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