Print this page
3525 Persistent L2ARC
| 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) 2013 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 /*
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;
790 790 }
791 791 cvd->vdev_id = mvd->vdev_id;
792 792 vdev_add_child(pvd, cvd);
793 793 vdev_top_update(cvd->vdev_top, cvd->vdev_top);
794 794
795 795 if (cvd == cvd->vdev_top)
796 796 vdev_top_transfer(mvd, cvd);
797 797
798 798 ASSERT(mvd->vdev_children == 0);
799 799 vdev_free(mvd);
800 800 }
801 801
802 802 int
803 803 vdev_metaslab_init(vdev_t *vd, uint64_t txg)
804 804 {
805 805 spa_t *spa = vd->vdev_spa;
806 806 objset_t *mos = spa->spa_meta_objset;
807 807 uint64_t m;
808 808 uint64_t oldc = vd->vdev_ms_count;
809 809 uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
810 810 metaslab_t **mspp;
811 811 int error;
812 812
813 813 ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER));
814 814
815 815 /*
816 816 * This vdev is not being allocated from yet or is a hole.
817 817 */
818 818 if (vd->vdev_ms_shift == 0)
819 819 return (0);
820 820
821 821 ASSERT(!vd->vdev_ishole);
822 822
823 823 /*
824 824 * Compute the raidz-deflation ratio. Note, we hard-code
825 825 * in 128k (1 << 17) because it is the current "typical" blocksize.
826 826 * Even if SPA_MAXBLOCKSIZE changes, this algorithm must never change,
827 827 * or we will inconsistently account for existing bp's.
828 828 */
829 829 vd->vdev_deflate_ratio = (1 << 17) /
830 830 (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
831 831
832 832 ASSERT(oldc <= newc);
833 833
834 834 mspp = kmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
835 835
836 836 if (oldc != 0) {
837 837 bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp));
838 838 kmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
839 839 }
840 840
841 841 vd->vdev_ms = mspp;
842 842 vd->vdev_ms_count = newc;
843 843
844 844 for (m = oldc; m < newc; m++) {
845 845 uint64_t object = 0;
846 846
847 847 if (txg == 0) {
848 848 error = dmu_read(mos, vd->vdev_ms_array,
849 849 m * sizeof (uint64_t), sizeof (uint64_t), &object,
850 850 DMU_READ_PREFETCH);
851 851 if (error)
852 852 return (error);
853 853 }
854 854 vd->vdev_ms[m] = metaslab_init(vd->vdev_mg, m, object, txg);
855 855 }
856 856
857 857 if (txg == 0)
858 858 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER);
859 859
860 860 /*
861 861 * If the vdev is being removed we don't activate
862 862 * the metaslabs since we want to ensure that no new
863 863 * allocations are performed on this device.
864 864 */
865 865 if (oldc == 0 && !vd->vdev_removing)
866 866 metaslab_group_activate(vd->vdev_mg);
867 867
868 868 if (txg == 0)
869 869 spa_config_exit(spa, SCL_ALLOC, FTAG);
870 870
871 871 return (0);
872 872 }
873 873
874 874 void
875 875 vdev_metaslab_fini(vdev_t *vd)
876 876 {
877 877 uint64_t m;
878 878 uint64_t count = vd->vdev_ms_count;
879 879
880 880 if (vd->vdev_ms != NULL) {
881 881 metaslab_group_passivate(vd->vdev_mg);
882 882 for (m = 0; m < count; m++) {
883 883 metaslab_t *msp = vd->vdev_ms[m];
884 884
885 885 if (msp != NULL)
886 886 metaslab_fini(msp);
887 887 }
888 888 kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
889 889 vd->vdev_ms = NULL;
890 890 }
891 891 }
892 892
893 893 typedef struct vdev_probe_stats {
894 894 boolean_t vps_readable;
895 895 boolean_t vps_writeable;
896 896 int vps_flags;
897 897 } vdev_probe_stats_t;
898 898
899 899 static void
900 900 vdev_probe_done(zio_t *zio)
901 901 {
902 902 spa_t *spa = zio->io_spa;
903 903 vdev_t *vd = zio->io_vd;
904 904 vdev_probe_stats_t *vps = zio->io_private;
905 905
906 906 ASSERT(vd->vdev_probe_zio != NULL);
907 907
908 908 if (zio->io_type == ZIO_TYPE_READ) {
909 909 if (zio->io_error == 0)
910 910 vps->vps_readable = 1;
911 911 if (zio->io_error == 0 && spa_writeable(spa)) {
912 912 zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
913 913 zio->io_offset, zio->io_size, zio->io_data,
914 914 ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
915 915 ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
916 916 } else {
917 917 zio_buf_free(zio->io_data, zio->io_size);
918 918 }
919 919 } else if (zio->io_type == ZIO_TYPE_WRITE) {
920 920 if (zio->io_error == 0)
921 921 vps->vps_writeable = 1;
922 922 zio_buf_free(zio->io_data, zio->io_size);
923 923 } else if (zio->io_type == ZIO_TYPE_NULL) {
924 924 zio_t *pio;
925 925
926 926 vd->vdev_cant_read |= !vps->vps_readable;
927 927 vd->vdev_cant_write |= !vps->vps_writeable;
928 928
929 929 if (vdev_readable(vd) &&
930 930 (vdev_writeable(vd) || !spa_writeable(spa))) {
931 931 zio->io_error = 0;
932 932 } else {
933 933 ASSERT(zio->io_error != 0);
934 934 zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
935 935 spa, vd, NULL, 0, 0);
936 936 zio->io_error = SET_ERROR(ENXIO);
937 937 }
938 938
939 939 mutex_enter(&vd->vdev_probe_lock);
940 940 ASSERT(vd->vdev_probe_zio == zio);
941 941 vd->vdev_probe_zio = NULL;
942 942 mutex_exit(&vd->vdev_probe_lock);
943 943
944 944 while ((pio = zio_walk_parents(zio)) != NULL)
945 945 if (!vdev_accessible(vd, pio))
946 946 pio->io_error = SET_ERROR(ENXIO);
947 947
948 948 kmem_free(vps, sizeof (*vps));
949 949 }
950 950 }
951 951
952 952 /*
953 953 * Determine whether this device is accessible.
954 954 *
955 955 * Read and write to several known locations: the pad regions of each
956 956 * vdev label but the first, which we leave alone in case it contains
957 957 * a VTOC.
958 958 */
959 959 zio_t *
960 960 vdev_probe(vdev_t *vd, zio_t *zio)
961 961 {
962 962 spa_t *spa = vd->vdev_spa;
963 963 vdev_probe_stats_t *vps = NULL;
964 964 zio_t *pio;
965 965
966 966 ASSERT(vd->vdev_ops->vdev_op_leaf);
967 967
968 968 /*
969 969 * Don't probe the probe.
970 970 */
971 971 if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
972 972 return (NULL);
973 973
974 974 /*
975 975 * To prevent 'probe storms' when a device fails, we create
976 976 * just one probe i/o at a time. All zios that want to probe
977 977 * this vdev will become parents of the probe io.
978 978 */
979 979 mutex_enter(&vd->vdev_probe_lock);
980 980
981 981 if ((pio = vd->vdev_probe_zio) == NULL) {
982 982 vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
983 983
984 984 vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
985 985 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
986 986 ZIO_FLAG_TRYHARD;
987 987
988 988 if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
989 989 /*
990 990 * vdev_cant_read and vdev_cant_write can only
991 991 * transition from TRUE to FALSE when we have the
992 992 * SCL_ZIO lock as writer; otherwise they can only
993 993 * transition from FALSE to TRUE. This ensures that
994 994 * any zio looking at these values can assume that
995 995 * failures persist for the life of the I/O. That's
996 996 * important because when a device has intermittent
997 997 * connectivity problems, we want to ensure that
998 998 * they're ascribed to the device (ENXIO) and not
999 999 * the zio (EIO).
1000 1000 *
1001 1001 * Since we hold SCL_ZIO as writer here, clear both
1002 1002 * values so the probe can reevaluate from first
1003 1003 * principles.
1004 1004 */
1005 1005 vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
1006 1006 vd->vdev_cant_read = B_FALSE;
1007 1007 vd->vdev_cant_write = B_FALSE;
1008 1008 }
1009 1009
1010 1010 vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
1011 1011 vdev_probe_done, vps,
1012 1012 vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
1013 1013
1014 1014 /*
1015 1015 * We can't change the vdev state in this context, so we
1016 1016 * kick off an async task to do it on our behalf.
1017 1017 */
1018 1018 if (zio != NULL) {
1019 1019 vd->vdev_probe_wanted = B_TRUE;
1020 1020 spa_async_request(spa, SPA_ASYNC_PROBE);
1021 1021 }
1022 1022 }
1023 1023
1024 1024 if (zio != NULL)
1025 1025 zio_add_child(zio, pio);
1026 1026
1027 1027 mutex_exit(&vd->vdev_probe_lock);
1028 1028
1029 1029 if (vps == NULL) {
1030 1030 ASSERT(zio != NULL);
1031 1031 return (NULL);
1032 1032 }
1033 1033
1034 1034 for (int l = 1; l < VDEV_LABELS; l++) {
1035 1035 zio_nowait(zio_read_phys(pio, vd,
1036 1036 vdev_label_offset(vd->vdev_psize, l,
1037 1037 offsetof(vdev_label_t, vl_pad2)),
1038 1038 VDEV_PAD_SIZE, zio_buf_alloc(VDEV_PAD_SIZE),
1039 1039 ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1040 1040 ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1041 1041 }
1042 1042
1043 1043 if (zio == NULL)
1044 1044 return (pio);
1045 1045
1046 1046 zio_nowait(pio);
1047 1047 return (NULL);
1048 1048 }
1049 1049
1050 1050 static void
1051 1051 vdev_open_child(void *arg)
1052 1052 {
1053 1053 vdev_t *vd = arg;
1054 1054
1055 1055 vd->vdev_open_thread = curthread;
1056 1056 vd->vdev_open_error = vdev_open(vd);
1057 1057 vd->vdev_open_thread = NULL;
1058 1058 }
1059 1059
1060 1060 boolean_t
1061 1061 vdev_uses_zvols(vdev_t *vd)
1062 1062 {
1063 1063 if (vd->vdev_path && strncmp(vd->vdev_path, ZVOL_DIR,
1064 1064 strlen(ZVOL_DIR)) == 0)
1065 1065 return (B_TRUE);
1066 1066 for (int c = 0; c < vd->vdev_children; c++)
1067 1067 if (vdev_uses_zvols(vd->vdev_child[c]))
1068 1068 return (B_TRUE);
1069 1069 return (B_FALSE);
1070 1070 }
1071 1071
1072 1072 void
1073 1073 vdev_open_children(vdev_t *vd)
1074 1074 {
1075 1075 taskq_t *tq;
1076 1076 int children = vd->vdev_children;
1077 1077
1078 1078 /*
1079 1079 * in order to handle pools on top of zvols, do the opens
1080 1080 * in a single thread so that the same thread holds the
1081 1081 * spa_namespace_lock
1082 1082 */
1083 1083 if (vdev_uses_zvols(vd)) {
1084 1084 for (int c = 0; c < children; c++)
1085 1085 vd->vdev_child[c]->vdev_open_error =
1086 1086 vdev_open(vd->vdev_child[c]);
1087 1087 return;
1088 1088 }
1089 1089 tq = taskq_create("vdev_open", children, minclsyspri,
1090 1090 children, children, TASKQ_PREPOPULATE);
1091 1091
1092 1092 for (int c = 0; c < children; c++)
1093 1093 VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c],
1094 1094 TQ_SLEEP) != NULL);
1095 1095
1096 1096 taskq_destroy(tq);
1097 1097 }
1098 1098
1099 1099 /*
1100 1100 * Prepare a virtual device for access.
1101 1101 */
1102 1102 int
1103 1103 vdev_open(vdev_t *vd)
1104 1104 {
1105 1105 spa_t *spa = vd->vdev_spa;
1106 1106 int error;
1107 1107 uint64_t osize = 0;
1108 1108 uint64_t max_osize = 0;
1109 1109 uint64_t asize, max_asize, psize;
1110 1110 uint64_t ashift = 0;
1111 1111
1112 1112 ASSERT(vd->vdev_open_thread == curthread ||
1113 1113 spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1114 1114 ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1115 1115 vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1116 1116 vd->vdev_state == VDEV_STATE_OFFLINE);
1117 1117
1118 1118 vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1119 1119 vd->vdev_cant_read = B_FALSE;
1120 1120 vd->vdev_cant_write = B_FALSE;
1121 1121 vd->vdev_min_asize = vdev_get_min_asize(vd);
1122 1122
1123 1123 /*
1124 1124 * If this vdev is not removed, check its fault status. If it's
1125 1125 * faulted, bail out of the open.
1126 1126 */
1127 1127 if (!vd->vdev_removed && vd->vdev_faulted) {
1128 1128 ASSERT(vd->vdev_children == 0);
1129 1129 ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1130 1130 vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1131 1131 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1132 1132 vd->vdev_label_aux);
1133 1133 return (SET_ERROR(ENXIO));
1134 1134 } else if (vd->vdev_offline) {
1135 1135 ASSERT(vd->vdev_children == 0);
1136 1136 vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1137 1137 return (SET_ERROR(ENXIO));
1138 1138 }
1139 1139
1140 1140 error = vd->vdev_ops->vdev_op_open(vd, &osize, &max_osize, &ashift);
1141 1141
1142 1142 /*
1143 1143 * Reset the vdev_reopening flag so that we actually close
1144 1144 * the vdev on error.
1145 1145 */
1146 1146 vd->vdev_reopening = B_FALSE;
1147 1147 if (zio_injection_enabled && error == 0)
1148 1148 error = zio_handle_device_injection(vd, NULL, ENXIO);
1149 1149
1150 1150 if (error) {
1151 1151 if (vd->vdev_removed &&
1152 1152 vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1153 1153 vd->vdev_removed = B_FALSE;
1154 1154
1155 1155 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1156 1156 vd->vdev_stat.vs_aux);
1157 1157 return (error);
1158 1158 }
1159 1159
1160 1160 vd->vdev_removed = B_FALSE;
1161 1161
1162 1162 /*
1163 1163 * Recheck the faulted flag now that we have confirmed that
1164 1164 * the vdev is accessible. If we're faulted, bail.
1165 1165 */
1166 1166 if (vd->vdev_faulted) {
1167 1167 ASSERT(vd->vdev_children == 0);
1168 1168 ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1169 1169 vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1170 1170 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1171 1171 vd->vdev_label_aux);
1172 1172 return (SET_ERROR(ENXIO));
1173 1173 }
1174 1174
1175 1175 if (vd->vdev_degraded) {
1176 1176 ASSERT(vd->vdev_children == 0);
1177 1177 vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1178 1178 VDEV_AUX_ERR_EXCEEDED);
1179 1179 } else {
1180 1180 vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1181 1181 }
1182 1182
1183 1183 /*
1184 1184 * For hole or missing vdevs we just return success.
1185 1185 */
1186 1186 if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1187 1187 return (0);
1188 1188
1189 1189 for (int c = 0; c < vd->vdev_children; c++) {
1190 1190 if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1191 1191 vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1192 1192 VDEV_AUX_NONE);
1193 1193 break;
1194 1194 }
1195 1195 }
1196 1196
1197 1197 osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1198 1198 max_osize = P2ALIGN(max_osize, (uint64_t)sizeof (vdev_label_t));
1199 1199
1200 1200 if (vd->vdev_children == 0) {
1201 1201 if (osize < SPA_MINDEVSIZE) {
1202 1202 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1203 1203 VDEV_AUX_TOO_SMALL);
1204 1204 return (SET_ERROR(EOVERFLOW));
1205 1205 }
1206 1206 psize = osize;
1207 1207 asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1208 1208 max_asize = max_osize - (VDEV_LABEL_START_SIZE +
1209 1209 VDEV_LABEL_END_SIZE);
1210 1210 } else {
1211 1211 if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1212 1212 (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1213 1213 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1214 1214 VDEV_AUX_TOO_SMALL);
1215 1215 return (SET_ERROR(EOVERFLOW));
1216 1216 }
1217 1217 psize = 0;
1218 1218 asize = osize;
1219 1219 max_asize = max_osize;
1220 1220 }
1221 1221
1222 1222 vd->vdev_psize = psize;
1223 1223
1224 1224 /*
1225 1225 * Make sure the allocatable size hasn't shrunk.
1226 1226 */
1227 1227 if (asize < vd->vdev_min_asize) {
1228 1228 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1229 1229 VDEV_AUX_BAD_LABEL);
1230 1230 return (SET_ERROR(EINVAL));
1231 1231 }
1232 1232
1233 1233 if (vd->vdev_asize == 0) {
1234 1234 /*
1235 1235 * This is the first-ever open, so use the computed values.
1236 1236 * For testing purposes, a higher ashift can be requested.
1237 1237 */
1238 1238 vd->vdev_asize = asize;
1239 1239 vd->vdev_max_asize = max_asize;
1240 1240 vd->vdev_ashift = MAX(ashift, vd->vdev_ashift);
1241 1241 } else {
1242 1242 /*
1243 1243 * Detect if the alignment requirement has increased.
1244 1244 * We don't want to make the pool unavailable, just
1245 1245 * issue a warning instead.
1246 1246 */
1247 1247 if (ashift > vd->vdev_top->vdev_ashift &&
1248 1248 vd->vdev_ops->vdev_op_leaf) {
1249 1249 cmn_err(CE_WARN,
1250 1250 "Disk, '%s', has a block alignment that is "
1251 1251 "larger than the pool's alignment\n",
1252 1252 vd->vdev_path);
1253 1253 }
1254 1254 vd->vdev_max_asize = max_asize;
1255 1255 }
1256 1256
1257 1257 /*
1258 1258 * If all children are healthy and the asize has increased,
1259 1259 * then we've experienced dynamic LUN growth. If automatic
1260 1260 * expansion is enabled then use the additional space.
1261 1261 */
1262 1262 if (vd->vdev_state == VDEV_STATE_HEALTHY && asize > vd->vdev_asize &&
1263 1263 (vd->vdev_expanding || spa->spa_autoexpand))
1264 1264 vd->vdev_asize = asize;
1265 1265
1266 1266 vdev_set_min_asize(vd);
1267 1267
1268 1268 /*
1269 1269 * Ensure we can issue some IO before declaring the
1270 1270 * vdev open for business.
1271 1271 */
1272 1272 if (vd->vdev_ops->vdev_op_leaf &&
1273 1273 (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1274 1274 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1275 1275 VDEV_AUX_ERR_EXCEEDED);
1276 1276 return (error);
1277 1277 }
1278 1278
1279 1279 /*
1280 1280 * If a leaf vdev has a DTL, and seems healthy, then kick off a
1281 1281 * resilver. But don't do this if we are doing a reopen for a scrub,
1282 1282 * since this would just restart the scrub we are already doing.
1283 1283 */
1284 1284 if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1285 1285 vdev_resilver_needed(vd, NULL, NULL))
1286 1286 spa_async_request(spa, SPA_ASYNC_RESILVER);
1287 1287
1288 1288 return (0);
1289 1289 }
1290 1290
1291 1291 /*
1292 1292 * Called once the vdevs are all opened, this routine validates the label
1293 1293 * contents. This needs to be done before vdev_load() so that we don't
1294 1294 * inadvertently do repair I/Os to the wrong device.
1295 1295 *
1296 1296 * If 'strict' is false ignore the spa guid check. This is necessary because
1297 1297 * if the machine crashed during a re-guid the new guid might have been written
1298 1298 * to all of the vdev labels, but not the cached config. The strict check
1299 1299 * will be performed when the pool is opened again using the mos config.
1300 1300 *
1301 1301 * This function will only return failure if one of the vdevs indicates that it
1302 1302 * has since been destroyed or exported. This is only possible if
1303 1303 * /etc/zfs/zpool.cache was readonly at the time. Otherwise, the vdev state
1304 1304 * will be updated but the function will return 0.
1305 1305 */
1306 1306 int
1307 1307 vdev_validate(vdev_t *vd, boolean_t strict)
1308 1308 {
1309 1309 spa_t *spa = vd->vdev_spa;
1310 1310 nvlist_t *label;
1311 1311 uint64_t guid = 0, top_guid;
1312 1312 uint64_t state;
1313 1313
1314 1314 for (int c = 0; c < vd->vdev_children; c++)
1315 1315 if (vdev_validate(vd->vdev_child[c], strict) != 0)
1316 1316 return (SET_ERROR(EBADF));
1317 1317
1318 1318 /*
1319 1319 * If the device has already failed, or was marked offline, don't do
1320 1320 * any further validation. Otherwise, label I/O will fail and we will
1321 1321 * overwrite the previous state.
1322 1322 */
1323 1323 if (vd->vdev_ops->vdev_op_leaf && vdev_readable(vd)) {
1324 1324 uint64_t aux_guid = 0;
1325 1325 nvlist_t *nvl;
1326 1326 uint64_t txg = spa_last_synced_txg(spa) != 0 ?
1327 1327 spa_last_synced_txg(spa) : -1ULL;
1328 1328
1329 1329 if ((label = vdev_label_read_config(vd, txg)) == NULL) {
1330 1330 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1331 1331 VDEV_AUX_BAD_LABEL);
1332 1332 return (0);
1333 1333 }
1334 1334
1335 1335 /*
1336 1336 * Determine if this vdev has been split off into another
1337 1337 * pool. If so, then refuse to open it.
1338 1338 */
1339 1339 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID,
1340 1340 &aux_guid) == 0 && aux_guid == spa_guid(spa)) {
1341 1341 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1342 1342 VDEV_AUX_SPLIT_POOL);
1343 1343 nvlist_free(label);
1344 1344 return (0);
1345 1345 }
1346 1346
1347 1347 if (strict && (nvlist_lookup_uint64(label,
1348 1348 ZPOOL_CONFIG_POOL_GUID, &guid) != 0 ||
1349 1349 guid != spa_guid(spa))) {
1350 1350 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1351 1351 VDEV_AUX_CORRUPT_DATA);
1352 1352 nvlist_free(label);
1353 1353 return (0);
1354 1354 }
1355 1355
1356 1356 if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl)
1357 1357 != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID,
1358 1358 &aux_guid) != 0)
1359 1359 aux_guid = 0;
1360 1360
1361 1361 /*
1362 1362 * If this vdev just became a top-level vdev because its
1363 1363 * sibling was detached, it will have adopted the parent's
1364 1364 * vdev guid -- but the label may or may not be on disk yet.
1365 1365 * Fortunately, either version of the label will have the
1366 1366 * same top guid, so if we're a top-level vdev, we can
1367 1367 * safely compare to that instead.
1368 1368 *
1369 1369 * If we split this vdev off instead, then we also check the
1370 1370 * original pool's guid. We don't want to consider the vdev
1371 1371 * corrupt if it is partway through a split operation.
1372 1372 */
1373 1373 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID,
1374 1374 &guid) != 0 ||
1375 1375 nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID,
1376 1376 &top_guid) != 0 ||
1377 1377 ((vd->vdev_guid != guid && vd->vdev_guid != aux_guid) &&
1378 1378 (vd->vdev_guid != top_guid || vd != vd->vdev_top))) {
1379 1379 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1380 1380 VDEV_AUX_CORRUPT_DATA);
1381 1381 nvlist_free(label);
1382 1382 return (0);
1383 1383 }
1384 1384
1385 1385 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1386 1386 &state) != 0) {
1387 1387 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1388 1388 VDEV_AUX_CORRUPT_DATA);
1389 1389 nvlist_free(label);
1390 1390 return (0);
1391 1391 }
1392 1392
1393 1393 nvlist_free(label);
1394 1394
1395 1395 /*
1396 1396 * If this is a verbatim import, no need to check the
1397 1397 * state of the pool.
1398 1398 */
1399 1399 if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) &&
1400 1400 spa_load_state(spa) == SPA_LOAD_OPEN &&
1401 1401 state != POOL_STATE_ACTIVE)
1402 1402 return (SET_ERROR(EBADF));
1403 1403
1404 1404 /*
1405 1405 * If we were able to open and validate a vdev that was
1406 1406 * previously marked permanently unavailable, clear that state
1407 1407 * now.
1408 1408 */
1409 1409 if (vd->vdev_not_present)
1410 1410 vd->vdev_not_present = 0;
1411 1411 }
1412 1412
1413 1413 return (0);
1414 1414 }
1415 1415
1416 1416 /*
1417 1417 * Close a virtual device.
1418 1418 */
1419 1419 void
1420 1420 vdev_close(vdev_t *vd)
1421 1421 {
1422 1422 spa_t *spa = vd->vdev_spa;
1423 1423 vdev_t *pvd = vd->vdev_parent;
1424 1424
1425 1425 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1426 1426
1427 1427 /*
1428 1428 * If our parent is reopening, then we are as well, unless we are
1429 1429 * going offline.
1430 1430 */
1431 1431 if (pvd != NULL && pvd->vdev_reopening)
1432 1432 vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline);
1433 1433
1434 1434 vd->vdev_ops->vdev_op_close(vd);
1435 1435
1436 1436 vdev_cache_purge(vd);
1437 1437
1438 1438 /*
1439 1439 * We record the previous state before we close it, so that if we are
1440 1440 * doing a reopen(), we don't generate FMA ereports if we notice that
1441 1441 * it's still faulted.
1442 1442 */
1443 1443 vd->vdev_prevstate = vd->vdev_state;
1444 1444
1445 1445 if (vd->vdev_offline)
1446 1446 vd->vdev_state = VDEV_STATE_OFFLINE;
1447 1447 else
1448 1448 vd->vdev_state = VDEV_STATE_CLOSED;
1449 1449 vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1450 1450 }
1451 1451
1452 1452 void
1453 1453 vdev_hold(vdev_t *vd)
1454 1454 {
1455 1455 spa_t *spa = vd->vdev_spa;
1456 1456
1457 1457 ASSERT(spa_is_root(spa));
1458 1458 if (spa->spa_state == POOL_STATE_UNINITIALIZED)
1459 1459 return;
1460 1460
1461 1461 for (int c = 0; c < vd->vdev_children; c++)
1462 1462 vdev_hold(vd->vdev_child[c]);
1463 1463
1464 1464 if (vd->vdev_ops->vdev_op_leaf)
1465 1465 vd->vdev_ops->vdev_op_hold(vd);
1466 1466 }
1467 1467
1468 1468 void
1469 1469 vdev_rele(vdev_t *vd)
1470 1470 {
1471 1471 spa_t *spa = vd->vdev_spa;
1472 1472
1473 1473 ASSERT(spa_is_root(spa));
1474 1474 for (int c = 0; c < vd->vdev_children; c++)
1475 1475 vdev_rele(vd->vdev_child[c]);
1476 1476
1477 1477 if (vd->vdev_ops->vdev_op_leaf)
1478 1478 vd->vdev_ops->vdev_op_rele(vd);
1479 1479 }
1480 1480
1481 1481 /*
1482 1482 * Reopen all interior vdevs and any unopened leaves. We don't actually
1483 1483 * reopen leaf vdevs which had previously been opened as they might deadlock
1484 1484 * on the spa_config_lock. Instead we only obtain the leaf's physical size.
1485 1485 * If the leaf has never been opened then open it, as usual.
1486 1486 */
1487 1487 void
1488 1488 vdev_reopen(vdev_t *vd)
1489 1489 {
1490 1490 spa_t *spa = vd->vdev_spa;
1491 1491
1492 1492 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1493 1493
1494 1494 /* set the reopening flag unless we're taking the vdev offline */
1495 1495 vd->vdev_reopening = !vd->vdev_offline;
1496 1496 vdev_close(vd);
1497 1497 (void) vdev_open(vd);
|
↓ open down ↓ |
1497 lines elided |
↑ open up ↑ |
1498 1498
1499 1499 /*
1500 1500 * Call vdev_validate() here to make sure we have the same device.
1501 1501 * Otherwise, a device with an invalid label could be successfully
1502 1502 * opened in response to vdev_reopen().
1503 1503 */
1504 1504 if (vd->vdev_aux) {
1505 1505 (void) vdev_validate_aux(vd);
1506 1506 if (vdev_readable(vd) && vdev_writeable(vd) &&
1507 1507 vd->vdev_aux == &spa->spa_l2cache &&
1508 - !l2arc_vdev_present(vd))
1509 - l2arc_add_vdev(spa, vd);
1508 + !l2arc_vdev_present(vd)) {
1509 + /*
1510 + * When reopening we can assume persistent L2ARC is
1511 + * supported, since we've already opened the device
1512 + * in the past and prepended an L2ARC uberblock.
1513 + */
1514 + l2arc_add_vdev(spa, vd, B_TRUE);
1515 + }
1510 1516 } else {
1511 1517 (void) vdev_validate(vd, B_TRUE);
1512 1518 }
1513 1519
1514 1520 /*
1515 1521 * Reassess parent vdev's health.
1516 1522 */
1517 1523 vdev_propagate_state(vd);
1518 1524 }
1519 1525
1520 1526 int
1521 1527 vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1522 1528 {
1523 1529 int error;
1524 1530
1525 1531 /*
1526 1532 * Normally, partial opens (e.g. of a mirror) are allowed.
1527 1533 * For a create, however, we want to fail the request if
1528 1534 * there are any components we can't open.
1529 1535 */
1530 1536 error = vdev_open(vd);
1531 1537
1532 1538 if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
1533 1539 vdev_close(vd);
1534 1540 return (error ? error : ENXIO);
1535 1541 }
1536 1542
1537 1543 /*
1538 1544 * Recursively load DTLs and initialize all labels.
1539 1545 */
1540 1546 if ((error = vdev_dtl_load(vd)) != 0 ||
1541 1547 (error = vdev_label_init(vd, txg, isreplacing ?
1542 1548 VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
1543 1549 vdev_close(vd);
1544 1550 return (error);
1545 1551 }
1546 1552
1547 1553 return (0);
1548 1554 }
1549 1555
1550 1556 void
1551 1557 vdev_metaslab_set_size(vdev_t *vd)
1552 1558 {
1553 1559 /*
1554 1560 * Aim for roughly 200 metaslabs per vdev.
1555 1561 */
1556 1562 vd->vdev_ms_shift = highbit(vd->vdev_asize / 200);
1557 1563 vd->vdev_ms_shift = MAX(vd->vdev_ms_shift, SPA_MAXBLOCKSHIFT);
1558 1564 }
1559 1565
1560 1566 void
1561 1567 vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
1562 1568 {
1563 1569 ASSERT(vd == vd->vdev_top);
1564 1570 ASSERT(!vd->vdev_ishole);
1565 1571 ASSERT(ISP2(flags));
1566 1572 ASSERT(spa_writeable(vd->vdev_spa));
1567 1573
1568 1574 if (flags & VDD_METASLAB)
1569 1575 (void) txg_list_add(&vd->vdev_ms_list, arg, txg);
1570 1576
1571 1577 if (flags & VDD_DTL)
1572 1578 (void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
1573 1579
1574 1580 (void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
1575 1581 }
1576 1582
1577 1583 void
1578 1584 vdev_dirty_leaves(vdev_t *vd, int flags, uint64_t txg)
1579 1585 {
1580 1586 for (int c = 0; c < vd->vdev_children; c++)
1581 1587 vdev_dirty_leaves(vd->vdev_child[c], flags, txg);
1582 1588
1583 1589 if (vd->vdev_ops->vdev_op_leaf)
1584 1590 vdev_dirty(vd->vdev_top, flags, vd, txg);
1585 1591 }
1586 1592
1587 1593 /*
1588 1594 * DTLs.
1589 1595 *
1590 1596 * A vdev's DTL (dirty time log) is the set of transaction groups for which
1591 1597 * the vdev has less than perfect replication. There are four kinds of DTL:
1592 1598 *
1593 1599 * DTL_MISSING: txgs for which the vdev has no valid copies of the data
1594 1600 *
1595 1601 * DTL_PARTIAL: txgs for which data is available, but not fully replicated
1596 1602 *
1597 1603 * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
1598 1604 * scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
1599 1605 * txgs that was scrubbed.
1600 1606 *
1601 1607 * DTL_OUTAGE: txgs which cannot currently be read, whether due to
1602 1608 * persistent errors or just some device being offline.
1603 1609 * Unlike the other three, the DTL_OUTAGE map is not generally
1604 1610 * maintained; it's only computed when needed, typically to
1605 1611 * determine whether a device can be detached.
1606 1612 *
1607 1613 * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
1608 1614 * either has the data or it doesn't.
1609 1615 *
1610 1616 * For interior vdevs such as mirror and RAID-Z the picture is more complex.
1611 1617 * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
1612 1618 * if any child is less than fully replicated, then so is its parent.
1613 1619 * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
1614 1620 * comprising only those txgs which appear in 'maxfaults' or more children;
1615 1621 * those are the txgs we don't have enough replication to read. For example,
1616 1622 * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
1617 1623 * thus, its DTL_MISSING consists of the set of txgs that appear in more than
1618 1624 * two child DTL_MISSING maps.
1619 1625 *
1620 1626 * It should be clear from the above that to compute the DTLs and outage maps
1621 1627 * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
1622 1628 * Therefore, that is all we keep on disk. When loading the pool, or after
1623 1629 * a configuration change, we generate all other DTLs from first principles.
1624 1630 */
1625 1631 void
1626 1632 vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1627 1633 {
1628 1634 range_tree_t *rt = vd->vdev_dtl[t];
1629 1635
1630 1636 ASSERT(t < DTL_TYPES);
1631 1637 ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1632 1638 ASSERT(spa_writeable(vd->vdev_spa));
1633 1639
1634 1640 mutex_enter(rt->rt_lock);
1635 1641 if (!range_tree_contains(rt, txg, size))
1636 1642 range_tree_add(rt, txg, size);
1637 1643 mutex_exit(rt->rt_lock);
1638 1644 }
1639 1645
1640 1646 boolean_t
1641 1647 vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1642 1648 {
1643 1649 range_tree_t *rt = vd->vdev_dtl[t];
1644 1650 boolean_t dirty = B_FALSE;
1645 1651
1646 1652 ASSERT(t < DTL_TYPES);
1647 1653 ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1648 1654
1649 1655 mutex_enter(rt->rt_lock);
1650 1656 if (range_tree_space(rt) != 0)
1651 1657 dirty = range_tree_contains(rt, txg, size);
1652 1658 mutex_exit(rt->rt_lock);
1653 1659
1654 1660 return (dirty);
1655 1661 }
1656 1662
1657 1663 boolean_t
1658 1664 vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
1659 1665 {
1660 1666 range_tree_t *rt = vd->vdev_dtl[t];
1661 1667 boolean_t empty;
1662 1668
1663 1669 mutex_enter(rt->rt_lock);
1664 1670 empty = (range_tree_space(rt) == 0);
1665 1671 mutex_exit(rt->rt_lock);
1666 1672
1667 1673 return (empty);
1668 1674 }
1669 1675
1670 1676 /*
1671 1677 * Returns the lowest txg in the DTL range.
1672 1678 */
1673 1679 static uint64_t
1674 1680 vdev_dtl_min(vdev_t *vd)
1675 1681 {
1676 1682 range_seg_t *rs;
1677 1683
1678 1684 ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
1679 1685 ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
1680 1686 ASSERT0(vd->vdev_children);
1681 1687
1682 1688 rs = avl_first(&vd->vdev_dtl[DTL_MISSING]->rt_root);
1683 1689 return (rs->rs_start - 1);
1684 1690 }
1685 1691
1686 1692 /*
1687 1693 * Returns the highest txg in the DTL.
1688 1694 */
1689 1695 static uint64_t
1690 1696 vdev_dtl_max(vdev_t *vd)
1691 1697 {
1692 1698 range_seg_t *rs;
1693 1699
1694 1700 ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
1695 1701 ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
1696 1702 ASSERT0(vd->vdev_children);
1697 1703
1698 1704 rs = avl_last(&vd->vdev_dtl[DTL_MISSING]->rt_root);
1699 1705 return (rs->rs_end);
1700 1706 }
1701 1707
1702 1708 /*
1703 1709 * Determine if a resilvering vdev should remove any DTL entries from
1704 1710 * its range. If the vdev was resilvering for the entire duration of the
1705 1711 * scan then it should excise that range from its DTLs. Otherwise, this
1706 1712 * vdev is considered partially resilvered and should leave its DTL
1707 1713 * entries intact. The comment in vdev_dtl_reassess() describes how we
1708 1714 * excise the DTLs.
1709 1715 */
1710 1716 static boolean_t
1711 1717 vdev_dtl_should_excise(vdev_t *vd)
1712 1718 {
1713 1719 spa_t *spa = vd->vdev_spa;
1714 1720 dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1715 1721
1716 1722 ASSERT0(scn->scn_phys.scn_errors);
1717 1723 ASSERT0(vd->vdev_children);
1718 1724
1719 1725 if (vd->vdev_resilver_txg == 0 ||
1720 1726 range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0)
1721 1727 return (B_TRUE);
1722 1728
1723 1729 /*
1724 1730 * When a resilver is initiated the scan will assign the scn_max_txg
1725 1731 * value to the highest txg value that exists in all DTLs. If this
1726 1732 * device's max DTL is not part of this scan (i.e. it is not in
1727 1733 * the range (scn_min_txg, scn_max_txg] then it is not eligible
1728 1734 * for excision.
1729 1735 */
1730 1736 if (vdev_dtl_max(vd) <= scn->scn_phys.scn_max_txg) {
1731 1737 ASSERT3U(scn->scn_phys.scn_min_txg, <=, vdev_dtl_min(vd));
1732 1738 ASSERT3U(scn->scn_phys.scn_min_txg, <, vd->vdev_resilver_txg);
1733 1739 ASSERT3U(vd->vdev_resilver_txg, <=, scn->scn_phys.scn_max_txg);
1734 1740 return (B_TRUE);
1735 1741 }
1736 1742 return (B_FALSE);
1737 1743 }
1738 1744
1739 1745 /*
1740 1746 * Reassess DTLs after a config change or scrub completion.
1741 1747 */
1742 1748 void
1743 1749 vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
1744 1750 {
1745 1751 spa_t *spa = vd->vdev_spa;
1746 1752 avl_tree_t reftree;
1747 1753 int minref;
1748 1754
1749 1755 ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
1750 1756
1751 1757 for (int c = 0; c < vd->vdev_children; c++)
1752 1758 vdev_dtl_reassess(vd->vdev_child[c], txg,
1753 1759 scrub_txg, scrub_done);
1754 1760
1755 1761 if (vd == spa->spa_root_vdev || vd->vdev_ishole || vd->vdev_aux)
1756 1762 return;
1757 1763
1758 1764 if (vd->vdev_ops->vdev_op_leaf) {
1759 1765 dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1760 1766
1761 1767 mutex_enter(&vd->vdev_dtl_lock);
1762 1768
1763 1769 /*
1764 1770 * If we've completed a scan cleanly then determine
1765 1771 * if this vdev should remove any DTLs. We only want to
1766 1772 * excise regions on vdevs that were available during
1767 1773 * the entire duration of this scan.
1768 1774 */
1769 1775 if (scrub_txg != 0 &&
1770 1776 (spa->spa_scrub_started ||
1771 1777 (scn != NULL && scn->scn_phys.scn_errors == 0)) &&
1772 1778 vdev_dtl_should_excise(vd)) {
1773 1779 /*
1774 1780 * We completed a scrub up to scrub_txg. If we
1775 1781 * did it without rebooting, then the scrub dtl
1776 1782 * will be valid, so excise the old region and
1777 1783 * fold in the scrub dtl. Otherwise, leave the
1778 1784 * dtl as-is if there was an error.
1779 1785 *
1780 1786 * There's little trick here: to excise the beginning
1781 1787 * of the DTL_MISSING map, we put it into a reference
1782 1788 * tree and then add a segment with refcnt -1 that
1783 1789 * covers the range [0, scrub_txg). This means
1784 1790 * that each txg in that range has refcnt -1 or 0.
1785 1791 * We then add DTL_SCRUB with a refcnt of 2, so that
1786 1792 * entries in the range [0, scrub_txg) will have a
1787 1793 * positive refcnt -- either 1 or 2. We then convert
1788 1794 * the reference tree into the new DTL_MISSING map.
1789 1795 */
1790 1796 space_reftree_create(&reftree);
1791 1797 space_reftree_add_map(&reftree,
1792 1798 vd->vdev_dtl[DTL_MISSING], 1);
1793 1799 space_reftree_add_seg(&reftree, 0, scrub_txg, -1);
1794 1800 space_reftree_add_map(&reftree,
1795 1801 vd->vdev_dtl[DTL_SCRUB], 2);
1796 1802 space_reftree_generate_map(&reftree,
1797 1803 vd->vdev_dtl[DTL_MISSING], 1);
1798 1804 space_reftree_destroy(&reftree);
1799 1805 }
1800 1806 range_tree_vacate(vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
1801 1807 range_tree_walk(vd->vdev_dtl[DTL_MISSING],
1802 1808 range_tree_add, vd->vdev_dtl[DTL_PARTIAL]);
1803 1809 if (scrub_done)
1804 1810 range_tree_vacate(vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
1805 1811 range_tree_vacate(vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
1806 1812 if (!vdev_readable(vd))
1807 1813 range_tree_add(vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
1808 1814 else
1809 1815 range_tree_walk(vd->vdev_dtl[DTL_MISSING],
1810 1816 range_tree_add, vd->vdev_dtl[DTL_OUTAGE]);
1811 1817
1812 1818 /*
1813 1819 * If the vdev was resilvering and no longer has any
1814 1820 * DTLs then reset its resilvering flag.
1815 1821 */
1816 1822 if (vd->vdev_resilver_txg != 0 &&
1817 1823 range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0 &&
1818 1824 range_tree_space(vd->vdev_dtl[DTL_OUTAGE]) == 0)
1819 1825 vd->vdev_resilver_txg = 0;
1820 1826
1821 1827 mutex_exit(&vd->vdev_dtl_lock);
1822 1828
1823 1829 if (txg != 0)
1824 1830 vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
1825 1831 return;
1826 1832 }
1827 1833
1828 1834 mutex_enter(&vd->vdev_dtl_lock);
1829 1835 for (int t = 0; t < DTL_TYPES; t++) {
1830 1836 /* account for child's outage in parent's missing map */
1831 1837 int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
1832 1838 if (t == DTL_SCRUB)
1833 1839 continue; /* leaf vdevs only */
1834 1840 if (t == DTL_PARTIAL)
1835 1841 minref = 1; /* i.e. non-zero */
1836 1842 else if (vd->vdev_nparity != 0)
1837 1843 minref = vd->vdev_nparity + 1; /* RAID-Z */
1838 1844 else
1839 1845 minref = vd->vdev_children; /* any kind of mirror */
1840 1846 space_reftree_create(&reftree);
1841 1847 for (int c = 0; c < vd->vdev_children; c++) {
1842 1848 vdev_t *cvd = vd->vdev_child[c];
1843 1849 mutex_enter(&cvd->vdev_dtl_lock);
1844 1850 space_reftree_add_map(&reftree, cvd->vdev_dtl[s], 1);
1845 1851 mutex_exit(&cvd->vdev_dtl_lock);
1846 1852 }
1847 1853 space_reftree_generate_map(&reftree, vd->vdev_dtl[t], minref);
1848 1854 space_reftree_destroy(&reftree);
1849 1855 }
1850 1856 mutex_exit(&vd->vdev_dtl_lock);
1851 1857 }
1852 1858
1853 1859 int
1854 1860 vdev_dtl_load(vdev_t *vd)
1855 1861 {
1856 1862 spa_t *spa = vd->vdev_spa;
1857 1863 objset_t *mos = spa->spa_meta_objset;
1858 1864 int error = 0;
1859 1865
1860 1866 if (vd->vdev_ops->vdev_op_leaf && vd->vdev_dtl_object != 0) {
1861 1867 ASSERT(!vd->vdev_ishole);
1862 1868
1863 1869 error = space_map_open(&vd->vdev_dtl_sm, mos,
1864 1870 vd->vdev_dtl_object, 0, -1ULL, 0, &vd->vdev_dtl_lock);
1865 1871 if (error)
1866 1872 return (error);
1867 1873 ASSERT(vd->vdev_dtl_sm != NULL);
1868 1874
1869 1875 mutex_enter(&vd->vdev_dtl_lock);
1870 1876
1871 1877 /*
1872 1878 * Now that we've opened the space_map we need to update
1873 1879 * the in-core DTL.
1874 1880 */
1875 1881 space_map_update(vd->vdev_dtl_sm);
1876 1882
1877 1883 error = space_map_load(vd->vdev_dtl_sm,
1878 1884 vd->vdev_dtl[DTL_MISSING], SM_ALLOC);
1879 1885 mutex_exit(&vd->vdev_dtl_lock);
1880 1886
1881 1887 return (error);
1882 1888 }
1883 1889
1884 1890 for (int c = 0; c < vd->vdev_children; c++) {
1885 1891 error = vdev_dtl_load(vd->vdev_child[c]);
1886 1892 if (error != 0)
1887 1893 break;
1888 1894 }
1889 1895
1890 1896 return (error);
1891 1897 }
1892 1898
1893 1899 void
1894 1900 vdev_dtl_sync(vdev_t *vd, uint64_t txg)
1895 1901 {
1896 1902 spa_t *spa = vd->vdev_spa;
1897 1903 range_tree_t *rt = vd->vdev_dtl[DTL_MISSING];
1898 1904 objset_t *mos = spa->spa_meta_objset;
1899 1905 range_tree_t *rtsync;
1900 1906 kmutex_t rtlock;
1901 1907 dmu_tx_t *tx;
1902 1908 uint64_t object = space_map_object(vd->vdev_dtl_sm);
1903 1909
1904 1910 ASSERT(!vd->vdev_ishole);
1905 1911 ASSERT(vd->vdev_ops->vdev_op_leaf);
1906 1912
1907 1913 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1908 1914
1909 1915 if (vd->vdev_detached || vd->vdev_top->vdev_removing) {
1910 1916 mutex_enter(&vd->vdev_dtl_lock);
1911 1917 space_map_free(vd->vdev_dtl_sm, tx);
1912 1918 space_map_close(vd->vdev_dtl_sm);
1913 1919 vd->vdev_dtl_sm = NULL;
1914 1920 mutex_exit(&vd->vdev_dtl_lock);
1915 1921 dmu_tx_commit(tx);
1916 1922 return;
1917 1923 }
1918 1924
1919 1925 if (vd->vdev_dtl_sm == NULL) {
1920 1926 uint64_t new_object;
1921 1927
1922 1928 new_object = space_map_alloc(mos, tx);
1923 1929 VERIFY3U(new_object, !=, 0);
1924 1930
1925 1931 VERIFY0(space_map_open(&vd->vdev_dtl_sm, mos, new_object,
1926 1932 0, -1ULL, 0, &vd->vdev_dtl_lock));
1927 1933 ASSERT(vd->vdev_dtl_sm != NULL);
1928 1934 }
1929 1935
1930 1936 mutex_init(&rtlock, NULL, MUTEX_DEFAULT, NULL);
1931 1937
1932 1938 rtsync = range_tree_create(NULL, NULL, &rtlock);
1933 1939
1934 1940 mutex_enter(&rtlock);
1935 1941
1936 1942 mutex_enter(&vd->vdev_dtl_lock);
1937 1943 range_tree_walk(rt, range_tree_add, rtsync);
1938 1944 mutex_exit(&vd->vdev_dtl_lock);
1939 1945
1940 1946 space_map_truncate(vd->vdev_dtl_sm, tx);
1941 1947 space_map_write(vd->vdev_dtl_sm, rtsync, SM_ALLOC, tx);
1942 1948 range_tree_vacate(rtsync, NULL, NULL);
1943 1949
1944 1950 range_tree_destroy(rtsync);
1945 1951
1946 1952 mutex_exit(&rtlock);
1947 1953 mutex_destroy(&rtlock);
1948 1954
1949 1955 /*
1950 1956 * If the object for the space map has changed then dirty
1951 1957 * the top level so that we update the config.
1952 1958 */
1953 1959 if (object != space_map_object(vd->vdev_dtl_sm)) {
1954 1960 zfs_dbgmsg("txg %llu, spa %s, DTL old object %llu, "
1955 1961 "new object %llu", txg, spa_name(spa), object,
1956 1962 space_map_object(vd->vdev_dtl_sm));
1957 1963 vdev_config_dirty(vd->vdev_top);
1958 1964 }
1959 1965
1960 1966 dmu_tx_commit(tx);
1961 1967
1962 1968 mutex_enter(&vd->vdev_dtl_lock);
1963 1969 space_map_update(vd->vdev_dtl_sm);
1964 1970 mutex_exit(&vd->vdev_dtl_lock);
1965 1971 }
1966 1972
1967 1973 /*
1968 1974 * Determine whether the specified vdev can be offlined/detached/removed
1969 1975 * without losing data.
1970 1976 */
1971 1977 boolean_t
1972 1978 vdev_dtl_required(vdev_t *vd)
1973 1979 {
1974 1980 spa_t *spa = vd->vdev_spa;
1975 1981 vdev_t *tvd = vd->vdev_top;
1976 1982 uint8_t cant_read = vd->vdev_cant_read;
1977 1983 boolean_t required;
1978 1984
1979 1985 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1980 1986
1981 1987 if (vd == spa->spa_root_vdev || vd == tvd)
1982 1988 return (B_TRUE);
1983 1989
1984 1990 /*
1985 1991 * Temporarily mark the device as unreadable, and then determine
1986 1992 * whether this results in any DTL outages in the top-level vdev.
1987 1993 * If not, we can safely offline/detach/remove the device.
1988 1994 */
1989 1995 vd->vdev_cant_read = B_TRUE;
1990 1996 vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1991 1997 required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
1992 1998 vd->vdev_cant_read = cant_read;
1993 1999 vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1994 2000
1995 2001 if (!required && zio_injection_enabled)
1996 2002 required = !!zio_handle_device_injection(vd, NULL, ECHILD);
1997 2003
1998 2004 return (required);
1999 2005 }
2000 2006
2001 2007 /*
2002 2008 * Determine if resilver is needed, and if so the txg range.
2003 2009 */
2004 2010 boolean_t
2005 2011 vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
2006 2012 {
2007 2013 boolean_t needed = B_FALSE;
2008 2014 uint64_t thismin = UINT64_MAX;
2009 2015 uint64_t thismax = 0;
2010 2016
2011 2017 if (vd->vdev_children == 0) {
2012 2018 mutex_enter(&vd->vdev_dtl_lock);
2013 2019 if (range_tree_space(vd->vdev_dtl[DTL_MISSING]) != 0 &&
2014 2020 vdev_writeable(vd)) {
2015 2021
2016 2022 thismin = vdev_dtl_min(vd);
2017 2023 thismax = vdev_dtl_max(vd);
2018 2024 needed = B_TRUE;
2019 2025 }
2020 2026 mutex_exit(&vd->vdev_dtl_lock);
2021 2027 } else {
2022 2028 for (int c = 0; c < vd->vdev_children; c++) {
2023 2029 vdev_t *cvd = vd->vdev_child[c];
2024 2030 uint64_t cmin, cmax;
2025 2031
2026 2032 if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
2027 2033 thismin = MIN(thismin, cmin);
2028 2034 thismax = MAX(thismax, cmax);
2029 2035 needed = B_TRUE;
2030 2036 }
2031 2037 }
2032 2038 }
2033 2039
2034 2040 if (needed && minp) {
2035 2041 *minp = thismin;
2036 2042 *maxp = thismax;
2037 2043 }
2038 2044 return (needed);
2039 2045 }
2040 2046
2041 2047 void
2042 2048 vdev_load(vdev_t *vd)
2043 2049 {
2044 2050 /*
2045 2051 * Recursively load all children.
2046 2052 */
2047 2053 for (int c = 0; c < vd->vdev_children; c++)
2048 2054 vdev_load(vd->vdev_child[c]);
2049 2055
2050 2056 /*
2051 2057 * If this is a top-level vdev, initialize its metaslabs.
2052 2058 */
2053 2059 if (vd == vd->vdev_top && !vd->vdev_ishole &&
2054 2060 (vd->vdev_ashift == 0 || vd->vdev_asize == 0 ||
2055 2061 vdev_metaslab_init(vd, 0) != 0))
2056 2062 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2057 2063 VDEV_AUX_CORRUPT_DATA);
2058 2064
2059 2065 /*
2060 2066 * If this is a leaf vdev, load its DTL.
2061 2067 */
2062 2068 if (vd->vdev_ops->vdev_op_leaf && vdev_dtl_load(vd) != 0)
2063 2069 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2064 2070 VDEV_AUX_CORRUPT_DATA);
2065 2071 }
2066 2072
2067 2073 /*
2068 2074 * The special vdev case is used for hot spares and l2cache devices. Its
2069 2075 * sole purpose it to set the vdev state for the associated vdev. To do this,
2070 2076 * we make sure that we can open the underlying device, then try to read the
2071 2077 * label, and make sure that the label is sane and that it hasn't been
2072 2078 * repurposed to another pool.
2073 2079 */
2074 2080 int
2075 2081 vdev_validate_aux(vdev_t *vd)
2076 2082 {
2077 2083 nvlist_t *label;
2078 2084 uint64_t guid, version;
2079 2085 uint64_t state;
2080 2086
2081 2087 if (!vdev_readable(vd))
2082 2088 return (0);
2083 2089
2084 2090 if ((label = vdev_label_read_config(vd, -1ULL)) == NULL) {
2085 2091 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2086 2092 VDEV_AUX_CORRUPT_DATA);
2087 2093 return (-1);
2088 2094 }
2089 2095
2090 2096 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
2091 2097 !SPA_VERSION_IS_SUPPORTED(version) ||
2092 2098 nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
2093 2099 guid != vd->vdev_guid ||
2094 2100 nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
2095 2101 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2096 2102 VDEV_AUX_CORRUPT_DATA);
2097 2103 nvlist_free(label);
2098 2104 return (-1);
2099 2105 }
2100 2106
2101 2107 /*
2102 2108 * We don't actually check the pool state here. If it's in fact in
2103 2109 * use by another pool, we update this fact on the fly when requested.
2104 2110 */
2105 2111 nvlist_free(label);
2106 2112 return (0);
2107 2113 }
2108 2114
2109 2115 void
2110 2116 vdev_remove(vdev_t *vd, uint64_t txg)
2111 2117 {
2112 2118 spa_t *spa = vd->vdev_spa;
2113 2119 objset_t *mos = spa->spa_meta_objset;
2114 2120 dmu_tx_t *tx;
2115 2121
2116 2122 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2117 2123
2118 2124 if (vd->vdev_ms != NULL) {
2119 2125 for (int m = 0; m < vd->vdev_ms_count; m++) {
2120 2126 metaslab_t *msp = vd->vdev_ms[m];
2121 2127
2122 2128 if (msp == NULL || msp->ms_sm == NULL)
2123 2129 continue;
2124 2130
2125 2131 mutex_enter(&msp->ms_lock);
2126 2132 VERIFY0(space_map_allocated(msp->ms_sm));
2127 2133 space_map_free(msp->ms_sm, tx);
2128 2134 space_map_close(msp->ms_sm);
2129 2135 msp->ms_sm = NULL;
2130 2136 mutex_exit(&msp->ms_lock);
2131 2137 }
2132 2138 }
2133 2139
2134 2140 if (vd->vdev_ms_array) {
2135 2141 (void) dmu_object_free(mos, vd->vdev_ms_array, tx);
2136 2142 vd->vdev_ms_array = 0;
2137 2143 }
2138 2144 dmu_tx_commit(tx);
2139 2145 }
2140 2146
2141 2147 void
2142 2148 vdev_sync_done(vdev_t *vd, uint64_t txg)
2143 2149 {
2144 2150 metaslab_t *msp;
2145 2151 boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg));
2146 2152
2147 2153 ASSERT(!vd->vdev_ishole);
2148 2154
2149 2155 while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
2150 2156 metaslab_sync_done(msp, txg);
2151 2157
2152 2158 if (reassess)
2153 2159 metaslab_sync_reassess(vd->vdev_mg);
2154 2160 }
2155 2161
2156 2162 void
2157 2163 vdev_sync(vdev_t *vd, uint64_t txg)
2158 2164 {
2159 2165 spa_t *spa = vd->vdev_spa;
2160 2166 vdev_t *lvd;
2161 2167 metaslab_t *msp;
2162 2168 dmu_tx_t *tx;
2163 2169
2164 2170 ASSERT(!vd->vdev_ishole);
2165 2171
2166 2172 if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0) {
2167 2173 ASSERT(vd == vd->vdev_top);
2168 2174 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2169 2175 vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
2170 2176 DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
2171 2177 ASSERT(vd->vdev_ms_array != 0);
2172 2178 vdev_config_dirty(vd);
2173 2179 dmu_tx_commit(tx);
2174 2180 }
2175 2181
2176 2182 /*
2177 2183 * Remove the metadata associated with this vdev once it's empty.
2178 2184 */
2179 2185 if (vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing)
2180 2186 vdev_remove(vd, txg);
2181 2187
2182 2188 while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
2183 2189 metaslab_sync(msp, txg);
2184 2190 (void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
2185 2191 }
2186 2192
2187 2193 while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
2188 2194 vdev_dtl_sync(lvd, txg);
2189 2195
2190 2196 (void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
2191 2197 }
2192 2198
2193 2199 uint64_t
2194 2200 vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
2195 2201 {
2196 2202 return (vd->vdev_ops->vdev_op_asize(vd, psize));
2197 2203 }
2198 2204
2199 2205 /*
2200 2206 * Mark the given vdev faulted. A faulted vdev behaves as if the device could
2201 2207 * not be opened, and no I/O is attempted.
2202 2208 */
2203 2209 int
2204 2210 vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2205 2211 {
2206 2212 vdev_t *vd, *tvd;
2207 2213
2208 2214 spa_vdev_state_enter(spa, SCL_NONE);
2209 2215
2210 2216 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2211 2217 return (spa_vdev_state_exit(spa, NULL, ENODEV));
2212 2218
2213 2219 if (!vd->vdev_ops->vdev_op_leaf)
2214 2220 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2215 2221
2216 2222 tvd = vd->vdev_top;
2217 2223
2218 2224 /*
2219 2225 * We don't directly use the aux state here, but if we do a
2220 2226 * vdev_reopen(), we need this value to be present to remember why we
2221 2227 * were faulted.
2222 2228 */
2223 2229 vd->vdev_label_aux = aux;
2224 2230
2225 2231 /*
2226 2232 * Faulted state takes precedence over degraded.
2227 2233 */
2228 2234 vd->vdev_delayed_close = B_FALSE;
2229 2235 vd->vdev_faulted = 1ULL;
2230 2236 vd->vdev_degraded = 0ULL;
2231 2237 vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
2232 2238
2233 2239 /*
2234 2240 * If this device has the only valid copy of the data, then
2235 2241 * back off and simply mark the vdev as degraded instead.
2236 2242 */
2237 2243 if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) {
2238 2244 vd->vdev_degraded = 1ULL;
2239 2245 vd->vdev_faulted = 0ULL;
2240 2246
2241 2247 /*
2242 2248 * If we reopen the device and it's not dead, only then do we
2243 2249 * mark it degraded.
2244 2250 */
2245 2251 vdev_reopen(tvd);
2246 2252
2247 2253 if (vdev_readable(vd))
2248 2254 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
2249 2255 }
2250 2256
2251 2257 return (spa_vdev_state_exit(spa, vd, 0));
2252 2258 }
2253 2259
2254 2260 /*
2255 2261 * Mark the given vdev degraded. A degraded vdev is purely an indication to the
2256 2262 * user that something is wrong. The vdev continues to operate as normal as far
2257 2263 * as I/O is concerned.
2258 2264 */
2259 2265 int
2260 2266 vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2261 2267 {
2262 2268 vdev_t *vd;
2263 2269
2264 2270 spa_vdev_state_enter(spa, SCL_NONE);
2265 2271
2266 2272 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2267 2273 return (spa_vdev_state_exit(spa, NULL, ENODEV));
2268 2274
2269 2275 if (!vd->vdev_ops->vdev_op_leaf)
2270 2276 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2271 2277
2272 2278 /*
2273 2279 * If the vdev is already faulted, then don't do anything.
2274 2280 */
2275 2281 if (vd->vdev_faulted || vd->vdev_degraded)
2276 2282 return (spa_vdev_state_exit(spa, NULL, 0));
2277 2283
2278 2284 vd->vdev_degraded = 1ULL;
2279 2285 if (!vdev_is_dead(vd))
2280 2286 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
2281 2287 aux);
2282 2288
2283 2289 return (spa_vdev_state_exit(spa, vd, 0));
2284 2290 }
2285 2291
2286 2292 /*
2287 2293 * Online the given vdev.
2288 2294 *
2289 2295 * If 'ZFS_ONLINE_UNSPARE' is set, it implies two things. First, any attached
2290 2296 * spare device should be detached when the device finishes resilvering.
2291 2297 * Second, the online should be treated like a 'test' online case, so no FMA
2292 2298 * events are generated if the device fails to open.
2293 2299 */
2294 2300 int
2295 2301 vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
2296 2302 {
2297 2303 vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
2298 2304
2299 2305 spa_vdev_state_enter(spa, SCL_NONE);
2300 2306
2301 2307 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2302 2308 return (spa_vdev_state_exit(spa, NULL, ENODEV));
2303 2309
2304 2310 if (!vd->vdev_ops->vdev_op_leaf)
2305 2311 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2306 2312
2307 2313 tvd = vd->vdev_top;
2308 2314 vd->vdev_offline = B_FALSE;
2309 2315 vd->vdev_tmpoffline = B_FALSE;
2310 2316 vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
2311 2317 vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
2312 2318
2313 2319 /* XXX - L2ARC 1.0 does not support expansion */
2314 2320 if (!vd->vdev_aux) {
2315 2321 for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2316 2322 pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
2317 2323 }
2318 2324
2319 2325 vdev_reopen(tvd);
2320 2326 vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
2321 2327
2322 2328 if (!vd->vdev_aux) {
2323 2329 for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2324 2330 pvd->vdev_expanding = B_FALSE;
2325 2331 }
2326 2332
2327 2333 if (newstate)
2328 2334 *newstate = vd->vdev_state;
2329 2335 if ((flags & ZFS_ONLINE_UNSPARE) &&
2330 2336 !vdev_is_dead(vd) && vd->vdev_parent &&
2331 2337 vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2332 2338 vd->vdev_parent->vdev_child[0] == vd)
2333 2339 vd->vdev_unspare = B_TRUE;
2334 2340
2335 2341 if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
2336 2342
2337 2343 /* XXX - L2ARC 1.0 does not support expansion */
2338 2344 if (vd->vdev_aux)
2339 2345 return (spa_vdev_state_exit(spa, vd, ENOTSUP));
2340 2346 spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
2341 2347 }
2342 2348 return (spa_vdev_state_exit(spa, vd, 0));
2343 2349 }
2344 2350
2345 2351 static int
2346 2352 vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
2347 2353 {
2348 2354 vdev_t *vd, *tvd;
2349 2355 int error = 0;
2350 2356 uint64_t generation;
2351 2357 metaslab_group_t *mg;
2352 2358
2353 2359 top:
2354 2360 spa_vdev_state_enter(spa, SCL_ALLOC);
2355 2361
2356 2362 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2357 2363 return (spa_vdev_state_exit(spa, NULL, ENODEV));
2358 2364
2359 2365 if (!vd->vdev_ops->vdev_op_leaf)
2360 2366 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2361 2367
2362 2368 tvd = vd->vdev_top;
2363 2369 mg = tvd->vdev_mg;
2364 2370 generation = spa->spa_config_generation + 1;
2365 2371
2366 2372 /*
2367 2373 * If the device isn't already offline, try to offline it.
2368 2374 */
2369 2375 if (!vd->vdev_offline) {
2370 2376 /*
2371 2377 * If this device has the only valid copy of some data,
2372 2378 * don't allow it to be offlined. Log devices are always
2373 2379 * expendable.
2374 2380 */
2375 2381 if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2376 2382 vdev_dtl_required(vd))
2377 2383 return (spa_vdev_state_exit(spa, NULL, EBUSY));
2378 2384
2379 2385 /*
2380 2386 * If the top-level is a slog and it has had allocations
2381 2387 * then proceed. We check that the vdev's metaslab group
2382 2388 * is not NULL since it's possible that we may have just
2383 2389 * added this vdev but not yet initialized its metaslabs.
2384 2390 */
2385 2391 if (tvd->vdev_islog && mg != NULL) {
2386 2392 /*
2387 2393 * Prevent any future allocations.
2388 2394 */
2389 2395 metaslab_group_passivate(mg);
2390 2396 (void) spa_vdev_state_exit(spa, vd, 0);
2391 2397
2392 2398 error = spa_offline_log(spa);
2393 2399
2394 2400 spa_vdev_state_enter(spa, SCL_ALLOC);
2395 2401
2396 2402 /*
2397 2403 * Check to see if the config has changed.
2398 2404 */
2399 2405 if (error || generation != spa->spa_config_generation) {
2400 2406 metaslab_group_activate(mg);
2401 2407 if (error)
2402 2408 return (spa_vdev_state_exit(spa,
2403 2409 vd, error));
2404 2410 (void) spa_vdev_state_exit(spa, vd, 0);
2405 2411 goto top;
2406 2412 }
2407 2413 ASSERT0(tvd->vdev_stat.vs_alloc);
2408 2414 }
2409 2415
2410 2416 /*
2411 2417 * Offline this device and reopen its top-level vdev.
2412 2418 * If the top-level vdev is a log device then just offline
2413 2419 * it. Otherwise, if this action results in the top-level
2414 2420 * vdev becoming unusable, undo it and fail the request.
2415 2421 */
2416 2422 vd->vdev_offline = B_TRUE;
2417 2423 vdev_reopen(tvd);
2418 2424
2419 2425 if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2420 2426 vdev_is_dead(tvd)) {
2421 2427 vd->vdev_offline = B_FALSE;
2422 2428 vdev_reopen(tvd);
2423 2429 return (spa_vdev_state_exit(spa, NULL, EBUSY));
2424 2430 }
2425 2431
2426 2432 /*
2427 2433 * Add the device back into the metaslab rotor so that
2428 2434 * once we online the device it's open for business.
2429 2435 */
2430 2436 if (tvd->vdev_islog && mg != NULL)
2431 2437 metaslab_group_activate(mg);
2432 2438 }
2433 2439
2434 2440 vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
2435 2441
2436 2442 return (spa_vdev_state_exit(spa, vd, 0));
2437 2443 }
2438 2444
2439 2445 int
2440 2446 vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
2441 2447 {
2442 2448 int error;
2443 2449
2444 2450 mutex_enter(&spa->spa_vdev_top_lock);
2445 2451 error = vdev_offline_locked(spa, guid, flags);
2446 2452 mutex_exit(&spa->spa_vdev_top_lock);
2447 2453
2448 2454 return (error);
2449 2455 }
2450 2456
2451 2457 /*
2452 2458 * Clear the error counts associated with this vdev. Unlike vdev_online() and
2453 2459 * vdev_offline(), we assume the spa config is locked. We also clear all
2454 2460 * children. If 'vd' is NULL, then the user wants to clear all vdevs.
2455 2461 */
2456 2462 void
2457 2463 vdev_clear(spa_t *spa, vdev_t *vd)
2458 2464 {
2459 2465 vdev_t *rvd = spa->spa_root_vdev;
2460 2466
2461 2467 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2462 2468
2463 2469 if (vd == NULL)
2464 2470 vd = rvd;
2465 2471
2466 2472 vd->vdev_stat.vs_read_errors = 0;
2467 2473 vd->vdev_stat.vs_write_errors = 0;
2468 2474 vd->vdev_stat.vs_checksum_errors = 0;
2469 2475
2470 2476 for (int c = 0; c < vd->vdev_children; c++)
2471 2477 vdev_clear(spa, vd->vdev_child[c]);
2472 2478
2473 2479 /*
2474 2480 * If we're in the FAULTED state or have experienced failed I/O, then
2475 2481 * clear the persistent state and attempt to reopen the device. We
2476 2482 * also mark the vdev config dirty, so that the new faulted state is
2477 2483 * written out to disk.
2478 2484 */
2479 2485 if (vd->vdev_faulted || vd->vdev_degraded ||
2480 2486 !vdev_readable(vd) || !vdev_writeable(vd)) {
2481 2487
2482 2488 /*
2483 2489 * When reopening in reponse to a clear event, it may be due to
2484 2490 * a fmadm repair request. In this case, if the device is
2485 2491 * still broken, we want to still post the ereport again.
2486 2492 */
2487 2493 vd->vdev_forcefault = B_TRUE;
2488 2494
2489 2495 vd->vdev_faulted = vd->vdev_degraded = 0ULL;
2490 2496 vd->vdev_cant_read = B_FALSE;
2491 2497 vd->vdev_cant_write = B_FALSE;
2492 2498
2493 2499 vdev_reopen(vd == rvd ? rvd : vd->vdev_top);
2494 2500
2495 2501 vd->vdev_forcefault = B_FALSE;
2496 2502
2497 2503 if (vd != rvd && vdev_writeable(vd->vdev_top))
2498 2504 vdev_state_dirty(vd->vdev_top);
2499 2505
2500 2506 if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
2501 2507 spa_async_request(spa, SPA_ASYNC_RESILVER);
2502 2508
2503 2509 spa_event_notify(spa, vd, ESC_ZFS_VDEV_CLEAR);
2504 2510 }
2505 2511
2506 2512 /*
2507 2513 * When clearing a FMA-diagnosed fault, we always want to
2508 2514 * unspare the device, as we assume that the original spare was
2509 2515 * done in response to the FMA fault.
2510 2516 */
2511 2517 if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
2512 2518 vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2513 2519 vd->vdev_parent->vdev_child[0] == vd)
2514 2520 vd->vdev_unspare = B_TRUE;
2515 2521 }
2516 2522
2517 2523 boolean_t
2518 2524 vdev_is_dead(vdev_t *vd)
2519 2525 {
2520 2526 /*
2521 2527 * Holes and missing devices are always considered "dead".
2522 2528 * This simplifies the code since we don't have to check for
2523 2529 * these types of devices in the various code paths.
2524 2530 * Instead we rely on the fact that we skip over dead devices
2525 2531 * before issuing I/O to them.
2526 2532 */
2527 2533 return (vd->vdev_state < VDEV_STATE_DEGRADED || vd->vdev_ishole ||
2528 2534 vd->vdev_ops == &vdev_missing_ops);
2529 2535 }
2530 2536
2531 2537 boolean_t
2532 2538 vdev_readable(vdev_t *vd)
2533 2539 {
2534 2540 return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
2535 2541 }
2536 2542
2537 2543 boolean_t
2538 2544 vdev_writeable(vdev_t *vd)
2539 2545 {
2540 2546 return (!vdev_is_dead(vd) && !vd->vdev_cant_write);
2541 2547 }
2542 2548
2543 2549 boolean_t
2544 2550 vdev_allocatable(vdev_t *vd)
2545 2551 {
2546 2552 uint64_t state = vd->vdev_state;
2547 2553
2548 2554 /*
2549 2555 * We currently allow allocations from vdevs which may be in the
2550 2556 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
2551 2557 * fails to reopen then we'll catch it later when we're holding
2552 2558 * the proper locks. Note that we have to get the vdev state
2553 2559 * in a local variable because although it changes atomically,
2554 2560 * we're asking two separate questions about it.
2555 2561 */
2556 2562 return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
2557 2563 !vd->vdev_cant_write && !vd->vdev_ishole);
2558 2564 }
2559 2565
2560 2566 boolean_t
2561 2567 vdev_accessible(vdev_t *vd, zio_t *zio)
2562 2568 {
2563 2569 ASSERT(zio->io_vd == vd);
2564 2570
2565 2571 if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
2566 2572 return (B_FALSE);
2567 2573
2568 2574 if (zio->io_type == ZIO_TYPE_READ)
2569 2575 return (!vd->vdev_cant_read);
2570 2576
2571 2577 if (zio->io_type == ZIO_TYPE_WRITE)
2572 2578 return (!vd->vdev_cant_write);
2573 2579
2574 2580 return (B_TRUE);
2575 2581 }
2576 2582
2577 2583 /*
2578 2584 * Get statistics for the given vdev.
2579 2585 */
2580 2586 void
2581 2587 vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
2582 2588 {
2583 2589 vdev_t *rvd = vd->vdev_spa->spa_root_vdev;
2584 2590
2585 2591 mutex_enter(&vd->vdev_stat_lock);
2586 2592 bcopy(&vd->vdev_stat, vs, sizeof (*vs));
2587 2593 vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
2588 2594 vs->vs_state = vd->vdev_state;
2589 2595 vs->vs_rsize = vdev_get_min_asize(vd);
2590 2596 if (vd->vdev_ops->vdev_op_leaf)
2591 2597 vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
2592 2598 vs->vs_esize = vd->vdev_max_asize - vd->vdev_asize;
2593 2599 mutex_exit(&vd->vdev_stat_lock);
2594 2600
2595 2601 /*
2596 2602 * If we're getting stats on the root vdev, aggregate the I/O counts
2597 2603 * over all top-level vdevs (i.e. the direct children of the root).
2598 2604 */
2599 2605 if (vd == rvd) {
2600 2606 for (int c = 0; c < rvd->vdev_children; c++) {
2601 2607 vdev_t *cvd = rvd->vdev_child[c];
2602 2608 vdev_stat_t *cvs = &cvd->vdev_stat;
2603 2609
2604 2610 mutex_enter(&vd->vdev_stat_lock);
2605 2611 for (int t = 0; t < ZIO_TYPES; t++) {
2606 2612 vs->vs_ops[t] += cvs->vs_ops[t];
2607 2613 vs->vs_bytes[t] += cvs->vs_bytes[t];
2608 2614 }
2609 2615 cvs->vs_scan_removing = cvd->vdev_removing;
2610 2616 mutex_exit(&vd->vdev_stat_lock);
2611 2617 }
2612 2618 }
2613 2619 }
2614 2620
2615 2621 void
2616 2622 vdev_clear_stats(vdev_t *vd)
2617 2623 {
2618 2624 mutex_enter(&vd->vdev_stat_lock);
2619 2625 vd->vdev_stat.vs_space = 0;
2620 2626 vd->vdev_stat.vs_dspace = 0;
2621 2627 vd->vdev_stat.vs_alloc = 0;
2622 2628 mutex_exit(&vd->vdev_stat_lock);
2623 2629 }
2624 2630
2625 2631 void
2626 2632 vdev_scan_stat_init(vdev_t *vd)
2627 2633 {
2628 2634 vdev_stat_t *vs = &vd->vdev_stat;
2629 2635
2630 2636 for (int c = 0; c < vd->vdev_children; c++)
2631 2637 vdev_scan_stat_init(vd->vdev_child[c]);
2632 2638
2633 2639 mutex_enter(&vd->vdev_stat_lock);
2634 2640 vs->vs_scan_processed = 0;
2635 2641 mutex_exit(&vd->vdev_stat_lock);
2636 2642 }
2637 2643
2638 2644 void
2639 2645 vdev_stat_update(zio_t *zio, uint64_t psize)
2640 2646 {
2641 2647 spa_t *spa = zio->io_spa;
2642 2648 vdev_t *rvd = spa->spa_root_vdev;
2643 2649 vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
2644 2650 vdev_t *pvd;
2645 2651 uint64_t txg = zio->io_txg;
2646 2652 vdev_stat_t *vs = &vd->vdev_stat;
2647 2653 zio_type_t type = zio->io_type;
2648 2654 int flags = zio->io_flags;
2649 2655
2650 2656 /*
2651 2657 * If this i/o is a gang leader, it didn't do any actual work.
2652 2658 */
2653 2659 if (zio->io_gang_tree)
2654 2660 return;
2655 2661
2656 2662 if (zio->io_error == 0) {
2657 2663 /*
2658 2664 * If this is a root i/o, don't count it -- we've already
2659 2665 * counted the top-level vdevs, and vdev_get_stats() will
2660 2666 * aggregate them when asked. This reduces contention on
2661 2667 * the root vdev_stat_lock and implicitly handles blocks
2662 2668 * that compress away to holes, for which there is no i/o.
2663 2669 * (Holes never create vdev children, so all the counters
2664 2670 * remain zero, which is what we want.)
2665 2671 *
2666 2672 * Note: this only applies to successful i/o (io_error == 0)
2667 2673 * because unlike i/o counts, errors are not additive.
2668 2674 * When reading a ditto block, for example, failure of
2669 2675 * one top-level vdev does not imply a root-level error.
2670 2676 */
2671 2677 if (vd == rvd)
2672 2678 return;
2673 2679
2674 2680 ASSERT(vd == zio->io_vd);
2675 2681
2676 2682 if (flags & ZIO_FLAG_IO_BYPASS)
2677 2683 return;
2678 2684
2679 2685 mutex_enter(&vd->vdev_stat_lock);
2680 2686
2681 2687 if (flags & ZIO_FLAG_IO_REPAIR) {
2682 2688 if (flags & ZIO_FLAG_SCAN_THREAD) {
2683 2689 dsl_scan_phys_t *scn_phys =
2684 2690 &spa->spa_dsl_pool->dp_scan->scn_phys;
2685 2691 uint64_t *processed = &scn_phys->scn_processed;
2686 2692
2687 2693 /* XXX cleanup? */
2688 2694 if (vd->vdev_ops->vdev_op_leaf)
2689 2695 atomic_add_64(processed, psize);
2690 2696 vs->vs_scan_processed += psize;
2691 2697 }
2692 2698
2693 2699 if (flags & ZIO_FLAG_SELF_HEAL)
2694 2700 vs->vs_self_healed += psize;
2695 2701 }
2696 2702
2697 2703 vs->vs_ops[type]++;
2698 2704 vs->vs_bytes[type] += psize;
2699 2705
2700 2706 mutex_exit(&vd->vdev_stat_lock);
2701 2707 return;
2702 2708 }
2703 2709
2704 2710 if (flags & ZIO_FLAG_SPECULATIVE)
2705 2711 return;
2706 2712
2707 2713 /*
2708 2714 * If this is an I/O error that is going to be retried, then ignore the
2709 2715 * error. Otherwise, the user may interpret B_FAILFAST I/O errors as
2710 2716 * hard errors, when in reality they can happen for any number of
2711 2717 * innocuous reasons (bus resets, MPxIO link failure, etc).
2712 2718 */
2713 2719 if (zio->io_error == EIO &&
2714 2720 !(zio->io_flags & ZIO_FLAG_IO_RETRY))
2715 2721 return;
2716 2722
2717 2723 /*
2718 2724 * Intent logs writes won't propagate their error to the root
2719 2725 * I/O so don't mark these types of failures as pool-level
2720 2726 * errors.
2721 2727 */
2722 2728 if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
2723 2729 return;
2724 2730
2725 2731 mutex_enter(&vd->vdev_stat_lock);
2726 2732 if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
2727 2733 if (zio->io_error == ECKSUM)
2728 2734 vs->vs_checksum_errors++;
2729 2735 else
2730 2736 vs->vs_read_errors++;
2731 2737 }
2732 2738 if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
2733 2739 vs->vs_write_errors++;
2734 2740 mutex_exit(&vd->vdev_stat_lock);
2735 2741
2736 2742 if (type == ZIO_TYPE_WRITE && txg != 0 &&
2737 2743 (!(flags & ZIO_FLAG_IO_REPAIR) ||
2738 2744 (flags & ZIO_FLAG_SCAN_THREAD) ||
2739 2745 spa->spa_claiming)) {
2740 2746 /*
2741 2747 * This is either a normal write (not a repair), or it's
2742 2748 * a repair induced by the scrub thread, or it's a repair
2743 2749 * made by zil_claim() during spa_load() in the first txg.
2744 2750 * In the normal case, we commit the DTL change in the same
2745 2751 * txg as the block was born. In the scrub-induced repair
2746 2752 * case, we know that scrubs run in first-pass syncing context,
2747 2753 * so we commit the DTL change in spa_syncing_txg(spa).
2748 2754 * In the zil_claim() case, we commit in spa_first_txg(spa).
2749 2755 *
2750 2756 * We currently do not make DTL entries for failed spontaneous
2751 2757 * self-healing writes triggered by normal (non-scrubbing)
2752 2758 * reads, because we have no transactional context in which to
2753 2759 * do so -- and it's not clear that it'd be desirable anyway.
2754 2760 */
2755 2761 if (vd->vdev_ops->vdev_op_leaf) {
2756 2762 uint64_t commit_txg = txg;
2757 2763 if (flags & ZIO_FLAG_SCAN_THREAD) {
2758 2764 ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2759 2765 ASSERT(spa_sync_pass(spa) == 1);
2760 2766 vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
2761 2767 commit_txg = spa_syncing_txg(spa);
2762 2768 } else if (spa->spa_claiming) {
2763 2769 ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2764 2770 commit_txg = spa_first_txg(spa);
2765 2771 }
2766 2772 ASSERT(commit_txg >= spa_syncing_txg(spa));
2767 2773 if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
2768 2774 return;
2769 2775 for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2770 2776 vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
2771 2777 vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
2772 2778 }
2773 2779 if (vd != rvd)
2774 2780 vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
2775 2781 }
2776 2782 }
2777 2783
2778 2784 /*
2779 2785 * Update the in-core space usage stats for this vdev, its metaslab class,
2780 2786 * and the root vdev.
2781 2787 */
2782 2788 void
2783 2789 vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
2784 2790 int64_t space_delta)
2785 2791 {
2786 2792 int64_t dspace_delta = space_delta;
2787 2793 spa_t *spa = vd->vdev_spa;
2788 2794 vdev_t *rvd = spa->spa_root_vdev;
2789 2795 metaslab_group_t *mg = vd->vdev_mg;
2790 2796 metaslab_class_t *mc = mg ? mg->mg_class : NULL;
2791 2797
2792 2798 ASSERT(vd == vd->vdev_top);
2793 2799
2794 2800 /*
2795 2801 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
2796 2802 * factor. We must calculate this here and not at the root vdev
2797 2803 * because the root vdev's psize-to-asize is simply the max of its
2798 2804 * childrens', thus not accurate enough for us.
2799 2805 */
2800 2806 ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
2801 2807 ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
2802 2808 dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
2803 2809 vd->vdev_deflate_ratio;
2804 2810
2805 2811 mutex_enter(&vd->vdev_stat_lock);
2806 2812 vd->vdev_stat.vs_alloc += alloc_delta;
2807 2813 vd->vdev_stat.vs_space += space_delta;
2808 2814 vd->vdev_stat.vs_dspace += dspace_delta;
2809 2815 mutex_exit(&vd->vdev_stat_lock);
2810 2816
2811 2817 if (mc == spa_normal_class(spa)) {
2812 2818 mutex_enter(&rvd->vdev_stat_lock);
2813 2819 rvd->vdev_stat.vs_alloc += alloc_delta;
2814 2820 rvd->vdev_stat.vs_space += space_delta;
2815 2821 rvd->vdev_stat.vs_dspace += dspace_delta;
2816 2822 mutex_exit(&rvd->vdev_stat_lock);
2817 2823 }
2818 2824
2819 2825 if (mc != NULL) {
2820 2826 ASSERT(rvd == vd->vdev_parent);
2821 2827 ASSERT(vd->vdev_ms_count != 0);
2822 2828
2823 2829 metaslab_class_space_update(mc,
2824 2830 alloc_delta, defer_delta, space_delta, dspace_delta);
2825 2831 }
2826 2832 }
2827 2833
2828 2834 /*
2829 2835 * Mark a top-level vdev's config as dirty, placing it on the dirty list
2830 2836 * so that it will be written out next time the vdev configuration is synced.
2831 2837 * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
2832 2838 */
2833 2839 void
2834 2840 vdev_config_dirty(vdev_t *vd)
2835 2841 {
2836 2842 spa_t *spa = vd->vdev_spa;
2837 2843 vdev_t *rvd = spa->spa_root_vdev;
2838 2844 int c;
2839 2845
2840 2846 ASSERT(spa_writeable(spa));
2841 2847
2842 2848 /*
2843 2849 * If this is an aux vdev (as with l2cache and spare devices), then we
2844 2850 * update the vdev config manually and set the sync flag.
2845 2851 */
2846 2852 if (vd->vdev_aux != NULL) {
2847 2853 spa_aux_vdev_t *sav = vd->vdev_aux;
2848 2854 nvlist_t **aux;
2849 2855 uint_t naux;
2850 2856
2851 2857 for (c = 0; c < sav->sav_count; c++) {
2852 2858 if (sav->sav_vdevs[c] == vd)
2853 2859 break;
2854 2860 }
2855 2861
2856 2862 if (c == sav->sav_count) {
2857 2863 /*
2858 2864 * We're being removed. There's nothing more to do.
2859 2865 */
2860 2866 ASSERT(sav->sav_sync == B_TRUE);
2861 2867 return;
2862 2868 }
2863 2869
2864 2870 sav->sav_sync = B_TRUE;
2865 2871
2866 2872 if (nvlist_lookup_nvlist_array(sav->sav_config,
2867 2873 ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
2868 2874 VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
2869 2875 ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
2870 2876 }
2871 2877
2872 2878 ASSERT(c < naux);
2873 2879
2874 2880 /*
2875 2881 * Setting the nvlist in the middle if the array is a little
2876 2882 * sketchy, but it will work.
2877 2883 */
2878 2884 nvlist_free(aux[c]);
2879 2885 aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0);
2880 2886
2881 2887 return;
2882 2888 }
2883 2889
2884 2890 /*
2885 2891 * The dirty list is protected by the SCL_CONFIG lock. The caller
2886 2892 * must either hold SCL_CONFIG as writer, or must be the sync thread
2887 2893 * (which holds SCL_CONFIG as reader). There's only one sync thread,
2888 2894 * so this is sufficient to ensure mutual exclusion.
2889 2895 */
2890 2896 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2891 2897 (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2892 2898 spa_config_held(spa, SCL_CONFIG, RW_READER)));
2893 2899
2894 2900 if (vd == rvd) {
2895 2901 for (c = 0; c < rvd->vdev_children; c++)
2896 2902 vdev_config_dirty(rvd->vdev_child[c]);
2897 2903 } else {
2898 2904 ASSERT(vd == vd->vdev_top);
2899 2905
2900 2906 if (!list_link_active(&vd->vdev_config_dirty_node) &&
2901 2907 !vd->vdev_ishole)
2902 2908 list_insert_head(&spa->spa_config_dirty_list, vd);
2903 2909 }
2904 2910 }
2905 2911
2906 2912 void
2907 2913 vdev_config_clean(vdev_t *vd)
2908 2914 {
2909 2915 spa_t *spa = vd->vdev_spa;
2910 2916
2911 2917 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2912 2918 (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2913 2919 spa_config_held(spa, SCL_CONFIG, RW_READER)));
2914 2920
2915 2921 ASSERT(list_link_active(&vd->vdev_config_dirty_node));
2916 2922 list_remove(&spa->spa_config_dirty_list, vd);
2917 2923 }
2918 2924
2919 2925 /*
2920 2926 * Mark a top-level vdev's state as dirty, so that the next pass of
2921 2927 * spa_sync() can convert this into vdev_config_dirty(). We distinguish
2922 2928 * the state changes from larger config changes because they require
2923 2929 * much less locking, and are often needed for administrative actions.
2924 2930 */
2925 2931 void
2926 2932 vdev_state_dirty(vdev_t *vd)
2927 2933 {
2928 2934 spa_t *spa = vd->vdev_spa;
2929 2935
2930 2936 ASSERT(spa_writeable(spa));
2931 2937 ASSERT(vd == vd->vdev_top);
2932 2938
2933 2939 /*
2934 2940 * The state list is protected by the SCL_STATE lock. The caller
2935 2941 * must either hold SCL_STATE as writer, or must be the sync thread
2936 2942 * (which holds SCL_STATE as reader). There's only one sync thread,
2937 2943 * so this is sufficient to ensure mutual exclusion.
2938 2944 */
2939 2945 ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2940 2946 (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2941 2947 spa_config_held(spa, SCL_STATE, RW_READER)));
2942 2948
2943 2949 if (!list_link_active(&vd->vdev_state_dirty_node) && !vd->vdev_ishole)
2944 2950 list_insert_head(&spa->spa_state_dirty_list, vd);
2945 2951 }
2946 2952
2947 2953 void
2948 2954 vdev_state_clean(vdev_t *vd)
2949 2955 {
2950 2956 spa_t *spa = vd->vdev_spa;
2951 2957
2952 2958 ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2953 2959 (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2954 2960 spa_config_held(spa, SCL_STATE, RW_READER)));
2955 2961
2956 2962 ASSERT(list_link_active(&vd->vdev_state_dirty_node));
2957 2963 list_remove(&spa->spa_state_dirty_list, vd);
2958 2964 }
2959 2965
2960 2966 /*
2961 2967 * Propagate vdev state up from children to parent.
2962 2968 */
2963 2969 void
2964 2970 vdev_propagate_state(vdev_t *vd)
2965 2971 {
2966 2972 spa_t *spa = vd->vdev_spa;
2967 2973 vdev_t *rvd = spa->spa_root_vdev;
2968 2974 int degraded = 0, faulted = 0;
2969 2975 int corrupted = 0;
2970 2976 vdev_t *child;
2971 2977
2972 2978 if (vd->vdev_children > 0) {
2973 2979 for (int c = 0; c < vd->vdev_children; c++) {
2974 2980 child = vd->vdev_child[c];
2975 2981
2976 2982 /*
2977 2983 * Don't factor holes into the decision.
2978 2984 */
2979 2985 if (child->vdev_ishole)
2980 2986 continue;
2981 2987
2982 2988 if (!vdev_readable(child) ||
2983 2989 (!vdev_writeable(child) && spa_writeable(spa))) {
2984 2990 /*
2985 2991 * Root special: if there is a top-level log
2986 2992 * device, treat the root vdev as if it were
2987 2993 * degraded.
2988 2994 */
2989 2995 if (child->vdev_islog && vd == rvd)
2990 2996 degraded++;
2991 2997 else
2992 2998 faulted++;
2993 2999 } else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
2994 3000 degraded++;
2995 3001 }
2996 3002
2997 3003 if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
2998 3004 corrupted++;
2999 3005 }
3000 3006
3001 3007 vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
3002 3008
3003 3009 /*
3004 3010 * Root special: if there is a top-level vdev that cannot be
3005 3011 * opened due to corrupted metadata, then propagate the root
3006 3012 * vdev's aux state as 'corrupt' rather than 'insufficient
3007 3013 * replicas'.
