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, 2017 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;
120 #define VDEV_REMOVAL_ZAP_OBJS "lzap"
122 static void spa_vdev_remove_thread(void *arg
);
125 spa_sync_removing_state(spa_t
*spa
, dmu_tx_t
*tx
)
127 VERIFY0(zap_update(spa
->spa_dsl_pool
->dp_meta_objset
,
128 DMU_POOL_DIRECTORY_OBJECT
,
129 DMU_POOL_REMOVING
, sizeof (uint64_t),
130 sizeof (spa
->spa_removing_phys
) / sizeof (uint64_t),
131 &spa
->spa_removing_phys
, tx
));
135 spa_nvlist_lookup_by_guid(nvlist_t
**nvpp
, int count
, uint64_t target_guid
)
137 for (int i
= 0; i
< count
; i
++) {
139 fnvlist_lookup_uint64(nvpp
[i
], ZPOOL_CONFIG_GUID
);
141 if (guid
== target_guid
)
149 spa_vdev_remove_aux(nvlist_t
*config
, char *name
, nvlist_t
**dev
, int count
,
150 nvlist_t
*dev_to_remove
)
152 nvlist_t
**newdev
= NULL
;
155 newdev
= kmem_alloc((count
- 1) * sizeof (void *), KM_SLEEP
);
157 for (int i
= 0, j
= 0; i
< count
; i
++) {
158 if (dev
[i
] == dev_to_remove
)
160 VERIFY(nvlist_dup(dev
[i
], &newdev
[j
++], KM_SLEEP
) == 0);
163 VERIFY(nvlist_remove(config
, name
, DATA_TYPE_NVLIST_ARRAY
) == 0);
164 VERIFY(nvlist_add_nvlist_array(config
, name
, newdev
, count
- 1) == 0);
166 for (int i
= 0; i
< count
- 1; i
++)
167 nvlist_free(newdev
[i
]);
170 kmem_free(newdev
, (count
- 1) * sizeof (void *));
173 static spa_vdev_removal_t
*
174 spa_vdev_removal_create(vdev_t
*vd
)
176 spa_vdev_removal_t
*svr
= kmem_zalloc(sizeof (*svr
), KM_SLEEP
);
177 mutex_init(&svr
->svr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
178 cv_init(&svr
->svr_cv
, NULL
, CV_DEFAULT
, NULL
);
179 svr
->svr_allocd_segs
= range_tree_create(NULL
, NULL
);
180 svr
->svr_vdev_id
= vd
->vdev_id
;
182 for (int i
= 0; i
< TXG_SIZE
; i
++) {
183 svr
->svr_frees
[i
] = range_tree_create(NULL
, NULL
);
184 list_create(&svr
->svr_new_segments
[i
],
185 sizeof (vdev_indirect_mapping_entry_t
),
186 offsetof(vdev_indirect_mapping_entry_t
, vime_node
));
193 spa_vdev_removal_destroy(spa_vdev_removal_t
*svr
)
195 for (int i
= 0; i
< TXG_SIZE
; i
++) {
196 ASSERT0(svr
->svr_bytes_done
[i
]);
197 ASSERT0(svr
->svr_max_offset_to_sync
[i
]);
198 range_tree_destroy(svr
->svr_frees
[i
]);
199 list_destroy(&svr
->svr_new_segments
[i
]);
202 range_tree_destroy(svr
->svr_allocd_segs
);
203 mutex_destroy(&svr
->svr_lock
);
204 cv_destroy(&svr
->svr_cv
);
205 kmem_free(svr
, sizeof (*svr
));
209 * This is called as a synctask in the txg in which we will mark this vdev
210 * as removing (in the config stored in the MOS).
212 * It begins the evacuation of a toplevel vdev by:
213 * - initializing the spa_removing_phys which tracks this removal
214 * - computing the amount of space to remove for accounting purposes
215 * - dirtying all dbufs in the spa_config_object
216 * - creating the spa_vdev_removal
217 * - starting the spa_vdev_remove_thread
220 vdev_remove_initiate_sync(void *arg
, dmu_tx_t
*tx
)
222 int vdev_id
= (uintptr_t)arg
;
223 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
224 vdev_t
*vd
= vdev_lookup_top(spa
, vdev_id
);
225 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
226 objset_t
*mos
= spa
->spa_dsl_pool
->dp_meta_objset
;
227 spa_vdev_removal_t
*svr
= NULL
;
228 ASSERTV(uint64_t txg
= dmu_tx_get_txg(tx
));
230 ASSERT3P(vd
->vdev_ops
, !=, &vdev_raidz_ops
);
231 svr
= spa_vdev_removal_create(vd
);
233 ASSERT(vd
->vdev_removing
);
234 ASSERT3P(vd
->vdev_indirect_mapping
, ==, NULL
);
236 spa_feature_incr(spa
, SPA_FEATURE_DEVICE_REMOVAL
, tx
);
237 if (spa_feature_is_enabled(spa
, SPA_FEATURE_OBSOLETE_COUNTS
)) {
239 * By activating the OBSOLETE_COUNTS feature, we prevent
240 * the pool from being downgraded and ensure that the
241 * refcounts are precise.
243 spa_feature_incr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
245 VERIFY0(zap_add(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
246 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE
, sizeof (one
), 1,
248 ASSERT3U(vdev_obsolete_counts_are_precise(vd
), !=, 0);
251 vic
->vic_mapping_object
= vdev_indirect_mapping_alloc(mos
, tx
);
252 vd
->vdev_indirect_mapping
=
253 vdev_indirect_mapping_open(mos
, vic
->vic_mapping_object
);
254 vic
->vic_births_object
= vdev_indirect_births_alloc(mos
, tx
);
255 vd
->vdev_indirect_births
=
256 vdev_indirect_births_open(mos
, vic
->vic_births_object
);
257 spa
->spa_removing_phys
.sr_removing_vdev
= vd
->vdev_id
;
258 spa
->spa_removing_phys
.sr_start_time
= gethrestime_sec();
259 spa
->spa_removing_phys
.sr_end_time
= 0;
260 spa
->spa_removing_phys
.sr_state
= DSS_SCANNING
;
261 spa
->spa_removing_phys
.sr_to_copy
= 0;
262 spa
->spa_removing_phys
.sr_copied
= 0;
265 * Note: We can't use vdev_stat's vs_alloc for sr_to_copy, because
266 * there may be space in the defer tree, which is free, but still
267 * counted in vs_alloc.
269 for (uint64_t i
= 0; i
< vd
->vdev_ms_count
; i
++) {
270 metaslab_t
*ms
= vd
->vdev_ms
[i
];
271 if (ms
->ms_sm
== NULL
)
275 * Sync tasks happen before metaslab_sync(), therefore
276 * smp_alloc and sm_alloc must be the same.
278 ASSERT3U(space_map_allocated(ms
->ms_sm
), ==,
279 ms
->ms_sm
->sm_phys
->smp_alloc
);
281 spa
->spa_removing_phys
.sr_to_copy
+=
282 space_map_allocated(ms
->ms_sm
);
285 * Space which we are freeing this txg does not need to
288 spa
->spa_removing_phys
.sr_to_copy
-=
289 range_tree_space(ms
->ms_freeingtree
);
291 ASSERT0(range_tree_space(ms
->ms_freedtree
));
292 for (int t
= 0; t
< TXG_SIZE
; t
++)
293 ASSERT0(range_tree_space(ms
->ms_alloctree
[t
]));
297 * Sync tasks are called before metaslab_sync(), so there should
298 * be no already-synced metaslabs in the TXG_CLEAN list.
300 ASSERT3P(txg_list_head(&vd
->vdev_ms_list
, TXG_CLEAN(txg
)), ==, NULL
);
302 spa_sync_removing_state(spa
, tx
);
305 * All blocks that we need to read the most recent mapping must be
306 * stored on concrete vdevs. Therefore, we must dirty anything that
307 * is read before spa_remove_init(). Specifically, the
308 * spa_config_object. (Note that although we already modified the
309 * spa_config_object in spa_sync_removing_state, that may not have
310 * modified all blocks of the object.)
312 dmu_object_info_t doi
;
313 VERIFY0(dmu_object_info(mos
, DMU_POOL_DIRECTORY_OBJECT
, &doi
));
314 for (uint64_t offset
= 0; offset
< doi
.doi_max_offset
; ) {
316 VERIFY0(dmu_buf_hold(mos
, DMU_POOL_DIRECTORY_OBJECT
,
317 offset
, FTAG
, &dbuf
, 0));
318 dmu_buf_will_dirty(dbuf
, tx
);
319 offset
+= dbuf
->db_size
;
320 dmu_buf_rele(dbuf
, FTAG
);
324 * Now that we've allocated the im_object, dirty the vdev to ensure
325 * that the object gets written to the config on disk.
327 vdev_config_dirty(vd
);
329 zfs_dbgmsg("starting removal thread for vdev %llu (%p) in txg %llu "
330 "im_obj=%llu", vd
->vdev_id
, vd
, dmu_tx_get_txg(tx
),
331 vic
->vic_mapping_object
);
333 spa_history_log_internal(spa
, "vdev remove started", tx
,
334 "%s vdev %llu %s", spa_name(spa
), vd
->vdev_id
,
335 (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
337 * Setting spa_vdev_removal causes subsequent frees to call
338 * free_from_removing_vdev(). Note that we don't need any locking
339 * because we are the sync thread, and metaslab_free_impl() is only
340 * called from syncing context (potentially from a zio taskq thread,
341 * but in any case only when there are outstanding free i/os, which
344 ASSERT3P(spa
->spa_vdev_removal
, ==, NULL
);
345 spa
->spa_vdev_removal
= svr
;
346 svr
->svr_thread
= thread_create(NULL
, 0,
347 spa_vdev_remove_thread
, spa
, 0, &p0
, TS_RUN
, minclsyspri
);
351 * When we are opening a pool, we must read the mapping for each
352 * indirect vdev in order from most recently removed to least
353 * recently removed. We do this because the blocks for the mapping
354 * of older indirect vdevs may be stored on more recently removed vdevs.
355 * In order to read each indirect mapping object, we must have
356 * initialized all more recently removed vdevs.
359 spa_remove_init(spa_t
*spa
)
363 error
= zap_lookup(spa
->spa_dsl_pool
->dp_meta_objset
,
364 DMU_POOL_DIRECTORY_OBJECT
,
365 DMU_POOL_REMOVING
, sizeof (uint64_t),
366 sizeof (spa
->spa_removing_phys
) / sizeof (uint64_t),
367 &spa
->spa_removing_phys
);
369 if (error
== ENOENT
) {
370 spa
->spa_removing_phys
.sr_state
= DSS_NONE
;
371 spa
->spa_removing_phys
.sr_removing_vdev
= -1;
372 spa
->spa_removing_phys
.sr_prev_indirect_vdev
= -1;
373 spa
->spa_indirect_vdevs_loaded
= B_TRUE
;
375 } else if (error
!= 0) {
379 if (spa
->spa_removing_phys
.sr_state
== DSS_SCANNING
) {
381 * We are currently removing a vdev. Create and
382 * initialize a spa_vdev_removal_t from the bonus
383 * buffer of the removing vdevs vdev_im_object, and
384 * initialize its partial mapping.
386 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
387 vdev_t
*vd
= vdev_lookup_top(spa
,
388 spa
->spa_removing_phys
.sr_removing_vdev
);
391 spa_config_exit(spa
, SCL_STATE
, FTAG
);
395 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
397 ASSERT(vdev_is_concrete(vd
));
398 spa_vdev_removal_t
*svr
= spa_vdev_removal_create(vd
);
399 ASSERT3U(svr
->svr_vdev_id
, ==, vd
->vdev_id
);
400 ASSERT(vd
->vdev_removing
);
402 vd
->vdev_indirect_mapping
= vdev_indirect_mapping_open(
403 spa
->spa_meta_objset
, vic
->vic_mapping_object
);
404 vd
->vdev_indirect_births
= vdev_indirect_births_open(
405 spa
->spa_meta_objset
, vic
->vic_births_object
);
406 spa_config_exit(spa
, SCL_STATE
, FTAG
);
408 spa
->spa_vdev_removal
= svr
;
411 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
412 uint64_t indirect_vdev_id
=
413 spa
->spa_removing_phys
.sr_prev_indirect_vdev
;
414 while (indirect_vdev_id
!= UINT64_MAX
) {
415 vdev_t
*vd
= vdev_lookup_top(spa
, indirect_vdev_id
);
416 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
418 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
419 vd
->vdev_indirect_mapping
= vdev_indirect_mapping_open(
420 spa
->spa_meta_objset
, vic
->vic_mapping_object
);
421 vd
->vdev_indirect_births
= vdev_indirect_births_open(
422 spa
->spa_meta_objset
, vic
->vic_births_object
);
424 indirect_vdev_id
= vic
->vic_prev_indirect_vdev
;
426 spa_config_exit(spa
, SCL_STATE
, FTAG
);
429 * Now that we've loaded all the indirect mappings, we can allow
430 * reads from other blocks (e.g. via predictive prefetch).
432 spa
->spa_indirect_vdevs_loaded
= B_TRUE
;
437 spa_restart_removal(spa_t
*spa
)
439 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
445 * In general when this function is called there is no
446 * removal thread running. The only scenario where this
447 * is not true is during spa_import() where this function
448 * is called twice [once from spa_import_impl() and
449 * spa_async_resume()]. Thus, in the scenario where we
450 * import a pool that has an ongoing removal we don't
451 * want to spawn a second thread.
453 if (svr
->svr_thread
!= NULL
)
456 if (!spa_writeable(spa
))
459 zfs_dbgmsg("restarting removal of %llu", svr
->svr_vdev_id
);
460 svr
->svr_thread
= thread_create(NULL
, 0, spa_vdev_remove_thread
, spa
,
461 0, &p0
, TS_RUN
, minclsyspri
);
465 * Process freeing from a device which is in the middle of being removed.
466 * We must handle this carefully so that we attempt to copy freed data,
467 * and we correctly free already-copied data.
470 free_from_removing_vdev(vdev_t
*vd
, uint64_t offset
, uint64_t size
,
473 spa_t
*spa
= vd
->vdev_spa
;
474 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
475 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
476 uint64_t max_offset_yet
= 0;
478 ASSERT(vd
->vdev_indirect_config
.vic_mapping_object
!= 0);
479 ASSERT3U(vd
->vdev_indirect_config
.vic_mapping_object
, ==,
480 vdev_indirect_mapping_object(vim
));
481 ASSERT3U(vd
->vdev_id
, ==, svr
->svr_vdev_id
);
482 ASSERT3U(spa_syncing_txg(spa
), ==, txg
);
484 mutex_enter(&svr
->svr_lock
);
487 * Remove the segment from the removing vdev's spacemap. This
488 * ensures that we will not attempt to copy this space (if the
489 * removal thread has not yet visited it), and also ensures
490 * that we know what is actually allocated on the new vdevs
491 * (needed if we cancel the removal).
493 * Note: we must do the metaslab_free_concrete() with the svr_lock
494 * held, so that the remove_thread can not load this metaslab and then
495 * visit this offset between the time that we metaslab_free_concrete()
496 * and when we check to see if it has been visited.
498 metaslab_free_concrete(vd
, offset
, size
, txg
);
500 uint64_t synced_size
= 0;
501 uint64_t synced_offset
= 0;
502 uint64_t max_offset_synced
= vdev_indirect_mapping_max_offset(vim
);
503 if (offset
< max_offset_synced
) {
505 * The mapping for this offset is already on disk.
506 * Free from the new location.
508 * Note that we use svr_max_synced_offset because it is
509 * updated atomically with respect to the in-core mapping.
510 * By contrast, vim_max_offset is not.
512 * This block may be split between a synced entry and an
513 * in-flight or unvisited entry. Only process the synced
514 * portion of it here.
516 synced_size
= MIN(size
, max_offset_synced
- offset
);
517 synced_offset
= offset
;
519 ASSERT3U(max_offset_yet
, <=, max_offset_synced
);
520 max_offset_yet
= max_offset_synced
;
522 DTRACE_PROBE3(remove__free__synced
,
525 uint64_t, synced_size
);
528 offset
+= synced_size
;
532 * Look at all in-flight txgs starting from the currently syncing one
533 * and see if a section of this free is being copied. By starting from
534 * this txg and iterating forward, we might find that this region
535 * was copied in two different txgs and handle it appropriately.
537 for (int i
= 0; i
< TXG_CONCURRENT_STATES
; i
++) {
538 int txgoff
= (txg
+ i
) & TXG_MASK
;
539 if (size
> 0 && offset
< svr
->svr_max_offset_to_sync
[txgoff
]) {
541 * The mapping for this offset is in flight, and
542 * will be synced in txg+i.
544 uint64_t inflight_size
= MIN(size
,
545 svr
->svr_max_offset_to_sync
[txgoff
] - offset
);
547 DTRACE_PROBE4(remove__free__inflight
,
550 uint64_t, inflight_size
,
554 * We copy data in order of increasing offset.
555 * Therefore the max_offset_to_sync[] must increase
556 * (or be zero, indicating that nothing is being
557 * copied in that txg).
559 if (svr
->svr_max_offset_to_sync
[txgoff
] != 0) {
560 ASSERT3U(svr
->svr_max_offset_to_sync
[txgoff
],
563 svr
->svr_max_offset_to_sync
[txgoff
];
567 * We've already committed to copying this segment:
568 * we have allocated space elsewhere in the pool for
569 * it and have an IO outstanding to copy the data. We
570 * cannot free the space before the copy has
571 * completed, or else the copy IO might overwrite any
572 * new data. To free that space, we record the
573 * segment in the appropriate svr_frees tree and free
574 * the mapped space later, in the txg where we have
575 * completed the copy and synced the mapping (see
576 * vdev_mapping_sync).
578 range_tree_add(svr
->svr_frees
[txgoff
],
579 offset
, inflight_size
);
580 size
-= inflight_size
;
581 offset
+= inflight_size
;
584 * This space is already accounted for as being
585 * done, because it is being copied in txg+i.
586 * However, if i!=0, then it is being copied in
587 * a future txg. If we crash after this txg
588 * syncs but before txg+i syncs, then the space
589 * will be free. Therefore we must account
590 * for the space being done in *this* txg
591 * (when it is freed) rather than the future txg
592 * (when it will be copied).
594 ASSERT3U(svr
->svr_bytes_done
[txgoff
], >=,
596 svr
->svr_bytes_done
[txgoff
] -= inflight_size
;
597 svr
->svr_bytes_done
[txg
& TXG_MASK
] += inflight_size
;
600 ASSERT0(svr
->svr_max_offset_to_sync
[TXG_CLEAN(txg
) & TXG_MASK
]);
604 * The copy thread has not yet visited this offset. Ensure
608 DTRACE_PROBE3(remove__free__unvisited
,
613 if (svr
->svr_allocd_segs
!= NULL
)
614 range_tree_clear(svr
->svr_allocd_segs
, offset
, size
);
617 * Since we now do not need to copy this data, for
618 * accounting purposes we have done our job and can count
621 svr
->svr_bytes_done
[txg
& TXG_MASK
] += size
;
623 mutex_exit(&svr
->svr_lock
);
626 * Now that we have dropped svr_lock, process the synced portion
629 if (synced_size
> 0) {
630 vdev_indirect_mark_obsolete(vd
, synced_offset
, synced_size
,
633 * Note: this can only be called from syncing context,
634 * and the vdev_indirect_mapping is only changed from the
635 * sync thread, so we don't need svr_lock while doing
636 * metaslab_free_impl_cb.
638 vdev_indirect_ops
.vdev_op_remap(vd
, synced_offset
, synced_size
,
639 metaslab_free_impl_cb
, &txg
);
644 * Stop an active removal and update the spa_removing phys.
647 spa_finish_removal(spa_t
*spa
, dsl_scan_state_t state
, dmu_tx_t
*tx
)
649 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
650 ASSERT3U(dmu_tx_get_txg(tx
), ==, spa_syncing_txg(spa
));
652 /* Ensure the removal thread has completed before we free the svr. */
653 spa_vdev_remove_suspend(spa
);
655 ASSERT(state
== DSS_FINISHED
|| state
== DSS_CANCELED
);
657 if (state
== DSS_FINISHED
) {
658 spa_removing_phys_t
*srp
= &spa
->spa_removing_phys
;
659 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
660 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
662 if (srp
->sr_prev_indirect_vdev
!= UINT64_MAX
) {
664 pvd
= vdev_lookup_top(spa
,
665 srp
->sr_prev_indirect_vdev
);
666 ASSERT3P(pvd
->vdev_ops
, ==, &vdev_indirect_ops
);
669 vic
->vic_prev_indirect_vdev
= srp
->sr_prev_indirect_vdev
;
670 srp
->sr_prev_indirect_vdev
= vd
->vdev_id
;
672 spa
->spa_removing_phys
.sr_state
= state
;
673 spa
->spa_removing_phys
.sr_end_time
= gethrestime_sec();
675 spa
->spa_vdev_removal
= NULL
;
676 spa_vdev_removal_destroy(svr
);
678 spa_sync_removing_state(spa
, tx
);
680 vdev_config_dirty(spa
->spa_root_vdev
);
684 free_mapped_segment_cb(void *arg
, uint64_t offset
, uint64_t size
)
687 vdev_indirect_mark_obsolete(vd
, offset
, size
,
688 vd
->vdev_spa
->spa_syncing_txg
);
689 vdev_indirect_ops
.vdev_op_remap(vd
, offset
, size
,
690 metaslab_free_impl_cb
, &vd
->vdev_spa
->spa_syncing_txg
);
694 * On behalf of the removal thread, syncs an incremental bit more of
695 * the indirect mapping to disk and updates the in-memory mapping.
696 * Called as a sync task in every txg that the removal thread makes progress.
699 vdev_mapping_sync(void *arg
, dmu_tx_t
*tx
)
701 spa_vdev_removal_t
*svr
= arg
;
702 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
703 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
704 ASSERTV(vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
);
705 uint64_t txg
= dmu_tx_get_txg(tx
);
706 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
708 ASSERT(vic
->vic_mapping_object
!= 0);
709 ASSERT3U(txg
, ==, spa_syncing_txg(spa
));
711 vdev_indirect_mapping_add_entries(vim
,
712 &svr
->svr_new_segments
[txg
& TXG_MASK
], tx
);
713 vdev_indirect_births_add_entry(vd
->vdev_indirect_births
,
714 vdev_indirect_mapping_max_offset(vim
), dmu_tx_get_txg(tx
), tx
);
717 * Free the copied data for anything that was freed while the
718 * mapping entries were in flight.
720 mutex_enter(&svr
->svr_lock
);
721 range_tree_vacate(svr
->svr_frees
[txg
& TXG_MASK
],
722 free_mapped_segment_cb
, vd
);
723 ASSERT3U(svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
], >=,
724 vdev_indirect_mapping_max_offset(vim
));
725 svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] = 0;
726 mutex_exit(&svr
->svr_lock
);
728 spa_sync_removing_state(spa
, tx
);
731 typedef struct vdev_copy_segment_arg
{
733 dva_t
*vcsa_dest_dva
;
735 range_tree_t
*vcsa_obsolete_segs
;
736 } vdev_copy_segment_arg_t
;
739 unalloc_seg(void *arg
, uint64_t start
, uint64_t size
)
741 vdev_copy_segment_arg_t
*vcsa
= arg
;
742 spa_t
*spa
= vcsa
->vcsa_spa
;
743 blkptr_t bp
= { { { {0} } } };
745 BP_SET_BIRTH(&bp
, TXG_INITIAL
, TXG_INITIAL
);
746 BP_SET_LSIZE(&bp
, size
);
747 BP_SET_PSIZE(&bp
, size
);
748 BP_SET_COMPRESS(&bp
, ZIO_COMPRESS_OFF
);
749 BP_SET_CHECKSUM(&bp
, ZIO_CHECKSUM_OFF
);
750 BP_SET_TYPE(&bp
, DMU_OT_NONE
);
751 BP_SET_LEVEL(&bp
, 0);
752 BP_SET_DEDUP(&bp
, 0);
753 BP_SET_BYTEORDER(&bp
, ZFS_HOST_BYTEORDER
);
755 DVA_SET_VDEV(&bp
.blk_dva
[0], DVA_GET_VDEV(vcsa
->vcsa_dest_dva
));
756 DVA_SET_OFFSET(&bp
.blk_dva
[0],
757 DVA_GET_OFFSET(vcsa
->vcsa_dest_dva
) + start
);
758 DVA_SET_ASIZE(&bp
.blk_dva
[0], size
);
760 zio_free(spa
, vcsa
->vcsa_txg
, &bp
);
764 * All reads and writes associated with a call to spa_vdev_copy_segment()
768 spa_vdev_copy_segment_done(zio_t
*zio
)
770 vdev_copy_segment_arg_t
*vcsa
= zio
->io_private
;
772 range_tree_vacate(vcsa
->vcsa_obsolete_segs
,
774 range_tree_destroy(vcsa
->vcsa_obsolete_segs
);
775 kmem_free(vcsa
, sizeof (*vcsa
));
777 spa_config_exit(zio
->io_spa
, SCL_STATE
, zio
->io_spa
);
781 * The write of the new location is done.
784 spa_vdev_copy_segment_write_done(zio_t
*zio
)
786 vdev_copy_arg_t
*vca
= zio
->io_private
;
788 abd_free(zio
->io_abd
);
790 mutex_enter(&vca
->vca_lock
);
791 vca
->vca_outstanding_bytes
-= zio
->io_size
;
792 cv_signal(&vca
->vca_cv
);
793 mutex_exit(&vca
->vca_lock
);
797 * The read of the old location is done. The parent zio is the write to
798 * the new location. Allow it to start.
801 spa_vdev_copy_segment_read_done(zio_t
*zio
)
803 zio_nowait(zio_unique_parent(zio
));
807 * If the old and new vdevs are mirrors, we will read both sides of the old
808 * mirror, and write each copy to the corresponding side of the new mirror.
809 * If the old and new vdevs have a different number of children, we will do
810 * this as best as possible. Since we aren't verifying checksums, this
811 * ensures that as long as there's a good copy of the data, we'll have a
812 * good copy after the removal, even if there's silent damage to one side
813 * of the mirror. If we're removing a mirror that has some silent damage,
814 * we'll have exactly the same damage in the new location (assuming that
815 * the new location is also a mirror).
817 * We accomplish this by creating a tree of zio_t's, with as many writes as
818 * there are "children" of the new vdev (a non-redundant vdev counts as one
819 * child, a 2-way mirror has 2 children, etc). Each write has an associated
820 * read from a child of the old vdev. Typically there will be the same
821 * number of children of the old and new vdevs. However, if there are more
822 * children of the new vdev, some child(ren) of the old vdev will be issued
823 * multiple reads. If there are more children of the old vdev, some copies
826 * For example, the tree of zio_t's for a 2-way mirror is:
830 * write(new vdev, child 0) write(new vdev, child 1)
832 * read(old vdev, child 0) read(old vdev, child 1)
834 * Child zio's complete before their parents complete. However, zio's
835 * created with zio_vdev_child_io() may be issued before their children
836 * complete. In this case we need to make sure that the children (reads)
837 * complete before the parents (writes) are *issued*. We do this by not
838 * calling zio_nowait() on each write until its corresponding read has
841 * The spa_config_lock must be held while zio's created by
842 * zio_vdev_child_io() are in progress, to ensure that the vdev tree does
843 * not change (e.g. due to a concurrent "zpool attach/detach"). The "null"
844 * zio is needed to release the spa_config_lock after all the reads and
845 * writes complete. (Note that we can't grab the config lock for each read,
846 * because it is not reentrant - we could deadlock with a thread waiting
850 spa_vdev_copy_one_child(vdev_copy_arg_t
*vca
, zio_t
*nzio
,
851 vdev_t
*source_vd
, uint64_t source_offset
,
852 vdev_t
*dest_child_vd
, uint64_t dest_offset
, int dest_id
, uint64_t size
)
854 ASSERT3U(spa_config_held(nzio
->io_spa
, SCL_ALL
, RW_READER
), !=, 0);
856 mutex_enter(&vca
->vca_lock
);
857 vca
->vca_outstanding_bytes
+= size
;
858 mutex_exit(&vca
->vca_lock
);
860 abd_t
*abd
= abd_alloc_for_io(size
, B_FALSE
);
862 vdev_t
*source_child_vd
;
863 if (source_vd
->vdev_ops
== &vdev_mirror_ops
&& dest_id
!= -1) {
865 * Source and dest are both mirrors. Copy from the same
866 * child id as we are copying to (wrapping around if there
867 * are more dest children than source children).
870 source_vd
->vdev_child
[dest_id
% source_vd
->vdev_children
];
872 source_child_vd
= source_vd
;
875 zio_t
*write_zio
= zio_vdev_child_io(nzio
, NULL
,
876 dest_child_vd
, dest_offset
, abd
, size
,
877 ZIO_TYPE_WRITE
, ZIO_PRIORITY_REMOVAL
,
879 spa_vdev_copy_segment_write_done
, vca
);
881 zio_nowait(zio_vdev_child_io(write_zio
, NULL
,
882 source_child_vd
, source_offset
, abd
, size
,
883 ZIO_TYPE_READ
, ZIO_PRIORITY_REMOVAL
,
885 spa_vdev_copy_segment_read_done
, vca
));
889 * Allocate a new location for this segment, and create the zio_t's to
890 * read from the old location and write to the new location.
893 spa_vdev_copy_segment(vdev_t
*vd
, range_tree_t
*segs
,
894 uint64_t maxalloc
, uint64_t txg
,
895 vdev_copy_arg_t
*vca
, zio_alloc_list_t
*zal
)
897 metaslab_group_t
*mg
= vd
->vdev_mg
;
898 spa_t
*spa
= vd
->vdev_spa
;
899 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
900 vdev_indirect_mapping_entry_t
*entry
;
902 uint64_t start
= range_tree_min(segs
);
904 ASSERT3U(maxalloc
, <=, SPA_MAXBLOCKSIZE
);
906 uint64_t size
= range_tree_span(segs
);
907 if (range_tree_span(segs
) > maxalloc
) {
909 * We can't allocate all the segments. Prefer to end
910 * the allocation at the end of a segment, thus avoiding
911 * additional split blocks.
915 search
.rs_start
= start
+ maxalloc
;
916 search
.rs_end
= search
.rs_start
;
917 range_seg_t
*rs
= avl_find(&segs
->rt_root
, &search
, &where
);
919 rs
= avl_nearest(&segs
->rt_root
, where
, AVL_BEFORE
);
921 rs
= AVL_PREV(&segs
->rt_root
, rs
);
924 size
= rs
->rs_end
- start
;
927 * There are no segments that end before maxalloc.
928 * I.e. the first segment is larger than maxalloc,
929 * so we must split it.
934 ASSERT3U(size
, <=, maxalloc
);
936 int error
= metaslab_alloc_dva(spa
, mg
->mg_class
, size
,
937 &dst
, 0, NULL
, txg
, 0, zal
);
942 * Determine the ranges that are not actually needed. Offsets are
943 * relative to the start of the range to be copied (i.e. relative to the
944 * local variable "start").
946 range_tree_t
*obsolete_segs
= range_tree_create(NULL
, NULL
);
948 range_seg_t
*rs
= avl_first(&segs
->rt_root
);
949 ASSERT3U(rs
->rs_start
, ==, start
);
950 uint64_t prev_seg_end
= rs
->rs_end
;
951 while ((rs
= AVL_NEXT(&segs
->rt_root
, rs
)) != NULL
) {
952 if (rs
->rs_start
>= start
+ size
) {
955 range_tree_add(obsolete_segs
,
956 prev_seg_end
- start
,
957 rs
->rs_start
- prev_seg_end
);
959 prev_seg_end
= rs
->rs_end
;
961 /* We don't end in the middle of an obsolete range */
962 ASSERT3U(start
+ size
, <=, prev_seg_end
);
964 range_tree_clear(segs
, start
, size
);
967 * We can't have any padding of the allocated size, otherwise we will
968 * misunderstand what's allocated, and the size of the mapping.
969 * The caller ensures this will be true by passing in a size that is
970 * aligned to the worst (highest) ashift in the pool.
972 ASSERT3U(DVA_GET_ASIZE(&dst
), ==, size
);
974 entry
= kmem_zalloc(sizeof (vdev_indirect_mapping_entry_t
), KM_SLEEP
);
975 DVA_MAPPING_SET_SRC_OFFSET(&entry
->vime_mapping
, start
);
976 entry
->vime_mapping
.vimep_dst
= dst
;
977 if (spa_feature_is_enabled(spa
, SPA_FEATURE_OBSOLETE_COUNTS
)) {
978 entry
->vime_obsolete_count
= range_tree_space(obsolete_segs
);
981 vdev_copy_segment_arg_t
*vcsa
= kmem_zalloc(sizeof (*vcsa
), KM_SLEEP
);
982 vcsa
->vcsa_dest_dva
= &entry
->vime_mapping
.vimep_dst
;
983 vcsa
->vcsa_obsolete_segs
= obsolete_segs
;
984 vcsa
->vcsa_spa
= spa
;
985 vcsa
->vcsa_txg
= txg
;
988 * See comment before spa_vdev_copy_one_child().
990 spa_config_enter(spa
, SCL_STATE
, spa
, RW_READER
);
991 zio_t
*nzio
= zio_null(spa
->spa_txg_zio
[txg
& TXG_MASK
], spa
, NULL
,
992 spa_vdev_copy_segment_done
, vcsa
, 0);
993 vdev_t
*dest_vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dst
));
994 if (dest_vd
->vdev_ops
== &vdev_mirror_ops
) {
995 for (int i
= 0; i
< dest_vd
->vdev_children
; i
++) {
996 vdev_t
*child
= dest_vd
->vdev_child
[i
];
997 spa_vdev_copy_one_child(vca
, nzio
, vd
, start
,
998 child
, DVA_GET_OFFSET(&dst
), i
, size
);
1001 spa_vdev_copy_one_child(vca
, nzio
, vd
, start
,
1002 dest_vd
, DVA_GET_OFFSET(&dst
), -1, size
);
1006 list_insert_tail(&svr
->svr_new_segments
[txg
& TXG_MASK
], entry
);
1007 ASSERT3U(start
+ size
, <=, vd
->vdev_ms_count
<< vd
->vdev_ms_shift
);
1008 vdev_dirty(vd
, 0, NULL
, txg
);
1014 * Complete the removal of a toplevel vdev. This is called as a
1015 * synctask in the same txg that we will sync out the new config (to the
1016 * MOS object) which indicates that this vdev is indirect.
1019 vdev_remove_complete_sync(void *arg
, dmu_tx_t
*tx
)
1021 spa_vdev_removal_t
*svr
= arg
;
1022 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1023 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1025 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
1027 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1028 ASSERT0(svr
->svr_bytes_done
[i
]);
1031 ASSERT3U(spa
->spa_removing_phys
.sr_copied
, ==,
1032 spa
->spa_removing_phys
.sr_to_copy
);
1034 vdev_destroy_spacemaps(vd
, tx
);
1036 /* destroy leaf zaps, if any */
1037 ASSERT3P(svr
->svr_zaplist
, !=, NULL
);
1038 for (nvpair_t
*pair
= nvlist_next_nvpair(svr
->svr_zaplist
, NULL
);
1040 pair
= nvlist_next_nvpair(svr
->svr_zaplist
, pair
)) {
1041 vdev_destroy_unlink_zap(vd
, fnvpair_value_uint64(pair
), tx
);
1043 fnvlist_free(svr
->svr_zaplist
);
1045 spa_finish_removal(dmu_tx_pool(tx
)->dp_spa
, DSS_FINISHED
, tx
);
1046 /* vd->vdev_path is not available here */
1047 spa_history_log_internal(spa
, "vdev remove completed", tx
,
1048 "%s vdev %llu", spa_name(spa
), vd
->vdev_id
);
1052 vdev_remove_enlist_zaps(vdev_t
*vd
, nvlist_t
*zlist
)
1054 ASSERT3P(zlist
, !=, NULL
);
1055 ASSERT3P(vd
->vdev_ops
, !=, &vdev_raidz_ops
);
1057 if (vd
->vdev_leaf_zap
!= 0) {
1059 (void) snprintf(zkey
, sizeof (zkey
), "%s-%llu",
1060 VDEV_REMOVAL_ZAP_OBJS
, (u_longlong_t
)vd
->vdev_leaf_zap
);
1061 fnvlist_add_uint64(zlist
, zkey
, vd
->vdev_leaf_zap
);
1064 for (uint64_t id
= 0; id
< vd
->vdev_children
; id
++) {
1065 vdev_remove_enlist_zaps(vd
->vdev_child
[id
], zlist
);
1070 vdev_remove_replace_with_indirect(vdev_t
*vd
, uint64_t txg
)
1074 spa_t
*spa
= vd
->vdev_spa
;
1075 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1078 * First, build a list of leaf zaps to be destroyed.
1079 * This is passed to the sync context thread,
1080 * which does the actual unlinking.
1082 svr
->svr_zaplist
= fnvlist_alloc();
1083 vdev_remove_enlist_zaps(vd
, svr
->svr_zaplist
);
1085 ivd
= vdev_add_parent(vd
, &vdev_indirect_ops
);
1086 ivd
->vdev_removing
= 0;
1088 vd
->vdev_leaf_zap
= 0;
1090 vdev_remove_child(ivd
, vd
);
1091 vdev_compact_children(ivd
);
1093 ASSERT(!list_link_active(&vd
->vdev_state_dirty_node
));
1095 tx
= dmu_tx_create_assigned(spa
->spa_dsl_pool
, txg
);
1096 dsl_sync_task_nowait(spa
->spa_dsl_pool
, vdev_remove_complete_sync
, svr
,
1097 0, ZFS_SPACE_CHECK_NONE
, tx
);
1101 * Indicate that this thread has exited.
1102 * After this, we can not use svr.
1104 mutex_enter(&svr
->svr_lock
);
1105 svr
->svr_thread
= NULL
;
1106 cv_broadcast(&svr
->svr_cv
);
1107 mutex_exit(&svr
->svr_lock
);
1111 * Complete the removal of a toplevel vdev. This is called in open
1112 * context by the removal thread after we have copied all vdev's data.
1115 vdev_remove_complete(spa_t
*spa
)
1120 * Wait for any deferred frees to be synced before we call
1121 * vdev_metaslab_fini()
1123 txg_wait_synced(spa
->spa_dsl_pool
, 0);
1124 txg
= spa_vdev_enter(spa
);
1125 vdev_t
*vd
= vdev_lookup_top(spa
, spa
->spa_vdev_removal
->svr_vdev_id
);
1127 sysevent_t
*ev
= spa_event_create(spa
, vd
, NULL
,
1128 ESC_ZFS_VDEV_REMOVE_DEV
);
1130 zfs_dbgmsg("finishing device removal for vdev %llu in txg %llu",
1134 * Discard allocation state.
1136 if (vd
->vdev_mg
!= NULL
) {
1137 vdev_metaslab_fini(vd
);
1138 metaslab_group_destroy(vd
->vdev_mg
);
1141 ASSERT0(vd
->vdev_stat
.vs_space
);
1142 ASSERT0(vd
->vdev_stat
.vs_dspace
);
1144 vdev_remove_replace_with_indirect(vd
, txg
);
1147 * We now release the locks, allowing spa_sync to run and finish the
1148 * removal via vdev_remove_complete_sync in syncing context.
1150 * Note that we hold on to the vdev_t that has been replaced. Since
1151 * it isn't part of the vdev tree any longer, it can't be concurrently
1152 * manipulated, even while we don't have the config lock.
1154 (void) spa_vdev_exit(spa
, NULL
, txg
, 0);
1157 * Top ZAP should have been transferred to the indirect vdev in
1158 * vdev_remove_replace_with_indirect.
1160 ASSERT0(vd
->vdev_top_zap
);
1163 * Leaf ZAP should have been moved in vdev_remove_replace_with_indirect.
1165 ASSERT0(vd
->vdev_leaf_zap
);
1167 txg
= spa_vdev_enter(spa
);
1168 (void) vdev_label_init(vd
, 0, VDEV_LABEL_REMOVE
);
1170 * Request to update the config and the config cachefile.
1172 vdev_config_dirty(spa
->spa_root_vdev
);
1173 (void) spa_vdev_exit(spa
, vd
, txg
, 0);
1180 * Evacuates a segment of size at most max_alloc from the vdev
1181 * via repeated calls to spa_vdev_copy_segment. If an allocation
1182 * fails, the pool is probably too fragmented to handle such a
1183 * large size, so decrease max_alloc so that the caller will not try
1184 * this size again this txg.
1187 spa_vdev_copy_impl(vdev_t
*vd
, spa_vdev_removal_t
*svr
, vdev_copy_arg_t
*vca
,
1188 uint64_t *max_alloc
, dmu_tx_t
*tx
)
1190 uint64_t txg
= dmu_tx_get_txg(tx
);
1191 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1193 mutex_enter(&svr
->svr_lock
);
1196 * Determine how big of a chunk to copy. We can allocate up
1197 * to max_alloc bytes, and we can span up to vdev_removal_max_span
1198 * bytes of unallocated space at a time. "segs" will track the
1199 * allocated segments that we are copying. We may also be copying
1200 * free segments (of up to vdev_removal_max_span bytes).
1202 range_tree_t
*segs
= range_tree_create(NULL
, NULL
);
1204 range_seg_t
*rs
= range_tree_first(svr
->svr_allocd_segs
);
1209 uint64_t seg_length
;
1211 if (range_tree_is_empty(segs
)) {
1212 /* need to truncate the first seg based on max_alloc */
1214 MIN(rs
->rs_end
- rs
->rs_start
, *max_alloc
);
1216 if (rs
->rs_start
- range_tree_max(segs
) >
1217 vdev_removal_max_span
) {
1219 * Including this segment would cause us to
1220 * copy a larger unneeded chunk than is allowed.
1223 } else if (rs
->rs_end
- range_tree_min(segs
) >
1226 * This additional segment would extend past
1227 * max_alloc. Rather than splitting this
1228 * segment, leave it for the next mapping.
1232 seg_length
= rs
->rs_end
- rs
->rs_start
;
1236 range_tree_add(segs
, rs
->rs_start
, seg_length
);
1237 range_tree_remove(svr
->svr_allocd_segs
,
1238 rs
->rs_start
, seg_length
);
1241 if (range_tree_is_empty(segs
)) {
1242 mutex_exit(&svr
->svr_lock
);
1243 range_tree_destroy(segs
);
1247 if (svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] == 0) {
1248 dsl_sync_task_nowait(dmu_tx_pool(tx
), vdev_mapping_sync
,
1249 svr
, 0, ZFS_SPACE_CHECK_NONE
, tx
);
1252 svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] = range_tree_max(segs
);
1255 * Note: this is the amount of *allocated* space
1256 * that we are taking care of each txg.
1258 svr
->svr_bytes_done
[txg
& TXG_MASK
] += range_tree_space(segs
);
1260 mutex_exit(&svr
->svr_lock
);
1262 zio_alloc_list_t zal
;
1263 metaslab_trace_init(&zal
);
1264 uint64_t thismax
= SPA_MAXBLOCKSIZE
;
1265 while (!range_tree_is_empty(segs
)) {
1266 int error
= spa_vdev_copy_segment(vd
,
1267 segs
, thismax
, txg
, vca
, &zal
);
1269 if (error
== ENOSPC
) {
1271 * Cut our segment in half, and don't try this
1272 * segment size again this txg. Note that the
1273 * allocation size must be aligned to the highest
1274 * ashift in the pool, so that the allocation will
1275 * not be padded out to a multiple of the ashift,
1276 * which could cause us to think that this mapping
1277 * is larger than we intended.
1279 ASSERT3U(spa
->spa_max_ashift
, >=, SPA_MINBLOCKSHIFT
);
1280 ASSERT3U(spa
->spa_max_ashift
, ==, spa
->spa_min_ashift
);
1281 uint64_t attempted
=
1282 MIN(range_tree_span(segs
), thismax
);
1283 thismax
= P2ROUNDUP(attempted
/ 2,
1284 1 << spa
->spa_max_ashift
);
1286 * The minimum-size allocation can not fail.
1288 ASSERT3U(attempted
, >, 1 << spa
->spa_max_ashift
);
1289 *max_alloc
= attempted
- (1 << spa
->spa_max_ashift
);
1294 * We've performed an allocation, so reset the
1297 metaslab_trace_fini(&zal
);
1298 metaslab_trace_init(&zal
);
1301 metaslab_trace_fini(&zal
);
1302 range_tree_destroy(segs
);
1306 * The removal thread operates in open context. It iterates over all
1307 * allocated space in the vdev, by loading each metaslab's spacemap.
1308 * For each contiguous segment of allocated space (capping the segment
1309 * size at SPA_MAXBLOCKSIZE), we:
1310 * - Allocate space for it on another vdev.
1311 * - Create a new mapping from the old location to the new location
1312 * (as a record in svr_new_segments).
1313 * - Initiate a physical read zio to get the data off the removing disk.
1314 * - In the read zio's done callback, initiate a physical write zio to
1315 * write it to the new vdev.
1316 * Note that all of this will take effect when a particular TXG syncs.
1317 * The sync thread ensures that all the phys reads and writes for the syncing
1318 * TXG have completed (see spa_txg_zio) and writes the new mappings to disk
1319 * (see vdev_mapping_sync()).
1322 spa_vdev_remove_thread(void *arg
)
1325 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1326 vdev_copy_arg_t vca
;
1327 uint64_t max_alloc
= zfs_remove_max_segment
;
1328 uint64_t last_txg
= 0;
1330 spa_config_enter(spa
, SCL_CONFIG
, FTAG
, RW_READER
);
1331 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1332 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
1333 uint64_t start_offset
= vdev_indirect_mapping_max_offset(vim
);
1335 ASSERT3P(vd
->vdev_ops
, !=, &vdev_indirect_ops
);
1336 ASSERT(vdev_is_concrete(vd
));
1337 ASSERT(vd
->vdev_removing
);
1338 ASSERT(vd
->vdev_indirect_config
.vic_mapping_object
!= 0);
1339 ASSERT(vim
!= NULL
);
1341 mutex_init(&vca
.vca_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1342 cv_init(&vca
.vca_cv
, NULL
, CV_DEFAULT
, NULL
);
1343 vca
.vca_outstanding_bytes
= 0;
1345 mutex_enter(&svr
->svr_lock
);
1348 * Start from vim_max_offset so we pick up where we left off
1349 * if we are restarting the removal after opening the pool.
1352 for (msi
= start_offset
>> vd
->vdev_ms_shift
;
1353 msi
< vd
->vdev_ms_count
&& !svr
->svr_thread_exit
; msi
++) {
1354 metaslab_t
*msp
= vd
->vdev_ms
[msi
];
1355 ASSERT3U(msi
, <=, vd
->vdev_ms_count
);
1357 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1359 mutex_enter(&msp
->ms_sync_lock
);
1360 mutex_enter(&msp
->ms_lock
);
1363 * Assert nothing in flight -- ms_*tree is empty.
1365 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1366 ASSERT0(range_tree_space(msp
->ms_alloctree
[i
]));
1370 * If the metaslab has ever been allocated from (ms_sm!=NULL),
1371 * read the allocated segments from the space map object
1372 * into svr_allocd_segs. Since we do this while holding
1373 * svr_lock and ms_sync_lock, concurrent frees (which
1374 * would have modified the space map) will wait for us
1375 * to finish loading the spacemap, and then take the
1376 * appropriate action (see free_from_removing_vdev()).
1378 if (msp
->ms_sm
!= NULL
) {
1379 space_map_t
*sm
= NULL
;
1382 * We have to open a new space map here, because
1383 * ms_sm's sm_length and sm_alloc may not reflect
1384 * what's in the object contents, if we are in between
1385 * metaslab_sync() and metaslab_sync_done().
1387 VERIFY0(space_map_open(&sm
,
1388 spa
->spa_dsl_pool
->dp_meta_objset
,
1389 msp
->ms_sm
->sm_object
, msp
->ms_sm
->sm_start
,
1390 msp
->ms_sm
->sm_size
, msp
->ms_sm
->sm_shift
));
1391 space_map_update(sm
);
1392 VERIFY0(space_map_load(sm
, svr
->svr_allocd_segs
,
1394 space_map_close(sm
);
1396 range_tree_walk(msp
->ms_freeingtree
,
1397 range_tree_remove
, svr
->svr_allocd_segs
);
1400 * When we are resuming from a paused removal (i.e.
1401 * when importing a pool with a removal in progress),
1402 * discard any state that we have already processed.
1404 range_tree_clear(svr
->svr_allocd_segs
, 0, start_offset
);
1406 mutex_exit(&msp
->ms_lock
);
1407 mutex_exit(&msp
->ms_sync_lock
);
1410 zfs_dbgmsg("copying %llu segments for metaslab %llu",
1411 avl_numnodes(&svr
->svr_allocd_segs
->rt_root
),
1414 while (!svr
->svr_thread_exit
&&
1415 range_tree_space(svr
->svr_allocd_segs
) != 0) {
1417 mutex_exit(&svr
->svr_lock
);
1420 * We need to periodically drop the config lock so that
1421 * writers can get in. Additionally, we can't wait
1422 * for a txg to sync while holding a config lock
1423 * (since a waiting writer could cause a 3-way deadlock
1424 * with the sync thread, which also gets a config
1425 * lock for reader). So we can't hold the config lock
1426 * while calling dmu_tx_assign().
1428 spa_config_exit(spa
, SCL_CONFIG
, FTAG
);
1430 mutex_enter(&vca
.vca_lock
);
1431 while (vca
.vca_outstanding_bytes
>
1432 zfs_remove_max_copy_bytes
) {
1433 cv_wait(&vca
.vca_cv
, &vca
.vca_lock
);
1435 mutex_exit(&vca
.vca_lock
);
1438 dmu_tx_create_dd(spa_get_dsl(spa
)->dp_mos_dir
);
1439 dmu_tx_hold_space(tx
, SPA_MAXBLOCKSIZE
);
1440 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
1441 uint64_t txg
= dmu_tx_get_txg(tx
);
1444 * Reacquire the vdev_config lock. The vdev_t
1445 * that we're removing may have changed, e.g. due
1446 * to a vdev_attach or vdev_detach.
1448 spa_config_enter(spa
, SCL_CONFIG
, FTAG
, RW_READER
);
1449 vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1451 if (txg
!= last_txg
)
1452 max_alloc
= zfs_remove_max_segment
;
1455 spa_vdev_copy_impl(vd
, svr
, &vca
, &max_alloc
, tx
);
1458 mutex_enter(&svr
->svr_lock
);
1462 mutex_exit(&svr
->svr_lock
);
1464 spa_config_exit(spa
, SCL_CONFIG
, FTAG
);
1467 * Wait for all copies to finish before cleaning up the vca.
1469 txg_wait_synced(spa
->spa_dsl_pool
, 0);
1470 ASSERT0(vca
.vca_outstanding_bytes
);
1472 mutex_destroy(&vca
.vca_lock
);
1473 cv_destroy(&vca
.vca_cv
);
1475 if (svr
->svr_thread_exit
) {
1476 mutex_enter(&svr
->svr_lock
);
1477 range_tree_vacate(svr
->svr_allocd_segs
, NULL
, NULL
);
1478 svr
->svr_thread
= NULL
;
1479 cv_broadcast(&svr
->svr_cv
);
1480 mutex_exit(&svr
->svr_lock
);
1482 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1483 vdev_remove_complete(spa
);
1488 spa_vdev_remove_suspend(spa_t
*spa
)
1490 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1495 mutex_enter(&svr
->svr_lock
);
1496 svr
->svr_thread_exit
= B_TRUE
;
1497 while (svr
->svr_thread
!= NULL
)
1498 cv_wait(&svr
->svr_cv
, &svr
->svr_lock
);
1499 svr
->svr_thread_exit
= B_FALSE
;
1500 mutex_exit(&svr
->svr_lock
);
1505 spa_vdev_remove_cancel_check(void *arg
, dmu_tx_t
*tx
)
1507 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1509 if (spa
->spa_vdev_removal
== NULL
)
1510 return (ENOTACTIVE
);
1515 * Cancel a removal by freeing all entries from the partial mapping
1516 * and marking the vdev as no longer being removing.
1520 spa_vdev_remove_cancel_sync(void *arg
, dmu_tx_t
*tx
)
1522 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1523 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1524 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1525 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
1526 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
1527 objset_t
*mos
= spa
->spa_meta_objset
;
1529 ASSERT3P(svr
->svr_thread
, ==, NULL
);
1531 spa_feature_decr(spa
, SPA_FEATURE_DEVICE_REMOVAL
, tx
);
1532 if (vdev_obsolete_counts_are_precise(vd
)) {
1533 spa_feature_decr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
1534 VERIFY0(zap_remove(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
1535 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE
, tx
));
1538 if (vdev_obsolete_sm_object(vd
) != 0) {
1539 ASSERT(vd
->vdev_obsolete_sm
!= NULL
);
1540 ASSERT3U(vdev_obsolete_sm_object(vd
), ==,
1541 space_map_object(vd
->vdev_obsolete_sm
));
1543 space_map_free(vd
->vdev_obsolete_sm
, tx
);
1544 VERIFY0(zap_remove(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
1545 VDEV_TOP_ZAP_INDIRECT_OBSOLETE_SM
, tx
));
1546 space_map_close(vd
->vdev_obsolete_sm
);
1547 vd
->vdev_obsolete_sm
= NULL
;
1548 spa_feature_decr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
1550 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1551 ASSERT(list_is_empty(&svr
->svr_new_segments
[i
]));
1552 ASSERT3U(svr
->svr_max_offset_to_sync
[i
], <=,
1553 vdev_indirect_mapping_max_offset(vim
));
1556 for (uint64_t msi
= 0; msi
< vd
->vdev_ms_count
; msi
++) {
1557 metaslab_t
*msp
= vd
->vdev_ms
[msi
];
1559 if (msp
->ms_start
>= vdev_indirect_mapping_max_offset(vim
))
1562 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1564 mutex_enter(&msp
->ms_lock
);
1567 * Assert nothing in flight -- ms_*tree is empty.
1569 for (int i
= 0; i
< TXG_SIZE
; i
++)
1570 ASSERT0(range_tree_space(msp
->ms_alloctree
[i
]));
1571 for (int i
= 0; i
< TXG_DEFER_SIZE
; i
++)
1572 ASSERT0(range_tree_space(msp
->ms_defertree
[i
]));
1573 ASSERT0(range_tree_space(msp
->ms_freedtree
));
1575 if (msp
->ms_sm
!= NULL
) {
1577 * Assert that the in-core spacemap has the same
1578 * length as the on-disk one, so we can use the
1579 * existing in-core spacemap to load it from disk.
1581 ASSERT3U(msp
->ms_sm
->sm_alloc
, ==,
1582 msp
->ms_sm
->sm_phys
->smp_alloc
);
1583 ASSERT3U(msp
->ms_sm
->sm_length
, ==,
1584 msp
->ms_sm
->sm_phys
->smp_objsize
);
1586 mutex_enter(&svr
->svr_lock
);
1587 VERIFY0(space_map_load(msp
->ms_sm
,
1588 svr
->svr_allocd_segs
, SM_ALLOC
));
1589 range_tree_walk(msp
->ms_freeingtree
,
1590 range_tree_remove
, svr
->svr_allocd_segs
);
1593 * Clear everything past what has been synced,
1594 * because we have not allocated mappings for it yet.
1596 uint64_t syncd
= vdev_indirect_mapping_max_offset(vim
);
1597 uint64_t sm_end
= msp
->ms_sm
->sm_start
+
1598 msp
->ms_sm
->sm_size
;
1600 range_tree_clear(svr
->svr_allocd_segs
,
1601 syncd
, sm_end
- syncd
);
1603 mutex_exit(&svr
->svr_lock
);
1605 mutex_exit(&msp
->ms_lock
);
1607 mutex_enter(&svr
->svr_lock
);
1608 range_tree_vacate(svr
->svr_allocd_segs
,
1609 free_mapped_segment_cb
, vd
);
1610 mutex_exit(&svr
->svr_lock
);
1614 * Note: this must happen after we invoke free_mapped_segment_cb,
1615 * because it adds to the obsolete_segments.
1617 range_tree_vacate(vd
->vdev_obsolete_segments
, NULL
, NULL
);
1619 ASSERT3U(vic
->vic_mapping_object
, ==,
1620 vdev_indirect_mapping_object(vd
->vdev_indirect_mapping
));
1621 vdev_indirect_mapping_close(vd
->vdev_indirect_mapping
);
1622 vd
->vdev_indirect_mapping
= NULL
;
1623 vdev_indirect_mapping_free(mos
, vic
->vic_mapping_object
, tx
);
1624 vic
->vic_mapping_object
= 0;
1626 ASSERT3U(vic
->vic_births_object
, ==,
1627 vdev_indirect_births_object(vd
->vdev_indirect_births
));
1628 vdev_indirect_births_close(vd
->vdev_indirect_births
);
1629 vd
->vdev_indirect_births
= NULL
;
1630 vdev_indirect_births_free(mos
, vic
->vic_births_object
, tx
);
1631 vic
->vic_births_object
= 0;
1634 * We may have processed some frees from the removing vdev in this
1635 * txg, thus increasing svr_bytes_done; discard that here to
1636 * satisfy the assertions in spa_vdev_removal_destroy().
1637 * Note that future txg's can not have any bytes_done, because
1638 * future TXG's are only modified from open context, and we have
1639 * already shut down the copying thread.
1641 svr
->svr_bytes_done
[dmu_tx_get_txg(tx
) & TXG_MASK
] = 0;
1642 spa_finish_removal(spa
, DSS_CANCELED
, tx
);
1644 vd
->vdev_removing
= B_FALSE
;
1645 vdev_config_dirty(vd
);
1647 zfs_dbgmsg("canceled device removal for vdev %llu in %llu",
1648 vd
->vdev_id
, dmu_tx_get_txg(tx
));
1649 spa_history_log_internal(spa
, "vdev remove canceled", tx
,
1650 "%s vdev %llu %s", spa_name(spa
),
1651 vd
->vdev_id
, (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
1655 spa_vdev_remove_cancel(spa_t
*spa
)
1657 spa_vdev_remove_suspend(spa
);
1659 if (spa
->spa_vdev_removal
== NULL
)
1660 return (ENOTACTIVE
);
1662 uint64_t vdid
= spa
->spa_vdev_removal
->svr_vdev_id
;
1664 int error
= dsl_sync_task(spa
->spa_name
, spa_vdev_remove_cancel_check
,
1665 spa_vdev_remove_cancel_sync
, NULL
, 0, ZFS_SPACE_CHECK_NONE
);
1668 spa_config_enter(spa
, SCL_ALLOC
| SCL_VDEV
, FTAG
, RW_WRITER
);
1669 vdev_t
*vd
= vdev_lookup_top(spa
, vdid
);
1670 metaslab_group_activate(vd
->vdev_mg
);
1671 spa_config_exit(spa
, SCL_ALLOC
| SCL_VDEV
, FTAG
);
1678 * Called every sync pass of every txg if there's a svr.
1681 svr_sync(spa_t
*spa
, dmu_tx_t
*tx
)
1683 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1684 int txgoff
= dmu_tx_get_txg(tx
) & TXG_MASK
;
1687 * This check is necessary so that we do not dirty the
1688 * DIRECTORY_OBJECT via spa_sync_removing_state() when there
1689 * is nothing to do. Dirtying it every time would prevent us
1690 * from syncing-to-convergence.
1692 if (svr
->svr_bytes_done
[txgoff
] == 0)
1696 * Update progress accounting.
1698 spa
->spa_removing_phys
.sr_copied
+= svr
->svr_bytes_done
[txgoff
];
1699 svr
->svr_bytes_done
[txgoff
] = 0;
1701 spa_sync_removing_state(spa
, tx
);
1705 vdev_remove_make_hole_and_free(vdev_t
*vd
)
1707 uint64_t id
= vd
->vdev_id
;
1708 spa_t
*spa
= vd
->vdev_spa
;
1709 vdev_t
*rvd
= spa
->spa_root_vdev
;
1710 boolean_t last_vdev
= (id
== (rvd
->vdev_children
- 1));
1712 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
1713 ASSERT(spa_config_held(spa
, SCL_ALL
, RW_WRITER
) == SCL_ALL
);
1718 vdev_compact_children(rvd
);
1720 vd
= vdev_alloc_common(spa
, id
, 0, &vdev_hole_ops
);
1721 vdev_add_child(rvd
, vd
);
1723 vdev_config_dirty(rvd
);
1726 * Reassess the health of our root vdev.
1732 * Remove a log device. The config lock is held for the specified TXG.
1735 spa_vdev_remove_log(vdev_t
*vd
, uint64_t *txg
)
1737 metaslab_group_t
*mg
= vd
->vdev_mg
;
1738 spa_t
*spa
= vd
->vdev_spa
;
1741 ASSERT(vd
->vdev_islog
);
1742 ASSERT(vd
== vd
->vdev_top
);
1745 * Stop allocating from this vdev.
1747 metaslab_group_passivate(mg
);
1750 * Wait for the youngest allocations and frees to sync,
1751 * and then wait for the deferral of those frees to finish.
1753 spa_vdev_config_exit(spa
, NULL
,
1754 *txg
+ TXG_CONCURRENT_STATES
+ TXG_DEFER_SIZE
, 0, FTAG
);
1757 * Evacuate the device. We don't hold the config lock as writer
1758 * since we need to do I/O but we do keep the
1759 * spa_namespace_lock held. Once this completes the device
1760 * should no longer have any blocks allocated on it.
1762 if (vd
->vdev_islog
) {
1763 if (vd
->vdev_stat
.vs_alloc
!= 0)
1764 error
= spa_reset_logs(spa
);
1767 *txg
= spa_vdev_config_enter(spa
);
1770 metaslab_group_activate(mg
);
1773 ASSERT0(vd
->vdev_stat
.vs_alloc
);
1776 * The evacuation succeeded. Remove any remaining MOS metadata
1777 * associated with this vdev, and wait for these changes to sync.
1779 vd
->vdev_removing
= B_TRUE
;
1781 vdev_dirty_leaves(vd
, VDD_DTL
, *txg
);
1782 vdev_config_dirty(vd
);
1784 spa_history_log_internal(spa
, "vdev remove", NULL
,
1785 "%s vdev %llu (log) %s", spa_name(spa
), vd
->vdev_id
,
1786 (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
1788 spa_vdev_config_exit(spa
, NULL
, *txg
, 0, FTAG
);
1790 *txg
= spa_vdev_config_enter(spa
);
1792 sysevent_t
*ev
= spa_event_create(spa
, vd
, NULL
,
1793 ESC_ZFS_VDEV_REMOVE_DEV
);
1794 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
1795 ASSERT(spa_config_held(spa
, SCL_ALL
, RW_WRITER
) == SCL_ALL
);
1797 /* The top ZAP should have been destroyed by vdev_remove_empty. */
1798 ASSERT0(vd
->vdev_top_zap
);
1799 /* The leaf ZAP should have been destroyed by vdev_dtl_sync. */
1800 ASSERT0(vd
->vdev_leaf_zap
);
1802 (void) vdev_label_init(vd
, 0, VDEV_LABEL_REMOVE
);
1804 if (list_link_active(&vd
->vdev_state_dirty_node
))
1805 vdev_state_clean(vd
);
1806 if (list_link_active(&vd
->vdev_config_dirty_node
))
1807 vdev_config_clean(vd
);
1810 * Clean up the vdev namespace.
1812 vdev_remove_make_hole_and_free(vd
);
1821 spa_vdev_remove_top_check(vdev_t
*vd
)
1823 spa_t
*spa
= vd
->vdev_spa
;
1825 if (vd
!= vd
->vdev_top
)
1826 return (SET_ERROR(ENOTSUP
));
1828 if (!spa_feature_is_enabled(spa
, SPA_FEATURE_DEVICE_REMOVAL
))
1829 return (SET_ERROR(ENOTSUP
));
1832 * There has to be enough free space to remove the
1833 * device and leave double the "slop" space (i.e. we
1834 * must leave at least 3% of the pool free, in addition to
1835 * the normal slop space).
1837 if (dsl_dir_space_available(spa
->spa_dsl_pool
->dp_root_dir
,
1839 vd
->vdev_stat
.vs_dspace
+ spa_get_slop_space(spa
)) {
1840 return (SET_ERROR(ENOSPC
));
1844 * There can not be a removal in progress.
1846 if (spa
->spa_removing_phys
.sr_state
== DSS_SCANNING
)
1847 return (SET_ERROR(EBUSY
));
1850 * The device must have all its data.
1852 if (!vdev_dtl_empty(vd
, DTL_MISSING
) ||
1853 !vdev_dtl_empty(vd
, DTL_OUTAGE
))
1854 return (SET_ERROR(EBUSY
));
1857 * The device must be healthy.
1859 if (!vdev_readable(vd
))
1860 return (SET_ERROR(EIO
));
1863 * All vdevs in normal class must have the same ashift.
1865 if (spa
->spa_max_ashift
!= spa
->spa_min_ashift
) {
1866 return (SET_ERROR(EINVAL
));
1870 * All vdevs in normal class must have the same ashift
1873 vdev_t
*rvd
= spa
->spa_root_vdev
;
1874 int num_indirect
= 0;
1875 for (uint64_t id
= 0; id
< rvd
->vdev_children
; id
++) {
1876 vdev_t
*cvd
= rvd
->vdev_child
[id
];
1877 if (cvd
->vdev_ashift
!= 0 && !cvd
->vdev_islog
)
1878 ASSERT3U(cvd
->vdev_ashift
, ==, spa
->spa_max_ashift
);
1879 if (cvd
->vdev_ops
== &vdev_indirect_ops
)
1881 if (!vdev_is_concrete(cvd
))
1883 if (cvd
->vdev_ops
== &vdev_raidz_ops
)
1884 return (SET_ERROR(EINVAL
));
1886 * Need the mirror to be mirror of leaf vdevs only
1888 if (cvd
->vdev_ops
== &vdev_mirror_ops
) {
1889 for (uint64_t cid
= 0;
1890 cid
< cvd
->vdev_children
; cid
++) {
1891 if (!cvd
->vdev_child
[cid
]->vdev_ops
->
1893 return (SET_ERROR(EINVAL
));
1902 * Initiate removal of a top-level vdev, reducing the total space in the pool.
1903 * The config lock is held for the specified TXG. Once initiated,
1904 * evacuation of all allocated space (copying it to other vdevs) happens
1905 * in the background (see spa_vdev_remove_thread()), and can be canceled
1906 * (see spa_vdev_remove_cancel()). If successful, the vdev will
1907 * be transformed to an indirect vdev (see spa_vdev_remove_complete()).
1910 spa_vdev_remove_top(vdev_t
*vd
, uint64_t *txg
)
1912 spa_t
*spa
= vd
->vdev_spa
;
1916 * Check for errors up-front, so that we don't waste time
1917 * passivating the metaslab group and clearing the ZIL if there
1920 error
= spa_vdev_remove_top_check(vd
);
1925 * Stop allocating from this vdev. Note that we must check
1926 * that this is not the only device in the pool before
1927 * passivating, otherwise we will not be able to make
1928 * progress because we can't allocate from any vdevs.
1929 * The above check for sufficient free space serves this
1932 metaslab_group_t
*mg
= vd
->vdev_mg
;
1933 metaslab_group_passivate(mg
);
1936 * Wait for the youngest allocations and frees to sync,
1937 * and then wait for the deferral of those frees to finish.
1939 spa_vdev_config_exit(spa
, NULL
,
1940 *txg
+ TXG_CONCURRENT_STATES
+ TXG_DEFER_SIZE
, 0, FTAG
);
1943 * We must ensure that no "stubby" log blocks are allocated
1944 * on the device to be removed. These blocks could be
1945 * written at any time, including while we are in the middle
1948 error
= spa_reset_logs(spa
);
1950 *txg
= spa_vdev_config_enter(spa
);
1953 * Things might have changed while the config lock was dropped
1954 * (e.g. space usage). Check for errors again.
1957 error
= spa_vdev_remove_top_check(vd
);
1960 metaslab_group_activate(mg
);
1964 vd
->vdev_removing
= B_TRUE
;
1966 vdev_dirty_leaves(vd
, VDD_DTL
, *txg
);
1967 vdev_config_dirty(vd
);
1968 dmu_tx_t
*tx
= dmu_tx_create_assigned(spa
->spa_dsl_pool
, *txg
);
1969 dsl_sync_task_nowait(spa
->spa_dsl_pool
,
1970 vdev_remove_initiate_sync
,
1971 (void *)(uintptr_t)vd
->vdev_id
, 0, ZFS_SPACE_CHECK_NONE
, tx
);
1978 * Remove a device from the pool.
1980 * Removing a device from the vdev namespace requires several steps
1981 * and can take a significant amount of time. As a result we use
1982 * the spa_vdev_config_[enter/exit] functions which allow us to
1983 * grab and release the spa_config_lock while still holding the namespace
1984 * lock. During each step the configuration is synced out.
1987 spa_vdev_remove(spa_t
*spa
, uint64_t guid
, boolean_t unspare
)
1990 nvlist_t
**spares
, **l2cache
, *nv
;
1992 uint_t nspares
, nl2cache
;
1994 boolean_t locked
= MUTEX_HELD(&spa_namespace_lock
);
1995 sysevent_t
*ev
= NULL
;
1997 ASSERT(spa_writeable(spa
));
2000 txg
= spa_vdev_enter(spa
);
2002 vd
= spa_lookup_by_guid(spa
, guid
, B_FALSE
);
2004 if (spa
->spa_spares
.sav_vdevs
!= NULL
&&
2005 nvlist_lookup_nvlist_array(spa
->spa_spares
.sav_config
,
2006 ZPOOL_CONFIG_SPARES
, &spares
, &nspares
) == 0 &&
2007 (nv
= spa_nvlist_lookup_by_guid(spares
, nspares
, guid
)) != NULL
) {
2009 * Only remove the hot spare if it's not currently in use
2012 if (vd
== NULL
|| unspare
) {
2014 vd
= spa_lookup_by_guid(spa
, guid
, B_TRUE
);
2015 ev
= spa_event_create(spa
, vd
, NULL
,
2016 ESC_ZFS_VDEV_REMOVE_AUX
);
2018 char *nvstr
= fnvlist_lookup_string(nv
,
2020 spa_history_log_internal(spa
, "vdev remove", NULL
,
2021 "%s vdev (%s) %s", spa_name(spa
),
2022 VDEV_TYPE_SPARE
, nvstr
);
2023 spa_vdev_remove_aux(spa
->spa_spares
.sav_config
,
2024 ZPOOL_CONFIG_SPARES
, spares
, nspares
, nv
);
2025 spa_load_spares(spa
);
2026 spa
->spa_spares
.sav_sync
= B_TRUE
;
2028 error
= SET_ERROR(EBUSY
);
2030 } else if (spa
->spa_l2cache
.sav_vdevs
!= NULL
&&
2031 nvlist_lookup_nvlist_array(spa
->spa_l2cache
.sav_config
,
2032 ZPOOL_CONFIG_L2CACHE
, &l2cache
, &nl2cache
) == 0 &&
2033 (nv
= spa_nvlist_lookup_by_guid(l2cache
, nl2cache
, guid
)) != NULL
) {
2034 char *nvstr
= fnvlist_lookup_string(nv
, ZPOOL_CONFIG_PATH
);
2035 spa_history_log_internal(spa
, "vdev remove", NULL
,
2036 "%s vdev (%s) %s", spa_name(spa
), VDEV_TYPE_L2CACHE
, nvstr
);
2038 * Cache devices can always be removed.
2040 vd
= spa_lookup_by_guid(spa
, guid
, B_TRUE
);
2041 ev
= spa_event_create(spa
, vd
, NULL
, ESC_ZFS_VDEV_REMOVE_AUX
);
2042 spa_vdev_remove_aux(spa
->spa_l2cache
.sav_config
,
2043 ZPOOL_CONFIG_L2CACHE
, l2cache
, nl2cache
, nv
);
2044 spa_load_l2cache(spa
);
2045 spa
->spa_l2cache
.sav_sync
= B_TRUE
;
2046 } else if (vd
!= NULL
&& vd
->vdev_islog
) {
2048 error
= spa_vdev_remove_log(vd
, &txg
);
2049 } else if (vd
!= NULL
) {
2051 error
= spa_vdev_remove_top(vd
, &txg
);
2054 * There is no vdev of any kind with the specified guid.
2056 error
= SET_ERROR(ENOENT
);
2060 error
= spa_vdev_exit(spa
, NULL
, txg
, error
);
2069 spa_removal_get_stats(spa_t
*spa
, pool_removal_stat_t
*prs
)
2071 prs
->prs_state
= spa
->spa_removing_phys
.sr_state
;
2073 if (prs
->prs_state
== DSS_NONE
)
2074 return (SET_ERROR(ENOENT
));
2076 prs
->prs_removing_vdev
= spa
->spa_removing_phys
.sr_removing_vdev
;
2077 prs
->prs_start_time
= spa
->spa_removing_phys
.sr_start_time
;
2078 prs
->prs_end_time
= spa
->spa_removing_phys
.sr_end_time
;
2079 prs
->prs_to_copy
= spa
->spa_removing_phys
.sr_to_copy
;
2080 prs
->prs_copied
= spa
->spa_removing_phys
.sr_copied
;
2082 if (spa
->spa_vdev_removal
!= NULL
) {
2083 for (int i
= 0; i
< TXG_SIZE
; i
++) {
2085 spa
->spa_vdev_removal
->svr_bytes_done
[i
];
2089 prs
->prs_mapping_memory
= 0;
2090 uint64_t indirect_vdev_id
=
2091 spa
->spa_removing_phys
.sr_prev_indirect_vdev
;
2092 while (indirect_vdev_id
!= -1) {
2093 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[indirect_vdev_id
];
2094 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
2095 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
2097 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
2098 prs
->prs_mapping_memory
+= vdev_indirect_mapping_size(vim
);
2099 indirect_vdev_id
= vic
->vic_prev_indirect_vdev
;
2105 #if defined(_KERNEL)
2106 module_param(zfs_remove_max_segment
, int, 0644);
2107 MODULE_PARM_DESC(zfs_remove_max_segment
,
2108 "Largest contiguous segment to allocate when removing device");
2110 module_param(vdev_removal_max_span
, int, 0644);
2111 MODULE_PARM_DESC(vdev_removal_max_span
,
2112 "Largest span of free chunks a remap segment can span");
2114 EXPORT_SYMBOL(free_from_removing_vdev
);
2115 EXPORT_SYMBOL(spa_removal_get_stats
);
2116 EXPORT_SYMBOL(spa_remove_init
);
2117 EXPORT_SYMBOL(spa_restart_removal
);
2118 EXPORT_SYMBOL(spa_vdev_removal_destroy
);
2119 EXPORT_SYMBOL(spa_vdev_remove
);
2120 EXPORT_SYMBOL(spa_vdev_remove_cancel
);
2121 EXPORT_SYMBOL(spa_vdev_remove_suspend
);
2122 EXPORT_SYMBOL(svr_sync
);