vdev_ops->vdev_op_leaf && (error = vdev_dtl_load(vd)) != 0) {
3377 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
3378 		    VDEV_AUX_CORRUPT_DATA);
3379 		vdev_dbgmsg(vd, "vdev_load: vdev_dtl_load failed "
3380 		    "[error=%d]", error);
3381 		return (error);
3382 	}
3383 
3384 	uint64_t obsolete_sm_object;
3385 	error = vdev_obsolete_sm_object(vd, &obsolete_sm_object);
3386 	if (error == 0 && obsolete_sm_object != 0) {
3387 		objset_t *mos = vd->vdev_spa->spa_meta_objset;
3388 		ASSERT(vd->vdev_asize != 0);
3389 		ASSERT3P(vd->vdev_obsolete_sm, ==, NULL);
3390 
3391 		if ((error = space_map_open(&vd->vdev_obsolete_sm, mos,
3392 		    obsolete_sm_object, 0, vd->vdev_asize, 0))) {
3393 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
3394 			    VDEV_AUX_CORRUPT_DATA);
3395 			vdev_dbgmsg(vd, "vdev_load: space_map_open failed for "
3396 			    "obsolete spacemap (obj %llu) [error=%d]",
3397 			    (u_longlong_t)obsolete_sm_object, error);
3398 			return (error);
3399 		}
3400 	} else if (error != 0) {
3401 		vdev_dbgmsg(vd, "vdev_load: failed to retrieve obsolete "
3402 		    "space map object from vdev ZAP [error=%d]", error);
3403 		return (error);
3404 	}
3405 
3406 	return (0);
3407 }
3408 
3409 /*
3410  * The special vdev case is used for hot spares and l2cache devices.  Its
3411  * sole purpose it to set the vdev state for the associated vdev.  To do this,
3412  * we make sure that we can open the underlying device, then try to read the
3413  * label, and make sure that the label is sane and that it hasn't been
3414  * repurposed to another pool.
3415  */
3416 int
3417 vdev_validate_aux(vdev_t *vd)
3418 {
3419 	nvlist_t *label;
3420 	uint64_t guid, version;
3421 	uint64_t state;
3422 
3423 	if (!vdev_readable(vd))
3424 		return (0);
3425 
3426 	if ((label = vdev_label_read_config(vd, -1ULL)) == NULL) {
3427 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
3428 		    VDEV_AUX_CORRUPT_DATA);
3429 		return (-1);
3430 	}
3431 
3432 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
3433 	    !SPA_VERSION_IS_SUPPORTED(version) ||
3434 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
3435 	    guid != vd->vdev_guid ||
3436 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
3437 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
3438 		    VDEV_AUX_CORRUPT_DATA);
3439 		nvlist_free(label);
3440 		return (-1);
3441 	}
3442 
3443 	/*
3444 	 * We don't actually check the pool state here.  If it's in fact in
3445 	 * use by another pool, we update this fact on the fly when requested.
3446 	 */
3447 	nvlist_free(label);
3448 	return (0);
3449 }
3450 
3451 static void
3452 vdev_destroy_ms_flush_data(vdev_t *vd, dmu_tx_t *tx)
3453 {
3454 	objset_t *mos = spa_meta_objset(vd->vdev_spa);
3455 
3456 	if (vd->vdev_top_zap == 0)
3457 		return;
3458 
3459 	uint64_t object = 0;
3460 	int err = zap_lookup(mos, vd->vdev_top_zap,
3461 	    VDEV_TOP_ZAP_MS_UNFLUSHED_PHYS_TXGS, sizeof (uint64_t), 1, &object);
3462 	if (err == ENOENT)
3463 		return;
3464 	VERIFY0(err);
3465 
3466 	VERIFY0(dmu_object_free(mos, object, tx));
3467 	VERIFY0(zap_remove(mos, vd->vdev_top_zap,
3468 	    VDEV_TOP_ZAP_MS_UNFLUSHED_PHYS_TXGS, tx));
3469 }
3470 
3471 /*
3472  * Free the objects used to store this vdev's spacemaps, and the array
3473  * that points to them.
3474  */
3475 void
3476 vdev_destroy_spacemaps(vdev_t *vd, dmu_tx_t *tx)
3477 {
3478 	if (vd->vdev_ms_array == 0)
3479 		return;
3480 
3481 	objset_t *mos = vd->vdev_spa->spa_meta_objset;
3482 	uint64_t array_count = vd->vdev_asize >> vd->vdev_ms_shift;
3483 	size_t array_bytes = array_count * sizeof (uint64_t);
3484 	uint64_t *smobj_array = kmem_alloc(array_bytes, KM_SLEEP);
3485 	VERIFY0(dmu_read(mos, vd->vdev_ms_array, 0,
3486 	    array_bytes, smobj_array, 0));
3487 
3488 	for (uint64_t i = 0; i < array_count; i++) {
3489 		uint64_t smobj = smobj_array[i];
3490 		if (smobj == 0)
3491 			continue;
3492 
3493 		space_map_free_obj(mos, smobj, tx);
3494 	}
3495 
3496 	kmem_free(smobj_array, array_bytes);
3497 	VERIFY0(dmu_object_free(mos, vd->vdev_ms_array, tx));
3498 	vdev_destroy_ms_flush_data(vd, tx);
3499 	vd->vdev_ms_array = 0;
3500 }
3501 
3502 static void
3503 vdev_remove_empty_log(vdev_t *vd, uint64_t txg)
3504 {
3505 	spa_t *spa = vd->vdev_spa;
3506 
3507 	ASSERT(vd->vdev_islog);
3508 	ASSERT(vd == vd->vdev_top);
3509 	ASSERT3U(txg, ==, spa_syncing_txg(spa));
3510 
3511 	dmu_tx_t *tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
3512 
3513 	vdev_destroy_spacemaps(vd, tx);
3514 	if (vd->vdev_top_zap != 0) {
3515 		vdev_destroy_unlink_zap(vd, vd->vdev_top_zap, tx);
3516 		vd->vdev_top_zap = 0;
3517 	}
3518 
3519 	dmu_tx_commit(tx);
3520 }
3521 
3522 void
3523 vdev_sync_done(vdev_t *vd, uint64_t txg)
3524 {
3525 	metaslab_t *msp;
3526 	boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg));
3527 
3528 	ASSERT(vdev_is_concrete(vd));
3529 
3530 	while ((msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
3531 	    != NULL)
3532 		metaslab_sync_done(msp, txg);
3533 
3534 	if (reassess)
3535 		metaslab_sync_reassess(vd->vdev_mg);
3536 }
3537 
3538 void
3539 vdev_sync(vdev_t *vd, uint64_t txg)
3540 {
3541 	spa_t *spa = vd->vdev_spa;
3542 	vdev_t *lvd;
3543 	metaslab_t *msp;
3544 
3545 	ASSERT3U(txg, ==, spa->spa_syncing_txg);
3546 	dmu_tx_t *tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
3547 	if (range_tree_space(vd->vdev_obsolete_segments) > 0) {
3548 		ASSERT(vd->vdev_removing ||
3549 		    vd->vdev_ops == &vdev_indirect_ops);
3550 
3551 		vdev_indirect_sync_obsolete(vd, tx);
3552 
3553 		/*
3554 		 * If the vdev is indirect, it can't have dirty
3555 		 * metaslabs or DTLs.
3556 		 */
3557 		if (vd->vdev_ops == &vdev_indirect_ops) {
3558 			ASSERT(txg_list_empty(&vd->vdev_ms_list, txg));
3559 			ASSERT(txg_list_empty(&vd->vdev_dtl_list, txg));
3560 			dmu_tx_commit(tx);
3561 			return;
3562 		}
3563 	}
3564 
3565 	ASSERT(vdev_is_concrete(vd));
3566 
3567 	if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0 &&
3568 	    !vd->vdev_removing) {
3569 		ASSERT(vd == vd->vdev_top);
3570 		ASSERT0(vd->vdev_indirect_config.vic_mapping_object);
3571 		vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
3572 		    DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
3573 		ASSERT(vd->vdev_ms_array != 0);
3574 		vdev_config_dirty(vd);
3575 	}
3576 
3577 	while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
3578 		metaslab_sync(msp, txg);
3579 		(void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
3580 	}
3581 
3582 	while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
3583 		vdev_dtl_sync(lvd, txg);
3584 
3585 	/*
3586 	 * If this is an empty log device being removed, destroy the
3587 	 * metadata associated with it.
3588 	 */
3589 	if (vd->vdev_islog && vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing)
3590 		vdev_remove_empty_log(vd, txg);
3591 
3592 	(void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
3593 	dmu_tx_commit(tx);
3594 }
3595 
3596 uint64_t
3597 vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
3598 {
3599 	return (vd->vdev_ops->vdev_op_asize(vd, psize));
3600 }
3601 
3602 /*
3603  * Mark the given vdev faulted.  A faulted vdev behaves as if the device could
3604  * not be opened, and no I/O is attempted.
3605  */
3606 int
3607 vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
3608 {
3609 	vdev_t *vd, *tvd;
3610 
3611 	spa_vdev_state_enter(spa, SCL_NONE);
3612 
3613 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
3614 		return (spa_vdev_state_exit(spa, NULL, SET_ERROR(ENODEV)));
3615 
3616 	if (!vd->vdev_ops->vdev_op_leaf)
3617 		return (spa_vdev_state_exit(spa, NULL, SET_ERROR(ENOTSUP)));
3618 
3619 	tvd = vd->vdev_top;
3620 
3621 	/*
3622 	 * If user did a 'zpool offline -f' then make the fault persist across
3623 	 * reboots.