3008 3014 */
3009 3015 if (corrupted && vd == rvd &&
3010 3016 rvd->vdev_state == VDEV_STATE_CANT_OPEN)
3011 3017 vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
3012 3018 VDEV_AUX_CORRUPT_DATA);
3013 3019 }
3014 3020
3015 3021 if (vd->vdev_parent)
3016 3022 vdev_propagate_state(vd->vdev_parent);
3017 3023 }
3018 3024
3019 3025 /*
3020 3026 * Set a vdev's state. If this is during an open, we don't update the parent
3021 3027 * state, because we're in the process of opening children depth-first.
3022 3028 * Otherwise, we propagate the change to the parent.
3023 3029 *
3024 3030 * If this routine places a device in a faulted state, an appropriate ereport is
3025 3031 * generated.
3026 3032 */
3027 3033 void
3028 3034 vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
3029 3035 {
3030 3036 uint64_t save_state;
3031 3037 spa_t *spa = vd->vdev_spa;
3032 3038
3033 3039 if (state == vd->vdev_state) {
3034 3040 vd->vdev_stat.vs_aux = aux;
3035 3041 return;
3036 3042 }
3037 3043
3038 3044 save_state = vd->vdev_state;
3039 3045
3040 3046 vd->vdev_state = state;
3041 3047 vd->vdev_stat.vs_aux = aux;
3042 3048
3043 3049 /*
3044 3050 * If we are setting the vdev state to anything but an open state, then
3045 3051 * always close the underlying device unless the device has requested
3046 3052 * a delayed close (i.e. we're about to remove or fault the device).
3047 3053 * Otherwise, we keep accessible but invalid devices open forever.
3048 3054 * We don't call vdev_close() itself, because that implies some extra
3049 3055 * checks (offline, etc) that we don't want here. This is limited to
3050 3056 * leaf devices, because otherwise closing the device will affect other
3051 3057 * children.
3052 3058 */
3053 3059 if (!vd->vdev_delayed_close && vdev_is_dead(vd) &&
3054 3060 vd->vdev_ops->vdev_op_leaf)
3055 3061 vd->vdev_ops->vdev_op_close(vd);
3056 3062
3057 3063 /*
3058 3064 * If we have brought this vdev back into service, we need
3059 3065 * to notify fmd so that it can gracefully repair any outstanding
3060 3066 * cases due to a missing device. We do this in all cases, even those
3061 3067 * that probably don't correlate to a repaired fault. This is sure to
3062 3068 * catch all cases, and we let the zfs-retire agent sort it out. If
3063 3069 * this is a transient state it's OK, as the retire agent will
3064 3070 * double-check the state of the vdev before repairing it.
3065 3071 */
3066 3072 if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf &&
3067 3073 vd->vdev_prevstate != state)
3068 3074 zfs_post_state_change(spa, vd);
3069 3075
3070 3076 if (vd->vdev_removed &&
3071 3077 state == VDEV_STATE_CANT_OPEN &&
3072 3078 (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
3073 3079 /*
3074 3080 * If the previous state is set to VDEV_STATE_REMOVED, then this
3075 3081 * device was previously marked removed and someone attempted to
3076 3082 * reopen it. If this failed due to a nonexistent device, then
3077 3083 * keep the device in the REMOVED state. We also let this be if
3078 3084 * it is one of our special test online cases, which is only
3079 3085 * attempting to online the device and shouldn't generate an FMA
3080 3086 * fault.
3081 3087 */
3082 3088 vd->vdev_state = VDEV_STATE_REMOVED;
3083 3089 vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
3084 3090 } else if (state == VDEV_STATE_REMOVED) {
3085 3091 vd->vdev_removed = B_TRUE;
3086 3092 } else if (state == VDEV_STATE_CANT_OPEN) {
3087 3093 /*
3088 3094 * If we fail to open a vdev during an import or recovery, we
3089 3095 * mark it as "not available", which signifies that it was
3090 3096 * never there to begin with. Failure to open such a device
3091 3097 * is not considered an error.
3092 3098 */
3093 3099 if ((spa_load_state(spa) == SPA_LOAD_IMPORT ||
3094 3100 spa_load_state(spa) == SPA_LOAD_RECOVER) &&
3095 3101 vd->vdev_ops->vdev_op_leaf)
3096 3102 vd->vdev_not_present = 1;
3097 3103
3098 3104 /*
3099 3105 * Post the appropriate ereport. If the 'prevstate' field is
3100 3106 * set to something other than VDEV_STATE_UNKNOWN, it indicates
3101 3107 * that this is part of a vdev_reopen(). In this case, we don't
3102 3108 * want to post the ereport if the device was already in the
3103 3109 * CANT_OPEN state beforehand.
3104 3110 *
3105 3111 * If the 'checkremove' flag is set, then this is an attempt to
3106 3112 * online the device in response to an insertion event. If we
3107 3113 * hit this case, then we have detected an insertion event for a
3108 3114 * faulted or offline device that wasn't in the removed state.
3109 3115 * In this scenario, we don't post an ereport because we are
3110 3116 * about to replace the device, or attempt an online with
3111 3117 * vdev_forcefault, which will generate the fault for us.
3112 3118 */
3113 3119 if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
3114 3120 !vd->vdev_not_present && !vd->vdev_checkremove &&
3115 3121 vd != spa->spa_root_vdev) {
3116 3122 const char *class;
3117 3123
3118 3124 switch (aux) {
3119 3125 case VDEV_AUX_OPEN_FAILED:
3120 3126 class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
3121 3127 break;
3122 3128 case VDEV_AUX_CORRUPT_DATA:
3123 3129 class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
3124 3130 break;
3125 3131 case VDEV_AUX_NO_REPLICAS:
3126 3132 class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
3127 3133 break;
3128 3134 case VDEV_AUX_BAD_GUID_SUM:
3129 3135 class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
3130 3136 break;
3131 3137 case VDEV_AUX_TOO_SMALL:
3132 3138 class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
3133 3139 break;
3134 3140 case VDEV_AUX_BAD_LABEL:
3135 3141 class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
3136 3142 break;
3137 3143 default:
3138 3144 class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
3139 3145 }
3140 3146
3141 3147 zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
3142 3148 }
3143 3149
3144 3150 /* Erase any notion of persistent removed state */
3145 3151 vd->vdev_removed = B_FALSE;
3146 3152 } else {
3147 3153 vd->vdev_removed = B_FALSE;
3148 3154 }
3149 3155
3150 3156 if (!isopen && vd->vdev_parent)
3151 3157 vdev_propagate_state(vd->vdev_parent);
3152 3158 }
3153 3159
3154 3160 /*
3155 3161 * Check the vdev configuration to ensure that it's capable of supporting
3156 3162 * a root pool. Currently, we do not support RAID-Z or partial configuration.
3157 3163 * In addition, only a single top-level vdev is allowed and none of the leaves
3158 3164 * can be wholedisks.
3159 3165 */
3160 3166 boolean_t
3161 3167 vdev_is_bootable(vdev_t *vd)
3162 3168 {
3163 3169 if (!vd->vdev_ops->vdev_op_leaf) {
3164 3170 char *vdev_type = vd->vdev_ops->vdev_op_type;
3165 3171
3166 3172 if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
3167 3173 vd->vdev_children > 1) {
3168 3174 return (B_FALSE);
3169 3175 } else if (strcmp(vdev_type, VDEV_TYPE_RAIDZ) == 0 ||
3170 3176 strcmp(vdev_type, VDEV_TYPE_MISSING) == 0) {
3171 3177 return (B_FALSE);
3172 3178 }
3173 3179 } else if (vd->vdev_wholedisk == 1) {
3174 3180 return (B_FALSE);
3175 3181 }
3176 3182
3177 3183 for (int c = 0; c < vd->vdev_children; c++) {
3178 3184 if (!vdev_is_bootable(vd->vdev_child[c]))
3179 3185 return (B_FALSE);
3180 3186 }
3181 3187 return (B_TRUE);
3182 3188 }
3183 3189
3184 3190 /*
3185 3191 * Load the state from the original vdev tree (ovd) which
3186 3192 * we've retrieved from the MOS config object. If the original
3187 3193 * vdev was offline or faulted then we transfer that state to the
3188 3194 * device in the current vdev tree (nvd).
3189 3195 */
3190 3196 void
3191 3197 vdev_load_log_state(vdev_t *nvd, vdev_t *ovd)
3192 3198 {
3193 3199 spa_t *spa = nvd->vdev_spa;
3194 3200
3195 3201 ASSERT(nvd->vdev_top->vdev_islog);
3196 3202 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3197 3203 ASSERT3U(nvd->vdev_guid, ==, ovd->vdev_guid);
3198 3204
3199 3205 for (int c = 0; c < nvd->vdev_children; c++)
3200 3206 vdev_load_log_state(nvd->vdev_child[c], ovd->vdev_child[c]);
3201 3207
3202 3208 if (nvd->vdev_ops->vdev_op_leaf) {
3203 3209 /*
3204 3210 * Restore the persistent vdev state
3205 3211 */
3206 3212 nvd->vdev_offline = ovd->vdev_offline;
3207 3213 nvd->vdev_faulted = ovd->vdev_faulted;
3208 3214 nvd->vdev_degraded = ovd->vdev_degraded;
3209 3215 nvd->vdev_removed = ovd->vdev_removed;
3210 3216 }
3211 3217 }
3212 3218
3213 3219 /*
3214 3220 * Determine if a log device has valid content. If the vdev was
3215 3221 * removed or faulted in the MOS config then we know that
3216 3222 * the content on the log device has already been written to the pool.
3217 3223 */
3218 3224 boolean_t
3219 3225 vdev_log_state_valid(vdev_t *vd)
3220 3226 {
3221 3227 if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted &&
3222 3228 !vd->vdev_removed)
3223 3229 return (B_TRUE);
3224 3230
3225 3231 for (int c = 0; c < vd->vdev_children; c++)
3226 3232 if (vdev_log_state_valid(vd->vdev_child[c]))
3227 3233 return (B_TRUE);
3228 3234
3229 3235 return (B_FALSE);
3230 3236 }
3231 3237
3232 3238 /*
3233 3239 * Expand a vdev if possible.
3234 3240 */
3235 3241 void
3236 3242 vdev_expand(vdev_t *vd, uint64_t txg)
3237 3243 {
3238 3244 ASSERT(vd->vdev_top == vd);
3239 3245 ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
3240 3246
3241 3247 if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count) {
3242 3248 VERIFY(vdev_metaslab_init(vd, txg) == 0);
3243 3249 vdev_config_dirty(vd);
3244 3250 }
3245 3251 }
3246 3252
3247 3253 /*
3248 3254 * Split a vdev.
3249 3255 */
3250 3256 void
3251 3257 vdev_split(vdev_t *vd)
3252 3258 {
3253 3259 vdev_t *cvd, *pvd = vd->vdev_parent;
3254 3260
3255 3261 vdev_remove_child(pvd, vd);
3256 3262 vdev_compact_children(pvd);
3257 3263
3258 3264 cvd = pvd->vdev_child[0];
3259 3265 if (pvd->vdev_children == 1) {
3260 3266 vdev_remove_parent(cvd);
3261 3267 cvd->vdev_splitting = B_TRUE;
3262 3268 }
3263 3269 vdev_propagate_state(cvd);
3264 3270 }
3265 3271
3266 3272 void
3267 3273 vdev_deadman(vdev_t *vd)
3268 3274 {
3269 3275 for (int c = 0; c < vd->vdev_children; c++) {
3270 3276 vdev_t *cvd = vd->vdev_child[c];
3271 3277
3272 3278 vdev_deadman(cvd);
3273 3279 }
3274 3280
3275 3281 if (vd->vdev_ops->vdev_op_leaf) {
3276 3282 vdev_queue_t *vq = &vd->vdev_queue;
3277 3283
3278 3284 mutex_enter(&vq->vq_lock);
3279 3285 if (avl_numnodes(&vq->vq_active_tree) > 0) {
3280 3286 spa_t *spa = vd->vdev_spa;
3281 3287 zio_t *fio;
3282 3288 uint64_t delta;
3283 3289
3284 3290 /*
3285 3291 * Look at the head of all the pending queues,
3286 3292 * if any I/O has been outstanding for longer than
3287 3293 * the spa_deadman_synctime we panic the system.
3288 3294 */
3289 3295 fio = avl_first(&vq->vq_active_tree);
3290 3296 delta = gethrtime() - fio->io_timestamp;
3291 3297 if (delta > spa_deadman_synctime(spa)) {
3292 3298 zfs_dbgmsg("SLOW IO: zio timestamp %lluns, "
3293 3299 "delta %lluns, last io %lluns",
3294 3300 fio->io_timestamp, delta,
3295 3301 vq->vq_io_complete_ts);
3296 3302 fm_panic("I/O to pool '%s' appears to be "
3297 3303 "hung.", spa_name(spa));
3298 3304 }
3299 3305 }
3300 3306 mutex_exit(&vq->vq_lock);
3301 3307 }
3302 3308 }
|
↓ open down ↓ |
1783 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX