Print this page
7938 disable LBA weighting on files and SSDs
Reviewed by: Yuri Pankov <yuripv@gmx.com>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/fs/zfs/vdev.c
+++ new/usr/src/uts/common/fs/zfs/vdev.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21
22 22 /*
23 23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 24 * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
25 25 * Copyright 2017 Nexenta Systems, Inc.
26 26 * Copyright (c) 2014 Integros [integros.com]
27 27 * Copyright 2016 Toomas Soome <tsoome@me.com>
28 28 * Copyright 2017 Joyent, Inc.
29 29 */
30 30
31 31 #include <sys/zfs_context.h>
32 32 #include <sys/fm/fs/zfs.h>
33 33 #include <sys/spa.h>
34 34 #include <sys/spa_impl.h>
35 35 #include <sys/dmu.h>
36 36 #include <sys/dmu_tx.h>
37 37 #include <sys/vdev_impl.h>
38 38 #include <sys/uberblock_impl.h>
39 39 #include <sys/metaslab.h>
40 40 #include <sys/metaslab_impl.h>
41 41 #include <sys/space_map.h>
42 42 #include <sys/space_reftree.h>
43 43 #include <sys/zio.h>
44 44 #include <sys/zap.h>
45 45 #include <sys/fs/zfs.h>
46 46 #include <sys/arc.h>
47 47 #include <sys/zil.h>
48 48 #include <sys/dsl_scan.h>
49 49 #include <sys/abd.h>
50 50
51 51 /*
52 52 * Virtual device management.
53 53 */
54 54
55 55 static vdev_ops_t *vdev_ops_table[] = {
56 56 &vdev_root_ops,
57 57 &vdev_raidz_ops,
58 58 &vdev_mirror_ops,
59 59 &vdev_replacing_ops,
60 60 &vdev_spare_ops,
61 61 &vdev_disk_ops,
62 62 &vdev_file_ops,
63 63 &vdev_missing_ops,
64 64 &vdev_hole_ops,
65 65 NULL
66 66 };
67 67
68 68 /* maximum scrub/resilver I/O queue per leaf vdev */
69 69 int zfs_scrub_limit = 10;
70 70
71 71 /*
72 72 * When a vdev is added, it will be divided into approximately (but no
73 73 * more than) this number of metaslabs.
74 74 */
75 75 int metaslabs_per_vdev = 200;
76 76
77 77 /*
78 78 * Given a vdev type, return the appropriate ops vector.
79 79 */
80 80 static vdev_ops_t *
81 81 vdev_getops(const char *type)
82 82 {
83 83 vdev_ops_t *ops, **opspp;
84 84
85 85 for (opspp = vdev_ops_table; (ops = *opspp) != NULL; opspp++)
86 86 if (strcmp(ops->vdev_op_type, type) == 0)
87 87 break;
88 88
89 89 return (ops);
90 90 }
91 91
92 92 /*
93 93 * Default asize function: return the MAX of psize with the asize of
94 94 * all children. This is what's used by anything other than RAID-Z.
95 95 */
96 96 uint64_t
97 97 vdev_default_asize(vdev_t *vd, uint64_t psize)
98 98 {
99 99 uint64_t asize = P2ROUNDUP(psize, 1ULL << vd->vdev_top->vdev_ashift);
100 100 uint64_t csize;
101 101
102 102 for (int c = 0; c < vd->vdev_children; c++) {
103 103 csize = vdev_psize_to_asize(vd->vdev_child[c], psize);
104 104 asize = MAX(asize, csize);
105 105 }
106 106
107 107 return (asize);
108 108 }
109 109
110 110 /*
111 111 * Get the minimum allocatable size. We define the allocatable size as
112 112 * the vdev's asize rounded to the nearest metaslab. This allows us to
113 113 * replace or attach devices which don't have the same physical size but
114 114 * can still satisfy the same number of allocations.
115 115 */
116 116 uint64_t
117 117 vdev_get_min_asize(vdev_t *vd)
118 118 {
119 119 vdev_t *pvd = vd->vdev_parent;
120 120
121 121 /*
122 122 * If our parent is NULL (inactive spare or cache) or is the root,
123 123 * just return our own asize.
124 124 */
125 125 if (pvd == NULL)
126 126 return (vd->vdev_asize);
127 127
128 128 /*
129 129 * The top-level vdev just returns the allocatable size rounded
130 130 * to the nearest metaslab.
131 131 */
132 132 if (vd == vd->vdev_top)
133 133 return (P2ALIGN(vd->vdev_asize, 1ULL << vd->vdev_ms_shift));
134 134
135 135 /*
136 136 * The allocatable space for a raidz vdev is N * sizeof(smallest child),
137 137 * so each child must provide at least 1/Nth of its asize.
138 138 */
139 139 if (pvd->vdev_ops == &vdev_raidz_ops)
140 140 return ((pvd->vdev_min_asize + pvd->vdev_children - 1) /
141 141 pvd->vdev_children);
142 142
143 143 return (pvd->vdev_min_asize);
144 144 }
145 145
146 146 void
147 147 vdev_set_min_asize(vdev_t *vd)
148 148 {
149 149 vd->vdev_min_asize = vdev_get_min_asize(vd);
150 150
151 151 for (int c = 0; c < vd->vdev_children; c++)
152 152 vdev_set_min_asize(vd->vdev_child[c]);
153 153 }
154 154
155 155 vdev_t *
156 156 vdev_lookup_top(spa_t *spa, uint64_t vdev)
157 157 {
158 158 vdev_t *rvd = spa->spa_root_vdev;
159 159
160 160 ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
161 161
162 162 if (vdev < rvd->vdev_children) {
163 163 ASSERT(rvd->vdev_child[vdev] != NULL);
164 164 return (rvd->vdev_child[vdev]);
165 165 }
166 166
167 167 return (NULL);
168 168 }
169 169
170 170 vdev_t *
171 171 vdev_lookup_by_guid(vdev_t *vd, uint64_t guid)
172 172 {
173 173 vdev_t *mvd;
174 174
175 175 if (vd->vdev_guid == guid)
176 176 return (vd);
177 177
178 178 for (int c = 0; c < vd->vdev_children; c++)
179 179 if ((mvd = vdev_lookup_by_guid(vd->vdev_child[c], guid)) !=
180 180 NULL)
181 181 return (mvd);
182 182
183 183 return (NULL);
184 184 }
185 185
186 186 static int
187 187 vdev_count_leaves_impl(vdev_t *vd)
188 188 {
189 189 int n = 0;
190 190
191 191 if (vd->vdev_ops->vdev_op_leaf)
192 192 return (1);
193 193
194 194 for (int c = 0; c < vd->vdev_children; c++)
195 195 n += vdev_count_leaves_impl(vd->vdev_child[c]);
196 196
197 197 return (n);
198 198 }
199 199
200 200 int
201 201 vdev_count_leaves(spa_t *spa)
202 202 {
203 203 return (vdev_count_leaves_impl(spa->spa_root_vdev));
204 204 }
205 205
206 206 void
207 207 vdev_add_child(vdev_t *pvd, vdev_t *cvd)
208 208 {
209 209 size_t oldsize, newsize;
210 210 uint64_t id = cvd->vdev_id;
211 211 vdev_t **newchild;
212 212 spa_t *spa = cvd->vdev_spa;
213 213
214 214 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
215 215 ASSERT(cvd->vdev_parent == NULL);
216 216
217 217 cvd->vdev_parent = pvd;
218 218
219 219 if (pvd == NULL)
220 220 return;
221 221
222 222 ASSERT(id >= pvd->vdev_children || pvd->vdev_child[id] == NULL);
223 223
224 224 oldsize = pvd->vdev_children * sizeof (vdev_t *);
225 225 pvd->vdev_children = MAX(pvd->vdev_children, id + 1);
226 226 newsize = pvd->vdev_children * sizeof (vdev_t *);
227 227
228 228 newchild = kmem_zalloc(newsize, KM_SLEEP);
229 229 if (pvd->vdev_child != NULL) {
230 230 bcopy(pvd->vdev_child, newchild, oldsize);
231 231 kmem_free(pvd->vdev_child, oldsize);
232 232 }
233 233
234 234 pvd->vdev_child = newchild;
235 235 pvd->vdev_child[id] = cvd;
236 236
237 237 cvd->vdev_top = (pvd->vdev_top ? pvd->vdev_top: cvd);
238 238 ASSERT(cvd->vdev_top->vdev_parent->vdev_parent == NULL);
239 239
240 240 /*
241 241 * Walk up all ancestors to update guid sum.
242 242 */
243 243 for (; pvd != NULL; pvd = pvd->vdev_parent)
244 244 pvd->vdev_guid_sum += cvd->vdev_guid_sum;
245 245 }
246 246
247 247 void
248 248 vdev_remove_child(vdev_t *pvd, vdev_t *cvd)
249 249 {
250 250 int c;
251 251 uint_t id = cvd->vdev_id;
252 252
253 253 ASSERT(cvd->vdev_parent == pvd);
254 254
255 255 if (pvd == NULL)
256 256 return;
257 257
258 258 ASSERT(id < pvd->vdev_children);
259 259 ASSERT(pvd->vdev_child[id] == cvd);
260 260
261 261 pvd->vdev_child[id] = NULL;
262 262 cvd->vdev_parent = NULL;
263 263
264 264 for (c = 0; c < pvd->vdev_children; c++)
265 265 if (pvd->vdev_child[c])
266 266 break;
267 267
268 268 if (c == pvd->vdev_children) {
269 269 kmem_free(pvd->vdev_child, c * sizeof (vdev_t *));
270 270 pvd->vdev_child = NULL;
271 271 pvd->vdev_children = 0;
272 272 }
273 273
274 274 /*
275 275 * Walk up all ancestors to update guid sum.
276 276 */
277 277 for (; pvd != NULL; pvd = pvd->vdev_parent)
278 278 pvd->vdev_guid_sum -= cvd->vdev_guid_sum;
279 279 }
280 280
281 281 /*
282 282 * Remove any holes in the child array.
283 283 */
284 284 void
285 285 vdev_compact_children(vdev_t *pvd)
286 286 {
287 287 vdev_t **newchild, *cvd;
288 288 int oldc = pvd->vdev_children;
289 289 int newc;
290 290
291 291 ASSERT(spa_config_held(pvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
292 292
293 293 for (int c = newc = 0; c < oldc; c++)
294 294 if (pvd->vdev_child[c])
295 295 newc++;
296 296
297 297 newchild = kmem_alloc(newc * sizeof (vdev_t *), KM_SLEEP);
298 298
299 299 for (int c = newc = 0; c < oldc; c++) {
300 300 if ((cvd = pvd->vdev_child[c]) != NULL) {
301 301 newchild[newc] = cvd;
302 302 cvd->vdev_id = newc++;
303 303 }
304 304 }
305 305
306 306 kmem_free(pvd->vdev_child, oldc * sizeof (vdev_t *));
307 307 pvd->vdev_child = newchild;
308 308 pvd->vdev_children = newc;
309 309 }
310 310
311 311 /*
312 312 * Allocate and minimally initialize a vdev_t.
313 313 */
314 314 vdev_t *
315 315 vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
316 316 {
317 317 vdev_t *vd;
318 318
319 319 vd = kmem_zalloc(sizeof (vdev_t), KM_SLEEP);
320 320
321 321 if (spa->spa_root_vdev == NULL) {
322 322 ASSERT(ops == &vdev_root_ops);
323 323 spa->spa_root_vdev = vd;
324 324 spa->spa_load_guid = spa_generate_guid(NULL);
325 325 }
326 326
327 327 if (guid == 0 && ops != &vdev_hole_ops) {
328 328 if (spa->spa_root_vdev == vd) {
329 329 /*
330 330 * The root vdev's guid will also be the pool guid,
331 331 * which must be unique among all pools.
332 332 */
333 333 guid = spa_generate_guid(NULL);
334 334 } else {
335 335 /*
336 336 * Any other vdev's guid must be unique within the pool.
337 337 */
338 338 guid = spa_generate_guid(spa);
339 339 }
340 340 ASSERT(!spa_guid_exists(spa_guid(spa), guid));
341 341 }
342 342
343 343 vd->vdev_spa = spa;
344 344 vd->vdev_id = id;
345 345 vd->vdev_guid = guid;
346 346 vd->vdev_guid_sum = guid;
347 347 vd->vdev_ops = ops;
348 348 vd->vdev_state = VDEV_STATE_CLOSED;
349 349 vd->vdev_ishole = (ops == &vdev_hole_ops);
350 350
351 351 mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_DEFAULT, NULL);
352 352 mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL);
353 353 mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL);
354 354 mutex_init(&vd->vdev_queue_lock, NULL, MUTEX_DEFAULT, NULL);
355 355 for (int t = 0; t < DTL_TYPES; t++) {
356 356 vd->vdev_dtl[t] = range_tree_create(NULL, NULL,
357 357 &vd->vdev_dtl_lock);
358 358 }
359 359 txg_list_create(&vd->vdev_ms_list, spa,
360 360 offsetof(struct metaslab, ms_txg_node));
361 361 txg_list_create(&vd->vdev_dtl_list, spa,
362 362 offsetof(struct vdev, vdev_dtl_node));
363 363 vd->vdev_stat.vs_timestamp = gethrtime();
364 364 vdev_queue_init(vd);
365 365 vdev_cache_init(vd);
366 366
367 367 return (vd);
368 368 }
369 369
370 370 /*
371 371 * Allocate a new vdev. The 'alloctype' is used to control whether we are
372 372 * creating a new vdev or loading an existing one - the behavior is slightly
373 373 * different for each case.
374 374 */
375 375 int
376 376 vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id,
377 377 int alloctype)
378 378 {
379 379 vdev_ops_t *ops;
380 380 char *type;
381 381 uint64_t guid = 0, islog, nparity;
382 382 vdev_t *vd;
383 383
384 384 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
385 385
386 386 if (nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) != 0)
387 387 return (SET_ERROR(EINVAL));
388 388
389 389 if ((ops = vdev_getops(type)) == NULL)
390 390 return (SET_ERROR(EINVAL));
391 391
392 392 /*
393 393 * If this is a load, get the vdev guid from the nvlist.
394 394 * Otherwise, vdev_alloc_common() will generate one for us.
395 395 */
396 396 if (alloctype == VDEV_ALLOC_LOAD) {
397 397 uint64_t label_id;
398 398
399 399 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ID, &label_id) ||
400 400 label_id != id)
401 401 return (SET_ERROR(EINVAL));
402 402
403 403 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
404 404 return (SET_ERROR(EINVAL));
405 405 } else if (alloctype == VDEV_ALLOC_SPARE) {
406 406 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
407 407 return (SET_ERROR(EINVAL));
408 408 } else if (alloctype == VDEV_ALLOC_L2CACHE) {
409 409 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
410 410 return (SET_ERROR(EINVAL));
411 411 } else if (alloctype == VDEV_ALLOC_ROOTPOOL) {
412 412 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
413 413 return (SET_ERROR(EINVAL));
414 414 }
415 415
416 416 /*
417 417 * The first allocated vdev must be of type 'root'.
418 418 */
419 419 if (ops != &vdev_root_ops && spa->spa_root_vdev == NULL)
420 420 return (SET_ERROR(EINVAL));
421 421
422 422 /*
423 423 * Determine whether we're a log vdev.
424 424 */
425 425 islog = 0;
426 426 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &islog);
427 427 if (islog && spa_version(spa) < SPA_VERSION_SLOGS)
428 428 return (SET_ERROR(ENOTSUP));
429 429
430 430 if (ops == &vdev_hole_ops && spa_version(spa) < SPA_VERSION_HOLES)
431 431 return (SET_ERROR(ENOTSUP));
432 432
433 433 /*
434 434 * Set the nparity property for RAID-Z vdevs.
435 435 */
436 436 nparity = -1ULL;
437 437 if (ops == &vdev_raidz_ops) {
438 438 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY,
439 439 &nparity) == 0) {
440 440 if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY)
441 441 return (SET_ERROR(EINVAL));
442 442 /*
443 443 * Previous versions could only support 1 or 2 parity
444 444 * device.
445 445 */
446 446 if (nparity > 1 &&
447 447 spa_version(spa) < SPA_VERSION_RAIDZ2)
448 448 return (SET_ERROR(ENOTSUP));
449 449 if (nparity > 2 &&
450 450 spa_version(spa) < SPA_VERSION_RAIDZ3)
451 451 return (SET_ERROR(ENOTSUP));
452 452 } else {
453 453 /*
454 454 * We require the parity to be specified for SPAs that
455 455 * support multiple parity levels.
456 456 */
457 457 if (spa_version(spa) >= SPA_VERSION_RAIDZ2)
458 458 return (SET_ERROR(EINVAL));
459 459 /*
460 460 * Otherwise, we default to 1 parity device for RAID-Z.
461 461 */
462 462 nparity = 1;
463 463 }
464 464 } else {
465 465 nparity = 0;
466 466 }
467 467 ASSERT(nparity != -1ULL);
468 468
469 469 vd = vdev_alloc_common(spa, id, guid, ops);
470 470
471 471 vd->vdev_islog = islog;
472 472 vd->vdev_nparity = nparity;
473 473
474 474 if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &vd->vdev_path) == 0)
475 475 vd->vdev_path = spa_strdup(vd->vdev_path);
476 476 if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &vd->vdev_devid) == 0)
477 477 vd->vdev_devid = spa_strdup(vd->vdev_devid);
478 478 if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH,
479 479 &vd->vdev_physpath) == 0)
480 480 vd->vdev_physpath = spa_strdup(vd->vdev_physpath);
481 481 if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &vd->vdev_fru) == 0)
482 482 vd->vdev_fru = spa_strdup(vd->vdev_fru);
483 483
484 484 /*
485 485 * Set the whole_disk property. If it's not specified, leave the value
486 486 * as -1.
