Print this page
5269 zfs: zpool import slow
While importing a pool all objsets are enumerated twice, once to check
the zil log chains and once to claim them. On pools with many datasets
this process might take a substantial amount of time.
Speed up the process by parallelizing it utilizing a taskq. The number
of parallel tasks is limited to 4 times the number of leaf vdevs.
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/fs/zfs/vdev.c
+++ new/usr/src/uts/common/fs/zfs/vdev.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21
22 22 /*
23 23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 24 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
25 25 * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
26 26 */
27 27
28 28 #include <sys/zfs_context.h>
29 29 #include <sys/fm/fs/zfs.h>
30 30 #include <sys/spa.h>
31 31 #include <sys/spa_impl.h>
32 32 #include <sys/dmu.h>
33 33 #include <sys/dmu_tx.h>
34 34 #include <sys/vdev_impl.h>
35 35 #include <sys/uberblock_impl.h>
36 36 #include <sys/metaslab.h>
37 37 #include <sys/metaslab_impl.h>
38 38 #include <sys/space_map.h>
39 39 #include <sys/space_reftree.h>
40 40 #include <sys/zio.h>
41 41 #include <sys/zap.h>
42 42 #include <sys/fs/zfs.h>
43 43 #include <sys/arc.h>
44 44 #include <sys/zil.h>
45 45 #include <sys/dsl_scan.h>
46 46
47 47 /*
48 48 * Virtual device management.
49 49 */
50 50
51 51 static vdev_ops_t *vdev_ops_table[] = {
52 52 &vdev_root_ops,
53 53 &vdev_raidz_ops,
54 54 &vdev_mirror_ops,
55 55 &vdev_replacing_ops,
56 56 &vdev_spare_ops,
57 57 &vdev_disk_ops,
58 58 &vdev_file_ops,
59 59 &vdev_missing_ops,
60 60 &vdev_hole_ops,
61 61 NULL
62 62 };
63 63
64 64 /* maximum scrub/resilver I/O queue per leaf vdev */
65 65 int zfs_scrub_limit = 10;
66 66
67 67 /*
68 68 * Given a vdev type, return the appropriate ops vector.
69 69 */
70 70 static vdev_ops_t *
71 71 vdev_getops(const char *type)
72 72 {
73 73 vdev_ops_t *ops, **opspp;
74 74
75 75 for (opspp = vdev_ops_table; (ops = *opspp) != NULL; opspp++)
76 76 if (strcmp(ops->vdev_op_type, type) == 0)
77 77 break;
78 78
79 79 return (ops);
80 80 }
81 81
82 82 /*
83 83 * Default asize function: return the MAX of psize with the asize of
84 84 * all children. This is what's used by anything other than RAID-Z.
85 85 */
86 86 uint64_t
87 87 vdev_default_asize(vdev_t *vd, uint64_t psize)
88 88 {
89 89 uint64_t asize = P2ROUNDUP(psize, 1ULL << vd->vdev_top->vdev_ashift);
90 90 uint64_t csize;
91 91
92 92 for (int c = 0; c < vd->vdev_children; c++) {
93 93 csize = vdev_psize_to_asize(vd->vdev_child[c], psize);
94 94 asize = MAX(asize, csize);
95 95 }
96 96
97 97 return (asize);
98 98 }
99 99
100 100 /*
101 101 * Get the minimum allocatable size. We define the allocatable size as
102 102 * the vdev's asize rounded to the nearest metaslab. This allows us to
103 103 * replace or attach devices which don't have the same physical size but
104 104 * can still satisfy the same number of allocations.
105 105 */
106 106 uint64_t
107 107 vdev_get_min_asize(vdev_t *vd)
108 108 {
109 109 vdev_t *pvd = vd->vdev_parent;
110 110
111 111 /*
112 112 * If our parent is NULL (inactive spare or cache) or is the root,
113 113 * just return our own asize.
114 114 */
115 115 if (pvd == NULL)
116 116 return (vd->vdev_asize);
117 117
118 118 /*
119 119 * The top-level vdev just returns the allocatable size rounded
120 120 * to the nearest metaslab.
121 121 */
122 122 if (vd == vd->vdev_top)
123 123 return (P2ALIGN(vd->vdev_asize, 1ULL << vd->vdev_ms_shift));
124 124
125 125 /*
126 126 * The allocatable space for a raidz vdev is N * sizeof(smallest child),
127 127 * so each child must provide at least 1/Nth of its asize.
128 128 */
129 129 if (pvd->vdev_ops == &vdev_raidz_ops)
130 130 return (pvd->vdev_min_asize / pvd->vdev_children);
131 131
132 132 return (pvd->vdev_min_asize);
133 133 }
134 134
135 135 void
136 136 vdev_set_min_asize(vdev_t *vd)
137 137 {
138 138 vd->vdev_min_asize = vdev_get_min_asize(vd);
139 139
140 140 for (int c = 0; c < vd->vdev_children; c++)
141 141 vdev_set_min_asize(vd->vdev_child[c]);
142 142 }
143 143
144 144 vdev_t *
145 145 vdev_lookup_top(spa_t *spa, uint64_t vdev)
146 146 {
147 147 vdev_t *rvd = spa->spa_root_vdev;
148 148
149 149 ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
150 150
151 151 if (vdev < rvd->vdev_children) {
152 152 ASSERT(rvd->vdev_child[vdev] != NULL);
153 153 return (rvd->vdev_child[vdev]);
154 154 }
155 155
156 156 return (NULL);
157 157 }
158 158
159 159 vdev_t *
160 160 vdev_lookup_by_guid(vdev_t *vd, uint64_t guid)
161 161 {
162 162 vdev_t *mvd;
163 163
164 164 if (vd->vdev_guid == guid)
|
↓ open down ↓ |
164 lines elided |
↑ open up ↑ |
165 165 return (vd);
166 166
167 167 for (int c = 0; c < vd->vdev_children; c++)
168 168 if ((mvd = vdev_lookup_by_guid(vd->vdev_child[c], guid)) !=
169 169 NULL)
170 170 return (mvd);
171 171
172 172 return (NULL);
173 173 }
174 174
175 +static int
176 +vdev_count_leaves_impl(vdev_t *vd)
177 +{
178 + vdev_t *mvd;
179 + int n = 0;
180 +
181 + if (vd->vdev_children == 0)
182 + return (1);
183 +
184 + for (int c = 0; c < vd->vdev_children; c++)
185 + n += vdev_count_leaves_impl(vd->vdev_child[c]);
186 +
187 + return (n);
188 +}
189 +
190 +int
191 +vdev_count_leaves(spa_t *spa)
192 +{
193 + return (vdev_count_leaves_impl(spa->spa_root_vdev));
194 +}
195 +
175 196 void
176 197 vdev_add_child(vdev_t *pvd, vdev_t *cvd)
177 198 {
178 199 size_t oldsize, newsize;
179 200 uint64_t id = cvd->vdev_id;
180 201 vdev_t **newchild;
181 202
182 203 ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
183 204 ASSERT(cvd->vdev_parent == NULL);
184 205
185 206 cvd->vdev_parent = pvd;
186 207
187 208 if (pvd == NULL)
188 209 return;
189 210
190 211 ASSERT(id >= pvd->vdev_children || pvd->vdev_child[id] == NULL);
191 212
192 213 oldsize = pvd->vdev_children * sizeof (vdev_t *);
193 214 pvd->vdev_children = MAX(pvd->vdev_children, id + 1);
194 215 newsize = pvd->vdev_children * sizeof (vdev_t *);
195 216
196 217 newchild = kmem_zalloc(newsize, KM_SLEEP);
197 218 if (pvd->vdev_child != NULL) {
198 219 bcopy(pvd->vdev_child, newchild, oldsize);
199 220 kmem_free(pvd->vdev_child, oldsize);
200 221 }
201 222
202 223 pvd->vdev_child = newchild;
203 224 pvd->vdev_child[id] = cvd;
204 225
205 226 cvd->vdev_top = (pvd->vdev_top ? pvd->vdev_top: cvd);
206 227 ASSERT(cvd->vdev_top->vdev_parent->vdev_parent == NULL);
207 228
208 229 /*
209 230 * Walk up all ancestors to update guid sum.
210 231 */
211 232 for (; pvd != NULL; pvd = pvd->vdev_parent)
212 233 pvd->vdev_guid_sum += cvd->vdev_guid_sum;
213 234 }
214 235
215 236 void
216 237 vdev_remove_child(vdev_t *pvd, vdev_t *cvd)
217 238 {
218 239 int c;
219 240 uint_t id = cvd->vdev_id;
220 241
221 242 ASSERT(cvd->vdev_parent == pvd);
222 243
223 244 if (pvd == NULL)
224 245 return;
225 246
226 247 ASSERT(id < pvd->vdev_children);
227 248 ASSERT(pvd->vdev_child[id] == cvd);
228 249
229 250 pvd->vdev_child[id] = NULL;
230 251 cvd->vdev_parent = NULL;
231 252
232 253 for (c = 0; c < pvd->vdev_children; c++)
233 254 if (pvd->vdev_child[c])
234 255 break;
235 256
236 257 if (c == pvd->vdev_children) {
237 258 kmem_free(pvd->vdev_child, c * sizeof (vdev_t *));
238 259 pvd->vdev_child = NULL;
239 260 pvd->vdev_children = 0;
240 261 }
241 262
242 263 /*
243 264 * Walk up all ancestors to update guid sum.
244 265 */
245 266 for (; pvd != NULL; pvd = pvd->vdev_parent)
246 267 pvd->vdev_guid_sum -= cvd->vdev_guid_sum;
247 268 }
248 269
249 270 /*
250 271 * Remove any holes in the child array.
251 272 */
252 273 void
253 274 vdev_compact_children(vdev_t *pvd)
254 275 {
255 276 vdev_t **newchild, *cvd;
256 277 int oldc = pvd->vdev_children;
257 278 int newc;
258 279
259 280 ASSERT(spa_config_held(pvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
260 281
261 282 for (int c = newc = 0; c < oldc; c++)
262 283 if (pvd->vdev_child[c])
263 284 newc++;
264 285
265 286 newchild = kmem_alloc(newc * sizeof (vdev_t *), KM_SLEEP);
266 287
267 288 for (int c = newc = 0; c < oldc; c++) {
268 289 if ((cvd = pvd->vdev_child[c]) != NULL) {
269 290 newchild[newc] = cvd;
270 291 cvd->vdev_id = newc++;
271 292 }
272 293 }
273 294
274 295 kmem_free(pvd->vdev_child, oldc * sizeof (vdev_t *));
275 296 pvd->vdev_child = newchild;
276 297 pvd->vdev_children = newc;
277 298 }
278 299
279 300 /*
280 301 * Allocate and minimally initialize a vdev_t.
281 302 */
282 303 vdev_t *
283 304 vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
284 305 {
285 306 vdev_t *vd;
286 307
287 308 vd = kmem_zalloc(sizeof (vdev_t), KM_SLEEP);
288 309
289 310 if (spa->spa_root_vdev == NULL) {
290 311 ASSERT(ops == &vdev_root_ops);
291 312 spa->spa_root_vdev = vd;
292 313 spa->spa_load_guid = spa_generate_guid(NULL);
293 314 }
294 315
295 316 if (guid == 0 && ops != &vdev_hole_ops) {
296 317 if (spa->spa_root_vdev == vd) {
297 318 /*
298 319 * The root vdev's guid will also be the pool guid,
299 320 * which must be unique among all pools.
300 321 */
301 322 guid = spa_generate_guid(NULL);
302 323 } else {
303 324 /*
304 325 * Any other vdev's guid must be unique within the pool.
305 326 */
306 327 guid = spa_generate_guid(spa);
307 328 }
308 329 ASSERT(!spa_guid_exists(spa_guid(spa), guid));
309 330 }
310 331
311 332 vd->vdev_spa = spa;
312 333 vd->vdev_id = id;
313 334 vd->vdev_guid = guid;
314 335 vd->vdev_guid_sum = guid;
315 336 vd->vdev_ops = ops;
316 337 vd->vdev_state = VDEV_STATE_CLOSED;
317 338 vd->vdev_ishole = (ops == &vdev_hole_ops);
318 339
319 340 mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_DEFAULT, NULL);
320 341 mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL);
321 342 mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL);
322 343 for (int t = 0; t < DTL_TYPES; t++) {
323 344 vd->vdev_dtl[t] = range_tree_create(NULL, NULL,
324 345 &vd->vdev_dtl_lock);
325 346 }
326 347 txg_list_create(&vd->vdev_ms_list,
327 348 offsetof(struct metaslab, ms_txg_node));
328 349 txg_list_create(&vd->vdev_dtl_list,
329 350 offsetof(struct vdev, vdev_dtl_node));
330 351 vd->vdev_stat.vs_timestamp = gethrtime();
331 352 vdev_queue_init(vd);
332 353 vdev_cache_init(vd);
333 354
334 355 return (vd);
335 356 }
336 357
337 358 /*
338 359 * Allocate a new vdev. The 'alloctype' is used to control whether we are
339 360 * creating a new vdev or loading an existing one - the behavior is slightly
340 361 * different for each case.
341 362 */
342 363 int
343 364 vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id,
344 365 int alloctype)
345 366 {
346 367 vdev_ops_t *ops;
347 368 char *type;
348 369 uint64_t guid = 0, islog, nparity;
349 370 vdev_t *vd;
350 371
351 372 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
352 373
353 374 if (nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) != 0)
354 375 return (SET_ERROR(EINVAL));
355 376
356 377 if ((ops = vdev_getops(type)) == NULL)
357 378 return (SET_ERROR(EINVAL));
358 379
359 380 /*
360 381 * If this is a load, get the vdev guid from the nvlist.
361 382 * Otherwise, vdev_alloc_common() will generate one for us.
362 383 */
363 384 if (alloctype == VDEV_ALLOC_LOAD) {
364 385 uint64_t label_id;
365 386
366 387 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ID, &label_id) ||
367 388 label_id != id)
368 389 return (SET_ERROR(EINVAL));
369 390
370 391 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
371 392 return (SET_ERROR(EINVAL));
372 393 } else if (alloctype == VDEV_ALLOC_SPARE) {
373 394 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
374 395 return (SET_ERROR(EINVAL));
375 396 } else if (alloctype == VDEV_ALLOC_L2CACHE) {
376 397 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
377 398 return (SET_ERROR(EINVAL));
378 399 } else if (alloctype == VDEV_ALLOC_ROOTPOOL) {
379 400 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
380 401 return (SET_ERROR(EINVAL));
381 402 }
382 403
383 404 /*
384 405 * The first allocated vdev must be of type 'root'.
385 406 */
386 407 if (ops != &vdev_root_ops && spa->spa_root_vdev == NULL)
387 408 return (SET_ERROR(EINVAL));
388 409
389 410 /*
390 411 * Determine whether we're a log vdev.
391 412 */
392 413 islog = 0;
393 414 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &islog);
394 415 if (islog && spa_version(spa) < SPA_VERSION_SLOGS)
395 416 return (SET_ERROR(ENOTSUP));
396 417
397 418 if (ops == &vdev_hole_ops && spa_version(spa) < SPA_VERSION_HOLES)
398 419 return (SET_ERROR(ENOTSUP));
399 420
400 421 /*
401 422 * Set the nparity property for RAID-Z vdevs.
402 423 */
403 424 nparity = -1ULL;
404 425 if (ops == &vdev_raidz_ops) {
405 426 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY,
406 427 &nparity) == 0) {
407 428 if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY)
408 429 return (SET_ERROR(EINVAL));
409 430 /*
410 431 * Previous versions could only support 1 or 2 parity
411 432 * device.
412 433 */
413 434 if (nparity > 1 &&
414 435 spa_version(spa) < SPA_VERSION_RAIDZ2)
415 436 return (SET_ERROR(ENOTSUP));
416 437 if (nparity > 2 &&
417 438 spa_version(spa) < SPA_VERSION_RAIDZ3)
418 439 return (SET_ERROR(ENOTSUP));
419 440 } else {
420 441 /*
421 442 * We require the parity to be specified for SPAs that
422 443 * support multiple parity levels.
423 444 */
424 445 if (spa_version(spa) >= SPA_VERSION_RAIDZ2)
425 446 return (SET_ERROR(EINVAL));
426 447 /*
427 448 * Otherwise, we default to 1 parity device for RAID-Z.