3624 	 */
3625 	if (aux == VDEV_AUX_EXTERNAL_PERSIST) {
3626 		/*
3627 		 * There are two kinds of forced faults: temporary and
3628 		 * persistent.  Temporary faults go away at pool import, while
3629 		 * persistent faults stay set.  Both types of faults can be
3630 		 * cleared with a zpool clear.
3631 		 *
3632 		 * We tell if a vdev is persistently faulted by looking at the
3633 		 * ZPOOL_CONFIG_AUX_STATE nvpair.  If it's set to "external" at
3634 		 * import then it's a persistent fault.  Otherwise, it's
3635 		 * temporary.  We get ZPOOL_CONFIG_AUX_STATE set to "external"
3636 		 * by setting vd.vdev_stat.vs_aux to VDEV_AUX_EXTERNAL.  This
3637 		 * tells vdev_config_generate() (which gets run later) to set
3638 		 * ZPOOL_CONFIG_AUX_STATE to "external" in the nvlist.
3639 		 */
3640 		vd->vdev_stat.vs_aux = VDEV_AUX_EXTERNAL;
3641 		vd->vdev_tmpoffline = B_FALSE;
3642 		aux = VDEV_AUX_EXTERNAL;
3643 	} else {
3644 		vd->vdev_tmpoffline = B_TRUE;
3645 	}
3646 
3647 	/*
3648 	 * We don't directly use the aux state here, but if we do a
3649 	 * vdev_reopen(), we need this value to be present to remember why we
3650 	 * were faulted.
3651 	 */
3652 	vd->vdev_label_aux = aux;
3653 
3654 	/*
3655 	 * Faulted state takes precedence over degraded.
3656 	 */
3657 	vd->vdev_delayed_close = B_FALSE;
3658 	vd->vdev_faulted = 1ULL;
3659 	vd->vdev_degraded = 0ULL;
3660 	vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
3661 
3662 	/*
3663 	 * If this device has the only valid copy of the data, then
3664 	 * back off and simply mark the vdev as degraded instead.
3665 	 */
3666 	if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) {
3667 		vd->vdev_degraded = 1ULL;
3668 		vd->vdev_faulted = 0ULL;
3669 
3670 		/*
3671 		 * If we reopen the device and it's not dead, only then do we
3672 		 * mark it degraded.
3673 		 */
3674 		vdev_reopen(tvd);
3675 
3676 		if (vdev_readable(vd))
3677 			vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
3678 	}
3679 
3680 	return (spa_vdev_state_exit(spa, vd, 0));
3681 }
3682 
3683 /*
3684  * Mark the given vdev degraded.  A degraded vdev is purely an indication to the
3685  * user that something is wrong.  The vdev continues to operate as normal as far
3686  * as I/O is concerned.
3687  */
3688 int
3689 vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
3690 {
3691 	vdev_t *vd;
3692 
3693 	spa_vdev_state_enter(spa, SCL_NONE);
3694 
3695 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
3696 		return (spa_vdev_state_exit(spa, NULL, SET_ERROR(ENODEV)));
3697 
3698 	if (!vd->vdev_ops->vdev_op_leaf)
3699 		return (spa_vdev_state_exit(spa, NULL, SET_ERROR(ENOTSUP)));
3700 
3701 	/*
3702 	 * If the vdev is already faulted, then don't do anything.
3703 	 */
3704 	if (vd->vdev_faulted || vd->vdev_degraded)
3705 		return (spa_vdev_state_exit(spa, NULL, 0));
3706 
3707 	vd->vdev_degraded = 1ULL;
3708 	if (!vdev_is_dead(vd))
3709 		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
3710 		    aux);
3711 
3712 	return (spa_vdev_state_exit(spa, vd, 0));
3713 }
3714 
3715 /*
3716  * Online the given vdev.
3717  *
3718  * If 'ZFS_ONLINE_UNSPARE' is set, it implies two things.  First, any attached
3719  * spare device should be detached when the device finishes resilvering.
3720  * Second, the online should be treated like a 'test' online case, so no FMA
3721  * events are generated if the device fails to open.
3722  */
3723 int
3724 vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
3725 {
3726 	vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
3727 	boolean_t wasoffline;
3728 	vdev_state_t oldstate;
3729 
3730 	spa_vdev_state_enter(spa, SCL_NONE);
3731 
3732 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
3733 		return (spa_vdev_state_exit(spa, NULL, SET_ERROR(ENODEV)));
3734 
3735 	if (!vd->vdev_ops->vdev_op_leaf)
3736 		return (spa_vdev_state_exit(spa, NULL, SET_ERROR(ENOTSUP)));
3737 
3738 	wasoffline = (vd->vdev_offline || vd->vdev_tmpoffline);
3739 	oldstate = vd->vdev_state;
3740 
3741 	tvd = vd->vdev_top;
3742 	vd->vdev_offline = B_FALSE;
3743 	vd->vdev_tmpoffline = B_FALSE;
3744 	vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
3745 	vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
3746 
3747 	/* XXX - L2ARC 1.0 does not support expansion */
3748 	if (!vd->vdev_aux) {
3749 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
3750 			pvd->vdev_expanding = !!((flags & ZFS_ONLINE_EXPAND) ||
3751 			    spa->spa_autoexpand);
3752 		vd->vdev_expansion_time = gethrestime_sec();
3753 	}
3754 
3755 	vdev_reopen(tvd);
3756 	vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
3757 
3758 	if (!vd->vdev_aux) {
3759 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
3760 			pvd->vdev_expanding = B_FALSE;
3761 	}
3762 
3763 	if (newstate)
3764 		*newstate = vd->vdev_state;
3765 	if ((flags & ZFS_ONLINE_UNSPARE) &&
3766 	    !vdev_is_dead(vd) && vd->vdev_parent &&
3767 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
3768 	    vd->vdev_parent->vdev_child[0] == vd)
3769 		vd->vdev_unspare = B_TRUE;
3770 
3771 	if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
3772 
3773 		/* XXX - L2ARC 1.0 does not support expansion */
3774 		if (vd->vdev_aux)
3775 			return (spa_vdev_state_exit(spa, vd, ENOTSUP));
3776 		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
3777 	}
3778 
3779 	/* Restart initializing if necessary */
3780 	mutex_enter(&vd->vdev_initialize_lock);
3781 	if (vdev_writeable(vd) &&
3782 	    vd->vdev_initialize_thread == NULL &&
3783 	    vd->vdev_initialize_state == VDEV_INITIALIZE_ACTIVE) {
3784 		(void) vdev_initialize(vd);
3785 	}
3786 	mutex_exit(&vd->vdev_initialize_lock);
3787 
3788 	/*
3789 	 * Restart trimming if necessary. We do not restart trimming for cache
3790 	 * devices here. This is triggered by l2arc_rebuild_vdev()
3791 	 * asynchronously for the whole device or in l2arc_evict() as it evicts
3792 	 * space for upcoming writes.
3793 	 */
3794 	mutex_enter(&vd->vdev_trim_lock);
3795 	if (vdev_writeable(vd) && !vd->vdev_isl2cache &&
3796 	    vd->vdev_trim_thread == NULL &&
3797 	    vd->vdev_trim_state == VDEV_TRIM_ACTIVE) {
3798 		(void) vdev_trim(vd, vd->vdev_trim_rate, vd->vdev_trim_partial,
3799 		    vd->vdev_trim_secure);
3800 	}
3801 	mutex_exit(&vd->vdev_trim_lock);
3802 
3803 	if (wasoffline ||
3804 	    (oldstate < VDEV_STATE_DEGRADED &&
3805 	    vd->vdev_state >= VDEV_STATE_DEGRADED))
3806 		spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_ONLINE);
3807 
3808 	return (spa_vdev_state_exit(spa, vd, 0));
3809 }
3810 
3811 static int
3812 vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
3813 {
3814 	vdev_t *vd, *tvd;
3815 	int error = 0;
3816 	uint64_t generation;
3817 	metaslab_group_t *mg;
3818 
3819 top:
3820 	spa_vdev_state_enter(spa, SCL_ALLOC);
3821 
3822 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
3823 		return (spa_vdev_state_exit(spa, NULL, SET_ERROR(ENODEV)));
3824 
3825 	if (!vd->vdev_ops->vdev_op_leaf)
3826 		return (spa_vdev_state_exit(spa, NULL, SET_ERROR(ENOTSUP)));
3827 
3828 	if (vd->vdev_ops == &vdev_draid_spare_ops)
3829 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
3830 
3831 	tvd = vd->vdev_top;
3832 	mg = tvd->vdev_mg;
3833 	generation = spa->spa_config_generation + 1;
3834 
3835 	/*
3836 	 * If the device isn't already offline, try to offline it.
3837 	 */
3838 	if (!vd->vdev_offline) {
3839 		/*
3840 		 * If this device has the only valid copy of some data,
3841 		 * don't allow it to be offlined. Log devices are always
3842 		 * expendable.
3843 		 */
3844 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
3845 		    vdev_dtl_required(vd))
3846 			return (spa_vdev_state_exit(spa, NULL,
3847 			    SET_ERROR(EBUSY)));
3848 
3849 		/*
3850 		 * If the top-level is a slog and it has had allocations
3851 		 * then proceed.  We check that the vdev's metaslab group
3852 		 * is not NULL since it's possible that we may have just
3853 		 * added this vdev but not yet initialized its metaslabs.
3854 		 */
3855 		if (tvd->vdev_islog && mg != NULL) {
3856 			/*
3857 			 * Prevent any future allocations.
3858 			 */
3859 			metaslab_group_passivate(mg);
3860 			(void) spa_vdev_state_exit(spa, vd, 0);
3861 
3862 			error = spa_reset_logs(spa);
3863 
3864 			/*
3865 			 * If the log device was successfully reset but has
3866 			 * checkpointed data, do not offline it.
3867 			 */
3868 			if (error == 0 &&
3869 			    tvd->vdev_checkpoint_sm != NULL) {
3870 				ASSERT3U(space_map_allocated(
3871 				    tvd->vdev_checkpoint_sm), !=, 0);
3872 				error = ZFS_ERR_CHECKPOINT_EXISTS;
3873 			}
3874 
3875 			spa_vdev_state_enter(spa, SCL_ALLOC);
3876 
3877 			/*
3878 			 * Check to see if the config has changed.
3879 			 */
3880 			if (error || generation != spa->spa_config_generation) {
3881 				metaslab_group_activate(mg);
3882 				if (error)
3883 					return (spa_vdev_state_exit(spa,
3884 					    vd, error));
3885 				(void) spa_vdev_state_exit(spa, vd, 0);
3886 				goto top;
3887 			}
3888 			ASSERT0(tvd->vdev_stat.vs_alloc);
3889 		}
3890 
3891 		/*
3892 		 * Offline this device and reopen its top-level vdev.
3893 		 * If the top-level vdev is a log device then just offline
3894 		 * it. Otherwise, if this action results in the top-level
3895 		 * vdev becoming unusable, undo it and fail the request.
3896 		 */
3897 		vd->vdev_offline = B_TRUE;
3898 		vdev_reopen(tvd);
3899 
3900 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
3901 		    vdev_is_dead(tvd)) {
3902 			vd->vdev_offline = B_FALSE;
3903 			vdev_reopen(tvd);
3904 			return (spa_vdev_state_exit(spa, NULL,
3905 			    SET_ERROR(EBUSY)));
3906 		}
3907 
3908 		/*
3909 		 * Add the device back into the metaslab rotor so that
3910 		 * once we online the device it's open for business.
3911 		 */
3912 		if (tvd->vdev_islog && mg != NULL)
3913 			metaslab_group_activate(mg);
3914 	}
3915 
3916 	vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
3917 
3918 	return (spa_vdev_state_exit(spa, vd, 0));
3919 }
3920 
3921 int
3922 vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
3923 {
3924 	int error;
3925 
3926 	mutex_enter(&spa->spa_vdev_top_lock);
3927 	error = vdev_offline_locked(spa, guid, flags);
3928 	mutex_exit(&spa->spa_vdev_top_lock);
3929 
3930 	return (error);
3931 }
3932 
3933 /*
3934  * Clear the error counts associated with this vdev.  Unlike vdev_online() and
3935  * vdev_offline(), we assume the spa config is locked.  We also clear all
3936  * children.  If 'vd' is NULL, then the user wants to clear all vdevs.
3937  */
3938 void
3939 vdev_clear(spa_t *spa, vdev_t *vd)
3940 {
3941 	vdev_t *rvd = spa->spa_root_vdev;
3942 
3943 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3944 
3945 	if (vd == NULL)
3946 		vd = rvd;
3947 
3948 	vd->vdev_stat.vs_read_errors = 0;
3949 	vd->vdev_stat.vs_write_errors = 0;
3950 	vd->vdev_stat.vs_checksum_errors = 0;
3951 	vd->vdev_stat.vs_slow_ios = 0;
3952 
3953 	for (int c = 0; c < vd->vdev_children; c++)
3954 		vdev_clear(spa, vd->vdev_child[c]);
3955 
3956 	/*
3957 	 * It makes no sense to "clear" an indirect vdev.
3958 	 */
3959 	if (!vdev_is_concrete(vd))
3960 		return;
3961 
3962 	/*
3963 	 * If we're in the FAULTED state or have experienced failed I/O, then
3964 	 * clear the persistent state and attempt to reopen the device.  We
3965 	 * also mark the vdev config dirty, so that the new faulted state is
3966 	 * written out to disk.
3967 	 */
3968 	if (vd->vdev_faulted || vd->vdev_degraded ||
3969 	    !vdev_readable(vd) || !vdev_writeable(vd)) {
3970 		/*
3971 		 * When reopening in response to a clear event, it may be due to
3972 		 * a fmadm repair request.  In this case, if the device is
3973 		 * still broken, we want to still post the ereport again.
3974 		 */
3975 		vd->vdev_forcefault = B_TRUE;
3976 
3977 		vd->vdev_faulted = vd->vdev_degraded = 0ULL;
3978 		vd->vdev_cant_read = B_FALSE;
3979 		vd->vdev_cant_write = B_FALSE;
3980 		vd->vdev_stat.vs_aux = 0;
3981 
3982 		vdev_reopen(vd == rvd ? rvd : vd->vdev_top);
3983 
3984 		vd->vdev_forcefault = B_FALSE;
3985 
3986 		if (vd != rvd && vdev_writeable(vd->vdev_top))
3987 			vdev_state_dirty(vd->vdev_top);
3988 
3989 		/* If a resilver isn't required, check if vdevs can be culled */
3990 		if (vd->vdev_aux == NULL && !vdev_is_dead(vd) &&
3991 		    !dsl_scan_resilvering(spa->spa_dsl_pool) &&
3992 		    !dsl_scan_resilver_scheduled(spa->spa_dsl_pool))
3993 			spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
3994 
3995 		spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_CLEAR);
3996 	}
3997 
3998 	/*
3999 	 * When clearing a FMA-diagnosed fault, we always want to
4000 	 * unspare the device, as we assume that the original spare was
4001 	 * done in response to the FMA fault.
4002 	 */
4003 	if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
4004 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
4005 	    vd->vdev_parent->vdev_child[0] == vd)
4006 		vd->vdev_unspare = B_TRUE;
4007 }
4008 
4009 boolean_t
4010 vdev_is_dead(vdev_t *vd)
4011 {
4012 	/*
4013 	 * Holes and missing devices are always considered "dead".
4014 	 * This simplifies the code since we don't have to check for
4015 	 * these types of devices in the various code paths.
4016 	 * Instead we rely on the fact that we skip over dead devices
4017 	 * before issuing I/O to them.
4018 	 */
4019 	return (vd->vdev_state < VDEV_STATE_DEGRADED ||
4020 	    vd->vdev_ops == &vdev_hole_ops ||
4021 	    vd->vdev_ops == &vdev_missing_ops);
4022 }
4023 
4024 boolean_t
4025 vdev_readable(vdev_t *vd)
4026 {
4027 	return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
4028 }
4029 
4030 boolean_t
4031 vdev_writeable(vdev_t *vd)
4032 {
4033 	return (!vdev_is_dead(vd) && !vd->vdev_cant_write &&
4034 	    vdev_is_concrete(vd));
4035 }
4036 
4037 boolean_t
4038 vdev_allocatable(vdev_t *vd)
4039 {
4040 	uint64_t state = vd->vdev_state;
4041 
4042 	/*
4043 	 * We currently allow allocations from vdevs which may be in the
4044 	 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
4045 	 * fails to reopen then we'll catch it later when we're holding
4046 	 * the proper locks.  Note that we have to get the vdev state
4047 	 * in a local variable because although it changes atomically,
4048 	 * we're asking two separate questions about it.
4049 	 */
4050 	return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
4051 	    !vd->vdev_cant_write && vdev_is_concrete(vd) &&
4052 	    vd->vdev_mg->mg_initialized);
4053 }
4054 
4055 boolean_t
4056 vdev_accessible(vdev_t *vd, zio_t *zio)
4057 {
4058 	ASSERT(zio->io_vd == vd);
4059 
4060 	if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
4061 		return (B_FALSE);
4062 
4063 	if (zio->io_type == ZIO_TYPE_READ)
4064 		return (!vd->vdev_cant_read);
4065 
4066 	if (zio->io_type == ZIO_TYPE_WRITE)
4067 		return (!vd->vdev_cant_write);
4068 
4069 	return (B_TRUE);
4070 }
4071 
4072 static void
4073 vdev_get_child_stat(vdev_t *cvd, vdev_stat_t *vs, vdev_stat_t *cvs)
4074 {
4075 	/*
4076 	 * Exclude the dRAID spare when aggregating to avoid double counting
4077 	 * the ops and bytes.  These IOs are counted by the physical leaves.
4078 	 */
4079 	if (cvd->vdev_ops == &vdev_draid_spare_ops)
4080 		return;
4081 
4082 	for (int t = 0; t < VS_ZIO_TYPES; t++) {
4083 		vs->vs_ops[t] += cvs->vs_ops[t];
4084 		vs->vs_bytes[t] += cvs->vs_bytes[t];
4085 	}
4086 
4087 	cvs->vs_scan_removing = cvd->vdev_removing;
4088 }
4089 
4090 /*
4091  * Get extended stats
4092  */
4093 static void
4094 vdev_get_child_stat_ex(vdev_t *cvd, vdev_stat_ex_t *vsx, vdev_stat_ex_t *cvsx)
4095 {
4096 	int t, b;
4097 	for (t = 0; t < ZIO_TYPES; t++) {
4098 		for (b = 0; b < ARRAY_SIZE(vsx->vsx_disk_histo[0]); b++)
4099 			vsx->vsx_disk_histo[t][b] += cvsx->vsx_disk_histo[t][b];
4100 
4101 		for (b = 0; b < ARRAY_SIZE(vsx->vsx_total_histo[0]); b++) {
4102 			vsx->vsx_total_histo[t][b] +=
4103 			    cvsx->vsx_total_histo[t][b];
4104 		}
4105 	}
4106 
4107 	for (t = 0; t < ZIO_PRIORITY_NUM_QUEUEABLE; t++) {
4108 		for (b = 0; b < ARRAY_SIZE(vsx->vsx_queue_histo[0]); b++) {
4109 			vsx->vsx_queue_histo[t][b] +=
4110 			    cvsx->vsx_queue_histo[t][b];
4111 		}
4112 		vsx->vsx_active_queue[t] += cvsx->vsx_active_queue[t];
4113 		vsx->vsx_pend_queue[t] += cvsx->vsx_pend_queue[t];
4114 
4115 		for (b = 0; b < ARRAY_SIZE(vsx->vsx_ind_histo[0]); b++)
4116 			vsx->vsx_ind_histo[t][b] += cvsx->vsx_ind_histo[t][b];
4117 
4118 		for (b = 0; b < ARRAY_SIZE(vsx->vsx_agg_histo[0]); b++)
4119 			vsx->vsx_agg_histo[t][b] += cvsx->vsx_agg_histo[t][b];
4120 	}
4121 
4122 }
4123 
4124 boolean_t
4125 vdev_is_spacemap_addressable(vdev_t *vd)
4126 {
4127 	if (spa_feature_is_active(vd->vdev_spa, SPA_FEATURE_SPACEMAP_V2))
4128 		return (B_TRUE);
4129 
4130 	/*
4131 	 * If double-word space map entries are not enabled we assume
4132 	 * 47 bits of the space map entry are dedicated to the entry's
4133 	 * offset (see SM_OFFSET_BITS in space_map.h). We then use that
4134 	 * to calculate the maximum address that can be described by a
4135 	 * space map entry for the given device.
4136 	 */
4137 	uint64_t shift = vd->vdev_ashift + SM_OFFSET_BITS;
4138 
4139 	if (shift >= 63) /* detect potential overflow */
4140 		return (B_TRUE);
4141 
4142 	return (vd->vdev_asize < (1ULL << shift));
4143 }
4144 
4145 /*
4146  * Get statistics for the given vdev.
4147  */
4148 static void
4149 vdev_get_stats_ex_impl(vdev_t *vd, vdev_stat_t *vs, vdev_stat_ex_t *vsx)
4150 {
4151 	int t;
4152 	/*
4153 	 * If we're getting stats on the root vdev, aggregate the I/O counts
4154 	 * over all top-level vdevs (i.e. the direct children of the root).
4155 	 */
4156 	if (!vd->vdev_ops->vdev_op_leaf) {
4157 		if (vs) {
4158 			memset(vs->vs_ops, 0, sizeof (vs->vs_ops));
4159 			memset(vs->vs_bytes, 0, sizeof (vs->vs_bytes));
4160 		}
4161 		if (vsx)
4162 			memset(vsx, 0, sizeof (*vsx));
4163 
4164 		for (int c = 0; c < vd->vdev_children; c++) {
4165 			vdev_t *cvd = vd->vdev_child[c];
4166 			vdev_stat_t *cvs = &cvd->vdev_stat;
4167 			vdev_stat_ex_t *cvsx = &cvd->vdev_stat_ex;
4168 
4169 			vdev_get_stats_ex_impl(cvd, cvs, cvsx);
4170 			if (vs)
4171 				vdev_get_child_stat(cvd, vs, cvs);
4172 			if (vsx)
4173 				vdev_get_child_stat_ex(cvd, vsx, cvsx);
4174 		}
4175 	} else {
4176 		/*
4177 		 * We're a leaf.  Just copy our ZIO active queue stats in.  The
4178 		 * other leaf stats are updated in vdev_stat_update().
4179 		 */
4180 		if (!vsx)
4181 			return;
4182 
4183 		memcpy(vsx, &vd->vdev_stat_ex, sizeof (vd->vdev_stat_ex));
4184 
4185 		for (t = 0; t < ARRAY_SIZE(vd->vdev_queue.vq_class); t++) {
4186 			vsx->vsx_active_queue[t] =
4187 			    vd->vdev_queue.vq_class[t].vqc_active;
4188 			vsx->vsx_pend_queue[t] = avl_numnodes(
4189 			    &vd->vdev_queue.vq_class[t].vqc_queued_tree);
4190 		}
4191 	}
4192 }
4193 
4194 void
4195 vdev_get_stats_ex(vdev_t *vd, vdev_stat_t *vs, vdev_stat_ex_t *vsx)
4196 {
4197 	vdev_t *tvd = vd->vdev_top;
4198 	mutex_enter(&vd->vdev_stat_lock);
4199 	if (vs) {
4200 		bcopy(&vd->vdev_stat, vs, sizeof (*vs));
4201 		vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
4202 		vs->vs_state = vd->vdev_state;
4203 		vs->vs_rsize = vdev_get_min_asize(vd);
4204 
4205 		if (vd->vdev_ops->vdev_op_leaf) {
4206 			vs->vs_rsize += VDEV_LABEL_START_SIZE +
4207 			    VDEV_LABEL_END_SIZE;
4208 			/*
4209 			 * Report initializing progress. Since we don't
4210 			 * have the initializing locks held, this is only
4211 			 * an estimate (although a fairly accurate one).
4212 			 */
4213 			vs->vs_initialize_bytes_done =
4214 			    vd->vdev_initialize_bytes_done;
4215 			vs->vs_initialize_bytes_est =
4216 			    vd->vdev_initialize_bytes_est;
4217 			vs->vs_initialize_state = vd->vdev_initialize_state;
4218 			vs->vs_initialize_action_time =
4219 			    vd->vdev_initialize_action_time;
4220 
4221 			/*
4222 			 * Report manual TRIM progress. Since we don't have
4223 			 * the manual TRIM locks held, this is only an
4224 			 * estimate (although fairly accurate one).
4225 			 */
4226 			vs->vs_trim_notsup = !vd->vdev_has_trim;
4227 			vs->vs_trim_bytes_done = vd->vdev_trim_bytes_done;
4228 			vs->vs_trim_bytes_est = vd->vdev_trim_bytes_est;
4229 			vs->vs_trim_state = vd->vdev_trim_state;
4230 			vs->vs_trim_action_time = vd->vdev_trim_action_time;
4231 
4232 			/* Set when there is a deferred resilver. */
4233 			vs->vs_resilver_deferred = vd->vdev_resilver_deferred;
4234 		}
4235 
4236 		/*
4237 		 * Report expandable space on top-level, non-auxiliary devices
4238 		 * only. The expandable space is reported in terms of metaslab
4239 		 * sized units since that determines how much space the pool
4240 		 * can expand.
4241 		 */
4242 		if (vd->vdev_aux == NULL && tvd != NULL) {
4243 			vs->vs_esize = P2ALIGN(
4244 			    vd->vdev_max_asize - vd->vdev_asize,
4245 			    1ULL << tvd->vdev_ms_shift);
4246 		}
4247 
4248 		vs->vs_configured_ashift = vd->vdev_top != NULL
4249 		    ? vd->vdev_top->vdev_ashift : vd->vdev_ashift;
4250 		vs->vs_logical_ashift = vd->vdev_logical_ashift;
4251 		vs->vs_physical_ashift = vd->vdev_physical_ashift;
4252 
4253 		/*
4254 		 * Report fragmentation and rebuild progress for top-level,
4255 		 * non-auxiliary, concrete devices.
4256 		 */
4257 		if (vd->vdev_aux == NULL && vd == vd->vdev_top &&
4258 		    vdev_is_concrete(vd)) {
4259 			vs->vs_fragmentation = (vd->vdev_mg != NULL) ?
4260 			    vd->vdev_mg->mg_fragmentation : 0;
4261 		}
4262 	}
4263 
4264 	vdev_get_stats_ex_impl(vd, vs, vsx);
4265 	mutex_exit(&vd->vdev_stat_lock);
4266 }
4267 
4268 void
4269 vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
4270 {
4271 	return (vdev_get_stats_ex(vd, vs, NULL));
4272 }
4273 
4274 void
4275 vdev_clear_stats(vdev_t *vd)
4276 {
4277 	mutex_enter(&vd->vdev_stat_lock);
4278 	vd->vdev_stat.vs_space = 0;
4279 	vd->vdev_stat.vs_dspace = 0;
4280 	vd->vdev_stat.vs_alloc = 0;
4281 	mutex_exit(&vd->vdev_stat_lock);
4282 }
4283 
4284 void
4285 vdev_scan_stat_init(vdev_t *vd)
4286 {
4287 	vdev_stat_t *vs = &vd->vdev_stat;
4288 
4289 	for (int c = 0; c < vd->vdev_children; c++)
4290 		vdev_scan_stat_init(vd->vdev_child[c]);
4291 
4292 	mutex_enter(&vd->vdev_stat_lock);
4293 	vs->vs_scan_processed = 0;
4294 	mutex_exit(&vd->vdev_stat_lock);
4295 }
4296 
4297 void
4298 vdev_stat_update(zio_t *zio, uint64_t psize)
4299 {
4300 	spa_t *spa = zio->io_spa;
4301 	vdev_t *rvd = spa->spa_root_vdev;
4302 	vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
4303 	vdev_t *pvd;
4304 	uint64_t txg = zio->io_txg;
4305 	vdev_stat_t *vs = &vd->vdev_stat;
4306 	vdev_stat_ex_t *vsx = &vd->vdev_stat_ex;
4307 	zio_type_t type = zio->io_type;
4308 	int flags = zio->io_flags;
4309 
4310 	/*
4311 	 * If this i/o is a gang leader, it didn't do any actual work.
4312 	 */
4313 	if (zio->io_gang_tree)
4314 		return;
4315 
4316 	if (zio->io_error == 0) {
4317 		/*
4318 		 * If this is a root i/o, don't count it -- we've already
4319 		 * counted the top-level vdevs, and vdev_get_stats() will
4320 		 * aggregate them when asked.  This reduces contention on
4321 		 * the root vdev_stat_lock and implicitly handles blocks
4322 		 * that compress away to holes, for which there is no i/o.
4323 		 * (Holes never create vdev children, so all the counters
4324 		 * remain zero, which is what we want.)
4325 		 *
4326 		 * Note: this only applies to successful i/o (io_error == 0)
4327 		 * because unlike i/o counts, errors are not additive.
4328 		 * When reading a ditto block, for example, failure of
4329 		 * one top-level vdev does not imply a root-level error.
4330 		 */
4331 		if (vd == rvd)
4332 			return;
4333 
4334 		ASSERT(vd == zio->io_vd);
4335 
4336 		if (flags & ZIO_FLAG_IO_BYPASS)
4337 			return;
4338 
4339 		mutex_enter(&vd->vdev_stat_lock);
4340 
4341 		if (flags & ZIO_FLAG_IO_REPAIR) {
4342 			/*
4343 			 * Repair is the result of a resilver issued by the
4344 			 * scan thread (spa_sync).
4345 			 */
4346 			if (flags & ZIO_FLAG_SCAN_THREAD) {
4347 				dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
4348 				dsl_scan_phys_t *scn_phys = &scn->scn_phys;
4349 				uint64_t *processed = &scn_phys->scn_processed;
4350 
4351 				if (vd->vdev_ops->vdev_op_leaf)
4352 					atomic_add_64(processed, psize);
4353 				vs->vs_scan_processed += psize;
4354 			}
4355 
4356 			/*
4357 			 * Repair is the result of a rebuild issued by the
4358 			 * rebuild thread (vdev_rebuild_thread).  To avoid
4359 			 * double counting repaired bytes the virtual dRAID
4360 			 * spare vdev is excluded from the processed bytes.
4361 			 */
4362 			if (zio->io_priority == ZIO_PRIORITY_REBUILD) {
4363 				vdev_t *tvd = vd->vdev_top;
4364 				vdev_rebuild_t *vr = &tvd->vdev_rebuild_config;
4365 				vdev_rebuild_phys_t *vrp = &vr->vr_rebuild_phys;
4366 				uint64_t *rebuilt = &vrp->vrp_bytes_rebuilt;
4367 
4368 				if (vd->vdev_ops->vdev_op_leaf &&
4369 				    vd->vdev_ops != &vdev_draid_spare_ops) {
4370 					atomic_add_64(rebuilt, psize);
4371 				}
4372 				vs->vs_rebuild_processed += psize;
4373 			}
4374 
4375 			if (flags & ZIO_FLAG_SELF_HEAL)
4376 				vs->vs_self_healed += psize;
4377 		}
4378 
4379 		/*
4380 		 * The bytes/ops/histograms are recorded at the leaf level and
4381 		 * aggregated into the higher level vdevs in vdev_get_stats().
4382 		 */
4383 		if (vd->vdev_ops->vdev_op_leaf &&
4384 		    (zio->io_priority < ZIO_PRIORITY_NUM_QUEUEABLE)) {
4385 			zio_type_t vs_type = type;
4386 			zio_priority_t priority = zio->io_priority;
4387 
4388 			/*
4389 			 * TRIM ops and bytes are reported to user space as
4390 			 * ZIO_TYPE_IOCTL.  This is done to preserve the
4391 			 * vdev_stat_t structure layout for user space.
4392 			 */
4393 			if (type == ZIO_TYPE_TRIM)
4394 				vs_type = ZIO_TYPE_IOCTL;
4395 
4396 			/*
4397 			 * Solely for the purposes of 'zpool iostat -lqrw'
4398 			 * reporting use the priority to catagorize the IO.
4399 			 * Only the following are reported to user space:
4400 			 *
4401 			 *   ZIO_PRIORITY_SYNC_READ,
4402 			 *   ZIO_PRIORITY_SYNC_WRITE,
4403 			 *   ZIO_PRIORITY_ASYNC_READ,
4404 			 *   ZIO_PRIORITY_ASYNC_WRITE,
4405 			 *   ZIO_PRIORITY_SCRUB,
4406 			 *   ZIO_PRIORITY_TRIM.
4407 			 */
4408 			if (priority == ZIO_PRIORITY_REBUILD) {
4409 				priority = ((type == ZIO_TYPE_WRITE) ?
4410 				    ZIO_PRIORITY_ASYNC_WRITE :
4411 				    ZIO_PRIORITY_SCRUB);
4412 			} else if (priority == ZIO_PRIORITY_INITIALIZING) {
4413 				ASSERT3U(type, ==, ZIO_TYPE_WRITE);
4414 				priority = ZIO_PRIORITY_ASYNC_WRITE;
4415 			} else if (priority == ZIO_PRIORITY_REMOVAL) {
4416 				priority = ((type == ZIO_TYPE_WRITE) ?
4417 				    ZIO_PRIORITY_ASYNC_WRITE :
4418 				    ZIO_PRIORITY_ASYNC_READ);
4419 			}
4420 
4421 			vs->vs_ops[vs_type]++;
4422 			vs->vs_bytes[vs_type] += psize;
4423 
4424 			if (flags & ZIO_FLAG_DELEGATED) {
4425 				vsx->vsx_agg_histo[priority]
4426 				    [RQ_HISTO(zio->io_size)]++;
4427 			} else {
4428 				vsx->vsx_ind_histo[priority]
4429 				    [RQ_HISTO(zio->io_size)]++;
4430 			}
4431 
4432 			if (zio->io_delta && zio->io_delay) {
4433 				vsx->vsx_queue_histo[priority]
4434 				    [L_HISTO(zio->io_delta - zio->io_delay)]++;
4435 				vsx->vsx_disk_histo[type]
4436 				    [L_HISTO(zio->io_delay)]++;
4437 				vsx->vsx_total_histo[type]
4438 				    [L_HISTO(zio->io_delta)]++;
4439 			}
4440 		}
4441 
4442 		mutex_exit(&vd->vdev_stat_lock);
4443 		return;
4444 	}
4445 
4446 	if (flags & ZIO_FLAG_SPECULATIVE)
4447 		return;
4448 
4449 	/*
4450 	 * If this is an I/O error that is going to be retried, then ignore the
4451 	 * error.  Otherwise, the user may interpret B_FAILFAST I/O errors as
4452 	 * hard errors, when in reality they can happen for any number of
4453 	 * innocuous reasons (bus resets, MPxIO link failure, etc).
4454 	 */
4455 	if (zio->io_error == EIO &&
4456 	    !(zio->io_flags & ZIO_FLAG_IO_RETRY))
4457 		return;
4458 
4459 	/*
4460 	 * Intent logs writes won't propagate their error to the root
4461 	 * I/O so don't mark these types of failures as pool-level
4462 	 * errors.
4463 	 */
4464 	if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
4465 		return;
4466 
4467 	if (type == ZIO_TYPE_WRITE && txg != 0 &&
4468 	    (!(flags & ZIO_FLAG_IO_REPAIR) ||
4469 	    (flags & ZIO_FLAG_SCAN_THREAD) ||
4470 	    spa->spa_claiming)) {
4471 		/*
4472 		 * This is either a normal write (not a repair), or it's
4473 		 * a repair induced by the scrub thread, or it's a repair
4474 		 * made by zil_claim() during spa_load() in the first txg.
4475 		 * In the normal case, we commit the DTL change in the same
4476 		 * txg as the block was born.  In the scrub-induced repair
4477 		 * case, we know that scrubs run in first-pass syncing context,
4478 		 * so we commit the DTL change in spa_syncing_txg(spa).
4479 		 * In the zil_claim() case, we commit in spa_first_txg(spa).
4480 		 *
4481 		 * We currently do not make DTL entries for failed spontaneous
4482 		 * self-healing writes triggered by normal (non-scrubbing)
4483 		 * reads, because we have no transactional context in which to
4484 		 * do so -- and it's not clear that it'd be desirable anyway.
4485 		 */
4486 		if (vd->vdev_ops->vdev_op_leaf) {
4487 			uint64_t commit_txg = txg;
4488 			if (flags & ZIO_FLAG_SCAN_THREAD) {
4489 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
4490 				ASSERT(spa_sync_pass(spa) == 1);
4491 				vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
4492 				commit_txg = spa_syncing_txg(spa);
4493 			} else if (spa->spa_claiming) {
4494 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
4495 				commit_txg = spa_first_txg(spa);
4496 			}
4497 			ASSERT(commit_txg >= spa_syncing_txg(spa));
4498 			if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
4499 				return;
4500 			for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
4501 				vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
4502 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
4503 		}
4504 		if (vd != rvd)
4505 			vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
4506 	}
4507 }
4508 
4509 int64_t
4510 vdev_deflated_space(vdev_t *vd, int64_t space)
4511 {
4512 	ASSERT((space & (SPA_MINBLOCKSIZE-1)) == 0);
4513 	ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
4514 
4515 	return ((space >> SPA_MINBLOCKSHIFT) * vd->vdev_deflate_ratio);
4516 }
4517 
4518 /*
4519  * Update the in-core space usage stats for this vdev, its metaslab class,
4520  * and the root vdev.
