1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21
22 /*
23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
25 */
26
27 #include <sys/zfs_context.h>
28 #include <sys/spa_impl.h>
29 #include <sys/dmu.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/zap.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/metaslab.h>
34 #include <sys/metaslab_impl.h>
35 #include <sys/uberblock_impl.h>
36 #include <sys/txg.h>
37 #include <sys/avl.h>
38 #include <sys/bpobj.h>
39 #include <sys/dsl_pool.h>
40 #include <sys/dsl_synctask.h>
41 #include <sys/dsl_dir.h>
42 #include <sys/arc.h>
43 #include <sys/zfeature.h>
44 #include <sys/vdev_indirect_births.h>
45 #include <sys/vdev_indirect_mapping.h>
46 #include <sys/abd.h>
47 #include <sys/vdev_initialize.h>
48
49 /*
50 * This file contains the necessary logic to remove vdevs from a
51 * storage pool. Currently, the only devices that can be removed
52 * are log, cache, and spare devices; and top level vdevs from a pool
53 * w/o raidz. (Note that members of a mirror can also be removed
54 * by the detach operation.)
55 *
56 * Log vdevs are removed by evacuating them and then turning the vdev
57 * into a hole vdev while holding spa config locks.
58 *
59 * Top level vdevs are removed and converted into an indirect vdev via
60 * a multi-step process:
61 *
62 * - Disable allocations from this device (spa_vdev_remove_top).
63 *
64 * - From a new thread (spa_vdev_remove_thread), copy data from
65 * the removing vdev to a different vdev. The copy happens in open
66 * context (spa_vdev_copy_impl) and issues a sync task
67 * (vdev_mapping_sync) so the sync thread can update the partial
68 * indirect mappings in core and on disk.
69 *
70 * - If a free happens during a removal, it is freed from the
71 * removing vdev, and if it has already been copied, from the new
72 * location as well (free_from_removing_vdev).
73 *
74 * - After the removal is completed, the copy thread converts the vdev
75 * into an indirect vdev (vdev_remove_complete) before instructing
76 * the sync thread to destroy the space maps and finish the removal
77 * (spa_finish_removal).
78 */
79
80 typedef struct vdev_copy_arg {
81 metaslab_t *vca_msp;
82 uint64_t vca_outstanding_bytes;
83 kcondvar_t vca_cv;
84 kmutex_t vca_lock;
85 } vdev_copy_arg_t;
86
87 /*
88 * The maximum amount of memory we can use for outstanding i/o while
89 * doing a device removal. This determines how much i/o we can have
90 * in flight concurrently.
91 */
92 int zfs_remove_max_copy_bytes = 64 * 1024 * 1024;
93
94 /*
95 * The largest contiguous segment that we will attempt to allocate when
96 * removing a device. This can be no larger than SPA_MAXBLOCKSIZE. If
97 * there is a performance problem with attempting to allocate large blocks,
98 * consider decreasing this.
99 *
100 * Note: we will issue I/Os of up to this size. The mpt driver does not
101 * respond well to I/Os larger than 1MB, so we set this to 1MB. (When
102 * mpt processes an I/O larger than 1MB, it needs to do an allocation of
103 * 2 physically contiguous pages; if this allocation fails, mpt will drop
104 * the I/O and hang the device.)
105 */
106 int zfs_remove_max_segment = 1024 * 1024;
107
108 /*
109 * Allow a remap segment to span free chunks of at most this size. The main
110 * impact of a larger span is that we will read and write larger, more
111 * contiguous chunks, with more "unnecessary" data -- trading off bandwidth
112 * for iops. The value here was chosen to align with
113 * zfs_vdev_read_gap_limit, which is a similar concept when doing regular
114 * reads (but there's no reason it has to be the same).
115 *
116 * Additionally, a higher span will have the following relatively minor
117 * effects:
118 * - the mapping will be smaller, since one entry can cover more allocated
119 * segments
120 * - more of the fragmentation in the removing device will be preserved
121 * - we'll do larger allocations, which may fail and fall back on smaller
122 * allocations
123 */
124 int vdev_removal_max_span = 32 * 1024;
125
126 /*
127 * This is used by the test suite so that it can ensure that certain
128 * actions happen while in the middle of a removal.
129 */
130 uint64_t zfs_remove_max_bytes_pause = UINT64_MAX;
131
132 #define VDEV_REMOVAL_ZAP_OBJS "lzap"
133
134 static void spa_vdev_remove_thread(void *arg);
135
136 static void
137 spa_sync_removing_state(spa_t *spa, dmu_tx_t *tx)
138 {
139 VERIFY0(zap_update(spa->spa_dsl_pool->dp_meta_objset,
140 DMU_POOL_DIRECTORY_OBJECT,
141 DMU_POOL_REMOVING, sizeof (uint64_t),
142 sizeof (spa->spa_removing_phys) / sizeof (uint64_t),
143 &spa->spa_removing_phys, tx));
144 }
145
146 static nvlist_t *
147 spa_nvlist_lookup_by_guid(nvlist_t **nvpp, int count, uint64_t target_guid)
148 {
149 for (int i = 0; i < count; i++) {
150 uint64_t guid =
151 fnvlist_lookup_uint64(nvpp[i], ZPOOL_CONFIG_GUID);
152
153 if (guid == target_guid)
154 return (nvpp[i]);
155 }
156
157 return (NULL);
158 }
159
160 static void
161 spa_vdev_remove_aux(nvlist_t *config, char *name, nvlist_t **dev, int count,
162 nvlist_t *dev_to_remove)
163 {
164 nvlist_t **newdev = NULL;
165
166 if (count > 1)
167 newdev = kmem_alloc((count - 1) * sizeof (void *), KM_SLEEP);
168
169 for (int i = 0, j = 0; i < count; i++) {
170 if (dev[i] == dev_to_remove)
171 continue;
172 VERIFY(nvlist_dup(dev[i], &newdev[j++], KM_SLEEP) == 0);
173 }
174
175 VERIFY(nvlist_remove(config, name, DATA_TYPE_NVLIST_ARRAY) == 0);
176 VERIFY(nvlist_add_nvlist_array(config, name, newdev, count - 1) == 0);
177
178 for (int i = 0; i < count - 1; i++)
179 nvlist_free(newdev[i]);
180
181 if (count > 1)
182 kmem_free(newdev, (count - 1) * sizeof (void *));
183 }
184
185 static spa_vdev_removal_t *
186 spa_vdev_removal_create(vdev_t *vd)
187 {
188 spa_vdev_removal_t *svr = kmem_zalloc(sizeof (*svr), KM_SLEEP);
189 mutex_init(&svr->svr_lock, NULL, MUTEX_DEFAULT, NULL);
190 cv_init(&svr->svr_cv, NULL, CV_DEFAULT, NULL);
191 svr->svr_allocd_segs = range_tree_create(NULL, NULL);
192 svr->svr_vdev_id = vd->vdev_id;
193
194 for (int i = 0; i < TXG_SIZE; i++) {
195 svr->svr_frees[i] = range_tree_create(NULL, NULL);
196 list_create(&svr->svr_new_segments[i],
197 sizeof (vdev_indirect_mapping_entry_t),
198 offsetof(vdev_indirect_mapping_entry_t, vime_node));
199 }
200
201 return (svr);
202 }
203
204 void
205 spa_vdev_removal_destroy(spa_vdev_removal_t *svr)
206 {
207 for (int i = 0; i < TXG_SIZE; i++) {
208 ASSERT0(svr->svr_bytes_done[i]);
209 ASSERT0(svr->svr_max_offset_to_sync[i]);
210 range_tree_destroy(svr->svr_frees[i]);
211 list_destroy(&svr->svr_new_segments[i]);
212 }
213
214 range_tree_destroy(svr->svr_allocd_segs);
215 mutex_destroy(&svr->svr_lock);
216 cv_destroy(&svr->svr_cv);
217 kmem_free(svr, sizeof (*svr));
218 }
219
220 /*
221 * This is called as a synctask in the txg in which we will mark this vdev
222 * as removing (in the config stored in the MOS).
223 *
224 * It begins the evacuation of a toplevel vdev by:
225 * - initializing the spa_removing_phys which tracks this removal
226 * - computing the amount of space to remove for accounting purposes
227 * - dirtying all dbufs in the spa_config_object
228 * - creating the spa_vdev_removal
229 * - starting the spa_vdev_remove_thread
230 */
231 static void
232 vdev_remove_initiate_sync(void *arg, dmu_tx_t *tx)
233 {
234 int vdev_id = (uintptr_t)arg;
235 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
236 vdev_t *vd = vdev_lookup_top(spa, vdev_id);
237 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
238 objset_t *mos = spa->spa_dsl_pool->dp_meta_objset;
239 spa_vdev_removal_t *svr = NULL;
240 uint64_t txg = dmu_tx_get_txg(tx);
241
242 ASSERT3P(vd->vdev_ops, !=, &vdev_raidz_ops);
243 svr = spa_vdev_removal_create(vd);
244
245 ASSERT(vd->vdev_removing);
246 ASSERT3P(vd->vdev_indirect_mapping, ==, NULL);
247
248 spa_feature_incr(spa, SPA_FEATURE_DEVICE_REMOVAL, tx);
249 if (spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS)) {
250 /*
251 * By activating the OBSOLETE_COUNTS feature, we prevent
252 * the pool from being downgraded and ensure that the
253 * refcounts are precise.
254 */
255 spa_feature_incr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
256 uint64_t one = 1;
257 VERIFY0(zap_add(spa->spa_meta_objset, vd->vdev_top_zap,
258 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE, sizeof (one), 1,
259 &one, tx));
260 ASSERT3U(vdev_obsolete_counts_are_precise(vd), !=, 0);
261 }
262
263 vic->vic_mapping_object = vdev_indirect_mapping_alloc(mos, tx);
264 vd->vdev_indirect_mapping =
265 vdev_indirect_mapping_open(mos, vic->vic_mapping_object);
266 vic->vic_births_object = vdev_indirect_births_alloc(mos, tx);
267 vd->vdev_indirect_births =
268 vdev_indirect_births_open(mos, vic->vic_births_object);
269 spa->spa_removing_phys.sr_removing_vdev = vd->vdev_id;
270 spa->spa_removing_phys.sr_start_time = gethrestime_sec();
271 spa->spa_removing_phys.sr_end_time = 0;
272 spa->spa_removing_phys.sr_state = DSS_SCANNING;
273 spa->spa_removing_phys.sr_to_copy = 0;
274 spa->spa_removing_phys.sr_copied = 0;
275
276 /*
277 * Note: We can't use vdev_stat's vs_alloc for sr_to_copy, because
278 * there may be space in the defer tree, which is free, but still
279 * counted in vs_alloc.
280 */
281 for (uint64_t i = 0; i < vd->vdev_ms_count; i++) {
282 metaslab_t *ms = vd->vdev_ms[i];
283 if (ms->ms_sm == NULL)
284 continue;
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 spa->spa_removing_phys.sr_to_copy +=
294 space_map_allocated(ms->ms_sm);
295
296 /*
297 * Space which we are freeing this txg does not need to
298 * be copied.
299 */
300 spa->spa_removing_phys.sr_to_copy -=
301 range_tree_space(ms->ms_freeing);
302
303 ASSERT0(range_tree_space(ms->ms_freed));
304 for (int t = 0; t < TXG_SIZE; t++)
305 ASSERT0(range_tree_space(ms->ms_allocating[t]));
306 }
307
308 /*
309 * Sync tasks are called before metaslab_sync(), so there should
310 * be no already-synced metaslabs in the TXG_CLEAN list.
311 */
312 ASSERT3P(txg_list_head(&vd->vdev_ms_list, TXG_CLEAN(txg)), ==, NULL);
313
314 spa_sync_removing_state(spa, tx);
315
316 /*
317 * All blocks that we need to read the most recent mapping must be
318 * stored on concrete vdevs. Therefore, we must dirty anything that
319 * is read before spa_remove_init(). Specifically, the
320 * spa_config_object. (Note that although we already modified the
321 * spa_config_object in spa_sync_removing_state, that may not have
322 * modified all blocks of the object.)
323 */
324 dmu_object_info_t doi;
325 VERIFY0(dmu_object_info(mos, DMU_POOL_DIRECTORY_OBJECT, &doi));
326 for (uint64_t offset = 0; offset < doi.doi_max_offset; ) {
327 dmu_buf_t *dbuf;
328 VERIFY0(dmu_buf_hold(mos, DMU_POOL_DIRECTORY_OBJECT,
329 offset, FTAG, &dbuf, 0));
330 dmu_buf_will_dirty(dbuf, tx);
331 offset += dbuf->db_size;
332 dmu_buf_rele(dbuf, FTAG);
333 }
334
335 /*
336 * Now that we've allocated the im_object, dirty the vdev to ensure
337 * that the object gets written to the config on disk.
338 */
339 vdev_config_dirty(vd);
340
341 zfs_dbgmsg("starting removal thread for vdev %llu (%p) in txg %llu "
342 "im_obj=%llu", vd->vdev_id, vd, dmu_tx_get_txg(tx),
343 vic->vic_mapping_object);
344
345 spa_history_log_internal(spa, "vdev remove started", tx,
346 "%s vdev %llu %s", spa_name(spa), vd->vdev_id,
347 (vd->vdev_path != NULL) ? vd->vdev_path : "-");
348 /*
349 * Setting spa_vdev_removal causes subsequent frees to call
350 * free_from_removing_vdev(). Note that we don't need any locking
351 * because we are the sync thread, and metaslab_free_impl() is only
352 * called from syncing context (potentially from a zio taskq thread,
353 * but in any case only when there are outstanding free i/os, which
354 * there are not).
355 */
356 ASSERT3P(spa->spa_vdev_removal, ==, NULL);
357 spa->spa_vdev_removal = svr;
358 svr->svr_thread = thread_create(NULL, 0,
359 spa_vdev_remove_thread, spa, 0, &p0, TS_RUN, minclsyspri);
360 }
361
362 /*
363 * When we are opening a pool, we must read the mapping for each
364 * indirect vdev in order from most recently removed to least
365 * recently removed. We do this because the blocks for the mapping
366 * of older indirect vdevs may be stored on more recently removed vdevs.
367 * In order to read each indirect mapping object, we must have
368 * initialized all more recently removed vdevs.
369 */
370 int
371 spa_remove_init(spa_t *spa)
372 {
373 int error;
374
375 error = zap_lookup(spa->spa_dsl_pool->dp_meta_objset,
376 DMU_POOL_DIRECTORY_OBJECT,
377 DMU_POOL_REMOVING, sizeof (uint64_t),
378 sizeof (spa->spa_removing_phys) / sizeof (uint64_t),
379 &spa->spa_removing_phys);
380
381 if (error == ENOENT) {
382 spa->spa_removing_phys.sr_state = DSS_NONE;
383 spa->spa_removing_phys.sr_removing_vdev = -1;
384 spa->spa_removing_phys.sr_prev_indirect_vdev = -1;
385 spa->spa_indirect_vdevs_loaded = B_TRUE;
386 return (0);
387 } else if (error != 0) {
388 return (error);
389 }
390
391 if (spa->spa_removing_phys.sr_state == DSS_SCANNING) {
392 /*
393 * We are currently removing a vdev. Create and
394 * initialize a spa_vdev_removal_t from the bonus
395 * buffer of the removing vdevs vdev_im_object, and
396 * initialize its partial mapping.
397 */
398 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
399 vdev_t *vd = vdev_lookup_top(spa,
400 spa->spa_removing_phys.sr_removing_vdev);
401
402 if (vd == NULL) {
403 spa_config_exit(spa, SCL_STATE, FTAG);
404 return (EINVAL);
405 }
406
407 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
408
409 ASSERT(vdev_is_concrete(vd));
410 spa_vdev_removal_t *svr = spa_vdev_removal_create(vd);
411 ASSERT3U(svr->svr_vdev_id, ==, vd->vdev_id);
412 ASSERT(vd->vdev_removing);
413
414 vd->vdev_indirect_mapping = vdev_indirect_mapping_open(
415 spa->spa_meta_objset, vic->vic_mapping_object);
416 vd->vdev_indirect_births = vdev_indirect_births_open(
417 spa->spa_meta_objset, vic->vic_births_object);
418 spa_config_exit(spa, SCL_STATE, FTAG);
419
420 spa->spa_vdev_removal = svr;
421 }
422
423 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
424 uint64_t indirect_vdev_id =
425 spa->spa_removing_phys.sr_prev_indirect_vdev;
426 while (indirect_vdev_id != UINT64_MAX) {
427 vdev_t *vd = vdev_lookup_top(spa, indirect_vdev_id);
428 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
429
430 ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
431 vd->vdev_indirect_mapping = vdev_indirect_mapping_open(
432 spa->spa_meta_objset, vic->vic_mapping_object);
433 vd->vdev_indirect_births = vdev_indirect_births_open(
434 spa->spa_meta_objset, vic->vic_births_object);
435
436 indirect_vdev_id = vic->vic_prev_indirect_vdev;
437 }
438 spa_config_exit(spa, SCL_STATE, FTAG);
439
440 /*
441 * Now that we've loaded all the indirect mappings, we can allow
442 * reads from other blocks (e.g. via predictive prefetch).
443 */
444 spa->spa_indirect_vdevs_loaded = B_TRUE;
445 return (0);
446 }
447
448 void
449 spa_restart_removal(spa_t *spa)
450 {
451 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
452
453 if (svr == NULL)
454 return;
455
456 /*
457 * In general when this function is called there is no
458 * removal thread running. The only scenario where this
459 * is not true is during spa_import() where this function
460 * is called twice [once from spa_import_impl() and
461 * spa_async_resume()]. Thus, in the scenario where we
462 * import a pool that has an ongoing removal we don't
463 * want to spawn a second thread.
464 */
465 if (svr->svr_thread != NULL)
466 return;
467
468 if (!spa_writeable(spa))
469 return;
470
471 zfs_dbgmsg("restarting removal of %llu", svr->svr_vdev_id);
472 svr->svr_thread = thread_create(NULL, 0, spa_vdev_remove_thread, spa,
473 0, &p0, TS_RUN, minclsyspri);
474 }
475
476 /*
477 * Process freeing from a device which is in the middle of being removed.
478 * We must handle this carefully so that we attempt to copy freed data,
479 * and we correctly free already-copied data.
480 */
481 void
482 free_from_removing_vdev(vdev_t *vd, uint64_t offset, uint64_t size)
483 {
484 spa_t *spa = vd->vdev_spa;
485 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
486 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
487 uint64_t txg = spa_syncing_txg(spa);
488 uint64_t max_offset_yet = 0;
489
490 ASSERT(vd->vdev_indirect_config.vic_mapping_object != 0);
491 ASSERT3U(vd->vdev_indirect_config.vic_mapping_object, ==,
492 vdev_indirect_mapping_object(vim));
493 ASSERT3U(vd->vdev_id, ==, svr->svr_vdev_id);
494
495 mutex_enter(&svr->svr_lock);
496
497 /*
498 * Remove the segment from the removing vdev's spacemap. This
499 * ensures that we will not attempt to copy this space (if the
500 * removal thread has not yet visited it), and also ensures
501 * that we know what is actually allocated on the new vdevs
502 * (needed if we cancel the removal).
503 *
504 * Note: we must do the metaslab_free_concrete() with the svr_lock
505 * held, so that the remove_thread can not load this metaslab and then
506 * visit this offset between the time that we metaslab_free_concrete()
507 * and when we check to see if it has been visited.
508 *
509 * Note: The checkpoint flag is set to false as having/taking
510 * a checkpoint and removing a device can't happen at the same
511 * time.
512 */
513 ASSERT(!spa_has_checkpoint(spa));
514 metaslab_free_concrete(vd, offset, size, B_FALSE);
515
516 uint64_t synced_size = 0;
517 uint64_t synced_offset = 0;
518 uint64_t max_offset_synced = vdev_indirect_mapping_max_offset(vim);
519 if (offset < max_offset_synced) {
520 /*
521 * The mapping for this offset is already on disk.
522 * Free from the new location.
523 *
524 * Note that we use svr_max_synced_offset because it is
525 * updated atomically with respect to the in-core mapping.
526 * By contrast, vim_max_offset is not.
527 *
528 * This block may be split between a synced entry and an
529 * in-flight or unvisited entry. Only process the synced
530 * portion of it here.
531 */
532 synced_size = MIN(size, max_offset_synced - offset);
533 synced_offset = offset;
534
535 ASSERT3U(max_offset_yet, <=, max_offset_synced);
536 max_offset_yet = max_offset_synced;
537
538 DTRACE_PROBE3(remove__free__synced,
539 spa_t *, spa,
540 uint64_t, offset,
541 uint64_t, synced_size);
542
543 size -= synced_size;
544 offset += synced_size;
545 }
546
547 /*
548 * Look at all in-flight txgs starting from the currently syncing one
549 * and see if a section of this free is being copied. By starting from
550 * this txg and iterating forward, we might find that this region
551 * was copied in two different txgs and handle it appropriately.
552 */
553 for (int i = 0; i < TXG_CONCURRENT_STATES; i++) {
554 int txgoff = (txg + i) & TXG_MASK;
555 if (size > 0 && offset < svr->svr_max_offset_to_sync[txgoff]) {
556 /*
557 * The mapping for this offset is in flight, and
558 * will be synced in txg+i.
559 */
560 uint64_t inflight_size = MIN(size,
561 svr->svr_max_offset_to_sync[txgoff] - offset);
562
563 DTRACE_PROBE4(remove__free__inflight,
564 spa_t *, spa,
565 uint64_t, offset,
566 uint64_t, inflight_size,
567 uint64_t, txg + i);
568
569 /*
570 * We copy data in order of increasing offset.
571 * Therefore the max_offset_to_sync[] must increase
572 * (or be zero, indicating that nothing is being
573 * copied in that txg).
574 */
575 if (svr->svr_max_offset_to_sync[txgoff] != 0) {
576 ASSERT3U(svr->svr_max_offset_to_sync[txgoff],
577 >=, max_offset_yet);
578 max_offset_yet =
579 svr->svr_max_offset_to_sync[txgoff];
580 }
581
582 /*
583 * We've already committed to copying this segment:
584 * we have allocated space elsewhere in the pool for
585 * it and have an IO outstanding to copy the data. We
586 * cannot free the space before the copy has
587 * completed, or else the copy IO might overwrite any
588 * new data. To free that space, we record the
589 * segment in the appropriate svr_frees tree and free
590 * the mapped space later, in the txg where we have
591 * completed the copy and synced the mapping (see
592 * vdev_mapping_sync).
593 */
594 range_tree_add(svr->svr_frees[txgoff],
595 offset, inflight_size);
596 size -= inflight_size;
597 offset += inflight_size;
598
599 /*
600 * This space is already accounted for as being
601 * done, because it is being copied in txg+i.
602 * However, if i!=0, then it is being copied in
603 * a future txg. If we crash after this txg
604 * syncs but before txg+i syncs, then the space
605 * will be free. Therefore we must account
606 * for the space being done in *this* txg
607 * (when it is freed) rather than the future txg
608 * (when it will be copied).
609 */
610 ASSERT3U(svr->svr_bytes_done[txgoff], >=,
611 inflight_size);
612 svr->svr_bytes_done[txgoff] -= inflight_size;
613 svr->svr_bytes_done[txg & TXG_MASK] += inflight_size;
614 }
615 }
616 ASSERT0(svr->svr_max_offset_to_sync[TXG_CLEAN(txg) & TXG_MASK]);
617
618 if (size > 0) {
619 /*
620 * The copy thread has not yet visited this offset. Ensure
621 * that it doesn't.
622 */
623
624 DTRACE_PROBE3(remove__free__unvisited,
625 spa_t *, spa,
626 uint64_t, offset,
627 uint64_t, size);
628
629 if (svr->svr_allocd_segs != NULL)
630 range_tree_clear(svr->svr_allocd_segs, offset, size);
631
632 /*
633 * Since we now do not need to copy this data, for
634 * accounting purposes we have done our job and can count
635 * it as completed.
636 */
637 svr->svr_bytes_done[txg & TXG_MASK] += size;
638 }
639 mutex_exit(&svr->svr_lock);
640
641 /*
642 * Now that we have dropped svr_lock, process the synced portion
643 * of this free.
644 */
645 if (synced_size > 0) {
646 vdev_indirect_mark_obsolete(vd, synced_offset, synced_size);
647
648 /*
649 * Note: this can only be called from syncing context,
650 * and the vdev_indirect_mapping is only changed from the
651 * sync thread, so we don't need svr_lock while doing
652 * metaslab_free_impl_cb.
653 */
654 boolean_t checkpoint = B_FALSE;
655 vdev_indirect_ops.vdev_op_remap(vd, synced_offset, synced_size,
656 metaslab_free_impl_cb, &checkpoint);
657 }
658 }
659
660 /*
661 * Stop an active removal and update the spa_removing phys.
662 */
663 static void
664 spa_finish_removal(spa_t *spa, dsl_scan_state_t state, dmu_tx_t *tx)
665 {
666 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
667 ASSERT3U(dmu_tx_get_txg(tx), ==, spa_syncing_txg(spa));
668
669 /* Ensure the removal thread has completed before we free the svr. */
670 spa_vdev_remove_suspend(spa);
671
672 ASSERT(state == DSS_FINISHED || state == DSS_CANCELED);
673
674 if (state == DSS_FINISHED) {
675 spa_removing_phys_t *srp = &spa->spa_removing_phys;
676 vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
677 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
678
679 if (srp->sr_prev_indirect_vdev != UINT64_MAX) {
680 vdev_t *pvd = vdev_lookup_top(spa,
681 srp->sr_prev_indirect_vdev);
682 ASSERT3P(pvd->vdev_ops, ==, &vdev_indirect_ops);
683 }
684
685 vic->vic_prev_indirect_vdev = srp->sr_prev_indirect_vdev;
686 srp->sr_prev_indirect_vdev = vd->vdev_id;
687 }
688 spa->spa_removing_phys.sr_state = state;
689 spa->spa_removing_phys.sr_end_time = gethrestime_sec();
690
691 spa->spa_vdev_removal = NULL;
692 spa_vdev_removal_destroy(svr);
693
694 spa_sync_removing_state(spa, tx);
695
696 vdev_config_dirty(spa->spa_root_vdev);
697 }
698
699 static void
700 free_mapped_segment_cb(void *arg, uint64_t offset, uint64_t size)
701 {
702 vdev_t *vd = arg;
703 vdev_indirect_mark_obsolete(vd, offset, size);
704 boolean_t checkpoint = B_FALSE;
705 vdev_indirect_ops.vdev_op_remap(vd, offset, size,
706 metaslab_free_impl_cb, &checkpoint);
707 }
708
709 /*
710 * On behalf of the removal thread, syncs an incremental bit more of
711 * the indirect mapping to disk and updates the in-memory mapping.
712 * Called as a sync task in every txg that the removal thread makes progress.
713 */
714 static void
715 vdev_mapping_sync(void *arg, dmu_tx_t *tx)
716 {
717 spa_vdev_removal_t *svr = arg;
718 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
719 vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
720 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
721 uint64_t txg = dmu_tx_get_txg(tx);
722 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
723
724 ASSERT(vic->vic_mapping_object != 0);
725 ASSERT3U(txg, ==, spa_syncing_txg(spa));
726
727 vdev_indirect_mapping_add_entries(vim,
728 &svr->svr_new_segments[txg & TXG_MASK], tx);
729 vdev_indirect_births_add_entry(vd->vdev_indirect_births,
730 vdev_indirect_mapping_max_offset(vim), dmu_tx_get_txg(tx), tx);
731
732 /*
733 * Free the copied data for anything that was freed while the
734 * mapping entries were in flight.
735 */
736 mutex_enter(&svr->svr_lock);
737 range_tree_vacate(svr->svr_frees[txg & TXG_MASK],
738 free_mapped_segment_cb, vd);
739 ASSERT3U(svr->svr_max_offset_to_sync[txg & TXG_MASK], >=,
740 vdev_indirect_mapping_max_offset(vim));
741 svr->svr_max_offset_to_sync[txg & TXG_MASK] = 0;
742 mutex_exit(&svr->svr_lock);
743
744 spa_sync_removing_state(spa, tx);
745 }
746
747 typedef struct vdev_copy_segment_arg {
748 spa_t *vcsa_spa;
749 dva_t *vcsa_dest_dva;
750 uint64_t vcsa_txg;
751 range_tree_t *vcsa_obsolete_segs;
752 } vdev_copy_segment_arg_t;
753
754 static void
755 unalloc_seg(void *arg, uint64_t start, uint64_t size)
756 {
757 vdev_copy_segment_arg_t *vcsa = arg;
758 spa_t *spa = vcsa->vcsa_spa;
759 blkptr_t bp = { 0 };
760
761 BP_SET_BIRTH(&bp, TXG_INITIAL, TXG_INITIAL);
762 BP_SET_LSIZE(&bp, size);
763 BP_SET_PSIZE(&bp, size);
764 BP_SET_COMPRESS(&bp, ZIO_COMPRESS_OFF);
765 BP_SET_CHECKSUM(&bp, ZIO_CHECKSUM_OFF);
766 BP_SET_TYPE(&bp, DMU_OT_NONE);
767 BP_SET_LEVEL(&bp, 0);
768 BP_SET_DEDUP(&bp, 0);
769 BP_SET_BYTEORDER(&bp, ZFS_HOST_BYTEORDER);
770
771 DVA_SET_VDEV(&bp.blk_dva[0], DVA_GET_VDEV(vcsa->vcsa_dest_dva));
772 DVA_SET_OFFSET(&bp.blk_dva[0],
773 DVA_GET_OFFSET(vcsa->vcsa_dest_dva) + start);
774 DVA_SET_ASIZE(&bp.blk_dva[0], size);
775
776 zio_free(spa, vcsa->vcsa_txg, &bp);
777 }
778
779 /*
780 * All reads and writes associated with a call to spa_vdev_copy_segment()
781 * are done.
782 */
783 static void
784 spa_vdev_copy_segment_done(zio_t *zio)
785 {
786 vdev_copy_segment_arg_t *vcsa = zio->io_private;
787
788 range_tree_vacate(vcsa->vcsa_obsolete_segs,
789 unalloc_seg, vcsa);
790 range_tree_destroy(vcsa->vcsa_obsolete_segs);
791 kmem_free(vcsa, sizeof (*vcsa));
792
793 spa_config_exit(zio->io_spa, SCL_STATE, zio->io_spa);
794 }
795
796 /*
797 * The write of the new location is done.
798 */
799 static void
800 spa_vdev_copy_segment_write_done(zio_t *zio)
801 {
802 vdev_copy_arg_t *vca = zio->io_private;
803
804 abd_free(zio->io_abd);
805
806 mutex_enter(&vca->vca_lock);
807 vca->vca_outstanding_bytes -= zio->io_size;
808 cv_signal(&vca->vca_cv);
809 mutex_exit(&vca->vca_lock);
810 }
811
812 /*
813 * The read of the old location is done. The parent zio is the write to
814 * the new location. Allow it to start.
815 */
816 static void
817 spa_vdev_copy_segment_read_done(zio_t *zio)
818 {
819 zio_nowait(zio_unique_parent(zio));
820 }
821
822 /*
823 * If the old and new vdevs are mirrors, we will read both sides of the old
824 * mirror, and write each copy to the corresponding side of the new mirror.
825 * If the old and new vdevs have a different number of children, we will do
826 * this as best as possible. Since we aren't verifying checksums, this
827 * ensures that as long as there's a good copy of the data, we'll have a
828 * good copy after the removal, even if there's silent damage to one side
829 * of the mirror. If we're removing a mirror that has some silent damage,
830 * we'll have exactly the same damage in the new location (assuming that
831 * the new location is also a mirror).
832 *
833 * We accomplish this by creating a tree of zio_t's, with as many writes as
834 * there are "children" of the new vdev (a non-redundant vdev counts as one
835 * child, a 2-way mirror has 2 children, etc). Each write has an associated
836 * read from a child of the old vdev. Typically there will be the same
837 * number of children of the old and new vdevs. However, if there are more
838 * children of the new vdev, some child(ren) of the old vdev will be issued
839 * multiple reads. If there are more children of the old vdev, some copies
840 * will be dropped.
841 *
842 * For example, the tree of zio_t's for a 2-way mirror is:
843 *
844 * null
845 * / \
846 * write(new vdev, child 0) write(new vdev, child 1)
847 * | |
848 * read(old vdev, child 0) read(old vdev, child 1)
849 *
850 * Child zio's complete before their parents complete. However, zio's
851 * created with zio_vdev_child_io() may be issued before their children
852 * complete. In this case we need to make sure that the children (reads)
853 * complete before the parents (writes) are *issued*. We do this by not
854 * calling zio_nowait() on each write until its corresponding read has
855 * completed.
856 *
857 * The spa_config_lock must be held while zio's created by
858 * zio_vdev_child_io() are in progress, to ensure that the vdev tree does
859 * not change (e.g. due to a concurrent "zpool attach/detach"). The "null"
860 * zio is needed to release the spa_config_lock after all the reads and
861 * writes complete. (Note that we can't grab the config lock for each read,
862 * because it is not reentrant - we could deadlock with a thread waiting
863 * for a write lock.)
864 */
865 static void
866 spa_vdev_copy_one_child(vdev_copy_arg_t *vca, zio_t *nzio,
867 vdev_t *source_vd, uint64_t source_offset,
868 vdev_t *dest_child_vd, uint64_t dest_offset, int dest_id, uint64_t size)
869 {
870 ASSERT3U(spa_config_held(nzio->io_spa, SCL_ALL, RW_READER), !=, 0);
871
872 mutex_enter(&vca->vca_lock);
873 vca->vca_outstanding_bytes += size;
874 mutex_exit(&vca->vca_lock);
875
876 abd_t *abd = abd_alloc_for_io(size, B_FALSE);
877
878 vdev_t *source_child_vd;
879 if (source_vd->vdev_ops == &vdev_mirror_ops && dest_id != -1) {
880 /*
881 * Source and dest are both mirrors. Copy from the same
882 * child id as we are copying to (wrapping around if there
883 * are more dest children than source children).
884 */
885 source_child_vd =
886 source_vd->vdev_child[dest_id % source_vd->vdev_children];
887 } else {
888 source_child_vd = source_vd;
889 }
890
891 zio_t *write_zio = zio_vdev_child_io(nzio, NULL,
892 dest_child_vd, dest_offset, abd, size,
893 ZIO_TYPE_WRITE, ZIO_PRIORITY_REMOVAL,
894 ZIO_FLAG_CANFAIL,
895 spa_vdev_copy_segment_write_done, vca);
896
897 zio_nowait(zio_vdev_child_io(write_zio, NULL,
898 source_child_vd, source_offset, abd, size,
899 ZIO_TYPE_READ, ZIO_PRIORITY_REMOVAL,
900 ZIO_FLAG_CANFAIL,
901 spa_vdev_copy_segment_read_done, vca));
902 }
903
904 /*
905 * Allocate a new location for this segment, and create the zio_t's to
906 * read from the old location and write to the new location.
907 */
908 static int
909 spa_vdev_copy_segment(vdev_t *vd, range_tree_t *segs,
910 uint64_t maxalloc, uint64_t txg,
911 vdev_copy_arg_t *vca, zio_alloc_list_t *zal)
912 {
913 metaslab_group_t *mg = vd->vdev_mg;
914 spa_t *spa = vd->vdev_spa;
915 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
916 vdev_indirect_mapping_entry_t *entry;
917 dva_t dst = { 0 };
918 uint64_t start = range_tree_min(segs);
919
920 ASSERT3U(maxalloc, <=, SPA_MAXBLOCKSIZE);
921
922 uint64_t size = range_tree_span(segs);
923 if (range_tree_span(segs) > maxalloc) {
924 /*
925 * We can't allocate all the segments. Prefer to end
926 * the allocation at the end of a segment, thus avoiding
927 * additional split blocks.
928 */
929 range_seg_t search;
930 avl_index_t where;
931 search.rs_start = start + maxalloc;
932 search.rs_end = search.rs_start;
933 range_seg_t *rs = avl_find(&segs->rt_root, &search, &where);
934 if (rs == NULL) {
935 rs = avl_nearest(&segs->rt_root, where, AVL_BEFORE);
936 } else {
937 rs = AVL_PREV(&segs->rt_root, rs);
938 }
939 if (rs != NULL) {
940 size = rs->rs_end - start;
941 } else {
942 /*
943 * There are no segments that end before maxalloc.
944 * I.e. the first segment is larger than maxalloc,
945 * so we must split it.
946 */
947 size = maxalloc;
948 }
949 }
950 ASSERT3U(size, <=, maxalloc);
951
952 /*
953 * An allocation class might not have any remaining vdevs or space
954 */
955 metaslab_class_t *mc = mg->mg_class;
956 if (mc != spa_normal_class(spa) && mc->mc_groups <= 1)
957 mc = spa_normal_class(spa);
958 int error = metaslab_alloc_dva(spa, mc, size, &dst, 0, NULL, txg, 0,
959 zal, 0);
960 if (error == ENOSPC && mc != spa_normal_class(spa)) {
961 error = metaslab_alloc_dva(spa, spa_normal_class(spa), size,
962 &dst, 0, NULL, txg, 0, zal, 0);
963 }
964 if (error != 0)
965 return (error);
966
967 /*
968 * Determine the ranges that are not actually needed. Offsets are
969 * relative to the start of the range to be copied (i.e. relative to the
970 * local variable "start").
971 */
972 range_tree_t *obsolete_segs = range_tree_create(NULL, NULL);
973
974 range_seg_t *rs = avl_first(&segs->rt_root);
975 ASSERT3U(rs->rs_start, ==, start);
976 uint64_t prev_seg_end = rs->rs_end;
977 while ((rs = AVL_NEXT(&segs->rt_root, rs)) != NULL) {
978 if (rs->rs_start >= start + size) {
979 break;
980 } else {
981 range_tree_add(obsolete_segs,
982 prev_seg_end - start,
983 rs->rs_start - prev_seg_end);
984 }
985 prev_seg_end = rs->rs_end;
986 }
987 /* We don't end in the middle of an obsolete range */
988 ASSERT3U(start + size, <=, prev_seg_end);
989
990 range_tree_clear(segs, start, size);
991
992 /*
993 * We can't have any padding of the allocated size, otherwise we will
994 * misunderstand what's allocated, and the size of the mapping.
995 * The caller ensures this will be true by passing in a size that is
996 * aligned to the worst (highest) ashift in the pool.
997 */
998 ASSERT3U(DVA_GET_ASIZE(&dst), ==, size);
999
1000 entry = kmem_zalloc(sizeof (vdev_indirect_mapping_entry_t), KM_SLEEP);
1001 DVA_MAPPING_SET_SRC_OFFSET(&entry->vime_mapping, start);
1002 entry->vime_mapping.vimep_dst = dst;
1003 if (spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS)) {
1004 entry->vime_obsolete_count = range_tree_space(obsolete_segs);
1005 }
1006
1007 vdev_copy_segment_arg_t *vcsa = kmem_zalloc(sizeof (*vcsa), KM_SLEEP);
1008 vcsa->vcsa_dest_dva = &entry->vime_mapping.vimep_dst;
1009 vcsa->vcsa_obsolete_segs = obsolete_segs;
1010 vcsa->vcsa_spa = spa;
1011 vcsa->vcsa_txg = txg;
1012
1013 /*
1014 * See comment before spa_vdev_copy_one_child().
1015 */
1016 spa_config_enter(spa, SCL_STATE, spa, RW_READER);
1017 zio_t *nzio = zio_null(spa->spa_txg_zio[txg & TXG_MASK], spa, NULL,
1018 spa_vdev_copy_segment_done, vcsa, 0);
1019 vdev_t *dest_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dst));
1020 if (dest_vd->vdev_ops == &vdev_mirror_ops) {
1021 for (int i = 0; i < dest_vd->vdev_children; i++) {
1022 vdev_t *child = dest_vd->vdev_child[i];
1023 spa_vdev_copy_one_child(vca, nzio, vd, start,
1024 child, DVA_GET_OFFSET(&dst), i, size);
1025 }
1026 } else {
1027 spa_vdev_copy_one_child(vca, nzio, vd, start,
1028 dest_vd, DVA_GET_OFFSET(&dst), -1, size);
1029 }
1030 zio_nowait(nzio);
1031
1032 list_insert_tail(&svr->svr_new_segments[txg & TXG_MASK], entry);
1033 ASSERT3U(start + size, <=, vd->vdev_ms_count << vd->vdev_ms_shift);
1034 vdev_dirty(vd, 0, NULL, txg);
1035
1036 return (0);
1037 }
1038
1039 /*
1040 * Complete the removal of a toplevel vdev. This is called as a
1041 * synctask in the same txg that we will sync out the new config (to the
1042 * MOS object) which indicates that this vdev is indirect.
1043 */
1044 static void
1045 vdev_remove_complete_sync(void *arg, dmu_tx_t *tx)
1046 {
1047 spa_vdev_removal_t *svr = arg;
1048 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
1049 vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
1050
1051 ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
1052
1053 for (int i = 0; i < TXG_SIZE; i++) {
1054 ASSERT0(svr->svr_bytes_done[i]);
1055 }
1056
1057 ASSERT3U(spa->spa_removing_phys.sr_copied, ==,
1058 spa->spa_removing_phys.sr_to_copy);
1059
1060 vdev_destroy_spacemaps(vd, tx);
1061
1062 /* destroy leaf zaps, if any */
1063 ASSERT3P(svr->svr_zaplist, !=, NULL);
1064 for (nvpair_t *pair = nvlist_next_nvpair(svr->svr_zaplist, NULL);
1065 pair != NULL;
1066 pair = nvlist_next_nvpair(svr->svr_zaplist, pair)) {
1067 vdev_destroy_unlink_zap(vd, fnvpair_value_uint64(pair), tx);
1068 }
1069 fnvlist_free(svr->svr_zaplist);
1070
1071 spa_finish_removal(dmu_tx_pool(tx)->dp_spa, DSS_FINISHED, tx);
1072 /* vd->vdev_path is not available here */
1073 spa_history_log_internal(spa, "vdev remove completed", tx,
1074 "%s vdev %llu", spa_name(spa), vd->vdev_id);
1075 }
1076
1077 static void
1078 vdev_remove_enlist_zaps(vdev_t *vd, nvlist_t *zlist)
1079 {
1080 ASSERT3P(zlist, !=, NULL);
1081 ASSERT3P(vd->vdev_ops, !=, &vdev_raidz_ops);
1082
1083 if (vd->vdev_leaf_zap != 0) {
1084 char zkey[32];
1085 (void) snprintf(zkey, sizeof (zkey), "%s-%"PRIu64,
1086 VDEV_REMOVAL_ZAP_OBJS, vd->vdev_leaf_zap);
1087 fnvlist_add_uint64(zlist, zkey, vd->vdev_leaf_zap);
1088 }
1089
1090 for (uint64_t id = 0; id < vd->vdev_children; id++) {
1091 vdev_remove_enlist_zaps(vd->vdev_child[id], zlist);
1092 }
1093 }
1094
1095 static void
1096 vdev_remove_replace_with_indirect(vdev_t *vd, uint64_t txg)
1097 {
1098 vdev_t *ivd;
1099 dmu_tx_t *tx;
1100 spa_t *spa = vd->vdev_spa;
1101 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1102
1103 /*
1104 * First, build a list of leaf zaps to be destroyed.
1105 * This is passed to the sync context thread,
1106 * which does the actual unlinking.
1107 */
1108 svr->svr_zaplist = fnvlist_alloc();
1109 vdev_remove_enlist_zaps(vd, svr->svr_zaplist);
1110
1111 ivd = vdev_add_parent(vd, &vdev_indirect_ops);
1112 ivd->vdev_removing = 0;
1113
1114 vd->vdev_leaf_zap = 0;
1115
1116 vdev_remove_child(ivd, vd);
1117 vdev_compact_children(ivd);
1118
1119 ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
1120
1121 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1122 dsl_sync_task_nowait(spa->spa_dsl_pool, vdev_remove_complete_sync, svr,
1123 0, ZFS_SPACE_CHECK_NONE, tx);
1124 dmu_tx_commit(tx);
1125
1126 /*
1127 * Indicate that this thread has exited.
1128 * After this, we can not use svr.
1129 */
1130 mutex_enter(&svr->svr_lock);
1131 svr->svr_thread = NULL;
1132 cv_broadcast(&svr->svr_cv);
1133 mutex_exit(&svr->svr_lock);
1134 }
1135
1136 /*
1137 * Complete the removal of a toplevel vdev. This is called in open
1138 * context by the removal thread after we have copied all vdev's data.
1139 */
1140 static void
1141 vdev_remove_complete(spa_t *spa)
1142 {
1143 uint64_t txg;
1144
1145 /*
1146 * Wait for any deferred frees to be synced before we call
1147 * vdev_metaslab_fini()
1148 */
1149 txg_wait_synced(spa->spa_dsl_pool, 0);
1150 txg = spa_vdev_enter(spa);
1151 vdev_t *vd = vdev_lookup_top(spa, spa->spa_vdev_removal->svr_vdev_id);
1152 ASSERT3P(vd->vdev_initialize_thread, ==, NULL);
1153
1154 sysevent_t *ev = spa_event_create(spa, vd, NULL,
1155 ESC_ZFS_VDEV_REMOVE_DEV);
1156
1157 zfs_dbgmsg("finishing device removal for vdev %llu in txg %llu",
1158 vd->vdev_id, txg);
1159
1160 /*
1161 * Discard allocation state.
1162 */
1163 if (vd->vdev_mg != NULL) {
1164 vdev_metaslab_fini(vd);
1165 metaslab_group_destroy(vd->vdev_mg);
1166 vd->vdev_mg = NULL;
1167 }
1168 ASSERT0(vd->vdev_stat.vs_space);
1169 ASSERT0(vd->vdev_stat.vs_dspace);
1170
1171 vdev_remove_replace_with_indirect(vd, txg);
1172
1173 /*
1174 * We now release the locks, allowing spa_sync to run and finish the
1175 * removal via vdev_remove_complete_sync in syncing context.
1176 *
1177 * Note that we hold on to the vdev_t that has been replaced. Since
1178 * it isn't part of the vdev tree any longer, it can't be concurrently
1179 * manipulated, even while we don't have the config lock.
1180 */
1181 (void) spa_vdev_exit(spa, NULL, txg, 0);
1182
1183 /*
1184 * Top ZAP should have been transferred to the indirect vdev in
1185 * vdev_remove_replace_with_indirect.
1186 */
1187 ASSERT0(vd->vdev_top_zap);
1188
1189 /*
1190 * Leaf ZAP should have been moved in vdev_remove_replace_with_indirect.
1191 */
1192 ASSERT0(vd->vdev_leaf_zap);
1193
1194 txg = spa_vdev_enter(spa);
1195 (void) vdev_label_init(vd, 0, VDEV_LABEL_REMOVE);
1196 /*
1197 * Request to update the config and the config cachefile.
1198 */
1199 vdev_config_dirty(spa->spa_root_vdev);
1200 (void) spa_vdev_exit(spa, vd, txg, 0);
1201
1202 spa_event_post(ev);
1203 }
1204
1205 /*
1206 * Evacuates a segment of size at most max_alloc from the vdev
1207 * via repeated calls to spa_vdev_copy_segment. If an allocation
1208 * fails, the pool is probably too fragmented to handle such a
1209 * large size, so decrease max_alloc so that the caller will not try
1210 * this size again this txg.
1211 */
1212 static void
1213 spa_vdev_copy_impl(vdev_t *vd, spa_vdev_removal_t *svr, vdev_copy_arg_t *vca,
1214 uint64_t *max_alloc, dmu_tx_t *tx)
1215 {
1216 uint64_t txg = dmu_tx_get_txg(tx);
1217 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
1218
1219 mutex_enter(&svr->svr_lock);
1220
1221 /*
1222 * Determine how big of a chunk to copy. We can allocate up
1223 * to max_alloc bytes, and we can span up to vdev_removal_max_span
1224 * bytes of unallocated space at a time. "segs" will track the
1225 * allocated segments that we are copying. We may also be copying
1226 * free segments (of up to vdev_removal_max_span bytes).
1227 */
1228 range_tree_t *segs = range_tree_create(NULL, NULL);
1229 for (;;) {
1230 range_seg_t *rs = avl_first(&svr->svr_allocd_segs->rt_root);
1231 if (rs == NULL)
1232 break;
1233
1234 uint64_t seg_length;
1235
1236 if (range_tree_is_empty(segs)) {
1237 /* need to truncate the first seg based on max_alloc */
1238 seg_length =
1239 MIN(rs->rs_end - rs->rs_start, *max_alloc);
1240 } else {
1241 if (rs->rs_start - range_tree_max(segs) >
1242 vdev_removal_max_span) {
1243 /*
1244 * Including this segment would cause us to
1245 * copy a larger unneeded chunk than is allowed.
1246 */
1247 break;
1248 } else if (rs->rs_end - range_tree_min(segs) >
1249 *max_alloc) {
1250 /*
1251 * This additional segment would extend past
1252 * max_alloc. Rather than splitting this
1253 * segment, leave it for the next mapping.
1254 */
1255 break;
1256 } else {
1257 seg_length = rs->rs_end - rs->rs_start;
1258 }
1259 }
1260
1261 range_tree_add(segs, rs->rs_start, seg_length);
1262 range_tree_remove(svr->svr_allocd_segs,
1263 rs->rs_start, seg_length);
1264 }
1265
1266 if (range_tree_is_empty(segs)) {
1267 mutex_exit(&svr->svr_lock);
1268 range_tree_destroy(segs);
1269 return;
1270 }
1271
1272 if (svr->svr_max_offset_to_sync[txg & TXG_MASK] == 0) {
1273 dsl_sync_task_nowait(dmu_tx_pool(tx), vdev_mapping_sync,
1274 svr, 0, ZFS_SPACE_CHECK_NONE, tx);
1275 }
1276
1277 svr->svr_max_offset_to_sync[txg & TXG_MASK] = range_tree_max(segs);
1278
1279 /*
1280 * Note: this is the amount of *allocated* space
1281 * that we are taking care of each txg.
1282 */
1283 svr->svr_bytes_done[txg & TXG_MASK] += range_tree_space(segs);
1284
1285 mutex_exit(&svr->svr_lock);
1286
1287 zio_alloc_list_t zal;
1288 metaslab_trace_init(&zal);
1289 uint64_t thismax = SPA_MAXBLOCKSIZE;
1290 while (!range_tree_is_empty(segs)) {
1291 int error = spa_vdev_copy_segment(vd,
1292 segs, thismax, txg, vca, &zal);
1293
1294 if (error == ENOSPC) {
1295 /*
1296 * Cut our segment in half, and don't try this
1297 * segment size again this txg. Note that the
1298 * allocation size must be aligned to the highest
1299 * ashift in the pool, so that the allocation will
1300 * not be padded out to a multiple of the ashift,
1301 * which could cause us to think that this mapping
1302 * is larger than we intended.
1303 */
1304 ASSERT3U(spa->spa_max_ashift, >=, SPA_MINBLOCKSHIFT);
1305 ASSERT3U(spa->spa_max_ashift, ==, spa->spa_min_ashift);
1306 uint64_t attempted =
1307 MIN(range_tree_span(segs), thismax);
1308 thismax = P2ROUNDUP(attempted / 2,
1309 1 << spa->spa_max_ashift);
1310 /*
1311 * The minimum-size allocation can not fail.
1312 */
1313 ASSERT3U(attempted, >, 1 << spa->spa_max_ashift);
1314 *max_alloc = attempted - (1 << spa->spa_max_ashift);
1315 } else {
1316 ASSERT0(error);
1317
1318 /*
1319 * We've performed an allocation, so reset the
1320 * alloc trace list.
1321 */
1322 metaslab_trace_fini(&zal);
1323 metaslab_trace_init(&zal);
1324 }
1325 }
1326 metaslab_trace_fini(&zal);
1327 range_tree_destroy(segs);
1328 }
1329
1330 /*
1331 * The removal thread operates in open context. It iterates over all
1332 * allocated space in the vdev, by loading each metaslab's spacemap.
1333 * For each contiguous segment of allocated space (capping the segment
1334 * size at SPA_MAXBLOCKSIZE), we:
1335 * - Allocate space for it on another vdev.
1336 * - Create a new mapping from the old location to the new location
1337 * (as a record in svr_new_segments).
1338 * - Initiate a logical read zio to get the data off the removing disk.
1339 * - In the read zio's done callback, initiate a logical write zio to
1340 * write it to the new vdev.
1341 * Note that all of this will take effect when a particular TXG syncs.
1342 * The sync thread ensures that all the phys reads and writes for the syncing
1343 * TXG have completed (see spa_txg_zio) and writes the new mappings to disk
1344 * (see vdev_mapping_sync()).
1345 */
1346 static void
1347 spa_vdev_remove_thread(void *arg)
1348 {
1349 spa_t *spa = arg;
1350 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1351 vdev_copy_arg_t vca;
1352 uint64_t max_alloc = zfs_remove_max_segment;
1353 uint64_t last_txg = 0;
1354
1355 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
1356 vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
1357 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
1358 uint64_t start_offset = vdev_indirect_mapping_max_offset(vim);
1359
1360 ASSERT3P(vd->vdev_ops, !=, &vdev_indirect_ops);
1361 ASSERT(vdev_is_concrete(vd));
1362 ASSERT(vd->vdev_removing);
1363 ASSERT(vd->vdev_indirect_config.vic_mapping_object != 0);
1364 ASSERT(vim != NULL);
1365
1366 mutex_init(&vca.vca_lock, NULL, MUTEX_DEFAULT, NULL);
1367 cv_init(&vca.vca_cv, NULL, CV_DEFAULT, NULL);
1368 vca.vca_outstanding_bytes = 0;
1369
1370 mutex_enter(&svr->svr_lock);
1371
1372 /*
1373 * Start from vim_max_offset so we pick up where we left off
1374 * if we are restarting the removal after opening the pool.
1375 */
1376 uint64_t msi;
1377 for (msi = start_offset >> vd->vdev_ms_shift;
1378 msi < vd->vdev_ms_count && !svr->svr_thread_exit; msi++) {
1379 metaslab_t *msp = vd->vdev_ms[msi];
1380 ASSERT3U(msi, <=, vd->vdev_ms_count);
1381
1382 ASSERT0(range_tree_space(svr->svr_allocd_segs));
1383
1384 mutex_enter(&msp->ms_sync_lock);
1385 mutex_enter(&msp->ms_lock);
1386
1387 /*
1388 * Assert nothing in flight -- ms_*tree is empty.
1389 */
1390 for (int i = 0; i < TXG_SIZE; i++) {
1391 ASSERT0(range_tree_space(msp->ms_allocating[i]));
1392 }
1393
1394 /*
1395 * If the metaslab has ever been allocated from (ms_sm!=NULL),
1396 * read the allocated segments from the space map object
1397 * into svr_allocd_segs. Since we do this while holding
1398 * svr_lock and ms_sync_lock, concurrent frees (which
1399 * would have modified the space map) will wait for us
1400 * to finish loading the spacemap, and then take the
1401 * appropriate action (see free_from_removing_vdev()).
1402 */
1403 if (msp->ms_sm != NULL) {
1404 space_map_t *sm = NULL;
1405
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 range_tree_walk(msp->ms_freeing,
1422 range_tree_remove, svr->svr_allocd_segs);
1423
1424 /*
1425 * When we are resuming from a paused removal (i.e.
1426 * when importing a pool with a removal in progress),
1427 * discard any state that we have already processed.
1428 */
1429 range_tree_clear(svr->svr_allocd_segs, 0, start_offset);
1430 }
1431 mutex_exit(&msp->ms_lock);
1432 mutex_exit(&msp->ms_sync_lock);
1433
1434 vca.vca_msp = msp;
1435 zfs_dbgmsg("copying %llu segments for metaslab %llu",
1436 avl_numnodes(&svr->svr_allocd_segs->rt_root),
1437 msp->ms_id);
1438
1439 while (!svr->svr_thread_exit &&
1440 !range_tree_is_empty(svr->svr_allocd_segs)) {
1441
1442 mutex_exit(&svr->svr_lock);
1443
1444 /*
1445 * We need to periodically drop the config lock so that
1446 * writers can get in. Additionally, we can't wait
1447 * for a txg to sync while holding a config lock
1448 * (since a waiting writer could cause a 3-way deadlock
1449 * with the sync thread, which also gets a config
1450 * lock for reader). So we can't hold the config lock
1451 * while calling dmu_tx_assign().
1452 */
1453 spa_config_exit(spa, SCL_CONFIG, FTAG);
1454
1455 /*
1456 * This delay will pause the removal around the point
1457 * specified by zfs_remove_max_bytes_pause. We do this
1458 * solely from the test suite or during debugging.
1459 */
1460 uint64_t bytes_copied =
1461 spa->spa_removing_phys.sr_copied;
1462 for (int i = 0; i < TXG_SIZE; i++)
1463 bytes_copied += svr->svr_bytes_done[i];
1464 while (zfs_remove_max_bytes_pause <= bytes_copied &&
1465 !svr->svr_thread_exit)
1466 delay(hz);
1467
1468 mutex_enter(&vca.vca_lock);
1469 while (vca.vca_outstanding_bytes >
1470 zfs_remove_max_copy_bytes) {
1471 cv_wait(&vca.vca_cv, &vca.vca_lock);
1472 }
1473 mutex_exit(&vca.vca_lock);
1474
1475 dmu_tx_t *tx =
1476 dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
1477
1478 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
1479 uint64_t txg = dmu_tx_get_txg(tx);
1480
1481 /*
1482 * Reacquire the vdev_config lock. The vdev_t
1483 * that we're removing may have changed, e.g. due
1484 * to a vdev_attach or vdev_detach.
1485 */
1486 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
1487 vd = vdev_lookup_top(spa, svr->svr_vdev_id);
1488
1489 if (txg != last_txg)
1490 max_alloc = zfs_remove_max_segment;
1491 last_txg = txg;
1492
1493 spa_vdev_copy_impl(vd, svr, &vca, &max_alloc, tx);
1494
1495 dmu_tx_commit(tx);
1496 mutex_enter(&svr->svr_lock);
1497 }
1498 }
1499
1500 mutex_exit(&svr->svr_lock);
1501
1502 spa_config_exit(spa, SCL_CONFIG, FTAG);
1503
1504 /*
1505 * Wait for all copies to finish before cleaning up the vca.
1506 */
1507 txg_wait_synced(spa->spa_dsl_pool, 0);
1508 ASSERT0(vca.vca_outstanding_bytes);
1509
1510 mutex_destroy(&vca.vca_lock);
1511 cv_destroy(&vca.vca_cv);
1512
1513 if (svr->svr_thread_exit) {
1514 mutex_enter(&svr->svr_lock);
1515 range_tree_vacate(svr->svr_allocd_segs, NULL, NULL);
1516 svr->svr_thread = NULL;
1517 cv_broadcast(&svr->svr_cv);
1518 mutex_exit(&svr->svr_lock);
1519 } else {
1520 ASSERT0(range_tree_space(svr->svr_allocd_segs));
1521 vdev_remove_complete(spa);
1522 }
1523 }
1524
1525 void
1526 spa_vdev_remove_suspend(spa_t *spa)
1527 {
1528 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1529
1530 if (svr == NULL)
1531 return;
1532
1533 mutex_enter(&svr->svr_lock);
1534 svr->svr_thread_exit = B_TRUE;
1535 while (svr->svr_thread != NULL)
1536 cv_wait(&svr->svr_cv, &svr->svr_lock);
1537 svr->svr_thread_exit = B_FALSE;
1538 mutex_exit(&svr->svr_lock);
1539 }
1540
1541 /* ARGSUSED */
1542 static int
1543 spa_vdev_remove_cancel_check(void *arg, dmu_tx_t *tx)
1544 {
1545 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
1546
1547 if (spa->spa_vdev_removal == NULL)
1548 return (ENOTACTIVE);
1549 return (0);
1550 }
1551
1552 /*
1553 * Cancel a removal by freeing all entries from the partial mapping
1554 * and marking the vdev as no longer being removing.
1555 */
1556 /* ARGSUSED */
1557 static void
1558 spa_vdev_remove_cancel_sync(void *arg, dmu_tx_t *tx)
1559 {
1560 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
1561 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1562 vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
1563 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
1564 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
1565 objset_t *mos = spa->spa_meta_objset;
1566
1567 ASSERT3P(svr->svr_thread, ==, NULL);
1568
1569 spa_feature_decr(spa, SPA_FEATURE_DEVICE_REMOVAL, tx);
1570 if (vdev_obsolete_counts_are_precise(vd)) {
1571 spa_feature_decr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
1572 VERIFY0(zap_remove(spa->spa_meta_objset, vd->vdev_top_zap,
1573 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE, tx));
1574 }
1575
1576 if (vdev_obsolete_sm_object(vd) != 0) {
1577 ASSERT(vd->vdev_obsolete_sm != NULL);
1578 ASSERT3U(vdev_obsolete_sm_object(vd), ==,
1579 space_map_object(vd->vdev_obsolete_sm));
1580
1581 space_map_free(vd->vdev_obsolete_sm, tx);
1582 VERIFY0(zap_remove(spa->spa_meta_objset, vd->vdev_top_zap,
1583 VDEV_TOP_ZAP_INDIRECT_OBSOLETE_SM, tx));
1584 space_map_close(vd->vdev_obsolete_sm);
1585 vd->vdev_obsolete_sm = NULL;
1586 spa_feature_decr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
1587 }
1588 for (int i = 0; i < TXG_SIZE; i++) {
1589 ASSERT(list_is_empty(&svr->svr_new_segments[i]));
1590 ASSERT3U(svr->svr_max_offset_to_sync[i], <=,
1591 vdev_indirect_mapping_max_offset(vim));
1592 }
1593
1594 for (uint64_t msi = 0; msi < vd->vdev_ms_count; msi++) {
1595 metaslab_t *msp = vd->vdev_ms[msi];
1596
1597 if (msp->ms_start >= vdev_indirect_mapping_max_offset(vim))
1598 break;
1599
1600 ASSERT0(range_tree_space(svr->svr_allocd_segs));
1601
1602 mutex_enter(&msp->ms_lock);
1603
1604 /*
1605 * Assert nothing in flight -- ms_*tree is empty.
1606 */
1607 for (int i = 0; i < TXG_SIZE; i++)
1608 ASSERT0(range_tree_space(msp->ms_allocating[i]));
1609 for (int i = 0; i < TXG_DEFER_SIZE; i++)
1610 ASSERT0(range_tree_space(msp->ms_defer[i]));
1611 ASSERT0(range_tree_space(msp->ms_freed));
1612
1613 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 mutex_enter(&svr->svr_lock);
1625 VERIFY0(space_map_load(msp->ms_sm,
1626 svr->svr_allocd_segs, SM_ALLOC));
1627 range_tree_walk(msp->ms_freeing,
1628 range_tree_remove, svr->svr_allocd_segs);
1629
1630 /*
1631 * Clear everything past what has been synced,
1632 * because we have not allocated mappings for it yet.
1633 */
1634 uint64_t syncd = vdev_indirect_mapping_max_offset(vim);
1635 uint64_t sm_end = msp->ms_sm->sm_start +
1636 msp->ms_sm->sm_size;
1637 if (sm_end > syncd)
1638 range_tree_clear(svr->svr_allocd_segs,
1639 syncd, sm_end - syncd);
1640
1641 mutex_exit(&svr->svr_lock);
1642 }
1643 mutex_exit(&msp->ms_lock);
1644
1645 mutex_enter(&svr->svr_lock);
1646 range_tree_vacate(svr->svr_allocd_segs,
1647 free_mapped_segment_cb, vd);
1648 mutex_exit(&svr->svr_lock);
1649 }
1650
1651 /*
1652 * Note: this must happen after we invoke free_mapped_segment_cb,
1653 * because it adds to the obsolete_segments.
1654 */
1655 range_tree_vacate(vd->vdev_obsolete_segments, NULL, NULL);
1656
1657 ASSERT3U(vic->vic_mapping_object, ==,
1658 vdev_indirect_mapping_object(vd->vdev_indirect_mapping));
1659 vdev_indirect_mapping_close(vd->vdev_indirect_mapping);
1660 vd->vdev_indirect_mapping = NULL;
1661 vdev_indirect_mapping_free(mos, vic->vic_mapping_object, tx);
1662 vic->vic_mapping_object = 0;
1663
1664 ASSERT3U(vic->vic_births_object, ==,
1665 vdev_indirect_births_object(vd->vdev_indirect_births));
1666 vdev_indirect_births_close(vd->vdev_indirect_births);
1667 vd->vdev_indirect_births = NULL;
1668 vdev_indirect_births_free(mos, vic->vic_births_object, tx);
1669 vic->vic_births_object = 0;
1670
1671 /*
1672 * We may have processed some frees from the removing vdev in this
1673 * txg, thus increasing svr_bytes_done; discard that here to
1674 * satisfy the assertions in spa_vdev_removal_destroy().
1675 * Note that future txg's can not have any bytes_done, because
1676 * future TXG's are only modified from open context, and we have
1677 * already shut down the copying thread.
1678 */
1679 svr->svr_bytes_done[dmu_tx_get_txg(tx) & TXG_MASK] = 0;
1680 spa_finish_removal(spa, DSS_CANCELED, tx);
1681
1682 vd->vdev_removing = B_FALSE;
1683 vdev_config_dirty(vd);
1684
1685 zfs_dbgmsg("canceled device removal for vdev %llu in %llu",
1686 vd->vdev_id, dmu_tx_get_txg(tx));
1687 spa_history_log_internal(spa, "vdev remove canceled", tx,
1688 "%s vdev %llu %s", spa_name(spa),
1689 vd->vdev_id, (vd->vdev_path != NULL) ? vd->vdev_path : "-");
1690 }
1691
1692 int
1693 spa_vdev_remove_cancel(spa_t *spa)
1694 {
1695 spa_vdev_remove_suspend(spa);
1696
1697 if (spa->spa_vdev_removal == NULL)
1698 return (ENOTACTIVE);
1699
1700 uint64_t vdid = spa->spa_vdev_removal->svr_vdev_id;
1701
1702 int error = dsl_sync_task(spa->spa_name, spa_vdev_remove_cancel_check,
1703 spa_vdev_remove_cancel_sync, NULL, 0,
1704 ZFS_SPACE_CHECK_EXTRA_RESERVED);
1705
1706 if (error == 0) {
1707 spa_config_enter(spa, SCL_ALLOC | SCL_VDEV, FTAG, RW_WRITER);
1708 vdev_t *vd = vdev_lookup_top(spa, vdid);
1709 metaslab_group_activate(vd->vdev_mg);
1710 spa_config_exit(spa, SCL_ALLOC | SCL_VDEV, FTAG);
1711 }
1712
1713 return (error);
1714 }
1715
1716 /*
1717 * Called every sync pass of every txg if there's a svr.
1718 */
1719 void
1720 svr_sync(spa_t *spa, dmu_tx_t *tx)
1721 {
1722 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1723 int txgoff = dmu_tx_get_txg(tx) & TXG_MASK;
1724
1725 /*
1726 * This check is necessary so that we do not dirty the
1727 * DIRECTORY_OBJECT via spa_sync_removing_state() when there
1728 * is nothing to do. Dirtying it every time would prevent us
1729 * from syncing-to-convergence.
1730 */
1731 if (svr->svr_bytes_done[txgoff] == 0)
1732 return;
1733
1734 /*
1735 * Update progress accounting.
1736 */
1737 spa->spa_removing_phys.sr_copied += svr->svr_bytes_done[txgoff];
1738 svr->svr_bytes_done[txgoff] = 0;
1739
1740 spa_sync_removing_state(spa, tx);
1741 }
1742
1743 static void
1744 vdev_remove_make_hole_and_free(vdev_t *vd)
1745 {
1746 uint64_t id = vd->vdev_id;
1747 spa_t *spa = vd->vdev_spa;
1748 vdev_t *rvd = spa->spa_root_vdev;
1749 boolean_t last_vdev = (id == (rvd->vdev_children - 1));
1750
1751 ASSERT(MUTEX_HELD(&spa_namespace_lock));
1752 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1753
1754 vdev_free(vd);
1755
1756 if (last_vdev) {
1757 vdev_compact_children(rvd);
1758 } else {
1759 vd = vdev_alloc_common(spa, id, 0, &vdev_hole_ops);
1760 vdev_add_child(rvd, vd);
1761 }
1762 vdev_config_dirty(rvd);
1763
1764 /*
1765 * Reassess the health of our root vdev.
1766 */
1767 vdev_reopen(rvd);
1768 }
1769
1770 /*
1771 * Remove a log device. The config lock is held for the specified TXG.
1772 */
1773 static int
1774 spa_vdev_remove_log(vdev_t *vd, uint64_t *txg)
1775 {
1776 metaslab_group_t *mg = vd->vdev_mg;
1777 spa_t *spa = vd->vdev_spa;
1778 int error = 0;
1779
1780 ASSERT(vd->vdev_islog);
1781 ASSERT(vd == vd->vdev_top);
1782
1783 /*
1784 * Stop allocating from this vdev.
1785 */
1786 metaslab_group_passivate(mg);
1787
1788 /*
1789 * Wait for the youngest allocations and frees to sync,
1790 * and then wait for the deferral of those frees to finish.
1791 */
1792 spa_vdev_config_exit(spa, NULL,
1793 *txg + TXG_CONCURRENT_STATES + TXG_DEFER_SIZE, 0, FTAG);
1794
1795 /*
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
1798 * spa_namespace_lock held. Once this completes the device
1799 * should no longer have any blocks allocated on it.
1800 */
1801 if (vd->vdev_islog) {
1802 if (vd->vdev_stat.vs_alloc != 0)
1803 error = spa_reset_logs(spa);
1804 }
1805
1806 *txg = spa_vdev_config_enter(spa);
1807
1808 if (error != 0) {
1809 metaslab_group_activate(mg);
1810 return (error);
1811 }
1812 ASSERT0(vd->vdev_stat.vs_alloc);
1813
1814 /*
1815 * The evacuation succeeded. Remove any remaining MOS metadata
1816 * associated with this vdev, and wait for these changes to sync.
1817 */
1818 vd->vdev_removing = B_TRUE;
1819
1820 vdev_dirty_leaves(vd, VDD_DTL, *txg);
1821 vdev_config_dirty(vd);
1822
1823 spa_history_log_internal(spa, "vdev remove", NULL,
1824 "%s vdev %llu (log) %s", spa_name(spa), vd->vdev_id,
1825 (vd->vdev_path != NULL) ? vd->vdev_path : "-");
1826
1827 /* Make sure these changes are sync'ed */
1828 spa_vdev_config_exit(spa, NULL, *txg, 0, FTAG);
1829
1830 /* Stop initializing */
1831 (void) vdev_initialize_stop_all(vd, VDEV_INITIALIZE_CANCELED);
1832
1833 *txg = spa_vdev_config_enter(spa);
1834
1835 sysevent_t *ev = spa_event_create(spa, vd, NULL,
1836 ESC_ZFS_VDEV_REMOVE_DEV);
1837 ASSERT(MUTEX_HELD(&spa_namespace_lock));
1838 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1839
1840 /* The top ZAP should have been destroyed by vdev_remove_empty. */
1841 ASSERT0(vd->vdev_top_zap);
1842 /* The leaf ZAP should have been destroyed by vdev_dtl_sync. */
1843 ASSERT0(vd->vdev_leaf_zap);
1844
1845 (void) vdev_label_init(vd, 0, VDEV_LABEL_REMOVE);
1846
1847 if (list_link_active(&vd->vdev_state_dirty_node))
1848 vdev_state_clean(vd);
1849 if (list_link_active(&vd->vdev_config_dirty_node))
1850 vdev_config_clean(vd);
1851
1852 /*
1853 * Clean up the vdev namespace.
1854 */
1855 vdev_remove_make_hole_and_free(vd);
1856
1857 if (ev != NULL)
1858 spa_event_post(ev);
1859
1860 return (0);
1861 }
1862
1863 static int
1864 spa_vdev_remove_top_check(vdev_t *vd)
1865 {
1866 spa_t *spa = vd->vdev_spa;
1867
1868 if (vd != vd->vdev_top)
1869 return (SET_ERROR(ENOTSUP));
1870
1871 if (!spa_feature_is_enabled(spa, SPA_FEATURE_DEVICE_REMOVAL))
1872 return (SET_ERROR(ENOTSUP));
1873
1874 /* available space in the pool's normal class */
1875 uint64_t available = dsl_dir_space_available(
1876 spa->spa_dsl_pool->dp_root_dir, NULL, 0, B_TRUE);
1877
1878 metaslab_class_t *mc = vd->vdev_mg->mg_class;
1879
1880 /*
1881 * When removing a vdev from an allocation class that has
1882 * remaining vdevs, include available space from the class.
1883 */
1884 if (mc != spa_normal_class(spa) && mc->mc_groups > 1) {
1885 uint64_t class_avail = metaslab_class_get_space(mc) -
1886 metaslab_class_get_alloc(mc);
1887
1888 /* add class space, adjusted for overhead */
1889 available += (class_avail * 94) / 100;
1890 }
1891
1892 /*
1893 * There has to be enough free space to remove the
1894 * device and leave double the "slop" space (i.e. we
1895 * must leave at least 3% of the pool free, in addition to
1896 * the normal slop space).
1897 */
1898 if (available < vd->vdev_stat.vs_dspace + spa_get_slop_space(spa)) {
1899 return (SET_ERROR(ENOSPC));
1900 }
1901
1902 /*
1903 * There can not be a removal in progress.
1904 */
1905 if (spa->spa_removing_phys.sr_state == DSS_SCANNING)
1906 return (SET_ERROR(EBUSY));
1907
1908 /*
1909 * The device must have all its data.
1910 */
1911 if (!vdev_dtl_empty(vd, DTL_MISSING) ||
1912 !vdev_dtl_empty(vd, DTL_OUTAGE))
1913 return (SET_ERROR(EBUSY));
1914
1915 /*
1916 * The device must be healthy.
1917 */
1918 if (!vdev_readable(vd))
1919 return (SET_ERROR(EIO));
1920
1921 /*
1922 * All vdevs in normal class must have the same ashift.
1923 */
1924 if (spa->spa_max_ashift != spa->spa_min_ashift) {
1925 return (SET_ERROR(EINVAL));
1926 }
1927
1928 /*
1929 * All vdevs in normal class must have the same ashift
1930 * and not be raidz.
1931 */
1932 vdev_t *rvd = spa->spa_root_vdev;
1933 int num_indirect = 0;
1934 for (uint64_t id = 0; id < rvd->vdev_children; id++) {
1935 vdev_t *cvd = rvd->vdev_child[id];
1936 if (cvd->vdev_ashift != 0 && !cvd->vdev_islog)
1937 ASSERT3U(cvd->vdev_ashift, ==, spa->spa_max_ashift);
1938 if (cvd->vdev_ops == &vdev_indirect_ops)
1939 num_indirect++;
1940 if (!vdev_is_concrete(cvd))
1941 continue;
1942 if (cvd->vdev_ops == &vdev_raidz_ops)
1943 return (SET_ERROR(EINVAL));
1944 /*
1945 * Need the mirror to be mirror of leaf vdevs only
1946 */
1947 if (cvd->vdev_ops == &vdev_mirror_ops) {
1948 for (uint64_t cid = 0;
1949 cid < cvd->vdev_children; cid++) {
1950 vdev_t *tmp = cvd->vdev_child[cid];
1951 if (!tmp->vdev_ops->vdev_op_leaf)
1952 return (SET_ERROR(EINVAL));
1953 }
1954 }
1955 }
1956
1957 return (0);
1958 }
1959
1960 /*
1961 * Initiate removal of a top-level vdev, reducing the total space in the pool.
1962 * The config lock is held for the specified TXG. Once initiated,
1963 * evacuation of all allocated space (copying it to other vdevs) happens
1964 * in the background (see spa_vdev_remove_thread()), and can be canceled
1965 * (see spa_vdev_remove_cancel()). If successful, the vdev will
1966 * be transformed to an indirect vdev (see spa_vdev_remove_complete()).
1967 */
1968 static int
1969 spa_vdev_remove_top(vdev_t *vd, uint64_t *txg)
1970 {
1971 spa_t *spa = vd->vdev_spa;
1972 int error;
1973
1974 /*
1975 * Check for errors up-front, so that we don't waste time
1976 * passivating the metaslab group and clearing the ZIL if there
1977 * are errors.
1978 */
1979 error = spa_vdev_remove_top_check(vd);
1980 if (error != 0)
1981 return (error);
1982
1983 /*
1984 * Stop allocating from this vdev. Note that we must check
1985 * that this is not the only device in the pool before
1986 * passivating, otherwise we will not be able to make
1987 * progress because we can't allocate from any vdevs.
1988 * The above check for sufficient free space serves this
1989 * purpose.
1990 */
1991 metaslab_group_t *mg = vd->vdev_mg;
1992 metaslab_group_passivate(mg);
1993
1994 /*
1995 * Wait for the youngest allocations and frees to sync,
1996 * and then wait for the deferral of those frees to finish.
1997 */
1998 spa_vdev_config_exit(spa, NULL,
1999 *txg + TXG_CONCURRENT_STATES + TXG_DEFER_SIZE, 0, FTAG);
2000
2001 /*
2002 * We must ensure that no "stubby" log blocks are allocated
2003 * on the device to be removed. These blocks could be
2004 * written at any time, including while we are in the middle
2005 * of copying them.
2006 */
2007 error = spa_reset_logs(spa);
2008
2009 /*
2010 * We stop any initializing that is currently in progress but leave
2011 * the state as "active". This will allow the initializing to resume
2012 * if the removal is canceled sometime later.
2013 */
2014 vdev_initialize_stop_all(vd, VDEV_INITIALIZE_ACTIVE);
2015
2016 *txg = spa_vdev_config_enter(spa);
2017
2018 /*
2019 * Things might have changed while the config lock was dropped
2020 * (e.g. space usage). Check for errors again.
2021 */
2022 if (error == 0)
2023 error = spa_vdev_remove_top_check(vd);
2024
2025 if (error != 0) {
2026 metaslab_group_activate(mg);
2027 spa_async_request(spa, SPA_ASYNC_INITIALIZE_RESTART);
2028 return (error);
2029 }
2030
2031 vd->vdev_removing = B_TRUE;
2032
2033 vdev_dirty_leaves(vd, VDD_DTL, *txg);
2034 vdev_config_dirty(vd);
2035 dmu_tx_t *tx = dmu_tx_create_assigned(spa->spa_dsl_pool, *txg);
2036 dsl_sync_task_nowait(spa->spa_dsl_pool,
2037 vdev_remove_initiate_sync,
2038 (void *)(uintptr_t)vd->vdev_id, 0, ZFS_SPACE_CHECK_NONE, tx);
2039 dmu_tx_commit(tx);
2040
2041 return (0);
2042 }
2043
2044 /*
2045 * Remove a device from the pool.
2046 *
2047 * Removing a device from the vdev namespace requires several steps
2048 * and can take a significant amount of time. As a result we use
2049 * the spa_vdev_config_[enter/exit] functions which allow us to
2050 * grab and release the spa_config_lock while still holding the namespace
2051 * lock. During each step the configuration is synced out.
2052 */
2053 int
2054 spa_vdev_remove(spa_t *spa, uint64_t guid, boolean_t unspare)
2055 {
2056 vdev_t *vd;
2057 nvlist_t **spares, **l2cache, *nv;
2058 uint64_t txg = 0;
2059 uint_t nspares, nl2cache;
2060 int error = 0;
2061 boolean_t locked = MUTEX_HELD(&spa_namespace_lock);
2062 sysevent_t *ev = NULL;
2063
2064 ASSERT(spa_writeable(spa));
2065
2066 if (!locked)
2067 txg = spa_vdev_enter(spa);
2068
2069 ASSERT(MUTEX_HELD(&spa_namespace_lock));
2070 if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
2071 error = (spa_has_checkpoint(spa)) ?
2072 ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT;
2073
2074 if (!locked)
2075 return (spa_vdev_exit(spa, NULL, txg, error));
2076
2077 return (error);
2078 }
2079
2080 vd = spa_lookup_by_guid(spa, guid, B_FALSE);
2081
2082 if (spa->spa_spares.sav_vdevs != NULL &&
2083 nvlist_lookup_nvlist_array(spa->spa_spares.sav_config,
2084 ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0 &&
2085 (nv = spa_nvlist_lookup_by_guid(spares, nspares, guid)) != NULL) {
2086 /*
2087 * Only remove the hot spare if it's not currently in use
2088 * in this pool.
2089 */
2090 if (vd == NULL || unspare) {
2091 char *nvstr = fnvlist_lookup_string(nv,
2092 ZPOOL_CONFIG_PATH);
2093 spa_history_log_internal(spa, "vdev remove", NULL,
2094 "%s vdev (%s) %s", spa_name(spa),
2095 VDEV_TYPE_SPARE, nvstr);
2096 if (vd == NULL)
2097 vd = spa_lookup_by_guid(spa, guid, B_TRUE);
2098 ev = spa_event_create(spa, vd, NULL,
2099 ESC_ZFS_VDEV_REMOVE_AUX);
2100 spa_vdev_remove_aux(spa->spa_spares.sav_config,
2101 ZPOOL_CONFIG_SPARES, spares, nspares, nv);
2102 spa_load_spares(spa);
2103 spa->spa_spares.sav_sync = B_TRUE;
2104 } else {
2105 error = SET_ERROR(EBUSY);
2106 }
2107 } else if (spa->spa_l2cache.sav_vdevs != NULL &&
2108 nvlist_lookup_nvlist_array(spa->spa_l2cache.sav_config,
2109 ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0 &&
2110 (nv = spa_nvlist_lookup_by_guid(l2cache, nl2cache, guid)) != NULL) {
2111 char *nvstr = fnvlist_lookup_string(nv, ZPOOL_CONFIG_PATH);
2112 spa_history_log_internal(spa, "vdev remove", NULL,
2113 "%s vdev (%s) %s", spa_name(spa), VDEV_TYPE_L2CACHE, nvstr);
2114 /*
2115 * Cache devices can always be removed.
2116 */
2117 vd = spa_lookup_by_guid(spa, guid, B_TRUE);
2118 ev = spa_event_create(spa, vd, NULL, ESC_ZFS_VDEV_REMOVE_AUX);
2119 spa_vdev_remove_aux(spa->spa_l2cache.sav_config,
2120 ZPOOL_CONFIG_L2CACHE, l2cache, nl2cache, nv);
2121 spa_load_l2cache(spa);
2122 spa->spa_l2cache.sav_sync = B_TRUE;
2123 } else if (vd != NULL && vd->vdev_islog) {
2124 ASSERT(!locked);
2125 error = spa_vdev_remove_log(vd, &txg);
2126 } else if (vd != NULL) {
2127 ASSERT(!locked);
2128 error = spa_vdev_remove_top(vd, &txg);
2129 } else {
2130 /*
2131 * There is no vdev of any kind with the specified guid.
2132 */
2133 error = SET_ERROR(ENOENT);
2134 }
2135
2136 if (!locked)
2137 error = spa_vdev_exit(spa, NULL, txg, error);
2138
2139 if (ev != NULL) {
2140 if (error != 0) {
2141 spa_event_discard(ev);
2142 } else {
2143 spa_event_post(ev);
2144 }
2145 }
2146
2147 return (error);
2148 }
2149
2150 int
2151 spa_removal_get_stats(spa_t *spa, pool_removal_stat_t *prs)
2152 {
2153 prs->prs_state = spa->spa_removing_phys.sr_state;
2154
2155 if (prs->prs_state == DSS_NONE)
2156 return (SET_ERROR(ENOENT));
2157
2158 prs->prs_removing_vdev = spa->spa_removing_phys.sr_removing_vdev;
2159 prs->prs_start_time = spa->spa_removing_phys.sr_start_time;
2160 prs->prs_end_time = spa->spa_removing_phys.sr_end_time;
2161 prs->prs_to_copy = spa->spa_removing_phys.sr_to_copy;
2162 prs->prs_copied = spa->spa_removing_phys.sr_copied;
2163
2164 if (spa->spa_vdev_removal != NULL) {
2165 for (int i = 0; i < TXG_SIZE; i++) {
2166 prs->prs_copied +=
2167 spa->spa_vdev_removal->svr_bytes_done[i];
2168 }
2169 }
2170
2171 prs->prs_mapping_memory = 0;
2172 uint64_t indirect_vdev_id =
2173 spa->spa_removing_phys.sr_prev_indirect_vdev;
2174 while (indirect_vdev_id != -1) {
2175 vdev_t *vd = spa->spa_root_vdev->vdev_child[indirect_vdev_id];
2176 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
2177 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
2178
2179 ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
2180 prs->prs_mapping_memory += vdev_indirect_mapping_size(vim);
2181 indirect_vdev_id = vic->vic_prev_indirect_vdev;
2182 }
2183
2184 return (0);
2185 }