Print this page
4932 vdev incorrectly expanding on last mirror child -> top-level vdev
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
| 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)
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
175 175 void
176 176 vdev_add_child(vdev_t *pvd, vdev_t *cvd)
177 177 {
178 178 size_t oldsize, newsize;
179 179 uint64_t id = cvd->vdev_id;
180 180 vdev_t **newchild;
181 181
182 182 ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
183 183 ASSERT(cvd->vdev_parent == NULL);
184 184
185 185 cvd->vdev_parent = pvd;
186 186
187 187 if (pvd == NULL)
188 188 return;
189 189
190 190 ASSERT(id >= pvd->vdev_children || pvd->vdev_child[id] == NULL);
191 191
192 192 oldsize = pvd->vdev_children * sizeof (vdev_t *);
193 193 pvd->vdev_children = MAX(pvd->vdev_children, id + 1);
194 194 newsize = pvd->vdev_children * sizeof (vdev_t *);
195 195
196 196 newchild = kmem_zalloc(newsize, KM_SLEEP);
197 197 if (pvd->vdev_child != NULL) {
198 198 bcopy(pvd->vdev_child, newchild, oldsize);
199 199 kmem_free(pvd->vdev_child, oldsize);
200 200 }
201 201
202 202 pvd->vdev_child = newchild;
203 203 pvd->vdev_child[id] = cvd;
204 204
205 205 cvd->vdev_top = (pvd->vdev_top ? pvd->vdev_top: cvd);
206 206 ASSERT(cvd->vdev_top->vdev_parent->vdev_parent == NULL);
207 207
208 208 /*
209 209 * Walk up all ancestors to update guid sum.
210 210 */
211 211 for (; pvd != NULL; pvd = pvd->vdev_parent)
212 212 pvd->vdev_guid_sum += cvd->vdev_guid_sum;
213 213 }
214 214
215 215 void
216 216 vdev_remove_child(vdev_t *pvd, vdev_t *cvd)
217 217 {
218 218 int c;
219 219 uint_t id = cvd->vdev_id;
220 220
221 221 ASSERT(cvd->vdev_parent == pvd);
222 222
223 223 if (pvd == NULL)
224 224 return;
225 225
226 226 ASSERT(id < pvd->vdev_children);
227 227 ASSERT(pvd->vdev_child[id] == cvd);
228 228
229 229 pvd->vdev_child[id] = NULL;
230 230 cvd->vdev_parent = NULL;
231 231
232 232 for (c = 0; c < pvd->vdev_children; c++)
233 233 if (pvd->vdev_child[c])
234 234 break;
235 235
236 236 if (c == pvd->vdev_children) {
237 237 kmem_free(pvd->vdev_child, c * sizeof (vdev_t *));
238 238 pvd->vdev_child = NULL;
239 239 pvd->vdev_children = 0;
240 240 }
241 241
242 242 /*
243 243 * Walk up all ancestors to update guid sum.
244 244 */
245 245 for (; pvd != NULL; pvd = pvd->vdev_parent)
246 246 pvd->vdev_guid_sum -= cvd->vdev_guid_sum;
247 247 }
248 248
249 249 /*
250 250 * Remove any holes in the child array.
251 251 */
252 252 void
253 253 vdev_compact_children(vdev_t *pvd)
254 254 {
255 255 vdev_t **newchild, *cvd;
256 256 int oldc = pvd->vdev_children;
257 257 int newc;
258 258
259 259 ASSERT(spa_config_held(pvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
260 260
261 261 for (int c = newc = 0; c < oldc; c++)
262 262 if (pvd->vdev_child[c])
263 263 newc++;
264 264
265 265 newchild = kmem_alloc(newc * sizeof (vdev_t *), KM_SLEEP);
266 266
267 267 for (int c = newc = 0; c < oldc; c++) {
268 268 if ((cvd = pvd->vdev_child[c]) != NULL) {
269 269 newchild[newc] = cvd;
270 270 cvd->vdev_id = newc++;
271 271 }
272 272 }
273 273
274 274 kmem_free(pvd->vdev_child, oldc * sizeof (vdev_t *));
275 275 pvd->vdev_child = newchild;
276 276 pvd->vdev_children = newc;
277 277 }
278 278
279 279 /*
280 280 * Allocate and minimally initialize a vdev_t.
281 281 */
282 282 vdev_t *
283 283 vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
284 284 {
285 285 vdev_t *vd;
286 286
287 287 vd = kmem_zalloc(sizeof (vdev_t), KM_SLEEP);
288 288
289 289 if (spa->spa_root_vdev == NULL) {
290 290 ASSERT(ops == &vdev_root_ops);
291 291 spa->spa_root_vdev = vd;
292 292 spa->spa_load_guid = spa_generate_guid(NULL);
293 293 }
294 294
295 295 if (guid == 0 && ops != &vdev_hole_ops) {
296 296 if (spa->spa_root_vdev == vd) {
297 297 /*
298 298 * The root vdev's guid will also be the pool guid,
299 299 * which must be unique among all pools.
300 300 */
301 301 guid = spa_generate_guid(NULL);
302 302 } else {
303 303 /*
304 304 * Any other vdev's guid must be unique within the pool.
305 305 */
306 306 guid = spa_generate_guid(spa);
307 307 }
308 308 ASSERT(!spa_guid_exists(spa_guid(spa), guid));
309 309 }
310 310
311 311 vd->vdev_spa = spa;
312 312 vd->vdev_id = id;
313 313 vd->vdev_guid = guid;
314 314 vd->vdev_guid_sum = guid;
315 315 vd->vdev_ops = ops;
316 316 vd->vdev_state = VDEV_STATE_CLOSED;
317 317 vd->vdev_ishole = (ops == &vdev_hole_ops);
318 318
319 319 mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_DEFAULT, NULL);
320 320 mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL);
321 321 mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL);
322 322 for (int t = 0; t < DTL_TYPES; t++) {
323 323 vd->vdev_dtl[t] = range_tree_create(NULL, NULL,
324 324 &vd->vdev_dtl_lock);
325 325 }
326 326 txg_list_create(&vd->vdev_ms_list,
327 327 offsetof(struct metaslab, ms_txg_node));
328 328 txg_list_create(&vd->vdev_dtl_list,
329 329 offsetof(struct vdev, vdev_dtl_node));
330 330 vd->vdev_stat.vs_timestamp = gethrtime();
331 331 vdev_queue_init(vd);
332 332 vdev_cache_init(vd);
333 333
334 334 return (vd);
335 335 }
336 336
337 337 /*
338 338 * Allocate a new vdev. The 'alloctype' is used to control whether we are
339 339 * creating a new vdev or loading an existing one - the behavior is slightly
340 340 * different for each case.
341 341 */
342 342 int
343 343 vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id,
344 344 int alloctype)
345 345 {
346 346 vdev_ops_t *ops;
347 347 char *type;
348 348 uint64_t guid = 0, islog, nparity;
349 349 vdev_t *vd;
350 350
351 351 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
352 352
353 353 if (nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) != 0)
354 354 return (SET_ERROR(EINVAL));
355 355
356 356 if ((ops = vdev_getops(type)) == NULL)
357 357 return (SET_ERROR(EINVAL));
358 358
359 359 /*
360 360 * If this is a load, get the vdev guid from the nvlist.
361 361 * Otherwise, vdev_alloc_common() will generate one for us.
362 362 */
363 363 if (alloctype == VDEV_ALLOC_LOAD) {
364 364 uint64_t label_id;
365 365
366 366 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ID, &label_id) ||
367 367 label_id != id)
368 368 return (SET_ERROR(EINVAL));
369 369
370 370 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
371 371 return (SET_ERROR(EINVAL));
372 372 } else if (alloctype == VDEV_ALLOC_SPARE) {
373 373 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
374 374 return (SET_ERROR(EINVAL));
375 375 } else if (alloctype == VDEV_ALLOC_L2CACHE) {
376 376 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
377 377 return (SET_ERROR(EINVAL));
378 378 } else if (alloctype == VDEV_ALLOC_ROOTPOOL) {
379 379 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
380 380 return (SET_ERROR(EINVAL));
381 381 }
382 382
383 383 /*
384 384 * The first allocated vdev must be of type 'root'.
385 385 */
386 386 if (ops != &vdev_root_ops && spa->spa_root_vdev == NULL)
387 387 return (SET_ERROR(EINVAL));
388 388
389 389 /*
390 390 * Determine whether we're a log vdev.
391 391 */
392 392 islog = 0;
393 393 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &islog);
394 394 if (islog && spa_version(spa) < SPA_VERSION_SLOGS)
395 395 return (SET_ERROR(ENOTSUP));
396 396
397 397 if (ops == &vdev_hole_ops && spa_version(spa) < SPA_VERSION_HOLES)
398 398 return (SET_ERROR(ENOTSUP));
399 399
400 400 /*
401 401 * Set the nparity property for RAID-Z vdevs.
402 402 */
403 403 nparity = -1ULL;
404 404 if (ops == &vdev_raidz_ops) {
405 405 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY,
406 406 &nparity) == 0) {
407 407 if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY)
408 408 return (SET_ERROR(EINVAL));
409 409 /*
410 410 * Previous versions could only support 1 or 2 parity
411 411 * device.
412 412 */
413 413 if (nparity > 1 &&
414 414 spa_version(spa) < SPA_VERSION_RAIDZ2)
415 415 return (SET_ERROR(ENOTSUP));
416 416 if (nparity > 2 &&
417 417 spa_version(spa) < SPA_VERSION_RAIDZ3)
418 418 return (SET_ERROR(ENOTSUP));
419 419 } else {
420 420 /*
421 421 * We require the parity to be specified for SPAs that
422 422 * support multiple parity levels.
423 423 */
424 424 if (spa_version(spa) >= SPA_VERSION_RAIDZ2)
425 425 return (SET_ERROR(EINVAL));
426 426 /*
427 427 * Otherwise, we default to 1 parity device for RAID-Z.
428 428 */
429 429 nparity = 1;
430 430 }
431 431 } else {
432 432 nparity = 0;
433 433 }
434 434 ASSERT(nparity != -1ULL);
435 435
436 436 vd = vdev_alloc_common(spa, id, guid, ops);
437 437
438 438 vd->vdev_islog = islog;
439 439 vd->vdev_nparity = nparity;
440 440
441 441 if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &vd->vdev_path) == 0)
442 442 vd->vdev_path = spa_strdup(vd->vdev_path);
443 443 if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &vd->vdev_devid) == 0)
444 444 vd->vdev_devid = spa_strdup(vd->vdev_devid);
445 445 if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH,
446 446 &vd->vdev_physpath) == 0)
447 447 vd->vdev_physpath = spa_strdup(vd->vdev_physpath);
448 448 if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &vd->vdev_fru) == 0)
449 449 vd->vdev_fru = spa_strdup(vd->vdev_fru);
450 450
451 451 /*
452 452 * Set the whole_disk property. If it's not specified, leave the value
453 453 * as -1.
454 454 */
455 455 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
456 456 &vd->vdev_wholedisk) != 0)
457 457 vd->vdev_wholedisk = -1ULL;
458 458
459 459 /*
460 460 * Look for the 'not present' flag. This will only be set if the device
461 461 * was not present at the time of import.
462 462 */
463 463 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT,
464 464 &vd->vdev_not_present);
465 465
466 466 /*
467 467 * Get the alignment requirement.
468 468 */
469 469 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT, &vd->vdev_ashift);
470 470
471 471 /*
472 472 * Retrieve the vdev creation time.
473 473 */
474 474 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_CREATE_TXG,
475 475 &vd->vdev_crtxg);
476 476
477 477 /*
478 478 * If we're a top-level vdev, try to load the allocation parameters.
479 479 */
480 480 if (parent && !parent->vdev_parent &&
481 481 (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
482 482 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
483 483 &vd->vdev_ms_array);
484 484 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
485 485 &vd->vdev_ms_shift);
486 486 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE,
487 487 &vd->vdev_asize);
488 488 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVING,
489 489 &vd->vdev_removing);
490 490 }
491 491
492 492 if (parent && !parent->vdev_parent && alloctype != VDEV_ALLOC_ATTACH) {
493 493 ASSERT(alloctype == VDEV_ALLOC_LOAD ||
494 494 alloctype == VDEV_ALLOC_ADD ||
495 495 alloctype == VDEV_ALLOC_SPLIT ||
496 496 alloctype == VDEV_ALLOC_ROOTPOOL);
497 497 vd->vdev_mg = metaslab_group_create(islog ?
498 498 spa_log_class(spa) : spa_normal_class(spa), vd);
499 499 }
500 500
501 501 /*
502 502 * If we're a leaf vdev, try to load the DTL object and other state.
503 503 */
504 504 if (vd->vdev_ops->vdev_op_leaf &&
505 505 (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE ||
506 506 alloctype == VDEV_ALLOC_ROOTPOOL)) {
507 507 if (alloctype == VDEV_ALLOC_LOAD) {
508 508 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL,
509 509 &vd->vdev_dtl_object);
510 510 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE,
511 511 &vd->vdev_unspare);
512 512 }
513 513
514 514 if (alloctype == VDEV_ALLOC_ROOTPOOL) {
515 515 uint64_t spare = 0;
516 516
517 517 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE,
518 518 &spare) == 0 && spare)
519 519 spa_spare_add(vd);
520 520 }
521 521
522 522 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE,
523 523 &vd->vdev_offline);
524 524
525 525 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_RESILVER_TXG,
526 526 &vd->vdev_resilver_txg);
527 527
528 528 /*
529 529 * When importing a pool, we want to ignore the persistent fault
530 530 * state, as the diagnosis made on another system may not be
531 531 * valid in the current context. Local vdevs will
532 532 * remain in the faulted state.
533 533 */
534 534 if (spa_load_state(spa) == SPA_LOAD_OPEN) {
535 535 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED,
536 536 &vd->vdev_faulted);
537 537 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED,
538 538 &vd->vdev_degraded);
539 539 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED,
540 540 &vd->vdev_removed);
541 541
542 542 if (vd->vdev_faulted || vd->vdev_degraded) {
543 543 char *aux;
544 544
545 545 vd->vdev_label_aux =
546 546 VDEV_AUX_ERR_EXCEEDED;
547 547 if (nvlist_lookup_string(nv,
548 548 ZPOOL_CONFIG_AUX_STATE, &aux) == 0 &&
549 549 strcmp(aux, "external") == 0)
550 550 vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
551 551 }
552 552 }
553 553 }
554 554
555 555 /*
556 556 * Add ourselves to the parent's list of children.
557 557 */
558 558 vdev_add_child(parent, vd);
559 559
560 560 *vdp = vd;
561 561
562 562 return (0);
563 563 }
564 564
565 565 void
566 566 vdev_free(vdev_t *vd)
567 567 {
568 568 spa_t *spa = vd->vdev_spa;
569 569
570 570 /*
571 571 * vdev_free() implies closing the vdev first. This is simpler than
572 572 * trying to ensure complicated semantics for all callers.
573 573 */
574 574 vdev_close(vd);
575 575
576 576 ASSERT(!list_link_active(&vd->vdev_config_dirty_node));
577 577 ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
578 578
579 579 /*
580 580 * Free all children.
581 581 */
582 582 for (int c = 0; c < vd->vdev_children; c++)
583 583 vdev_free(vd->vdev_child[c]);
584 584
585 585 ASSERT(vd->vdev_child == NULL);
586 586 ASSERT(vd->vdev_guid_sum == vd->vdev_guid);
587 587
588 588 /*
589 589 * Discard allocation state.
590 590 */
591 591 if (vd->vdev_mg != NULL) {
592 592 vdev_metaslab_fini(vd);
593 593 metaslab_group_destroy(vd->vdev_mg);
594 594 }
595 595
596 596 ASSERT0(vd->vdev_stat.vs_space);
597 597 ASSERT0(vd->vdev_stat.vs_dspace);
598 598 ASSERT0(vd->vdev_stat.vs_alloc);
599 599
600 600 /*
601 601 * Remove this vdev from its parent's child list.
602 602 */
603 603 vdev_remove_child(vd->vdev_parent, vd);
604 604
605 605 ASSERT(vd->vdev_parent == NULL);
606 606
607 607 /*
608 608 * Clean up vdev structure.
609 609 */
610 610 vdev_queue_fini(vd);
611 611 vdev_cache_fini(vd);
612 612
613 613 if (vd->vdev_path)
614 614 spa_strfree(vd->vdev_path);
615 615 if (vd->vdev_devid)
616 616 spa_strfree(vd->vdev_devid);
617 617 if (vd->vdev_physpath)
618 618 spa_strfree(vd->vdev_physpath);
619 619 if (vd->vdev_fru)
620 620 spa_strfree(vd->vdev_fru);
621 621
622 622 if (vd->vdev_isspare)
623 623 spa_spare_remove(vd);
624 624 if (vd->vdev_isl2cache)
625 625 spa_l2cache_remove(vd);
626 626
627 627 txg_list_destroy(&vd->vdev_ms_list);
628 628 txg_list_destroy(&vd->vdev_dtl_list);
629 629
630 630 mutex_enter(&vd->vdev_dtl_lock);
631 631 space_map_close(vd->vdev_dtl_sm);
632 632 for (int t = 0; t < DTL_TYPES; t++) {
633 633 range_tree_vacate(vd->vdev_dtl[t], NULL, NULL);
634 634 range_tree_destroy(vd->vdev_dtl[t]);
635 635 }
636 636 mutex_exit(&vd->vdev_dtl_lock);
637 637
638 638 mutex_destroy(&vd->vdev_dtl_lock);
639 639 mutex_destroy(&vd->vdev_stat_lock);
640 640 mutex_destroy(&vd->vdev_probe_lock);
641 641
642 642 if (vd == spa->spa_root_vdev)
643 643 spa->spa_root_vdev = NULL;
644 644
645 645 kmem_free(vd, sizeof (vdev_t));
646 646 }
647 647
648 648 /*
649 649 * Transfer top-level vdev state from svd to tvd.
650 650 */
651 651 static void
652 652 vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
653 653 {
654 654 spa_t *spa = svd->vdev_spa;
655 655 metaslab_t *msp;
656 656 vdev_t *vd;
657 657 int t;
658 658
659 659 ASSERT(tvd == tvd->vdev_top);
660 660
661 661 tvd->vdev_ms_array = svd->vdev_ms_array;
662 662 tvd->vdev_ms_shift = svd->vdev_ms_shift;
663 663 tvd->vdev_ms_count = svd->vdev_ms_count;
664 664
665 665 svd->vdev_ms_array = 0;
666 666 svd->vdev_ms_shift = 0;
667 667 svd->vdev_ms_count = 0;
668 668
669 669 if (tvd->vdev_mg)
670 670 ASSERT3P(tvd->vdev_mg, ==, svd->vdev_mg);
671 671 tvd->vdev_mg = svd->vdev_mg;
672 672 tvd->vdev_ms = svd->vdev_ms;
673 673
674 674 svd->vdev_mg = NULL;
675 675 svd->vdev_ms = NULL;
676 676
677 677 if (tvd->vdev_mg != NULL)
678 678 tvd->vdev_mg->mg_vd = tvd;
679 679
680 680 tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
681 681 tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
682 682 tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
683 683
684 684 svd->vdev_stat.vs_alloc = 0;
685 685 svd->vdev_stat.vs_space = 0;
686 686 svd->vdev_stat.vs_dspace = 0;
687 687
688 688 for (t = 0; t < TXG_SIZE; t++) {
689 689 while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
690 690 (void) txg_list_add(&tvd->vdev_ms_list, msp, t);
691 691 while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
692 692 (void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
693 693 if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
694 694 (void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
695 695 }
696 696
697 697 if (list_link_active(&svd->vdev_config_dirty_node)) {
698 698 vdev_config_clean(svd);
699 699 vdev_config_dirty(tvd);
700 700 }
701 701
702 702 if (list_link_active(&svd->vdev_state_dirty_node)) {
703 703 vdev_state_clean(svd);
704 704 vdev_state_dirty(tvd);
705 705 }
706 706
707 707 tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
708 708 svd->vdev_deflate_ratio = 0;
709 709
710 710 tvd->vdev_islog = svd->vdev_islog;
711 711 svd->vdev_islog = 0;
712 712 }
713 713
714 714 static void
715 715 vdev_top_update(vdev_t *tvd, vdev_t *vd)
716 716 {
717 717 if (vd == NULL)
718 718 return;
719 719
720 720 vd->vdev_top = tvd;
721 721
722 722 for (int c = 0; c < vd->vdev_children; c++)
723 723 vdev_top_update(tvd, vd->vdev_child[c]);
724 724 }
725 725
726 726 /*
727 727 * Add a mirror/replacing vdev above an existing vdev.
728 728 */
729 729 vdev_t *
730 730 vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
731 731 {
732 732 spa_t *spa = cvd->vdev_spa;
733 733 vdev_t *pvd = cvd->vdev_parent;
734 734 vdev_t *mvd;
735 735
736 736 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
737 737
738 738 mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
739 739
740 740 mvd->vdev_asize = cvd->vdev_asize;
741 741 mvd->vdev_min_asize = cvd->vdev_min_asize;
742 742 mvd->vdev_max_asize = cvd->vdev_max_asize;
743 743 mvd->vdev_ashift = cvd->vdev_ashift;
744 744 mvd->vdev_state = cvd->vdev_state;
745 745 mvd->vdev_crtxg = cvd->vdev_crtxg;
746 746
747 747 vdev_remove_child(pvd, cvd);
748 748 vdev_add_child(pvd, mvd);
749 749 cvd->vdev_id = mvd->vdev_children;
750 750 vdev_add_child(mvd, cvd);
751 751 vdev_top_update(cvd->vdev_top, cvd->vdev_top);
752 752
753 753 if (mvd == mvd->vdev_top)
754 754 vdev_top_transfer(cvd, mvd);
755 755
756 756 return (mvd);
757 757 }
758 758
759 759 /*
760 760 * Remove a 1-way mirror/replacing vdev from the tree.
761 761 */
762 762 void
763 763 vdev_remove_parent(vdev_t *cvd)
764 764 {
765 765 vdev_t *mvd = cvd->vdev_parent;
766 766 vdev_t *pvd = mvd->vdev_parent;
767 767
768 768 ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
769 769
770 770 ASSERT(mvd->vdev_children == 1);
771 771 ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
772 772 mvd->vdev_ops == &vdev_replacing_ops ||
773 773 mvd->vdev_ops == &vdev_spare_ops);
774 774 cvd->vdev_ashift = mvd->vdev_ashift;
775 775
776 776 vdev_remove_child(mvd, cvd);
777 777 vdev_remove_child(pvd, mvd);
778 778
779 779 /*
|
↓ open down ↓ |
779 lines elided |
↑ open up ↑ |
780 780 * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
781 781 * Otherwise, we could have detached an offline device, and when we
782 782 * go to import the pool we'll think we have two top-level vdevs,
783 783 * instead of a different version of the same top-level vdev.
784 784 */
785 785 if (mvd->vdev_top == mvd) {
786 786 uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
787 787 cvd->vdev_orig_guid = cvd->vdev_guid;
788 788 cvd->vdev_guid += guid_delta;
789 789 cvd->vdev_guid_sum += guid_delta;
790 +
791 + /*
792 + * If pool not set for autoexpand, we need to also preserve
793 + * mvd's asize to prevent automatic expansion of cvd.
794 + * Otherwise if we are adjusting the mirror by attaching and
795 + * detaching children of non-uniform sizes, the mirror could
796 + * autoexpand, unexpectedly requiring larger devices to
797 + * re-establish the mirror.
798 + */
799 + if (!cvd->vdev_spa->spa_autoexpand)
800 + cvd->vdev_asize = mvd->vdev_asize;
790 801 }
791 802 cvd->vdev_id = mvd->vdev_id;
792 803 vdev_add_child(pvd, cvd);
793 804 vdev_top_update(cvd->vdev_top, cvd->vdev_top);
794 805
795 806 if (cvd == cvd->vdev_top)
796 807 vdev_top_transfer(mvd, cvd);
797 808
798 809 ASSERT(mvd->vdev_children == 0);
799 810 vdev_free(mvd);
800 811 }
801 812
802 813 int
803 814 vdev_metaslab_init(vdev_t *vd, uint64_t txg)
804 815 {
805 816 spa_t *spa = vd->vdev_spa;
806 817 objset_t *mos = spa->spa_meta_objset;
807 818 uint64_t m;
808 819 uint64_t oldc = vd->vdev_ms_count;
809 820 uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
810 821 metaslab_t **mspp;
811 822 int error;
812 823
813 824 ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER));
814 825
815 826 /*
816 827 * This vdev is not being allocated from yet or is a hole.
817 828 */
818 829 if (vd->vdev_ms_shift == 0)
819 830 return (0);
820 831
821 832 ASSERT(!vd->vdev_ishole);
822 833
823 834 /*
824 835 * Compute the raidz-deflation ratio. Note, we hard-code
825 836 * in 128k (1 << 17) because it is the current "typical" blocksize.
826 837 * Even if SPA_MAXBLOCKSIZE changes, this algorithm must never change,
827 838 * or we will inconsistently account for existing bp's.
828 839 */
829 840 vd->vdev_deflate_ratio = (1 << 17) /
830 841 (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
831 842
832 843 ASSERT(oldc <= newc);
833 844
834 845 mspp = kmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
835 846
836 847 if (oldc != 0) {
837 848 bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp));
838 849 kmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
839 850 }
840 851
841 852 vd->vdev_ms = mspp;
842 853 vd->vdev_ms_count = newc;
843 854
844 855 for (m = oldc; m < newc; m++) {
845 856 uint64_t object = 0;
846 857
847 858 if (txg == 0) {
848 859 error = dmu_read(mos, vd->vdev_ms_array,
849 860 m * sizeof (uint64_t), sizeof (uint64_t), &object,
850 861 DMU_READ_PREFETCH);
851 862 if (error)
852 863 return (error);
853 864 }
854 865 vd->vdev_ms[m] = metaslab_init(vd->vdev_mg, m, object, txg);
855 866 }
856 867
857 868 if (txg == 0)
858 869 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER);
859 870
860 871 /*
861 872 * If the vdev is being removed we don't activate
862 873 * the metaslabs since we want to ensure that no new
863 874 * allocations are performed on this device.
864 875 */
865 876 if (oldc == 0 && !vd->vdev_removing)
866 877 metaslab_group_activate(vd->vdev_mg);
867 878
868 879 if (txg == 0)
869 880 spa_config_exit(spa, SCL_ALLOC, FTAG);
870 881
871 882 return (0);
872 883 }
873 884
874 885 void
875 886 vdev_metaslab_fini(vdev_t *vd)
876 887 {
877 888 uint64_t m;
878 889 uint64_t count = vd->vdev_ms_count;
879 890
880 891 if (vd->vdev_ms != NULL) {
881 892 metaslab_group_passivate(vd->vdev_mg);
882 893 for (m = 0; m < count; m++) {
883 894 metaslab_t *msp = vd->vdev_ms[m];
884 895
885 896 if (msp != NULL)
886 897 metaslab_fini(msp);
887 898 }
888 899 kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
889 900 vd->vdev_ms = NULL;
890 901 }
891 902 }
892 903
893 904 typedef struct vdev_probe_stats {
894 905 boolean_t vps_readable;
895 906 boolean_t vps_writeable;
896 907 int vps_flags;
897 908 } vdev_probe_stats_t;
898 909
899 910 static void
900 911 vdev_probe_done(zio_t *zio)
901 912 {
902 913 spa_t *spa = zio->io_spa;
903 914 vdev_t *vd = zio->io_vd;
904 915 vdev_probe_stats_t *vps = zio->io_private;
905 916
906 917 ASSERT(vd->vdev_probe_zio != NULL);
907 918
908 919 if (zio->io_type == ZIO_TYPE_READ) {
909 920 if (zio->io_error == 0)
910 921 vps->vps_readable = 1;
911 922 if (zio->io_error == 0 && spa_writeable(spa)) {
912 923 zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
913 924 zio->io_offset, zio->io_size, zio->io_data,
914 925 ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
915 926 ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
916 927 } else {
917 928 zio_buf_free(zio->io_data, zio->io_size);
918 929 }
919 930 } else if (zio->io_type == ZIO_TYPE_WRITE) {
920 931 if (zio->io_error == 0)
921 932 vps->vps_writeable = 1;
922 933 zio_buf_free(zio->io_data, zio->io_size);
923 934 } else if (zio->io_type == ZIO_TYPE_NULL) {
924 935 zio_t *pio;
925 936
926 937 vd->vdev_cant_read |= !vps->vps_readable;
927 938 vd->vdev_cant_write |= !vps->vps_writeable;
928 939
929 940 if (vdev_readable(vd) &&
930 941 (vdev_writeable(vd) || !spa_writeable(spa))) {
931 942 zio->io_error = 0;
932 943 } else {
933 944 ASSERT(zio->io_error != 0);
934 945 zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
935 946 spa, vd, NULL, 0, 0);
936 947 zio->io_error = SET_ERROR(ENXIO);
937 948 }
938 949
939 950 mutex_enter(&vd->vdev_probe_lock);
940 951 ASSERT(vd->vdev_probe_zio == zio);
941 952 vd->vdev_probe_zio = NULL;
942 953 mutex_exit(&vd->vdev_probe_lock);
943 954
944 955 while ((pio = zio_walk_parents(zio)) != NULL)
945 956 if (!vdev_accessible(vd, pio))
946 957 pio->io_error = SET_ERROR(ENXIO);
947 958
948 959 kmem_free(vps, sizeof (*vps));
949 960 }
950 961 }
951 962
952 963 /*
953 964 * Determine whether this device is accessible.
954 965 *
955 966 * Read and write to several known locations: the pad regions of each
956 967 * vdev label but the first, which we leave alone in case it contains
957 968 * a VTOC.
958 969 */
959 970 zio_t *
960 971 vdev_probe(vdev_t *vd, zio_t *zio)
961 972 {
962 973 spa_t *spa = vd->vdev_spa;
963 974 vdev_probe_stats_t *vps = NULL;
964 975 zio_t *pio;
965 976
966 977 ASSERT(vd->vdev_ops->vdev_op_leaf);
967 978
968 979 /*
969 980 * Don't probe the probe.
970 981 */
971 982 if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
972 983 return (NULL);
973 984
974 985 /*
975 986 * To prevent 'probe storms' when a device fails, we create
976 987 * just one probe i/o at a time. All zios that want to probe
977 988 * this vdev will become parents of the probe io.
978 989 */
979 990 mutex_enter(&vd->vdev_probe_lock);
980 991
981 992 if ((pio = vd->vdev_probe_zio) == NULL) {
982 993 vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
983 994
984 995 vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
985 996 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
986 997 ZIO_FLAG_TRYHARD;
987 998
988 999 if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
989 1000 /*
990 1001 * vdev_cant_read and vdev_cant_write can only
991 1002 * transition from TRUE to FALSE when we have the
992 1003 * SCL_ZIO lock as writer; otherwise they can only
993 1004 * transition from FALSE to TRUE. This ensures that
994 1005 * any zio looking at these values can assume that
995 1006 * failures persist for the life of the I/O. That's
996 1007 * important because when a device has intermittent
997 1008 * connectivity problems, we want to ensure that
998 1009 * they're ascribed to the device (ENXIO) and not
999 1010 * the zio (EIO).
1000 1011 *
1001 1012 * Since we hold SCL_ZIO as writer here, clear both
1002 1013 * values so the probe can reevaluate from first
1003 1014 * principles.
1004 1015 */
1005 1016 vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
1006 1017 vd->vdev_cant_read = B_FALSE;
1007 1018 vd->vdev_cant_write = B_FALSE;
1008 1019 }
1009 1020
1010 1021 vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
1011 1022 vdev_probe_done, vps,
1012 1023 vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
1013 1024
1014 1025 /*
1015 1026 * We can't change the vdev state in this context, so we
1016 1027 * kick off an async task to do it on our behalf.
1017 1028 */
1018 1029 if (zio != NULL) {
1019 1030 vd->vdev_probe_wanted = B_TRUE;
1020 1031 spa_async_request(spa, SPA_ASYNC_PROBE);
1021 1032 }
1022 1033 }
1023 1034
1024 1035 if (zio != NULL)
1025 1036 zio_add_child(zio, pio);
1026 1037
1027 1038 mutex_exit(&vd->vdev_probe_lock);
1028 1039
1029 1040 if (vps == NULL) {
1030 1041 ASSERT(zio != NULL);
1031 1042 return (NULL);
1032 1043 }
1033 1044
1034 1045 for (int l = 1; l < VDEV_LABELS; l++) {
1035 1046 zio_nowait(zio_read_phys(pio, vd,
1036 1047 vdev_label_offset(vd->vdev_psize, l,
1037 1048 offsetof(vdev_label_t, vl_pad2)),
1038 1049 VDEV_PAD_SIZE, zio_buf_alloc(VDEV_PAD_SIZE),
1039 1050 ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1040 1051 ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1041 1052 }
1042 1053
1043 1054 if (zio == NULL)
1044 1055 return (pio);
1045 1056
1046 1057 zio_nowait(pio);
1047 1058 return (NULL);
1048 1059 }
1049 1060
1050 1061 static void
1051 1062 vdev_open_child(void *arg)
1052 1063 {
1053 1064 vdev_t *vd = arg;
1054 1065
1055 1066 vd->vdev_open_thread = curthread;
1056 1067 vd->vdev_open_error = vdev_open(vd);
1057 1068 vd->vdev_open_thread = NULL;
1058 1069 }
1059 1070
1060 1071 boolean_t
1061 1072 vdev_uses_zvols(vdev_t *vd)
1062 1073 {
1063 1074 if (vd->vdev_path && strncmp(vd->vdev_path, ZVOL_DIR,
1064 1075 strlen(ZVOL_DIR)) == 0)
1065 1076 return (B_TRUE);
1066 1077 for (int c = 0; c < vd->vdev_children; c++)
1067 1078 if (vdev_uses_zvols(vd->vdev_child[c]))
1068 1079 return (B_TRUE);
1069 1080 return (B_FALSE);
1070 1081 }
1071 1082
1072 1083 void
1073 1084 vdev_open_children(vdev_t *vd)
1074 1085 {
1075 1086 taskq_t *tq;
1076 1087 int children = vd->vdev_children;
1077 1088
1078 1089 /*
1079 1090 * in order to handle pools on top of zvols, do the opens
1080 1091 * in a single thread so that the same thread holds the
1081 1092 * spa_namespace_lock
1082 1093 */
1083 1094 if (vdev_uses_zvols(vd)) {
1084 1095 for (int c = 0; c < children; c++)
1085 1096 vd->vdev_child[c]->vdev_open_error =
1086 1097 vdev_open(vd->vdev_child[c]);
1087 1098 return;
1088 1099 }
1089 1100 tq = taskq_create("vdev_open", children, minclsyspri,
1090 1101 children, children, TASKQ_PREPOPULATE);
1091 1102
1092 1103 for (int c = 0; c < children; c++)
1093 1104 VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c],
1094 1105 TQ_SLEEP) != NULL);
1095 1106
1096 1107 taskq_destroy(tq);
1097 1108 }
1098 1109
1099 1110 /*
1100 1111 * Prepare a virtual device for access.
1101 1112 */
1102 1113 int
1103 1114 vdev_open(vdev_t *vd)
1104 1115 {
1105 1116 spa_t *spa = vd->vdev_spa;
1106 1117 int error;
1107 1118 uint64_t osize = 0;
1108 1119 uint64_t max_osize = 0;
1109 1120 uint64_t asize, max_asize, psize;
1110 1121 uint64_t ashift = 0;
1111 1122
1112 1123 ASSERT(vd->vdev_open_thread == curthread ||
1113 1124 spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1114 1125 ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1115 1126 vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1116 1127 vd->vdev_state == VDEV_STATE_OFFLINE);
1117 1128
1118 1129 vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1119 1130 vd->vdev_cant_read = B_FALSE;
1120 1131 vd->vdev_cant_write = B_FALSE;
1121 1132 vd->vdev_min_asize = vdev_get_min_asize(vd);
1122 1133
1123 1134 /*
1124 1135 * If this vdev is not removed, check its fault status. If it's
1125 1136 * faulted, bail out of the open.
1126 1137 */
1127 1138 if (!vd->vdev_removed && vd->vdev_faulted) {
1128 1139 ASSERT(vd->vdev_children == 0);
1129 1140 ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1130 1141 vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1131 1142 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1132 1143 vd->vdev_label_aux);
1133 1144 return (SET_ERROR(ENXIO));
1134 1145 } else if (vd->vdev_offline) {
1135 1146 ASSERT(vd->vdev_children == 0);
1136 1147 vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1137 1148 return (SET_ERROR(ENXIO));
1138 1149 }
1139 1150
1140 1151 error = vd->vdev_ops->vdev_op_open(vd, &osize, &max_osize, &ashift);
1141 1152
1142 1153 /*
1143 1154 * Reset the vdev_reopening flag so that we actually close
1144 1155 * the vdev on error.
1145 1156 */
1146 1157 vd->vdev_reopening = B_FALSE;
1147 1158 if (zio_injection_enabled && error == 0)
1148 1159 error = zio_handle_device_injection(vd, NULL, ENXIO);
1149 1160
1150 1161 if (error) {
1151 1162 if (vd->vdev_removed &&
1152 1163 vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1153 1164 vd->vdev_removed = B_FALSE;
1154 1165
1155 1166 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1156 1167 vd->vdev_stat.vs_aux);
1157 1168 return (error);
1158 1169 }
1159 1170
1160 1171 vd->vdev_removed = B_FALSE;
1161 1172
1162 1173 /*
1163 1174 * Recheck the faulted flag now that we have confirmed that
1164 1175 * the vdev is accessible. If we're faulted, bail.
1165 1176 */
1166 1177 if (vd->vdev_faulted) {
1167 1178 ASSERT(vd->vdev_children == 0);
1168 1179 ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1169 1180 vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1170 1181 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1171 1182 vd->vdev_label_aux);
1172 1183 return (SET_ERROR(ENXIO));
1173 1184 }
1174 1185
1175 1186 if (vd->vdev_degraded) {
1176 1187 ASSERT(vd->vdev_children == 0);
1177 1188 vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1178 1189 VDEV_AUX_ERR_EXCEEDED);
1179 1190 } else {
1180 1191 vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1181 1192 }
1182 1193
1183 1194 /*
1184 1195 * For hole or missing vdevs we just return success.
1185 1196 */
1186 1197 if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1187 1198 return (0);
1188 1199
1189 1200 for (int c = 0; c < vd->vdev_children; c++) {
1190 1201 if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1191 1202 vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1192 1203 VDEV_AUX_NONE);
1193 1204 break;
1194 1205 }
1195 1206 }
1196 1207
1197 1208 osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1198 1209 max_osize = P2ALIGN(max_osize, (uint64_t)sizeof (vdev_label_t));
1199 1210
1200 1211 if (vd->vdev_children == 0) {
1201 1212 if (osize < SPA_MINDEVSIZE) {
1202 1213 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1203 1214 VDEV_AUX_TOO_SMALL);
1204 1215 return (SET_ERROR(EOVERFLOW));
1205 1216 }
1206 1217 psize = osize;
1207 1218 asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1208 1219 max_asize = max_osize - (VDEV_LABEL_START_SIZE +
1209 1220 VDEV_LABEL_END_SIZE);
1210 1221 } else {
1211 1222 if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1212 1223 (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1213 1224 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1214 1225 VDEV_AUX_TOO_SMALL);
1215 1226 return (SET_ERROR(EOVERFLOW));
1216 1227 }
1217 1228 psize = 0;
1218 1229 asize = osize;
1219 1230 max_asize = max_osize;
1220 1231 }
1221 1232
1222 1233 vd->vdev_psize = psize;
1223 1234
1224 1235 /*
1225 1236 * Make sure the allocatable size hasn't shrunk.
1226 1237 */
1227 1238 if (asize < vd->vdev_min_asize) {
1228 1239 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1229 1240 VDEV_AUX_BAD_LABEL);
1230 1241 return (SET_ERROR(EINVAL));
1231 1242 }
1232 1243
1233 1244 if (vd->vdev_asize == 0) {
1234 1245 /*
1235 1246 * This is the first-ever open, so use the computed values.
1236 1247 * For testing purposes, a higher ashift can be requested.
1237 1248 */
1238 1249 vd->vdev_asize = asize;
1239 1250 vd->vdev_max_asize = max_asize;
1240 1251 vd->vdev_ashift = MAX(ashift, vd->vdev_ashift);
1241 1252 } else {
1242 1253 /*
1243 1254 * Detect if the alignment requirement has increased.
1244 1255 * We don't want to make the pool unavailable, just
1245 1256 * issue a warning instead.
1246 1257 */
1247 1258 if (ashift > vd->vdev_top->vdev_ashift &&
1248 1259 vd->vdev_ops->vdev_op_leaf) {
1249 1260 cmn_err(CE_WARN,
1250 1261 "Disk, '%s', has a block alignment that is "
1251 1262 "larger than the pool's alignment\n",
1252 1263 vd->vdev_path);
1253 1264 }
1254 1265 vd->vdev_max_asize = max_asize;
1255 1266 }
1256 1267
1257 1268 /*
1258 1269 * If all children are healthy and the asize has increased,
1259 1270 * then we've experienced dynamic LUN growth. If automatic
1260 1271 * expansion is enabled then use the additional space.
1261 1272 */
1262 1273 if (vd->vdev_state == VDEV_STATE_HEALTHY && asize > vd->vdev_asize &&
1263 1274 (vd->vdev_expanding || spa->spa_autoexpand))
1264 1275 vd->vdev_asize = asize;
1265 1276
1266 1277 vdev_set_min_asize(vd);
1267 1278
1268 1279 /*
1269 1280 * Ensure we can issue some IO before declaring the
1270 1281 * vdev open for business.
1271 1282 */
1272 1283 if (vd->vdev_ops->vdev_op_leaf &&
1273 1284 (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1274 1285 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1275 1286 VDEV_AUX_ERR_EXCEEDED);
1276 1287 return (error);
1277 1288 }
1278 1289
1279 1290 /*
1280 1291 * If a leaf vdev has a DTL, and seems healthy, then kick off a
1281 1292 * resilver. But don't do this if we are doing a reopen for a scrub,
1282 1293 * since this would just restart the scrub we are already doing.
1283 1294 */
1284 1295 if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1285 1296 vdev_resilver_needed(vd, NULL, NULL))
1286 1297 spa_async_request(spa, SPA_ASYNC_RESILVER);
1287 1298
1288 1299 return (0);
1289 1300 }
1290 1301
1291 1302 /*
1292 1303 * Called once the vdevs are all opened, this routine validates the label
1293 1304 * contents. This needs to be done before vdev_load() so that we don't
1294 1305 * inadvertently do repair I/Os to the wrong device.
1295 1306 *
1296 1307 * If 'strict' is false ignore the spa guid check. This is necessary because
1297 1308 * if the machine crashed during a re-guid the new guid might have been written
1298 1309 * to all of the vdev labels, but not the cached config. The strict check
1299 1310 * will be performed when the pool is opened again using the mos config.
1300 1311 *
1301 1312 * This function will only return failure if one of the vdevs indicates that it
1302 1313 * has since been destroyed or exported. This is only possible if
1303 1314 * /etc/zfs/zpool.cache was readonly at the time. Otherwise, the vdev state
1304 1315 * will be updated but the function will return 0.
1305 1316 */
1306 1317 int
1307 1318 vdev_validate(vdev_t *vd, boolean_t strict)
1308 1319 {
1309 1320 spa_t *spa = vd->vdev_spa;
1310 1321 nvlist_t *label;
1311 1322 uint64_t guid = 0, top_guid;
1312 1323 uint64_t state;
1313 1324
1314 1325 for (int c = 0; c < vd->vdev_children; c++)
1315 1326 if (vdev_validate(vd->vdev_child[c], strict) != 0)
1316 1327 return (SET_ERROR(EBADF));
1317 1328
1318 1329 /*
1319 1330 * If the device has already failed, or was marked offline, don't do
1320 1331 * any further validation. Otherwise, label I/O will fail and we will
1321 1332 * overwrite the previous state.
1322 1333 */
1323 1334 if (vd->vdev_ops->vdev_op_leaf && vdev_readable(vd)) {
1324 1335 uint64_t aux_guid = 0;
1325 1336 nvlist_t *nvl;
1326 1337 uint64_t txg = spa_last_synced_txg(spa) != 0 ?
1327 1338 spa_last_synced_txg(spa) : -1ULL;
1328 1339
1329 1340 if ((label = vdev_label_read_config(vd, txg)) == NULL) {
1330 1341 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1331 1342 VDEV_AUX_BAD_LABEL);
1332 1343 return (0);
1333 1344 }
1334 1345
1335 1346 /*
1336 1347 * Determine if this vdev has been split off into another
1337 1348 * pool. If so, then refuse to open it.
1338 1349 */
1339 1350 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID,
1340 1351 &aux_guid) == 0 && aux_guid == spa_guid(spa)) {
1341 1352 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1342 1353 VDEV_AUX_SPLIT_POOL);
1343 1354 nvlist_free(label);
1344 1355 return (0);
1345 1356 }
1346 1357
1347 1358 if (strict && (nvlist_lookup_uint64(label,
1348 1359 ZPOOL_CONFIG_POOL_GUID, &guid) != 0 ||
1349 1360 guid != spa_guid(spa))) {
1350 1361 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1351 1362 VDEV_AUX_CORRUPT_DATA);
1352 1363 nvlist_free(label);
1353 1364 return (0);
1354 1365 }
1355 1366
1356 1367 if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl)
1357 1368 != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID,
1358 1369 &aux_guid) != 0)
1359 1370 aux_guid = 0;
1360 1371
1361 1372 /*
1362 1373 * If this vdev just became a top-level vdev because its
1363 1374 * sibling was detached, it will have adopted the parent's
1364 1375 * vdev guid -- but the label may or may not be on disk yet.
1365 1376 * Fortunately, either version of the label will have the
1366 1377 * same top guid, so if we're a top-level vdev, we can
1367 1378 * safely compare to that instead.
1368 1379 *
1369 1380 * If we split this vdev off instead, then we also check the
1370 1381 * original pool's guid. We don't want to consider the vdev
1371 1382 * corrupt if it is partway through a split operation.
1372 1383 */
1373 1384 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID,
1374 1385 &guid) != 0 ||
1375 1386 nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID,
1376 1387 &top_guid) != 0 ||
1377 1388 ((vd->vdev_guid != guid && vd->vdev_guid != aux_guid) &&
1378 1389 (vd->vdev_guid != top_guid || vd != vd->vdev_top))) {
1379 1390 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1380 1391 VDEV_AUX_CORRUPT_DATA);
1381 1392 nvlist_free(label);
1382 1393 return (0);
1383 1394 }
1384 1395
1385 1396 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1386 1397 &state) != 0) {
1387 1398 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1388 1399 VDEV_AUX_CORRUPT_DATA);
1389 1400 nvlist_free(label);
1390 1401 return (0);
1391 1402 }
1392 1403
1393 1404 nvlist_free(label);
1394 1405
1395 1406 /*
1396 1407 * If this is a verbatim import, no need to check the
1397 1408 * state of the pool.
1398 1409 */
1399 1410 if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) &&
1400 1411 spa_load_state(spa) == SPA_LOAD_OPEN &&
1401 1412 state != POOL_STATE_ACTIVE)
1402 1413 return (SET_ERROR(EBADF));
1403 1414
1404 1415 /*
1405 1416 * If we were able to open and validate a vdev that was
1406 1417 * previously marked permanently unavailable, clear that state
1407 1418 * now.
1408 1419 */
1409 1420 if (vd->vdev_not_present)
1410 1421 vd->vdev_not_present = 0;
1411 1422 }
1412 1423
1413 1424 return (0);
1414 1425 }
1415 1426
1416 1427 /*
1417 1428 * Close a virtual device.
1418 1429 */
1419 1430 void
1420 1431 vdev_close(vdev_t *vd)
1421 1432 {
1422 1433 spa_t *spa = vd->vdev_spa;
1423 1434 vdev_t *pvd = vd->vdev_parent;
1424 1435
1425 1436 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1426 1437
1427 1438 /*
1428 1439 * If our parent is reopening, then we are as well, unless we are
1429 1440 * going offline.
1430 1441 */
1431 1442 if (pvd != NULL && pvd->vdev_reopening)
1432 1443 vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline);
1433 1444
1434 1445 vd->vdev_ops->vdev_op_close(vd);
1435 1446
1436 1447 vdev_cache_purge(vd);
1437 1448
1438 1449 /*
1439 1450 * We record the previous state before we close it, so that if we are
1440 1451 * doing a reopen(), we don't generate FMA ereports if we notice that
1441 1452 * it's still faulted.
1442 1453 */
1443 1454 vd->vdev_prevstate = vd->vdev_state;
1444 1455
1445 1456 if (vd->vdev_offline)
1446 1457 vd->vdev_state = VDEV_STATE_OFFLINE;
1447 1458 else
1448 1459 vd->vdev_state = VDEV_STATE_CLOSED;
1449 1460 vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1450 1461 }
1451 1462
1452 1463 void
1453 1464 vdev_hold(vdev_t *vd)
1454 1465 {
1455 1466 spa_t *spa = vd->vdev_spa;
1456 1467
1457 1468 ASSERT(spa_is_root(spa));
1458 1469 if (spa->spa_state == POOL_STATE_UNINITIALIZED)
1459 1470 return;
1460 1471
1461 1472 for (int c = 0; c < vd->vdev_children; c++)
1462 1473 vdev_hold(vd->vdev_child[c]);
1463 1474
1464 1475 if (vd->vdev_ops->vdev_op_leaf)
1465 1476 vd->vdev_ops->vdev_op_hold(vd);
1466 1477 }
1467 1478
1468 1479 void
1469 1480 vdev_rele(vdev_t *vd)
1470 1481 {
1471 1482 spa_t *spa = vd->vdev_spa;
1472 1483
1473 1484 ASSERT(spa_is_root(spa));
1474 1485 for (int c = 0; c < vd->vdev_children; c++)
1475 1486 vdev_rele(vd->vdev_child[c]);
1476 1487
1477 1488 if (vd->vdev_ops->vdev_op_leaf)
1478 1489 vd->vdev_ops->vdev_op_rele(vd);
1479 1490 }
1480 1491
1481 1492 /*
1482 1493 * Reopen all interior vdevs and any unopened leaves. We don't actually
1483 1494 * reopen leaf vdevs which had previously been opened as they might deadlock
1484 1495 * on the spa_config_lock. Instead we only obtain the leaf's physical size.
1485 1496 * If the leaf has never been opened then open it, as usual.
1486 1497 */
1487 1498 void
1488 1499 vdev_reopen(vdev_t *vd)
1489 1500 {
1490 1501 spa_t *spa = vd->vdev_spa;
1491 1502
1492 1503 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1493 1504
1494 1505 /* set the reopening flag unless we're taking the vdev offline */
1495 1506 vd->vdev_reopening = !vd->vdev_offline;
1496 1507 vdev_close(vd);
1497 1508 (void) vdev_open(vd);
1498 1509
1499 1510 /*
1500 1511 * Call vdev_validate() here to make sure we have the same device.
1501 1512 * Otherwise, a device with an invalid label could be successfully
1502 1513 * opened in response to vdev_reopen().
1503 1514 */
1504 1515 if (vd->vdev_aux) {
1505 1516 (void) vdev_validate_aux(vd);
1506 1517 if (vdev_readable(vd) && vdev_writeable(vd) &&
1507 1518 vd->vdev_aux == &spa->spa_l2cache &&
1508 1519 !l2arc_vdev_present(vd))
1509 1520 l2arc_add_vdev(spa, vd);
1510 1521 } else {
1511 1522 (void) vdev_validate(vd, B_TRUE);
1512 1523 }
1513 1524
1514 1525 /*
1515 1526 * Reassess parent vdev's health.
1516 1527 */
1517 1528 vdev_propagate_state(vd);
1518 1529 }
1519 1530
1520 1531 int
1521 1532 vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1522 1533 {
1523 1534 int error;
1524 1535
1525 1536 /*
1526 1537 * Normally, partial opens (e.g. of a mirror) are allowed.
1527 1538 * For a create, however, we want to fail the request if
1528 1539 * there are any components we can't open.
1529 1540 */
1530 1541 error = vdev_open(vd);
1531 1542
1532 1543 if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
1533 1544 vdev_close(vd);
1534 1545 return (error ? error : ENXIO);
1535 1546 }
1536 1547
1537 1548 /*
1538 1549 * Recursively load DTLs and initialize all labels.
1539 1550 */
1540 1551 if ((error = vdev_dtl_load(vd)) != 0 ||
1541 1552 (error = vdev_label_init(vd, txg, isreplacing ?
1542 1553 VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
1543 1554 vdev_close(vd);
1544 1555 return (error);
1545 1556 }
1546 1557
1547 1558 return (0);
1548 1559 }
1549 1560
1550 1561 void
1551 1562 vdev_metaslab_set_size(vdev_t *vd)
1552 1563 {
1553 1564 /*
1554 1565 * Aim for roughly 200 metaslabs per vdev.
1555 1566 */
1556 1567 vd->vdev_ms_shift = highbit64(vd->vdev_asize / 200);
1557 1568 vd->vdev_ms_shift = MAX(vd->vdev_ms_shift, SPA_MAXBLOCKSHIFT);
1558 1569 }
1559 1570
1560 1571 void
1561 1572 vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
1562 1573 {
1563 1574 ASSERT(vd == vd->vdev_top);
1564 1575 ASSERT(!vd->vdev_ishole);
1565 1576 ASSERT(ISP2(flags));
1566 1577 ASSERT(spa_writeable(vd->vdev_spa));
1567 1578
1568 1579 if (flags & VDD_METASLAB)
1569 1580 (void) txg_list_add(&vd->vdev_ms_list, arg, txg);
1570 1581
1571 1582 if (flags & VDD_DTL)
1572 1583 (void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
1573 1584
1574 1585 (void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
1575 1586 }
1576 1587
1577 1588 void
1578 1589 vdev_dirty_leaves(vdev_t *vd, int flags, uint64_t txg)
1579 1590 {
1580 1591 for (int c = 0; c < vd->vdev_children; c++)
1581 1592 vdev_dirty_leaves(vd->vdev_child[c], flags, txg);
1582 1593
1583 1594 if (vd->vdev_ops->vdev_op_leaf)
1584 1595 vdev_dirty(vd->vdev_top, flags, vd, txg);
1585 1596 }
1586 1597
1587 1598 /*
1588 1599 * DTLs.
1589 1600 *
1590 1601 * A vdev's DTL (dirty time log) is the set of transaction groups for which
1591 1602 * the vdev has less than perfect replication. There are four kinds of DTL:
1592 1603 *
1593 1604 * DTL_MISSING: txgs for which the vdev has no valid copies of the data
1594 1605 *
1595 1606 * DTL_PARTIAL: txgs for which data is available, but not fully replicated
1596 1607 *
1597 1608 * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
1598 1609 * scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
1599 1610 * txgs that was scrubbed.
1600 1611 *
1601 1612 * DTL_OUTAGE: txgs which cannot currently be read, whether due to
1602 1613 * persistent errors or just some device being offline.
1603 1614 * Unlike the other three, the DTL_OUTAGE map is not generally
1604 1615 * maintained; it's only computed when needed, typically to
1605 1616 * determine whether a device can be detached.
1606 1617 *
1607 1618 * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
1608 1619 * either has the data or it doesn't.
1609 1620 *
1610 1621 * For interior vdevs such as mirror and RAID-Z the picture is more complex.
1611 1622 * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
1612 1623 * if any child is less than fully replicated, then so is its parent.
1613 1624 * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
1614 1625 * comprising only those txgs which appear in 'maxfaults' or more children;
1615 1626 * those are the txgs we don't have enough replication to read. For example,
1616 1627 * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
1617 1628 * thus, its DTL_MISSING consists of the set of txgs that appear in more than
1618 1629 * two child DTL_MISSING maps.
1619 1630 *
1620 1631 * It should be clear from the above that to compute the DTLs and outage maps
1621 1632 * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
1622 1633 * Therefore, that is all we keep on disk. When loading the pool, or after
1623 1634 * a configuration change, we generate all other DTLs from first principles.
1624 1635 */
1625 1636 void
1626 1637 vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1627 1638 {
1628 1639 range_tree_t *rt = vd->vdev_dtl[t];
1629 1640
1630 1641 ASSERT(t < DTL_TYPES);
1631 1642 ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1632 1643 ASSERT(spa_writeable(vd->vdev_spa));
1633 1644
1634 1645 mutex_enter(rt->rt_lock);
1635 1646 if (!range_tree_contains(rt, txg, size))
1636 1647 range_tree_add(rt, txg, size);
1637 1648 mutex_exit(rt->rt_lock);
1638 1649 }
1639 1650
1640 1651 boolean_t
1641 1652 vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1642 1653 {
1643 1654 range_tree_t *rt = vd->vdev_dtl[t];
1644 1655 boolean_t dirty = B_FALSE;
1645 1656
1646 1657 ASSERT(t < DTL_TYPES);
1647 1658 ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1648 1659
1649 1660 mutex_enter(rt->rt_lock);
1650 1661 if (range_tree_space(rt) != 0)
1651 1662 dirty = range_tree_contains(rt, txg, size);
1652 1663 mutex_exit(rt->rt_lock);
1653 1664
1654 1665 return (dirty);
1655 1666 }
1656 1667
1657 1668 boolean_t
1658 1669 vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
1659 1670 {
1660 1671 range_tree_t *rt = vd->vdev_dtl[t];
1661 1672 boolean_t empty;
1662 1673
1663 1674 mutex_enter(rt->rt_lock);
1664 1675 empty = (range_tree_space(rt) == 0);
1665 1676 mutex_exit(rt->rt_lock);
1666 1677
1667 1678 return (empty);
1668 1679 }
1669 1680
1670 1681 /*
1671 1682 * Returns the lowest txg in the DTL range.
1672 1683 */
1673 1684 static uint64_t
1674 1685 vdev_dtl_min(vdev_t *vd)
1675 1686 {
1676 1687 range_seg_t *rs;
1677 1688
1678 1689 ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
1679 1690 ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
1680 1691 ASSERT0(vd->vdev_children);
1681 1692
1682 1693 rs = avl_first(&vd->vdev_dtl[DTL_MISSING]->rt_root);
1683 1694 return (rs->rs_start - 1);
1684 1695 }
1685 1696
1686 1697 /*
1687 1698 * Returns the highest txg in the DTL.
1688 1699 */
1689 1700 static uint64_t
1690 1701 vdev_dtl_max(vdev_t *vd)
1691 1702 {
1692 1703 range_seg_t *rs;
1693 1704
1694 1705 ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
1695 1706 ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
1696 1707 ASSERT0(vd->vdev_children);
1697 1708
1698 1709 rs = avl_last(&vd->vdev_dtl[DTL_MISSING]->rt_root);
1699 1710 return (rs->rs_end);
1700 1711 }
1701 1712
1702 1713 /*
1703 1714 * Determine if a resilvering vdev should remove any DTL entries from
1704 1715 * its range. If the vdev was resilvering for the entire duration of the
1705 1716 * scan then it should excise that range from its DTLs. Otherwise, this
1706 1717 * vdev is considered partially resilvered and should leave its DTL
1707 1718 * entries intact. The comment in vdev_dtl_reassess() describes how we
1708 1719 * excise the DTLs.
1709 1720 */
1710 1721 static boolean_t
1711 1722 vdev_dtl_should_excise(vdev_t *vd)
1712 1723 {
1713 1724 spa_t *spa = vd->vdev_spa;
1714 1725 dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1715 1726
1716 1727 ASSERT0(scn->scn_phys.scn_errors);
1717 1728 ASSERT0(vd->vdev_children);
1718 1729
1719 1730 if (vd->vdev_resilver_txg == 0 ||
1720 1731 range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0)
1721 1732 return (B_TRUE);
1722 1733
1723 1734 /*
1724 1735 * When a resilver is initiated the scan will assign the scn_max_txg
1725 1736 * value to the highest txg value that exists in all DTLs. If this
1726 1737 * device's max DTL is not part of this scan (i.e. it is not in
1727 1738 * the range (scn_min_txg, scn_max_txg] then it is not eligible
1728 1739 * for excision.
1729 1740 */
1730 1741 if (vdev_dtl_max(vd) <= scn->scn_phys.scn_max_txg) {
1731 1742 ASSERT3U(scn->scn_phys.scn_min_txg, <=, vdev_dtl_min(vd));
1732 1743 ASSERT3U(scn->scn_phys.scn_min_txg, <, vd->vdev_resilver_txg);
1733 1744 ASSERT3U(vd->vdev_resilver_txg, <=, scn->scn_phys.scn_max_txg);
1734 1745 return (B_TRUE);
1735 1746 }
1736 1747 return (B_FALSE);
1737 1748 }
1738 1749
1739 1750 /*
1740 1751 * Reassess DTLs after a config change or scrub completion.
1741 1752 */
1742 1753 void
1743 1754 vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
1744 1755 {
1745 1756 spa_t *spa = vd->vdev_spa;
1746 1757 avl_tree_t reftree;
1747 1758 int minref;
1748 1759
1749 1760 ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
1750 1761
1751 1762 for (int c = 0; c < vd->vdev_children; c++)
1752 1763 vdev_dtl_reassess(vd->vdev_child[c], txg,
1753 1764 scrub_txg, scrub_done);
1754 1765
1755 1766 if (vd == spa->spa_root_vdev || vd->vdev_ishole || vd->vdev_aux)
1756 1767 return;
1757 1768
1758 1769 if (vd->vdev_ops->vdev_op_leaf) {
1759 1770 dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1760 1771
1761 1772 mutex_enter(&vd->vdev_dtl_lock);
1762 1773
1763 1774 /*
1764 1775 * If we've completed a scan cleanly then determine
1765 1776 * if this vdev should remove any DTLs. We only want to
1766 1777 * excise regions on vdevs that were available during
1767 1778 * the entire duration of this scan.
1768 1779 */
1769 1780 if (scrub_txg != 0 &&
1770 1781 (spa->spa_scrub_started ||
1771 1782 (scn != NULL && scn->scn_phys.scn_errors == 0)) &&
1772 1783 vdev_dtl_should_excise(vd)) {
1773 1784 /*
1774 1785 * We completed a scrub up to scrub_txg. If we
1775 1786 * did it without rebooting, then the scrub dtl
1776 1787 * will be valid, so excise the old region and
1777 1788 * fold in the scrub dtl. Otherwise, leave the
1778 1789 * dtl as-is if there was an error.
1779 1790 *
1780 1791 * There's little trick here: to excise the beginning
1781 1792 * of the DTL_MISSING map, we put it into a reference
1782 1793 * tree and then add a segment with refcnt -1 that
1783 1794 * covers the range [0, scrub_txg). This means
1784 1795 * that each txg in that range has refcnt -1 or 0.
1785 1796 * We then add DTL_SCRUB with a refcnt of 2, so that
1786 1797 * entries in the range [0, scrub_txg) will have a
1787 1798 * positive refcnt -- either 1 or 2. We then convert
1788 1799 * the reference tree into the new DTL_MISSING map.
1789 1800 */
1790 1801 space_reftree_create(&reftree);
1791 1802 space_reftree_add_map(&reftree,
1792 1803 vd->vdev_dtl[DTL_MISSING], 1);
1793 1804 space_reftree_add_seg(&reftree, 0, scrub_txg, -1);
1794 1805 space_reftree_add_map(&reftree,
1795 1806 vd->vdev_dtl[DTL_SCRUB], 2);
1796 1807 space_reftree_generate_map(&reftree,
1797 1808 vd->vdev_dtl[DTL_MISSING], 1);
1798 1809 space_reftree_destroy(&reftree);
1799 1810 }
1800 1811 range_tree_vacate(vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
1801 1812 range_tree_walk(vd->vdev_dtl[DTL_MISSING],
1802 1813 range_tree_add, vd->vdev_dtl[DTL_PARTIAL]);
1803 1814 if (scrub_done)
1804 1815 range_tree_vacate(vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
1805 1816 range_tree_vacate(vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
1806 1817 if (!vdev_readable(vd))
1807 1818 range_tree_add(vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
1808 1819 else
1809 1820 range_tree_walk(vd->vdev_dtl[DTL_MISSING],
1810 1821 range_tree_add, vd->vdev_dtl[DTL_OUTAGE]);
1811 1822
1812 1823 /*
1813 1824 * If the vdev was resilvering and no longer has any
1814 1825 * DTLs then reset its resilvering flag.
1815 1826 */
1816 1827 if (vd->vdev_resilver_txg != 0 &&
1817 1828 range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0 &&
1818 1829 range_tree_space(vd->vdev_dtl[DTL_OUTAGE]) == 0)
1819 1830 vd->vdev_resilver_txg = 0;
1820 1831
1821 1832 mutex_exit(&vd->vdev_dtl_lock);
1822 1833
1823 1834 if (txg != 0)
1824 1835 vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
1825 1836 return;
1826 1837 }
1827 1838
1828 1839 mutex_enter(&vd->vdev_dtl_lock);
1829 1840 for (int t = 0; t < DTL_TYPES; t++) {
1830 1841 /* account for child's outage in parent's missing map */
1831 1842 int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
1832 1843 if (t == DTL_SCRUB)
1833 1844 continue; /* leaf vdevs only */
1834 1845 if (t == DTL_PARTIAL)
1835 1846 minref = 1; /* i.e. non-zero */
1836 1847 else if (vd->vdev_nparity != 0)
1837 1848 minref = vd->vdev_nparity + 1; /* RAID-Z */
1838 1849 else
1839 1850 minref = vd->vdev_children; /* any kind of mirror */
1840 1851 space_reftree_create(&reftree);
1841 1852 for (int c = 0; c < vd->vdev_children; c++) {
1842 1853 vdev_t *cvd = vd->vdev_child[c];
1843 1854 mutex_enter(&cvd->vdev_dtl_lock);
1844 1855 space_reftree_add_map(&reftree, cvd->vdev_dtl[s], 1);
1845 1856 mutex_exit(&cvd->vdev_dtl_lock);
1846 1857 }
1847 1858 space_reftree_generate_map(&reftree, vd->vdev_dtl[t], minref);
1848 1859 space_reftree_destroy(&reftree);
1849 1860 }
1850 1861 mutex_exit(&vd->vdev_dtl_lock);
1851 1862 }
1852 1863
1853 1864 int
1854 1865 vdev_dtl_load(vdev_t *vd)
1855 1866 {
1856 1867 spa_t *spa = vd->vdev_spa;
1857 1868 objset_t *mos = spa->spa_meta_objset;
1858 1869 int error = 0;
1859 1870
1860 1871 if (vd->vdev_ops->vdev_op_leaf && vd->vdev_dtl_object != 0) {
1861 1872 ASSERT(!vd->vdev_ishole);
1862 1873
1863 1874 error = space_map_open(&vd->vdev_dtl_sm, mos,
1864 1875 vd->vdev_dtl_object, 0, -1ULL, 0, &vd->vdev_dtl_lock);
1865 1876 if (error)
1866 1877 return (error);
1867 1878 ASSERT(vd->vdev_dtl_sm != NULL);
1868 1879
1869 1880 mutex_enter(&vd->vdev_dtl_lock);
1870 1881
1871 1882 /*
1872 1883 * Now that we've opened the space_map we need to update
1873 1884 * the in-core DTL.
1874 1885 */
1875 1886 space_map_update(vd->vdev_dtl_sm);
1876 1887
1877 1888 error = space_map_load(vd->vdev_dtl_sm,
1878 1889 vd->vdev_dtl[DTL_MISSING], SM_ALLOC);
1879 1890 mutex_exit(&vd->vdev_dtl_lock);
1880 1891
1881 1892 return (error);
1882 1893 }
1883 1894
1884 1895 for (int c = 0; c < vd->vdev_children; c++) {
1885 1896 error = vdev_dtl_load(vd->vdev_child[c]);
1886 1897 if (error != 0)
1887 1898 break;
1888 1899 }
1889 1900
1890 1901 return (error);
1891 1902 }
1892 1903
1893 1904 void
1894 1905 vdev_dtl_sync(vdev_t *vd, uint64_t txg)
1895 1906 {
1896 1907 spa_t *spa = vd->vdev_spa;
1897 1908 range_tree_t *rt = vd->vdev_dtl[DTL_MISSING];
1898 1909 objset_t *mos = spa->spa_meta_objset;
1899 1910 range_tree_t *rtsync;
1900 1911 kmutex_t rtlock;
1901 1912 dmu_tx_t *tx;
1902 1913 uint64_t object = space_map_object(vd->vdev_dtl_sm);
1903 1914
1904 1915 ASSERT(!vd->vdev_ishole);
1905 1916 ASSERT(vd->vdev_ops->vdev_op_leaf);
1906 1917
1907 1918 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1908 1919
1909 1920 if (vd->vdev_detached || vd->vdev_top->vdev_removing) {
1910 1921 mutex_enter(&vd->vdev_dtl_lock);
1911 1922 space_map_free(vd->vdev_dtl_sm, tx);
1912 1923 space_map_close(vd->vdev_dtl_sm);
1913 1924 vd->vdev_dtl_sm = NULL;
1914 1925 mutex_exit(&vd->vdev_dtl_lock);
1915 1926 dmu_tx_commit(tx);
1916 1927 return;
1917 1928 }
1918 1929
1919 1930 if (vd->vdev_dtl_sm == NULL) {
1920 1931 uint64_t new_object;
1921 1932
1922 1933 new_object = space_map_alloc(mos, tx);
1923 1934 VERIFY3U(new_object, !=, 0);
1924 1935
1925 1936 VERIFY0(space_map_open(&vd->vdev_dtl_sm, mos, new_object,
1926 1937 0, -1ULL, 0, &vd->vdev_dtl_lock));
1927 1938 ASSERT(vd->vdev_dtl_sm != NULL);
1928 1939 }
1929 1940
1930 1941 mutex_init(&rtlock, NULL, MUTEX_DEFAULT, NULL);
1931 1942
1932 1943 rtsync = range_tree_create(NULL, NULL, &rtlock);
1933 1944
1934 1945 mutex_enter(&rtlock);
1935 1946
1936 1947 mutex_enter(&vd->vdev_dtl_lock);
1937 1948 range_tree_walk(rt, range_tree_add, rtsync);
1938 1949 mutex_exit(&vd->vdev_dtl_lock);
1939 1950
1940 1951 space_map_truncate(vd->vdev_dtl_sm, tx);
1941 1952 space_map_write(vd->vdev_dtl_sm, rtsync, SM_ALLOC, tx);
1942 1953 range_tree_vacate(rtsync, NULL, NULL);
1943 1954
1944 1955 range_tree_destroy(rtsync);
1945 1956
1946 1957 mutex_exit(&rtlock);
1947 1958 mutex_destroy(&rtlock);
1948 1959
1949 1960 /*
1950 1961 * If the object for the space map has changed then dirty
1951 1962 * the top level so that we update the config.
1952 1963 */
1953 1964 if (object != space_map_object(vd->vdev_dtl_sm)) {
1954 1965 zfs_dbgmsg("txg %llu, spa %s, DTL old object %llu, "
1955 1966 "new object %llu", txg, spa_name(spa), object,
1956 1967 space_map_object(vd->vdev_dtl_sm));
1957 1968 vdev_config_dirty(vd->vdev_top);
1958 1969 }
1959 1970
1960 1971 dmu_tx_commit(tx);
1961 1972
1962 1973 mutex_enter(&vd->vdev_dtl_lock);
1963 1974 space_map_update(vd->vdev_dtl_sm);
1964 1975 mutex_exit(&vd->vdev_dtl_lock);
1965 1976 }
1966 1977
1967 1978 /*
1968 1979 * Determine whether the specified vdev can be offlined/detached/removed
1969 1980 * without losing data.
1970 1981 */
1971 1982 boolean_t
1972 1983 vdev_dtl_required(vdev_t *vd)
1973 1984 {
1974 1985 spa_t *spa = vd->vdev_spa;
1975 1986 vdev_t *tvd = vd->vdev_top;
1976 1987 uint8_t cant_read = vd->vdev_cant_read;
1977 1988 boolean_t required;
1978 1989
1979 1990 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1980 1991
1981 1992 if (vd == spa->spa_root_vdev || vd == tvd)
1982 1993 return (B_TRUE);
1983 1994
1984 1995 /*
1985 1996 * Temporarily mark the device as unreadable, and then determine
1986 1997 * whether this results in any DTL outages in the top-level vdev.
1987 1998 * If not, we can safely offline/detach/remove the device.
1988 1999 */
1989 2000 vd->vdev_cant_read = B_TRUE;
1990 2001 vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1991 2002 required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
1992 2003 vd->vdev_cant_read = cant_read;
1993 2004 vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1994 2005
1995 2006 if (!required && zio_injection_enabled)
1996 2007 required = !!zio_handle_device_injection(vd, NULL, ECHILD);
1997 2008
1998 2009 return (required);
1999 2010 }
2000 2011
2001 2012 /*
2002 2013 * Determine if resilver is needed, and if so the txg range.
2003 2014 */
2004 2015 boolean_t
2005 2016 vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
2006 2017 {
2007 2018 boolean_t needed = B_FALSE;
2008 2019 uint64_t thismin = UINT64_MAX;
2009 2020 uint64_t thismax = 0;
2010 2021
2011 2022 if (vd->vdev_children == 0) {
2012 2023 mutex_enter(&vd->vdev_dtl_lock);
2013 2024 if (range_tree_space(vd->vdev_dtl[DTL_MISSING]) != 0 &&
2014 2025 vdev_writeable(vd)) {
2015 2026
2016 2027 thismin = vdev_dtl_min(vd);
2017 2028 thismax = vdev_dtl_max(vd);
2018 2029 needed = B_TRUE;
2019 2030 }
2020 2031 mutex_exit(&vd->vdev_dtl_lock);
2021 2032 } else {
2022 2033 for (int c = 0; c < vd->vdev_children; c++) {
2023 2034 vdev_t *cvd = vd->vdev_child[c];
2024 2035 uint64_t cmin, cmax;
2025 2036
2026 2037 if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
2027 2038 thismin = MIN(thismin, cmin);
2028 2039 thismax = MAX(thismax, cmax);
2029 2040 needed = B_TRUE;
2030 2041 }
2031 2042 }
2032 2043 }
2033 2044
2034 2045 if (needed && minp) {
2035 2046 *minp = thismin;
2036 2047 *maxp = thismax;
2037 2048 }
2038 2049 return (needed);
2039 2050 }
2040 2051
2041 2052 void
2042 2053 vdev_load(vdev_t *vd)
2043 2054 {
2044 2055 /*
2045 2056 * Recursively load all children.
2046 2057 */
2047 2058 for (int c = 0; c < vd->vdev_children; c++)
2048 2059 vdev_load(vd->vdev_child[c]);
2049 2060
2050 2061 /*
2051 2062 * If this is a top-level vdev, initialize its metaslabs.
2052 2063 */
2053 2064 if (vd == vd->vdev_top && !vd->vdev_ishole &&
2054 2065 (vd->vdev_ashift == 0 || vd->vdev_asize == 0 ||
2055 2066 vdev_metaslab_init(vd, 0) != 0))
2056 2067 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2057 2068 VDEV_AUX_CORRUPT_DATA);
2058 2069
2059 2070 /*
2060 2071 * If this is a leaf vdev, load its DTL.
2061 2072 */
2062 2073 if (vd->vdev_ops->vdev_op_leaf && vdev_dtl_load(vd) != 0)
2063 2074 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2064 2075 VDEV_AUX_CORRUPT_DATA);
2065 2076 }
2066 2077
2067 2078 /*
2068 2079 * The special vdev case is used for hot spares and l2cache devices. Its
2069 2080 * sole purpose it to set the vdev state for the associated vdev. To do this,
2070 2081 * we make sure that we can open the underlying device, then try to read the
2071 2082 * label, and make sure that the label is sane and that it hasn't been
2072 2083 * repurposed to another pool.
2073 2084 */
2074 2085 int
2075 2086 vdev_validate_aux(vdev_t *vd)
2076 2087 {
2077 2088 nvlist_t *label;
2078 2089 uint64_t guid, version;
2079 2090 uint64_t state;
2080 2091
2081 2092 if (!vdev_readable(vd))
2082 2093 return (0);
2083 2094
2084 2095 if ((label = vdev_label_read_config(vd, -1ULL)) == NULL) {
2085 2096 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2086 2097 VDEV_AUX_CORRUPT_DATA);
2087 2098 return (-1);
2088 2099 }
2089 2100
2090 2101 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
2091 2102 !SPA_VERSION_IS_SUPPORTED(version) ||
2092 2103 nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
2093 2104 guid != vd->vdev_guid ||
2094 2105 nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
2095 2106 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2096 2107 VDEV_AUX_CORRUPT_DATA);
2097 2108 nvlist_free(label);
2098 2109 return (-1);
2099 2110 }
2100 2111
2101 2112 /*
2102 2113 * We don't actually check the pool state here. If it's in fact in
2103 2114 * use by another pool, we update this fact on the fly when requested.
2104 2115 */
2105 2116 nvlist_free(label);
2106 2117 return (0);
2107 2118 }
2108 2119
2109 2120 void
2110 2121 vdev_remove(vdev_t *vd, uint64_t txg)
2111 2122 {
2112 2123 spa_t *spa = vd->vdev_spa;
2113 2124 objset_t *mos = spa->spa_meta_objset;
2114 2125 dmu_tx_t *tx;
2115 2126
2116 2127 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2117 2128
2118 2129 if (vd->vdev_ms != NULL) {
2119 2130 for (int m = 0; m < vd->vdev_ms_count; m++) {
2120 2131 metaslab_t *msp = vd->vdev_ms[m];
2121 2132
2122 2133 if (msp == NULL || msp->ms_sm == NULL)
2123 2134 continue;
2124 2135
2125 2136 mutex_enter(&msp->ms_lock);
2126 2137 VERIFY0(space_map_allocated(msp->ms_sm));
2127 2138 space_map_free(msp->ms_sm, tx);
2128 2139 space_map_close(msp->ms_sm);
2129 2140 msp->ms_sm = NULL;
2130 2141 mutex_exit(&msp->ms_lock);
2131 2142 }
2132 2143 }
2133 2144
2134 2145 if (vd->vdev_ms_array) {
2135 2146 (void) dmu_object_free(mos, vd->vdev_ms_array, tx);
2136 2147 vd->vdev_ms_array = 0;
2137 2148 }
2138 2149 dmu_tx_commit(tx);
2139 2150 }
2140 2151
2141 2152 void
2142 2153 vdev_sync_done(vdev_t *vd, uint64_t txg)
2143 2154 {
2144 2155 metaslab_t *msp;
2145 2156 boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg));
2146 2157
2147 2158 ASSERT(!vd->vdev_ishole);
2148 2159
2149 2160 while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
2150 2161 metaslab_sync_done(msp, txg);
2151 2162
2152 2163 if (reassess)
2153 2164 metaslab_sync_reassess(vd->vdev_mg);
2154 2165 }
2155 2166
2156 2167 void
2157 2168 vdev_sync(vdev_t *vd, uint64_t txg)
2158 2169 {
2159 2170 spa_t *spa = vd->vdev_spa;
2160 2171 vdev_t *lvd;
2161 2172 metaslab_t *msp;
2162 2173 dmu_tx_t *tx;
2163 2174
2164 2175 ASSERT(!vd->vdev_ishole);
2165 2176
2166 2177 if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0) {
2167 2178 ASSERT(vd == vd->vdev_top);
2168 2179 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2169 2180 vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
2170 2181 DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
2171 2182 ASSERT(vd->vdev_ms_array != 0);
2172 2183 vdev_config_dirty(vd);
2173 2184 dmu_tx_commit(tx);
2174 2185 }
2175 2186
2176 2187 /*
2177 2188 * Remove the metadata associated with this vdev once it's empty.
2178 2189 */
2179 2190 if (vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing)
2180 2191 vdev_remove(vd, txg);
2181 2192
2182 2193 while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
2183 2194 metaslab_sync(msp, txg);
2184 2195 (void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
2185 2196 }
2186 2197
2187 2198 while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
2188 2199 vdev_dtl_sync(lvd, txg);
2189 2200
2190 2201 (void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
2191 2202 }
2192 2203
2193 2204 uint64_t
2194 2205 vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
2195 2206 {
2196 2207 return (vd->vdev_ops->vdev_op_asize(vd, psize));
2197 2208 }
2198 2209
2199 2210 /*
2200 2211 * Mark the given vdev faulted. A faulted vdev behaves as if the device could
2201 2212 * not be opened, and no I/O is attempted.
2202 2213 */
2203 2214 int
2204 2215 vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2205 2216 {
2206 2217 vdev_t *vd, *tvd;
2207 2218
2208 2219 spa_vdev_state_enter(spa, SCL_NONE);
2209 2220
2210 2221 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2211 2222 return (spa_vdev_state_exit(spa, NULL, ENODEV));
2212 2223
2213 2224 if (!vd->vdev_ops->vdev_op_leaf)
2214 2225 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2215 2226
2216 2227 tvd = vd->vdev_top;
2217 2228
2218 2229 /*
2219 2230 * We don't directly use the aux state here, but if we do a
2220 2231 * vdev_reopen(), we need this value to be present to remember why we
2221 2232 * were faulted.
2222 2233 */
2223 2234 vd->vdev_label_aux = aux;
2224 2235
2225 2236 /*
2226 2237 * Faulted state takes precedence over degraded.
2227 2238 */
2228 2239 vd->vdev_delayed_close = B_FALSE;
2229 2240 vd->vdev_faulted = 1ULL;
2230 2241 vd->vdev_degraded = 0ULL;
2231 2242 vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
2232 2243
2233 2244 /*
2234 2245 * If this device has the only valid copy of the data, then
2235 2246 * back off and simply mark the vdev as degraded instead.
2236 2247 */
2237 2248 if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) {
2238 2249 vd->vdev_degraded = 1ULL;
2239 2250 vd->vdev_faulted = 0ULL;
2240 2251
2241 2252 /*
2242 2253 * If we reopen the device and it's not dead, only then do we
2243 2254 * mark it degraded.
2244 2255 */
2245 2256 vdev_reopen(tvd);
2246 2257
2247 2258 if (vdev_readable(vd))
2248 2259 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
2249 2260 }
2250 2261
2251 2262 return (spa_vdev_state_exit(spa, vd, 0));
2252 2263 }
2253 2264
2254 2265 /*
2255 2266 * Mark the given vdev degraded. A degraded vdev is purely an indication to the
2256 2267 * user that something is wrong. The vdev continues to operate as normal as far
2257 2268 * as I/O is concerned.
2258 2269 */
2259 2270 int
2260 2271 vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2261 2272 {
2262 2273 vdev_t *vd;
2263 2274
2264 2275 spa_vdev_state_enter(spa, SCL_NONE);
2265 2276
2266 2277 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2267 2278 return (spa_vdev_state_exit(spa, NULL, ENODEV));
2268 2279
2269 2280 if (!vd->vdev_ops->vdev_op_leaf)
2270 2281 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2271 2282
2272 2283 /*
2273 2284 * If the vdev is already faulted, then don't do anything.
2274 2285 */
2275 2286 if (vd->vdev_faulted || vd->vdev_degraded)
2276 2287 return (spa_vdev_state_exit(spa, NULL, 0));
2277 2288
2278 2289 vd->vdev_degraded = 1ULL;
2279 2290 if (!vdev_is_dead(vd))
2280 2291 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
2281 2292 aux);
2282 2293
2283 2294 return (spa_vdev_state_exit(spa, vd, 0));
2284 2295 }
2285 2296
2286 2297 /*
2287 2298 * Online the given vdev.
2288 2299 *
2289 2300 * If 'ZFS_ONLINE_UNSPARE' is set, it implies two things. First, any attached
2290 2301 * spare device should be detached when the device finishes resilvering.
2291 2302 * Second, the online should be treated like a 'test' online case, so no FMA
2292 2303 * events are generated if the device fails to open.
2293 2304 */
2294 2305 int
2295 2306 vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
2296 2307 {
2297 2308 vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
2298 2309
2299 2310 spa_vdev_state_enter(spa, SCL_NONE);
2300 2311
2301 2312 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2302 2313 return (spa_vdev_state_exit(spa, NULL, ENODEV));
2303 2314
2304 2315 if (!vd->vdev_ops->vdev_op_leaf)
2305 2316 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2306 2317
2307 2318 tvd = vd->vdev_top;
2308 2319 vd->vdev_offline = B_FALSE;
2309 2320 vd->vdev_tmpoffline = B_FALSE;
2310 2321 vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
2311 2322 vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
2312 2323
2313 2324 /* XXX - L2ARC 1.0 does not support expansion */
2314 2325 if (!vd->vdev_aux) {
2315 2326 for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2316 2327 pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
2317 2328 }
2318 2329
2319 2330 vdev_reopen(tvd);
2320 2331 vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
2321 2332
2322 2333 if (!vd->vdev_aux) {
2323 2334 for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2324 2335 pvd->vdev_expanding = B_FALSE;
2325 2336 }
2326 2337
2327 2338 if (newstate)
2328 2339 *newstate = vd->vdev_state;
2329 2340 if ((flags & ZFS_ONLINE_UNSPARE) &&
2330 2341 !vdev_is_dead(vd) && vd->vdev_parent &&
2331 2342 vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2332 2343 vd->vdev_parent->vdev_child[0] == vd)
2333 2344 vd->vdev_unspare = B_TRUE;
2334 2345
2335 2346 if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
2336 2347
2337 2348 /* XXX - L2ARC 1.0 does not support expansion */
2338 2349 if (vd->vdev_aux)
2339 2350 return (spa_vdev_state_exit(spa, vd, ENOTSUP));
2340 2351 spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
2341 2352 }
2342 2353 return (spa_vdev_state_exit(spa, vd, 0));
2343 2354 }
2344 2355
2345 2356 static int
2346 2357 vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
2347 2358 {
2348 2359 vdev_t *vd, *tvd;
2349 2360 int error = 0;
2350 2361 uint64_t generation;
2351 2362 metaslab_group_t *mg;
2352 2363
2353 2364 top:
2354 2365 spa_vdev_state_enter(spa, SCL_ALLOC);
2355 2366
2356 2367 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2357 2368 return (spa_vdev_state_exit(spa, NULL, ENODEV));
2358 2369
2359 2370 if (!vd->vdev_ops->vdev_op_leaf)
2360 2371 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2361 2372
2362 2373 tvd = vd->vdev_top;
2363 2374 mg = tvd->vdev_mg;
2364 2375 generation = spa->spa_config_generation + 1;
2365 2376
2366 2377 /*
2367 2378 * If the device isn't already offline, try to offline it.
2368 2379 */
2369 2380 if (!vd->vdev_offline) {
2370 2381 /*
2371 2382 * If this device has the only valid copy of some data,
2372 2383 * don't allow it to be offlined. Log devices are always
2373 2384 * expendable.
2374 2385 */
2375 2386 if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2376 2387 vdev_dtl_required(vd))
2377 2388 return (spa_vdev_state_exit(spa, NULL, EBUSY));
2378 2389
2379 2390 /*
2380 2391 * If the top-level is a slog and it has had allocations
2381 2392 * then proceed. We check that the vdev's metaslab group
2382 2393 * is not NULL since it's possible that we may have just
2383 2394 * added this vdev but not yet initialized its metaslabs.
2384 2395 */
2385 2396 if (tvd->vdev_islog && mg != NULL) {
2386 2397 /*
2387 2398 * Prevent any future allocations.
2388 2399 */
2389 2400 metaslab_group_passivate(mg);
2390 2401 (void) spa_vdev_state_exit(spa, vd, 0);
2391 2402
2392 2403 error = spa_offline_log(spa);
2393 2404
2394 2405 spa_vdev_state_enter(spa, SCL_ALLOC);
2395 2406
2396 2407 /*
2397 2408 * Check to see if the config has changed.
2398 2409 */
2399 2410 if (error || generation != spa->spa_config_generation) {
2400 2411 metaslab_group_activate(mg);
2401 2412 if (error)
2402 2413 return (spa_vdev_state_exit(spa,
2403 2414 vd, error));
2404 2415 (void) spa_vdev_state_exit(spa, vd, 0);
2405 2416 goto top;
2406 2417 }
2407 2418 ASSERT0(tvd->vdev_stat.vs_alloc);
2408 2419 }
2409 2420
2410 2421 /*
2411 2422 * Offline this device and reopen its top-level vdev.
2412 2423 * If the top-level vdev is a log device then just offline
2413 2424 * it. Otherwise, if this action results in the top-level
2414 2425 * vdev becoming unusable, undo it and fail the request.
2415 2426 */
2416 2427 vd->vdev_offline = B_TRUE;
2417 2428 vdev_reopen(tvd);
2418 2429
2419 2430 if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2420 2431 vdev_is_dead(tvd)) {
2421 2432 vd->vdev_offline = B_FALSE;
2422 2433 vdev_reopen(tvd);
2423 2434 return (spa_vdev_state_exit(spa, NULL, EBUSY));
2424 2435 }
2425 2436
2426 2437 /*
2427 2438 * Add the device back into the metaslab rotor so that
2428 2439 * once we online the device it's open for business.
2429 2440 */
2430 2441 if (tvd->vdev_islog && mg != NULL)
2431 2442 metaslab_group_activate(mg);
2432 2443 }
2433 2444
2434 2445 vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
2435 2446
2436 2447 return (spa_vdev_state_exit(spa, vd, 0));
2437 2448 }
2438 2449
2439 2450 int
2440 2451 vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
2441 2452 {
2442 2453 int error;
2443 2454
2444 2455 mutex_enter(&spa->spa_vdev_top_lock);
2445 2456 error = vdev_offline_locked(spa, guid, flags);
2446 2457 mutex_exit(&spa->spa_vdev_top_lock);
2447 2458
2448 2459 return (error);
2449 2460 }
2450 2461
2451 2462 /*
2452 2463 * Clear the error counts associated with this vdev. Unlike vdev_online() and
2453 2464 * vdev_offline(), we assume the spa config is locked. We also clear all
2454 2465 * children. If 'vd' is NULL, then the user wants to clear all vdevs.
2455 2466 */
2456 2467 void
2457 2468 vdev_clear(spa_t *spa, vdev_t *vd)
2458 2469 {
2459 2470 vdev_t *rvd = spa->spa_root_vdev;
2460 2471
2461 2472 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2462 2473
2463 2474 if (vd == NULL)
2464 2475 vd = rvd;
2465 2476
2466 2477 vd->vdev_stat.vs_read_errors = 0;
2467 2478 vd->vdev_stat.vs_write_errors = 0;
2468 2479 vd->vdev_stat.vs_checksum_errors = 0;
2469 2480
2470 2481 for (int c = 0; c < vd->vdev_children; c++)
2471 2482 vdev_clear(spa, vd->vdev_child[c]);
2472 2483
2473 2484 /*
2474 2485 * If we're in the FAULTED state or have experienced failed I/O, then
2475 2486 * clear the persistent state and attempt to reopen the device. We
2476 2487 * also mark the vdev config dirty, so that the new faulted state is
2477 2488 * written out to disk.
2478 2489 */
2479 2490 if (vd->vdev_faulted || vd->vdev_degraded ||
2480 2491 !vdev_readable(vd) || !vdev_writeable(vd)) {
2481 2492
2482 2493 /*
2483 2494 * When reopening in reponse to a clear event, it may be due to
2484 2495 * a fmadm repair request. In this case, if the device is
2485 2496 * still broken, we want to still post the ereport again.
2486 2497 */
2487 2498 vd->vdev_forcefault = B_TRUE;
2488 2499
2489 2500 vd->vdev_faulted = vd->vdev_degraded = 0ULL;
2490 2501 vd->vdev_cant_read = B_FALSE;
2491 2502 vd->vdev_cant_write = B_FALSE;
2492 2503
2493 2504 vdev_reopen(vd == rvd ? rvd : vd->vdev_top);
2494 2505
2495 2506 vd->vdev_forcefault = B_FALSE;
2496 2507
2497 2508 if (vd != rvd && vdev_writeable(vd->vdev_top))
2498 2509 vdev_state_dirty(vd->vdev_top);
2499 2510
2500 2511 if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
2501 2512 spa_async_request(spa, SPA_ASYNC_RESILVER);
2502 2513
2503 2514 spa_event_notify(spa, vd, ESC_ZFS_VDEV_CLEAR);
2504 2515 }
2505 2516
2506 2517 /*
2507 2518 * When clearing a FMA-diagnosed fault, we always want to
2508 2519 * unspare the device, as we assume that the original spare was
2509 2520 * done in response to the FMA fault.
2510 2521 */
2511 2522 if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
2512 2523 vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2513 2524 vd->vdev_parent->vdev_child[0] == vd)
2514 2525 vd->vdev_unspare = B_TRUE;
2515 2526 }
2516 2527
2517 2528 boolean_t
2518 2529 vdev_is_dead(vdev_t *vd)
2519 2530 {
2520 2531 /*
2521 2532 * Holes and missing devices are always considered "dead".
2522 2533 * This simplifies the code since we don't have to check for
2523 2534 * these types of devices in the various code paths.
2524 2535 * Instead we rely on the fact that we skip over dead devices
2525 2536 * before issuing I/O to them.
2526 2537 */
2527 2538 return (vd->vdev_state < VDEV_STATE_DEGRADED || vd->vdev_ishole ||
2528 2539 vd->vdev_ops == &vdev_missing_ops);
2529 2540 }
2530 2541
2531 2542 boolean_t
2532 2543 vdev_readable(vdev_t *vd)
2533 2544 {
2534 2545 return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
2535 2546 }
2536 2547
2537 2548 boolean_t
2538 2549 vdev_writeable(vdev_t *vd)
2539 2550 {
2540 2551 return (!vdev_is_dead(vd) && !vd->vdev_cant_write);
2541 2552 }
2542 2553
2543 2554 boolean_t
2544 2555 vdev_allocatable(vdev_t *vd)
2545 2556 {
2546 2557 uint64_t state = vd->vdev_state;
2547 2558
2548 2559 /*
2549 2560 * We currently allow allocations from vdevs which may be in the
2550 2561 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
2551 2562 * fails to reopen then we'll catch it later when we're holding
2552 2563 * the proper locks. Note that we have to get the vdev state
2553 2564 * in a local variable because although it changes atomically,
2554 2565 * we're asking two separate questions about it.
2555 2566 */
2556 2567 return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
2557 2568 !vd->vdev_cant_write && !vd->vdev_ishole);
2558 2569 }
2559 2570
2560 2571 boolean_t
2561 2572 vdev_accessible(vdev_t *vd, zio_t *zio)
2562 2573 {
2563 2574 ASSERT(zio->io_vd == vd);
2564 2575
2565 2576 if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
2566 2577 return (B_FALSE);
2567 2578
2568 2579 if (zio->io_type == ZIO_TYPE_READ)
2569 2580 return (!vd->vdev_cant_read);
2570 2581
2571 2582 if (zio->io_type == ZIO_TYPE_WRITE)
2572 2583 return (!vd->vdev_cant_write);
2573 2584
2574 2585 return (B_TRUE);
2575 2586 }
2576 2587
2577 2588 /*
2578 2589 * Get statistics for the given vdev.
2579 2590 */
2580 2591 void
2581 2592 vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
2582 2593 {
2583 2594 vdev_t *rvd = vd->vdev_spa->spa_root_vdev;
2584 2595
2585 2596 mutex_enter(&vd->vdev_stat_lock);
2586 2597 bcopy(&vd->vdev_stat, vs, sizeof (*vs));
2587 2598 vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
2588 2599 vs->vs_state = vd->vdev_state;
2589 2600 vs->vs_rsize = vdev_get_min_asize(vd);
2590 2601 if (vd->vdev_ops->vdev_op_leaf)
2591 2602 vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
2592 2603 vs->vs_esize = vd->vdev_max_asize - vd->vdev_asize;
2593 2604 mutex_exit(&vd->vdev_stat_lock);
2594 2605
2595 2606 /*
2596 2607 * If we're getting stats on the root vdev, aggregate the I/O counts
2597 2608 * over all top-level vdevs (i.e. the direct children of the root).
2598 2609 */
2599 2610 if (vd == rvd) {
2600 2611 for (int c = 0; c < rvd->vdev_children; c++) {
2601 2612 vdev_t *cvd = rvd->vdev_child[c];
2602 2613 vdev_stat_t *cvs = &cvd->vdev_stat;
2603 2614
2604 2615 mutex_enter(&vd->vdev_stat_lock);
2605 2616 for (int t = 0; t < ZIO_TYPES; t++) {
2606 2617 vs->vs_ops[t] += cvs->vs_ops[t];
2607 2618 vs->vs_bytes[t] += cvs->vs_bytes[t];
2608 2619 }
2609 2620 cvs->vs_scan_removing = cvd->vdev_removing;
2610 2621 mutex_exit(&vd->vdev_stat_lock);
2611 2622 }
2612 2623 }
2613 2624 }
2614 2625
2615 2626 void
2616 2627 vdev_clear_stats(vdev_t *vd)
2617 2628 {
2618 2629 mutex_enter(&vd->vdev_stat_lock);
2619 2630 vd->vdev_stat.vs_space = 0;
2620 2631 vd->vdev_stat.vs_dspace = 0;
2621 2632 vd->vdev_stat.vs_alloc = 0;
2622 2633 mutex_exit(&vd->vdev_stat_lock);
2623 2634 }
2624 2635
2625 2636 void
2626 2637 vdev_scan_stat_init(vdev_t *vd)
2627 2638 {
2628 2639 vdev_stat_t *vs = &vd->vdev_stat;
2629 2640
2630 2641 for (int c = 0; c < vd->vdev_children; c++)
2631 2642 vdev_scan_stat_init(vd->vdev_child[c]);
2632 2643
2633 2644 mutex_enter(&vd->vdev_stat_lock);
2634 2645 vs->vs_scan_processed = 0;
2635 2646 mutex_exit(&vd->vdev_stat_lock);
2636 2647 }
2637 2648
2638 2649 void
2639 2650 vdev_stat_update(zio_t *zio, uint64_t psize)
2640 2651 {
2641 2652 spa_t *spa = zio->io_spa;
2642 2653 vdev_t *rvd = spa->spa_root_vdev;
2643 2654 vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
2644 2655 vdev_t *pvd;
2645 2656 uint64_t txg = zio->io_txg;
2646 2657 vdev_stat_t *vs = &vd->vdev_stat;
2647 2658 zio_type_t type = zio->io_type;
2648 2659 int flags = zio->io_flags;
2649 2660
2650 2661 /*
2651 2662 * If this i/o is a gang leader, it didn't do any actual work.
2652 2663 */
2653 2664 if (zio->io_gang_tree)
2654 2665 return;
2655 2666
2656 2667 if (zio->io_error == 0) {
2657 2668 /*
2658 2669 * If this is a root i/o, don't count it -- we've already
2659 2670 * counted the top-level vdevs, and vdev_get_stats() will
2660 2671 * aggregate them when asked. This reduces contention on
2661 2672 * the root vdev_stat_lock and implicitly handles blocks
2662 2673 * that compress away to holes, for which there is no i/o.
2663 2674 * (Holes never create vdev children, so all the counters
2664 2675 * remain zero, which is what we want.)
2665 2676 *
2666 2677 * Note: this only applies to successful i/o (io_error == 0)
2667 2678 * because unlike i/o counts, errors are not additive.
2668 2679 * When reading a ditto block, for example, failure of
2669 2680 * one top-level vdev does not imply a root-level error.
2670 2681 */
2671 2682 if (vd == rvd)
2672 2683 return;
2673 2684
2674 2685 ASSERT(vd == zio->io_vd);
2675 2686
2676 2687 if (flags & ZIO_FLAG_IO_BYPASS)
2677 2688 return;
2678 2689
2679 2690 mutex_enter(&vd->vdev_stat_lock);
2680 2691
2681 2692 if (flags & ZIO_FLAG_IO_REPAIR) {
2682 2693 if (flags & ZIO_FLAG_SCAN_THREAD) {
2683 2694 dsl_scan_phys_t *scn_phys =
2684 2695 &spa->spa_dsl_pool->dp_scan->scn_phys;
2685 2696 uint64_t *processed = &scn_phys->scn_processed;
2686 2697
2687 2698 /* XXX cleanup? */
2688 2699 if (vd->vdev_ops->vdev_op_leaf)
2689 2700 atomic_add_64(processed, psize);
2690 2701 vs->vs_scan_processed += psize;
2691 2702 }
2692 2703
2693 2704 if (flags & ZIO_FLAG_SELF_HEAL)
2694 2705 vs->vs_self_healed += psize;
2695 2706 }
2696 2707
2697 2708 vs->vs_ops[type]++;
2698 2709 vs->vs_bytes[type] += psize;
2699 2710
2700 2711 mutex_exit(&vd->vdev_stat_lock);
2701 2712 return;
2702 2713 }
2703 2714
2704 2715 if (flags & ZIO_FLAG_SPECULATIVE)
2705 2716 return;
2706 2717
2707 2718 /*
2708 2719 * If this is an I/O error that is going to be retried, then ignore the
2709 2720 * error. Otherwise, the user may interpret B_FAILFAST I/O errors as
2710 2721 * hard errors, when in reality they can happen for any number of
2711 2722 * innocuous reasons (bus resets, MPxIO link failure, etc).
2712 2723 */
2713 2724 if (zio->io_error == EIO &&
2714 2725 !(zio->io_flags & ZIO_FLAG_IO_RETRY))
2715 2726 return;
2716 2727
2717 2728 /*
2718 2729 * Intent logs writes won't propagate their error to the root
2719 2730 * I/O so don't mark these types of failures as pool-level
2720 2731 * errors.
2721 2732 */
2722 2733 if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
2723 2734 return;
2724 2735
2725 2736 mutex_enter(&vd->vdev_stat_lock);
2726 2737 if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
2727 2738 if (zio->io_error == ECKSUM)
2728 2739 vs->vs_checksum_errors++;
2729 2740 else
2730 2741 vs->vs_read_errors++;
2731 2742 }
2732 2743 if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
2733 2744 vs->vs_write_errors++;
2734 2745 mutex_exit(&vd->vdev_stat_lock);
2735 2746
2736 2747 if (type == ZIO_TYPE_WRITE && txg != 0 &&
2737 2748 (!(flags & ZIO_FLAG_IO_REPAIR) ||
2738 2749 (flags & ZIO_FLAG_SCAN_THREAD) ||
2739 2750 spa->spa_claiming)) {
2740 2751 /*
2741 2752 * This is either a normal write (not a repair), or it's
2742 2753 * a repair induced by the scrub thread, or it's a repair
2743 2754 * made by zil_claim() during spa_load() in the first txg.
2744 2755 * In the normal case, we commit the DTL change in the same
2745 2756 * txg as the block was born. In the scrub-induced repair
2746 2757 * case, we know that scrubs run in first-pass syncing context,
2747 2758 * so we commit the DTL change in spa_syncing_txg(spa).
2748 2759 * In the zil_claim() case, we commit in spa_first_txg(spa).
2749 2760 *
2750 2761 * We currently do not make DTL entries for failed spontaneous
2751 2762 * self-healing writes triggered by normal (non-scrubbing)
2752 2763 * reads, because we have no transactional context in which to
2753 2764 * do so -- and it's not clear that it'd be desirable anyway.
2754 2765 */
2755 2766 if (vd->vdev_ops->vdev_op_leaf) {
2756 2767 uint64_t commit_txg = txg;
2757 2768 if (flags & ZIO_FLAG_SCAN_THREAD) {
2758 2769 ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2759 2770 ASSERT(spa_sync_pass(spa) == 1);
2760 2771 vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
2761 2772 commit_txg = spa_syncing_txg(spa);
2762 2773 } else if (spa->spa_claiming) {
2763 2774 ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2764 2775 commit_txg = spa_first_txg(spa);
2765 2776 }
2766 2777 ASSERT(commit_txg >= spa_syncing_txg(spa));
2767 2778 if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
2768 2779 return;
2769 2780 for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2770 2781 vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
2771 2782 vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
2772 2783 }
2773 2784 if (vd != rvd)
2774 2785 vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
2775 2786 }
2776 2787 }
2777 2788
2778 2789 /*
2779 2790 * Update the in-core space usage stats for this vdev, its metaslab class,
2780 2791 * and the root vdev.
2781 2792 */
2782 2793 void
2783 2794 vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
2784 2795 int64_t space_delta)
2785 2796 {
2786 2797 int64_t dspace_delta = space_delta;
2787 2798 spa_t *spa = vd->vdev_spa;
2788 2799 vdev_t *rvd = spa->spa_root_vdev;
2789 2800 metaslab_group_t *mg = vd->vdev_mg;
2790 2801 metaslab_class_t *mc = mg ? mg->mg_class : NULL;
2791 2802
2792 2803 ASSERT(vd == vd->vdev_top);
2793 2804
2794 2805 /*
2795 2806 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
2796 2807 * factor. We must calculate this here and not at the root vdev
2797 2808 * because the root vdev's psize-to-asize is simply the max of its
2798 2809 * childrens', thus not accurate enough for us.
2799 2810 */
2800 2811 ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
2801 2812 ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
2802 2813 dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
2803 2814 vd->vdev_deflate_ratio;
2804 2815
2805 2816 mutex_enter(&vd->vdev_stat_lock);
2806 2817 vd->vdev_stat.vs_alloc += alloc_delta;
2807 2818 vd->vdev_stat.vs_space += space_delta;
2808 2819 vd->vdev_stat.vs_dspace += dspace_delta;
2809 2820 mutex_exit(&vd->vdev_stat_lock);
2810 2821
2811 2822 if (mc == spa_normal_class(spa)) {
2812 2823 mutex_enter(&rvd->vdev_stat_lock);
2813 2824 rvd->vdev_stat.vs_alloc += alloc_delta;
2814 2825 rvd->vdev_stat.vs_space += space_delta;
2815 2826 rvd->vdev_stat.vs_dspace += dspace_delta;
2816 2827 mutex_exit(&rvd->vdev_stat_lock);
2817 2828 }
2818 2829
2819 2830 if (mc != NULL) {
2820 2831 ASSERT(rvd == vd->vdev_parent);
2821 2832 ASSERT(vd->vdev_ms_count != 0);
2822 2833
2823 2834 metaslab_class_space_update(mc,
2824 2835 alloc_delta, defer_delta, space_delta, dspace_delta);
2825 2836 }
2826 2837 }
2827 2838
2828 2839 /*
2829 2840 * Mark a top-level vdev's config as dirty, placing it on the dirty list
2830 2841 * so that it will be written out next time the vdev configuration is synced.
2831 2842 * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
2832 2843 */
2833 2844 void
2834 2845 vdev_config_dirty(vdev_t *vd)
2835 2846 {
2836 2847 spa_t *spa = vd->vdev_spa;
2837 2848 vdev_t *rvd = spa->spa_root_vdev;
2838 2849 int c;
2839 2850
2840 2851 ASSERT(spa_writeable(spa));
2841 2852
2842 2853 /*
2843 2854 * If this is an aux vdev (as with l2cache and spare devices), then we
2844 2855 * update the vdev config manually and set the sync flag.
2845 2856 */
2846 2857 if (vd->vdev_aux != NULL) {
2847 2858 spa_aux_vdev_t *sav = vd->vdev_aux;
2848 2859 nvlist_t **aux;
2849 2860 uint_t naux;
2850 2861
2851 2862 for (c = 0; c < sav->sav_count; c++) {
2852 2863 if (sav->sav_vdevs[c] == vd)
2853 2864 break;
2854 2865 }
2855 2866
2856 2867 if (c == sav->sav_count) {
2857 2868 /*
2858 2869 * We're being removed. There's nothing more to do.
2859 2870 */
2860 2871 ASSERT(sav->sav_sync == B_TRUE);
2861 2872 return;
2862 2873 }
2863 2874
2864 2875 sav->sav_sync = B_TRUE;
2865 2876
2866 2877 if (nvlist_lookup_nvlist_array(sav->sav_config,
2867 2878 ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
2868 2879 VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
2869 2880 ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
2870 2881 }
2871 2882
2872 2883 ASSERT(c < naux);
2873 2884
2874 2885 /*
2875 2886 * Setting the nvlist in the middle if the array is a little
2876 2887 * sketchy, but it will work.
2877 2888 */
2878 2889 nvlist_free(aux[c]);
2879 2890 aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0);
2880 2891
2881 2892 return;
2882 2893 }
2883 2894
2884 2895 /*
2885 2896 * The dirty list is protected by the SCL_CONFIG lock. The caller
2886 2897 * must either hold SCL_CONFIG as writer, or must be the sync thread
2887 2898 * (which holds SCL_CONFIG as reader). There's only one sync thread,
2888 2899 * so this is sufficient to ensure mutual exclusion.
2889 2900 */
2890 2901 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2891 2902 (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2892 2903 spa_config_held(spa, SCL_CONFIG, RW_READER)));
2893 2904
2894 2905 if (vd == rvd) {
2895 2906 for (c = 0; c < rvd->vdev_children; c++)
2896 2907 vdev_config_dirty(rvd->vdev_child[c]);
2897 2908 } else {
2898 2909 ASSERT(vd == vd->vdev_top);
2899 2910
2900 2911 if (!list_link_active(&vd->vdev_config_dirty_node) &&
2901 2912 !vd->vdev_ishole)
2902 2913 list_insert_head(&spa->spa_config_dirty_list, vd);
2903 2914 }
2904 2915 }
2905 2916
2906 2917 void
2907 2918 vdev_config_clean(vdev_t *vd)
2908 2919 {
2909 2920 spa_t *spa = vd->vdev_spa;
2910 2921
2911 2922 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2912 2923 (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2913 2924 spa_config_held(spa, SCL_CONFIG, RW_READER)));
2914 2925
2915 2926 ASSERT(list_link_active(&vd->vdev_config_dirty_node));
2916 2927 list_remove(&spa->spa_config_dirty_list, vd);
2917 2928 }
2918 2929
2919 2930 /*
2920 2931 * Mark a top-level vdev's state as dirty, so that the next pass of
2921 2932 * spa_sync() can convert this into vdev_config_dirty(). We distinguish
2922 2933 * the state changes from larger config changes because they require
2923 2934 * much less locking, and are often needed for administrative actions.
