4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
27 #include <sys/zfs_context.h>
28 #include <sys/spa_impl.h>
30 #include <sys/dmu_tx.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/metaslab.h>
34 #include <sys/metaslab_impl.h>
35 #include <sys/uberblock_impl.h>
38 #include <sys/bpobj.h>
39 #include <sys/dsl_pool.h>
40 #include <sys/dsl_synctask.h>
41 #include <sys/dsl_dir.h>
43 #include <sys/zfeature.h>
44 #include <sys/vdev_indirect_births.h>
45 #include <sys/vdev_indirect_mapping.h>
47 #include <sys/trace_vdev.h>
50 * This file contains the necessary logic to remove vdevs from a
51 * storage pool. Currently, the only devices that can be removed
52 * are log, cache, and spare devices; and top level vdevs from a pool
53 * w/o raidz or mirrors. (Note that members of a mirror can be removed
54 * by the detach operation.)
56 * Log vdevs are removed by evacuating them and then turning the vdev
57 * into a hole vdev while holding spa config locks.
59 * Top level vdevs are removed and converted into an indirect vdev via
60 * a multi-step process:
62 * - Disable allocations from this device (spa_vdev_remove_top).
64 * - From a new thread (spa_vdev_remove_thread), copy data from
65 * the removing vdev to a different vdev. The copy happens in open
66 * context (spa_vdev_copy_impl) and issues a sync task
67 * (vdev_mapping_sync) so the sync thread can update the partial
68 * indirect mappings in core and on disk.
70 * - If a free happens during a removal, it is freed from the
71 * removing vdev, and if it has already been copied, from the new
72 * location as well (free_from_removing_vdev).
74 * - After the removal is completed, the copy thread converts the vdev
75 * into an indirect vdev (vdev_remove_complete) before instructing
76 * the sync thread to destroy the space maps and finish the removal
77 * (spa_finish_removal).
80 typedef struct vdev_copy_arg
{
82 uint64_t vca_outstanding_bytes
;
88 * The maximum amount of memory we can use for outstanding i/o while
89 * doing a device removal. This determines how much i/o we can have
90 * in flight concurrently.
92 int zfs_remove_max_copy_bytes
= 64 * 1024 * 1024;
95 * The largest contiguous segment that we will attempt to allocate when
96 * removing a device. This can be no larger than SPA_MAXBLOCKSIZE. If
97 * there is a performance problem with attempting to allocate large blocks,
98 * consider decreasing this.
100 int zfs_remove_max_segment
= SPA_MAXBLOCKSIZE
;
103 * Allow a remap segment to span free chunks of at most this size. The main
104 * impact of a larger span is that we will read and write larger, more
105 * contiguous chunks, with more "unnecessary" data -- trading off bandwidth
106 * for iops. The value here was chosen to align with
107 * zfs_vdev_read_gap_limit, which is a similar concept when doing regular
108 * reads (but there's no reason it has to be the same).
110 * Additionally, a higher span will have the following relatively minor
112 * - the mapping will be smaller, since one entry can cover more allocated
114 * - more of the fragmentation in the removing device will be preserved
115 * - we'll do larger allocations, which may fail and fall back on smaller
118 int vdev_removal_max_span
= 32 * 1024;
121 * This is used by the test suite so that it can ensure that certain
122 * actions happen while in the middle of a removal.
124 unsigned long zfs_remove_max_bytes_pause
= -1UL;
126 #define VDEV_REMOVAL_ZAP_OBJS "lzap"
128 static void spa_vdev_remove_thread(void *arg
);
131 spa_sync_removing_state(spa_t
*spa
, dmu_tx_t
*tx
)
133 VERIFY0(zap_update(spa
->spa_dsl_pool
->dp_meta_objset
,
134 DMU_POOL_DIRECTORY_OBJECT
,
135 DMU_POOL_REMOVING
, sizeof (uint64_t),
136 sizeof (spa
->spa_removing_phys
) / sizeof (uint64_t),
137 &spa
->spa_removing_phys
, tx
));
141 spa_nvlist_lookup_by_guid(nvlist_t
**nvpp
, int count
, uint64_t target_guid
)
143 for (int i
= 0; i
< count
; i
++) {
145 fnvlist_lookup_uint64(nvpp
[i
], ZPOOL_CONFIG_GUID
);
147 if (guid
== target_guid
)
155 spa_vdev_remove_aux(nvlist_t
*config
, char *name
, nvlist_t
**dev
, int count
,
156 nvlist_t
*dev_to_remove
)
158 nvlist_t
**newdev
= NULL
;
161 newdev
= kmem_alloc((count
- 1) * sizeof (void *), KM_SLEEP
);
163 for (int i
= 0, j
= 0; i
< count
; i
++) {
164 if (dev
[i
] == dev_to_remove
)
166 VERIFY(nvlist_dup(dev
[i
], &newdev
[j
++], KM_SLEEP
) == 0);
169 VERIFY(nvlist_remove(config
, name
, DATA_TYPE_NVLIST_ARRAY
) == 0);
170 VERIFY(nvlist_add_nvlist_array(config
, name
, newdev
, count
- 1) == 0);
172 for (int i
= 0; i
< count
- 1; i
++)
173 nvlist_free(newdev
[i
]);
176 kmem_free(newdev
, (count
- 1) * sizeof (void *));
179 static spa_vdev_removal_t
*
180 spa_vdev_removal_create(vdev_t
*vd
)
182 spa_vdev_removal_t
*svr
= kmem_zalloc(sizeof (*svr
), KM_SLEEP
);
183 mutex_init(&svr
->svr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
184 cv_init(&svr
->svr_cv
, NULL
, CV_DEFAULT
, NULL
);
185 svr
->svr_allocd_segs
= range_tree_create(NULL
, NULL
);
186 svr
->svr_vdev_id
= vd
->vdev_id
;
188 for (int i
= 0; i
< TXG_SIZE
; i
++) {
189 svr
->svr_frees
[i
] = range_tree_create(NULL
, NULL
);
190 list_create(&svr
->svr_new_segments
[i
],
191 sizeof (vdev_indirect_mapping_entry_t
),
192 offsetof(vdev_indirect_mapping_entry_t
, vime_node
));
199 spa_vdev_removal_destroy(spa_vdev_removal_t
*svr
)
201 for (int i
= 0; i
< TXG_SIZE
; i
++) {
202 ASSERT0(svr
->svr_bytes_done
[i
]);
203 ASSERT0(svr
->svr_max_offset_to_sync
[i
]);
204 range_tree_destroy(svr
->svr_frees
[i
]);
205 list_destroy(&svr
->svr_new_segments
[i
]);
208 range_tree_destroy(svr
->svr_allocd_segs
);
209 mutex_destroy(&svr
->svr_lock
);
210 cv_destroy(&svr
->svr_cv
);
211 kmem_free(svr
, sizeof (*svr
));
215 * This is called as a synctask in the txg in which we will mark this vdev
216 * as removing (in the config stored in the MOS).
218 * It begins the evacuation of a toplevel vdev by:
219 * - initializing the spa_removing_phys which tracks this removal
220 * - computing the amount of space to remove for accounting purposes
221 * - dirtying all dbufs in the spa_config_object
222 * - creating the spa_vdev_removal
223 * - starting the spa_vdev_remove_thread
226 vdev_remove_initiate_sync(void *arg
, dmu_tx_t
*tx
)
228 int vdev_id
= (uintptr_t)arg
;
229 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
230 vdev_t
*vd
= vdev_lookup_top(spa
, vdev_id
);
231 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
232 objset_t
*mos
= spa
->spa_dsl_pool
->dp_meta_objset
;
233 spa_vdev_removal_t
*svr
= NULL
;
234 ASSERTV(uint64_t txg
= dmu_tx_get_txg(tx
));
236 ASSERT3P(vd
->vdev_ops
, !=, &vdev_raidz_ops
);
237 svr
= spa_vdev_removal_create(vd
);
239 ASSERT(vd
->vdev_removing
);
240 ASSERT3P(vd
->vdev_indirect_mapping
, ==, NULL
);
242 spa_feature_incr(spa
, SPA_FEATURE_DEVICE_REMOVAL
, tx
);
243 if (spa_feature_is_enabled(spa
, SPA_FEATURE_OBSOLETE_COUNTS
)) {
245 * By activating the OBSOLETE_COUNTS feature, we prevent
246 * the pool from being downgraded and ensure that the
247 * refcounts are precise.
249 spa_feature_incr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
251 VERIFY0(zap_add(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
252 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE
, sizeof (one
), 1,
254 ASSERT3U(vdev_obsolete_counts_are_precise(vd
), !=, 0);
257 vic
->vic_mapping_object
= vdev_indirect_mapping_alloc(mos
, tx
);
258 vd
->vdev_indirect_mapping
=
259 vdev_indirect_mapping_open(mos
, vic
->vic_mapping_object
);
260 vic
->vic_births_object
= vdev_indirect_births_alloc(mos
, tx
);
261 vd
->vdev_indirect_births
=
262 vdev_indirect_births_open(mos
, vic
->vic_births_object
);
263 spa
->spa_removing_phys
.sr_removing_vdev
= vd
->vdev_id
;
264 spa
->spa_removing_phys
.sr_start_time
= gethrestime_sec();
265 spa
->spa_removing_phys
.sr_end_time
= 0;
266 spa
->spa_removing_phys
.sr_state
= DSS_SCANNING
;
267 spa
->spa_removing_phys
.sr_to_copy
= 0;
268 spa
->spa_removing_phys
.sr_copied
= 0;
271 * Note: We can't use vdev_stat's vs_alloc for sr_to_copy, because
272 * there may be space in the defer tree, which is free, but still
273 * counted in vs_alloc.
275 for (uint64_t i
= 0; i
< vd
->vdev_ms_count
; i
++) {
276 metaslab_t
*ms
= vd
->vdev_ms
[i
];
277 if (ms
->ms_sm
== NULL
)
281 * Sync tasks happen before metaslab_sync(), therefore
282 * smp_alloc and sm_alloc must be the same.
284 ASSERT3U(space_map_allocated(ms
->ms_sm
), ==,
285 ms
->ms_sm
->sm_phys
->smp_alloc
);
287 spa
->spa_removing_phys
.sr_to_copy
+=
288 space_map_allocated(ms
->ms_sm
);
291 * Space which we are freeing this txg does not need to
294 spa
->spa_removing_phys
.sr_to_copy
-=
295 range_tree_space(ms
->ms_freeing
);
297 ASSERT0(range_tree_space(ms
->ms_freed
));
298 for (int t
= 0; t
< TXG_SIZE
; t
++)
299 ASSERT0(range_tree_space(ms
->ms_allocating
[t
]));
303 * Sync tasks are called before metaslab_sync(), so there should
304 * be no already-synced metaslabs in the TXG_CLEAN list.
306 ASSERT3P(txg_list_head(&vd
->vdev_ms_list
, TXG_CLEAN(txg
)), ==, NULL
);
308 spa_sync_removing_state(spa
, tx
);
311 * All blocks that we need to read the most recent mapping must be
312 * stored on concrete vdevs. Therefore, we must dirty anything that
313 * is read before spa_remove_init(). Specifically, the
314 * spa_config_object. (Note that although we already modified the
315 * spa_config_object in spa_sync_removing_state, that may not have
316 * modified all blocks of the object.)
318 dmu_object_info_t doi
;
319 VERIFY0(dmu_object_info(mos
, DMU_POOL_DIRECTORY_OBJECT
, &doi
));
320 for (uint64_t offset
= 0; offset
< doi
.doi_max_offset
; ) {
322 VERIFY0(dmu_buf_hold(mos
, DMU_POOL_DIRECTORY_OBJECT
,
323 offset
, FTAG
, &dbuf
, 0));
324 dmu_buf_will_dirty(dbuf
, tx
);
325 offset
+= dbuf
->db_size
;
326 dmu_buf_rele(dbuf
, FTAG
);
330 * Now that we've allocated the im_object, dirty the vdev to ensure
331 * that the object gets written to the config on disk.
333 vdev_config_dirty(vd
);
335 zfs_dbgmsg("starting removal thread for vdev %llu (%p) in txg %llu "
336 "im_obj=%llu", vd
->vdev_id
, vd
, dmu_tx_get_txg(tx
),
337 vic
->vic_mapping_object
);
339 spa_history_log_internal(spa
, "vdev remove started", tx
,
340 "%s vdev %llu %s", spa_name(spa
), vd
->vdev_id
,
341 (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
343 * Setting spa_vdev_removal causes subsequent frees to call
344 * free_from_removing_vdev(). Note that we don't need any locking
345 * because we are the sync thread, and metaslab_free_impl() is only
346 * called from syncing context (potentially from a zio taskq thread,
347 * but in any case only when there are outstanding free i/os, which
350 ASSERT3P(spa
->spa_vdev_removal
, ==, NULL
);
351 spa
->spa_vdev_removal
= svr
;
352 svr
->svr_thread
= thread_create(NULL
, 0,
353 spa_vdev_remove_thread
, spa
, 0, &p0
, TS_RUN
, minclsyspri
);
357 * When we are opening a pool, we must read the mapping for each
358 * indirect vdev in order from most recently removed to least
359 * recently removed. We do this because the blocks for the mapping
360 * of older indirect vdevs may be stored on more recently removed vdevs.
361 * In order to read each indirect mapping object, we must have
362 * initialized all more recently removed vdevs.
365 spa_remove_init(spa_t
*spa
)
369 error
= zap_lookup(spa
->spa_dsl_pool
->dp_meta_objset
,
370 DMU_POOL_DIRECTORY_OBJECT
,
371 DMU_POOL_REMOVING
, sizeof (uint64_t),
372 sizeof (spa
->spa_removing_phys
) / sizeof (uint64_t),
373 &spa
->spa_removing_phys
);
375 if (error
== ENOENT
) {
376 spa
->spa_removing_phys
.sr_state
= DSS_NONE
;
377 spa
->spa_removing_phys
.sr_removing_vdev
= -1;
378 spa
->spa_removing_phys
.sr_prev_indirect_vdev
= -1;
379 spa
->spa_indirect_vdevs_loaded
= B_TRUE
;
381 } else if (error
!= 0) {
385 if (spa
->spa_removing_phys
.sr_state
== DSS_SCANNING
) {
387 * We are currently removing a vdev. Create and
388 * initialize a spa_vdev_removal_t from the bonus
389 * buffer of the removing vdevs vdev_im_object, and
390 * initialize its partial mapping.
392 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
393 vdev_t
*vd
= vdev_lookup_top(spa
,
394 spa
->spa_removing_phys
.sr_removing_vdev
);
397 spa_config_exit(spa
, SCL_STATE
, FTAG
);
401 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
403 ASSERT(vdev_is_concrete(vd
));
404 spa_vdev_removal_t
*svr
= spa_vdev_removal_create(vd
);
405 ASSERT3U(svr
->svr_vdev_id
, ==, vd
->vdev_id
);
406 ASSERT(vd
->vdev_removing
);
408 vd
->vdev_indirect_mapping
= vdev_indirect_mapping_open(
409 spa
->spa_meta_objset
, vic
->vic_mapping_object
);
410 vd
->vdev_indirect_births
= vdev_indirect_births_open(
411 spa
->spa_meta_objset
, vic
->vic_births_object
);
412 spa_config_exit(spa
, SCL_STATE
, FTAG
);
414 spa
->spa_vdev_removal
= svr
;
417 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
418 uint64_t indirect_vdev_id
=
419 spa
->spa_removing_phys
.sr_prev_indirect_vdev
;
420 while (indirect_vdev_id
!= UINT64_MAX
) {
421 vdev_t
*vd
= vdev_lookup_top(spa
, indirect_vdev_id
);
422 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
424 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
425 vd
->vdev_indirect_mapping
= vdev_indirect_mapping_open(
426 spa
->spa_meta_objset
, vic
->vic_mapping_object
);
427 vd
->vdev_indirect_births
= vdev_indirect_births_open(
428 spa
->spa_meta_objset
, vic
->vic_births_object
);
430 indirect_vdev_id
= vic
->vic_prev_indirect_vdev
;
432 spa_config_exit(spa
, SCL_STATE
, FTAG
);
435 * Now that we've loaded all the indirect mappings, we can allow
436 * reads from other blocks (e.g. via predictive prefetch).
438 spa
->spa_indirect_vdevs_loaded
= B_TRUE
;
443 spa_restart_removal(spa_t
*spa
)
445 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
451 * In general when this function is called there is no
452 * removal thread running. The only scenario where this
453 * is not true is during spa_import() where this function
454 * is called twice [once from spa_import_impl() and
455 * spa_async_resume()]. Thus, in the scenario where we
456 * import a pool that has an ongoing removal we don't
457 * want to spawn a second thread.
459 if (svr
->svr_thread
!= NULL
)
462 if (!spa_writeable(spa
))
465 zfs_dbgmsg("restarting removal of %llu", svr
->svr_vdev_id
);
466 svr
->svr_thread
= thread_create(NULL
, 0, spa_vdev_remove_thread
, spa
,
467 0, &p0
, TS_RUN
, minclsyspri
);
471 * Process freeing from a device which is in the middle of being removed.
472 * We must handle this carefully so that we attempt to copy freed data,
473 * and we correctly free already-copied data.
476 free_from_removing_vdev(vdev_t
*vd
, uint64_t offset
, uint64_t size
)
478 spa_t
*spa
= vd
->vdev_spa
;
479 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
480 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
481 uint64_t txg
= spa_syncing_txg(spa
);
482 uint64_t max_offset_yet
= 0;
484 ASSERT(vd
->vdev_indirect_config
.vic_mapping_object
!= 0);
485 ASSERT3U(vd
->vdev_indirect_config
.vic_mapping_object
, ==,
486 vdev_indirect_mapping_object(vim
));
487 ASSERT3U(vd
->vdev_id
, ==, svr
->svr_vdev_id
);
489 mutex_enter(&svr
->svr_lock
);
492 * Remove the segment from the removing vdev's spacemap. This
493 * ensures that we will not attempt to copy this space (if the
494 * removal thread has not yet visited it), and also ensures
495 * that we know what is actually allocated on the new vdevs
496 * (needed if we cancel the removal).
498 * Note: we must do the metaslab_free_concrete() with the svr_lock
499 * held, so that the remove_thread can not load this metaslab and then
500 * visit this offset between the time that we metaslab_free_concrete()
501 * and when we check to see if it has been visited.
503 * Note: The checkpoint flag is set to false as having/taking
504 * a checkpoint and removing a device can't happen at the same
507 ASSERT(!spa_has_checkpoint(spa
));
508 metaslab_free_concrete(vd
, offset
, size
, B_FALSE
);
510 uint64_t synced_size
= 0;
511 uint64_t synced_offset
= 0;
512 uint64_t max_offset_synced
= vdev_indirect_mapping_max_offset(vim
);
513 if (offset
< max_offset_synced
) {
515 * The mapping for this offset is already on disk.
516 * Free from the new location.
518 * Note that we use svr_max_synced_offset because it is
519 * updated atomically with respect to the in-core mapping.
520 * By contrast, vim_max_offset is not.
522 * This block may be split between a synced entry and an
523 * in-flight or unvisited entry. Only process the synced
524 * portion of it here.
526 synced_size
= MIN(size
, max_offset_synced
- offset
);
527 synced_offset
= offset
;
529 ASSERT3U(max_offset_yet
, <=, max_offset_synced
);
530 max_offset_yet
= max_offset_synced
;
532 DTRACE_PROBE3(remove__free__synced
,
535 uint64_t, synced_size
);
538 offset
+= synced_size
;
542 * Look at all in-flight txgs starting from the currently syncing one
543 * and see if a section of this free is being copied. By starting from
544 * this txg and iterating forward, we might find that this region
545 * was copied in two different txgs and handle it appropriately.
547 for (int i
= 0; i
< TXG_CONCURRENT_STATES
; i
++) {
548 int txgoff
= (txg
+ i
) & TXG_MASK
;
549 if (size
> 0 && offset
< svr
->svr_max_offset_to_sync
[txgoff
]) {
551 * The mapping for this offset is in flight, and
552 * will be synced in txg+i.
554 uint64_t inflight_size
= MIN(size
,
555 svr
->svr_max_offset_to_sync
[txgoff
] - offset
);
557 DTRACE_PROBE4(remove__free__inflight
,
560 uint64_t, inflight_size
,
564 * We copy data in order of increasing offset.
565 * Therefore the max_offset_to_sync[] must increase
566 * (or be zero, indicating that nothing is being
567 * copied in that txg).
569 if (svr
->svr_max_offset_to_sync
[txgoff
] != 0) {
570 ASSERT3U(svr
->svr_max_offset_to_sync
[txgoff
],
573 svr
->svr_max_offset_to_sync
[txgoff
];
577 * We've already committed to copying this segment:
578 * we have allocated space elsewhere in the pool for
579 * it and have an IO outstanding to copy the data. We
580 * cannot free the space before the copy has
581 * completed, or else the copy IO might overwrite any
582 * new data. To free that space, we record the
583 * segment in the appropriate svr_frees tree and free
584 * the mapped space later, in the txg where we have
585 * completed the copy and synced the mapping (see
586 * vdev_mapping_sync).
588 range_tree_add(svr
->svr_frees
[txgoff
],
589 offset
, inflight_size
);
590 size
-= inflight_size
;
591 offset
+= inflight_size
;
594 * This space is already accounted for as being
595 * done, because it is being copied in txg+i.
596 * However, if i!=0, then it is being copied in
597 * a future txg. If we crash after this txg
598 * syncs but before txg+i syncs, then the space
599 * will be free. Therefore we must account
600 * for the space being done in *this* txg
601 * (when it is freed) rather than the future txg
602 * (when it will be copied).
604 ASSERT3U(svr
->svr_bytes_done
[txgoff
], >=,
606 svr
->svr_bytes_done
[txgoff
] -= inflight_size
;
607 svr
->svr_bytes_done
[txg
& TXG_MASK
] += inflight_size
;
610 ASSERT0(svr
->svr_max_offset_to_sync
[TXG_CLEAN(txg
) & TXG_MASK
]);
614 * The copy thread has not yet visited this offset. Ensure
618 DTRACE_PROBE3(remove__free__unvisited
,
623 if (svr
->svr_allocd_segs
!= NULL
)
624 range_tree_clear(svr
->svr_allocd_segs
, offset
, size
);
627 * Since we now do not need to copy this data, for
628 * accounting purposes we have done our job and can count
631 svr
->svr_bytes_done
[txg
& TXG_MASK
] += size
;
633 mutex_exit(&svr
->svr_lock
);
636 * Now that we have dropped svr_lock, process the synced portion
639 if (synced_size
> 0) {
640 vdev_indirect_mark_obsolete(vd
, synced_offset
, synced_size
);
643 * Note: this can only be called from syncing context,
644 * and the vdev_indirect_mapping is only changed from the
645 * sync thread, so we don't need svr_lock while doing
646 * metaslab_free_impl_cb.
648 boolean_t checkpoint
= B_FALSE
;
649 vdev_indirect_ops
.vdev_op_remap(vd
, synced_offset
, synced_size
,
650 metaslab_free_impl_cb
, &checkpoint
);
655 * Stop an active removal and update the spa_removing phys.
658 spa_finish_removal(spa_t
*spa
, dsl_scan_state_t state
, dmu_tx_t
*tx
)
660 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
661 ASSERT3U(dmu_tx_get_txg(tx
), ==, spa_syncing_txg(spa
));
663 /* Ensure the removal thread has completed before we free the svr. */
664 spa_vdev_remove_suspend(spa
);
666 ASSERT(state
== DSS_FINISHED
|| state
== DSS_CANCELED
);
668 if (state
== DSS_FINISHED
) {
669 spa_removing_phys_t
*srp
= &spa
->spa_removing_phys
;
670 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
671 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
673 if (srp
->sr_prev_indirect_vdev
!= UINT64_MAX
) {
675 pvd
= vdev_lookup_top(spa
,
676 srp
->sr_prev_indirect_vdev
);
677 ASSERT3P(pvd
->vdev_ops
, ==, &vdev_indirect_ops
);
680 vic
->vic_prev_indirect_vdev
= srp
->sr_prev_indirect_vdev
;
681 srp
->sr_prev_indirect_vdev
= vd
->vdev_id
;
683 spa
->spa_removing_phys
.sr_state
= state
;
684 spa
->spa_removing_phys
.sr_end_time
= gethrestime_sec();
686 spa
->spa_vdev_removal
= NULL
;
687 spa_vdev_removal_destroy(svr
);
689 spa_sync_removing_state(spa
, tx
);
691 vdev_config_dirty(spa
->spa_root_vdev
);
695 free_mapped_segment_cb(void *arg
, uint64_t offset
, uint64_t size
)
698 vdev_indirect_mark_obsolete(vd
, offset
, size
);
699 boolean_t checkpoint
= B_FALSE
;
700 vdev_indirect_ops
.vdev_op_remap(vd
, offset
, size
,
701 metaslab_free_impl_cb
, &checkpoint
);
705 * On behalf of the removal thread, syncs an incremental bit more of
706 * the indirect mapping to disk and updates the in-memory mapping.
707 * Called as a sync task in every txg that the removal thread makes progress.
710 vdev_mapping_sync(void *arg
, dmu_tx_t
*tx
)
712 spa_vdev_removal_t
*svr
= arg
;
713 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
714 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
715 ASSERTV(vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
);
716 uint64_t txg
= dmu_tx_get_txg(tx
);
717 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
719 ASSERT(vic
->vic_mapping_object
!= 0);
720 ASSERT3U(txg
, ==, spa_syncing_txg(spa
));
722 vdev_indirect_mapping_add_entries(vim
,
723 &svr
->svr_new_segments
[txg
& TXG_MASK
], tx
);
724 vdev_indirect_births_add_entry(vd
->vdev_indirect_births
,
725 vdev_indirect_mapping_max_offset(vim
), dmu_tx_get_txg(tx
), tx
);
728 * Free the copied data for anything that was freed while the
729 * mapping entries were in flight.
731 mutex_enter(&svr
->svr_lock
);
732 range_tree_vacate(svr
->svr_frees
[txg
& TXG_MASK
],
733 free_mapped_segment_cb
, vd
);
734 ASSERT3U(svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
], >=,
735 vdev_indirect_mapping_max_offset(vim
));
736 svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] = 0;
737 mutex_exit(&svr
->svr_lock
);
739 spa_sync_removing_state(spa
, tx
);
742 typedef struct vdev_copy_segment_arg
{
744 dva_t
*vcsa_dest_dva
;
746 range_tree_t
*vcsa_obsolete_segs
;
747 } vdev_copy_segment_arg_t
;
750 unalloc_seg(void *arg
, uint64_t start
, uint64_t size
)
752 vdev_copy_segment_arg_t
*vcsa
= arg
;
753 spa_t
*spa
= vcsa
->vcsa_spa
;
754 blkptr_t bp
= { { { {0} } } };
756 BP_SET_BIRTH(&bp
, TXG_INITIAL
, TXG_INITIAL
);
757 BP_SET_LSIZE(&bp
, size
);
758 BP_SET_PSIZE(&bp
, size
);
759 BP_SET_COMPRESS(&bp
, ZIO_COMPRESS_OFF
);
760 BP_SET_CHECKSUM(&bp
, ZIO_CHECKSUM_OFF
);
761 BP_SET_TYPE(&bp
, DMU_OT_NONE
);
762 BP_SET_LEVEL(&bp
, 0);
763 BP_SET_DEDUP(&bp
, 0);
764 BP_SET_BYTEORDER(&bp
, ZFS_HOST_BYTEORDER
);
766 DVA_SET_VDEV(&bp
.blk_dva
[0], DVA_GET_VDEV(vcsa
->vcsa_dest_dva
));
767 DVA_SET_OFFSET(&bp
.blk_dva
[0],
768 DVA_GET_OFFSET(vcsa
->vcsa_dest_dva
) + start
);
769 DVA_SET_ASIZE(&bp
.blk_dva
[0], size
);
771 zio_free(spa
, vcsa
->vcsa_txg
, &bp
);
775 * All reads and writes associated with a call to spa_vdev_copy_segment()
779 spa_vdev_copy_segment_done(zio_t
*zio
)
781 vdev_copy_segment_arg_t
*vcsa
= zio
->io_private
;
783 range_tree_vacate(vcsa
->vcsa_obsolete_segs
,
785 range_tree_destroy(vcsa
->vcsa_obsolete_segs
);
786 kmem_free(vcsa
, sizeof (*vcsa
));
788 spa_config_exit(zio
->io_spa
, SCL_STATE
, zio
->io_spa
);
792 * The write of the new location is done.
795 spa_vdev_copy_segment_write_done(zio_t
*zio
)
797 vdev_copy_arg_t
*vca
= zio
->io_private
;
799 abd_free(zio
->io_abd
);
801 mutex_enter(&vca
->vca_lock
);
802 vca
->vca_outstanding_bytes
-= zio
->io_size
;
803 cv_signal(&vca
->vca_cv
);
804 mutex_exit(&vca
->vca_lock
);
808 * The read of the old location is done. The parent zio is the write to
809 * the new location. Allow it to start.
812 spa_vdev_copy_segment_read_done(zio_t
*zio
)
814 zio_nowait(zio_unique_parent(zio
));
818 * If the old and new vdevs are mirrors, we will read both sides of the old
819 * mirror, and write each copy to the corresponding side of the new mirror.
820 * If the old and new vdevs have a different number of children, we will do
821 * this as best as possible. Since we aren't verifying checksums, this
822 * ensures that as long as there's a good copy of the data, we'll have a
823 * good copy after the removal, even if there's silent damage to one side
824 * of the mirror. If we're removing a mirror that has some silent damage,
825 * we'll have exactly the same damage in the new location (assuming that
826 * the new location is also a mirror).
828 * We accomplish this by creating a tree of zio_t's, with as many writes as
829 * there are "children" of the new vdev (a non-redundant vdev counts as one
830 * child, a 2-way mirror has 2 children, etc). Each write has an associated
831 * read from a child of the old vdev. Typically there will be the same
832 * number of children of the old and new vdevs. However, if there are more
833 * children of the new vdev, some child(ren) of the old vdev will be issued
834 * multiple reads. If there are more children of the old vdev, some copies
837 * For example, the tree of zio_t's for a 2-way mirror is:
841 * write(new vdev, child 0) write(new vdev, child 1)
843 * read(old vdev, child 0) read(old vdev, child 1)
845 * Child zio's complete before their parents complete. However, zio's
846 * created with zio_vdev_child_io() may be issued before their children
847 * complete. In this case we need to make sure that the children (reads)
848 * complete before the parents (writes) are *issued*. We do this by not
849 * calling zio_nowait() on each write until its corresponding read has
852 * The spa_config_lock must be held while zio's created by
853 * zio_vdev_child_io() are in progress, to ensure that the vdev tree does
854 * not change (e.g. due to a concurrent "zpool attach/detach"). The "null"
855 * zio is needed to release the spa_config_lock after all the reads and
856 * writes complete. (Note that we can't grab the config lock for each read,
857 * because it is not reentrant - we could deadlock with a thread waiting
861 spa_vdev_copy_one_child(vdev_copy_arg_t
*vca
, zio_t
*nzio
,
862 vdev_t
*source_vd
, uint64_t source_offset
,
863 vdev_t
*dest_child_vd
, uint64_t dest_offset
, int dest_id
, uint64_t size
)
865 ASSERT3U(spa_config_held(nzio
->io_spa
, SCL_ALL
, RW_READER
), !=, 0);
867 mutex_enter(&vca
->vca_lock
);
868 vca
->vca_outstanding_bytes
+= size
;
869 mutex_exit(&vca
->vca_lock
);
871 abd_t
*abd
= abd_alloc_for_io(size
, B_FALSE
);
873 vdev_t
*source_child_vd
;
874 if (source_vd
->vdev_ops
== &vdev_mirror_ops
&& dest_id
!= -1) {
876 * Source and dest are both mirrors. Copy from the same
877 * child id as we are copying to (wrapping around if there
878 * are more dest children than source children).
881 source_vd
->vdev_child
[dest_id
% source_vd
->vdev_children
];
883 source_child_vd
= source_vd
;
886 zio_t
*write_zio
= zio_vdev_child_io(nzio
, NULL
,
887 dest_child_vd
, dest_offset
, abd
, size
,
888 ZIO_TYPE_WRITE
, ZIO_PRIORITY_REMOVAL
,
890 spa_vdev_copy_segment_write_done
, vca
);
892 zio_nowait(zio_vdev_child_io(write_zio
, NULL
,
893 source_child_vd
, source_offset
, abd
, size
,
894 ZIO_TYPE_READ
, ZIO_PRIORITY_REMOVAL
,
896 spa_vdev_copy_segment_read_done
, vca
));
900 * Allocate a new location for this segment, and create the zio_t's to
901 * read from the old location and write to the new location.
904 spa_vdev_copy_segment(vdev_t
*vd
, range_tree_t
*segs
,
905 uint64_t maxalloc
, uint64_t txg
,
906 vdev_copy_arg_t
*vca
, zio_alloc_list_t
*zal
)
908 metaslab_group_t
*mg
= vd
->vdev_mg
;
909 spa_t
*spa
= vd
->vdev_spa
;
910 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
911 vdev_indirect_mapping_entry_t
*entry
;
913 uint64_t start
= range_tree_min(segs
);
915 ASSERT3U(maxalloc
, <=, SPA_MAXBLOCKSIZE
);
917 uint64_t size
= range_tree_span(segs
);
918 if (range_tree_span(segs
) > maxalloc
) {
920 * We can't allocate all the segments. Prefer to end
921 * the allocation at the end of a segment, thus avoiding
922 * additional split blocks.
926 search
.rs_start
= start
+ maxalloc
;
927 search
.rs_end
= search
.rs_start
;
928 range_seg_t
*rs
= avl_find(&segs
->rt_root
, &search
, &where
);
930 rs
= avl_nearest(&segs
->rt_root
, where
, AVL_BEFORE
);
932 rs
= AVL_PREV(&segs
->rt_root
, rs
);
935 size
= rs
->rs_end
- start
;
938 * There are no segments that end before maxalloc.
939 * I.e. the first segment is larger than maxalloc,
940 * so we must split it.
945 ASSERT3U(size
, <=, maxalloc
);
948 * An allocation class might not have any remaining vdevs or space
950 metaslab_class_t
*mc
= mg
->mg_class
;
951 if (mc
!= spa_normal_class(spa
) && mc
->mc_groups
<= 1)
952 mc
= spa_normal_class(spa
);
953 int error
= metaslab_alloc_dva(spa
, mc
, size
, &dst
, 0, NULL
, txg
, 0,
955 if (error
== ENOSPC
&& mc
!= spa_normal_class(spa
)) {
956 error
= metaslab_alloc_dva(spa
, spa_normal_class(spa
), size
,
957 &dst
, 0, NULL
, txg
, 0, zal
, 0);
963 * Determine the ranges that are not actually needed. Offsets are
964 * relative to the start of the range to be copied (i.e. relative to the
965 * local variable "start").
967 range_tree_t
*obsolete_segs
= range_tree_create(NULL
, NULL
);
969 range_seg_t
*rs
= avl_first(&segs
->rt_root
);
970 ASSERT3U(rs
->rs_start
, ==, start
);
971 uint64_t prev_seg_end
= rs
->rs_end
;
972 while ((rs
= AVL_NEXT(&segs
->rt_root
, rs
)) != NULL
) {
973 if (rs
->rs_start
>= start
+ size
) {
976 range_tree_add(obsolete_segs
,
977 prev_seg_end
- start
,
978 rs
->rs_start
- prev_seg_end
);
980 prev_seg_end
= rs
->rs_end
;
982 /* We don't end in the middle of an obsolete range */
983 ASSERT3U(start
+ size
, <=, prev_seg_end
);
985 range_tree_clear(segs
, start
, size
);
988 * We can't have any padding of the allocated size, otherwise we will
989 * misunderstand what's allocated, and the size of the mapping.
990 * The caller ensures this will be true by passing in a size that is
991 * aligned to the worst (highest) ashift in the pool.
993 ASSERT3U(DVA_GET_ASIZE(&dst
), ==, size
);
995 entry
= kmem_zalloc(sizeof (vdev_indirect_mapping_entry_t
), KM_SLEEP
);
996 DVA_MAPPING_SET_SRC_OFFSET(&entry
->vime_mapping
, start
);
997 entry
->vime_mapping
.vimep_dst
= dst
;
998 if (spa_feature_is_enabled(spa
, SPA_FEATURE_OBSOLETE_COUNTS
)) {
999 entry
->vime_obsolete_count
= range_tree_space(obsolete_segs
);
1002 vdev_copy_segment_arg_t
*vcsa
= kmem_zalloc(sizeof (*vcsa
), KM_SLEEP
);
1003 vcsa
->vcsa_dest_dva
= &entry
->vime_mapping
.vimep_dst
;
1004 vcsa
->vcsa_obsolete_segs
= obsolete_segs
;
1005 vcsa
->vcsa_spa
= spa
;
1006 vcsa
->vcsa_txg
= txg
;
1009 * See comment before spa_vdev_copy_one_child().
1011 spa_config_enter(spa
, SCL_STATE
, spa
, RW_READER
);
1012 zio_t
*nzio
= zio_null(spa
->spa_txg_zio
[txg
& TXG_MASK
], spa
, NULL
,
1013 spa_vdev_copy_segment_done
, vcsa
, 0);
1014 vdev_t
*dest_vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dst
));
1015 if (dest_vd
->vdev_ops
== &vdev_mirror_ops
) {
1016 for (int i
= 0; i
< dest_vd
->vdev_children
; i
++) {
1017 vdev_t
*child
= dest_vd
->vdev_child
[i
];
1018 spa_vdev_copy_one_child(vca
, nzio
, vd
, start
,
1019 child
, DVA_GET_OFFSET(&dst
), i
, size
);
1022 spa_vdev_copy_one_child(vca
, nzio
, vd
, start
,
1023 dest_vd
, DVA_GET_OFFSET(&dst
), -1, size
);
1027 list_insert_tail(&svr
->svr_new_segments
[txg
& TXG_MASK
], entry
);
1028 ASSERT3U(start
+ size
, <=, vd
->vdev_ms_count
<< vd
->vdev_ms_shift
);
1029 vdev_dirty(vd
, 0, NULL
, txg
);
1035 * Complete the removal of a toplevel vdev. This is called as a
1036 * synctask in the same txg that we will sync out the new config (to the
1037 * MOS object) which indicates that this vdev is indirect.
1040 vdev_remove_complete_sync(void *arg
, dmu_tx_t
*tx
)
1042 spa_vdev_removal_t
*svr
= arg
;
1043 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1044 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1046 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
1048 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1049 ASSERT0(svr
->svr_bytes_done
[i
]);
1052 ASSERT3U(spa
->spa_removing_phys
.sr_copied
, ==,
1053 spa
->spa_removing_phys
.sr_to_copy
);
1055 vdev_destroy_spacemaps(vd
, tx
);
1057 /* destroy leaf zaps, if any */
1058 ASSERT3P(svr
->svr_zaplist
, !=, NULL
);
1059 for (nvpair_t
*pair
= nvlist_next_nvpair(svr
->svr_zaplist
, NULL
);
1061 pair
= nvlist_next_nvpair(svr
->svr_zaplist
, pair
)) {
1062 vdev_destroy_unlink_zap(vd
, fnvpair_value_uint64(pair
), tx
);
1064 fnvlist_free(svr
->svr_zaplist
);
1066 spa_finish_removal(dmu_tx_pool(tx
)->dp_spa
, DSS_FINISHED
, tx
);
1067 /* vd->vdev_path is not available here */
1068 spa_history_log_internal(spa
, "vdev remove completed", tx
,
1069 "%s vdev %llu", spa_name(spa
), vd
->vdev_id
);
1073 vdev_remove_enlist_zaps(vdev_t
*vd
, nvlist_t
*zlist
)
1075 ASSERT3P(zlist
, !=, NULL
);
1076 ASSERT3P(vd
->vdev_ops
, !=, &vdev_raidz_ops
);
1078 if (vd
->vdev_leaf_zap
!= 0) {
1080 (void) snprintf(zkey
, sizeof (zkey
), "%s-%llu",
1081 VDEV_REMOVAL_ZAP_OBJS
, (u_longlong_t
)vd
->vdev_leaf_zap
);
1082 fnvlist_add_uint64(zlist
, zkey
, vd
->vdev_leaf_zap
);
1085 for (uint64_t id
= 0; id
< vd
->vdev_children
; id
++) {
1086 vdev_remove_enlist_zaps(vd
->vdev_child
[id
], zlist
);
1091 vdev_remove_replace_with_indirect(vdev_t
*vd
, uint64_t txg
)
1095 spa_t
*spa
= vd
->vdev_spa
;
1096 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1099 * First, build a list of leaf zaps to be destroyed.
1100 * This is passed to the sync context thread,
1101 * which does the actual unlinking.
1103 svr
->svr_zaplist
= fnvlist_alloc();
1104 vdev_remove_enlist_zaps(vd
, svr
->svr_zaplist
);
1106 ivd
= vdev_add_parent(vd
, &vdev_indirect_ops
);
1107 ivd
->vdev_removing
= 0;
1109 vd
->vdev_leaf_zap
= 0;
1111 vdev_remove_child(ivd
, vd
);
1112 vdev_compact_children(ivd
);
1114 ASSERT(!list_link_active(&vd
->vdev_state_dirty_node
));
1116 tx
= dmu_tx_create_assigned(spa
->spa_dsl_pool
, txg
);
1117 dsl_sync_task_nowait(spa
->spa_dsl_pool
, vdev_remove_complete_sync
, svr
,
1118 0, ZFS_SPACE_CHECK_NONE
, tx
);
1122 * Indicate that this thread has exited.
1123 * After this, we can not use svr.
1125 mutex_enter(&svr
->svr_lock
);
1126 svr
->svr_thread
= NULL
;
1127 cv_broadcast(&svr
->svr_cv
);
1128 mutex_exit(&svr
->svr_lock
);
1132 * Complete the removal of a toplevel vdev. This is called in open
1133 * context by the removal thread after we have copied all vdev's data.
1136 vdev_remove_complete(spa_t
*spa
)
1141 * Wait for any deferred frees to be synced before we call
1142 * vdev_metaslab_fini()
1144 txg_wait_synced(spa
->spa_dsl_pool
, 0);
1145 txg
= spa_vdev_enter(spa
);
1146 vdev_t
*vd
= vdev_lookup_top(spa
, spa
->spa_vdev_removal
->svr_vdev_id
);
1148 sysevent_t
*ev
= spa_event_create(spa
, vd
, NULL
,
1149 ESC_ZFS_VDEV_REMOVE_DEV
);
1151 zfs_dbgmsg("finishing device removal for vdev %llu in txg %llu",
1155 * Discard allocation state.
1157 if (vd
->vdev_mg
!= NULL
) {
1158 vdev_metaslab_fini(vd
);
1159 metaslab_group_destroy(vd
->vdev_mg
);
1162 ASSERT0(vd
->vdev_stat
.vs_space
);
1163 ASSERT0(vd
->vdev_stat
.vs_dspace
);
1165 vdev_remove_replace_with_indirect(vd
, txg
);
1168 * We now release the locks, allowing spa_sync to run and finish the
1169 * removal via vdev_remove_complete_sync in syncing context.
1171 * Note that we hold on to the vdev_t that has been replaced. Since
1172 * it isn't part of the vdev tree any longer, it can't be concurrently
1173 * manipulated, even while we don't have the config lock.
1175 (void) spa_vdev_exit(spa
, NULL
, txg
, 0);
1178 * Top ZAP should have been transferred to the indirect vdev in
1179 * vdev_remove_replace_with_indirect.
1181 ASSERT0(vd
->vdev_top_zap
);
1184 * Leaf ZAP should have been moved in vdev_remove_replace_with_indirect.
1186 ASSERT0(vd
->vdev_leaf_zap
);
1188 txg
= spa_vdev_enter(spa
);
1189 (void) vdev_label_init(vd
, 0, VDEV_LABEL_REMOVE
);
1191 * Request to update the config and the config cachefile.
1193 vdev_config_dirty(spa
->spa_root_vdev
);
1194 (void) spa_vdev_exit(spa
, vd
, txg
, 0);
1201 * Evacuates a segment of size at most max_alloc from the vdev
1202 * via repeated calls to spa_vdev_copy_segment. If an allocation
1203 * fails, the pool is probably too fragmented to handle such a
1204 * large size, so decrease max_alloc so that the caller will not try
1205 * this size again this txg.
1208 spa_vdev_copy_impl(vdev_t
*vd
, spa_vdev_removal_t
*svr
, vdev_copy_arg_t
*vca
,
1209 uint64_t *max_alloc
, dmu_tx_t
*tx
)
1211 uint64_t txg
= dmu_tx_get_txg(tx
);
1212 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1214 mutex_enter(&svr
->svr_lock
);
1217 * Determine how big of a chunk to copy. We can allocate up
1218 * to max_alloc bytes, and we can span up to vdev_removal_max_span
1219 * bytes of unallocated space at a time. "segs" will track the
1220 * allocated segments that we are copying. We may also be copying
1221 * free segments (of up to vdev_removal_max_span bytes).
1223 range_tree_t
*segs
= range_tree_create(NULL
, NULL
);
1225 range_seg_t
*rs
= range_tree_first(svr
->svr_allocd_segs
);
1230 uint64_t seg_length
;
1232 if (range_tree_is_empty(segs
)) {
1233 /* need to truncate the first seg based on max_alloc */
1235 MIN(rs
->rs_end
- rs
->rs_start
, *max_alloc
);
1237 if (rs
->rs_start
- range_tree_max(segs
) >
1238 vdev_removal_max_span
) {
1240 * Including this segment would cause us to
1241 * copy a larger unneeded chunk than is allowed.
1244 } else if (rs
->rs_end
- range_tree_min(segs
) >
1247 * This additional segment would extend past
1248 * max_alloc. Rather than splitting this
1249 * segment, leave it for the next mapping.
1253 seg_length
= rs
->rs_end
- rs
->rs_start
;
1257 range_tree_add(segs
, rs
->rs_start
, seg_length
);
1258 range_tree_remove(svr
->svr_allocd_segs
,
1259 rs
->rs_start
, seg_length
);
1262 if (range_tree_is_empty(segs
)) {
1263 mutex_exit(&svr
->svr_lock
);
1264 range_tree_destroy(segs
);
1268 if (svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] == 0) {
1269 dsl_sync_task_nowait(dmu_tx_pool(tx
), vdev_mapping_sync
,
1270 svr
, 0, ZFS_SPACE_CHECK_NONE
, tx
);
1273 svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] = range_tree_max(segs
);
1276 * Note: this is the amount of *allocated* space
1277 * that we are taking care of each txg.
1279 svr
->svr_bytes_done
[txg
& TXG_MASK
] += range_tree_space(segs
);
1281 mutex_exit(&svr
->svr_lock
);
1283 zio_alloc_list_t zal
;
1284 metaslab_trace_init(&zal
);
1285 uint64_t thismax
= SPA_MAXBLOCKSIZE
;
1286 while (!range_tree_is_empty(segs
)) {
1287 int error
= spa_vdev_copy_segment(vd
,
1288 segs
, thismax
, txg
, vca
, &zal
);
1290 if (error
== ENOSPC
) {
1292 * Cut our segment in half, and don't try this
1293 * segment size again this txg. Note that the
1294 * allocation size must be aligned to the highest
1295 * ashift in the pool, so that the allocation will
1296 * not be padded out to a multiple of the ashift,
1297 * which could cause us to think that this mapping
1298 * is larger than we intended.
1300 ASSERT3U(spa
->spa_max_ashift
, >=, SPA_MINBLOCKSHIFT
);
1301 ASSERT3U(spa
->spa_max_ashift
, ==, spa
->spa_min_ashift
);
1302 uint64_t attempted
=
1303 MIN(range_tree_span(segs
), thismax
);
1304 thismax
= P2ROUNDUP(attempted
/ 2,
1305 1 << spa
->spa_max_ashift
);
1307 * The minimum-size allocation can not fail.
1309 ASSERT3U(attempted
, >, 1 << spa
->spa_max_ashift
);
1310 *max_alloc
= attempted
- (1 << spa
->spa_max_ashift
);
1315 * We've performed an allocation, so reset the
1318 metaslab_trace_fini(&zal
);
1319 metaslab_trace_init(&zal
);
1322 metaslab_trace_fini(&zal
);
1323 range_tree_destroy(segs
);
1327 * The removal thread operates in open context. It iterates over all
1328 * allocated space in the vdev, by loading each metaslab's spacemap.
1329 * For each contiguous segment of allocated space (capping the segment
1330 * size at SPA_MAXBLOCKSIZE), we:
1331 * - Allocate space for it on another vdev.
1332 * - Create a new mapping from the old location to the new location
1333 * (as a record in svr_new_segments).
1334 * - Initiate a physical read zio to get the data off the removing disk.
1335 * - In the read zio's done callback, initiate a physical write zio to
1336 * write it to the new vdev.
1337 * Note that all of this will take effect when a particular TXG syncs.
1338 * The sync thread ensures that all the phys reads and writes for the syncing
1339 * TXG have completed (see spa_txg_zio) and writes the new mappings to disk
1340 * (see vdev_mapping_sync()).
1343 spa_vdev_remove_thread(void *arg
)
1346 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1347 vdev_copy_arg_t vca
;
1348 uint64_t max_alloc
= zfs_remove_max_segment
;
1349 uint64_t last_txg
= 0;
1351 spa_config_enter(spa
, SCL_CONFIG
, FTAG
, RW_READER
);
1352 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1353 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
1354 uint64_t start_offset
= vdev_indirect_mapping_max_offset(vim
);
1356 ASSERT3P(vd
->vdev_ops
, !=, &vdev_indirect_ops
);
1357 ASSERT(vdev_is_concrete(vd
));
1358 ASSERT(vd
->vdev_removing
);
1359 ASSERT(vd
->vdev_indirect_config
.vic_mapping_object
!= 0);
1360 ASSERT(vim
!= NULL
);
1362 mutex_init(&vca
.vca_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1363 cv_init(&vca
.vca_cv
, NULL
, CV_DEFAULT
, NULL
);
1364 vca
.vca_outstanding_bytes
= 0;
1366 mutex_enter(&svr
->svr_lock
);
1369 * Start from vim_max_offset so we pick up where we left off
1370 * if we are restarting the removal after opening the pool.
1373 for (msi
= start_offset
>> vd
->vdev_ms_shift
;
1374 msi
< vd
->vdev_ms_count
&& !svr
->svr_thread_exit
; msi
++) {
1375 metaslab_t
*msp
= vd
->vdev_ms
[msi
];
1376 ASSERT3U(msi
, <=, vd
->vdev_ms_count
);
1378 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1380 mutex_enter(&msp
->ms_sync_lock
);
1381 mutex_enter(&msp
->ms_lock
);
1384 * Assert nothing in flight -- ms_*tree is empty.
1386 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1387 ASSERT0(range_tree_space(msp
->ms_allocating
[i
]));
1391 * If the metaslab has ever been allocated from (ms_sm!=NULL),
1392 * read the allocated segments from the space map object
1393 * into svr_allocd_segs. Since we do this while holding
1394 * svr_lock and ms_sync_lock, concurrent frees (which
1395 * would have modified the space map) will wait for us
1396 * to finish loading the spacemap, and then take the
1397 * appropriate action (see free_from_removing_vdev()).
1399 if (msp
->ms_sm
!= NULL
) {
1400 space_map_t
*sm
= NULL
;
1403 * We have to open a new space map here, because
1404 * ms_sm's sm_length and sm_alloc may not reflect
1405 * what's in the object contents, if we are in between
1406 * metaslab_sync() and metaslab_sync_done().
1408 VERIFY0(space_map_open(&sm
,
1409 spa
->spa_dsl_pool
->dp_meta_objset
,
1410 msp
->ms_sm
->sm_object
, msp
->ms_sm
->sm_start
,
1411 msp
->ms_sm
->sm_size
, msp
->ms_sm
->sm_shift
));
1412 space_map_update(sm
);
1413 VERIFY0(space_map_load(sm
, svr
->svr_allocd_segs
,
1415 space_map_close(sm
);
1417 range_tree_walk(msp
->ms_freeing
,
1418 range_tree_remove
, svr
->svr_allocd_segs
);
1421 * When we are resuming from a paused removal (i.e.
1422 * when importing a pool with a removal in progress),
1423 * discard any state that we have already processed.
1425 range_tree_clear(svr
->svr_allocd_segs
, 0, start_offset
);
1427 mutex_exit(&msp
->ms_lock
);
1428 mutex_exit(&msp
->ms_sync_lock
);
1431 zfs_dbgmsg("copying %llu segments for metaslab %llu",
1432 avl_numnodes(&svr
->svr_allocd_segs
->rt_root
),
1435 while (!svr
->svr_thread_exit
&&
1436 !range_tree_is_empty(svr
->svr_allocd_segs
)) {
1438 mutex_exit(&svr
->svr_lock
);
1441 * We need to periodically drop the config lock so that
1442 * writers can get in. Additionally, we can't wait
1443 * for a txg to sync while holding a config lock
1444 * (since a waiting writer could cause a 3-way deadlock
1445 * with the sync thread, which also gets a config
1446 * lock for reader). So we can't hold the config lock
1447 * while calling dmu_tx_assign().
1449 spa_config_exit(spa
, SCL_CONFIG
, FTAG
);
1452 * This delay will pause the removal around the point
1453 * specified by zfs_remove_max_bytes_pause. We do this
1454 * solely from the test suite or during debugging.
1456 uint64_t bytes_copied
=
1457 spa
->spa_removing_phys
.sr_copied
;
1458 for (int i
= 0; i
< TXG_SIZE
; i
++)
1459 bytes_copied
+= svr
->svr_bytes_done
[i
];
1460 while (zfs_remove_max_bytes_pause
<= bytes_copied
&&
1461 !svr
->svr_thread_exit
)
1464 mutex_enter(&vca
.vca_lock
);
1465 while (vca
.vca_outstanding_bytes
>
1466 zfs_remove_max_copy_bytes
) {
1467 cv_wait(&vca
.vca_cv
, &vca
.vca_lock
);
1469 mutex_exit(&vca
.vca_lock
);
1472 dmu_tx_create_dd(spa_get_dsl(spa
)->dp_mos_dir
);
1473 dmu_tx_hold_space(tx
, SPA_MAXBLOCKSIZE
);
1474 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
1475 uint64_t txg
= dmu_tx_get_txg(tx
);
1478 * Reacquire the vdev_config lock. The vdev_t
1479 * that we're removing may have changed, e.g. due
1480 * to a vdev_attach or vdev_detach.
1482 spa_config_enter(spa
, SCL_CONFIG
, FTAG
, RW_READER
);
1483 vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1485 if (txg
!= last_txg
)
1486 max_alloc
= zfs_remove_max_segment
;
1489 spa_vdev_copy_impl(vd
, svr
, &vca
, &max_alloc
, tx
);
1492 mutex_enter(&svr
->svr_lock
);
1496 mutex_exit(&svr
->svr_lock
);
1498 spa_config_exit(spa
, SCL_CONFIG
, FTAG
);
1501 * Wait for all copies to finish before cleaning up the vca.
1503 txg_wait_synced(spa
->spa_dsl_pool
, 0);
1504 ASSERT0(vca
.vca_outstanding_bytes
);
1506 mutex_destroy(&vca
.vca_lock
);
1507 cv_destroy(&vca
.vca_cv
);
1509 if (svr
->svr_thread_exit
) {
1510 mutex_enter(&svr
->svr_lock
);
1511 range_tree_vacate(svr
->svr_allocd_segs
, NULL
, NULL
);
1512 svr
->svr_thread
= NULL
;
1513 cv_broadcast(&svr
->svr_cv
);
1514 mutex_exit(&svr
->svr_lock
);
1516 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1517 vdev_remove_complete(spa
);
1522 spa_vdev_remove_suspend(spa_t
*spa
)
1524 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1529 mutex_enter(&svr
->svr_lock
);
1530 svr
->svr_thread_exit
= B_TRUE
;
1531 while (svr
->svr_thread
!= NULL
)
1532 cv_wait(&svr
->svr_cv
, &svr
->svr_lock
);
1533 svr
->svr_thread_exit
= B_FALSE
;
1534 mutex_exit(&svr
->svr_lock
);
1539 spa_vdev_remove_cancel_check(void *arg
, dmu_tx_t
*tx
)
1541 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1543 if (spa
->spa_vdev_removal
== NULL
)
1544 return (ENOTACTIVE
);
1549 * Cancel a removal by freeing all entries from the partial mapping
1550 * and marking the vdev as no longer being removing.
1554 spa_vdev_remove_cancel_sync(void *arg
, dmu_tx_t
*tx
)
1556 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1557 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1558 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1559 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
1560 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
1561 objset_t
*mos
= spa
->spa_meta_objset
;
1563 ASSERT3P(svr
->svr_thread
, ==, NULL
);
1565 spa_feature_decr(spa
, SPA_FEATURE_DEVICE_REMOVAL
, tx
);
1566 if (vdev_obsolete_counts_are_precise(vd
)) {
1567 spa_feature_decr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
1568 VERIFY0(zap_remove(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
1569 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE
, tx
));
1572 if (vdev_obsolete_sm_object(vd
) != 0) {
1573 ASSERT(vd
->vdev_obsolete_sm
!= NULL
);
1574 ASSERT3U(vdev_obsolete_sm_object(vd
), ==,
1575 space_map_object(vd
->vdev_obsolete_sm
));
1577 space_map_free(vd
->vdev_obsolete_sm
, tx
);
1578 VERIFY0(zap_remove(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
1579 VDEV_TOP_ZAP_INDIRECT_OBSOLETE_SM
, tx
));
1580 space_map_close(vd
->vdev_obsolete_sm
);
1581 vd
->vdev_obsolete_sm
= NULL
;
1582 spa_feature_decr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
1584 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1585 ASSERT(list_is_empty(&svr
->svr_new_segments
[i
]));
1586 ASSERT3U(svr
->svr_max_offset_to_sync
[i
], <=,
1587 vdev_indirect_mapping_max_offset(vim
));
1590 for (uint64_t msi
= 0; msi
< vd
->vdev_ms_count
; msi
++) {
1591 metaslab_t
*msp
= vd
->vdev_ms
[msi
];
1593 if (msp
->ms_start
>= vdev_indirect_mapping_max_offset(vim
))
1596 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1598 mutex_enter(&msp
->ms_lock
);
1601 * Assert nothing in flight -- ms_*tree is empty.
1603 for (int i
= 0; i
< TXG_SIZE
; i
++)
1604 ASSERT0(range_tree_space(msp
->ms_allocating
[i
]));
1605 for (int i
= 0; i
< TXG_DEFER_SIZE
; i
++)
1606 ASSERT0(range_tree_space(msp
->ms_defer
[i
]));
1607 ASSERT0(range_tree_space(msp
->ms_freed
));
1609 if (msp
->ms_sm
!= NULL
) {
1611 * Assert that the in-core spacemap has the same
1612 * length as the on-disk one, so we can use the
1613 * existing in-core spacemap to load it from disk.
1615 ASSERT3U(msp
->ms_sm
->sm_alloc
, ==,
1616 msp
->ms_sm
->sm_phys
->smp_alloc
);
1617 ASSERT3U(msp
->ms_sm
->sm_length
, ==,
1618 msp
->ms_sm
->sm_phys
->smp_objsize
);
1620 mutex_enter(&svr
->svr_lock
);
1621 VERIFY0(space_map_load(msp
->ms_sm
,
1622 svr
->svr_allocd_segs
, SM_ALLOC
));
1623 range_tree_walk(msp
->ms_freeing
,
1624 range_tree_remove
, svr
->svr_allocd_segs
);
1627 * Clear everything past what has been synced,
1628 * because we have not allocated mappings for it yet.
1630 uint64_t syncd
= vdev_indirect_mapping_max_offset(vim
);
1631 uint64_t sm_end
= msp
->ms_sm
->sm_start
+
1632 msp
->ms_sm
->sm_size
;
1634 range_tree_clear(svr
->svr_allocd_segs
,
1635 syncd
, sm_end
- syncd
);
1637 mutex_exit(&svr
->svr_lock
);
1639 mutex_exit(&msp
->ms_lock
);
1641 mutex_enter(&svr
->svr_lock
);
1642 range_tree_vacate(svr
->svr_allocd_segs
,
1643 free_mapped_segment_cb
, vd
);
1644 mutex_exit(&svr
->svr_lock
);
1648 * Note: this must happen after we invoke free_mapped_segment_cb,
1649 * because it adds to the obsolete_segments.
1651 range_tree_vacate(vd
->vdev_obsolete_segments
, NULL
, NULL
);
1653 ASSERT3U(vic
->vic_mapping_object
, ==,
1654 vdev_indirect_mapping_object(vd
->vdev_indirect_mapping
));
1655 vdev_indirect_mapping_close(vd
->vdev_indirect_mapping
);
1656 vd
->vdev_indirect_mapping
= NULL
;
1657 vdev_indirect_mapping_free(mos
, vic
->vic_mapping_object
, tx
);
1658 vic
->vic_mapping_object
= 0;
1660 ASSERT3U(vic
->vic_births_object
, ==,
1661 vdev_indirect_births_object(vd
->vdev_indirect_births
));
1662 vdev_indirect_births_close(vd
->vdev_indirect_births
);
1663 vd
->vdev_indirect_births
= NULL
;
1664 vdev_indirect_births_free(mos
, vic
->vic_births_object
, tx
);
1665 vic
->vic_births_object
= 0;
1668 * We may have processed some frees from the removing vdev in this
1669 * txg, thus increasing svr_bytes_done; discard that here to
1670 * satisfy the assertions in spa_vdev_removal_destroy().
1671 * Note that future txg's can not have any bytes_done, because
1672 * future TXG's are only modified from open context, and we have
1673 * already shut down the copying thread.
1675 svr
->svr_bytes_done
[dmu_tx_get_txg(tx
) & TXG_MASK
] = 0;
1676 spa_finish_removal(spa
, DSS_CANCELED
, tx
);
1678 vd
->vdev_removing
= B_FALSE
;
1679 vdev_config_dirty(vd
);
1681 zfs_dbgmsg("canceled device removal for vdev %llu in %llu",
1682 vd
->vdev_id
, dmu_tx_get_txg(tx
));
1683 spa_history_log_internal(spa
, "vdev remove canceled", tx
,
1684 "%s vdev %llu %s", spa_name(spa
),
1685 vd
->vdev_id
, (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
1689 spa_vdev_remove_cancel(spa_t
*spa
)
1691 spa_vdev_remove_suspend(spa
);
1693 if (spa
->spa_vdev_removal
== NULL
)
1694 return (ENOTACTIVE
);
1696 uint64_t vdid
= spa
->spa_vdev_removal
->svr_vdev_id
;
1698 int error
= dsl_sync_task(spa
->spa_name
, spa_vdev_remove_cancel_check
,
1699 spa_vdev_remove_cancel_sync
, NULL
, 0,
1700 ZFS_SPACE_CHECK_EXTRA_RESERVED
);
1703 spa_config_enter(spa
, SCL_ALLOC
| SCL_VDEV
, FTAG
, RW_WRITER
);
1704 vdev_t
*vd
= vdev_lookup_top(spa
, vdid
);
1705 metaslab_group_activate(vd
->vdev_mg
);
1706 spa_config_exit(spa
, SCL_ALLOC
| SCL_VDEV
, FTAG
);
1713 * Called every sync pass of every txg if there's a svr.
1716 svr_sync(spa_t
*spa
, dmu_tx_t
*tx
)
1718 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1719 int txgoff
= dmu_tx_get_txg(tx
) & TXG_MASK
;
1722 * This check is necessary so that we do not dirty the
1723 * DIRECTORY_OBJECT via spa_sync_removing_state() when there
1724 * is nothing to do. Dirtying it every time would prevent us
1725 * from syncing-to-convergence.
1727 if (svr
->svr_bytes_done
[txgoff
] == 0)
1731 * Update progress accounting.
1733 spa
->spa_removing_phys
.sr_copied
+= svr
->svr_bytes_done
[txgoff
];
1734 svr
->svr_bytes_done
[txgoff
] = 0;
1736 spa_sync_removing_state(spa
, tx
);
1740 vdev_remove_make_hole_and_free(vdev_t
*vd
)
1742 uint64_t id
= vd
->vdev_id
;
1743 spa_t
*spa
= vd
->vdev_spa
;
1744 vdev_t
*rvd
= spa
->spa_root_vdev
;
1745 boolean_t last_vdev
= (id
== (rvd
->vdev_children
- 1));
1747 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
1748 ASSERT(spa_config_held(spa
, SCL_ALL
, RW_WRITER
) == SCL_ALL
);
1753 vdev_compact_children(rvd
);
1755 vd
= vdev_alloc_common(spa
, id
, 0, &vdev_hole_ops
);
1756 vdev_add_child(rvd
, vd
);
1758 vdev_config_dirty(rvd
);
1761 * Reassess the health of our root vdev.
1767 * Remove a log device. The config lock is held for the specified TXG.
1770 spa_vdev_remove_log(vdev_t
*vd
, uint64_t *txg
)
1772 metaslab_group_t
*mg
= vd
->vdev_mg
;
1773 spa_t
*spa
= vd
->vdev_spa
;
1776 ASSERT(vd
->vdev_islog
);
1777 ASSERT(vd
== vd
->vdev_top
);
1780 * Stop allocating from this vdev.
1782 metaslab_group_passivate(mg
);
1785 * Wait for the youngest allocations and frees to sync,
1786 * and then wait for the deferral of those frees to finish.
1788 spa_vdev_config_exit(spa
, NULL
,
1789 *txg
+ TXG_CONCURRENT_STATES
+ TXG_DEFER_SIZE
, 0, FTAG
);
1792 * Evacuate the device. We don't hold the config lock as writer
1793 * since we need to do I/O but we do keep the
1794 * spa_namespace_lock held. Once this completes the device
1795 * should no longer have any blocks allocated on it.
1797 if (vd
->vdev_islog
) {
1798 if (vd
->vdev_stat
.vs_alloc
!= 0)
1799 error
= spa_reset_logs(spa
);
1802 *txg
= spa_vdev_config_enter(spa
);
1805 metaslab_group_activate(mg
);
1808 ASSERT0(vd
->vdev_stat
.vs_alloc
);
1811 * The evacuation succeeded. Remove any remaining MOS metadata
1812 * associated with this vdev, and wait for these changes to sync.
1814 vd
->vdev_removing
= B_TRUE
;
1816 vdev_dirty_leaves(vd
, VDD_DTL
, *txg
);
1817 vdev_config_dirty(vd
);
1819 spa_history_log_internal(spa
, "vdev remove", NULL
,
1820 "%s vdev %llu (log) %s", spa_name(spa
), vd
->vdev_id
,
1821 (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
1823 spa_vdev_config_exit(spa
, NULL
, *txg
, 0, FTAG
);
1825 *txg
= spa_vdev_config_enter(spa
);
1827 sysevent_t
*ev
= spa_event_create(spa
, vd
, NULL
,
1828 ESC_ZFS_VDEV_REMOVE_DEV
);
1829 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
1830 ASSERT(spa_config_held(spa
, SCL_ALL
, RW_WRITER
) == SCL_ALL
);
1832 /* The top ZAP should have been destroyed by vdev_remove_empty. */
1833 ASSERT0(vd
->vdev_top_zap
);
1834 /* The leaf ZAP should have been destroyed by vdev_dtl_sync. */
1835 ASSERT0(vd
->vdev_leaf_zap
);
1837 (void) vdev_label_init(vd
, 0, VDEV_LABEL_REMOVE
);
1839 if (list_link_active(&vd
->vdev_state_dirty_node
))
1840 vdev_state_clean(vd
);
1841 if (list_link_active(&vd
->vdev_config_dirty_node
))
1842 vdev_config_clean(vd
);
1845 * Clean up the vdev namespace.
1847 vdev_remove_make_hole_and_free(vd
);
1856 spa_vdev_remove_top_check(vdev_t
*vd
)
1858 spa_t
*spa
= vd
->vdev_spa
;
1860 if (vd
!= vd
->vdev_top
)
1861 return (SET_ERROR(ENOTSUP
));
1863 if (!spa_feature_is_enabled(spa
, SPA_FEATURE_DEVICE_REMOVAL
))
1864 return (SET_ERROR(ENOTSUP
));
1866 /* available space in the pool's normal class */
1867 uint64_t available
= dsl_dir_space_available(
1868 spa
->spa_dsl_pool
->dp_root_dir
, NULL
, 0, B_TRUE
);
1870 metaslab_class_t
*mc
= vd
->vdev_mg
->mg_class
;
1873 * When removing a vdev from an allocation class that has
1874 * remaining vdevs, include available space from the class.
1876 if (mc
!= spa_normal_class(spa
) && mc
->mc_groups
> 1) {
1877 uint64_t class_avail
= metaslab_class_get_space(mc
) -
1878 metaslab_class_get_alloc(mc
);
1880 /* add class space, adjusted for overhead */
1881 available
+= (class_avail
* 94) / 100;
1885 * There has to be enough free space to remove the
1886 * device and leave double the "slop" space (i.e. we
1887 * must leave at least 3% of the pool free, in addition to
1888 * the normal slop space).
1890 if (available
< vd
->vdev_stat
.vs_dspace
+ spa_get_slop_space(spa
)) {
1891 return (SET_ERROR(ENOSPC
));
1895 * There can not be a removal in progress.
1897 if (spa
->spa_removing_phys
.sr_state
== DSS_SCANNING
)
1898 return (SET_ERROR(EBUSY
));
1901 * The device must have all its data.
1903 if (!vdev_dtl_empty(vd
, DTL_MISSING
) ||
1904 !vdev_dtl_empty(vd
, DTL_OUTAGE
))
1905 return (SET_ERROR(EBUSY
));
1908 * The device must be healthy.
1910 if (!vdev_readable(vd
))
1911 return (SET_ERROR(EIO
));
1914 * All vdevs in normal class must have the same ashift.
1916 if (spa
->spa_max_ashift
!= spa
->spa_min_ashift
) {
1917 return (SET_ERROR(EINVAL
));
1921 * All vdevs in normal class must have the same ashift
1924 vdev_t
*rvd
= spa
->spa_root_vdev
;
1925 int num_indirect
= 0;
1926 for (uint64_t id
= 0; id
< rvd
->vdev_children
; id
++) {
1927 vdev_t
*cvd
= rvd
->vdev_child
[id
];
1928 if (cvd
->vdev_ashift
!= 0 && !cvd
->vdev_islog
)
1929 ASSERT3U(cvd
->vdev_ashift
, ==, spa
->spa_max_ashift
);
1930 if (cvd
->vdev_ops
== &vdev_indirect_ops
)
1932 if (!vdev_is_concrete(cvd
))
1934 if (cvd
->vdev_ops
== &vdev_raidz_ops
)
1935 return (SET_ERROR(EINVAL
));
1937 * Need the mirror to be mirror of leaf vdevs only
1939 if (cvd
->vdev_ops
== &vdev_mirror_ops
) {
1940 for (uint64_t cid
= 0;
1941 cid
< cvd
->vdev_children
; cid
++) {
1942 if (!cvd
->vdev_child
[cid
]->vdev_ops
->
1944 return (SET_ERROR(EINVAL
));
1953 * Initiate removal of a top-level vdev, reducing the total space in the pool.
1954 * The config lock is held for the specified TXG. Once initiated,
1955 * evacuation of all allocated space (copying it to other vdevs) happens
1956 * in the background (see spa_vdev_remove_thread()), and can be canceled
1957 * (see spa_vdev_remove_cancel()). If successful, the vdev will
1958 * be transformed to an indirect vdev (see spa_vdev_remove_complete()).
1961 spa_vdev_remove_top(vdev_t
*vd
, uint64_t *txg
)
1963 spa_t
*spa
= vd
->vdev_spa
;
1967 * Check for errors up-front, so that we don't waste time
1968 * passivating the metaslab group and clearing the ZIL if there
1971 error
= spa_vdev_remove_top_check(vd
);
1976 * Stop allocating from this vdev. Note that we must check
1977 * that this is not the only device in the pool before
1978 * passivating, otherwise we will not be able to make
1979 * progress because we can't allocate from any vdevs.
1980 * The above check for sufficient free space serves this
1983 metaslab_group_t
*mg
= vd
->vdev_mg
;
1984 metaslab_group_passivate(mg
);
1987 * Wait for the youngest allocations and frees to sync,
1988 * and then wait for the deferral of those frees to finish.
1990 spa_vdev_config_exit(spa
, NULL
,
1991 *txg
+ TXG_CONCURRENT_STATES
+ TXG_DEFER_SIZE
, 0, FTAG
);
1994 * We must ensure that no "stubby" log blocks are allocated
1995 * on the device to be removed. These blocks could be
1996 * written at any time, including while we are in the middle
1999 error
= spa_reset_logs(spa
);
2001 *txg
= spa_vdev_config_enter(spa
);
2004 * Things might have changed while the config lock was dropped
2005 * (e.g. space usage). Check for errors again.
2008 error
= spa_vdev_remove_top_check(vd
);
2011 metaslab_group_activate(mg
);
2015 vd
->vdev_removing
= B_TRUE
;
2017 vdev_dirty_leaves(vd
, VDD_DTL
, *txg
);
2018 vdev_config_dirty(vd
);
2019 dmu_tx_t
*tx
= dmu_tx_create_assigned(spa
->spa_dsl_pool
, *txg
);
2020 dsl_sync_task_nowait(spa
->spa_dsl_pool
,
2021 vdev_remove_initiate_sync
,
2022 (void *)(uintptr_t)vd
->vdev_id
, 0, ZFS_SPACE_CHECK_NONE
, tx
);
2029 * Remove a device from the pool.
2031 * Removing a device from the vdev namespace requires several steps
2032 * and can take a significant amount of time. As a result we use
2033 * the spa_vdev_config_[enter/exit] functions which allow us to
2034 * grab and release the spa_config_lock while still holding the namespace
2035 * lock. During each step the configuration is synced out.
2038 spa_vdev_remove(spa_t
*spa
, uint64_t guid
, boolean_t unspare
)
2041 nvlist_t
**spares
, **l2cache
, *nv
;
2043 uint_t nspares
, nl2cache
;
2045 boolean_t locked
= MUTEX_HELD(&spa_namespace_lock
);
2046 sysevent_t
*ev
= NULL
;
2048 ASSERT(spa_writeable(spa
));
2051 txg
= spa_vdev_enter(spa
);
2053 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
2054 if (spa_feature_is_active(spa
, SPA_FEATURE_POOL_CHECKPOINT
)) {
2055 error
= (spa_has_checkpoint(spa
)) ?
2056 ZFS_ERR_CHECKPOINT_EXISTS
: ZFS_ERR_DISCARDING_CHECKPOINT
;
2059 return (spa_vdev_exit(spa
, NULL
, txg
, error
));
2064 vd
= spa_lookup_by_guid(spa
, guid
, B_FALSE
);
2066 if (spa
->spa_spares
.sav_vdevs
!= NULL
&&
2067 nvlist_lookup_nvlist_array(spa
->spa_spares
.sav_config
,
2068 ZPOOL_CONFIG_SPARES
, &spares
, &nspares
) == 0 &&
2069 (nv
= spa_nvlist_lookup_by_guid(spares
, nspares
, guid
)) != NULL
) {
2071 * Only remove the hot spare if it's not currently in use
2074 if (vd
== NULL
|| unspare
) {
2076 vd
= spa_lookup_by_guid(spa
, guid
, B_TRUE
);
2077 ev
= spa_event_create(spa
, vd
, NULL
,
2078 ESC_ZFS_VDEV_REMOVE_AUX
);
2080 char *nvstr
= fnvlist_lookup_string(nv
,
2082 spa_history_log_internal(spa
, "vdev remove", NULL
,
2083 "%s vdev (%s) %s", spa_name(spa
),
2084 VDEV_TYPE_SPARE
, nvstr
);
2085 spa_vdev_remove_aux(spa
->spa_spares
.sav_config
,
2086 ZPOOL_CONFIG_SPARES
, spares
, nspares
, nv
);
2087 spa_load_spares(spa
);
2088 spa
->spa_spares
.sav_sync
= B_TRUE
;
2090 error
= SET_ERROR(EBUSY
);
2092 } else if (spa
->spa_l2cache
.sav_vdevs
!= NULL
&&
2093 nvlist_lookup_nvlist_array(spa
->spa_l2cache
.sav_config
,
2094 ZPOOL_CONFIG_L2CACHE
, &l2cache
, &nl2cache
) == 0 &&
2095 (nv
= spa_nvlist_lookup_by_guid(l2cache
, nl2cache
, guid
)) != NULL
) {
2096 char *nvstr
= fnvlist_lookup_string(nv
, ZPOOL_CONFIG_PATH
);
2097 spa_history_log_internal(spa
, "vdev remove", NULL
,
2098 "%s vdev (%s) %s", spa_name(spa
), VDEV_TYPE_L2CACHE
, nvstr
);
2100 * Cache devices can always be removed.
2102 vd
= spa_lookup_by_guid(spa
, guid
, B_TRUE
);
2103 ev
= spa_event_create(spa
, vd
, NULL
, ESC_ZFS_VDEV_REMOVE_AUX
);
2104 spa_vdev_remove_aux(spa
->spa_l2cache
.sav_config
,
2105 ZPOOL_CONFIG_L2CACHE
, l2cache
, nl2cache
, nv
);
2106 spa_load_l2cache(spa
);
2107 spa
->spa_l2cache
.sav_sync
= B_TRUE
;
2108 } else if (vd
!= NULL
&& vd
->vdev_islog
) {
2110 error
= spa_vdev_remove_log(vd
, &txg
);
2111 } else if (vd
!= NULL
) {
2113 error
= spa_vdev_remove_top(vd
, &txg
);
2116 * There is no vdev of any kind with the specified guid.
2118 error
= SET_ERROR(ENOENT
);
2122 error
= spa_vdev_exit(spa
, NULL
, txg
, error
);
2131 spa_removal_get_stats(spa_t
*spa
, pool_removal_stat_t
*prs
)
2133 prs
->prs_state
= spa
->spa_removing_phys
.sr_state
;
2135 if (prs
->prs_state
== DSS_NONE
)
2136 return (SET_ERROR(ENOENT
));
2138 prs
->prs_removing_vdev
= spa
->spa_removing_phys
.sr_removing_vdev
;
2139 prs
->prs_start_time
= spa
->spa_removing_phys
.sr_start_time
;
2140 prs
->prs_end_time
= spa
->spa_removing_phys
.sr_end_time
;
2141 prs
->prs_to_copy
= spa
->spa_removing_phys
.sr_to_copy
;
2142 prs
->prs_copied
= spa
->spa_removing_phys
.sr_copied
;
2144 if (spa
->spa_vdev_removal
!= NULL
) {
2145 for (int i
= 0; i
< TXG_SIZE
; i
++) {
2147 spa
->spa_vdev_removal
->svr_bytes_done
[i
];
2151 prs
->prs_mapping_memory
= 0;
2152 uint64_t indirect_vdev_id
=
2153 spa
->spa_removing_phys
.sr_prev_indirect_vdev
;
2154 while (indirect_vdev_id
!= -1) {
2155 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[indirect_vdev_id
];
2156 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
2157 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
2159 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
2160 prs
->prs_mapping_memory
+= vdev_indirect_mapping_size(vim
);
2161 indirect_vdev_id
= vic
->vic_prev_indirect_vdev
;
2167 #if defined(_KERNEL)
2168 module_param(zfs_remove_max_segment
, int, 0644);
2169 MODULE_PARM_DESC(zfs_remove_max_segment
,
2170 "Largest contiguous segment to allocate when removing device");
2172 module_param(vdev_removal_max_span
, int, 0644);
2173 MODULE_PARM_DESC(vdev_removal_max_span
,
2174 "Largest span of free chunks a remap segment can span");
2177 module_param(zfs_remove_max_bytes_pause
, ulong
, 0644);
2178 MODULE_PARM_DESC(zfs_remove_max_bytes_pause
,
2179 "Pause device removal after this many bytes are copied "
2180 "(debug use only - causes removal to hang)");
2183 EXPORT_SYMBOL(free_from_removing_vdev
);
2184 EXPORT_SYMBOL(spa_removal_get_stats
);
2185 EXPORT_SYMBOL(spa_remove_init
);
2186 EXPORT_SYMBOL(spa_restart_removal
);
2187 EXPORT_SYMBOL(spa_vdev_removal_destroy
);
2188 EXPORT_SYMBOL(spa_vdev_remove
);
2189 EXPORT_SYMBOL(spa_vdev_remove_cancel
);
2190 EXPORT_SYMBOL(spa_vdev_remove_suspend
);
2191 EXPORT_SYMBOL(svr_sync
);