]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - sys/cddl/contrib/opensolaris/uts/common/fs/zfs/vdev.c
MFV r331695, 331700: 9166 zfs storage pool checkpoint
[FreeBSD/FreeBSD.git] / sys / cddl / contrib / opensolaris / uts / common / fs / zfs / vdev.c
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21
22 /*
23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
25  * Copyright 2017 Nexenta Systems, Inc.
26  * Copyright 2013 Martin Matuska <mm@FreeBSD.org>. All rights reserved.
27  * Copyright (c) 2014 Integros [integros.com]
28  * Copyright 2016 Toomas Soome <tsoome@me.com>
29  * Copyright 2017 Joyent, Inc.
30  */
31
32 #include <sys/zfs_context.h>
33 #include <sys/fm/fs/zfs.h>
34 #include <sys/spa.h>
35 #include <sys/spa_impl.h>
36 #include <sys/bpobj.h>
37 #include <sys/dmu.h>
38 #include <sys/dmu_tx.h>
39 #include <sys/dsl_dir.h>
40 #include <sys/vdev_impl.h>
41 #include <sys/uberblock_impl.h>
42 #include <sys/metaslab.h>
43 #include <sys/metaslab_impl.h>
44 #include <sys/space_map.h>
45 #include <sys/space_reftree.h>
46 #include <sys/zio.h>
47 #include <sys/zap.h>
48 #include <sys/fs/zfs.h>
49 #include <sys/arc.h>
50 #include <sys/zil.h>
51 #include <sys/dsl_scan.h>
52 #include <sys/abd.h>
53 #include <sys/trim_map.h>
54
55 SYSCTL_DECL(_vfs_zfs);
56 SYSCTL_NODE(_vfs_zfs, OID_AUTO, vdev, CTLFLAG_RW, 0, "ZFS VDEV");
57
58 /*
59  * Virtual device management.
60  */
61
62 /*
63  * The limit for ZFS to automatically increase a top-level vdev's ashift
64  * from logical ashift to physical ashift.
65  *
66  * Example: one or more 512B emulation child vdevs
67  *          child->vdev_ashift = 9 (512 bytes)
68  *          child->vdev_physical_ashift = 12 (4096 bytes)
69  *          zfs_max_auto_ashift = 11 (2048 bytes)
70  *          zfs_min_auto_ashift = 9 (512 bytes)
71  *
72  * On pool creation or the addition of a new top-level vdev, ZFS will
73  * increase the ashift of the top-level vdev to 2048 as limited by
74  * zfs_max_auto_ashift.
75  *
76  * Example: one or more 512B emulation child vdevs
77  *          child->vdev_ashift = 9 (512 bytes)
78  *          child->vdev_physical_ashift = 12 (4096 bytes)
79  *          zfs_max_auto_ashift = 13 (8192 bytes)
80  *          zfs_min_auto_ashift = 9 (512 bytes)
81  *
82  * On pool creation or the addition of a new top-level vdev, ZFS will
83  * increase the ashift of the top-level vdev to 4096 to match the
84  * max vdev_physical_ashift.
85  *
86  * Example: one or more 512B emulation child vdevs
87  *          child->vdev_ashift = 9 (512 bytes)
88  *          child->vdev_physical_ashift = 9 (512 bytes)
89  *          zfs_max_auto_ashift = 13 (8192 bytes)
90  *          zfs_min_auto_ashift = 12 (4096 bytes)
91  *
92  * On pool creation or the addition of a new top-level vdev, ZFS will
93  * increase the ashift of the top-level vdev to 4096 to match the
94  * zfs_min_auto_ashift.
95  */
96 static uint64_t zfs_max_auto_ashift = SPA_MAXASHIFT;
97 static uint64_t zfs_min_auto_ashift = SPA_MINASHIFT;
98
99 static int
100 sysctl_vfs_zfs_max_auto_ashift(SYSCTL_HANDLER_ARGS)
101 {
102         uint64_t val;
103         int err;
104
105         val = zfs_max_auto_ashift;
106         err = sysctl_handle_64(oidp, &val, 0, req);
107         if (err != 0 || req->newptr == NULL)
108                 return (err);
109
110         if (val > SPA_MAXASHIFT || val < zfs_min_auto_ashift)
111                 return (EINVAL);
112
113         zfs_max_auto_ashift = val;
114
115         return (0);
116 }
117 SYSCTL_PROC(_vfs_zfs, OID_AUTO, max_auto_ashift,
118     CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
119     sysctl_vfs_zfs_max_auto_ashift, "QU",
120     "Max ashift used when optimising for logical -> physical sectors size on "
121     "new top-level vdevs.");
122
123 static int
124 sysctl_vfs_zfs_min_auto_ashift(SYSCTL_HANDLER_ARGS)
125 {
126         uint64_t val;
127         int err;
128
129         val = zfs_min_auto_ashift;
130         err = sysctl_handle_64(oidp, &val, 0, req);
131         if (err != 0 || req->newptr == NULL)
132                 return (err);
133
134         if (val < SPA_MINASHIFT || val > zfs_max_auto_ashift)
135                 return (EINVAL);
136
137         zfs_min_auto_ashift = val;
138
139         return (0);
140 }
141 SYSCTL_PROC(_vfs_zfs, OID_AUTO, min_auto_ashift,
142     CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
143     sysctl_vfs_zfs_min_auto_ashift, "QU",
144     "Min ashift used when creating new top-level vdevs.");
145
146 static vdev_ops_t *vdev_ops_table[] = {
147         &vdev_root_ops,
148         &vdev_raidz_ops,
149         &vdev_mirror_ops,
150         &vdev_replacing_ops,
151         &vdev_spare_ops,
152 #ifdef _KERNEL
153         &vdev_geom_ops,
154 #else
155         &vdev_disk_ops,
156 #endif
157         &vdev_file_ops,
158         &vdev_missing_ops,
159         &vdev_hole_ops,
160         &vdev_indirect_ops,
161         NULL
162 };
163
164
165 /* maximum number of metaslabs per top-level vdev */
166 int vdev_max_ms_count = 200;
167 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, max_ms_count, CTLFLAG_RDTUN,
168     &vdev_max_ms_count, 0,
169     "Maximum number of metaslabs per top-level vdev");
170
171 /* minimum amount of metaslabs per top-level vdev */
172 int vdev_min_ms_count = 16;
173 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, min_ms_count, CTLFLAG_RDTUN,
174     &vdev_min_ms_count, 0,
175     "Minimum number of metaslabs per top-level vdev");
176
177 /* see comment in vdev_metaslab_set_size() */
178 int vdev_default_ms_shift = 29;
179 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, default_ms_shift, CTLFLAG_RDTUN,
180     &vdev_default_ms_shift, 0,
181     "Shift between vdev size and number of metaslabs");
182
183 boolean_t vdev_validate_skip = B_FALSE;
184
185 /*
186  * Since the DTL space map of a vdev is not expected to have a lot of
187  * entries, we default its block size to 4K.
188  */
189 int vdev_dtl_sm_blksz = (1 << 12);
190 SYSCTL_INT(_vfs_zfs, OID_AUTO, dtl_sm_blksz, CTLFLAG_RDTUN,
191     &vdev_dtl_sm_blksz, 0,
192     "Block size for DTL space map.  Power of 2 and greater than 4096.");
193
194 /*
195  * vdev-wide space maps that have lots of entries written to them at
196  * the end of each transaction can benefit from a higher I/O bandwidth
197  * (e.g. vdev_obsolete_sm), thus we default their block size to 128K.
198  */
199 int vdev_standard_sm_blksz = (1 << 17);
200 SYSCTL_INT(_vfs_zfs, OID_AUTO, standard_sm_blksz, CTLFLAG_RDTUN,
201     &vdev_standard_sm_blksz, 0,
202     "Block size for standard space map.  Power of 2 and greater than 4096.");
203
204 /*PRINTFLIKE2*/
205 void
206 vdev_dbgmsg(vdev_t *vd, const char *fmt, ...)
207 {
208         va_list adx;
209         char buf[256];
210
211         va_start(adx, fmt);
212         (void) vsnprintf(buf, sizeof (buf), fmt, adx);
213         va_end(adx);
214
215         if (vd->vdev_path != NULL) {
216                 zfs_dbgmsg("%s vdev '%s': %s", vd->vdev_ops->vdev_op_type,
217                     vd->vdev_path, buf);
218         } else {
219                 zfs_dbgmsg("%s-%llu vdev (guid %llu): %s",
220                     vd->vdev_ops->vdev_op_type,
221                     (u_longlong_t)vd->vdev_id,
222                     (u_longlong_t)vd->vdev_guid, buf);
223         }
224 }
225
226 void
227 vdev_dbgmsg_print_tree(vdev_t *vd, int indent)
228 {
229         char state[20];
230
231         if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops) {
232                 zfs_dbgmsg("%*svdev %u: %s", indent, "", vd->vdev_id,
233                     vd->vdev_ops->vdev_op_type);
234                 return;
235         }
236
237         switch (vd->vdev_state) {
238         case VDEV_STATE_UNKNOWN:
239                 (void) snprintf(state, sizeof (state), "unknown");
240                 break;
241         case VDEV_STATE_CLOSED:
242                 (void) snprintf(state, sizeof (state), "closed");
243                 break;
244         case VDEV_STATE_OFFLINE:
245                 (void) snprintf(state, sizeof (state), "offline");
246                 break;
247         case VDEV_STATE_REMOVED:
248                 (void) snprintf(state, sizeof (state), "removed");
249                 break;
250         case VDEV_STATE_CANT_OPEN:
251                 (void) snprintf(state, sizeof (state), "can't open");
252                 break;
253         case VDEV_STATE_FAULTED:
254                 (void) snprintf(state, sizeof (state), "faulted");
255                 break;
256         case VDEV_STATE_DEGRADED:
257                 (void) snprintf(state, sizeof (state), "degraded");
258                 break;
259         case VDEV_STATE_HEALTHY:
260                 (void) snprintf(state, sizeof (state), "healthy");
261                 break;
262         default:
263                 (void) snprintf(state, sizeof (state), "<state %u>",
264                     (uint_t)vd->vdev_state);
265         }
266
267         zfs_dbgmsg("%*svdev %u: %s%s, guid: %llu, path: %s, %s", indent,
268             "", (int)vd->vdev_id, vd->vdev_ops->vdev_op_type,
269             vd->vdev_islog ? " (log)" : "",
270             (u_longlong_t)vd->vdev_guid,
271             vd->vdev_path ? vd->vdev_path : "N/A", state);
272
273         for (uint64_t i = 0; i < vd->vdev_children; i++)
274                 vdev_dbgmsg_print_tree(vd->vdev_child[i], indent + 2);
275 }
276
277 /*
278  * Given a vdev type, return the appropriate ops vector.
279  */
280 static vdev_ops_t *
281 vdev_getops(const char *type)
282 {
283         vdev_ops_t *ops, **opspp;
284
285         for (opspp = vdev_ops_table; (ops = *opspp) != NULL; opspp++)
286                 if (strcmp(ops->vdev_op_type, type) == 0)
287                         break;
288
289         return (ops);
290 }
291
292 /*
293  * Default asize function: return the MAX of psize with the asize of
294  * all children.  This is what's used by anything other than RAID-Z.
295  */
296 uint64_t
297 vdev_default_asize(vdev_t *vd, uint64_t psize)
298 {
299         uint64_t asize = P2ROUNDUP(psize, 1ULL << vd->vdev_top->vdev_ashift);
300         uint64_t csize;
301
302         for (int c = 0; c < vd->vdev_children; c++) {
303                 csize = vdev_psize_to_asize(vd->vdev_child[c], psize);
304                 asize = MAX(asize, csize);
305         }
306
307         return (asize);
308 }
309
310 /*
311  * Get the minimum allocatable size. We define the allocatable size as
312  * the vdev's asize rounded to the nearest metaslab. This allows us to
313  * replace or attach devices which don't have the same physical size but
314  * can still satisfy the same number of allocations.
315  */
316 uint64_t
317 vdev_get_min_asize(vdev_t *vd)
318 {
319         vdev_t *pvd = vd->vdev_parent;
320
321         /*
322          * If our parent is NULL (inactive spare or cache) or is the root,
323          * just return our own asize.
324          */
325         if (pvd == NULL)
326                 return (vd->vdev_asize);
327
328         /*
329          * The top-level vdev just returns the allocatable size rounded
330          * to the nearest metaslab.
331          */
332         if (vd == vd->vdev_top)
333                 return (P2ALIGN(vd->vdev_asize, 1ULL << vd->vdev_ms_shift));
334
335         /*
336          * The allocatable space for a raidz vdev is N * sizeof(smallest child),
337          * so each child must provide at least 1/Nth of its asize.
338          */
339         if (pvd->vdev_ops == &vdev_raidz_ops)
340                 return ((pvd->vdev_min_asize + pvd->vdev_children - 1) /
341                     pvd->vdev_children);
342
343         return (pvd->vdev_min_asize);
344 }
345
346 void
347 vdev_set_min_asize(vdev_t *vd)
348 {
349         vd->vdev_min_asize = vdev_get_min_asize(vd);
350
351         for (int c = 0; c < vd->vdev_children; c++)
352                 vdev_set_min_asize(vd->vdev_child[c]);
353 }
354
355 vdev_t *
356 vdev_lookup_top(spa_t *spa, uint64_t vdev)
357 {
358         vdev_t *rvd = spa->spa_root_vdev;
359
360         ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
361
362         if (vdev < rvd->vdev_children) {
363                 ASSERT(rvd->vdev_child[vdev] != NULL);
364                 return (rvd->vdev_child[vdev]);
365         }
366
367         return (NULL);
368 }
369
370 vdev_t *
371 vdev_lookup_by_guid(vdev_t *vd, uint64_t guid)
372 {
373         vdev_t *mvd;
374
375         if (vd->vdev_guid == guid)
376                 return (vd);
377
378         for (int c = 0; c < vd->vdev_children; c++)
379                 if ((mvd = vdev_lookup_by_guid(vd->vdev_child[c], guid)) !=
380                     NULL)
381                         return (mvd);
382
383         return (NULL);
384 }
385
386 static int
387 vdev_count_leaves_impl(vdev_t *vd)
388 {
389         int n = 0;
390
391         if (vd->vdev_ops->vdev_op_leaf)
392                 return (1);
393
394         for (int c = 0; c < vd->vdev_children; c++)
395                 n += vdev_count_leaves_impl(vd->vdev_child[c]);
396
397         return (n);
398 }
399
400 int
401 vdev_count_leaves(spa_t *spa)
402 {
403         return (vdev_count_leaves_impl(spa->spa_root_vdev));
404 }
405
406 void
407 vdev_add_child(vdev_t *pvd, vdev_t *cvd)
408 {
409         size_t oldsize, newsize;
410         uint64_t id = cvd->vdev_id;
411         vdev_t **newchild;
412         spa_t *spa = cvd->vdev_spa;
413
414         ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
415         ASSERT(cvd->vdev_parent == NULL);
416
417         cvd->vdev_parent = pvd;
418
419         if (pvd == NULL)
420                 return;
421
422         ASSERT(id >= pvd->vdev_children || pvd->vdev_child[id] == NULL);
423
424         oldsize = pvd->vdev_children * sizeof (vdev_t *);
425         pvd->vdev_children = MAX(pvd->vdev_children, id + 1);
426         newsize = pvd->vdev_children * sizeof (vdev_t *);
427
428         newchild = kmem_zalloc(newsize, KM_SLEEP);
429         if (pvd->vdev_child != NULL) {
430                 bcopy(pvd->vdev_child, newchild, oldsize);
431                 kmem_free(pvd->vdev_child, oldsize);
432         }
433
434         pvd->vdev_child = newchild;
435         pvd->vdev_child[id] = cvd;
436
437         cvd->vdev_top = (pvd->vdev_top ? pvd->vdev_top: cvd);
438         ASSERT(cvd->vdev_top->vdev_parent->vdev_parent == NULL);
439
440         /*
441          * Walk up all ancestors to update guid sum.
442          */
443         for (; pvd != NULL; pvd = pvd->vdev_parent)
444                 pvd->vdev_guid_sum += cvd->vdev_guid_sum;
445 }
446
447 void
448 vdev_remove_child(vdev_t *pvd, vdev_t *cvd)
449 {
450         int c;
451         uint_t id = cvd->vdev_id;
452
453         ASSERT(cvd->vdev_parent == pvd);
454
455         if (pvd == NULL)
456                 return;
457
458         ASSERT(id < pvd->vdev_children);
459         ASSERT(pvd->vdev_child[id] == cvd);
460
461         pvd->vdev_child[id] = NULL;
462         cvd->vdev_parent = NULL;
463
464         for (c = 0; c < pvd->vdev_children; c++)
465                 if (pvd->vdev_child[c])
466                         break;
467
468         if (c == pvd->vdev_children) {
469                 kmem_free(pvd->vdev_child, c * sizeof (vdev_t *));
470                 pvd->vdev_child = NULL;
471                 pvd->vdev_children = 0;
472         }
473
474         /*
475          * Walk up all ancestors to update guid sum.
476          */
477         for (; pvd != NULL; pvd = pvd->vdev_parent)
478                 pvd->vdev_guid_sum -= cvd->vdev_guid_sum;
479 }
480
481 /*
482  * Remove any holes in the child array.
483  */
484 void
485 vdev_compact_children(vdev_t *pvd)
486 {
487         vdev_t **newchild, *cvd;
488         int oldc = pvd->vdev_children;
489         int newc;
490
491         ASSERT(spa_config_held(pvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
492
493         for (int c = newc = 0; c < oldc; c++)
494                 if (pvd->vdev_child[c])
495                         newc++;
496
497         newchild = kmem_alloc(newc * sizeof (vdev_t *), KM_SLEEP);
498
499         for (int c = newc = 0; c < oldc; c++) {
500                 if ((cvd = pvd->vdev_child[c]) != NULL) {
501                         newchild[newc] = cvd;
502                         cvd->vdev_id = newc++;
503                 }
504         }
505
506         kmem_free(pvd->vdev_child, oldc * sizeof (vdev_t *));
507         pvd->vdev_child = newchild;
508         pvd->vdev_children = newc;
509 }
510
511 /*
512  * Allocate and minimally initialize a vdev_t.
513  */
514 vdev_t *
515 vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
516 {
517         vdev_t *vd;
518         vdev_indirect_config_t *vic;
519
520         vd = kmem_zalloc(sizeof (vdev_t), KM_SLEEP);
521         vic = &vd->vdev_indirect_config;
522
523         if (spa->spa_root_vdev == NULL) {
524                 ASSERT(ops == &vdev_root_ops);
525                 spa->spa_root_vdev = vd;
526                 spa->spa_load_guid = spa_generate_guid(NULL);
527         }
528
529         if (guid == 0 && ops != &vdev_hole_ops) {
530                 if (spa->spa_root_vdev == vd) {
531                         /*
532                          * The root vdev's guid will also be the pool guid,
533                          * which must be unique among all pools.
534                          */
535                         guid = spa_generate_guid(NULL);
536                 } else {
537                         /*
538                          * Any other vdev's guid must be unique within the pool.
539                          */
540                         guid = spa_generate_guid(spa);
541                 }
542                 ASSERT(!spa_guid_exists(spa_guid(spa), guid));
543         }
544
545         vd->vdev_spa = spa;
546         vd->vdev_id = id;
547         vd->vdev_guid = guid;
548         vd->vdev_guid_sum = guid;
549         vd->vdev_ops = ops;
550         vd->vdev_state = VDEV_STATE_CLOSED;
551         vd->vdev_ishole = (ops == &vdev_hole_ops);
552         vic->vic_prev_indirect_vdev = UINT64_MAX;
553
554         rw_init(&vd->vdev_indirect_rwlock, NULL, RW_DEFAULT, NULL);
555         mutex_init(&vd->vdev_obsolete_lock, NULL, MUTEX_DEFAULT, NULL);
556         vd->vdev_obsolete_segments = range_tree_create(NULL, NULL);
557
558         mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_DEFAULT, NULL);
559         mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL);
560         mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL);
561         mutex_init(&vd->vdev_queue_lock, NULL, MUTEX_DEFAULT, NULL);
562         for (int t = 0; t < DTL_TYPES; t++) {
563                 vd->vdev_dtl[t] = range_tree_create(NULL, NULL);
564         }
565         txg_list_create(&vd->vdev_ms_list, spa,
566             offsetof(struct metaslab, ms_txg_node));
567         txg_list_create(&vd->vdev_dtl_list, spa,
568             offsetof(struct vdev, vdev_dtl_node));
569         vd->vdev_stat.vs_timestamp = gethrtime();
570         vdev_queue_init(vd);
571         vdev_cache_init(vd);
572
573         return (vd);
574 }
575
576 /*
577  * Allocate a new vdev.  The 'alloctype' is used to control whether we are
578  * creating a new vdev or loading an existing one - the behavior is slightly
579  * different for each case.
580  */
581 int
582 vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id,
583     int alloctype)
584 {
585         vdev_ops_t *ops;
586         char *type;
587         uint64_t guid = 0, islog, nparity;
588         vdev_t *vd;
589         vdev_indirect_config_t *vic;
590
591         ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
592
593         if (nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) != 0)
594                 return (SET_ERROR(EINVAL));
595
596         if ((ops = vdev_getops(type)) == NULL)
597                 return (SET_ERROR(EINVAL));
598
599         /*
600          * If this is a load, get the vdev guid from the nvlist.
601          * Otherwise, vdev_alloc_common() will generate one for us.
602          */
603         if (alloctype == VDEV_ALLOC_LOAD) {
604                 uint64_t label_id;
605
606                 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ID, &label_id) ||
607                     label_id != id)
608                         return (SET_ERROR(EINVAL));
609
610                 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
611                         return (SET_ERROR(EINVAL));
612         } else if (alloctype == VDEV_ALLOC_SPARE) {
613                 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
614                         return (SET_ERROR(EINVAL));
615         } else if (alloctype == VDEV_ALLOC_L2CACHE) {
616                 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
617                         return (SET_ERROR(EINVAL));
618         } else if (alloctype == VDEV_ALLOC_ROOTPOOL) {
619                 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
620                         return (SET_ERROR(EINVAL));
621         }
622
623         /*
624          * The first allocated vdev must be of type 'root'.
625          */
626         if (ops != &vdev_root_ops && spa->spa_root_vdev == NULL)
627                 return (SET_ERROR(EINVAL));
628
629         /*
630          * Determine whether we're a log vdev.
631          */
632         islog = 0;
633         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &islog);
634         if (islog && spa_version(spa) < SPA_VERSION_SLOGS)
635                 return (SET_ERROR(ENOTSUP));
636
637         if (ops == &vdev_hole_ops && spa_version(spa) < SPA_VERSION_HOLES)
638                 return (SET_ERROR(ENOTSUP));
639
640         /*
641          * Set the nparity property for RAID-Z vdevs.
642          */
643         nparity = -1ULL;
644         if (ops == &vdev_raidz_ops) {
645                 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY,
646                     &nparity) == 0) {
647                         if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY)
648                                 return (SET_ERROR(EINVAL));
649                         /*
650                          * Previous versions could only support 1 or 2 parity
651                          * device.
652                          */
653                         if (nparity > 1 &&
654                             spa_version(spa) < SPA_VERSION_RAIDZ2)
655                                 return (SET_ERROR(ENOTSUP));
656                         if (nparity > 2 &&
657                             spa_version(spa) < SPA_VERSION_RAIDZ3)
658                                 return (SET_ERROR(ENOTSUP));
659                 } else {
660                         /*
661                          * We require the parity to be specified for SPAs that
662                          * support multiple parity levels.
663                          */
664                         if (spa_version(spa) >= SPA_VERSION_RAIDZ2)
665                                 return (SET_ERROR(EINVAL));
666                         /*
667                          * Otherwise, we default to 1 parity device for RAID-Z.
668                          */
669                         nparity = 1;
670                 }
671         } else {
672                 nparity = 0;
673         }
674         ASSERT(nparity != -1ULL);
675
676         vd = vdev_alloc_common(spa, id, guid, ops);
677         vic = &vd->vdev_indirect_config;
678
679         vd->vdev_islog = islog;
680         vd->vdev_nparity = nparity;
681
682         if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &vd->vdev_path) == 0)
683                 vd->vdev_path = spa_strdup(vd->vdev_path);
684         if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &vd->vdev_devid) == 0)
685                 vd->vdev_devid = spa_strdup(vd->vdev_devid);
686         if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH,
687             &vd->vdev_physpath) == 0)
688                 vd->vdev_physpath = spa_strdup(vd->vdev_physpath);
689         if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &vd->vdev_fru) == 0)
690                 vd->vdev_fru = spa_strdup(vd->vdev_fru);
691
692         /*
693          * Set the whole_disk property.  If it's not specified, leave the value
694          * as -1.
695          */
696         if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
697             &vd->vdev_wholedisk) != 0)
698                 vd->vdev_wholedisk = -1ULL;
699
700         ASSERT0(vic->vic_mapping_object);
701         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_INDIRECT_OBJECT,
702             &vic->vic_mapping_object);
703         ASSERT0(vic->vic_births_object);
704         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_INDIRECT_BIRTHS,
705             &vic->vic_births_object);
706         ASSERT3U(vic->vic_prev_indirect_vdev, ==, UINT64_MAX);
707         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_PREV_INDIRECT_VDEV,
708             &vic->vic_prev_indirect_vdev);
709
710         /*
711          * Look for the 'not present' flag.  This will only be set if the device
712          * was not present at the time of import.
713          */
714         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT,
715             &vd->vdev_not_present);
716
717         /*
718          * Get the alignment requirement.
719          */
720         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT, &vd->vdev_ashift);
721
722         /*
723          * Retrieve the vdev creation time.
724          */
725         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_CREATE_TXG,
726             &vd->vdev_crtxg);
727
728         /*
729          * If we're a top-level vdev, try to load the allocation parameters.
730          */
731         if (parent && !parent->vdev_parent &&
732             (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
733                 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
734                     &vd->vdev_ms_array);
735                 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
736                     &vd->vdev_ms_shift);
737                 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE,
738                     &vd->vdev_asize);
739                 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVING,
740                     &vd->vdev_removing);
741                 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_VDEV_TOP_ZAP,
742                     &vd->vdev_top_zap);
743         } else {
744                 ASSERT0(vd->vdev_top_zap);
745         }
746
747         if (parent && !parent->vdev_parent && alloctype != VDEV_ALLOC_ATTACH) {
748                 ASSERT(alloctype == VDEV_ALLOC_LOAD ||
749                     alloctype == VDEV_ALLOC_ADD ||
750                     alloctype == VDEV_ALLOC_SPLIT ||
751                     alloctype == VDEV_ALLOC_ROOTPOOL);
752                 vd->vdev_mg = metaslab_group_create(islog ?
753                     spa_log_class(spa) : spa_normal_class(spa), vd);
754         }
755
756         if (vd->vdev_ops->vdev_op_leaf &&
757             (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
758                 (void) nvlist_lookup_uint64(nv,
759                     ZPOOL_CONFIG_VDEV_LEAF_ZAP, &vd->vdev_leaf_zap);
760         } else {
761                 ASSERT0(vd->vdev_leaf_zap);
762         }
763
764         /*
765          * If we're a leaf vdev, try to load the DTL object and other state.
766          */
767
768         if (vd->vdev_ops->vdev_op_leaf &&
769             (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE ||
770             alloctype == VDEV_ALLOC_ROOTPOOL)) {
771                 if (alloctype == VDEV_ALLOC_LOAD) {
772                         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL,
773                             &vd->vdev_dtl_object);
774                         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE,
775                             &vd->vdev_unspare);
776                 }
777
778                 if (alloctype == VDEV_ALLOC_ROOTPOOL) {
779                         uint64_t spare = 0;
780
781                         if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE,
782                             &spare) == 0 && spare)
783                                 spa_spare_add(vd);
784                 }
785
786                 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE,
787                     &vd->vdev_offline);
788
789                 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_RESILVER_TXG,
790                     &vd->vdev_resilver_txg);
791
792                 /*
793                  * When importing a pool, we want to ignore the persistent fault
794                  * state, as the diagnosis made on another system may not be
795                  * valid in the current context.  Local vdevs will
796                  * remain in the faulted state.
797                  */
798                 if (spa_load_state(spa) == SPA_LOAD_OPEN) {
799                         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED,
800                             &vd->vdev_faulted);
801                         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED,
802                             &vd->vdev_degraded);
803                         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED,
804                             &vd->vdev_removed);
805
806                         if (vd->vdev_faulted || vd->vdev_degraded) {
807                                 char *aux;
808
809                                 vd->vdev_label_aux =
810                                     VDEV_AUX_ERR_EXCEEDED;
811                                 if (nvlist_lookup_string(nv,
812                                     ZPOOL_CONFIG_AUX_STATE, &aux) == 0 &&
813                                     strcmp(aux, "external") == 0)
814                                         vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
815                         }
816                 }
817         }
818
819         /*
820          * Add ourselves to the parent's list of children.
821          */
822         vdev_add_child(parent, vd);
823
824         *vdp = vd;
825
826         return (0);
827 }
828
829 void
830 vdev_free(vdev_t *vd)
831 {
832         spa_t *spa = vd->vdev_spa;
833
834         /*
835          * vdev_free() implies closing the vdev first.  This is simpler than
836          * trying to ensure complicated semantics for all callers.
837          */
838         vdev_close(vd);
839
840         ASSERT(!list_link_active(&vd->vdev_config_dirty_node));
841         ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
842
843         /*
844          * Free all children.
845          */
846         for (int c = 0; c < vd->vdev_children; c++)
847                 vdev_free(vd->vdev_child[c]);
848
849         ASSERT(vd->vdev_child == NULL);
850         ASSERT(vd->vdev_guid_sum == vd->vdev_guid);
851
852         /*
853          * Discard allocation state.
854          */
855         if (vd->vdev_mg != NULL) {
856                 vdev_metaslab_fini(vd);
857                 metaslab_group_destroy(vd->vdev_mg);
858         }
859
860         ASSERT0(vd->vdev_stat.vs_space);
861         ASSERT0(vd->vdev_stat.vs_dspace);
862         ASSERT0(vd->vdev_stat.vs_alloc);
863
864         /*
865          * Remove this vdev from its parent's child list.
866          */
867         vdev_remove_child(vd->vdev_parent, vd);
868
869         ASSERT(vd->vdev_parent == NULL);
870
871         /*
872          * Clean up vdev structure.
873          */
874         vdev_queue_fini(vd);
875         vdev_cache_fini(vd);
876
877         if (vd->vdev_path)
878                 spa_strfree(vd->vdev_path);
879         if (vd->vdev_devid)
880                 spa_strfree(vd->vdev_devid);
881         if (vd->vdev_physpath)
882                 spa_strfree(vd->vdev_physpath);
883         if (vd->vdev_fru)
884                 spa_strfree(vd->vdev_fru);
885
886         if (vd->vdev_isspare)
887                 spa_spare_remove(vd);
888         if (vd->vdev_isl2cache)
889                 spa_l2cache_remove(vd);
890
891         txg_list_destroy(&vd->vdev_ms_list);
892         txg_list_destroy(&vd->vdev_dtl_list);
893
894         mutex_enter(&vd->vdev_dtl_lock);
895         space_map_close(vd->vdev_dtl_sm);
896         for (int t = 0; t < DTL_TYPES; t++) {
897                 range_tree_vacate(vd->vdev_dtl[t], NULL, NULL);
898                 range_tree_destroy(vd->vdev_dtl[t]);
899         }
900         mutex_exit(&vd->vdev_dtl_lock);
901
902         EQUIV(vd->vdev_indirect_births != NULL,
903             vd->vdev_indirect_mapping != NULL);
904         if (vd->vdev_indirect_births != NULL) {
905                 vdev_indirect_mapping_close(vd->vdev_indirect_mapping);
906                 vdev_indirect_births_close(vd->vdev_indirect_births);
907         }
908
909         if (vd->vdev_obsolete_sm != NULL) {
910                 ASSERT(vd->vdev_removing ||
911                     vd->vdev_ops == &vdev_indirect_ops);
912                 space_map_close(vd->vdev_obsolete_sm);
913                 vd->vdev_obsolete_sm = NULL;
914         }
915         range_tree_destroy(vd->vdev_obsolete_segments);
916         rw_destroy(&vd->vdev_indirect_rwlock);
917         mutex_destroy(&vd->vdev_obsolete_lock);
918
919         mutex_destroy(&vd->vdev_queue_lock);
920         mutex_destroy(&vd->vdev_dtl_lock);
921         mutex_destroy(&vd->vdev_stat_lock);
922         mutex_destroy(&vd->vdev_probe_lock);
923
924         if (vd == spa->spa_root_vdev)
925                 spa->spa_root_vdev = NULL;
926
927         kmem_free(vd, sizeof (vdev_t));
928 }
929
930 /*
931  * Transfer top-level vdev state from svd to tvd.
932  */
933 static void
934 vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
935 {
936         spa_t *spa = svd->vdev_spa;
937         metaslab_t *msp;
938         vdev_t *vd;
939         int t;
940
941         ASSERT(tvd == tvd->vdev_top);
942
943         tvd->vdev_ms_array = svd->vdev_ms_array;
944         tvd->vdev_ms_shift = svd->vdev_ms_shift;
945         tvd->vdev_ms_count = svd->vdev_ms_count;
946         tvd->vdev_top_zap = svd->vdev_top_zap;
947
948         svd->vdev_ms_array = 0;
949         svd->vdev_ms_shift = 0;
950         svd->vdev_ms_count = 0;
951         svd->vdev_top_zap = 0;
952
953         if (tvd->vdev_mg)
954                 ASSERT3P(tvd->vdev_mg, ==, svd->vdev_mg);
955         tvd->vdev_mg = svd->vdev_mg;
956         tvd->vdev_ms = svd->vdev_ms;
957
958         svd->vdev_mg = NULL;
959         svd->vdev_ms = NULL;
960
961         if (tvd->vdev_mg != NULL)
962                 tvd->vdev_mg->mg_vd = tvd;
963
964         tvd->vdev_checkpoint_sm = svd->vdev_checkpoint_sm;
965         svd->vdev_checkpoint_sm = NULL;
966
967         tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
968         tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
969         tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
970
971         svd->vdev_stat.vs_alloc = 0;
972         svd->vdev_stat.vs_space = 0;
973         svd->vdev_stat.vs_dspace = 0;
974
975         for (t = 0; t < TXG_SIZE; t++) {
976                 while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
977                         (void) txg_list_add(&tvd->vdev_ms_list, msp, t);
978                 while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
979                         (void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
980                 if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
981                         (void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
982         }
983
984         if (list_link_active(&svd->vdev_config_dirty_node)) {
985                 vdev_config_clean(svd);
986                 vdev_config_dirty(tvd);
987         }
988
989         if (list_link_active(&svd->vdev_state_dirty_node)) {
990                 vdev_state_clean(svd);
991                 vdev_state_dirty(tvd);
992         }
993
994         tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
995         svd->vdev_deflate_ratio = 0;
996
997         tvd->vdev_islog = svd->vdev_islog;
998         svd->vdev_islog = 0;
999 }
1000
1001 static void
1002 vdev_top_update(vdev_t *tvd, vdev_t *vd)
1003 {
1004         if (vd == NULL)
1005                 return;
1006
1007         vd->vdev_top = tvd;
1008
1009         for (int c = 0; c < vd->vdev_children; c++)
1010                 vdev_top_update(tvd, vd->vdev_child[c]);
1011 }
1012
1013 /*
1014  * Add a mirror/replacing vdev above an existing vdev.
1015  */
1016 vdev_t *
1017 vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
1018 {
1019         spa_t *spa = cvd->vdev_spa;
1020         vdev_t *pvd = cvd->vdev_parent;
1021         vdev_t *mvd;
1022
1023         ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1024
1025         mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
1026
1027         mvd->vdev_asize = cvd->vdev_asize;
1028         mvd->vdev_min_asize = cvd->vdev_min_asize;
1029         mvd->vdev_max_asize = cvd->vdev_max_asize;
1030         mvd->vdev_psize = cvd->vdev_psize;
1031         mvd->vdev_ashift = cvd->vdev_ashift;
1032         mvd->vdev_logical_ashift = cvd->vdev_logical_ashift;
1033         mvd->vdev_physical_ashift = cvd->vdev_physical_ashift;
1034         mvd->vdev_state = cvd->vdev_state;
1035         mvd->vdev_crtxg = cvd->vdev_crtxg;
1036
1037         vdev_remove_child(pvd, cvd);
1038         vdev_add_child(pvd, mvd);
1039         cvd->vdev_id = mvd->vdev_children;
1040         vdev_add_child(mvd, cvd);
1041         vdev_top_update(cvd->vdev_top, cvd->vdev_top);
1042
1043         if (mvd == mvd->vdev_top)
1044                 vdev_top_transfer(cvd, mvd);
1045
1046         return (mvd);
1047 }
1048
1049 /*
1050  * Remove a 1-way mirror/replacing vdev from the tree.
1051  */
1052 void
1053 vdev_remove_parent(vdev_t *cvd)
1054 {
1055         vdev_t *mvd = cvd->vdev_parent;
1056         vdev_t *pvd = mvd->vdev_parent;
1057
1058         ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1059
1060         ASSERT(mvd->vdev_children == 1);
1061         ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
1062             mvd->vdev_ops == &vdev_replacing_ops ||
1063             mvd->vdev_ops == &vdev_spare_ops);
1064         cvd->vdev_ashift = mvd->vdev_ashift;
1065         cvd->vdev_logical_ashift = mvd->vdev_logical_ashift;
1066         cvd->vdev_physical_ashift = mvd->vdev_physical_ashift;
1067
1068         vdev_remove_child(mvd, cvd);
1069         vdev_remove_child(pvd, mvd);
1070
1071         /*
1072          * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
1073          * Otherwise, we could have detached an offline device, and when we
1074          * go to import the pool we'll think we have two top-level vdevs,
1075          * instead of a different version of the same top-level vdev.
1076          */
1077         if (mvd->vdev_top == mvd) {
1078                 uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
1079                 cvd->vdev_orig_guid = cvd->vdev_guid;
1080                 cvd->vdev_guid += guid_delta;
1081                 cvd->vdev_guid_sum += guid_delta;
1082         }
1083         cvd->vdev_id = mvd->vdev_id;
1084         vdev_add_child(pvd, cvd);
1085         vdev_top_update(cvd->vdev_top, cvd->vdev_top);
1086
1087         if (cvd == cvd->vdev_top)
1088                 vdev_top_transfer(mvd, cvd);
1089
1090         ASSERT(mvd->vdev_children == 0);
1091         vdev_free(mvd);
1092 }
1093
1094 int
1095 vdev_metaslab_init(vdev_t *vd, uint64_t txg)
1096 {
1097         spa_t *spa = vd->vdev_spa;
1098         objset_t *mos = spa->spa_meta_objset;
1099         uint64_t m;
1100         uint64_t oldc = vd->vdev_ms_count;
1101         uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
1102         metaslab_t **mspp;
1103         int error;
1104
1105         ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER));
1106
1107         /*
1108          * This vdev is not being allocated from yet or is a hole.
1109          */
1110         if (vd->vdev_ms_shift == 0)
1111                 return (0);
1112
1113         ASSERT(!vd->vdev_ishole);
1114
1115         ASSERT(oldc <= newc);
1116
1117         mspp = kmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
1118
1119         if (oldc != 0) {
1120                 bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp));
1121                 kmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
1122         }
1123
1124         vd->vdev_ms = mspp;
1125         vd->vdev_ms_count = newc;
1126
1127         for (m = oldc; m < newc; m++) {
1128                 uint64_t object = 0;
1129
1130                 /*
1131                  * vdev_ms_array may be 0 if we are creating the "fake"
1132                  * metaslabs for an indirect vdev for zdb's leak detection.
1133                  * See zdb_leak_init().
1134                  */
1135                 if (txg == 0 && vd->vdev_ms_array != 0) {
1136                         error = dmu_read(mos, vd->vdev_ms_array,
1137                             m * sizeof (uint64_t), sizeof (uint64_t), &object,
1138                             DMU_READ_PREFETCH);
1139                         if (error != 0) {
1140                                 vdev_dbgmsg(vd, "unable to read the metaslab "
1141                                     "array [error=%d]", error);
1142                                 return (error);
1143                         }
1144                 }
1145
1146                 error = metaslab_init(vd->vdev_mg, m, object, txg,
1147                     &(vd->vdev_ms[m]));
1148                 if (error != 0) {
1149                         vdev_dbgmsg(vd, "metaslab_init failed [error=%d]",
1150                             error);
1151                         return (error);
1152                 }
1153         }
1154
1155         if (txg == 0)
1156                 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER);
1157
1158         /*
1159          * If the vdev is being removed we don't activate
1160          * the metaslabs since we want to ensure that no new
1161          * allocations are performed on this device.
1162          */
1163         if (oldc == 0 && !vd->vdev_removing)
1164                 metaslab_group_activate(vd->vdev_mg);
1165
1166         if (txg == 0)
1167                 spa_config_exit(spa, SCL_ALLOC, FTAG);
1168
1169         return (0);
1170 }
1171
1172 void
1173 vdev_metaslab_fini(vdev_t *vd)
1174 {
1175         if (vd->vdev_checkpoint_sm != NULL) {
1176                 ASSERT(spa_feature_is_active(vd->vdev_spa,
1177                     SPA_FEATURE_POOL_CHECKPOINT));
1178                 space_map_close(vd->vdev_checkpoint_sm);
1179                 /*
1180                  * Even though we close the space map, we need to set its
1181                  * pointer to NULL. The reason is that vdev_metaslab_fini()
1182                  * may be called multiple times for certain operations
1183                  * (i.e. when destroying a pool) so we need to ensure that
1184                  * this clause never executes twice. This logic is similar
1185                  * to the one used for the vdev_ms clause below.
1186                  */
1187                 vd->vdev_checkpoint_sm = NULL;
1188         }
1189
1190         if (vd->vdev_ms != NULL) {
1191                 uint64_t count = vd->vdev_ms_count;
1192
1193                 metaslab_group_passivate(vd->vdev_mg);
1194                 for (uint64_t m = 0; m < count; m++) {
1195                         metaslab_t *msp = vd->vdev_ms[m];
1196
1197                         if (msp != NULL)
1198                                 metaslab_fini(msp);
1199                 }
1200                 kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
1201                 vd->vdev_ms = NULL;
1202
1203                 vd->vdev_ms_count = 0;
1204         }
1205         ASSERT0(vd->vdev_ms_count);
1206 }
1207
1208 typedef struct vdev_probe_stats {
1209         boolean_t       vps_readable;
1210         boolean_t       vps_writeable;
1211         int             vps_flags;
1212 } vdev_probe_stats_t;
1213
1214 static void
1215 vdev_probe_done(zio_t *zio)
1216 {
1217         spa_t *spa = zio->io_spa;
1218         vdev_t *vd = zio->io_vd;
1219         vdev_probe_stats_t *vps = zio->io_private;
1220
1221         ASSERT(vd->vdev_probe_zio != NULL);
1222
1223         if (zio->io_type == ZIO_TYPE_READ) {
1224                 if (zio->io_error == 0)
1225                         vps->vps_readable = 1;
1226                 if (zio->io_error == 0 && spa_writeable(spa)) {
1227                         zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
1228                             zio->io_offset, zio->io_size, zio->io_abd,
1229                             ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1230                             ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
1231                 } else {
1232                         abd_free(zio->io_abd);
1233                 }
1234         } else if (zio->io_type == ZIO_TYPE_WRITE) {
1235                 if (zio->io_error == 0)
1236                         vps->vps_writeable = 1;
1237                 abd_free(zio->io_abd);
1238         } else if (zio->io_type == ZIO_TYPE_NULL) {
1239                 zio_t *pio;
1240
1241                 vd->vdev_cant_read |= !vps->vps_readable;
1242                 vd->vdev_cant_write |= !vps->vps_writeable;
1243
1244                 if (vdev_readable(vd) &&
1245                     (vdev_writeable(vd) || !spa_writeable(spa))) {
1246                         zio->io_error = 0;
1247                 } else {
1248                         ASSERT(zio->io_error != 0);
1249                         vdev_dbgmsg(vd, "failed probe");
1250                         zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
1251                             spa, vd, NULL, 0, 0);
1252                         zio->io_error = SET_ERROR(ENXIO);
1253                 }
1254
1255                 mutex_enter(&vd->vdev_probe_lock);
1256                 ASSERT(vd->vdev_probe_zio == zio);
1257                 vd->vdev_probe_zio = NULL;
1258                 mutex_exit(&vd->vdev_probe_lock);
1259
1260                 zio_link_t *zl = NULL;
1261                 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
1262                         if (!vdev_accessible(vd, pio))
1263                                 pio->io_error = SET_ERROR(ENXIO);
1264
1265                 kmem_free(vps, sizeof (*vps));
1266         }
1267 }
1268
1269 /*
1270  * Determine whether this device is accessible.
1271  *
1272  * Read and write to several known locations: the pad regions of each
1273  * vdev label but the first, which we leave alone in case it contains
1274  * a VTOC.
1275  */
1276 zio_t *
1277 vdev_probe(vdev_t *vd, zio_t *zio)
1278 {
1279         spa_t *spa = vd->vdev_spa;
1280         vdev_probe_stats_t *vps = NULL;
1281         zio_t *pio;
1282
1283         ASSERT(vd->vdev_ops->vdev_op_leaf);
1284
1285         /*
1286          * Don't probe the probe.
1287          */
1288         if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
1289                 return (NULL);
1290
1291         /*
1292          * To prevent 'probe storms' when a device fails, we create
1293          * just one probe i/o at a time.  All zios that want to probe
1294          * this vdev will become parents of the probe io.
1295          */
1296         mutex_enter(&vd->vdev_probe_lock);
1297
1298         if ((pio = vd->vdev_probe_zio) == NULL) {
1299                 vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
1300
1301                 vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
1302                     ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
1303                     ZIO_FLAG_TRYHARD;
1304
1305                 if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
1306                         /*
1307                          * vdev_cant_read and vdev_cant_write can only
1308                          * transition from TRUE to FALSE when we have the
1309                          * SCL_ZIO lock as writer; otherwise they can only
1310                          * transition from FALSE to TRUE.  This ensures that
1311                          * any zio looking at these values can assume that
1312                          * failures persist for the life of the I/O.  That's
1313                          * important because when a device has intermittent
1314                          * connectivity problems, we want to ensure that
1315                          * they're ascribed to the device (ENXIO) and not
1316                          * the zio (EIO).
1317                          *
1318                          * Since we hold SCL_ZIO as writer here, clear both
1319                          * values so the probe can reevaluate from first
1320                          * principles.
1321                          */
1322                         vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
1323                         vd->vdev_cant_read = B_FALSE;
1324                         vd->vdev_cant_write = B_FALSE;
1325                 }
1326
1327                 vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
1328                     vdev_probe_done, vps,
1329                     vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
1330
1331                 /*
1332                  * We can't change the vdev state in this context, so we
1333                  * kick off an async task to do it on our behalf.
1334                  */
1335                 if (zio != NULL) {
1336                         vd->vdev_probe_wanted = B_TRUE;
1337                         spa_async_request(spa, SPA_ASYNC_PROBE);
1338                 }
1339         }
1340
1341         if (zio != NULL)
1342                 zio_add_child(zio, pio);
1343
1344         mutex_exit(&vd->vdev_probe_lock);
1345
1346         if (vps == NULL) {
1347                 ASSERT(zio != NULL);
1348                 return (NULL);
1349         }
1350
1351         for (int l = 1; l < VDEV_LABELS; l++) {
1352                 zio_nowait(zio_read_phys(pio, vd,
1353                     vdev_label_offset(vd->vdev_psize, l,
1354                     offsetof(vdev_label_t, vl_pad2)), VDEV_PAD_SIZE,
1355                     abd_alloc_for_io(VDEV_PAD_SIZE, B_TRUE),
1356                     ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1357                     ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1358         }
1359
1360         if (zio == NULL)
1361                 return (pio);
1362
1363         zio_nowait(pio);
1364         return (NULL);
1365 }
1366
1367 static void
1368 vdev_open_child(void *arg)
1369 {
1370         vdev_t *vd = arg;
1371
1372         vd->vdev_open_thread = curthread;
1373         vd->vdev_open_error = vdev_open(vd);
1374         vd->vdev_open_thread = NULL;
1375 }
1376
1377 boolean_t
1378 vdev_uses_zvols(vdev_t *vd)
1379 {
1380         if (vd->vdev_path && strncmp(vd->vdev_path, ZVOL_DIR,
1381             strlen(ZVOL_DIR)) == 0)
1382                 return (B_TRUE);
1383         for (int c = 0; c < vd->vdev_children; c++)
1384                 if (vdev_uses_zvols(vd->vdev_child[c]))
1385                         return (B_TRUE);
1386         return (B_FALSE);
1387 }
1388
1389 void
1390 vdev_open_children(vdev_t *vd)
1391 {
1392         taskq_t *tq;
1393         int children = vd->vdev_children;
1394
1395         /*
1396          * in order to handle pools on top of zvols, do the opens
1397          * in a single thread so that the same thread holds the
1398          * spa_namespace_lock
1399          */
1400         if (B_TRUE || vdev_uses_zvols(vd)) {
1401                 for (int c = 0; c < children; c++)
1402                         vd->vdev_child[c]->vdev_open_error =
1403                             vdev_open(vd->vdev_child[c]);
1404                 return;
1405         }
1406         tq = taskq_create("vdev_open", children, minclsyspri,
1407             children, children, TASKQ_PREPOPULATE);
1408
1409         for (int c = 0; c < children; c++)
1410                 VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c],
1411                     TQ_SLEEP) != 0);
1412
1413         taskq_destroy(tq);
1414 }
1415
1416 /*
1417  * Compute the raidz-deflation ratio.  Note, we hard-code
1418  * in 128k (1 << 17) because it is the "typical" blocksize.
1419  * Even though SPA_MAXBLOCKSIZE changed, this algorithm can not change,
1420  * otherwise it would inconsistently account for existing bp's.
1421  */
1422 static void
1423 vdev_set_deflate_ratio(vdev_t *vd)
1424 {
1425         if (vd == vd->vdev_top && !vd->vdev_ishole && vd->vdev_ashift != 0) {
1426                 vd->vdev_deflate_ratio = (1 << 17) /
1427                     (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
1428         }
1429 }
1430
1431 /*
1432  * Prepare a virtual device for access.
1433  */
1434 int
1435 vdev_open(vdev_t *vd)
1436 {
1437         spa_t *spa = vd->vdev_spa;
1438         int error;
1439         uint64_t osize = 0;
1440         uint64_t max_osize = 0;
1441         uint64_t asize, max_asize, psize;
1442         uint64_t logical_ashift = 0;
1443         uint64_t physical_ashift = 0;
1444
1445         ASSERT(vd->vdev_open_thread == curthread ||
1446             spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1447         ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1448             vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1449             vd->vdev_state == VDEV_STATE_OFFLINE);
1450
1451         vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1452         vd->vdev_cant_read = B_FALSE;
1453         vd->vdev_cant_write = B_FALSE;
1454         vd->vdev_notrim = B_FALSE;
1455         vd->vdev_min_asize = vdev_get_min_asize(vd);
1456
1457         /*
1458          * If this vdev is not removed, check its fault status.  If it's
1459          * faulted, bail out of the open.
1460          */
1461         if (!vd->vdev_removed && vd->vdev_faulted) {
1462                 ASSERT(vd->vdev_children == 0);
1463                 ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1464                     vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1465                 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1466                     vd->vdev_label_aux);
1467                 return (SET_ERROR(ENXIO));
1468         } else if (vd->vdev_offline) {
1469                 ASSERT(vd->vdev_children == 0);
1470                 vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1471                 return (SET_ERROR(ENXIO));
1472         }
1473
1474         error = vd->vdev_ops->vdev_op_open(vd, &osize, &max_osize,
1475             &logical_ashift, &physical_ashift);
1476
1477         /*
1478          * Reset the vdev_reopening flag so that we actually close
1479          * the vdev on error.
1480          */
1481         vd->vdev_reopening = B_FALSE;
1482         if (zio_injection_enabled && error == 0)
1483                 error = zio_handle_device_injection(vd, NULL, ENXIO);
1484
1485         if (error) {
1486                 if (vd->vdev_removed &&
1487                     vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1488                         vd->vdev_removed = B_FALSE;
1489
1490                 if (vd->vdev_stat.vs_aux == VDEV_AUX_CHILDREN_OFFLINE) {
1491                         vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE,
1492                             vd->vdev_stat.vs_aux);
1493                 } else {
1494                         vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1495                             vd->vdev_stat.vs_aux);
1496                 }
1497                 return (error);
1498         }
1499
1500         vd->vdev_removed = B_FALSE;
1501
1502         /*
1503          * Recheck the faulted flag now that we have confirmed that
1504          * the vdev is accessible.  If we're faulted, bail.
1505          */
1506         if (vd->vdev_faulted) {
1507                 ASSERT(vd->vdev_children == 0);
1508                 ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1509                     vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1510                 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1511                     vd->vdev_label_aux);
1512                 return (SET_ERROR(ENXIO));
1513         }
1514
1515         if (vd->vdev_degraded) {
1516                 ASSERT(vd->vdev_children == 0);
1517                 vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1518                     VDEV_AUX_ERR_EXCEEDED);
1519         } else {
1520                 vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1521         }
1522
1523         /*
1524          * For hole or missing vdevs we just return success.
1525          */
1526         if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1527                 return (0);
1528
1529         if (zfs_trim_enabled && !vd->vdev_notrim && vd->vdev_ops->vdev_op_leaf)
1530                 trim_map_create(vd);
1531
1532         for (int c = 0; c < vd->vdev_children; c++) {
1533                 if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1534                         vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1535                             VDEV_AUX_NONE);
1536                         break;
1537                 }
1538         }
1539
1540         osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1541         max_osize = P2ALIGN(max_osize, (uint64_t)sizeof (vdev_label_t));
1542
1543         if (vd->vdev_children == 0) {
1544                 if (osize < SPA_MINDEVSIZE) {
1545                         vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1546                             VDEV_AUX_TOO_SMALL);
1547                         return (SET_ERROR(EOVERFLOW));
1548                 }
1549                 psize = osize;
1550                 asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1551                 max_asize = max_osize - (VDEV_LABEL_START_SIZE +
1552                     VDEV_LABEL_END_SIZE);
1553         } else {
1554                 if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1555                     (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1556                         vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1557                             VDEV_AUX_TOO_SMALL);
1558                         return (SET_ERROR(EOVERFLOW));
1559                 }
1560                 psize = 0;
1561                 asize = osize;
1562                 max_asize = max_osize;
1563         }
1564
1565         vd->vdev_psize = psize;
1566
1567         /*
1568          * Make sure the allocatable size hasn't shrunk too much.
1569          */
1570         if (asize < vd->vdev_min_asize) {
1571                 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1572                     VDEV_AUX_BAD_LABEL);
1573                 return (SET_ERROR(EINVAL));
1574         }
1575
1576         vd->vdev_physical_ashift =
1577             MAX(physical_ashift, vd->vdev_physical_ashift);
1578         vd->vdev_logical_ashift = MAX(logical_ashift, vd->vdev_logical_ashift);
1579         vd->vdev_ashift = MAX(vd->vdev_logical_ashift, vd->vdev_ashift);
1580
1581         if (vd->vdev_logical_ashift > SPA_MAXASHIFT) {
1582                 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1583                     VDEV_AUX_ASHIFT_TOO_BIG);
1584                 return (EINVAL);
1585         }
1586
1587         if (vd->vdev_asize == 0) {
1588                 /*
1589                  * This is the first-ever open, so use the computed values.
1590                  * For testing purposes, a higher ashift can be requested.
1591                  */
1592                 vd->vdev_asize = asize;
1593                 vd->vdev_max_asize = max_asize;
1594         } else {
1595                 /*
1596                  * Make sure the alignment requirement hasn't increased.
1597                  */
1598                 if (vd->vdev_ashift > vd->vdev_top->vdev_ashift &&
1599                     vd->vdev_ops->vdev_op_leaf) {
1600                         vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1601                             VDEV_AUX_BAD_LABEL);
1602                         return (EINVAL);
1603                 }
1604                 vd->vdev_max_asize = max_asize;
1605         }
1606
1607         /*
1608          * If all children are healthy we update asize if either:
1609          * The asize has increased, due to a device expansion caused by dynamic
1610          * LUN growth or vdev replacement, and automatic expansion is enabled;
1611          * making the additional space available.
1612          *
1613          * The asize has decreased, due to a device shrink usually caused by a
1614          * vdev replace with a smaller device. This ensures that calculations
1615          * based of max_asize and asize e.g. esize are always valid. It's safe
1616          * to do this as we've already validated that asize is greater than
1617          * vdev_min_asize.
1618          */
1619         if (vd->vdev_state == VDEV_STATE_HEALTHY &&
1620             ((asize > vd->vdev_asize &&
1621             (vd->vdev_expanding || spa->spa_autoexpand)) ||
1622             (asize < vd->vdev_asize)))
1623                 vd->vdev_asize = asize;
1624
1625         vdev_set_min_asize(vd);
1626
1627         /*
1628          * Ensure we can issue some IO before declaring the
1629          * vdev open for business.
1630          */
1631         if (vd->vdev_ops->vdev_op_leaf &&
1632             (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1633                 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1634                     VDEV_AUX_ERR_EXCEEDED);
1635                 return (error);
1636         }
1637
1638         /*
1639          * Track the min and max ashift values for normal data devices.
1640          */
1641         if (vd->vdev_top == vd && vd->vdev_ashift != 0 &&
1642             !vd->vdev_islog && vd->vdev_aux == NULL) {
1643                 if (vd->vdev_ashift > spa->spa_max_ashift)
1644                         spa->spa_max_ashift = vd->vdev_ashift;
1645                 if (vd->vdev_ashift < spa->spa_min_ashift)
1646                         spa->spa_min_ashift = vd->vdev_ashift;
1647         }
1648
1649         /*
1650          * If a leaf vdev has a DTL, and seems healthy, then kick off a
1651          * resilver.  But don't do this if we are doing a reopen for a scrub,
1652          * since this would just restart the scrub we are already doing.
1653          */
1654         if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1655             vdev_resilver_needed(vd, NULL, NULL))
1656                 spa_async_request(spa, SPA_ASYNC_RESILVER);
1657
1658         return (0);
1659 }
1660
1661 /*
1662  * Called once the vdevs are all opened, this routine validates the label
1663  * contents. This needs to be done before vdev_load() so that we don't
1664  * inadvertently do repair I/Os to the wrong device.
1665  *
1666  * This function will only return failure if one of the vdevs indicates that it
1667  * has since been destroyed or exported.  This is only possible if
1668  * /etc/zfs/zpool.cache was readonly at the time.  Otherwise, the vdev state
1669  * will be updated but the function will return 0.
1670  */
1671 int
1672 vdev_validate(vdev_t *vd)
1673 {
1674         spa_t *spa = vd->vdev_spa;
1675         nvlist_t *label;
1676         uint64_t guid = 0, aux_guid = 0, top_guid;
1677         uint64_t state;
1678         nvlist_t *nvl;
1679         uint64_t txg;
1680
1681         if (vdev_validate_skip)
1682                 return (0);
1683
1684         for (uint64_t c = 0; c < vd->vdev_children; c++)
1685                 if (vdev_validate(vd->vdev_child[c]) != 0)
1686                         return (SET_ERROR(EBADF));
1687
1688         /*
1689          * If the device has already failed, or was marked offline, don't do
1690          * any further validation.  Otherwise, label I/O will fail and we will
1691          * overwrite the previous state.
1692          */
1693         if (!vd->vdev_ops->vdev_op_leaf || !vdev_readable(vd))
1694                 return (0);
1695
1696         /*
1697          * If we are performing an extreme rewind, we allow for a label that
1698          * was modified at a point after the current txg.
1699          */
1700         if (spa->spa_extreme_rewind || spa_last_synced_txg(spa) == 0)
1701                 txg = UINT64_MAX;
1702         else
1703                 txg = spa_last_synced_txg(spa);
1704
1705         if ((label = vdev_label_read_config(vd, txg)) == NULL) {
1706                 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1707                     VDEV_AUX_BAD_LABEL);
1708                 vdev_dbgmsg(vd, "vdev_validate: failed reading config");
1709                 return (0);
1710         }
1711
1712         /*
1713          * Determine if this vdev has been split off into another
1714          * pool.  If so, then refuse to open it.
1715          */
1716         if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID,
1717             &aux_guid) == 0 && aux_guid == spa_guid(spa)) {
1718                 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1719                     VDEV_AUX_SPLIT_POOL);
1720                 nvlist_free(label);
1721                 vdev_dbgmsg(vd, "vdev_validate: vdev split into other pool");
1722                 return (0);
1723         }
1724
1725         if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_GUID, &guid) != 0) {
1726                 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1727                     VDEV_AUX_CORRUPT_DATA);
1728                 nvlist_free(label);
1729                 vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
1730                     ZPOOL_CONFIG_POOL_GUID);
1731                 return (0);
1732         }
1733
1734         /*
1735          * If config is not trusted then ignore the spa guid check. This is
1736          * necessary because if the machine crashed during a re-guid the new
1737          * guid might have been written to all of the vdev labels, but not the
1738          * cached config. The check will be performed again once we have the
1739          * trusted config from the MOS.
1740          */
1741         if (spa->spa_trust_config && guid != spa_guid(spa)) {
1742                 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1743                     VDEV_AUX_CORRUPT_DATA);
1744                 nvlist_free(label);
1745                 vdev_dbgmsg(vd, "vdev_validate: vdev label pool_guid doesn't "
1746                     "match config (%llu != %llu)", (u_longlong_t)guid,
1747                     (u_longlong_t)spa_guid(spa));
1748                 return (0);
1749         }
1750
1751         if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl)
1752             != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID,
1753             &aux_guid) != 0)
1754                 aux_guid = 0;
1755
1756         if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0) {
1757                 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1758                     VDEV_AUX_CORRUPT_DATA);
1759                 nvlist_free(label);
1760                 vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
1761                     ZPOOL_CONFIG_GUID);
1762                 return (0);
1763         }
1764
1765         if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID, &top_guid)
1766             != 0) {
1767                 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1768                     VDEV_AUX_CORRUPT_DATA);
1769                 nvlist_free(label);
1770                 vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
1771                     ZPOOL_CONFIG_TOP_GUID);
1772                 return (0);
1773         }
1774
1775         /*
1776          * If this vdev just became a top-level vdev because its sibling was
1777          * detached, it will have adopted the parent's vdev guid -- but the
1778          * label may or may not be on disk yet. Fortunately, either version
1779          * of the label will have the same top guid, so if we're a top-level
1780          * vdev, we can safely compare to that instead.
1781          * However, if the config comes from a cachefile that failed to update
1782          * after the detach, a top-level vdev will appear as a non top-level
1783          * vdev in the config. Also relax the constraints if we perform an
1784          * extreme rewind.
1785          *
1786          * If we split this vdev off instead, then we also check the
1787          * original pool's guid. We don't want to consider the vdev
1788          * corrupt if it is partway through a split operation.
1789          */
1790         if (vd->vdev_guid != guid && vd->vdev_guid != aux_guid) {
1791                 boolean_t mismatch = B_FALSE;
1792                 if (spa->spa_trust_config && !spa->spa_extreme_rewind) {
1793                         if (vd != vd->vdev_top || vd->vdev_guid != top_guid)
1794                                 mismatch = B_TRUE;
1795                 } else {
1796                         if (vd->vdev_guid != top_guid &&
1797                             vd->vdev_top->vdev_guid != guid)
1798                                 mismatch = B_TRUE;
1799                 }
1800
1801                 if (mismatch) {
1802                         vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1803                             VDEV_AUX_CORRUPT_DATA);
1804                         nvlist_free(label);
1805                         vdev_dbgmsg(vd, "vdev_validate: config guid "
1806                             "doesn't match label guid");
1807                         vdev_dbgmsg(vd, "CONFIG: guid %llu, top_guid %llu",
1808                             (u_longlong_t)vd->vdev_guid,
1809                             (u_longlong_t)vd->vdev_top->vdev_guid);
1810                         vdev_dbgmsg(vd, "LABEL: guid %llu, top_guid %llu, "
1811                             "aux_guid %llu", (u_longlong_t)guid,
1812                             (u_longlong_t)top_guid, (u_longlong_t)aux_guid);
1813                         return (0);
1814                 }
1815         }
1816
1817         if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1818             &state) != 0) {
1819                 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1820                     VDEV_AUX_CORRUPT_DATA);
1821                 nvlist_free(label);
1822                 vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
1823                     ZPOOL_CONFIG_POOL_STATE);
1824                 return (0);
1825         }
1826
1827         nvlist_free(label);
1828
1829         /*
1830          * If this is a verbatim import, no need to check the
1831          * state of the pool.
1832          */
1833         if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) &&
1834             spa_load_state(spa) == SPA_LOAD_OPEN &&
1835             state != POOL_STATE_ACTIVE) {
1836                 vdev_dbgmsg(vd, "vdev_validate: invalid pool state (%llu) "
1837                     "for spa %s", (u_longlong_t)state, spa->spa_name);
1838                 return (SET_ERROR(EBADF));
1839         }
1840
1841         /*
1842          * If we were able to open and validate a vdev that was
1843          * previously marked permanently unavailable, clear that state
1844          * now.
1845          */
1846         if (vd->vdev_not_present)
1847                 vd->vdev_not_present = 0;
1848
1849         return (0);
1850 }
1851
1852 static void
1853 vdev_copy_path_impl(vdev_t *svd, vdev_t *dvd)
1854 {
1855         if (svd->vdev_path != NULL && dvd->vdev_path != NULL) {
1856                 if (strcmp(svd->vdev_path, dvd->vdev_path) != 0) {
1857                         zfs_dbgmsg("vdev_copy_path: vdev %llu: path changed "
1858                             "from '%s' to '%s'", (u_longlong_t)dvd->vdev_guid,
1859                             dvd->vdev_path, svd->vdev_path);
1860                         spa_strfree(dvd->vdev_path);
1861                         dvd->vdev_path = spa_strdup(svd->vdev_path);
1862                 }
1863         } else if (svd->vdev_path != NULL) {
1864                 dvd->vdev_path = spa_strdup(svd->vdev_path);
1865                 zfs_dbgmsg("vdev_copy_path: vdev %llu: path set to '%s'",
1866                     (u_longlong_t)dvd->vdev_guid, dvd->vdev_path);
1867         }
1868 }
1869
1870 /*
1871  * Recursively copy vdev paths from one vdev to another. Source and destination
1872  * vdev trees must have same geometry otherwise return error. Intended to copy
1873  * paths from userland config into MOS config.
1874  */
1875 int
1876 vdev_copy_path_strict(vdev_t *svd, vdev_t *dvd)
1877 {
1878         if ((svd->vdev_ops == &vdev_missing_ops) ||
1879             (svd->vdev_ishole && dvd->vdev_ishole) ||
1880             (dvd->vdev_ops == &vdev_indirect_ops))
1881                 return (0);
1882
1883         if (svd->vdev_ops != dvd->vdev_ops) {
1884                 vdev_dbgmsg(svd, "vdev_copy_path: vdev type mismatch: %s != %s",
1885                     svd->vdev_ops->vdev_op_type, dvd->vdev_ops->vdev_op_type);
1886                 return (SET_ERROR(EINVAL));
1887         }
1888
1889         if (svd->vdev_guid != dvd->vdev_guid) {
1890                 vdev_dbgmsg(svd, "vdev_copy_path: guids mismatch (%llu != "
1891                     "%llu)", (u_longlong_t)svd->vdev_guid,
1892                     (u_longlong_t)dvd->vdev_guid);
1893                 return (SET_ERROR(EINVAL));
1894         }
1895
1896         if (svd->vdev_children != dvd->vdev_children) {
1897                 vdev_dbgmsg(svd, "vdev_copy_path: children count mismatch: "
1898                     "%llu != %llu", (u_longlong_t)svd->vdev_children,
1899                     (u_longlong_t)dvd->vdev_children);
1900                 return (SET_ERROR(EINVAL));
1901         }
1902
1903         for (uint64_t i = 0; i < svd->vdev_children; i++) {
1904                 int error = vdev_copy_path_strict(svd->vdev_child[i],
1905                     dvd->vdev_child[i]);
1906                 if (error != 0)
1907                         return (error);
1908         }
1909
1910         if (svd->vdev_ops->vdev_op_leaf)
1911                 vdev_copy_path_impl(svd, dvd);
1912
1913         return (0);
1914 }
1915
1916 static void
1917 vdev_copy_path_search(vdev_t *stvd, vdev_t *dvd)
1918 {
1919         ASSERT(stvd->vdev_top == stvd);
1920         ASSERT3U(stvd->vdev_id, ==, dvd->vdev_top->vdev_id);
1921
1922         for (uint64_t i = 0; i < dvd->vdev_children; i++) {
1923                 vdev_copy_path_search(stvd, dvd->vdev_child[i]);
1924         }
1925
1926         if (!dvd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(dvd))
1927                 return;
1928
1929         /*
1930          * The idea here is that while a vdev can shift positions within
1931          * a top vdev (when replacing, attaching mirror, etc.) it cannot
1932          * step outside of it.
1933          */
1934         vdev_t *vd = vdev_lookup_by_guid(stvd, dvd->vdev_guid);
1935
1936         if (vd == NULL || vd->vdev_ops != dvd->vdev_ops)
1937                 return;
1938
1939         ASSERT(vd->vdev_ops->vdev_op_leaf);
1940
1941         vdev_copy_path_impl(vd, dvd);
1942 }
1943
1944 /*
1945  * Recursively copy vdev paths from one root vdev to another. Source and
1946  * destination vdev trees may differ in geometry. For each destination leaf
1947  * vdev, search a vdev with the same guid and top vdev id in the source.
1948  * Intended to copy paths from userland config into MOS config.
1949  */
1950 void
1951 vdev_copy_path_relaxed(vdev_t *srvd, vdev_t *drvd)
1952 {
1953         uint64_t children = MIN(srvd->vdev_children, drvd->vdev_children);
1954         ASSERT(srvd->vdev_ops == &vdev_root_ops);
1955         ASSERT(drvd->vdev_ops == &vdev_root_ops);
1956
1957         for (uint64_t i = 0; i < children; i++) {
1958                 vdev_copy_path_search(srvd->vdev_child[i],
1959                     drvd->vdev_child[i]);
1960         }
1961 }
1962
1963 /*
1964  * Close a virtual device.
1965  */
1966 void
1967 vdev_close(vdev_t *vd)
1968 {
1969         spa_t *spa = vd->vdev_spa;
1970         vdev_t *pvd = vd->vdev_parent;
1971
1972         ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1973
1974         /*
1975          * If our parent is reopening, then we are as well, unless we are
1976          * going offline.
1977          */
1978         if (pvd != NULL && pvd->vdev_reopening)
1979                 vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline);
1980
1981         vd->vdev_ops->vdev_op_close(vd);
1982
1983         vdev_cache_purge(vd);
1984
1985         if (vd->vdev_ops->vdev_op_leaf)
1986                 trim_map_destroy(vd);
1987
1988         /*
1989          * We record the previous state before we close it, so that if we are
1990          * doing a reopen(), we don't generate FMA ereports if we notice that
1991          * it's still faulted.
1992          */
1993         vd->vdev_prevstate = vd->vdev_state;
1994
1995         if (vd->vdev_offline)
1996                 vd->vdev_state = VDEV_STATE_OFFLINE;
1997         else
1998                 vd->vdev_state = VDEV_STATE_CLOSED;
1999         vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
2000 }
2001
2002 void
2003 vdev_hold(vdev_t *vd)
2004 {
2005         spa_t *spa = vd->vdev_spa;
2006
2007         ASSERT(spa_is_root(spa));
2008         if (spa->spa_state == POOL_STATE_UNINITIALIZED)
2009                 return;
2010
2011         for (int c = 0; c < vd->vdev_children; c++)
2012                 vdev_hold(vd->vdev_child[c]);
2013
2014         if (vd->vdev_ops->vdev_op_leaf)
2015                 vd->vdev_ops->vdev_op_hold(vd);
2016 }
2017
2018 void
2019 vdev_rele(vdev_t *vd)
2020 {
2021         spa_t *spa = vd->vdev_spa;
2022
2023         ASSERT(spa_is_root(spa));
2024         for (int c = 0; c < vd->vdev_children; c++)
2025                 vdev_rele(vd->vdev_child[c]);
2026
2027         if (vd->vdev_ops->vdev_op_leaf)
2028                 vd->vdev_ops->vdev_op_rele(vd);
2029 }
2030
2031 /*
2032  * Reopen all interior vdevs and any unopened leaves.  We don't actually
2033  * reopen leaf vdevs which had previously been opened as they might deadlock
2034  * on the spa_config_lock.  Instead we only obtain the leaf's physical size.
2035  * If the leaf has never been opened then open it, as usual.
2036  */
2037 void
2038 vdev_reopen(vdev_t *vd)
2039 {
2040         spa_t *spa = vd->vdev_spa;
2041
2042         ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2043
2044         /* set the reopening flag unless we're taking the vdev offline */
2045         vd->vdev_reopening = !vd->vdev_offline;
2046         vdev_close(vd);
2047         (void) vdev_open(vd);
2048
2049         /*
2050          * Call vdev_validate() here to make sure we have the same device.
2051          * Otherwise, a device with an invalid label could be successfully
2052          * opened in response to vdev_reopen().
2053          */
2054         if (vd->vdev_aux) {
2055                 (void) vdev_validate_aux(vd);
2056                 if (vdev_readable(vd) && vdev_writeable(vd) &&
2057                     vd->vdev_aux == &spa->spa_l2cache &&
2058                     !l2arc_vdev_present(vd))
2059                         l2arc_add_vdev(spa, vd);
2060         } else {
2061                 (void) vdev_validate(vd);
2062         }
2063
2064         /*
2065          * Reassess parent vdev's health.
2066          */
2067         vdev_propagate_state(vd);
2068 }
2069
2070 int
2071 vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
2072 {
2073         int error;
2074
2075         /*
2076          * Normally, partial opens (e.g. of a mirror) are allowed.
2077          * For a create, however, we want to fail the request if
2078          * there are any components we can't open.
2079          */
2080         error = vdev_open(vd);
2081
2082         if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
2083                 vdev_close(vd);
2084                 return (error ? error : ENXIO);
2085         }
2086
2087         /*
2088          * Recursively load DTLs and initialize all labels.
2089          */
2090         if ((error = vdev_dtl_load(vd)) != 0 ||
2091             (error = vdev_label_init(vd, txg, isreplacing ?
2092             VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
2093                 vdev_close(vd);
2094                 return (error);
2095         }
2096
2097         return (0);
2098 }
2099
2100 void
2101 vdev_metaslab_set_size(vdev_t *vd)
2102 {
2103         uint64_t asize = vd->vdev_asize;
2104         uint64_t ms_shift = 0;
2105
2106         /*
2107          * For vdevs that are bigger than 8G the metaslab size varies in
2108          * a way that the number of metaslabs increases in powers of two,
2109          * linearly in terms of vdev_asize, starting from 16 metaslabs.
2110          * So for vdev_asize of 8G we get 16 metaslabs, for 16G, we get 32,
2111          * and so on, until we hit the maximum metaslab count limit
2112          * [vdev_max_ms_count] from which point the metaslab count stays
2113          * the same.
2114          */
2115         ms_shift = vdev_default_ms_shift;
2116
2117         if ((asize >> ms_shift) < vdev_min_ms_count) {
2118                 /*
2119                  * For devices that are less than 8G we want to have
2120                  * exactly 16 metaslabs. We don't want less as integer
2121                  * division rounds down, so less metaslabs mean more
2122                  * wasted space. We don't want more as these vdevs are
2123                  * small and in the likely event that we are running
2124                  * out of space, the SPA will have a hard time finding
2125                  * space due to fragmentation.
2126                  */
2127                 ms_shift = highbit64(asize / vdev_min_ms_count);
2128                 ms_shift = MAX(ms_shift, SPA_MAXBLOCKSHIFT);
2129
2130         } else if ((asize >> ms_shift) > vdev_max_ms_count) {
2131                 ms_shift = highbit64(asize / vdev_max_ms_count);
2132         }
2133
2134         vd->vdev_ms_shift = ms_shift;
2135         ASSERT3U(vd->vdev_ms_shift, >=, SPA_MAXBLOCKSHIFT);
2136 }
2137
2138 /*
2139  * Maximize performance by inflating the configured ashift for top level
2140  * vdevs to be as close to the physical ashift as possible while maintaining
2141  * administrator defined limits and ensuring it doesn't go below the
2142  * logical ashift.
2143  */
2144 void
2145 vdev_ashift_optimize(vdev_t *vd)
2146 {
2147         if (vd == vd->vdev_top) {
2148                 if (vd->vdev_ashift < vd->vdev_physical_ashift) {
2149                         vd->vdev_ashift = MIN(
2150                             MAX(zfs_max_auto_ashift, vd->vdev_ashift),
2151                             MAX(zfs_min_auto_ashift, vd->vdev_physical_ashift));
2152                 } else {
2153                         /*
2154                          * Unusual case where logical ashift > physical ashift
2155                          * so we can't cap the calculated ashift based on max
2156                          * ashift as that would cause failures.
2157                          * We still check if we need to increase it to match
2158                          * the min ashift.
2159                          */
2160                         vd->vdev_ashift = MAX(zfs_min_auto_ashift,
2161                             vd->vdev_ashift);
2162                 }
2163         }
2164 }
2165
2166 void
2167 vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
2168 {
2169         ASSERT(vd == vd->vdev_top);
2170         /* indirect vdevs don't have metaslabs or dtls */
2171         ASSERT(vdev_is_concrete(vd) || flags == 0);
2172         ASSERT(ISP2(flags));
2173         ASSERT(spa_writeable(vd->vdev_spa));
2174
2175         if (flags & VDD_METASLAB)
2176                 (void) txg_list_add(&vd->vdev_ms_list, arg, txg);
2177
2178         if (flags & VDD_DTL)
2179                 (void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
2180
2181         (void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
2182 }
2183
2184 void
2185 vdev_dirty_leaves(vdev_t *vd, int flags, uint64_t txg)
2186 {
2187         for (int c = 0; c < vd->vdev_children; c++)
2188                 vdev_dirty_leaves(vd->vdev_child[c], flags, txg);
2189
2190         if (vd->vdev_ops->vdev_op_leaf)
2191                 vdev_dirty(vd->vdev_top, flags, vd, txg);
2192 }
2193
2194 /*
2195  * DTLs.
2196  *
2197  * A vdev's DTL (dirty time log) is the set of transaction groups for which
2198  * the vdev has less than perfect replication.  There are four kinds of DTL:
2199  *
2200  * DTL_MISSING: txgs for which the vdev has no valid copies of the data
2201  *
2202  * DTL_PARTIAL: txgs for which data is available, but not fully replicated
2203  *
2204  * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
2205  *      scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
2206  *      txgs that was scrubbed.
2207  *
2208  * DTL_OUTAGE: txgs which cannot currently be read, whether due to
2209  *      persistent errors or just some device being offline.
2210  *      Unlike the other three, the DTL_OUTAGE map is not generally
2211  *      maintained; it's only computed when needed, typically to
2212  *      determine whether a device can be detached.
2213  *
2214  * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
2215  * either has the data or it doesn't.
2216  *
2217  * For interior vdevs such as mirror and RAID-Z the picture is more complex.
2218  * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
2219  * if any child is less than fully replicated, then so is its parent.
2220  * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
2221  * comprising only those txgs which appear in 'maxfaults' or more children;
2222  * those are the txgs we don't have enough replication to read.  For example,
2223  * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
2224  * thus, its DTL_MISSING consists of the set of txgs that appear in more than
2225  * two child DTL_MISSING maps.
2226  *
2227  * It should be clear from the above that to compute the DTLs and outage maps
2228  * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
2229  * Therefore, that is all we keep on disk.  When loading the pool, or after
2230  * a configuration change, we generate all other DTLs from first principles.
2231  */
2232 void
2233 vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
2234 {
2235         range_tree_t *rt = vd->vdev_dtl[t];
2236
2237         ASSERT(t < DTL_TYPES);
2238         ASSERT(vd != vd->vdev_spa->spa_root_vdev);
2239         ASSERT(spa_writeable(vd->vdev_spa));
2240
2241         mutex_enter(&vd->vdev_dtl_lock);
2242         if (!range_tree_contains(rt, txg, size))
2243                 range_tree_add(rt, txg, size);
2244         mutex_exit(&vd->vdev_dtl_lock);
2245 }
2246
2247 boolean_t
2248 vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
2249 {
2250         range_tree_t *rt = vd->vdev_dtl[t];
2251         boolean_t dirty = B_FALSE;
2252
2253         ASSERT(t < DTL_TYPES);
2254         ASSERT(vd != vd->vdev_spa->spa_root_vdev);
2255
2256         /*
2257          * While we are loading the pool, the DTLs have not been loaded yet.
2258          * Ignore the DTLs and try all devices.  This avoids a recursive
2259          * mutex enter on the vdev_dtl_lock, and also makes us try hard
2260          * when loading the pool (relying on the checksum to ensure that
2261          * we get the right data -- note that we while loading, we are
2262          * only reading the MOS, which is always checksummed).
2263          */
2264         if (vd->vdev_spa->spa_load_state != SPA_LOAD_NONE)
2265                 return (B_FALSE);
2266
2267         mutex_enter(&vd->vdev_dtl_lock);
2268         if (!range_tree_is_empty(rt))
2269                 dirty = range_tree_contains(rt, txg, size);
2270         mutex_exit(&vd->vdev_dtl_lock);
2271
2272         return (dirty);
2273 }
2274
2275 boolean_t
2276 vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
2277 {
2278         range_tree_t *rt = vd->vdev_dtl[t];
2279         boolean_t empty;
2280
2281         mutex_enter(&vd->vdev_dtl_lock);
2282         empty = range_tree_is_empty(rt);
2283         mutex_exit(&vd->vdev_dtl_lock);
2284
2285         return (empty);
2286 }
2287
2288 /*
2289  * Returns the lowest txg in the DTL range.
2290  */
2291 static uint64_t
2292 vdev_dtl_min(vdev_t *vd)
2293 {
2294         range_seg_t *rs;
2295
2296         ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
2297         ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
2298         ASSERT0(vd->vdev_children);
2299
2300         rs = avl_first(&vd->vdev_dtl[DTL_MISSING]->rt_root);
2301         return (rs->rs_start - 1);
2302 }
2303
2304 /*
2305  * Returns the highest txg in the DTL.
2306  */
2307 static uint64_t
2308 vdev_dtl_max(vdev_t *vd)
2309 {
2310         range_seg_t *rs;
2311
2312         ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
2313         ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
2314         ASSERT0(vd->vdev_children);
2315
2316         rs = avl_last(&vd->vdev_dtl[DTL_MISSING]->rt_root);
2317         return (rs->rs_end);
2318 }
2319
2320 /*
2321  * Determine if a resilvering vdev should remove any DTL entries from
2322  * its range. If the vdev was resilvering for the entire duration of the
2323  * scan then it should excise that range from its DTLs. Otherwise, this
2324  * vdev is considered partially resilvered and should leave its DTL
2325  * entries intact. The comment in vdev_dtl_reassess() describes how we
2326  * excise the DTLs.
2327  */
2328 static boolean_t
2329 vdev_dtl_should_excise(vdev_t *vd)
2330 {
2331         spa_t *spa = vd->vdev_spa;
2332         dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
2333
2334         ASSERT0(scn->scn_phys.scn_errors);
2335         ASSERT0(vd->vdev_children);
2336
2337         if (vd->vdev_state < VDEV_STATE_DEGRADED)
2338                 return (B_FALSE);
2339
2340         if (vd->vdev_resilver_txg == 0 ||
2341             range_tree_is_empty(vd->vdev_dtl[DTL_MISSING]))
2342                 return (B_TRUE);
2343
2344         /*
2345          * When a resilver is initiated the scan will assign the scn_max_txg
2346          * value to the highest txg value that exists in all DTLs. If this
2347          * device's max DTL is not part of this scan (i.e. it is not in
2348          * the range (scn_min_txg, scn_max_txg] then it is not eligible
2349          * for excision.
2350          */
2351         if (vdev_dtl_max(vd) <= scn->scn_phys.scn_max_txg) {
2352                 ASSERT3U(scn->scn_phys.scn_min_txg, <=, vdev_dtl_min(vd));
2353                 ASSERT3U(scn->scn_phys.scn_min_txg, <, vd->vdev_resilver_txg);
2354                 ASSERT3U(vd->vdev_resilver_txg, <=, scn->scn_phys.scn_max_txg);
2355                 return (B_TRUE);
2356         }
2357         return (B_FALSE);
2358 }
2359
2360 /*
2361  * Reassess DTLs after a config change or scrub completion.
2362  */
2363 void
2364 vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
2365 {
2366         spa_t *spa = vd->vdev_spa;
2367         avl_tree_t reftree;
2368         int minref;
2369
2370         ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
2371
2372         for (int c = 0; c < vd->vdev_children; c++)
2373                 vdev_dtl_reassess(vd->vdev_child[c], txg,
2374                     scrub_txg, scrub_done);
2375
2376         if (vd == spa->spa_root_vdev || !vdev_is_concrete(vd) || vd->vdev_aux)
2377                 return;
2378
2379         if (vd->vdev_ops->vdev_op_leaf) {
2380                 dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
2381
2382                 mutex_enter(&vd->vdev_dtl_lock);
2383
2384                 /*
2385                  * If we've completed a scan cleanly then determine
2386                  * if this vdev should remove any DTLs. We only want to
2387                  * excise regions on vdevs that were available during
2388                  * the entire duration of this scan.
2389                  */
2390                 if (scrub_txg != 0 &&
2391                     (spa->spa_scrub_started ||
2392                     (scn != NULL && scn->scn_phys.scn_errors == 0)) &&
2393                     vdev_dtl_should_excise(vd)) {
2394                         /*
2395                          * We completed a scrub up to scrub_txg.  If we
2396                          * did it without rebooting, then the scrub dtl
2397                          * will be valid, so excise the old region and
2398                          * fold in the scrub dtl.  Otherwise, leave the
2399                          * dtl as-is if there was an error.
2400                          *
2401                          * There's little trick here: to excise the beginning
2402                          * of the DTL_MISSING map, we put it into a reference
2403                          * tree and then add a segment with refcnt -1 that
2404                          * covers the range [0, scrub_txg).  This means
2405                          * that each txg in that range has refcnt -1 or 0.
2406                          * We then add DTL_SCRUB with a refcnt of 2, so that
2407                          * entries in the range [0, scrub_txg) will have a
2408                          * positive refcnt -- either 1 or 2.  We then convert
2409                          * the reference tree into the new DTL_MISSING map.
2410                          */
2411                         space_reftree_create(&reftree);
2412                         space_reftree_add_map(&reftree,
2413                             vd->vdev_dtl[DTL_MISSING], 1);
2414                         space_reftree_add_seg(&reftree, 0, scrub_txg, -1);
2415                         space_reftree_add_map(&reftree,
2416                             vd->vdev_dtl[DTL_SCRUB], 2);
2417                         space_reftree_generate_map(&reftree,
2418                             vd->vdev_dtl[DTL_MISSING], 1);
2419                         space_reftree_destroy(&reftree);
2420                 }
2421                 range_tree_vacate(vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
2422                 range_tree_walk(vd->vdev_dtl[DTL_MISSING],
2423                     range_tree_add, vd->vdev_dtl[DTL_PARTIAL]);
2424                 if (scrub_done)
2425                         range_tree_vacate(vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
2426                 range_tree_vacate(vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
2427                 if (!vdev_readable(vd))
2428                         range_tree_add(vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
2429                 else
2430                         range_tree_walk(vd->vdev_dtl[DTL_MISSING],
2431                             range_tree_add, vd->vdev_dtl[DTL_OUTAGE]);
2432
2433                 /*
2434                  * If the vdev was resilvering and no longer has any
2435                  * DTLs then reset its resilvering flag and dirty
2436                  * the top level so that we persist the change.
2437                  */
2438                 if (vd->vdev_resilver_txg != 0 &&
2439                     range_tree_is_empty(vd->vdev_dtl[DTL_MISSING]) &&
2440                     range_tree_is_empty(vd->vdev_dtl[DTL_OUTAGE])) {
2441                         vd->vdev_resilver_txg = 0;
2442                         vdev_config_dirty(vd->vdev_top);
2443                 }
2444
2445                 mutex_exit(&vd->vdev_dtl_lock);
2446
2447                 if (txg != 0)
2448                         vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
2449                 return;
2450         }
2451
2452         mutex_enter(&vd->vdev_dtl_lock);
2453         for (int t = 0; t < DTL_TYPES; t++) {
2454                 /* account for child's outage in parent's missing map */
2455                 int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
2456                 if (t == DTL_SCRUB)
2457                         continue;                       /* leaf vdevs only */
2458                 if (t == DTL_PARTIAL)
2459                         minref = 1;                     /* i.e. non-zero */
2460                 else if (vd->vdev_nparity != 0)
2461                         minref = vd->vdev_nparity + 1;  /* RAID-Z */
2462                 else
2463                         minref = vd->vdev_children;     /* any kind of mirror */
2464                 space_reftree_create(&reftree);
2465                 for (int c = 0; c < vd->vdev_children; c++) {
2466                         vdev_t *cvd = vd->vdev_child[c];
2467                         mutex_enter(&cvd->vdev_dtl_lock);
2468                         space_reftree_add_map(&reftree, cvd->vdev_dtl[s], 1);
2469                         mutex_exit(&cvd->vdev_dtl_lock);
2470                 }
2471                 space_reftree_generate_map(&reftree, vd->vdev_dtl[t], minref);
2472                 space_reftree_destroy(&reftree);
2473         }
2474         mutex_exit(&vd->vdev_dtl_lock);
2475 }
2476
2477 int
2478 vdev_dtl_load(vdev_t *vd)
2479 {
2480         spa_t *spa = vd->vdev_spa;
2481         objset_t *mos = spa->spa_meta_objset;
2482         int error = 0;
2483
2484         if (vd->vdev_ops->vdev_op_leaf && vd->vdev_dtl_object != 0) {
2485                 ASSERT(vdev_is_concrete(vd));
2486
2487                 error = space_map_open(&vd->vdev_dtl_sm, mos,
2488                     vd->vdev_dtl_object, 0, -1ULL, 0);
2489                 if (error)
2490                         return (error);
2491                 ASSERT(vd->vdev_dtl_sm != NULL);
2492
2493                 mutex_enter(&vd->vdev_dtl_lock);
2494
2495                 /*
2496                  * Now that we've opened the space_map we need to update
2497                  * the in-core DTL.
2498                  */
2499                 space_map_update(vd->vdev_dtl_sm);
2500
2501                 error = space_map_load(vd->vdev_dtl_sm,
2502                     vd->vdev_dtl[DTL_MISSING], SM_ALLOC);
2503                 mutex_exit(&vd->vdev_dtl_lock);
2504
2505                 return (error);
2506         }
2507
2508         for (int c = 0; c < vd->vdev_children; c++) {
2509                 error = vdev_dtl_load(vd->vdev_child[c]);
2510                 if (error != 0)
2511                         break;
2512         }
2513
2514         return (error);
2515 }
2516
2517 void
2518 vdev_destroy_unlink_zap(vdev_t *vd, uint64_t zapobj, dmu_tx_t *tx)
2519 {
2520         spa_t *spa = vd->vdev_spa;
2521
2522         VERIFY0(zap_destroy(spa->spa_meta_objset, zapobj, tx));
2523         VERIFY0(zap_remove_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
2524             zapobj, tx));
2525 }
2526
2527 uint64_t
2528 vdev_create_link_zap(vdev_t *vd, dmu_tx_t *tx)
2529 {
2530         spa_t *spa = vd->vdev_spa;
2531         uint64_t zap = zap_create(spa->spa_meta_objset, DMU_OTN_ZAP_METADATA,
2532             DMU_OT_NONE, 0, tx);
2533
2534         ASSERT(zap != 0);
2535         VERIFY0(zap_add_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
2536             zap, tx));
2537
2538         return (zap);
2539 }
2540
2541 void
2542 vdev_construct_zaps(vdev_t *vd, dmu_tx_t *tx)
2543 {
2544         if (vd->vdev_ops != &vdev_hole_ops &&
2545             vd->vdev_ops != &vdev_missing_ops &&
2546             vd->vdev_ops != &vdev_root_ops &&
2547             !vd->vdev_top->vdev_removing) {
2548                 if (vd->vdev_ops->vdev_op_leaf && vd->vdev_leaf_zap == 0) {
2549                         vd->vdev_leaf_zap = vdev_create_link_zap(vd, tx);
2550                 }
2551                 if (vd == vd->vdev_top && vd->vdev_top_zap == 0) {
2552                         vd->vdev_top_zap = vdev_create_link_zap(vd, tx);
2553                 }
2554         }
2555         for (uint64_t i = 0; i < vd->vdev_children; i++) {
2556                 vdev_construct_zaps(vd->vdev_child[i], tx);
2557         }
2558 }
2559
2560 void
2561 vdev_dtl_sync(vdev_t *vd, uint64_t txg)
2562 {
2563         spa_t *spa = vd->vdev_spa;
2564         range_tree_t *rt = vd->vdev_dtl[DTL_MISSING];
2565         objset_t *mos = spa->spa_meta_objset;
2566         range_tree_t *rtsync;
2567         dmu_tx_t *tx;
2568         uint64_t object = space_map_object(vd->vdev_dtl_sm);
2569
2570         ASSERT(vdev_is_concrete(vd));
2571         ASSERT(vd->vdev_ops->vdev_op_leaf);
2572
2573         tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2574
2575         if (vd->vdev_detached || vd->vdev_top->vdev_removing) {
2576                 mutex_enter(&vd->vdev_dtl_lock);
2577                 space_map_free(vd->vdev_dtl_sm, tx);
2578                 space_map_close(vd->vdev_dtl_sm);
2579                 vd->vdev_dtl_sm = NULL;
2580                 mutex_exit(&vd->vdev_dtl_lock);
2581
2582                 /*
2583                  * We only destroy the leaf ZAP for detached leaves or for
2584                  * removed log devices. Removed data devices handle leaf ZAP
2585                  * cleanup later, once cancellation is no longer possible.
2586                  */
2587                 if (vd->vdev_leaf_zap != 0 && (vd->vdev_detached ||
2588                     vd->vdev_top->vdev_islog)) {
2589                         vdev_destroy_unlink_zap(vd, vd->vdev_leaf_zap, tx);
2590                         vd->vdev_leaf_zap = 0;
2591                 }
2592
2593                 dmu_tx_commit(tx);
2594                 return;
2595         }
2596
2597         if (vd->vdev_dtl_sm == NULL) {
2598                 uint64_t new_object;
2599
2600                 new_object = space_map_alloc(mos, vdev_dtl_sm_blksz, tx);
2601                 VERIFY3U(new_object, !=, 0);
2602
2603                 VERIFY0(space_map_open(&vd->vdev_dtl_sm, mos, new_object,
2604                     0, -1ULL, 0));
2605                 ASSERT(vd->vdev_dtl_sm != NULL);
2606         }
2607
2608         rtsync = range_tree_create(NULL, NULL);
2609
2610         mutex_enter(&vd->vdev_dtl_lock);
2611         range_tree_walk(rt, range_tree_add, rtsync);
2612         mutex_exit(&vd->vdev_dtl_lock);
2613
2614         space_map_truncate(vd->vdev_dtl_sm, vdev_dtl_sm_blksz, tx);
2615         space_map_write(vd->vdev_dtl_sm, rtsync, SM_ALLOC, tx);
2616         range_tree_vacate(rtsync, NULL, NULL);
2617
2618         range_tree_destroy(rtsync);
2619
2620         /*
2621          * If the object for the space map has changed then dirty
2622          * the top level so that we update the config.
2623          */
2624         if (object != space_map_object(vd->vdev_dtl_sm)) {
2625                 vdev_dbgmsg(vd, "txg %llu, spa %s, DTL old object %llu, "
2626                     "new object %llu", (u_longlong_t)txg, spa_name(spa),
2627                     (u_longlong_t)object,
2628                     (u_longlong_t)space_map_object(vd->vdev_dtl_sm));
2629                 vdev_config_dirty(vd->vdev_top);
2630         }
2631
2632         dmu_tx_commit(tx);
2633
2634         mutex_enter(&vd->vdev_dtl_lock);
2635         space_map_update(vd->vdev_dtl_sm);
2636         mutex_exit(&vd->vdev_dtl_lock);
2637 }
2638
2639 /*
2640  * Determine whether the specified vdev can be offlined/detached/removed
2641  * without losing data.
2642  */
2643 boolean_t
2644 vdev_dtl_required(vdev_t *vd)
2645 {
2646         spa_t *spa = vd->vdev_spa;
2647         vdev_t *tvd = vd->vdev_top;
2648         uint8_t cant_read = vd->vdev_cant_read;
2649         boolean_t required;
2650
2651         ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2652
2653         if (vd == spa->spa_root_vdev || vd == tvd)
2654                 return (B_TRUE);
2655
2656         /*
2657          * Temporarily mark the device as unreadable, and then determine
2658          * whether this results in any DTL outages in the top-level vdev.
2659          * If not, we can safely offline/detach/remove the device.
2660          */
2661         vd->vdev_cant_read = B_TRUE;
2662         vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
2663         required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
2664         vd->vdev_cant_read = cant_read;
2665         vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
2666
2667         if (!required && zio_injection_enabled)
2668                 required = !!zio_handle_device_injection(vd, NULL, ECHILD);
2669
2670         return (required);
2671 }
2672
2673 /*
2674  * Determine if resilver is needed, and if so the txg range.
2675  */
2676 boolean_t
2677 vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
2678 {
2679         boolean_t needed = B_FALSE;
2680         uint64_t thismin = UINT64_MAX;
2681         uint64_t thismax = 0;
2682
2683         if (vd->vdev_children == 0) {
2684                 mutex_enter(&vd->vdev_dtl_lock);
2685                 if (!range_tree_is_empty(vd->vdev_dtl[DTL_MISSING]) &&
2686                     vdev_writeable(vd)) {
2687
2688                         thismin = vdev_dtl_min(vd);
2689                         thismax = vdev_dtl_max(vd);
2690                         needed = B_TRUE;
2691                 }
2692                 mutex_exit(&vd->vdev_dtl_lock);
2693         } else {
2694                 for (int c = 0; c < vd->vdev_children; c++) {
2695                         vdev_t *cvd = vd->vdev_child[c];
2696                         uint64_t cmin, cmax;
2697
2698                         if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
2699                                 thismin = MIN(thismin, cmin);
2700                                 thismax = MAX(thismax, cmax);
2701                                 needed = B_TRUE;
2702                         }
2703                 }
2704         }
2705
2706         if (needed && minp) {
2707                 *minp = thismin;
2708                 *maxp = thismax;
2709         }
2710         return (needed);
2711 }
2712
2713 /*
2714  * Gets the checkpoint space map object from the vdev's ZAP.
2715  * Returns the spacemap object, or 0 if it wasn't in the ZAP
2716  * or the ZAP doesn't exist yet.
2717  */
2718 int
2719 vdev_checkpoint_sm_object(vdev_t *vd)
2720 {
2721         ASSERT0(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER));
2722         if (vd->vdev_top_zap == 0) {
2723                 return (0);
2724         }
2725
2726         uint64_t sm_obj = 0;
2727         int err = zap_lookup(spa_meta_objset(vd->vdev_spa), vd->vdev_top_zap,
2728             VDEV_TOP_ZAP_POOL_CHECKPOINT_SM, sizeof (uint64_t), 1, &sm_obj);
2729
2730         ASSERT(err == 0 || err == ENOENT);
2731
2732         return (sm_obj);
2733 }
2734
2735 int
2736 vdev_load(vdev_t *vd)
2737 {
2738         int error = 0;
2739         /*
2740          * Recursively load all children.
2741          */
2742         for (int c = 0; c < vd->vdev_children; c++) {
2743                 error = vdev_load(vd->vdev_child[c]);
2744                 if (error != 0) {
2745                         return (error);
2746                 }
2747         }
2748
2749         vdev_set_deflate_ratio(vd);
2750
2751         /*
2752          * If this is a top-level vdev, initialize its metaslabs.
2753          */
2754         if (vd == vd->vdev_top && vdev_is_concrete(vd)) {
2755                 if (vd->vdev_ashift == 0 || vd->vdev_asize == 0) {
2756                         vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2757                             VDEV_AUX_CORRUPT_DATA);
2758                         vdev_dbgmsg(vd, "vdev_load: invalid size. ashift=%llu, "
2759                             "asize=%llu", (u_longlong_t)vd->vdev_ashift,
2760                             (u_longlong_t)vd->vdev_asize);
2761                         return (SET_ERROR(ENXIO));
2762                 } else if ((error = vdev_metaslab_init(vd, 0)) != 0) {
2763                         vdev_dbgmsg(vd, "vdev_load: metaslab_init failed "
2764                             "[error=%d]", error);
2765                         vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2766                             VDEV_AUX_CORRUPT_DATA);
2767                         return (error);
2768                 }
2769
2770                 uint64_t checkpoint_sm_obj = vdev_checkpoint_sm_object(vd);
2771                 if (checkpoint_sm_obj != 0) {
2772                         objset_t *mos = spa_meta_objset(vd->vdev_spa);
2773                         ASSERT(vd->vdev_asize != 0);
2774                         ASSERT3P(vd->vdev_checkpoint_sm, ==, NULL);
2775
2776                         if ((error = space_map_open(&vd->vdev_checkpoint_sm,
2777                             mos, checkpoint_sm_obj, 0, vd->vdev_asize,
2778                             vd->vdev_ashift))) {
2779                                 vdev_dbgmsg(vd, "vdev_load: space_map_open "
2780                                     "failed for checkpoint spacemap (obj %llu) "
2781                                     "[error=%d]",
2782                                     (u_longlong_t)checkpoint_sm_obj, error);
2783                                 return (error);
2784                         }
2785                         ASSERT3P(vd->vdev_checkpoint_sm, !=, NULL);
2786                         space_map_update(vd->vdev_checkpoint_sm);
2787
2788                         /*
2789                          * Since the checkpoint_sm contains free entries
2790                          * exclusively we can use sm_alloc to indicate the
2791                          * culmulative checkpointed space that has been freed.
2792                          */
2793                         vd->vdev_stat.vs_checkpoint_space =
2794                             -vd->vdev_checkpoint_sm->sm_alloc;
2795                         vd->vdev_spa->spa_checkpoint_info.sci_dspace +=
2796                             vd->vdev_stat.vs_checkpoint_space;
2797                 }
2798         }
2799
2800         /*
2801          * If this is a leaf vdev, load its DTL.
2802          */
2803         if (vd->vdev_ops->vdev_op_leaf && (error = vdev_dtl_load(vd)) != 0) {
2804                 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2805                     VDEV_AUX_CORRUPT_DATA);
2806                 vdev_dbgmsg(vd, "vdev_load: vdev_dtl_load failed "
2807                     "[error=%d]", error);
2808                 return (error);
2809         }
2810
2811         uint64_t obsolete_sm_object = vdev_obsolete_sm_object(vd);
2812         if (obsolete_sm_object != 0) {
2813                 objset_t *mos = vd->vdev_spa->spa_meta_objset;
2814                 ASSERT(vd->vdev_asize != 0);
2815                 ASSERT3P(vd->vdev_obsolete_sm, ==, NULL);
2816
2817                 if ((error = space_map_open(&vd->vdev_obsolete_sm, mos,
2818                     obsolete_sm_object, 0, vd->vdev_asize, 0))) {
2819                         vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2820                             VDEV_AUX_CORRUPT_DATA);
2821                         vdev_dbgmsg(vd, "vdev_load: space_map_open failed for "
2822                             "obsolete spacemap (obj %llu) [error=%d]",
2823                             (u_longlong_t)obsolete_sm_object, error);
2824                         return (error);
2825                 }
2826                 space_map_update(vd->vdev_obsolete_sm);
2827         }
2828
2829         return (0);
2830 }
2831
2832 /*
2833  * The special vdev case is used for hot spares and l2cache devices.  Its
2834  * sole purpose it to set the vdev state for the associated vdev.  To do this,
2835  * we make sure that we can open the underlying device, then try to read the
2836  * label, and make sure that the label is sane and that it hasn't been
2837  * repurposed to another pool.
2838  */
2839 int
2840 vdev_validate_aux(vdev_t *vd)
2841 {
2842         nvlist_t *label;
2843         uint64_t guid, version;
2844         uint64_t state;
2845
2846         if (!vdev_readable(vd))
2847                 return (0);
2848
2849         if ((label = vdev_label_read_config(vd, -1ULL)) == NULL) {
2850                 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2851                     VDEV_AUX_CORRUPT_DATA);
2852                 return (-1);
2853         }
2854
2855         if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
2856             !SPA_VERSION_IS_SUPPORTED(version) ||
2857             nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
2858             guid != vd->vdev_guid ||
2859             nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
2860                 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2861                     VDEV_AUX_CORRUPT_DATA);
2862                 nvlist_free(label);
2863                 return (-1);
2864         }
2865
2866         /*
2867          * We don't actually check the pool state here.  If it's in fact in
2868          * use by another pool, we update this fact on the fly when requested.
2869          */
2870         nvlist_free(label);
2871         return (0);
2872 }
2873
2874 /*
2875  * Free the objects used to store this vdev's spacemaps, and the array
2876  * that points to them.
2877  */
2878 void
2879 vdev_destroy_spacemaps(vdev_t *vd, dmu_tx_t *tx)
2880 {
2881         if (vd->vdev_ms_array == 0)
2882                 return;
2883
2884         objset_t *mos = vd->vdev_spa->spa_meta_objset;
2885         uint64_t array_count = vd->vdev_asize >> vd->vdev_ms_shift;
2886         size_t array_bytes = array_count * sizeof (uint64_t);
2887         uint64_t *smobj_array = kmem_alloc(array_bytes, KM_SLEEP);
2888         VERIFY0(dmu_read(mos, vd->vdev_ms_array, 0,
2889             array_bytes, smobj_array, 0));
2890
2891         for (uint64_t i = 0; i < array_count; i++) {
2892                 uint64_t smobj = smobj_array[i];
2893                 if (smobj == 0)
2894                         continue;
2895
2896                 space_map_free_obj(mos, smobj, tx);
2897         }
2898
2899         kmem_free(smobj_array, array_bytes);
2900         VERIFY0(dmu_object_free(mos, vd->vdev_ms_array, tx));
2901         vd->vdev_ms_array = 0;
2902 }
2903
2904 static void
2905 vdev_remove_empty(vdev_t *vd, uint64_t txg)
2906 {
2907         spa_t *spa = vd->vdev_spa;
2908         dmu_tx_t *tx;
2909
2910         ASSERT(vd == vd->vdev_top);
2911         ASSERT3U(txg, ==, spa_syncing_txg(spa));
2912
2913         if (vd->vdev_ms != NULL) {
2914                 metaslab_group_t *mg = vd->vdev_mg;
2915
2916                 metaslab_group_histogram_verify(mg);
2917                 metaslab_class_histogram_verify(mg->mg_class);
2918
2919                 for (int m = 0; m < vd->vdev_ms_count; m++) {
2920                         metaslab_t *msp = vd->vdev_ms[m];
2921
2922                         if (msp == NULL || msp->ms_sm == NULL)
2923                                 continue;
2924
2925                         mutex_enter(&msp->ms_lock);
2926                         /*
2927                          * If the metaslab was not loaded when the vdev
2928                          * was removed then the histogram accounting may
2929                          * not be accurate. Update the histogram information
2930                          * here so that we ensure that the metaslab group
2931                          * and metaslab class are up-to-date.
2932                          */
2933                         metaslab_group_histogram_remove(mg, msp);
2934
2935                         VERIFY0(space_map_allocated(msp->ms_sm));
2936                         space_map_close(msp->ms_sm);
2937                         msp->ms_sm = NULL;
2938                         mutex_exit(&msp->ms_lock);
2939                 }
2940
2941                 if (vd->vdev_checkpoint_sm != NULL) {
2942                         ASSERT(spa_has_checkpoint(spa));
2943                         space_map_close(vd->vdev_checkpoint_sm);
2944                         vd->vdev_checkpoint_sm = NULL;
2945                 }
2946
2947                 metaslab_group_histogram_verify(mg);
2948                 metaslab_class_histogram_verify(mg->mg_class);
2949                 for (int i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
2950                         ASSERT0(mg->mg_histogram[i]);
2951         }
2952
2953         tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2954         vdev_destroy_spacemaps(vd, tx);
2955
2956         if (vd->vdev_islog && vd->vdev_top_zap != 0) {
2957                 vdev_destroy_unlink_zap(vd, vd->vdev_top_zap, tx);
2958                 vd->vdev_top_zap = 0;
2959         }
2960         dmu_tx_commit(tx);
2961 }
2962
2963 void
2964 vdev_sync_done(vdev_t *vd, uint64_t txg)
2965 {
2966         metaslab_t *msp;
2967         boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg));
2968
2969         ASSERT(vdev_is_concrete(vd));
2970
2971         while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
2972                 metaslab_sync_done(msp, txg);
2973
2974         if (reassess)
2975                 metaslab_sync_reassess(vd->vdev_mg);
2976 }
2977
2978 void
2979 vdev_sync(vdev_t *vd, uint64_t txg)
2980 {
2981         spa_t *spa = vd->vdev_spa;
2982         vdev_t *lvd;
2983         metaslab_t *msp;
2984         dmu_tx_t *tx;
2985
2986         if (range_tree_space(vd->vdev_obsolete_segments) > 0) {
2987                 dmu_tx_t *tx;
2988
2989                 ASSERT(vd->vdev_removing ||
2990                     vd->vdev_ops == &vdev_indirect_ops);
2991
2992                 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2993                 vdev_indirect_sync_obsolete(vd, tx);
2994                 dmu_tx_commit(tx);
2995
2996                 /*
2997                  * If the vdev is indirect, it can't have dirty
2998                  * metaslabs or DTLs.
2999                  */
3000                 if (vd->vdev_ops == &vdev_indirect_ops) {
3001                         ASSERT(txg_list_empty(&vd->vdev_ms_list, txg));
3002                         ASSERT(txg_list_empty(&vd->vdev_dtl_list, txg));
3003                         return;
3004                 }
3005         }
3006
3007         ASSERT(vdev_is_concrete(vd));
3008
3009         if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0 &&
3010             !vd->vdev_removing) {
3011                 ASSERT(vd == vd->vdev_top);
3012                 ASSERT0(vd->vdev_indirect_config.vic_mapping_object);
3013                 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
3014                 vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
3015                     DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
3016                 ASSERT(vd->vdev_ms_array != 0);
3017                 vdev_config_dirty(vd);
3018                 dmu_tx_commit(tx);
3019         }
3020
3021         while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
3022                 metaslab_sync(msp, txg);
3023                 (void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
3024         }
3025
3026         while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
3027                 vdev_dtl_sync(lvd, txg);
3028
3029         /*
3030          * Remove the metadata associated with this vdev once it's empty.
3031          * Note that this is typically used for log/cache device removal;
3032          * we don't empty toplevel vdevs when removing them.  But if
3033          * a toplevel happens to be emptied, this is not harmful.
3034          */
3035         if (vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing) {
3036                 vdev_remove_empty(vd, txg);
3037         }
3038
3039         (void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
3040 }
3041
3042 uint64_t
3043 vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
3044 {
3045         return (vd->vdev_ops->vdev_op_asize(vd, psize));
3046 }
3047
3048 /*
3049  * Mark the given vdev faulted.  A faulted vdev behaves as if the device could
3050  * not be opened, and no I/O is attempted.
3051  */
3052 int
3053 vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
3054 {
3055         vdev_t *vd, *tvd;
3056
3057         spa_vdev_state_enter(spa, SCL_NONE);
3058
3059         if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
3060                 return (spa_vdev_state_exit(spa, NULL, ENODEV));
3061
3062         if (!vd->vdev_ops->vdev_op_leaf)
3063                 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
3064
3065         tvd = vd->vdev_top;
3066
3067         /*
3068          * We don't directly use the aux state here, but if we do a
3069          * vdev_reopen(), we need this value to be present to remember why we
3070          * were faulted.
3071          */
3072         vd->vdev_label_aux = aux;
3073
3074         /*
3075          * Faulted state takes precedence over degraded.
3076          */
3077         vd->vdev_delayed_close = B_FALSE;
3078         vd->vdev_faulted = 1ULL;
3079         vd->vdev_degraded = 0ULL;
3080         vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
3081
3082         /*
3083          * If this device has the only valid copy of the data, then
3084          * back off and simply mark the vdev as degraded instead.
3085          */
3086         if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) {
3087                 vd->vdev_degraded = 1ULL;
3088                 vd->vdev_faulted = 0ULL;
3089
3090                 /*
3091                  * If we reopen the device and it's not dead, only then do we
3092                  * mark it degraded.
3093                  */
3094                 vdev_reopen(tvd);
3095
3096                 if (vdev_readable(vd))
3097                         vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
3098         }
3099
3100         return (spa_vdev_state_exit(spa, vd, 0));
3101 }
3102
3103 /*
3104  * Mark the given vdev degraded.  A degraded vdev is purely an indication to the
3105  * user that something is wrong.  The vdev continues to operate as normal as far
3106  * as I/O is concerned.
3107  */
3108 int
3109 vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
3110 {
3111         vdev_t *vd;
3112
3113         spa_vdev_state_enter(spa, SCL_NONE);
3114
3115         if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
3116                 return (spa_vdev_state_exit(spa, NULL, ENODEV));
3117
3118         if (!vd->vdev_ops->vdev_op_leaf)
3119                 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
3120
3121         /*
3122          * If the vdev is already faulted, then don't do anything.
3123          */
3124         if (vd->vdev_faulted || vd->vdev_degraded)
3125                 return (spa_vdev_state_exit(spa, NULL, 0));
3126
3127         vd->vdev_degraded = 1ULL;
3128         if (!vdev_is_dead(vd))
3129                 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
3130                     aux);
3131
3132         return (spa_vdev_state_exit(spa, vd, 0));
3133 }
3134
3135 /*
3136  * Online the given vdev.
3137  *
3138  * If 'ZFS_ONLINE_UNSPARE' is set, it implies two things.  First, any attached
3139  * spare device should be detached when the device finishes resilvering.
3140  * Second, the online should be treated like a 'test' online case, so no FMA
3141  * events are generated if the device fails to open.
3142  */
3143 int
3144 vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
3145 {
3146         vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
3147         boolean_t wasoffline;
3148         vdev_state_t oldstate;
3149
3150         spa_vdev_state_enter(spa, SCL_NONE);
3151
3152         if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
3153                 return (spa_vdev_state_exit(spa, NULL, ENODEV));
3154
3155         if (!vd->vdev_ops->vdev_op_leaf)
3156                 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
3157
3158         wasoffline = (vd->vdev_offline || vd->vdev_tmpoffline);
3159         oldstate = vd->vdev_state;
3160
3161         tvd = vd->vdev_top;
3162         vd->vdev_offline = B_FALSE;
3163         vd->vdev_tmpoffline = B_FALSE;
3164         vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
3165         vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
3166
3167         /* XXX - L2ARC 1.0 does not support expansion */
3168         if (!vd->vdev_aux) {
3169                 for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
3170                         pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
3171         }
3172
3173         vdev_reopen(tvd);
3174         vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
3175
3176         if (!vd->vdev_aux) {
3177                 for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
3178                         pvd->vdev_expanding = B_FALSE;
3179         }
3180
3181         if (newstate)
3182                 *newstate = vd->vdev_state;
3183         if ((flags & ZFS_ONLINE_UNSPARE) &&
3184             !vdev_is_dead(vd) && vd->vdev_parent &&
3185             vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
3186             vd->vdev_parent->vdev_child[0] == vd)
3187                 vd->vdev_unspare = B_TRUE;
3188
3189         if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
3190
3191                 /* XXX - L2ARC 1.0 does not support expansion */
3192                 if (vd->vdev_aux)
3193                         return (spa_vdev_state_exit(spa, vd, ENOTSUP));
3194                 spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
3195         }
3196
3197         if (wasoffline ||
3198             (oldstate < VDEV_STATE_DEGRADED &&
3199             vd->vdev_state >= VDEV_STATE_DEGRADED))
3200                 spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_ONLINE);
3201
3202         return (spa_vdev_state_exit(spa, vd, 0));
3203 }
3204
3205 static int
3206 vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
3207 {
3208         vdev_t *vd, *tvd;
3209         int error = 0;
3210         uint64_t generation;
3211         metaslab_group_t *mg;
3212
3213 top:
3214         spa_vdev_state_enter(spa, SCL_ALLOC);
3215
3216         if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
3217                 return (spa_vdev_state_exit(spa, NULL, ENODEV));
3218
3219         if (!vd->vdev_ops->vdev_op_leaf)
3220                 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
3221
3222         tvd = vd->vdev_top;
3223         mg = tvd->vdev_mg;
3224         generation = spa->spa_config_generation + 1;
3225
3226         /*
3227          * If the device isn't already offline, try to offline it.
3228          */
3229         if (!vd->vdev_offline) {
3230                 /*
3231                  * If this device has the only valid copy of some data,
3232                  * don't allow it to be offlined. Log devices are always
3233                  * expendable.
3234                  */
3235                 if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
3236                     vdev_dtl_required(vd))
3237                         return (spa_vdev_state_exit(spa, NULL, EBUSY));
3238
3239                 /*
3240                  * If the top-level is a slog and it has had allocations
3241                  * then proceed.  We check that the vdev's metaslab group
3242                  * is not NULL since it's possible that we may have just
3243                  * added this vdev but not yet initialized its metaslabs.
3244                  */
3245                 if (tvd->vdev_islog && mg != NULL) {
3246                         /*
3247                          * Prevent any future allocations.
3248                          */
3249                         metaslab_group_passivate(mg);
3250                         (void) spa_vdev_state_exit(spa, vd, 0);
3251
3252                         error = spa_reset_logs(spa);
3253
3254                         /*
3255                          * If the log device was successfully reset but has
3256                          * checkpointed data, do not offline it.
3257                          */
3258                         if (error == 0 &&
3259                             tvd->vdev_checkpoint_sm != NULL) {
3260                                 ASSERT3U(tvd->vdev_checkpoint_sm->sm_alloc,
3261                                     !=, 0);
3262                                 error = ZFS_ERR_CHECKPOINT_EXISTS;
3263                         }
3264
3265                         spa_vdev_state_enter(spa, SCL_ALLOC);
3266
3267                         /*
3268                          * Check to see if the config has changed.
3269                          */
3270                         if (error || generation != spa->spa_config_generation) {
3271                                 metaslab_group_activate(mg);
3272                                 if (error)
3273                                         return (spa_vdev_state_exit(spa,
3274                                             vd, error));
3275                                 (void) spa_vdev_state_exit(spa, vd, 0);
3276                                 goto top;
3277                         }
3278                         ASSERT0(tvd->vdev_stat.vs_alloc);
3279                 }
3280
3281                 /*
3282                  * Offline this device and reopen its top-level vdev.
3283                  * If the top-level vdev is a log device then just offline
3284                  * it. Otherwise, if this action results in the top-level
3285                  * vdev becoming unusable, undo it and fail the request.
3286                  */
3287                 vd->vdev_offline = B_TRUE;
3288                 vdev_reopen(tvd);
3289
3290                 if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
3291                     vdev_is_dead(tvd)) {
3292                         vd->vdev_offline = B_FALSE;
3293                         vdev_reopen(tvd);
3294                         return (spa_vdev_state_exit(spa, NULL, EBUSY));
3295                 }
3296
3297                 /*
3298                  * Add the device back into the metaslab rotor so that
3299                  * once we online the device it's open for business.
3300                  */
3301                 if (tvd->vdev_islog && mg != NULL)
3302                         metaslab_group_activate(mg);
3303         }
3304
3305         vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
3306
3307         return (spa_vdev_state_exit(spa, vd, 0));
3308 }
3309
3310 int
3311 vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
3312 {
3313         int error;
3314
3315         mutex_enter(&spa->spa_vdev_top_lock);
3316         error = vdev_offline_locked(spa, guid, flags);
3317         mutex_exit(&spa->spa_vdev_top_lock);
3318
3319         return (error);
3320 }
3321
3322 /*
3323  * Clear the error counts associated with this vdev.  Unlike vdev_online() and
3324  * vdev_offline(), we assume the spa config is locked.  We also clear all
3325  * children.  If 'vd' is NULL, then the user wants to clear all vdevs.
3326  */
3327 void
3328 vdev_clear(spa_t *spa, vdev_t *vd)
3329 {
3330         vdev_t *rvd = spa->spa_root_vdev;
3331
3332         ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3333
3334         if (vd == NULL)
3335                 vd = rvd;
3336
3337         vd->vdev_stat.vs_read_errors = 0;
3338         vd->vdev_stat.vs_write_errors = 0;
3339         vd->vdev_stat.vs_checksum_errors = 0;
3340
3341         for (int c = 0; c < vd->vdev_children; c++)
3342                 vdev_clear(spa, vd->vdev_child[c]);
3343
3344         if (vd == rvd) {
3345                 for (int c = 0; c < spa->spa_l2cache.sav_count; c++)
3346                         vdev_clear(spa, spa->spa_l2cache.sav_vdevs[c]);
3347
3348                 for (int c = 0; c < spa->spa_spares.sav_count; c++)
3349                         vdev_clear(spa, spa->spa_spares.sav_vdevs[c]);
3350         }
3351
3352         /*
3353          * It makes no sense to "clear" an indirect vdev.
3354          */
3355         if (!vdev_is_concrete(vd))
3356                 return;
3357
3358         /*
3359          * If we're in the FAULTED state or have experienced failed I/O, then
3360          * clear the persistent state and attempt to reopen the device.  We
3361          * also mark the vdev config dirty, so that the new faulted state is
3362          * written out to disk.
3363          */
3364         if (vd->vdev_faulted || vd->vdev_degraded ||
3365             !vdev_readable(vd) || !vdev_writeable(vd)) {
3366
3367                 /*
3368                  * When reopening in reponse to a clear event, it may be due to
3369                  * a fmadm repair request.  In this case, if the device is
3370                  * still broken, we want to still post the ereport again.
3371                  */
3372                 vd->vdev_forcefault = B_TRUE;
3373
3374                 vd->vdev_faulted = vd->vdev_degraded = 0ULL;
3375                 vd->vdev_cant_read = B_FALSE;
3376                 vd->vdev_cant_write = B_FALSE;
3377
3378                 vdev_reopen(vd == rvd ? rvd : vd->vdev_top);
3379
3380                 vd->vdev_forcefault = B_FALSE;
3381
3382                 if (vd != rvd && vdev_writeable(vd->vdev_top))
3383                         vdev_state_dirty(vd->vdev_top);
3384
3385                 if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
3386                         spa_async_request(spa, SPA_ASYNC_RESILVER);
3387
3388                 spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_CLEAR);
3389         }
3390
3391         /*
3392          * When clearing a FMA-diagnosed fault, we always want to
3393          * unspare the device, as we assume that the original spare was
3394          * done in response to the FMA fault.
3395          */
3396         if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
3397             vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
3398             vd->vdev_parent->vdev_child[0] == vd)
3399                 vd->vdev_unspare = B_TRUE;
3400 }
3401
3402 boolean_t
3403 vdev_is_dead(vdev_t *vd)
3404 {
3405         /*
3406          * Holes and missing devices are always considered "dead".
3407          * This simplifies the code since we don't have to check for
3408          * these types of devices in the various code paths.
3409          * Instead we rely on the fact that we skip over dead devices
3410          * before issuing I/O to them.
3411          */
3412         return (vd->vdev_state < VDEV_STATE_DEGRADED ||
3413             vd->vdev_ops == &vdev_hole_ops ||
3414             vd->vdev_ops == &vdev_missing_ops);
3415 }
3416
3417 boolean_t
3418 vdev_readable(vdev_t *vd)
3419 {
3420         return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
3421 }
3422
3423 boolean_t
3424 vdev_writeable(vdev_t *vd)
3425 {
3426         return (!vdev_is_dead(vd) && !vd->vdev_cant_write &&
3427             vdev_is_concrete(vd));
3428 }
3429
3430 boolean_t
3431 vdev_allocatable(vdev_t *vd)
3432 {
3433         uint64_t state = vd->vdev_state;
3434
3435         /*
3436          * We currently allow allocations from vdevs which may be in the
3437          * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
3438          * fails to reopen then we'll catch it later when we're holding
3439          * the proper locks.  Note that we have to get the vdev state
3440          * in a local variable because although it changes atomically,
3441          * we're asking two separate questions about it.
3442          */
3443         return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
3444             !vd->vdev_cant_write && vdev_is_concrete(vd) &&
3445             vd->vdev_mg->mg_initialized);
3446 }
3447
3448 boolean_t
3449 vdev_accessible(vdev_t *vd, zio_t *zio)
3450 {
3451         ASSERT(zio->io_vd == vd);
3452
3453         if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
3454                 return (B_FALSE);
3455
3456         if (zio->io_type == ZIO_TYPE_READ)
3457                 return (!vd->vdev_cant_read);
3458
3459         if (zio->io_type == ZIO_TYPE_WRITE)
3460                 return (!vd->vdev_cant_write);
3461
3462         return (B_TRUE);
3463 }
3464
3465 boolean_t
3466 vdev_is_spacemap_addressable(vdev_t *vd)
3467 {
3468         /*
3469          * Assuming 47 bits of the space map entry dedicated for the entry's
3470          * offset (see description in space_map.h), we calculate the maximum
3471          * address that can be described by a space map entry for the given
3472          * device.
3473          */
3474         uint64_t shift = vd->vdev_ashift + 47;
3475
3476         if (shift >= 63) /* detect potential overflow */
3477                 return (B_TRUE);
3478
3479         return (vd->vdev_asize < (1ULL << shift));
3480 }
3481
3482 /*
3483  * Get statistics for the given vdev.
3484  */
3485 void
3486 vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
3487 {
3488         spa_t *spa = vd->vdev_spa;
3489         vdev_t *rvd = spa->spa_root_vdev;
3490         vdev_t *tvd = vd->vdev_top;
3491
3492         ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
3493
3494         mutex_enter(&vd->vdev_stat_lock);
3495         bcopy(&vd->vdev_stat, vs, sizeof (*vs));
3496         vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
3497         vs->vs_state = vd->vdev_state;
3498         vs->vs_rsize = vdev_get_min_asize(vd);
3499         if (vd->vdev_ops->vdev_op_leaf)
3500                 vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
3501         /*
3502          * Report expandable space on top-level, non-auxillary devices only.
3503          * The expandable space is reported in terms of metaslab sized units
3504          * since that determines how much space the pool can expand.
3505          */
3506         if (vd->vdev_aux == NULL && tvd != NULL && vd->vdev_max_asize != 0) {
3507                 vs->vs_esize = P2ALIGN(vd->vdev_max_asize - vd->vdev_asize -
3508                     spa->spa_bootsize, 1ULL << tvd->vdev_ms_shift);
3509         }
3510         vs->vs_configured_ashift = vd->vdev_top != NULL
3511             ? vd->vdev_top->vdev_ashift : vd->vdev_ashift;
3512         vs->vs_logical_ashift = vd->vdev_logical_ashift;
3513         vs->vs_physical_ashift = vd->vdev_physical_ashift;
3514         if (vd->vdev_aux == NULL && vd == vd->vdev_top &&
3515             vdev_is_concrete(vd)) {
3516                 vs->vs_fragmentation = vd->vdev_mg->mg_fragmentation;
3517         }
3518
3519         /*
3520          * If we're getting stats on the root vdev, aggregate the I/O counts
3521          * over all top-level vdevs (i.e. the direct children of the root).
3522          */
3523         if (vd == rvd) {
3524                 for (int c = 0; c < rvd->vdev_children; c++) {
3525                         vdev_t *cvd = rvd->vdev_child[c];
3526                         vdev_stat_t *cvs = &cvd->vdev_stat;
3527
3528                         for (int t = 0; t < ZIO_TYPES; t++) {
3529                                 vs->vs_ops[t] += cvs->vs_ops[t];
3530                                 vs->vs_bytes[t] += cvs->vs_bytes[t];
3531                         }
3532                         cvs->vs_scan_removing = cvd->vdev_removing;
3533                 }
3534         }
3535         mutex_exit(&vd->vdev_stat_lock);
3536 }
3537
3538 void
3539 vdev_clear_stats(vdev_t *vd)
3540 {
3541         mutex_enter(&vd->vdev_stat_lock);
3542         vd->vdev_stat.vs_space = 0;
3543         vd->vdev_stat.vs_dspace = 0;
3544         vd->vdev_stat.vs_alloc = 0;
3545         mutex_exit(&vd->vdev_stat_lock);
3546 }
3547
3548 void
3549 vdev_scan_stat_init(vdev_t *vd)
3550 {
3551         vdev_stat_t *vs = &vd->vdev_stat;
3552
3553         for (int c = 0; c < vd->vdev_children; c++)
3554                 vdev_scan_stat_init(vd->vdev_child[c]);
3555
3556         mutex_enter(&vd->vdev_stat_lock);
3557         vs->vs_scan_processed = 0;
3558         mutex_exit(&vd->vdev_stat_lock);
3559 }
3560
3561 void
3562 vdev_stat_update(zio_t *zio, uint64_t psize)
3563 {
3564         spa_t *spa = zio->io_spa;
3565         vdev_t *rvd = spa->spa_root_vdev;
3566         vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
3567         vdev_t *pvd;
3568         uint64_t txg = zio->io_txg;
3569         vdev_stat_t *vs = &vd->vdev_stat;
3570         zio_type_t type = zio->io_type;
3571         int flags = zio->io_flags;
3572
3573         /*
3574          * If this i/o is a gang leader, it didn't do any actual work.
3575          */
3576         if (zio->io_gang_tree)
3577                 return;
3578
3579         if (zio->io_error == 0) {
3580                 /*
3581                  * If this is a root i/o, don't count it -- we've already
3582                  * counted the top-level vdevs, and vdev_get_stats() will
3583                  * aggregate them when asked.  This reduces contention on
3584                  * the root vdev_stat_lock and implicitly handles blocks
3585                  * that compress away to holes, for which there is no i/o.
3586                  * (Holes never create vdev children, so all the counters
3587                  * remain zero, which is what we want.)
3588                  *
3589                  * Note: this only applies to successful i/o (io_error == 0)
3590                  * because unlike i/o counts, errors are not additive.
3591                  * When reading a ditto block, for example, failure of
3592                  * one top-level vdev does not imply a root-level error.
3593                  */
3594                 if (vd == rvd)
3595                         return;
3596
3597                 ASSERT(vd == zio->io_vd);
3598
3599                 if (flags & ZIO_FLAG_IO_BYPASS)
3600                         return;
3601
3602                 mutex_enter(&vd->vdev_stat_lock);
3603
3604                 if (flags & ZIO_FLAG_IO_REPAIR) {
3605                         if (flags & ZIO_FLAG_SCAN_THREAD) {
3606                                 dsl_scan_phys_t *scn_phys =
3607                                     &spa->spa_dsl_pool->dp_scan->scn_phys;
3608                                 uint64_t *processed = &scn_phys->scn_processed;
3609
3610                                 /* XXX cleanup? */
3611                                 if (vd->vdev_ops->vdev_op_leaf)
3612                                         atomic_add_64(processed, psize);
3613                                 vs->vs_scan_processed += psize;
3614                         }
3615
3616                         if (flags & ZIO_FLAG_SELF_HEAL)
3617                                 vs->vs_self_healed += psize;
3618                 }
3619
3620                 vs->vs_ops[type]++;
3621                 vs->vs_bytes[type] += psize;
3622
3623                 mutex_exit(&vd->vdev_stat_lock);
3624                 return;
3625         }
3626
3627         if (flags & ZIO_FLAG_SPECULATIVE)
3628                 return;
3629
3630         /*
3631          * If this is an I/O error that is going to be retried, then ignore the
3632          * error.  Otherwise, the user may interpret B_FAILFAST I/O errors as
3633          * hard errors, when in reality they can happen for any number of
3634          * innocuous reasons (bus resets, MPxIO link failure, etc).
3635          */
3636         if (zio->io_error == EIO &&
3637             !(zio->io_flags & ZIO_FLAG_IO_RETRY))
3638                 return;
3639
3640         /*
3641          * Intent logs writes won't propagate their error to the root
3642          * I/O so don't mark these types of failures as pool-level
3643          * errors.
3644          */
3645         if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
3646                 return;
3647
3648         mutex_enter(&vd->vdev_stat_lock);
3649         if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
3650                 if (zio->io_error == ECKSUM)
3651                         vs->vs_checksum_errors++;
3652                 else
3653                         vs->vs_read_errors++;
3654         }
3655         if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
3656                 vs->vs_write_errors++;
3657         mutex_exit(&vd->vdev_stat_lock);
3658
3659         if (spa->spa_load_state == SPA_LOAD_NONE &&
3660             type == ZIO_TYPE_WRITE && txg != 0 &&
3661             (!(flags & ZIO_FLAG_IO_REPAIR) ||
3662             (flags & ZIO_FLAG_SCAN_THREAD) ||
3663             spa->spa_claiming)) {
3664                 /*
3665                  * This is either a normal write (not a repair), or it's
3666                  * a repair induced by the scrub thread, or it's a repair
3667                  * made by zil_claim() during spa_load() in the first txg.
3668                  * In the normal case, we commit the DTL change in the same
3669                  * txg as the block was born.  In the scrub-induced repair
3670                  * case, we know that scrubs run in first-pass syncing context,
3671                  * so we commit the DTL change in spa_syncing_txg(spa).
3672                  * In the zil_claim() case, we commit in spa_first_txg(spa).
3673                  *
3674                  * We currently do not make DTL entries for failed spontaneous
3675                  * self-healing writes triggered by normal (non-scrubbing)
3676                  * reads, because we have no transactional context in which to
3677                  * do so -- and it's not clear that it'd be desirable anyway.
3678                  */
3679                 if (vd->vdev_ops->vdev_op_leaf) {
3680                         uint64_t commit_txg = txg;
3681                         if (flags & ZIO_FLAG_SCAN_THREAD) {
3682                                 ASSERT(flags & ZIO_FLAG_IO_REPAIR);
3683                                 ASSERT(spa_sync_pass(spa) == 1);
3684                                 vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
3685                                 commit_txg = spa_syncing_txg(spa);
3686                         } else if (spa->spa_claiming) {
3687                                 ASSERT(flags & ZIO_FLAG_IO_REPAIR);
3688                                 commit_txg = spa_first_txg(spa);
3689                         }
3690                         ASSERT(commit_txg >= spa_syncing_txg(spa));
3691                         if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
3692                                 return;
3693                         for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
3694                                 vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
3695                         vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
3696                 }
3697                 if (vd != rvd)
3698                         vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
3699         }
3700 }
3701
3702 /*
3703  * Update the in-core space usage stats for this vdev, its metaslab class,
3704  * and the root vdev.
3705  */
3706 void
3707 vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
3708     int64_t space_delta)
3709 {
3710         int64_t dspace_delta = space_delta;
3711         spa_t *spa = vd->vdev_spa;
3712         vdev_t *rvd = spa->spa_root_vdev;
3713         metaslab_group_t *mg = vd->vdev_mg;
3714         metaslab_class_t *mc = mg ? mg->mg_class : NULL;
3715
3716         ASSERT(vd == vd->vdev_top);
3717
3718         /*
3719          * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
3720          * factor.  We must calculate this here and not at the root vdev
3721          * because the root vdev's psize-to-asize is simply the max of its
3722          * childrens', thus not accurate enough for us.
3723          */
3724         ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
3725         ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
3726         dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
3727             vd->vdev_deflate_ratio;
3728
3729         mutex_enter(&vd->vdev_stat_lock);
3730         vd->vdev_stat.vs_alloc += alloc_delta;
3731         vd->vdev_stat.vs_space += space_delta;
3732         vd->vdev_stat.vs_dspace += dspace_delta;
3733         mutex_exit(&vd->vdev_stat_lock);
3734
3735         if (mc == spa_normal_class(spa)) {
3736                 mutex_enter(&rvd->vdev_stat_lock);
3737                 rvd->vdev_stat.vs_alloc += alloc_delta;
3738                 rvd->vdev_stat.vs_space += space_delta;
3739                 rvd->vdev_stat.vs_dspace += dspace_delta;
3740                 mutex_exit(&rvd->vdev_stat_lock);
3741         }
3742
3743         if (mc != NULL) {
3744                 ASSERT(rvd == vd->vdev_parent);
3745                 ASSERT(vd->vdev_ms_count != 0);
3746
3747                 metaslab_class_space_update(mc,
3748                     alloc_delta, defer_delta, space_delta, dspace_delta);
3749         }
3750 }
3751
3752 /*
3753  * Mark a top-level vdev's config as dirty, placing it on the dirty list
3754  * so that it will be written out next time the vdev configuration is synced.
3755  * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
3756  */
3757 void
3758 vdev_config_dirty(vdev_t *vd)
3759 {
3760         spa_t *spa = vd->vdev_spa;
3761         vdev_t *rvd = spa->spa_root_vdev;
3762         int c;
3763
3764         ASSERT(spa_writeable(spa));
3765
3766         /*
3767          * If this is an aux vdev (as with l2cache and spare devices), then we
3768          * update the vdev config manually and set the sync flag.
3769          */
3770         if (vd->vdev_aux != NULL) {
3771                 spa_aux_vdev_t *sav = vd->vdev_aux;
3772                 nvlist_t **aux;
3773                 uint_t naux;
3774
3775                 for (c = 0; c < sav->sav_count; c++) {
3776                         if (sav->sav_vdevs[c] == vd)
3777                                 break;
3778                 }
3779
3780                 if (c == sav->sav_count) {
3781                         /*
3782                          * We're being removed.  There's nothing more to do.
3783                          */
3784                         ASSERT(sav->sav_sync == B_TRUE);
3785                         return;
3786                 }
3787
3788                 sav->sav_sync = B_TRUE;
3789
3790                 if (nvlist_lookup_nvlist_array(sav->sav_config,
3791                     ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
3792                         VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
3793                             ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
3794                 }
3795
3796                 ASSERT(c < naux);
3797
3798                 /*
3799                  * Setting the nvlist in the middle if the array is a little
3800                  * sketchy, but it will work.
3801                  */
3802                 nvlist_free(aux[c]);
3803                 aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0);
3804
3805                 return;
3806         }
3807
3808         /*
3809          * The dirty list is protected by the SCL_CONFIG lock.  The caller
3810          * must either hold SCL_CONFIG as writer, or must be the sync thread
3811          * (which holds SCL_CONFIG as reader).  There's only one sync thread,
3812          * so this is sufficient to ensure mutual exclusion.
3813          */
3814         ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
3815             (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3816             spa_config_held(spa, SCL_CONFIG, RW_READER)));
3817
3818         if (vd == rvd) {
3819                 for (c = 0; c < rvd->vdev_children; c++)
3820                         vdev_config_dirty(rvd->vdev_child[c]);
3821         } else {
3822                 ASSERT(vd == vd->vdev_top);
3823
3824                 if (!list_link_active(&vd->vdev_config_dirty_node) &&
3825                     vdev_is_concrete(vd)) {
3826                         list_insert_head(&spa->spa_config_dirty_list, vd);
3827                 }
3828         }
3829 }
3830
3831 void
3832 vdev_config_clean(vdev_t *vd)
3833 {
3834         spa_t *spa = vd->vdev_spa;
3835
3836         ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
3837             (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3838             spa_config_held(spa, SCL_CONFIG, RW_READER)));
3839
3840         ASSERT(list_link_active(&vd->vdev_config_dirty_node));
3841         list_remove(&spa->spa_config_dirty_list, vd);
3842 }
3843
3844 /*
3845  * Mark a top-level vdev's state as dirty, so that the next pass of
3846  * spa_sync() can convert this into vdev_config_dirty().  We distinguish
3847  * the state changes from larger config changes because they require
3848  * much less locking, and are often needed for administrative actions.
3849  */
3850 void
3851 vdev_state_dirty(vdev_t *vd)
3852 {
3853         spa_t *spa = vd->vdev_spa;
3854
3855         ASSERT(spa_writeable(spa));
3856         ASSERT(vd == vd->vdev_top);
3857
3858         /*
3859          * The state list is protected by the SCL_STATE lock.  The caller
3860          * must either hold SCL_STATE as writer, or must be the sync thread
3861          * (which holds SCL_STATE as reader).  There's only one sync thread,
3862          * so this is sufficient to ensure mutual exclusion.
3863          */
3864         ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
3865             (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3866             spa_config_held(spa, SCL_STATE, RW_READER)));
3867
3868         if (!list_link_active(&vd->vdev_state_dirty_node) &&
3869             vdev_is_concrete(vd))
3870                 list_insert_head(&spa->spa_state_dirty_list, vd);
3871 }
3872
3873 void
3874 vdev_state_clean(vdev_t *vd)
3875 {
3876         spa_t *spa = vd->vdev_spa;
3877
3878         ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
3879             (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3880             spa_config_held(spa, SCL_STATE, RW_READER)));
3881
3882         ASSERT(list_link_active(&vd->vdev_state_dirty_node));
3883         list_remove(&spa->spa_state_dirty_list, vd);
3884 }
3885
3886 /*
3887  * Propagate vdev state up from children to parent.
3888  */
3889 void
3890 vdev_propagate_state(vdev_t *vd)
3891 {
3892         spa_t *spa = vd->vdev_spa;
3893         vdev_t *rvd = spa->spa_root_vdev;
3894         int degraded = 0, faulted = 0;
3895         int corrupted = 0;
3896         vdev_t *child;
3897
3898         if (vd->vdev_children > 0) {
3899                 for (int c = 0; c < vd->vdev_children; c++) {
3900                         child = vd->vdev_child[c];
3901
3902                         /*
3903                          * Don't factor holes or indirect vdevs into the
3904                          * decision.
3905                          */
3906                         if (!vdev_is_concrete(child))
3907                                 continue;
3908
3909                         if (!vdev_readable(child) ||
3910                             (!vdev_writeable(child) && spa_writeable(spa))) {
3911                                 /*
3912                                  * Root special: if there is a top-level log
3913                                  * device, treat the root vdev as if it were
3914                                  * degraded.
3915                                  */
3916                                 if (child->vdev_islog && vd == rvd)
3917                                         degraded++;
3918                                 else
3919                                         faulted++;
3920                         } else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
3921                                 degraded++;
3922                         }
3923
3924                         if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
3925                                 corrupted++;
3926                 }
3927
3928                 vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
3929
3930                 /*
3931                  * Root special: if there is a top-level vdev that cannot be
3932                  * opened due to corrupted metadata, then propagate the root
3933                  * vdev's aux state as 'corrupt' rather than 'insufficient
3934                  * replicas'.
3935                  */
3936                 if (corrupted && vd == rvd &&
3937                     rvd->vdev_state == VDEV_STATE_CANT_OPEN)
3938                         vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
3939                             VDEV_AUX_CORRUPT_DATA);
3940         }
3941
3942         if (vd->vdev_parent)
3943                 vdev_propagate_state(vd->vdev_parent);
3944 }
3945
3946 /*
3947  * Set a vdev's state.  If this is during an open, we don't update the parent
3948  * state, because we're in the process of opening children depth-first.
3949  * Otherwise, we propagate the change to the parent.
3950  *
3951  * If this routine places a device in a faulted state, an appropriate ereport is
3952  * generated.
3953  */
3954 void
3955 vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
3956 {
3957         uint64_t save_state;
3958         spa_t *spa = vd->vdev_spa;
3959
3960         if (state == vd->vdev_state) {
3961                 vd->vdev_stat.vs_aux = aux;
3962                 return;
3963         }
3964
3965         save_state = vd->vdev_state;
3966
3967         vd->vdev_state = state;
3968         vd->vdev_stat.vs_aux = aux;
3969
3970         /*
3971          * If we are setting the vdev state to anything but an open state, then
3972          * always close the underlying device unless the device has requested
3973          * a delayed close (i.e. we're about to remove or fault the device).
3974          * Otherwise, we keep accessible but invalid devices open forever.
3975          * We don't call vdev_close() itself, because that implies some extra
3976          * checks (offline, etc) that we don't want here.  This is limited to
3977          * leaf devices, because otherwise closing the device will affect other
3978          * children.
3979          */
3980         if (!vd->vdev_delayed_close && vdev_is_dead(vd) &&
3981             vd->vdev_ops->vdev_op_leaf)
3982                 vd->vdev_ops->vdev_op_close(vd);
3983
3984         if (vd->vdev_removed &&
3985             state == VDEV_STATE_CANT_OPEN &&
3986             (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
3987                 /*
3988                  * If the previous state is set to VDEV_STATE_REMOVED, then this
3989                  * device was previously marked removed and someone attempted to
3990                  * reopen it.  If this failed due to a nonexistent device, then
3991                  * keep the device in the REMOVED state.  We also let this be if
3992                  * it is one of our special test online cases, which is only
3993                  * attempting to online the device and shouldn't generate an FMA
3994                  * fault.
3995                  */
3996                 vd->vdev_state = VDEV_STATE_REMOVED;
3997                 vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
3998         } else if (state == VDEV_STATE_REMOVED) {
3999                 vd->vdev_removed = B_TRUE;
4000         } else if (state == VDEV_STATE_CANT_OPEN) {
4001                 /*
4002                  * If we fail to open a vdev during an import or recovery, we
4003                  * mark it as "not available", which signifies that it was
4004                  * never there to begin with.  Failure to open such a device
4005                  * is not considered an error.
4006                  */
4007                 if ((spa_load_state(spa) == SPA_LOAD_IMPORT ||
4008                     spa_load_state(spa) == SPA_LOAD_RECOVER) &&
4009                     vd->vdev_ops->vdev_op_leaf)
4010                         vd->vdev_not_present = 1;
4011
4012                 /*
4013                  * Post the appropriate ereport.  If the 'prevstate' field is
4014                  * set to something other than VDEV_STATE_UNKNOWN, it indicates
4015                  * that this is part of a vdev_reopen().  In this case, we don't
4016                  * want to post the ereport if the device was already in the
4017                  * CANT_OPEN state beforehand.
4018                  *
4019                  * If the 'checkremove' flag is set, then this is an attempt to
4020                  * online the device in response to an insertion event.  If we
4021                  * hit this case, then we have detected an insertion event for a
4022                  * faulted or offline device that wasn't in the removed state.
4023                  * In this scenario, we don't post an ereport because we are
4024                  * about to replace the device, or attempt an online with
4025                  * vdev_forcefault, which will generate the fault for us.
4026                  */
4027                 if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
4028                     !vd->vdev_not_present && !vd->vdev_checkremove &&
4029                     vd != spa->spa_root_vdev) {
4030                         const char *class;
4031
4032                         switch (aux) {
4033                         case VDEV_AUX_OPEN_FAILED:
4034                                 class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
4035                                 break;
4036                         case VDEV_AUX_CORRUPT_DATA:
4037                                 class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
4038                                 break;
4039                         case VDEV_AUX_NO_REPLICAS:
4040                                 class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
4041                                 break;
4042                         case VDEV_AUX_BAD_GUID_SUM:
4043                                 class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
4044                                 break;
4045                         case VDEV_AUX_TOO_SMALL:
4046                                 class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
4047                                 break;
4048                         case VDEV_AUX_BAD_LABEL:
4049                                 class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
4050                                 break;
4051                         default:
4052                                 class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
4053                         }
4054
4055                         zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
4056                 }
4057
4058                 /* Erase any notion of persistent removed state */
4059                 vd->vdev_removed = B_FALSE;
4060         } else {
4061                 vd->vdev_removed = B_FALSE;
4062         }
4063
4064         /*
4065         * Notify the fmd of the state change.  Be verbose and post
4066         * notifications even for stuff that's not important; the fmd agent can
4067         * sort it out.  Don't emit state change events for non-leaf vdevs since
4068         * they can't change state on their own.  The FMD can check their state
4069         * if it wants to when it sees that a leaf vdev had a state change.
4070         */
4071         if (vd->vdev_ops->vdev_op_leaf)
4072                 zfs_post_state_change(spa, vd);
4073
4074         if (!isopen && vd->vdev_parent)
4075                 vdev_propagate_state(vd->vdev_parent);
4076 }
4077
4078 boolean_t
4079 vdev_children_are_offline(vdev_t *vd)
4080 {
4081         ASSERT(!vd->vdev_ops->vdev_op_leaf);
4082
4083         for (uint64_t i = 0; i < vd->vdev_children; i++) {
4084                 if (vd->vdev_child[i]->vdev_state != VDEV_STATE_OFFLINE)
4085                         return (B_FALSE);
4086         }
4087
4088         return (B_TRUE);
4089 }
4090
4091 /*
4092  * Check the vdev configuration to ensure that it's capable of supporting
4093  * a root pool. We do not support partial configuration.
4094  * In addition, only a single top-level vdev is allowed.
4095  *
4096  * FreeBSD does not have above limitations.
4097  */
4098 boolean_t
4099 vdev_is_bootable(vdev_t *vd)
4100 {
4101 #ifdef illumos
4102         if (!vd->vdev_ops->vdev_op_leaf) {
4103                 char *vdev_type = vd->vdev_ops->vdev_op_type;
4104
4105                 if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
4106                     vd->vdev_children > 1) {
4107                         return (B_FALSE);
4108                 } else if (strcmp(vdev_type, VDEV_TYPE_MISSING) == 0 ||
4109                     strcmp(vdev_type, VDEV_TYPE_INDIRECT) == 0) {
4110                         return (B_FALSE);
4111                 }
4112         }
4113
4114         for (int c = 0; c < vd->vdev_children; c++) {
4115                 if (!vdev_is_bootable(vd->vdev_child[c]))
4116                         return (B_FALSE);
4117         }
4118 #endif  /* illumos */
4119         return (B_TRUE);
4120 }
4121
4122 boolean_t
4123 vdev_is_concrete(vdev_t *vd)
4124 {
4125         vdev_ops_t *ops = vd->vdev_ops;
4126         if (ops == &vdev_indirect_ops || ops == &vdev_hole_ops ||
4127             ops == &vdev_missing_ops || ops == &vdev_root_ops) {
4128                 return (B_FALSE);
4129         } else {
4130                 return (B_TRUE);
4131         }
4132 }
4133
4134 /*
4135  * Determine if a log device has valid content.  If the vdev was
4136  * removed or faulted in the MOS config then we know that
4137  * the content on the log device has already been written to the pool.
4138  */
4139 boolean_t
4140 vdev_log_state_valid(vdev_t *vd)
4141 {
4142         if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted &&
4143             !vd->vdev_removed)
4144                 return (B_TRUE);
4145
4146         for (int c = 0; c < vd->vdev_children; c++)
4147                 if (vdev_log_state_valid(vd->vdev_child[c]))
4148                         return (B_TRUE);
4149
4150         return (B_FALSE);
4151 }
4152
4153 /*
4154  * Expand a vdev if possible.
4155  */
4156 void
4157 vdev_expand(vdev_t *vd, uint64_t txg)
4158 {
4159         ASSERT(vd->vdev_top == vd);
4160         ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
4161
4162         vdev_set_deflate_ratio(vd);
4163
4164         if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count &&
4165             vdev_is_concrete(vd)) {
4166                 VERIFY(vdev_metaslab_init(vd, txg) == 0);
4167                 vdev_config_dirty(vd);
4168         }
4169 }
4170
4171 /*
4172  * Split a vdev.
4173  */
4174 void
4175 vdev_split(vdev_t *vd)
4176 {
4177         vdev_t *cvd, *pvd = vd->vdev_parent;
4178
4179         vdev_remove_child(pvd, vd);
4180         vdev_compact_children(pvd);
4181
4182         cvd = pvd->vdev_child[0];
4183         if (pvd->vdev_children == 1) {
4184                 vdev_remove_parent(cvd);
4185                 cvd->vdev_splitting = B_TRUE;
4186         }
4187         vdev_propagate_state(cvd);
4188 }
4189
4190 void
4191 vdev_deadman(vdev_t *vd)
4192 {
4193         for (int c = 0; c < vd->vdev_children; c++) {
4194                 vdev_t *cvd = vd->vdev_child[c];
4195
4196                 vdev_deadman(cvd);
4197         }
4198
4199         if (vd->vdev_ops->vdev_op_leaf) {
4200                 vdev_queue_t *vq = &vd->vdev_queue;
4201
4202                 mutex_enter(&vq->vq_lock);
4203                 if (avl_numnodes(&vq->vq_active_tree) > 0) {
4204                         spa_t *spa = vd->vdev_spa;
4205                         zio_t *fio;
4206                         uint64_t delta;
4207
4208                         /*
4209                          * Look at the head of all the pending queues,
4210                          * if any I/O has been outstanding for longer than
4211                          * the spa_deadman_synctime we panic the system.
4212                          */
4213                         fio = avl_first(&vq->vq_active_tree);
4214                         delta = gethrtime() - fio->io_timestamp;
4215                         if (delta > spa_deadman_synctime(spa)) {
4216                                 vdev_dbgmsg(vd, "SLOW IO: zio timestamp "
4217                                     "%lluns, delta %lluns, last io %lluns",
4218                                     fio->io_timestamp, (u_longlong_t)delta,
4219                                     vq->vq_io_complete_ts);
4220                                 fm_panic("I/O to pool '%s' appears to be "
4221                                     "hung on vdev guid %llu at '%s'.",
4222                                     spa_name(spa),
4223                                     (long long unsigned int) vd->vdev_guid,
4224                                     vd->vdev_path);
4225                         }
4226                 }
4227                 mutex_exit(&vq->vq_lock);
4228         }
4229 }