2924 2935 */
2925 2936 void
2926 2937 vdev_state_dirty(vdev_t *vd)
2927 2938 {
2928 2939 spa_t *spa = vd->vdev_spa;
2929 2940
2930 2941 ASSERT(spa_writeable(spa));
2931 2942 ASSERT(vd == vd->vdev_top);
2932 2943
2933 2944 /*
2934 2945 * The state list is protected by the SCL_STATE lock. The caller
2935 2946 * must either hold SCL_STATE as writer, or must be the sync thread
2936 2947 * (which holds SCL_STATE as reader). There's only one sync thread,
2937 2948 * so this is sufficient to ensure mutual exclusion.
2938 2949 */
2939 2950 ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2940 2951 (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2941 2952 spa_config_held(spa, SCL_STATE, RW_READER)));
2942 2953
2943 2954 if (!list_link_active(&vd->vdev_state_dirty_node) && !vd->vdev_ishole)
2944 2955 list_insert_head(&spa->spa_state_dirty_list, vd);
2945 2956 }
2946 2957
2947 2958 void
2948 2959 vdev_state_clean(vdev_t *vd)
2949 2960 {
2950 2961 spa_t *spa = vd->vdev_spa;
2951 2962
2952 2963 ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2953 2964 (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2954 2965 spa_config_held(spa, SCL_STATE, RW_READER)));
2955 2966
2956 2967 ASSERT(list_link_active(&vd->vdev_state_dirty_node));
2957 2968 list_remove(&spa->spa_state_dirty_list, vd);
2958 2969 }
2959 2970
2960 2971 /*
2961 2972 * Propagate vdev state up from children to parent.
2962 2973 */
2963 2974 void
2964 2975 vdev_propagate_state(vdev_t *vd)
2965 2976 {
2966 2977 spa_t *spa = vd->vdev_spa;
2967 2978 vdev_t *rvd = spa->spa_root_vdev;
2968 2979 int degraded = 0, faulted = 0;
2969 2980 int corrupted = 0;
2970 2981 vdev_t *child;
2971 2982
2972 2983 if (vd->vdev_children > 0) {
2973 2984 for (int c = 0; c < vd->vdev_children; c++) {
2974 2985 child = vd->vdev_child[c];
2975 2986
2976 2987 /*
2977 2988 * Don't factor holes into the decision.
2978 2989 */
2979 2990 if (child->vdev_ishole)
2980 2991 continue;
2981 2992
2982 2993 if (!vdev_readable(child) ||
2983 2994 (!vdev_writeable(child) && spa_writeable(spa))) {
2984 2995 /*
2985 2996 * Root special: if there is a top-level log
2986 2997 * device, treat the root vdev as if it were
2987 2998 * degraded.
2988 2999 */
2989 3000 if (child->vdev_islog && vd == rvd)
2990 3001 degraded++;
2991 3002 else
2992 3003 faulted++;
2993 3004 } else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
2994 3005 degraded++;
2995 3006 }
2996 3007
2997 3008 if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
2998 3009 corrupted++;
2999 3010 }
3000 3011
3001 3012 vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
3002 3013
3003 3014 /*
3004 3015 * Root special: if there is a top-level vdev that cannot be
3005 3016 * opened due to corrupted metadata, then propagate the root
3006 3017 * vdev's aux state as 'corrupt' rather than 'insufficient
3007 3018 * replicas'.
3008 3019 */
3009 3020 if (corrupted && vd == rvd &&
3010 3021 rvd->vdev_state == VDEV_STATE_CANT_OPEN)
3011 3022 vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
3012 3023 VDEV_AUX_CORRUPT_DATA);
3013 3024 }
3014 3025
3015 3026 if (vd->vdev_parent)
3016 3027 vdev_propagate_state(vd->vdev_parent);
3017 3028 }
3018 3029
3019 3030 /*
3020 3031 * Set a vdev's state. If this is during an open, we don't update the parent
3021 3032 * state, because we're in the process of opening children depth-first.
3022 3033 * Otherwise, we propagate the change to the parent.
3023 3034 *
3024 3035 * If this routine places a device in a faulted state, an appropriate ereport is
3025 3036 * generated.
3026 3037 */
3027 3038 void
3028 3039 vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
3029 3040 {
3030 3041 uint64_t save_state;
3031 3042 spa_t *spa = vd->vdev_spa;
3032 3043
3033 3044 if (state == vd->vdev_state) {
3034 3045 vd->vdev_stat.vs_aux = aux;
3035 3046 return;
3036 3047 }
3037 3048
3038 3049 save_state = vd->vdev_state;
3039 3050
3040 3051 vd->vdev_state = state;
3041 3052 vd->vdev_stat.vs_aux = aux;
3042 3053
3043 3054 /*
3044 3055 * If we are setting the vdev state to anything but an open state, then
3045 3056 * always close the underlying device unless the device has requested
3046 3057 * a delayed close (i.e. we're about to remove or fault the device).
3047 3058 * Otherwise, we keep accessible but invalid devices open forever.
3048 3059 * We don't call vdev_close() itself, because that implies some extra
3049 3060 * checks (offline, etc) that we don't want here. This is limited to
3050 3061 * leaf devices, because otherwise closing the device will affect other
3051 3062 * children.
3052 3063 */
3053 3064 if (!vd->vdev_delayed_close && vdev_is_dead(vd) &&
3054 3065 vd->vdev_ops->vdev_op_leaf)
3055 3066 vd->vdev_ops->vdev_op_close(vd);
3056 3067
3057 3068 /*
3058 3069 * If we have brought this vdev back into service, we need
3059 3070 * to notify fmd so that it can gracefully repair any outstanding
3060 3071 * cases due to a missing device. We do this in all cases, even those
3061 3072 * that probably don't correlate to a repaired fault. This is sure to
3062 3073 * catch all cases, and we let the zfs-retire agent sort it out. If
3063 3074 * this is a transient state it's OK, as the retire agent will
3064 3075 * double-check the state of the vdev before repairing it.
3065 3076 */
3066 3077 if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf &&
3067 3078 vd->vdev_prevstate != state)
3068 3079 zfs_post_state_change(spa, vd);
3069 3080
3070 3081 if (vd->vdev_removed &&
3071 3082 state == VDEV_STATE_CANT_OPEN &&
3072 3083 (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
3073 3084 /*
3074 3085 * If the previous state is set to VDEV_STATE_REMOVED, then this
3075 3086 * device was previously marked removed and someone attempted to
3076 3087 * reopen it. If this failed due to a nonexistent device, then
3077 3088 * keep the device in the REMOVED state. We also let this be if
3078 3089 * it is one of our special test online cases, which is only
3079 3090 * attempting to online the device and shouldn't generate an FMA
3080 3091 * fault.
3081 3092 */
3082 3093 vd->vdev_state = VDEV_STATE_REMOVED;
3083 3094 vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
3084 3095 } else if (state == VDEV_STATE_REMOVED) {
3085 3096 vd->vdev_removed = B_TRUE;
3086 3097 } else if (state == VDEV_STATE_CANT_OPEN) {
3087 3098 /*
3088 3099 * If we fail to open a vdev during an import or recovery, we
3089 3100 * mark it as "not available", which signifies that it was
3090 3101 * never there to begin with. Failure to open such a device
3091 3102 * is not considered an error.
3092 3103 */
3093 3104 if ((spa_load_state(spa) == SPA_LOAD_IMPORT ||
3094 3105 spa_load_state(spa) == SPA_LOAD_RECOVER) &&
3095 3106 vd->vdev_ops->vdev_op_leaf)
3096 3107 vd->vdev_not_present = 1;
3097 3108
3098 3109 /*
3099 3110 * Post the appropriate ereport. If the 'prevstate' field is
3100 3111 * set to something other than VDEV_STATE_UNKNOWN, it indicates
3101 3112 * that this is part of a vdev_reopen(). In this case, we don't
3102 3113 * want to post the ereport if the device was already in the
3103 3114 * CANT_OPEN state beforehand.
3104 3115 *
3105 3116 * If the 'checkremove' flag is set, then this is an attempt to
3106 3117 * online the device in response to an insertion event. If we
3107 3118 * hit this case, then we have detected an insertion event for a
3108 3119 * faulted or offline device that wasn't in the removed state.
3109 3120 * In this scenario, we don't post an ereport because we are
3110 3121 * about to replace the device, or attempt an online with
3111 3122 * vdev_forcefault, which will generate the fault for us.
3112 3123 */
3113 3124 if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
3114 3125 !vd->vdev_not_present && !vd->vdev_checkremove &&
3115 3126 vd != spa->spa_root_vdev) {
3116 3127 const char *class;
3117 3128
3118 3129 switch (aux) {
3119 3130 case VDEV_AUX_OPEN_FAILED:
3120 3131 class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
3121 3132 break;
3122 3133 case VDEV_AUX_CORRUPT_DATA:
3123 3134 class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
3124 3135 break;
3125 3136 case VDEV_AUX_NO_REPLICAS:
3126 3137 class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
3127 3138 break;
3128 3139 case VDEV_AUX_BAD_GUID_SUM:
3129 3140 class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
3130 3141 break;
3131 3142 case VDEV_AUX_TOO_SMALL:
3132 3143 class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
3133 3144 break;
3134 3145 case VDEV_AUX_BAD_LABEL:
3135 3146 class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
3136 3147 break;
3137 3148 default:
3138 3149 class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
3139 3150 }
3140 3151
3141 3152 zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
3142 3153 }
3143 3154
3144 3155 /* Erase any notion of persistent removed state */
3145 3156 vd->vdev_removed = B_FALSE;
3146 3157 } else {
3147 3158 vd->vdev_removed = B_FALSE;
3148 3159 }
3149 3160
3150 3161 if (!isopen && vd->vdev_parent)
3151 3162 vdev_propagate_state(vd->vdev_parent);
3152 3163 }
3153 3164
3154 3165 /*
3155 3166 * Check the vdev configuration to ensure that it's capable of supporting
3156 3167 * a root pool. Currently, we do not support RAID-Z or partial configuration.
3157 3168 * In addition, only a single top-level vdev is allowed and none of the leaves
3158 3169 * can be wholedisks.
3159 3170 */
3160 3171 boolean_t
3161 3172 vdev_is_bootable(vdev_t *vd)
3162 3173 {
3163 3174 if (!vd->vdev_ops->vdev_op_leaf) {
3164 3175 char *vdev_type = vd->vdev_ops->vdev_op_type;
3165 3176
3166 3177 if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
3167 3178 vd->vdev_children > 1) {
3168 3179 return (B_FALSE);
3169 3180 } else if (strcmp(vdev_type, VDEV_TYPE_RAIDZ) == 0 ||
3170 3181 strcmp(vdev_type, VDEV_TYPE_MISSING) == 0) {
3171 3182 return (B_FALSE);
3172 3183 }
3173 3184 } else if (vd->vdev_wholedisk == 1) {
3174 3185 return (B_FALSE);
3175 3186 }
3176 3187
3177 3188 for (int c = 0; c < vd->vdev_children; c++) {
3178 3189 if (!vdev_is_bootable(vd->vdev_child[c]))
3179 3190 return (B_FALSE);
3180 3191 }
3181 3192 return (B_TRUE);
3182 3193 }
3183 3194
3184 3195 /*
3185 3196 * Load the state from the original vdev tree (ovd) which
3186 3197 * we've retrieved from the MOS config object. If the original
3187 3198 * vdev was offline or faulted then we transfer that state to the
3188 3199 * device in the current vdev tree (nvd).
3189 3200 */
3190 3201 void
3191 3202 vdev_load_log_state(vdev_t *nvd, vdev_t *ovd)
3192 3203 {
3193 3204 spa_t *spa = nvd->vdev_spa;
3194 3205
3195 3206 ASSERT(nvd->vdev_top->vdev_islog);
3196 3207 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3197 3208 ASSERT3U(nvd->vdev_guid, ==, ovd->vdev_guid);
3198 3209
3199 3210 for (int c = 0; c < nvd->vdev_children; c++)
3200 3211 vdev_load_log_state(nvd->vdev_child[c], ovd->vdev_child[c]);
3201 3212
3202 3213 if (nvd->vdev_ops->vdev_op_leaf) {
3203 3214 /*
3204 3215 * Restore the persistent vdev state
3205 3216 */
3206 3217 nvd->vdev_offline = ovd->vdev_offline;
3207 3218 nvd->vdev_faulted = ovd->vdev_faulted;
3208 3219 nvd->vdev_degraded = ovd->vdev_degraded;
3209 3220 nvd->vdev_removed = ovd->vdev_removed;
3210 3221 }
3211 3222 }
3212 3223
3213 3224 /*
3214 3225 * Determine if a log device has valid content. If the vdev was
3215 3226 * removed or faulted in the MOS config then we know that
3216 3227 * the content on the log device has already been written to the pool.
3217 3228 */
3218 3229 boolean_t
3219 3230 vdev_log_state_valid(vdev_t *vd)
3220 3231 {
3221 3232 if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted &&
3222 3233 !vd->vdev_removed)
3223 3234 return (B_TRUE);
3224 3235
3225 3236 for (int c = 0; c < vd->vdev_children; c++)
3226 3237 if (vdev_log_state_valid(vd->vdev_child[c]))
3227 3238 return (B_TRUE);
3228 3239
3229 3240 return (B_FALSE);
3230 3241 }
3231 3242
3232 3243 /*
3233 3244 * Expand a vdev if possible.
3234 3245 */
3235 3246 void
3236 3247 vdev_expand(vdev_t *vd, uint64_t txg)
3237 3248 {
3238 3249 ASSERT(vd->vdev_top == vd);
3239 3250 ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
3240 3251
3241 3252 if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count) {
3242 3253 VERIFY(vdev_metaslab_init(vd, txg) == 0);
3243 3254 vdev_config_dirty(vd);
3244 3255 }
3245 3256 }
3246 3257
3247 3258 /*
3248 3259 * Split a vdev.
3249 3260 */
3250 3261 void
3251 3262 vdev_split(vdev_t *vd)
3252 3263 {
3253 3264 vdev_t *cvd, *pvd = vd->vdev_parent;
3254 3265
3255 3266 vdev_remove_child(pvd, vd);
3256 3267 vdev_compact_children(pvd);
3257 3268
3258 3269 cvd = pvd->vdev_child[0];
3259 3270 if (pvd->vdev_children == 1) {
3260 3271 vdev_remove_parent(cvd);
3261 3272 cvd->vdev_splitting = B_TRUE;
3262 3273 }
3263 3274 vdev_propagate_state(cvd);
3264 3275 }
3265 3276
3266 3277 void
3267 3278 vdev_deadman(vdev_t *vd)
3268 3279 {
3269 3280 for (int c = 0; c < vd->vdev_children; c++) {
3270 3281 vdev_t *cvd = vd->vdev_child[c];
3271 3282
3272 3283 vdev_deadman(cvd);
3273 3284 }
3274 3285
3275 3286 if (vd->vdev_ops->vdev_op_leaf) {
3276 3287 vdev_queue_t *vq = &vd->vdev_queue;
3277 3288
3278 3289 mutex_enter(&vq->vq_lock);
3279 3290 if (avl_numnodes(&vq->vq_active_tree) > 0) {
3280 3291 spa_t *spa = vd->vdev_spa;
3281 3292 zio_t *fio;
3282 3293 uint64_t delta;
3283 3294
3284 3295 /*
3285 3296 * Look at the head of all the pending queues,
3286 3297 * if any I/O has been outstanding for longer than
3287 3298 * the spa_deadman_synctime we panic the system.
3288 3299 */
3289 3300 fio = avl_first(&vq->vq_active_tree);
3290 3301 delta = gethrtime() - fio->io_timestamp;
3291 3302 if (delta > spa_deadman_synctime(spa)) {
3292 3303 zfs_dbgmsg("SLOW IO: zio timestamp %lluns, "
3293 3304 "delta %lluns, last io %lluns",
3294 3305 fio->io_timestamp, delta,
3295 3306 vq->vq_io_complete_ts);
3296 3307 fm_panic("I/O to pool '%s' appears to be "
3297 3308 "hung.", spa_name(spa));
3298 3309 }
3299 3310 }
3300 3311 mutex_exit(&vq->vq_lock);
3301 3312 }
3302 3313 }
|
↓ open down ↓ |
2503 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX