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