4521  */
4522 void
4523 vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
4524     int64_t space_delta)
4525 {
4526 	int64_t dspace_delta;
4527 	spa_t *spa = vd->vdev_spa;
4528 	vdev_t *rvd = spa->spa_root_vdev;
4529 
4530 	ASSERT(vd == vd->vdev_top);
4531 
4532 	/*
4533 	 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
4534 	 * factor.  We must calculate this here and not at the root vdev
4535 	 * because the root vdev's psize-to-asize is simply the max of its
4536 	 * children's, thus not accurate enough for us.
4537 	 */
4538 	dspace_delta = vdev_deflated_space(vd, space_delta);
4539 
4540 	mutex_enter(&vd->vdev_stat_lock);
4541 	/* ensure we won't underflow */
4542 	if (alloc_delta < 0) {
4543 		ASSERT3U(vd->vdev_stat.vs_alloc, >=, -alloc_delta);
4544 	}
4545 
4546 	vd->vdev_stat.vs_alloc += alloc_delta;
4547 	vd->vdev_stat.vs_space += space_delta;
4548 	vd->vdev_stat.vs_dspace += dspace_delta;
4549 	mutex_exit(&vd->vdev_stat_lock);
4550 
4551 	/* every class but log contributes to root space stats */
4552 	if (vd->vdev_mg != NULL && !vd->vdev_islog) {
4553 		ASSERT(!vd->vdev_isl2cache);
4554 		mutex_enter(&rvd->vdev_stat_lock);
4555 		rvd->vdev_stat.vs_alloc += alloc_delta;
4556 		rvd->vdev_stat.vs_space += space_delta;
4557 		rvd->vdev_stat.vs_dspace += dspace_delta;
4558 		mutex_exit(&rvd->vdev_stat_lock);
4559 	}
4560 	/* Note: metaslab_class_space_update moved to metaslab_space_update */
4561 }
4562 
4563 /*
4564  * Mark a top-level vdev's config as dirty, placing it on the dirty list
4565  * so that it will be written out next time the vdev configuration is synced.
4566  * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
4567  */
4568 void
4569 vdev_config_dirty(vdev_t *vd)
4570 {
4571 	spa_t *spa = vd->vdev_spa;
4572 	vdev_t *rvd = spa->spa_root_vdev;
4573 	int c;
4574 
4575 	ASSERT(spa_writeable(spa));
4576 
4577 	/*
4578 	 * If this is an aux vdev (as with l2cache and spare devices), then we
4579 	 * update the vdev config manually and set the sync flag.
4580 	 */
4581 	if (vd->vdev_aux != NULL) {
4582 		spa_aux_vdev_t *sav = vd->vdev_aux;
4583 		nvlist_t **aux;
4584 		uint_t naux;
4585 
4586 		for (c = 0; c < sav->sav_count; c++) {
4587 			if (sav->sav_vdevs[c] == vd)
4588 				break;
4589 		}
4590 
4591 		if (c == sav->sav_count) {
4592 			/*
4593 			 * We're being removed.  There's nothing more to do.
4594 			 */
4595 			ASSERT(sav->sav_sync == B_TRUE);
4596 			return;
4597 		}
4598 
4599 		sav->sav_sync = B_TRUE;
4600 
4601 		if (nvlist_lookup_nvlist_array(sav->sav_config,
4602 		    ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
4603 			VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
4604 			    ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
4605 		}
4606 
4607 		ASSERT(c < naux);
4608 
4609 		/*
4610 		 * Setting the nvlist in the middle if the array is a little
4611 		 * sketchy, but it will work.
4612 		 */
4613 		nvlist_free(aux[c]);
4614 		aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0);
4615 
4616 		return;
4617 	}
4618 
4619 	/*
4620 	 * The dirty list is protected by the SCL_CONFIG lock.  The caller
4621 	 * must either hold SCL_CONFIG as writer, or must be the sync thread
4622 	 * (which holds SCL_CONFIG as reader).  There's only one sync thread,
4623 	 * so this is sufficient to ensure mutual exclusion.
4624 	 */
4625 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
4626 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
4627 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
4628 
4629 	if (vd == rvd) {
4630 		for (c = 0; c < rvd->vdev_children; c++)
4631 			vdev_config_dirty(rvd->vdev_child[c]);
4632 	} else {
4633 		ASSERT(vd == vd->vdev_top);
4634 
4635 		if (!list_link_active(&vd->vdev_config_dirty_node) &&
4636 		    vdev_is_concrete(vd)) {
4637 			list_insert_head(&spa->spa_config_dirty_list, vd);
4638 		}
4639 	}
4640 }
4641 
4642 void
4643 vdev_config_clean(vdev_t *vd)
4644 {
4645 	spa_t *spa = vd->vdev_spa;
4646 
4647 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
4648 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
4649 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
4650 
4651 	ASSERT(list_link_active(&vd->vdev_config_dirty_node));
4652 	list_remove(&spa->spa_config_dirty_list, vd);
4653 }
4654 
4655 /*
4656  * Mark a top-level vdev's state as dirty, so that the next pass of
4657  * spa_sync() can convert this into vdev_config_dirty().  We distinguish
4658  * the state changes from larger config changes because they require
4659  * much less locking, and are often needed for administrative actions.
4660  */
4661 void
4662 vdev_state_dirty(vdev_t *vd)
4663 {
4664 	spa_t *spa = vd->vdev_spa;
4665 
4666 	ASSERT(spa_writeable(spa));
4667 	ASSERT(vd == vd->vdev_top);
4668 
4669 	/*
4670 	 * The state list is protected by the SCL_STATE lock.  The caller
4671 	 * must either hold SCL_STATE as writer, or must be the sync thread
4672 	 * (which holds SCL_STATE as reader).  There's only one sync thread,
4673 	 * so this is sufficient to ensure mutual exclusion.
4674 	 */
4675 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
4676 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
4677 	    spa_config_held(spa, SCL_STATE, RW_READER)));
4678 
4679 	if (!list_link_active(&vd->vdev_state_dirty_node) &&
4680 	    vdev_is_concrete(vd))
4681 		list_insert_head(&spa->spa_state_dirty_list, vd);
4682 }
4683 
4684 void
4685 vdev_state_clean(vdev_t *vd)
4686 {
4687 	spa_t *spa = vd->vdev_spa;
4688 
4689 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
4690 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
4691 	    spa_config_held(spa, SCL_STATE, RW_READER)));
4692 
4693 	ASSERT(list_link_active(&vd->vdev_state_dirty_node));
4694 	list_remove(&spa->spa_state_dirty_list, vd);
4695 }
4696 
4697 /*
4698  * Propagate vdev state up from children to parent.
4699  */
4700 void
4701 vdev_propagate_state(vdev_t *vd)
4702 {
4703 	spa_t *spa = vd->vdev_spa;
4704 	vdev_t *rvd = spa->spa_root_vdev;
4705 	int degraded = 0, faulted = 0;
4706 	int corrupted = 0;
4707 	vdev_t *child;
4708 
4709 	if (vd->vdev_children > 0) {
4710 		for (int c = 0; c < vd->vdev_children; c++) {
4711 			child = vd->vdev_child[c];
4712 
4713 			/*
4714 			 * Don't factor holes or indirect vdevs into the
4715 			 * decision.
4716 			 */
4717 			if (!vdev_is_concrete(child))
4718 				continue;
4719 
4720 			if (!vdev_readable(child) ||
4721 			    (!vdev_writeable(child) && spa_writeable(spa))) {
4722 				/*
4723 				 * Root special: if there is a top-level log
4724 				 * device, treat the root vdev as if it were
4725 				 * degraded.
4726 				 */
4727 				if (child->vdev_islog && vd == rvd)
4728 					degraded++;
4729 				else
4730 					faulted++;
4731 			} else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
4732 				degraded++;
4733 			}
4734 
4735 			if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
4736 				corrupted++;
4737 		}
4738 
4739 		vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
4740 
4741 		/*
4742 		 * Root special: if there is a top-level vdev that cannot be
4743 		 * opened due to corrupted metadata, then propagate the root
4744 		 * vdev's aux state as 'corrupt' rather than 'insufficient
4745 		 * replicas'.
4746 		 */
4747 		if (corrupted && vd == rvd &&
4748 		    rvd->vdev_state == VDEV_STATE_CANT_OPEN)
4749 			vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
4750 			    VDEV_AUX_CORRUPT_DATA);
4751 	}
4752 
4753 	if (vd->vdev_parent)
4754 		vdev_propagate_state(vd->vdev_parent);
4755 }
4756 
4757 /*
4758  * Set a vdev's state.  If this is during an open, we don't update the parent
4759  * state, because we're in the process of opening children depth-first.
4760  * Otherwise, we propagate the change to the parent.
4761  *
4762  * If this routine places a device in a faulted state, an appropriate ereport is
4763  * generated.
4764  */
4765 void
4766 vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
4767 {
4768 	uint64_t save_state;
4769 	spa_t *spa = vd->vdev_spa;
4770 
4771 	if (state == vd->vdev_state) {
4772 		/*
4773 		 * Since vdev_offline() code path is already in an offline
4774 		 * state we can miss a statechange event to OFFLINE. Check
4775 		 * the previous state to catch this condition.
4776 		 */
4777 		if (vd->vdev_ops->vdev_op_leaf &&
4778 		    (state == VDEV_STATE_OFFLINE) &&
4779 		    (vd->vdev_prevstate >= VDEV_STATE_FAULTED)) {
4780 			/* post an offline state change */
4781 			zfs_post_state_change(spa, vd, vd->vdev_prevstate);
4782 		}
4783 		vd->vdev_stat.vs_aux = aux;
4784 		return;
4785 	}
4786 
4787 	save_state = vd->vdev_state;
4788 
4789 	vd->vdev_state = state;
4790 	vd->vdev_stat.vs_aux = aux;
4791 
4792 	/*
4793 	 * If we are setting the vdev state to anything but an open state, then
4794 	 * always close the underlying device unless the device has requested
4795 	 * a delayed close (i.e. we're about to remove or fault the device).
4796 	 * Otherwise, we keep accessible but invalid devices open forever.
4797 	 * We don't call vdev_close() itself, because that implies some extra
4798 	 * checks (offline, etc) that we don't want here.  This is limited to
4799 	 * leaf devices, because otherwise closing the device will affect other
4800 	 * children.
4801 	 */
4802 	if (!vd->vdev_delayed_close && vdev_is_dead(vd) &&
4803 	    vd->vdev_ops->vdev_op_leaf)
4804 		vd->vdev_ops->vdev_op_close(vd);
4805 
4806 	if (vd->vdev_removed &&
4807 	    state == VDEV_STATE_CANT_OPEN &&
4808 	    (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
4809 		/*
4810 		 * If the previous state is set to VDEV_STATE_REMOVED, then this
4811 		 * device was previously marked removed and someone attempted to
4812 		 * reopen it.  If this failed due to a nonexistent device, then
4813 		 * keep the device in the REMOVED state.  We also let this be if
4814 		 * it is one of our special test online cases, which is only
4815 		 * attempting to online the device and shouldn't generate an FMA
4816 		 * fault.
4817 		 */
4818 		vd->vdev_state = VDEV_STATE_REMOVED;
4819 		vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
4820 	} else if (state == VDEV_STATE_REMOVED) {
4821 		vd->vdev_removed = B_TRUE;
4822 	} else if (state == VDEV_STATE_CANT_OPEN) {
4823 		/*
4824 		 * If we fail to open a vdev during an import or recovery, we
4825 		 * mark it as "not available", which signifies that it was
4826 		 * never there to begin with.  Failure to open such a device
4827 		 * is not considered an error.
4828 		 */
4829 		if ((spa_load_state(spa) == SPA_LOAD_IMPORT ||
4830 		    spa_load_state(spa) == SPA_LOAD_RECOVER) &&
4831 		    vd->vdev_ops->vdev_op_leaf)
4832 			vd->vdev_not_present = 1;
4833 
4834 		/*
4835 		 * Post the appropriate ereport.  If the 'prevstate' field is
4836 		 * set to something other than VDEV_STATE_UNKNOWN, it indicates
4837 		 * that this is part of a vdev_reopen().  In this case, we don't
4838 		 * want to post the ereport if the device was already in the
4839 		 * CANT_OPEN state beforehand.
4840 		 *
4841 		 * If the 'checkremove' flag is set, then this is an attempt to
4842 		 * online the device in response to an insertion event.  If we
4843 		 * hit this case, then we have detected an insertion event for a
4844 		 * faulted or offline device that wasn't in the removed state.
4845 		 * In this scenario, we don't post an ereport because we are
4846 		 * about to replace the device, or attempt an online with
4847 		 * vdev_forcefault, which will generate the fault for us.
4848 		 */
4849 		if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
4850 		    !vd->vdev_not_present && !vd->vdev_checkremove &&
4851 		    vd != spa->spa_root_vdev) {
4852 			const char *class;
4853 
4854 			switch (aux) {
4855 			case VDEV_AUX_OPEN_FAILED:
4856 				class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
4857 				break;
4858 			case VDEV_AUX_CORRUPT_DATA:
4859 				class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
4860 				break;
4861 			case VDEV_AUX_NO_REPLICAS:
4862 				class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
4863 				break;
4864 			case VDEV_AUX_BAD_GUID_SUM:
4865 				class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
4866 				break;
4867 			case VDEV_AUX_TOO_SMALL:
4868 				class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
4869 				break;
4870 			case VDEV_AUX_BAD_LABEL:
4871 				class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
4872 				break;
4873 			case VDEV_AUX_BAD_ASHIFT:
4874 				class = FM_EREPORT_ZFS_DEVICE_BAD_ASHIFT;
4875 				break;
4876 			default:
4877 				class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
4878 			}
4879 
4880 			(void) zfs_ereport_post(class, spa, vd, NULL, NULL,
4881 			    save_state);
4882 		}
4883 
4884 		/* Erase any notion of persistent removed state */
4885 		vd->vdev_removed = B_FALSE;
4886 	} else {
4887 		vd->vdev_removed = B_FALSE;
4888 	}
4889 
4890 	/*
4891 	 * Notify ZED of any significant state-change on a leaf vdev.
4892 	 *
4893 	 */
4894 	if (vd->vdev_ops->vdev_op_leaf) {
4895 		/* preserve original state from a vdev_reopen() */
4896 		if ((vd->vdev_prevstate != VDEV_STATE_UNKNOWN) &&
4897 		    (vd->vdev_prevstate != vd->vdev_state) &&
4898 		    (save_state <= VDEV_STATE_CLOSED))
4899 			save_state = vd->vdev_prevstate;
4900 
4901 		/* filter out state change due to initial vdev_open */
4902 		if (save_state > VDEV_STATE_CLOSED)
4903 			zfs_post_state_change(spa, vd, save_state);
4904 	}
4905 
4906 	if (!isopen && vd->vdev_parent)
4907 		vdev_propagate_state(vd->vdev_parent);
4908 }
4909 
4910 boolean_t
4911 vdev_children_are_offline(vdev_t *vd)
4912 {
4913 	ASSERT(!vd->vdev_ops->vdev_op_leaf);
4914 
4915 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
4916 		if (vd->vdev_child[i]->vdev_state != VDEV_STATE_OFFLINE)
4917 			return (B_FALSE);
4918 	}
4919 
4920 	return (B_TRUE);
4921 }
4922 
4923 /*
4924  * Check the vdev configuration to ensure that it's capable of supporting
4925  * a root pool. We do not support partial configuration.
4926  */
4927 boolean_t
4928 vdev_is_bootable(vdev_t *vd)
4929 {
4930 	if (!vd->vdev_ops->vdev_op_leaf) {
4931 		const char *vdev_type = vd->vdev_ops->vdev_op_type;
4932 
4933 		if (strcmp(vdev_type, VDEV_TYPE_MISSING) == 0 ||
4934 		    strcmp(vdev_type, VDEV_TYPE_INDIRECT) == 0) {
4935 			return (B_FALSE);
4936 		}
4937 	}
4938 
4939 	for (int c = 0; c < vd->vdev_children; c++) {
4940 		if (!vdev_is_bootable(vd->vdev_child[c]))
4941 			return (B_FALSE);
4942 	}
4943 	return (B_TRUE);
4944 }
4945 
4946 boolean_t
4947 vdev_is_concrete(vdev_t *vd)
4948 {
4949 	vdev_ops_t *ops = vd->vdev_ops;
4950 	if (ops == &vdev_indirect_ops || ops == &vdev_hole_ops ||
4951 	    ops == &vdev_missing_ops || ops == &vdev_root_ops) {
4952 		return (B_FALSE);
4953 	} else {
4954 		return (B_TRUE);
4955 	}
4956 }
4957 
4958 /*
4959  * Determine if a log device has valid content.  If the vdev was
4960  * removed or faulted in the MOS config then we know that
4961  * the content on the log device has already been written to the pool.
4962  */
4963 boolean_t
4964 vdev_log_state_valid(vdev_t *vd)
4965 {
4966 	if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted &&
4967 	    !vd->vdev_removed)
4968 		return (B_TRUE);
4969 
4970 	for (int c = 0; c < vd->vdev_children; c++)
4971 		if (vdev_log_state_valid(vd->vdev_child[c]))
4972 			return (B_TRUE);
4973 
4974 	return (B_FALSE);
4975 }
4976 
4977 /*
4978  * Expand a vdev if possible.
4979  */
4980 void
4981 vdev_expand(vdev_t *vd, uint64_t txg)
4982 {
4983 	ASSERT(vd->vdev_top == vd);
4984 	ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
4985 	ASSERT(vdev_is_concrete(vd));
4986 
4987 	vdev_set_deflate_ratio(vd);
4988 
4989 	if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count &&
4990 	    vdev_is_concrete(vd)) {
4991 		vdev_metaslab_group_create(vd);
4992 		VERIFY(vdev_metaslab_init(vd, txg) == 0);
4993 		vdev_config_dirty(vd);
4994 	}
4995 }
4996 
4997 /*
4998  * Split a vdev.
4999  */
5000 void
5001 vdev_split(vdev_t *vd)
5002 {
5003 	vdev_t *cvd, *pvd = vd->vdev_parent;
5004 
5005 	vdev_remove_child(pvd, vd);
5006 	vdev_compact_children(pvd);
5007 
5008 	cvd = pvd->vdev_child[0];
5009 	if (pvd->vdev_children == 1) {
5010 		vdev_remove_parent(cvd);
5011 		cvd->vdev_splitting = B_TRUE;
5012 	}
5013 	vdev_propagate_state(cvd);
5014 }
5015 
5016 void
5017 vdev_deadman(vdev_t *vd, char *tag)
5018 {
5019 	for (int c = 0; c < vd->vdev_children; c++) {
5020 		vdev_t *cvd = vd->vdev_child[c];
5021 
5022 		vdev_deadman(cvd, tag);
5023 	}
5024 
5025 	if (vd->vdev_ops->vdev_op_leaf) {
5026 		vdev_queue_t *vq = &vd->vdev_queue;
5027 
5028 		mutex_enter(&vq->vq_lock);
5029 		if (avl_numnodes(&vq->vq_active_tree) > 0) {
5030 			spa_t *spa = vd->vdev_spa;
5031 			zio_t *fio;
5032 			uint64_t delta;
5033 
5034 			zfs_dbgmsg("slow vdev: %s has %d active IOs",
5035 			    vd->vdev_path, avl_numnodes(&vq->vq_active_tree));
5036 
5037 			/*
5038 			 * Look at the head of all the pending queues,
5039 			 * if any I/O has been outstanding for longer than
5040 			 * the spa_deadman_synctime invoke the deadman logic.
5041 			 */
5042 			fio = avl_first(&vq->vq_active_tree);
5043 			delta = gethrtime() - fio->io_timestamp;
5044 			if (delta > spa_deadman_synctime(spa))
5045 				zio_deadman(fio, tag);
5046 		}
5047 		mutex_exit(&vq->vq_lock);
5048 	}
5049 }
5050 
5051 void
5052 vdev_defer_resilver(vdev_t *vd)
5053 {
5054 	ASSERT(vd->vdev_ops->vdev_op_leaf);
5055 
5056 	vd->vdev_resilver_deferred = B_TRUE;
5057 	vd->vdev_spa->spa_resilver_deferred = B_TRUE;
5058 }
5059 
5060 /*
5061  * Clears the resilver deferred flag on all leaf devs under vd. Returns
5062  * B_TRUE if we have devices that need to be resilvered and are available to
5063  * accept resilver I/Os.
5064  */
5065 boolean_t
5066 vdev_clear_resilver_deferred(vdev_t *vd, dmu_tx_t *tx)
5067 {
5068 	boolean_t resilver_needed = B_FALSE;
5069 	spa_t *spa = vd->vdev_spa;
5070 
5071 	for (int c = 0; c < vd->vdev_children; c++) {
5072 		vdev_t *cvd = vd->vdev_child[c];
5073 		resilver_needed |= vdev_clear_resilver_deferred(cvd, tx);
5074 	}
5075 
5076 	if (vd == spa->spa_root_vdev &&
5077 	    spa_feature_is_active(spa, SPA_FEATURE_RESILVER_DEFER)) {
5078 		spa_feature_decr(spa, SPA_FEATURE_RESILVER_DEFER, tx);
5079 		vdev_config_dirty(vd);
5080 		spa->spa_resilver_deferred = B_FALSE;
5081 		return (resilver_needed);
5082 	}
5083 
5084 	if (!vdev_is_concrete(vd) || vd->vdev_aux ||
5085 	    !vd->vdev_ops->vdev_op_leaf)
5086 		return (resilver_needed);
5087 
5088 	vd->vdev_resilver_deferred = B_FALSE;
5089 
5090 	return (!vdev_is_dead(vd) && !vd->vdev_offline &&
5091 	    vdev_resilver_needed(vd, NULL, NULL));
5092 }
5093 
5094 boolean_t
5095 vdev_xlate_is_empty(range_seg64_t *rs)
5096 {
5097 	return (rs->rs_start == rs->rs_end);
5098 }
5099 
5100 /*
5101  * Translate a logical range to the first contiguous physical range for the
5102  * specified vdev_t.  This function is initially called with a leaf vdev and
5103  * will walk each parent vdev until it reaches a top-level vdev. Once the
5104  * top-level is reached the physical range is initialized and the recursive
5105  * function begins to unwind. As it unwinds it calls the parent's vdev
5106  * specific translation function to do the real conversion.
5107  */
5108 void
5109 vdev_xlate(vdev_t *vd, const range_seg64_t *logical_rs,
5110     range_seg64_t *physical_rs, range_seg64_t *remain_rs)
5111 {
5112 	/*
5113 	 * Walk up the vdev tree
5114 	 */
5115 	if (vd != vd->vdev_top) {
5116 		vdev_xlate(vd->vdev_parent, logical_rs, physical_rs,
5117 		    remain_rs);
5118 	} else {
5119 		/*
5120 		 * We've reached the top-level vdev, initialize the physical
5121 		 * range to the logical range and set an empty remaining
5122 		 * range then start to unwind.
5123 		 */
5124 		physical_rs->rs_start = logical_rs->rs_start;
5125 		physical_rs->rs_end = logical_rs->rs_end;
5126 
5127 		remain_rs->rs_start = logical_rs->rs_start;
5128 		remain_rs->rs_end = logical_rs->rs_start;
5129 
5130 		return;
5131 	}
5132 
5133 	vdev_t *pvd = vd->vdev_parent;
5134 	ASSERT3P(pvd, !=, NULL);
5135 	ASSERT3P(pvd->vdev_ops->vdev_op_xlate, !=, NULL);
5136 
5137 	/*
5138 	 * As this recursive function unwinds, translate the logical
5139 	 * range into its physical and any remaining components by calling
5140 	 * the vdev specific translate function.
5141 	 */
5142 	range_seg64_t intermediate = { 0 };
5143 	pvd->vdev_ops->vdev_op_xlate(vd, physical_rs, &intermediate, remain_rs);
5144 
5145 	physical_rs->rs_start = intermediate.rs_start;
5146 	physical_rs->rs_end = intermediate.rs_end;
5147 }
5148 
5149 void
5150 vdev_xlate_walk(vdev_t *vd, const range_seg64_t *logical_rs,
5151     vdev_xlate_func_t *func, void *arg)
5152 {
5153 	range_seg64_t iter_rs = *logical_rs;
5154 	range_seg64_t physical_rs;
5155 	range_seg64_t remain_rs;
5156 
5157 	while (!vdev_xlate_is_empty(&iter_rs)) {
5158 
5159 		vdev_xlate(vd, &iter_rs, &physical_rs, &remain_rs);
5160 
5161 		/*
5162 		 * With raidz and dRAID, it's possible that the logical range
5163 		 * does not live on this leaf vdev. Only when there is a non-
5164 		 * zero physical size call the provided function.
5165 		 */
5166 		if (!vdev_xlate_is_empty(&physical_rs))
5167 			func(arg, &physical_rs);
5168 
5169 		iter_rs = remain_rs;
5170 	}
5171 }
5172 
5173 /*
5174  * Look at the vdev tree and determine whether any devices are currently being
5175  * replaced.
5176  */
5177 boolean_t
5178 vdev_replace_in_progress(vdev_t *vdev)
5179 {
5180 	ASSERT(spa_config_held(vdev->vdev_spa, SCL_ALL, RW_READER) != 0);
5181 
5182 	if (vdev->vdev_ops == &vdev_replacing_ops)
5183 		return (B_TRUE);
5184 
5185 	/*
5186 	 * A 'spare' vdev indicates that we have a replace in progress, unless
5187 	 * it has exactly two children, and the second, the hot spare, has
5188 	 * finished being resilvered.
5189 	 */
5190 	if (vdev->vdev_ops == &vdev_spare_ops && (vdev->vdev_children > 2 ||
5191 	    !vdev_dtl_empty(vdev->vdev_child[1], DTL_MISSING)))
5192 		return (B_TRUE);
5193 
5194 	for (int i = 0; i < vdev->vdev_children; i++) {
5195 		if (vdev_replace_in_progress(vdev->vdev_child[i]))
5196 			return (B_TRUE);
5197 	}
5198 
5199 	return (B_FALSE);
5200 }
5201 
5202 EXPORT_SYMBOL(vdev_fault);
5203 EXPORT_SYMBOL(vdev_degrade);
5204 EXPORT_SYMBOL(vdev_online);
5205 EXPORT_SYMBOL(vdev_offline);
5206 EXPORT_SYMBOL(vdev_clear);
5207 
5208 /* BEGIN CSTYLED */
5209 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, default_ms_count, INT, ZMOD_RW,
5210 	"Target number of metaslabs per top-level vdev");
5211 
5212 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, default_ms_shift, INT, ZMOD_RW,
5213 	"Default limit for metaslab size");
5214 
5215 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, min_ms_count, INT, ZMOD_RW,
5216 	"Minimum number of metaslabs per top-level vdev");
5217 
5218 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, ms_count_limit, INT, ZMOD_RW,
5219 	"Practical upper limit of total metaslabs per top-level vdev");
5220 
5221 ZFS_MODULE_PARAM(zfs, zfs_, slow_io_events_per_second, UINT, ZMOD_RW,
5222 	"Rate limit slow IO (delay) events to this many per second");
5223 
5224 ZFS_MODULE_PARAM(zfs, zfs_, checksum_events_per_second, UINT, ZMOD_RW,
5225 	"Rate limit checksum events to this many checksum errors per second "
5226 	"(do not set below zed threshold).");
5227 
5228 ZFS_MODULE_PARAM(zfs, zfs_, scan_ignore_errors, INT, ZMOD_RW,
5229 	"Ignore errors during resilver/scrub");
5230 
5231 ZFS_MODULE_PARAM(zfs_vdev, vdev_, validate_skip, INT, ZMOD_RW,
5232 	"Bypass vdev_validate()");
5233 
5234 ZFS_MODULE_PARAM(zfs, zfs_, nocacheflush, INT, ZMOD_RW,
5235 	"Disable cache flushes");
5236 
5237 ZFS_MODULE_PARAM_CALL(zfs_vdev, zfs_vdev_, min_auto_ashift,
5238 	param_set_min_auto_ashift, param_get_ulong, ZMOD_RW,
5239 	"Minimum ashift used when creating new top-level vdevs");
5240 
5241 ZFS_MODULE_PARAM_CALL(zfs_vdev, zfs_vdev_, max_auto_ashift,
5242 	param_set_max_auto_ashift, param_get_ulong, ZMOD_RW,
5243 	"Maximum ashift used when optimizing for logical -> physical sector "
5244 	"size on new top-level vdevs");
5245 /* END CSTYLED */
5246