487 487 */
488 488 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
489 489 &vd->vdev_wholedisk) != 0)
490 490 vd->vdev_wholedisk = -1ULL;
491 491
492 492 /*
493 493 * Look for the 'not present' flag. This will only be set if the device
494 494 * was not present at the time of import.
495 495 */
496 496 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT,
497 497 &vd->vdev_not_present);
498 498
499 499 /*
500 500 * Get the alignment requirement.
501 501 */
502 502 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT, &vd->vdev_ashift);
503 503
504 504 /*
505 505 * Retrieve the vdev creation time.
506 506 */
507 507 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_CREATE_TXG,
508 508 &vd->vdev_crtxg);
509 509
510 510 /*
511 511 * If we're a top-level vdev, try to load the allocation parameters.
512 512 */
513 513 if (parent && !parent->vdev_parent &&
514 514 (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
515 515 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
516 516 &vd->vdev_ms_array);
517 517 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
518 518 &vd->vdev_ms_shift);
519 519 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE,
520 520 &vd->vdev_asize);
521 521 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVING,
522 522 &vd->vdev_removing);
523 523 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_VDEV_TOP_ZAP,
524 524 &vd->vdev_top_zap);
525 525 } else {
526 526 ASSERT0(vd->vdev_top_zap);
527 527 }
528 528
529 529 if (parent && !parent->vdev_parent && alloctype != VDEV_ALLOC_ATTACH) {
530 530 ASSERT(alloctype == VDEV_ALLOC_LOAD ||
531 531 alloctype == VDEV_ALLOC_ADD ||
532 532 alloctype == VDEV_ALLOC_SPLIT ||
533 533 alloctype == VDEV_ALLOC_ROOTPOOL);
534 534 vd->vdev_mg = metaslab_group_create(islog ?
535 535 spa_log_class(spa) : spa_normal_class(spa), vd);
536 536 }
537 537
538 538 if (vd->vdev_ops->vdev_op_leaf &&
539 539 (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
540 540 (void) nvlist_lookup_uint64(nv,
541 541 ZPOOL_CONFIG_VDEV_LEAF_ZAP, &vd->vdev_leaf_zap);
542 542 } else {
543 543 ASSERT0(vd->vdev_leaf_zap);
544 544 }
545 545
546 546 /*
547 547 * If we're a leaf vdev, try to load the DTL object and other state.
548 548 */
549 549
550 550 if (vd->vdev_ops->vdev_op_leaf &&
551 551 (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE ||
552 552 alloctype == VDEV_ALLOC_ROOTPOOL)) {
553 553 if (alloctype == VDEV_ALLOC_LOAD) {
554 554 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL,
555 555 &vd->vdev_dtl_object);
556 556 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE,
557 557 &vd->vdev_unspare);
558 558 }
559 559
560 560 if (alloctype == VDEV_ALLOC_ROOTPOOL) {
561 561 uint64_t spare = 0;
562 562
563 563 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE,
564 564 &spare) == 0 && spare)
565 565 spa_spare_add(vd);
566 566 }
567 567
568 568 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE,
569 569 &vd->vdev_offline);
570 570
571 571 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_RESILVER_TXG,
572 572 &vd->vdev_resilver_txg);
573 573
574 574 /*
575 575 * When importing a pool, we want to ignore the persistent fault
576 576 * state, as the diagnosis made on another system may not be
577 577 * valid in the current context. Local vdevs will
578 578 * remain in the faulted state.
579 579 */
580 580 if (spa_load_state(spa) == SPA_LOAD_OPEN) {
581 581 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED,
582 582 &vd->vdev_faulted);
583 583 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED,
584 584 &vd->vdev_degraded);
585 585 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED,
586 586 &vd->vdev_removed);
587 587
588 588 if (vd->vdev_faulted || vd->vdev_degraded) {
589 589 char *aux;
590 590
591 591 vd->vdev_label_aux =
592 592 VDEV_AUX_ERR_EXCEEDED;
593 593 if (nvlist_lookup_string(nv,
594 594 ZPOOL_CONFIG_AUX_STATE, &aux) == 0 &&
595 595 strcmp(aux, "external") == 0)
596 596 vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
597 597 }
598 598 }
599 599 }
600 600
601 601 /*
602 602 * Add ourselves to the parent's list of children.
603 603 */
604 604 vdev_add_child(parent, vd);
605 605
606 606 *vdp = vd;
607 607
608 608 return (0);
609 609 }
610 610
611 611 void
612 612 vdev_free(vdev_t *vd)
613 613 {
614 614 spa_t *spa = vd->vdev_spa;
615 615
616 616 /*
617 617 * vdev_free() implies closing the vdev first. This is simpler than
618 618 * trying to ensure complicated semantics for all callers.
619 619 */
620 620 vdev_close(vd);
621 621
622 622 ASSERT(!list_link_active(&vd->vdev_config_dirty_node));
623 623 ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
624 624
625 625 /*
626 626 * Free all children.
627 627 */
628 628 for (int c = 0; c < vd->vdev_children; c++)
629 629 vdev_free(vd->vdev_child[c]);
630 630
631 631 ASSERT(vd->vdev_child == NULL);
632 632 ASSERT(vd->vdev_guid_sum == vd->vdev_guid);
633 633
634 634 /*
635 635 * Discard allocation state.
636 636 */
637 637 if (vd->vdev_mg != NULL) {
638 638 vdev_metaslab_fini(vd);
639 639 metaslab_group_destroy(vd->vdev_mg);
640 640 }
641 641
642 642 ASSERT0(vd->vdev_stat.vs_space);
643 643 ASSERT0(vd->vdev_stat.vs_dspace);
644 644 ASSERT0(vd->vdev_stat.vs_alloc);
645 645
646 646 /*
647 647 * Remove this vdev from its parent's child list.
648 648 */
649 649 vdev_remove_child(vd->vdev_parent, vd);
650 650
651 651 ASSERT(vd->vdev_parent == NULL);
652 652
653 653 /*
654 654 * Clean up vdev structure.
655 655 */
656 656 vdev_queue_fini(vd);
657 657 vdev_cache_fini(vd);
658 658
659 659 if (vd->vdev_path)
660 660 spa_strfree(vd->vdev_path);
661 661 if (vd->vdev_devid)
662 662 spa_strfree(vd->vdev_devid);
663 663 if (vd->vdev_physpath)
664 664 spa_strfree(vd->vdev_physpath);
665 665 if (vd->vdev_fru)
666 666 spa_strfree(vd->vdev_fru);
667 667
668 668 if (vd->vdev_isspare)
669 669 spa_spare_remove(vd);
670 670 if (vd->vdev_isl2cache)
671 671 spa_l2cache_remove(vd);
672 672
673 673 txg_list_destroy(&vd->vdev_ms_list);
674 674 txg_list_destroy(&vd->vdev_dtl_list);
675 675
676 676 mutex_enter(&vd->vdev_dtl_lock);
677 677 space_map_close(vd->vdev_dtl_sm);
678 678 for (int t = 0; t < DTL_TYPES; t++) {
679 679 range_tree_vacate(vd->vdev_dtl[t], NULL, NULL);
680 680 range_tree_destroy(vd->vdev_dtl[t]);
681 681 }
682 682 mutex_exit(&vd->vdev_dtl_lock);
683 683
684 684 mutex_destroy(&vd->vdev_queue_lock);
685 685 mutex_destroy(&vd->vdev_dtl_lock);
686 686 mutex_destroy(&vd->vdev_stat_lock);
687 687 mutex_destroy(&vd->vdev_probe_lock);
688 688
689 689 if (vd == spa->spa_root_vdev)
690 690 spa->spa_root_vdev = NULL;
691 691
692 692 kmem_free(vd, sizeof (vdev_t));
693 693 }
694 694
695 695 /*
696 696 * Transfer top-level vdev state from svd to tvd.
697 697 */
698 698 static void
699 699 vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
700 700 {
701 701 spa_t *spa = svd->vdev_spa;
702 702 metaslab_t *msp;
703 703 vdev_t *vd;
704 704 int t;
705 705
706 706 ASSERT(tvd == tvd->vdev_top);
707 707
708 708 tvd->vdev_ms_array = svd->vdev_ms_array;
709 709 tvd->vdev_ms_shift = svd->vdev_ms_shift;
710 710 tvd->vdev_ms_count = svd->vdev_ms_count;
711 711 tvd->vdev_top_zap = svd->vdev_top_zap;
712 712
713 713 svd->vdev_ms_array = 0;
714 714 svd->vdev_ms_shift = 0;
715 715 svd->vdev_ms_count = 0;
716 716 svd->vdev_top_zap = 0;
717 717
718 718 if (tvd->vdev_mg)
719 719 ASSERT3P(tvd->vdev_mg, ==, svd->vdev_mg);
720 720 tvd->vdev_mg = svd->vdev_mg;
721 721 tvd->vdev_ms = svd->vdev_ms;
722 722
723 723 svd->vdev_mg = NULL;
724 724 svd->vdev_ms = NULL;
725 725
726 726 if (tvd->vdev_mg != NULL)
727 727 tvd->vdev_mg->mg_vd = tvd;
728 728
729 729 tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
730 730 tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
731 731 tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
732 732
733 733 svd->vdev_stat.vs_alloc = 0;
734 734 svd->vdev_stat.vs_space = 0;
735 735 svd->vdev_stat.vs_dspace = 0;
736 736
737 737 for (t = 0; t < TXG_SIZE; t++) {
738 738 while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
739 739 (void) txg_list_add(&tvd->vdev_ms_list, msp, t);
740 740 while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
741 741 (void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
742 742 if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
743 743 (void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
744 744 }
745 745
746 746 if (list_link_active(&svd->vdev_config_dirty_node)) {
747 747 vdev_config_clean(svd);
748 748 vdev_config_dirty(tvd);
749 749 }
750 750
751 751 if (list_link_active(&svd->vdev_state_dirty_node)) {
752 752 vdev_state_clean(svd);
753 753 vdev_state_dirty(tvd);
754 754 }
755 755
756 756 tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
757 757 svd->vdev_deflate_ratio = 0;
758 758
759 759 tvd->vdev_islog = svd->vdev_islog;
760 760 svd->vdev_islog = 0;
761 761 }
762 762
763 763 static void
764 764 vdev_top_update(vdev_t *tvd, vdev_t *vd)
765 765 {
766 766 if (vd == NULL)
767 767 return;
768 768
769 769 vd->vdev_top = tvd;
770 770
771 771 for (int c = 0; c < vd->vdev_children; c++)
772 772 vdev_top_update(tvd, vd->vdev_child[c]);
773 773 }
774 774
775 775 /*
776 776 * Add a mirror/replacing vdev above an existing vdev.
777 777 */
778 778 vdev_t *
779 779 vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
780 780 {
781 781 spa_t *spa = cvd->vdev_spa;
782 782 vdev_t *pvd = cvd->vdev_parent;
783 783 vdev_t *mvd;
784 784
785 785 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
786 786
787 787 mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
788 788
789 789 mvd->vdev_asize = cvd->vdev_asize;
790 790 mvd->vdev_min_asize = cvd->vdev_min_asize;
791 791 mvd->vdev_max_asize = cvd->vdev_max_asize;
792 792 mvd->vdev_ashift = cvd->vdev_ashift;
793 793 mvd->vdev_state = cvd->vdev_state;
794 794 mvd->vdev_crtxg = cvd->vdev_crtxg;
795 795
796 796 vdev_remove_child(pvd, cvd);
797 797 vdev_add_child(pvd, mvd);
798 798 cvd->vdev_id = mvd->vdev_children;
799 799 vdev_add_child(mvd, cvd);
800 800 vdev_top_update(cvd->vdev_top, cvd->vdev_top);
801 801
802 802 if (mvd == mvd->vdev_top)
803 803 vdev_top_transfer(cvd, mvd);
804 804
805 805 return (mvd);
806 806 }
807 807
808 808 /*
809 809 * Remove a 1-way mirror/replacing vdev from the tree.
810 810 */
811 811 void
812 812 vdev_remove_parent(vdev_t *cvd)
813 813 {
814 814 vdev_t *mvd = cvd->vdev_parent;
815 815 vdev_t *pvd = mvd->vdev_parent;
816 816
817 817 ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
818 818
819 819 ASSERT(mvd->vdev_children == 1);
820 820 ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
821 821 mvd->vdev_ops == &vdev_replacing_ops ||
822 822 mvd->vdev_ops == &vdev_spare_ops);
823 823 cvd->vdev_ashift = mvd->vdev_ashift;
824 824
825 825 vdev_remove_child(mvd, cvd);
826 826 vdev_remove_child(pvd, mvd);
827 827
828 828 /*
829 829 * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
830 830 * Otherwise, we could have detached an offline device, and when we
831 831 * go to import the pool we'll think we have two top-level vdevs,
832 832 * instead of a different version of the same top-level vdev.
833 833 */
834 834 if (mvd->vdev_top == mvd) {
835 835 uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
836 836 cvd->vdev_orig_guid = cvd->vdev_guid;
837 837 cvd->vdev_guid += guid_delta;
838 838 cvd->vdev_guid_sum += guid_delta;
839 839 }
840 840 cvd->vdev_id = mvd->vdev_id;
841 841 vdev_add_child(pvd, cvd);
842 842 vdev_top_update(cvd->vdev_top, cvd->vdev_top);
843 843
844 844 if (cvd == cvd->vdev_top)
845 845 vdev_top_transfer(mvd, cvd);
846 846
847 847 ASSERT(mvd->vdev_children == 0);
848 848 vdev_free(mvd);
849 849 }
850 850
851 851 int
852 852 vdev_metaslab_init(vdev_t *vd, uint64_t txg)
853 853 {
854 854 spa_t *spa = vd->vdev_spa;
855 855 objset_t *mos = spa->spa_meta_objset;
856 856 uint64_t m;
857 857 uint64_t oldc = vd->vdev_ms_count;
858 858 uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
859 859 metaslab_t **mspp;
860 860 int error;
861 861
862 862 ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER));
863 863
864 864 /*
865 865 * This vdev is not being allocated from yet or is a hole.
866 866 */
867 867 if (vd->vdev_ms_shift == 0)
868 868 return (0);
869 869
870 870 ASSERT(!vd->vdev_ishole);
871 871
872 872 /*
873 873 * Compute the raidz-deflation ratio. Note, we hard-code
874 874 * in 128k (1 << 17) because it is the "typical" blocksize.
875 875 * Even though SPA_MAXBLOCKSIZE changed, this algorithm can not change,
876 876 * otherwise it would inconsistently account for existing bp's.
877 877 */
878 878 vd->vdev_deflate_ratio = (1 << 17) /
879 879 (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
880 880
881 881 ASSERT(oldc <= newc);
882 882
883 883 mspp = kmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
884 884
885 885 if (oldc != 0) {
886 886 bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp));
887 887 kmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
888 888 }
889 889
890 890 vd->vdev_ms = mspp;
891 891 vd->vdev_ms_count = newc;
892 892
893 893 for (m = oldc; m < newc; m++) {
894 894 uint64_t object = 0;
895 895
896 896 if (txg == 0) {
897 897 error = dmu_read(mos, vd->vdev_ms_array,
898 898 m * sizeof (uint64_t), sizeof (uint64_t), &object,
899 899 DMU_READ_PREFETCH);
900 900 if (error)
901 901 return (error);
902 902 }
903 903
904 904 error = metaslab_init(vd->vdev_mg, m, object, txg,
905 905 &(vd->vdev_ms[m]));
906 906 if (error)
907 907 return (error);
908 908 }
909 909
910 910 if (txg == 0)
911 911 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER);
912 912
913 913 /*
914 914 * If the vdev is being removed we don't activate
915 915 * the metaslabs since we want to ensure that no new
916 916 * allocations are performed on this device.
917 917 */
918 918 if (oldc == 0 && !vd->vdev_removing)
919 919 metaslab_group_activate(vd->vdev_mg);
920 920
921 921 if (txg == 0)
922 922 spa_config_exit(spa, SCL_ALLOC, FTAG);
923 923
924 924 return (0);
925 925 }
926 926
927 927 void
928 928 vdev_metaslab_fini(vdev_t *vd)
929 929 {
930 930 uint64_t m;
931 931 uint64_t count = vd->vdev_ms_count;
932 932
933 933 if (vd->vdev_ms != NULL) {
934 934 metaslab_group_passivate(vd->vdev_mg);
935 935 for (m = 0; m < count; m++) {
936 936 metaslab_t *msp = vd->vdev_ms[m];
937 937
938 938 if (msp != NULL)
939 939 metaslab_fini(msp);
940 940 }
941 941 kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
942 942 vd->vdev_ms = NULL;
943 943 }
944 944 }
945 945
946 946 typedef struct vdev_probe_stats {
947 947 boolean_t vps_readable;
948 948 boolean_t vps_writeable;
949 949 int vps_flags;
950 950 } vdev_probe_stats_t;
951 951
952 952 static void
953 953 vdev_probe_done(zio_t *zio)
954 954 {
955 955 spa_t *spa = zio->io_spa;
956 956 vdev_t *vd = zio->io_vd;
957 957 vdev_probe_stats_t *vps = zio->io_private;
958 958
959 959 ASSERT(vd->vdev_probe_zio != NULL);
960 960
961 961 if (zio->io_type == ZIO_TYPE_READ) {
962 962 if (zio->io_error == 0)
963 963 vps->vps_readable = 1;
964 964 if (zio->io_error == 0 && spa_writeable(spa)) {
965 965 zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
966 966 zio->io_offset, zio->io_size, zio->io_abd,
967 967 ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
968 968 ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
969 969 } else {
970 970 abd_free(zio->io_abd);
971 971 }
972 972 } else if (zio->io_type == ZIO_TYPE_WRITE) {
973 973 if (zio->io_error == 0)
974 974 vps->vps_writeable = 1;
975 975 abd_free(zio->io_abd);
976 976 } else if (zio->io_type == ZIO_TYPE_NULL) {
977 977 zio_t *pio;
978 978
979 979 vd->vdev_cant_read |= !vps->vps_readable;
980 980 vd->vdev_cant_write |= !vps->vps_writeable;
981 981
982 982 if (vdev_readable(vd) &&
983 983 (vdev_writeable(vd) || !spa_writeable(spa))) {
984 984 zio->io_error = 0;
985 985 } else {
986 986 ASSERT(zio->io_error != 0);
987 987 zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
988 988 spa, vd, NULL, 0, 0);
989 989 zio->io_error = SET_ERROR(ENXIO);
990 990 }
991 991
992 992 mutex_enter(&vd->vdev_probe_lock);
993 993 ASSERT(vd->vdev_probe_zio == zio);
994 994 vd->vdev_probe_zio = NULL;
995 995 mutex_exit(&vd->vdev_probe_lock);
996 996
997 997 zio_link_t *zl = NULL;
998 998 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
999 999 if (!vdev_accessible(vd, pio))
1000 1000 pio->io_error = SET_ERROR(ENXIO);
1001 1001
1002 1002 kmem_free(vps, sizeof (*vps));
1003 1003 }
1004 1004 }
1005 1005
1006 1006 /*
1007 1007 * Determine whether this device is accessible.
