Print this page
10592 misc. metaslab and vdev related ZoL bug fixes
Portions contributed by: Jerry Jelinek <jerry.jelinek@joyent.com>
Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed by: Giuseppe Di Natale <guss80@gmail.com>
Reviewed by: George Melikov <mail@gmelikov.ru>
Reviewed by: Paul Dagnelie <pcd@delphix.com>
Reviewed by: Matt Ahrens <mahrens@delphix.com>
Reviewed by: Pavel Zakharov <pavel.zakharov@delphix.com>
Reviewed by: Tony Hutter <hutter2@llnl.gov>
Reviewed by: Kody Kantor <kody.kantor@joyent.com>
Approved by: Dan McDonald <danmcd@joyent.com>
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/fs/zfs/vdev_removal.c
+++ new/usr/src/uts/common/fs/zfs/vdev_removal.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, 2018 by Delphix. All rights reserved.
25 25 */
26 26
27 27 #include <sys/zfs_context.h>
28 28 #include <sys/spa_impl.h>
29 29 #include <sys/dmu.h>
30 30 #include <sys/dmu_tx.h>
31 31 #include <sys/zap.h>
32 32 #include <sys/vdev_impl.h>
33 33 #include <sys/metaslab.h>
34 34 #include <sys/metaslab_impl.h>
35 35 #include <sys/uberblock_impl.h>
36 36 #include <sys/txg.h>
37 37 #include <sys/avl.h>
38 38 #include <sys/bpobj.h>
39 39 #include <sys/dsl_pool.h>
40 40 #include <sys/dsl_synctask.h>
41 41 #include <sys/dsl_dir.h>
42 42 #include <sys/arc.h>
43 43 #include <sys/zfeature.h>
44 44 #include <sys/vdev_indirect_births.h>
45 45 #include <sys/vdev_indirect_mapping.h>
46 46 #include <sys/abd.h>
47 47 #include <sys/vdev_initialize.h>
48 48
49 49 /*
50 50 * This file contains the necessary logic to remove vdevs from a
51 51 * storage pool. Currently, the only devices that can be removed
52 52 * are log, cache, and spare devices; and top level vdevs from a pool
53 53 * w/o raidz. (Note that members of a mirror can also be removed
54 54 * by the detach operation.)
55 55 *
56 56 * Log vdevs are removed by evacuating them and then turning the vdev
57 57 * into a hole vdev while holding spa config locks.
58 58 *
59 59 * Top level vdevs are removed and converted into an indirect vdev via
60 60 * a multi-step process:
61 61 *
62 62 * - Disable allocations from this device (spa_vdev_remove_top).
63 63 *
64 64 * - From a new thread (spa_vdev_remove_thread), copy data from
65 65 * the removing vdev to a different vdev. The copy happens in open
66 66 * context (spa_vdev_copy_impl) and issues a sync task
67 67 * (vdev_mapping_sync) so the sync thread can update the partial
68 68 * indirect mappings in core and on disk.
69 69 *
70 70 * - If a free happens during a removal, it is freed from the
71 71 * removing vdev, and if it has already been copied, from the new
72 72 * location as well (free_from_removing_vdev).
73 73 *
74 74 * - After the removal is completed, the copy thread converts the vdev
75 75 * into an indirect vdev (vdev_remove_complete) before instructing
76 76 * the sync thread to destroy the space maps and finish the removal
77 77 * (spa_finish_removal).
78 78 */
79 79
80 80 typedef struct vdev_copy_arg {
81 81 metaslab_t *vca_msp;
82 82 uint64_t vca_outstanding_bytes;
83 83 kcondvar_t vca_cv;
84 84 kmutex_t vca_lock;
85 85 } vdev_copy_arg_t;
86 86
87 87 /*
88 88 * The maximum amount of memory we can use for outstanding i/o while
89 89 * doing a device removal. This determines how much i/o we can have
90 90 * in flight concurrently.
91 91 */
92 92 int zfs_remove_max_copy_bytes = 64 * 1024 * 1024;
93 93
94 94 /*
95 95 * The largest contiguous segment that we will attempt to allocate when
96 96 * removing a device. This can be no larger than SPA_MAXBLOCKSIZE. If
97 97 * there is a performance problem with attempting to allocate large blocks,
98 98 * consider decreasing this.
99 99 *
100 100 * Note: we will issue I/Os of up to this size. The mpt driver does not
101 101 * respond well to I/Os larger than 1MB, so we set this to 1MB. (When
102 102 * mpt processes an I/O larger than 1MB, it needs to do an allocation of
103 103 * 2 physically contiguous pages; if this allocation fails, mpt will drop
104 104 * the I/O and hang the device.)
105 105 */
106 106 int zfs_remove_max_segment = 1024 * 1024;
107 107
108 108 /*
109 109 * Allow a remap segment to span free chunks of at most this size. The main
110 110 * impact of a larger span is that we will read and write larger, more
111 111 * contiguous chunks, with more "unnecessary" data -- trading off bandwidth
112 112 * for iops. The value here was chosen to align with
113 113 * zfs_vdev_read_gap_limit, which is a similar concept when doing regular
114 114 * reads (but there's no reason it has to be the same).
115 115 *
116 116 * Additionally, a higher span will have the following relatively minor
117 117 * effects:
118 118 * - the mapping will be smaller, since one entry can cover more allocated
119 119 * segments
120 120 * - more of the fragmentation in the removing device will be preserved
121 121 * - we'll do larger allocations, which may fail and fall back on smaller
122 122 * allocations
123 123 */
124 124 int vdev_removal_max_span = 32 * 1024;
125 125
126 126 /*
127 127 * This is used by the test suite so that it can ensure that certain
128 128 * actions happen while in the middle of a removal.
129 129 */
130 130 uint64_t zfs_remove_max_bytes_pause = UINT64_MAX;
131 131
132 132 #define VDEV_REMOVAL_ZAP_OBJS "lzap"
133 133
134 134 static void spa_vdev_remove_thread(void *arg);
135 135
136 136 static void
137 137 spa_sync_removing_state(spa_t *spa, dmu_tx_t *tx)
138 138 {
139 139 VERIFY0(zap_update(spa->spa_dsl_pool->dp_meta_objset,
140 140 DMU_POOL_DIRECTORY_OBJECT,
141 141 DMU_POOL_REMOVING, sizeof (uint64_t),
142 142 sizeof (spa->spa_removing_phys) / sizeof (uint64_t),
143 143 &spa->spa_removing_phys, tx));
144 144 }
145 145
146 146 static nvlist_t *
147 147 spa_nvlist_lookup_by_guid(nvlist_t **nvpp, int count, uint64_t target_guid)
148 148 {
149 149 for (int i = 0; i < count; i++) {
150 150 uint64_t guid =
151 151 fnvlist_lookup_uint64(nvpp[i], ZPOOL_CONFIG_GUID);
152 152
153 153 if (guid == target_guid)
154 154 return (nvpp[i]);
155 155 }
156 156
157 157 return (NULL);
158 158 }
159 159
160 160 static void
161 161 spa_vdev_remove_aux(nvlist_t *config, char *name, nvlist_t **dev, int count,
162 162 nvlist_t *dev_to_remove)
163 163 {
164 164 nvlist_t **newdev = NULL;
165 165
166 166 if (count > 1)
167 167 newdev = kmem_alloc((count - 1) * sizeof (void *), KM_SLEEP);
168 168
169 169 for (int i = 0, j = 0; i < count; i++) {
170 170 if (dev[i] == dev_to_remove)
171 171 continue;
172 172 VERIFY(nvlist_dup(dev[i], &newdev[j++], KM_SLEEP) == 0);
173 173 }
174 174
175 175 VERIFY(nvlist_remove(config, name, DATA_TYPE_NVLIST_ARRAY) == 0);
176 176 VERIFY(nvlist_add_nvlist_array(config, name, newdev, count - 1) == 0);
177 177
178 178 for (int i = 0; i < count - 1; i++)
179 179 nvlist_free(newdev[i]);
180 180
181 181 if (count > 1)
182 182 kmem_free(newdev, (count - 1) * sizeof (void *));
183 183 }
184 184
185 185 static spa_vdev_removal_t *
186 186 spa_vdev_removal_create(vdev_t *vd)
187 187 {
188 188 spa_vdev_removal_t *svr = kmem_zalloc(sizeof (*svr), KM_SLEEP);
189 189 mutex_init(&svr->svr_lock, NULL, MUTEX_DEFAULT, NULL);
190 190 cv_init(&svr->svr_cv, NULL, CV_DEFAULT, NULL);
191 191 svr->svr_allocd_segs = range_tree_create(NULL, NULL);
192 192 svr->svr_vdev_id = vd->vdev_id;
193 193
194 194 for (int i = 0; i < TXG_SIZE; i++) {
195 195 svr->svr_frees[i] = range_tree_create(NULL, NULL);
196 196 list_create(&svr->svr_new_segments[i],
197 197 sizeof (vdev_indirect_mapping_entry_t),
198 198 offsetof(vdev_indirect_mapping_entry_t, vime_node));
199 199 }
200 200
201 201 return (svr);
202 202 }
203 203
204 204 void
205 205 spa_vdev_removal_destroy(spa_vdev_removal_t *svr)
206 206 {
207 207 for (int i = 0; i < TXG_SIZE; i++) {
208 208 ASSERT0(svr->svr_bytes_done[i]);
209 209 ASSERT0(svr->svr_max_offset_to_sync[i]);
210 210 range_tree_destroy(svr->svr_frees[i]);
211 211 list_destroy(&svr->svr_new_segments[i]);
212 212 }
213 213
214 214 range_tree_destroy(svr->svr_allocd_segs);
215 215 mutex_destroy(&svr->svr_lock);
216 216 cv_destroy(&svr->svr_cv);
217 217 kmem_free(svr, sizeof (*svr));
218 218 }
219 219
220 220 /*
221 221 * This is called as a synctask in the txg in which we will mark this vdev
222 222 * as removing (in the config stored in the MOS).
223 223 *
224 224 * It begins the evacuation of a toplevel vdev by:
225 225 * - initializing the spa_removing_phys which tracks this removal
226 226 * - computing the amount of space to remove for accounting purposes
227 227 * - dirtying all dbufs in the spa_config_object
228 228 * - creating the spa_vdev_removal
229 229 * - starting the spa_vdev_remove_thread
230 230 */
231 231 static void
232 232 vdev_remove_initiate_sync(void *arg, dmu_tx_t *tx)
233 233 {
234 234 int vdev_id = (uintptr_t)arg;
235 235 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
236 236 vdev_t *vd = vdev_lookup_top(spa, vdev_id);
237 237 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
238 238 objset_t *mos = spa->spa_dsl_pool->dp_meta_objset;
239 239 spa_vdev_removal_t *svr = NULL;
240 240 uint64_t txg = dmu_tx_get_txg(tx);
241 241
242 242 ASSERT3P(vd->vdev_ops, !=, &vdev_raidz_ops);
243 243 svr = spa_vdev_removal_create(vd);
244 244
245 245 ASSERT(vd->vdev_removing);
246 246 ASSERT3P(vd->vdev_indirect_mapping, ==, NULL);
247 247
248 248 spa_feature_incr(spa, SPA_FEATURE_DEVICE_REMOVAL, tx);
249 249 if (spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS)) {
250 250 /*
251 251 * By activating the OBSOLETE_COUNTS feature, we prevent
252 252 * the pool from being downgraded and ensure that the
253 253 * refcounts are precise.
254 254 */
255 255 spa_feature_incr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
256 256 uint64_t one = 1;
257 257 VERIFY0(zap_add(spa->spa_meta_objset, vd->vdev_top_zap,
258 258 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE, sizeof (one), 1,
259 259 &one, tx));
260 260 ASSERT3U(vdev_obsolete_counts_are_precise(vd), !=, 0);
261 261 }
262 262
263 263 vic->vic_mapping_object = vdev_indirect_mapping_alloc(mos, tx);
264 264 vd->vdev_indirect_mapping =
265 265 vdev_indirect_mapping_open(mos, vic->vic_mapping_object);
266 266 vic->vic_births_object = vdev_indirect_births_alloc(mos, tx);
267 267 vd->vdev_indirect_births =
268 268 vdev_indirect_births_open(mos, vic->vic_births_object);
269 269 spa->spa_removing_phys.sr_removing_vdev = vd->vdev_id;
270 270 spa->spa_removing_phys.sr_start_time = gethrestime_sec();
271 271 spa->spa_removing_phys.sr_end_time = 0;
272 272 spa->spa_removing_phys.sr_state = DSS_SCANNING;
273 273 spa->spa_removing_phys.sr_to_copy = 0;
274 274 spa->spa_removing_phys.sr_copied = 0;
275 275
|
↓ open down ↓ |
275 lines elided |
↑ open up ↑ |
276 276 /*
277 277 * Note: We can't use vdev_stat's vs_alloc for sr_to_copy, because
278 278 * there may be space in the defer tree, which is free, but still
279 279 * counted in vs_alloc.
280 280 */
281 281 for (uint64_t i = 0; i < vd->vdev_ms_count; i++) {
282 282 metaslab_t *ms = vd->vdev_ms[i];
283 283 if (ms->ms_sm == NULL)
284 284 continue;
285 285
286 - /*
287 - * Sync tasks happen before metaslab_sync(), therefore
288 - * smp_alloc and sm_alloc must be the same.
289 - */
290 - ASSERT3U(space_map_allocated(ms->ms_sm), ==,
291 - ms->ms_sm->sm_phys->smp_alloc);
292 -
293 286 spa->spa_removing_phys.sr_to_copy +=
294 - space_map_allocated(ms->ms_sm);
287 + metaslab_allocated_space(ms);
295 288
296 289 /*
297 290 * Space which we are freeing this txg does not need to
298 291 * be copied.
299 292 */
300 293 spa->spa_removing_phys.sr_to_copy -=
301 294 range_tree_space(ms->ms_freeing);
302 295
303 296 ASSERT0(range_tree_space(ms->ms_freed));
304 297 for (int t = 0; t < TXG_SIZE; t++)
305 298 ASSERT0(range_tree_space(ms->ms_allocating[t]));
306 299 }
307 300
308 301 /*
309 302 * Sync tasks are called before metaslab_sync(), so there should
310 303 * be no already-synced metaslabs in the TXG_CLEAN list.
311 304 */
312 305 ASSERT3P(txg_list_head(&vd->vdev_ms_list, TXG_CLEAN(txg)), ==, NULL);
313 306
314 307 spa_sync_removing_state(spa, tx);
315 308
316 309 /*
317 310 * All blocks that we need to read the most recent mapping must be
318 311 * stored on concrete vdevs. Therefore, we must dirty anything that
319 312 * is read before spa_remove_init(). Specifically, the
320 313 * spa_config_object. (Note that although we already modified the
321 314 * spa_config_object in spa_sync_removing_state, that may not have
322 315 * modified all blocks of the object.)
323 316 */
324 317 dmu_object_info_t doi;
325 318 VERIFY0(dmu_object_info(mos, DMU_POOL_DIRECTORY_OBJECT, &doi));
326 319 for (uint64_t offset = 0; offset < doi.doi_max_offset; ) {
327 320 dmu_buf_t *dbuf;
328 321 VERIFY0(dmu_buf_hold(mos, DMU_POOL_DIRECTORY_OBJECT,
329 322 offset, FTAG, &dbuf, 0));
330 323 dmu_buf_will_dirty(dbuf, tx);
331 324 offset += dbuf->db_size;
332 325 dmu_buf_rele(dbuf, FTAG);
333 326 }
334 327
335 328 /*
336 329 * Now that we've allocated the im_object, dirty the vdev to ensure
337 330 * that the object gets written to the config on disk.
338 331 */
339 332 vdev_config_dirty(vd);
340 333
341 334 zfs_dbgmsg("starting removal thread for vdev %llu (%p) in txg %llu "
342 335 "im_obj=%llu", vd->vdev_id, vd, dmu_tx_get_txg(tx),
343 336 vic->vic_mapping_object);
344 337
345 338 spa_history_log_internal(spa, "vdev remove started", tx,
346 339 "%s vdev %llu %s", spa_name(spa), vd->vdev_id,
347 340 (vd->vdev_path != NULL) ? vd->vdev_path : "-");
348 341 /*
349 342 * Setting spa_vdev_removal causes subsequent frees to call
350 343 * free_from_removing_vdev(). Note that we don't need any locking
351 344 * because we are the sync thread, and metaslab_free_impl() is only
352 345 * called from syncing context (potentially from a zio taskq thread,
353 346 * but in any case only when there are outstanding free i/os, which
354 347 * there are not).
355 348 */
356 349 ASSERT3P(spa->spa_vdev_removal, ==, NULL);
357 350 spa->spa_vdev_removal = svr;
358 351 svr->svr_thread = thread_create(NULL, 0,
359 352 spa_vdev_remove_thread, spa, 0, &p0, TS_RUN, minclsyspri);
360 353 }
361 354
362 355 /*
363 356 * When we are opening a pool, we must read the mapping for each
364 357 * indirect vdev in order from most recently removed to least
365 358 * recently removed. We do this because the blocks for the mapping
366 359 * of older indirect vdevs may be stored on more recently removed vdevs.
367 360 * In order to read each indirect mapping object, we must have
368 361 * initialized all more recently removed vdevs.
369 362 */
370 363 int
371 364 spa_remove_init(spa_t *spa)
372 365 {
373 366 int error;
374 367
375 368 error = zap_lookup(spa->spa_dsl_pool->dp_meta_objset,
376 369 DMU_POOL_DIRECTORY_OBJECT,
377 370 DMU_POOL_REMOVING, sizeof (uint64_t),
378 371 sizeof (spa->spa_removing_phys) / sizeof (uint64_t),
379 372 &spa->spa_removing_phys);
380 373
381 374 if (error == ENOENT) {
382 375 spa->spa_removing_phys.sr_state = DSS_NONE;
383 376 spa->spa_removing_phys.sr_removing_vdev = -1;
384 377 spa->spa_removing_phys.sr_prev_indirect_vdev = -1;
385 378 spa->spa_indirect_vdevs_loaded = B_TRUE;
386 379 return (0);
387 380 } else if (error != 0) {
388 381 return (error);
389 382 }
390 383
391 384 if (spa->spa_removing_phys.sr_state == DSS_SCANNING) {
392 385 /*
393 386 * We are currently removing a vdev. Create and
394 387 * initialize a spa_vdev_removal_t from the bonus
395 388 * buffer of the removing vdevs vdev_im_object, and
396 389 * initialize its partial mapping.
397 390 */
398 391 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
399 392 vdev_t *vd = vdev_lookup_top(spa,
400 393 spa->spa_removing_phys.sr_removing_vdev);
401 394
402 395 if (vd == NULL) {
403 396 spa_config_exit(spa, SCL_STATE, FTAG);
404 397 return (EINVAL);
405 398 }
406 399
407 400 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
408 401
409 402 ASSERT(vdev_is_concrete(vd));
410 403 spa_vdev_removal_t *svr = spa_vdev_removal_create(vd);
411 404 ASSERT3U(svr->svr_vdev_id, ==, vd->vdev_id);
412 405 ASSERT(vd->vdev_removing);
413 406
414 407 vd->vdev_indirect_mapping = vdev_indirect_mapping_open(
415 408 spa->spa_meta_objset, vic->vic_mapping_object);
416 409 vd->vdev_indirect_births = vdev_indirect_births_open(
417 410 spa->spa_meta_objset, vic->vic_births_object);
418 411 spa_config_exit(spa, SCL_STATE, FTAG);
419 412
420 413 spa->spa_vdev_removal = svr;
421 414 }
422 415
423 416 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
424 417 uint64_t indirect_vdev_id =
425 418 spa->spa_removing_phys.sr_prev_indirect_vdev;
426 419 while (indirect_vdev_id != UINT64_MAX) {
427 420 vdev_t *vd = vdev_lookup_top(spa, indirect_vdev_id);
428 421 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
429 422
430 423 ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
431 424 vd->vdev_indirect_mapping = vdev_indirect_mapping_open(
432 425 spa->spa_meta_objset, vic->vic_mapping_object);
433 426 vd->vdev_indirect_births = vdev_indirect_births_open(
434 427 spa->spa_meta_objset, vic->vic_births_object);
435 428
436 429 indirect_vdev_id = vic->vic_prev_indirect_vdev;
437 430 }
438 431 spa_config_exit(spa, SCL_STATE, FTAG);
439 432
440 433 /*
441 434 * Now that we've loaded all the indirect mappings, we can allow
442 435 * reads from other blocks (e.g. via predictive prefetch).
443 436 */
444 437 spa->spa_indirect_vdevs_loaded = B_TRUE;
445 438 return (0);
446 439 }
447 440
448 441 void
449 442 spa_restart_removal(spa_t *spa)
450 443 {
451 444 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
452 445
453 446 if (svr == NULL)
454 447 return;
455 448
456 449 /*
457 450 * In general when this function is called there is no
458 451 * removal thread running. The only scenario where this
459 452 * is not true is during spa_import() where this function
460 453 * is called twice [once from spa_import_impl() and
461 454 * spa_async_resume()]. Thus, in the scenario where we
462 455 * import a pool that has an ongoing removal we don't
463 456 * want to spawn a second thread.
464 457 */
465 458 if (svr->svr_thread != NULL)
466 459 return;
467 460
468 461 if (!spa_writeable(spa))
469 462 return;
470 463
471 464 zfs_dbgmsg("restarting removal of %llu", svr->svr_vdev_id);
472 465 svr->svr_thread = thread_create(NULL, 0, spa_vdev_remove_thread, spa,
473 466 0, &p0, TS_RUN, minclsyspri);
474 467 }
475 468
476 469 /*
477 470 * Process freeing from a device which is in the middle of being removed.
478 471 * We must handle this carefully so that we attempt to copy freed data,
479 472 * and we correctly free already-copied data.
480 473 */
481 474 void
482 475 free_from_removing_vdev(vdev_t *vd, uint64_t offset, uint64_t size)
483 476 {
484 477 spa_t *spa = vd->vdev_spa;
485 478 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
486 479 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
487 480 uint64_t txg = spa_syncing_txg(spa);
488 481 uint64_t max_offset_yet = 0;
489 482
490 483 ASSERT(vd->vdev_indirect_config.vic_mapping_object != 0);
491 484 ASSERT3U(vd->vdev_indirect_config.vic_mapping_object, ==,
492 485 vdev_indirect_mapping_object(vim));
493 486 ASSERT3U(vd->vdev_id, ==, svr->svr_vdev_id);
494 487
495 488 mutex_enter(&svr->svr_lock);
496 489
497 490 /*
498 491 * Remove the segment from the removing vdev's spacemap. This
499 492 * ensures that we will not attempt to copy this space (if the
500 493 * removal thread has not yet visited it), and also ensures
501 494 * that we know what is actually allocated on the new vdevs
502 495 * (needed if we cancel the removal).
503 496 *
504 497 * Note: we must do the metaslab_free_concrete() with the svr_lock
505 498 * held, so that the remove_thread can not load this metaslab and then
506 499 * visit this offset between the time that we metaslab_free_concrete()
507 500 * and when we check to see if it has been visited.
508 501 *
509 502 * Note: The checkpoint flag is set to false as having/taking
510 503 * a checkpoint and removing a device can't happen at the same
511 504 * time.
512 505 */
513 506 ASSERT(!spa_has_checkpoint(spa));
514 507 metaslab_free_concrete(vd, offset, size, B_FALSE);
515 508
516 509 uint64_t synced_size = 0;
517 510 uint64_t synced_offset = 0;
518 511 uint64_t max_offset_synced = vdev_indirect_mapping_max_offset(vim);
519 512 if (offset < max_offset_synced) {
520 513 /*
521 514 * The mapping for this offset is already on disk.
522 515 * Free from the new location.
523 516 *
524 517 * Note that we use svr_max_synced_offset because it is
525 518 * updated atomically with respect to the in-core mapping.
526 519 * By contrast, vim_max_offset is not.
527 520 *
528 521 * This block may be split between a synced entry and an
529 522 * in-flight or unvisited entry. Only process the synced
530 523 * portion of it here.
531 524 */
532 525 synced_size = MIN(size, max_offset_synced - offset);
533 526 synced_offset = offset;
534 527
535 528 ASSERT3U(max_offset_yet, <=, max_offset_synced);
536 529 max_offset_yet = max_offset_synced;
537 530
538 531 DTRACE_PROBE3(remove__free__synced,
539 532 spa_t *, spa,
540 533 uint64_t, offset,
541 534 uint64_t, synced_size);
542 535
543 536 size -= synced_size;
544 537 offset += synced_size;
545 538 }
546 539
547 540 /*
548 541 * Look at all in-flight txgs starting from the currently syncing one
549 542 * and see if a section of this free is being copied. By starting from
550 543 * this txg and iterating forward, we might find that this region
551 544 * was copied in two different txgs and handle it appropriately.
552 545 */
553 546 for (int i = 0; i < TXG_CONCURRENT_STATES; i++) {
554 547 int txgoff = (txg + i) & TXG_MASK;
555 548 if (size > 0 && offset < svr->svr_max_offset_to_sync[txgoff]) {
556 549 /*
557 550 * The mapping for this offset is in flight, and
558 551 * will be synced in txg+i.
559 552 */
560 553 uint64_t inflight_size = MIN(size,
561 554 svr->svr_max_offset_to_sync[txgoff] - offset);
562 555
563 556 DTRACE_PROBE4(remove__free__inflight,
564 557 spa_t *, spa,
565 558 uint64_t, offset,
566 559 uint64_t, inflight_size,
567 560 uint64_t, txg + i);
568 561
569 562 /*
570 563 * We copy data in order of increasing offset.
571 564 * Therefore the max_offset_to_sync[] must increase
572 565 * (or be zero, indicating that nothing is being
573 566 * copied in that txg).
574 567 */
575 568 if (svr->svr_max_offset_to_sync[txgoff] != 0) {
576 569 ASSERT3U(svr->svr_max_offset_to_sync[txgoff],
577 570 >=, max_offset_yet);
578 571 max_offset_yet =
579 572 svr->svr_max_offset_to_sync[txgoff];
580 573 }
581 574
582 575 /*
583 576 * We've already committed to copying this segment:
584 577 * we have allocated space elsewhere in the pool for
585 578 * it and have an IO outstanding to copy the data. We
586 579 * cannot free the space before the copy has
587 580 * completed, or else the copy IO might overwrite any
588 581 * new data. To free that space, we record the
589 582 * segment in the appropriate svr_frees tree and free
590 583 * the mapped space later, in the txg where we have
591 584 * completed the copy and synced the mapping (see
592 585 * vdev_mapping_sync).
593 586 */
594 587 range_tree_add(svr->svr_frees[txgoff],
595 588 offset, inflight_size);
596 589 size -= inflight_size;
597 590 offset += inflight_size;
598 591
599 592 /*
600 593 * This space is already accounted for as being
601 594 * done, because it is being copied in txg+i.
602 595 * However, if i!=0, then it is being copied in
603 596 * a future txg. If we crash after this txg
604 597 * syncs but before txg+i syncs, then the space
605 598 * will be free. Therefore we must account
606 599 * for the space being done in *this* txg
607 600 * (when it is freed) rather than the future txg
608 601 * (when it will be copied).
609 602 */
610 603 ASSERT3U(svr->svr_bytes_done[txgoff], >=,
611 604 inflight_size);
612 605 svr->svr_bytes_done[txgoff] -= inflight_size;
613 606 svr->svr_bytes_done[txg & TXG_MASK] += inflight_size;
614 607 }
615 608 }
616 609 ASSERT0(svr->svr_max_offset_to_sync[TXG_CLEAN(txg) & TXG_MASK]);
617 610
618 611 if (size > 0) {
619 612 /*
620 613 * The copy thread has not yet visited this offset. Ensure
621 614 * that it doesn't.
622 615 */
623 616
624 617 DTRACE_PROBE3(remove__free__unvisited,
625 618 spa_t *, spa,
626 619 uint64_t, offset,
627 620 uint64_t, size);
628 621
629 622 if (svr->svr_allocd_segs != NULL)
630 623 range_tree_clear(svr->svr_allocd_segs, offset, size);
631 624
632 625 /*
633 626 * Since we now do not need to copy this data, for
634 627 * accounting purposes we have done our job and can count
635 628 * it as completed.
636 629 */
637 630 svr->svr_bytes_done[txg & TXG_MASK] += size;
638 631 }
639 632 mutex_exit(&svr->svr_lock);
640 633
641 634 /*
642 635 * Now that we have dropped svr_lock, process the synced portion
643 636 * of this free.
644 637 */
645 638 if (synced_size > 0) {
646 639 vdev_indirect_mark_obsolete(vd, synced_offset, synced_size);
647 640
648 641 /*
649 642 * Note: this can only be called from syncing context,
650 643 * and the vdev_indirect_mapping is only changed from the
651 644 * sync thread, so we don't need svr_lock while doing
652 645 * metaslab_free_impl_cb.
653 646 */
654 647 boolean_t checkpoint = B_FALSE;
655 648 vdev_indirect_ops.vdev_op_remap(vd, synced_offset, synced_size,
656 649 metaslab_free_impl_cb, &checkpoint);
657 650 }
658 651 }
659 652
660 653 /*
661 654 * Stop an active removal and update the spa_removing phys.
662 655 */
663 656 static void
664 657 spa_finish_removal(spa_t *spa, dsl_scan_state_t state, dmu_tx_t *tx)
665 658 {
666 659 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
667 660 ASSERT3U(dmu_tx_get_txg(tx), ==, spa_syncing_txg(spa));
668 661
669 662 /* Ensure the removal thread has completed before we free the svr. */
670 663 spa_vdev_remove_suspend(spa);
671 664
672 665 ASSERT(state == DSS_FINISHED || state == DSS_CANCELED);
673 666
674 667 if (state == DSS_FINISHED) {
675 668 spa_removing_phys_t *srp = &spa->spa_removing_phys;
676 669 vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
677 670 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
678 671
679 672 if (srp->sr_prev_indirect_vdev != UINT64_MAX) {
680 673 vdev_t *pvd = vdev_lookup_top(spa,
681 674 srp->sr_prev_indirect_vdev);
682 675 ASSERT3P(pvd->vdev_ops, ==, &vdev_indirect_ops);
683 676 }
684 677
685 678 vic->vic_prev_indirect_vdev = srp->sr_prev_indirect_vdev;
686 679 srp->sr_prev_indirect_vdev = vd->vdev_id;
687 680 }
688 681 spa->spa_removing_phys.sr_state = state;
689 682 spa->spa_removing_phys.sr_end_time = gethrestime_sec();
690 683
691 684 spa->spa_vdev_removal = NULL;
692 685 spa_vdev_removal_destroy(svr);
693 686
694 687 spa_sync_removing_state(spa, tx);
695 688
696 689 vdev_config_dirty(spa->spa_root_vdev);
697 690 }
698 691
699 692 static void
700 693 free_mapped_segment_cb(void *arg, uint64_t offset, uint64_t size)
701 694 {
702 695 vdev_t *vd = arg;
703 696 vdev_indirect_mark_obsolete(vd, offset, size);
704 697 boolean_t checkpoint = B_FALSE;
705 698 vdev_indirect_ops.vdev_op_remap(vd, offset, size,
706 699 metaslab_free_impl_cb, &checkpoint);
707 700 }
708 701
709 702 /*
710 703 * On behalf of the removal thread, syncs an incremental bit more of
711 704 * the indirect mapping to disk and updates the in-memory mapping.
712 705 * Called as a sync task in every txg that the removal thread makes progress.
713 706 */
714 707 static void
715 708 vdev_mapping_sync(void *arg, dmu_tx_t *tx)
716 709 {
717 710 spa_vdev_removal_t *svr = arg;
718 711 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
719 712 vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
720 713 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
721 714 uint64_t txg = dmu_tx_get_txg(tx);
722 715 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
723 716
724 717 ASSERT(vic->vic_mapping_object != 0);
725 718 ASSERT3U(txg, ==, spa_syncing_txg(spa));
726 719
727 720 vdev_indirect_mapping_add_entries(vim,
728 721 &svr->svr_new_segments[txg & TXG_MASK], tx);
729 722 vdev_indirect_births_add_entry(vd->vdev_indirect_births,
730 723 vdev_indirect_mapping_max_offset(vim), dmu_tx_get_txg(tx), tx);
731 724
732 725 /*
733 726 * Free the copied data for anything that was freed while the
734 727 * mapping entries were in flight.
735 728 */
736 729 mutex_enter(&svr->svr_lock);
737 730 range_tree_vacate(svr->svr_frees[txg & TXG_MASK],
738 731 free_mapped_segment_cb, vd);
739 732 ASSERT3U(svr->svr_max_offset_to_sync[txg & TXG_MASK], >=,
740 733 vdev_indirect_mapping_max_offset(vim));
741 734 svr->svr_max_offset_to_sync[txg & TXG_MASK] = 0;
742 735 mutex_exit(&svr->svr_lock);
743 736
744 737 spa_sync_removing_state(spa, tx);
745 738 }
746 739
747 740 typedef struct vdev_copy_segment_arg {
748 741 spa_t *vcsa_spa;
749 742 dva_t *vcsa_dest_dva;
750 743 uint64_t vcsa_txg;
751 744 range_tree_t *vcsa_obsolete_segs;
752 745 } vdev_copy_segment_arg_t;
753 746
754 747 static void
755 748 unalloc_seg(void *arg, uint64_t start, uint64_t size)
756 749 {
757 750 vdev_copy_segment_arg_t *vcsa = arg;
758 751 spa_t *spa = vcsa->vcsa_spa;
759 752 blkptr_t bp = { 0 };
760 753
761 754 BP_SET_BIRTH(&bp, TXG_INITIAL, TXG_INITIAL);
762 755 BP_SET_LSIZE(&bp, size);
763 756 BP_SET_PSIZE(&bp, size);
764 757 BP_SET_COMPRESS(&bp, ZIO_COMPRESS_OFF);
765 758 BP_SET_CHECKSUM(&bp, ZIO_CHECKSUM_OFF);
766 759 BP_SET_TYPE(&bp, DMU_OT_NONE);
767 760 BP_SET_LEVEL(&bp, 0);
768 761 BP_SET_DEDUP(&bp, 0);
769 762 BP_SET_BYTEORDER(&bp, ZFS_HOST_BYTEORDER);
770 763
771 764 DVA_SET_VDEV(&bp.blk_dva[0], DVA_GET_VDEV(vcsa->vcsa_dest_dva));
772 765 DVA_SET_OFFSET(&bp.blk_dva[0],
773 766 DVA_GET_OFFSET(vcsa->vcsa_dest_dva) + start);
774 767 DVA_SET_ASIZE(&bp.blk_dva[0], size);
775 768
776 769 zio_free(spa, vcsa->vcsa_txg, &bp);
777 770 }
778 771
779 772 /*
780 773 * All reads and writes associated with a call to spa_vdev_copy_segment()
781 774 * are done.
782 775 */
783 776 static void
784 777 spa_vdev_copy_segment_done(zio_t *zio)
785 778 {
786 779 vdev_copy_segment_arg_t *vcsa = zio->io_private;
787 780
788 781 range_tree_vacate(vcsa->vcsa_obsolete_segs,
789 782 unalloc_seg, vcsa);
790 783 range_tree_destroy(vcsa->vcsa_obsolete_segs);
791 784 kmem_free(vcsa, sizeof (*vcsa));
792 785
793 786 spa_config_exit(zio->io_spa, SCL_STATE, zio->io_spa);
794 787 }
795 788
796 789 /*
797 790 * The write of the new location is done.
798 791 */
799 792 static void
800 793 spa_vdev_copy_segment_write_done(zio_t *zio)
801 794 {
802 795 vdev_copy_arg_t *vca = zio->io_private;
803 796
804 797 abd_free(zio->io_abd);
805 798
806 799 mutex_enter(&vca->vca_lock);
807 800 vca->vca_outstanding_bytes -= zio->io_size;
808 801 cv_signal(&vca->vca_cv);
809 802 mutex_exit(&vca->vca_lock);
810 803 }
811 804
812 805 /*
813 806 * The read of the old location is done. The parent zio is the write to
814 807 * the new location. Allow it to start.
815 808 */
816 809 static void
817 810 spa_vdev_copy_segment_read_done(zio_t *zio)
818 811 {
819 812 zio_nowait(zio_unique_parent(zio));
820 813 }
821 814
822 815 /*
823 816 * If the old and new vdevs are mirrors, we will read both sides of the old
824 817 * mirror, and write each copy to the corresponding side of the new mirror.
825 818 * If the old and new vdevs have a different number of children, we will do
826 819 * this as best as possible. Since we aren't verifying checksums, this
827 820 * ensures that as long as there's a good copy of the data, we'll have a
828 821 * good copy after the removal, even if there's silent damage to one side
829 822 * of the mirror. If we're removing a mirror that has some silent damage,
830 823 * we'll have exactly the same damage in the new location (assuming that
831 824 * the new location is also a mirror).
832 825 *
833 826 * We accomplish this by creating a tree of zio_t's, with as many writes as
834 827 * there are "children" of the new vdev (a non-redundant vdev counts as one
835 828 * child, a 2-way mirror has 2 children, etc). Each write has an associated
836 829 * read from a child of the old vdev. Typically there will be the same
837 830 * number of children of the old and new vdevs. However, if there are more
838 831 * children of the new vdev, some child(ren) of the old vdev will be issued
839 832 * multiple reads. If there are more children of the old vdev, some copies
840 833 * will be dropped.
841 834 *
842 835 * For example, the tree of zio_t's for a 2-way mirror is:
843 836 *
844 837 * null
845 838 * / \
846 839 * write(new vdev, child 0) write(new vdev, child 1)
847 840 * | |
848 841 * read(old vdev, child 0) read(old vdev, child 1)
849 842 *
850 843 * Child zio's complete before their parents complete. However, zio's
851 844 * created with zio_vdev_child_io() may be issued before their children
852 845 * complete. In this case we need to make sure that the children (reads)
853 846 * complete before the parents (writes) are *issued*. We do this by not
854 847 * calling zio_nowait() on each write until its corresponding read has
855 848 * completed.
856 849 *
857 850 * The spa_config_lock must be held while zio's created by
858 851 * zio_vdev_child_io() are in progress, to ensure that the vdev tree does
859 852 * not change (e.g. due to a concurrent "zpool attach/detach"). The "null"
860 853 * zio is needed to release the spa_config_lock after all the reads and
861 854 * writes complete. (Note that we can't grab the config lock for each read,
862 855 * because it is not reentrant - we could deadlock with a thread waiting
863 856 * for a write lock.)
864 857 */
865 858 static void
866 859 spa_vdev_copy_one_child(vdev_copy_arg_t *vca, zio_t *nzio,
867 860 vdev_t *source_vd, uint64_t source_offset,
868 861 vdev_t *dest_child_vd, uint64_t dest_offset, int dest_id, uint64_t size)
869 862 {
870 863 ASSERT3U(spa_config_held(nzio->io_spa, SCL_ALL, RW_READER), !=, 0);
871 864
872 865 mutex_enter(&vca->vca_lock);
873 866 vca->vca_outstanding_bytes += size;
874 867 mutex_exit(&vca->vca_lock);
875 868
876 869 abd_t *abd = abd_alloc_for_io(size, B_FALSE);
877 870
878 871 vdev_t *source_child_vd;
879 872 if (source_vd->vdev_ops == &vdev_mirror_ops && dest_id != -1) {
880 873 /*
881 874 * Source and dest are both mirrors. Copy from the same
882 875 * child id as we are copying to (wrapping around if there
883 876 * are more dest children than source children).
884 877 */
885 878 source_child_vd =
886 879 source_vd->vdev_child[dest_id % source_vd->vdev_children];
887 880 } else {
888 881 source_child_vd = source_vd;
889 882 }
890 883
891 884 zio_t *write_zio = zio_vdev_child_io(nzio, NULL,
892 885 dest_child_vd, dest_offset, abd, size,
893 886 ZIO_TYPE_WRITE, ZIO_PRIORITY_REMOVAL,
894 887 ZIO_FLAG_CANFAIL,
895 888 spa_vdev_copy_segment_write_done, vca);
896 889
897 890 zio_nowait(zio_vdev_child_io(write_zio, NULL,
898 891 source_child_vd, source_offset, abd, size,
899 892 ZIO_TYPE_READ, ZIO_PRIORITY_REMOVAL,
900 893 ZIO_FLAG_CANFAIL,
901 894 spa_vdev_copy_segment_read_done, vca));
902 895 }
903 896
904 897 /*
905 898 * Allocate a new location for this segment, and create the zio_t's to
906 899 * read from the old location and write to the new location.
907 900 */
908 901 static int
909 902 spa_vdev_copy_segment(vdev_t *vd, range_tree_t *segs,
910 903 uint64_t maxalloc, uint64_t txg,
911 904 vdev_copy_arg_t *vca, zio_alloc_list_t *zal)
912 905 {
913 906 metaslab_group_t *mg = vd->vdev_mg;
914 907 spa_t *spa = vd->vdev_spa;
915 908 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
916 909 vdev_indirect_mapping_entry_t *entry;
917 910 dva_t dst = { 0 };
918 911 uint64_t start = range_tree_min(segs);
919 912
920 913 ASSERT3U(maxalloc, <=, SPA_MAXBLOCKSIZE);
921 914
922 915 uint64_t size = range_tree_span(segs);
923 916 if (range_tree_span(segs) > maxalloc) {
924 917 /*
925 918 * We can't allocate all the segments. Prefer to end
926 919 * the allocation at the end of a segment, thus avoiding
927 920 * additional split blocks.
928 921 */
929 922 range_seg_t search;
930 923 avl_index_t where;
931 924 search.rs_start = start + maxalloc;
932 925 search.rs_end = search.rs_start;
933 926 range_seg_t *rs = avl_find(&segs->rt_root, &search, &where);
934 927 if (rs == NULL) {
935 928 rs = avl_nearest(&segs->rt_root, where, AVL_BEFORE);
936 929 } else {
937 930 rs = AVL_PREV(&segs->rt_root, rs);
938 931 }
939 932 if (rs != NULL) {
940 933 size = rs->rs_end - start;
941 934 } else {
942 935 /*
943 936 * There are no segments that end before maxalloc.
944 937 * I.e. the first segment is larger than maxalloc,
945 938 * so we must split it.
946 939 */
947 940 size = maxalloc;
948 941 }
949 942 }
950 943 ASSERT3U(size, <=, maxalloc);
951 944
952 945 /*
953 946 * An allocation class might not have any remaining vdevs or space
954 947 */
955 948 metaslab_class_t *mc = mg->mg_class;
956 949 if (mc != spa_normal_class(spa) && mc->mc_groups <= 1)
957 950 mc = spa_normal_class(spa);
958 951 int error = metaslab_alloc_dva(spa, mc, size, &dst, 0, NULL, txg, 0,
959 952 zal, 0);
960 953 if (error == ENOSPC && mc != spa_normal_class(spa)) {
961 954 error = metaslab_alloc_dva(spa, spa_normal_class(spa), size,
962 955 &dst, 0, NULL, txg, 0, zal, 0);
963 956 }
964 957 if (error != 0)
965 958 return (error);
966 959
967 960 /*
968 961 * Determine the ranges that are not actually needed. Offsets are
969 962 * relative to the start of the range to be copied (i.e. relative to the
970 963 * local variable "start").
971 964 */
972 965 range_tree_t *obsolete_segs = range_tree_create(NULL, NULL);
973 966
974 967 range_seg_t *rs = avl_first(&segs->rt_root);
975 968 ASSERT3U(rs->rs_start, ==, start);
976 969 uint64_t prev_seg_end = rs->rs_end;
977 970 while ((rs = AVL_NEXT(&segs->rt_root, rs)) != NULL) {
978 971 if (rs->rs_start >= start + size) {
979 972 break;
980 973 } else {
981 974 range_tree_add(obsolete_segs,
982 975 prev_seg_end - start,
983 976 rs->rs_start - prev_seg_end);
984 977 }
985 978 prev_seg_end = rs->rs_end;
986 979 }
987 980 /* We don't end in the middle of an obsolete range */
988 981 ASSERT3U(start + size, <=, prev_seg_end);
989 982
990 983 range_tree_clear(segs, start, size);
991 984
992 985 /*
993 986 * We can't have any padding of the allocated size, otherwise we will
994 987 * misunderstand what's allocated, and the size of the mapping.
995 988 * The caller ensures this will be true by passing in a size that is
996 989 * aligned to the worst (highest) ashift in the pool.
997 990 */
998 991 ASSERT3U(DVA_GET_ASIZE(&dst), ==, size);
999 992
1000 993 entry = kmem_zalloc(sizeof (vdev_indirect_mapping_entry_t), KM_SLEEP);
1001 994 DVA_MAPPING_SET_SRC_OFFSET(&entry->vime_mapping, start);
1002 995 entry->vime_mapping.vimep_dst = dst;
1003 996 if (spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS)) {
1004 997 entry->vime_obsolete_count = range_tree_space(obsolete_segs);
1005 998 }
1006 999
1007 1000 vdev_copy_segment_arg_t *vcsa = kmem_zalloc(sizeof (*vcsa), KM_SLEEP);
1008 1001 vcsa->vcsa_dest_dva = &entry->vime_mapping.vimep_dst;
1009 1002 vcsa->vcsa_obsolete_segs = obsolete_segs;
1010 1003 vcsa->vcsa_spa = spa;
1011 1004 vcsa->vcsa_txg = txg;
1012 1005
1013 1006 /*
1014 1007 * See comment before spa_vdev_copy_one_child().
1015 1008 */
1016 1009 spa_config_enter(spa, SCL_STATE, spa, RW_READER);
1017 1010 zio_t *nzio = zio_null(spa->spa_txg_zio[txg & TXG_MASK], spa, NULL,
1018 1011 spa_vdev_copy_segment_done, vcsa, 0);
1019 1012 vdev_t *dest_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dst));
1020 1013 if (dest_vd->vdev_ops == &vdev_mirror_ops) {
1021 1014 for (int i = 0; i < dest_vd->vdev_children; i++) {
1022 1015 vdev_t *child = dest_vd->vdev_child[i];
1023 1016 spa_vdev_copy_one_child(vca, nzio, vd, start,
1024 1017 child, DVA_GET_OFFSET(&dst), i, size);
1025 1018 }
1026 1019 } else {
1027 1020 spa_vdev_copy_one_child(vca, nzio, vd, start,
1028 1021 dest_vd, DVA_GET_OFFSET(&dst), -1, size);
1029 1022 }
1030 1023 zio_nowait(nzio);
1031 1024
1032 1025 list_insert_tail(&svr->svr_new_segments[txg & TXG_MASK], entry);
1033 1026 ASSERT3U(start + size, <=, vd->vdev_ms_count << vd->vdev_ms_shift);
1034 1027 vdev_dirty(vd, 0, NULL, txg);
1035 1028
1036 1029 return (0);
1037 1030 }
1038 1031
1039 1032 /*
1040 1033 * Complete the removal of a toplevel vdev. This is called as a
1041 1034 * synctask in the same txg that we will sync out the new config (to the
1042 1035 * MOS object) which indicates that this vdev is indirect.
1043 1036 */
1044 1037 static void
1045 1038 vdev_remove_complete_sync(void *arg, dmu_tx_t *tx)
1046 1039 {
1047 1040 spa_vdev_removal_t *svr = arg;
1048 1041 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
1049 1042 vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
1050 1043
1051 1044 ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
1052 1045
1053 1046 for (int i = 0; i < TXG_SIZE; i++) {
1054 1047 ASSERT0(svr->svr_bytes_done[i]);
1055 1048 }
1056 1049
1057 1050 ASSERT3U(spa->spa_removing_phys.sr_copied, ==,
1058 1051 spa->spa_removing_phys.sr_to_copy);
1059 1052
1060 1053 vdev_destroy_spacemaps(vd, tx);
1061 1054
1062 1055 /* destroy leaf zaps, if any */
1063 1056 ASSERT3P(svr->svr_zaplist, !=, NULL);
1064 1057 for (nvpair_t *pair = nvlist_next_nvpair(svr->svr_zaplist, NULL);
1065 1058 pair != NULL;
1066 1059 pair = nvlist_next_nvpair(svr->svr_zaplist, pair)) {
1067 1060 vdev_destroy_unlink_zap(vd, fnvpair_value_uint64(pair), tx);
1068 1061 }
1069 1062 fnvlist_free(svr->svr_zaplist);
1070 1063
1071 1064 spa_finish_removal(dmu_tx_pool(tx)->dp_spa, DSS_FINISHED, tx);
1072 1065 /* vd->vdev_path is not available here */
1073 1066 spa_history_log_internal(spa, "vdev remove completed", tx,
1074 1067 "%s vdev %llu", spa_name(spa), vd->vdev_id);
1075 1068 }
1076 1069
1077 1070 static void
1078 1071 vdev_remove_enlist_zaps(vdev_t *vd, nvlist_t *zlist)
1079 1072 {
1080 1073 ASSERT3P(zlist, !=, NULL);
1081 1074 ASSERT3P(vd->vdev_ops, !=, &vdev_raidz_ops);
1082 1075
1083 1076 if (vd->vdev_leaf_zap != 0) {
1084 1077 char zkey[32];
1085 1078 (void) snprintf(zkey, sizeof (zkey), "%s-%"PRIu64,
1086 1079 VDEV_REMOVAL_ZAP_OBJS, vd->vdev_leaf_zap);
1087 1080 fnvlist_add_uint64(zlist, zkey, vd->vdev_leaf_zap);
1088 1081 }
1089 1082
1090 1083 for (uint64_t id = 0; id < vd->vdev_children; id++) {
1091 1084 vdev_remove_enlist_zaps(vd->vdev_child[id], zlist);
1092 1085 }
1093 1086 }
1094 1087
1095 1088 static void
1096 1089 vdev_remove_replace_with_indirect(vdev_t *vd, uint64_t txg)
1097 1090 {
1098 1091 vdev_t *ivd;
1099 1092 dmu_tx_t *tx;
1100 1093 spa_t *spa = vd->vdev_spa;
1101 1094 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1102 1095
1103 1096 /*
1104 1097 * First, build a list of leaf zaps to be destroyed.
1105 1098 * This is passed to the sync context thread,
1106 1099 * which does the actual unlinking.
1107 1100 */
1108 1101 svr->svr_zaplist = fnvlist_alloc();
1109 1102 vdev_remove_enlist_zaps(vd, svr->svr_zaplist);
1110 1103
1111 1104 ivd = vdev_add_parent(vd, &vdev_indirect_ops);
1112 1105 ivd->vdev_removing = 0;
1113 1106
1114 1107 vd->vdev_leaf_zap = 0;
1115 1108
1116 1109 vdev_remove_child(ivd, vd);
1117 1110 vdev_compact_children(ivd);
1118 1111
1119 1112 ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
1120 1113
1121 1114 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1122 1115 dsl_sync_task_nowait(spa->spa_dsl_pool, vdev_remove_complete_sync, svr,
1123 1116 0, ZFS_SPACE_CHECK_NONE, tx);
1124 1117 dmu_tx_commit(tx);
1125 1118
1126 1119 /*
1127 1120 * Indicate that this thread has exited.
1128 1121 * After this, we can not use svr.
1129 1122 */
1130 1123 mutex_enter(&svr->svr_lock);
1131 1124 svr->svr_thread = NULL;
1132 1125 cv_broadcast(&svr->svr_cv);
1133 1126 mutex_exit(&svr->svr_lock);
1134 1127 }
1135 1128
1136 1129 /*
1137 1130 * Complete the removal of a toplevel vdev. This is called in open
1138 1131 * context by the removal thread after we have copied all vdev's data.
1139 1132 */
1140 1133 static void
1141 1134 vdev_remove_complete(spa_t *spa)
1142 1135 {
1143 1136 uint64_t txg;
1144 1137
1145 1138 /*
1146 1139 * Wait for any deferred frees to be synced before we call
1147 1140 * vdev_metaslab_fini()
1148 1141 */
1149 1142 txg_wait_synced(spa->spa_dsl_pool, 0);
1150 1143 txg = spa_vdev_enter(spa);
1151 1144 vdev_t *vd = vdev_lookup_top(spa, spa->spa_vdev_removal->svr_vdev_id);
1152 1145 ASSERT3P(vd->vdev_initialize_thread, ==, NULL);
1153 1146
1154 1147 sysevent_t *ev = spa_event_create(spa, vd, NULL,
1155 1148 ESC_ZFS_VDEV_REMOVE_DEV);
1156 1149
1157 1150 zfs_dbgmsg("finishing device removal for vdev %llu in txg %llu",
1158 1151 vd->vdev_id, txg);
1159 1152
1160 1153 /*
1161 1154 * Discard allocation state.
1162 1155 */
1163 1156 if (vd->vdev_mg != NULL) {
1164 1157 vdev_metaslab_fini(vd);
1165 1158 metaslab_group_destroy(vd->vdev_mg);
1166 1159 vd->vdev_mg = NULL;
1167 1160 }
1168 1161 ASSERT0(vd->vdev_stat.vs_space);
1169 1162 ASSERT0(vd->vdev_stat.vs_dspace);
1170 1163
1171 1164 vdev_remove_replace_with_indirect(vd, txg);
1172 1165
1173 1166 /*
1174 1167 * We now release the locks, allowing spa_sync to run and finish the
1175 1168 * removal via vdev_remove_complete_sync in syncing context.
1176 1169 *
1177 1170 * Note that we hold on to the vdev_t that has been replaced. Since
1178 1171 * it isn't part of the vdev tree any longer, it can't be concurrently
1179 1172 * manipulated, even while we don't have the config lock.
1180 1173 */
1181 1174 (void) spa_vdev_exit(spa, NULL, txg, 0);
1182 1175
1183 1176 /*
1184 1177 * Top ZAP should have been transferred to the indirect vdev in
1185 1178 * vdev_remove_replace_with_indirect.
1186 1179 */
1187 1180 ASSERT0(vd->vdev_top_zap);
1188 1181
1189 1182 /*
1190 1183 * Leaf ZAP should have been moved in vdev_remove_replace_with_indirect.
1191 1184 */
1192 1185 ASSERT0(vd->vdev_leaf_zap);
1193 1186
1194 1187 txg = spa_vdev_enter(spa);
1195 1188 (void) vdev_label_init(vd, 0, VDEV_LABEL_REMOVE);
1196 1189 /*
1197 1190 * Request to update the config and the config cachefile.
1198 1191 */
1199 1192 vdev_config_dirty(spa->spa_root_vdev);
1200 1193 (void) spa_vdev_exit(spa, vd, txg, 0);
1201 1194
1202 1195 spa_event_post(ev);
1203 1196 }
1204 1197
1205 1198 /*
1206 1199 * Evacuates a segment of size at most max_alloc from the vdev
1207 1200 * via repeated calls to spa_vdev_copy_segment. If an allocation
1208 1201 * fails, the pool is probably too fragmented to handle such a
1209 1202 * large size, so decrease max_alloc so that the caller will not try
1210 1203 * this size again this txg.
1211 1204 */
1212 1205 static void
1213 1206 spa_vdev_copy_impl(vdev_t *vd, spa_vdev_removal_t *svr, vdev_copy_arg_t *vca,
1214 1207 uint64_t *max_alloc, dmu_tx_t *tx)
1215 1208 {
1216 1209 uint64_t txg = dmu_tx_get_txg(tx);
1217 1210 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
1218 1211
1219 1212 mutex_enter(&svr->svr_lock);
1220 1213
1221 1214 /*
1222 1215 * Determine how big of a chunk to copy. We can allocate up
1223 1216 * to max_alloc bytes, and we can span up to vdev_removal_max_span
1224 1217 * bytes of unallocated space at a time. "segs" will track the
1225 1218 * allocated segments that we are copying. We may also be copying
1226 1219 * free segments (of up to vdev_removal_max_span bytes).
1227 1220 */
1228 1221 range_tree_t *segs = range_tree_create(NULL, NULL);
1229 1222 for (;;) {
1230 1223 range_seg_t *rs = avl_first(&svr->svr_allocd_segs->rt_root);
1231 1224 if (rs == NULL)
1232 1225 break;
1233 1226
1234 1227 uint64_t seg_length;
1235 1228
1236 1229 if (range_tree_is_empty(segs)) {
1237 1230 /* need to truncate the first seg based on max_alloc */
1238 1231 seg_length =
1239 1232 MIN(rs->rs_end - rs->rs_start, *max_alloc);
1240 1233 } else {
1241 1234 if (rs->rs_start - range_tree_max(segs) >
1242 1235 vdev_removal_max_span) {
1243 1236 /*
1244 1237 * Including this segment would cause us to
1245 1238 * copy a larger unneeded chunk than is allowed.
1246 1239 */
1247 1240 break;
1248 1241 } else if (rs->rs_end - range_tree_min(segs) >
1249 1242 *max_alloc) {
1250 1243 /*
1251 1244 * This additional segment would extend past
1252 1245 * max_alloc. Rather than splitting this
1253 1246 * segment, leave it for the next mapping.
1254 1247 */
1255 1248 break;
1256 1249 } else {
1257 1250 seg_length = rs->rs_end - rs->rs_start;
1258 1251 }
1259 1252 }
1260 1253
1261 1254 range_tree_add(segs, rs->rs_start, seg_length);
1262 1255 range_tree_remove(svr->svr_allocd_segs,
1263 1256 rs->rs_start, seg_length);
1264 1257 }
1265 1258
1266 1259 if (range_tree_is_empty(segs)) {
1267 1260 mutex_exit(&svr->svr_lock);
1268 1261 range_tree_destroy(segs);
1269 1262 return;
1270 1263 }
1271 1264
1272 1265 if (svr->svr_max_offset_to_sync[txg & TXG_MASK] == 0) {
1273 1266 dsl_sync_task_nowait(dmu_tx_pool(tx), vdev_mapping_sync,
1274 1267 svr, 0, ZFS_SPACE_CHECK_NONE, tx);
1275 1268 }
1276 1269
1277 1270 svr->svr_max_offset_to_sync[txg & TXG_MASK] = range_tree_max(segs);
1278 1271
1279 1272 /*
1280 1273 * Note: this is the amount of *allocated* space
1281 1274 * that we are taking care of each txg.
1282 1275 */
1283 1276 svr->svr_bytes_done[txg & TXG_MASK] += range_tree_space(segs);
1284 1277
1285 1278 mutex_exit(&svr->svr_lock);
1286 1279
1287 1280 zio_alloc_list_t zal;
1288 1281 metaslab_trace_init(&zal);
1289 1282 uint64_t thismax = SPA_MAXBLOCKSIZE;
1290 1283 while (!range_tree_is_empty(segs)) {
1291 1284 int error = spa_vdev_copy_segment(vd,
1292 1285 segs, thismax, txg, vca, &zal);
1293 1286
1294 1287 if (error == ENOSPC) {
1295 1288 /*
1296 1289 * Cut our segment in half, and don't try this
1297 1290 * segment size again this txg. Note that the
1298 1291 * allocation size must be aligned to the highest
1299 1292 * ashift in the pool, so that the allocation will
1300 1293 * not be padded out to a multiple of the ashift,
1301 1294 * which could cause us to think that this mapping
1302 1295 * is larger than we intended.
1303 1296 */
1304 1297 ASSERT3U(spa->spa_max_ashift, >=, SPA_MINBLOCKSHIFT);
1305 1298 ASSERT3U(spa->spa_max_ashift, ==, spa->spa_min_ashift);
1306 1299 uint64_t attempted =
1307 1300 MIN(range_tree_span(segs), thismax);
1308 1301 thismax = P2ROUNDUP(attempted / 2,
1309 1302 1 << spa->spa_max_ashift);
1310 1303 /*
1311 1304 * The minimum-size allocation can not fail.
1312 1305 */
1313 1306 ASSERT3U(attempted, >, 1 << spa->spa_max_ashift);
1314 1307 *max_alloc = attempted - (1 << spa->spa_max_ashift);
1315 1308 } else {
1316 1309 ASSERT0(error);
1317 1310
1318 1311 /*
1319 1312 * We've performed an allocation, so reset the
1320 1313 * alloc trace list.
1321 1314 */
1322 1315 metaslab_trace_fini(&zal);
1323 1316 metaslab_trace_init(&zal);
1324 1317 }
1325 1318 }
1326 1319 metaslab_trace_fini(&zal);
1327 1320 range_tree_destroy(segs);
1328 1321 }
1329 1322
1330 1323 /*
1331 1324 * The removal thread operates in open context. It iterates over all
1332 1325 * allocated space in the vdev, by loading each metaslab's spacemap.
1333 1326 * For each contiguous segment of allocated space (capping the segment
1334 1327 * size at SPA_MAXBLOCKSIZE), we:
1335 1328 * - Allocate space for it on another vdev.
1336 1329 * - Create a new mapping from the old location to the new location
1337 1330 * (as a record in svr_new_segments).
1338 1331 * - Initiate a logical read zio to get the data off the removing disk.
1339 1332 * - In the read zio's done callback, initiate a logical write zio to
1340 1333 * write it to the new vdev.
1341 1334 * Note that all of this will take effect when a particular TXG syncs.
1342 1335 * The sync thread ensures that all the phys reads and writes for the syncing
1343 1336 * TXG have completed (see spa_txg_zio) and writes the new mappings to disk
1344 1337 * (see vdev_mapping_sync()).
1345 1338 */
1346 1339 static void
1347 1340 spa_vdev_remove_thread(void *arg)
1348 1341 {
1349 1342 spa_t *spa = arg;
1350 1343 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1351 1344 vdev_copy_arg_t vca;
1352 1345 uint64_t max_alloc = zfs_remove_max_segment;
1353 1346 uint64_t last_txg = 0;
1354 1347
1355 1348 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
1356 1349 vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
1357 1350 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
1358 1351 uint64_t start_offset = vdev_indirect_mapping_max_offset(vim);
1359 1352
1360 1353 ASSERT3P(vd->vdev_ops, !=, &vdev_indirect_ops);
1361 1354 ASSERT(vdev_is_concrete(vd));
1362 1355 ASSERT(vd->vdev_removing);
1363 1356 ASSERT(vd->vdev_indirect_config.vic_mapping_object != 0);
1364 1357 ASSERT(vim != NULL);
1365 1358
1366 1359 mutex_init(&vca.vca_lock, NULL, MUTEX_DEFAULT, NULL);
1367 1360 cv_init(&vca.vca_cv, NULL, CV_DEFAULT, NULL);
1368 1361 vca.vca_outstanding_bytes = 0;
1369 1362
1370 1363 mutex_enter(&svr->svr_lock);
1371 1364
1372 1365 /*
1373 1366 * Start from vim_max_offset so we pick up where we left off
1374 1367 * if we are restarting the removal after opening the pool.
1375 1368 */
1376 1369 uint64_t msi;
1377 1370 for (msi = start_offset >> vd->vdev_ms_shift;
1378 1371 msi < vd->vdev_ms_count && !svr->svr_thread_exit; msi++) {
1379 1372 metaslab_t *msp = vd->vdev_ms[msi];
1380 1373 ASSERT3U(msi, <=, vd->vdev_ms_count);
1381 1374
1382 1375 ASSERT0(range_tree_space(svr->svr_allocd_segs));
1383 1376
1384 1377 mutex_enter(&msp->ms_sync_lock);
1385 1378 mutex_enter(&msp->ms_lock);
1386 1379
1387 1380 /*
1388 1381 * Assert nothing in flight -- ms_*tree is empty.
1389 1382 */
1390 1383 for (int i = 0; i < TXG_SIZE; i++) {
1391 1384 ASSERT0(range_tree_space(msp->ms_allocating[i]));
1392 1385 }
1393 1386
|
↓ open down ↓ |
1089 lines elided |
↑ open up ↑ |
1394 1387 /*
1395 1388 * If the metaslab has ever been allocated from (ms_sm!=NULL),
1396 1389 * read the allocated segments from the space map object
1397 1390 * into svr_allocd_segs. Since we do this while holding
1398 1391 * svr_lock and ms_sync_lock, concurrent frees (which
1399 1392 * would have modified the space map) will wait for us
1400 1393 * to finish loading the spacemap, and then take the
1401 1394 * appropriate action (see free_from_removing_vdev()).
1402 1395 */
1403 1396 if (msp->ms_sm != NULL) {
1404 - space_map_t *sm = NULL;
1397 + VERIFY0(space_map_load(msp->ms_sm,
1398 + svr->svr_allocd_segs, SM_ALLOC));
1405 1399
1406 - /*
1407 - * We have to open a new space map here, because
1408 - * ms_sm's sm_length and sm_alloc may not reflect
1409 - * what's in the object contents, if we are in between
1410 - * metaslab_sync() and metaslab_sync_done().
1411 - */
1412 - VERIFY0(space_map_open(&sm,
1413 - spa->spa_dsl_pool->dp_meta_objset,
1414 - msp->ms_sm->sm_object, msp->ms_sm->sm_start,
1415 - msp->ms_sm->sm_size, msp->ms_sm->sm_shift));
1416 - space_map_update(sm);
1417 - VERIFY0(space_map_load(sm, svr->svr_allocd_segs,
1418 - SM_ALLOC));
1419 - space_map_close(sm);
1420 -
1421 1400 range_tree_walk(msp->ms_freeing,
1422 1401 range_tree_remove, svr->svr_allocd_segs);
1423 1402
1424 1403 /*
1425 1404 * When we are resuming from a paused removal (i.e.
1426 1405 * when importing a pool with a removal in progress),
1427 1406 * discard any state that we have already processed.
1428 1407 */
1429 1408 range_tree_clear(svr->svr_allocd_segs, 0, start_offset);
1430 1409 }
1431 1410 mutex_exit(&msp->ms_lock);
1432 1411 mutex_exit(&msp->ms_sync_lock);
1433 1412
1434 1413 vca.vca_msp = msp;
1435 1414 zfs_dbgmsg("copying %llu segments for metaslab %llu",
1436 1415 avl_numnodes(&svr->svr_allocd_segs->rt_root),
1437 1416 msp->ms_id);
1438 1417
1439 1418 while (!svr->svr_thread_exit &&
1440 1419 !range_tree_is_empty(svr->svr_allocd_segs)) {
1441 1420
1442 1421 mutex_exit(&svr->svr_lock);
1443 1422
1444 1423 /*
1445 1424 * We need to periodically drop the config lock so that
1446 1425 * writers can get in. Additionally, we can't wait
1447 1426 * for a txg to sync while holding a config lock
1448 1427 * (since a waiting writer could cause a 3-way deadlock
1449 1428 * with the sync thread, which also gets a config
1450 1429 * lock for reader). So we can't hold the config lock
1451 1430 * while calling dmu_tx_assign().
1452 1431 */
1453 1432 spa_config_exit(spa, SCL_CONFIG, FTAG);
1454 1433
1455 1434 /*
1456 1435 * This delay will pause the removal around the point
1457 1436 * specified by zfs_remove_max_bytes_pause. We do this
1458 1437 * solely from the test suite or during debugging.
1459 1438 */
1460 1439 uint64_t bytes_copied =
1461 1440 spa->spa_removing_phys.sr_copied;
1462 1441 for (int i = 0; i < TXG_SIZE; i++)
1463 1442 bytes_copied += svr->svr_bytes_done[i];
1464 1443 while (zfs_remove_max_bytes_pause <= bytes_copied &&
1465 1444 !svr->svr_thread_exit)
1466 1445 delay(hz);
1467 1446
1468 1447 mutex_enter(&vca.vca_lock);
1469 1448 while (vca.vca_outstanding_bytes >
1470 1449 zfs_remove_max_copy_bytes) {
1471 1450 cv_wait(&vca.vca_cv, &vca.vca_lock);
1472 1451 }
1473 1452 mutex_exit(&vca.vca_lock);
1474 1453
1475 1454 dmu_tx_t *tx =
1476 1455 dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
1477 1456
1478 1457 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
1479 1458 uint64_t txg = dmu_tx_get_txg(tx);
1480 1459
1481 1460 /*
1482 1461 * Reacquire the vdev_config lock. The vdev_t
1483 1462 * that we're removing may have changed, e.g. due
1484 1463 * to a vdev_attach or vdev_detach.
1485 1464 */
1486 1465 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
1487 1466 vd = vdev_lookup_top(spa, svr->svr_vdev_id);
1488 1467
1489 1468 if (txg != last_txg)
1490 1469 max_alloc = zfs_remove_max_segment;
1491 1470 last_txg = txg;
1492 1471
1493 1472 spa_vdev_copy_impl(vd, svr, &vca, &max_alloc, tx);
1494 1473
1495 1474 dmu_tx_commit(tx);
1496 1475 mutex_enter(&svr->svr_lock);
1497 1476 }
1498 1477 }
1499 1478
1500 1479 mutex_exit(&svr->svr_lock);
1501 1480
1502 1481 spa_config_exit(spa, SCL_CONFIG, FTAG);
1503 1482
1504 1483 /*
1505 1484 * Wait for all copies to finish before cleaning up the vca.
1506 1485 */
1507 1486 txg_wait_synced(spa->spa_dsl_pool, 0);
1508 1487 ASSERT0(vca.vca_outstanding_bytes);
1509 1488
1510 1489 mutex_destroy(&vca.vca_lock);
1511 1490 cv_destroy(&vca.vca_cv);
1512 1491
1513 1492 if (svr->svr_thread_exit) {
1514 1493 mutex_enter(&svr->svr_lock);
1515 1494 range_tree_vacate(svr->svr_allocd_segs, NULL, NULL);
1516 1495 svr->svr_thread = NULL;
1517 1496 cv_broadcast(&svr->svr_cv);
1518 1497 mutex_exit(&svr->svr_lock);
1519 1498 } else {
1520 1499 ASSERT0(range_tree_space(svr->svr_allocd_segs));
1521 1500 vdev_remove_complete(spa);
1522 1501 }
1523 1502 }
1524 1503
1525 1504 void
1526 1505 spa_vdev_remove_suspend(spa_t *spa)
1527 1506 {
1528 1507 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1529 1508
1530 1509 if (svr == NULL)
1531 1510 return;
1532 1511
1533 1512 mutex_enter(&svr->svr_lock);
1534 1513 svr->svr_thread_exit = B_TRUE;
1535 1514 while (svr->svr_thread != NULL)
1536 1515 cv_wait(&svr->svr_cv, &svr->svr_lock);
1537 1516 svr->svr_thread_exit = B_FALSE;
1538 1517 mutex_exit(&svr->svr_lock);
1539 1518 }
1540 1519
1541 1520 /* ARGSUSED */
1542 1521 static int
1543 1522 spa_vdev_remove_cancel_check(void *arg, dmu_tx_t *tx)
1544 1523 {
1545 1524 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
1546 1525
1547 1526 if (spa->spa_vdev_removal == NULL)
1548 1527 return (ENOTACTIVE);
1549 1528 return (0);
1550 1529 }
1551 1530
1552 1531 /*
1553 1532 * Cancel a removal by freeing all entries from the partial mapping
1554 1533 * and marking the vdev as no longer being removing.
1555 1534 */
1556 1535 /* ARGSUSED */
1557 1536 static void
1558 1537 spa_vdev_remove_cancel_sync(void *arg, dmu_tx_t *tx)
1559 1538 {
1560 1539 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
1561 1540 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1562 1541 vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
1563 1542 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
1564 1543 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
1565 1544 objset_t *mos = spa->spa_meta_objset;
1566 1545
1567 1546 ASSERT3P(svr->svr_thread, ==, NULL);
1568 1547
1569 1548 spa_feature_decr(spa, SPA_FEATURE_DEVICE_REMOVAL, tx);
1570 1549 if (vdev_obsolete_counts_are_precise(vd)) {
1571 1550 spa_feature_decr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
1572 1551 VERIFY0(zap_remove(spa->spa_meta_objset, vd->vdev_top_zap,
1573 1552 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE, tx));
1574 1553 }
1575 1554
1576 1555 if (vdev_obsolete_sm_object(vd) != 0) {
1577 1556 ASSERT(vd->vdev_obsolete_sm != NULL);
1578 1557 ASSERT3U(vdev_obsolete_sm_object(vd), ==,
1579 1558 space_map_object(vd->vdev_obsolete_sm));
1580 1559
1581 1560 space_map_free(vd->vdev_obsolete_sm, tx);
1582 1561 VERIFY0(zap_remove(spa->spa_meta_objset, vd->vdev_top_zap,
1583 1562 VDEV_TOP_ZAP_INDIRECT_OBSOLETE_SM, tx));
1584 1563 space_map_close(vd->vdev_obsolete_sm);
1585 1564 vd->vdev_obsolete_sm = NULL;
1586 1565 spa_feature_decr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
1587 1566 }
1588 1567 for (int i = 0; i < TXG_SIZE; i++) {
1589 1568 ASSERT(list_is_empty(&svr->svr_new_segments[i]));
1590 1569 ASSERT3U(svr->svr_max_offset_to_sync[i], <=,
1591 1570 vdev_indirect_mapping_max_offset(vim));
1592 1571 }
1593 1572
1594 1573 for (uint64_t msi = 0; msi < vd->vdev_ms_count; msi++) {
1595 1574 metaslab_t *msp = vd->vdev_ms[msi];
1596 1575
1597 1576 if (msp->ms_start >= vdev_indirect_mapping_max_offset(vim))
1598 1577 break;
1599 1578
1600 1579 ASSERT0(range_tree_space(svr->svr_allocd_segs));
1601 1580
1602 1581 mutex_enter(&msp->ms_lock);
1603 1582
|
↓ open down ↓ |
173 lines elided |
↑ open up ↑ |
1604 1583 /*
1605 1584 * Assert nothing in flight -- ms_*tree is empty.
1606 1585 */
1607 1586 for (int i = 0; i < TXG_SIZE; i++)
1608 1587 ASSERT0(range_tree_space(msp->ms_allocating[i]));
1609 1588 for (int i = 0; i < TXG_DEFER_SIZE; i++)
1610 1589 ASSERT0(range_tree_space(msp->ms_defer[i]));
1611 1590 ASSERT0(range_tree_space(msp->ms_freed));
1612 1591
1613 1592 if (msp->ms_sm != NULL) {
1614 - /*
1615 - * Assert that the in-core spacemap has the same
1616 - * length as the on-disk one, so we can use the
1617 - * existing in-core spacemap to load it from disk.
1618 - */
1619 - ASSERT3U(msp->ms_sm->sm_alloc, ==,
1620 - msp->ms_sm->sm_phys->smp_alloc);
1621 - ASSERT3U(msp->ms_sm->sm_length, ==,
1622 - msp->ms_sm->sm_phys->smp_objsize);
1623 -
1624 1593 mutex_enter(&svr->svr_lock);
1625 1594 VERIFY0(space_map_load(msp->ms_sm,
1626 1595 svr->svr_allocd_segs, SM_ALLOC));
1627 1596 range_tree_walk(msp->ms_freeing,
1628 1597 range_tree_remove, svr->svr_allocd_segs);
1629 1598
1630 1599 /*
1631 1600 * Clear everything past what has been synced,
1632 1601 * because we have not allocated mappings for it yet.
1633 1602 */
1634 1603 uint64_t syncd = vdev_indirect_mapping_max_offset(vim);
1635 1604 uint64_t sm_end = msp->ms_sm->sm_start +
1636 1605 msp->ms_sm->sm_size;
1637 1606 if (sm_end > syncd)
1638 1607 range_tree_clear(svr->svr_allocd_segs,
1639 1608 syncd, sm_end - syncd);
1640 1609
1641 1610 mutex_exit(&svr->svr_lock);
1642 1611 }
1643 1612 mutex_exit(&msp->ms_lock);
1644 1613
1645 1614 mutex_enter(&svr->svr_lock);
1646 1615 range_tree_vacate(svr->svr_allocd_segs,
1647 1616 free_mapped_segment_cb, vd);
1648 1617 mutex_exit(&svr->svr_lock);
1649 1618 }
1650 1619
1651 1620 /*
1652 1621 * Note: this must happen after we invoke free_mapped_segment_cb,
1653 1622 * because it adds to the obsolete_segments.
1654 1623 */
1655 1624 range_tree_vacate(vd->vdev_obsolete_segments, NULL, NULL);
1656 1625
1657 1626 ASSERT3U(vic->vic_mapping_object, ==,
1658 1627 vdev_indirect_mapping_object(vd->vdev_indirect_mapping));
1659 1628 vdev_indirect_mapping_close(vd->vdev_indirect_mapping);
1660 1629 vd->vdev_indirect_mapping = NULL;
1661 1630 vdev_indirect_mapping_free(mos, vic->vic_mapping_object, tx);
1662 1631 vic->vic_mapping_object = 0;
1663 1632
1664 1633 ASSERT3U(vic->vic_births_object, ==,
1665 1634 vdev_indirect_births_object(vd->vdev_indirect_births));
1666 1635 vdev_indirect_births_close(vd->vdev_indirect_births);
1667 1636 vd->vdev_indirect_births = NULL;
1668 1637 vdev_indirect_births_free(mos, vic->vic_births_object, tx);
1669 1638 vic->vic_births_object = 0;
1670 1639
1671 1640 /*
1672 1641 * We may have processed some frees from the removing vdev in this
1673 1642 * txg, thus increasing svr_bytes_done; discard that here to
1674 1643 * satisfy the assertions in spa_vdev_removal_destroy().
1675 1644 * Note that future txg's can not have any bytes_done, because
1676 1645 * future TXG's are only modified from open context, and we have
1677 1646 * already shut down the copying thread.
1678 1647 */
1679 1648 svr->svr_bytes_done[dmu_tx_get_txg(tx) & TXG_MASK] = 0;
1680 1649 spa_finish_removal(spa, DSS_CANCELED, tx);
1681 1650
1682 1651 vd->vdev_removing = B_FALSE;
1683 1652 vdev_config_dirty(vd);
1684 1653
1685 1654 zfs_dbgmsg("canceled device removal for vdev %llu in %llu",
1686 1655 vd->vdev_id, dmu_tx_get_txg(tx));
1687 1656 spa_history_log_internal(spa, "vdev remove canceled", tx,
1688 1657 "%s vdev %llu %s", spa_name(spa),
1689 1658 vd->vdev_id, (vd->vdev_path != NULL) ? vd->vdev_path : "-");
1690 1659 }
1691 1660
1692 1661 int
1693 1662 spa_vdev_remove_cancel(spa_t *spa)
1694 1663 {
1695 1664 spa_vdev_remove_suspend(spa);
1696 1665
1697 1666 if (spa->spa_vdev_removal == NULL)
1698 1667 return (ENOTACTIVE);
1699 1668
1700 1669 uint64_t vdid = spa->spa_vdev_removal->svr_vdev_id;
1701 1670
1702 1671 int error = dsl_sync_task(spa->spa_name, spa_vdev_remove_cancel_check,
1703 1672 spa_vdev_remove_cancel_sync, NULL, 0,
1704 1673 ZFS_SPACE_CHECK_EXTRA_RESERVED);
1705 1674
|
↓ open down ↓ |
72 lines elided |
↑ open up ↑ |
1706 1675 if (error == 0) {
1707 1676 spa_config_enter(spa, SCL_ALLOC | SCL_VDEV, FTAG, RW_WRITER);
1708 1677 vdev_t *vd = vdev_lookup_top(spa, vdid);
1709 1678 metaslab_group_activate(vd->vdev_mg);
1710 1679 spa_config_exit(spa, SCL_ALLOC | SCL_VDEV, FTAG);
1711 1680 }
1712 1681
1713 1682 return (error);
1714 1683 }
1715 1684
1716 -/*
1717 - * Called every sync pass of every txg if there's a svr.
1718 - */
1719 1685 void
1720 1686 svr_sync(spa_t *spa, dmu_tx_t *tx)
1721 1687 {
1722 1688 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1723 1689 int txgoff = dmu_tx_get_txg(tx) & TXG_MASK;
1724 1690
1725 1691 /*
1726 1692 * This check is necessary so that we do not dirty the
1727 1693 * DIRECTORY_OBJECT via spa_sync_removing_state() when there
1728 1694 * is nothing to do. Dirtying it every time would prevent us
1729 1695 * from syncing-to-convergence.
1730 1696 */
1731 1697 if (svr->svr_bytes_done[txgoff] == 0)
1732 1698 return;
1733 1699
1734 1700 /*
1735 1701 * Update progress accounting.
1736 1702 */
1737 1703 spa->spa_removing_phys.sr_copied += svr->svr_bytes_done[txgoff];
1738 1704 svr->svr_bytes_done[txgoff] = 0;
1739 1705
1740 1706 spa_sync_removing_state(spa, tx);
1741 1707 }
1742 1708
1743 1709 static void
1744 1710 vdev_remove_make_hole_and_free(vdev_t *vd)
1745 1711 {
1746 1712 uint64_t id = vd->vdev_id;
1747 1713 spa_t *spa = vd->vdev_spa;
1748 1714 vdev_t *rvd = spa->spa_root_vdev;
1749 1715 boolean_t last_vdev = (id == (rvd->vdev_children - 1));
1750 1716
1751 1717 ASSERT(MUTEX_HELD(&spa_namespace_lock));
1752 1718 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1753 1719
1754 1720 vdev_free(vd);
1755 1721
1756 1722 if (last_vdev) {
1757 1723 vdev_compact_children(rvd);
1758 1724 } else {
1759 1725 vd = vdev_alloc_common(spa, id, 0, &vdev_hole_ops);
1760 1726 vdev_add_child(rvd, vd);
1761 1727 }
1762 1728 vdev_config_dirty(rvd);
1763 1729
1764 1730 /*
1765 1731 * Reassess the health of our root vdev.
1766 1732 */
1767 1733 vdev_reopen(rvd);
1768 1734 }
1769 1735
1770 1736 /*
1771 1737 * Remove a log device. The config lock is held for the specified TXG.
|
↓ open down ↓ |
43 lines elided |
↑ open up ↑ |
1772 1738 */
1773 1739 static int
1774 1740 spa_vdev_remove_log(vdev_t *vd, uint64_t *txg)
1775 1741 {
1776 1742 metaslab_group_t *mg = vd->vdev_mg;
1777 1743 spa_t *spa = vd->vdev_spa;
1778 1744 int error = 0;
1779 1745
1780 1746 ASSERT(vd->vdev_islog);
1781 1747 ASSERT(vd == vd->vdev_top);
1748 + ASSERT(MUTEX_HELD(&spa_namespace_lock));
1782 1749
1783 1750 /*
1784 1751 * Stop allocating from this vdev.
1785 1752 */
1786 1753 metaslab_group_passivate(mg);
1787 1754
1788 1755 /*
1789 1756 * Wait for the youngest allocations and frees to sync,
1790 1757 * and then wait for the deferral of those frees to finish.
1791 1758 */
1792 1759 spa_vdev_config_exit(spa, NULL,
1793 1760 *txg + TXG_CONCURRENT_STATES + TXG_DEFER_SIZE, 0, FTAG);
1794 1761
1795 1762 /*
1796 - * Evacuate the device. We don't hold the config lock as writer
1797 - * since we need to do I/O but we do keep the
1763 + * Evacuate the device. We don't hold the config lock as
1764 + * writer since we need to do I/O but we do keep the
1798 1765 * spa_namespace_lock held. Once this completes the device
1799 1766 * should no longer have any blocks allocated on it.
1800 1767 */
1801 - if (vd->vdev_islog) {
1802 - if (vd->vdev_stat.vs_alloc != 0)
1803 - error = spa_reset_logs(spa);
1804 - }
1768 + ASSERT(MUTEX_HELD(&spa_namespace_lock));
1769 + if (vd->vdev_stat.vs_alloc != 0)
1770 + error = spa_reset_logs(spa);
1805 1771
1806 1772 *txg = spa_vdev_config_enter(spa);
1807 1773
1808 1774 if (error != 0) {
1809 1775 metaslab_group_activate(mg);
1810 1776 return (error);
1811 1777 }
1812 1778 ASSERT0(vd->vdev_stat.vs_alloc);
1813 1779
1814 1780 /*
1815 1781 * The evacuation succeeded. Remove any remaining MOS metadata
1816 1782 * associated with this vdev, and wait for these changes to sync.
1817 1783 */
1818 1784 vd->vdev_removing = B_TRUE;
1819 1785
1820 1786 vdev_dirty_leaves(vd, VDD_DTL, *txg);
1821 1787 vdev_config_dirty(vd);
1822 1788
1789 + vdev_metaslab_fini(vd);
1790 +
1823 1791 spa_history_log_internal(spa, "vdev remove", NULL,
1824 1792 "%s vdev %llu (log) %s", spa_name(spa), vd->vdev_id,
1825 1793 (vd->vdev_path != NULL) ? vd->vdev_path : "-");
1826 1794
1827 1795 /* Make sure these changes are sync'ed */
1828 1796 spa_vdev_config_exit(spa, NULL, *txg, 0, FTAG);
1829 1797
1830 1798 /* Stop initializing */
1831 1799 (void) vdev_initialize_stop_all(vd, VDEV_INITIALIZE_CANCELED);
1832 1800
1833 1801 *txg = spa_vdev_config_enter(spa);
1834 1802
1835 1803 sysevent_t *ev = spa_event_create(spa, vd, NULL,
1836 1804 ESC_ZFS_VDEV_REMOVE_DEV);
1837 1805 ASSERT(MUTEX_HELD(&spa_namespace_lock));
1838 1806 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1839 1807
1840 1808 /* The top ZAP should have been destroyed by vdev_remove_empty. */
1841 1809 ASSERT0(vd->vdev_top_zap);
|
↓ open down ↓ |
9 lines elided |
↑ open up ↑ |
1842 1810 /* The leaf ZAP should have been destroyed by vdev_dtl_sync. */
1843 1811 ASSERT0(vd->vdev_leaf_zap);
1844 1812
1845 1813 (void) vdev_label_init(vd, 0, VDEV_LABEL_REMOVE);
1846 1814
1847 1815 if (list_link_active(&vd->vdev_state_dirty_node))
1848 1816 vdev_state_clean(vd);
1849 1817 if (list_link_active(&vd->vdev_config_dirty_node))
1850 1818 vdev_config_clean(vd);
1851 1819
1820 + ASSERT0(vd->vdev_stat.vs_alloc);
1821 +
1852 1822 /*
1853 1823 * Clean up the vdev namespace.
1854 1824 */
1855 1825 vdev_remove_make_hole_and_free(vd);
1856 1826
1857 1827 if (ev != NULL)
1858 1828 spa_event_post(ev);
1859 1829
1860 1830 return (0);
1861 1831 }
1862 1832
1863 1833 static int
1864 1834 spa_vdev_remove_top_check(vdev_t *vd)
1865 1835 {
1866 1836 spa_t *spa = vd->vdev_spa;
1867 1837
1868 1838 if (vd != vd->vdev_top)
1869 1839 return (SET_ERROR(ENOTSUP));
1870 1840
1871 1841 if (!spa_feature_is_enabled(spa, SPA_FEATURE_DEVICE_REMOVAL))
1872 1842 return (SET_ERROR(ENOTSUP));
1873 1843
1874 1844 /* available space in the pool's normal class */
1875 1845 uint64_t available = dsl_dir_space_available(
1876 1846 spa->spa_dsl_pool->dp_root_dir, NULL, 0, B_TRUE);
1877 1847
1878 1848 metaslab_class_t *mc = vd->vdev_mg->mg_class;
1879 1849
1880 1850 /*
1881 1851 * When removing a vdev from an allocation class that has
1882 1852 * remaining vdevs, include available space from the class.
1883 1853 */
1884 1854 if (mc != spa_normal_class(spa) && mc->mc_groups > 1) {
1885 1855 uint64_t class_avail = metaslab_class_get_space(mc) -
1886 1856 metaslab_class_get_alloc(mc);
1887 1857
1888 1858 /* add class space, adjusted for overhead */
1889 1859 available += (class_avail * 94) / 100;
1890 1860 }
1891 1861
1892 1862 /*
1893 1863 * There has to be enough free space to remove the
1894 1864 * device and leave double the "slop" space (i.e. we
1895 1865 * must leave at least 3% of the pool free, in addition to
1896 1866 * the normal slop space).
1897 1867 */
1898 1868 if (available < vd->vdev_stat.vs_dspace + spa_get_slop_space(spa)) {
1899 1869 return (SET_ERROR(ENOSPC));
1900 1870 }
1901 1871
1902 1872 /*
1903 1873 * There can not be a removal in progress.
1904 1874 */
1905 1875 if (spa->spa_removing_phys.sr_state == DSS_SCANNING)
1906 1876 return (SET_ERROR(EBUSY));
1907 1877
1908 1878 /*
1909 1879 * The device must have all its data.
1910 1880 */
1911 1881 if (!vdev_dtl_empty(vd, DTL_MISSING) ||
1912 1882 !vdev_dtl_empty(vd, DTL_OUTAGE))
1913 1883 return (SET_ERROR(EBUSY));
1914 1884
1915 1885 /*
1916 1886 * The device must be healthy.
1917 1887 */
1918 1888 if (!vdev_readable(vd))
1919 1889 return (SET_ERROR(EIO));
1920 1890
1921 1891 /*
1922 1892 * All vdevs in normal class must have the same ashift.
1923 1893 */
1924 1894 if (spa->spa_max_ashift != spa->spa_min_ashift) {
1925 1895 return (SET_ERROR(EINVAL));
1926 1896 }
1927 1897
1928 1898 /*
1929 1899 * All vdevs in normal class must have the same ashift
1930 1900 * and not be raidz.
1931 1901 */
1932 1902 vdev_t *rvd = spa->spa_root_vdev;
1933 1903 int num_indirect = 0;
1934 1904 for (uint64_t id = 0; id < rvd->vdev_children; id++) {
1935 1905 vdev_t *cvd = rvd->vdev_child[id];
1936 1906 if (cvd->vdev_ashift != 0 && !cvd->vdev_islog)
1937 1907 ASSERT3U(cvd->vdev_ashift, ==, spa->spa_max_ashift);
1938 1908 if (cvd->vdev_ops == &vdev_indirect_ops)
1939 1909 num_indirect++;
1940 1910 if (!vdev_is_concrete(cvd))
1941 1911 continue;
1942 1912 if (cvd->vdev_ops == &vdev_raidz_ops)
1943 1913 return (SET_ERROR(EINVAL));
1944 1914 /*
1945 1915 * Need the mirror to be mirror of leaf vdevs only
1946 1916 */
1947 1917 if (cvd->vdev_ops == &vdev_mirror_ops) {
1948 1918 for (uint64_t cid = 0;
1949 1919 cid < cvd->vdev_children; cid++) {
1950 1920 vdev_t *tmp = cvd->vdev_child[cid];
1951 1921 if (!tmp->vdev_ops->vdev_op_leaf)
1952 1922 return (SET_ERROR(EINVAL));
1953 1923 }
1954 1924 }
1955 1925 }
1956 1926
1957 1927 return (0);
1958 1928 }
1959 1929
1960 1930 /*
1961 1931 * Initiate removal of a top-level vdev, reducing the total space in the pool.
1962 1932 * The config lock is held for the specified TXG. Once initiated,
1963 1933 * evacuation of all allocated space (copying it to other vdevs) happens
1964 1934 * in the background (see spa_vdev_remove_thread()), and can be canceled
1965 1935 * (see spa_vdev_remove_cancel()). If successful, the vdev will
1966 1936 * be transformed to an indirect vdev (see spa_vdev_remove_complete()).
1967 1937 */
1968 1938 static int
1969 1939 spa_vdev_remove_top(vdev_t *vd, uint64_t *txg)
1970 1940 {
1971 1941 spa_t *spa = vd->vdev_spa;
1972 1942 int error;
1973 1943
1974 1944 /*
1975 1945 * Check for errors up-front, so that we don't waste time
1976 1946 * passivating the metaslab group and clearing the ZIL if there
1977 1947 * are errors.
1978 1948 */
1979 1949 error = spa_vdev_remove_top_check(vd);
1980 1950 if (error != 0)
1981 1951 return (error);
1982 1952
1983 1953 /*
1984 1954 * Stop allocating from this vdev. Note that we must check
1985 1955 * that this is not the only device in the pool before
1986 1956 * passivating, otherwise we will not be able to make
1987 1957 * progress because we can't allocate from any vdevs.
1988 1958 * The above check for sufficient free space serves this
1989 1959 * purpose.
1990 1960 */
1991 1961 metaslab_group_t *mg = vd->vdev_mg;
1992 1962 metaslab_group_passivate(mg);
1993 1963
1994 1964 /*
1995 1965 * Wait for the youngest allocations and frees to sync,
1996 1966 * and then wait for the deferral of those frees to finish.
1997 1967 */
1998 1968 spa_vdev_config_exit(spa, NULL,
1999 1969 *txg + TXG_CONCURRENT_STATES + TXG_DEFER_SIZE, 0, FTAG);
2000 1970
2001 1971 /*
2002 1972 * We must ensure that no "stubby" log blocks are allocated
2003 1973 * on the device to be removed. These blocks could be
2004 1974 * written at any time, including while we are in the middle
2005 1975 * of copying them.
2006 1976 */
2007 1977 error = spa_reset_logs(spa);
2008 1978
2009 1979 /*
2010 1980 * We stop any initializing that is currently in progress but leave
2011 1981 * the state as "active". This will allow the initializing to resume
2012 1982 * if the removal is canceled sometime later.
2013 1983 */
2014 1984 vdev_initialize_stop_all(vd, VDEV_INITIALIZE_ACTIVE);
2015 1985
2016 1986 *txg = spa_vdev_config_enter(spa);
2017 1987
2018 1988 /*
2019 1989 * Things might have changed while the config lock was dropped
2020 1990 * (e.g. space usage). Check for errors again.
2021 1991 */
2022 1992 if (error == 0)
2023 1993 error = spa_vdev_remove_top_check(vd);
2024 1994
2025 1995 if (error != 0) {
2026 1996 metaslab_group_activate(mg);
2027 1997 spa_async_request(spa, SPA_ASYNC_INITIALIZE_RESTART);
2028 1998 return (error);
2029 1999 }
2030 2000
2031 2001 vd->vdev_removing = B_TRUE;
2032 2002
2033 2003 vdev_dirty_leaves(vd, VDD_DTL, *txg);
2034 2004 vdev_config_dirty(vd);
2035 2005 dmu_tx_t *tx = dmu_tx_create_assigned(spa->spa_dsl_pool, *txg);
2036 2006 dsl_sync_task_nowait(spa->spa_dsl_pool,
2037 2007 vdev_remove_initiate_sync,
2038 2008 (void *)(uintptr_t)vd->vdev_id, 0, ZFS_SPACE_CHECK_NONE, tx);
2039 2009 dmu_tx_commit(tx);
2040 2010
2041 2011 return (0);
2042 2012 }
2043 2013
2044 2014 /*
2045 2015 * Remove a device from the pool.
2046 2016 *
2047 2017 * Removing a device from the vdev namespace requires several steps
2048 2018 * and can take a significant amount of time. As a result we use
2049 2019 * the spa_vdev_config_[enter/exit] functions which allow us to
2050 2020 * grab and release the spa_config_lock while still holding the namespace
2051 2021 * lock. During each step the configuration is synced out.
2052 2022 */
2053 2023 int
2054 2024 spa_vdev_remove(spa_t *spa, uint64_t guid, boolean_t unspare)
2055 2025 {
2056 2026 vdev_t *vd;
2057 2027 nvlist_t **spares, **l2cache, *nv;
2058 2028 uint64_t txg = 0;
2059 2029 uint_t nspares, nl2cache;
2060 2030 int error = 0;
2061 2031 boolean_t locked = MUTEX_HELD(&spa_namespace_lock);
2062 2032 sysevent_t *ev = NULL;
2063 2033
2064 2034 ASSERT(spa_writeable(spa));
2065 2035
2066 2036 if (!locked)
2067 2037 txg = spa_vdev_enter(spa);
2068 2038
2069 2039 ASSERT(MUTEX_HELD(&spa_namespace_lock));
2070 2040 if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
2071 2041 error = (spa_has_checkpoint(spa)) ?
2072 2042 ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT;
2073 2043
2074 2044 if (!locked)
2075 2045 return (spa_vdev_exit(spa, NULL, txg, error));
2076 2046
2077 2047 return (error);
2078 2048 }
2079 2049
2080 2050 vd = spa_lookup_by_guid(spa, guid, B_FALSE);
2081 2051
2082 2052 if (spa->spa_spares.sav_vdevs != NULL &&
2083 2053 nvlist_lookup_nvlist_array(spa->spa_spares.sav_config,
2084 2054 ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0 &&
2085 2055 (nv = spa_nvlist_lookup_by_guid(spares, nspares, guid)) != NULL) {
2086 2056 /*
2087 2057 * Only remove the hot spare if it's not currently in use
2088 2058 * in this pool.
2089 2059 */
2090 2060 if (vd == NULL || unspare) {
2091 2061 char *nvstr = fnvlist_lookup_string(nv,
2092 2062 ZPOOL_CONFIG_PATH);
2093 2063 spa_history_log_internal(spa, "vdev remove", NULL,
2094 2064 "%s vdev (%s) %s", spa_name(spa),
2095 2065 VDEV_TYPE_SPARE, nvstr);
2096 2066 if (vd == NULL)
2097 2067 vd = spa_lookup_by_guid(spa, guid, B_TRUE);
2098 2068 ev = spa_event_create(spa, vd, NULL,
2099 2069 ESC_ZFS_VDEV_REMOVE_AUX);
2100 2070 spa_vdev_remove_aux(spa->spa_spares.sav_config,
2101 2071 ZPOOL_CONFIG_SPARES, spares, nspares, nv);
2102 2072 spa_load_spares(spa);
2103 2073 spa->spa_spares.sav_sync = B_TRUE;
2104 2074 } else {
2105 2075 error = SET_ERROR(EBUSY);
2106 2076 }
2107 2077 } else if (spa->spa_l2cache.sav_vdevs != NULL &&
2108 2078 nvlist_lookup_nvlist_array(spa->spa_l2cache.sav_config,
2109 2079 ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0 &&
2110 2080 (nv = spa_nvlist_lookup_by_guid(l2cache, nl2cache, guid)) != NULL) {
2111 2081 char *nvstr = fnvlist_lookup_string(nv, ZPOOL_CONFIG_PATH);
2112 2082 spa_history_log_internal(spa, "vdev remove", NULL,
2113 2083 "%s vdev (%s) %s", spa_name(spa), VDEV_TYPE_L2CACHE, nvstr);
2114 2084 /*
2115 2085 * Cache devices can always be removed.
2116 2086 */
2117 2087 vd = spa_lookup_by_guid(spa, guid, B_TRUE);
2118 2088 ev = spa_event_create(spa, vd, NULL, ESC_ZFS_VDEV_REMOVE_AUX);
2119 2089 spa_vdev_remove_aux(spa->spa_l2cache.sav_config,
2120 2090 ZPOOL_CONFIG_L2CACHE, l2cache, nl2cache, nv);
2121 2091 spa_load_l2cache(spa);
2122 2092 spa->spa_l2cache.sav_sync = B_TRUE;
2123 2093 } else if (vd != NULL && vd->vdev_islog) {
2124 2094 ASSERT(!locked);
2125 2095 error = spa_vdev_remove_log(vd, &txg);
2126 2096 } else if (vd != NULL) {
2127 2097 ASSERT(!locked);
2128 2098 error = spa_vdev_remove_top(vd, &txg);
2129 2099 } else {
2130 2100 /*
2131 2101 * There is no vdev of any kind with the specified guid.
2132 2102 */
2133 2103 error = SET_ERROR(ENOENT);
2134 2104 }
2135 2105
2136 2106 if (!locked)
2137 2107 error = spa_vdev_exit(spa, NULL, txg, error);
2138 2108
2139 2109 if (ev != NULL) {
2140 2110 if (error != 0) {
2141 2111 spa_event_discard(ev);
2142 2112 } else {
2143 2113 spa_event_post(ev);
2144 2114 }
2145 2115 }
2146 2116
2147 2117 return (error);
2148 2118 }
2149 2119
2150 2120 int
2151 2121 spa_removal_get_stats(spa_t *spa, pool_removal_stat_t *prs)
2152 2122 {
2153 2123 prs->prs_state = spa->spa_removing_phys.sr_state;
2154 2124
2155 2125 if (prs->prs_state == DSS_NONE)
2156 2126 return (SET_ERROR(ENOENT));
2157 2127
2158 2128 prs->prs_removing_vdev = spa->spa_removing_phys.sr_removing_vdev;
2159 2129 prs->prs_start_time = spa->spa_removing_phys.sr_start_time;
2160 2130 prs->prs_end_time = spa->spa_removing_phys.sr_end_time;
2161 2131 prs->prs_to_copy = spa->spa_removing_phys.sr_to_copy;
2162 2132 prs->prs_copied = spa->spa_removing_phys.sr_copied;
2163 2133
2164 2134 if (spa->spa_vdev_removal != NULL) {
2165 2135 for (int i = 0; i < TXG_SIZE; i++) {
2166 2136 prs->prs_copied +=
2167 2137 spa->spa_vdev_removal->svr_bytes_done[i];
2168 2138 }
2169 2139 }
2170 2140
2171 2141 prs->prs_mapping_memory = 0;
2172 2142 uint64_t indirect_vdev_id =
2173 2143 spa->spa_removing_phys.sr_prev_indirect_vdev;
2174 2144 while (indirect_vdev_id != -1) {
2175 2145 vdev_t *vd = spa->spa_root_vdev->vdev_child[indirect_vdev_id];
2176 2146 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
2177 2147 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
2178 2148
2179 2149 ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
2180 2150 prs->prs_mapping_memory += vdev_indirect_mapping_size(vim);
2181 2151 indirect_vdev_id = vic->vic_prev_indirect_vdev;
2182 2152 }
2183 2153
2184 2154 return (0);
2185 2155 }
|
↓ open down ↓ |
324 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX