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, 2018 by Delphix. All rights reserved.
27 #include <sys/zfs_context.h>
28 #include <sys/spa_impl.h>
30 #include <sys/dmu_tx.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/metaslab.h>
34 #include <sys/metaslab_impl.h>
35 #include <sys/uberblock_impl.h>
38 #include <sys/bpobj.h>
39 #include <sys/dsl_pool.h>
40 #include <sys/dsl_synctask.h>
41 #include <sys/dsl_dir.h>
43 #include <sys/zfeature.h>
44 #include <sys/vdev_indirect_births.h>
45 #include <sys/vdev_indirect_mapping.h>
47 #include <sys/trace_vdev.h>
50 * This file contains the necessary logic to remove vdevs from a
51 * storage pool. Currently, the only devices that can be removed
52 * are log, cache, and spare devices; and top level vdevs from a pool
53 * w/o raidz or mirrors. (Note that members of a mirror can be removed
54 * by the detach operation.)
56 * Log vdevs are removed by evacuating them and then turning the vdev
57 * into a hole vdev while holding spa config locks.
59 * Top level vdevs are removed and converted into an indirect vdev via
60 * a multi-step process:
62 * - Disable allocations from this device (spa_vdev_remove_top).
64 * - From a new thread (spa_vdev_remove_thread), copy data from
65 * the removing vdev to a different vdev. The copy happens in open
66 * context (spa_vdev_copy_impl) and issues a sync task
67 * (vdev_mapping_sync) so the sync thread can update the partial
68 * indirect mappings in core and on disk.
70 * - If a free happens during a removal, it is freed from the
71 * removing vdev, and if it has already been copied, from the new
72 * location as well (free_from_removing_vdev).
74 * - After the removal is completed, the copy thread converts the vdev
75 * into an indirect vdev (vdev_remove_complete) before instructing
76 * the sync thread to destroy the space maps and finish the removal
77 * (spa_finish_removal).
80 typedef struct vdev_copy_arg
{
82 uint64_t vca_outstanding_bytes
;
88 * The maximum amount of memory we can use for outstanding i/o while
89 * doing a device removal. This determines how much i/o we can have
90 * in flight concurrently.
92 int zfs_remove_max_copy_bytes
= 64 * 1024 * 1024;
95 * The largest contiguous segment that we will attempt to allocate when
96 * removing a device. This can be no larger than SPA_MAXBLOCKSIZE. If
97 * there is a performance problem with attempting to allocate large blocks,
98 * consider decreasing this.
100 int zfs_remove_max_segment
= SPA_MAXBLOCKSIZE
;
103 * Allow a remap segment to span free chunks of at most this size. The main
104 * impact of a larger span is that we will read and write larger, more
105 * contiguous chunks, with more "unnecessary" data -- trading off bandwidth
106 * for iops. The value here was chosen to align with
107 * zfs_vdev_read_gap_limit, which is a similar concept when doing regular
108 * reads (but there's no reason it has to be the same).
110 * Additionally, a higher span will have the following relatively minor
112 * - the mapping will be smaller, since one entry can cover more allocated
114 * - more of the fragmentation in the removing device will be preserved
115 * - we'll do larger allocations, which may fail and fall back on smaller
118 int vdev_removal_max_span
= 32 * 1024;
121 * This is used by the test suite so that it can ensure that certain
122 * actions happen while in the middle of a removal.
124 unsigned long zfs_remove_max_bytes_pause
= -1UL;
126 #define VDEV_REMOVAL_ZAP_OBJS "lzap"
128 static void spa_vdev_remove_thread(void *arg
);
131 spa_sync_removing_state(spa_t
*spa
, dmu_tx_t
*tx
)
133 VERIFY0(zap_update(spa
->spa_dsl_pool
->dp_meta_objset
,
134 DMU_POOL_DIRECTORY_OBJECT
,
135 DMU_POOL_REMOVING
, sizeof (uint64_t),
136 sizeof (spa
->spa_removing_phys
) / sizeof (uint64_t),
137 &spa
->spa_removing_phys
, tx
));
141 spa_nvlist_lookup_by_guid(nvlist_t
**nvpp
, int count
, uint64_t target_guid
)
143 for (int i
= 0; i
< count
; i
++) {
145 fnvlist_lookup_uint64(nvpp
[i
], ZPOOL_CONFIG_GUID
);
147 if (guid
== target_guid
)
155 spa_vdev_remove_aux(nvlist_t
*config
, char *name
, nvlist_t
**dev
, int count
,
156 nvlist_t
*dev_to_remove
)
158 nvlist_t
**newdev
= NULL
;
161 newdev
= kmem_alloc((count
- 1) * sizeof (void *), KM_SLEEP
);
163 for (int i
= 0, j
= 0; i
< count
; i
++) {
164 if (dev
[i
] == dev_to_remove
)
166 VERIFY(nvlist_dup(dev
[i
], &newdev
[j
++], KM_SLEEP
) == 0);
169 VERIFY(nvlist_remove(config
, name
, DATA_TYPE_NVLIST_ARRAY
) == 0);
170 VERIFY(nvlist_add_nvlist_array(config
, name
, newdev
, count
- 1) == 0);
172 for (int i
= 0; i
< count
- 1; i
++)
173 nvlist_free(newdev
[i
]);
176 kmem_free(newdev
, (count
- 1) * sizeof (void *));
179 static spa_vdev_removal_t
*
180 spa_vdev_removal_create(vdev_t
*vd
)
182 spa_vdev_removal_t
*svr
= kmem_zalloc(sizeof (*svr
), KM_SLEEP
);
183 mutex_init(&svr
->svr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
184 cv_init(&svr
->svr_cv
, NULL
, CV_DEFAULT
, NULL
);
185 svr
->svr_allocd_segs
= range_tree_create(NULL
, NULL
);
186 svr
->svr_vdev_id
= vd
->vdev_id
;
188 for (int i
= 0; i
< TXG_SIZE
; i
++) {
189 svr
->svr_frees
[i
] = range_tree_create(NULL
, NULL
);
190 list_create(&svr
->svr_new_segments
[i
],
191 sizeof (vdev_indirect_mapping_entry_t
),
192 offsetof(vdev_indirect_mapping_entry_t
, vime_node
));
199 spa_vdev_removal_destroy(spa_vdev_removal_t
*svr
)
201 for (int i
= 0; i
< TXG_SIZE
; i
++) {
202 ASSERT0(svr
->svr_bytes_done
[i
]);
203 ASSERT0(svr
->svr_max_offset_to_sync
[i
]);
204 range_tree_destroy(svr
->svr_frees
[i
]);
205 list_destroy(&svr
->svr_new_segments
[i
]);
208 range_tree_destroy(svr
->svr_allocd_segs
);
209 mutex_destroy(&svr
->svr_lock
);
210 cv_destroy(&svr
->svr_cv
);
211 kmem_free(svr
, sizeof (*svr
));
215 * This is called as a synctask in the txg in which we will mark this vdev
216 * as removing (in the config stored in the MOS).
218 * It begins the evacuation of a toplevel vdev by:
219 * - initializing the spa_removing_phys which tracks this removal
220 * - computing the amount of space to remove for accounting purposes
221 * - dirtying all dbufs in the spa_config_object
222 * - creating the spa_vdev_removal
223 * - starting the spa_vdev_remove_thread
226 vdev_remove_initiate_sync(void *arg
, dmu_tx_t
*tx
)
228 int vdev_id
= (uintptr_t)arg
;
229 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
230 vdev_t
*vd
= vdev_lookup_top(spa
, vdev_id
);
231 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
232 objset_t
*mos
= spa
->spa_dsl_pool
->dp_meta_objset
;
233 spa_vdev_removal_t
*svr
= NULL
;
234 ASSERTV(uint64_t txg
= dmu_tx_get_txg(tx
));
236 ASSERT3P(vd
->vdev_ops
, !=, &vdev_raidz_ops
);
237 svr
= spa_vdev_removal_create(vd
);
239 ASSERT(vd
->vdev_removing
);
240 ASSERT3P(vd
->vdev_indirect_mapping
, ==, NULL
);
242 spa_feature_incr(spa
, SPA_FEATURE_DEVICE_REMOVAL
, tx
);
243 if (spa_feature_is_enabled(spa
, SPA_FEATURE_OBSOLETE_COUNTS
)) {
245 * By activating the OBSOLETE_COUNTS feature, we prevent
246 * the pool from being downgraded and ensure that the
247 * refcounts are precise.
249 spa_feature_incr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
251 VERIFY0(zap_add(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
252 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE
, sizeof (one
), 1,
254 ASSERTV(boolean_t are_precise
);
255 ASSERT0(vdev_obsolete_counts_are_precise(vd
, &are_precise
));
256 ASSERT3B(are_precise
, ==, B_TRUE
);
259 vic
->vic_mapping_object
= vdev_indirect_mapping_alloc(mos
, tx
);
260 vd
->vdev_indirect_mapping
=
261 vdev_indirect_mapping_open(mos
, vic
->vic_mapping_object
);
262 vic
->vic_births_object
= vdev_indirect_births_alloc(mos
, tx
);
263 vd
->vdev_indirect_births
=
264 vdev_indirect_births_open(mos
, vic
->vic_births_object
);
265 spa
->spa_removing_phys
.sr_removing_vdev
= vd
->vdev_id
;
266 spa
->spa_removing_phys
.sr_start_time
= gethrestime_sec();
267 spa
->spa_removing_phys
.sr_end_time
= 0;
268 spa
->spa_removing_phys
.sr_state
= DSS_SCANNING
;
269 spa
->spa_removing_phys
.sr_to_copy
= 0;
270 spa
->spa_removing_phys
.sr_copied
= 0;
273 * Note: We can't use vdev_stat's vs_alloc for sr_to_copy, because
274 * there may be space in the defer tree, which is free, but still
275 * counted in vs_alloc.
277 for (uint64_t i
= 0; i
< vd
->vdev_ms_count
; i
++) {
278 metaslab_t
*ms
= vd
->vdev_ms
[i
];
279 if (ms
->ms_sm
== NULL
)
283 * Sync tasks happen before metaslab_sync(), therefore
284 * smp_alloc and sm_alloc must be the same.
286 ASSERT3U(space_map_allocated(ms
->ms_sm
), ==,
287 ms
->ms_sm
->sm_phys
->smp_alloc
);
289 spa
->spa_removing_phys
.sr_to_copy
+=
290 space_map_allocated(ms
->ms_sm
);
293 * Space which we are freeing this txg does not need to
296 spa
->spa_removing_phys
.sr_to_copy
-=
297 range_tree_space(ms
->ms_freeing
);
299 ASSERT0(range_tree_space(ms
->ms_freed
));
300 for (int t
= 0; t
< TXG_SIZE
; t
++)
301 ASSERT0(range_tree_space(ms
->ms_allocating
[t
]));
305 * Sync tasks are called before metaslab_sync(), so there should
306 * be no already-synced metaslabs in the TXG_CLEAN list.
308 ASSERT3P(txg_list_head(&vd
->vdev_ms_list
, TXG_CLEAN(txg
)), ==, NULL
);
310 spa_sync_removing_state(spa
, tx
);
313 * All blocks that we need to read the most recent mapping must be
314 * stored on concrete vdevs. Therefore, we must dirty anything that
315 * is read before spa_remove_init(). Specifically, the
316 * spa_config_object. (Note that although we already modified the
317 * spa_config_object in spa_sync_removing_state, that may not have
318 * modified all blocks of the object.)
320 dmu_object_info_t doi
;
321 VERIFY0(dmu_object_info(mos
, DMU_POOL_DIRECTORY_OBJECT
, &doi
));
322 for (uint64_t offset
= 0; offset
< doi
.doi_max_offset
; ) {
324 VERIFY0(dmu_buf_hold(mos
, DMU_POOL_DIRECTORY_OBJECT
,
325 offset
, FTAG
, &dbuf
, 0));
326 dmu_buf_will_dirty(dbuf
, tx
);
327 offset
+= dbuf
->db_size
;
328 dmu_buf_rele(dbuf
, FTAG
);
332 * Now that we've allocated the im_object, dirty the vdev to ensure
333 * that the object gets written to the config on disk.
335 vdev_config_dirty(vd
);
337 zfs_dbgmsg("starting removal thread for vdev %llu (%p) in txg %llu "
338 "im_obj=%llu", vd
->vdev_id
, vd
, dmu_tx_get_txg(tx
),
339 vic
->vic_mapping_object
);
341 spa_history_log_internal(spa
, "vdev remove started", tx
,
342 "%s vdev %llu %s", spa_name(spa
), vd
->vdev_id
,
343 (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
345 * Setting spa_vdev_removal causes subsequent frees to call
346 * free_from_removing_vdev(). Note that we don't need any locking
347 * because we are the sync thread, and metaslab_free_impl() is only
348 * called from syncing context (potentially from a zio taskq thread,
349 * but in any case only when there are outstanding free i/os, which
352 ASSERT3P(spa
->spa_vdev_removal
, ==, NULL
);
353 spa
->spa_vdev_removal
= svr
;
354 svr
->svr_thread
= thread_create(NULL
, 0,
355 spa_vdev_remove_thread
, spa
, 0, &p0
, TS_RUN
, minclsyspri
);
359 * When we are opening a pool, we must read the mapping for each
360 * indirect vdev in order from most recently removed to least
361 * recently removed. We do this because the blocks for the mapping
362 * of older indirect vdevs may be stored on more recently removed vdevs.
363 * In order to read each indirect mapping object, we must have
364 * initialized all more recently removed vdevs.
367 spa_remove_init(spa_t
*spa
)
371 error
= zap_lookup(spa
->spa_dsl_pool
->dp_meta_objset
,
372 DMU_POOL_DIRECTORY_OBJECT
,
373 DMU_POOL_REMOVING
, sizeof (uint64_t),
374 sizeof (spa
->spa_removing_phys
) / sizeof (uint64_t),
375 &spa
->spa_removing_phys
);
377 if (error
== ENOENT
) {
378 spa
->spa_removing_phys
.sr_state
= DSS_NONE
;
379 spa
->spa_removing_phys
.sr_removing_vdev
= -1;
380 spa
->spa_removing_phys
.sr_prev_indirect_vdev
= -1;
381 spa
->spa_indirect_vdevs_loaded
= B_TRUE
;
383 } else if (error
!= 0) {
387 if (spa
->spa_removing_phys
.sr_state
== DSS_SCANNING
) {
389 * We are currently removing a vdev. Create and
390 * initialize a spa_vdev_removal_t from the bonus
391 * buffer of the removing vdevs vdev_im_object, and
392 * initialize its partial mapping.
394 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
395 vdev_t
*vd
= vdev_lookup_top(spa
,
396 spa
->spa_removing_phys
.sr_removing_vdev
);
399 spa_config_exit(spa
, SCL_STATE
, FTAG
);
403 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
405 ASSERT(vdev_is_concrete(vd
));
406 spa_vdev_removal_t
*svr
= spa_vdev_removal_create(vd
);
407 ASSERT3U(svr
->svr_vdev_id
, ==, vd
->vdev_id
);
408 ASSERT(vd
->vdev_removing
);
410 vd
->vdev_indirect_mapping
= vdev_indirect_mapping_open(
411 spa
->spa_meta_objset
, vic
->vic_mapping_object
);
412 vd
->vdev_indirect_births
= vdev_indirect_births_open(
413 spa
->spa_meta_objset
, vic
->vic_births_object
);
414 spa_config_exit(spa
, SCL_STATE
, FTAG
);
416 spa
->spa_vdev_removal
= svr
;
419 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
420 uint64_t indirect_vdev_id
=
421 spa
->spa_removing_phys
.sr_prev_indirect_vdev
;
422 while (indirect_vdev_id
!= UINT64_MAX
) {
423 vdev_t
*vd
= vdev_lookup_top(spa
, indirect_vdev_id
);
424 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
426 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
427 vd
->vdev_indirect_mapping
= vdev_indirect_mapping_open(
428 spa
->spa_meta_objset
, vic
->vic_mapping_object
);
429 vd
->vdev_indirect_births
= vdev_indirect_births_open(
430 spa
->spa_meta_objset
, vic
->vic_births_object
);
432 indirect_vdev_id
= vic
->vic_prev_indirect_vdev
;
434 spa_config_exit(spa
, SCL_STATE
, FTAG
);
437 * Now that we've loaded all the indirect mappings, we can allow
438 * reads from other blocks (e.g. via predictive prefetch).
440 spa
->spa_indirect_vdevs_loaded
= B_TRUE
;
445 spa_restart_removal(spa_t
*spa
)
447 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
453 * In general when this function is called there is no
454 * removal thread running. The only scenario where this
455 * is not true is during spa_import() where this function
456 * is called twice [once from spa_import_impl() and
457 * spa_async_resume()]. Thus, in the scenario where we
458 * import a pool that has an ongoing removal we don't
459 * want to spawn a second thread.
461 if (svr
->svr_thread
!= NULL
)
464 if (!spa_writeable(spa
))
467 zfs_dbgmsg("restarting removal of %llu", svr
->svr_vdev_id
);
468 svr
->svr_thread
= thread_create(NULL
, 0, spa_vdev_remove_thread
, spa
,
469 0, &p0
, TS_RUN
, minclsyspri
);
473 * Process freeing from a device which is in the middle of being removed.
474 * We must handle this carefully so that we attempt to copy freed data,
475 * and we correctly free already-copied data.
478 free_from_removing_vdev(vdev_t
*vd
, uint64_t offset
, uint64_t size
)
480 spa_t
*spa
= vd
->vdev_spa
;
481 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
482 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
483 uint64_t txg
= spa_syncing_txg(spa
);
484 uint64_t max_offset_yet
= 0;
486 ASSERT(vd
->vdev_indirect_config
.vic_mapping_object
!= 0);
487 ASSERT3U(vd
->vdev_indirect_config
.vic_mapping_object
, ==,
488 vdev_indirect_mapping_object(vim
));
489 ASSERT3U(vd
->vdev_id
, ==, svr
->svr_vdev_id
);
491 mutex_enter(&svr
->svr_lock
);
494 * Remove the segment from the removing vdev's spacemap. This
495 * ensures that we will not attempt to copy this space (if the
496 * removal thread has not yet visited it), and also ensures
497 * that we know what is actually allocated on the new vdevs
498 * (needed if we cancel the removal).
500 * Note: we must do the metaslab_free_concrete() with the svr_lock
501 * held, so that the remove_thread can not load this metaslab and then
502 * visit this offset between the time that we metaslab_free_concrete()
503 * and when we check to see if it has been visited.
505 * Note: The checkpoint flag is set to false as having/taking
506 * a checkpoint and removing a device can't happen at the same
509 ASSERT(!spa_has_checkpoint(spa
));
510 metaslab_free_concrete(vd
, offset
, size
, B_FALSE
);
512 uint64_t synced_size
= 0;
513 uint64_t synced_offset
= 0;
514 uint64_t max_offset_synced
= vdev_indirect_mapping_max_offset(vim
);
515 if (offset
< max_offset_synced
) {
517 * The mapping for this offset is already on disk.
518 * Free from the new location.
520 * Note that we use svr_max_synced_offset because it is
521 * updated atomically with respect to the in-core mapping.
522 * By contrast, vim_max_offset is not.
524 * This block may be split between a synced entry and an
525 * in-flight or unvisited entry. Only process the synced
526 * portion of it here.
528 synced_size
= MIN(size
, max_offset_synced
- offset
);
529 synced_offset
= offset
;
531 ASSERT3U(max_offset_yet
, <=, max_offset_synced
);
532 max_offset_yet
= max_offset_synced
;
534 DTRACE_PROBE3(remove__free__synced
,
537 uint64_t, synced_size
);
540 offset
+= synced_size
;
544 * Look at all in-flight txgs starting from the currently syncing one
545 * and see if a section of this free is being copied. By starting from
546 * this txg and iterating forward, we might find that this region
547 * was copied in two different txgs and handle it appropriately.
549 for (int i
= 0; i
< TXG_CONCURRENT_STATES
; i
++) {
550 int txgoff
= (txg
+ i
) & TXG_MASK
;
551 if (size
> 0 && offset
< svr
->svr_max_offset_to_sync
[txgoff
]) {
553 * The mapping for this offset is in flight, and
554 * will be synced in txg+i.
556 uint64_t inflight_size
= MIN(size
,
557 svr
->svr_max_offset_to_sync
[txgoff
] - offset
);
559 DTRACE_PROBE4(remove__free__inflight
,
562 uint64_t, inflight_size
,
566 * We copy data in order of increasing offset.
567 * Therefore the max_offset_to_sync[] must increase
568 * (or be zero, indicating that nothing is being
569 * copied in that txg).
571 if (svr
->svr_max_offset_to_sync
[txgoff
] != 0) {
572 ASSERT3U(svr
->svr_max_offset_to_sync
[txgoff
],
575 svr
->svr_max_offset_to_sync
[txgoff
];
579 * We've already committed to copying this segment:
580 * we have allocated space elsewhere in the pool for
581 * it and have an IO outstanding to copy the data. We
582 * cannot free the space before the copy has
583 * completed, or else the copy IO might overwrite any
584 * new data. To free that space, we record the
585 * segment in the appropriate svr_frees tree and free
586 * the mapped space later, in the txg where we have
587 * completed the copy and synced the mapping (see
588 * vdev_mapping_sync).
590 range_tree_add(svr
->svr_frees
[txgoff
],
591 offset
, inflight_size
);
592 size
-= inflight_size
;
593 offset
+= inflight_size
;
596 * This space is already accounted for as being
597 * done, because it is being copied in txg+i.
598 * However, if i!=0, then it is being copied in
599 * a future txg. If we crash after this txg
600 * syncs but before txg+i syncs, then the space
601 * will be free. Therefore we must account
602 * for the space being done in *this* txg
603 * (when it is freed) rather than the future txg
604 * (when it will be copied).
606 ASSERT3U(svr
->svr_bytes_done
[txgoff
], >=,
608 svr
->svr_bytes_done
[txgoff
] -= inflight_size
;
609 svr
->svr_bytes_done
[txg
& TXG_MASK
] += inflight_size
;
612 ASSERT0(svr
->svr_max_offset_to_sync
[TXG_CLEAN(txg
) & TXG_MASK
]);
616 * The copy thread has not yet visited this offset. Ensure
620 DTRACE_PROBE3(remove__free__unvisited
,
625 if (svr
->svr_allocd_segs
!= NULL
)
626 range_tree_clear(svr
->svr_allocd_segs
, offset
, size
);
629 * Since we now do not need to copy this data, for
630 * accounting purposes we have done our job and can count
633 svr
->svr_bytes_done
[txg
& TXG_MASK
] += size
;
635 mutex_exit(&svr
->svr_lock
);
638 * Now that we have dropped svr_lock, process the synced portion
641 if (synced_size
> 0) {
642 vdev_indirect_mark_obsolete(vd
, synced_offset
, synced_size
);
645 * Note: this can only be called from syncing context,
646 * and the vdev_indirect_mapping is only changed from the
647 * sync thread, so we don't need svr_lock while doing
648 * metaslab_free_impl_cb.
650 boolean_t checkpoint
= B_FALSE
;
651 vdev_indirect_ops
.vdev_op_remap(vd
, synced_offset
, synced_size
,
652 metaslab_free_impl_cb
, &checkpoint
);
657 * Stop an active removal and update the spa_removing phys.
660 spa_finish_removal(spa_t
*spa
, dsl_scan_state_t state
, dmu_tx_t
*tx
)
662 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
663 ASSERT3U(dmu_tx_get_txg(tx
), ==, spa_syncing_txg(spa
));
665 /* Ensure the removal thread has completed before we free the svr. */
666 spa_vdev_remove_suspend(spa
);
668 ASSERT(state
== DSS_FINISHED
|| state
== DSS_CANCELED
);
670 if (state
== DSS_FINISHED
) {
671 spa_removing_phys_t
*srp
= &spa
->spa_removing_phys
;
672 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
673 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
675 if (srp
->sr_prev_indirect_vdev
!= UINT64_MAX
) {
677 pvd
= vdev_lookup_top(spa
,
678 srp
->sr_prev_indirect_vdev
);
679 ASSERT3P(pvd
->vdev_ops
, ==, &vdev_indirect_ops
);
682 vic
->vic_prev_indirect_vdev
= srp
->sr_prev_indirect_vdev
;
683 srp
->sr_prev_indirect_vdev
= vd
->vdev_id
;
685 spa
->spa_removing_phys
.sr_state
= state
;
686 spa
->spa_removing_phys
.sr_end_time
= gethrestime_sec();
688 spa
->spa_vdev_removal
= NULL
;
689 spa_vdev_removal_destroy(svr
);
691 spa_sync_removing_state(spa
, tx
);
693 vdev_config_dirty(spa
->spa_root_vdev
);
697 free_mapped_segment_cb(void *arg
, uint64_t offset
, uint64_t size
)
700 vdev_indirect_mark_obsolete(vd
, offset
, size
);
701 boolean_t checkpoint
= B_FALSE
;
702 vdev_indirect_ops
.vdev_op_remap(vd
, offset
, size
,
703 metaslab_free_impl_cb
, &checkpoint
);
707 * On behalf of the removal thread, syncs an incremental bit more of
708 * the indirect mapping to disk and updates the in-memory mapping.
709 * Called as a sync task in every txg that the removal thread makes progress.
712 vdev_mapping_sync(void *arg
, dmu_tx_t
*tx
)
714 spa_vdev_removal_t
*svr
= arg
;
715 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
716 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
717 ASSERTV(vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
);
718 uint64_t txg
= dmu_tx_get_txg(tx
);
719 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
721 ASSERT(vic
->vic_mapping_object
!= 0);
722 ASSERT3U(txg
, ==, spa_syncing_txg(spa
));
724 vdev_indirect_mapping_add_entries(vim
,
725 &svr
->svr_new_segments
[txg
& TXG_MASK
], tx
);
726 vdev_indirect_births_add_entry(vd
->vdev_indirect_births
,
727 vdev_indirect_mapping_max_offset(vim
), dmu_tx_get_txg(tx
), tx
);
730 * Free the copied data for anything that was freed while the
731 * mapping entries were in flight.
733 mutex_enter(&svr
->svr_lock
);
734 range_tree_vacate(svr
->svr_frees
[txg
& TXG_MASK
],
735 free_mapped_segment_cb
, vd
);
736 ASSERT3U(svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
], >=,
737 vdev_indirect_mapping_max_offset(vim
));
738 svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] = 0;
739 mutex_exit(&svr
->svr_lock
);
741 spa_sync_removing_state(spa
, tx
);
744 typedef struct vdev_copy_segment_arg
{
746 dva_t
*vcsa_dest_dva
;
748 range_tree_t
*vcsa_obsolete_segs
;
749 } vdev_copy_segment_arg_t
;
752 unalloc_seg(void *arg
, uint64_t start
, uint64_t size
)
754 vdev_copy_segment_arg_t
*vcsa
= arg
;
755 spa_t
*spa
= vcsa
->vcsa_spa
;
756 blkptr_t bp
= { { { {0} } } };
758 BP_SET_BIRTH(&bp
, TXG_INITIAL
, TXG_INITIAL
);
759 BP_SET_LSIZE(&bp
, size
);
760 BP_SET_PSIZE(&bp
, size
);
761 BP_SET_COMPRESS(&bp
, ZIO_COMPRESS_OFF
);
762 BP_SET_CHECKSUM(&bp
, ZIO_CHECKSUM_OFF
);
763 BP_SET_TYPE(&bp
, DMU_OT_NONE
);
764 BP_SET_LEVEL(&bp
, 0);
765 BP_SET_DEDUP(&bp
, 0);
766 BP_SET_BYTEORDER(&bp
, ZFS_HOST_BYTEORDER
);
768 DVA_SET_VDEV(&bp
.blk_dva
[0], DVA_GET_VDEV(vcsa
->vcsa_dest_dva
));
769 DVA_SET_OFFSET(&bp
.blk_dva
[0],
770 DVA_GET_OFFSET(vcsa
->vcsa_dest_dva
) + start
);
771 DVA_SET_ASIZE(&bp
.blk_dva
[0], size
);
773 zio_free(spa
, vcsa
->vcsa_txg
, &bp
);
777 * All reads and writes associated with a call to spa_vdev_copy_segment()
781 spa_vdev_copy_segment_done(zio_t
*zio
)
783 vdev_copy_segment_arg_t
*vcsa
= zio
->io_private
;
785 range_tree_vacate(vcsa
->vcsa_obsolete_segs
,
787 range_tree_destroy(vcsa
->vcsa_obsolete_segs
);
788 kmem_free(vcsa
, sizeof (*vcsa
));
790 spa_config_exit(zio
->io_spa
, SCL_STATE
, zio
->io_spa
);
794 * The write of the new location is done.
797 spa_vdev_copy_segment_write_done(zio_t
*zio
)
799 vdev_copy_arg_t
*vca
= zio
->io_private
;
801 abd_free(zio
->io_abd
);
803 mutex_enter(&vca
->vca_lock
);
804 vca
->vca_outstanding_bytes
-= zio
->io_size
;
805 cv_signal(&vca
->vca_cv
);
806 mutex_exit(&vca
->vca_lock
);
810 * The read of the old location is done. The parent zio is the write to
811 * the new location. Allow it to start.
814 spa_vdev_copy_segment_read_done(zio_t
*zio
)
816 zio_nowait(zio_unique_parent(zio
));
820 * If the old and new vdevs are mirrors, we will read both sides of the old
821 * mirror, and write each copy to the corresponding side of the new mirror.
822 * If the old and new vdevs have a different number of children, we will do
823 * this as best as possible. Since we aren't verifying checksums, this
824 * ensures that as long as there's a good copy of the data, we'll have a
825 * good copy after the removal, even if there's silent damage to one side
826 * of the mirror. If we're removing a mirror that has some silent damage,
827 * we'll have exactly the same damage in the new location (assuming that
828 * the new location is also a mirror).
830 * We accomplish this by creating a tree of zio_t's, with as many writes as
831 * there are "children" of the new vdev (a non-redundant vdev counts as one
832 * child, a 2-way mirror has 2 children, etc). Each write has an associated
833 * read from a child of the old vdev. Typically there will be the same
834 * number of children of the old and new vdevs. However, if there are more
835 * children of the new vdev, some child(ren) of the old vdev will be issued
836 * multiple reads. If there are more children of the old vdev, some copies
839 * For example, the tree of zio_t's for a 2-way mirror is:
843 * write(new vdev, child 0) write(new vdev, child 1)
845 * read(old vdev, child 0) read(old vdev, child 1)
847 * Child zio's complete before their parents complete. However, zio's
848 * created with zio_vdev_child_io() may be issued before their children
849 * complete. In this case we need to make sure that the children (reads)
850 * complete before the parents (writes) are *issued*. We do this by not
851 * calling zio_nowait() on each write until its corresponding read has
854 * The spa_config_lock must be held while zio's created by
855 * zio_vdev_child_io() are in progress, to ensure that the vdev tree does
856 * not change (e.g. due to a concurrent "zpool attach/detach"). The "null"
857 * zio is needed to release the spa_config_lock after all the reads and
858 * writes complete. (Note that we can't grab the config lock for each read,
859 * because it is not reentrant - we could deadlock with a thread waiting
863 spa_vdev_copy_one_child(vdev_copy_arg_t
*vca
, zio_t
*nzio
,
864 vdev_t
*source_vd
, uint64_t source_offset
,
865 vdev_t
*dest_child_vd
, uint64_t dest_offset
, int dest_id
, uint64_t size
)
867 ASSERT3U(spa_config_held(nzio
->io_spa
, SCL_ALL
, RW_READER
), !=, 0);
869 mutex_enter(&vca
->vca_lock
);
870 vca
->vca_outstanding_bytes
+= size
;
871 mutex_exit(&vca
->vca_lock
);
873 abd_t
*abd
= abd_alloc_for_io(size
, B_FALSE
);
875 vdev_t
*source_child_vd
;
876 if (source_vd
->vdev_ops
== &vdev_mirror_ops
&& dest_id
!= -1) {
878 * Source and dest are both mirrors. Copy from the same
879 * child id as we are copying to (wrapping around if there
880 * are more dest children than source children).
883 source_vd
->vdev_child
[dest_id
% source_vd
->vdev_children
];
885 source_child_vd
= source_vd
;
888 zio_t
*write_zio
= zio_vdev_child_io(nzio
, NULL
,
889 dest_child_vd
, dest_offset
, abd
, size
,
890 ZIO_TYPE_WRITE
, ZIO_PRIORITY_REMOVAL
,
892 spa_vdev_copy_segment_write_done
, vca
);
894 zio_nowait(zio_vdev_child_io(write_zio
, NULL
,
895 source_child_vd
, source_offset
, abd
, size
,
896 ZIO_TYPE_READ
, ZIO_PRIORITY_REMOVAL
,
898 spa_vdev_copy_segment_read_done
, vca
));
902 * Allocate a new location for this segment, and create the zio_t's to
903 * read from the old location and write to the new location.
906 spa_vdev_copy_segment(vdev_t
*vd
, range_tree_t
*segs
,
907 uint64_t maxalloc
, uint64_t txg
,
908 vdev_copy_arg_t
*vca
, zio_alloc_list_t
*zal
)
910 metaslab_group_t
*mg
= vd
->vdev_mg
;
911 spa_t
*spa
= vd
->vdev_spa
;
912 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
913 vdev_indirect_mapping_entry_t
*entry
;
915 uint64_t start
= range_tree_min(segs
);
917 ASSERT3U(maxalloc
, <=, SPA_MAXBLOCKSIZE
);
919 uint64_t size
= range_tree_span(segs
);
920 if (range_tree_span(segs
) > maxalloc
) {
922 * We can't allocate all the segments. Prefer to end
923 * the allocation at the end of a segment, thus avoiding
924 * additional split blocks.
928 search
.rs_start
= start
+ maxalloc
;
929 search
.rs_end
= search
.rs_start
;
930 range_seg_t
*rs
= avl_find(&segs
->rt_root
, &search
, &where
);
932 rs
= avl_nearest(&segs
->rt_root
, where
, AVL_BEFORE
);
934 rs
= AVL_PREV(&segs
->rt_root
, rs
);
937 size
= rs
->rs_end
- start
;
940 * There are no segments that end before maxalloc.
941 * I.e. the first segment is larger than maxalloc,
942 * so we must split it.
947 ASSERT3U(size
, <=, maxalloc
);
950 * An allocation class might not have any remaining vdevs or space
952 metaslab_class_t
*mc
= mg
->mg_class
;
953 if (mc
!= spa_normal_class(spa
) && mc
->mc_groups
<= 1)
954 mc
= spa_normal_class(spa
);
955 int error
= metaslab_alloc_dva(spa
, mc
, size
, &dst
, 0, NULL
, txg
, 0,
957 if (error
== ENOSPC
&& mc
!= spa_normal_class(spa
)) {
958 error
= metaslab_alloc_dva(spa
, spa_normal_class(spa
), size
,
959 &dst
, 0, NULL
, txg
, 0, zal
, 0);
965 * Determine the ranges that are not actually needed. Offsets are
966 * relative to the start of the range to be copied (i.e. relative to the
967 * local variable "start").
969 range_tree_t
*obsolete_segs
= range_tree_create(NULL
, NULL
);
971 range_seg_t
*rs
= avl_first(&segs
->rt_root
);
972 ASSERT3U(rs
->rs_start
, ==, start
);
973 uint64_t prev_seg_end
= rs
->rs_end
;
974 while ((rs
= AVL_NEXT(&segs
->rt_root
, rs
)) != NULL
) {
975 if (rs
->rs_start
>= start
+ size
) {
978 range_tree_add(obsolete_segs
,
979 prev_seg_end
- start
,
980 rs
->rs_start
- prev_seg_end
);
982 prev_seg_end
= rs
->rs_end
;
984 /* We don't end in the middle of an obsolete range */
985 ASSERT3U(start
+ size
, <=, prev_seg_end
);
987 range_tree_clear(segs
, start
, size
);
990 * We can't have any padding of the allocated size, otherwise we will
991 * misunderstand what's allocated, and the size of the mapping.
992 * The caller ensures this will be true by passing in a size that is
993 * aligned to the worst (highest) ashift in the pool.
995 ASSERT3U(DVA_GET_ASIZE(&dst
), ==, size
);
997 entry
= kmem_zalloc(sizeof (vdev_indirect_mapping_entry_t
), KM_SLEEP
);
998 DVA_MAPPING_SET_SRC_OFFSET(&entry
->vime_mapping
, start
);
999 entry
->vime_mapping
.vimep_dst
= dst
;
1000 if (spa_feature_is_enabled(spa
, SPA_FEATURE_OBSOLETE_COUNTS
)) {
1001 entry
->vime_obsolete_count
= range_tree_space(obsolete_segs
);
1004 vdev_copy_segment_arg_t
*vcsa
= kmem_zalloc(sizeof (*vcsa
), KM_SLEEP
);
1005 vcsa
->vcsa_dest_dva
= &entry
->vime_mapping
.vimep_dst
;
1006 vcsa
->vcsa_obsolete_segs
= obsolete_segs
;
1007 vcsa
->vcsa_spa
= spa
;
1008 vcsa
->vcsa_txg
= txg
;
1011 * See comment before spa_vdev_copy_one_child().
1013 spa_config_enter(spa
, SCL_STATE
, spa
, RW_READER
);
1014 zio_t
*nzio
= zio_null(spa
->spa_txg_zio
[txg
& TXG_MASK
], spa
, NULL
,
1015 spa_vdev_copy_segment_done
, vcsa
, 0);
1016 vdev_t
*dest_vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dst
));
1017 if (dest_vd
->vdev_ops
== &vdev_mirror_ops
) {
1018 for (int i
= 0; i
< dest_vd
->vdev_children
; i
++) {
1019 vdev_t
*child
= dest_vd
->vdev_child
[i
];
1020 spa_vdev_copy_one_child(vca
, nzio
, vd
, start
,
1021 child
, DVA_GET_OFFSET(&dst
), i
, size
);
1024 spa_vdev_copy_one_child(vca
, nzio
, vd
, start
,
1025 dest_vd
, DVA_GET_OFFSET(&dst
), -1, size
);
1029 list_insert_tail(&svr
->svr_new_segments
[txg
& TXG_MASK
], entry
);
1030 ASSERT3U(start
+ size
, <=, vd
->vdev_ms_count
<< vd
->vdev_ms_shift
);
1031 vdev_dirty(vd
, 0, NULL
, txg
);
1037 * Complete the removal of a toplevel vdev. This is called as a
1038 * synctask in the same txg that we will sync out the new config (to the
1039 * MOS object) which indicates that this vdev is indirect.
1042 vdev_remove_complete_sync(void *arg
, dmu_tx_t
*tx
)
1044 spa_vdev_removal_t
*svr
= arg
;
1045 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1046 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1048 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
1050 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1051 ASSERT0(svr
->svr_bytes_done
[i
]);
1054 ASSERT3U(spa
->spa_removing_phys
.sr_copied
, ==,
1055 spa
->spa_removing_phys
.sr_to_copy
);
1057 vdev_destroy_spacemaps(vd
, tx
);
1059 /* destroy leaf zaps, if any */
1060 ASSERT3P(svr
->svr_zaplist
, !=, NULL
);
1061 for (nvpair_t
*pair
= nvlist_next_nvpair(svr
->svr_zaplist
, NULL
);
1063 pair
= nvlist_next_nvpair(svr
->svr_zaplist
, pair
)) {
1064 vdev_destroy_unlink_zap(vd
, fnvpair_value_uint64(pair
), tx
);
1066 fnvlist_free(svr
->svr_zaplist
);
1068 spa_finish_removal(dmu_tx_pool(tx
)->dp_spa
, DSS_FINISHED
, tx
);
1069 /* vd->vdev_path is not available here */
1070 spa_history_log_internal(spa
, "vdev remove completed", tx
,
1071 "%s vdev %llu", spa_name(spa
), vd
->vdev_id
);
1075 vdev_remove_enlist_zaps(vdev_t
*vd
, nvlist_t
*zlist
)
1077 ASSERT3P(zlist
, !=, NULL
);
1078 ASSERT3P(vd
->vdev_ops
, !=, &vdev_raidz_ops
);
1080 if (vd
->vdev_leaf_zap
!= 0) {
1082 (void) snprintf(zkey
, sizeof (zkey
), "%s-%llu",
1083 VDEV_REMOVAL_ZAP_OBJS
, (u_longlong_t
)vd
->vdev_leaf_zap
);
1084 fnvlist_add_uint64(zlist
, zkey
, vd
->vdev_leaf_zap
);
1087 for (uint64_t id
= 0; id
< vd
->vdev_children
; id
++) {
1088 vdev_remove_enlist_zaps(vd
->vdev_child
[id
], zlist
);
1093 vdev_remove_replace_with_indirect(vdev_t
*vd
, uint64_t txg
)
1097 spa_t
*spa
= vd
->vdev_spa
;
1098 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1101 * First, build a list of leaf zaps to be destroyed.
1102 * This is passed to the sync context thread,
1103 * which does the actual unlinking.
1105 svr
->svr_zaplist
= fnvlist_alloc();
1106 vdev_remove_enlist_zaps(vd
, svr
->svr_zaplist
);
1108 ivd
= vdev_add_parent(vd
, &vdev_indirect_ops
);
1109 ivd
->vdev_removing
= 0;
1111 vd
->vdev_leaf_zap
= 0;
1113 vdev_remove_child(ivd
, vd
);
1114 vdev_compact_children(ivd
);
1116 ASSERT(!list_link_active(&vd
->vdev_state_dirty_node
));
1118 tx
= dmu_tx_create_assigned(spa
->spa_dsl_pool
, txg
);
1119 dsl_sync_task_nowait(spa
->spa_dsl_pool
, vdev_remove_complete_sync
, svr
,
1120 0, ZFS_SPACE_CHECK_NONE
, tx
);
1124 * Indicate that this thread has exited.
1125 * After this, we can not use svr.
1127 mutex_enter(&svr
->svr_lock
);
1128 svr
->svr_thread
= NULL
;
1129 cv_broadcast(&svr
->svr_cv
);
1130 mutex_exit(&svr
->svr_lock
);
1134 * Complete the removal of a toplevel vdev. This is called in open
1135 * context by the removal thread after we have copied all vdev's data.
1138 vdev_remove_complete(spa_t
*spa
)
1143 * Wait for any deferred frees to be synced before we call
1144 * vdev_metaslab_fini()
1146 txg_wait_synced(spa
->spa_dsl_pool
, 0);
1147 txg
= spa_vdev_enter(spa
);
1148 vdev_t
*vd
= vdev_lookup_top(spa
, spa
->spa_vdev_removal
->svr_vdev_id
);
1150 sysevent_t
*ev
= spa_event_create(spa
, vd
, NULL
,
1151 ESC_ZFS_VDEV_REMOVE_DEV
);
1153 zfs_dbgmsg("finishing device removal for vdev %llu in txg %llu",
1157 * Discard allocation state.
1159 if (vd
->vdev_mg
!= NULL
) {
1160 vdev_metaslab_fini(vd
);
1161 metaslab_group_destroy(vd
->vdev_mg
);
1164 ASSERT0(vd
->vdev_stat
.vs_space
);
1165 ASSERT0(vd
->vdev_stat
.vs_dspace
);
1167 vdev_remove_replace_with_indirect(vd
, txg
);
1170 * We now release the locks, allowing spa_sync to run and finish the
1171 * removal via vdev_remove_complete_sync in syncing context.
1173 * Note that we hold on to the vdev_t that has been replaced. Since
1174 * it isn't part of the vdev tree any longer, it can't be concurrently
1175 * manipulated, even while we don't have the config lock.
1177 (void) spa_vdev_exit(spa
, NULL
, txg
, 0);
1180 * Top ZAP should have been transferred to the indirect vdev in
1181 * vdev_remove_replace_with_indirect.
1183 ASSERT0(vd
->vdev_top_zap
);
1186 * Leaf ZAP should have been moved in vdev_remove_replace_with_indirect.
1188 ASSERT0(vd
->vdev_leaf_zap
);
1190 txg
= spa_vdev_enter(spa
);
1191 (void) vdev_label_init(vd
, 0, VDEV_LABEL_REMOVE
);
1193 * Request to update the config and the config cachefile.
1195 vdev_config_dirty(spa
->spa_root_vdev
);
1196 (void) spa_vdev_exit(spa
, vd
, txg
, 0);
1203 * Evacuates a segment of size at most max_alloc from the vdev
1204 * via repeated calls to spa_vdev_copy_segment. If an allocation
1205 * fails, the pool is probably too fragmented to handle such a
1206 * large size, so decrease max_alloc so that the caller will not try
1207 * this size again this txg.
1210 spa_vdev_copy_impl(vdev_t
*vd
, spa_vdev_removal_t
*svr
, vdev_copy_arg_t
*vca
,
1211 uint64_t *max_alloc
, dmu_tx_t
*tx
)
1213 uint64_t txg
= dmu_tx_get_txg(tx
);
1214 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1216 mutex_enter(&svr
->svr_lock
);
1219 * Determine how big of a chunk to copy. We can allocate up
1220 * to max_alloc bytes, and we can span up to vdev_removal_max_span
1221 * bytes of unallocated space at a time. "segs" will track the
1222 * allocated segments that we are copying. We may also be copying
1223 * free segments (of up to vdev_removal_max_span bytes).
1225 range_tree_t
*segs
= range_tree_create(NULL
, NULL
);
1227 range_seg_t
*rs
= range_tree_first(svr
->svr_allocd_segs
);
1232 uint64_t seg_length
;
1234 if (range_tree_is_empty(segs
)) {
1235 /* need to truncate the first seg based on max_alloc */
1237 MIN(rs
->rs_end
- rs
->rs_start
, *max_alloc
);
1239 if (rs
->rs_start
- range_tree_max(segs
) >
1240 vdev_removal_max_span
) {
1242 * Including this segment would cause us to
1243 * copy a larger unneeded chunk than is allowed.
1246 } else if (rs
->rs_end
- range_tree_min(segs
) >
1249 * This additional segment would extend past
1250 * max_alloc. Rather than splitting this
1251 * segment, leave it for the next mapping.
1255 seg_length
= rs
->rs_end
- rs
->rs_start
;
1259 range_tree_add(segs
, rs
->rs_start
, seg_length
);
1260 range_tree_remove(svr
->svr_allocd_segs
,
1261 rs
->rs_start
, seg_length
);
1264 if (range_tree_is_empty(segs
)) {
1265 mutex_exit(&svr
->svr_lock
);
1266 range_tree_destroy(segs
);
1270 if (svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] == 0) {
1271 dsl_sync_task_nowait(dmu_tx_pool(tx
), vdev_mapping_sync
,
1272 svr
, 0, ZFS_SPACE_CHECK_NONE
, tx
);
1275 svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] = range_tree_max(segs
);
1278 * Note: this is the amount of *allocated* space
1279 * that we are taking care of each txg.
1281 svr
->svr_bytes_done
[txg
& TXG_MASK
] += range_tree_space(segs
);
1283 mutex_exit(&svr
->svr_lock
);
1285 zio_alloc_list_t zal
;
1286 metaslab_trace_init(&zal
);
1287 uint64_t thismax
= SPA_MAXBLOCKSIZE
;
1288 while (!range_tree_is_empty(segs
)) {
1289 int error
= spa_vdev_copy_segment(vd
,
1290 segs
, thismax
, txg
, vca
, &zal
);
1292 if (error
== ENOSPC
) {
1294 * Cut our segment in half, and don't try this
1295 * segment size again this txg. Note that the
1296 * allocation size must be aligned to the highest
1297 * ashift in the pool, so that the allocation will
1298 * not be padded out to a multiple of the ashift,
1299 * which could cause us to think that this mapping
1300 * is larger than we intended.
1302 ASSERT3U(spa
->spa_max_ashift
, >=, SPA_MINBLOCKSHIFT
);
1303 ASSERT3U(spa
->spa_max_ashift
, ==, spa
->spa_min_ashift
);
1304 uint64_t attempted
=
1305 MIN(range_tree_span(segs
), thismax
);
1306 thismax
= P2ROUNDUP(attempted
/ 2,
1307 1 << spa
->spa_max_ashift
);
1309 * The minimum-size allocation can not fail.
1311 ASSERT3U(attempted
, >, 1 << spa
->spa_max_ashift
);
1312 *max_alloc
= attempted
- (1 << spa
->spa_max_ashift
);
1317 * We've performed an allocation, so reset the
1320 metaslab_trace_fini(&zal
);
1321 metaslab_trace_init(&zal
);
1324 metaslab_trace_fini(&zal
);
1325 range_tree_destroy(segs
);
1329 * The removal thread operates in open context. It iterates over all
1330 * allocated space in the vdev, by loading each metaslab's spacemap.
1331 * For each contiguous segment of allocated space (capping the segment
1332 * size at SPA_MAXBLOCKSIZE), we:
1333 * - Allocate space for it on another vdev.
1334 * - Create a new mapping from the old location to the new location
1335 * (as a record in svr_new_segments).
1336 * - Initiate a physical read zio to get the data off the removing disk.
1337 * - In the read zio's done callback, initiate a physical write zio to
1338 * write it to the new vdev.
1339 * Note that all of this will take effect when a particular TXG syncs.
1340 * The sync thread ensures that all the phys reads and writes for the syncing
1341 * TXG have completed (see spa_txg_zio) and writes the new mappings to disk
1342 * (see vdev_mapping_sync()).
1345 spa_vdev_remove_thread(void *arg
)
1348 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1349 vdev_copy_arg_t vca
;
1350 uint64_t max_alloc
= zfs_remove_max_segment
;
1351 uint64_t last_txg
= 0;
1353 spa_config_enter(spa
, SCL_CONFIG
, FTAG
, RW_READER
);
1354 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1355 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
1356 uint64_t start_offset
= vdev_indirect_mapping_max_offset(vim
);
1358 ASSERT3P(vd
->vdev_ops
, !=, &vdev_indirect_ops
);
1359 ASSERT(vdev_is_concrete(vd
));
1360 ASSERT(vd
->vdev_removing
);
1361 ASSERT(vd
->vdev_indirect_config
.vic_mapping_object
!= 0);
1362 ASSERT(vim
!= NULL
);
1364 mutex_init(&vca
.vca_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1365 cv_init(&vca
.vca_cv
, NULL
, CV_DEFAULT
, NULL
);
1366 vca
.vca_outstanding_bytes
= 0;
1368 mutex_enter(&svr
->svr_lock
);
1371 * Start from vim_max_offset so we pick up where we left off
1372 * if we are restarting the removal after opening the pool.
1375 for (msi
= start_offset
>> vd
->vdev_ms_shift
;
1376 msi
< vd
->vdev_ms_count
&& !svr
->svr_thread_exit
; msi
++) {
1377 metaslab_t
*msp
= vd
->vdev_ms
[msi
];
1378 ASSERT3U(msi
, <=, vd
->vdev_ms_count
);
1380 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1382 mutex_enter(&msp
->ms_sync_lock
);
1383 mutex_enter(&msp
->ms_lock
);
1386 * Assert nothing in flight -- ms_*tree is empty.
1388 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1389 ASSERT0(range_tree_space(msp
->ms_allocating
[i
]));
1393 * If the metaslab has ever been allocated from (ms_sm!=NULL),
1394 * read the allocated segments from the space map object
1395 * into svr_allocd_segs. Since we do this while holding
1396 * svr_lock and ms_sync_lock, concurrent frees (which
1397 * would have modified the space map) will wait for us
1398 * to finish loading the spacemap, and then take the
1399 * appropriate action (see free_from_removing_vdev()).
1401 if (msp
->ms_sm
!= NULL
) {
1402 space_map_t
*sm
= NULL
;
1405 * We have to open a new space map here, because
1406 * ms_sm's sm_length and sm_alloc may not reflect
1407 * what's in the object contents, if we are in between
1408 * metaslab_sync() and metaslab_sync_done().
1410 VERIFY0(space_map_open(&sm
,
1411 spa
->spa_dsl_pool
->dp_meta_objset
,
1412 msp
->ms_sm
->sm_object
, msp
->ms_sm
->sm_start
,
1413 msp
->ms_sm
->sm_size
, msp
->ms_sm
->sm_shift
));
1414 space_map_update(sm
);
1415 VERIFY0(space_map_load(sm
, svr
->svr_allocd_segs
,
1417 space_map_close(sm
);
1419 range_tree_walk(msp
->ms_freeing
,
1420 range_tree_remove
, svr
->svr_allocd_segs
);
1423 * When we are resuming from a paused removal (i.e.
1424 * when importing a pool with a removal in progress),
1425 * discard any state that we have already processed.
1427 range_tree_clear(svr
->svr_allocd_segs
, 0, start_offset
);
1429 mutex_exit(&msp
->ms_lock
);
1430 mutex_exit(&msp
->ms_sync_lock
);
1433 zfs_dbgmsg("copying %llu segments for metaslab %llu",
1434 avl_numnodes(&svr
->svr_allocd_segs
->rt_root
),
1437 while (!svr
->svr_thread_exit
&&
1438 !range_tree_is_empty(svr
->svr_allocd_segs
)) {
1440 mutex_exit(&svr
->svr_lock
);
1443 * We need to periodically drop the config lock so that
1444 * writers can get in. Additionally, we can't wait
1445 * for a txg to sync while holding a config lock
1446 * (since a waiting writer could cause a 3-way deadlock
1447 * with the sync thread, which also gets a config
1448 * lock for reader). So we can't hold the config lock
1449 * while calling dmu_tx_assign().
1451 spa_config_exit(spa
, SCL_CONFIG
, FTAG
);
1454 * This delay will pause the removal around the point
1455 * specified by zfs_remove_max_bytes_pause. We do this
1456 * solely from the test suite or during debugging.
1458 uint64_t bytes_copied
=
1459 spa
->spa_removing_phys
.sr_copied
;
1460 for (int i
= 0; i
< TXG_SIZE
; i
++)
1461 bytes_copied
+= svr
->svr_bytes_done
[i
];
1462 while (zfs_remove_max_bytes_pause
<= bytes_copied
&&
1463 !svr
->svr_thread_exit
)
1466 mutex_enter(&vca
.vca_lock
);
1467 while (vca
.vca_outstanding_bytes
>
1468 zfs_remove_max_copy_bytes
) {
1469 cv_wait(&vca
.vca_cv
, &vca
.vca_lock
);
1471 mutex_exit(&vca
.vca_lock
);
1474 dmu_tx_create_dd(spa_get_dsl(spa
)->dp_mos_dir
);
1475 dmu_tx_hold_space(tx
, SPA_MAXBLOCKSIZE
);
1476 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
1477 uint64_t txg
= dmu_tx_get_txg(tx
);
1480 * Reacquire the vdev_config lock. The vdev_t
1481 * that we're removing may have changed, e.g. due
1482 * to a vdev_attach or vdev_detach.
1484 spa_config_enter(spa
, SCL_CONFIG
, FTAG
, RW_READER
);
1485 vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1487 if (txg
!= last_txg
)
1488 max_alloc
= zfs_remove_max_segment
;
1491 spa_vdev_copy_impl(vd
, svr
, &vca
, &max_alloc
, tx
);
1494 mutex_enter(&svr
->svr_lock
);
1498 mutex_exit(&svr
->svr_lock
);
1500 spa_config_exit(spa
, SCL_CONFIG
, FTAG
);
1503 * Wait for all copies to finish before cleaning up the vca.
1505 txg_wait_synced(spa
->spa_dsl_pool
, 0);
1506 ASSERT0(vca
.vca_outstanding_bytes
);
1508 mutex_destroy(&vca
.vca_lock
);
1509 cv_destroy(&vca
.vca_cv
);
1511 if (svr
->svr_thread_exit
) {
1512 mutex_enter(&svr
->svr_lock
);
1513 range_tree_vacate(svr
->svr_allocd_segs
, NULL
, NULL
);
1514 svr
->svr_thread
= NULL
;
1515 cv_broadcast(&svr
->svr_cv
);
1516 mutex_exit(&svr
->svr_lock
);
1518 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1519 vdev_remove_complete(spa
);
1524 spa_vdev_remove_suspend(spa_t
*spa
)
1526 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1531 mutex_enter(&svr
->svr_lock
);
1532 svr
->svr_thread_exit
= B_TRUE
;
1533 while (svr
->svr_thread
!= NULL
)
1534 cv_wait(&svr
->svr_cv
, &svr
->svr_lock
);
1535 svr
->svr_thread_exit
= B_FALSE
;
1536 mutex_exit(&svr
->svr_lock
);
1541 spa_vdev_remove_cancel_check(void *arg
, dmu_tx_t
*tx
)
1543 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1545 if (spa
->spa_vdev_removal
== NULL
)
1546 return (ENOTACTIVE
);
1551 * Cancel a removal by freeing all entries from the partial mapping
1552 * and marking the vdev as no longer being removing.
1556 spa_vdev_remove_cancel_sync(void *arg
, dmu_tx_t
*tx
)
1558 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1559 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1560 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1561 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
1562 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
1563 objset_t
*mos
= spa
->spa_meta_objset
;
1565 ASSERT3P(svr
->svr_thread
, ==, NULL
);
1567 spa_feature_decr(spa
, SPA_FEATURE_DEVICE_REMOVAL
, tx
);
1569 boolean_t are_precise
;
1570 VERIFY0(vdev_obsolete_counts_are_precise(vd
, &are_precise
));
1572 spa_feature_decr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
1573 VERIFY0(zap_remove(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
1574 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE
, tx
));
1577 uint64_t obsolete_sm_object
;
1578 VERIFY0(vdev_obsolete_sm_object(vd
, &obsolete_sm_object
));
1579 if (obsolete_sm_object
!= 0) {
1580 ASSERT(vd
->vdev_obsolete_sm
!= NULL
);
1581 ASSERT3U(obsolete_sm_object
, ==,
1582 space_map_object(vd
->vdev_obsolete_sm
));
1584 space_map_free(vd
->vdev_obsolete_sm
, tx
);
1585 VERIFY0(zap_remove(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
1586 VDEV_TOP_ZAP_INDIRECT_OBSOLETE_SM
, tx
));
1587 space_map_close(vd
->vdev_obsolete_sm
);
1588 vd
->vdev_obsolete_sm
= NULL
;
1589 spa_feature_decr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
1591 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1592 ASSERT(list_is_empty(&svr
->svr_new_segments
[i
]));
1593 ASSERT3U(svr
->svr_max_offset_to_sync
[i
], <=,
1594 vdev_indirect_mapping_max_offset(vim
));
1597 for (uint64_t msi
= 0; msi
< vd
->vdev_ms_count
; msi
++) {
1598 metaslab_t
*msp
= vd
->vdev_ms
[msi
];
1600 if (msp
->ms_start
>= vdev_indirect_mapping_max_offset(vim
))
1603 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1605 mutex_enter(&msp
->ms_lock
);
1608 * Assert nothing in flight -- ms_*tree is empty.
1610 for (int i
= 0; i
< TXG_SIZE
; i
++)
1611 ASSERT0(range_tree_space(msp
->ms_allocating
[i
]));
1612 for (int i
= 0; i
< TXG_DEFER_SIZE
; i
++)
1613 ASSERT0(range_tree_space(msp
->ms_defer
[i
]));
1614 ASSERT0(range_tree_space(msp
->ms_freed
));
1616 if (msp
->ms_sm
!= NULL
) {
1618 * Assert that the in-core spacemap has the same
1619 * length as the on-disk one, so we can use the
1620 * existing in-core spacemap to load it from disk.
1622 ASSERT3U(msp
->ms_sm
->sm_alloc
, ==,
1623 msp
->ms_sm
->sm_phys
->smp_alloc
);
1624 ASSERT3U(msp
->ms_sm
->sm_length
, ==,
1625 msp
->ms_sm
->sm_phys
->smp_objsize
);
1627 mutex_enter(&svr
->svr_lock
);
1628 VERIFY0(space_map_load(msp
->ms_sm
,
1629 svr
->svr_allocd_segs
, SM_ALLOC
));
1630 range_tree_walk(msp
->ms_freeing
,
1631 range_tree_remove
, svr
->svr_allocd_segs
);
1634 * Clear everything past what has been synced,
1635 * because we have not allocated mappings for it yet.
1637 uint64_t syncd
= vdev_indirect_mapping_max_offset(vim
);
1638 uint64_t sm_end
= msp
->ms_sm
->sm_start
+
1639 msp
->ms_sm
->sm_size
;
1641 range_tree_clear(svr
->svr_allocd_segs
,
1642 syncd
, sm_end
- syncd
);
1644 mutex_exit(&svr
->svr_lock
);
1646 mutex_exit(&msp
->ms_lock
);
1648 mutex_enter(&svr
->svr_lock
);
1649 range_tree_vacate(svr
->svr_allocd_segs
,
1650 free_mapped_segment_cb
, vd
);
1651 mutex_exit(&svr
->svr_lock
);
1655 * Note: this must happen after we invoke free_mapped_segment_cb,
1656 * because it adds to the obsolete_segments.
1658 range_tree_vacate(vd
->vdev_obsolete_segments
, NULL
, NULL
);
1660 ASSERT3U(vic
->vic_mapping_object
, ==,
1661 vdev_indirect_mapping_object(vd
->vdev_indirect_mapping
));
1662 vdev_indirect_mapping_close(vd
->vdev_indirect_mapping
);
1663 vd
->vdev_indirect_mapping
= NULL
;
1664 vdev_indirect_mapping_free(mos
, vic
->vic_mapping_object
, tx
);
1665 vic
->vic_mapping_object
= 0;
1667 ASSERT3U(vic
->vic_births_object
, ==,
1668 vdev_indirect_births_object(vd
->vdev_indirect_births
));
1669 vdev_indirect_births_close(vd
->vdev_indirect_births
);
1670 vd
->vdev_indirect_births
= NULL
;
1671 vdev_indirect_births_free(mos
, vic
->vic_births_object
, tx
);
1672 vic
->vic_births_object
= 0;
1675 * We may have processed some frees from the removing vdev in this
1676 * txg, thus increasing svr_bytes_done; discard that here to
1677 * satisfy the assertions in spa_vdev_removal_destroy().
1678 * Note that future txg's can not have any bytes_done, because
1679 * future TXG's are only modified from open context, and we have
1680 * already shut down the copying thread.
1682 svr
->svr_bytes_done
[dmu_tx_get_txg(tx
) & TXG_MASK
] = 0;
1683 spa_finish_removal(spa
, DSS_CANCELED
, tx
);
1685 vd
->vdev_removing
= B_FALSE
;
1686 vdev_config_dirty(vd
);
1688 zfs_dbgmsg("canceled device removal for vdev %llu in %llu",
1689 vd
->vdev_id
, dmu_tx_get_txg(tx
));
1690 spa_history_log_internal(spa
, "vdev remove canceled", tx
,
1691 "%s vdev %llu %s", spa_name(spa
),
1692 vd
->vdev_id
, (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
1696 spa_vdev_remove_cancel(spa_t
*spa
)
1698 spa_vdev_remove_suspend(spa
);
1700 if (spa
->spa_vdev_removal
== NULL
)
1701 return (ENOTACTIVE
);
1703 uint64_t vdid
= spa
->spa_vdev_removal
->svr_vdev_id
;
1705 int error
= dsl_sync_task(spa
->spa_name
, spa_vdev_remove_cancel_check
,
1706 spa_vdev_remove_cancel_sync
, NULL
, 0,
1707 ZFS_SPACE_CHECK_EXTRA_RESERVED
);
1710 spa_config_enter(spa
, SCL_ALLOC
| SCL_VDEV
, FTAG
, RW_WRITER
);
1711 vdev_t
*vd
= vdev_lookup_top(spa
, vdid
);
1712 metaslab_group_activate(vd
->vdev_mg
);
1713 spa_config_exit(spa
, SCL_ALLOC
| SCL_VDEV
, FTAG
);
1720 * Called every sync pass of every txg if there's a svr.
1723 svr_sync(spa_t
*spa
, dmu_tx_t
*tx
)
1725 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1726 int txgoff
= dmu_tx_get_txg(tx
) & TXG_MASK
;
1729 * This check is necessary so that we do not dirty the
1730 * DIRECTORY_OBJECT via spa_sync_removing_state() when there
1731 * is nothing to do. Dirtying it every time would prevent us
1732 * from syncing-to-convergence.
1734 if (svr
->svr_bytes_done
[txgoff
] == 0)
1738 * Update progress accounting.
1740 spa
->spa_removing_phys
.sr_copied
+= svr
->svr_bytes_done
[txgoff
];
1741 svr
->svr_bytes_done
[txgoff
] = 0;
1743 spa_sync_removing_state(spa
, tx
);
1747 vdev_remove_make_hole_and_free(vdev_t
*vd
)
1749 uint64_t id
= vd
->vdev_id
;
1750 spa_t
*spa
= vd
->vdev_spa
;
1751 vdev_t
*rvd
= spa
->spa_root_vdev
;
1752 boolean_t last_vdev
= (id
== (rvd
->vdev_children
- 1));
1754 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
1755 ASSERT(spa_config_held(spa
, SCL_ALL
, RW_WRITER
) == SCL_ALL
);
1760 vdev_compact_children(rvd
);
1762 vd
= vdev_alloc_common(spa
, id
, 0, &vdev_hole_ops
);
1763 vdev_add_child(rvd
, vd
);
1765 vdev_config_dirty(rvd
);
1768 * Reassess the health of our root vdev.
1774 * Remove a log device. The config lock is held for the specified TXG.
1777 spa_vdev_remove_log(vdev_t
*vd
, uint64_t *txg
)
1779 metaslab_group_t
*mg
= vd
->vdev_mg
;
1780 spa_t
*spa
= vd
->vdev_spa
;
1783 ASSERT(vd
->vdev_islog
);
1784 ASSERT(vd
== vd
->vdev_top
);
1787 * Stop allocating from this vdev.
1789 metaslab_group_passivate(mg
);
1792 * Wait for the youngest allocations and frees to sync,
1793 * and then wait for the deferral of those frees to finish.
1795 spa_vdev_config_exit(spa
, NULL
,
1796 *txg
+ TXG_CONCURRENT_STATES
+ TXG_DEFER_SIZE
, 0, FTAG
);
1799 * Evacuate the device. We don't hold the config lock as writer
1800 * since we need to do I/O but we do keep the
1801 * spa_namespace_lock held. Once this completes the device
1802 * should no longer have any blocks allocated on it.
1804 if (vd
->vdev_islog
) {
1805 if (vd
->vdev_stat
.vs_alloc
!= 0)
1806 error
= spa_reset_logs(spa
);
1809 *txg
= spa_vdev_config_enter(spa
);
1812 metaslab_group_activate(mg
);
1815 ASSERT0(vd
->vdev_stat
.vs_alloc
);
1818 * The evacuation succeeded. Remove any remaining MOS metadata
1819 * associated with this vdev, and wait for these changes to sync.
1821 vd
->vdev_removing
= B_TRUE
;
1823 vdev_dirty_leaves(vd
, VDD_DTL
, *txg
);
1824 vdev_config_dirty(vd
);
1826 spa_history_log_internal(spa
, "vdev remove", NULL
,
1827 "%s vdev %llu (log) %s", spa_name(spa
), vd
->vdev_id
,
1828 (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
1830 spa_vdev_config_exit(spa
, NULL
, *txg
, 0, FTAG
);
1832 *txg
= spa_vdev_config_enter(spa
);
1834 sysevent_t
*ev
= spa_event_create(spa
, vd
, NULL
,
1835 ESC_ZFS_VDEV_REMOVE_DEV
);
1836 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
1837 ASSERT(spa_config_held(spa
, SCL_ALL
, RW_WRITER
) == SCL_ALL
);
1839 /* The top ZAP should have been destroyed by vdev_remove_empty. */
1840 ASSERT0(vd
->vdev_top_zap
);
1841 /* The leaf ZAP should have been destroyed by vdev_dtl_sync. */
1842 ASSERT0(vd
->vdev_leaf_zap
);
1844 (void) vdev_label_init(vd
, 0, VDEV_LABEL_REMOVE
);
1846 if (list_link_active(&vd
->vdev_state_dirty_node
))
1847 vdev_state_clean(vd
);
1848 if (list_link_active(&vd
->vdev_config_dirty_node
))
1849 vdev_config_clean(vd
);
1852 * Clean up the vdev namespace.
1854 vdev_remove_make_hole_and_free(vd
);
1863 spa_vdev_remove_top_check(vdev_t
*vd
)
1865 spa_t
*spa
= vd
->vdev_spa
;
1867 if (vd
!= vd
->vdev_top
)
1868 return (SET_ERROR(ENOTSUP
));
1870 if (!spa_feature_is_enabled(spa
, SPA_FEATURE_DEVICE_REMOVAL
))
1871 return (SET_ERROR(ENOTSUP
));
1873 /* available space in the pool's normal class */
1874 uint64_t available
= dsl_dir_space_available(
1875 spa
->spa_dsl_pool
->dp_root_dir
, NULL
, 0, B_TRUE
);
1877 metaslab_class_t
*mc
= vd
->vdev_mg
->mg_class
;
1880 * When removing a vdev from an allocation class that has
1881 * remaining vdevs, include available space from the class.
1883 if (mc
!= spa_normal_class(spa
) && mc
->mc_groups
> 1) {
1884 uint64_t class_avail
= metaslab_class_get_space(mc
) -
1885 metaslab_class_get_alloc(mc
);
1887 /* add class space, adjusted for overhead */
1888 available
+= (class_avail
* 94) / 100;
1892 * There has to be enough free space to remove the
1893 * device and leave double the "slop" space (i.e. we
1894 * must leave at least 3% of the pool free, in addition to
1895 * the normal slop space).
1897 if (available
< vd
->vdev_stat
.vs_dspace
+ spa_get_slop_space(spa
)) {
1898 return (SET_ERROR(ENOSPC
));
1902 * There can not be a removal in progress.
1904 if (spa
->spa_removing_phys
.sr_state
== DSS_SCANNING
)
1905 return (SET_ERROR(EBUSY
));
1908 * The device must have all its data.
1910 if (!vdev_dtl_empty(vd
, DTL_MISSING
) ||
1911 !vdev_dtl_empty(vd
, DTL_OUTAGE
))
1912 return (SET_ERROR(EBUSY
));
1915 * The device must be healthy.
1917 if (!vdev_readable(vd
))
1918 return (SET_ERROR(EIO
));
1921 * All vdevs in normal class must have the same ashift.
1923 if (spa
->spa_max_ashift
!= spa
->spa_min_ashift
) {
1924 return (SET_ERROR(EINVAL
));
1928 * All vdevs in normal class must have the same ashift
1931 vdev_t
*rvd
= spa
->spa_root_vdev
;
1932 int num_indirect
= 0;
1933 for (uint64_t id
= 0; id
< rvd
->vdev_children
; id
++) {
1934 vdev_t
*cvd
= rvd
->vdev_child
[id
];
1935 if (cvd
->vdev_ashift
!= 0 && !cvd
->vdev_islog
)
1936 ASSERT3U(cvd
->vdev_ashift
, ==, spa
->spa_max_ashift
);
1937 if (cvd
->vdev_ops
== &vdev_indirect_ops
)
1939 if (!vdev_is_concrete(cvd
))
1941 if (cvd
->vdev_ops
== &vdev_raidz_ops
)
1942 return (SET_ERROR(EINVAL
));
1944 * Need the mirror to be mirror of leaf vdevs only
1946 if (cvd
->vdev_ops
== &vdev_mirror_ops
) {
1947 for (uint64_t cid
= 0;
1948 cid
< cvd
->vdev_children
; cid
++) {
1949 if (!cvd
->vdev_child
[cid
]->vdev_ops
->
1951 return (SET_ERROR(EINVAL
));
1960 * Initiate removal of a top-level vdev, reducing the total space in the pool.
1961 * The config lock is held for the specified TXG. Once initiated,
1962 * evacuation of all allocated space (copying it to other vdevs) happens
1963 * in the background (see spa_vdev_remove_thread()), and can be canceled
1964 * (see spa_vdev_remove_cancel()). If successful, the vdev will
1965 * be transformed to an indirect vdev (see spa_vdev_remove_complete()).
1968 spa_vdev_remove_top(vdev_t
*vd
, uint64_t *txg
)
1970 spa_t
*spa
= vd
->vdev_spa
;
1974 * Check for errors up-front, so that we don't waste time
1975 * passivating the metaslab group and clearing the ZIL if there
1978 error
= spa_vdev_remove_top_check(vd
);
1983 * Stop allocating from this vdev. Note that we must check
1984 * that this is not the only device in the pool before
1985 * passivating, otherwise we will not be able to make
1986 * progress because we can't allocate from any vdevs.
1987 * The above check for sufficient free space serves this
1990 metaslab_group_t
*mg
= vd
->vdev_mg
;
1991 metaslab_group_passivate(mg
);
1994 * Wait for the youngest allocations and frees to sync,
1995 * and then wait for the deferral of those frees to finish.
1997 spa_vdev_config_exit(spa
, NULL
,
1998 *txg
+ TXG_CONCURRENT_STATES
+ TXG_DEFER_SIZE
, 0, FTAG
);
2001 * We must ensure that no "stubby" log blocks are allocated
2002 * on the device to be removed. These blocks could be
2003 * written at any time, including while we are in the middle
2006 error
= spa_reset_logs(spa
);
2008 *txg
= spa_vdev_config_enter(spa
);
2011 * Things might have changed while the config lock was dropped
2012 * (e.g. space usage). Check for errors again.
2015 error
= spa_vdev_remove_top_check(vd
);
2018 metaslab_group_activate(mg
);
2022 vd
->vdev_removing
= B_TRUE
;
2024 vdev_dirty_leaves(vd
, VDD_DTL
, *txg
);
2025 vdev_config_dirty(vd
);
2026 dmu_tx_t
*tx
= dmu_tx_create_assigned(spa
->spa_dsl_pool
, *txg
);
2027 dsl_sync_task_nowait(spa
->spa_dsl_pool
,
2028 vdev_remove_initiate_sync
,
2029 (void *)(uintptr_t)vd
->vdev_id
, 0, ZFS_SPACE_CHECK_NONE
, tx
);
2036 * Remove a device from the pool.
2038 * Removing a device from the vdev namespace requires several steps
2039 * and can take a significant amount of time. As a result we use
2040 * the spa_vdev_config_[enter/exit] functions which allow us to
2041 * grab and release the spa_config_lock while still holding the namespace
2042 * lock. During each step the configuration is synced out.
2045 spa_vdev_remove(spa_t
*spa
, uint64_t guid
, boolean_t unspare
)
2048 nvlist_t
**spares
, **l2cache
, *nv
;
2050 uint_t nspares
, nl2cache
;
2052 boolean_t locked
= MUTEX_HELD(&spa_namespace_lock
);
2053 sysevent_t
*ev
= NULL
;
2055 ASSERT(spa_writeable(spa
));
2058 txg
= spa_vdev_enter(spa
);
2060 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
2061 if (spa_feature_is_active(spa
, SPA_FEATURE_POOL_CHECKPOINT
)) {
2062 error
= (spa_has_checkpoint(spa
)) ?
2063 ZFS_ERR_CHECKPOINT_EXISTS
: ZFS_ERR_DISCARDING_CHECKPOINT
;
2066 return (spa_vdev_exit(spa
, NULL
, txg
, error
));
2071 vd
= spa_lookup_by_guid(spa
, guid
, B_FALSE
);
2073 if (spa
->spa_spares
.sav_vdevs
!= NULL
&&
2074 nvlist_lookup_nvlist_array(spa
->spa_spares
.sav_config
,
2075 ZPOOL_CONFIG_SPARES
, &spares
, &nspares
) == 0 &&
2076 (nv
= spa_nvlist_lookup_by_guid(spares
, nspares
, guid
)) != NULL
) {
2078 * Only remove the hot spare if it's not currently in use
2081 if (vd
== NULL
|| unspare
) {
2083 vd
= spa_lookup_by_guid(spa
, guid
, B_TRUE
);
2084 ev
= spa_event_create(spa
, vd
, NULL
,
2085 ESC_ZFS_VDEV_REMOVE_AUX
);
2087 char *nvstr
= fnvlist_lookup_string(nv
,
2089 spa_history_log_internal(spa
, "vdev remove", NULL
,
2090 "%s vdev (%s) %s", spa_name(spa
),
2091 VDEV_TYPE_SPARE
, nvstr
);
2092 spa_vdev_remove_aux(spa
->spa_spares
.sav_config
,
2093 ZPOOL_CONFIG_SPARES
, spares
, nspares
, nv
);
2094 spa_load_spares(spa
);
2095 spa
->spa_spares
.sav_sync
= B_TRUE
;
2097 error
= SET_ERROR(EBUSY
);
2099 } else if (spa
->spa_l2cache
.sav_vdevs
!= NULL
&&
2100 nvlist_lookup_nvlist_array(spa
->spa_l2cache
.sav_config
,
2101 ZPOOL_CONFIG_L2CACHE
, &l2cache
, &nl2cache
) == 0 &&
2102 (nv
= spa_nvlist_lookup_by_guid(l2cache
, nl2cache
, guid
)) != NULL
) {
2103 char *nvstr
= fnvlist_lookup_string(nv
, ZPOOL_CONFIG_PATH
);
2104 spa_history_log_internal(spa
, "vdev remove", NULL
,
2105 "%s vdev (%s) %s", spa_name(spa
), VDEV_TYPE_L2CACHE
, nvstr
);
2107 * Cache devices can always be removed.
2109 vd
= spa_lookup_by_guid(spa
, guid
, B_TRUE
);
2110 ev
= spa_event_create(spa
, vd
, NULL
, ESC_ZFS_VDEV_REMOVE_AUX
);
2111 spa_vdev_remove_aux(spa
->spa_l2cache
.sav_config
,
2112 ZPOOL_CONFIG_L2CACHE
, l2cache
, nl2cache
, nv
);
2113 spa_load_l2cache(spa
);
2114 spa
->spa_l2cache
.sav_sync
= B_TRUE
;
2115 } else if (vd
!= NULL
&& vd
->vdev_islog
) {
2117 error
= spa_vdev_remove_log(vd
, &txg
);
2118 } else if (vd
!= NULL
) {
2120 error
= spa_vdev_remove_top(vd
, &txg
);
2123 * There is no vdev of any kind with the specified guid.
2125 error
= SET_ERROR(ENOENT
);
2129 error
= spa_vdev_exit(spa
, NULL
, txg
, error
);
2138 spa_removal_get_stats(spa_t
*spa
, pool_removal_stat_t
*prs
)
2140 prs
->prs_state
= spa
->spa_removing_phys
.sr_state
;
2142 if (prs
->prs_state
== DSS_NONE
)
2143 return (SET_ERROR(ENOENT
));
2145 prs
->prs_removing_vdev
= spa
->spa_removing_phys
.sr_removing_vdev
;
2146 prs
->prs_start_time
= spa
->spa_removing_phys
.sr_start_time
;
2147 prs
->prs_end_time
= spa
->spa_removing_phys
.sr_end_time
;
2148 prs
->prs_to_copy
= spa
->spa_removing_phys
.sr_to_copy
;
2149 prs
->prs_copied
= spa
->spa_removing_phys
.sr_copied
;
2151 if (spa
->spa_vdev_removal
!= NULL
) {
2152 for (int i
= 0; i
< TXG_SIZE
; i
++) {
2154 spa
->spa_vdev_removal
->svr_bytes_done
[i
];
2158 prs
->prs_mapping_memory
= 0;
2159 uint64_t indirect_vdev_id
=
2160 spa
->spa_removing_phys
.sr_prev_indirect_vdev
;
2161 while (indirect_vdev_id
!= -1) {
2162 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[indirect_vdev_id
];
2163 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
2164 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
2166 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
2167 prs
->prs_mapping_memory
+= vdev_indirect_mapping_size(vim
);
2168 indirect_vdev_id
= vic
->vic_prev_indirect_vdev
;
2174 #if defined(_KERNEL)
2175 module_param(zfs_remove_max_segment
, int, 0644);
2176 MODULE_PARM_DESC(zfs_remove_max_segment
,
2177 "Largest contiguous segment to allocate when removing device");
2179 module_param(vdev_removal_max_span
, int, 0644);
2180 MODULE_PARM_DESC(vdev_removal_max_span
,
2181 "Largest span of free chunks a remap segment can span");
2184 module_param(zfs_remove_max_bytes_pause
, ulong
, 0644);
2185 MODULE_PARM_DESC(zfs_remove_max_bytes_pause
,
2186 "Pause device removal after this many bytes are copied "
2187 "(debug use only - causes removal to hang)");
2190 EXPORT_SYMBOL(free_from_removing_vdev
);
2191 EXPORT_SYMBOL(spa_removal_get_stats
);
2192 EXPORT_SYMBOL(spa_remove_init
);
2193 EXPORT_SYMBOL(spa_restart_removal
);
2194 EXPORT_SYMBOL(spa_vdev_removal_destroy
);
2195 EXPORT_SYMBOL(spa_vdev_remove
);
2196 EXPORT_SYMBOL(spa_vdev_remove_cancel
);
2197 EXPORT_SYMBOL(spa_vdev_remove_suspend
);
2198 EXPORT_SYMBOL(svr_sync
);