1008 1008 *
1009 1009 * Read and write to several known locations: the pad regions of each
1010 1010 * vdev label but the first, which we leave alone in case it contains
1011 1011 * a VTOC.
1012 1012 */
1013 1013 zio_t *
1014 1014 vdev_probe(vdev_t *vd, zio_t *zio)
1015 1015 {
1016 1016 spa_t *spa = vd->vdev_spa;
1017 1017 vdev_probe_stats_t *vps = NULL;
1018 1018 zio_t *pio;
1019 1019
1020 1020 ASSERT(vd->vdev_ops->vdev_op_leaf);
1021 1021
1022 1022 /*
1023 1023 * Don't probe the probe.
1024 1024 */
1025 1025 if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
1026 1026 return (NULL);
1027 1027
1028 1028 /*
1029 1029 * To prevent 'probe storms' when a device fails, we create
1030 1030 * just one probe i/o at a time. All zios that want to probe
1031 1031 * this vdev will become parents of the probe io.
1032 1032 */
1033 1033 mutex_enter(&vd->vdev_probe_lock);
1034 1034
1035 1035 if ((pio = vd->vdev_probe_zio) == NULL) {
1036 1036 vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
1037 1037
1038 1038 vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
1039 1039 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
1040 1040 ZIO_FLAG_TRYHARD;
1041 1041
1042 1042 if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
1043 1043 /*
1044 1044 * vdev_cant_read and vdev_cant_write can only
1045 1045 * transition from TRUE to FALSE when we have the
1046 1046 * SCL_ZIO lock as writer; otherwise they can only
1047 1047 * transition from FALSE to TRUE. This ensures that
1048 1048 * any zio looking at these values can assume that
1049 1049 * failures persist for the life of the I/O. That's
1050 1050 * important because when a device has intermittent
1051 1051 * connectivity problems, we want to ensure that
1052 1052 * they're ascribed to the device (ENXIO) and not
1053 1053 * the zio (EIO).
1054 1054 *
1055 1055 * Since we hold SCL_ZIO as writer here, clear both
1056 1056 * values so the probe can reevaluate from first
1057 1057 * principles.
1058 1058 */
1059 1059 vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
1060 1060 vd->vdev_cant_read = B_FALSE;
1061 1061 vd->vdev_cant_write = B_FALSE;
1062 1062 }
1063 1063
1064 1064 vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
1065 1065 vdev_probe_done, vps,
1066 1066 vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
1067 1067
1068 1068 /*
1069 1069 * We can't change the vdev state in this context, so we
1070 1070 * kick off an async task to do it on our behalf.
1071 1071 */
1072 1072 if (zio != NULL) {
1073 1073 vd->vdev_probe_wanted = B_TRUE;
1074 1074 spa_async_request(spa, SPA_ASYNC_PROBE);
1075 1075 }
1076 1076 }
1077 1077
1078 1078 if (zio != NULL)
1079 1079 zio_add_child(zio, pio);
1080 1080
1081 1081 mutex_exit(&vd->vdev_probe_lock);
1082 1082
1083 1083 if (vps == NULL) {
1084 1084 ASSERT(zio != NULL);
1085 1085 return (NULL);
1086 1086 }
1087 1087
1088 1088 for (int l = 1; l < VDEV_LABELS; l++) {
1089 1089 zio_nowait(zio_read_phys(pio, vd,
1090 1090 vdev_label_offset(vd->vdev_psize, l,
1091 1091 offsetof(vdev_label_t, vl_pad2)), VDEV_PAD_SIZE,
1092 1092 abd_alloc_for_io(VDEV_PAD_SIZE, B_TRUE),
1093 1093 ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1094 1094 ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1095 1095 }
1096 1096
1097 1097 if (zio == NULL)
1098 1098 return (pio);
1099 1099
1100 1100 zio_nowait(pio);
1101 1101 return (NULL);
|
↓ open down ↓ |
1101 lines elided |
↑ open up ↑ |
1102 1102 }
1103 1103
1104 1104 static void
1105 1105 vdev_open_child(void *arg)
1106 1106 {
1107 1107 vdev_t *vd = arg;
1108 1108
1109 1109 vd->vdev_open_thread = curthread;
1110 1110 vd->vdev_open_error = vdev_open(vd);
1111 1111 vd->vdev_open_thread = NULL;
1112 + vd->vdev_parent->vdev_nonrot &= vd->vdev_nonrot;
1112 1113 }
1113 1114
1114 1115 boolean_t
1115 1116 vdev_uses_zvols(vdev_t *vd)
1116 1117 {
1117 1118 if (vd->vdev_path && strncmp(vd->vdev_path, ZVOL_DIR,
1118 1119 strlen(ZVOL_DIR)) == 0)
1119 1120 return (B_TRUE);
1120 1121 for (int c = 0; c < vd->vdev_children; c++)
1121 1122 if (vdev_uses_zvols(vd->vdev_child[c]))
1122 1123 return (B_TRUE);
1123 1124 return (B_FALSE);
1124 1125 }
1125 1126
1126 1127 void
1127 1128 vdev_open_children(vdev_t *vd)
1128 1129 {
1129 1130 taskq_t *tq;
1130 1131 int children = vd->vdev_children;
1131 1132
1133 + vd->vdev_nonrot = B_TRUE;
1134 +
1132 1135 /*
1133 1136 * in order to handle pools on top of zvols, do the opens
1134 1137 * in a single thread so that the same thread holds the
1135 1138 * spa_namespace_lock
1136 1139 */
1137 1140 if (vdev_uses_zvols(vd)) {
1138 - for (int c = 0; c < children; c++)
1141 + for (int c = 0; c < children; c++) {
1139 1142 vd->vdev_child[c]->vdev_open_error =
1140 1143 vdev_open(vd->vdev_child[c]);
1144 + vd->vdev_nonrot &= vd->vdev_child[c]->vdev_nonrot;
1145 + }
1141 1146 return;
1142 1147 }
1143 1148 tq = taskq_create("vdev_open", children, minclsyspri,
1144 1149 children, children, TASKQ_PREPOPULATE);
1145 1150
1146 1151 for (int c = 0; c < children; c++)
1147 1152 VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c],
1148 1153 TQ_SLEEP) != NULL);
1149 1154
1150 1155 taskq_destroy(tq);
1156 +
1157 + for (int c = 0; c < children; c++)
1158 + vd->vdev_nonrot &= vd->vdev_child[c]->vdev_nonrot;
1151 1159 }
1152 1160
1153 1161 /*
1154 1162 * Prepare a virtual device for access.
1155 1163 */
1156 1164 int
1157 1165 vdev_open(vdev_t *vd)
1158 1166 {
1159 1167 spa_t *spa = vd->vdev_spa;
1160 1168 int error;
1161 1169 uint64_t osize = 0;
1162 1170 uint64_t max_osize = 0;
1163 1171 uint64_t asize, max_asize, psize;
1164 1172 uint64_t ashift = 0;
1165 1173
1166 1174 ASSERT(vd->vdev_open_thread == curthread ||
1167 1175 spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1168 1176 ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1169 1177 vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1170 1178 vd->vdev_state == VDEV_STATE_OFFLINE);
1171 1179
1172 1180 vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1173 1181 vd->vdev_cant_read = B_FALSE;
1174 1182 vd->vdev_cant_write = B_FALSE;
1175 1183 vd->vdev_min_asize = vdev_get_min_asize(vd);
1176 1184
1177 1185 /*
1178 1186 * If this vdev is not removed, check its fault status. If it's
1179 1187 * faulted, bail out of the open.
1180 1188 */
1181 1189 if (!vd->vdev_removed && vd->vdev_faulted) {
1182 1190 ASSERT(vd->vdev_children == 0);
1183 1191 ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1184 1192 vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1185 1193 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1186 1194 vd->vdev_label_aux);
1187 1195 return (SET_ERROR(ENXIO));
1188 1196 } else if (vd->vdev_offline) {
1189 1197 ASSERT(vd->vdev_children == 0);
1190 1198 vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1191 1199 return (SET_ERROR(ENXIO));
1192 1200 }
1193 1201
1194 1202 error = vd->vdev_ops->vdev_op_open(vd, &osize, &max_osize, &ashift);
1195 1203
1196 1204 /*
1197 1205 * Reset the vdev_reopening flag so that we actually close
1198 1206 * the vdev on error.
1199 1207 */
1200 1208 vd->vdev_reopening = B_FALSE;
1201 1209 if (zio_injection_enabled && error == 0)
1202 1210 error = zio_handle_device_injection(vd, NULL, ENXIO);
1203 1211
1204 1212 if (error) {
1205 1213 if (vd->vdev_removed &&
1206 1214 vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1207 1215 vd->vdev_removed = B_FALSE;
1208 1216
1209 1217 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1210 1218 vd->vdev_stat.vs_aux);
1211 1219 return (error);
1212 1220 }
1213 1221
1214 1222 vd->vdev_removed = B_FALSE;
1215 1223
1216 1224 /*
1217 1225 * Recheck the faulted flag now that we have confirmed that
1218 1226 * the vdev is accessible. If we're faulted, bail.
1219 1227 */
1220 1228 if (vd->vdev_faulted) {
1221 1229 ASSERT(vd->vdev_children == 0);
1222 1230 ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1223 1231 vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1224 1232 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1225 1233 vd->vdev_label_aux);
1226 1234 return (SET_ERROR(ENXIO));
1227 1235 }
1228 1236
1229 1237 if (vd->vdev_degraded) {
1230 1238 ASSERT(vd->vdev_children == 0);
1231 1239 vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1232 1240 VDEV_AUX_ERR_EXCEEDED);
1233 1241 } else {
1234 1242 vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1235 1243 }
1236 1244
1237 1245 /*
1238 1246 * For hole or missing vdevs we just return success.
1239 1247 */
1240 1248 if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1241 1249 return (0);
1242 1250
1243 1251 for (int c = 0; c < vd->vdev_children; c++) {
1244 1252 if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1245 1253 vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1246 1254 VDEV_AUX_NONE);
1247 1255 break;
1248 1256 }
1249 1257 }
1250 1258
1251 1259 osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1252 1260 max_osize = P2ALIGN(max_osize, (uint64_t)sizeof (vdev_label_t));
1253 1261
1254 1262 if (vd->vdev_children == 0) {
1255 1263 if (osize < SPA_MINDEVSIZE) {
1256 1264 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1257 1265 VDEV_AUX_TOO_SMALL);
1258 1266 return (SET_ERROR(EOVERFLOW));
1259 1267 }
1260 1268 psize = osize;
1261 1269 asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1262 1270 max_asize = max_osize - (VDEV_LABEL_START_SIZE +
1263 1271 VDEV_LABEL_END_SIZE);
1264 1272 } else {
1265 1273 if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1266 1274 (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1267 1275 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1268 1276 VDEV_AUX_TOO_SMALL);
1269 1277 return (SET_ERROR(EOVERFLOW));
1270 1278 }
1271 1279 psize = 0;
1272 1280 asize = osize;
1273 1281 max_asize = max_osize;
1274 1282 }
1275 1283
1276 1284 vd->vdev_psize = psize;
1277 1285
1278 1286 /*
1279 1287 * Make sure the allocatable size hasn't shrunk too much.
1280 1288 */
1281 1289 if (asize < vd->vdev_min_asize) {
1282 1290 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1283 1291 VDEV_AUX_BAD_LABEL);
1284 1292 return (SET_ERROR(EINVAL));
1285 1293 }
1286 1294
1287 1295 if (vd->vdev_asize == 0) {
1288 1296 /*
1289 1297 * This is the first-ever open, so use the computed values.
1290 1298 * For testing purposes, a higher ashift can be requested.
1291 1299 */
1292 1300 vd->vdev_asize = asize;
1293 1301 vd->vdev_max_asize = max_asize;
1294 1302 vd->vdev_ashift = MAX(ashift, vd->vdev_ashift);
1295 1303 } else {
1296 1304 /*
1297 1305 * Detect if the alignment requirement has increased.
1298 1306 * We don't want to make the pool unavailable, just
1299 1307 * issue a warning instead.
1300 1308 */
1301 1309 if (ashift > vd->vdev_top->vdev_ashift &&
1302 1310 vd->vdev_ops->vdev_op_leaf) {
1303 1311 cmn_err(CE_WARN,
1304 1312 "Disk, '%s', has a block alignment that is "
1305 1313 "larger than the pool's alignment\n",
1306 1314 vd->vdev_path);
1307 1315 }
1308 1316 vd->vdev_max_asize = max_asize;
1309 1317 }
1310 1318
1311 1319 /*
1312 1320 * If all children are healthy we update asize if either:
1313 1321 * The asize has increased, due to a device expansion caused by dynamic
1314 1322 * LUN growth or vdev replacement, and automatic expansion is enabled;
1315 1323 * making the additional space available.
1316 1324 *
1317 1325 * The asize has decreased, due to a device shrink usually caused by a
1318 1326 * vdev replace with a smaller device. This ensures that calculations
1319 1327 * based of max_asize and asize e.g. esize are always valid. It's safe
1320 1328 * to do this as we've already validated that asize is greater than
1321 1329 * vdev_min_asize.
1322 1330 */
1323 1331 if (vd->vdev_state == VDEV_STATE_HEALTHY &&
1324 1332 ((asize > vd->vdev_asize &&
1325 1333 (vd->vdev_expanding || spa->spa_autoexpand)) ||
1326 1334 (asize < vd->vdev_asize)))
1327 1335 vd->vdev_asize = asize;
1328 1336
1329 1337 vdev_set_min_asize(vd);
1330 1338
1331 1339 /*
1332 1340 * Ensure we can issue some IO before declaring the
1333 1341 * vdev open for business.
1334 1342 */
1335 1343 if (vd->vdev_ops->vdev_op_leaf &&
1336 1344 (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1337 1345 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1338 1346 VDEV_AUX_ERR_EXCEEDED);
1339 1347 return (error);
1340 1348 }
1341 1349
1342 1350 /*
1343 1351 * Track the min and max ashift values for normal data devices.
1344 1352 */
1345 1353 if (vd->vdev_top == vd && vd->vdev_ashift != 0 &&
1346 1354 !vd->vdev_islog && vd->vdev_aux == NULL) {
1347 1355 if (vd->vdev_ashift > spa->spa_max_ashift)
1348 1356 spa->spa_max_ashift = vd->vdev_ashift;
1349 1357 if (vd->vdev_ashift < spa->spa_min_ashift)
1350 1358 spa->spa_min_ashift = vd->vdev_ashift;
1351 1359 }
1352 1360
1353 1361 /*
1354 1362 * If a leaf vdev has a DTL, and seems healthy, then kick off a
1355 1363 * resilver. But don't do this if we are doing a reopen for a scrub,
1356 1364 * since this would just restart the scrub we are already doing.
1357 1365 */
1358 1366 if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1359 1367 vdev_resilver_needed(vd, NULL, NULL))
1360 1368 spa_async_request(spa, SPA_ASYNC_RESILVER);
1361 1369
1362 1370 return (0);
1363 1371 }
1364 1372
1365 1373 /*
1366 1374 * Called once the vdevs are all opened, this routine validates the label
1367 1375 * contents. This needs to be done before vdev_load() so that we don't
1368 1376 * inadvertently do repair I/Os to the wrong device.
1369 1377 *
1370 1378 * If 'strict' is false ignore the spa guid check. This is necessary because
1371 1379 * if the machine crashed during a re-guid the new guid might have been written
1372 1380 * to all of the vdev labels, but not the cached config. The strict check
1373 1381 * will be performed when the pool is opened again using the mos config.
1374 1382 *
1375 1383 * This function will only return failure if one of the vdevs indicates that it
1376 1384 * has since been destroyed or exported. This is only possible if
1377 1385 * /etc/zfs/zpool.cache was readonly at the time. Otherwise, the vdev state
1378 1386 * will be updated but the function will return 0.
1379 1387 */
1380 1388 int
1381 1389 vdev_validate(vdev_t *vd, boolean_t strict)
1382 1390 {
1383 1391 spa_t *spa = vd->vdev_spa;
1384 1392 nvlist_t *label;
1385 1393 uint64_t guid = 0, top_guid;
1386 1394 uint64_t state;
1387 1395
1388 1396 for (int c = 0; c < vd->vdev_children; c++)
1389 1397 if (vdev_validate(vd->vdev_child[c], strict) != 0)
1390 1398 return (SET_ERROR(EBADF));
1391 1399
1392 1400 /*
1393 1401 * If the device has already failed, or was marked offline, don't do
1394 1402 * any further validation. Otherwise, label I/O will fail and we will
1395 1403 * overwrite the previous state.
1396 1404 */
1397 1405 if (vd->vdev_ops->vdev_op_leaf && vdev_readable(vd)) {
1398 1406 uint64_t aux_guid = 0;
1399 1407 nvlist_t *nvl;
1400 1408 uint64_t txg = spa_last_synced_txg(spa) != 0 ?
1401 1409 spa_last_synced_txg(spa) : -1ULL;
1402 1410
1403 1411 if ((label = vdev_label_read_config(vd, txg)) == NULL) {
1404 1412 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1405 1413 VDEV_AUX_BAD_LABEL);
1406 1414 return (0);
1407 1415 }
1408 1416
1409 1417 /*
1410 1418 * Determine if this vdev has been split off into another
1411 1419 * pool. If so, then refuse to open it.
1412 1420 */
1413 1421 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID,
1414 1422 &aux_guid) == 0 && aux_guid == spa_guid(spa)) {
1415 1423 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1416 1424 VDEV_AUX_SPLIT_POOL);
1417 1425 nvlist_free(label);
1418 1426 return (0);
1419 1427 }
1420 1428
1421 1429 if (strict && (nvlist_lookup_uint64(label,
1422 1430 ZPOOL_CONFIG_POOL_GUID, &guid) != 0 ||
1423 1431 guid != spa_guid(spa))) {
1424 1432 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1425 1433 VDEV_AUX_CORRUPT_DATA);
1426 1434 nvlist_free(label);
1427 1435 return (0);
1428 1436 }
1429 1437
1430 1438 if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl)
1431 1439 != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID,
1432 1440 &aux_guid) != 0)
1433 1441 aux_guid = 0;
1434 1442
1435 1443 /*
1436 1444 * If this vdev just became a top-level vdev because its
1437 1445 * sibling was detached, it will have adopted the parent's
1438 1446 * vdev guid -- but the label may or may not be on disk yet.
1439 1447 * Fortunately, either version of the label will have the
1440 1448 * same top guid, so if we're a top-level vdev, we can
1441 1449 * safely compare to that instead.
1442 1450 *
1443 1451 * If we split this vdev off instead, then we also check the
1444 1452 * original pool's guid. We don't want to consider the vdev
1445 1453 * corrupt if it is partway through a split operation.
1446 1454 */
1447 1455 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID,
1448 1456 &guid) != 0 ||
1449 1457 nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID,
1450 1458 &top_guid) != 0 ||
1451 1459 ((vd->vdev_guid != guid && vd->vdev_guid != aux_guid) &&
1452 1460 (vd->vdev_guid != top_guid || vd != vd->vdev_top))) {
1453 1461 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1454 1462 VDEV_AUX_CORRUPT_DATA);
1455 1463 nvlist_free(label);
1456 1464 return (0);
1457 1465 }
1458 1466
1459 1467 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1460 1468 &state) != 0) {
1461 1469 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1462 1470 VDEV_AUX_CORRUPT_DATA);
1463 1471 nvlist_free(label);
1464 1472 return (0);
1465 1473 }
1466 1474
1467 1475 nvlist_free(label);
1468 1476
1469 1477 /*
1470 1478 * If this is a verbatim import, no need to check the
1471 1479 * state of the pool.
1472 1480 */
1473 1481 if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) &&
1474 1482 spa_load_state(spa) == SPA_LOAD_OPEN &&
1475 1483 state != POOL_STATE_ACTIVE)
1476 1484 return (SET_ERROR(EBADF));
1477 1485
1478 1486 /*
1479 1487 * If we were able to open and validate a vdev that was
1480 1488 * previously marked permanently unavailable, clear that state
1481 1489 * now.
1482 1490 */
1483 1491 if (vd->vdev_not_present)
1484 1492 vd->vdev_not_present = 0;
1485 1493 }
1486 1494
1487 1495 return (0);
1488 1496 }
1489 1497
1490 1498 /*
1491 1499 * Close a virtual device.
1492 1500 */
1493 1501 void
1494 1502 vdev_close(vdev_t *vd)
1495 1503 {
1496 1504 spa_t *spa = vd->vdev_spa;
1497 1505 vdev_t *pvd = vd->vdev_parent;
1498 1506
1499 1507 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1500 1508
1501 1509 /*
1502 1510 * If our parent is reopening, then we are as well, unless we are
1503 1511 * going offline.
1504 1512 */
1505 1513 if (pvd != NULL && pvd->vdev_reopening)
1506 1514 vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline);
1507 1515
1508 1516 vd->vdev_ops->vdev_op_close(vd);
1509 1517
1510 1518 vdev_cache_purge(vd);
1511 1519
1512 1520 /*
1513 1521 * We record the previous state before we close it, so that if we are
1514 1522 * doing a reopen(), we don't generate FMA ereports if we notice that
1515 1523 * it's still faulted.
1516 1524 */
1517 1525 vd->vdev_prevstate = vd->vdev_state;
1518 1526
1519 1527 if (vd->vdev_offline)
1520 1528 vd->vdev_state = VDEV_STATE_OFFLINE;
1521 1529 else
1522 1530 vd->vdev_state = VDEV_STATE_CLOSED;
1523 1531 vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1524 1532 }
1525 1533
1526 1534 void
1527 1535 vdev_hold(vdev_t *vd)
1528 1536 {
1529 1537 spa_t *spa = vd->vdev_spa;
1530 1538
1531 1539 ASSERT(spa_is_root(spa));
1532 1540 if (spa->spa_state == POOL_STATE_UNINITIALIZED)
1533 1541 return;
1534 1542
1535 1543 for (int c = 0; c < vd->vdev_children; c++)
1536 1544 vdev_hold(vd->vdev_child[c]);
1537 1545
1538 1546 if (vd->vdev_ops->vdev_op_leaf)
1539 1547 vd->vdev_ops->vdev_op_hold(vd);
1540 1548 }
1541 1549
1542 1550 void
1543 1551 vdev_rele(vdev_t *vd)
1544 1552 {
1545 1553 spa_t *spa = vd->vdev_spa;
1546 1554
1547 1555 ASSERT(spa_is_root(spa));
1548 1556 for (int c = 0; c < vd->vdev_children; c++)
1549 1557 vdev_rele(vd->vdev_child[c]);
1550 1558
1551 1559 if (vd->vdev_ops->vdev_op_leaf)
1552 1560 vd->vdev_ops->vdev_op_rele(vd);
1553 1561 }
1554 1562
1555 1563 /*
1556 1564 * Reopen all interior vdevs and any unopened leaves. We don't actually
1557 1565 * reopen leaf vdevs which had previously been opened as they might deadlock
1558 1566 * on the spa_config_lock. Instead we only obtain the leaf's physical size.
1559 1567 * If the leaf has never been opened then open it, as usual.
1560 1568 */
1561 1569 void
1562 1570 vdev_reopen(vdev_t *vd)
1563 1571 {
1564 1572 spa_t *spa = vd->vdev_spa;
1565 1573
1566 1574 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1567 1575
1568 1576 /* set the reopening flag unless we're taking the vdev offline */
1569 1577 vd->vdev_reopening = !vd->vdev_offline;
1570 1578 vdev_close(vd);
1571 1579 (void) vdev_open(vd);
1572 1580
1573 1581 /*
1574 1582 * Call vdev_validate() here to make sure we have the same device.
1575 1583 * Otherwise, a device with an invalid label could be successfully
1576 1584 * opened in response to vdev_reopen().
1577 1585 */
1578 1586 if (vd->vdev_aux) {
1579 1587 (void) vdev_validate_aux(vd);
1580 1588 if (vdev_readable(vd) && vdev_writeable(vd) &&
1581 1589 vd->vdev_aux == &spa->spa_l2cache &&
1582 1590 !l2arc_vdev_present(vd))
1583 1591 l2arc_add_vdev(spa, vd);
1584 1592 } else {
1585 1593 (void) vdev_validate(vd, B_TRUE);
1586 1594 }
1587 1595
1588 1596 /*
1589 1597 * Reassess parent vdev's health.
1590 1598 */
1591 1599 vdev_propagate_state(vd);
1592 1600 }
1593 1601
1594 1602 int
1595 1603 vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1596 1604 {
1597 1605 int error;
1598 1606
1599 1607 /*
1600 1608 * Normally, partial opens (e.g. of a mirror) are allowed.
1601 1609 * For a create, however, we want to fail the request if
1602 1610 * there are any components we can't open.
1603 1611 */
1604 1612 error = vdev_open(vd);
1605 1613
1606 1614 if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
1607 1615 vdev_close(vd);
1608 1616 return (error ? error : ENXIO);
1609 1617 }
1610 1618
1611 1619 /*
1612 1620 * Recursively load DTLs and initialize all labels.
1613 1621 */
1614 1622 if ((error = vdev_dtl_load(vd)) != 0 ||
1615 1623 (error = vdev_label_init(vd, txg, isreplacing ?
1616 1624 VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
1617 1625 vdev_close(vd);
1618 1626 return (error);
1619 1627 }
1620 1628
1621 1629 return (0);
1622 1630 }
1623 1631
1624 1632 void
1625 1633 vdev_metaslab_set_size(vdev_t *vd)
1626 1634 {
1627 1635 /*
1628 1636 * Aim for roughly metaslabs_per_vdev (default 200) metaslabs per vdev.
1629 1637 */
1630 1638 vd->vdev_ms_shift = highbit64(vd->vdev_asize / metaslabs_per_vdev);
1631 1639 vd->vdev_ms_shift = MAX(vd->vdev_ms_shift, SPA_MAXBLOCKSHIFT);
1632 1640 }
1633 1641
1634 1642 void
1635 1643 vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
1636 1644 {
1637 1645 ASSERT(vd == vd->vdev_top);
1638 1646 ASSERT(!vd->vdev_ishole);
1639 1647 ASSERT(ISP2(flags));
1640 1648 ASSERT(spa_writeable(vd->vdev_spa));
1641 1649
1642 1650 if (flags & VDD_METASLAB)
1643 1651 (void) txg_list_add(&vd->vdev_ms_list, arg, txg);
1644 1652
1645 1653 if (flags & VDD_DTL)
1646 1654 (void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
1647 1655
1648 1656 (void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
1649 1657 }
1650 1658
1651 1659 void
1652 1660 vdev_dirty_leaves(vdev_t *vd, int flags, uint64_t txg)
1653 1661 {
1654 1662 for (int c = 0; c < vd->vdev_children; c++)
1655 1663 vdev_dirty_leaves(vd->vdev_child[c], flags, txg);
1656 1664
1657 1665 if (vd->vdev_ops->vdev_op_leaf)
1658 1666 vdev_dirty(vd->vdev_top, flags, vd, txg);
1659 1667 }
1660 1668
1661 1669 /*
1662 1670 * DTLs.
1663 1671 *
1664 1672 * A vdev's DTL (dirty time log) is the set of transaction groups for which
1665 1673 * the vdev has less than perfect replication. There are four kinds of DTL:
1666 1674 *
1667 1675 * DTL_MISSING: txgs for which the vdev has no valid copies of the data
1668 1676 *
1669 1677 * DTL_PARTIAL: txgs for which data is available, but not fully replicated
1670 1678 *
1671 1679 * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
1672 1680 * scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
1673 1681 * txgs that was scrubbed.
1674 1682 *
1675 1683 * DTL_OUTAGE: txgs which cannot currently be read, whether due to
1676 1684 * persistent errors or just some device being offline.
1677 1685 * Unlike the other three, the DTL_OUTAGE map is not generally
1678 1686 * maintained; it's only computed when needed, typically to
1679 1687 * determine whether a device can be detached.
1680 1688 *
1681 1689 * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
1682 1690 * either has the data or it doesn't.
1683 1691 *
1684 1692 * For interior vdevs such as mirror and RAID-Z the picture is more complex.
1685 1693 * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
1686 1694 * if any child is less than fully replicated, then so is its parent.
1687 1695 * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
1688 1696 * comprising only those txgs which appear in 'maxfaults' or more children;
1689 1697 * those are the txgs we don't have enough replication to read. For example,
1690 1698 * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
1691 1699 * thus, its DTL_MISSING consists of the set of txgs that appear in more than
1692 1700 * two child DTL_MISSING maps.
1693 1701 *
1694 1702 * It should be clear from the above that to compute the DTLs and outage maps
1695 1703 * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
1696 1704 * Therefore, that is all we keep on disk. When loading the pool, or after
1697 1705 * a configuration change, we generate all other DTLs from first principles.
1698 1706 */
1699 1707 void
1700 1708 vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1701 1709 {
1702 1710 range_tree_t *rt = vd->vdev_dtl[t];
1703 1711
1704 1712 ASSERT(t < DTL_TYPES);
1705 1713 ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1706 1714 ASSERT(spa_writeable(vd->vdev_spa));
1707 1715
1708 1716 mutex_enter(rt->rt_lock);
1709 1717 if (!range_tree_contains(rt, txg, size))
1710 1718 range_tree_add(rt, txg, size);
1711 1719 mutex_exit(rt->rt_lock);
1712 1720 }
1713 1721
1714 1722 boolean_t
1715 1723 vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1716 1724 {
1717 1725 range_tree_t *rt = vd->vdev_dtl[t];
1718 1726 boolean_t dirty = B_FALSE;
1719 1727
1720 1728 ASSERT(t < DTL_TYPES);
1721 1729 ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1722 1730
1723 1731 mutex_enter(rt->rt_lock);
1724 1732 if (range_tree_space(rt) != 0)
1725 1733 dirty = range_tree_contains(rt, txg, size);
1726 1734 mutex_exit(rt->rt_lock);
1727 1735
1728 1736 return (dirty);
1729 1737 }
1730 1738
1731 1739 boolean_t
1732 1740 vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
1733 1741 {
1734 1742 range_tree_t *rt = vd->vdev_dtl[t];
1735 1743 boolean_t empty;
1736 1744
1737 1745 mutex_enter(rt->rt_lock);
1738 1746 empty = (range_tree_space(rt) == 0);
1739 1747 mutex_exit(rt->rt_lock);
1740 1748
1741 1749 return (empty);
1742 1750 }
1743 1751
1744 1752 /*
1745 1753 * Returns the lowest txg in the DTL range.
1746 1754 */
1747 1755 static uint64_t
1748 1756 vdev_dtl_min(vdev_t *vd)
1749 1757 {
1750 1758 range_seg_t *rs;
1751 1759
1752 1760 ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
1753 1761 ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
1754 1762 ASSERT0(vd->vdev_children);
1755 1763
1756 1764 rs = avl_first(&vd->vdev_dtl[DTL_MISSING]->rt_root);
1757 1765 return (rs->rs_start - 1);
1758 1766 }
1759 1767
1760 1768 /*
1761 1769 * Returns the highest txg in the DTL.
1762 1770 */
1763 1771 static uint64_t
1764 1772 vdev_dtl_max(vdev_t *vd)
1765 1773 {
1766 1774 range_seg_t *rs;
1767 1775
1768 1776 ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
1769 1777 ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
1770 1778 ASSERT0(vd->vdev_children);
1771 1779
1772 1780 rs = avl_last(&vd->vdev_dtl[DTL_MISSING]->rt_root);
1773 1781 return (rs->rs_end);
1774 1782 }
1775 1783
1776 1784 /*
1777 1785 * Determine if a resilvering vdev should remove any DTL entries from
1778 1786 * its range. If the vdev was resilvering for the entire duration of the
1779 1787 * scan then it should excise that range from its DTLs. Otherwise, this
1780 1788 * vdev is considered partially resilvered and should leave its DTL
1781 1789 * entries intact. The comment in vdev_dtl_reassess() describes how we
1782 1790 * excise the DTLs.
1783 1791 */
1784 1792 static boolean_t
1785 1793 vdev_dtl_should_excise(vdev_t *vd)
1786 1794 {
1787 1795 spa_t *spa = vd->vdev_spa;
1788 1796 dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1789 1797
1790 1798 ASSERT0(scn->scn_phys.scn_errors);
1791 1799 ASSERT0(vd->vdev_children);
1792 1800
1793 1801 if (vd->vdev_state < VDEV_STATE_DEGRADED)
1794 1802 return (B_FALSE);
1795 1803
1796 1804 if (vd->vdev_resilver_txg == 0 ||
1797 1805 range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0)
1798 1806 return (B_TRUE);
1799 1807
1800 1808 /*
1801 1809 * When a resilver is initiated the scan will assign the scn_max_txg
1802 1810 * value to the highest txg value that exists in all DTLs. If this
1803 1811 * device's max DTL is not part of this scan (i.e. it is not in
1804 1812 * the range (scn_min_txg, scn_max_txg] then it is not eligible
1805 1813 * for excision.
1806 1814 */
1807 1815 if (vdev_dtl_max(vd) <= scn->scn_phys.scn_max_txg) {
1808 1816 ASSERT3U(scn->scn_phys.scn_min_txg, <=, vdev_dtl_min(vd));
1809 1817 ASSERT3U(scn->scn_phys.scn_min_txg, <, vd->vdev_resilver_txg);
1810 1818 ASSERT3U(vd->vdev_resilver_txg, <=, scn->scn_phys.scn_max_txg);
1811 1819 return (B_TRUE);
1812 1820 }
1813 1821 return (B_FALSE);
1814 1822 }
1815 1823
1816 1824 /*
1817 1825 * Reassess DTLs after a config change or scrub completion.
1818 1826 */
1819 1827 void
1820 1828 vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
1821 1829 {
1822 1830 spa_t *spa = vd->vdev_spa;
1823 1831 avl_tree_t reftree;
1824 1832 int minref;
1825 1833
1826 1834 ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
1827 1835
1828 1836 for (int c = 0; c < vd->vdev_children; c++)
1829 1837 vdev_dtl_reassess(vd->vdev_child[c], txg,
1830 1838 scrub_txg, scrub_done);
1831 1839
1832 1840 if (vd == spa->spa_root_vdev || vd->vdev_ishole || vd->vdev_aux)
1833 1841 return;
1834 1842
1835 1843 if (vd->vdev_ops->vdev_op_leaf) {
1836 1844 dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1837 1845
1838 1846 mutex_enter(&vd->vdev_dtl_lock);
1839 1847
1840 1848 /*
1841 1849 * If we've completed a scan cleanly then determine
1842 1850 * if this vdev should remove any DTLs. We only want to
1843 1851 * excise regions on vdevs that were available during
1844 1852 * the entire duration of this scan.
1845 1853 */
1846 1854 if (scrub_txg != 0 &&
1847 1855 (spa->spa_scrub_started ||
1848 1856 (scn != NULL && scn->scn_phys.scn_errors == 0)) &&
1849 1857 vdev_dtl_should_excise(vd)) {
1850 1858 /*
1851 1859 * We completed a scrub up to scrub_txg. If we
1852 1860 * did it without rebooting, then the scrub dtl
1853 1861 * will be valid, so excise the old region and
1854 1862 * fold in the scrub dtl. Otherwise, leave the
1855 1863 * dtl as-is if there was an error.
1856 1864 *
1857 1865 * There's little trick here: to excise the beginning
1858 1866 * of the DTL_MISSING map, we put it into a reference
1859 1867 * tree and then add a segment with refcnt -1 that
1860 1868 * covers the range [0, scrub_txg). This means
1861 1869 * that each txg in that range has refcnt -1 or 0.
1862 1870 * We then add DTL_SCRUB with a refcnt of 2, so that
1863 1871 * entries in the range [0, scrub_txg) will have a
1864 1872 * positive refcnt -- either 1 or 2. We then convert
1865 1873 * the reference tree into the new DTL_MISSING map.
1866 1874 */
1867 1875 space_reftree_create(&reftree);
1868 1876 space_reftree_add_map(&reftree,
1869 1877 vd->vdev_dtl[DTL_MISSING], 1);
1870 1878 space_reftree_add_seg(&reftree, 0, scrub_txg, -1);
1871 1879 space_reftree_add_map(&reftree,
1872 1880 vd->vdev_dtl[DTL_SCRUB], 2);
1873 1881 space_reftree_generate_map(&reftree,
1874 1882 vd->vdev_dtl[DTL_MISSING], 1);
1875 1883 space_reftree_destroy(&reftree);
1876 1884 }
1877 1885 range_tree_vacate(vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
1878 1886 range_tree_walk(vd->vdev_dtl[DTL_MISSING],
1879 1887 range_tree_add, vd->vdev_dtl[DTL_PARTIAL]);
1880 1888 if (scrub_done)
1881 1889 range_tree_vacate(vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
1882 1890 range_tree_vacate(vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
1883 1891 if (!vdev_readable(vd))
1884 1892 range_tree_add(vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
1885 1893 else
1886 1894 range_tree_walk(vd->vdev_dtl[DTL_MISSING],
1887 1895 range_tree_add, vd->vdev_dtl[DTL_OUTAGE]);
1888 1896
1889 1897 /*
1890 1898 * If the vdev was resilvering and no longer has any
1891 1899 * DTLs then reset its resilvering flag.
1892 1900 */
1893 1901 if (vd->vdev_resilver_txg != 0 &&
1894 1902 range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0 &&
1895 1903 range_tree_space(vd->vdev_dtl[DTL_OUTAGE]) == 0)
1896 1904 vd->vdev_resilver_txg = 0;
1897 1905
1898 1906 mutex_exit(&vd->vdev_dtl_lock);
1899 1907
1900 1908 if (txg != 0)
1901 1909 vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
1902 1910 return;
1903 1911 }
1904 1912
1905 1913 mutex_enter(&vd->vdev_dtl_lock);
1906 1914 for (int t = 0; t < DTL_TYPES; t++) {
1907 1915 /* account for child's outage in parent's missing map */
1908 1916 int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
1909 1917 if (t == DTL_SCRUB)
1910 1918 continue; /* leaf vdevs only */
1911 1919 if (t == DTL_PARTIAL)
1912 1920 minref = 1; /* i.e. non-zero */
1913 1921 else if (vd->vdev_nparity != 0)
1914 1922 minref = vd->vdev_nparity + 1; /* RAID-Z */
1915 1923 else
1916 1924 minref = vd->vdev_children; /* any kind of mirror */
1917 1925 space_reftree_create(&reftree);
1918 1926 for (int c = 0; c < vd->vdev_children; c++) {
1919 1927 vdev_t *cvd = vd->vdev_child[c];
1920 1928 mutex_enter(&cvd->vdev_dtl_lock);
1921 1929 space_reftree_add_map(&reftree, cvd->vdev_dtl[s], 1);
1922 1930 mutex_exit(&cvd->vdev_dtl_lock);
1923 1931 }
1924 1932 space_reftree_generate_map(&reftree, vd->vdev_dtl[t], minref);
1925 1933 space_reftree_destroy(&reftree);
1926 1934 }
1927 1935 mutex_exit(&vd->vdev_dtl_lock);
1928 1936 }
1929 1937
1930 1938 int
1931 1939 vdev_dtl_load(vdev_t *vd)
1932 1940 {
1933 1941 spa_t *spa = vd->vdev_spa;
1934 1942 objset_t *mos = spa->spa_meta_objset;
1935 1943 int error = 0;
1936 1944
1937 1945 if (vd->vdev_ops->vdev_op_leaf && vd->vdev_dtl_object != 0) {
1938 1946 ASSERT(!vd->vdev_ishole);
1939 1947
1940 1948 error = space_map_open(&vd->vdev_dtl_sm, mos,
1941 1949 vd->vdev_dtl_object, 0, -1ULL, 0, &vd->vdev_dtl_lock);
1942 1950 if (error)
1943 1951 return (error);
1944 1952 ASSERT(vd->vdev_dtl_sm != NULL);
1945 1953
1946 1954 mutex_enter(&vd->vdev_dtl_lock);
1947 1955
1948 1956 /*
1949 1957 * Now that we've opened the space_map we need to update
1950 1958 * the in-core DTL.
1951 1959 */
1952 1960 space_map_update(vd->vdev_dtl_sm);
1953 1961
1954 1962 error = space_map_load(vd->vdev_dtl_sm,
1955 1963 vd->vdev_dtl[DTL_MISSING], SM_ALLOC);
1956 1964 mutex_exit(&vd->vdev_dtl_lock);
1957 1965
1958 1966 return (error);
1959 1967 }
1960 1968
1961 1969 for (int c = 0; c < vd->vdev_children; c++) {
1962 1970 error = vdev_dtl_load(vd->vdev_child[c]);
1963 1971 if (error != 0)
1964 1972 break;
1965 1973 }
1966 1974
1967 1975 return (error);
1968 1976 }
1969 1977
1970 1978 void
1971 1979 vdev_destroy_unlink_zap(vdev_t *vd, uint64_t zapobj, dmu_tx_t *tx)
1972 1980 {
1973 1981 spa_t *spa = vd->vdev_spa;
1974 1982
1975 1983 VERIFY0(zap_destroy(spa->spa_meta_objset, zapobj, tx));
1976 1984 VERIFY0(zap_remove_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
1977 1985 zapobj, tx));
1978 1986 }
1979 1987
1980 1988 uint64_t
1981 1989 vdev_create_link_zap(vdev_t *vd, dmu_tx_t *tx)
1982 1990 {
1983 1991 spa_t *spa = vd->vdev_spa;
1984 1992 uint64_t zap = zap_create(spa->spa_meta_objset, DMU_OTN_ZAP_METADATA,
1985 1993 DMU_OT_NONE, 0, tx);
1986 1994
1987 1995 ASSERT(zap != 0);
1988 1996 VERIFY0(zap_add_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
1989 1997 zap, tx));
1990 1998
1991 1999 return (zap);
1992 2000 }
1993 2001
1994 2002 void
1995 2003 vdev_construct_zaps(vdev_t *vd, dmu_tx_t *tx)
1996 2004 {
1997 2005 if (vd->vdev_ops != &vdev_hole_ops &&
1998 2006 vd->vdev_ops != &vdev_missing_ops &&
1999 2007 vd->vdev_ops != &vdev_root_ops &&
2000 2008 !vd->vdev_top->vdev_removing) {
2001 2009 if (vd->vdev_ops->vdev_op_leaf && vd->vdev_leaf_zap == 0) {
2002 2010 vd->vdev_leaf_zap = vdev_create_link_zap(vd, tx);
2003 2011 }
2004 2012 if (vd == vd->vdev_top && vd->vdev_top_zap == 0) {
2005 2013 vd->vdev_top_zap = vdev_create_link_zap(vd, tx);
2006 2014 }
2007 2015 }
2008 2016 for (uint64_t i = 0; i < vd->vdev_children; i++) {
2009 2017 vdev_construct_zaps(vd->vdev_child[i], tx);
2010 2018 }
2011 2019 }
2012 2020
2013 2021 void
2014 2022 vdev_dtl_sync(vdev_t *vd, uint64_t txg)
2015 2023 {
2016 2024 spa_t *spa = vd->vdev_spa;
2017 2025 range_tree_t *rt = vd->vdev_dtl[DTL_MISSING];
2018 2026 objset_t *mos = spa->spa_meta_objset;
2019 2027 range_tree_t *rtsync;
2020 2028 kmutex_t rtlock;
2021 2029 dmu_tx_t *tx;
2022 2030 uint64_t object = space_map_object(vd->vdev_dtl_sm);
2023 2031
2024 2032 ASSERT(!vd->vdev_ishole);
2025 2033 ASSERT(vd->vdev_ops->vdev_op_leaf);
2026 2034
2027 2035 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2028 2036
2029 2037 if (vd->vdev_detached || vd->vdev_top->vdev_removing) {
2030 2038 mutex_enter(&vd->vdev_dtl_lock);
2031 2039 space_map_free(vd->vdev_dtl_sm, tx);
2032 2040 space_map_close(vd->vdev_dtl_sm);
2033 2041 vd->vdev_dtl_sm = NULL;
2034 2042 mutex_exit(&vd->vdev_dtl_lock);
2035 2043
2036 2044 /*
2037 2045 * We only destroy the leaf ZAP for detached leaves or for
2038 2046 * removed log devices. Removed data devices handle leaf ZAP
2039 2047 * cleanup later, once cancellation is no longer possible.
2040 2048 */
2041 2049 if (vd->vdev_leaf_zap != 0 && (vd->vdev_detached ||
2042 2050 vd->vdev_top->vdev_islog)) {
2043 2051 vdev_destroy_unlink_zap(vd, vd->vdev_leaf_zap, tx);
2044 2052 vd->vdev_leaf_zap = 0;
2045 2053 }
2046 2054
2047 2055 dmu_tx_commit(tx);
2048 2056 return;
2049 2057 }
2050 2058
2051 2059 if (vd->vdev_dtl_sm == NULL) {
2052 2060 uint64_t new_object;
2053 2061
2054 2062 new_object = space_map_alloc(mos, tx);
2055 2063 VERIFY3U(new_object, !=, 0);
2056 2064
2057 2065 VERIFY0(space_map_open(&vd->vdev_dtl_sm, mos, new_object,
2058 2066 0, -1ULL, 0, &vd->vdev_dtl_lock));
2059 2067 ASSERT(vd->vdev_dtl_sm != NULL);
2060 2068 }
2061 2069
2062 2070 mutex_init(&rtlock, NULL, MUTEX_DEFAULT, NULL);
2063 2071
2064 2072 rtsync = range_tree_create(NULL, NULL, &rtlock);
2065 2073
2066 2074 mutex_enter(&rtlock);
2067 2075
2068 2076 mutex_enter(&vd->vdev_dtl_lock);
2069 2077 range_tree_walk(rt, range_tree_add, rtsync);
2070 2078 mutex_exit(&vd->vdev_dtl_lock);
2071 2079
2072 2080 space_map_truncate(vd->vdev_dtl_sm, tx);
2073 2081 space_map_write(vd->vdev_dtl_sm, rtsync, SM_ALLOC, tx);
2074 2082 range_tree_vacate(rtsync, NULL, NULL);
2075 2083
2076 2084 range_tree_destroy(rtsync);
2077 2085
2078 2086 mutex_exit(&rtlock);
2079 2087 mutex_destroy(&rtlock);
2080 2088
2081 2089 /*
2082 2090 * If the object for the space map has changed then dirty
2083 2091 * the top level so that we update the config.
2084 2092 */
2085 2093 if (object != space_map_object(vd->vdev_dtl_sm)) {
2086 2094 zfs_dbgmsg("txg %llu, spa %s, DTL old object %llu, "
2087 2095 "new object %llu", txg, spa_name(spa), object,
2088 2096 space_map_object(vd->vdev_dtl_sm));
2089 2097 vdev_config_dirty(vd->vdev_top);
2090 2098 }
2091 2099
2092 2100 dmu_tx_commit(tx);
2093 2101
2094 2102 mutex_enter(&vd->vdev_dtl_lock);
2095 2103 space_map_update(vd->vdev_dtl_sm);
2096 2104 mutex_exit(&vd->vdev_dtl_lock);
2097 2105 }
2098 2106
2099 2107 /*
2100 2108 * Determine whether the specified vdev can be offlined/detached/removed
2101 2109 * without losing data.
2102 2110 */
2103 2111 boolean_t
2104 2112 vdev_dtl_required(vdev_t *vd)
2105 2113 {
2106 2114 spa_t *spa = vd->vdev_spa;
2107 2115 vdev_t *tvd = vd->vdev_top;
2108 2116 uint8_t cant_read = vd->vdev_cant_read;
2109 2117 boolean_t required;
2110 2118
2111 2119 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2112 2120
2113 2121 if (vd == spa->spa_root_vdev || vd == tvd)
2114 2122 return (B_TRUE);
2115 2123
2116 2124 /*
2117 2125 * Temporarily mark the device as unreadable, and then determine
2118 2126 * whether this results in any DTL outages in the top-level vdev.
2119 2127 * If not, we can safely offline/detach/remove the device.
2120 2128 */
2121 2129 vd->vdev_cant_read = B_TRUE;
2122 2130 vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
2123 2131 required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
2124 2132 vd->vdev_cant_read = cant_read;
2125 2133 vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
2126 2134
2127 2135 if (!required && zio_injection_enabled)
2128 2136 required = !!zio_handle_device_injection(vd, NULL, ECHILD);
2129 2137
2130 2138 return (required);
2131 2139 }
2132 2140
2133 2141 /*
2134 2142 * Determine if resilver is needed, and if so the txg range.
2135 2143 */
2136 2144 boolean_t
2137 2145 vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
2138 2146 {
2139 2147 boolean_t needed = B_FALSE;
2140 2148 uint64_t thismin = UINT64_MAX;
2141 2149 uint64_t thismax = 0;
2142 2150
2143 2151 if (vd->vdev_children == 0) {
2144 2152 mutex_enter(&vd->vdev_dtl_lock);
2145 2153 if (range_tree_space(vd->vdev_dtl[DTL_MISSING]) != 0 &&
2146 2154 vdev_writeable(vd)) {
2147 2155
2148 2156 thismin = vdev_dtl_min(vd);
2149 2157 thismax = vdev_dtl_max(vd);
2150 2158 needed = B_TRUE;
2151 2159 }
2152 2160 mutex_exit(&vd->vdev_dtl_lock);
2153 2161 } else {
2154 2162 for (int c = 0; c < vd->vdev_children; c++) {
2155 2163 vdev_t *cvd = vd->vdev_child[c];
2156 2164 uint64_t cmin, cmax;
2157 2165
2158 2166 if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
2159 2167 thismin = MIN(thismin, cmin);
2160 2168 thismax = MAX(thismax, cmax);
2161 2169 needed = B_TRUE;
2162 2170 }
2163 2171 }
2164 2172 }
2165 2173
2166 2174 if (needed && minp) {
2167 2175 *minp = thismin;
2168 2176 *maxp = thismax;
2169 2177 }
2170 2178 return (needed);
2171 2179 }
2172 2180
2173 2181 void
2174 2182 vdev_load(vdev_t *vd)
2175 2183 {
2176 2184 /*
2177 2185 * Recursively load all children.
2178 2186 */
2179 2187 for (int c = 0; c < vd->vdev_children; c++)
2180 2188 vdev_load(vd->vdev_child[c]);
2181 2189
2182 2190 /*
2183 2191 * If this is a top-level vdev, initialize its metaslabs.
2184 2192 */
2185 2193 if (vd == vd->vdev_top && !vd->vdev_ishole &&
2186 2194 (vd->vdev_ashift == 0 || vd->vdev_asize == 0 ||
2187 2195 vdev_metaslab_init(vd, 0) != 0))
2188 2196 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2189 2197 VDEV_AUX_CORRUPT_DATA);
2190 2198
2191 2199 /*
2192 2200 * If this is a leaf vdev, load its DTL.
2193 2201 */
2194 2202 if (vd->vdev_ops->vdev_op_leaf && vdev_dtl_load(vd) != 0)
2195 2203 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2196 2204 VDEV_AUX_CORRUPT_DATA);
2197 2205 }
2198 2206
2199 2207 /*
2200 2208 * The special vdev case is used for hot spares and l2cache devices. Its
2201 2209 * sole purpose it to set the vdev state for the associated vdev. To do this,
2202 2210 * we make sure that we can open the underlying device, then try to read the
2203 2211 * label, and make sure that the label is sane and that it hasn't been
2204 2212 * repurposed to another pool.
2205 2213 */
2206 2214 int
2207 2215 vdev_validate_aux(vdev_t *vd)
2208 2216 {
2209 2217 nvlist_t *label;
2210 2218 uint64_t guid, version;
2211 2219 uint64_t state;
2212 2220
2213 2221 if (!vdev_readable(vd))
2214 2222 return (0);
2215 2223
2216 2224 if ((label = vdev_label_read_config(vd, -1ULL)) == NULL) {
2217 2225 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2218 2226 VDEV_AUX_CORRUPT_DATA);
2219 2227 return (-1);
2220 2228 }
2221 2229
2222 2230 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
2223 2231 !SPA_VERSION_IS_SUPPORTED(version) ||
2224 2232 nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
2225 2233 guid != vd->vdev_guid ||
2226 2234 nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
2227 2235 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2228 2236 VDEV_AUX_CORRUPT_DATA);
2229 2237 nvlist_free(label);
2230 2238 return (-1);
2231 2239 }
2232 2240
2233 2241 /*
2234 2242 * We don't actually check the pool state here. If it's in fact in
2235 2243 * use by another pool, we update this fact on the fly when requested.
2236 2244 */
2237 2245 nvlist_free(label);
2238 2246 return (0);
2239 2247 }
2240 2248
2241 2249 void
2242 2250 vdev_remove(vdev_t *vd, uint64_t txg)
2243 2251 {
2244 2252 spa_t *spa = vd->vdev_spa;
2245 2253 objset_t *mos = spa->spa_meta_objset;
2246 2254 dmu_tx_t *tx;
2247 2255
2248 2256 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2249 2257 ASSERT(vd == vd->vdev_top);
2250 2258 ASSERT3U(txg, ==, spa_syncing_txg(spa));
2251 2259
2252 2260 if (vd->vdev_ms != NULL) {
2253 2261 metaslab_group_t *mg = vd->vdev_mg;
2254 2262
2255 2263 metaslab_group_histogram_verify(mg);
2256 2264 metaslab_class_histogram_verify(mg->mg_class);
2257 2265
2258 2266 for (int m = 0; m < vd->vdev_ms_count; m++) {
2259 2267 metaslab_t *msp = vd->vdev_ms[m];
2260 2268
2261 2269 if (msp == NULL || msp->ms_sm == NULL)
2262 2270 continue;
2263 2271
2264 2272 mutex_enter(&msp->ms_lock);
2265 2273 /*
2266 2274 * If the metaslab was not loaded when the vdev
2267 2275 * was removed then the histogram accounting may
2268 2276 * not be accurate. Update the histogram information
2269 2277 * here so that we ensure that the metaslab group
2270 2278 * and metaslab class are up-to-date.
2271 2279 */
2272 2280 metaslab_group_histogram_remove(mg, msp);
2273 2281
2274 2282 VERIFY0(space_map_allocated(msp->ms_sm));
2275 2283 space_map_free(msp->ms_sm, tx);
2276 2284 space_map_close(msp->ms_sm);
2277 2285 msp->ms_sm = NULL;
2278 2286 mutex_exit(&msp->ms_lock);
2279 2287 }
2280 2288
2281 2289 metaslab_group_histogram_verify(mg);
2282 2290 metaslab_class_histogram_verify(mg->mg_class);
2283 2291 for (int i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
2284 2292 ASSERT0(mg->mg_histogram[i]);
2285 2293
2286 2294 }
2287 2295
2288 2296 if (vd->vdev_ms_array) {
2289 2297 (void) dmu_object_free(mos, vd->vdev_ms_array, tx);
2290 2298 vd->vdev_ms_array = 0;
2291 2299 }
2292 2300
2293 2301 if (vd->vdev_islog && vd->vdev_top_zap != 0) {
2294 2302 vdev_destroy_unlink_zap(vd, vd->vdev_top_zap, tx);
2295 2303 vd->vdev_top_zap = 0;
2296 2304 }
2297 2305 dmu_tx_commit(tx);
2298 2306 }
2299 2307
2300 2308 void
2301 2309 vdev_sync_done(vdev_t *vd, uint64_t txg)
2302 2310 {
2303 2311 metaslab_t *msp;
2304 2312 boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg));
2305 2313
2306 2314 ASSERT(!vd->vdev_ishole);
2307 2315
2308 2316 while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
2309 2317 metaslab_sync_done(msp, txg);
2310 2318
2311 2319 if (reassess)
2312 2320 metaslab_sync_reassess(vd->vdev_mg);
2313 2321 }
2314 2322
2315 2323 void
2316 2324 vdev_sync(vdev_t *vd, uint64_t txg)
2317 2325 {
2318 2326 spa_t *spa = vd->vdev_spa;
2319 2327 vdev_t *lvd;
2320 2328 metaslab_t *msp;
2321 2329 dmu_tx_t *tx;
2322 2330
2323 2331 ASSERT(!vd->vdev_ishole);
2324 2332
2325 2333 if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0) {
2326 2334 ASSERT(vd == vd->vdev_top);
2327 2335 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2328 2336 vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
2329 2337 DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
2330 2338 ASSERT(vd->vdev_ms_array != 0);
2331 2339 vdev_config_dirty(vd);
2332 2340 dmu_tx_commit(tx);
2333 2341 }
2334 2342
2335 2343 /*
2336 2344 * Remove the metadata associated with this vdev once it's empty.
2337 2345 */
2338 2346 if (vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing)
2339 2347 vdev_remove(vd, txg);
2340 2348
2341 2349 while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
2342 2350 metaslab_sync(msp, txg);
2343 2351 (void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
2344 2352 }
2345 2353
2346 2354 while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
2347 2355 vdev_dtl_sync(lvd, txg);
2348 2356
2349 2357 (void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
2350 2358 }
2351 2359
2352 2360 uint64_t
2353 2361 vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
2354 2362 {
2355 2363 return (vd->vdev_ops->vdev_op_asize(vd, psize));
2356 2364 }
2357 2365
2358 2366 /*
2359 2367 * Mark the given vdev faulted. A faulted vdev behaves as if the device could
2360 2368 * not be opened, and no I/O is attempted.
2361 2369 */
2362 2370 int
2363 2371 vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2364 2372 {
2365 2373 vdev_t *vd, *tvd;
2366 2374
2367 2375 spa_vdev_state_enter(spa, SCL_NONE);
2368 2376
2369 2377 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2370 2378 return (spa_vdev_state_exit(spa, NULL, ENODEV));
2371 2379
2372 2380 if (!vd->vdev_ops->vdev_op_leaf)
2373 2381 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2374 2382
2375 2383 tvd = vd->vdev_top;
2376 2384
2377 2385 /*
2378 2386 * We don't directly use the aux state here, but if we do a
2379 2387 * vdev_reopen(), we need this value to be present to remember why we
2380 2388 * were faulted.
2381 2389 */
2382 2390 vd->vdev_label_aux = aux;
2383 2391
2384 2392 /*
2385 2393 * Faulted state takes precedence over degraded.
2386 2394 */
2387 2395 vd->vdev_delayed_close = B_FALSE;
2388 2396 vd->vdev_faulted = 1ULL;
2389 2397 vd->vdev_degraded = 0ULL;
2390 2398 vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
2391 2399
2392 2400 /*
2393 2401 * If this device has the only valid copy of the data, then
2394 2402 * back off and simply mark the vdev as degraded instead.
2395 2403 */
2396 2404 if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) {
2397 2405 vd->vdev_degraded = 1ULL;
2398 2406 vd->vdev_faulted = 0ULL;
2399 2407
2400 2408 /*
2401 2409 * If we reopen the device and it's not dead, only then do we
2402 2410 * mark it degraded.
2403 2411 */
2404 2412 vdev_reopen(tvd);
2405 2413
2406 2414 if (vdev_readable(vd))
2407 2415 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
2408 2416 }
2409 2417
2410 2418 return (spa_vdev_state_exit(spa, vd, 0));
2411 2419 }
2412 2420
2413 2421 /*
2414 2422 * Mark the given vdev degraded. A degraded vdev is purely an indication to the
2415 2423 * user that something is wrong. The vdev continues to operate as normal as far
2416 2424 * as I/O is concerned.
2417 2425 */
2418 2426 int
2419 2427 vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2420 2428 {
2421 2429 vdev_t *vd;
2422 2430
2423 2431 spa_vdev_state_enter(spa, SCL_NONE);
2424 2432
2425 2433 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2426 2434 return (spa_vdev_state_exit(spa, NULL, ENODEV));
2427 2435
2428 2436 if (!vd->vdev_ops->vdev_op_leaf)
2429 2437 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2430 2438
2431 2439 /*
2432 2440 * If the vdev is already faulted, then don't do anything.
2433 2441 */
2434 2442 if (vd->vdev_faulted || vd->vdev_degraded)
2435 2443 return (spa_vdev_state_exit(spa, NULL, 0));
2436 2444
2437 2445 vd->vdev_degraded = 1ULL;
2438 2446 if (!vdev_is_dead(vd))
2439 2447 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
2440 2448 aux);
2441 2449
2442 2450 return (spa_vdev_state_exit(spa, vd, 0));
2443 2451 }
2444 2452
2445 2453 /*
2446 2454 * Online the given vdev.
2447 2455 *
2448 2456 * If 'ZFS_ONLINE_UNSPARE' is set, it implies two things. First, any attached
2449 2457 * spare device should be detached when the device finishes resilvering.
2450 2458 * Second, the online should be treated like a 'test' online case, so no FMA
2451 2459 * events are generated if the device fails to open.
2452 2460 */
2453 2461 int
2454 2462 vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
2455 2463 {
2456 2464 vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
2457 2465 boolean_t wasoffline;
2458 2466 vdev_state_t oldstate;
2459 2467
2460 2468 spa_vdev_state_enter(spa, SCL_NONE);
2461 2469
2462 2470 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2463 2471 return (spa_vdev_state_exit(spa, NULL, ENODEV));
2464 2472
2465 2473 if (!vd->vdev_ops->vdev_op_leaf)
2466 2474 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2467 2475
2468 2476 wasoffline = (vd->vdev_offline || vd->vdev_tmpoffline);
2469 2477 oldstate = vd->vdev_state;
2470 2478
2471 2479 tvd = vd->vdev_top;
2472 2480 vd->vdev_offline = B_FALSE;
2473 2481 vd->vdev_tmpoffline = B_FALSE;
2474 2482 vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
2475 2483 vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
2476 2484
2477 2485 /* XXX - L2ARC 1.0 does not support expansion */
2478 2486 if (!vd->vdev_aux) {
2479 2487 for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2480 2488 pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
2481 2489 }
2482 2490
2483 2491 vdev_reopen(tvd);
2484 2492 vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
2485 2493
2486 2494 if (!vd->vdev_aux) {
2487 2495 for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2488 2496 pvd->vdev_expanding = B_FALSE;
2489 2497 }
2490 2498
2491 2499 if (newstate)
2492 2500 *newstate = vd->vdev_state;
2493 2501 if ((flags & ZFS_ONLINE_UNSPARE) &&
2494 2502 !vdev_is_dead(vd) && vd->vdev_parent &&
2495 2503 vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2496 2504 vd->vdev_parent->vdev_child[0] == vd)
2497 2505 vd->vdev_unspare = B_TRUE;
2498 2506
2499 2507 if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
2500 2508
2501 2509 /* XXX - L2ARC 1.0 does not support expansion */
2502 2510 if (vd->vdev_aux)
2503 2511 return (spa_vdev_state_exit(spa, vd, ENOTSUP));
2504 2512 spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
2505 2513 }
2506 2514
2507 2515 if (wasoffline ||
2508 2516 (oldstate < VDEV_STATE_DEGRADED &&
2509 2517 vd->vdev_state >= VDEV_STATE_DEGRADED))
2510 2518 spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_ONLINE);
2511 2519
2512 2520 return (spa_vdev_state_exit(spa, vd, 0));
2513 2521 }
2514 2522
2515 2523 static int
2516 2524 vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
2517 2525 {
2518 2526 vdev_t *vd, *tvd;
2519 2527 int error = 0;
2520 2528 uint64_t generation;
2521 2529 metaslab_group_t *mg;
2522 2530
2523 2531 top:
2524 2532 spa_vdev_state_enter(spa, SCL_ALLOC);
2525 2533
2526 2534 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2527 2535 return (spa_vdev_state_exit(spa, NULL, ENODEV));
2528 2536
2529 2537 if (!vd->vdev_ops->vdev_op_leaf)
2530 2538 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2531 2539
2532 2540 tvd = vd->vdev_top;
2533 2541 mg = tvd->vdev_mg;
2534 2542 generation = spa->spa_config_generation + 1;
2535 2543
2536 2544 /*
2537 2545 * If the device isn't already offline, try to offline it.
2538 2546 */
2539 2547 if (!vd->vdev_offline) {
2540 2548 /*
2541 2549 * If this device has the only valid copy of some data,
2542 2550 * don't allow it to be offlined. Log devices are always
2543 2551 * expendable.
2544 2552 */
2545 2553 if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2546 2554 vdev_dtl_required(vd))
2547 2555 return (spa_vdev_state_exit(spa, NULL, EBUSY));
2548 2556
2549 2557 /*
2550 2558 * If the top-level is a slog and it has had allocations
2551 2559 * then proceed. We check that the vdev's metaslab group
2552 2560 * is not NULL since it's possible that we may have just
2553 2561 * added this vdev but not yet initialized its metaslabs.
2554 2562 */
2555 2563 if (tvd->vdev_islog && mg != NULL) {
2556 2564 /*
2557 2565 * Prevent any future allocations.
2558 2566 */
2559 2567 metaslab_group_passivate(mg);
2560 2568 (void) spa_vdev_state_exit(spa, vd, 0);
2561 2569
2562 2570 error = spa_offline_log(spa);
2563 2571
2564 2572 spa_vdev_state_enter(spa, SCL_ALLOC);
2565 2573
2566 2574 /*
2567 2575 * Check to see if the config has changed.
2568 2576 */
2569 2577 if (error || generation != spa->spa_config_generation) {
2570 2578 metaslab_group_activate(mg);
2571 2579 if (error)
2572 2580 return (spa_vdev_state_exit(spa,
2573 2581 vd, error));
2574 2582 (void) spa_vdev_state_exit(spa, vd, 0);
2575 2583 goto top;
2576 2584 }
2577 2585 ASSERT0(tvd->vdev_stat.vs_alloc);
2578 2586 }
2579 2587
2580 2588 /*
2581 2589 * Offline this device and reopen its top-level vdev.
2582 2590 * If the top-level vdev is a log device then just offline
2583 2591 * it. Otherwise, if this action results in the top-level
2584 2592 * vdev becoming unusable, undo it and fail the request.
2585 2593 */
2586 2594 vd->vdev_offline = B_TRUE;
2587 2595 vdev_reopen(tvd);
2588 2596
2589 2597 if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2590 2598 vdev_is_dead(tvd)) {
2591 2599 vd->vdev_offline = B_FALSE;
2592 2600 vdev_reopen(tvd);
2593 2601 return (spa_vdev_state_exit(spa, NULL, EBUSY));
2594 2602 }
2595 2603
2596 2604 /*
2597 2605 * Add the device back into the metaslab rotor so that
2598 2606 * once we online the device it's open for business.
2599 2607 */
2600 2608 if (tvd->vdev_islog && mg != NULL)
2601 2609 metaslab_group_activate(mg);
2602 2610 }
2603 2611
2604 2612 vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
2605 2613
2606 2614 return (spa_vdev_state_exit(spa, vd, 0));
2607 2615 }
2608 2616
2609 2617 int
2610 2618 vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
2611 2619 {
2612 2620 int error;
2613 2621
2614 2622 mutex_enter(&spa->spa_vdev_top_lock);
2615 2623 error = vdev_offline_locked(spa, guid, flags);
2616 2624 mutex_exit(&spa->spa_vdev_top_lock);
2617 2625
2618 2626 return (error);
2619 2627 }
2620 2628
2621 2629 /*
2622 2630 * Clear the error counts associated with this vdev. Unlike vdev_online() and
2623 2631 * vdev_offline(), we assume the spa config is locked. We also clear all
2624 2632 * children. If 'vd' is NULL, then the user wants to clear all vdevs.
2625 2633 */
2626 2634 void
2627 2635 vdev_clear(spa_t *spa, vdev_t *vd)
2628 2636 {
2629 2637 vdev_t *rvd = spa->spa_root_vdev;
2630 2638
2631 2639 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2632 2640
2633 2641 if (vd == NULL)
2634 2642 vd = rvd;
2635 2643
2636 2644 vd->vdev_stat.vs_read_errors = 0;
2637 2645 vd->vdev_stat.vs_write_errors = 0;
2638 2646 vd->vdev_stat.vs_checksum_errors = 0;
2639 2647
2640 2648 for (int c = 0; c < vd->vdev_children; c++)
2641 2649 vdev_clear(spa, vd->vdev_child[c]);
2642 2650
2643 2651 /*
2644 2652 * If we're in the FAULTED state or have experienced failed I/O, then
2645 2653 * clear the persistent state and attempt to reopen the device. We
2646 2654 * also mark the vdev config dirty, so that the new faulted state is
2647 2655 * written out to disk.
2648 2656 */
2649 2657 if (vd->vdev_faulted || vd->vdev_degraded ||
2650 2658 !vdev_readable(vd) || !vdev_writeable(vd)) {
2651 2659
2652 2660 /*
2653 2661 * When reopening in reponse to a clear event, it may be due to
2654 2662 * a fmadm repair request. In this case, if the device is
2655 2663 * still broken, we want to still post the ereport again.
2656 2664 */
2657 2665 vd->vdev_forcefault = B_TRUE;
2658 2666
2659 2667 vd->vdev_faulted = vd->vdev_degraded = 0ULL;
2660 2668 vd->vdev_cant_read = B_FALSE;
2661 2669 vd->vdev_cant_write = B_FALSE;
2662 2670
2663 2671 vdev_reopen(vd == rvd ? rvd : vd->vdev_top);
2664 2672
2665 2673 vd->vdev_forcefault = B_FALSE;
2666 2674
2667 2675 if (vd != rvd && vdev_writeable(vd->vdev_top))
2668 2676 vdev_state_dirty(vd->vdev_top);
2669 2677
2670 2678 if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
2671 2679 spa_async_request(spa, SPA_ASYNC_RESILVER);
2672 2680
2673 2681 spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_CLEAR);
2674 2682 }
2675 2683
2676 2684 /*
2677 2685 * When clearing a FMA-diagnosed fault, we always want to
2678 2686 * unspare the device, as we assume that the original spare was
2679 2687 * done in response to the FMA fault.
2680 2688 */
2681 2689 if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
2682 2690 vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2683 2691 vd->vdev_parent->vdev_child[0] == vd)
2684 2692 vd->vdev_unspare = B_TRUE;
2685 2693 }
2686 2694
2687 2695 boolean_t
2688 2696 vdev_is_dead(vdev_t *vd)
2689 2697 {
2690 2698 /*
2691 2699 * Holes and missing devices are always considered "dead".
2692 2700 * This simplifies the code since we don't have to check for
2693 2701 * these types of devices in the various code paths.
2694 2702 * Instead we rely on the fact that we skip over dead devices
2695 2703 * before issuing I/O to them.
2696 2704 */
2697 2705 return (vd->vdev_state < VDEV_STATE_DEGRADED || vd->vdev_ishole ||
2698 2706 vd->vdev_ops == &vdev_missing_ops);
2699 2707 }
2700 2708
2701 2709 boolean_t
2702 2710 vdev_readable(vdev_t *vd)
2703 2711 {
2704 2712 return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
2705 2713 }
2706 2714
2707 2715 boolean_t
2708 2716 vdev_writeable(vdev_t *vd)
2709 2717 {
2710 2718 return (!vdev_is_dead(vd) && !vd->vdev_cant_write);
2711 2719 }
2712 2720
2713 2721 boolean_t
2714 2722 vdev_allocatable(vdev_t *vd)
2715 2723 {
2716 2724 uint64_t state = vd->vdev_state;
2717 2725
2718 2726 /*
2719 2727 * We currently allow allocations from vdevs which may be in the
2720 2728 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
2721 2729 * fails to reopen then we'll catch it later when we're holding
2722 2730 * the proper locks. Note that we have to get the vdev state
2723 2731 * in a local variable because although it changes atomically,
2724 2732 * we're asking two separate questions about it.
2725 2733 */
2726 2734 return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
2727 2735 !vd->vdev_cant_write && !vd->vdev_ishole &&
2728 2736 vd->vdev_mg->mg_initialized);
2729 2737 }
2730 2738
2731 2739 boolean_t
2732 2740 vdev_accessible(vdev_t *vd, zio_t *zio)
2733 2741 {
2734 2742 ASSERT(zio->io_vd == vd);
2735 2743
2736 2744 if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
2737 2745 return (B_FALSE);
2738 2746
2739 2747 if (zio->io_type == ZIO_TYPE_READ)
2740 2748 return (!vd->vdev_cant_read);
2741 2749
2742 2750 if (zio->io_type == ZIO_TYPE_WRITE)
2743 2751 return (!vd->vdev_cant_write);
2744 2752
2745 2753 return (B_TRUE);
2746 2754 }
2747 2755
2748 2756 /*
2749 2757 * Get statistics for the given vdev.
2750 2758 */
2751 2759 void
2752 2760 vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
2753 2761 {
2754 2762 spa_t *spa = vd->vdev_spa;
2755 2763 vdev_t *rvd = spa->spa_root_vdev;
2756 2764 vdev_t *tvd = vd->vdev_top;
2757 2765
2758 2766 ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
2759 2767
2760 2768 mutex_enter(&vd->vdev_stat_lock);
2761 2769 bcopy(&vd->vdev_stat, vs, sizeof (*vs));
2762 2770 vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
2763 2771 vs->vs_state = vd->vdev_state;
2764 2772 vs->vs_rsize = vdev_get_min_asize(vd);
2765 2773 if (vd->vdev_ops->vdev_op_leaf)
2766 2774 vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
2767 2775 /*
2768 2776 * Report expandable space on top-level, non-auxillary devices only.
2769 2777 * The expandable space is reported in terms of metaslab sized units
2770 2778 * since that determines how much space the pool can expand.
2771 2779 */
2772 2780 if (vd->vdev_aux == NULL && tvd != NULL) {
2773 2781 vs->vs_esize = P2ALIGN(vd->vdev_max_asize - vd->vdev_asize -
2774 2782 spa->spa_bootsize, 1ULL << tvd->vdev_ms_shift);
2775 2783 }
2776 2784 if (vd->vdev_aux == NULL && vd == vd->vdev_top && !vd->vdev_ishole) {
2777 2785 vs->vs_fragmentation = vd->vdev_mg->mg_fragmentation;
2778 2786 }
2779 2787
2780 2788 /*
2781 2789 * If we're getting stats on the root vdev, aggregate the I/O counts
2782 2790 * over all top-level vdevs (i.e. the direct children of the root).
2783 2791 */
2784 2792 if (vd == rvd) {
2785 2793 for (int c = 0; c < rvd->vdev_children; c++) {
2786 2794 vdev_t *cvd = rvd->vdev_child[c];
2787 2795 vdev_stat_t *cvs = &cvd->vdev_stat;
2788 2796
2789 2797 for (int t = 0; t < ZIO_TYPES; t++) {
2790 2798 vs->vs_ops[t] += cvs->vs_ops[t];
2791 2799 vs->vs_bytes[t] += cvs->vs_bytes[t];
2792 2800 }
2793 2801 cvs->vs_scan_removing = cvd->vdev_removing;
2794 2802 }
2795 2803 }
2796 2804 mutex_exit(&vd->vdev_stat_lock);
2797 2805 }
2798 2806
2799 2807 void
2800 2808 vdev_clear_stats(vdev_t *vd)
2801 2809 {
2802 2810 mutex_enter(&vd->vdev_stat_lock);
2803 2811 vd->vdev_stat.vs_space = 0;
2804 2812 vd->vdev_stat.vs_dspace = 0;
2805 2813 vd->vdev_stat.vs_alloc = 0;
2806 2814 mutex_exit(&vd->vdev_stat_lock);
2807 2815 }
2808 2816
2809 2817 void
2810 2818 vdev_scan_stat_init(vdev_t *vd)
2811 2819 {
2812 2820 vdev_stat_t *vs = &vd->vdev_stat;
2813 2821
2814 2822 for (int c = 0; c < vd->vdev_children; c++)
2815 2823 vdev_scan_stat_init(vd->vdev_child[c]);
2816 2824
2817 2825 mutex_enter(&vd->vdev_stat_lock);
2818 2826 vs->vs_scan_processed = 0;
2819 2827 mutex_exit(&vd->vdev_stat_lock);
2820 2828 }
2821 2829
2822 2830 void
2823 2831 vdev_stat_update(zio_t *zio, uint64_t psize)
2824 2832 {
2825 2833 spa_t *spa = zio->io_spa;
2826 2834 vdev_t *rvd = spa->spa_root_vdev;
2827 2835 vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
2828 2836 vdev_t *pvd;
2829 2837 uint64_t txg = zio->io_txg;
2830 2838 vdev_stat_t *vs = &vd->vdev_stat;
2831 2839 zio_type_t type = zio->io_type;
2832 2840 int flags = zio->io_flags;
2833 2841
2834 2842 /*
2835 2843 * If this i/o is a gang leader, it didn't do any actual work.
2836 2844 */
2837 2845 if (zio->io_gang_tree)
2838 2846 return;
2839 2847
2840 2848 if (zio->io_error == 0) {
2841 2849 /*
2842 2850 * If this is a root i/o, don't count it -- we've already
2843 2851 * counted the top-level vdevs, and vdev_get_stats() will
2844 2852 * aggregate them when asked. This reduces contention on
2845 2853 * the root vdev_stat_lock and implicitly handles blocks
2846 2854 * that compress away to holes, for which there is no i/o.
2847 2855 * (Holes never create vdev children, so all the counters
2848 2856 * remain zero, which is what we want.)
2849 2857 *
2850 2858 * Note: this only applies to successful i/o (io_error == 0)
2851 2859 * because unlike i/o counts, errors are not additive.
2852 2860 * When reading a ditto block, for example, failure of
2853 2861 * one top-level vdev does not imply a root-level error.
2854 2862 */
2855 2863 if (vd == rvd)
2856 2864 return;
2857 2865
2858 2866 ASSERT(vd == zio->io_vd);
2859 2867
2860 2868 if (flags & ZIO_FLAG_IO_BYPASS)
2861 2869 return;
2862 2870
2863 2871 mutex_enter(&vd->vdev_stat_lock);
2864 2872
2865 2873 if (flags & ZIO_FLAG_IO_REPAIR) {
2866 2874 if (flags & ZIO_FLAG_SCAN_THREAD) {
2867 2875 dsl_scan_phys_t *scn_phys =
2868 2876 &spa->spa_dsl_pool->dp_scan->scn_phys;
2869 2877 uint64_t *processed = &scn_phys->scn_processed;
2870 2878
2871 2879 /* XXX cleanup? */
2872 2880 if (vd->vdev_ops->vdev_op_leaf)
2873 2881 atomic_add_64(processed, psize);
2874 2882 vs->vs_scan_processed += psize;
2875 2883 }
2876 2884
2877 2885 if (flags & ZIO_FLAG_SELF_HEAL)
2878 2886 vs->vs_self_healed += psize;
2879 2887 }
2880 2888
2881 2889 vs->vs_ops[type]++;
2882 2890 vs->vs_bytes[type] += psize;
2883 2891
2884 2892 mutex_exit(&vd->vdev_stat_lock);
2885 2893 return;
2886 2894 }
2887 2895
2888 2896 if (flags & ZIO_FLAG_SPECULATIVE)
2889 2897 return;
2890 2898
2891 2899 /*
2892 2900 * If this is an I/O error that is going to be retried, then ignore the
2893 2901 * error. Otherwise, the user may interpret B_FAILFAST I/O errors as
2894 2902 * hard errors, when in reality they can happen for any number of
2895 2903 * innocuous reasons (bus resets, MPxIO link failure, etc).
2896 2904 */
2897 2905 if (zio->io_error == EIO &&
2898 2906 !(zio->io_flags & ZIO_FLAG_IO_RETRY))
2899 2907 return;
2900 2908
2901 2909 /*
2902 2910 * Intent logs writes won't propagate their error to the root
2903 2911 * I/O so don't mark these types of failures as pool-level
2904 2912 * errors.
2905 2913 */
2906 2914 if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
2907 2915 return;
2908 2916
2909 2917 mutex_enter(&vd->vdev_stat_lock);
2910 2918 if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
2911 2919 if (zio->io_error == ECKSUM)
2912 2920 vs->vs_checksum_errors++;
2913 2921 else
2914 2922 vs->vs_read_errors++;
2915 2923 }
2916 2924 if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
2917 2925 vs->vs_write_errors++;
2918 2926 mutex_exit(&vd->vdev_stat_lock);
2919 2927
2920 2928 if (type == ZIO_TYPE_WRITE && txg != 0 &&
2921 2929 (!(flags & ZIO_FLAG_IO_REPAIR) ||
2922 2930 (flags & ZIO_FLAG_SCAN_THREAD) ||
2923 2931 spa->spa_claiming)) {
2924 2932 /*
2925 2933 * This is either a normal write (not a repair), or it's
2926 2934 * a repair induced by the scrub thread, or it's a repair
2927 2935 * made by zil_claim() during spa_load() in the first txg.
2928 2936 * In the normal case, we commit the DTL change in the same
2929 2937 * txg as the block was born. In the scrub-induced repair
2930 2938 * case, we know that scrubs run in first-pass syncing context,
2931 2939 * so we commit the DTL change in spa_syncing_txg(spa).
2932 2940 * In the zil_claim() case, we commit in spa_first_txg(spa).
2933 2941 *
2934 2942 * We currently do not make DTL entries for failed spontaneous
2935 2943 * self-healing writes triggered by normal (non-scrubbing)
2936 2944 * reads, because we have no transactional context in which to
2937 2945 * do so -- and it's not clear that it'd be desirable anyway.
2938 2946 */
2939 2947 if (vd->vdev_ops->vdev_op_leaf) {
2940 2948 uint64_t commit_txg = txg;
2941 2949 if (flags & ZIO_FLAG_SCAN_THREAD) {
2942 2950 ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2943 2951 ASSERT(spa_sync_pass(spa) == 1);
2944 2952 vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
2945 2953 commit_txg = spa_syncing_txg(spa);
2946 2954 } else if (spa->spa_claiming) {
2947 2955 ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2948 2956 commit_txg = spa_first_txg(spa);
2949 2957 }
2950 2958 ASSERT(commit_txg >= spa_syncing_txg(spa));
2951 2959 if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
2952 2960 return;
2953 2961 for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2954 2962 vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
2955 2963 vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
2956 2964 }
2957 2965 if (vd != rvd)
2958 2966 vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
2959 2967 }
2960 2968 }
2961 2969
2962 2970 /*
2963 2971 * Update the in-core space usage stats for this vdev, its metaslab class,
2964 2972 * and the root vdev.
2965 2973 */
2966 2974 void
2967 2975 vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
2968 2976 int64_t space_delta)
2969 2977 {
2970 2978 int64_t dspace_delta = space_delta;
2971 2979 spa_t *spa = vd->vdev_spa;
2972 2980 vdev_t *rvd = spa->spa_root_vdev;
2973 2981 metaslab_group_t *mg = vd->vdev_mg;
2974 2982 metaslab_class_t *mc = mg ? mg->mg_class : NULL;
2975 2983
2976 2984 ASSERT(vd == vd->vdev_top);
2977 2985
2978 2986 /*
2979 2987 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
2980 2988 * factor. We must calculate this here and not at the root vdev
2981 2989 * because the root vdev's psize-to-asize is simply the max of its
2982 2990 * childrens', thus not accurate enough for us.
2983 2991 */
2984 2992 ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
2985 2993 ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
2986 2994 dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
2987 2995 vd->vdev_deflate_ratio;
2988 2996
2989 2997 mutex_enter(&vd->vdev_stat_lock);
2990 2998 vd->vdev_stat.vs_alloc += alloc_delta;
2991 2999 vd->vdev_stat.vs_space += space_delta;
2992 3000 vd->vdev_stat.vs_dspace += dspace_delta;
2993 3001 mutex_exit(&vd->vdev_stat_lock);
2994 3002
2995 3003 if (mc == spa_normal_class(spa)) {
2996 3004 mutex_enter(&rvd->vdev_stat_lock);
2997 3005 rvd->vdev_stat.vs_alloc += alloc_delta;
2998 3006 rvd->vdev_stat.vs_space += space_delta;
2999 3007 rvd->vdev_stat.vs_dspace += dspace_delta;
3000 3008 mutex_exit(&rvd->vdev_stat_lock);
3001 3009 }
3002 3010
3003 3011 if (mc != NULL) {
3004 3012 ASSERT(rvd == vd->vdev_parent);
3005 3013 ASSERT(vd->vdev_ms_count != 0);
3006 3014
3007 3015 metaslab_class_space_update(mc,
3008 3016 alloc_delta, defer_delta, space_delta, dspace_delta);
3009 3017 }
3010 3018 }
3011 3019
3012 3020 /*
3013 3021 * Mark a top-level vdev's config as dirty, placing it on the dirty list
3014 3022 * so that it will be written out next time the vdev configuration is synced.
3015 3023 * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
3016 3024 */
3017 3025 void
3018 3026 vdev_config_dirty(vdev_t *vd)
3019 3027 {
3020 3028 spa_t *spa = vd->vdev_spa;
3021 3029 vdev_t *rvd = spa->spa_root_vdev;
3022 3030 int c;
3023 3031
3024 3032 ASSERT(spa_writeable(spa));
3025 3033
3026 3034 /*
3027 3035 * If this is an aux vdev (as with l2cache and spare devices), then we
3028 3036 * update the vdev config manually and set the sync flag.
3029 3037 */
3030 3038 if (vd->vdev_aux != NULL) {
3031 3039 spa_aux_vdev_t *sav = vd->vdev_aux;
3032 3040 nvlist_t **aux;
3033 3041 uint_t naux;
3034 3042
3035 3043 for (c = 0; c < sav->sav_count; c++) {
3036 3044 if (sav->sav_vdevs[c] == vd)
3037 3045 break;
3038 3046 }
3039 3047
3040 3048 if (c == sav->sav_count) {
3041 3049 /*
3042 3050 * We're being removed. There's nothing more to do.
3043 3051 */
3044 3052 ASSERT(sav->sav_sync == B_TRUE);
3045 3053 return;
3046 3054 }
3047 3055
3048 3056 sav->sav_sync = B_TRUE;
3049 3057
3050 3058 if (nvlist_lookup_nvlist_array(sav->sav_config,
3051 3059 ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
3052 3060 VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
3053 3061 ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
3054 3062 }
3055 3063
3056 3064 ASSERT(c < naux);
3057 3065
3058 3066 /*
3059 3067 * Setting the nvlist in the middle if the array is a little
3060 3068 * sketchy, but it will work.
3061 3069 */
3062 3070 nvlist_free(aux[c]);
3063 3071 aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0);
3064 3072
3065 3073 return;
3066 3074 }
3067 3075
3068 3076 /*
3069 3077 * The dirty list is protected by the SCL_CONFIG lock. The caller
3070 3078 * must either hold SCL_CONFIG as writer, or must be the sync thread
3071 3079 * (which holds SCL_CONFIG as reader). There's only one sync thread,
3072 3080 * so this is sufficient to ensure mutual exclusion.
3073 3081 */
3074 3082 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
3075 3083 (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3076 3084 spa_config_held(spa, SCL_CONFIG, RW_READER)));
3077 3085
3078 3086 if (vd == rvd) {
3079 3087 for (c = 0; c < rvd->vdev_children; c++)
3080 3088 vdev_config_dirty(rvd->vdev_child[c]);
3081 3089 } else {
3082 3090 ASSERT(vd == vd->vdev_top);
3083 3091
3084 3092 if (!list_link_active(&vd->vdev_config_dirty_node) &&
3085 3093 !vd->vdev_ishole)
3086 3094 list_insert_head(&spa->spa_config_dirty_list, vd);
3087 3095 }
3088 3096 }
3089 3097
3090 3098 void
3091 3099 vdev_config_clean(vdev_t *vd)
3092 3100 {
3093 3101 spa_t *spa = vd->vdev_spa;
3094 3102
3095 3103 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
3096 3104 (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3097 3105 spa_config_held(spa, SCL_CONFIG, RW_READER)));
3098 3106
3099 3107 ASSERT(list_link_active(&vd->vdev_config_dirty_node));
3100 3108 list_remove(&spa->spa_config_dirty_list, vd);
3101 3109 }
3102 3110
3103 3111 /*
3104 3112 * Mark a top-level vdev's state as dirty, so that the next pass of
3105 3113 * spa_sync() can convert this into vdev_config_dirty(). We distinguish
3106 3114 * the state changes from larger config changes because they require
3107 3115 * much less locking, and are often needed for administrative actions.
3108 3116 */
3109 3117 void
3110 3118 vdev_state_dirty(vdev_t *vd)
3111 3119 {
3112 3120 spa_t *spa = vd->vdev_spa;
3113 3121
3114 3122 ASSERT(spa_writeable(spa));
3115 3123 ASSERT(vd == vd->vdev_top);
3116 3124
3117 3125 /*
3118 3126 * The state list is protected by the SCL_STATE lock. The caller
3119 3127 * must either hold SCL_STATE as writer, or must be the sync thread
3120 3128 * (which holds SCL_STATE as reader). There's only one sync thread,
3121 3129 * so this is sufficient to ensure mutual exclusion.
3122 3130 */
3123 3131 ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
3124 3132 (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3125 3133 spa_config_held(spa, SCL_STATE, RW_READER)));
3126 3134
3127 3135 if (!list_link_active(&vd->vdev_state_dirty_node) && !vd->vdev_ishole)
3128 3136 list_insert_head(&spa->spa_state_dirty_list, vd);
3129 3137 }
3130 3138
3131 3139 void
3132 3140 vdev_state_clean(vdev_t *vd)
3133 3141 {
3134 3142 spa_t *spa = vd->vdev_spa;
3135 3143
3136 3144 ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
3137 3145 (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3138 3146 spa_config_held(spa, SCL_STATE, RW_READER)));
3139 3147
3140 3148 ASSERT(list_link_active(&vd->vdev_state_dirty_node));
3141 3149 list_remove(&spa->spa_state_dirty_list, vd);
3142 3150 }
3143 3151
3144 3152 /*
3145 3153 * Propagate vdev state up from children to parent.
3146 3154 */
3147 3155 void
3148 3156 vdev_propagate_state(vdev_t *vd)
3149 3157 {
3150 3158 spa_t *spa = vd->vdev_spa;
3151 3159 vdev_t *rvd = spa->spa_root_vdev;
3152 3160 int degraded = 0, faulted = 0;
3153 3161 int corrupted = 0;
3154 3162 vdev_t *child;
3155 3163
3156 3164 if (vd->vdev_children > 0) {
3157 3165 for (int c = 0; c < vd->vdev_children; c++) {
3158 3166 child = vd->vdev_child[c];
3159 3167
3160 3168 /*
3161 3169 * Don't factor holes into the decision.
3162 3170 */
3163 3171 if (child->vdev_ishole)
3164 3172 continue;
3165 3173
3166 3174 if (!vdev_readable(child) ||
3167 3175 (!vdev_writeable(child) && spa_writeable(spa))) {
3168 3176 /*
3169 3177 * Root special: if there is a top-level log
3170 3178 * device, treat the root vdev as if it were
3171 3179 * degraded.
3172 3180 */
3173 3181 if (child->vdev_islog && vd == rvd)
3174 3182 degraded++;
3175 3183 else
3176 3184 faulted++;
3177 3185 } else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
3178 3186 degraded++;
3179 3187 }
3180 3188
3181 3189 if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
3182 3190 corrupted++;
3183 3191 }
3184 3192
3185 3193 vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
3186 3194
3187 3195 /*
3188 3196 * Root special: if there is a top-level vdev that cannot be
3189 3197 * opened due to corrupted metadata, then propagate the root
3190 3198 * vdev's aux state as 'corrupt' rather than 'insufficient
3191 3199 * replicas'.
3192 3200 */
3193 3201 if (corrupted && vd == rvd &&
3194 3202 rvd->vdev_state == VDEV_STATE_CANT_OPEN)
3195 3203 vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
3196 3204 VDEV_AUX_CORRUPT_DATA);
3197 3205 }
3198 3206
3199 3207 if (vd->vdev_parent)
3200 3208 vdev_propagate_state(vd->vdev_parent);
3201 3209 }
3202 3210
3203 3211 /*
3204 3212 * Set a vdev's state. If this is during an open, we don't update the parent
3205 3213 * state, because we're in the process of opening children depth-first.
3206 3214 * Otherwise, we propagate the change to the parent.
3207 3215 *
3208 3216 * If this routine places a device in a faulted state, an appropriate ereport is
3209 3217 * generated.
3210 3218 */
3211 3219 void
3212 3220 vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
3213 3221 {
3214 3222 uint64_t save_state;
3215 3223 spa_t *spa = vd->vdev_spa;
3216 3224
3217 3225 if (state == vd->vdev_state) {
3218 3226 vd->vdev_stat.vs_aux = aux;
3219 3227 return;
3220 3228 }
3221 3229
3222 3230 save_state = vd->vdev_state;
3223 3231
3224 3232 vd->vdev_state = state;
3225 3233 vd->vdev_stat.vs_aux = aux;
3226 3234
3227 3235 /*
3228 3236 * If we are setting the vdev state to anything but an open state, then
3229 3237 * always close the underlying device unless the device has requested
3230 3238 * a delayed close (i.e. we're about to remove or fault the device).
3231 3239 * Otherwise, we keep accessible but invalid devices open forever.
3232 3240 * We don't call vdev_close() itself, because that implies some extra
3233 3241 * checks (offline, etc) that we don't want here. This is limited to
3234 3242 * leaf devices, because otherwise closing the device will affect other
3235 3243 * children.
3236 3244 */
3237 3245 if (!vd->vdev_delayed_close && vdev_is_dead(vd) &&
3238 3246 vd->vdev_ops->vdev_op_leaf)
3239 3247 vd->vdev_ops->vdev_op_close(vd);
3240 3248
3241 3249 /*
3242 3250 * If we have brought this vdev back into service, we need
3243 3251 * to notify fmd so that it can gracefully repair any outstanding
3244 3252 * cases due to a missing device. We do this in all cases, even those
3245 3253 * that probably don't correlate to a repaired fault. This is sure to
3246 3254 * catch all cases, and we let the zfs-retire agent sort it out. If
3247 3255 * this is a transient state it's OK, as the retire agent will
3248 3256 * double-check the state of the vdev before repairing it.
3249 3257 */
3250 3258 if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf &&
3251 3259 vd->vdev_prevstate != state)
3252 3260 zfs_post_state_change(spa, vd);
3253 3261
3254 3262 if (vd->vdev_removed &&
3255 3263 state == VDEV_STATE_CANT_OPEN &&
3256 3264 (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
3257 3265 /*
3258 3266 * If the previous state is set to VDEV_STATE_REMOVED, then this
3259 3267 * device was previously marked removed and someone attempted to
3260 3268 * reopen it. If this failed due to a nonexistent device, then
3261 3269 * keep the device in the REMOVED state. We also let this be if
3262 3270 * it is one of our special test online cases, which is only
3263 3271 * attempting to online the device and shouldn't generate an FMA
3264 3272 * fault.
3265 3273 */
3266 3274 vd->vdev_state = VDEV_STATE_REMOVED;
3267 3275 vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
3268 3276 } else if (state == VDEV_STATE_REMOVED) {
3269 3277 vd->vdev_removed = B_TRUE;
3270 3278 } else if (state == VDEV_STATE_CANT_OPEN) {
3271 3279 /*
3272 3280 * If we fail to open a vdev during an import or recovery, we
3273 3281 * mark it as "not available", which signifies that it was
3274 3282 * never there to begin with. Failure to open such a device
3275 3283 * is not considered an error.
3276 3284 */
3277 3285 if ((spa_load_state(spa) == SPA_LOAD_IMPORT ||
3278 3286 spa_load_state(spa) == SPA_LOAD_RECOVER) &&
3279 3287 vd->vdev_ops->vdev_op_leaf)
3280 3288 vd->vdev_not_present = 1;
3281 3289
3282 3290 /*
3283 3291 * Post the appropriate ereport. If the 'prevstate' field is
3284 3292 * set to something other than VDEV_STATE_UNKNOWN, it indicates
3285 3293 * that this is part of a vdev_reopen(). In this case, we don't
3286 3294 * want to post the ereport if the device was already in the
3287 3295 * CANT_OPEN state beforehand.
3288 3296 *
3289 3297 * If the 'checkremove' flag is set, then this is an attempt to
3290 3298 * online the device in response to an insertion event. If we
3291 3299 * hit this case, then we have detected an insertion event for a
3292 3300 * faulted or offline device that wasn't in the removed state.
3293 3301 * In this scenario, we don't post an ereport because we are
3294 3302 * about to replace the device, or attempt an online with
3295 3303 * vdev_forcefault, which will generate the fault for us.
3296 3304 */
3297 3305 if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
3298 3306 !vd->vdev_not_present && !vd->vdev_checkremove &&
3299 3307 vd != spa->spa_root_vdev) {
3300 3308 const char *class;
3301 3309
3302 3310 switch (aux) {
3303 3311 case VDEV_AUX_OPEN_FAILED:
3304 3312 class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
3305 3313 break;
3306 3314 case VDEV_AUX_CORRUPT_DATA:
3307 3315 class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
3308 3316 break;
3309 3317 case VDEV_AUX_NO_REPLICAS:
3310 3318 class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
3311 3319 break;
3312 3320 case VDEV_AUX_BAD_GUID_SUM:
3313 3321 class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
3314 3322 break;
3315 3323 case VDEV_AUX_TOO_SMALL:
3316 3324 class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
3317 3325 break;
3318 3326 case VDEV_AUX_BAD_LABEL:
3319 3327 class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
3320 3328 break;
3321 3329 default:
3322 3330 class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
3323 3331 }
3324 3332
3325 3333 zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
3326 3334 }
3327 3335
3328 3336 /* Erase any notion of persistent removed state */
3329 3337 vd->vdev_removed = B_FALSE;
3330 3338 } else {
3331 3339 vd->vdev_removed = B_FALSE;
3332 3340 }
3333 3341
3334 3342 if (!isopen && vd->vdev_parent)
3335 3343 vdev_propagate_state(vd->vdev_parent);
3336 3344 }
3337 3345
3338 3346 /*
3339 3347 * Check the vdev configuration to ensure that it's capable of supporting
3340 3348 * a root pool. We do not support partial configuration.
3341 3349 * In addition, only a single top-level vdev is allowed.
3342 3350 */
3343 3351 boolean_t
3344 3352 vdev_is_bootable(vdev_t *vd)
3345 3353 {
3346 3354 if (!vd->vdev_ops->vdev_op_leaf) {
3347 3355 char *vdev_type = vd->vdev_ops->vdev_op_type;
3348 3356
3349 3357 if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
3350 3358 vd->vdev_children > 1) {
3351 3359 return (B_FALSE);
3352 3360 } else if (strcmp(vdev_type, VDEV_TYPE_MISSING) == 0) {
3353 3361 return (B_FALSE);
3354 3362 }
3355 3363 }
3356 3364
3357 3365 for (int c = 0; c < vd->vdev_children; c++) {
3358 3366 if (!vdev_is_bootable(vd->vdev_child[c]))
3359 3367 return (B_FALSE);
3360 3368 }
3361 3369 return (B_TRUE);
3362 3370 }
3363 3371
3364 3372 /*
3365 3373 * Load the state from the original vdev tree (ovd) which
3366 3374 * we've retrieved from the MOS config object. If the original
3367 3375 * vdev was offline or faulted then we transfer that state to the
3368 3376 * device in the current vdev tree (nvd).
3369 3377 */
3370 3378 void
3371 3379 vdev_load_log_state(vdev_t *nvd, vdev_t *ovd)
3372 3380 {
3373 3381 spa_t *spa = nvd->vdev_spa;
3374 3382
3375 3383 ASSERT(nvd->vdev_top->vdev_islog);
3376 3384 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3377 3385 ASSERT3U(nvd->vdev_guid, ==, ovd->vdev_guid);
3378 3386
3379 3387 for (int c = 0; c < nvd->vdev_children; c++)
3380 3388 vdev_load_log_state(nvd->vdev_child[c], ovd->vdev_child[c]);
3381 3389
3382 3390 if (nvd->vdev_ops->vdev_op_leaf) {
3383 3391 /*
3384 3392 * Restore the persistent vdev state
3385 3393 */
3386 3394 nvd->vdev_offline = ovd->vdev_offline;
3387 3395 nvd->vdev_faulted = ovd->vdev_faulted;
3388 3396 nvd->vdev_degraded = ovd->vdev_degraded;
3389 3397 nvd->vdev_removed = ovd->vdev_removed;
3390 3398 }
3391 3399 }
3392 3400
3393 3401 /*
3394 3402 * Determine if a log device has valid content. If the vdev was
3395 3403 * removed or faulted in the MOS config then we know that
3396 3404 * the content on the log device has already been written to the pool.
3397 3405 */
3398 3406 boolean_t
3399 3407 vdev_log_state_valid(vdev_t *vd)
3400 3408 {
3401 3409 if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted &&
3402 3410 !vd->vdev_removed)
3403 3411 return (B_TRUE);
3404 3412
3405 3413 for (int c = 0; c < vd->vdev_children; c++)
3406 3414 if (vdev_log_state_valid(vd->vdev_child[c]))
3407 3415 return (B_TRUE);
3408 3416
3409 3417 return (B_FALSE);
3410 3418 }
3411 3419
3412 3420 /*
3413 3421 * Expand a vdev if possible.
3414 3422 */
3415 3423 void
3416 3424 vdev_expand(vdev_t *vd, uint64_t txg)
3417 3425 {
3418 3426 ASSERT(vd->vdev_top == vd);
3419 3427 ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
3420 3428
3421 3429 if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count) {
3422 3430 VERIFY(vdev_metaslab_init(vd, txg) == 0);
3423 3431 vdev_config_dirty(vd);
3424 3432 }
3425 3433 }
3426 3434
3427 3435 /*
3428 3436 * Split a vdev.
3429 3437 */
3430 3438 void
3431 3439 vdev_split(vdev_t *vd)
3432 3440 {
3433 3441 vdev_t *cvd, *pvd = vd->vdev_parent;
3434 3442
3435 3443 vdev_remove_child(pvd, vd);
3436 3444 vdev_compact_children(pvd);
3437 3445
3438 3446 cvd = pvd->vdev_child[0];
3439 3447 if (pvd->vdev_children == 1) {
3440 3448 vdev_remove_parent(cvd);
3441 3449 cvd->vdev_splitting = B_TRUE;
3442 3450 }
3443 3451 vdev_propagate_state(cvd);
3444 3452 }
3445 3453
3446 3454 void
3447 3455 vdev_deadman(vdev_t *vd)
3448 3456 {
3449 3457 for (int c = 0; c < vd->vdev_children; c++) {
3450 3458 vdev_t *cvd = vd->vdev_child[c];
3451 3459
3452 3460 vdev_deadman(cvd);
3453 3461 }
3454 3462
3455 3463 if (vd->vdev_ops->vdev_op_leaf) {
3456 3464 vdev_queue_t *vq = &vd->vdev_queue;
3457 3465
3458 3466 mutex_enter(&vq->vq_lock);
3459 3467 if (avl_numnodes(&vq->vq_active_tree) > 0) {
3460 3468 spa_t *spa = vd->vdev_spa;
3461 3469 zio_t *fio;
3462 3470 uint64_t delta;
3463 3471
3464 3472 /*
3465 3473 * Look at the head of all the pending queues,
3466 3474 * if any I/O has been outstanding for longer than
3467 3475 * the spa_deadman_synctime we panic the system.
3468 3476 */
3469 3477 fio = avl_first(&vq->vq_active_tree);
3470 3478 delta = gethrtime() - fio->io_timestamp;
3471 3479 if (delta > spa_deadman_synctime(spa)) {
3472 3480 zfs_dbgmsg("SLOW IO: zio timestamp %lluns, "
3473 3481 "delta %lluns, last io %lluns",
3474 3482 fio->io_timestamp, delta,
3475 3483 vq->vq_io_complete_ts);
3476 3484 fm_panic("I/O to pool '%s' appears to be "
3477 3485 "hung.", spa_name(spa));
3478 3486 }
3479 3487 }
3480 3488 mutex_exit(&vq->vq_lock);
3481 3489 }
3482 3490 }
|
↓ open down ↓ |
2322 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX