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/vdev_initialize.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. (Note that members of a mirror can also 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 * Note: we will issue I/Os of up to this size. The mpt driver does not
101 * respond well to I/Os larger than 1MB, so we set this to 1MB. (When
102 * mpt processes an I/O larger than 1MB, it needs to do an allocation of
103 * 2 physically contiguous pages; if this allocation fails, mpt will drop
104 * the I/O and hang the device.)
106 int zfs_remove_max_segment
= 1024 * 1024;
109 * Allow a remap segment to span free chunks of at most this size. The main
110 * impact of a larger span is that we will read and write larger, more
111 * contiguous chunks, with more "unnecessary" data -- trading off bandwidth
112 * for iops. The value here was chosen to align with
113 * zfs_vdev_read_gap_limit, which is a similar concept when doing regular
114 * reads (but there's no reason it has to be the same).
116 * Additionally, a higher span will have the following relatively minor
118 * - the mapping will be smaller, since one entry can cover more allocated
120 * - more of the fragmentation in the removing device will be preserved
121 * - we'll do larger allocations, which may fail and fall back on smaller
124 int vdev_removal_max_span
= 32 * 1024;
127 * This is used by the test suite so that it can ensure that certain
128 * actions happen while in the middle of a removal.
130 uint64_t zfs_remove_max_bytes_pause
= UINT64_MAX
;
132 #define VDEV_REMOVAL_ZAP_OBJS "lzap"
134 static void spa_vdev_remove_thread(void *arg
);
137 spa_sync_removing_state(spa_t
*spa
, dmu_tx_t
*tx
)
139 VERIFY0(zap_update(spa
->spa_dsl_pool
->dp_meta_objset
,
140 DMU_POOL_DIRECTORY_OBJECT
,
141 DMU_POOL_REMOVING
, sizeof (uint64_t),
142 sizeof (spa
->spa_removing_phys
) / sizeof (uint64_t),
143 &spa
->spa_removing_phys
, tx
));
147 spa_nvlist_lookup_by_guid(nvlist_t
**nvpp
, int count
, uint64_t target_guid
)
149 for (int i
= 0; i
< count
; i
++) {
151 fnvlist_lookup_uint64(nvpp
[i
], ZPOOL_CONFIG_GUID
);
153 if (guid
== target_guid
)
161 spa_vdev_remove_aux(nvlist_t
*config
, char *name
, nvlist_t
**dev
, int count
,
162 nvlist_t
*dev_to_remove
)
164 nvlist_t
**newdev
= NULL
;
167 newdev
= kmem_alloc((count
- 1) * sizeof (void *), KM_SLEEP
);
169 for (int i
= 0, j
= 0; i
< count
; i
++) {
170 if (dev
[i
] == dev_to_remove
)
172 VERIFY(nvlist_dup(dev
[i
], &newdev
[j
++], KM_SLEEP
) == 0);
175 VERIFY(nvlist_remove(config
, name
, DATA_TYPE_NVLIST_ARRAY
) == 0);
176 VERIFY(nvlist_add_nvlist_array(config
, name
, newdev
, count
- 1) == 0);
178 for (int i
= 0; i
< count
- 1; i
++)
179 nvlist_free(newdev
[i
]);
182 kmem_free(newdev
, (count
- 1) * sizeof (void *));
185 static spa_vdev_removal_t
*
186 spa_vdev_removal_create(vdev_t
*vd
)
188 spa_vdev_removal_t
*svr
= kmem_zalloc(sizeof (*svr
), KM_SLEEP
);
189 mutex_init(&svr
->svr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
190 cv_init(&svr
->svr_cv
, NULL
, CV_DEFAULT
, NULL
);
191 svr
->svr_allocd_segs
= range_tree_create(NULL
, NULL
);
192 svr
->svr_vdev_id
= vd
->vdev_id
;
194 for (int i
= 0; i
< TXG_SIZE
; i
++) {
195 svr
->svr_frees
[i
] = range_tree_create(NULL
, NULL
);
196 list_create(&svr
->svr_new_segments
[i
],
197 sizeof (vdev_indirect_mapping_entry_t
),
198 offsetof(vdev_indirect_mapping_entry_t
, vime_node
));
205 spa_vdev_removal_destroy(spa_vdev_removal_t
*svr
)
207 for (int i
= 0; i
< TXG_SIZE
; i
++) {
208 ASSERT0(svr
->svr_bytes_done
[i
]);
209 ASSERT0(svr
->svr_max_offset_to_sync
[i
]);
210 range_tree_destroy(svr
->svr_frees
[i
]);
211 list_destroy(&svr
->svr_new_segments
[i
]);
214 range_tree_destroy(svr
->svr_allocd_segs
);
215 mutex_destroy(&svr
->svr_lock
);
216 cv_destroy(&svr
->svr_cv
);
217 kmem_free(svr
, sizeof (*svr
));
221 * This is called as a synctask in the txg in which we will mark this vdev
222 * as removing (in the config stored in the MOS).
224 * It begins the evacuation of a toplevel vdev by:
225 * - initializing the spa_removing_phys which tracks this removal
226 * - computing the amount of space to remove for accounting purposes
227 * - dirtying all dbufs in the spa_config_object
228 * - creating the spa_vdev_removal
229 * - starting the spa_vdev_remove_thread
232 vdev_remove_initiate_sync(void *arg
, dmu_tx_t
*tx
)
234 int vdev_id
= (uintptr_t)arg
;
235 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
236 vdev_t
*vd
= vdev_lookup_top(spa
, vdev_id
);
237 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
238 objset_t
*mos
= spa
->spa_dsl_pool
->dp_meta_objset
;
239 spa_vdev_removal_t
*svr
= NULL
;
240 uint64_t txg
= dmu_tx_get_txg(tx
);
242 ASSERT3P(vd
->vdev_ops
, !=, &vdev_raidz_ops
);
243 svr
= spa_vdev_removal_create(vd
);
245 ASSERT(vd
->vdev_removing
);
246 ASSERT3P(vd
->vdev_indirect_mapping
, ==, NULL
);
248 spa_feature_incr(spa
, SPA_FEATURE_DEVICE_REMOVAL
, tx
);
249 if (spa_feature_is_enabled(spa
, SPA_FEATURE_OBSOLETE_COUNTS
)) {
251 * By activating the OBSOLETE_COUNTS feature, we prevent
252 * the pool from being downgraded and ensure that the
253 * refcounts are precise.
255 spa_feature_incr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
257 VERIFY0(zap_add(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
258 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE
, sizeof (one
), 1,
260 ASSERT3U(vdev_obsolete_counts_are_precise(vd
), !=, 0);
263 vic
->vic_mapping_object
= vdev_indirect_mapping_alloc(mos
, tx
);
264 vd
->vdev_indirect_mapping
=
265 vdev_indirect_mapping_open(mos
, vic
->vic_mapping_object
);
266 vic
->vic_births_object
= vdev_indirect_births_alloc(mos
, tx
);
267 vd
->vdev_indirect_births
=
268 vdev_indirect_births_open(mos
, vic
->vic_births_object
);
269 spa
->spa_removing_phys
.sr_removing_vdev
= vd
->vdev_id
;
270 spa
->spa_removing_phys
.sr_start_time
= gethrestime_sec();
271 spa
->spa_removing_phys
.sr_end_time
= 0;
272 spa
->spa_removing_phys
.sr_state
= DSS_SCANNING
;
273 spa
->spa_removing_phys
.sr_to_copy
= 0;
274 spa
->spa_removing_phys
.sr_copied
= 0;
277 * Note: We can't use vdev_stat's vs_alloc for sr_to_copy, because
278 * there may be space in the defer tree, which is free, but still
279 * counted in vs_alloc.
281 for (uint64_t i
= 0; i
< vd
->vdev_ms_count
; i
++) {
282 metaslab_t
*ms
= vd
->vdev_ms
[i
];
283 if (ms
->ms_sm
== NULL
)
287 * Sync tasks happen before metaslab_sync(), therefore
288 * smp_alloc and sm_alloc must be the same.
290 ASSERT3U(space_map_allocated(ms
->ms_sm
), ==,
291 ms
->ms_sm
->sm_phys
->smp_alloc
);
293 spa
->spa_removing_phys
.sr_to_copy
+=
294 space_map_allocated(ms
->ms_sm
);
297 * Space which we are freeing this txg does not need to
300 spa
->spa_removing_phys
.sr_to_copy
-=
301 range_tree_space(ms
->ms_freeing
);
303 ASSERT0(range_tree_space(ms
->ms_freed
));
304 for (int t
= 0; t
< TXG_SIZE
; t
++)
305 ASSERT0(range_tree_space(ms
->ms_allocating
[t
]));
309 * Sync tasks are called before metaslab_sync(), so there should
310 * be no already-synced metaslabs in the TXG_CLEAN list.
312 ASSERT3P(txg_list_head(&vd
->vdev_ms_list
, TXG_CLEAN(txg
)), ==, NULL
);
314 spa_sync_removing_state(spa
, tx
);
317 * All blocks that we need to read the most recent mapping must be
318 * stored on concrete vdevs. Therefore, we must dirty anything that
319 * is read before spa_remove_init(). Specifically, the
320 * spa_config_object. (Note that although we already modified the
321 * spa_config_object in spa_sync_removing_state, that may not have
322 * modified all blocks of the object.)
324 dmu_object_info_t doi
;
325 VERIFY0(dmu_object_info(mos
, DMU_POOL_DIRECTORY_OBJECT
, &doi
));
326 for (uint64_t offset
= 0; offset
< doi
.doi_max_offset
; ) {
328 VERIFY0(dmu_buf_hold(mos
, DMU_POOL_DIRECTORY_OBJECT
,
329 offset
, FTAG
, &dbuf
, 0));
330 dmu_buf_will_dirty(dbuf
, tx
);
331 offset
+= dbuf
->db_size
;
332 dmu_buf_rele(dbuf
, FTAG
);
336 * Now that we've allocated the im_object, dirty the vdev to ensure
337 * that the object gets written to the config on disk.
339 vdev_config_dirty(vd
);
341 zfs_dbgmsg("starting removal thread for vdev %llu (%p) in txg %llu "
342 "im_obj=%llu", vd
->vdev_id
, vd
, dmu_tx_get_txg(tx
),
343 vic
->vic_mapping_object
);
345 spa_history_log_internal(spa
, "vdev remove started", tx
,
346 "%s vdev %llu %s", spa_name(spa
), vd
->vdev_id
,
347 (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
349 * Setting spa_vdev_removal causes subsequent frees to call
350 * free_from_removing_vdev(). Note that we don't need any locking
351 * because we are the sync thread, and metaslab_free_impl() is only
352 * called from syncing context (potentially from a zio taskq thread,
353 * but in any case only when there are outstanding free i/os, which
356 ASSERT3P(spa
->spa_vdev_removal
, ==, NULL
);
357 spa
->spa_vdev_removal
= svr
;
358 svr
->svr_thread
= thread_create(NULL
, 0,
359 spa_vdev_remove_thread
, spa
, 0, &p0
, TS_RUN
, minclsyspri
);
363 * When we are opening a pool, we must read the mapping for each
364 * indirect vdev in order from most recently removed to least
365 * recently removed. We do this because the blocks for the mapping
366 * of older indirect vdevs may be stored on more recently removed vdevs.
367 * In order to read each indirect mapping object, we must have
368 * initialized all more recently removed vdevs.
371 spa_remove_init(spa_t
*spa
)
375 error
= zap_lookup(spa
->spa_dsl_pool
->dp_meta_objset
,
376 DMU_POOL_DIRECTORY_OBJECT
,
377 DMU_POOL_REMOVING
, sizeof (uint64_t),
378 sizeof (spa
->spa_removing_phys
) / sizeof (uint64_t),
379 &spa
->spa_removing_phys
);
381 if (error
== ENOENT
) {
382 spa
->spa_removing_phys
.sr_state
= DSS_NONE
;
383 spa
->spa_removing_phys
.sr_removing_vdev
= -1;
384 spa
->spa_removing_phys
.sr_prev_indirect_vdev
= -1;
385 spa
->spa_indirect_vdevs_loaded
= B_TRUE
;
387 } else if (error
!= 0) {
391 if (spa
->spa_removing_phys
.sr_state
== DSS_SCANNING
) {
393 * We are currently removing a vdev. Create and
394 * initialize a spa_vdev_removal_t from the bonus
395 * buffer of the removing vdevs vdev_im_object, and
396 * initialize its partial mapping.
398 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
399 vdev_t
*vd
= vdev_lookup_top(spa
,
400 spa
->spa_removing_phys
.sr_removing_vdev
);
403 spa_config_exit(spa
, SCL_STATE
, FTAG
);
407 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
409 ASSERT(vdev_is_concrete(vd
));
410 spa_vdev_removal_t
*svr
= spa_vdev_removal_create(vd
);
411 ASSERT3U(svr
->svr_vdev_id
, ==, vd
->vdev_id
);
412 ASSERT(vd
->vdev_removing
);
414 vd
->vdev_indirect_mapping
= vdev_indirect_mapping_open(
415 spa
->spa_meta_objset
, vic
->vic_mapping_object
);
416 vd
->vdev_indirect_births
= vdev_indirect_births_open(
417 spa
->spa_meta_objset
, vic
->vic_births_object
);
418 spa_config_exit(spa
, SCL_STATE
, FTAG
);
420 spa
->spa_vdev_removal
= svr
;
423 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
424 uint64_t indirect_vdev_id
=
425 spa
->spa_removing_phys
.sr_prev_indirect_vdev
;
426 while (indirect_vdev_id
!= UINT64_MAX
) {
427 vdev_t
*vd
= vdev_lookup_top(spa
, indirect_vdev_id
);
428 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
430 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
431 vd
->vdev_indirect_mapping
= vdev_indirect_mapping_open(
432 spa
->spa_meta_objset
, vic
->vic_mapping_object
);
433 vd
->vdev_indirect_births
= vdev_indirect_births_open(
434 spa
->spa_meta_objset
, vic
->vic_births_object
);
436 indirect_vdev_id
= vic
->vic_prev_indirect_vdev
;
438 spa_config_exit(spa
, SCL_STATE
, FTAG
);
441 * Now that we've loaded all the indirect mappings, we can allow
442 * reads from other blocks (e.g. via predictive prefetch).
444 spa
->spa_indirect_vdevs_loaded
= B_TRUE
;
449 spa_restart_removal(spa_t
*spa
)
451 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
457 * In general when this function is called there is no
458 * removal thread running. The only scenario where this
459 * is not true is during spa_import() where this function
460 * is called twice [once from spa_import_impl() and
461 * spa_async_resume()]. Thus, in the scenario where we
462 * import a pool that has an ongoing removal we don't
463 * want to spawn a second thread.
465 if (svr
->svr_thread
!= NULL
)
468 if (!spa_writeable(spa
))
471 zfs_dbgmsg("restarting removal of %llu", svr
->svr_vdev_id
);
472 svr
->svr_thread
= thread_create(NULL
, 0, spa_vdev_remove_thread
, spa
,
473 0, &p0
, TS_RUN
, minclsyspri
);
477 * Process freeing from a device which is in the middle of being removed.
478 * We must handle this carefully so that we attempt to copy freed data,
479 * and we correctly free already-copied data.
482 free_from_removing_vdev(vdev_t
*vd
, uint64_t offset
, uint64_t size
)
484 spa_t
*spa
= vd
->vdev_spa
;
485 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
486 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
487 uint64_t txg
= spa_syncing_txg(spa
);
488 uint64_t max_offset_yet
= 0;
490 ASSERT(vd
->vdev_indirect_config
.vic_mapping_object
!= 0);
491 ASSERT3U(vd
->vdev_indirect_config
.vic_mapping_object
, ==,
492 vdev_indirect_mapping_object(vim
));
493 ASSERT3U(vd
->vdev_id
, ==, svr
->svr_vdev_id
);
495 mutex_enter(&svr
->svr_lock
);
498 * Remove the segment from the removing vdev's spacemap. This
499 * ensures that we will not attempt to copy this space (if the
500 * removal thread has not yet visited it), and also ensures
501 * that we know what is actually allocated on the new vdevs
502 * (needed if we cancel the removal).
504 * Note: we must do the metaslab_free_concrete() with the svr_lock
505 * held, so that the remove_thread can not load this metaslab and then
506 * visit this offset between the time that we metaslab_free_concrete()
507 * and when we check to see if it has been visited.
509 * Note: The checkpoint flag is set to false as having/taking
510 * a checkpoint and removing a device can't happen at the same
513 ASSERT(!spa_has_checkpoint(spa
));
514 metaslab_free_concrete(vd
, offset
, size
, B_FALSE
);
516 uint64_t synced_size
= 0;
517 uint64_t synced_offset
= 0;
518 uint64_t max_offset_synced
= vdev_indirect_mapping_max_offset(vim
);
519 if (offset
< max_offset_synced
) {
521 * The mapping for this offset is already on disk.
522 * Free from the new location.
524 * Note that we use svr_max_synced_offset because it is
525 * updated atomically with respect to the in-core mapping.
526 * By contrast, vim_max_offset is not.
528 * This block may be split between a synced entry and an
529 * in-flight or unvisited entry. Only process the synced
530 * portion of it here.
532 synced_size
= MIN(size
, max_offset_synced
- offset
);
533 synced_offset
= offset
;
535 ASSERT3U(max_offset_yet
, <=, max_offset_synced
);
536 max_offset_yet
= max_offset_synced
;
538 DTRACE_PROBE3(remove__free__synced
,
541 uint64_t, synced_size
);
544 offset
+= synced_size
;
548 * Look at all in-flight txgs starting from the currently syncing one
549 * and see if a section of this free is being copied. By starting from
550 * this txg and iterating forward, we might find that this region
551 * was copied in two different txgs and handle it appropriately.
553 for (int i
= 0; i
< TXG_CONCURRENT_STATES
; i
++) {
554 int txgoff
= (txg
+ i
) & TXG_MASK
;
555 if (size
> 0 && offset
< svr
->svr_max_offset_to_sync
[txgoff
]) {
557 * The mapping for this offset is in flight, and
558 * will be synced in txg+i.
560 uint64_t inflight_size
= MIN(size
,
561 svr
->svr_max_offset_to_sync
[txgoff
] - offset
);
563 DTRACE_PROBE4(remove__free__inflight
,
566 uint64_t, inflight_size
,
570 * We copy data in order of increasing offset.
571 * Therefore the max_offset_to_sync[] must increase
572 * (or be zero, indicating that nothing is being
573 * copied in that txg).
575 if (svr
->svr_max_offset_to_sync
[txgoff
] != 0) {
576 ASSERT3U(svr
->svr_max_offset_to_sync
[txgoff
],
579 svr
->svr_max_offset_to_sync
[txgoff
];
583 * We've already committed to copying this segment:
584 * we have allocated space elsewhere in the pool for
585 * it and have an IO outstanding to copy the data. We
586 * cannot free the space before the copy has
587 * completed, or else the copy IO might overwrite any
588 * new data. To free that space, we record the
589 * segment in the appropriate svr_frees tree and free
590 * the mapped space later, in the txg where we have
591 * completed the copy and synced the mapping (see
592 * vdev_mapping_sync).
594 range_tree_add(svr
->svr_frees
[txgoff
],
595 offset
, inflight_size
);
596 size
-= inflight_size
;
597 offset
+= inflight_size
;
600 * This space is already accounted for as being
601 * done, because it is being copied in txg+i.
602 * However, if i!=0, then it is being copied in
603 * a future txg. If we crash after this txg
604 * syncs but before txg+i syncs, then the space
605 * will be free. Therefore we must account
606 * for the space being done in *this* txg
607 * (when it is freed) rather than the future txg
608 * (when it will be copied).
610 ASSERT3U(svr
->svr_bytes_done
[txgoff
], >=,
612 svr
->svr_bytes_done
[txgoff
] -= inflight_size
;
613 svr
->svr_bytes_done
[txg
& TXG_MASK
] += inflight_size
;
616 ASSERT0(svr
->svr_max_offset_to_sync
[TXG_CLEAN(txg
) & TXG_MASK
]);
620 * The copy thread has not yet visited this offset. Ensure
624 DTRACE_PROBE3(remove__free__unvisited
,
629 if (svr
->svr_allocd_segs
!= NULL
)
630 range_tree_clear(svr
->svr_allocd_segs
, offset
, size
);
633 * Since we now do not need to copy this data, for
634 * accounting purposes we have done our job and can count
637 svr
->svr_bytes_done
[txg
& TXG_MASK
] += size
;
639 mutex_exit(&svr
->svr_lock
);
642 * Now that we have dropped svr_lock, process the synced portion
645 if (synced_size
> 0) {
646 vdev_indirect_mark_obsolete(vd
, synced_offset
, synced_size
);
649 * Note: this can only be called from syncing context,
650 * and the vdev_indirect_mapping is only changed from the
651 * sync thread, so we don't need svr_lock while doing
652 * metaslab_free_impl_cb.
654 boolean_t checkpoint
= B_FALSE
;
655 vdev_indirect_ops
.vdev_op_remap(vd
, synced_offset
, synced_size
,
656 metaslab_free_impl_cb
, &checkpoint
);
661 * Stop an active removal and update the spa_removing phys.
664 spa_finish_removal(spa_t
*spa
, dsl_scan_state_t state
, dmu_tx_t
*tx
)
666 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
667 ASSERT3U(dmu_tx_get_txg(tx
), ==, spa_syncing_txg(spa
));
669 /* Ensure the removal thread has completed before we free the svr. */
670 spa_vdev_remove_suspend(spa
);
672 ASSERT(state
== DSS_FINISHED
|| state
== DSS_CANCELED
);
674 if (state
== DSS_FINISHED
) {
675 spa_removing_phys_t
*srp
= &spa
->spa_removing_phys
;
676 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
677 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
679 if (srp
->sr_prev_indirect_vdev
!= UINT64_MAX
) {
680 vdev_t
*pvd
= vdev_lookup_top(spa
,
681 srp
->sr_prev_indirect_vdev
);
682 ASSERT3P(pvd
->vdev_ops
, ==, &vdev_indirect_ops
);
685 vic
->vic_prev_indirect_vdev
= srp
->sr_prev_indirect_vdev
;
686 srp
->sr_prev_indirect_vdev
= vd
->vdev_id
;
688 spa
->spa_removing_phys
.sr_state
= state
;
689 spa
->spa_removing_phys
.sr_end_time
= gethrestime_sec();
691 spa
->spa_vdev_removal
= NULL
;
692 spa_vdev_removal_destroy(svr
);
694 spa_sync_removing_state(spa
, tx
);
696 vdev_config_dirty(spa
->spa_root_vdev
);
700 free_mapped_segment_cb(void *arg
, uint64_t offset
, uint64_t size
)
703 vdev_indirect_mark_obsolete(vd
, offset
, size
);
704 boolean_t checkpoint
= B_FALSE
;
705 vdev_indirect_ops
.vdev_op_remap(vd
, offset
, size
,
706 metaslab_free_impl_cb
, &checkpoint
);
710 * On behalf of the removal thread, syncs an incremental bit more of
711 * the indirect mapping to disk and updates the in-memory mapping.
712 * Called as a sync task in every txg that the removal thread makes progress.
715 vdev_mapping_sync(void *arg
, dmu_tx_t
*tx
)
717 spa_vdev_removal_t
*svr
= arg
;
718 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
719 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
720 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
721 uint64_t txg
= dmu_tx_get_txg(tx
);
722 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
724 ASSERT(vic
->vic_mapping_object
!= 0);
725 ASSERT3U(txg
, ==, spa_syncing_txg(spa
));
727 vdev_indirect_mapping_add_entries(vim
,
728 &svr
->svr_new_segments
[txg
& TXG_MASK
], tx
);
729 vdev_indirect_births_add_entry(vd
->vdev_indirect_births
,
730 vdev_indirect_mapping_max_offset(vim
), dmu_tx_get_txg(tx
), tx
);
733 * Free the copied data for anything that was freed while the
734 * mapping entries were in flight.
736 mutex_enter(&svr
->svr_lock
);
737 range_tree_vacate(svr
->svr_frees
[txg
& TXG_MASK
],
738 free_mapped_segment_cb
, vd
);
739 ASSERT3U(svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
], >=,
740 vdev_indirect_mapping_max_offset(vim
));
741 svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] = 0;
742 mutex_exit(&svr
->svr_lock
);
744 spa_sync_removing_state(spa
, tx
);
747 typedef struct vdev_copy_segment_arg
{
749 dva_t
*vcsa_dest_dva
;
751 range_tree_t
*vcsa_obsolete_segs
;
752 } vdev_copy_segment_arg_t
;
755 unalloc_seg(void *arg
, uint64_t start
, uint64_t size
)
757 vdev_copy_segment_arg_t
*vcsa
= arg
;
758 spa_t
*spa
= vcsa
->vcsa_spa
;
761 BP_SET_BIRTH(&bp
, TXG_INITIAL
, TXG_INITIAL
);
762 BP_SET_LSIZE(&bp
, size
);
763 BP_SET_PSIZE(&bp
, size
);
764 BP_SET_COMPRESS(&bp
, ZIO_COMPRESS_OFF
);
765 BP_SET_CHECKSUM(&bp
, ZIO_CHECKSUM_OFF
);
766 BP_SET_TYPE(&bp
, DMU_OT_NONE
);
767 BP_SET_LEVEL(&bp
, 0);
768 BP_SET_DEDUP(&bp
, 0);
769 BP_SET_BYTEORDER(&bp
, ZFS_HOST_BYTEORDER
);
771 DVA_SET_VDEV(&bp
.blk_dva
[0], DVA_GET_VDEV(vcsa
->vcsa_dest_dva
));
772 DVA_SET_OFFSET(&bp
.blk_dva
[0],
773 DVA_GET_OFFSET(vcsa
->vcsa_dest_dva
) + start
);
774 DVA_SET_ASIZE(&bp
.blk_dva
[0], size
);
776 zio_free(spa
, vcsa
->vcsa_txg
, &bp
);
780 * All reads and writes associated with a call to spa_vdev_copy_segment()
784 spa_vdev_copy_segment_done(zio_t
*zio
)
786 vdev_copy_segment_arg_t
*vcsa
= zio
->io_private
;
788 range_tree_vacate(vcsa
->vcsa_obsolete_segs
,
790 range_tree_destroy(vcsa
->vcsa_obsolete_segs
);
791 kmem_free(vcsa
, sizeof (*vcsa
));
793 spa_config_exit(zio
->io_spa
, SCL_STATE
, zio
->io_spa
);
797 * The write of the new location is done.
800 spa_vdev_copy_segment_write_done(zio_t
*zio
)
802 vdev_copy_arg_t
*vca
= zio
->io_private
;
804 abd_free(zio
->io_abd
);
806 mutex_enter(&vca
->vca_lock
);
807 vca
->vca_outstanding_bytes
-= zio
->io_size
;
808 cv_signal(&vca
->vca_cv
);
809 mutex_exit(&vca
->vca_lock
);
813 * The read of the old location is done. The parent zio is the write to
814 * the new location. Allow it to start.
817 spa_vdev_copy_segment_read_done(zio_t
*zio
)
819 zio_nowait(zio_unique_parent(zio
));
823 * If the old and new vdevs are mirrors, we will read both sides of the old
824 * mirror, and write each copy to the corresponding side of the new mirror.
825 * If the old and new vdevs have a different number of children, we will do
826 * this as best as possible. Since we aren't verifying checksums, this
827 * ensures that as long as there's a good copy of the data, we'll have a
828 * good copy after the removal, even if there's silent damage to one side
829 * of the mirror. If we're removing a mirror that has some silent damage,
830 * we'll have exactly the same damage in the new location (assuming that
831 * the new location is also a mirror).
833 * We accomplish this by creating a tree of zio_t's, with as many writes as
834 * there are "children" of the new vdev (a non-redundant vdev counts as one
835 * child, a 2-way mirror has 2 children, etc). Each write has an associated
836 * read from a child of the old vdev. Typically there will be the same
837 * number of children of the old and new vdevs. However, if there are more
838 * children of the new vdev, some child(ren) of the old vdev will be issued
839 * multiple reads. If there are more children of the old vdev, some copies
842 * For example, the tree of zio_t's for a 2-way mirror is:
846 * write(new vdev, child 0) write(new vdev, child 1)
848 * read(old vdev, child 0) read(old vdev, child 1)
850 * Child zio's complete before their parents complete. However, zio's
851 * created with zio_vdev_child_io() may be issued before their children
852 * complete. In this case we need to make sure that the children (reads)
853 * complete before the parents (writes) are *issued*. We do this by not
854 * calling zio_nowait() on each write until its corresponding read has
857 * The spa_config_lock must be held while zio's created by
858 * zio_vdev_child_io() are in progress, to ensure that the vdev tree does
859 * not change (e.g. due to a concurrent "zpool attach/detach"). The "null"
860 * zio is needed to release the spa_config_lock after all the reads and
861 * writes complete. (Note that we can't grab the config lock for each read,
862 * because it is not reentrant - we could deadlock with a thread waiting
866 spa_vdev_copy_one_child(vdev_copy_arg_t
*vca
, zio_t
*nzio
,
867 vdev_t
*source_vd
, uint64_t source_offset
,
868 vdev_t
*dest_child_vd
, uint64_t dest_offset
, int dest_id
, uint64_t size
)
870 ASSERT3U(spa_config_held(nzio
->io_spa
, SCL_ALL
, RW_READER
), !=, 0);
872 mutex_enter(&vca
->vca_lock
);
873 vca
->vca_outstanding_bytes
+= size
;
874 mutex_exit(&vca
->vca_lock
);
876 abd_t
*abd
= abd_alloc_for_io(size
, B_FALSE
);
878 vdev_t
*source_child_vd
;
879 if (source_vd
->vdev_ops
== &vdev_mirror_ops
&& dest_id
!= -1) {
881 * Source and dest are both mirrors. Copy from the same
882 * child id as we are copying to (wrapping around if there
883 * are more dest children than source children).
886 source_vd
->vdev_child
[dest_id
% source_vd
->vdev_children
];
888 source_child_vd
= source_vd
;
891 zio_t
*write_zio
= zio_vdev_child_io(nzio
, NULL
,
892 dest_child_vd
, dest_offset
, abd
, size
,
893 ZIO_TYPE_WRITE
, ZIO_PRIORITY_REMOVAL
,
895 spa_vdev_copy_segment_write_done
, vca
);
897 zio_nowait(zio_vdev_child_io(write_zio
, NULL
,
898 source_child_vd
, source_offset
, abd
, size
,
899 ZIO_TYPE_READ
, ZIO_PRIORITY_REMOVAL
,
901 spa_vdev_copy_segment_read_done
, vca
));
905 * Allocate a new location for this segment, and create the zio_t's to
906 * read from the old location and write to the new location.
909 spa_vdev_copy_segment(vdev_t
*vd
, range_tree_t
*segs
,
910 uint64_t maxalloc
, uint64_t txg
,
911 vdev_copy_arg_t
*vca
, zio_alloc_list_t
*zal
)
913 metaslab_group_t
*mg
= vd
->vdev_mg
;
914 spa_t
*spa
= vd
->vdev_spa
;
915 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
916 vdev_indirect_mapping_entry_t
*entry
;
918 uint64_t start
= range_tree_min(segs
);
920 ASSERT3U(maxalloc
, <=, SPA_MAXBLOCKSIZE
);
922 uint64_t size
= range_tree_span(segs
);
923 if (range_tree_span(segs
) > maxalloc
) {
925 * We can't allocate all the segments. Prefer to end
926 * the allocation at the end of a segment, thus avoiding
927 * additional split blocks.
931 search
.rs_start
= start
+ maxalloc
;
932 search
.rs_end
= search
.rs_start
;
933 range_seg_t
*rs
= avl_find(&segs
->rt_root
, &search
, &where
);
935 rs
= avl_nearest(&segs
->rt_root
, where
, AVL_BEFORE
);
937 rs
= AVL_PREV(&segs
->rt_root
, rs
);
940 size
= rs
->rs_end
- start
;
943 * There are no segments that end before maxalloc.
944 * I.e. the first segment is larger than maxalloc,
945 * so we must split it.
950 ASSERT3U(size
, <=, maxalloc
);
953 * We use allocator 0 for this I/O because we don't expect device remap
954 * to be the steady state of the system, so parallelizing is not as
955 * critical as it is for other allocation types. We also want to ensure
956 * that the IOs are allocated together as much as possible, to reduce
959 int error
= metaslab_alloc_dva(spa
, mg
->mg_class
, size
,
960 &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-%"PRIu64
,
1083 VDEV_REMOVAL_ZAP_OBJS
, 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
);
1149 ASSERT3P(vd
->vdev_initialize_thread
, ==, NULL
);
1151 sysevent_t
*ev
= spa_event_create(spa
, vd
, NULL
,
1152 ESC_ZFS_VDEV_REMOVE_DEV
);
1154 zfs_dbgmsg("finishing device removal for vdev %llu in txg %llu",
1158 * Discard allocation state.
1160 if (vd
->vdev_mg
!= NULL
) {
1161 vdev_metaslab_fini(vd
);
1162 metaslab_group_destroy(vd
->vdev_mg
);
1165 ASSERT0(vd
->vdev_stat
.vs_space
);
1166 ASSERT0(vd
->vdev_stat
.vs_dspace
);
1168 vdev_remove_replace_with_indirect(vd
, txg
);
1171 * We now release the locks, allowing spa_sync to run and finish the
1172 * removal via vdev_remove_complete_sync in syncing context.
1174 * Note that we hold on to the vdev_t that has been replaced. Since
1175 * it isn't part of the vdev tree any longer, it can't be concurrently
1176 * manipulated, even while we don't have the config lock.
1178 (void) spa_vdev_exit(spa
, NULL
, txg
, 0);
1181 * Top ZAP should have been transferred to the indirect vdev in
1182 * vdev_remove_replace_with_indirect.
1184 ASSERT0(vd
->vdev_top_zap
);
1187 * Leaf ZAP should have been moved in vdev_remove_replace_with_indirect.
1189 ASSERT0(vd
->vdev_leaf_zap
);
1191 txg
= spa_vdev_enter(spa
);
1192 (void) vdev_label_init(vd
, 0, VDEV_LABEL_REMOVE
);
1194 * Request to update the config and the config cachefile.
1196 vdev_config_dirty(spa
->spa_root_vdev
);
1197 (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
= avl_first(&svr
->svr_allocd_segs
->rt_root
);
1231 uint64_t seg_length
;
1233 if (range_tree_is_empty(segs
)) {
1234 /* need to truncate the first seg based on max_alloc */
1236 MIN(rs
->rs_end
- rs
->rs_start
, *max_alloc
);
1238 if (rs
->rs_start
- range_tree_max(segs
) >
1239 vdev_removal_max_span
) {
1241 * Including this segment would cause us to
1242 * copy a larger unneeded chunk than is allowed.
1245 } else if (rs
->rs_end
- range_tree_min(segs
) >
1248 * This additional segment would extend past
1249 * max_alloc. Rather than splitting this
1250 * segment, leave it for the next mapping.
1254 seg_length
= rs
->rs_end
- rs
->rs_start
;
1258 range_tree_add(segs
, rs
->rs_start
, seg_length
);
1259 range_tree_remove(svr
->svr_allocd_segs
,
1260 rs
->rs_start
, seg_length
);
1263 if (range_tree_is_empty(segs
)) {
1264 mutex_exit(&svr
->svr_lock
);
1265 range_tree_destroy(segs
);
1269 if (svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] == 0) {
1270 dsl_sync_task_nowait(dmu_tx_pool(tx
), vdev_mapping_sync
,
1271 svr
, 0, ZFS_SPACE_CHECK_NONE
, tx
);
1274 svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] = range_tree_max(segs
);
1277 * Note: this is the amount of *allocated* space
1278 * that we are taking care of each txg.
1280 svr
->svr_bytes_done
[txg
& TXG_MASK
] += range_tree_space(segs
);
1282 mutex_exit(&svr
->svr_lock
);
1284 zio_alloc_list_t zal
;
1285 metaslab_trace_init(&zal
);
1286 uint64_t thismax
= SPA_MAXBLOCKSIZE
;
1287 while (!range_tree_is_empty(segs
)) {
1288 int error
= spa_vdev_copy_segment(vd
,
1289 segs
, thismax
, txg
, vca
, &zal
);
1291 if (error
== ENOSPC
) {
1293 * Cut our segment in half, and don't try this
1294 * segment size again this txg. Note that the
1295 * allocation size must be aligned to the highest
1296 * ashift in the pool, so that the allocation will
1297 * not be padded out to a multiple of the ashift,
1298 * which could cause us to think that this mapping
1299 * is larger than we intended.
1301 ASSERT3U(spa
->spa_max_ashift
, >=, SPA_MINBLOCKSHIFT
);
1302 ASSERT3U(spa
->spa_max_ashift
, ==, spa
->spa_min_ashift
);
1303 uint64_t attempted
=
1304 MIN(range_tree_span(segs
), thismax
);
1305 thismax
= P2ROUNDUP(attempted
/ 2,
1306 1 << spa
->spa_max_ashift
);
1308 * The minimum-size allocation can not fail.
1310 ASSERT3U(attempted
, >, 1 << spa
->spa_max_ashift
);
1311 *max_alloc
= attempted
- (1 << spa
->spa_max_ashift
);
1316 * We've performed an allocation, so reset the
1319 metaslab_trace_fini(&zal
);
1320 metaslab_trace_init(&zal
);
1323 metaslab_trace_fini(&zal
);
1324 range_tree_destroy(segs
);
1328 * The removal thread operates in open context. It iterates over all
1329 * allocated space in the vdev, by loading each metaslab's spacemap.
1330 * For each contiguous segment of allocated space (capping the segment
1331 * size at SPA_MAXBLOCKSIZE), we:
1332 * - Allocate space for it on another vdev.
1333 * - Create a new mapping from the old location to the new location
1334 * (as a record in svr_new_segments).
1335 * - Initiate a logical read zio to get the data off the removing disk.
1336 * - In the read zio's done callback, initiate a logical write zio to
1337 * write it to the new vdev.
1338 * Note that all of this will take effect when a particular TXG syncs.
1339 * The sync thread ensures that all the phys reads and writes for the syncing
1340 * TXG have completed (see spa_txg_zio) and writes the new mappings to disk
1341 * (see vdev_mapping_sync()).
1344 spa_vdev_remove_thread(void *arg
)
1347 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1348 vdev_copy_arg_t vca
;
1349 uint64_t max_alloc
= zfs_remove_max_segment
;
1350 uint64_t last_txg
= 0;
1352 spa_config_enter(spa
, SCL_CONFIG
, FTAG
, RW_READER
);
1353 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1354 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
1355 uint64_t start_offset
= vdev_indirect_mapping_max_offset(vim
);
1357 ASSERT3P(vd
->vdev_ops
, !=, &vdev_indirect_ops
);
1358 ASSERT(vdev_is_concrete(vd
));
1359 ASSERT(vd
->vdev_removing
);
1360 ASSERT(vd
->vdev_indirect_config
.vic_mapping_object
!= 0);
1361 ASSERT(vim
!= NULL
);
1363 mutex_init(&vca
.vca_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1364 cv_init(&vca
.vca_cv
, NULL
, CV_DEFAULT
, NULL
);
1365 vca
.vca_outstanding_bytes
= 0;
1367 mutex_enter(&svr
->svr_lock
);
1370 * Start from vim_max_offset so we pick up where we left off
1371 * if we are restarting the removal after opening the pool.
1374 for (msi
= start_offset
>> vd
->vdev_ms_shift
;
1375 msi
< vd
->vdev_ms_count
&& !svr
->svr_thread_exit
; msi
++) {
1376 metaslab_t
*msp
= vd
->vdev_ms
[msi
];
1377 ASSERT3U(msi
, <=, vd
->vdev_ms_count
);
1379 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1381 mutex_enter(&msp
->ms_sync_lock
);
1382 mutex_enter(&msp
->ms_lock
);
1385 * Assert nothing in flight -- ms_*tree is empty.
1387 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1388 ASSERT0(range_tree_space(msp
->ms_allocating
[i
]));
1392 * If the metaslab has ever been allocated from (ms_sm!=NULL),
1393 * read the allocated segments from the space map object
1394 * into svr_allocd_segs. Since we do this while holding
1395 * svr_lock and ms_sync_lock, concurrent frees (which
1396 * would have modified the space map) will wait for us
1397 * to finish loading the spacemap, and then take the
1398 * appropriate action (see free_from_removing_vdev()).
1400 if (msp
->ms_sm
!= NULL
) {
1401 space_map_t
*sm
= NULL
;
1404 * We have to open a new space map here, because
1405 * ms_sm's sm_length and sm_alloc may not reflect
1406 * what's in the object contents, if we are in between
1407 * metaslab_sync() and metaslab_sync_done().
1409 VERIFY0(space_map_open(&sm
,
1410 spa
->spa_dsl_pool
->dp_meta_objset
,
1411 msp
->ms_sm
->sm_object
, msp
->ms_sm
->sm_start
,
1412 msp
->ms_sm
->sm_size
, msp
->ms_sm
->sm_shift
));
1413 space_map_update(sm
);
1414 VERIFY0(space_map_load(sm
, svr
->svr_allocd_segs
,
1416 space_map_close(sm
);
1418 range_tree_walk(msp
->ms_freeing
,
1419 range_tree_remove
, svr
->svr_allocd_segs
);
1422 * When we are resuming from a paused removal (i.e.
1423 * when importing a pool with a removal in progress),
1424 * discard any state that we have already processed.
1426 range_tree_clear(svr
->svr_allocd_segs
, 0, start_offset
);
1428 mutex_exit(&msp
->ms_lock
);
1429 mutex_exit(&msp
->ms_sync_lock
);
1432 zfs_dbgmsg("copying %llu segments for metaslab %llu",
1433 avl_numnodes(&svr
->svr_allocd_segs
->rt_root
),
1436 while (!svr
->svr_thread_exit
&&
1437 !range_tree_is_empty(svr
->svr_allocd_segs
)) {
1439 mutex_exit(&svr
->svr_lock
);
1442 * We need to periodically drop the config lock so that
1443 * writers can get in. Additionally, we can't wait
1444 * for a txg to sync while holding a config lock
1445 * (since a waiting writer could cause a 3-way deadlock
1446 * with the sync thread, which also gets a config
1447 * lock for reader). So we can't hold the config lock
1448 * while calling dmu_tx_assign().
1450 spa_config_exit(spa
, SCL_CONFIG
, FTAG
);
1453 * This delay will pause the removal around the point
1454 * specified by zfs_remove_max_bytes_pause. We do this
1455 * solely from the test suite or during debugging.
1457 uint64_t bytes_copied
=
1458 spa
->spa_removing_phys
.sr_copied
;
1459 for (int i
= 0; i
< TXG_SIZE
; i
++)
1460 bytes_copied
+= svr
->svr_bytes_done
[i
];
1461 while (zfs_remove_max_bytes_pause
<= bytes_copied
&&
1462 !svr
->svr_thread_exit
)
1465 mutex_enter(&vca
.vca_lock
);
1466 while (vca
.vca_outstanding_bytes
>
1467 zfs_remove_max_copy_bytes
) {
1468 cv_wait(&vca
.vca_cv
, &vca
.vca_lock
);
1470 mutex_exit(&vca
.vca_lock
);
1473 dmu_tx_create_dd(spa_get_dsl(spa
)->dp_mos_dir
);
1475 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
1476 uint64_t txg
= dmu_tx_get_txg(tx
);
1479 * Reacquire the vdev_config lock. The vdev_t
1480 * that we're removing may have changed, e.g. due
1481 * to a vdev_attach or vdev_detach.
1483 spa_config_enter(spa
, SCL_CONFIG
, FTAG
, RW_READER
);
1484 vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1486 if (txg
!= last_txg
)
1487 max_alloc
= zfs_remove_max_segment
;
1490 spa_vdev_copy_impl(vd
, svr
, &vca
, &max_alloc
, tx
);
1493 mutex_enter(&svr
->svr_lock
);
1497 mutex_exit(&svr
->svr_lock
);
1499 spa_config_exit(spa
, SCL_CONFIG
, FTAG
);
1502 * Wait for all copies to finish before cleaning up the vca.
1504 txg_wait_synced(spa
->spa_dsl_pool
, 0);
1505 ASSERT0(vca
.vca_outstanding_bytes
);
1507 mutex_destroy(&vca
.vca_lock
);
1508 cv_destroy(&vca
.vca_cv
);
1510 if (svr
->svr_thread_exit
) {
1511 mutex_enter(&svr
->svr_lock
);
1512 range_tree_vacate(svr
->svr_allocd_segs
, NULL
, NULL
);
1513 svr
->svr_thread
= NULL
;
1514 cv_broadcast(&svr
->svr_cv
);
1515 mutex_exit(&svr
->svr_lock
);
1517 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1518 vdev_remove_complete(spa
);
1523 spa_vdev_remove_suspend(spa_t
*spa
)
1525 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1530 mutex_enter(&svr
->svr_lock
);
1531 svr
->svr_thread_exit
= B_TRUE
;
1532 while (svr
->svr_thread
!= NULL
)
1533 cv_wait(&svr
->svr_cv
, &svr
->svr_lock
);
1534 svr
->svr_thread_exit
= B_FALSE
;
1535 mutex_exit(&svr
->svr_lock
);
1540 spa_vdev_remove_cancel_check(void *arg
, dmu_tx_t
*tx
)
1542 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1544 if (spa
->spa_vdev_removal
== NULL
)
1545 return (ENOTACTIVE
);
1550 * Cancel a removal by freeing all entries from the partial mapping
1551 * and marking the vdev as no longer being removing.
1555 spa_vdev_remove_cancel_sync(void *arg
, dmu_tx_t
*tx
)
1557 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1558 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1559 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1560 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
1561 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
1562 objset_t
*mos
= spa
->spa_meta_objset
;
1564 ASSERT3P(svr
->svr_thread
, ==, NULL
);
1566 spa_feature_decr(spa
, SPA_FEATURE_DEVICE_REMOVAL
, tx
);
1567 if (vdev_obsolete_counts_are_precise(vd
)) {
1568 spa_feature_decr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
1569 VERIFY0(zap_remove(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
1570 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE
, tx
));
1573 if (vdev_obsolete_sm_object(vd
) != 0) {
1574 ASSERT(vd
->vdev_obsolete_sm
!= NULL
);
1575 ASSERT3U(vdev_obsolete_sm_object(vd
), ==,
1576 space_map_object(vd
->vdev_obsolete_sm
));
1578 space_map_free(vd
->vdev_obsolete_sm
, tx
);
1579 VERIFY0(zap_remove(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
1580 VDEV_TOP_ZAP_INDIRECT_OBSOLETE_SM
, tx
));
1581 space_map_close(vd
->vdev_obsolete_sm
);
1582 vd
->vdev_obsolete_sm
= NULL
;
1583 spa_feature_decr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
1585 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1586 ASSERT(list_is_empty(&svr
->svr_new_segments
[i
]));
1587 ASSERT3U(svr
->svr_max_offset_to_sync
[i
], <=,
1588 vdev_indirect_mapping_max_offset(vim
));
1591 for (uint64_t msi
= 0; msi
< vd
->vdev_ms_count
; msi
++) {
1592 metaslab_t
*msp
= vd
->vdev_ms
[msi
];
1594 if (msp
->ms_start
>= vdev_indirect_mapping_max_offset(vim
))
1597 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1599 mutex_enter(&msp
->ms_lock
);
1602 * Assert nothing in flight -- ms_*tree is empty.
1604 for (int i
= 0; i
< TXG_SIZE
; i
++)
1605 ASSERT0(range_tree_space(msp
->ms_allocating
[i
]));
1606 for (int i
= 0; i
< TXG_DEFER_SIZE
; i
++)
1607 ASSERT0(range_tree_space(msp
->ms_defer
[i
]));
1608 ASSERT0(range_tree_space(msp
->ms_freed
));
1610 if (msp
->ms_sm
!= NULL
) {
1612 * Assert that the in-core spacemap has the same
1613 * length as the on-disk one, so we can use the
1614 * existing in-core spacemap to load it from disk.
1616 ASSERT3U(msp
->ms_sm
->sm_alloc
, ==,
1617 msp
->ms_sm
->sm_phys
->smp_alloc
);
1618 ASSERT3U(msp
->ms_sm
->sm_length
, ==,
1619 msp
->ms_sm
->sm_phys
->smp_objsize
);
1621 mutex_enter(&svr
->svr_lock
);
1622 VERIFY0(space_map_load(msp
->ms_sm
,
1623 svr
->svr_allocd_segs
, SM_ALLOC
));
1624 range_tree_walk(msp
->ms_freeing
,
1625 range_tree_remove
, svr
->svr_allocd_segs
);
1628 * Clear everything past what has been synced,
1629 * because we have not allocated mappings for it yet.
1631 uint64_t syncd
= vdev_indirect_mapping_max_offset(vim
);
1632 uint64_t sm_end
= msp
->ms_sm
->sm_start
+
1633 msp
->ms_sm
->sm_size
;
1635 range_tree_clear(svr
->svr_allocd_segs
,
1636 syncd
, sm_end
- syncd
);
1638 mutex_exit(&svr
->svr_lock
);
1640 mutex_exit(&msp
->ms_lock
);
1642 mutex_enter(&svr
->svr_lock
);
1643 range_tree_vacate(svr
->svr_allocd_segs
,
1644 free_mapped_segment_cb
, vd
);
1645 mutex_exit(&svr
->svr_lock
);
1649 * Note: this must happen after we invoke free_mapped_segment_cb,
1650 * because it adds to the obsolete_segments.
1652 range_tree_vacate(vd
->vdev_obsolete_segments
, NULL
, NULL
);
1654 ASSERT3U(vic
->vic_mapping_object
, ==,
1655 vdev_indirect_mapping_object(vd
->vdev_indirect_mapping
));
1656 vdev_indirect_mapping_close(vd
->vdev_indirect_mapping
);
1657 vd
->vdev_indirect_mapping
= NULL
;
1658 vdev_indirect_mapping_free(mos
, vic
->vic_mapping_object
, tx
);
1659 vic
->vic_mapping_object
= 0;
1661 ASSERT3U(vic
->vic_births_object
, ==,
1662 vdev_indirect_births_object(vd
->vdev_indirect_births
));
1663 vdev_indirect_births_close(vd
->vdev_indirect_births
);
1664 vd
->vdev_indirect_births
= NULL
;
1665 vdev_indirect_births_free(mos
, vic
->vic_births_object
, tx
);
1666 vic
->vic_births_object
= 0;
1669 * We may have processed some frees from the removing vdev in this
1670 * txg, thus increasing svr_bytes_done; discard that here to
1671 * satisfy the assertions in spa_vdev_removal_destroy().
1672 * Note that future txg's can not have any bytes_done, because
1673 * future TXG's are only modified from open context, and we have
1674 * already shut down the copying thread.
1676 svr
->svr_bytes_done
[dmu_tx_get_txg(tx
) & TXG_MASK
] = 0;
1677 spa_finish_removal(spa
, DSS_CANCELED
, tx
);
1679 vd
->vdev_removing
= B_FALSE
;
1680 vdev_config_dirty(vd
);
1682 zfs_dbgmsg("canceled device removal for vdev %llu in %llu",
1683 vd
->vdev_id
, dmu_tx_get_txg(tx
));
1684 spa_history_log_internal(spa
, "vdev remove canceled", tx
,
1685 "%s vdev %llu %s", spa_name(spa
),
1686 vd
->vdev_id
, (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
1690 spa_vdev_remove_cancel(spa_t
*spa
)
1692 spa_vdev_remove_suspend(spa
);
1694 if (spa
->spa_vdev_removal
== NULL
)
1695 return (ENOTACTIVE
);
1697 uint64_t vdid
= spa
->spa_vdev_removal
->svr_vdev_id
;
1699 int error
= dsl_sync_task(spa
->spa_name
, spa_vdev_remove_cancel_check
,
1700 spa_vdev_remove_cancel_sync
, NULL
, 0,
1701 ZFS_SPACE_CHECK_EXTRA_RESERVED
);
1704 spa_config_enter(spa
, SCL_ALLOC
| SCL_VDEV
, FTAG
, RW_WRITER
);
1705 vdev_t
*vd
= vdev_lookup_top(spa
, vdid
);
1706 metaslab_group_activate(vd
->vdev_mg
);
1707 spa_config_exit(spa
, SCL_ALLOC
| SCL_VDEV
, FTAG
);
1714 * Called every sync pass of every txg if there's a svr.
1717 svr_sync(spa_t
*spa
, dmu_tx_t
*tx
)
1719 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1720 int txgoff
= dmu_tx_get_txg(tx
) & TXG_MASK
;
1723 * This check is necessary so that we do not dirty the
1724 * DIRECTORY_OBJECT via spa_sync_removing_state() when there
1725 * is nothing to do. Dirtying it every time would prevent us
1726 * from syncing-to-convergence.
1728 if (svr
->svr_bytes_done
[txgoff
] == 0)
1732 * Update progress accounting.
1734 spa
->spa_removing_phys
.sr_copied
+= svr
->svr_bytes_done
[txgoff
];
1735 svr
->svr_bytes_done
[txgoff
] = 0;
1737 spa_sync_removing_state(spa
, tx
);
1741 vdev_remove_make_hole_and_free(vdev_t
*vd
)
1743 uint64_t id
= vd
->vdev_id
;
1744 spa_t
*spa
= vd
->vdev_spa
;
1745 vdev_t
*rvd
= spa
->spa_root_vdev
;
1746 boolean_t last_vdev
= (id
== (rvd
->vdev_children
- 1));
1748 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
1749 ASSERT(spa_config_held(spa
, SCL_ALL
, RW_WRITER
) == SCL_ALL
);
1754 vdev_compact_children(rvd
);
1756 vd
= vdev_alloc_common(spa
, id
, 0, &vdev_hole_ops
);
1757 vdev_add_child(rvd
, vd
);
1759 vdev_config_dirty(rvd
);
1762 * Reassess the health of our root vdev.
1768 * Remove a log device. The config lock is held for the specified TXG.
1771 spa_vdev_remove_log(vdev_t
*vd
, uint64_t *txg
)
1773 metaslab_group_t
*mg
= vd
->vdev_mg
;
1774 spa_t
*spa
= vd
->vdev_spa
;
1777 ASSERT(vd
->vdev_islog
);
1778 ASSERT(vd
== vd
->vdev_top
);
1781 * Stop allocating from this vdev.
1783 metaslab_group_passivate(mg
);
1786 * Wait for the youngest allocations and frees to sync,
1787 * and then wait for the deferral of those frees to finish.
1789 spa_vdev_config_exit(spa
, NULL
,
1790 *txg
+ TXG_CONCURRENT_STATES
+ TXG_DEFER_SIZE
, 0, FTAG
);
1793 * Evacuate the device. We don't hold the config lock as writer
1794 * since we need to do I/O but we do keep the
1795 * spa_namespace_lock held. Once this completes the device
1796 * should no longer have any blocks allocated on it.
1798 if (vd
->vdev_islog
) {
1799 if (vd
->vdev_stat
.vs_alloc
!= 0)
1800 error
= spa_reset_logs(spa
);
1803 *txg
= spa_vdev_config_enter(spa
);
1806 metaslab_group_activate(mg
);
1809 ASSERT0(vd
->vdev_stat
.vs_alloc
);
1812 * The evacuation succeeded. Remove any remaining MOS metadata
1813 * associated with this vdev, and wait for these changes to sync.
1815 vd
->vdev_removing
= B_TRUE
;
1817 vdev_dirty_leaves(vd
, VDD_DTL
, *txg
);
1818 vdev_config_dirty(vd
);
1820 spa_history_log_internal(spa
, "vdev remove", NULL
,
1821 "%s vdev %llu (log) %s", spa_name(spa
), vd
->vdev_id
,
1822 (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
1824 /* Make sure these changes are sync'ed */
1825 spa_vdev_config_exit(spa
, NULL
, *txg
, 0, FTAG
);
1827 /* Stop initializing */
1828 (void) vdev_initialize_stop_all(vd
, VDEV_INITIALIZE_CANCELED
);
1830 *txg
= spa_vdev_config_enter(spa
);
1832 sysevent_t
*ev
= spa_event_create(spa
, vd
, NULL
,
1833 ESC_ZFS_VDEV_REMOVE_DEV
);
1834 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
1835 ASSERT(spa_config_held(spa
, SCL_ALL
, RW_WRITER
) == SCL_ALL
);
1837 /* The top ZAP should have been destroyed by vdev_remove_empty. */
1838 ASSERT0(vd
->vdev_top_zap
);
1839 /* The leaf ZAP should have been destroyed by vdev_dtl_sync. */
1840 ASSERT0(vd
->vdev_leaf_zap
);
1842 (void) vdev_label_init(vd
, 0, VDEV_LABEL_REMOVE
);
1844 if (list_link_active(&vd
->vdev_state_dirty_node
))
1845 vdev_state_clean(vd
);
1846 if (list_link_active(&vd
->vdev_config_dirty_node
))
1847 vdev_config_clean(vd
);
1850 * Clean up the vdev namespace.
1852 vdev_remove_make_hole_and_free(vd
);
1861 spa_vdev_remove_top_check(vdev_t
*vd
)
1863 spa_t
*spa
= vd
->vdev_spa
;
1865 if (vd
!= vd
->vdev_top
)
1866 return (SET_ERROR(ENOTSUP
));
1868 if (!spa_feature_is_enabled(spa
, SPA_FEATURE_DEVICE_REMOVAL
))
1869 return (SET_ERROR(ENOTSUP
));
1872 * There has to be enough free space to remove the
1873 * device and leave double the "slop" space (i.e. we
1874 * must leave at least 3% of the pool free, in addition to
1875 * the normal slop space).
1877 if (dsl_dir_space_available(spa
->spa_dsl_pool
->dp_root_dir
,
1879 vd
->vdev_stat
.vs_dspace
+ spa_get_slop_space(spa
)) {
1880 return (SET_ERROR(ENOSPC
));
1884 * There can not be a removal in progress.
1886 if (spa
->spa_removing_phys
.sr_state
== DSS_SCANNING
)
1887 return (SET_ERROR(EBUSY
));
1890 * The device must have all its data.
1892 if (!vdev_dtl_empty(vd
, DTL_MISSING
) ||
1893 !vdev_dtl_empty(vd
, DTL_OUTAGE
))
1894 return (SET_ERROR(EBUSY
));
1897 * The device must be healthy.
1899 if (!vdev_readable(vd
))
1900 return (SET_ERROR(EIO
));
1903 * All vdevs in normal class must have the same ashift.
1905 if (spa
->spa_max_ashift
!= spa
->spa_min_ashift
) {
1906 return (SET_ERROR(EINVAL
));
1910 * All vdevs in normal class must have the same ashift
1913 vdev_t
*rvd
= spa
->spa_root_vdev
;
1914 int num_indirect
= 0;
1915 for (uint64_t id
= 0; id
< rvd
->vdev_children
; id
++) {
1916 vdev_t
*cvd
= rvd
->vdev_child
[id
];
1917 if (cvd
->vdev_ashift
!= 0 && !cvd
->vdev_islog
)
1918 ASSERT3U(cvd
->vdev_ashift
, ==, spa
->spa_max_ashift
);
1919 if (cvd
->vdev_ops
== &vdev_indirect_ops
)
1921 if (!vdev_is_concrete(cvd
))
1923 if (cvd
->vdev_ops
== &vdev_raidz_ops
)
1924 return (SET_ERROR(EINVAL
));
1926 * Need the mirror to be mirror of leaf vdevs only
1928 if (cvd
->vdev_ops
== &vdev_mirror_ops
) {
1929 for (uint64_t cid
= 0;
1930 cid
< cvd
->vdev_children
; cid
++) {
1931 vdev_t
*tmp
= cvd
->vdev_child
[cid
];
1932 if (!tmp
->vdev_ops
->vdev_op_leaf
)
1933 return (SET_ERROR(EINVAL
));
1942 * Initiate removal of a top-level vdev, reducing the total space in the pool.
1943 * The config lock is held for the specified TXG. Once initiated,
1944 * evacuation of all allocated space (copying it to other vdevs) happens
1945 * in the background (see spa_vdev_remove_thread()), and can be canceled
1946 * (see spa_vdev_remove_cancel()). If successful, the vdev will
1947 * be transformed to an indirect vdev (see spa_vdev_remove_complete()).
1950 spa_vdev_remove_top(vdev_t
*vd
, uint64_t *txg
)
1952 spa_t
*spa
= vd
->vdev_spa
;
1956 * Check for errors up-front, so that we don't waste time
1957 * passivating the metaslab group and clearing the ZIL if there
1960 error
= spa_vdev_remove_top_check(vd
);
1965 * Stop allocating from this vdev. Note that we must check
1966 * that this is not the only device in the pool before
1967 * passivating, otherwise we will not be able to make
1968 * progress because we can't allocate from any vdevs.
1969 * The above check for sufficient free space serves this
1972 metaslab_group_t
*mg
= vd
->vdev_mg
;
1973 metaslab_group_passivate(mg
);
1976 * Wait for the youngest allocations and frees to sync,
1977 * and then wait for the deferral of those frees to finish.
1979 spa_vdev_config_exit(spa
, NULL
,
1980 *txg
+ TXG_CONCURRENT_STATES
+ TXG_DEFER_SIZE
, 0, FTAG
);
1983 * We must ensure that no "stubby" log blocks are allocated
1984 * on the device to be removed. These blocks could be
1985 * written at any time, including while we are in the middle
1988 error
= spa_reset_logs(spa
);
1991 * We stop any initializing that is currently in progress but leave
1992 * the state as "active". This will allow the initializing to resume
1993 * if the removal is canceled sometime later.
1995 vdev_initialize_stop_all(vd
, VDEV_INITIALIZE_ACTIVE
);
1997 *txg
= spa_vdev_config_enter(spa
);
2000 * Things might have changed while the config lock was dropped
2001 * (e.g. space usage). Check for errors again.
2004 error
= spa_vdev_remove_top_check(vd
);
2007 metaslab_group_activate(mg
);
2008 spa_async_request(spa
, SPA_ASYNC_INITIALIZE_RESTART
);
2012 vd
->vdev_removing
= B_TRUE
;
2014 vdev_dirty_leaves(vd
, VDD_DTL
, *txg
);
2015 vdev_config_dirty(vd
);
2016 dmu_tx_t
*tx
= dmu_tx_create_assigned(spa
->spa_dsl_pool
, *txg
);
2017 dsl_sync_task_nowait(spa
->spa_dsl_pool
,
2018 vdev_remove_initiate_sync
,
2019 (void *)(uintptr_t)vd
->vdev_id
, 0, ZFS_SPACE_CHECK_NONE
, tx
);
2026 * Remove a device from the pool.
2028 * Removing a device from the vdev namespace requires several steps
2029 * and can take a significant amount of time. As a result we use
2030 * the spa_vdev_config_[enter/exit] functions which allow us to
2031 * grab and release the spa_config_lock while still holding the namespace
2032 * lock. During each step the configuration is synced out.
2035 spa_vdev_remove(spa_t
*spa
, uint64_t guid
, boolean_t unspare
)
2038 nvlist_t
**spares
, **l2cache
, *nv
;
2040 uint_t nspares
, nl2cache
;
2042 boolean_t locked
= MUTEX_HELD(&spa_namespace_lock
);
2043 sysevent_t
*ev
= NULL
;
2045 ASSERT(spa_writeable(spa
));
2048 txg
= spa_vdev_enter(spa
);
2050 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
2051 if (spa_feature_is_active(spa
, SPA_FEATURE_POOL_CHECKPOINT
)) {
2052 error
= (spa_has_checkpoint(spa
)) ?
2053 ZFS_ERR_CHECKPOINT_EXISTS
: ZFS_ERR_DISCARDING_CHECKPOINT
;
2056 return (spa_vdev_exit(spa
, NULL
, txg
, error
));
2061 vd
= spa_lookup_by_guid(spa
, guid
, B_FALSE
);
2063 if (spa
->spa_spares
.sav_vdevs
!= NULL
&&
2064 nvlist_lookup_nvlist_array(spa
->spa_spares
.sav_config
,
2065 ZPOOL_CONFIG_SPARES
, &spares
, &nspares
) == 0 &&
2066 (nv
= spa_nvlist_lookup_by_guid(spares
, nspares
, guid
)) != NULL
) {
2068 * Only remove the hot spare if it's not currently in use
2071 if (vd
== NULL
|| unspare
) {
2072 char *nvstr
= fnvlist_lookup_string(nv
,
2074 spa_history_log_internal(spa
, "vdev remove", NULL
,
2075 "%s vdev (%s) %s", spa_name(spa
),
2076 VDEV_TYPE_SPARE
, nvstr
);
2078 vd
= spa_lookup_by_guid(spa
, guid
, B_TRUE
);
2079 ev
= spa_event_create(spa
, vd
, NULL
,
2080 ESC_ZFS_VDEV_REMOVE_AUX
);
2081 spa_vdev_remove_aux(spa
->spa_spares
.sav_config
,
2082 ZPOOL_CONFIG_SPARES
, spares
, nspares
, nv
);
2083 spa_load_spares(spa
);
2084 spa
->spa_spares
.sav_sync
= B_TRUE
;
2086 error
= SET_ERROR(EBUSY
);
2088 } else if (spa
->spa_l2cache
.sav_vdevs
!= NULL
&&
2089 nvlist_lookup_nvlist_array(spa
->spa_l2cache
.sav_config
,
2090 ZPOOL_CONFIG_L2CACHE
, &l2cache
, &nl2cache
) == 0 &&
2091 (nv
= spa_nvlist_lookup_by_guid(l2cache
, nl2cache
, guid
)) != NULL
) {
2092 char *nvstr
= fnvlist_lookup_string(nv
, ZPOOL_CONFIG_PATH
);
2093 spa_history_log_internal(spa
, "vdev remove", NULL
,
2094 "%s vdev (%s) %s", spa_name(spa
), VDEV_TYPE_L2CACHE
, nvstr
);
2096 * Cache devices can always be removed.
2098 vd
= spa_lookup_by_guid(spa
, guid
, B_TRUE
);
2099 ev
= spa_event_create(spa
, vd
, NULL
, ESC_ZFS_VDEV_REMOVE_AUX
);
2100 spa_vdev_remove_aux(spa
->spa_l2cache
.sav_config
,
2101 ZPOOL_CONFIG_L2CACHE
, l2cache
, nl2cache
, nv
);
2102 spa_load_l2cache(spa
);
2103 spa
->spa_l2cache
.sav_sync
= B_TRUE
;
2104 } else if (vd
!= NULL
&& vd
->vdev_islog
) {
2106 error
= spa_vdev_remove_log(vd
, &txg
);
2107 } else if (vd
!= NULL
) {
2109 error
= spa_vdev_remove_top(vd
, &txg
);
2112 * There is no vdev of any kind with the specified guid.
2114 error
= SET_ERROR(ENOENT
);
2118 error
= spa_vdev_exit(spa
, NULL
, txg
, error
);
2122 spa_event_discard(ev
);
2132 spa_removal_get_stats(spa_t
*spa
, pool_removal_stat_t
*prs
)
2134 prs
->prs_state
= spa
->spa_removing_phys
.sr_state
;
2136 if (prs
->prs_state
== DSS_NONE
)
2137 return (SET_ERROR(ENOENT
));
2139 prs
->prs_removing_vdev
= spa
->spa_removing_phys
.sr_removing_vdev
;
2140 prs
->prs_start_time
= spa
->spa_removing_phys
.sr_start_time
;
2141 prs
->prs_end_time
= spa
->spa_removing_phys
.sr_end_time
;
2142 prs
->prs_to_copy
= spa
->spa_removing_phys
.sr_to_copy
;
2143 prs
->prs_copied
= spa
->spa_removing_phys
.sr_copied
;
2145 if (spa
->spa_vdev_removal
!= NULL
) {
2146 for (int i
= 0; i
< TXG_SIZE
; i
++) {
2148 spa
->spa_vdev_removal
->svr_bytes_done
[i
];
2152 prs
->prs_mapping_memory
= 0;
2153 uint64_t indirect_vdev_id
=
2154 spa
->spa_removing_phys
.sr_prev_indirect_vdev
;
2155 while (indirect_vdev_id
!= -1) {
2156 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[indirect_vdev_id
];
2157 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
2158 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
2160 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
2161 prs
->prs_mapping_memory
+= vdev_indirect_mapping_size(vim
);
2162 indirect_vdev_id
= vic
->vic_prev_indirect_vdev
;