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]
23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2011, 2020 by Delphix. All rights reserved.
25 * Copyright (c) 2019, loli10K <ezomori.nozomu@gmail.com>. All rights reserved.
28 #include <sys/zfs_context.h>
29 #include <sys/spa_impl.h>
31 #include <sys/dmu_tx.h>
33 #include <sys/vdev_impl.h>
34 #include <sys/metaslab.h>
35 #include <sys/metaslab_impl.h>
36 #include <sys/uberblock_impl.h>
39 #include <sys/bpobj.h>
40 #include <sys/dsl_pool.h>
41 #include <sys/dsl_synctask.h>
42 #include <sys/dsl_dir.h>
44 #include <sys/zfeature.h>
45 #include <sys/vdev_indirect_births.h>
46 #include <sys/vdev_indirect_mapping.h>
48 #include <sys/vdev_initialize.h>
49 #include <sys/vdev_trim.h>
50 #include <sys/trace_zfs.h>
53 * This file contains the necessary logic to remove vdevs from a
54 * storage pool. Currently, the only devices that can be removed
55 * are log, cache, and spare devices; and top level vdevs from a pool
56 * w/o raidz or mirrors. (Note that members of a mirror can be removed
57 * by the detach operation.)
59 * Log vdevs are removed by evacuating them and then turning the vdev
60 * into a hole vdev while holding spa config locks.
62 * Top level vdevs are removed and converted into an indirect vdev via
63 * a multi-step process:
65 * - Disable allocations from this device (spa_vdev_remove_top).
67 * - From a new thread (spa_vdev_remove_thread), copy data from
68 * the removing vdev to a different vdev. The copy happens in open
69 * context (spa_vdev_copy_impl) and issues a sync task
70 * (vdev_mapping_sync) so the sync thread can update the partial
71 * indirect mappings in core and on disk.
73 * - If a free happens during a removal, it is freed from the
74 * removing vdev, and if it has already been copied, from the new
75 * location as well (free_from_removing_vdev).
77 * - After the removal is completed, the copy thread converts the vdev
78 * into an indirect vdev (vdev_remove_complete) before instructing
79 * the sync thread to destroy the space maps and finish the removal
80 * (spa_finish_removal).
83 typedef struct vdev_copy_arg {
85 uint64_t vca_outstanding_bytes;
86 uint64_t vca_read_error_bytes;
87 uint64_t vca_write_error_bytes;
93 * The maximum amount of memory we can use for outstanding i/o while
94 * doing a device removal. This determines how much i/o we can have
95 * in flight concurrently.
97 int zfs_remove_max_copy_bytes = 64 * 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 * See also the accessor function spa_remove_max_segment().
107 int zfs_remove_max_segment = SPA_MAXBLOCKSIZE;
110 * Ignore hard IO errors during device removal. When set if a device
111 * encounters hard IO error during the removal process the removal will
112 * not be cancelled. This can result in a normally recoverable block
113 * becoming permanently damaged and is not recommended.
115 int zfs_removal_ignore_errors = 0;
118 * Allow a remap segment to span free chunks of at most this size. The main
119 * impact of a larger span is that we will read and write larger, more
120 * contiguous chunks, with more "unnecessary" data -- trading off bandwidth
121 * for iops. The value here was chosen to align with
122 * zfs_vdev_read_gap_limit, which is a similar concept when doing regular
123 * reads (but there's no reason it has to be the same).
125 * Additionally, a higher span will have the following relatively minor
127 * - the mapping will be smaller, since one entry can cover more allocated
129 * - more of the fragmentation in the removing device will be preserved
130 * - we'll do larger allocations, which may fail and fall back on smaller
133 int vdev_removal_max_span = 32 * 1024;
136 * This is used by the test suite so that it can ensure that certain
137 * actions happen while in the middle of a removal.
139 int zfs_removal_suspend_progress = 0;
141 #define VDEV_REMOVAL_ZAP_OBJS "lzap"
143 static void spa_vdev_remove_thread(void *arg);
144 static int spa_vdev_remove_cancel_impl(spa_t *spa);
147 spa_sync_removing_state(spa_t *spa, dmu_tx_t *tx)
149 VERIFY0(zap_update(spa->spa_dsl_pool->dp_meta_objset,
150 DMU_POOL_DIRECTORY_OBJECT,
151 DMU_POOL_REMOVING, sizeof (uint64_t),
152 sizeof (spa->spa_removing_phys) / sizeof (uint64_t),
153 &spa->spa_removing_phys, tx));
157 spa_nvlist_lookup_by_guid(nvlist_t **nvpp, int count, uint64_t target_guid)
159 for (int i = 0; i < count; i++) {
161 fnvlist_lookup_uint64(nvpp[i], ZPOOL_CONFIG_GUID);
163 if (guid == target_guid)
171 spa_vdev_remove_aux(nvlist_t *config, char *name, nvlist_t **dev, int count,
172 nvlist_t *dev_to_remove)
174 nvlist_t **newdev = NULL;
177 newdev = kmem_alloc((count - 1) * sizeof (void *), KM_SLEEP);
179 for (int i = 0, j = 0; i < count; i++) {
180 if (dev[i] == dev_to_remove)
182 VERIFY(nvlist_dup(dev[i], &newdev[j++], KM_SLEEP) == 0);
185 VERIFY(nvlist_remove(config, name, DATA_TYPE_NVLIST_ARRAY) == 0);
186 VERIFY(nvlist_add_nvlist_array(config, name, newdev, count - 1) == 0);
188 for (int i = 0; i < count - 1; i++)
189 nvlist_free(newdev[i]);
192 kmem_free(newdev, (count - 1) * sizeof (void *));
195 static spa_vdev_removal_t *
196 spa_vdev_removal_create(vdev_t *vd)
198 spa_vdev_removal_t *svr = kmem_zalloc(sizeof (*svr), KM_SLEEP);
199 mutex_init(&svr->svr_lock, NULL, MUTEX_DEFAULT, NULL);
200 cv_init(&svr->svr_cv, NULL, CV_DEFAULT, NULL);
201 svr->svr_allocd_segs = range_tree_create(NULL, RANGE_SEG64, NULL, 0, 0);
202 svr->svr_vdev_id = vd->vdev_id;
204 for (int i = 0; i < TXG_SIZE; i++) {
205 svr->svr_frees[i] = range_tree_create(NULL, RANGE_SEG64, NULL,
207 list_create(&svr->svr_new_segments[i],
208 sizeof (vdev_indirect_mapping_entry_t),
209 offsetof(vdev_indirect_mapping_entry_t, vime_node));
216 spa_vdev_removal_destroy(spa_vdev_removal_t *svr)
218 for (int i = 0; i < TXG_SIZE; i++) {
219 ASSERT0(svr->svr_bytes_done[i]);
220 ASSERT0(svr->svr_max_offset_to_sync[i]);
221 range_tree_destroy(svr->svr_frees[i]);
222 list_destroy(&svr->svr_new_segments[i]);
225 range_tree_destroy(svr->svr_allocd_segs);
226 mutex_destroy(&svr->svr_lock);
227 cv_destroy(&svr->svr_cv);
228 kmem_free(svr, sizeof (*svr));
232 * This is called as a synctask in the txg in which we will mark this vdev
233 * as removing (in the config stored in the MOS).
235 * It begins the evacuation of a toplevel vdev by:
236 * - initializing the spa_removing_phys which tracks this removal
237 * - computing the amount of space to remove for accounting purposes
238 * - dirtying all dbufs in the spa_config_object
239 * - creating the spa_vdev_removal
240 * - starting the spa_vdev_remove_thread
243 vdev_remove_initiate_sync(void *arg, dmu_tx_t *tx)
245 int vdev_id = (uintptr_t)arg;
246 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
247 vdev_t *vd = vdev_lookup_top(spa, vdev_id);
248 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
249 objset_t *mos = spa->spa_dsl_pool->dp_meta_objset;
250 spa_vdev_removal_t *svr = NULL;
251 uint64_t txg __maybe_unused = dmu_tx_get_txg(tx);
253 ASSERT3P(vd->vdev_ops, !=, &vdev_raidz_ops);
254 svr = spa_vdev_removal_create(vd);
256 ASSERT(vd->vdev_removing);
257 ASSERT3P(vd->vdev_indirect_mapping, ==, NULL);
259 spa_feature_incr(spa, SPA_FEATURE_DEVICE_REMOVAL, tx);
260 if (spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS)) {
262 * By activating the OBSOLETE_COUNTS feature, we prevent
263 * the pool from being downgraded and ensure that the
264 * refcounts are precise.
266 spa_feature_incr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
268 VERIFY0(zap_add(spa->spa_meta_objset, vd->vdev_top_zap,
269 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE, sizeof (one), 1,
271 boolean_t are_precise __maybe_unused;
272 ASSERT0(vdev_obsolete_counts_are_precise(vd, &are_precise));
273 ASSERT3B(are_precise, ==, B_TRUE);
276 vic->vic_mapping_object = vdev_indirect_mapping_alloc(mos, tx);
277 vd->vdev_indirect_mapping =
278 vdev_indirect_mapping_open(mos, vic->vic_mapping_object);
279 vic->vic_births_object = vdev_indirect_births_alloc(mos, tx);
280 vd->vdev_indirect_births =
281 vdev_indirect_births_open(mos, vic->vic_births_object);
282 spa->spa_removing_phys.sr_removing_vdev = vd->vdev_id;
283 spa->spa_removing_phys.sr_start_time = gethrestime_sec();
284 spa->spa_removing_phys.sr_end_time = 0;
285 spa->spa_removing_phys.sr_state = DSS_SCANNING;
286 spa->spa_removing_phys.sr_to_copy = 0;
287 spa->spa_removing_phys.sr_copied = 0;
290 * Note: We can't use vdev_stat's vs_alloc for sr_to_copy, because
291 * there may be space in the defer tree, which is free, but still
292 * counted in vs_alloc.
294 for (uint64_t i = 0; i < vd->vdev_ms_count; i++) {
295 metaslab_t *ms = vd->vdev_ms[i];
296 if (ms->ms_sm == NULL)
299 spa->spa_removing_phys.sr_to_copy +=
300 metaslab_allocated_space(ms);
303 * Space which we are freeing this txg does not need to
306 spa->spa_removing_phys.sr_to_copy -=
307 range_tree_space(ms->ms_freeing);
309 ASSERT0(range_tree_space(ms->ms_freed));
310 for (int t = 0; t < TXG_SIZE; t++)
311 ASSERT0(range_tree_space(ms->ms_allocating[t]));
315 * Sync tasks are called before metaslab_sync(), so there should
316 * be no already-synced metaslabs in the TXG_CLEAN list.
318 ASSERT3P(txg_list_head(&vd->vdev_ms_list, TXG_CLEAN(txg)), ==, NULL);
320 spa_sync_removing_state(spa, tx);
323 * All blocks that we need to read the most recent mapping must be
324 * stored on concrete vdevs. Therefore, we must dirty anything that
325 * is read before spa_remove_init(). Specifically, the
326 * spa_config_object. (Note that although we already modified the
327 * spa_config_object in spa_sync_removing_state, that may not have
328 * modified all blocks of the object.)
330 dmu_object_info_t doi;
331 VERIFY0(dmu_object_info(mos, DMU_POOL_DIRECTORY_OBJECT, &doi));
332 for (uint64_t offset = 0; offset < doi.doi_max_offset; ) {
334 VERIFY0(dmu_buf_hold(mos, DMU_POOL_DIRECTORY_OBJECT,
335 offset, FTAG, &dbuf, 0));
336 dmu_buf_will_dirty(dbuf, tx);
337 offset += dbuf->db_size;
338 dmu_buf_rele(dbuf, FTAG);
342 * Now that we've allocated the im_object, dirty the vdev to ensure
343 * that the object gets written to the config on disk.
345 vdev_config_dirty(vd);
347 zfs_dbgmsg("starting removal thread for vdev %llu (%px) in txg %llu "
348 "im_obj=%llu", vd->vdev_id, vd, dmu_tx_get_txg(tx),
349 vic->vic_mapping_object);
351 spa_history_log_internal(spa, "vdev remove started", tx,
352 "%s vdev %llu %s", spa_name(spa), (u_longlong_t)vd->vdev_id,
353 (vd->vdev_path != NULL) ? vd->vdev_path : "-");
355 * Setting spa_vdev_removal causes subsequent frees to call
356 * free_from_removing_vdev(). Note that we don't need any locking
357 * because we are the sync thread, and metaslab_free_impl() is only
358 * called from syncing context (potentially from a zio taskq thread,
359 * but in any case only when there are outstanding free i/os, which
362 ASSERT3P(spa->spa_vdev_removal, ==, NULL);
363 spa->spa_vdev_removal = svr;
364 svr->svr_thread = thread_create(NULL, 0,
365 spa_vdev_remove_thread, spa, 0, &p0, TS_RUN, minclsyspri);
369 * When we are opening a pool, we must read the mapping for each
370 * indirect vdev in order from most recently removed to least
371 * recently removed. We do this because the blocks for the mapping
372 * of older indirect vdevs may be stored on more recently removed vdevs.
373 * In order to read each indirect mapping object, we must have
374 * initialized all more recently removed vdevs.
377 spa_remove_init(spa_t *spa)
381 error = zap_lookup(spa->spa_dsl_pool->dp_meta_objset,
382 DMU_POOL_DIRECTORY_OBJECT,
383 DMU_POOL_REMOVING, sizeof (uint64_t),
384 sizeof (spa->spa_removing_phys) / sizeof (uint64_t),
385 &spa->spa_removing_phys);
387 if (error == ENOENT) {
388 spa->spa_removing_phys.sr_state = DSS_NONE;
389 spa->spa_removing_phys.sr_removing_vdev = -1;
390 spa->spa_removing_phys.sr_prev_indirect_vdev = -1;
391 spa->spa_indirect_vdevs_loaded = B_TRUE;
393 } else if (error != 0) {
397 if (spa->spa_removing_phys.sr_state == DSS_SCANNING) {
399 * We are currently removing a vdev. Create and
400 * initialize a spa_vdev_removal_t from the bonus
401 * buffer of the removing vdevs vdev_im_object, and
402 * initialize its partial mapping.
404 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
405 vdev_t *vd = vdev_lookup_top(spa,
406 spa->spa_removing_phys.sr_removing_vdev);
409 spa_config_exit(spa, SCL_STATE, FTAG);
413 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
415 ASSERT(vdev_is_concrete(vd));
416 spa_vdev_removal_t *svr = spa_vdev_removal_create(vd);
417 ASSERT3U(svr->svr_vdev_id, ==, vd->vdev_id);
418 ASSERT(vd->vdev_removing);
420 vd->vdev_indirect_mapping = vdev_indirect_mapping_open(
421 spa->spa_meta_objset, vic->vic_mapping_object);
422 vd->vdev_indirect_births = vdev_indirect_births_open(
423 spa->spa_meta_objset, vic->vic_births_object);
424 spa_config_exit(spa, SCL_STATE, FTAG);
426 spa->spa_vdev_removal = svr;
429 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
430 uint64_t indirect_vdev_id =
431 spa->spa_removing_phys.sr_prev_indirect_vdev;
432 while (indirect_vdev_id != UINT64_MAX) {
433 vdev_t *vd = vdev_lookup_top(spa, indirect_vdev_id);
434 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
436 ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
437 vd->vdev_indirect_mapping = vdev_indirect_mapping_open(
438 spa->spa_meta_objset, vic->vic_mapping_object);
439 vd->vdev_indirect_births = vdev_indirect_births_open(
440 spa->spa_meta_objset, vic->vic_births_object);
442 indirect_vdev_id = vic->vic_prev_indirect_vdev;
444 spa_config_exit(spa, SCL_STATE, FTAG);
447 * Now that we've loaded all the indirect mappings, we can allow
448 * reads from other blocks (e.g. via predictive prefetch).
450 spa->spa_indirect_vdevs_loaded = B_TRUE;
455 spa_restart_removal(spa_t *spa)
457 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
463 * In general when this function is called there is no
464 * removal thread running. The only scenario where this
465 * is not true is during spa_import() where this function
466 * is called twice [once from spa_import_impl() and
467 * spa_async_resume()]. Thus, in the scenario where we
468 * import a pool that has an ongoing removal we don't
469 * want to spawn a second thread.
471 if (svr->svr_thread != NULL)
474 if (!spa_writeable(spa))
477 zfs_dbgmsg("restarting removal of %llu", svr->svr_vdev_id);
478 svr->svr_thread = thread_create(NULL, 0, spa_vdev_remove_thread, spa,
479 0, &p0, TS_RUN, minclsyspri);
483 * Process freeing from a device which is in the middle of being removed.
484 * We must handle this carefully so that we attempt to copy freed data,
485 * and we correctly free already-copied data.
488 free_from_removing_vdev(vdev_t *vd, uint64_t offset, uint64_t size)
490 spa_t *spa = vd->vdev_spa;
491 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
492 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
493 uint64_t txg = spa_syncing_txg(spa);
494 uint64_t max_offset_yet = 0;
496 ASSERT(vd->vdev_indirect_config.vic_mapping_object != 0);
497 ASSERT3U(vd->vdev_indirect_config.vic_mapping_object, ==,
498 vdev_indirect_mapping_object(vim));
499 ASSERT3U(vd->vdev_id, ==, svr->svr_vdev_id);
501 mutex_enter(&svr->svr_lock);
504 * Remove the segment from the removing vdev's spacemap. This
505 * ensures that we will not attempt to copy this space (if the
506 * removal thread has not yet visited it), and also ensures
507 * that we know what is actually allocated on the new vdevs
508 * (needed if we cancel the removal).
510 * Note: we must do the metaslab_free_concrete() with the svr_lock
511 * held, so that the remove_thread can not load this metaslab and then
512 * visit this offset between the time that we metaslab_free_concrete()
513 * and when we check to see if it has been visited.
515 * Note: The checkpoint flag is set to false as having/taking
516 * a checkpoint and removing a device can't happen at the same
519 ASSERT(!spa_has_checkpoint(spa));
520 metaslab_free_concrete(vd, offset, size, B_FALSE);
522 uint64_t synced_size = 0;
523 uint64_t synced_offset = 0;
524 uint64_t max_offset_synced = vdev_indirect_mapping_max_offset(vim);
525 if (offset < max_offset_synced) {
527 * The mapping for this offset is already on disk.
528 * Free from the new location.
530 * Note that we use svr_max_synced_offset because it is
531 * updated atomically with respect to the in-core mapping.
532 * By contrast, vim_max_offset is not.
534 * This block may be split between a synced entry and an
535 * in-flight or unvisited entry. Only process the synced
536 * portion of it here.
538 synced_size = MIN(size, max_offset_synced - offset);
539 synced_offset = offset;
541 ASSERT3U(max_offset_yet, <=, max_offset_synced);
542 max_offset_yet = max_offset_synced;
544 DTRACE_PROBE3(remove__free__synced,
547 uint64_t, synced_size);
550 offset += synced_size;
554 * Look at all in-flight txgs starting from the currently syncing one
555 * and see if a section of this free is being copied. By starting from
556 * this txg and iterating forward, we might find that this region
557 * was copied in two different txgs and handle it appropriately.
559 for (int i = 0; i < TXG_CONCURRENT_STATES; i++) {
560 int txgoff = (txg + i) & TXG_MASK;
561 if (size > 0 && offset < svr->svr_max_offset_to_sync[txgoff]) {
563 * The mapping for this offset is in flight, and
564 * will be synced in txg+i.
566 uint64_t inflight_size = MIN(size,
567 svr->svr_max_offset_to_sync[txgoff] - offset);
569 DTRACE_PROBE4(remove__free__inflight,
572 uint64_t, inflight_size,
576 * We copy data in order of increasing offset.
577 * Therefore the max_offset_to_sync[] must increase
578 * (or be zero, indicating that nothing is being
579 * copied in that txg).
581 if (svr->svr_max_offset_to_sync[txgoff] != 0) {
582 ASSERT3U(svr->svr_max_offset_to_sync[txgoff],
585 svr->svr_max_offset_to_sync[txgoff];
589 * We've already committed to copying this segment:
590 * we have allocated space elsewhere in the pool for
591 * it and have an IO outstanding to copy the data. We
592 * cannot free the space before the copy has
593 * completed, or else the copy IO might overwrite any
594 * new data. To free that space, we record the
595 * segment in the appropriate svr_frees tree and free
596 * the mapped space later, in the txg where we have
597 * completed the copy and synced the mapping (see
598 * vdev_mapping_sync).
600 range_tree_add(svr->svr_frees[txgoff],
601 offset, inflight_size);
602 size -= inflight_size;
603 offset += inflight_size;
606 * This space is already accounted for as being
607 * done, because it is being copied in txg+i.
608 * However, if i!=0, then it is being copied in
609 * a future txg. If we crash after this txg
610 * syncs but before txg+i syncs, then the space
611 * will be free. Therefore we must account
612 * for the space being done in *this* txg
613 * (when it is freed) rather than the future txg
614 * (when it will be copied).
616 ASSERT3U(svr->svr_bytes_done[txgoff], >=,
618 svr->svr_bytes_done[txgoff] -= inflight_size;
619 svr->svr_bytes_done[txg & TXG_MASK] += inflight_size;
622 ASSERT0(svr->svr_max_offset_to_sync[TXG_CLEAN(txg) & TXG_MASK]);
626 * The copy thread has not yet visited this offset. Ensure
630 DTRACE_PROBE3(remove__free__unvisited,
635 if (svr->svr_allocd_segs != NULL)
636 range_tree_clear(svr->svr_allocd_segs, offset, size);
639 * Since we now do not need to copy this data, for
640 * accounting purposes we have done our job and can count
643 svr->svr_bytes_done[txg & TXG_MASK] += size;
645 mutex_exit(&svr->svr_lock);
648 * Now that we have dropped svr_lock, process the synced portion
651 if (synced_size > 0) {
652 vdev_indirect_mark_obsolete(vd, synced_offset, synced_size);
655 * Note: this can only be called from syncing context,
656 * and the vdev_indirect_mapping is only changed from the
657 * sync thread, so we don't need svr_lock while doing
658 * metaslab_free_impl_cb.
660 boolean_t checkpoint = B_FALSE;
661 vdev_indirect_ops.vdev_op_remap(vd, synced_offset, synced_size,
662 metaslab_free_impl_cb, &checkpoint);
667 * Stop an active removal and update the spa_removing phys.
670 spa_finish_removal(spa_t *spa, dsl_scan_state_t state, dmu_tx_t *tx)
672 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
673 ASSERT3U(dmu_tx_get_txg(tx), ==, spa_syncing_txg(spa));
675 /* Ensure the removal thread has completed before we free the svr. */
676 spa_vdev_remove_suspend(spa);
678 ASSERT(state == DSS_FINISHED || state == DSS_CANCELED);
680 if (state == DSS_FINISHED) {
681 spa_removing_phys_t *srp = &spa->spa_removing_phys;
682 vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
683 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
685 if (srp->sr_prev_indirect_vdev != -1) {
687 pvd = vdev_lookup_top(spa,
688 srp->sr_prev_indirect_vdev);
689 ASSERT3P(pvd->vdev_ops, ==, &vdev_indirect_ops);
692 vic->vic_prev_indirect_vdev = srp->sr_prev_indirect_vdev;
693 srp->sr_prev_indirect_vdev = vd->vdev_id;
695 spa->spa_removing_phys.sr_state = state;
696 spa->spa_removing_phys.sr_end_time = gethrestime_sec();
698 spa->spa_vdev_removal = NULL;
699 spa_vdev_removal_destroy(svr);
701 spa_sync_removing_state(spa, tx);
702 spa_notify_waiters(spa);
704 vdev_config_dirty(spa->spa_root_vdev);
708 free_mapped_segment_cb(void *arg, uint64_t offset, uint64_t size)
711 vdev_indirect_mark_obsolete(vd, offset, size);
712 boolean_t checkpoint = B_FALSE;
713 vdev_indirect_ops.vdev_op_remap(vd, offset, size,
714 metaslab_free_impl_cb, &checkpoint);
718 * On behalf of the removal thread, syncs an incremental bit more of
719 * the indirect mapping to disk and updates the in-memory mapping.
720 * Called as a sync task in every txg that the removal thread makes progress.
723 vdev_mapping_sync(void *arg, dmu_tx_t *tx)
725 spa_vdev_removal_t *svr = arg;
726 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
727 vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
728 vdev_indirect_config_t *vic __maybe_unused = &vd->vdev_indirect_config;
729 uint64_t txg = dmu_tx_get_txg(tx);
730 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
732 ASSERT(vic->vic_mapping_object != 0);
733 ASSERT3U(txg, ==, spa_syncing_txg(spa));
735 vdev_indirect_mapping_add_entries(vim,
736 &svr->svr_new_segments[txg & TXG_MASK], tx);
737 vdev_indirect_births_add_entry(vd->vdev_indirect_births,
738 vdev_indirect_mapping_max_offset(vim), dmu_tx_get_txg(tx), tx);
741 * Free the copied data for anything that was freed while the
742 * mapping entries were in flight.
744 mutex_enter(&svr->svr_lock);
745 range_tree_vacate(svr->svr_frees[txg & TXG_MASK],
746 free_mapped_segment_cb, vd);
747 ASSERT3U(svr->svr_max_offset_to_sync[txg & TXG_MASK], >=,
748 vdev_indirect_mapping_max_offset(vim));
749 svr->svr_max_offset_to_sync[txg & TXG_MASK] = 0;
750 mutex_exit(&svr->svr_lock);
752 spa_sync_removing_state(spa, tx);
755 typedef struct vdev_copy_segment_arg {
757 dva_t *vcsa_dest_dva;
759 range_tree_t *vcsa_obsolete_segs;
760 } vdev_copy_segment_arg_t;
763 unalloc_seg(void *arg, uint64_t start, uint64_t size)
765 vdev_copy_segment_arg_t *vcsa = arg;
766 spa_t *spa = vcsa->vcsa_spa;
767 blkptr_t bp = { { { {0} } } };
769 BP_SET_BIRTH(&bp, TXG_INITIAL, TXG_INITIAL);
770 BP_SET_LSIZE(&bp, size);
771 BP_SET_PSIZE(&bp, size);
772 BP_SET_COMPRESS(&bp, ZIO_COMPRESS_OFF);
773 BP_SET_CHECKSUM(&bp, ZIO_CHECKSUM_OFF);
774 BP_SET_TYPE(&bp, DMU_OT_NONE);
775 BP_SET_LEVEL(&bp, 0);
776 BP_SET_DEDUP(&bp, 0);
777 BP_SET_BYTEORDER(&bp, ZFS_HOST_BYTEORDER);
779 DVA_SET_VDEV(&bp.blk_dva[0], DVA_GET_VDEV(vcsa->vcsa_dest_dva));
780 DVA_SET_OFFSET(&bp.blk_dva[0],
781 DVA_GET_OFFSET(vcsa->vcsa_dest_dva) + start);
782 DVA_SET_ASIZE(&bp.blk_dva[0], size);
784 zio_free(spa, vcsa->vcsa_txg, &bp);
788 * All reads and writes associated with a call to spa_vdev_copy_segment()
792 spa_vdev_copy_segment_done(zio_t *zio)
794 vdev_copy_segment_arg_t *vcsa = zio->io_private;
796 range_tree_vacate(vcsa->vcsa_obsolete_segs,
798 range_tree_destroy(vcsa->vcsa_obsolete_segs);
799 kmem_free(vcsa, sizeof (*vcsa));
801 spa_config_exit(zio->io_spa, SCL_STATE, zio->io_spa);
805 * The write of the new location is done.
808 spa_vdev_copy_segment_write_done(zio_t *zio)
810 vdev_copy_arg_t *vca = zio->io_private;
812 abd_free(zio->io_abd);
814 mutex_enter(&vca->vca_lock);
815 vca->vca_outstanding_bytes -= zio->io_size;
817 if (zio->io_error != 0)
818 vca->vca_write_error_bytes += zio->io_size;
820 cv_signal(&vca->vca_cv);
821 mutex_exit(&vca->vca_lock);
825 * The read of the old location is done. The parent zio is the write to
826 * the new location. Allow it to start.
829 spa_vdev_copy_segment_read_done(zio_t *zio)
831 vdev_copy_arg_t *vca = zio->io_private;
833 if (zio->io_error != 0) {
834 mutex_enter(&vca->vca_lock);
835 vca->vca_read_error_bytes += zio->io_size;
836 mutex_exit(&vca->vca_lock);
839 zio_nowait(zio_unique_parent(zio));
843 * If the old and new vdevs are mirrors, we will read both sides of the old
844 * mirror, and write each copy to the corresponding side of the new mirror.
845 * If the old and new vdevs have a different number of children, we will do
846 * this as best as possible. Since we aren't verifying checksums, this
847 * ensures that as long as there's a good copy of the data, we'll have a
848 * good copy after the removal, even if there's silent damage to one side
849 * of the mirror. If we're removing a mirror that has some silent damage,
850 * we'll have exactly the same damage in the new location (assuming that
851 * the new location is also a mirror).
853 * We accomplish this by creating a tree of zio_t's, with as many writes as
854 * there are "children" of the new vdev (a non-redundant vdev counts as one
855 * child, a 2-way mirror has 2 children, etc). Each write has an associated
856 * read from a child of the old vdev. Typically there will be the same
857 * number of children of the old and new vdevs. However, if there are more
858 * children of the new vdev, some child(ren) of the old vdev will be issued
859 * multiple reads. If there are more children of the old vdev, some copies
862 * For example, the tree of zio_t's for a 2-way mirror is:
866 * write(new vdev, child 0) write(new vdev, child 1)
868 * read(old vdev, child 0) read(old vdev, child 1)
870 * Child zio's complete before their parents complete. However, zio's
871 * created with zio_vdev_child_io() may be issued before their children
872 * complete. In this case we need to make sure that the children (reads)
873 * complete before the parents (writes) are *issued*. We do this by not
874 * calling zio_nowait() on each write until its corresponding read has
877 * The spa_config_lock must be held while zio's created by
878 * zio_vdev_child_io() are in progress, to ensure that the vdev tree does
879 * not change (e.g. due to a concurrent "zpool attach/detach"). The "null"
880 * zio is needed to release the spa_config_lock after all the reads and
881 * writes complete. (Note that we can't grab the config lock for each read,
882 * because it is not reentrant - we could deadlock with a thread waiting
886 spa_vdev_copy_one_child(vdev_copy_arg_t *vca, zio_t *nzio,
887 vdev_t *source_vd, uint64_t source_offset,
888 vdev_t *dest_child_vd, uint64_t dest_offset, int dest_id, uint64_t size)
890 ASSERT3U(spa_config_held(nzio->io_spa, SCL_ALL, RW_READER), !=, 0);
893 * If the destination child in unwritable then there is no point
894 * in issuing the source reads which cannot be written.
896 if (!vdev_writeable(dest_child_vd))
899 mutex_enter(&vca->vca_lock);
900 vca->vca_outstanding_bytes += size;
901 mutex_exit(&vca->vca_lock);
903 abd_t *abd = abd_alloc_for_io(size, B_FALSE);
905 vdev_t *source_child_vd = NULL;
906 if (source_vd->vdev_ops == &vdev_mirror_ops && dest_id != -1) {
908 * Source and dest are both mirrors. Copy from the same
909 * child id as we are copying to (wrapping around if there
910 * are more dest children than source children). If the
911 * preferred source child is unreadable select another.
913 for (int i = 0; i < source_vd->vdev_children; i++) {
914 source_child_vd = source_vd->vdev_child[
915 (dest_id + i) % source_vd->vdev_children];
916 if (vdev_readable(source_child_vd))
920 source_child_vd = source_vd;
924 * There should always be at least one readable source child or
925 * the pool would be in a suspended state. Somehow selecting an
926 * unreadable child would result in IO errors, the removal process
927 * being cancelled, and the pool reverting to its pre-removal state.
929 ASSERT3P(source_child_vd, !=, NULL);
931 zio_t *write_zio = zio_vdev_child_io(nzio, NULL,
932 dest_child_vd, dest_offset, abd, size,
933 ZIO_TYPE_WRITE, ZIO_PRIORITY_REMOVAL,
935 spa_vdev_copy_segment_write_done, vca);
937 zio_nowait(zio_vdev_child_io(write_zio, NULL,
938 source_child_vd, source_offset, abd, size,
939 ZIO_TYPE_READ, ZIO_PRIORITY_REMOVAL,
941 spa_vdev_copy_segment_read_done, vca));
945 * Allocate a new location for this segment, and create the zio_t's to
946 * read from the old location and write to the new location.
949 spa_vdev_copy_segment(vdev_t *vd, range_tree_t *segs,
950 uint64_t maxalloc, uint64_t txg,
951 vdev_copy_arg_t *vca, zio_alloc_list_t *zal)
953 metaslab_group_t *mg = vd->vdev_mg;
954 spa_t *spa = vd->vdev_spa;
955 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
956 vdev_indirect_mapping_entry_t *entry;
958 uint64_t start = range_tree_min(segs);
959 ASSERT0(P2PHASE(start, 1 << spa->spa_min_ashift));
961 ASSERT3U(maxalloc, <=, SPA_MAXBLOCKSIZE);
962 ASSERT0(P2PHASE(maxalloc, 1 << spa->spa_min_ashift));
964 uint64_t size = range_tree_span(segs);
965 if (range_tree_span(segs) > maxalloc) {
967 * We can't allocate all the segments. Prefer to end
968 * the allocation at the end of a segment, thus avoiding
969 * additional split blocks.
971 range_seg_max_t search;
972 zfs_btree_index_t where;
973 rs_set_start(&search, segs, start + maxalloc);
974 rs_set_end(&search, segs, start + maxalloc);
975 (void) zfs_btree_find(&segs->rt_root, &search, &where);
976 range_seg_t *rs = zfs_btree_prev(&segs->rt_root, &where,
979 size = rs_get_end(rs, segs) - start;
982 * There are no segments that end before maxalloc.
983 * I.e. the first segment is larger than maxalloc,
984 * so we must split it.
989 ASSERT3U(size, <=, maxalloc);
990 ASSERT0(P2PHASE(size, 1 << spa->spa_min_ashift));
993 * An allocation class might not have any remaining vdevs or space
995 metaslab_class_t *mc = mg->mg_class;
996 if (mc != spa_normal_class(spa) && mc->mc_groups <= 1)
997 mc = spa_normal_class(spa);
998 int error = metaslab_alloc_dva(spa, mc, size, &dst, 0, NULL, txg, 0,
1000 if (error == ENOSPC && mc != spa_normal_class(spa)) {
1001 error = metaslab_alloc_dva(spa, spa_normal_class(spa), size,
1002 &dst, 0, NULL, txg, 0, zal, 0);
1008 * Determine the ranges that are not actually needed. Offsets are
1009 * relative to the start of the range to be copied (i.e. relative to the
1010 * local variable "start").
1012 range_tree_t *obsolete_segs = range_tree_create(NULL, RANGE_SEG64, NULL,
1015 zfs_btree_index_t where;
1016 range_seg_t *rs = zfs_btree_first(&segs->rt_root, &where);
1017 ASSERT3U(rs_get_start(rs, segs), ==, start);
1018 uint64_t prev_seg_end = rs_get_end(rs, segs);
1019 while ((rs = zfs_btree_next(&segs->rt_root, &where, &where)) != NULL) {
1020 if (rs_get_start(rs, segs) >= start + size) {
1023 range_tree_add(obsolete_segs,
1024 prev_seg_end - start,
1025 rs_get_start(rs, segs) - prev_seg_end);
1027 prev_seg_end = rs_get_end(rs, segs);
1029 /* We don't end in the middle of an obsolete range */
1030 ASSERT3U(start + size, <=, prev_seg_end);
1032 range_tree_clear(segs, start, size);
1035 * We can't have any padding of the allocated size, otherwise we will
1036 * misunderstand what's allocated, and the size of the mapping. We
1037 * prevent padding by ensuring that all devices in the pool have the
1038 * same ashift, and the allocation size is a multiple of the ashift.
1040 VERIFY3U(DVA_GET_ASIZE(&dst), ==, size);
1042 entry = kmem_zalloc(sizeof (vdev_indirect_mapping_entry_t), KM_SLEEP);
1043 DVA_MAPPING_SET_SRC_OFFSET(&entry->vime_mapping, start);
1044 entry->vime_mapping.vimep_dst = dst;
1045 if (spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS)) {
1046 entry->vime_obsolete_count = range_tree_space(obsolete_segs);
1049 vdev_copy_segment_arg_t *vcsa = kmem_zalloc(sizeof (*vcsa), KM_SLEEP);
1050 vcsa->vcsa_dest_dva = &entry->vime_mapping.vimep_dst;
1051 vcsa->vcsa_obsolete_segs = obsolete_segs;
1052 vcsa->vcsa_spa = spa;
1053 vcsa->vcsa_txg = txg;
1056 * See comment before spa_vdev_copy_one_child().
1058 spa_config_enter(spa, SCL_STATE, spa, RW_READER);
1059 zio_t *nzio = zio_null(spa->spa_txg_zio[txg & TXG_MASK], spa, NULL,
1060 spa_vdev_copy_segment_done, vcsa, 0);
1061 vdev_t *dest_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dst));
1062 if (dest_vd->vdev_ops == &vdev_mirror_ops) {
1063 for (int i = 0; i < dest_vd->vdev_children; i++) {
1064 vdev_t *child = dest_vd->vdev_child[i];
1065 spa_vdev_copy_one_child(vca, nzio, vd, start,
1066 child, DVA_GET_OFFSET(&dst), i, size);
1069 spa_vdev_copy_one_child(vca, nzio, vd, start,
1070 dest_vd, DVA_GET_OFFSET(&dst), -1, size);
1074 list_insert_tail(&svr->svr_new_segments[txg & TXG_MASK], entry);
1075 ASSERT3U(start + size, <=, vd->vdev_ms_count << vd->vdev_ms_shift);
1076 vdev_dirty(vd, 0, NULL, txg);
1082 * Complete the removal of a toplevel vdev. This is called as a
1083 * synctask in the same txg that we will sync out the new config (to the
1084 * MOS object) which indicates that this vdev is indirect.
1087 vdev_remove_complete_sync(void *arg, dmu_tx_t *tx)
1089 spa_vdev_removal_t *svr = arg;
1090 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
1091 vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
1093 ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
1095 for (int i = 0; i < TXG_SIZE; i++) {
1096 ASSERT0(svr->svr_bytes_done[i]);
1099 ASSERT3U(spa->spa_removing_phys.sr_copied, ==,
1100 spa->spa_removing_phys.sr_to_copy);
1102 vdev_destroy_spacemaps(vd, tx);
1104 /* destroy leaf zaps, if any */
1105 ASSERT3P(svr->svr_zaplist, !=, NULL);
1106 for (nvpair_t *pair = nvlist_next_nvpair(svr->svr_zaplist, NULL);
1108 pair = nvlist_next_nvpair(svr->svr_zaplist, pair)) {
1109 vdev_destroy_unlink_zap(vd, fnvpair_value_uint64(pair), tx);
1111 fnvlist_free(svr->svr_zaplist);
1113 spa_finish_removal(dmu_tx_pool(tx)->dp_spa, DSS_FINISHED, tx);
1114 /* vd->vdev_path is not available here */
1115 spa_history_log_internal(spa, "vdev remove completed", tx,
1116 "%s vdev %llu", spa_name(spa), (u_longlong_t)vd->vdev_id);
1120 vdev_remove_enlist_zaps(vdev_t *vd, nvlist_t *zlist)
1122 ASSERT3P(zlist, !=, NULL);
1123 ASSERT3P(vd->vdev_ops, !=, &vdev_raidz_ops);
1125 if (vd->vdev_leaf_zap != 0) {
1127 (void) snprintf(zkey, sizeof (zkey), "%s-%llu",
1128 VDEV_REMOVAL_ZAP_OBJS, (u_longlong_t)vd->vdev_leaf_zap);
1129 fnvlist_add_uint64(zlist, zkey, vd->vdev_leaf_zap);
1132 for (uint64_t id = 0; id < vd->vdev_children; id++) {
1133 vdev_remove_enlist_zaps(vd->vdev_child[id], zlist);
1138 vdev_remove_replace_with_indirect(vdev_t *vd, uint64_t txg)
1142 spa_t *spa = vd->vdev_spa;
1143 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1146 * First, build a list of leaf zaps to be destroyed.
1147 * This is passed to the sync context thread,
1148 * which does the actual unlinking.
1150 svr->svr_zaplist = fnvlist_alloc();
1151 vdev_remove_enlist_zaps(vd, svr->svr_zaplist);
1153 ivd = vdev_add_parent(vd, &vdev_indirect_ops);
1154 ivd->vdev_removing = 0;
1156 vd->vdev_leaf_zap = 0;
1158 vdev_remove_child(ivd, vd);
1159 vdev_compact_children(ivd);
1161 ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
1163 mutex_enter(&svr->svr_lock);
1164 svr->svr_thread = NULL;
1165 cv_broadcast(&svr->svr_cv);
1166 mutex_exit(&svr->svr_lock);
1168 /* After this, we can not use svr. */
1169 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1170 dsl_sync_task_nowait(spa->spa_dsl_pool,
1171 vdev_remove_complete_sync, svr, tx);
1176 * Complete the removal of a toplevel vdev. This is called in open
1177 * context by the removal thread after we have copied all vdev's data.
1180 vdev_remove_complete(spa_t *spa)
1185 * Wait for any deferred frees to be synced before we call
1186 * vdev_metaslab_fini()
1188 txg_wait_synced(spa->spa_dsl_pool, 0);
1189 txg = spa_vdev_enter(spa);
1190 vdev_t *vd = vdev_lookup_top(spa, spa->spa_vdev_removal->svr_vdev_id);
1191 ASSERT3P(vd->vdev_initialize_thread, ==, NULL);
1192 ASSERT3P(vd->vdev_trim_thread, ==, NULL);
1193 ASSERT3P(vd->vdev_autotrim_thread, ==, NULL);
1195 sysevent_t *ev = spa_event_create(spa, vd, NULL,
1196 ESC_ZFS_VDEV_REMOVE_DEV);
1198 zfs_dbgmsg("finishing device removal for vdev %llu in txg %llu",
1202 * Discard allocation state.
1204 if (vd->vdev_mg != NULL) {
1205 vdev_metaslab_fini(vd);
1206 metaslab_group_destroy(vd->vdev_mg);
1208 spa_log_sm_set_blocklimit(spa);
1210 ASSERT0(vd->vdev_stat.vs_space);
1211 ASSERT0(vd->vdev_stat.vs_dspace);
1213 vdev_remove_replace_with_indirect(vd, txg);
1216 * We now release the locks, allowing spa_sync to run and finish the
1217 * removal via vdev_remove_complete_sync in syncing context.
1219 * Note that we hold on to the vdev_t that has been replaced. Since
1220 * it isn't part of the vdev tree any longer, it can't be concurrently
1221 * manipulated, even while we don't have the config lock.
1223 (void) spa_vdev_exit(spa, NULL, txg, 0);
1226 * Top ZAP should have been transferred to the indirect vdev in
1227 * vdev_remove_replace_with_indirect.
1229 ASSERT0(vd->vdev_top_zap);
1232 * Leaf ZAP should have been moved in vdev_remove_replace_with_indirect.
1234 ASSERT0(vd->vdev_leaf_zap);
1236 txg = spa_vdev_enter(spa);
1237 (void) vdev_label_init(vd, 0, VDEV_LABEL_REMOVE);
1239 * Request to update the config and the config cachefile.
1241 vdev_config_dirty(spa->spa_root_vdev);
1242 (void) spa_vdev_exit(spa, vd, txg, 0);
1249 * Evacuates a segment of size at most max_alloc from the vdev
1250 * via repeated calls to spa_vdev_copy_segment. If an allocation
1251 * fails, the pool is probably too fragmented to handle such a
1252 * large size, so decrease max_alloc so that the caller will not try
1253 * this size again this txg.
1256 spa_vdev_copy_impl(vdev_t *vd, spa_vdev_removal_t *svr, vdev_copy_arg_t *vca,
1257 uint64_t *max_alloc, dmu_tx_t *tx)
1259 uint64_t txg = dmu_tx_get_txg(tx);
1260 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
1262 mutex_enter(&svr->svr_lock);
1265 * Determine how big of a chunk to copy. We can allocate up
1266 * to max_alloc bytes, and we can span up to vdev_removal_max_span
1267 * bytes of unallocated space at a time. "segs" will track the
1268 * allocated segments that we are copying. We may also be copying
1269 * free segments (of up to vdev_removal_max_span bytes).
1271 range_tree_t *segs = range_tree_create(NULL, RANGE_SEG64, NULL, 0, 0);
1273 range_tree_t *rt = svr->svr_allocd_segs;
1274 range_seg_t *rs = range_tree_first(rt);
1279 uint64_t seg_length;
1281 if (range_tree_is_empty(segs)) {
1282 /* need to truncate the first seg based on max_alloc */
1283 seg_length = MIN(rs_get_end(rs, rt) - rs_get_start(rs,
1286 if (rs_get_start(rs, rt) - range_tree_max(segs) >
1287 vdev_removal_max_span) {
1289 * Including this segment would cause us to
1290 * copy a larger unneeded chunk than is allowed.
1293 } else if (rs_get_end(rs, rt) - range_tree_min(segs) >
1296 * This additional segment would extend past
1297 * max_alloc. Rather than splitting this
1298 * segment, leave it for the next mapping.
1302 seg_length = rs_get_end(rs, rt) -
1303 rs_get_start(rs, rt);
1307 range_tree_add(segs, rs_get_start(rs, rt), seg_length);
1308 range_tree_remove(svr->svr_allocd_segs,
1309 rs_get_start(rs, rt), seg_length);
1312 if (range_tree_is_empty(segs)) {
1313 mutex_exit(&svr->svr_lock);
1314 range_tree_destroy(segs);
1318 if (svr->svr_max_offset_to_sync[txg & TXG_MASK] == 0) {
1319 dsl_sync_task_nowait(dmu_tx_pool(tx), vdev_mapping_sync,
1323 svr->svr_max_offset_to_sync[txg & TXG_MASK] = range_tree_max(segs);
1326 * Note: this is the amount of *allocated* space
1327 * that we are taking care of each txg.
1329 svr->svr_bytes_done[txg & TXG_MASK] += range_tree_space(segs);
1331 mutex_exit(&svr->svr_lock);
1333 zio_alloc_list_t zal;
1334 metaslab_trace_init(&zal);
1335 uint64_t thismax = SPA_MAXBLOCKSIZE;
1336 while (!range_tree_is_empty(segs)) {
1337 int error = spa_vdev_copy_segment(vd,
1338 segs, thismax, txg, vca, &zal);
1340 if (error == ENOSPC) {
1342 * Cut our segment in half, and don't try this
1343 * segment size again this txg. Note that the
1344 * allocation size must be aligned to the highest
1345 * ashift in the pool, so that the allocation will
1346 * not be padded out to a multiple of the ashift,
1347 * which could cause us to think that this mapping
1348 * is larger than we intended.
1350 ASSERT3U(spa->spa_max_ashift, >=, SPA_MINBLOCKSHIFT);
1351 ASSERT3U(spa->spa_max_ashift, ==, spa->spa_min_ashift);
1352 uint64_t attempted =
1353 MIN(range_tree_span(segs), thismax);
1354 thismax = P2ROUNDUP(attempted / 2,
1355 1 << spa->spa_max_ashift);
1357 * The minimum-size allocation can not fail.
1359 ASSERT3U(attempted, >, 1 << spa->spa_max_ashift);
1360 *max_alloc = attempted - (1 << spa->spa_max_ashift);
1365 * We've performed an allocation, so reset the
1368 metaslab_trace_fini(&zal);
1369 metaslab_trace_init(&zal);
1372 metaslab_trace_fini(&zal);
1373 range_tree_destroy(segs);
1377 * The size of each removal mapping is limited by the tunable
1378 * zfs_remove_max_segment, but we must adjust this to be a multiple of the
1379 * pool's ashift, so that we don't try to split individual sectors regardless
1380 * of the tunable value. (Note that device removal requires that all devices
1381 * have the same ashift, so there's no difference between spa_min_ashift and
1382 * spa_max_ashift.) The raw tunable should not be used elsewhere.
1385 spa_remove_max_segment(spa_t *spa)
1387 return (P2ROUNDUP(zfs_remove_max_segment, 1 << spa->spa_max_ashift));
1391 * The removal thread operates in open context. It iterates over all
1392 * allocated space in the vdev, by loading each metaslab's spacemap.
1393 * For each contiguous segment of allocated space (capping the segment
1394 * size at SPA_MAXBLOCKSIZE), we:
1395 * - Allocate space for it on another vdev.
1396 * - Create a new mapping from the old location to the new location
1397 * (as a record in svr_new_segments).
1398 * - Initiate a physical read zio to get the data off the removing disk.
1399 * - In the read zio's done callback, initiate a physical write zio to
1400 * write it to the new vdev.
1401 * Note that all of this will take effect when a particular TXG syncs.
1402 * The sync thread ensures that all the phys reads and writes for the syncing
1403 * TXG have completed (see spa_txg_zio) and writes the new mappings to disk
1404 * (see vdev_mapping_sync()).
1407 spa_vdev_remove_thread(void *arg)
1410 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1411 vdev_copy_arg_t vca;
1412 uint64_t max_alloc = spa_remove_max_segment(spa);
1413 uint64_t last_txg = 0;
1415 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
1416 vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
1417 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
1418 uint64_t start_offset = vdev_indirect_mapping_max_offset(vim);
1420 ASSERT3P(vd->vdev_ops, !=, &vdev_indirect_ops);
1421 ASSERT(vdev_is_concrete(vd));
1422 ASSERT(vd->vdev_removing);
1423 ASSERT(vd->vdev_indirect_config.vic_mapping_object != 0);
1424 ASSERT(vim != NULL);
1426 mutex_init(&vca.vca_lock, NULL, MUTEX_DEFAULT, NULL);
1427 cv_init(&vca.vca_cv, NULL, CV_DEFAULT, NULL);
1428 vca.vca_outstanding_bytes = 0;
1429 vca.vca_read_error_bytes = 0;
1430 vca.vca_write_error_bytes = 0;
1432 mutex_enter(&svr->svr_lock);
1435 * Start from vim_max_offset so we pick up where we left off
1436 * if we are restarting the removal after opening the pool.
1439 for (msi = start_offset >> vd->vdev_ms_shift;
1440 msi < vd->vdev_ms_count && !svr->svr_thread_exit; msi++) {
1441 metaslab_t *msp = vd->vdev_ms[msi];
1442 ASSERT3U(msi, <=, vd->vdev_ms_count);
1444 ASSERT0(range_tree_space(svr->svr_allocd_segs));
1446 mutex_enter(&msp->ms_sync_lock);
1447 mutex_enter(&msp->ms_lock);
1450 * Assert nothing in flight -- ms_*tree is empty.
1452 for (int i = 0; i < TXG_SIZE; i++) {
1453 ASSERT0(range_tree_space(msp->ms_allocating[i]));
1457 * If the metaslab has ever been allocated from (ms_sm!=NULL),
1458 * read the allocated segments from the space map object
1459 * into svr_allocd_segs. Since we do this while holding
1460 * svr_lock and ms_sync_lock, concurrent frees (which
1461 * would have modified the space map) will wait for us
1462 * to finish loading the spacemap, and then take the
1463 * appropriate action (see free_from_removing_vdev()).
1465 if (msp->ms_sm != NULL) {
1466 VERIFY0(space_map_load(msp->ms_sm,
1467 svr->svr_allocd_segs, SM_ALLOC));
1469 range_tree_walk(msp->ms_unflushed_allocs,
1470 range_tree_add, svr->svr_allocd_segs);
1471 range_tree_walk(msp->ms_unflushed_frees,
1472 range_tree_remove, svr->svr_allocd_segs);
1473 range_tree_walk(msp->ms_freeing,
1474 range_tree_remove, svr->svr_allocd_segs);
1477 * When we are resuming from a paused removal (i.e.
1478 * when importing a pool with a removal in progress),
1479 * discard any state that we have already processed.
1481 range_tree_clear(svr->svr_allocd_segs, 0, start_offset);
1483 mutex_exit(&msp->ms_lock);
1484 mutex_exit(&msp->ms_sync_lock);
1487 zfs_dbgmsg("copying %llu segments for metaslab %llu",
1488 zfs_btree_numnodes(&svr->svr_allocd_segs->rt_root),
1491 while (!svr->svr_thread_exit &&
1492 !range_tree_is_empty(svr->svr_allocd_segs)) {
1494 mutex_exit(&svr->svr_lock);
1497 * We need to periodically drop the config lock so that
1498 * writers can get in. Additionally, we can't wait
1499 * for a txg to sync while holding a config lock
1500 * (since a waiting writer could cause a 3-way deadlock
1501 * with the sync thread, which also gets a config
1502 * lock for reader). So we can't hold the config lock
1503 * while calling dmu_tx_assign().
1505 spa_config_exit(spa, SCL_CONFIG, FTAG);
1508 * This delay will pause the removal around the point
1509 * specified by zfs_removal_suspend_progress. We do this
1510 * solely from the test suite or during debugging.
1512 uint64_t bytes_copied =
1513 spa->spa_removing_phys.sr_copied;
1514 for (int i = 0; i < TXG_SIZE; i++)
1515 bytes_copied += svr->svr_bytes_done[i];
1516 while (zfs_removal_suspend_progress &&
1517 !svr->svr_thread_exit)
1520 mutex_enter(&vca.vca_lock);
1521 while (vca.vca_outstanding_bytes >
1522 zfs_remove_max_copy_bytes) {
1523 cv_wait(&vca.vca_cv, &vca.vca_lock);
1525 mutex_exit(&vca.vca_lock);
1528 dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
1530 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
1531 uint64_t txg = dmu_tx_get_txg(tx);
1534 * Reacquire the vdev_config lock. The vdev_t
1535 * that we're removing may have changed, e.g. due
1536 * to a vdev_attach or vdev_detach.
1538 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
1539 vd = vdev_lookup_top(spa, svr->svr_vdev_id);
1541 if (txg != last_txg)
1542 max_alloc = spa_remove_max_segment(spa);
1545 spa_vdev_copy_impl(vd, svr, &vca, &max_alloc, tx);
1548 mutex_enter(&svr->svr_lock);
1551 mutex_enter(&vca.vca_lock);
1552 if (zfs_removal_ignore_errors == 0 &&
1553 (vca.vca_read_error_bytes > 0 ||
1554 vca.vca_write_error_bytes > 0)) {
1555 svr->svr_thread_exit = B_TRUE;
1557 mutex_exit(&vca.vca_lock);
1560 mutex_exit(&svr->svr_lock);
1562 spa_config_exit(spa, SCL_CONFIG, FTAG);
1565 * Wait for all copies to finish before cleaning up the vca.
1567 txg_wait_synced(spa->spa_dsl_pool, 0);
1568 ASSERT0(vca.vca_outstanding_bytes);
1570 mutex_destroy(&vca.vca_lock);
1571 cv_destroy(&vca.vca_cv);
1573 if (svr->svr_thread_exit) {
1574 mutex_enter(&svr->svr_lock);
1575 range_tree_vacate(svr->svr_allocd_segs, NULL, NULL);
1576 svr->svr_thread = NULL;
1577 cv_broadcast(&svr->svr_cv);
1578 mutex_exit(&svr->svr_lock);
1581 * During the removal process an unrecoverable read or write
1582 * error was encountered. The removal process must be
1583 * cancelled or this damage may become permanent.
1585 if (zfs_removal_ignore_errors == 0 &&
1586 (vca.vca_read_error_bytes > 0 ||
1587 vca.vca_write_error_bytes > 0)) {
1588 zfs_dbgmsg("canceling removal due to IO errors: "
1589 "[read_error_bytes=%llu] [write_error_bytes=%llu]",
1590 vca.vca_read_error_bytes,
1591 vca.vca_write_error_bytes);
1592 spa_vdev_remove_cancel_impl(spa);
1595 ASSERT0(range_tree_space(svr->svr_allocd_segs));
1596 vdev_remove_complete(spa);
1603 spa_vdev_remove_suspend(spa_t *spa)
1605 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1610 mutex_enter(&svr->svr_lock);
1611 svr->svr_thread_exit = B_TRUE;
1612 while (svr->svr_thread != NULL)
1613 cv_wait(&svr->svr_cv, &svr->svr_lock);
1614 svr->svr_thread_exit = B_FALSE;
1615 mutex_exit(&svr->svr_lock);
1620 spa_vdev_remove_cancel_check(void *arg, dmu_tx_t *tx)
1622 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
1624 if (spa->spa_vdev_removal == NULL)
1625 return (ENOTACTIVE);
1630 * Cancel a removal by freeing all entries from the partial mapping
1631 * and marking the vdev as no longer being removing.
1635 spa_vdev_remove_cancel_sync(void *arg, dmu_tx_t *tx)
1637 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
1638 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1639 vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
1640 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
1641 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
1642 objset_t *mos = spa->spa_meta_objset;
1644 ASSERT3P(svr->svr_thread, ==, NULL);
1646 spa_feature_decr(spa, SPA_FEATURE_DEVICE_REMOVAL, tx);
1648 boolean_t are_precise;
1649 VERIFY0(vdev_obsolete_counts_are_precise(vd, &are_precise));
1651 spa_feature_decr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
1652 VERIFY0(zap_remove(spa->spa_meta_objset, vd->vdev_top_zap,
1653 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE, tx));
1656 uint64_t obsolete_sm_object;
1657 VERIFY0(vdev_obsolete_sm_object(vd, &obsolete_sm_object));
1658 if (obsolete_sm_object != 0) {
1659 ASSERT(vd->vdev_obsolete_sm != NULL);
1660 ASSERT3U(obsolete_sm_object, ==,
1661 space_map_object(vd->vdev_obsolete_sm));
1663 space_map_free(vd->vdev_obsolete_sm, tx);
1664 VERIFY0(zap_remove(spa->spa_meta_objset, vd->vdev_top_zap,
1665 VDEV_TOP_ZAP_INDIRECT_OBSOLETE_SM, tx));
1666 space_map_close(vd->vdev_obsolete_sm);
1667 vd->vdev_obsolete_sm = NULL;
1668 spa_feature_decr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
1670 for (int i = 0; i < TXG_SIZE; i++) {
1671 ASSERT(list_is_empty(&svr->svr_new_segments[i]));
1672 ASSERT3U(svr->svr_max_offset_to_sync[i], <=,
1673 vdev_indirect_mapping_max_offset(vim));
1676 for (uint64_t msi = 0; msi < vd->vdev_ms_count; msi++) {
1677 metaslab_t *msp = vd->vdev_ms[msi];
1679 if (msp->ms_start >= vdev_indirect_mapping_max_offset(vim))
1682 ASSERT0(range_tree_space(svr->svr_allocd_segs));
1684 mutex_enter(&msp->ms_lock);
1687 * Assert nothing in flight -- ms_*tree is empty.
1689 for (int i = 0; i < TXG_SIZE; i++)
1690 ASSERT0(range_tree_space(msp->ms_allocating[i]));
1691 for (int i = 0; i < TXG_DEFER_SIZE; i++)
1692 ASSERT0(range_tree_space(msp->ms_defer[i]));
1693 ASSERT0(range_tree_space(msp->ms_freed));
1695 if (msp->ms_sm != NULL) {
1696 mutex_enter(&svr->svr_lock);
1697 VERIFY0(space_map_load(msp->ms_sm,
1698 svr->svr_allocd_segs, SM_ALLOC));
1700 range_tree_walk(msp->ms_unflushed_allocs,
1701 range_tree_add, svr->svr_allocd_segs);
1702 range_tree_walk(msp->ms_unflushed_frees,
1703 range_tree_remove, svr->svr_allocd_segs);
1704 range_tree_walk(msp->ms_freeing,
1705 range_tree_remove, svr->svr_allocd_segs);
1708 * Clear everything past what has been synced,
1709 * because we have not allocated mappings for it yet.
1711 uint64_t syncd = vdev_indirect_mapping_max_offset(vim);
1712 uint64_t sm_end = msp->ms_sm->sm_start +
1713 msp->ms_sm->sm_size;
1715 range_tree_clear(svr->svr_allocd_segs,
1716 syncd, sm_end - syncd);
1718 mutex_exit(&svr->svr_lock);
1720 mutex_exit(&msp->ms_lock);
1722 mutex_enter(&svr->svr_lock);
1723 range_tree_vacate(svr->svr_allocd_segs,
1724 free_mapped_segment_cb, vd);
1725 mutex_exit(&svr->svr_lock);
1729 * Note: this must happen after we invoke free_mapped_segment_cb,
1730 * because it adds to the obsolete_segments.
1732 range_tree_vacate(vd->vdev_obsolete_segments, NULL, NULL);
1734 ASSERT3U(vic->vic_mapping_object, ==,
1735 vdev_indirect_mapping_object(vd->vdev_indirect_mapping));
1736 vdev_indirect_mapping_close(vd->vdev_indirect_mapping);
1737 vd->vdev_indirect_mapping = NULL;
1738 vdev_indirect_mapping_free(mos, vic->vic_mapping_object, tx);
1739 vic->vic_mapping_object = 0;
1741 ASSERT3U(vic->vic_births_object, ==,
1742 vdev_indirect_births_object(vd->vdev_indirect_births));
1743 vdev_indirect_births_close(vd->vdev_indirect_births);
1744 vd->vdev_indirect_births = NULL;
1745 vdev_indirect_births_free(mos, vic->vic_births_object, tx);
1746 vic->vic_births_object = 0;
1749 * We may have processed some frees from the removing vdev in this
1750 * txg, thus increasing svr_bytes_done; discard that here to
1751 * satisfy the assertions in spa_vdev_removal_destroy().
1752 * Note that future txg's can not have any bytes_done, because
1753 * future TXG's are only modified from open context, and we have
1754 * already shut down the copying thread.
1756 svr->svr_bytes_done[dmu_tx_get_txg(tx) & TXG_MASK] = 0;
1757 spa_finish_removal(spa, DSS_CANCELED, tx);
1759 vd->vdev_removing = B_FALSE;
1760 vdev_config_dirty(vd);
1762 zfs_dbgmsg("canceled device removal for vdev %llu in %llu",
1763 vd->vdev_id, dmu_tx_get_txg(tx));
1764 spa_history_log_internal(spa, "vdev remove canceled", tx,
1765 "%s vdev %llu %s", spa_name(spa),
1766 (u_longlong_t)vd->vdev_id,
1767 (vd->vdev_path != NULL) ? vd->vdev_path : "-");
1771 spa_vdev_remove_cancel_impl(spa_t *spa)
1773 uint64_t vdid = spa->spa_vdev_removal->svr_vdev_id;
1775 int error = dsl_sync_task(spa->spa_name, spa_vdev_remove_cancel_check,
1776 spa_vdev_remove_cancel_sync, NULL, 0,
1777 ZFS_SPACE_CHECK_EXTRA_RESERVED);
1780 spa_config_enter(spa, SCL_ALLOC | SCL_VDEV, FTAG, RW_WRITER);
1781 vdev_t *vd = vdev_lookup_top(spa, vdid);
1782 metaslab_group_activate(vd->vdev_mg);
1783 spa_config_exit(spa, SCL_ALLOC | SCL_VDEV, FTAG);
1790 spa_vdev_remove_cancel(spa_t *spa)
1792 spa_vdev_remove_suspend(spa);
1794 if (spa->spa_vdev_removal == NULL)
1795 return (ENOTACTIVE);
1797 return (spa_vdev_remove_cancel_impl(spa));
1801 svr_sync(spa_t *spa, dmu_tx_t *tx)
1803 spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1804 int txgoff = dmu_tx_get_txg(tx) & TXG_MASK;
1810 * This check is necessary so that we do not dirty the
1811 * DIRECTORY_OBJECT via spa_sync_removing_state() when there
1812 * is nothing to do. Dirtying it every time would prevent us
1813 * from syncing-to-convergence.
1815 if (svr->svr_bytes_done[txgoff] == 0)
1819 * Update progress accounting.
1821 spa->spa_removing_phys.sr_copied += svr->svr_bytes_done[txgoff];
1822 svr->svr_bytes_done[txgoff] = 0;
1824 spa_sync_removing_state(spa, tx);
1828 vdev_remove_make_hole_and_free(vdev_t *vd)
1830 uint64_t id = vd->vdev_id;
1831 spa_t *spa = vd->vdev_spa;
1832 vdev_t *rvd = spa->spa_root_vdev;
1834 ASSERT(MUTEX_HELD(&spa_namespace_lock));
1835 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1839 vd = vdev_alloc_common(spa, id, 0, &vdev_hole_ops);
1840 vdev_add_child(rvd, vd);
1841 vdev_config_dirty(rvd);
1844 * Reassess the health of our root vdev.
1850 * Remove a log device. The config lock is held for the specified TXG.
1853 spa_vdev_remove_log(vdev_t *vd, uint64_t *txg)
1855 metaslab_group_t *mg = vd->vdev_mg;
1856 spa_t *spa = vd->vdev_spa;
1859 ASSERT(vd->vdev_islog);
1860 ASSERT(vd == vd->vdev_top);
1861 ASSERT(MUTEX_HELD(&spa_namespace_lock));
1864 * Stop allocating from this vdev.
1866 metaslab_group_passivate(mg);
1869 * Wait for the youngest allocations and frees to sync,
1870 * and then wait for the deferral of those frees to finish.
1872 spa_vdev_config_exit(spa, NULL,
1873 *txg + TXG_CONCURRENT_STATES + TXG_DEFER_SIZE, 0, FTAG);
1876 * Cancel any initialize or TRIM which was in progress.
1878 vdev_initialize_stop_all(vd, VDEV_INITIALIZE_CANCELED);
1879 vdev_trim_stop_all(vd, VDEV_TRIM_CANCELED);
1880 vdev_autotrim_stop_wait(vd);
1883 * Evacuate the device. We don't hold the config lock as
1884 * writer since we need to do I/O but we do keep the
1885 * spa_namespace_lock held. Once this completes the device
1886 * should no longer have any blocks allocated on it.
1888 ASSERT(MUTEX_HELD(&spa_namespace_lock));
1889 if (vd->vdev_stat.vs_alloc != 0)
1890 error = spa_reset_logs(spa);
1892 *txg = spa_vdev_config_enter(spa);
1895 metaslab_group_activate(mg);
1898 ASSERT0(vd->vdev_stat.vs_alloc);
1901 * The evacuation succeeded. Remove any remaining MOS metadata
1902 * associated with this vdev, and wait for these changes to sync.
1904 vd->vdev_removing = B_TRUE;
1906 vdev_dirty_leaves(vd, VDD_DTL, *txg);
1907 vdev_config_dirty(vd);
1910 * When the log space map feature is enabled we look at
1911 * the vdev's top_zap to find the on-disk flush data of
1912 * the metaslab we just flushed. Thus, while removing a
1913 * log vdev we make sure to call vdev_metaslab_fini()
1914 * first, which removes all metaslabs of this vdev from
1915 * spa_metaslabs_by_flushed before vdev_remove_empty()
1916 * destroys the top_zap of this log vdev.
1918 * This avoids the scenario where we flush a metaslab
1919 * from the log vdev being removed that doesn't have a
1920 * top_zap and end up failing to lookup its on-disk flush
1923 * We don't call metaslab_group_destroy() right away
1924 * though (it will be called in vdev_free() later) as
1925 * during metaslab_sync() of metaslabs from other vdevs
1926 * we may touch the metaslab group of this vdev through
1927 * metaslab_class_histogram_verify()
1929 vdev_metaslab_fini(vd);
1930 spa_log_sm_set_blocklimit(spa);
1932 spa_vdev_config_exit(spa, NULL, *txg, 0, FTAG);
1933 *txg = spa_vdev_config_enter(spa);
1935 sysevent_t *ev = spa_event_create(spa, vd, NULL,
1936 ESC_ZFS_VDEV_REMOVE_DEV);
1937 ASSERT(MUTEX_HELD(&spa_namespace_lock));
1938 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1940 /* The top ZAP should have been destroyed by vdev_remove_empty. */
1941 ASSERT0(vd->vdev_top_zap);
1942 /* The leaf ZAP should have been destroyed by vdev_dtl_sync. */
1943 ASSERT0(vd->vdev_leaf_zap);
1945 (void) vdev_label_init(vd, 0, VDEV_LABEL_REMOVE);
1947 if (list_link_active(&vd->vdev_state_dirty_node))
1948 vdev_state_clean(vd);
1949 if (list_link_active(&vd->vdev_config_dirty_node))
1950 vdev_config_clean(vd);
1952 ASSERT0(vd->vdev_stat.vs_alloc);
1955 * Clean up the vdev namespace.
1957 vdev_remove_make_hole_and_free(vd);
1966 spa_vdev_remove_top_check(vdev_t *vd)
1968 spa_t *spa = vd->vdev_spa;
1970 if (vd != vd->vdev_top)
1971 return (SET_ERROR(ENOTSUP));
1973 if (!vdev_is_concrete(vd))
1974 return (SET_ERROR(ENOTSUP));
1976 if (!spa_feature_is_enabled(spa, SPA_FEATURE_DEVICE_REMOVAL))
1977 return (SET_ERROR(ENOTSUP));
1979 /* available space in the pool's normal class */
1980 uint64_t available = dsl_dir_space_available(
1981 spa->spa_dsl_pool->dp_root_dir, NULL, 0, B_TRUE);
1983 metaslab_class_t *mc = vd->vdev_mg->mg_class;
1986 * When removing a vdev from an allocation class that has
1987 * remaining vdevs, include available space from the class.
1989 if (mc != spa_normal_class(spa) && mc->mc_groups > 1) {
1990 uint64_t class_avail = metaslab_class_get_space(mc) -
1991 metaslab_class_get_alloc(mc);
1993 /* add class space, adjusted for overhead */
1994 available += (class_avail * 94) / 100;
1998 * There has to be enough free space to remove the
1999 * device and leave double the "slop" space (i.e. we
2000 * must leave at least 3% of the pool free, in addition to
2001 * the normal slop space).
2003 if (available < vd->vdev_stat.vs_dspace + spa_get_slop_space(spa)) {
2004 return (SET_ERROR(ENOSPC));
2008 * There can not be a removal in progress.
2010 if (spa->spa_removing_phys.sr_state == DSS_SCANNING)
2011 return (SET_ERROR(EBUSY));
2014 * The device must have all its data.
2016 if (!vdev_dtl_empty(vd, DTL_MISSING) ||
2017 !vdev_dtl_empty(vd, DTL_OUTAGE))
2018 return (SET_ERROR(EBUSY));
2021 * The device must be healthy.
2023 if (!vdev_readable(vd))
2024 return (SET_ERROR(EIO));
2027 * All vdevs in normal class must have the same ashift.
2029 if (spa->spa_max_ashift != spa->spa_min_ashift) {
2030 return (SET_ERROR(EINVAL));
2034 * A removed special/dedup vdev must have same ashift as normal class.
2036 ASSERT(!vd->vdev_islog);
2037 if (vd->vdev_alloc_bias != VDEV_BIAS_NONE &&
2038 vd->vdev_ashift != spa->spa_max_ashift) {
2039 return (SET_ERROR(EINVAL));
2043 * All vdevs in normal class must have the same ashift
2046 vdev_t *rvd = spa->spa_root_vdev;
2047 int num_indirect = 0;
2048 for (uint64_t id = 0; id < rvd->vdev_children; id++) {
2049 vdev_t *cvd = rvd->vdev_child[id];
2052 * A removed special/dedup vdev must have the same ashift
2053 * across all vdevs in its class.
2055 if (vd->vdev_alloc_bias != VDEV_BIAS_NONE &&
2056 cvd->vdev_alloc_bias == vd->vdev_alloc_bias &&
2057 cvd->vdev_ashift != vd->vdev_ashift) {
2058 return (SET_ERROR(EINVAL));
2060 if (cvd->vdev_ashift != 0 &&
2061 cvd->vdev_alloc_bias == VDEV_BIAS_NONE)
2062 ASSERT3U(cvd->vdev_ashift, ==, spa->spa_max_ashift);
2063 if (cvd->vdev_ops == &vdev_indirect_ops)
2065 if (!vdev_is_concrete(cvd))
2067 if (cvd->vdev_ops == &vdev_raidz_ops)
2068 return (SET_ERROR(EINVAL));
2070 * Need the mirror to be mirror of leaf vdevs only
2072 if (cvd->vdev_ops == &vdev_mirror_ops) {
2073 for (uint64_t cid = 0;
2074 cid < cvd->vdev_children; cid++) {
2075 if (!cvd->vdev_child[cid]->vdev_ops->
2077 return (SET_ERROR(EINVAL));
2086 * Initiate removal of a top-level vdev, reducing the total space in the pool.
2087 * The config lock is held for the specified TXG. Once initiated,
2088 * evacuation of all allocated space (copying it to other vdevs) happens
2089 * in the background (see spa_vdev_remove_thread()), and can be canceled
2090 * (see spa_vdev_remove_cancel()). If successful, the vdev will
2091 * be transformed to an indirect vdev (see spa_vdev_remove_complete()).
2094 spa_vdev_remove_top(vdev_t *vd, uint64_t *txg)
2096 spa_t *spa = vd->vdev_spa;
2100 * Check for errors up-front, so that we don't waste time
2101 * passivating the metaslab group and clearing the ZIL if there
2104 error = spa_vdev_remove_top_check(vd);
2109 * Stop allocating from this vdev. Note that we must check
2110 * that this is not the only device in the pool before
2111 * passivating, otherwise we will not be able to make
2112 * progress because we can't allocate from any vdevs.
2113 * The above check for sufficient free space serves this
2116 metaslab_group_t *mg = vd->vdev_mg;
2117 metaslab_group_passivate(mg);
2120 * Wait for the youngest allocations and frees to sync,
2121 * and then wait for the deferral of those frees to finish.
2123 spa_vdev_config_exit(spa, NULL,
2124 *txg + TXG_CONCURRENT_STATES + TXG_DEFER_SIZE, 0, FTAG);
2127 * We must ensure that no "stubby" log blocks are allocated
2128 * on the device to be removed. These blocks could be
2129 * written at any time, including while we are in the middle
2132 error = spa_reset_logs(spa);
2135 * We stop any initializing and TRIM that is currently in progress
2136 * but leave the state as "active". This will allow the process to
2137 * resume if the removal is canceled sometime later.
2139 vdev_initialize_stop_all(vd, VDEV_INITIALIZE_ACTIVE);
2140 vdev_trim_stop_all(vd, VDEV_TRIM_ACTIVE);
2141 vdev_autotrim_stop_wait(vd);
2143 *txg = spa_vdev_config_enter(spa);
2146 * Things might have changed while the config lock was dropped
2147 * (e.g. space usage). Check for errors again.
2150 error = spa_vdev_remove_top_check(vd);
2153 metaslab_group_activate(mg);
2154 spa_async_request(spa, SPA_ASYNC_INITIALIZE_RESTART);
2155 spa_async_request(spa, SPA_ASYNC_TRIM_RESTART);
2156 spa_async_request(spa, SPA_ASYNC_AUTOTRIM_RESTART);
2160 vd->vdev_removing = B_TRUE;
2162 vdev_dirty_leaves(vd, VDD_DTL, *txg);
2163 vdev_config_dirty(vd);
2164 dmu_tx_t *tx = dmu_tx_create_assigned(spa->spa_dsl_pool, *txg);
2165 dsl_sync_task_nowait(spa->spa_dsl_pool,
2166 vdev_remove_initiate_sync, (void *)(uintptr_t)vd->vdev_id, tx);
2173 * Remove a device from the pool.
2175 * Removing a device from the vdev namespace requires several steps
2176 * and can take a significant amount of time. As a result we use
2177 * the spa_vdev_config_[enter/exit] functions which allow us to
2178 * grab and release the spa_config_lock while still holding the namespace
2179 * lock. During each step the configuration is synced out.
2182 spa_vdev_remove(spa_t *spa, uint64_t guid, boolean_t unspare)
2185 nvlist_t **spares, **l2cache, *nv;
2187 uint_t nspares, nl2cache;
2188 int error = 0, error_log;
2189 boolean_t locked = MUTEX_HELD(&spa_namespace_lock);
2190 sysevent_t *ev = NULL;
2191 char *vd_type = NULL, *vd_path = NULL;
2193 ASSERT(spa_writeable(spa));
2196 txg = spa_vdev_enter(spa);
2198 ASSERT(MUTEX_HELD(&spa_namespace_lock));
2199 if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
2200 error = (spa_has_checkpoint(spa)) ?
2201 ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT;
2204 return (spa_vdev_exit(spa, NULL, txg, error));
2209 vd = spa_lookup_by_guid(spa, guid, B_FALSE);
2211 if (spa->spa_spares.sav_vdevs != NULL &&
2212 nvlist_lookup_nvlist_array(spa->spa_spares.sav_config,
2213 ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0 &&
2214 (nv = spa_nvlist_lookup_by_guid(spares, nspares, guid)) != NULL) {
2216 * Only remove the hot spare if it's not currently in use
2219 if (vd == NULL || unspare) {
2221 vd = spa_lookup_by_guid(spa, guid, B_TRUE);
2222 ev = spa_event_create(spa, vd, NULL,
2223 ESC_ZFS_VDEV_REMOVE_AUX);
2225 vd_type = VDEV_TYPE_SPARE;
2226 vd_path = spa_strdup(fnvlist_lookup_string(
2227 nv, ZPOOL_CONFIG_PATH));
2228 spa_vdev_remove_aux(spa->spa_spares.sav_config,
2229 ZPOOL_CONFIG_SPARES, spares, nspares, nv);
2230 spa_load_spares(spa);
2231 spa->spa_spares.sav_sync = B_TRUE;
2233 error = SET_ERROR(EBUSY);
2235 } else if (spa->spa_l2cache.sav_vdevs != NULL &&
2236 nvlist_lookup_nvlist_array(spa->spa_l2cache.sav_config,
2237 ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0 &&
2238 (nv = spa_nvlist_lookup_by_guid(l2cache, nl2cache, guid)) != NULL) {
2239 vd_type = VDEV_TYPE_L2CACHE;
2240 vd_path = spa_strdup(fnvlist_lookup_string(
2241 nv, ZPOOL_CONFIG_PATH));
2243 * Cache devices can always be removed.
2245 vd = spa_lookup_by_guid(spa, guid, B_TRUE);
2248 * Stop trimming the cache device. We need to release the
2249 * config lock to allow the syncing of TRIM transactions
2250 * without releasing the spa_namespace_lock. The same
2251 * strategy is employed in spa_vdev_remove_top().
2253 spa_vdev_config_exit(spa, NULL,
2254 txg + TXG_CONCURRENT_STATES + TXG_DEFER_SIZE, 0, FTAG);
2255 mutex_enter(&vd->vdev_trim_lock);
2256 vdev_trim_stop(vd, VDEV_TRIM_CANCELED, NULL);
2257 mutex_exit(&vd->vdev_trim_lock);
2258 txg = spa_vdev_config_enter(spa);
2260 ev = spa_event_create(spa, vd, NULL, ESC_ZFS_VDEV_REMOVE_AUX);
2261 spa_vdev_remove_aux(spa->spa_l2cache.sav_config,
2262 ZPOOL_CONFIG_L2CACHE, l2cache, nl2cache, nv);
2263 spa_load_l2cache(spa);
2264 spa->spa_l2cache.sav_sync = B_TRUE;
2265 } else if (vd != NULL && vd->vdev_islog) {
2267 vd_type = VDEV_TYPE_LOG;
2268 vd_path = spa_strdup((vd->vdev_path != NULL) ?
2269 vd->vdev_path : "-");
2270 error = spa_vdev_remove_log(vd, &txg);
2271 } else if (vd != NULL) {
2273 error = spa_vdev_remove_top(vd, &txg);
2276 * There is no vdev of any kind with the specified guid.
2278 error = SET_ERROR(ENOENT);
2284 error = spa_vdev_exit(spa, NULL, txg, error);
2287 * Logging must be done outside the spa config lock. Otherwise,
2288 * this code path could end up holding the spa config lock while
2289 * waiting for a txg_sync so it can write to the internal log.
2290 * Doing that would prevent the txg sync from actually happening,
2291 * causing a deadlock.
2293 if (error_log == 0 && vd_type != NULL && vd_path != NULL) {
2294 spa_history_log_internal(spa, "vdev remove", NULL,
2295 "%s vdev (%s) %s", spa_name(spa), vd_type, vd_path);
2297 if (vd_path != NULL)
2298 spa_strfree(vd_path);
2307 spa_removal_get_stats(spa_t *spa, pool_removal_stat_t *prs)
2309 prs->prs_state = spa->spa_removing_phys.sr_state;
2311 if (prs->prs_state == DSS_NONE)
2312 return (SET_ERROR(ENOENT));
2314 prs->prs_removing_vdev = spa->spa_removing_phys.sr_removing_vdev;
2315 prs->prs_start_time = spa->spa_removing_phys.sr_start_time;
2316 prs->prs_end_time = spa->spa_removing_phys.sr_end_time;
2317 prs->prs_to_copy = spa->spa_removing_phys.sr_to_copy;
2318 prs->prs_copied = spa->spa_removing_phys.sr_copied;
2320 prs->prs_mapping_memory = 0;
2321 uint64_t indirect_vdev_id =
2322 spa->spa_removing_phys.sr_prev_indirect_vdev;
2323 while (indirect_vdev_id != -1) {
2324 vdev_t *vd = spa->spa_root_vdev->vdev_child[indirect_vdev_id];
2325 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
2326 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
2328 ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
2329 prs->prs_mapping_memory += vdev_indirect_mapping_size(vim);
2330 indirect_vdev_id = vic->vic_prev_indirect_vdev;
2337 ZFS_MODULE_PARAM(zfs_vdev, zfs_, removal_ignore_errors, INT, ZMOD_RW,
2338 "Ignore hard IO errors when removing device");
2340 ZFS_MODULE_PARAM(zfs_vdev, zfs_, remove_max_segment, INT, ZMOD_RW,
2341 "Largest contiguous segment to allocate when removing device");
2343 ZFS_MODULE_PARAM(zfs_vdev, vdev_, removal_max_span, INT, ZMOD_RW,
2344 "Largest span of free chunks a remap segment can span");
2346 ZFS_MODULE_PARAM(zfs_vdev, zfs_, removal_suspend_progress, INT, ZMOD_RW,
2347 "Pause device removal after this many bytes are copied "
2348 "(debug use only - causes removal to hang)");
2351 EXPORT_SYMBOL(free_from_removing_vdev);
2352 EXPORT_SYMBOL(spa_removal_get_stats);
2353 EXPORT_SYMBOL(spa_remove_init);
2354 EXPORT_SYMBOL(spa_restart_removal);
2355 EXPORT_SYMBOL(spa_vdev_removal_destroy);
2356 EXPORT_SYMBOL(spa_vdev_remove);
2357 EXPORT_SYMBOL(spa_vdev_remove_cancel);
2358 EXPORT_SYMBOL(spa_vdev_remove_suspend);
2359 EXPORT_SYMBOL(svr_sync);