Merge commit '7d815089a43a963b49eaddf97e514194ec29805b'
[unleashed.git] / kernel / fs / zfs / vdev_removal.c
blobd00b5b35f743c7f0b5c21bfe3e2ea4eb499812a7
1 /*
2 * CDDL HEADER START
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.
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.
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]
19 * CDDL HEADER END
23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
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>
49 * This file contains the necessary logic to remove vdevs from a
50 * storage pool. Currently, the only devices that can be removed
51 * are log, cache, and spare devices; and top level vdevs from a pool
52 * w/o raidz. (Note that members of a mirror can also be removed
53 * by the detach operation.)
55 * Log vdevs are removed by evacuating them and then turning the vdev
56 * into a hole vdev while holding spa config locks.
58 * Top level vdevs are removed and converted into an indirect vdev via
59 * a multi-step process:
61 * - Disable allocations from this device (spa_vdev_remove_top).
63 * - From a new thread (spa_vdev_remove_thread), copy data from
64 * the removing vdev to a different vdev. The copy happens in open
65 * context (spa_vdev_copy_impl) and issues a sync task
66 * (vdev_mapping_sync) so the sync thread can update the partial
67 * indirect mappings in core and on disk.
69 * - If a free happens during a removal, it is freed from the
70 * removing vdev, and if it has already been copied, from the new
71 * location as well (free_from_removing_vdev).
73 * - After the removal is completed, the copy thread converts the vdev
74 * into an indirect vdev (vdev_remove_complete) before instructing
75 * the sync thread to destroy the space maps and finish the removal
76 * (spa_finish_removal).
79 typedef struct vdev_copy_arg {
80 metaslab_t *vca_msp;
81 uint64_t vca_outstanding_bytes;
82 kcondvar_t vca_cv;
83 kmutex_t vca_lock;
84 } vdev_copy_arg_t;
86 typedef struct vdev_copy_seg_arg {
87 vdev_copy_arg_t *vcsa_copy_arg;
88 uint64_t vcsa_txg;
89 dva_t *vcsa_dest_dva;
90 blkptr_t *vcsa_dest_bp;
91 } vdev_copy_seg_arg_t;
94 * The maximum amount of allowed data we're allowed to copy from a device
95 * at a time when removing it.
97 int zfs_remove_max_copy_bytes = 8 * 1024 * 1024;
100 * The largest contiguous segment that we will attempt to allocate when
101 * removing a device. This can be no larger than SPA_MAXBLOCKSIZE. If
102 * there is a performance problem with attempting to allocate large blocks,
103 * consider decreasing this.
105 * Note: we will issue I/Os of up to this size. The mpt driver does not
106 * respond well to I/Os larger than 1MB, so we set this to 1MB. (When
107 * mpt processes an I/O larger than 1MB, it needs to do an allocation of
108 * 2 physically contiguous pages; if this allocation fails, mpt will drop
109 * the I/O and hang the device.)
111 int zfs_remove_max_segment = 1024 * 1024;
114 * This is used by the test suite so that it can ensure that certain
115 * actions happen while in the middle of a removal.
117 uint64_t zfs_remove_max_bytes_pause = UINT64_MAX;
119 #define VDEV_REMOVAL_ZAP_OBJS "lzap"
121 static void spa_vdev_remove_thread(void *arg);
123 static void
124 spa_sync_removing_state(spa_t *spa, dmu_tx_t *tx)
126 VERIFY0(zap_update(spa->spa_dsl_pool->dp_meta_objset,
127 DMU_POOL_DIRECTORY_OBJECT,
128 DMU_POOL_REMOVING, sizeof (uint64_t),
129 sizeof (spa->spa_removing_phys) / sizeof (uint64_t),
130 &spa->spa_removing_phys, tx));
133 static nvlist_t *
134 spa_nvlist_lookup_by_guid(nvlist_t **nvpp, int count, uint64_t target_guid)
136 for (int i = 0; i < count; i++) {
137 uint64_t guid =
138 fnvlist_lookup_uint64(nvpp[i], ZPOOL_CONFIG_GUID);
140 if (guid == target_guid)
141 return (nvpp[i]);
144 return (NULL);
147 static void
148 spa_vdev_remove_aux(nvlist_t *config, char *name, nvlist_t **dev, int count,
149 nvlist_t *dev_to_remove)
151 nvlist_t **newdev = NULL;
153 if (count > 1)
154 newdev = kmem_alloc((count - 1) * sizeof (void *), KM_SLEEP);
156 for (int i = 0, j = 0; i < count; i++) {
157 if (dev[i] == dev_to_remove)
158 continue;
159 VERIFY(nvlist_dup(dev[i], &newdev[j++], KM_SLEEP) == 0);
162 VERIFY(nvlist_remove(config, name, DATA_TYPE_NVLIST_ARRAY) == 0);
163 VERIFY(nvlist_add_nvlist_array(config, name, newdev, count - 1) == 0);
165 for (int i = 0; i < count - 1; i++)
166 nvlist_free(newdev[i]);
168 if (count > 1)
169 kmem_free(newdev, (count - 1) * sizeof (void *));
172 static spa_vdev_removal_t *
173 spa_vdev_removal_create(vdev_t *vd)
175 spa_vdev_removal_t *svr = kmem_zalloc(sizeof (*svr), KM_SLEEP);
176 mutex_init(&svr->svr_lock, NULL, MUTEX_DEFAULT, NULL);
177 cv_init(&svr->svr_cv, NULL, CV_DEFAULT, NULL);
178 svr->svr_allocd_segs = range_tree_create(NULL, NULL);
179 svr->svr_vdev = vd;
181 for (int i = 0; i < TXG_SIZE; i++) {
182 svr->svr_frees[i] = range_tree_create(NULL, NULL);
183 list_create(&svr->svr_new_segments[i],
184 sizeof (vdev_indirect_mapping_entry_t),
185 offsetof(vdev_indirect_mapping_entry_t, vime_node));
188 return (svr);
191 void
192 spa_vdev_removal_destroy(spa_vdev_removal_t *svr)
194 for (int i = 0; i < TXG_SIZE; i++) {
195 ASSERT0(svr->svr_bytes_done[i]);
196 ASSERT0(svr->svr_max_offset_to_sync[i]);
197 range_tree_destroy(svr->svr_frees[i]);
198 list_destroy(&svr->svr_new_segments[i]);
201 range_tree_destroy(svr->svr_allocd_segs);
202 mutex_destroy(&svr->svr_lock);
203 cv_destroy(&svr->svr_cv);
204 kmem_free(svr, sizeof (*svr));
208 * This is called as a synctask in the txg in which we will mark this vdev
209 * as removing (in the config stored in the MOS).
211 * It begins the evacuation of a toplevel vdev by:
212 * - initializing the spa_removing_phys which tracks this removal
213 * - computing the amount of space to remove for accounting purposes
214 * - dirtying all dbufs in the spa_config_object
215 * - creating the spa_vdev_removal
216 * - starting the spa_vdev_remove_thread
218 static void
219 vdev_remove_initiate_sync(void *arg, dmu_tx_t *tx)
221 vdev_t *vd = arg;
222 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
223 spa_t *spa = vd->vdev_spa;
224 objset_t *mos = spa->spa_dsl_pool->dp_meta_objset;
225 spa_vdev_removal_t *svr = NULL;
226 uint64_t txg = dmu_tx_get_txg(tx);
228 ASSERT3P(vd->vdev_ops, !=, &vdev_raidz_ops);
229 svr = spa_vdev_removal_create(vd);
231 ASSERT(vd->vdev_removing);
232 ASSERT3P(vd->vdev_indirect_mapping, ==, NULL);
234 spa_feature_incr(spa, SPA_FEATURE_DEVICE_REMOVAL, tx);
235 if (spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS)) {
237 * By activating the OBSOLETE_COUNTS feature, we prevent
238 * the pool from being downgraded and ensure that the
239 * refcounts are precise.
241 spa_feature_incr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
242 uint64_t one = 1;
243 VERIFY0(zap_add(spa->spa_meta_objset, vd->vdev_top_zap,
244 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE, sizeof (one), 1,
245 &one, tx));
246 ASSERT3U(vdev_obsolete_counts_are_precise(vd), !=, 0);
249 vic->vic_mapping_object = vdev_indirect_mapping_alloc(mos, tx);
250 vd->vdev_indirect_mapping =
251 vdev_indirect_mapping_open(mos, vic->vic_mapping_object);
252 vic->vic_births_object = vdev_indirect_births_alloc(mos, tx);
253 vd->vdev_indirect_births =
254 vdev_indirect_births_open(mos, vic->vic_births_object);
255 spa->spa_removing_phys.sr_removing_vdev = vd->vdev_id;
256 spa->spa_removing_phys.sr_start_time = gethrestime_sec();
257 spa->spa_removing_phys.sr_end_time = 0;
258 spa->spa_removing_phys.sr_state = DSS_SCANNING;
259 spa->spa_removing_phys.sr_to_copy = 0;
260 spa->spa_removing_phys.sr_copied = 0;
263 * Note: We can't use vdev_stat's vs_alloc for sr_to_copy, because
264 * there may be space in the defer tree, which is free, but still
265 * counted in vs_alloc.
267 for (uint64_t i = 0; i < vd->vdev_ms_count; i++) {
268 metaslab_t *ms = vd->vdev_ms[i];
269 if (ms->ms_sm == NULL)
270 continue;
273 * Sync tasks happen before metaslab_sync(), therefore
274 * smp_alloc and sm_alloc must be the same.
276 ASSERT3U(space_map_allocated(ms->ms_sm), ==,
277 ms->ms_sm->sm_phys->smp_alloc);
279 spa->spa_removing_phys.sr_to_copy +=
280 space_map_allocated(ms->ms_sm);
283 * Space which we are freeing this txg does not need to
284 * be copied.
286 spa->spa_removing_phys.sr_to_copy -=
287 range_tree_space(ms->ms_freeing);
289 ASSERT0(range_tree_space(ms->ms_freed));
290 for (int t = 0; t < TXG_SIZE; t++)
291 ASSERT0(range_tree_space(ms->ms_allocating[t]));
295 * Sync tasks are called before metaslab_sync(), so there should
296 * be no already-synced metaslabs in the TXG_CLEAN list.
298 ASSERT3P(txg_list_head(&vd->vdev_ms_list, TXG_CLEAN(txg)), ==, NULL);
300 spa_sync_removing_state(spa, tx);
303 * All blocks that we need to read the most recent mapping must be
304 * stored on concrete vdevs. Therefore, we must dirty anything that
305 * is read before spa_remove_init(). Specifically, the
306 * spa_config_object. (Note that although we already modified the
307 * spa_config_object in spa_sync_removing_state, that may not have
308 * modified all blocks of the object.)
310 dmu_object_info_t doi;
311 VERIFY0(dmu_object_info(mos, DMU_POOL_DIRECTORY_OBJECT, &doi));
312 for (uint64_t offset = 0; offset < doi.doi_max_offset; ) {
313 dmu_buf_t *dbuf;
314 VERIFY0(dmu_buf_hold(mos, DMU_POOL_DIRECTORY_OBJECT,
315 offset, FTAG, &dbuf, 0));
316 dmu_buf_will_dirty(dbuf, tx);
317 offset += dbuf->db_size;
318 dmu_buf_rele(dbuf, FTAG);
322 * Now that we've allocated the im_object, dirty the vdev to ensure
323 * that the object gets written to the config on disk.
325 vdev_config_dirty(vd);
327 zfs_dbgmsg("starting removal thread for vdev %llu (%p) in txg %llu "
328 "im_obj=%llu", vd->vdev_id, vd, dmu_tx_get_txg(tx),
329 vic->vic_mapping_object);
331 spa_history_log_internal(spa, "vdev remove started", tx,
332 "%s vdev %llu %s", spa_name(spa), vd->vdev_id,
333 (vd->vdev_path != NULL) ? vd->vdev_path : "-");
335 * Setting spa_vdev_removal causes subsequent frees to call
336 * free_from_removing_vdev(). Note that we don't need any locking
337 * because we are the sync thread, and metaslab_free_impl() is only
338 * called from syncing context (potentially from a zio taskq thread,
339 * but in any case only when there are outstanding free i/os, which
340 * there are not).
342 ASSERT3P(spa->spa_vdev_removal, ==, NULL);
343 spa->spa_vdev_removal = svr;
344 svr->svr_thread = thread_create(NULL, 0,
345 spa_vdev_remove_thread, vd, 0, &p0, TS_RUN, minclsyspri);
349 * When we are opening a pool, we must read the mapping for each
350 * indirect vdev in order from most recently removed to least
351 * recently removed. We do this because the blocks for the mapping
352 * of older indirect vdevs may be stored on more recently removed vdevs.
353 * In order to read each indirect mapping object, we must have
354 * initialized all more recently removed vdevs.
357 spa_remove_init(spa_t *spa)
359 int error;
361 error = zap_lookup(spa->spa_dsl_pool->dp_meta_objset,
362 DMU_POOL_DIRECTORY_OBJECT,
363 DMU_POOL_REMOVING, sizeof (uint64_t),
364 sizeof (spa->spa_removing_phys) / sizeof (uint64_t),
365 &spa->spa_removing_phys);
367 if (error == ENOENT) {
368 spa->spa_removing_phys.sr_state = DSS_NONE;
369 spa->spa_removing_phys.sr_removing_vdev = -1;
370 spa->spa_removing_phys.sr_prev_indirect_vdev = -1;
371 return (0);
372 } else if (error != 0) {
373 return (error);
376 if (spa->spa_removing_phys.sr_state == DSS_SCANNING) {
378 * We are currently removing a vdev. Create and
379 * initialize a spa_vdev_removal_t from the bonus
380 * buffer of the removing vdevs vdev_im_object, and
381 * initialize its partial mapping.
383 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
384 vdev_t *vd = vdev_lookup_top(spa,
385 spa->spa_removing_phys.sr_removing_vdev);
386 spa_config_exit(spa, SCL_STATE, FTAG);
388 if (vd == NULL)
389 return (EINVAL);
391 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
393 ASSERT(vdev_is_concrete(vd));
394 spa_vdev_removal_t *svr = spa_vdev_removal_create(vd);
395 ASSERT(svr->svr_vdev->vdev_removing);
397 vd->vdev_indirect_mapping = vdev_indirect_mapping_open(
398 spa->spa_meta_objset, vic->vic_mapping_object);
399 vd->vdev_indirect_births = vdev_indirect_births_open(
400 spa->spa_meta_objset, vic->vic_births_object);
402 spa->spa_vdev_removal = svr;
405 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
406 uint64_t indirect_vdev_id =
407 spa->spa_removing_phys.sr_prev_indirect_vdev;
408 while (indirect_vdev_id != UINT64_MAX) {
409 vdev_t *vd = vdev_lookup_top(spa, indirect_vdev_id);
410 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
412 ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
413 vd->vdev_indirect_mapping = vdev_indirect_mapping_open(
414 spa->spa_meta_objset, vic->vic_mapping_object);
415 vd->vdev_indirect_births = vdev_indirect_births_open(
416 spa->spa_meta_objset, vic->vic_births_object);
418 indirect_vdev_id = vic->vic_prev_indirect_vdev;
420 spa_config_exit(spa, SCL_STATE, FTAG);
423 * Now that we've loaded all the indirect mappings, we can allow
424 * reads from other blocks (e.g. via predictive prefetch).
426 spa->spa_indirect_vdevs_loaded = B_TRUE;
427 return (0);
430 void
431 spa_restart_removal(spa_t *spa)
433 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
435 if (svr == NULL)
436 return;
439 * In general when this function is called there is no
440 * removal thread running. The only scenario where this
441 * is not true is during spa_import() where this function
442 * is called twice [once from spa_import_impl() and
443 * spa_async_resume()]. Thus, in the scenario where we
444 * import a pool that has an ongoing removal we don't
445 * want to spawn a second thread.
447 if (svr->svr_thread != NULL)
448 return;
450 if (!spa_writeable(spa))
451 return;
453 vdev_t *vd = svr->svr_vdev;
454 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
456 ASSERT3P(vd, !=, NULL);
457 ASSERT(vd->vdev_removing);
459 zfs_dbgmsg("restarting removal of %llu at count=%llu",
460 vd->vdev_id, vdev_indirect_mapping_num_entries(vim));
461 svr->svr_thread = thread_create(NULL, 0, spa_vdev_remove_thread, vd,
462 0, &p0, TS_RUN, minclsyspri);
466 * Process freeing from a device which is in the middle of being removed.
467 * We must handle this carefully so that we attempt to copy freed data,
468 * and we correctly free already-copied data.
470 void
471 free_from_removing_vdev(vdev_t *vd, uint64_t offset, uint64_t size)
473 spa_t *spa = vd->vdev_spa;
474 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
475 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
476 uint64_t txg = spa_syncing_txg(spa);
477 uint64_t max_offset_yet = 0;
479 ASSERT(vd->vdev_indirect_config.vic_mapping_object != 0);
480 ASSERT3U(vd->vdev_indirect_config.vic_mapping_object, ==,
481 vdev_indirect_mapping_object(vim));
482 ASSERT3P(vd, ==, svr->svr_vdev);
484 mutex_enter(&svr->svr_lock);
487 * Remove the segment from the removing vdev's spacemap. This
488 * ensures that we will not attempt to copy this space (if the
489 * removal thread has not yet visited it), and also ensures
490 * that we know what is actually allocated on the new vdevs
491 * (needed if we cancel the removal).
493 * Note: we must do the metaslab_free_concrete() with the svr_lock
494 * held, so that the remove_thread can not load this metaslab and then
495 * visit this offset between the time that we metaslab_free_concrete()
496 * and when we check to see if it has been visited.
498 * Note: The checkpoint flag is set to false as having/taking
499 * a checkpoint and removing a device can't happen at the same
500 * time.
502 ASSERT(!spa_has_checkpoint(spa));
503 metaslab_free_concrete(vd, offset, size, B_FALSE);
505 uint64_t synced_size = 0;
506 uint64_t synced_offset = 0;
507 uint64_t max_offset_synced = vdev_indirect_mapping_max_offset(vim);
508 if (offset < max_offset_synced) {
510 * The mapping for this offset is already on disk.
511 * Free from the new location.
513 * Note that we use svr_max_synced_offset because it is
514 * updated atomically with respect to the in-core mapping.
515 * By contrast, vim_max_offset is not.
517 * This block may be split between a synced entry and an
518 * in-flight or unvisited entry. Only process the synced
519 * portion of it here.
521 synced_size = MIN(size, max_offset_synced - offset);
522 synced_offset = offset;
524 ASSERT3U(max_offset_yet, <=, max_offset_synced);
525 max_offset_yet = max_offset_synced;
527 DTRACE_PROBE3(remove__free__synced,
528 spa_t *, spa,
529 uint64_t, offset,
530 uint64_t, synced_size);
532 size -= synced_size;
533 offset += synced_size;
537 * Look at all in-flight txgs starting from the currently syncing one
538 * and see if a section of this free is being copied. By starting from
539 * this txg and iterating forward, we might find that this region
540 * was copied in two different txgs and handle it appropriately.
542 for (int i = 0; i < TXG_CONCURRENT_STATES; i++) {
543 int txgoff = (txg + i) & TXG_MASK;
544 if (size > 0 && offset < svr->svr_max_offset_to_sync[txgoff]) {
546 * The mapping for this offset is in flight, and
547 * will be synced in txg+i.
549 uint64_t inflight_size = MIN(size,
550 svr->svr_max_offset_to_sync[txgoff] - offset);
552 DTRACE_PROBE4(remove__free__inflight,
553 spa_t *, spa,
554 uint64_t, offset,
555 uint64_t, inflight_size,
556 uint64_t, txg + i);
559 * We copy data in order of increasing offset.
560 * Therefore the max_offset_to_sync[] must increase
561 * (or be zero, indicating that nothing is being
562 * copied in that txg).
564 if (svr->svr_max_offset_to_sync[txgoff] != 0) {
565 ASSERT3U(svr->svr_max_offset_to_sync[txgoff],
566 >=, max_offset_yet);
567 max_offset_yet =
568 svr->svr_max_offset_to_sync[txgoff];
572 * We've already committed to copying this segment:
573 * we have allocated space elsewhere in the pool for
574 * it and have an IO outstanding to copy the data. We
575 * cannot free the space before the copy has
576 * completed, or else the copy IO might overwrite any
577 * new data. To free that space, we record the
578 * segment in the appropriate svr_frees tree and free
579 * the mapped space later, in the txg where we have
580 * completed the copy and synced the mapping (see
581 * vdev_mapping_sync).
583 range_tree_add(svr->svr_frees[txgoff],
584 offset, inflight_size);
585 size -= inflight_size;
586 offset += inflight_size;
589 * This space is already accounted for as being
590 * done, because it is being copied in txg+i.
591 * However, if i!=0, then it is being copied in
592 * a future txg. If we crash after this txg
593 * syncs but before txg+i syncs, then the space
594 * will be free. Therefore we must account
595 * for the space being done in *this* txg
596 * (when it is freed) rather than the future txg
597 * (when it will be copied).
599 ASSERT3U(svr->svr_bytes_done[txgoff], >=,
600 inflight_size);
601 svr->svr_bytes_done[txgoff] -= inflight_size;
602 svr->svr_bytes_done[txg & TXG_MASK] += inflight_size;
605 ASSERT0(svr->svr_max_offset_to_sync[TXG_CLEAN(txg) & TXG_MASK]);
607 if (size > 0) {
609 * The copy thread has not yet visited this offset. Ensure
610 * that it doesn't.
613 DTRACE_PROBE3(remove__free__unvisited,
614 spa_t *, spa,
615 uint64_t, offset,
616 uint64_t, size);
618 if (svr->svr_allocd_segs != NULL)
619 range_tree_clear(svr->svr_allocd_segs, offset, size);
622 * Since we now do not need to copy this data, for
623 * accounting purposes we have done our job and can count
624 * it as completed.
626 svr->svr_bytes_done[txg & TXG_MASK] += size;
628 mutex_exit(&svr->svr_lock);
631 * Now that we have dropped svr_lock, process the synced portion
632 * of this free.
634 if (synced_size > 0) {
635 vdev_indirect_mark_obsolete(vd, synced_offset, synced_size);
638 * Note: this can only be called from syncing context,
639 * and the vdev_indirect_mapping is only changed from the
640 * sync thread, so we don't need svr_lock while doing
641 * metaslab_free_impl_cb.
643 boolean_t checkpoint = B_FALSE;
644 vdev_indirect_ops.vdev_op_remap(vd, synced_offset, synced_size,
645 metaslab_free_impl_cb, &checkpoint);
650 * Stop an active removal and update the spa_removing phys.
652 static void
653 spa_finish_removal(spa_t *spa, dsl_scan_state_t state, dmu_tx_t *tx)
655 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
656 ASSERT3U(dmu_tx_get_txg(tx), ==, spa_syncing_txg(spa));
658 /* Ensure the removal thread has completed before we free the svr. */
659 spa_vdev_remove_suspend(spa);
661 ASSERT(state == DSS_FINISHED || state == DSS_CANCELED);
663 if (state == DSS_FINISHED) {
664 spa_removing_phys_t *srp = &spa->spa_removing_phys;
665 vdev_t *vd = svr->svr_vdev;
666 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
668 if (srp->sr_prev_indirect_vdev != UINT64_MAX) {
669 vdev_t *pvd = vdev_lookup_top(spa,
670 srp->sr_prev_indirect_vdev);
671 ASSERT3P(pvd->vdev_ops, ==, &vdev_indirect_ops);
674 vic->vic_prev_indirect_vdev = srp->sr_prev_indirect_vdev;
675 srp->sr_prev_indirect_vdev = vd->vdev_id;
677 spa->spa_removing_phys.sr_state = state;
678 spa->spa_removing_phys.sr_end_time = gethrestime_sec();
680 spa->spa_vdev_removal = NULL;
681 spa_vdev_removal_destroy(svr);
683 spa_sync_removing_state(spa, tx);
685 vdev_config_dirty(spa->spa_root_vdev);
688 static void
689 free_mapped_segment_cb(void *arg, uint64_t offset, uint64_t size)
691 vdev_t *vd = arg;
692 vdev_indirect_mark_obsolete(vd, offset, size);
693 boolean_t checkpoint = B_FALSE;
694 vdev_indirect_ops.vdev_op_remap(vd, offset, size,
695 metaslab_free_impl_cb, &checkpoint);
699 * On behalf of the removal thread, syncs an incremental bit more of
700 * the indirect mapping to disk and updates the in-memory mapping.
701 * Called as a sync task in every txg that the removal thread makes progress.
703 static void
704 vdev_mapping_sync(void *arg, dmu_tx_t *tx)
706 spa_vdev_removal_t *svr = arg;
707 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
708 vdev_t *vd = svr->svr_vdev;
709 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
710 uint64_t txg = dmu_tx_get_txg(tx);
711 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
713 ASSERT(vic->vic_mapping_object != 0);
714 ASSERT3U(txg, ==, spa_syncing_txg(spa));
716 vdev_indirect_mapping_add_entries(vim,
717 &svr->svr_new_segments[txg & TXG_MASK], tx);
718 vdev_indirect_births_add_entry(vd->vdev_indirect_births,
719 vdev_indirect_mapping_max_offset(vim), dmu_tx_get_txg(tx), tx);
722 * Free the copied data for anything that was freed while the
723 * mapping entries were in flight.
725 mutex_enter(&svr->svr_lock);
726 range_tree_vacate(svr->svr_frees[txg & TXG_MASK],
727 free_mapped_segment_cb, vd);
728 ASSERT3U(svr->svr_max_offset_to_sync[txg & TXG_MASK], >=,
729 vdev_indirect_mapping_max_offset(vim));
730 svr->svr_max_offset_to_sync[txg & TXG_MASK] = 0;
731 mutex_exit(&svr->svr_lock);
733 spa_sync_removing_state(spa, tx);
736 static void
737 spa_vdev_copy_segment_write_done(zio_t *zio)
739 vdev_copy_seg_arg_t *vcsa = zio->io_private;
740 vdev_copy_arg_t *vca = vcsa->vcsa_copy_arg;
741 spa_config_exit(zio->io_spa, SCL_STATE, FTAG);
742 abd_free(zio->io_abd);
744 mutex_enter(&vca->vca_lock);
745 vca->vca_outstanding_bytes -= zio->io_size;
746 cv_signal(&vca->vca_cv);
747 mutex_exit(&vca->vca_lock);
749 ASSERT0(zio->io_error);
750 kmem_free(vcsa->vcsa_dest_bp, sizeof (blkptr_t));
751 kmem_free(vcsa, sizeof (vdev_copy_seg_arg_t));
754 static void
755 spa_vdev_copy_segment_read_done(zio_t *zio)
757 vdev_copy_seg_arg_t *vcsa = zio->io_private;
758 dva_t *dest_dva = vcsa->vcsa_dest_dva;
759 uint64_t txg = vcsa->vcsa_txg;
760 spa_t *spa = zio->io_spa;
761 vdev_t *dest_vd = vdev_lookup_top(spa, DVA_GET_VDEV(dest_dva));
762 blkptr_t *bp = NULL;
763 dva_t *dva = NULL;
764 uint64_t size = zio->io_size;
766 ASSERT3P(dest_vd, !=, NULL);
767 ASSERT0(zio->io_error);
769 vcsa->vcsa_dest_bp = kmem_alloc(sizeof (blkptr_t), KM_SLEEP);
770 bp = vcsa->vcsa_dest_bp;
771 dva = bp->blk_dva;
773 BP_ZERO(bp);
775 /* initialize with dest_dva */
776 bcopy(dest_dva, dva, sizeof (dva_t));
777 BP_SET_BIRTH(bp, TXG_INITIAL, TXG_INITIAL);
779 BP_SET_LSIZE(bp, size);
780 BP_SET_PSIZE(bp, size);
781 BP_SET_COMPRESS(bp, ZIO_COMPRESS_OFF);
782 BP_SET_CHECKSUM(bp, ZIO_CHECKSUM_OFF);
783 BP_SET_TYPE(bp, DMU_OT_NONE);
784 BP_SET_LEVEL(bp, 0);
785 BP_SET_DEDUP(bp, 0);
786 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
788 zio_nowait(zio_rewrite(spa->spa_txg_zio[txg & TXG_MASK], spa,
789 txg, bp, zio->io_abd, size,
790 spa_vdev_copy_segment_write_done, vcsa,
791 ZIO_PRIORITY_REMOVAL, 0, NULL));
794 static int
795 spa_vdev_copy_segment(vdev_t *vd, uint64_t start, uint64_t size, uint64_t txg,
796 vdev_copy_arg_t *vca, zio_alloc_list_t *zal)
798 metaslab_group_t *mg = vd->vdev_mg;
799 spa_t *spa = vd->vdev_spa;
800 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
801 vdev_indirect_mapping_entry_t *entry;
802 vdev_copy_seg_arg_t *private;
803 dva_t dst = { 0 };
804 blkptr_t blk, *bp = &blk;
805 dva_t *dva = bp->blk_dva;
807 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
810 * We use allocator 0 for this I/O because we don't expect device remap
811 * to be the steady state of the system, so parallelizing is not as
812 * critical as it is for other allocation types. We also want to ensure
813 * that the IOs are allocated together as much as possible, to reduce
814 * mapping sizes.
816 int error = metaslab_alloc_dva(spa, mg->mg_class, size,
817 &dst, 0, NULL, txg, 0, zal, 0);
818 if (error != 0)
819 return (error);
822 * We can't have any padding of the allocated size, otherwise we will
823 * misunderstand what's allocated, and the size of the mapping.
824 * The caller ensures this will be true by passing in a size that is
825 * aligned to the worst (highest) ashift in the pool.
827 ASSERT3U(DVA_GET_ASIZE(&dst), ==, size);
829 mutex_enter(&vca->vca_lock);
830 vca->vca_outstanding_bytes += size;
831 mutex_exit(&vca->vca_lock);
833 entry = kmem_zalloc(sizeof (vdev_indirect_mapping_entry_t), KM_SLEEP);
834 DVA_MAPPING_SET_SRC_OFFSET(&entry->vime_mapping, start);
835 entry->vime_mapping.vimep_dst = dst;
837 private = kmem_alloc(sizeof (vdev_copy_seg_arg_t), KM_SLEEP);
838 private->vcsa_dest_dva = &entry->vime_mapping.vimep_dst;
839 private->vcsa_txg = txg;
840 private->vcsa_copy_arg = vca;
843 * This lock is eventually released by the donefunc for the
844 * zio_write_phys that finishes copying the data.
846 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
849 * Do logical I/O, letting the redundancy vdevs (like mirror)
850 * handle their own I/O instead of duplicating that code here.
852 BP_ZERO(bp);
854 DVA_SET_VDEV(&dva[0], vd->vdev_id);
855 DVA_SET_OFFSET(&dva[0], start);
856 DVA_SET_GANG(&dva[0], 0);
857 DVA_SET_ASIZE(&dva[0], vdev_psize_to_asize(vd, size));
859 BP_SET_BIRTH(bp, TXG_INITIAL, TXG_INITIAL);
861 BP_SET_LSIZE(bp, size);
862 BP_SET_PSIZE(bp, size);
863 BP_SET_COMPRESS(bp, ZIO_COMPRESS_OFF);
864 BP_SET_CHECKSUM(bp, ZIO_CHECKSUM_OFF);
865 BP_SET_TYPE(bp, DMU_OT_NONE);
866 BP_SET_LEVEL(bp, 0);
867 BP_SET_DEDUP(bp, 0);
868 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
870 zio_nowait(zio_read(spa->spa_txg_zio[txg & TXG_MASK], spa,
871 bp, abd_alloc_for_io(size, B_FALSE), size,
872 spa_vdev_copy_segment_read_done, private,
873 ZIO_PRIORITY_REMOVAL, 0, NULL));
875 list_insert_tail(&svr->svr_new_segments[txg & TXG_MASK], entry);
876 ASSERT3U(start + size, <=, vd->vdev_ms_count << vd->vdev_ms_shift);
877 vdev_dirty(vd, 0, NULL, txg);
879 return (0);
883 * Complete the removal of a toplevel vdev. This is called as a
884 * synctask in the same txg that we will sync out the new config (to the
885 * MOS object) which indicates that this vdev is indirect.
887 static void
888 vdev_remove_complete_sync(void *arg, dmu_tx_t *tx)
890 spa_vdev_removal_t *svr = arg;
891 vdev_t *vd = svr->svr_vdev;
892 spa_t *spa = vd->vdev_spa;
894 ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
896 for (int i = 0; i < TXG_SIZE; i++) {
897 ASSERT0(svr->svr_bytes_done[i]);
900 ASSERT3U(spa->spa_removing_phys.sr_copied, ==,
901 spa->spa_removing_phys.sr_to_copy);
903 vdev_destroy_spacemaps(vd, tx);
905 /* destroy leaf zaps, if any */
906 ASSERT3P(svr->svr_zaplist, !=, NULL);
907 for (nvpair_t *pair = nvlist_next_nvpair(svr->svr_zaplist, NULL);
908 pair != NULL;
909 pair = nvlist_next_nvpair(svr->svr_zaplist, pair)) {
910 vdev_destroy_unlink_zap(vd, fnvpair_value_uint64(pair), tx);
912 fnvlist_free(svr->svr_zaplist);
914 spa_finish_removal(dmu_tx_pool(tx)->dp_spa, DSS_FINISHED, tx);
915 /* vd->vdev_path is not available here */
916 spa_history_log_internal(spa, "vdev remove completed", tx,
917 "%s vdev %llu", spa_name(spa), vd->vdev_id);
920 static void
921 vdev_indirect_state_transfer(vdev_t *ivd, vdev_t *vd)
923 ivd->vdev_indirect_config = vd->vdev_indirect_config;
925 ASSERT3P(ivd->vdev_indirect_mapping, ==, NULL);
926 ASSERT(vd->vdev_indirect_mapping != NULL);
927 ivd->vdev_indirect_mapping = vd->vdev_indirect_mapping;
928 vd->vdev_indirect_mapping = NULL;
930 ASSERT3P(ivd->vdev_indirect_births, ==, NULL);
931 ASSERT(vd->vdev_indirect_births != NULL);
932 ivd->vdev_indirect_births = vd->vdev_indirect_births;
933 vd->vdev_indirect_births = NULL;
935 ASSERT0(range_tree_space(vd->vdev_obsolete_segments));
936 ASSERT0(range_tree_space(ivd->vdev_obsolete_segments));
938 if (vd->vdev_obsolete_sm != NULL) {
939 ASSERT3U(ivd->vdev_asize, ==, vd->vdev_asize);
942 * We cannot use space_map_{open,close} because we hold all
943 * the config locks as writer.
945 ASSERT3P(ivd->vdev_obsolete_sm, ==, NULL);
946 ivd->vdev_obsolete_sm = vd->vdev_obsolete_sm;
947 vd->vdev_obsolete_sm = NULL;
951 static void
952 vdev_remove_enlist_zaps(vdev_t *vd, nvlist_t *zlist)
954 ASSERT3P(zlist, !=, NULL);
955 ASSERT3P(vd->vdev_ops, !=, &vdev_raidz_ops);
957 if (vd->vdev_leaf_zap != 0) {
958 char zkey[32];
959 (void) snprintf(zkey, sizeof (zkey), "%s-%"PRIu64,
960 VDEV_REMOVAL_ZAP_OBJS, vd->vdev_leaf_zap);
961 fnvlist_add_uint64(zlist, zkey, vd->vdev_leaf_zap);
964 for (uint64_t id = 0; id < vd->vdev_children; id++) {
965 vdev_remove_enlist_zaps(vd->vdev_child[id], zlist);
969 static void
970 vdev_remove_replace_with_indirect(vdev_t *vd, uint64_t txg)
972 vdev_t *ivd;
973 dmu_tx_t *tx;
974 spa_t *spa = vd->vdev_spa;
975 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
978 * First, build a list of leaf zaps to be destroyed.
979 * This is passed to the sync context thread,
980 * which does the actual unlinking.
982 svr->svr_zaplist = fnvlist_alloc();
983 vdev_remove_enlist_zaps(vd, svr->svr_zaplist);
985 ivd = vdev_add_parent(vd, &vdev_indirect_ops);
987 vd->vdev_leaf_zap = 0;
989 vdev_remove_child(ivd, vd);
990 vdev_compact_children(ivd);
992 vdev_indirect_state_transfer(ivd, vd);
994 svr->svr_vdev = ivd;
996 ASSERT(!ivd->vdev_removing);
997 ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
999 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1000 dsl_sync_task_nowait(spa->spa_dsl_pool, vdev_remove_complete_sync, svr,
1001 0, ZFS_SPACE_CHECK_NONE, tx);
1002 dmu_tx_commit(tx);
1005 * Indicate that this thread has exited.
1006 * After this, we can not use svr.
1008 mutex_enter(&svr->svr_lock);
1009 svr->svr_thread = NULL;
1010 cv_broadcast(&svr->svr_cv);
1011 mutex_exit(&svr->svr_lock);
1015 * Complete the removal of a toplevel vdev. This is called in open
1016 * context by the removal thread after we have copied all vdev's data.
1018 static void
1019 vdev_remove_complete(vdev_t *vd)
1021 spa_t *spa = vd->vdev_spa;
1022 uint64_t txg;
1025 * Wait for any deferred frees to be synced before we call
1026 * vdev_metaslab_fini()
1028 txg_wait_synced(spa->spa_dsl_pool, 0);
1030 txg = spa_vdev_enter(spa);
1031 zfs_dbgmsg("finishing device removal for vdev %llu in txg %llu",
1032 vd->vdev_id, txg);
1035 * Discard allocation state.
1037 if (vd->vdev_mg != NULL) {
1038 vdev_metaslab_fini(vd);
1039 metaslab_group_destroy(vd->vdev_mg);
1040 vd->vdev_mg = NULL;
1042 ASSERT0(vd->vdev_stat.vs_space);
1043 ASSERT0(vd->vdev_stat.vs_dspace);
1045 vdev_remove_replace_with_indirect(vd, txg);
1048 * We now release the locks, allowing spa_sync to run and finish the
1049 * removal via vdev_remove_complete_sync in syncing context.
1051 (void) spa_vdev_exit(spa, NULL, txg, 0);
1054 * Top ZAP should have been transferred to the indirect vdev in
1055 * vdev_remove_replace_with_indirect.
1057 ASSERT0(vd->vdev_top_zap);
1060 * Leaf ZAP should have been moved in vdev_remove_replace_with_indirect.
1062 ASSERT0(vd->vdev_leaf_zap);
1064 txg = spa_vdev_enter(spa);
1065 (void) vdev_label_init(vd, 0, VDEV_LABEL_REMOVE);
1067 * Request to update the config and the config cachefile.
1069 vdev_config_dirty(spa->spa_root_vdev);
1070 (void) spa_vdev_exit(spa, vd, txg, 0);
1074 * Evacuates a segment of size at most max_alloc from the vdev
1075 * via repeated calls to spa_vdev_copy_segment. If an allocation
1076 * fails, the pool is probably too fragmented to handle such a
1077 * large size, so decrease max_alloc so that the caller will not try
1078 * this size again this txg.
1080 static void
1081 spa_vdev_copy_impl(spa_vdev_removal_t *svr, vdev_copy_arg_t *vca,
1082 uint64_t *max_alloc, dmu_tx_t *tx)
1084 uint64_t txg = dmu_tx_get_txg(tx);
1085 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
1087 mutex_enter(&svr->svr_lock);
1089 range_seg_t *rs = avl_first(&svr->svr_allocd_segs->rt_root);
1090 if (rs == NULL) {
1091 mutex_exit(&svr->svr_lock);
1092 return;
1094 uint64_t offset = rs->rs_start;
1095 uint64_t length = MIN(rs->rs_end - rs->rs_start, *max_alloc);
1097 range_tree_remove(svr->svr_allocd_segs, offset, length);
1099 if (svr->svr_max_offset_to_sync[txg & TXG_MASK] == 0) {
1100 dsl_sync_task_nowait(dmu_tx_pool(tx), vdev_mapping_sync,
1101 svr, 0, ZFS_SPACE_CHECK_NONE, tx);
1104 svr->svr_max_offset_to_sync[txg & TXG_MASK] = offset + length;
1107 * Note: this is the amount of *allocated* space
1108 * that we are taking care of each txg.
1110 svr->svr_bytes_done[txg & TXG_MASK] += length;
1112 mutex_exit(&svr->svr_lock);
1114 zio_alloc_list_t zal;
1115 metaslab_trace_init(&zal);
1116 uint64_t thismax = *max_alloc;
1117 while (length > 0) {
1118 uint64_t mylen = MIN(length, thismax);
1120 int error = spa_vdev_copy_segment(svr->svr_vdev,
1121 offset, mylen, txg, vca, &zal);
1123 if (error == ENOSPC) {
1125 * Cut our segment in half, and don't try this
1126 * segment size again this txg. Note that the
1127 * allocation size must be aligned to the highest
1128 * ashift in the pool, so that the allocation will
1129 * not be padded out to a multiple of the ashift,
1130 * which could cause us to think that this mapping
1131 * is larger than we intended.
1133 ASSERT3U(spa->spa_max_ashift, >=, SPA_MINBLOCKSHIFT);
1134 ASSERT3U(spa->spa_max_ashift, ==, spa->spa_min_ashift);
1135 thismax = P2ROUNDUP(mylen / 2,
1136 1 << spa->spa_max_ashift);
1137 ASSERT3U(thismax, <, mylen);
1139 * The minimum-size allocation can not fail.
1141 ASSERT3U(mylen, >, 1 << spa->spa_max_ashift);
1142 *max_alloc = mylen - (1 << spa->spa_max_ashift);
1143 } else {
1144 ASSERT0(error);
1145 length -= mylen;
1146 offset += mylen;
1149 * We've performed an allocation, so reset the
1150 * alloc trace list.
1152 metaslab_trace_fini(&zal);
1153 metaslab_trace_init(&zal);
1156 metaslab_trace_fini(&zal);
1160 * The removal thread operates in open context. It iterates over all
1161 * allocated space in the vdev, by loading each metaslab's spacemap.
1162 * For each contiguous segment of allocated space (capping the segment
1163 * size at SPA_MAXBLOCKSIZE), we:
1164 * - Allocate space for it on another vdev.
1165 * - Create a new mapping from the old location to the new location
1166 * (as a record in svr_new_segments).
1167 * - Initiate a logical read zio to get the data off the removing disk.
1168 * - In the read zio's done callback, initiate a logical write zio to
1169 * write it to the new vdev.
1170 * Note that all of this will take effect when a particular TXG syncs.
1171 * The sync thread ensures that all the phys reads and writes for the syncing
1172 * TXG have completed (see spa_txg_zio) and writes the new mappings to disk
1173 * (see vdev_mapping_sync()).
1175 static void
1176 spa_vdev_remove_thread(void *arg)
1178 vdev_t *vd = arg;
1179 spa_t *spa = vd->vdev_spa;
1180 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1181 vdev_copy_arg_t vca;
1182 uint64_t max_alloc = zfs_remove_max_segment;
1183 uint64_t last_txg = 0;
1184 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
1185 uint64_t start_offset = vdev_indirect_mapping_max_offset(vim);
1187 ASSERT3P(vd->vdev_ops, !=, &vdev_indirect_ops);
1188 ASSERT(vdev_is_concrete(vd));
1189 ASSERT(vd->vdev_removing);
1190 ASSERT(vd->vdev_indirect_config.vic_mapping_object != 0);
1191 ASSERT3P(svr->svr_vdev, ==, vd);
1192 ASSERT(vim != NULL);
1194 mutex_init(&vca.vca_lock, NULL, MUTEX_DEFAULT, NULL);
1195 cv_init(&vca.vca_cv, NULL, CV_DEFAULT, NULL);
1196 vca.vca_outstanding_bytes = 0;
1198 mutex_enter(&svr->svr_lock);
1201 * Start from vim_max_offset so we pick up where we left off
1202 * if we are restarting the removal after opening the pool.
1204 uint64_t msi;
1205 for (msi = start_offset >> vd->vdev_ms_shift;
1206 msi < vd->vdev_ms_count && !svr->svr_thread_exit; msi++) {
1207 metaslab_t *msp = vd->vdev_ms[msi];
1208 ASSERT3U(msi, <=, vd->vdev_ms_count);
1210 ASSERT0(range_tree_space(svr->svr_allocd_segs));
1212 mutex_enter(&msp->ms_sync_lock);
1213 mutex_enter(&msp->ms_lock);
1216 * Assert nothing in flight -- ms_*tree is empty.
1218 for (int i = 0; i < TXG_SIZE; i++) {
1219 ASSERT0(range_tree_space(msp->ms_allocating[i]));
1223 * If the metaslab has ever been allocated from (ms_sm!=NULL),
1224 * read the allocated segments from the space map object
1225 * into svr_allocd_segs. Since we do this while holding
1226 * svr_lock and ms_sync_lock, concurrent frees (which
1227 * would have modified the space map) will wait for us
1228 * to finish loading the spacemap, and then take the
1229 * appropriate action (see free_from_removing_vdev()).
1231 if (msp->ms_sm != NULL) {
1232 space_map_t *sm = NULL;
1235 * We have to open a new space map here, because
1236 * ms_sm's sm_length and sm_alloc may not reflect
1237 * what's in the object contents, if we are in between
1238 * metaslab_sync() and metaslab_sync_done().
1240 VERIFY0(space_map_open(&sm,
1241 spa->spa_dsl_pool->dp_meta_objset,
1242 msp->ms_sm->sm_object, msp->ms_sm->sm_start,
1243 msp->ms_sm->sm_size, msp->ms_sm->sm_shift));
1244 space_map_update(sm);
1245 VERIFY0(space_map_load(sm, svr->svr_allocd_segs,
1246 SM_ALLOC));
1247 space_map_close(sm);
1249 range_tree_walk(msp->ms_freeing,
1250 range_tree_remove, svr->svr_allocd_segs);
1253 * When we are resuming from a paused removal (i.e.
1254 * when importing a pool with a removal in progress),
1255 * discard any state that we have already processed.
1257 range_tree_clear(svr->svr_allocd_segs, 0, start_offset);
1259 mutex_exit(&msp->ms_lock);
1260 mutex_exit(&msp->ms_sync_lock);
1262 vca.vca_msp = msp;
1263 zfs_dbgmsg("copying %llu segments for metaslab %llu",
1264 avl_numnodes(&svr->svr_allocd_segs->rt_root),
1265 msp->ms_id);
1267 while (!svr->svr_thread_exit &&
1268 !range_tree_is_empty(svr->svr_allocd_segs)) {
1270 mutex_exit(&svr->svr_lock);
1273 * This delay will pause the removal around the point
1274 * specified by zfs_remove_max_bytes_pause. We do this
1275 * solely from the test suite or during debugging.
1277 uint64_t bytes_copied =
1278 spa->spa_removing_phys.sr_copied;
1279 for (int i = 0; i < TXG_SIZE; i++)
1280 bytes_copied += svr->svr_bytes_done[i];
1281 while (zfs_remove_max_bytes_pause <= bytes_copied &&
1282 !svr->svr_thread_exit)
1283 delay(hz);
1285 mutex_enter(&vca.vca_lock);
1286 while (vca.vca_outstanding_bytes >
1287 zfs_remove_max_copy_bytes) {
1288 cv_wait(&vca.vca_cv, &vca.vca_lock);
1290 mutex_exit(&vca.vca_lock);
1292 dmu_tx_t *tx =
1293 dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
1295 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
1296 uint64_t txg = dmu_tx_get_txg(tx);
1298 if (txg != last_txg)
1299 max_alloc = zfs_remove_max_segment;
1300 last_txg = txg;
1302 spa_vdev_copy_impl(svr, &vca, &max_alloc, tx);
1304 dmu_tx_commit(tx);
1305 mutex_enter(&svr->svr_lock);
1309 mutex_exit(&svr->svr_lock);
1311 * Wait for all copies to finish before cleaning up the vca.
1313 txg_wait_synced(spa->spa_dsl_pool, 0);
1314 ASSERT0(vca.vca_outstanding_bytes);
1316 mutex_destroy(&vca.vca_lock);
1317 cv_destroy(&vca.vca_cv);
1319 if (svr->svr_thread_exit) {
1320 mutex_enter(&svr->svr_lock);
1321 range_tree_vacate(svr->svr_allocd_segs, NULL, NULL);
1322 svr->svr_thread = NULL;
1323 cv_broadcast(&svr->svr_cv);
1324 mutex_exit(&svr->svr_lock);
1325 } else {
1326 ASSERT0(range_tree_space(svr->svr_allocd_segs));
1327 vdev_remove_complete(vd);
1331 void
1332 spa_vdev_remove_suspend(spa_t *spa)
1334 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1336 if (svr == NULL)
1337 return;
1339 mutex_enter(&svr->svr_lock);
1340 svr->svr_thread_exit = B_TRUE;
1341 while (svr->svr_thread != NULL)
1342 cv_wait(&svr->svr_cv, &svr->svr_lock);
1343 svr->svr_thread_exit = B_FALSE;
1344 mutex_exit(&svr->svr_lock);
1347 /* ARGSUSED */
1348 static int
1349 spa_vdev_remove_cancel_check(void *arg, dmu_tx_t *tx)
1351 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
1353 if (spa->spa_vdev_removal == NULL)
1354 return (ENOTACTIVE);
1355 return (0);
1359 * Cancel a removal by freeing all entries from the partial mapping
1360 * and marking the vdev as no longer being removing.
1362 /* ARGSUSED */
1363 static void
1364 spa_vdev_remove_cancel_sync(void *arg, dmu_tx_t *tx)
1366 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
1367 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1368 vdev_t *vd = svr->svr_vdev;
1369 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
1370 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
1371 objset_t *mos = spa->spa_meta_objset;
1373 ASSERT3P(svr->svr_thread, ==, NULL);
1375 spa_feature_decr(spa, SPA_FEATURE_DEVICE_REMOVAL, tx);
1376 if (vdev_obsolete_counts_are_precise(vd)) {
1377 spa_feature_decr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
1378 VERIFY0(zap_remove(spa->spa_meta_objset, vd->vdev_top_zap,
1379 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE, tx));
1382 if (vdev_obsolete_sm_object(vd) != 0) {
1383 ASSERT(vd->vdev_obsolete_sm != NULL);
1384 ASSERT3U(vdev_obsolete_sm_object(vd), ==,
1385 space_map_object(vd->vdev_obsolete_sm));
1387 space_map_free(vd->vdev_obsolete_sm, tx);
1388 VERIFY0(zap_remove(spa->spa_meta_objset, vd->vdev_top_zap,
1389 VDEV_TOP_ZAP_INDIRECT_OBSOLETE_SM, tx));
1390 space_map_close(vd->vdev_obsolete_sm);
1391 vd->vdev_obsolete_sm = NULL;
1392 spa_feature_decr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
1394 for (int i = 0; i < TXG_SIZE; i++) {
1395 ASSERT(list_is_empty(&svr->svr_new_segments[i]));
1396 ASSERT3U(svr->svr_max_offset_to_sync[i], <=,
1397 vdev_indirect_mapping_max_offset(vim));
1400 for (uint64_t msi = 0; msi < vd->vdev_ms_count; msi++) {
1401 metaslab_t *msp = vd->vdev_ms[msi];
1403 if (msp->ms_start >= vdev_indirect_mapping_max_offset(vim))
1404 break;
1406 ASSERT0(range_tree_space(svr->svr_allocd_segs));
1408 mutex_enter(&msp->ms_lock);
1411 * Assert nothing in flight -- ms_*tree is empty.
1413 for (int i = 0; i < TXG_SIZE; i++)
1414 ASSERT0(range_tree_space(msp->ms_allocating[i]));
1415 for (int i = 0; i < TXG_DEFER_SIZE; i++)
1416 ASSERT0(range_tree_space(msp->ms_defer[i]));
1417 ASSERT0(range_tree_space(msp->ms_freed));
1419 if (msp->ms_sm != NULL) {
1421 * Assert that the in-core spacemap has the same
1422 * length as the on-disk one, so we can use the
1423 * existing in-core spacemap to load it from disk.
1425 ASSERT3U(msp->ms_sm->sm_alloc, ==,
1426 msp->ms_sm->sm_phys->smp_alloc);
1427 ASSERT3U(msp->ms_sm->sm_length, ==,
1428 msp->ms_sm->sm_phys->smp_objsize);
1430 mutex_enter(&svr->svr_lock);
1431 VERIFY0(space_map_load(msp->ms_sm,
1432 svr->svr_allocd_segs, SM_ALLOC));
1433 range_tree_walk(msp->ms_freeing,
1434 range_tree_remove, svr->svr_allocd_segs);
1437 * Clear everything past what has been synced,
1438 * because we have not allocated mappings for it yet.
1440 uint64_t syncd = vdev_indirect_mapping_max_offset(vim);
1441 range_tree_clear(svr->svr_allocd_segs, syncd,
1442 msp->ms_sm->sm_start + msp->ms_sm->sm_size - syncd);
1444 mutex_exit(&svr->svr_lock);
1446 mutex_exit(&msp->ms_lock);
1448 mutex_enter(&svr->svr_lock);
1449 range_tree_vacate(svr->svr_allocd_segs,
1450 free_mapped_segment_cb, vd);
1451 mutex_exit(&svr->svr_lock);
1455 * Note: this must happen after we invoke free_mapped_segment_cb,
1456 * because it adds to the obsolete_segments.
1458 range_tree_vacate(vd->vdev_obsolete_segments, NULL, NULL);
1460 ASSERT3U(vic->vic_mapping_object, ==,
1461 vdev_indirect_mapping_object(vd->vdev_indirect_mapping));
1462 vdev_indirect_mapping_close(vd->vdev_indirect_mapping);
1463 vd->vdev_indirect_mapping = NULL;
1464 vdev_indirect_mapping_free(mos, vic->vic_mapping_object, tx);
1465 vic->vic_mapping_object = 0;
1467 ASSERT3U(vic->vic_births_object, ==,
1468 vdev_indirect_births_object(vd->vdev_indirect_births));
1469 vdev_indirect_births_close(vd->vdev_indirect_births);
1470 vd->vdev_indirect_births = NULL;
1471 vdev_indirect_births_free(mos, vic->vic_births_object, tx);
1472 vic->vic_births_object = 0;
1475 * We may have processed some frees from the removing vdev in this
1476 * txg, thus increasing svr_bytes_done; discard that here to
1477 * satisfy the assertions in spa_vdev_removal_destroy().
1478 * Note that future txg's can not have any bytes_done, because
1479 * future TXG's are only modified from open context, and we have
1480 * already shut down the copying thread.
1482 svr->svr_bytes_done[dmu_tx_get_txg(tx) & TXG_MASK] = 0;
1483 spa_finish_removal(spa, DSS_CANCELED, tx);
1485 vd->vdev_removing = B_FALSE;
1486 vdev_config_dirty(vd);
1488 zfs_dbgmsg("canceled device removal for vdev %llu in %llu",
1489 vd->vdev_id, dmu_tx_get_txg(tx));
1490 spa_history_log_internal(spa, "vdev remove canceled", tx,
1491 "%s vdev %llu %s", spa_name(spa),
1492 vd->vdev_id, (vd->vdev_path != NULL) ? vd->vdev_path : "-");
1496 spa_vdev_remove_cancel(spa_t *spa)
1498 spa_vdev_remove_suspend(spa);
1500 if (spa->spa_vdev_removal == NULL)
1501 return (ENOTACTIVE);
1503 uint64_t vdid = spa->spa_vdev_removal->svr_vdev->vdev_id;
1505 int error = dsl_sync_task(spa->spa_name, spa_vdev_remove_cancel_check,
1506 spa_vdev_remove_cancel_sync, NULL, 0,
1507 ZFS_SPACE_CHECK_EXTRA_RESERVED);
1509 if (error == 0) {
1510 spa_config_enter(spa, SCL_ALLOC | SCL_VDEV, FTAG, RW_WRITER);
1511 vdev_t *vd = vdev_lookup_top(spa, vdid);
1512 metaslab_group_activate(vd->vdev_mg);
1513 spa_config_exit(spa, SCL_ALLOC | SCL_VDEV, FTAG);
1516 return (error);
1520 * Called every sync pass of every txg if there's a svr.
1522 void
1523 svr_sync(spa_t *spa, dmu_tx_t *tx)
1525 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1526 int txgoff = dmu_tx_get_txg(tx) & TXG_MASK;
1529 * This check is necessary so that we do not dirty the
1530 * DIRECTORY_OBJECT via spa_sync_removing_state() when there
1531 * is nothing to do. Dirtying it every time would prevent us
1532 * from syncing-to-convergence.
1534 if (svr->svr_bytes_done[txgoff] == 0)
1535 return;
1538 * Update progress accounting.
1540 spa->spa_removing_phys.sr_copied += svr->svr_bytes_done[txgoff];
1541 svr->svr_bytes_done[txgoff] = 0;
1543 spa_sync_removing_state(spa, tx);
1546 static void
1547 vdev_remove_make_hole_and_free(vdev_t *vd)
1549 uint64_t id = vd->vdev_id;
1550 spa_t *spa = vd->vdev_spa;
1551 vdev_t *rvd = spa->spa_root_vdev;
1552 boolean_t last_vdev = (id == (rvd->vdev_children - 1));
1554 ASSERT(MUTEX_HELD(&spa_namespace_lock));
1555 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1557 vdev_free(vd);
1559 if (last_vdev) {
1560 vdev_compact_children(rvd);
1561 } else {
1562 vd = vdev_alloc_common(spa, id, 0, &vdev_hole_ops);
1563 vdev_add_child(rvd, vd);
1565 vdev_config_dirty(rvd);
1568 * Reassess the health of our root vdev.
1570 vdev_reopen(rvd);
1574 * Remove a log device. The config lock is held for the specified TXG.
1576 static int
1577 spa_vdev_remove_log(vdev_t *vd, uint64_t *txg)
1579 metaslab_group_t *mg = vd->vdev_mg;
1580 spa_t *spa = vd->vdev_spa;
1581 int error = 0;
1583 ASSERT(vd->vdev_islog);
1584 ASSERT(vd == vd->vdev_top);
1587 * Stop allocating from this vdev.
1589 metaslab_group_passivate(mg);
1592 * Wait for the youngest allocations and frees to sync,
1593 * and then wait for the deferral of those frees to finish.
1595 spa_vdev_config_exit(spa, NULL,
1596 *txg + TXG_CONCURRENT_STATES + TXG_DEFER_SIZE, 0, FTAG);
1599 * Evacuate the device. We don't hold the config lock as writer
1600 * since we need to do I/O but we do keep the
1601 * spa_namespace_lock held. Once this completes the device
1602 * should no longer have any blocks allocated on it.
1604 if (vd->vdev_islog) {
1605 if (vd->vdev_stat.vs_alloc != 0)
1606 error = spa_reset_logs(spa);
1609 *txg = spa_vdev_config_enter(spa);
1611 if (error != 0) {
1612 metaslab_group_activate(mg);
1613 return (error);
1615 ASSERT0(vd->vdev_stat.vs_alloc);
1618 * The evacuation succeeded. Remove any remaining MOS metadata
1619 * associated with this vdev, and wait for these changes to sync.
1621 vd->vdev_removing = B_TRUE;
1623 vdev_dirty_leaves(vd, VDD_DTL, *txg);
1624 vdev_config_dirty(vd);
1626 spa_history_log_internal(spa, "vdev remove", NULL,
1627 "%s vdev %llu (log) %s", spa_name(spa), vd->vdev_id,
1628 (vd->vdev_path != NULL) ? vd->vdev_path : "-");
1630 /* Make sure these changes are sync'ed */
1631 spa_vdev_config_exit(spa, NULL, *txg, 0, FTAG);
1633 *txg = spa_vdev_config_enter(spa);
1635 sysevent_t *ev = spa_event_create(spa, vd, NULL,
1636 ESC_ZFS_VDEV_REMOVE_DEV);
1637 ASSERT(MUTEX_HELD(&spa_namespace_lock));
1638 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1640 /* The top ZAP should have been destroyed by vdev_remove_empty. */
1641 ASSERT0(vd->vdev_top_zap);
1642 /* The leaf ZAP should have been destroyed by vdev_dtl_sync. */
1643 ASSERT0(vd->vdev_leaf_zap);
1645 (void) vdev_label_init(vd, 0, VDEV_LABEL_REMOVE);
1647 if (list_link_active(&vd->vdev_state_dirty_node))
1648 vdev_state_clean(vd);
1649 if (list_link_active(&vd->vdev_config_dirty_node))
1650 vdev_config_clean(vd);
1653 * Clean up the vdev namespace.
1655 vdev_remove_make_hole_and_free(vd);
1657 if (ev != NULL)
1658 spa_event_post(ev);
1660 return (0);
1663 static int
1664 spa_vdev_remove_top_check(vdev_t *vd)
1666 spa_t *spa = vd->vdev_spa;
1668 if (vd != vd->vdev_top)
1669 return (SET_ERROR(ENOTSUP));
1671 if (!spa_feature_is_enabled(spa, SPA_FEATURE_DEVICE_REMOVAL))
1672 return (SET_ERROR(ENOTSUP));
1675 * There has to be enough free space to remove the
1676 * device and leave double the "slop" space (i.e. we
1677 * must leave at least 3% of the pool free, in addition to
1678 * the normal slop space).
1680 if (dsl_dir_space_available(spa->spa_dsl_pool->dp_root_dir,
1681 NULL, 0, B_TRUE) <
1682 vd->vdev_stat.vs_dspace + spa_get_slop_space(spa)) {
1683 return (SET_ERROR(ENOSPC));
1687 * There can not be a removal in progress.
1689 if (spa->spa_removing_phys.sr_state == DSS_SCANNING)
1690 return (SET_ERROR(EBUSY));
1693 * The device must have all its data.
1695 if (!vdev_dtl_empty(vd, DTL_MISSING) ||
1696 !vdev_dtl_empty(vd, DTL_OUTAGE))
1697 return (SET_ERROR(EBUSY));
1700 * The device must be healthy.
1702 if (!vdev_readable(vd))
1703 return (SET_ERROR(EIO));
1706 * All vdevs in normal class must have the same ashift.
1708 if (spa->spa_max_ashift != spa->spa_min_ashift) {
1709 return (SET_ERROR(EINVAL));
1713 * All vdevs in normal class must have the same ashift
1714 * and not be raidz.
1716 vdev_t *rvd = spa->spa_root_vdev;
1717 int num_indirect = 0;
1718 for (uint64_t id = 0; id < rvd->vdev_children; id++) {
1719 vdev_t *cvd = rvd->vdev_child[id];
1720 if (cvd->vdev_ashift != 0 && !cvd->vdev_islog)
1721 ASSERT3U(cvd->vdev_ashift, ==, spa->spa_max_ashift);
1722 if (cvd->vdev_ops == &vdev_indirect_ops)
1723 num_indirect++;
1724 if (!vdev_is_concrete(cvd))
1725 continue;
1726 if (cvd->vdev_ops == &vdev_raidz_ops)
1727 return (SET_ERROR(EINVAL));
1729 * Need the mirror to be mirror of leaf vdevs only
1731 if (cvd->vdev_ops == &vdev_mirror_ops) {
1732 for (uint64_t cid = 0;
1733 cid < cvd->vdev_children; cid++) {
1734 vdev_t *tmp = cvd->vdev_child[cid];
1735 if (!tmp->vdev_ops->vdev_op_leaf)
1736 return (SET_ERROR(EINVAL));
1741 return (0);
1745 * Initiate removal of a top-level vdev, reducing the total space in the pool.
1746 * The config lock is held for the specified TXG. Once initiated,
1747 * evacuation of all allocated space (copying it to other vdevs) happens
1748 * in the background (see spa_vdev_remove_thread()), and can be canceled
1749 * (see spa_vdev_remove_cancel()). If successful, the vdev will
1750 * be transformed to an indirect vdev (see spa_vdev_remove_complete()).
1752 static int
1753 spa_vdev_remove_top(vdev_t *vd, uint64_t *txg)
1755 spa_t *spa = vd->vdev_spa;
1756 int error;
1759 * Check for errors up-front, so that we don't waste time
1760 * passivating the metaslab group and clearing the ZIL if there
1761 * are errors.
1763 error = spa_vdev_remove_top_check(vd);
1764 if (error != 0)
1765 return (error);
1768 * Stop allocating from this vdev. Note that we must check
1769 * that this is not the only device in the pool before
1770 * passivating, otherwise we will not be able to make
1771 * progress because we can't allocate from any vdevs.
1772 * The above check for sufficient free space serves this
1773 * purpose.
1775 metaslab_group_t *mg = vd->vdev_mg;
1776 metaslab_group_passivate(mg);
1779 * Wait for the youngest allocations and frees to sync,
1780 * and then wait for the deferral of those frees to finish.
1782 spa_vdev_config_exit(spa, NULL,
1783 *txg + TXG_CONCURRENT_STATES + TXG_DEFER_SIZE, 0, FTAG);
1786 * We must ensure that no "stubby" log blocks are allocated
1787 * on the device to be removed. These blocks could be
1788 * written at any time, including while we are in the middle
1789 * of copying them.
1791 error = spa_reset_logs(spa);
1793 *txg = spa_vdev_config_enter(spa);
1796 * Things might have changed while the config lock was dropped
1797 * (e.g. space usage). Check for errors again.
1799 if (error == 0)
1800 error = spa_vdev_remove_top_check(vd);
1802 if (error != 0) {
1803 metaslab_group_activate(mg);
1804 return (error);
1807 vd->vdev_removing = B_TRUE;
1809 vdev_dirty_leaves(vd, VDD_DTL, *txg);
1810 vdev_config_dirty(vd);
1811 dmu_tx_t *tx = dmu_tx_create_assigned(spa->spa_dsl_pool, *txg);
1812 dsl_sync_task_nowait(spa->spa_dsl_pool,
1813 vdev_remove_initiate_sync,
1814 vd, 0, ZFS_SPACE_CHECK_NONE, tx);
1815 dmu_tx_commit(tx);
1817 return (0);
1821 * Remove a device from the pool.
1823 * Removing a device from the vdev namespace requires several steps
1824 * and can take a significant amount of time. As a result we use
1825 * the spa_vdev_config_[enter/exit] functions which allow us to
1826 * grab and release the spa_config_lock while still holding the namespace
1827 * lock. During each step the configuration is synced out.
1830 spa_vdev_remove(spa_t *spa, uint64_t guid, boolean_t unspare)
1832 vdev_t *vd;
1833 nvlist_t **spares, **l2cache, *nv;
1834 uint64_t txg = 0;
1835 uint_t nspares, nl2cache;
1836 int error = 0;
1837 boolean_t locked = MUTEX_HELD(&spa_namespace_lock);
1838 sysevent_t *ev = NULL;
1840 ASSERT(spa_writeable(spa));
1842 if (!locked)
1843 txg = spa_vdev_enter(spa);
1845 ASSERT(MUTEX_HELD(&spa_namespace_lock));
1846 if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
1847 error = (spa_has_checkpoint(spa)) ?
1848 ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT;
1850 if (!locked)
1851 return (spa_vdev_exit(spa, NULL, txg, error));
1853 return (error);
1856 vd = spa_lookup_by_guid(spa, guid, B_FALSE);
1858 if (spa->spa_spares.sav_vdevs != NULL &&
1859 nvlist_lookup_nvlist_array(spa->spa_spares.sav_config,
1860 ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0 &&
1861 (nv = spa_nvlist_lookup_by_guid(spares, nspares, guid)) != NULL) {
1863 * Only remove the hot spare if it's not currently in use
1864 * in this pool.
1866 if (vd == NULL || unspare) {
1867 char *nvstr = fnvlist_lookup_string(nv,
1868 ZPOOL_CONFIG_PATH);
1869 spa_history_log_internal(spa, "vdev remove", NULL,
1870 "%s vdev (%s) %s", spa_name(spa),
1871 VDEV_TYPE_SPARE, nvstr);
1872 if (vd == NULL)
1873 vd = spa_lookup_by_guid(spa, guid, B_TRUE);
1874 ev = spa_event_create(spa, vd, NULL,
1875 ESC_ZFS_VDEV_REMOVE_AUX);
1876 spa_vdev_remove_aux(spa->spa_spares.sav_config,
1877 ZPOOL_CONFIG_SPARES, spares, nspares, nv);
1878 spa_load_spares(spa);
1879 spa->spa_spares.sav_sync = B_TRUE;
1880 } else {
1881 error = SET_ERROR(EBUSY);
1883 } else if (spa->spa_l2cache.sav_vdevs != NULL &&
1884 nvlist_lookup_nvlist_array(spa->spa_l2cache.sav_config,
1885 ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0 &&
1886 (nv = spa_nvlist_lookup_by_guid(l2cache, nl2cache, guid)) != NULL) {
1887 char *nvstr = fnvlist_lookup_string(nv, ZPOOL_CONFIG_PATH);
1888 spa_history_log_internal(spa, "vdev remove", NULL,
1889 "%s vdev (%s) %s", spa_name(spa), VDEV_TYPE_L2CACHE, nvstr);
1891 * Cache devices can always be removed.
1893 vd = spa_lookup_by_guid(spa, guid, B_TRUE);
1894 ev = spa_event_create(spa, vd, NULL, ESC_ZFS_VDEV_REMOVE_AUX);
1895 spa_vdev_remove_aux(spa->spa_l2cache.sav_config,
1896 ZPOOL_CONFIG_L2CACHE, l2cache, nl2cache, nv);
1897 spa_load_l2cache(spa);
1898 spa->spa_l2cache.sav_sync = B_TRUE;
1899 } else if (vd != NULL && vd->vdev_islog) {
1900 ASSERT(!locked);
1901 error = spa_vdev_remove_log(vd, &txg);
1902 } else if (vd != NULL) {
1903 ASSERT(!locked);
1904 error = spa_vdev_remove_top(vd, &txg);
1905 } else {
1907 * There is no vdev of any kind with the specified guid.
1909 error = SET_ERROR(ENOENT);
1912 if (!locked)
1913 error = spa_vdev_exit(spa, NULL, txg, error);
1915 if (ev != NULL) {
1916 if (error != 0) {
1917 spa_event_discard(ev);
1918 } else {
1919 spa_event_post(ev);
1923 return (error);
1927 spa_removal_get_stats(spa_t *spa, pool_removal_stat_t *prs)
1929 prs->prs_state = spa->spa_removing_phys.sr_state;
1931 if (prs->prs_state == DSS_NONE)
1932 return (SET_ERROR(ENOENT));
1934 prs->prs_removing_vdev = spa->spa_removing_phys.sr_removing_vdev;
1935 prs->prs_start_time = spa->spa_removing_phys.sr_start_time;
1936 prs->prs_end_time = spa->spa_removing_phys.sr_end_time;
1937 prs->prs_to_copy = spa->spa_removing_phys.sr_to_copy;
1938 prs->prs_copied = spa->spa_removing_phys.sr_copied;
1940 if (spa->spa_vdev_removal != NULL) {
1941 for (int i = 0; i < TXG_SIZE; i++) {
1942 prs->prs_copied +=
1943 spa->spa_vdev_removal->svr_bytes_done[i];
1947 prs->prs_mapping_memory = 0;
1948 uint64_t indirect_vdev_id =
1949 spa->spa_removing_phys.sr_prev_indirect_vdev;
1950 while (indirect_vdev_id != -1) {
1951 vdev_t *vd = spa->spa_root_vdev->vdev_child[indirect_vdev_id];
1952 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
1953 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
1955 ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
1956 prs->prs_mapping_memory += vdev_indirect_mapping_size(vim);
1957 indirect_vdev_id = vic->vic_prev_indirect_vdev;
1960 return (0);