428 449 */
429 450 nparity = 1;
430 451 }
431 452 } else {
432 453 nparity = 0;
433 454 }
434 455 ASSERT(nparity != -1ULL);
435 456
436 457 vd = vdev_alloc_common(spa, id, guid, ops);
437 458
438 459 vd->vdev_islog = islog;
439 460 vd->vdev_nparity = nparity;
440 461
441 462 if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &vd->vdev_path) == 0)
442 463 vd->vdev_path = spa_strdup(vd->vdev_path);
443 464 if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &vd->vdev_devid) == 0)
444 465 vd->vdev_devid = spa_strdup(vd->vdev_devid);
445 466 if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH,
446 467 &vd->vdev_physpath) == 0)
447 468 vd->vdev_physpath = spa_strdup(vd->vdev_physpath);
448 469 if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &vd->vdev_fru) == 0)
449 470 vd->vdev_fru = spa_strdup(vd->vdev_fru);
450 471
451 472 /*
452 473 * Set the whole_disk property. If it's not specified, leave the value
453 474 * as -1.
454 475 */
455 476 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
456 477 &vd->vdev_wholedisk) != 0)
457 478 vd->vdev_wholedisk = -1ULL;
458 479
459 480 /*
460 481 * Look for the 'not present' flag. This will only be set if the device
461 482 * was not present at the time of import.
462 483 */
463 484 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT,
464 485 &vd->vdev_not_present);
465 486
466 487 /*
467 488 * Get the alignment requirement.
468 489 */
469 490 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT, &vd->vdev_ashift);
470 491
471 492 /*
472 493 * Retrieve the vdev creation time.
473 494 */
474 495 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_CREATE_TXG,
475 496 &vd->vdev_crtxg);
476 497
477 498 /*
478 499 * If we're a top-level vdev, try to load the allocation parameters.
479 500 */
480 501 if (parent && !parent->vdev_parent &&
481 502 (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
482 503 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
483 504 &vd->vdev_ms_array);
484 505 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
485 506 &vd->vdev_ms_shift);
486 507 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE,
487 508 &vd->vdev_asize);
488 509 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVING,
489 510 &vd->vdev_removing);
490 511 }
491 512
492 513 if (parent && !parent->vdev_parent && alloctype != VDEV_ALLOC_ATTACH) {
493 514 ASSERT(alloctype == VDEV_ALLOC_LOAD ||
494 515 alloctype == VDEV_ALLOC_ADD ||
495 516 alloctype == VDEV_ALLOC_SPLIT ||
496 517 alloctype == VDEV_ALLOC_ROOTPOOL);
497 518 vd->vdev_mg = metaslab_group_create(islog ?
498 519 spa_log_class(spa) : spa_normal_class(spa), vd);
499 520 }
500 521
501 522 /*
502 523 * If we're a leaf vdev, try to load the DTL object and other state.
503 524 */
504 525 if (vd->vdev_ops->vdev_op_leaf &&
505 526 (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE ||
506 527 alloctype == VDEV_ALLOC_ROOTPOOL)) {
507 528 if (alloctype == VDEV_ALLOC_LOAD) {
508 529 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL,
509 530 &vd->vdev_dtl_object);
510 531 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE,
511 532 &vd->vdev_unspare);
512 533 }
513 534
514 535 if (alloctype == VDEV_ALLOC_ROOTPOOL) {
515 536 uint64_t spare = 0;
516 537
517 538 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE,
518 539 &spare) == 0 && spare)
519 540 spa_spare_add(vd);
520 541 }
521 542
522 543 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE,
523 544 &vd->vdev_offline);
524 545
525 546 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_RESILVER_TXG,
526 547 &vd->vdev_resilver_txg);
527 548
528 549 /*
529 550 * When importing a pool, we want to ignore the persistent fault
530 551 * state, as the diagnosis made on another system may not be
531 552 * valid in the current context. Local vdevs will
532 553 * remain in the faulted state.
533 554 */
534 555 if (spa_load_state(spa) == SPA_LOAD_OPEN) {
535 556 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED,
536 557 &vd->vdev_faulted);
537 558 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED,
538 559 &vd->vdev_degraded);
539 560 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED,
540 561 &vd->vdev_removed);
541 562
542 563 if (vd->vdev_faulted || vd->vdev_degraded) {
543 564 char *aux;
544 565
545 566 vd->vdev_label_aux =
546 567 VDEV_AUX_ERR_EXCEEDED;
547 568 if (nvlist_lookup_string(nv,
548 569 ZPOOL_CONFIG_AUX_STATE, &aux) == 0 &&
549 570 strcmp(aux, "external") == 0)
550 571 vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
551 572 }
552 573 }
553 574 }
554 575
555 576 /*
556 577 * Add ourselves to the parent's list of children.
557 578 */
558 579 vdev_add_child(parent, vd);
559 580
560 581 *vdp = vd;
561 582
562 583 return (0);
563 584 }
564 585
565 586 void
566 587 vdev_free(vdev_t *vd)
567 588 {
568 589 spa_t *spa = vd->vdev_spa;
569 590
570 591 /*
571 592 * vdev_free() implies closing the vdev first. This is simpler than
572 593 * trying to ensure complicated semantics for all callers.
573 594 */
574 595 vdev_close(vd);
575 596
576 597 ASSERT(!list_link_active(&vd->vdev_config_dirty_node));
577 598 ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
578 599
579 600 /*
580 601 * Free all children.
581 602 */
582 603 for (int c = 0; c < vd->vdev_children; c++)
583 604 vdev_free(vd->vdev_child[c]);
584 605
585 606 ASSERT(vd->vdev_child == NULL);
586 607 ASSERT(vd->vdev_guid_sum == vd->vdev_guid);
587 608
588 609 /*
589 610 * Discard allocation state.
590 611 */
591 612 if (vd->vdev_mg != NULL) {
592 613 vdev_metaslab_fini(vd);
593 614 metaslab_group_destroy(vd->vdev_mg);
594 615 }
595 616
596 617 ASSERT0(vd->vdev_stat.vs_space);
597 618 ASSERT0(vd->vdev_stat.vs_dspace);
598 619 ASSERT0(vd->vdev_stat.vs_alloc);
599 620
600 621 /*
601 622 * Remove this vdev from its parent's child list.
602 623 */
603 624 vdev_remove_child(vd->vdev_parent, vd);
604 625
605 626 ASSERT(vd->vdev_parent == NULL);
606 627
607 628 /*
608 629 * Clean up vdev structure.
609 630 */
610 631 vdev_queue_fini(vd);
611 632 vdev_cache_fini(vd);
612 633
613 634 if (vd->vdev_path)
614 635 spa_strfree(vd->vdev_path);
615 636 if (vd->vdev_devid)
616 637 spa_strfree(vd->vdev_devid);
617 638 if (vd->vdev_physpath)
618 639 spa_strfree(vd->vdev_physpath);
619 640 if (vd->vdev_fru)
620 641 spa_strfree(vd->vdev_fru);
621 642
622 643 if (vd->vdev_isspare)
623 644 spa_spare_remove(vd);
624 645 if (vd->vdev_isl2cache)
625 646 spa_l2cache_remove(vd);
626 647
627 648 txg_list_destroy(&vd->vdev_ms_list);
628 649 txg_list_destroy(&vd->vdev_dtl_list);
629 650
630 651 mutex_enter(&vd->vdev_dtl_lock);
631 652 space_map_close(vd->vdev_dtl_sm);
632 653 for (int t = 0; t < DTL_TYPES; t++) {
633 654 range_tree_vacate(vd->vdev_dtl[t], NULL, NULL);
634 655 range_tree_destroy(vd->vdev_dtl[t]);
635 656 }
636 657 mutex_exit(&vd->vdev_dtl_lock);
637 658
638 659 mutex_destroy(&vd->vdev_dtl_lock);
639 660 mutex_destroy(&vd->vdev_stat_lock);
640 661 mutex_destroy(&vd->vdev_probe_lock);
641 662
642 663 if (vd == spa->spa_root_vdev)
643 664 spa->spa_root_vdev = NULL;
644 665
645 666 kmem_free(vd, sizeof (vdev_t));
646 667 }
647 668
648 669 /*
649 670 * Transfer top-level vdev state from svd to tvd.
650 671 */
651 672 static void
652 673 vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
653 674 {
654 675 spa_t *spa = svd->vdev_spa;
655 676 metaslab_t *msp;
656 677 vdev_t *vd;
657 678 int t;
658 679
659 680 ASSERT(tvd == tvd->vdev_top);
660 681
661 682 tvd->vdev_ms_array = svd->vdev_ms_array;
662 683 tvd->vdev_ms_shift = svd->vdev_ms_shift;
663 684 tvd->vdev_ms_count = svd->vdev_ms_count;
664 685
665 686 svd->vdev_ms_array = 0;
666 687 svd->vdev_ms_shift = 0;
667 688 svd->vdev_ms_count = 0;
668 689
669 690 if (tvd->vdev_mg)
670 691 ASSERT3P(tvd->vdev_mg, ==, svd->vdev_mg);
671 692 tvd->vdev_mg = svd->vdev_mg;
672 693 tvd->vdev_ms = svd->vdev_ms;
673 694
674 695 svd->vdev_mg = NULL;
675 696 svd->vdev_ms = NULL;
676 697
677 698 if (tvd->vdev_mg != NULL)
678 699 tvd->vdev_mg->mg_vd = tvd;
679 700
680 701 tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
681 702 tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
682 703 tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
683 704
684 705 svd->vdev_stat.vs_alloc = 0;
685 706 svd->vdev_stat.vs_space = 0;
686 707 svd->vdev_stat.vs_dspace = 0;
687 708
688 709 for (t = 0; t < TXG_SIZE; t++) {
689 710 while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
690 711 (void) txg_list_add(&tvd->vdev_ms_list, msp, t);
691 712 while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
692 713 (void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
693 714 if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
694 715 (void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
695 716 }
696 717
697 718 if (list_link_active(&svd->vdev_config_dirty_node)) {
698 719 vdev_config_clean(svd);
699 720 vdev_config_dirty(tvd);
700 721 }
701 722
702 723 if (list_link_active(&svd->vdev_state_dirty_node)) {
703 724 vdev_state_clean(svd);
704 725 vdev_state_dirty(tvd);
705 726 }
706 727
707 728 tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
708 729 svd->vdev_deflate_ratio = 0;
709 730
710 731 tvd->vdev_islog = svd->vdev_islog;
711 732 svd->vdev_islog = 0;
712 733 }
713 734
714 735 static void
715 736 vdev_top_update(vdev_t *tvd, vdev_t *vd)
716 737 {
717 738 if (vd == NULL)
718 739 return;
719 740
720 741 vd->vdev_top = tvd;
721 742
722 743 for (int c = 0; c < vd->vdev_children; c++)
723 744 vdev_top_update(tvd, vd->vdev_child[c]);
724 745 }
725 746
726 747 /*
727 748 * Add a mirror/replacing vdev above an existing vdev.
728 749 */
729 750 vdev_t *
730 751 vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
731 752 {
732 753 spa_t *spa = cvd->vdev_spa;
733 754 vdev_t *pvd = cvd->vdev_parent;
734 755 vdev_t *mvd;
735 756
736 757 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
737 758
738 759 mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
739 760
740 761 mvd->vdev_asize = cvd->vdev_asize;
741 762 mvd->vdev_min_asize = cvd->vdev_min_asize;
742 763 mvd->vdev_max_asize = cvd->vdev_max_asize;
743 764 mvd->vdev_ashift = cvd->vdev_ashift;
744 765 mvd->vdev_state = cvd->vdev_state;
745 766 mvd->vdev_crtxg = cvd->vdev_crtxg;
746 767
747 768 vdev_remove_child(pvd, cvd);
748 769 vdev_add_child(pvd, mvd);
749 770 cvd->vdev_id = mvd->vdev_children;
750 771 vdev_add_child(mvd, cvd);
751 772 vdev_top_update(cvd->vdev_top, cvd->vdev_top);
752 773
753 774 if (mvd == mvd->vdev_top)
754 775 vdev_top_transfer(cvd, mvd);
755 776
756 777 return (mvd);
757 778 }
758 779
759 780 /*
760 781 * Remove a 1-way mirror/replacing vdev from the tree.
761 782 */
762 783 void
763 784 vdev_remove_parent(vdev_t *cvd)
764 785 {
765 786 vdev_t *mvd = cvd->vdev_parent;
766 787 vdev_t *pvd = mvd->vdev_parent;
767 788
768 789 ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
769 790
770 791 ASSERT(mvd->vdev_children == 1);
771 792 ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
772 793 mvd->vdev_ops == &vdev_replacing_ops ||
773 794 mvd->vdev_ops == &vdev_spare_ops);
774 795 cvd->vdev_ashift = mvd->vdev_ashift;
775 796
776 797 vdev_remove_child(mvd, cvd);
777 798 vdev_remove_child(pvd, mvd);
778 799
779 800 /*
780 801 * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
781 802 * Otherwise, we could have detached an offline device, and when we
782 803 * go to import the pool we'll think we have two top-level vdevs,
783 804 * instead of a different version of the same top-level vdev.
784 805 */
785 806 if (mvd->vdev_top == mvd) {
786 807 uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
787 808 cvd->vdev_orig_guid = cvd->vdev_guid;
788 809 cvd->vdev_guid += guid_delta;
789 810 cvd->vdev_guid_sum += guid_delta;
790 811 }
791 812 cvd->vdev_id = mvd->vdev_id;
792 813 vdev_add_child(pvd, cvd);
793 814 vdev_top_update(cvd->vdev_top, cvd->vdev_top);
794 815
795 816 if (cvd == cvd->vdev_top)
796 817 vdev_top_transfer(mvd, cvd);
797 818
798 819 ASSERT(mvd->vdev_children == 0);
799 820 vdev_free(mvd);
800 821 }
801 822
802 823 int
803 824 vdev_metaslab_init(vdev_t *vd, uint64_t txg)
804 825 {
805 826 spa_t *spa = vd->vdev_spa;
806 827 objset_t *mos = spa->spa_meta_objset;
807 828 uint64_t m;
808 829 uint64_t oldc = vd->vdev_ms_count;
809 830 uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
810 831 metaslab_t **mspp;
811 832 int error;
812 833
813 834 ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER));
814 835
815 836 /*
816 837 * This vdev is not being allocated from yet or is a hole.
817 838 */
818 839 if (vd->vdev_ms_shift == 0)
819 840 return (0);
820 841
821 842 ASSERT(!vd->vdev_ishole);
822 843
823 844 /*
824 845 * Compute the raidz-deflation ratio. Note, we hard-code
825 846 * in 128k (1 << 17) because it is the current "typical" blocksize.
826 847 * Even if SPA_MAXBLOCKSIZE changes, this algorithm must never change,
827 848 * or we will inconsistently account for existing bp's.
828 849 */
829 850 vd->vdev_deflate_ratio = (1 << 17) /
830 851 (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
831 852
832 853 ASSERT(oldc <= newc);
833 854
834 855 mspp = kmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
835 856
836 857 if (oldc != 0) {
837 858 bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp));
838 859 kmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
839 860 }
840 861
841 862 vd->vdev_ms = mspp;
842 863 vd->vdev_ms_count = newc;
843 864
844 865 for (m = oldc; m < newc; m++) {
845 866 uint64_t object = 0;
846 867
847 868 if (txg == 0) {
848 869 error = dmu_read(mos, vd->vdev_ms_array,
849 870 m * sizeof (uint64_t), sizeof (uint64_t), &object,
850 871 DMU_READ_PREFETCH);
851 872 if (error)
852 873 return (error);
853 874 }
854 875 vd->vdev_ms[m] = metaslab_init(vd->vdev_mg, m, object, txg);
855 876 }
856 877
857 878 if (txg == 0)
858 879 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER);
859 880
860 881 /*
861 882 * If the vdev is being removed we don't activate
862 883 * the metaslabs since we want to ensure that no new
863 884 * allocations are performed on this device.
864 885 */
865 886 if (oldc == 0 && !vd->vdev_removing)
866 887 metaslab_group_activate(vd->vdev_mg);
867 888
868 889 if (txg == 0)
869 890 spa_config_exit(spa, SCL_ALLOC, FTAG);
870 891
871 892 return (0);
872 893 }
873 894
874 895 void
875 896 vdev_metaslab_fini(vdev_t *vd)
876 897 {
877 898 uint64_t m;
878 899 uint64_t count = vd->vdev_ms_count;
879 900
880 901 if (vd->vdev_ms != NULL) {
881 902 metaslab_group_passivate(vd->vdev_mg);
882 903 for (m = 0; m < count; m++) {
883 904 metaslab_t *msp = vd->vdev_ms[m];
884 905
885 906 if (msp != NULL)
886 907 metaslab_fini(msp);
887 908 }
888 909 kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
889 910 vd->vdev_ms = NULL;
890 911 }
891 912 }
892 913
893 914 typedef struct vdev_probe_stats {
894 915 boolean_t vps_readable;
895 916 boolean_t vps_writeable;
896 917 int vps_flags;
897 918 } vdev_probe_stats_t;
898 919
899 920 static void
900 921 vdev_probe_done(zio_t *zio)
901 922 {
902 923 spa_t *spa = zio->io_spa;
903 924 vdev_t *vd = zio->io_vd;
904 925 vdev_probe_stats_t *vps = zio->io_private;
905 926
906 927 ASSERT(vd->vdev_probe_zio != NULL);
907 928
908 929 if (zio->io_type == ZIO_TYPE_READ) {
909 930 if (zio->io_error == 0)
910 931 vps->vps_readable = 1;
911 932 if (zio->io_error == 0 && spa_writeable(spa)) {
912 933 zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
913 934 zio->io_offset, zio->io_size, zio->io_data,
914 935 ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
915 936 ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
916 937 } else {
917 938 zio_buf_free(zio->io_data, zio->io_size);
918 939 }
919 940 } else if (zio->io_type == ZIO_TYPE_WRITE) {
920 941 if (zio->io_error == 0)
921 942 vps->vps_writeable = 1;
922 943 zio_buf_free(zio->io_data, zio->io_size);
923 944 } else if (zio->io_type == ZIO_TYPE_NULL) {
924 945 zio_t *pio;
925 946
926 947 vd->vdev_cant_read |= !vps->vps_readable;
927 948 vd->vdev_cant_write |= !vps->vps_writeable;
928 949
929 950 if (vdev_readable(vd) &&
930 951 (vdev_writeable(vd) || !spa_writeable(spa))) {
931 952 zio->io_error = 0;
932 953 } else {
933 954 ASSERT(zio->io_error != 0);
934 955 zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
935 956 spa, vd, NULL, 0, 0);
936 957 zio->io_error = SET_ERROR(ENXIO);
937 958 }
938 959
939 960 mutex_enter(&vd->vdev_probe_lock);
940 961 ASSERT(vd->vdev_probe_zio == zio);
941 962 vd->vdev_probe_zio = NULL;
942 963 mutex_exit(&vd->vdev_probe_lock);
943 964
944 965 while ((pio = zio_walk_parents(zio)) != NULL)
945 966 if (!vdev_accessible(vd, pio))
946 967 pio->io_error = SET_ERROR(ENXIO);
947 968
948 969 kmem_free(vps, sizeof (*vps));
949 970 }
950 971 }
951 972
952 973 /*
953 974 * Determine whether this device is accessible.
954 975 *
955 976 * Read and write to several known locations: the pad regions of each
956 977 * vdev label but the first, which we leave alone in case it contains
957 978 * a VTOC.
958 979 */
959 980 zio_t *
960 981 vdev_probe(vdev_t *vd, zio_t *zio)
961 982 {
962 983 spa_t *spa = vd->vdev_spa;
963 984 vdev_probe_stats_t *vps = NULL;
964 985 zio_t *pio;
965 986
966 987 ASSERT(vd->vdev_ops->vdev_op_leaf);
967 988
968 989 /*
969 990 * Don't probe the probe.
970 991 */
971 992 if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
972 993 return (NULL);
973 994
974 995 /*
975 996 * To prevent 'probe storms' when a device fails, we create
976 997 * just one probe i/o at a time. All zios that want to probe
977 998 * this vdev will become parents of the probe io.
978 999 */
979 1000 mutex_enter(&vd->vdev_probe_lock);
980 1001
981 1002 if ((pio = vd->vdev_probe_zio) == NULL) {
982 1003 vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
983 1004
984 1005 vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
985 1006 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
986 1007 ZIO_FLAG_TRYHARD;
987 1008
988 1009 if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
989 1010 /*
990 1011 * vdev_cant_read and vdev_cant_write can only
991 1012 * transition from TRUE to FALSE when we have the
992 1013 * SCL_ZIO lock as writer; otherwise they can only
993 1014 * transition from FALSE to TRUE. This ensures that
994 1015 * any zio looking at these values can assume that
995 1016 * failures persist for the life of the I/O. That's
996 1017 * important because when a device has intermittent
997 1018 * connectivity problems, we want to ensure that
998 1019 * they're ascribed to the device (ENXIO) and not
999 1020 * the zio (EIO).
1000 1021 *
1001 1022 * Since we hold SCL_ZIO as writer here, clear both
1002 1023 * values so the probe can reevaluate from first
1003 1024 * principles.
1004 1025 */
1005 1026 vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
1006 1027 vd->vdev_cant_read = B_FALSE;
1007 1028 vd->vdev_cant_write = B_FALSE;
1008 1029 }
1009 1030
1010 1031 vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
1011 1032 vdev_probe_done, vps,
1012 1033 vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
1013 1034
1014 1035 /*
1015 1036 * We can't change the vdev state in this context, so we
1016 1037 * kick off an async task to do it on our behalf.
1017 1038 */
1018 1039 if (zio != NULL) {
1019 1040 vd->vdev_probe_wanted = B_TRUE;
1020 1041 spa_async_request(spa, SPA_ASYNC_PROBE);
1021 1042 }
1022 1043 }
1023 1044
1024 1045 if (zio != NULL)
1025 1046 zio_add_child(zio, pio);
1026 1047
1027 1048 mutex_exit(&vd->vdev_probe_lock);
1028 1049
1029 1050 if (vps == NULL) {
1030 1051 ASSERT(zio != NULL);
1031 1052 return (NULL);
1032 1053 }
1033 1054
1034 1055 for (int l = 1; l < VDEV_LABELS; l++) {
1035 1056 zio_nowait(zio_read_phys(pio, vd,
1036 1057 vdev_label_offset(vd->vdev_psize, l,
1037 1058 offsetof(vdev_label_t, vl_pad2)),
1038 1059 VDEV_PAD_SIZE, zio_buf_alloc(VDEV_PAD_SIZE),
1039 1060 ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1040 1061 ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1041 1062 }
1042 1063
1043 1064 if (zio == NULL)
1044 1065 return (pio);
1045 1066
1046 1067 zio_nowait(pio);
1047 1068 return (NULL);
1048 1069 }
1049 1070
1050 1071 static void
1051 1072 vdev_open_child(void *arg)
1052 1073 {
1053 1074 vdev_t *vd = arg;
1054 1075
1055 1076 vd->vdev_open_thread = curthread;
1056 1077 vd->vdev_open_error = vdev_open(vd);
1057 1078 vd->vdev_open_thread = NULL;
1058 1079 }
1059 1080
1060 1081 boolean_t
1061 1082 vdev_uses_zvols(vdev_t *vd)
1062 1083 {
1063 1084 if (vd->vdev_path && strncmp(vd->vdev_path, ZVOL_DIR,
1064 1085 strlen(ZVOL_DIR)) == 0)
1065 1086 return (B_TRUE);
1066 1087 for (int c = 0; c < vd->vdev_children; c++)
1067 1088 if (vdev_uses_zvols(vd->vdev_child[c]))
1068 1089 return (B_TRUE);
1069 1090 return (B_FALSE);
1070 1091 }
1071 1092
1072 1093 void
1073 1094 vdev_open_children(vdev_t *vd)
1074 1095 {
1075 1096 taskq_t *tq;
1076 1097 int children = vd->vdev_children;
1077 1098
1078 1099 /*
1079 1100 * in order to handle pools on top of zvols, do the opens
1080 1101 * in a single thread so that the same thread holds the
1081 1102 * spa_namespace_lock
1082 1103 */
1083 1104 if (vdev_uses_zvols(vd)) {
1084 1105 for (int c = 0; c < children; c++)
1085 1106 vd->vdev_child[c]->vdev_open_error =
1086 1107 vdev_open(vd->vdev_child[c]);
1087 1108 return;
1088 1109 }
1089 1110 tq = taskq_create("vdev_open", children, minclsyspri,
1090 1111 children, children, TASKQ_PREPOPULATE);
1091 1112
1092 1113 for (int c = 0; c < children; c++)
1093 1114 VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c],
1094 1115 TQ_SLEEP) != NULL);
1095 1116
1096 1117 taskq_destroy(tq);
1097 1118 }
1098 1119
1099 1120 /*
1100 1121 * Prepare a virtual device for access.
1101 1122 */
1102 1123 int
1103 1124 vdev_open(vdev_t *vd)
1104 1125 {
1105 1126 spa_t *spa = vd->vdev_spa;
1106 1127 int error;
1107 1128 uint64_t osize = 0;
1108 1129 uint64_t max_osize = 0;
1109 1130 uint64_t asize, max_asize, psize;
1110 1131 uint64_t ashift = 0;
1111 1132
1112 1133 ASSERT(vd->vdev_open_thread == curthread ||
1113 1134 spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1114 1135 ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1115 1136 vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1116 1137 vd->vdev_state == VDEV_STATE_OFFLINE);
1117 1138
1118 1139 vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1119 1140 vd->vdev_cant_read = B_FALSE;
1120 1141 vd->vdev_cant_write = B_FALSE;
1121 1142 vd->vdev_min_asize = vdev_get_min_asize(vd);
1122 1143
1123 1144 /*
1124 1145 * If this vdev is not removed, check its fault status. If it's
1125 1146 * faulted, bail out of the open.
1126 1147 */
1127 1148 if (!vd->vdev_removed && vd->vdev_faulted) {
1128 1149 ASSERT(vd->vdev_children == 0);
1129 1150 ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1130 1151 vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1131 1152 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1132 1153 vd->vdev_label_aux);
1133 1154 return (SET_ERROR(ENXIO));
1134 1155 } else if (vd->vdev_offline) {
1135 1156 ASSERT(vd->vdev_children == 0);
1136 1157 vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1137 1158 return (SET_ERROR(ENXIO));
1138 1159 }
1139 1160
1140 1161 error = vd->vdev_ops->vdev_op_open(vd, &osize, &max_osize, &ashift);
1141 1162
1142 1163 /*
1143 1164 * Reset the vdev_reopening flag so that we actually close
1144 1165 * the vdev on error.
1145 1166 */
1146 1167 vd->vdev_reopening = B_FALSE;
1147 1168 if (zio_injection_enabled && error == 0)
1148 1169 error = zio_handle_device_injection(vd, NULL, ENXIO);
1149 1170
1150 1171 if (error) {
1151 1172 if (vd->vdev_removed &&
1152 1173 vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1153 1174 vd->vdev_removed = B_FALSE;
1154 1175
1155 1176 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1156 1177 vd->vdev_stat.vs_aux);
1157 1178 return (error);
1158 1179 }
1159 1180
1160 1181 vd->vdev_removed = B_FALSE;
1161 1182
1162 1183 /*
1163 1184 * Recheck the faulted flag now that we have confirmed that
1164 1185 * the vdev is accessible. If we're faulted, bail.
1165 1186 */
1166 1187 if (vd->vdev_faulted) {
1167 1188 ASSERT(vd->vdev_children == 0);
1168 1189 ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1169 1190 vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1170 1191 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1171 1192 vd->vdev_label_aux);
1172 1193 return (SET_ERROR(ENXIO));
1173 1194 }
1174 1195
1175 1196 if (vd->vdev_degraded) {
1176 1197 ASSERT(vd->vdev_children == 0);
1177 1198 vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1178 1199 VDEV_AUX_ERR_EXCEEDED);
1179 1200 } else {
1180 1201 vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1181 1202 }
1182 1203
1183 1204 /*
1184 1205 * For hole or missing vdevs we just return success.
1185 1206 */
1186 1207 if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1187 1208 return (0);
1188 1209
1189 1210 for (int c = 0; c < vd->vdev_children; c++) {
1190 1211 if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1191 1212 vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1192 1213 VDEV_AUX_NONE);
1193 1214 break;
1194 1215 }
1195 1216 }
1196 1217
1197 1218 osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1198 1219 max_osize = P2ALIGN(max_osize, (uint64_t)sizeof (vdev_label_t));
1199 1220
1200 1221 if (vd->vdev_children == 0) {
1201 1222 if (osize < SPA_MINDEVSIZE) {
1202 1223 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1203 1224 VDEV_AUX_TOO_SMALL);
1204 1225 return (SET_ERROR(EOVERFLOW));
1205 1226 }
1206 1227 psize = osize;
1207 1228 asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1208 1229 max_asize = max_osize - (VDEV_LABEL_START_SIZE +
1209 1230 VDEV_LABEL_END_SIZE);
1210 1231 } else {
1211 1232 if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1212 1233 (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1213 1234 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1214 1235 VDEV_AUX_TOO_SMALL);
1215 1236 return (SET_ERROR(EOVERFLOW));
1216 1237 }
1217 1238 psize = 0;
1218 1239 asize = osize;
1219 1240 max_asize = max_osize;
1220 1241 }
1221 1242
1222 1243 vd->vdev_psize = psize;
1223 1244
1224 1245 /*
1225 1246 * Make sure the allocatable size hasn't shrunk.
1226 1247 */
1227 1248 if (asize < vd->vdev_min_asize) {
1228 1249 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1229 1250 VDEV_AUX_BAD_LABEL);
1230 1251 return (SET_ERROR(EINVAL));
1231 1252 }
1232 1253
1233 1254 if (vd->vdev_asize == 0) {
1234 1255 /*
1235 1256 * This is the first-ever open, so use the computed values.
1236 1257 * For testing purposes, a higher ashift can be requested.
1237 1258 */
1238 1259 vd->vdev_asize = asize;
1239 1260 vd->vdev_max_asize = max_asize;
1240 1261 vd->vdev_ashift = MAX(ashift, vd->vdev_ashift);
1241 1262 } else {
1242 1263 /*
1243 1264 * Detect if the alignment requirement has increased.
1244 1265 * We don't want to make the pool unavailable, just
1245 1266 * issue a warning instead.
1246 1267 */
1247 1268 if (ashift > vd->vdev_top->vdev_ashift &&
1248 1269 vd->vdev_ops->vdev_op_leaf) {
1249 1270 cmn_err(CE_WARN,
1250 1271 "Disk, '%s', has a block alignment that is "
1251 1272 "larger than the pool's alignment\n",
1252 1273 vd->vdev_path);
1253 1274 }
1254 1275 vd->vdev_max_asize = max_asize;
1255 1276 }
1256 1277
1257 1278 /*
1258 1279 * If all children are healthy and the asize has increased,
1259 1280 * then we've experienced dynamic LUN growth. If automatic
1260 1281 * expansion is enabled then use the additional space.
1261 1282 */
1262 1283 if (vd->vdev_state == VDEV_STATE_HEALTHY && asize > vd->vdev_asize &&
1263 1284 (vd->vdev_expanding || spa->spa_autoexpand))
1264 1285 vd->vdev_asize = asize;
1265 1286
1266 1287 vdev_set_min_asize(vd);
1267 1288
1268 1289 /*
1269 1290 * Ensure we can issue some IO before declaring the
1270 1291 * vdev open for business.
1271 1292 */
1272 1293 if (vd->vdev_ops->vdev_op_leaf &&
1273 1294 (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1274 1295 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1275 1296 VDEV_AUX_ERR_EXCEEDED);
1276 1297 return (error);
1277 1298 }
1278 1299
1279 1300 /*
1280 1301 * If a leaf vdev has a DTL, and seems healthy, then kick off a
1281 1302 * resilver. But don't do this if we are doing a reopen for a scrub,
1282 1303 * since this would just restart the scrub we are already doing.
1283 1304 */
1284 1305 if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1285 1306 vdev_resilver_needed(vd, NULL, NULL))
1286 1307 spa_async_request(spa, SPA_ASYNC_RESILVER);
1287 1308
1288 1309 return (0);
1289 1310 }
1290 1311
1291 1312 /*
1292 1313 * Called once the vdevs are all opened, this routine validates the label
1293 1314 * contents. This needs to be done before vdev_load() so that we don't
1294 1315 * inadvertently do repair I/Os to the wrong device.
1295 1316 *
1296 1317 * If 'strict' is false ignore the spa guid check. This is necessary because
1297 1318 * if the machine crashed during a re-guid the new guid might have been written
1298 1319 * to all of the vdev labels, but not the cached config. The strict check
1299 1320 * will be performed when the pool is opened again using the mos config.
1300 1321 *
1301 1322 * This function will only return failure if one of the vdevs indicates that it
1302 1323 * has since been destroyed or exported. This is only possible if
1303 1324 * /etc/zfs/zpool.cache was readonly at the time. Otherwise, the vdev state
1304 1325 * will be updated but the function will return 0.
1305 1326 */
1306 1327 int
1307 1328 vdev_validate(vdev_t *vd, boolean_t strict)
1308 1329 {
1309 1330 spa_t *spa = vd->vdev_spa;
1310 1331 nvlist_t *label;
1311 1332 uint64_t guid = 0, top_guid;
1312 1333 uint64_t state;
1313 1334
1314 1335 for (int c = 0; c < vd->vdev_children; c++)
1315 1336 if (vdev_validate(vd->vdev_child[c], strict) != 0)
1316 1337 return (SET_ERROR(EBADF));
1317 1338
1318 1339 /*
1319 1340 * If the device has already failed, or was marked offline, don't do
1320 1341 * any further validation. Otherwise, label I/O will fail and we will
1321 1342 * overwrite the previous state.
1322 1343 */
1323 1344 if (vd->vdev_ops->vdev_op_leaf && vdev_readable(vd)) {
1324 1345 uint64_t aux_guid = 0;
1325 1346 nvlist_t *nvl;
1326 1347 uint64_t txg = spa_last_synced_txg(spa) != 0 ?
1327 1348 spa_last_synced_txg(spa) : -1ULL;
1328 1349
1329 1350 if ((label = vdev_label_read_config(vd, txg)) == NULL) {
1330 1351 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1331 1352 VDEV_AUX_BAD_LABEL);
1332 1353 return (0);
1333 1354 }
1334 1355
1335 1356 /*
1336 1357 * Determine if this vdev has been split off into another
1337 1358 * pool. If so, then refuse to open it.
1338 1359 */
1339 1360 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID,
1340 1361 &aux_guid) == 0 && aux_guid == spa_guid(spa)) {
1341 1362 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1342 1363 VDEV_AUX_SPLIT_POOL);
1343 1364 nvlist_free(label);
1344 1365 return (0);
1345 1366 }
1346 1367
1347 1368 if (strict && (nvlist_lookup_uint64(label,
1348 1369 ZPOOL_CONFIG_POOL_GUID, &guid) != 0 ||
1349 1370 guid != spa_guid(spa))) {
1350 1371 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1351 1372 VDEV_AUX_CORRUPT_DATA);
1352 1373 nvlist_free(label);
1353 1374 return (0);
1354 1375 }
1355 1376
1356 1377 if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl)
1357 1378 != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID,
1358 1379 &aux_guid) != 0)
1359 1380 aux_guid = 0;
1360 1381
1361 1382 /*
1362 1383 * If this vdev just became a top-level vdev because its
1363 1384 * sibling was detached, it will have adopted the parent's
1364 1385 * vdev guid -- but the label may or may not be on disk yet.
1365 1386 * Fortunately, either version of the label will have the
1366 1387 * same top guid, so if we're a top-level vdev, we can
1367 1388 * safely compare to that instead.
1368 1389 *
1369 1390 * If we split this vdev off instead, then we also check the
1370 1391 * original pool's guid. We don't want to consider the vdev
1371 1392 * corrupt if it is partway through a split operation.
1372 1393 */
1373 1394 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID,
1374 1395 &guid) != 0 ||
1375 1396 nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID,
1376 1397 &top_guid) != 0 ||
1377 1398 ((vd->vdev_guid != guid && vd->vdev_guid != aux_guid) &&
1378 1399 (vd->vdev_guid != top_guid || vd != vd->vdev_top))) {
1379 1400 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1380 1401 VDEV_AUX_CORRUPT_DATA);
1381 1402 nvlist_free(label);
1382 1403 return (0);
1383 1404 }
1384 1405
1385 1406 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1386 1407 &state) != 0) {
1387 1408 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1388 1409 VDEV_AUX_CORRUPT_DATA);
1389 1410 nvlist_free(label);
1390 1411 return (0);
1391 1412 }
1392 1413
1393 1414 nvlist_free(label);
1394 1415
1395 1416 /*
1396 1417 * If this is a verbatim import, no need to check the
1397 1418 * state of the pool.
1398 1419 */
1399 1420 if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) &&
1400 1421 spa_load_state(spa) == SPA_LOAD_OPEN &&
1401 1422 state != POOL_STATE_ACTIVE)
1402 1423 return (SET_ERROR(EBADF));
1403 1424
1404 1425 /*
1405 1426 * If we were able to open and validate a vdev that was
1406 1427 * previously marked permanently unavailable, clear that state
1407 1428 * now.
1408 1429 */
1409 1430 if (vd->vdev_not_present)
1410 1431 vd->vdev_not_present = 0;
1411 1432 }
1412 1433
1413 1434 return (0);
1414 1435 }
1415 1436
1416 1437 /*
1417 1438 * Close a virtual device.
1418 1439 */
1419 1440 void
1420 1441 vdev_close(vdev_t *vd)
1421 1442 {
1422 1443 spa_t *spa = vd->vdev_spa;
1423 1444 vdev_t *pvd = vd->vdev_parent;
1424 1445
1425 1446 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1426 1447
1427 1448 /*
1428 1449 * If our parent is reopening, then we are as well, unless we are
1429 1450 * going offline.
1430 1451 */
1431 1452 if (pvd != NULL && pvd->vdev_reopening)
1432 1453 vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline);
1433 1454
1434 1455 vd->vdev_ops->vdev_op_close(vd);
1435 1456
1436 1457 vdev_cache_purge(vd);
1437 1458
1438 1459 /*
1439 1460 * We record the previous state before we close it, so that if we are
1440 1461 * doing a reopen(), we don't generate FMA ereports if we notice that
1441 1462 * it's still faulted.
1442 1463 */
1443 1464 vd->vdev_prevstate = vd->vdev_state;
1444 1465
1445 1466 if (vd->vdev_offline)
1446 1467 vd->vdev_state = VDEV_STATE_OFFLINE;
1447 1468 else
1448 1469 vd->vdev_state = VDEV_STATE_CLOSED;
1449 1470 vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1450 1471 }
1451 1472
1452 1473 void
1453 1474 vdev_hold(vdev_t *vd)
1454 1475 {
1455 1476 spa_t *spa = vd->vdev_spa;
1456 1477
1457 1478 ASSERT(spa_is_root(spa));
1458 1479 if (spa->spa_state == POOL_STATE_UNINITIALIZED)
1459 1480 return;
1460 1481
1461 1482 for (int c = 0; c < vd->vdev_children; c++)
1462 1483 vdev_hold(vd->vdev_child[c]);
1463 1484
1464 1485 if (vd->vdev_ops->vdev_op_leaf)
1465 1486 vd->vdev_ops->vdev_op_hold(vd);
1466 1487 }
1467 1488
1468 1489 void
1469 1490 vdev_rele(vdev_t *vd)
1470 1491 {
1471 1492 spa_t *spa = vd->vdev_spa;
1472 1493
1473 1494 ASSERT(spa_is_root(spa));
1474 1495 for (int c = 0; c < vd->vdev_children; c++)
1475 1496 vdev_rele(vd->vdev_child[c]);
1476 1497
1477 1498 if (vd->vdev_ops->vdev_op_leaf)
1478 1499 vd->vdev_ops->vdev_op_rele(vd);
1479 1500 }
1480 1501
1481 1502 /*
1482 1503 * Reopen all interior vdevs and any unopened leaves. We don't actually
1483 1504 * reopen leaf vdevs which had previously been opened as they might deadlock
1484 1505 * on the spa_config_lock. Instead we only obtain the leaf's physical size.
1485 1506 * If the leaf has never been opened then open it, as usual.
1486 1507 */
1487 1508 void
1488 1509 vdev_reopen(vdev_t *vd)
1489 1510 {
1490 1511 spa_t *spa = vd->vdev_spa;
1491 1512
1492 1513 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1493 1514
1494 1515 /* set the reopening flag unless we're taking the vdev offline */
1495 1516 vd->vdev_reopening = !vd->vdev_offline;
1496 1517 vdev_close(vd);
1497 1518 (void) vdev_open(vd);
1498 1519
1499 1520 /*
1500 1521 * Call vdev_validate() here to make sure we have the same device.
1501 1522 * Otherwise, a device with an invalid label could be successfully
1502 1523 * opened in response to vdev_reopen().
1503 1524 */
1504 1525 if (vd->vdev_aux) {
1505 1526 (void) vdev_validate_aux(vd);
1506 1527 if (vdev_readable(vd) && vdev_writeable(vd) &&
1507 1528 vd->vdev_aux == &spa->spa_l2cache &&
1508 1529 !l2arc_vdev_present(vd))
1509 1530 l2arc_add_vdev(spa, vd);
1510 1531 } else {
1511 1532 (void) vdev_validate(vd, B_TRUE);
1512 1533 }
1513 1534
1514 1535 /*
1515 1536 * Reassess parent vdev's health.
1516 1537 */
1517 1538 vdev_propagate_state(vd);
1518 1539 }
1519 1540
1520 1541 int
1521 1542 vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1522 1543 {
1523 1544 int error;
1524 1545
1525 1546 /*
1526 1547 * Normally, partial opens (e.g. of a mirror) are allowed.
1527 1548 * For a create, however, we want to fail the request if
1528 1549 * there are any components we can't open.
1529 1550 */
1530 1551 error = vdev_open(vd);
1531 1552
1532 1553 if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
1533 1554 vdev_close(vd);
1534 1555 return (error ? error : ENXIO);
1535 1556 }
1536 1557
1537 1558 /*
1538 1559 * Recursively load DTLs and initialize all labels.
1539 1560 */
1540 1561 if ((error = vdev_dtl_load(vd)) != 0 ||
1541 1562 (error = vdev_label_init(vd, txg, isreplacing ?
1542 1563 VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
1543 1564 vdev_close(vd);
1544 1565 return (error);
1545 1566 }
1546 1567
1547 1568 return (0);
1548 1569 }
1549 1570
1550 1571 void
1551 1572 vdev_metaslab_set_size(vdev_t *vd)
1552 1573 {
1553 1574 /*
1554 1575 * Aim for roughly 200 metaslabs per vdev.
1555 1576 */
1556 1577 vd->vdev_ms_shift = highbit64(vd->vdev_asize / 200);
1557 1578 vd->vdev_ms_shift = MAX(vd->vdev_ms_shift, SPA_MAXBLOCKSHIFT);
1558 1579 }
1559 1580
1560 1581 void
1561 1582 vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
1562 1583 {
1563 1584 ASSERT(vd == vd->vdev_top);
1564 1585 ASSERT(!vd->vdev_ishole);
1565 1586 ASSERT(ISP2(flags));
1566 1587 ASSERT(spa_writeable(vd->vdev_spa));
1567 1588
1568 1589 if (flags & VDD_METASLAB)
1569 1590 (void) txg_list_add(&vd->vdev_ms_list, arg, txg);
1570 1591
1571 1592 if (flags & VDD_DTL)
1572 1593 (void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
1573 1594
1574 1595 (void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
1575 1596 }
1576 1597
1577 1598 void
1578 1599 vdev_dirty_leaves(vdev_t *vd, int flags, uint64_t txg)
1579 1600 {
1580 1601 for (int c = 0; c < vd->vdev_children; c++)
1581 1602 vdev_dirty_leaves(vd->vdev_child[c], flags, txg);
1582 1603
1583 1604 if (vd->vdev_ops->vdev_op_leaf)
1584 1605 vdev_dirty(vd->vdev_top, flags, vd, txg);
1585 1606 }
1586 1607
1587 1608 /*
1588 1609 * DTLs.
1589 1610 *
1590 1611 * A vdev's DTL (dirty time log) is the set of transaction groups for which
1591 1612 * the vdev has less than perfect replication. There are four kinds of DTL:
1592 1613 *
1593 1614 * DTL_MISSING: txgs for which the vdev has no valid copies of the data
1594 1615 *
1595 1616 * DTL_PARTIAL: txgs for which data is available, but not fully replicated
1596 1617 *
1597 1618 * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
1598 1619 * scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
1599 1620 * txgs that was scrubbed.
1600 1621 *
1601 1622 * DTL_OUTAGE: txgs which cannot currently be read, whether due to
1602 1623 * persistent errors or just some device being offline.
1603 1624 * Unlike the other three, the DTL_OUTAGE map is not generally
1604 1625 * maintained; it's only computed when needed, typically to
1605 1626 * determine whether a device can be detached.
1606 1627 *
1607 1628 * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
1608 1629 * either has the data or it doesn't.
1609 1630 *
1610 1631 * For interior vdevs such as mirror and RAID-Z the picture is more complex.
1611 1632 * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
1612 1633 * if any child is less than fully replicated, then so is its parent.
1613 1634 * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
1614 1635 * comprising only those txgs which appear in 'maxfaults' or more children;
1615 1636 * those are the txgs we don't have enough replication to read. For example,
1616 1637 * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
1617 1638 * thus, its DTL_MISSING consists of the set of txgs that appear in more than
1618 1639 * two child DTL_MISSING maps.
1619 1640 *
1620 1641 * It should be clear from the above that to compute the DTLs and outage maps
1621 1642 * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
1622 1643 * Therefore, that is all we keep on disk. When loading the pool, or after
1623 1644 * a configuration change, we generate all other DTLs from first principles.
1624 1645 */
1625 1646 void
1626 1647 vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1627 1648 {
1628 1649 range_tree_t *rt = vd->vdev_dtl[t];
1629 1650
1630 1651 ASSERT(t < DTL_TYPES);
1631 1652 ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1632 1653 ASSERT(spa_writeable(vd->vdev_spa));
1633 1654
1634 1655 mutex_enter(rt->rt_lock);
1635 1656 if (!range_tree_contains(rt, txg, size))
1636 1657 range_tree_add(rt, txg, size);
1637 1658 mutex_exit(rt->rt_lock);
1638 1659 }
1639 1660
1640 1661 boolean_t
1641 1662 vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1642 1663 {
1643 1664 range_tree_t *rt = vd->vdev_dtl[t];
1644 1665 boolean_t dirty = B_FALSE;
1645 1666
1646 1667 ASSERT(t < DTL_TYPES);
1647 1668 ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1648 1669
1649 1670 mutex_enter(rt->rt_lock);
1650 1671 if (range_tree_space(rt) != 0)
1651 1672 dirty = range_tree_contains(rt, txg, size);
1652 1673 mutex_exit(rt->rt_lock);
1653 1674
1654 1675 return (dirty);
1655 1676 }
1656 1677
1657 1678 boolean_t
1658 1679 vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
1659 1680 {
1660 1681 range_tree_t *rt = vd->vdev_dtl[t];
1661 1682 boolean_t empty;
1662 1683
1663 1684 mutex_enter(rt->rt_lock);
1664 1685 empty = (range_tree_space(rt) == 0);
1665 1686 mutex_exit(rt->rt_lock);
1666 1687
1667 1688 return (empty);
1668 1689 }
1669 1690
1670 1691 /*
1671 1692 * Returns the lowest txg in the DTL range.
1672 1693 */
1673 1694 static uint64_t
1674 1695 vdev_dtl_min(vdev_t *vd)
1675 1696 {
1676 1697 range_seg_t *rs;
1677 1698
1678 1699 ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
1679 1700 ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
1680 1701 ASSERT0(vd->vdev_children);
1681 1702
1682 1703 rs = avl_first(&vd->vdev_dtl[DTL_MISSING]->rt_root);
1683 1704 return (rs->rs_start - 1);
1684 1705 }
1685 1706
1686 1707 /*
1687 1708 * Returns the highest txg in the DTL.
1688 1709 */
1689 1710 static uint64_t
1690 1711 vdev_dtl_max(vdev_t *vd)
1691 1712 {
1692 1713 range_seg_t *rs;
1693 1714
1694 1715 ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
1695 1716 ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
1696 1717 ASSERT0(vd->vdev_children);
1697 1718
1698 1719 rs = avl_last(&vd->vdev_dtl[DTL_MISSING]->rt_root);
1699 1720 return (rs->rs_end);
1700 1721 }
1701 1722
1702 1723 /*
1703 1724 * Determine if a resilvering vdev should remove any DTL entries from
1704 1725 * its range. If the vdev was resilvering for the entire duration of the
1705 1726 * scan then it should excise that range from its DTLs. Otherwise, this
1706 1727 * vdev is considered partially resilvered and should leave its DTL
1707 1728 * entries intact. The comment in vdev_dtl_reassess() describes how we
1708 1729 * excise the DTLs.
1709 1730 */
1710 1731 static boolean_t
1711 1732 vdev_dtl_should_excise(vdev_t *vd)
1712 1733 {
1713 1734 spa_t *spa = vd->vdev_spa;
1714 1735 dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1715 1736
1716 1737 ASSERT0(scn->scn_phys.scn_errors);
1717 1738 ASSERT0(vd->vdev_children);
1718 1739
1719 1740 if (vd->vdev_resilver_txg == 0 ||
1720 1741 range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0)
1721 1742 return (B_TRUE);
1722 1743
1723 1744 /*
1724 1745 * When a resilver is initiated the scan will assign the scn_max_txg
1725 1746 * value to the highest txg value that exists in all DTLs. If this
1726 1747 * device's max DTL is not part of this scan (i.e. it is not in
1727 1748 * the range (scn_min_txg, scn_max_txg] then it is not eligible
1728 1749 * for excision.
1729 1750 */
1730 1751 if (vdev_dtl_max(vd) <= scn->scn_phys.scn_max_txg) {
1731 1752 ASSERT3U(scn->scn_phys.scn_min_txg, <=, vdev_dtl_min(vd));
1732 1753 ASSERT3U(scn->scn_phys.scn_min_txg, <, vd->vdev_resilver_txg);
1733 1754 ASSERT3U(vd->vdev_resilver_txg, <=, scn->scn_phys.scn_max_txg);
1734 1755 return (B_TRUE);
1735 1756 }
1736 1757 return (B_FALSE);
1737 1758 }
1738 1759
1739 1760 /*
1740 1761 * Reassess DTLs after a config change or scrub completion.
1741 1762 */
1742 1763 void
1743 1764 vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
1744 1765 {
1745 1766 spa_t *spa = vd->vdev_spa;
1746 1767 avl_tree_t reftree;
1747 1768 int minref;
1748 1769
1749 1770 ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
1750 1771
1751 1772 for (int c = 0; c < vd->vdev_children; c++)
1752 1773 vdev_dtl_reassess(vd->vdev_child[c], txg,
1753 1774 scrub_txg, scrub_done);
1754 1775
1755 1776 if (vd == spa->spa_root_vdev || vd->vdev_ishole || vd->vdev_aux)
1756 1777 return;
1757 1778
1758 1779 if (vd->vdev_ops->vdev_op_leaf) {
1759 1780 dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1760 1781
1761 1782 mutex_enter(&vd->vdev_dtl_lock);
1762 1783
1763 1784 /*
1764 1785 * If we've completed a scan cleanly then determine
1765 1786 * if this vdev should remove any DTLs. We only want to
1766 1787 * excise regions on vdevs that were available during
1767 1788 * the entire duration of this scan.
1768 1789 */
1769 1790 if (scrub_txg != 0 &&
1770 1791 (spa->spa_scrub_started ||
1771 1792 (scn != NULL && scn->scn_phys.scn_errors == 0)) &&
1772 1793 vdev_dtl_should_excise(vd)) {
1773 1794 /*
1774 1795 * We completed a scrub up to scrub_txg. If we
1775 1796 * did it without rebooting, then the scrub dtl
1776 1797 * will be valid, so excise the old region and
1777 1798 * fold in the scrub dtl. Otherwise, leave the
1778 1799 * dtl as-is if there was an error.
1779 1800 *
1780 1801 * There's little trick here: to excise the beginning
1781 1802 * of the DTL_MISSING map, we put it into a reference
1782 1803 * tree and then add a segment with refcnt -1 that
1783 1804 * covers the range [0, scrub_txg). This means
1784 1805 * that each txg in that range has refcnt -1 or 0.
1785 1806 * We then add DTL_SCRUB with a refcnt of 2, so that
1786 1807 * entries in the range [0, scrub_txg) will have a
1787 1808 * positive refcnt -- either 1 or 2. We then convert
1788 1809 * the reference tree into the new DTL_MISSING map.
1789 1810 */
1790 1811 space_reftree_create(&reftree);
1791 1812 space_reftree_add_map(&reftree,
1792 1813 vd->vdev_dtl[DTL_MISSING], 1);
1793 1814 space_reftree_add_seg(&reftree, 0, scrub_txg, -1);
1794 1815 space_reftree_add_map(&reftree,
1795 1816 vd->vdev_dtl[DTL_SCRUB], 2);
1796 1817 space_reftree_generate_map(&reftree,
1797 1818 vd->vdev_dtl[DTL_MISSING], 1);
1798 1819 space_reftree_destroy(&reftree);
1799 1820 }
1800 1821 range_tree_vacate(vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
1801 1822 range_tree_walk(vd->vdev_dtl[DTL_MISSING],
1802 1823 range_tree_add, vd->vdev_dtl[DTL_PARTIAL]);
1803 1824 if (scrub_done)
1804 1825 range_tree_vacate(vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
1805 1826 range_tree_vacate(vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
1806 1827 if (!vdev_readable(vd))
1807 1828 range_tree_add(vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
1808 1829 else
1809 1830 range_tree_walk(vd->vdev_dtl[DTL_MISSING],
1810 1831 range_tree_add, vd->vdev_dtl[DTL_OUTAGE]);
1811 1832
1812 1833 /*
1813 1834 * If the vdev was resilvering and no longer has any
1814 1835 * DTLs then reset its resilvering flag.
1815 1836 */
1816 1837 if (vd->vdev_resilver_txg != 0 &&
1817 1838 range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0 &&
1818 1839 range_tree_space(vd->vdev_dtl[DTL_OUTAGE]) == 0)
1819 1840 vd->vdev_resilver_txg = 0;
1820 1841
1821 1842 mutex_exit(&vd->vdev_dtl_lock);
1822 1843
1823 1844 if (txg != 0)
1824 1845 vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
1825 1846 return;
1826 1847 }
1827 1848
1828 1849 mutex_enter(&vd->vdev_dtl_lock);
1829 1850 for (int t = 0; t < DTL_TYPES; t++) {
1830 1851 /* account for child's outage in parent's missing map */
1831 1852 int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
1832 1853 if (t == DTL_SCRUB)
1833 1854 continue; /* leaf vdevs only */
1834 1855 if (t == DTL_PARTIAL)
1835 1856 minref = 1; /* i.e. non-zero */
1836 1857 else if (vd->vdev_nparity != 0)
1837 1858 minref = vd->vdev_nparity + 1; /* RAID-Z */
1838 1859 else
1839 1860 minref = vd->vdev_children; /* any kind of mirror */
1840 1861 space_reftree_create(&reftree);
1841 1862 for (int c = 0; c < vd->vdev_children; c++) {
1842 1863 vdev_t *cvd = vd->vdev_child[c];
1843 1864 mutex_enter(&cvd->vdev_dtl_lock);
1844 1865 space_reftree_add_map(&reftree, cvd->vdev_dtl[s], 1);
1845 1866 mutex_exit(&cvd->vdev_dtl_lock);
1846 1867 }
1847 1868 space_reftree_generate_map(&reftree, vd->vdev_dtl[t], minref);
1848 1869 space_reftree_destroy(&reftree);
1849 1870 }
1850 1871 mutex_exit(&vd->vdev_dtl_lock);
1851 1872 }
1852 1873
1853 1874 int
1854 1875 vdev_dtl_load(vdev_t *vd)
1855 1876 {
1856 1877 spa_t *spa = vd->vdev_spa;
1857 1878 objset_t *mos = spa->spa_meta_objset;
1858 1879 int error = 0;
1859 1880
1860 1881 if (vd->vdev_ops->vdev_op_leaf && vd->vdev_dtl_object != 0) {
1861 1882 ASSERT(!vd->vdev_ishole);
1862 1883
1863 1884 error = space_map_open(&vd->vdev_dtl_sm, mos,
1864 1885 vd->vdev_dtl_object, 0, -1ULL, 0, &vd->vdev_dtl_lock);
1865 1886 if (error)
1866 1887 return (error);
1867 1888 ASSERT(vd->vdev_dtl_sm != NULL);
1868 1889
1869 1890 mutex_enter(&vd->vdev_dtl_lock);
1870 1891
1871 1892 /*
1872 1893 * Now that we've opened the space_map we need to update
1873 1894 * the in-core DTL.
1874 1895 */
1875 1896 space_map_update(vd->vdev_dtl_sm);
1876 1897
1877 1898 error = space_map_load(vd->vdev_dtl_sm,
1878 1899 vd->vdev_dtl[DTL_MISSING], SM_ALLOC);
1879 1900 mutex_exit(&vd->vdev_dtl_lock);
1880 1901
1881 1902 return (error);
1882 1903 }
1883 1904
1884 1905 for (int c = 0; c < vd->vdev_children; c++) {
1885 1906 error = vdev_dtl_load(vd->vdev_child[c]);
1886 1907 if (error != 0)
1887 1908 break;
1888 1909 }
1889 1910
1890 1911 return (error);
1891 1912 }
1892 1913
1893 1914 void
1894 1915 vdev_dtl_sync(vdev_t *vd, uint64_t txg)
1895 1916 {
1896 1917 spa_t *spa = vd->vdev_spa;
1897 1918 range_tree_t *rt = vd->vdev_dtl[DTL_MISSING];
1898 1919 objset_t *mos = spa->spa_meta_objset;
1899 1920 range_tree_t *rtsync;
1900 1921 kmutex_t rtlock;
1901 1922 dmu_tx_t *tx;
1902 1923 uint64_t object = space_map_object(vd->vdev_dtl_sm);
1903 1924
1904 1925 ASSERT(!vd->vdev_ishole);
1905 1926 ASSERT(vd->vdev_ops->vdev_op_leaf);
1906 1927
1907 1928 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1908 1929
1909 1930 if (vd->vdev_detached || vd->vdev_top->vdev_removing) {
1910 1931 mutex_enter(&vd->vdev_dtl_lock);
1911 1932 space_map_free(vd->vdev_dtl_sm, tx);
1912 1933 space_map_close(vd->vdev_dtl_sm);
1913 1934 vd->vdev_dtl_sm = NULL;
1914 1935 mutex_exit(&vd->vdev_dtl_lock);
1915 1936 dmu_tx_commit(tx);
1916 1937 return;
1917 1938 }
1918 1939
1919 1940 if (vd->vdev_dtl_sm == NULL) {
1920 1941 uint64_t new_object;
1921 1942
1922 1943 new_object = space_map_alloc(mos, tx);
1923 1944 VERIFY3U(new_object, !=, 0);
1924 1945
1925 1946 VERIFY0(space_map_open(&vd->vdev_dtl_sm, mos, new_object,
1926 1947 0, -1ULL, 0, &vd->vdev_dtl_lock));
1927 1948 ASSERT(vd->vdev_dtl_sm != NULL);
1928 1949 }
1929 1950
1930 1951 mutex_init(&rtlock, NULL, MUTEX_DEFAULT, NULL);
1931 1952
1932 1953 rtsync = range_tree_create(NULL, NULL, &rtlock);
1933 1954
1934 1955 mutex_enter(&rtlock);
1935 1956
1936 1957 mutex_enter(&vd->vdev_dtl_lock);
1937 1958 range_tree_walk(rt, range_tree_add, rtsync);
1938 1959 mutex_exit(&vd->vdev_dtl_lock);
1939 1960
1940 1961 space_map_truncate(vd->vdev_dtl_sm, tx);
1941 1962 space_map_write(vd->vdev_dtl_sm, rtsync, SM_ALLOC, tx);
1942 1963 range_tree_vacate(rtsync, NULL, NULL);
1943 1964
1944 1965 range_tree_destroy(rtsync);
1945 1966
1946 1967 mutex_exit(&rtlock);
1947 1968 mutex_destroy(&rtlock);
1948 1969
1949 1970 /*
1950 1971 * If the object for the space map has changed then dirty
1951 1972 * the top level so that we update the config.
1952 1973 */
1953 1974 if (object != space_map_object(vd->vdev_dtl_sm)) {
1954 1975 zfs_dbgmsg("txg %llu, spa %s, DTL old object %llu, "
1955 1976 "new object %llu", txg, spa_name(spa), object,
1956 1977 space_map_object(vd->vdev_dtl_sm));
1957 1978 vdev_config_dirty(vd->vdev_top);
1958 1979 }
1959 1980
1960 1981 dmu_tx_commit(tx);
1961 1982
1962 1983 mutex_enter(&vd->vdev_dtl_lock);
1963 1984 space_map_update(vd->vdev_dtl_sm);
1964 1985 mutex_exit(&vd->vdev_dtl_lock);
1965 1986 }
1966 1987
1967 1988 /*
1968 1989 * Determine whether the specified vdev can be offlined/detached/removed
1969 1990 * without losing data.
1970 1991 */
1971 1992 boolean_t
1972 1993 vdev_dtl_required(vdev_t *vd)
1973 1994 {
1974 1995 spa_t *spa = vd->vdev_spa;
1975 1996 vdev_t *tvd = vd->vdev_top;
1976 1997 uint8_t cant_read = vd->vdev_cant_read;
1977 1998 boolean_t required;
1978 1999
1979 2000 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1980 2001
1981 2002 if (vd == spa->spa_root_vdev || vd == tvd)
1982 2003 return (B_TRUE);
1983 2004
1984 2005 /*
1985 2006 * Temporarily mark the device as unreadable, and then determine
1986 2007 * whether this results in any DTL outages in the top-level vdev.
1987 2008 * If not, we can safely offline/detach/remove the device.
1988 2009 */
1989 2010 vd->vdev_cant_read = B_TRUE;
1990 2011 vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1991 2012 required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
1992 2013 vd->vdev_cant_read = cant_read;
1993 2014 vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1994 2015
1995 2016 if (!required && zio_injection_enabled)
1996 2017 required = !!zio_handle_device_injection(vd, NULL, ECHILD);
1997 2018
1998 2019 return (required);
1999 2020 }
2000 2021
2001 2022 /*
2002 2023 * Determine if resilver is needed, and if so the txg range.
2003 2024 */
2004 2025 boolean_t
2005 2026 vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
2006 2027 {
2007 2028 boolean_t needed = B_FALSE;
2008 2029 uint64_t thismin = UINT64_MAX;
2009 2030 uint64_t thismax = 0;
2010 2031
2011 2032 if (vd->vdev_children == 0) {
2012 2033 mutex_enter(&vd->vdev_dtl_lock);
2013 2034 if (range_tree_space(vd->vdev_dtl[DTL_MISSING]) != 0 &&
2014 2035 vdev_writeable(vd)) {
2015 2036
2016 2037 thismin = vdev_dtl_min(vd);
2017 2038 thismax = vdev_dtl_max(vd);
2018 2039 needed = B_TRUE;
2019 2040 }
2020 2041 mutex_exit(&vd->vdev_dtl_lock);
2021 2042 } else {
2022 2043 for (int c = 0; c < vd->vdev_children; c++) {
2023 2044 vdev_t *cvd = vd->vdev_child[c];
2024 2045 uint64_t cmin, cmax;
2025 2046
2026 2047 if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
2027 2048 thismin = MIN(thismin, cmin);
2028 2049 thismax = MAX(thismax, cmax);
2029 2050 needed = B_TRUE;
2030 2051 }
2031 2052 }
2032 2053 }
2033 2054
2034 2055 if (needed && minp) {
2035 2056 *minp = thismin;
2036 2057 *maxp = thismax;
2037 2058 }
2038 2059 return (needed);
2039 2060 }
2040 2061
2041 2062 void
2042 2063 vdev_load(vdev_t *vd)
2043 2064 {
2044 2065 /*
2045 2066 * Recursively load all children.
2046 2067 */
2047 2068 for (int c = 0; c < vd->vdev_children; c++)
2048 2069 vdev_load(vd->vdev_child[c]);
2049 2070
2050 2071 /*
2051 2072 * If this is a top-level vdev, initialize its metaslabs.
2052 2073 */
2053 2074 if (vd == vd->vdev_top && !vd->vdev_ishole &&
2054 2075 (vd->vdev_ashift == 0 || vd->vdev_asize == 0 ||
2055 2076 vdev_metaslab_init(vd, 0) != 0))
2056 2077 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2057 2078 VDEV_AUX_CORRUPT_DATA);
2058 2079
2059 2080 /*
2060 2081 * If this is a leaf vdev, load its DTL.
2061 2082 */
2062 2083 if (vd->vdev_ops->vdev_op_leaf && vdev_dtl_load(vd) != 0)
2063 2084 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2064 2085 VDEV_AUX_CORRUPT_DATA);
2065 2086 }
2066 2087
2067 2088 /*
2068 2089 * The special vdev case is used for hot spares and l2cache devices. Its
2069 2090 * sole purpose it to set the vdev state for the associated vdev. To do this,
2070 2091 * we make sure that we can open the underlying device, then try to read the
2071 2092 * label, and make sure that the label is sane and that it hasn't been
2072 2093 * repurposed to another pool.
2073 2094 */
2074 2095 int
2075 2096 vdev_validate_aux(vdev_t *vd)
2076 2097 {
2077 2098 nvlist_t *label;
2078 2099 uint64_t guid, version;
2079 2100 uint64_t state;
2080 2101
2081 2102 if (!vdev_readable(vd))
2082 2103 return (0);
2083 2104
2084 2105 if ((label = vdev_label_read_config(vd, -1ULL)) == NULL) {
2085 2106 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2086 2107 VDEV_AUX_CORRUPT_DATA);
2087 2108 return (-1);
2088 2109 }
2089 2110
2090 2111 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
2091 2112 !SPA_VERSION_IS_SUPPORTED(version) ||
2092 2113 nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
2093 2114 guid != vd->vdev_guid ||
2094 2115 nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
2095 2116 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2096 2117 VDEV_AUX_CORRUPT_DATA);
2097 2118 nvlist_free(label);
2098 2119 return (-1);
2099 2120 }
2100 2121
2101 2122 /*
2102 2123 * We don't actually check the pool state here. If it's in fact in
2103 2124 * use by another pool, we update this fact on the fly when requested.
2104 2125 */
2105 2126 nvlist_free(label);
2106 2127 return (0);
2107 2128 }
2108 2129
2109 2130 void
2110 2131 vdev_remove(vdev_t *vd, uint64_t txg)
2111 2132 {
2112 2133 spa_t *spa = vd->vdev_spa;
2113 2134 objset_t *mos = spa->spa_meta_objset;
2114 2135 dmu_tx_t *tx;
2115 2136
2116 2137 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2117 2138
2118 2139 if (vd->vdev_ms != NULL) {
2119 2140 for (int m = 0; m < vd->vdev_ms_count; m++) {
2120 2141 metaslab_t *msp = vd->vdev_ms[m];
2121 2142
2122 2143 if (msp == NULL || msp->ms_sm == NULL)
2123 2144 continue;
2124 2145
2125 2146 mutex_enter(&msp->ms_lock);
2126 2147 VERIFY0(space_map_allocated(msp->ms_sm));
2127 2148 space_map_free(msp->ms_sm, tx);
2128 2149 space_map_close(msp->ms_sm);
2129 2150 msp->ms_sm = NULL;
2130 2151 mutex_exit(&msp->ms_lock);
2131 2152 }
2132 2153 }
2133 2154
2134 2155 if (vd->vdev_ms_array) {
2135 2156 (void) dmu_object_free(mos, vd->vdev_ms_array, tx);
2136 2157 vd->vdev_ms_array = 0;
2137 2158 }
2138 2159 dmu_tx_commit(tx);
2139 2160 }
2140 2161
2141 2162 void
2142 2163 vdev_sync_done(vdev_t *vd, uint64_t txg)
2143 2164 {
2144 2165 metaslab_t *msp;
2145 2166 boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg));
2146 2167
2147 2168 ASSERT(!vd->vdev_ishole);
2148 2169
2149 2170 while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
2150 2171 metaslab_sync_done(msp, txg);
2151 2172
2152 2173 if (reassess)
2153 2174 metaslab_sync_reassess(vd->vdev_mg);
2154 2175 }
2155 2176
2156 2177 void
2157 2178 vdev_sync(vdev_t *vd, uint64_t txg)
2158 2179 {
2159 2180 spa_t *spa = vd->vdev_spa;
2160 2181 vdev_t *lvd;
2161 2182 metaslab_t *msp;
2162 2183 dmu_tx_t *tx;
2163 2184
2164 2185 ASSERT(!vd->vdev_ishole);
2165 2186
2166 2187 if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0) {
2167 2188 ASSERT(vd == vd->vdev_top);
2168 2189 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2169 2190 vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
2170 2191 DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
2171 2192 ASSERT(vd->vdev_ms_array != 0);
2172 2193 vdev_config_dirty(vd);
2173 2194 dmu_tx_commit(tx);
2174 2195 }
2175 2196
2176 2197 /*
2177 2198 * Remove the metadata associated with this vdev once it's empty.
2178 2199 */
2179 2200 if (vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing)
2180 2201 vdev_remove(vd, txg);
2181 2202
2182 2203 while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
2183 2204 metaslab_sync(msp, txg);
2184 2205 (void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
2185 2206 }
2186 2207
2187 2208 while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
2188 2209 vdev_dtl_sync(lvd, txg);
2189 2210
2190 2211 (void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
2191 2212 }
2192 2213
2193 2214 uint64_t
2194 2215 vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
2195 2216 {
2196 2217 return (vd->vdev_ops->vdev_op_asize(vd, psize));
2197 2218 }
2198 2219
2199 2220 /*
2200 2221 * Mark the given vdev faulted. A faulted vdev behaves as if the device could
2201 2222 * not be opened, and no I/O is attempted.
2202 2223 */
2203 2224 int
2204 2225 vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2205 2226 {
2206 2227 vdev_t *vd, *tvd;
2207 2228
2208 2229 spa_vdev_state_enter(spa, SCL_NONE);
2209 2230
2210 2231 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2211 2232 return (spa_vdev_state_exit(spa, NULL, ENODEV));
2212 2233
2213 2234 if (!vd->vdev_ops->vdev_op_leaf)
2214 2235 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2215 2236
2216 2237 tvd = vd->vdev_top;
2217 2238
2218 2239 /*
2219 2240 * We don't directly use the aux state here, but if we do a
2220 2241 * vdev_reopen(), we need this value to be present to remember why we
2221 2242 * were faulted.
2222 2243 */
2223 2244 vd->vdev_label_aux = aux;
2224 2245
2225 2246 /*
2226 2247 * Faulted state takes precedence over degraded.
2227 2248 */
2228 2249 vd->vdev_delayed_close = B_FALSE;
2229 2250 vd->vdev_faulted = 1ULL;
2230 2251 vd->vdev_degraded = 0ULL;
2231 2252 vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
2232 2253
2233 2254 /*
2234 2255 * If this device has the only valid copy of the data, then
2235 2256 * back off and simply mark the vdev as degraded instead.
2236 2257 */
2237 2258 if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) {
2238 2259 vd->vdev_degraded = 1ULL;
2239 2260 vd->vdev_faulted = 0ULL;
2240 2261
2241 2262 /*
2242 2263 * If we reopen the device and it's not dead, only then do we
2243 2264 * mark it degraded.
2244 2265 */
2245 2266 vdev_reopen(tvd);
2246 2267
2247 2268 if (vdev_readable(vd))
2248 2269 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
2249 2270 }
2250 2271
2251 2272 return (spa_vdev_state_exit(spa, vd, 0));
2252 2273 }
2253 2274
2254 2275 /*
2255 2276 * Mark the given vdev degraded. A degraded vdev is purely an indication to the
2256 2277 * user that something is wrong. The vdev continues to operate as normal as far
2257 2278 * as I/O is concerned.
2258 2279 */
2259 2280 int
2260 2281 vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2261 2282 {
2262 2283 vdev_t *vd;
2263 2284
2264 2285 spa_vdev_state_enter(spa, SCL_NONE);
2265 2286
2266 2287 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2267 2288 return (spa_vdev_state_exit(spa, NULL, ENODEV));
2268 2289
2269 2290 if (!vd->vdev_ops->vdev_op_leaf)
2270 2291 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2271 2292
2272 2293 /*
2273 2294 * If the vdev is already faulted, then don't do anything.
2274 2295 */
2275 2296 if (vd->vdev_faulted || vd->vdev_degraded)
2276 2297 return (spa_vdev_state_exit(spa, NULL, 0));
2277 2298
2278 2299 vd->vdev_degraded = 1ULL;
2279 2300 if (!vdev_is_dead(vd))
2280 2301 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
2281 2302 aux);
2282 2303
2283 2304 return (spa_vdev_state_exit(spa, vd, 0));
2284 2305 }
2285 2306
2286 2307 /*
2287 2308 * Online the given vdev.
2288 2309 *
2289 2310 * If 'ZFS_ONLINE_UNSPARE' is set, it implies two things. First, any attached
2290 2311 * spare device should be detached when the device finishes resilvering.
2291 2312 * Second, the online should be treated like a 'test' online case, so no FMA
2292 2313 * events are generated if the device fails to open.
2293 2314 */
2294 2315 int
2295 2316 vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
2296 2317 {
2297 2318 vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
2298 2319
2299 2320 spa_vdev_state_enter(spa, SCL_NONE);
2300 2321
2301 2322 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2302 2323 return (spa_vdev_state_exit(spa, NULL, ENODEV));
2303 2324
2304 2325 if (!vd->vdev_ops->vdev_op_leaf)
2305 2326 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2306 2327
2307 2328 tvd = vd->vdev_top;
2308 2329 vd->vdev_offline = B_FALSE;
2309 2330 vd->vdev_tmpoffline = B_FALSE;
2310 2331 vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
2311 2332 vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
2312 2333
2313 2334 /* XXX - L2ARC 1.0 does not support expansion */
2314 2335 if (!vd->vdev_aux) {
2315 2336 for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2316 2337 pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
2317 2338 }
2318 2339
2319 2340 vdev_reopen(tvd);
2320 2341 vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
2321 2342
2322 2343 if (!vd->vdev_aux) {
2323 2344 for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2324 2345 pvd->vdev_expanding = B_FALSE;
2325 2346 }
2326 2347
2327 2348 if (newstate)
2328 2349 *newstate = vd->vdev_state;
2329 2350 if ((flags & ZFS_ONLINE_UNSPARE) &&
2330 2351 !vdev_is_dead(vd) && vd->vdev_parent &&
2331 2352 vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2332 2353 vd->vdev_parent->vdev_child[0] == vd)
2333 2354 vd->vdev_unspare = B_TRUE;
2334 2355
2335 2356 if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
2336 2357
2337 2358 /* XXX - L2ARC 1.0 does not support expansion */
2338 2359 if (vd->vdev_aux)
2339 2360 return (spa_vdev_state_exit(spa, vd, ENOTSUP));
2340 2361 spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
2341 2362 }
2342 2363 return (spa_vdev_state_exit(spa, vd, 0));
2343 2364 }
2344 2365
2345 2366 static int
2346 2367 vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
2347 2368 {
2348 2369 vdev_t *vd, *tvd;
2349 2370 int error = 0;
2350 2371 uint64_t generation;
2351 2372 metaslab_group_t *mg;
2352 2373
2353 2374 top:
2354 2375 spa_vdev_state_enter(spa, SCL_ALLOC);
2355 2376
2356 2377 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2357 2378 return (spa_vdev_state_exit(spa, NULL, ENODEV));
2358 2379
2359 2380 if (!vd->vdev_ops->vdev_op_leaf)
2360 2381 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2361 2382
2362 2383 tvd = vd->vdev_top;
2363 2384 mg = tvd->vdev_mg;
2364 2385 generation = spa->spa_config_generation + 1;
2365 2386
2366 2387 /*
2367 2388 * If the device isn't already offline, try to offline it.
2368 2389 */
2369 2390 if (!vd->vdev_offline) {
2370 2391 /*
2371 2392 * If this device has the only valid copy of some data,
2372 2393 * don't allow it to be offlined. Log devices are always
2373 2394 * expendable.
2374 2395 */
2375 2396 if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2376 2397 vdev_dtl_required(vd))
2377 2398 return (spa_vdev_state_exit(spa, NULL, EBUSY));
2378 2399
2379 2400 /*
2380 2401 * If the top-level is a slog and it has had allocations
2381 2402 * then proceed. We check that the vdev's metaslab group
2382 2403 * is not NULL since it's possible that we may have just
2383 2404 * added this vdev but not yet initialized its metaslabs.
2384 2405 */
2385 2406 if (tvd->vdev_islog && mg != NULL) {
2386 2407 /*
2387 2408 * Prevent any future allocations.
2388 2409 */
2389 2410 metaslab_group_passivate(mg);
2390 2411 (void) spa_vdev_state_exit(spa, vd, 0);
2391 2412
2392 2413 error = spa_offline_log(spa);
2393 2414
2394 2415 spa_vdev_state_enter(spa, SCL_ALLOC);
2395 2416
2396 2417 /*
2397 2418 * Check to see if the config has changed.
2398 2419 */
2399 2420 if (error || generation != spa->spa_config_generation) {
2400 2421 metaslab_group_activate(mg);
2401 2422 if (error)
2402 2423 return (spa_vdev_state_exit(spa,
2403 2424 vd, error));
2404 2425 (void) spa_vdev_state_exit(spa, vd, 0);
2405 2426 goto top;
2406 2427 }
2407 2428 ASSERT0(tvd->vdev_stat.vs_alloc);
2408 2429 }
2409 2430
2410 2431 /*
2411 2432 * Offline this device and reopen its top-level vdev.
2412 2433 * If the top-level vdev is a log device then just offline
2413 2434 * it. Otherwise, if this action results in the top-level
2414 2435 * vdev becoming unusable, undo it and fail the request.
2415 2436 */
2416 2437 vd->vdev_offline = B_TRUE;
2417 2438 vdev_reopen(tvd);
2418 2439
2419 2440 if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2420 2441 vdev_is_dead(tvd)) {
2421 2442 vd->vdev_offline = B_FALSE;
2422 2443 vdev_reopen(tvd);
2423 2444 return (spa_vdev_state_exit(spa, NULL, EBUSY));
2424 2445 }
2425 2446
2426 2447 /*
2427 2448 * Add the device back into the metaslab rotor so that
2428 2449 * once we online the device it's open for business.
2429 2450 */
2430 2451 if (tvd->vdev_islog && mg != NULL)
2431 2452 metaslab_group_activate(mg);
2432 2453 }
2433 2454
2434 2455 vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
2435 2456
2436 2457 return (spa_vdev_state_exit(spa, vd, 0));
2437 2458 }
2438 2459
2439 2460 int
2440 2461 vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
2441 2462 {
2442 2463 int error;
2443 2464
2444 2465 mutex_enter(&spa->spa_vdev_top_lock);
2445 2466 error = vdev_offline_locked(spa, guid, flags);
2446 2467 mutex_exit(&spa->spa_vdev_top_lock);
2447 2468
2448 2469 return (error);
2449 2470 }
2450 2471
2451 2472 /*
2452 2473 * Clear the error counts associated with this vdev. Unlike vdev_online() and
2453 2474 * vdev_offline(), we assume the spa config is locked. We also clear all
2454 2475 * children. If 'vd' is NULL, then the user wants to clear all vdevs.
2455 2476 */
2456 2477 void
2457 2478 vdev_clear(spa_t *spa, vdev_t *vd)
2458 2479 {
2459 2480 vdev_t *rvd = spa->spa_root_vdev;
2460 2481
2461 2482 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2462 2483
2463 2484 if (vd == NULL)
2464 2485 vd = rvd;
2465 2486
2466 2487 vd->vdev_stat.vs_read_errors = 0;
2467 2488 vd->vdev_stat.vs_write_errors = 0;
2468 2489 vd->vdev_stat.vs_checksum_errors = 0;
2469 2490
2470 2491 for (int c = 0; c < vd->vdev_children; c++)
2471 2492 vdev_clear(spa, vd->vdev_child[c]);
2472 2493
2473 2494 /*
2474 2495 * If we're in the FAULTED state or have experienced failed I/O, then
2475 2496 * clear the persistent state and attempt to reopen the device. We
2476 2497 * also mark the vdev config dirty, so that the new faulted state is
2477 2498 * written out to disk.
2478 2499 */
2479 2500 if (vd->vdev_faulted || vd->vdev_degraded ||
2480 2501 !vdev_readable(vd) || !vdev_writeable(vd)) {
2481 2502
2482 2503 /*
2483 2504 * When reopening in reponse to a clear event, it may be due to
2484 2505 * a fmadm repair request. In this case, if the device is
2485 2506 * still broken, we want to still post the ereport again.
2486 2507 */
2487 2508 vd->vdev_forcefault = B_TRUE;
2488 2509
2489 2510 vd->vdev_faulted = vd->vdev_degraded = 0ULL;
2490 2511 vd->vdev_cant_read = B_FALSE;
2491 2512 vd->vdev_cant_write = B_FALSE;
2492 2513
2493 2514 vdev_reopen(vd == rvd ? rvd : vd->vdev_top);
2494 2515
2495 2516 vd->vdev_forcefault = B_FALSE;
2496 2517
2497 2518 if (vd != rvd && vdev_writeable(vd->vdev_top))
2498 2519 vdev_state_dirty(vd->vdev_top);
2499 2520
2500 2521 if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
2501 2522 spa_async_request(spa, SPA_ASYNC_RESILVER);
2502 2523
2503 2524 spa_event_notify(spa, vd, ESC_ZFS_VDEV_CLEAR);
2504 2525 }
2505 2526
2506 2527 /*
2507 2528 * When clearing a FMA-diagnosed fault, we always want to
2508 2529 * unspare the device, as we assume that the original spare was
2509 2530 * done in response to the FMA fault.
2510 2531 */
2511 2532 if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
2512 2533 vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2513 2534 vd->vdev_parent->vdev_child[0] == vd)
2514 2535 vd->vdev_unspare = B_TRUE;
2515 2536 }
2516 2537
2517 2538 boolean_t
2518 2539 vdev_is_dead(vdev_t *vd)
2519 2540 {
2520 2541 /*
2521 2542 * Holes and missing devices are always considered "dead".
2522 2543 * This simplifies the code since we don't have to check for
2523 2544 * these types of devices in the various code paths.
2524 2545 * Instead we rely on the fact that we skip over dead devices
2525 2546 * before issuing I/O to them.
2526 2547 */
2527 2548 return (vd->vdev_state < VDEV_STATE_DEGRADED || vd->vdev_ishole ||
2528 2549 vd->vdev_ops == &vdev_missing_ops);
2529 2550 }
2530 2551
2531 2552 boolean_t
2532 2553 vdev_readable(vdev_t *vd)
2533 2554 {
2534 2555 return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
2535 2556 }
2536 2557
2537 2558 boolean_t
2538 2559 vdev_writeable(vdev_t *vd)
2539 2560 {
2540 2561 return (!vdev_is_dead(vd) && !vd->vdev_cant_write);
2541 2562 }
2542 2563
2543 2564 boolean_t
2544 2565 vdev_allocatable(vdev_t *vd)
2545 2566 {
2546 2567 uint64_t state = vd->vdev_state;
2547 2568
2548 2569 /*
2549 2570 * We currently allow allocations from vdevs which may be in the
2550 2571 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
2551 2572 * fails to reopen then we'll catch it later when we're holding
2552 2573 * the proper locks. Note that we have to get the vdev state
2553 2574 * in a local variable because although it changes atomically,
2554 2575 * we're asking two separate questions about it.
2555 2576 */
2556 2577 return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
2557 2578 !vd->vdev_cant_write && !vd->vdev_ishole);
2558 2579 }
2559 2580
2560 2581 boolean_t
2561 2582 vdev_accessible(vdev_t *vd, zio_t *zio)
2562 2583 {
2563 2584 ASSERT(zio->io_vd == vd);
2564 2585
2565 2586 if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
2566 2587 return (B_FALSE);
2567 2588
2568 2589 if (zio->io_type == ZIO_TYPE_READ)
2569 2590 return (!vd->vdev_cant_read);
2570 2591
2571 2592 if (zio->io_type == ZIO_TYPE_WRITE)
2572 2593 return (!vd->vdev_cant_write);
2573 2594
2574 2595 return (B_TRUE);
2575 2596 }
2576 2597
2577 2598 /*
2578 2599 * Get statistics for the given vdev.
2579 2600 */
2580 2601 void
2581 2602 vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
2582 2603 {
2583 2604 vdev_t *rvd = vd->vdev_spa->spa_root_vdev;
2584 2605
2585 2606 mutex_enter(&vd->vdev_stat_lock);
2586 2607 bcopy(&vd->vdev_stat, vs, sizeof (*vs));
2587 2608 vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
2588 2609 vs->vs_state = vd->vdev_state;
2589 2610 vs->vs_rsize = vdev_get_min_asize(vd);
2590 2611 if (vd->vdev_ops->vdev_op_leaf)
2591 2612 vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
2592 2613 vs->vs_esize = vd->vdev_max_asize - vd->vdev_asize;
2593 2614 mutex_exit(&vd->vdev_stat_lock);
2594 2615
2595 2616 /*
2596 2617 * If we're getting stats on the root vdev, aggregate the I/O counts
2597 2618 * over all top-level vdevs (i.e. the direct children of the root).
2598 2619 */
2599 2620 if (vd == rvd) {
2600 2621 for (int c = 0; c < rvd->vdev_children; c++) {
2601 2622 vdev_t *cvd = rvd->vdev_child[c];
2602 2623 vdev_stat_t *cvs = &cvd->vdev_stat;
2603 2624
2604 2625 mutex_enter(&vd->vdev_stat_lock);
2605 2626 for (int t = 0; t < ZIO_TYPES; t++) {
2606 2627 vs->vs_ops[t] += cvs->vs_ops[t];
2607 2628 vs->vs_bytes[t] += cvs->vs_bytes[t];
2608 2629 }
2609 2630 cvs->vs_scan_removing = cvd->vdev_removing;
2610 2631 mutex_exit(&vd->vdev_stat_lock);
2611 2632 }
2612 2633 }
2613 2634 }
2614 2635
2615 2636 void
2616 2637 vdev_clear_stats(vdev_t *vd)
2617 2638 {
2618 2639 mutex_enter(&vd->vdev_stat_lock);
2619 2640 vd->vdev_stat.vs_space = 0;
2620 2641 vd->vdev_stat.vs_dspace = 0;
2621 2642 vd->vdev_stat.vs_alloc = 0;
2622 2643 mutex_exit(&vd->vdev_stat_lock);
2623 2644 }
2624 2645
2625 2646 void
2626 2647 vdev_scan_stat_init(vdev_t *vd)
2627 2648 {
2628 2649 vdev_stat_t *vs = &vd->vdev_stat;
2629 2650
2630 2651 for (int c = 0; c < vd->vdev_children; c++)
2631 2652 vdev_scan_stat_init(vd->vdev_child[c]);
2632 2653
2633 2654 mutex_enter(&vd->vdev_stat_lock);
2634 2655 vs->vs_scan_processed = 0;
2635 2656 mutex_exit(&vd->vdev_stat_lock);
2636 2657 }
2637 2658
2638 2659 void
2639 2660 vdev_stat_update(zio_t *zio, uint64_t psize)
2640 2661 {
2641 2662 spa_t *spa = zio->io_spa;
2642 2663 vdev_t *rvd = spa->spa_root_vdev;
2643 2664 vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
2644 2665 vdev_t *pvd;
2645 2666 uint64_t txg = zio->io_txg;
2646 2667 vdev_stat_t *vs = &vd->vdev_stat;
2647 2668 zio_type_t type = zio->io_type;
2648 2669 int flags = zio->io_flags;
2649 2670
2650 2671 /*
2651 2672 * If this i/o is a gang leader, it didn't do any actual work.
2652 2673 */
2653 2674 if (zio->io_gang_tree)
2654 2675 return;
2655 2676
2656 2677 if (zio->io_error == 0) {
2657 2678 /*
2658 2679 * If this is a root i/o, don't count it -- we've already
2659 2680 * counted the top-level vdevs, and vdev_get_stats() will
2660 2681 * aggregate them when asked. This reduces contention on
2661 2682 * the root vdev_stat_lock and implicitly handles blocks
2662 2683 * that compress away to holes, for which there is no i/o.
2663 2684 * (Holes never create vdev children, so all the counters
2664 2685 * remain zero, which is what we want.)
2665 2686 *
2666 2687 * Note: this only applies to successful i/o (io_error == 0)
2667 2688 * because unlike i/o counts, errors are not additive.
2668 2689 * When reading a ditto block, for example, failure of
2669 2690 * one top-level vdev does not imply a root-level error.
2670 2691 */
2671 2692 if (vd == rvd)
2672 2693 return;
2673 2694
2674 2695 ASSERT(vd == zio->io_vd);
2675 2696
2676 2697 if (flags & ZIO_FLAG_IO_BYPASS)
2677 2698 return;
2678 2699
2679 2700 mutex_enter(&vd->vdev_stat_lock);
2680 2701
2681 2702 if (flags & ZIO_FLAG_IO_REPAIR) {
2682 2703 if (flags & ZIO_FLAG_SCAN_THREAD) {
2683 2704 dsl_scan_phys_t *scn_phys =
2684 2705 &spa->spa_dsl_pool->dp_scan->scn_phys;
2685 2706 uint64_t *processed = &scn_phys->scn_processed;
2686 2707
2687 2708 /* XXX cleanup? */
2688 2709 if (vd->vdev_ops->vdev_op_leaf)
2689 2710 atomic_add_64(processed, psize);
2690 2711 vs->vs_scan_processed += psize;
2691 2712 }
2692 2713
2693 2714 if (flags & ZIO_FLAG_SELF_HEAL)
2694 2715 vs->vs_self_healed += psize;
2695 2716 }
2696 2717
2697 2718 vs->vs_ops[type]++;
2698 2719 vs->vs_bytes[type] += psize;
2699 2720
2700 2721 mutex_exit(&vd->vdev_stat_lock);
2701 2722 return;
2702 2723 }
2703 2724
2704 2725 if (flags & ZIO_FLAG_SPECULATIVE)
2705 2726 return;
2706 2727
2707 2728 /*
2708 2729 * If this is an I/O error that is going to be retried, then ignore the
2709 2730 * error. Otherwise, the user may interpret B_FAILFAST I/O errors as
2710 2731 * hard errors, when in reality they can happen for any number of
2711 2732 * innocuous reasons (bus resets, MPxIO link failure, etc).
2712 2733 */
2713 2734 if (zio->io_error == EIO &&
2714 2735 !(zio->io_flags & ZIO_FLAG_IO_RETRY))
2715 2736 return;
2716 2737
2717 2738 /*
2718 2739 * Intent logs writes won't propagate their error to the root
2719 2740 * I/O so don't mark these types of failures as pool-level
2720 2741 * errors.
2721 2742 */
2722 2743 if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
2723 2744 return;
2724 2745
2725 2746 mutex_enter(&vd->vdev_stat_lock);
2726 2747 if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
2727 2748 if (zio->io_error == ECKSUM)
2728 2749 vs->vs_checksum_errors++;
2729 2750 else
2730 2751 vs->vs_read_errors++;
2731 2752 }
2732 2753 if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
2733 2754 vs->vs_write_errors++;
2734 2755 mutex_exit(&vd->vdev_stat_lock);
2735 2756
2736 2757 if (type == ZIO_TYPE_WRITE && txg != 0 &&
2737 2758 (!(flags & ZIO_FLAG_IO_REPAIR) ||
2738 2759 (flags & ZIO_FLAG_SCAN_THREAD) ||
2739 2760 spa->spa_claiming)) {
2740 2761 /*
2741 2762 * This is either a normal write (not a repair), or it's
2742 2763 * a repair induced by the scrub thread, or it's a repair
2743 2764 * made by zil_claim() during spa_load() in the first txg.
2744 2765 * In the normal case, we commit the DTL change in the same
2745 2766 * txg as the block was born. In the scrub-induced repair
2746 2767 * case, we know that scrubs run in first-pass syncing context,
2747 2768 * so we commit the DTL change in spa_syncing_txg(spa).
2748 2769 * In the zil_claim() case, we commit in spa_first_txg(spa).
2749 2770 *
2750 2771 * We currently do not make DTL entries for failed spontaneous
2751 2772 * self-healing writes triggered by normal (non-scrubbing)
2752 2773 * reads, because we have no transactional context in which to
2753 2774 * do so -- and it's not clear that it'd be desirable anyway.
2754 2775 */
2755 2776 if (vd->vdev_ops->vdev_op_leaf) {
2756 2777 uint64_t commit_txg = txg;
2757 2778 if (flags & ZIO_FLAG_SCAN_THREAD) {
2758 2779 ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2759 2780 ASSERT(spa_sync_pass(spa) == 1);
2760 2781 vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
2761 2782 commit_txg = spa_syncing_txg(spa);
2762 2783 } else if (spa->spa_claiming) {
2763 2784 ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2764 2785 commit_txg = spa_first_txg(spa);
2765 2786 }
2766 2787 ASSERT(commit_txg >= spa_syncing_txg(spa));
2767 2788 if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
2768 2789 return;
2769 2790 for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2770 2791 vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
2771 2792 vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
2772 2793 }
2773 2794 if (vd != rvd)
2774 2795 vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
2775 2796 }
2776 2797 }
2777 2798
2778 2799 /*
2779 2800 * Update the in-core space usage stats for this vdev, its metaslab class,
2780 2801 * and the root vdev.
2781 2802 */
2782 2803 void
2783 2804 vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
2784 2805 int64_t space_delta)
2785 2806 {
2786 2807 int64_t dspace_delta = space_delta;
2787 2808 spa_t *spa = vd->vdev_spa;
2788 2809 vdev_t *rvd = spa->spa_root_vdev;
2789 2810 metaslab_group_t *mg = vd->vdev_mg;
2790 2811 metaslab_class_t *mc = mg ? mg->mg_class : NULL;
2791 2812
2792 2813 ASSERT(vd == vd->vdev_top);
2793 2814
2794 2815 /*
2795 2816 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
2796 2817 * factor. We must calculate this here and not at the root vdev
2797 2818 * because the root vdev's psize-to-asize is simply the max of its
2798 2819 * childrens', thus not accurate enough for us.
2799 2820 */
2800 2821 ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
2801 2822 ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
2802 2823 dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
2803 2824 vd->vdev_deflate_ratio;
2804 2825
2805 2826 mutex_enter(&vd->vdev_stat_lock);
2806 2827 vd->vdev_stat.vs_alloc += alloc_delta;
2807 2828 vd->vdev_stat.vs_space += space_delta;
2808 2829 vd->vdev_stat.vs_dspace += dspace_delta;
2809 2830 mutex_exit(&vd->vdev_stat_lock);
2810 2831
2811 2832 if (mc == spa_normal_class(spa)) {
2812 2833 mutex_enter(&rvd->vdev_stat_lock);
2813 2834 rvd->vdev_stat.vs_alloc += alloc_delta;
2814 2835 rvd->vdev_stat.vs_space += space_delta;
2815 2836 rvd->vdev_stat.vs_dspace += dspace_delta;
2816 2837 mutex_exit(&rvd->vdev_stat_lock);
2817 2838 }
2818 2839
2819 2840 if (mc != NULL) {
2820 2841 ASSERT(rvd == vd->vdev_parent);
2821 2842 ASSERT(vd->vdev_ms_count != 0);
2822 2843
2823 2844 metaslab_class_space_update(mc,
2824 2845 alloc_delta, defer_delta, space_delta, dspace_delta);
2825 2846 }
2826 2847 }
2827 2848
2828 2849 /*
2829 2850 * Mark a top-level vdev's config as dirty, placing it on the dirty list
2830 2851 * so that it will be written out next time the vdev configuration is synced.
2831 2852 * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
2832 2853 */
2833 2854 void
2834 2855 vdev_config_dirty(vdev_t *vd)
2835 2856 {
2836 2857 spa_t *spa = vd->vdev_spa;
2837 2858 vdev_t *rvd = spa->spa_root_vdev;
2838 2859 int c;
2839 2860
2840 2861 ASSERT(spa_writeable(spa));
2841 2862
2842 2863 /*
2843 2864 * If this is an aux vdev (as with l2cache and spare devices), then we
2844 2865 * update the vdev config manually and set the sync flag.
2845 2866 */
2846 2867 if (vd->vdev_aux != NULL) {
2847 2868 spa_aux_vdev_t *sav = vd->vdev_aux;
2848 2869 nvlist_t **aux;
2849 2870 uint_t naux;
2850 2871
2851 2872 for (c = 0; c < sav->sav_count; c++) {
2852 2873 if (sav->sav_vdevs[c] == vd)
2853 2874 break;
2854 2875 }
2855 2876
2856 2877 if (c == sav->sav_count) {
2857 2878 /*
2858 2879 * We're being removed. There's nothing more to do.
2859 2880 */
2860 2881 ASSERT(sav->sav_sync == B_TRUE);
2861 2882 return;
2862 2883 }
2863 2884
2864 2885 sav->sav_sync = B_TRUE;
2865 2886
2866 2887 if (nvlist_lookup_nvlist_array(sav->sav_config,
2867 2888 ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
2868 2889 VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
2869 2890 ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
2870 2891 }
2871 2892
2872 2893 ASSERT(c < naux);
2873 2894
2874 2895 /*
2875 2896 * Setting the nvlist in the middle if the array is a little
2876 2897 * sketchy, but it will work.
2877 2898 */
2878 2899 nvlist_free(aux[c]);
2879 2900 aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0);
2880 2901
2881 2902 return;
2882 2903 }
2883 2904
2884 2905 /*
2885 2906 * The dirty list is protected by the SCL_CONFIG lock. The caller
2886 2907 * must either hold SCL_CONFIG as writer, or must be the sync thread
2887 2908 * (which holds SCL_CONFIG as reader). There's only one sync thread,
2888 2909 * so this is sufficient to ensure mutual exclusion.
2889 2910 */
2890 2911 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2891 2912 (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2892 2913 spa_config_held(spa, SCL_CONFIG, RW_READER)));
2893 2914
2894 2915 if (vd == rvd) {
2895 2916 for (c = 0; c < rvd->vdev_children; c++)
2896 2917 vdev_config_dirty(rvd->vdev_child[c]);
2897 2918 } else {
2898 2919 ASSERT(vd == vd->vdev_top);
2899 2920
2900 2921 if (!list_link_active(&vd->vdev_config_dirty_node) &&
2901 2922 !vd->vdev_ishole)
2902 2923 list_insert_head(&spa->spa_config_dirty_list, vd);
2903 2924 }
2904 2925 }
2905 2926
2906 2927 void
2907 2928 vdev_config_clean(vdev_t *vd)
2908 2929 {
2909 2930 spa_t *spa = vd->vdev_spa;
2910 2931
2911 2932 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2912 2933 (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2913 2934 spa_config_held(spa, SCL_CONFIG, RW_READER)));
2914 2935
2915 2936 ASSERT(list_link_active(&vd->vdev_config_dirty_node));
2916 2937 list_remove(&spa->spa_config_dirty_list, vd);
2917 2938 }
2918 2939
2919 2940 /*
2920 2941 * Mark a top-level vdev's state as dirty, so that the next pass of
2921 2942 * spa_sync() can convert this into vdev_config_dirty(). We distinguish
2922 2943 * the state changes from larger config changes because they require
2923 2944 * much less locking, and are often needed for administrative actions.
2924 2945 */
2925 2946 void
2926 2947 vdev_state_dirty(vdev_t *vd)
2927 2948 {
2928 2949 spa_t *spa = vd->vdev_spa;
2929 2950
2930 2951 ASSERT(spa_writeable(spa));
2931 2952 ASSERT(vd == vd->vdev_top);
2932 2953
2933 2954 /*
2934 2955 * The state list is protected by the SCL_STATE lock. The caller
2935 2956 * must either hold SCL_STATE as writer, or must be the sync thread
2936 2957 * (which holds SCL_STATE as reader). There's only one sync thread,
2937 2958 * so this is sufficient to ensure mutual exclusion.
2938 2959 */
2939 2960 ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2940 2961 (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2941 2962 spa_config_held(spa, SCL_STATE, RW_READER)));
2942 2963
2943 2964 if (!list_link_active(&vd->vdev_state_dirty_node) && !vd->vdev_ishole)
2944 2965 list_insert_head(&spa->spa_state_dirty_list, vd);
2945 2966 }
2946 2967
2947 2968 void
2948 2969 vdev_state_clean(vdev_t *vd)
2949 2970 {
2950 2971 spa_t *spa = vd->vdev_spa;
2951 2972
2952 2973 ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2953 2974 (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2954 2975 spa_config_held(spa, SCL_STATE, RW_READER)));
2955 2976
2956 2977 ASSERT(list_link_active(&vd->vdev_state_dirty_node));
2957 2978 list_remove(&spa->spa_state_dirty_list, vd);
2958 2979 }
2959 2980
2960 2981 /*
2961 2982 * Propagate vdev state up from children to parent.
2962 2983 */
2963 2984 void
2964 2985 vdev_propagate_state(vdev_t *vd)
2965 2986 {
2966 2987 spa_t *spa = vd->vdev_spa;
2967 2988 vdev_t *rvd = spa->spa_root_vdev;
2968 2989 int degraded = 0, faulted = 0;
2969 2990 int corrupted = 0;
2970 2991 vdev_t *child;
2971 2992
2972 2993 if (vd->vdev_children > 0) {
2973 2994 for (int c = 0; c < vd->vdev_children; c++) {
2974 2995 child = vd->vdev_child[c];
2975 2996
2976 2997 /*
2977 2998 * Don't factor holes into the decision.
2978 2999 */
2979 3000 if (child->vdev_ishole)
2980 3001 continue;
2981 3002
2982 3003 if (!vdev_readable(child) ||
2983 3004 (!vdev_writeable(child) && spa_writeable(spa))) {
2984 3005 /*
2985 3006 * Root special: if there is a top-level log
2986 3007 * device, treat the root vdev as if it were
2987 3008 * degraded.
2988 3009 */
2989 3010 if (child->vdev_islog && vd == rvd)
2990 3011 degraded++;
2991 3012 else
2992 3013 faulted++;
2993 3014 } else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
2994 3015 degraded++;
2995 3016 }
2996 3017
2997 3018 if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
2998 3019 corrupted++;
2999 3020 }
3000 3021
3001 3022 vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
3002 3023
3003 3024 /*
3004 3025 * Root special: if there is a top-level vdev that cannot be
3005 3026 * opened due to corrupted metadata, then propagate the root
3006 3027 * vdev's aux state as 'corrupt' rather than 'insufficient
3007 3028 * replicas'.
3008 3029 */
3009 3030 if (corrupted && vd == rvd &&
3010 3031 rvd->vdev_state == VDEV_STATE_CANT_OPEN)
3011 3032 vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
3012 3033 VDEV_AUX_CORRUPT_DATA);
3013 3034 }
3014 3035
3015 3036 if (vd->vdev_parent)
3016 3037 vdev_propagate_state(vd->vdev_parent);
3017 3038 }
3018 3039
3019 3040 /*
3020 3041 * Set a vdev's state. If this is during an open, we don't update the parent
3021 3042 * state, because we're in the process of opening children depth-first.
3022 3043 * Otherwise, we propagate the change to the parent.
3023 3044 *
3024 3045 * If this routine places a device in a faulted state, an appropriate ereport is
3025 3046 * generated.
3026 3047 */
3027 3048 void
3028 3049 vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
3029 3050 {
3030 3051 uint64_t save_state;
3031 3052 spa_t *spa = vd->vdev_spa;
3032 3053
3033 3054 if (state == vd->vdev_state) {
3034 3055 vd->vdev_stat.vs_aux = aux;
3035 3056 return;
3036 3057 }
3037 3058
3038 3059 save_state = vd->vdev_state;
3039 3060
3040 3061 vd->vdev_state = state;
3041 3062 vd->vdev_stat.vs_aux = aux;
3042 3063
3043 3064 /*
3044 3065 * If we are setting the vdev state to anything but an open state, then
3045 3066 * always close the underlying device unless the device has requested
3046 3067 * a delayed close (i.e. we're about to remove or fault the device).
3047 3068 * Otherwise, we keep accessible but invalid devices open forever.
3048 3069 * We don't call vdev_close() itself, because that implies some extra
3049 3070 * checks (offline, etc) that we don't want here. This is limited to
3050 3071 * leaf devices, because otherwise closing the device will affect other
3051 3072 * children.
3052 3073 */
3053 3074 if (!vd->vdev_delayed_close && vdev_is_dead(vd) &&
3054 3075 vd->vdev_ops->vdev_op_leaf)
3055 3076 vd->vdev_ops->vdev_op_close(vd);
3056 3077
3057 3078 /*
3058 3079 * If we have brought this vdev back into service, we need
3059 3080 * to notify fmd so that it can gracefully repair any outstanding
3060 3081 * cases due to a missing device. We do this in all cases, even those
3061 3082 * that probably don't correlate to a repaired fault. This is sure to
3062 3083 * catch all cases, and we let the zfs-retire agent sort it out. If
3063 3084 * this is a transient state it's OK, as the retire agent will
3064 3085 * double-check the state of the vdev before repairing it.
3065 3086 */
3066 3087 if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf &&
3067 3088 vd->vdev_prevstate != state)
3068 3089 zfs_post_state_change(spa, vd);
3069 3090
3070 3091 if (vd->vdev_removed &&
3071 3092 state == VDEV_STATE_CANT_OPEN &&
3072 3093 (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
3073 3094 /*
3074 3095 * If the previous state is set to VDEV_STATE_REMOVED, then this
3075 3096 * device was previously marked removed and someone attempted to
3076 3097 * reopen it. If this failed due to a nonexistent device, then
3077 3098 * keep the device in the REMOVED state. We also let this be if
3078 3099 * it is one of our special test online cases, which is only
3079 3100 * attempting to online the device and shouldn't generate an FMA
3080 3101 * fault.
3081 3102 */
3082 3103 vd->vdev_state = VDEV_STATE_REMOVED;
3083 3104 vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
3084 3105 } else if (state == VDEV_STATE_REMOVED) {
3085 3106 vd->vdev_removed = B_TRUE;
3086 3107 } else if (state == VDEV_STATE_CANT_OPEN) {
3087 3108 /*
3088 3109 * If we fail to open a vdev during an import or recovery, we
3089 3110 * mark it as "not available", which signifies that it was
3090 3111 * never there to begin with. Failure to open such a device
3091 3112 * is not considered an error.
3092 3113 */
3093 3114 if ((spa_load_state(spa) == SPA_LOAD_IMPORT ||
3094 3115 spa_load_state(spa) == SPA_LOAD_RECOVER) &&
3095 3116 vd->vdev_ops->vdev_op_leaf)
3096 3117 vd->vdev_not_present = 1;
3097 3118
3098 3119 /*
3099 3120 * Post the appropriate ereport. If the 'prevstate' field is
3100 3121 * set to something other than VDEV_STATE_UNKNOWN, it indicates
3101 3122 * that this is part of a vdev_reopen(). In this case, we don't
3102 3123 * want to post the ereport if the device was already in the
3103 3124 * CANT_OPEN state beforehand.
3104 3125 *
3105 3126 * If the 'checkremove' flag is set, then this is an attempt to
3106 3127 * online the device in response to an insertion event. If we
3107 3128 * hit this case, then we have detected an insertion event for a
3108 3129 * faulted or offline device that wasn't in the removed state.
3109 3130 * In this scenario, we don't post an ereport because we are
3110 3131 * about to replace the device, or attempt an online with
3111 3132 * vdev_forcefault, which will generate the fault for us.
3112 3133 */
3113 3134 if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
3114 3135 !vd->vdev_not_present && !vd->vdev_checkremove &&
3115 3136 vd != spa->spa_root_vdev) {
3116 3137 const char *class;
3117 3138
3118 3139 switch (aux) {
3119 3140 case VDEV_AUX_OPEN_FAILED:
3120 3141 class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
3121 3142 break;
3122 3143 case VDEV_AUX_CORRUPT_DATA:
3123 3144 class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
3124 3145 break;
3125 3146 case VDEV_AUX_NO_REPLICAS:
3126 3147 class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
3127 3148 break;
3128 3149 case VDEV_AUX_BAD_GUID_SUM:
3129 3150 class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
3130 3151 break;
3131 3152 case VDEV_AUX_TOO_SMALL:
3132 3153 class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
3133 3154 break;
3134 3155 case VDEV_AUX_BAD_LABEL:
3135 3156 class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
3136 3157 break;
3137 3158 default:
3138 3159 class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
3139 3160 }
3140 3161
3141 3162 zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
3142 3163 }
3143 3164
3144 3165 /* Erase any notion of persistent removed state */
3145 3166 vd->vdev_removed = B_FALSE;
3146 3167 } else {
3147 3168 vd->vdev_removed = B_FALSE;
3148 3169 }
3149 3170
3150 3171 if (!isopen && vd->vdev_parent)
3151 3172 vdev_propagate_state(vd->vdev_parent);
3152 3173 }
3153 3174
3154 3175 /*
3155 3176 * Check the vdev configuration to ensure that it's capable of supporting
3156 3177 * a root pool. Currently, we do not support RAID-Z or partial configuration.
3157 3178 * In addition, only a single top-level vdev is allowed and none of the leaves
3158 3179 * can be wholedisks.
3159 3180 */
3160 3181 boolean_t
3161 3182 vdev_is_bootable(vdev_t *vd)
3162 3183 {
3163 3184 if (!vd->vdev_ops->vdev_op_leaf) {
3164 3185 char *vdev_type = vd->vdev_ops->vdev_op_type;
3165 3186
3166 3187 if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
3167 3188 vd->vdev_children > 1) {
3168 3189 return (B_FALSE);
3169 3190 } else if (strcmp(vdev_type, VDEV_TYPE_RAIDZ) == 0 ||
3170 3191 strcmp(vdev_type, VDEV_TYPE_MISSING) == 0) {
3171 3192 return (B_FALSE);
3172 3193 }
3173 3194 } else if (vd->vdev_wholedisk == 1) {
3174 3195 return (B_FALSE);
3175 3196 }
3176 3197
3177 3198 for (int c = 0; c < vd->vdev_children; c++) {
3178 3199 if (!vdev_is_bootable(vd->vdev_child[c]))
3179 3200 return (B_FALSE);
3180 3201 }
3181 3202 return (B_TRUE);
3182 3203 }
3183 3204
3184 3205 /*
3185 3206 * Load the state from the original vdev tree (ovd) which
3186 3207 * we've retrieved from the MOS config object. If the original
3187 3208 * vdev was offline or faulted then we transfer that state to the
3188 3209 * device in the current vdev tree (nvd).
3189 3210 */
3190 3211 void
3191 3212 vdev_load_log_state(vdev_t *nvd, vdev_t *ovd)
3192 3213 {
3193 3214 spa_t *spa = nvd->vdev_spa;
3194 3215
3195 3216 ASSERT(nvd->vdev_top->vdev_islog);
3196 3217 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3197 3218 ASSERT3U(nvd->vdev_guid, ==, ovd->vdev_guid);
3198 3219
3199 3220 for (int c = 0; c < nvd->vdev_children; c++)
3200 3221 vdev_load_log_state(nvd->vdev_child[c], ovd->vdev_child[c]);
3201 3222
3202 3223 if (nvd->vdev_ops->vdev_op_leaf) {
3203 3224 /*
3204 3225 * Restore the persistent vdev state
3205 3226 */
3206 3227 nvd->vdev_offline = ovd->vdev_offline;
3207 3228 nvd->vdev_faulted = ovd->vdev_faulted;
3208 3229 nvd->vdev_degraded = ovd->vdev_degraded;
3209 3230 nvd->vdev_removed = ovd->vdev_removed;
3210 3231 }
3211 3232 }
3212 3233
3213 3234 /*
3214 3235 * Determine if a log device has valid content. If the vdev was
3215 3236 * removed or faulted in the MOS config then we know that
3216 3237 * the content on the log device has already been written to the pool.
3217 3238 */
3218 3239 boolean_t
3219 3240 vdev_log_state_valid(vdev_t *vd)
3220 3241 {
3221 3242 if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted &&
3222 3243 !vd->vdev_removed)
3223 3244 return (B_TRUE);
3224 3245
3225 3246 for (int c = 0; c < vd->vdev_children; c++)
3226 3247 if (vdev_log_state_valid(vd->vdev_child[c]))
3227 3248 return (B_TRUE);
3228 3249
3229 3250 return (B_FALSE);
3230 3251 }
3231 3252
3232 3253 /*
3233 3254 * Expand a vdev if possible.
3234 3255 */
3235 3256 void
3236 3257 vdev_expand(vdev_t *vd, uint64_t txg)
3237 3258 {
3238 3259 ASSERT(vd->vdev_top == vd);
3239 3260 ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
3240 3261
3241 3262 if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count) {
3242 3263 VERIFY(vdev_metaslab_init(vd, txg) == 0);
3243 3264 vdev_config_dirty(vd);
3244 3265 }
3245 3266 }
3246 3267
3247 3268 /*
3248 3269 * Split a vdev.
3249 3270 */
3250 3271 void
3251 3272 vdev_split(vdev_t *vd)
3252 3273 {
3253 3274 vdev_t *cvd, *pvd = vd->vdev_parent;
3254 3275
3255 3276 vdev_remove_child(pvd, vd);
3256 3277 vdev_compact_children(pvd);
3257 3278
3258 3279 cvd = pvd->vdev_child[0];
3259 3280 if (pvd->vdev_children == 1) {
3260 3281 vdev_remove_parent(cvd);
3261 3282 cvd->vdev_splitting = B_TRUE;
3262 3283 }
3263 3284 vdev_propagate_state(cvd);
3264 3285 }
3265 3286
3266 3287 void
3267 3288 vdev_deadman(vdev_t *vd)
3268 3289 {
3269 3290 for (int c = 0; c < vd->vdev_children; c++) {
3270 3291 vdev_t *cvd = vd->vdev_child[c];
3271 3292
3272 3293 vdev_deadman(cvd);
3273 3294 }
3274 3295
3275 3296 if (vd->vdev_ops->vdev_op_leaf) {
3276 3297 vdev_queue_t *vq = &vd->vdev_queue;
3277 3298
3278 3299 mutex_enter(&vq->vq_lock);
3279 3300 if (avl_numnodes(&vq->vq_active_tree) > 0) {
3280 3301 spa_t *spa = vd->vdev_spa;
3281 3302 zio_t *fio;
3282 3303 uint64_t delta;
3283 3304
3284 3305 /*
3285 3306 * Look at the head of all the pending queues,
3286 3307 * if any I/O has been outstanding for longer than
3287 3308 * the spa_deadman_synctime we panic the system.
3288 3309 */
3289 3310 fio = avl_first(&vq->vq_active_tree);
3290 3311 delta = gethrtime() - fio->io_timestamp;
3291 3312 if (delta > spa_deadman_synctime(spa)) {
3292 3313 zfs_dbgmsg("SLOW IO: zio timestamp %lluns, "
3293 3314 "delta %lluns, last io %lluns",
3294 3315 fio->io_timestamp, delta,
3295 3316 vq->vq_io_complete_ts);
3296 3317 fm_panic("I/O to pool '%s' appears to be "
3297 3318 "hung.", spa_name(spa));
3298 3319 }
3299 3320 }
3300 3321 mutex_exit(&vq->vq_lock);
3301 3322 }
3302 3323 }
|
↓ open down ↓ |
3118 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX