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
;
83 uint64_t vca_read_error_bytes
;
84 uint64_t vca_write_error_bytes
;
90 * The maximum amount of memory we can use for outstanding i/o while
91 * doing a device removal. This determines how much i/o we can have
92 * in flight concurrently.
94 int zfs_remove_max_copy_bytes
= 64 * 1024 * 1024;
97 * The largest contiguous segment that we will attempt to allocate when
98 * removing a device. This can be no larger than SPA_MAXBLOCKSIZE. If
99 * there is a performance problem with attempting to allocate large blocks,
100 * consider decreasing this.
102 int zfs_remove_max_segment
= SPA_MAXBLOCKSIZE
;
105 * Ignore hard IO errors during device removal. When set if a device
106 * encounters hard IO error during the removal process the removal will
107 * not be cancelled. This can result in a normally recoverable block
108 * becoming permanently damaged and is not recommended.
110 int zfs_removal_ignore_errors
= 0;
113 * Allow a remap segment to span free chunks of at most this size. The main
114 * impact of a larger span is that we will read and write larger, more
115 * contiguous chunks, with more "unnecessary" data -- trading off bandwidth
116 * for iops. The value here was chosen to align with
117 * zfs_vdev_read_gap_limit, which is a similar concept when doing regular
118 * reads (but there's no reason it has to be the same).
120 * Additionally, a higher span will have the following relatively minor
122 * - the mapping will be smaller, since one entry can cover more allocated
124 * - more of the fragmentation in the removing device will be preserved
125 * - we'll do larger allocations, which may fail and fall back on smaller
128 int vdev_removal_max_span
= 32 * 1024;
131 * This is used by the test suite so that it can ensure that certain
132 * actions happen while in the middle of a removal.
134 int zfs_removal_suspend_progress
= 0;
136 #define VDEV_REMOVAL_ZAP_OBJS "lzap"
138 static void spa_vdev_remove_thread(void *arg
);
139 static int spa_vdev_remove_cancel_impl(spa_t
*spa
);
142 spa_sync_removing_state(spa_t
*spa
, dmu_tx_t
*tx
)
144 VERIFY0(zap_update(spa
->spa_dsl_pool
->dp_meta_objset
,
145 DMU_POOL_DIRECTORY_OBJECT
,
146 DMU_POOL_REMOVING
, sizeof (uint64_t),
147 sizeof (spa
->spa_removing_phys
) / sizeof (uint64_t),
148 &spa
->spa_removing_phys
, tx
));
152 spa_nvlist_lookup_by_guid(nvlist_t
**nvpp
, int count
, uint64_t target_guid
)
154 for (int i
= 0; i
< count
; i
++) {
156 fnvlist_lookup_uint64(nvpp
[i
], ZPOOL_CONFIG_GUID
);
158 if (guid
== target_guid
)
166 spa_vdev_remove_aux(nvlist_t
*config
, char *name
, nvlist_t
**dev
, int count
,
167 nvlist_t
*dev_to_remove
)
169 nvlist_t
**newdev
= NULL
;
172 newdev
= kmem_alloc((count
- 1) * sizeof (void *), KM_SLEEP
);
174 for (int i
= 0, j
= 0; i
< count
; i
++) {
175 if (dev
[i
] == dev_to_remove
)
177 VERIFY(nvlist_dup(dev
[i
], &newdev
[j
++], KM_SLEEP
) == 0);
180 VERIFY(nvlist_remove(config
, name
, DATA_TYPE_NVLIST_ARRAY
) == 0);
181 VERIFY(nvlist_add_nvlist_array(config
, name
, newdev
, count
- 1) == 0);
183 for (int i
= 0; i
< count
- 1; i
++)
184 nvlist_free(newdev
[i
]);
187 kmem_free(newdev
, (count
- 1) * sizeof (void *));
190 static spa_vdev_removal_t
*
191 spa_vdev_removal_create(vdev_t
*vd
)
193 spa_vdev_removal_t
*svr
= kmem_zalloc(sizeof (*svr
), KM_SLEEP
);
194 mutex_init(&svr
->svr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
195 cv_init(&svr
->svr_cv
, NULL
, CV_DEFAULT
, NULL
);
196 svr
->svr_allocd_segs
= range_tree_create(NULL
, NULL
);
197 svr
->svr_vdev_id
= vd
->vdev_id
;
199 for (int i
= 0; i
< TXG_SIZE
; i
++) {
200 svr
->svr_frees
[i
] = range_tree_create(NULL
, NULL
);
201 list_create(&svr
->svr_new_segments
[i
],
202 sizeof (vdev_indirect_mapping_entry_t
),
203 offsetof(vdev_indirect_mapping_entry_t
, vime_node
));
210 spa_vdev_removal_destroy(spa_vdev_removal_t
*svr
)
212 for (int i
= 0; i
< TXG_SIZE
; i
++) {
213 ASSERT0(svr
->svr_bytes_done
[i
]);
214 ASSERT0(svr
->svr_max_offset_to_sync
[i
]);
215 range_tree_destroy(svr
->svr_frees
[i
]);
216 list_destroy(&svr
->svr_new_segments
[i
]);
219 range_tree_destroy(svr
->svr_allocd_segs
);
220 mutex_destroy(&svr
->svr_lock
);
221 cv_destroy(&svr
->svr_cv
);
222 kmem_free(svr
, sizeof (*svr
));
226 * This is called as a synctask in the txg in which we will mark this vdev
227 * as removing (in the config stored in the MOS).
229 * It begins the evacuation of a toplevel vdev by:
230 * - initializing the spa_removing_phys which tracks this removal
231 * - computing the amount of space to remove for accounting purposes
232 * - dirtying all dbufs in the spa_config_object
233 * - creating the spa_vdev_removal
234 * - starting the spa_vdev_remove_thread
237 vdev_remove_initiate_sync(void *arg
, dmu_tx_t
*tx
)
239 int vdev_id
= (uintptr_t)arg
;
240 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
241 vdev_t
*vd
= vdev_lookup_top(spa
, vdev_id
);
242 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
243 objset_t
*mos
= spa
->spa_dsl_pool
->dp_meta_objset
;
244 spa_vdev_removal_t
*svr
= NULL
;
245 ASSERTV(uint64_t txg
= dmu_tx_get_txg(tx
));
247 ASSERT3P(vd
->vdev_ops
, !=, &vdev_raidz_ops
);
248 svr
= spa_vdev_removal_create(vd
);
250 ASSERT(vd
->vdev_removing
);
251 ASSERT3P(vd
->vdev_indirect_mapping
, ==, NULL
);
253 spa_feature_incr(spa
, SPA_FEATURE_DEVICE_REMOVAL
, tx
);
254 if (spa_feature_is_enabled(spa
, SPA_FEATURE_OBSOLETE_COUNTS
)) {
256 * By activating the OBSOLETE_COUNTS feature, we prevent
257 * the pool from being downgraded and ensure that the
258 * refcounts are precise.
260 spa_feature_incr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
262 VERIFY0(zap_add(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
263 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE
, sizeof (one
), 1,
265 ASSERTV(boolean_t are_precise
);
266 ASSERT0(vdev_obsolete_counts_are_precise(vd
, &are_precise
));
267 ASSERT3B(are_precise
, ==, B_TRUE
);
270 vic
->vic_mapping_object
= vdev_indirect_mapping_alloc(mos
, tx
);
271 vd
->vdev_indirect_mapping
=
272 vdev_indirect_mapping_open(mos
, vic
->vic_mapping_object
);
273 vic
->vic_births_object
= vdev_indirect_births_alloc(mos
, tx
);
274 vd
->vdev_indirect_births
=
275 vdev_indirect_births_open(mos
, vic
->vic_births_object
);
276 spa
->spa_removing_phys
.sr_removing_vdev
= vd
->vdev_id
;
277 spa
->spa_removing_phys
.sr_start_time
= gethrestime_sec();
278 spa
->spa_removing_phys
.sr_end_time
= 0;
279 spa
->spa_removing_phys
.sr_state
= DSS_SCANNING
;
280 spa
->spa_removing_phys
.sr_to_copy
= 0;
281 spa
->spa_removing_phys
.sr_copied
= 0;
284 * Note: We can't use vdev_stat's vs_alloc for sr_to_copy, because
285 * there may be space in the defer tree, which is free, but still
286 * counted in vs_alloc.
288 for (uint64_t i
= 0; i
< vd
->vdev_ms_count
; i
++) {
289 metaslab_t
*ms
= vd
->vdev_ms
[i
];
290 if (ms
->ms_sm
== NULL
)
294 * Sync tasks happen before metaslab_sync(), therefore
295 * smp_alloc and sm_alloc must be the same.
297 ASSERT3U(space_map_allocated(ms
->ms_sm
), ==,
298 ms
->ms_sm
->sm_phys
->smp_alloc
);
300 spa
->spa_removing_phys
.sr_to_copy
+=
301 space_map_allocated(ms
->ms_sm
);
304 * Space which we are freeing this txg does not need to
307 spa
->spa_removing_phys
.sr_to_copy
-=
308 range_tree_space(ms
->ms_freeing
);
310 ASSERT0(range_tree_space(ms
->ms_freed
));
311 for (int t
= 0; t
< TXG_SIZE
; t
++)
312 ASSERT0(range_tree_space(ms
->ms_allocating
[t
]));
316 * Sync tasks are called before metaslab_sync(), so there should
317 * be no already-synced metaslabs in the TXG_CLEAN list.
319 ASSERT3P(txg_list_head(&vd
->vdev_ms_list
, TXG_CLEAN(txg
)), ==, NULL
);
321 spa_sync_removing_state(spa
, tx
);
324 * All blocks that we need to read the most recent mapping must be
325 * stored on concrete vdevs. Therefore, we must dirty anything that
326 * is read before spa_remove_init(). Specifically, the
327 * spa_config_object. (Note that although we already modified the
328 * spa_config_object in spa_sync_removing_state, that may not have
329 * modified all blocks of the object.)
331 dmu_object_info_t doi
;
332 VERIFY0(dmu_object_info(mos
, DMU_POOL_DIRECTORY_OBJECT
, &doi
));
333 for (uint64_t offset
= 0; offset
< doi
.doi_max_offset
; ) {
335 VERIFY0(dmu_buf_hold(mos
, DMU_POOL_DIRECTORY_OBJECT
,
336 offset
, FTAG
, &dbuf
, 0));
337 dmu_buf_will_dirty(dbuf
, tx
);
338 offset
+= dbuf
->db_size
;
339 dmu_buf_rele(dbuf
, FTAG
);
343 * Now that we've allocated the im_object, dirty the vdev to ensure
344 * that the object gets written to the config on disk.
346 vdev_config_dirty(vd
);
348 zfs_dbgmsg("starting removal thread for vdev %llu (%p) in txg %llu "
349 "im_obj=%llu", vd
->vdev_id
, vd
, dmu_tx_get_txg(tx
),
350 vic
->vic_mapping_object
);
352 spa_history_log_internal(spa
, "vdev remove started", tx
,
353 "%s vdev %llu %s", spa_name(spa
), vd
->vdev_id
,
354 (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
356 * Setting spa_vdev_removal causes subsequent frees to call
357 * free_from_removing_vdev(). Note that we don't need any locking
358 * because we are the sync thread, and metaslab_free_impl() is only
359 * called from syncing context (potentially from a zio taskq thread,
360 * but in any case only when there are outstanding free i/os, which
363 ASSERT3P(spa
->spa_vdev_removal
, ==, NULL
);
364 spa
->spa_vdev_removal
= svr
;
365 svr
->svr_thread
= thread_create(NULL
, 0,
366 spa_vdev_remove_thread
, spa
, 0, &p0
, TS_RUN
, minclsyspri
);
370 * When we are opening a pool, we must read the mapping for each
371 * indirect vdev in order from most recently removed to least
372 * recently removed. We do this because the blocks for the mapping
373 * of older indirect vdevs may be stored on more recently removed vdevs.
374 * In order to read each indirect mapping object, we must have
375 * initialized all more recently removed vdevs.
378 spa_remove_init(spa_t
*spa
)
382 error
= zap_lookup(spa
->spa_dsl_pool
->dp_meta_objset
,
383 DMU_POOL_DIRECTORY_OBJECT
,
384 DMU_POOL_REMOVING
, sizeof (uint64_t),
385 sizeof (spa
->spa_removing_phys
) / sizeof (uint64_t),
386 &spa
->spa_removing_phys
);
388 if (error
== ENOENT
) {
389 spa
->spa_removing_phys
.sr_state
= DSS_NONE
;
390 spa
->spa_removing_phys
.sr_removing_vdev
= -1;
391 spa
->spa_removing_phys
.sr_prev_indirect_vdev
= -1;
392 spa
->spa_indirect_vdevs_loaded
= B_TRUE
;
394 } else if (error
!= 0) {
398 if (spa
->spa_removing_phys
.sr_state
== DSS_SCANNING
) {
400 * We are currently removing a vdev. Create and
401 * initialize a spa_vdev_removal_t from the bonus
402 * buffer of the removing vdevs vdev_im_object, and
403 * initialize its partial mapping.
405 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
406 vdev_t
*vd
= vdev_lookup_top(spa
,
407 spa
->spa_removing_phys
.sr_removing_vdev
);
410 spa_config_exit(spa
, SCL_STATE
, FTAG
);
414 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
416 ASSERT(vdev_is_concrete(vd
));
417 spa_vdev_removal_t
*svr
= spa_vdev_removal_create(vd
);
418 ASSERT3U(svr
->svr_vdev_id
, ==, vd
->vdev_id
);
419 ASSERT(vd
->vdev_removing
);
421 vd
->vdev_indirect_mapping
= vdev_indirect_mapping_open(
422 spa
->spa_meta_objset
, vic
->vic_mapping_object
);
423 vd
->vdev_indirect_births
= vdev_indirect_births_open(
424 spa
->spa_meta_objset
, vic
->vic_births_object
);
425 spa_config_exit(spa
, SCL_STATE
, FTAG
);
427 spa
->spa_vdev_removal
= svr
;
430 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
431 uint64_t indirect_vdev_id
=
432 spa
->spa_removing_phys
.sr_prev_indirect_vdev
;
433 while (indirect_vdev_id
!= UINT64_MAX
) {
434 vdev_t
*vd
= vdev_lookup_top(spa
, indirect_vdev_id
);
435 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
437 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
438 vd
->vdev_indirect_mapping
= vdev_indirect_mapping_open(
439 spa
->spa_meta_objset
, vic
->vic_mapping_object
);
440 vd
->vdev_indirect_births
= vdev_indirect_births_open(
441 spa
->spa_meta_objset
, vic
->vic_births_object
);
443 indirect_vdev_id
= vic
->vic_prev_indirect_vdev
;
445 spa_config_exit(spa
, SCL_STATE
, FTAG
);
448 * Now that we've loaded all the indirect mappings, we can allow
449 * reads from other blocks (e.g. via predictive prefetch).
451 spa
->spa_indirect_vdevs_loaded
= B_TRUE
;
456 spa_restart_removal(spa_t
*spa
)
458 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
464 * In general when this function is called there is no
465 * removal thread running. The only scenario where this
466 * is not true is during spa_import() where this function
467 * is called twice [once from spa_import_impl() and
468 * spa_async_resume()]. Thus, in the scenario where we
469 * import a pool that has an ongoing removal we don't
470 * want to spawn a second thread.
472 if (svr
->svr_thread
!= NULL
)
475 if (!spa_writeable(spa
))
478 zfs_dbgmsg("restarting removal of %llu", svr
->svr_vdev_id
);
479 svr
->svr_thread
= thread_create(NULL
, 0, spa_vdev_remove_thread
, spa
,
480 0, &p0
, TS_RUN
, minclsyspri
);
484 * Process freeing from a device which is in the middle of being removed.
485 * We must handle this carefully so that we attempt to copy freed data,
486 * and we correctly free already-copied data.
489 free_from_removing_vdev(vdev_t
*vd
, uint64_t offset
, uint64_t size
)
491 spa_t
*spa
= vd
->vdev_spa
;
492 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
493 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
494 uint64_t txg
= spa_syncing_txg(spa
);
495 uint64_t max_offset_yet
= 0;
497 ASSERT(vd
->vdev_indirect_config
.vic_mapping_object
!= 0);
498 ASSERT3U(vd
->vdev_indirect_config
.vic_mapping_object
, ==,
499 vdev_indirect_mapping_object(vim
));
500 ASSERT3U(vd
->vdev_id
, ==, svr
->svr_vdev_id
);
502 mutex_enter(&svr
->svr_lock
);
505 * Remove the segment from the removing vdev's spacemap. This
506 * ensures that we will not attempt to copy this space (if the
507 * removal thread has not yet visited it), and also ensures
508 * that we know what is actually allocated on the new vdevs
509 * (needed if we cancel the removal).
511 * Note: we must do the metaslab_free_concrete() with the svr_lock
512 * held, so that the remove_thread can not load this metaslab and then
513 * visit this offset between the time that we metaslab_free_concrete()
514 * and when we check to see if it has been visited.
516 * Note: The checkpoint flag is set to false as having/taking
517 * a checkpoint and removing a device can't happen at the same
520 ASSERT(!spa_has_checkpoint(spa
));
521 metaslab_free_concrete(vd
, offset
, size
, B_FALSE
);
523 uint64_t synced_size
= 0;
524 uint64_t synced_offset
= 0;
525 uint64_t max_offset_synced
= vdev_indirect_mapping_max_offset(vim
);
526 if (offset
< max_offset_synced
) {
528 * The mapping for this offset is already on disk.
529 * Free from the new location.
531 * Note that we use svr_max_synced_offset because it is
532 * updated atomically with respect to the in-core mapping.
533 * By contrast, vim_max_offset is not.
535 * This block may be split between a synced entry and an
536 * in-flight or unvisited entry. Only process the synced
537 * portion of it here.
539 synced_size
= MIN(size
, max_offset_synced
- offset
);
540 synced_offset
= offset
;
542 ASSERT3U(max_offset_yet
, <=, max_offset_synced
);
543 max_offset_yet
= max_offset_synced
;
545 DTRACE_PROBE3(remove__free__synced
,
548 uint64_t, synced_size
);
551 offset
+= synced_size
;
555 * Look at all in-flight txgs starting from the currently syncing one
556 * and see if a section of this free is being copied. By starting from
557 * this txg and iterating forward, we might find that this region
558 * was copied in two different txgs and handle it appropriately.
560 for (int i
= 0; i
< TXG_CONCURRENT_STATES
; i
++) {
561 int txgoff
= (txg
+ i
) & TXG_MASK
;
562 if (size
> 0 && offset
< svr
->svr_max_offset_to_sync
[txgoff
]) {
564 * The mapping for this offset is in flight, and
565 * will be synced in txg+i.
567 uint64_t inflight_size
= MIN(size
,
568 svr
->svr_max_offset_to_sync
[txgoff
] - offset
);
570 DTRACE_PROBE4(remove__free__inflight
,
573 uint64_t, inflight_size
,
577 * We copy data in order of increasing offset.
578 * Therefore the max_offset_to_sync[] must increase
579 * (or be zero, indicating that nothing is being
580 * copied in that txg).
582 if (svr
->svr_max_offset_to_sync
[txgoff
] != 0) {
583 ASSERT3U(svr
->svr_max_offset_to_sync
[txgoff
],
586 svr
->svr_max_offset_to_sync
[txgoff
];
590 * We've already committed to copying this segment:
591 * we have allocated space elsewhere in the pool for
592 * it and have an IO outstanding to copy the data. We
593 * cannot free the space before the copy has
594 * completed, or else the copy IO might overwrite any
595 * new data. To free that space, we record the
596 * segment in the appropriate svr_frees tree and free
597 * the mapped space later, in the txg where we have
598 * completed the copy and synced the mapping (see
599 * vdev_mapping_sync).
601 range_tree_add(svr
->svr_frees
[txgoff
],
602 offset
, inflight_size
);
603 size
-= inflight_size
;
604 offset
+= inflight_size
;
607 * This space is already accounted for as being
608 * done, because it is being copied in txg+i.
609 * However, if i!=0, then it is being copied in
610 * a future txg. If we crash after this txg
611 * syncs but before txg+i syncs, then the space
612 * will be free. Therefore we must account
613 * for the space being done in *this* txg
614 * (when it is freed) rather than the future txg
615 * (when it will be copied).
617 ASSERT3U(svr
->svr_bytes_done
[txgoff
], >=,
619 svr
->svr_bytes_done
[txgoff
] -= inflight_size
;
620 svr
->svr_bytes_done
[txg
& TXG_MASK
] += inflight_size
;
623 ASSERT0(svr
->svr_max_offset_to_sync
[TXG_CLEAN(txg
) & TXG_MASK
]);
627 * The copy thread has not yet visited this offset. Ensure
631 DTRACE_PROBE3(remove__free__unvisited
,
636 if (svr
->svr_allocd_segs
!= NULL
)
637 range_tree_clear(svr
->svr_allocd_segs
, offset
, size
);
640 * Since we now do not need to copy this data, for
641 * accounting purposes we have done our job and can count
644 svr
->svr_bytes_done
[txg
& TXG_MASK
] += size
;
646 mutex_exit(&svr
->svr_lock
);
649 * Now that we have dropped svr_lock, process the synced portion
652 if (synced_size
> 0) {
653 vdev_indirect_mark_obsolete(vd
, synced_offset
, synced_size
);
656 * Note: this can only be called from syncing context,
657 * and the vdev_indirect_mapping is only changed from the
658 * sync thread, so we don't need svr_lock while doing
659 * metaslab_free_impl_cb.
661 boolean_t checkpoint
= B_FALSE
;
662 vdev_indirect_ops
.vdev_op_remap(vd
, synced_offset
, synced_size
,
663 metaslab_free_impl_cb
, &checkpoint
);
668 * Stop an active removal and update the spa_removing phys.
671 spa_finish_removal(spa_t
*spa
, dsl_scan_state_t state
, dmu_tx_t
*tx
)
673 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
674 ASSERT3U(dmu_tx_get_txg(tx
), ==, spa_syncing_txg(spa
));
676 /* Ensure the removal thread has completed before we free the svr. */
677 spa_vdev_remove_suspend(spa
);
679 ASSERT(state
== DSS_FINISHED
|| state
== DSS_CANCELED
);
681 if (state
== DSS_FINISHED
) {
682 spa_removing_phys_t
*srp
= &spa
->spa_removing_phys
;
683 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
684 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
686 if (srp
->sr_prev_indirect_vdev
!= -1) {
688 pvd
= vdev_lookup_top(spa
,
689 srp
->sr_prev_indirect_vdev
);
690 ASSERT3P(pvd
->vdev_ops
, ==, &vdev_indirect_ops
);
693 vic
->vic_prev_indirect_vdev
= srp
->sr_prev_indirect_vdev
;
694 srp
->sr_prev_indirect_vdev
= vd
->vdev_id
;
696 spa
->spa_removing_phys
.sr_state
= state
;
697 spa
->spa_removing_phys
.sr_end_time
= gethrestime_sec();
699 spa
->spa_vdev_removal
= NULL
;
700 spa_vdev_removal_destroy(svr
);
702 spa_sync_removing_state(spa
, tx
);
704 vdev_config_dirty(spa
->spa_root_vdev
);
708 free_mapped_segment_cb(void *arg
, uint64_t offset
, uint64_t size
)
711 vdev_indirect_mark_obsolete(vd
, offset
, size
);
712 boolean_t checkpoint
= B_FALSE
;
713 vdev_indirect_ops
.vdev_op_remap(vd
, offset
, size
,
714 metaslab_free_impl_cb
, &checkpoint
);
718 * On behalf of the removal thread, syncs an incremental bit more of
719 * the indirect mapping to disk and updates the in-memory mapping.
720 * Called as a sync task in every txg that the removal thread makes progress.
723 vdev_mapping_sync(void *arg
, dmu_tx_t
*tx
)
725 spa_vdev_removal_t
*svr
= arg
;
726 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
727 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
728 ASSERTV(vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
);
729 uint64_t txg
= dmu_tx_get_txg(tx
);
730 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
732 ASSERT(vic
->vic_mapping_object
!= 0);
733 ASSERT3U(txg
, ==, spa_syncing_txg(spa
));
735 vdev_indirect_mapping_add_entries(vim
,
736 &svr
->svr_new_segments
[txg
& TXG_MASK
], tx
);
737 vdev_indirect_births_add_entry(vd
->vdev_indirect_births
,
738 vdev_indirect_mapping_max_offset(vim
), dmu_tx_get_txg(tx
), tx
);
741 * Free the copied data for anything that was freed while the
742 * mapping entries were in flight.
744 mutex_enter(&svr
->svr_lock
);
745 range_tree_vacate(svr
->svr_frees
[txg
& TXG_MASK
],
746 free_mapped_segment_cb
, vd
);
747 ASSERT3U(svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
], >=,
748 vdev_indirect_mapping_max_offset(vim
));
749 svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] = 0;
750 mutex_exit(&svr
->svr_lock
);
752 spa_sync_removing_state(spa
, tx
);
755 typedef struct vdev_copy_segment_arg
{
757 dva_t
*vcsa_dest_dva
;
759 range_tree_t
*vcsa_obsolete_segs
;
760 } vdev_copy_segment_arg_t
;
763 unalloc_seg(void *arg
, uint64_t start
, uint64_t size
)
765 vdev_copy_segment_arg_t
*vcsa
= arg
;
766 spa_t
*spa
= vcsa
->vcsa_spa
;
767 blkptr_t bp
= { { { {0} } } };
769 BP_SET_BIRTH(&bp
, TXG_INITIAL
, TXG_INITIAL
);
770 BP_SET_LSIZE(&bp
, size
);
771 BP_SET_PSIZE(&bp
, size
);
772 BP_SET_COMPRESS(&bp
, ZIO_COMPRESS_OFF
);
773 BP_SET_CHECKSUM(&bp
, ZIO_CHECKSUM_OFF
);
774 BP_SET_TYPE(&bp
, DMU_OT_NONE
);
775 BP_SET_LEVEL(&bp
, 0);
776 BP_SET_DEDUP(&bp
, 0);
777 BP_SET_BYTEORDER(&bp
, ZFS_HOST_BYTEORDER
);
779 DVA_SET_VDEV(&bp
.blk_dva
[0], DVA_GET_VDEV(vcsa
->vcsa_dest_dva
));
780 DVA_SET_OFFSET(&bp
.blk_dva
[0],
781 DVA_GET_OFFSET(vcsa
->vcsa_dest_dva
) + start
);
782 DVA_SET_ASIZE(&bp
.blk_dva
[0], size
);
784 zio_free(spa
, vcsa
->vcsa_txg
, &bp
);
788 * All reads and writes associated with a call to spa_vdev_copy_segment()
792 spa_vdev_copy_segment_done(zio_t
*zio
)
794 vdev_copy_segment_arg_t
*vcsa
= zio
->io_private
;
796 range_tree_vacate(vcsa
->vcsa_obsolete_segs
,
798 range_tree_destroy(vcsa
->vcsa_obsolete_segs
);
799 kmem_free(vcsa
, sizeof (*vcsa
));
801 spa_config_exit(zio
->io_spa
, SCL_STATE
, zio
->io_spa
);
805 * The write of the new location is done.
808 spa_vdev_copy_segment_write_done(zio_t
*zio
)
810 vdev_copy_arg_t
*vca
= zio
->io_private
;
812 abd_free(zio
->io_abd
);
814 mutex_enter(&vca
->vca_lock
);
815 vca
->vca_outstanding_bytes
-= zio
->io_size
;
817 if (zio
->io_error
!= 0)
818 vca
->vca_write_error_bytes
+= zio
->io_size
;
820 cv_signal(&vca
->vca_cv
);
821 mutex_exit(&vca
->vca_lock
);
825 * The read of the old location is done. The parent zio is the write to
826 * the new location. Allow it to start.
829 spa_vdev_copy_segment_read_done(zio_t
*zio
)
831 vdev_copy_arg_t
*vca
= zio
->io_private
;
833 if (zio
->io_error
!= 0) {
834 mutex_enter(&vca
->vca_lock
);
835 vca
->vca_read_error_bytes
+= zio
->io_size
;
836 mutex_exit(&vca
->vca_lock
);
839 zio_nowait(zio_unique_parent(zio
));
843 * If the old and new vdevs are mirrors, we will read both sides of the old
844 * mirror, and write each copy to the corresponding side of the new mirror.
845 * If the old and new vdevs have a different number of children, we will do
846 * this as best as possible. Since we aren't verifying checksums, this
847 * ensures that as long as there's a good copy of the data, we'll have a
848 * good copy after the removal, even if there's silent damage to one side
849 * of the mirror. If we're removing a mirror that has some silent damage,
850 * we'll have exactly the same damage in the new location (assuming that
851 * the new location is also a mirror).
853 * We accomplish this by creating a tree of zio_t's, with as many writes as
854 * there are "children" of the new vdev (a non-redundant vdev counts as one
855 * child, a 2-way mirror has 2 children, etc). Each write has an associated
856 * read from a child of the old vdev. Typically there will be the same
857 * number of children of the old and new vdevs. However, if there are more
858 * children of the new vdev, some child(ren) of the old vdev will be issued
859 * multiple reads. If there are more children of the old vdev, some copies
862 * For example, the tree of zio_t's for a 2-way mirror is:
866 * write(new vdev, child 0) write(new vdev, child 1)
868 * read(old vdev, child 0) read(old vdev, child 1)
870 * Child zio's complete before their parents complete. However, zio's
871 * created with zio_vdev_child_io() may be issued before their children
872 * complete. In this case we need to make sure that the children (reads)
873 * complete before the parents (writes) are *issued*. We do this by not
874 * calling zio_nowait() on each write until its corresponding read has
877 * The spa_config_lock must be held while zio's created by
878 * zio_vdev_child_io() are in progress, to ensure that the vdev tree does
879 * not change (e.g. due to a concurrent "zpool attach/detach"). The "null"
880 * zio is needed to release the spa_config_lock after all the reads and
881 * writes complete. (Note that we can't grab the config lock for each read,
882 * because it is not reentrant - we could deadlock with a thread waiting
886 spa_vdev_copy_one_child(vdev_copy_arg_t
*vca
, zio_t
*nzio
,
887 vdev_t
*source_vd
, uint64_t source_offset
,
888 vdev_t
*dest_child_vd
, uint64_t dest_offset
, int dest_id
, uint64_t size
)
890 ASSERT3U(spa_config_held(nzio
->io_spa
, SCL_ALL
, RW_READER
), !=, 0);
893 * If the destination child in unwritable then there is no point
894 * in issuing the source reads which cannot be written.
896 if (!vdev_writeable(dest_child_vd
))
899 mutex_enter(&vca
->vca_lock
);
900 vca
->vca_outstanding_bytes
+= size
;
901 mutex_exit(&vca
->vca_lock
);
903 abd_t
*abd
= abd_alloc_for_io(size
, B_FALSE
);
905 vdev_t
*source_child_vd
= NULL
;
906 if (source_vd
->vdev_ops
== &vdev_mirror_ops
&& dest_id
!= -1) {
908 * Source and dest are both mirrors. Copy from the same
909 * child id as we are copying to (wrapping around if there
910 * are more dest children than source children). If the
911 * preferred source child is unreadable select another.
913 for (int i
= 0; i
< source_vd
->vdev_children
; i
++) {
914 source_child_vd
= source_vd
->vdev_child
[
915 (dest_id
+ i
) % source_vd
->vdev_children
];
916 if (vdev_readable(source_child_vd
))
920 source_child_vd
= source_vd
;
924 * There should always be at least one readable source child or
925 * the pool would be in a suspended state. Somehow selecting an
926 * unreadable child would result in IO errors, the removal process
927 * being cancelled, and the pool reverting to its pre-removal state.
929 ASSERT3P(source_child_vd
, !=, NULL
);
931 zio_t
*write_zio
= zio_vdev_child_io(nzio
, NULL
,
932 dest_child_vd
, dest_offset
, abd
, size
,
933 ZIO_TYPE_WRITE
, ZIO_PRIORITY_REMOVAL
,
935 spa_vdev_copy_segment_write_done
, vca
);
937 zio_nowait(zio_vdev_child_io(write_zio
, NULL
,
938 source_child_vd
, source_offset
, abd
, size
,
939 ZIO_TYPE_READ
, ZIO_PRIORITY_REMOVAL
,
941 spa_vdev_copy_segment_read_done
, vca
));
945 * Allocate a new location for this segment, and create the zio_t's to
946 * read from the old location and write to the new location.
949 spa_vdev_copy_segment(vdev_t
*vd
, range_tree_t
*segs
,
950 uint64_t maxalloc
, uint64_t txg
,
951 vdev_copy_arg_t
*vca
, zio_alloc_list_t
*zal
)
953 metaslab_group_t
*mg
= vd
->vdev_mg
;
954 spa_t
*spa
= vd
->vdev_spa
;
955 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
956 vdev_indirect_mapping_entry_t
*entry
;
958 uint64_t start
= range_tree_min(segs
);
960 ASSERT3U(maxalloc
, <=, SPA_MAXBLOCKSIZE
);
962 uint64_t size
= range_tree_span(segs
);
963 if (range_tree_span(segs
) > maxalloc
) {
965 * We can't allocate all the segments. Prefer to end
966 * the allocation at the end of a segment, thus avoiding
967 * additional split blocks.
971 search
.rs_start
= start
+ maxalloc
;
972 search
.rs_end
= search
.rs_start
;
973 range_seg_t
*rs
= avl_find(&segs
->rt_root
, &search
, &where
);
975 rs
= avl_nearest(&segs
->rt_root
, where
, AVL_BEFORE
);
977 rs
= AVL_PREV(&segs
->rt_root
, rs
);
980 size
= rs
->rs_end
- start
;
983 * There are no segments that end before maxalloc.
984 * I.e. the first segment is larger than maxalloc,
985 * so we must split it.
990 ASSERT3U(size
, <=, maxalloc
);
993 * An allocation class might not have any remaining vdevs or space
995 metaslab_class_t
*mc
= mg
->mg_class
;
996 if (mc
!= spa_normal_class(spa
) && mc
->mc_groups
<= 1)
997 mc
= spa_normal_class(spa
);
998 int error
= metaslab_alloc_dva(spa
, mc
, size
, &dst
, 0, NULL
, txg
, 0,
1000 if (error
== ENOSPC
&& mc
!= spa_normal_class(spa
)) {
1001 error
= metaslab_alloc_dva(spa
, spa_normal_class(spa
), size
,
1002 &dst
, 0, NULL
, txg
, 0, zal
, 0);
1008 * Determine the ranges that are not actually needed. Offsets are
1009 * relative to the start of the range to be copied (i.e. relative to the
1010 * local variable "start").
1012 range_tree_t
*obsolete_segs
= range_tree_create(NULL
, NULL
);
1014 range_seg_t
*rs
= avl_first(&segs
->rt_root
);
1015 ASSERT3U(rs
->rs_start
, ==, start
);
1016 uint64_t prev_seg_end
= rs
->rs_end
;
1017 while ((rs
= AVL_NEXT(&segs
->rt_root
, rs
)) != NULL
) {
1018 if (rs
->rs_start
>= start
+ size
) {
1021 range_tree_add(obsolete_segs
,
1022 prev_seg_end
- start
,
1023 rs
->rs_start
- prev_seg_end
);
1025 prev_seg_end
= rs
->rs_end
;
1027 /* We don't end in the middle of an obsolete range */
1028 ASSERT3U(start
+ size
, <=, prev_seg_end
);
1030 range_tree_clear(segs
, start
, size
);
1033 * We can't have any padding of the allocated size, otherwise we will
1034 * misunderstand what's allocated, and the size of the mapping.
1035 * The caller ensures this will be true by passing in a size that is
1036 * aligned to the worst (highest) ashift in the pool.
1038 ASSERT3U(DVA_GET_ASIZE(&dst
), ==, size
);
1040 entry
= kmem_zalloc(sizeof (vdev_indirect_mapping_entry_t
), KM_SLEEP
);
1041 DVA_MAPPING_SET_SRC_OFFSET(&entry
->vime_mapping
, start
);
1042 entry
->vime_mapping
.vimep_dst
= dst
;
1043 if (spa_feature_is_enabled(spa
, SPA_FEATURE_OBSOLETE_COUNTS
)) {
1044 entry
->vime_obsolete_count
= range_tree_space(obsolete_segs
);
1047 vdev_copy_segment_arg_t
*vcsa
= kmem_zalloc(sizeof (*vcsa
), KM_SLEEP
);
1048 vcsa
->vcsa_dest_dva
= &entry
->vime_mapping
.vimep_dst
;
1049 vcsa
->vcsa_obsolete_segs
= obsolete_segs
;
1050 vcsa
->vcsa_spa
= spa
;
1051 vcsa
->vcsa_txg
= txg
;
1054 * See comment before spa_vdev_copy_one_child().
1056 spa_config_enter(spa
, SCL_STATE
, spa
, RW_READER
);
1057 zio_t
*nzio
= zio_null(spa
->spa_txg_zio
[txg
& TXG_MASK
], spa
, NULL
,
1058 spa_vdev_copy_segment_done
, vcsa
, 0);
1059 vdev_t
*dest_vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dst
));
1060 if (dest_vd
->vdev_ops
== &vdev_mirror_ops
) {
1061 for (int i
= 0; i
< dest_vd
->vdev_children
; i
++) {
1062 vdev_t
*child
= dest_vd
->vdev_child
[i
];
1063 spa_vdev_copy_one_child(vca
, nzio
, vd
, start
,
1064 child
, DVA_GET_OFFSET(&dst
), i
, size
);
1067 spa_vdev_copy_one_child(vca
, nzio
, vd
, start
,
1068 dest_vd
, DVA_GET_OFFSET(&dst
), -1, size
);
1072 list_insert_tail(&svr
->svr_new_segments
[txg
& TXG_MASK
], entry
);
1073 ASSERT3U(start
+ size
, <=, vd
->vdev_ms_count
<< vd
->vdev_ms_shift
);
1074 vdev_dirty(vd
, 0, NULL
, txg
);
1080 * Complete the removal of a toplevel vdev. This is called as a
1081 * synctask in the same txg that we will sync out the new config (to the
1082 * MOS object) which indicates that this vdev is indirect.
1085 vdev_remove_complete_sync(void *arg
, dmu_tx_t
*tx
)
1087 spa_vdev_removal_t
*svr
= arg
;
1088 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1089 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1091 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
1093 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1094 ASSERT0(svr
->svr_bytes_done
[i
]);
1097 ASSERT3U(spa
->spa_removing_phys
.sr_copied
, ==,
1098 spa
->spa_removing_phys
.sr_to_copy
);
1100 vdev_destroy_spacemaps(vd
, tx
);
1102 /* destroy leaf zaps, if any */
1103 ASSERT3P(svr
->svr_zaplist
, !=, NULL
);
1104 for (nvpair_t
*pair
= nvlist_next_nvpair(svr
->svr_zaplist
, NULL
);
1106 pair
= nvlist_next_nvpair(svr
->svr_zaplist
, pair
)) {
1107 vdev_destroy_unlink_zap(vd
, fnvpair_value_uint64(pair
), tx
);
1109 fnvlist_free(svr
->svr_zaplist
);
1111 spa_finish_removal(dmu_tx_pool(tx
)->dp_spa
, DSS_FINISHED
, tx
);
1112 /* vd->vdev_path is not available here */
1113 spa_history_log_internal(spa
, "vdev remove completed", tx
,
1114 "%s vdev %llu", spa_name(spa
), vd
->vdev_id
);
1118 vdev_remove_enlist_zaps(vdev_t
*vd
, nvlist_t
*zlist
)
1120 ASSERT3P(zlist
, !=, NULL
);
1121 ASSERT3P(vd
->vdev_ops
, !=, &vdev_raidz_ops
);
1123 if (vd
->vdev_leaf_zap
!= 0) {
1125 (void) snprintf(zkey
, sizeof (zkey
), "%s-%llu",
1126 VDEV_REMOVAL_ZAP_OBJS
, (u_longlong_t
)vd
->vdev_leaf_zap
);
1127 fnvlist_add_uint64(zlist
, zkey
, vd
->vdev_leaf_zap
);
1130 for (uint64_t id
= 0; id
< vd
->vdev_children
; id
++) {
1131 vdev_remove_enlist_zaps(vd
->vdev_child
[id
], zlist
);
1136 vdev_remove_replace_with_indirect(vdev_t
*vd
, uint64_t txg
)
1140 spa_t
*spa
= vd
->vdev_spa
;
1141 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1144 * First, build a list of leaf zaps to be destroyed.
1145 * This is passed to the sync context thread,
1146 * which does the actual unlinking.
1148 svr
->svr_zaplist
= fnvlist_alloc();
1149 vdev_remove_enlist_zaps(vd
, svr
->svr_zaplist
);
1151 ivd
= vdev_add_parent(vd
, &vdev_indirect_ops
);
1152 ivd
->vdev_removing
= 0;
1154 vd
->vdev_leaf_zap
= 0;
1156 vdev_remove_child(ivd
, vd
);
1157 vdev_compact_children(ivd
);
1159 ASSERT(!list_link_active(&vd
->vdev_state_dirty_node
));
1161 mutex_enter(&svr
->svr_lock
);
1162 svr
->svr_thread
= NULL
;
1163 cv_broadcast(&svr
->svr_cv
);
1164 mutex_exit(&svr
->svr_lock
);
1166 /* After this, we can not use svr. */
1167 tx
= dmu_tx_create_assigned(spa
->spa_dsl_pool
, txg
);
1168 dsl_sync_task_nowait(spa
->spa_dsl_pool
, vdev_remove_complete_sync
, svr
,
1169 0, ZFS_SPACE_CHECK_NONE
, tx
);
1174 * Complete the removal of a toplevel vdev. This is called in open
1175 * context by the removal thread after we have copied all vdev's data.
1178 vdev_remove_complete(spa_t
*spa
)
1183 * Wait for any deferred frees to be synced before we call
1184 * vdev_metaslab_fini()
1186 txg_wait_synced(spa
->spa_dsl_pool
, 0);
1187 txg
= spa_vdev_enter(spa
);
1188 vdev_t
*vd
= vdev_lookup_top(spa
, spa
->spa_vdev_removal
->svr_vdev_id
);
1190 sysevent_t
*ev
= spa_event_create(spa
, vd
, NULL
,
1191 ESC_ZFS_VDEV_REMOVE_DEV
);
1193 zfs_dbgmsg("finishing device removal for vdev %llu in txg %llu",
1197 * Discard allocation state.
1199 if (vd
->vdev_mg
!= NULL
) {
1200 vdev_metaslab_fini(vd
);
1201 metaslab_group_destroy(vd
->vdev_mg
);
1204 ASSERT0(vd
->vdev_stat
.vs_space
);
1205 ASSERT0(vd
->vdev_stat
.vs_dspace
);
1207 vdev_remove_replace_with_indirect(vd
, txg
);
1210 * We now release the locks, allowing spa_sync to run and finish the
1211 * removal via vdev_remove_complete_sync in syncing context.
1213 * Note that we hold on to the vdev_t that has been replaced. Since
1214 * it isn't part of the vdev tree any longer, it can't be concurrently
1215 * manipulated, even while we don't have the config lock.
1217 (void) spa_vdev_exit(spa
, NULL
, txg
, 0);
1220 * Top ZAP should have been transferred to the indirect vdev in
1221 * vdev_remove_replace_with_indirect.
1223 ASSERT0(vd
->vdev_top_zap
);
1226 * Leaf ZAP should have been moved in vdev_remove_replace_with_indirect.
1228 ASSERT0(vd
->vdev_leaf_zap
);
1230 txg
= spa_vdev_enter(spa
);
1231 (void) vdev_label_init(vd
, 0, VDEV_LABEL_REMOVE
);
1233 * Request to update the config and the config cachefile.
1235 vdev_config_dirty(spa
->spa_root_vdev
);
1236 (void) spa_vdev_exit(spa
, vd
, txg
, 0);
1243 * Evacuates a segment of size at most max_alloc from the vdev
1244 * via repeated calls to spa_vdev_copy_segment. If an allocation
1245 * fails, the pool is probably too fragmented to handle such a
1246 * large size, so decrease max_alloc so that the caller will not try
1247 * this size again this txg.
1250 spa_vdev_copy_impl(vdev_t
*vd
, spa_vdev_removal_t
*svr
, vdev_copy_arg_t
*vca
,
1251 uint64_t *max_alloc
, dmu_tx_t
*tx
)
1253 uint64_t txg
= dmu_tx_get_txg(tx
);
1254 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1256 mutex_enter(&svr
->svr_lock
);
1259 * Determine how big of a chunk to copy. We can allocate up
1260 * to max_alloc bytes, and we can span up to vdev_removal_max_span
1261 * bytes of unallocated space at a time. "segs" will track the
1262 * allocated segments that we are copying. We may also be copying
1263 * free segments (of up to vdev_removal_max_span bytes).
1265 range_tree_t
*segs
= range_tree_create(NULL
, NULL
);
1267 range_seg_t
*rs
= range_tree_first(svr
->svr_allocd_segs
);
1272 uint64_t seg_length
;
1274 if (range_tree_is_empty(segs
)) {
1275 /* need to truncate the first seg based on max_alloc */
1277 MIN(rs
->rs_end
- rs
->rs_start
, *max_alloc
);
1279 if (rs
->rs_start
- range_tree_max(segs
) >
1280 vdev_removal_max_span
) {
1282 * Including this segment would cause us to
1283 * copy a larger unneeded chunk than is allowed.
1286 } else if (rs
->rs_end
- range_tree_min(segs
) >
1289 * This additional segment would extend past
1290 * max_alloc. Rather than splitting this
1291 * segment, leave it for the next mapping.
1295 seg_length
= rs
->rs_end
- rs
->rs_start
;
1299 range_tree_add(segs
, rs
->rs_start
, seg_length
);
1300 range_tree_remove(svr
->svr_allocd_segs
,
1301 rs
->rs_start
, seg_length
);
1304 if (range_tree_is_empty(segs
)) {
1305 mutex_exit(&svr
->svr_lock
);
1306 range_tree_destroy(segs
);
1310 if (svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] == 0) {
1311 dsl_sync_task_nowait(dmu_tx_pool(tx
), vdev_mapping_sync
,
1312 svr
, 0, ZFS_SPACE_CHECK_NONE
, tx
);
1315 svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] = range_tree_max(segs
);
1318 * Note: this is the amount of *allocated* space
1319 * that we are taking care of each txg.
1321 svr
->svr_bytes_done
[txg
& TXG_MASK
] += range_tree_space(segs
);
1323 mutex_exit(&svr
->svr_lock
);
1325 zio_alloc_list_t zal
;
1326 metaslab_trace_init(&zal
);
1327 uint64_t thismax
= SPA_MAXBLOCKSIZE
;
1328 while (!range_tree_is_empty(segs
)) {
1329 int error
= spa_vdev_copy_segment(vd
,
1330 segs
, thismax
, txg
, vca
, &zal
);
1332 if (error
== ENOSPC
) {
1334 * Cut our segment in half, and don't try this
1335 * segment size again this txg. Note that the
1336 * allocation size must be aligned to the highest
1337 * ashift in the pool, so that the allocation will
1338 * not be padded out to a multiple of the ashift,
1339 * which could cause us to think that this mapping
1340 * is larger than we intended.
1342 ASSERT3U(spa
->spa_max_ashift
, >=, SPA_MINBLOCKSHIFT
);
1343 ASSERT3U(spa
->spa_max_ashift
, ==, spa
->spa_min_ashift
);
1344 uint64_t attempted
=
1345 MIN(range_tree_span(segs
), thismax
);
1346 thismax
= P2ROUNDUP(attempted
/ 2,
1347 1 << spa
->spa_max_ashift
);
1349 * The minimum-size allocation can not fail.
1351 ASSERT3U(attempted
, >, 1 << spa
->spa_max_ashift
);
1352 *max_alloc
= attempted
- (1 << spa
->spa_max_ashift
);
1357 * We've performed an allocation, so reset the
1360 metaslab_trace_fini(&zal
);
1361 metaslab_trace_init(&zal
);
1364 metaslab_trace_fini(&zal
);
1365 range_tree_destroy(segs
);
1369 * The removal thread operates in open context. It iterates over all
1370 * allocated space in the vdev, by loading each metaslab's spacemap.
1371 * For each contiguous segment of allocated space (capping the segment
1372 * size at SPA_MAXBLOCKSIZE), we:
1373 * - Allocate space for it on another vdev.
1374 * - Create a new mapping from the old location to the new location
1375 * (as a record in svr_new_segments).
1376 * - Initiate a physical read zio to get the data off the removing disk.
1377 * - In the read zio's done callback, initiate a physical write zio to
1378 * write it to the new vdev.
1379 * Note that all of this will take effect when a particular TXG syncs.
1380 * The sync thread ensures that all the phys reads and writes for the syncing
1381 * TXG have completed (see spa_txg_zio) and writes the new mappings to disk
1382 * (see vdev_mapping_sync()).
1385 spa_vdev_remove_thread(void *arg
)
1388 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1389 vdev_copy_arg_t vca
;
1390 uint64_t max_alloc
= zfs_remove_max_segment
;
1391 uint64_t last_txg
= 0;
1393 spa_config_enter(spa
, SCL_CONFIG
, FTAG
, RW_READER
);
1394 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1395 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
1396 uint64_t start_offset
= vdev_indirect_mapping_max_offset(vim
);
1398 ASSERT3P(vd
->vdev_ops
, !=, &vdev_indirect_ops
);
1399 ASSERT(vdev_is_concrete(vd
));
1400 ASSERT(vd
->vdev_removing
);
1401 ASSERT(vd
->vdev_indirect_config
.vic_mapping_object
!= 0);
1402 ASSERT(vim
!= NULL
);
1404 mutex_init(&vca
.vca_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1405 cv_init(&vca
.vca_cv
, NULL
, CV_DEFAULT
, NULL
);
1406 vca
.vca_outstanding_bytes
= 0;
1407 vca
.vca_read_error_bytes
= 0;
1408 vca
.vca_write_error_bytes
= 0;
1410 mutex_enter(&svr
->svr_lock
);
1413 * Start from vim_max_offset so we pick up where we left off
1414 * if we are restarting the removal after opening the pool.
1417 for (msi
= start_offset
>> vd
->vdev_ms_shift
;
1418 msi
< vd
->vdev_ms_count
&& !svr
->svr_thread_exit
; msi
++) {
1419 metaslab_t
*msp
= vd
->vdev_ms
[msi
];
1420 ASSERT3U(msi
, <=, vd
->vdev_ms_count
);
1422 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1424 mutex_enter(&msp
->ms_sync_lock
);
1425 mutex_enter(&msp
->ms_lock
);
1428 * Assert nothing in flight -- ms_*tree is empty.
1430 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1431 ASSERT0(range_tree_space(msp
->ms_allocating
[i
]));
1435 * If the metaslab has ever been allocated from (ms_sm!=NULL),
1436 * read the allocated segments from the space map object
1437 * into svr_allocd_segs. Since we do this while holding
1438 * svr_lock and ms_sync_lock, concurrent frees (which
1439 * would have modified the space map) will wait for us
1440 * to finish loading the spacemap, and then take the
1441 * appropriate action (see free_from_removing_vdev()).
1443 if (msp
->ms_sm
!= NULL
) {
1444 space_map_t
*sm
= NULL
;
1447 * We have to open a new space map here, because
1448 * ms_sm's sm_length and sm_alloc may not reflect
1449 * what's in the object contents, if we are in between
1450 * metaslab_sync() and metaslab_sync_done().
1452 VERIFY0(space_map_open(&sm
,
1453 spa
->spa_dsl_pool
->dp_meta_objset
,
1454 msp
->ms_sm
->sm_object
, msp
->ms_sm
->sm_start
,
1455 msp
->ms_sm
->sm_size
, msp
->ms_sm
->sm_shift
));
1456 space_map_update(sm
);
1457 VERIFY0(space_map_load(sm
, svr
->svr_allocd_segs
,
1459 space_map_close(sm
);
1461 range_tree_walk(msp
->ms_freeing
,
1462 range_tree_remove
, svr
->svr_allocd_segs
);
1465 * When we are resuming from a paused removal (i.e.
1466 * when importing a pool with a removal in progress),
1467 * discard any state that we have already processed.
1469 range_tree_clear(svr
->svr_allocd_segs
, 0, start_offset
);
1471 mutex_exit(&msp
->ms_lock
);
1472 mutex_exit(&msp
->ms_sync_lock
);
1475 zfs_dbgmsg("copying %llu segments for metaslab %llu",
1476 avl_numnodes(&svr
->svr_allocd_segs
->rt_root
),
1479 while (!svr
->svr_thread_exit
&&
1480 !range_tree_is_empty(svr
->svr_allocd_segs
)) {
1482 mutex_exit(&svr
->svr_lock
);
1485 * We need to periodically drop the config lock so that
1486 * writers can get in. Additionally, we can't wait
1487 * for a txg to sync while holding a config lock
1488 * (since a waiting writer could cause a 3-way deadlock
1489 * with the sync thread, which also gets a config
1490 * lock for reader). So we can't hold the config lock
1491 * while calling dmu_tx_assign().
1493 spa_config_exit(spa
, SCL_CONFIG
, FTAG
);
1496 * This delay will pause the removal around the point
1497 * specified by zfs_removal_suspend_progress. We do this
1498 * solely from the test suite or during debugging.
1500 uint64_t bytes_copied
=
1501 spa
->spa_removing_phys
.sr_copied
;
1502 for (int i
= 0; i
< TXG_SIZE
; i
++)
1503 bytes_copied
+= svr
->svr_bytes_done
[i
];
1504 while (zfs_removal_suspend_progress
&&
1505 !svr
->svr_thread_exit
)
1508 mutex_enter(&vca
.vca_lock
);
1509 while (vca
.vca_outstanding_bytes
>
1510 zfs_remove_max_copy_bytes
) {
1511 cv_wait(&vca
.vca_cv
, &vca
.vca_lock
);
1513 mutex_exit(&vca
.vca_lock
);
1516 dmu_tx_create_dd(spa_get_dsl(spa
)->dp_mos_dir
);
1517 dmu_tx_hold_space(tx
, SPA_MAXBLOCKSIZE
);
1518 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
1519 uint64_t txg
= dmu_tx_get_txg(tx
);
1522 * Reacquire the vdev_config lock. The vdev_t
1523 * that we're removing may have changed, e.g. due
1524 * to a vdev_attach or vdev_detach.
1526 spa_config_enter(spa
, SCL_CONFIG
, FTAG
, RW_READER
);
1527 vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1529 if (txg
!= last_txg
)
1530 max_alloc
= zfs_remove_max_segment
;
1533 spa_vdev_copy_impl(vd
, svr
, &vca
, &max_alloc
, tx
);
1536 mutex_enter(&svr
->svr_lock
);
1539 mutex_enter(&vca
.vca_lock
);
1540 if (zfs_removal_ignore_errors
== 0 &&
1541 (vca
.vca_read_error_bytes
> 0 ||
1542 vca
.vca_write_error_bytes
> 0)) {
1543 svr
->svr_thread_exit
= B_TRUE
;
1545 mutex_exit(&vca
.vca_lock
);
1548 mutex_exit(&svr
->svr_lock
);
1550 spa_config_exit(spa
, SCL_CONFIG
, FTAG
);
1553 * Wait for all copies to finish before cleaning up the vca.
1555 txg_wait_synced(spa
->spa_dsl_pool
, 0);
1556 ASSERT0(vca
.vca_outstanding_bytes
);
1558 mutex_destroy(&vca
.vca_lock
);
1559 cv_destroy(&vca
.vca_cv
);
1561 if (svr
->svr_thread_exit
) {
1562 mutex_enter(&svr
->svr_lock
);
1563 range_tree_vacate(svr
->svr_allocd_segs
, NULL
, NULL
);
1564 svr
->svr_thread
= NULL
;
1565 cv_broadcast(&svr
->svr_cv
);
1566 mutex_exit(&svr
->svr_lock
);
1569 * During the removal process an unrecoverable read or write
1570 * error was encountered. The removal process must be
1571 * cancelled or this damage may become permanent.
1573 if (zfs_removal_ignore_errors
== 0 &&
1574 (vca
.vca_read_error_bytes
> 0 ||
1575 vca
.vca_write_error_bytes
> 0)) {
1576 zfs_dbgmsg("canceling removal due to IO errors: "
1577 "[read_error_bytes=%llu] [write_error_bytes=%llu]",
1578 vca
.vca_read_error_bytes
,
1579 vca
.vca_write_error_bytes
);
1580 spa_vdev_remove_cancel_impl(spa
);
1583 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1584 vdev_remove_complete(spa
);
1589 spa_vdev_remove_suspend(spa_t
*spa
)
1591 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1596 mutex_enter(&svr
->svr_lock
);
1597 svr
->svr_thread_exit
= B_TRUE
;
1598 while (svr
->svr_thread
!= NULL
)
1599 cv_wait(&svr
->svr_cv
, &svr
->svr_lock
);
1600 svr
->svr_thread_exit
= B_FALSE
;
1601 mutex_exit(&svr
->svr_lock
);
1606 spa_vdev_remove_cancel_check(void *arg
, dmu_tx_t
*tx
)
1608 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1610 if (spa
->spa_vdev_removal
== NULL
)
1611 return (ENOTACTIVE
);
1616 * Cancel a removal by freeing all entries from the partial mapping
1617 * and marking the vdev as no longer being removing.
1621 spa_vdev_remove_cancel_sync(void *arg
, dmu_tx_t
*tx
)
1623 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1624 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1625 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1626 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
1627 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
1628 objset_t
*mos
= spa
->spa_meta_objset
;
1630 ASSERT3P(svr
->svr_thread
, ==, NULL
);
1632 spa_feature_decr(spa
, SPA_FEATURE_DEVICE_REMOVAL
, tx
);
1634 boolean_t are_precise
;
1635 VERIFY0(vdev_obsolete_counts_are_precise(vd
, &are_precise
));
1637 spa_feature_decr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
1638 VERIFY0(zap_remove(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
1639 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE
, tx
));
1642 uint64_t obsolete_sm_object
;
1643 VERIFY0(vdev_obsolete_sm_object(vd
, &obsolete_sm_object
));
1644 if (obsolete_sm_object
!= 0) {
1645 ASSERT(vd
->vdev_obsolete_sm
!= NULL
);
1646 ASSERT3U(obsolete_sm_object
, ==,
1647 space_map_object(vd
->vdev_obsolete_sm
));
1649 space_map_free(vd
->vdev_obsolete_sm
, tx
);
1650 VERIFY0(zap_remove(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
1651 VDEV_TOP_ZAP_INDIRECT_OBSOLETE_SM
, tx
));
1652 space_map_close(vd
->vdev_obsolete_sm
);
1653 vd
->vdev_obsolete_sm
= NULL
;
1654 spa_feature_decr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
1656 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1657 ASSERT(list_is_empty(&svr
->svr_new_segments
[i
]));
1658 ASSERT3U(svr
->svr_max_offset_to_sync
[i
], <=,
1659 vdev_indirect_mapping_max_offset(vim
));
1662 for (uint64_t msi
= 0; msi
< vd
->vdev_ms_count
; msi
++) {
1663 metaslab_t
*msp
= vd
->vdev_ms
[msi
];
1665 if (msp
->ms_start
>= vdev_indirect_mapping_max_offset(vim
))
1668 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1670 mutex_enter(&msp
->ms_lock
);
1673 * Assert nothing in flight -- ms_*tree is empty.
1675 for (int i
= 0; i
< TXG_SIZE
; i
++)
1676 ASSERT0(range_tree_space(msp
->ms_allocating
[i
]));
1677 for (int i
= 0; i
< TXG_DEFER_SIZE
; i
++)
1678 ASSERT0(range_tree_space(msp
->ms_defer
[i
]));
1679 ASSERT0(range_tree_space(msp
->ms_freed
));
1681 if (msp
->ms_sm
!= NULL
) {
1683 * Assert that the in-core spacemap has the same
1684 * length as the on-disk one, so we can use the
1685 * existing in-core spacemap to load it from disk.
1687 ASSERT3U(msp
->ms_sm
->sm_alloc
, ==,
1688 msp
->ms_sm
->sm_phys
->smp_alloc
);
1689 ASSERT3U(msp
->ms_sm
->sm_length
, ==,
1690 msp
->ms_sm
->sm_phys
->smp_objsize
);
1692 mutex_enter(&svr
->svr_lock
);
1693 VERIFY0(space_map_load(msp
->ms_sm
,
1694 svr
->svr_allocd_segs
, SM_ALLOC
));
1695 range_tree_walk(msp
->ms_freeing
,
1696 range_tree_remove
, svr
->svr_allocd_segs
);
1699 * Clear everything past what has been synced,
1700 * because we have not allocated mappings for it yet.
1702 uint64_t syncd
= vdev_indirect_mapping_max_offset(vim
);
1703 uint64_t sm_end
= msp
->ms_sm
->sm_start
+
1704 msp
->ms_sm
->sm_size
;
1706 range_tree_clear(svr
->svr_allocd_segs
,
1707 syncd
, sm_end
- syncd
);
1709 mutex_exit(&svr
->svr_lock
);
1711 mutex_exit(&msp
->ms_lock
);
1713 mutex_enter(&svr
->svr_lock
);
1714 range_tree_vacate(svr
->svr_allocd_segs
,
1715 free_mapped_segment_cb
, vd
);
1716 mutex_exit(&svr
->svr_lock
);
1720 * Note: this must happen after we invoke free_mapped_segment_cb,
1721 * because it adds to the obsolete_segments.
1723 range_tree_vacate(vd
->vdev_obsolete_segments
, NULL
, NULL
);
1725 ASSERT3U(vic
->vic_mapping_object
, ==,
1726 vdev_indirect_mapping_object(vd
->vdev_indirect_mapping
));
1727 vdev_indirect_mapping_close(vd
->vdev_indirect_mapping
);
1728 vd
->vdev_indirect_mapping
= NULL
;
1729 vdev_indirect_mapping_free(mos
, vic
->vic_mapping_object
, tx
);
1730 vic
->vic_mapping_object
= 0;
1732 ASSERT3U(vic
->vic_births_object
, ==,
1733 vdev_indirect_births_object(vd
->vdev_indirect_births
));
1734 vdev_indirect_births_close(vd
->vdev_indirect_births
);
1735 vd
->vdev_indirect_births
= NULL
;
1736 vdev_indirect_births_free(mos
, vic
->vic_births_object
, tx
);
1737 vic
->vic_births_object
= 0;
1740 * We may have processed some frees from the removing vdev in this
1741 * txg, thus increasing svr_bytes_done; discard that here to
1742 * satisfy the assertions in spa_vdev_removal_destroy().
1743 * Note that future txg's can not have any bytes_done, because
1744 * future TXG's are only modified from open context, and we have
1745 * already shut down the copying thread.
1747 svr
->svr_bytes_done
[dmu_tx_get_txg(tx
) & TXG_MASK
] = 0;
1748 spa_finish_removal(spa
, DSS_CANCELED
, tx
);
1750 vd
->vdev_removing
= B_FALSE
;
1751 vdev_config_dirty(vd
);
1753 zfs_dbgmsg("canceled device removal for vdev %llu in %llu",
1754 vd
->vdev_id
, dmu_tx_get_txg(tx
));
1755 spa_history_log_internal(spa
, "vdev remove canceled", tx
,
1756 "%s vdev %llu %s", spa_name(spa
),
1757 vd
->vdev_id
, (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
1761 spa_vdev_remove_cancel_impl(spa_t
*spa
)
1763 uint64_t vdid
= spa
->spa_vdev_removal
->svr_vdev_id
;
1765 int error
= dsl_sync_task(spa
->spa_name
, spa_vdev_remove_cancel_check
,
1766 spa_vdev_remove_cancel_sync
, NULL
, 0,
1767 ZFS_SPACE_CHECK_EXTRA_RESERVED
);
1770 spa_config_enter(spa
, SCL_ALLOC
| SCL_VDEV
, FTAG
, RW_WRITER
);
1771 vdev_t
*vd
= vdev_lookup_top(spa
, vdid
);
1772 metaslab_group_activate(vd
->vdev_mg
);
1773 spa_config_exit(spa
, SCL_ALLOC
| SCL_VDEV
, FTAG
);
1780 spa_vdev_remove_cancel(spa_t
*spa
)
1782 spa_vdev_remove_suspend(spa
);
1784 if (spa
->spa_vdev_removal
== NULL
)
1785 return (ENOTACTIVE
);
1787 return (spa_vdev_remove_cancel_impl(spa
));
1791 * Called every sync pass of every txg if there's a svr.
1794 svr_sync(spa_t
*spa
, dmu_tx_t
*tx
)
1796 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1797 int txgoff
= dmu_tx_get_txg(tx
) & TXG_MASK
;
1800 * This check is necessary so that we do not dirty the
1801 * DIRECTORY_OBJECT via spa_sync_removing_state() when there
1802 * is nothing to do. Dirtying it every time would prevent us
1803 * from syncing-to-convergence.
1805 if (svr
->svr_bytes_done
[txgoff
] == 0)
1809 * Update progress accounting.
1811 spa
->spa_removing_phys
.sr_copied
+= svr
->svr_bytes_done
[txgoff
];
1812 svr
->svr_bytes_done
[txgoff
] = 0;
1814 spa_sync_removing_state(spa
, tx
);
1818 vdev_remove_make_hole_and_free(vdev_t
*vd
)
1820 uint64_t id
= vd
->vdev_id
;
1821 spa_t
*spa
= vd
->vdev_spa
;
1822 vdev_t
*rvd
= spa
->spa_root_vdev
;
1823 boolean_t last_vdev
= (id
== (rvd
->vdev_children
- 1));
1825 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
1826 ASSERT(spa_config_held(spa
, SCL_ALL
, RW_WRITER
) == SCL_ALL
);
1831 vdev_compact_children(rvd
);
1833 vd
= vdev_alloc_common(spa
, id
, 0, &vdev_hole_ops
);
1834 vdev_add_child(rvd
, vd
);
1836 vdev_config_dirty(rvd
);
1839 * Reassess the health of our root vdev.
1845 * Remove a log device. The config lock is held for the specified TXG.
1848 spa_vdev_remove_log(vdev_t
*vd
, uint64_t *txg
)
1850 metaslab_group_t
*mg
= vd
->vdev_mg
;
1851 spa_t
*spa
= vd
->vdev_spa
;
1854 ASSERT(vd
->vdev_islog
);
1855 ASSERT(vd
== vd
->vdev_top
);
1858 * Stop allocating from this vdev.
1860 metaslab_group_passivate(mg
);
1863 * Wait for the youngest allocations and frees to sync,
1864 * and then wait for the deferral of those frees to finish.
1866 spa_vdev_config_exit(spa
, NULL
,
1867 *txg
+ TXG_CONCURRENT_STATES
+ TXG_DEFER_SIZE
, 0, FTAG
);
1870 * Evacuate the device. We don't hold the config lock as writer
1871 * since we need to do I/O but we do keep the
1872 * spa_namespace_lock held. Once this completes the device
1873 * should no longer have any blocks allocated on it.
1875 if (vd
->vdev_islog
) {
1876 if (vd
->vdev_stat
.vs_alloc
!= 0)
1877 error
= spa_reset_logs(spa
);
1880 *txg
= spa_vdev_config_enter(spa
);
1883 metaslab_group_activate(mg
);
1886 ASSERT0(vd
->vdev_stat
.vs_alloc
);
1889 * The evacuation succeeded. Remove any remaining MOS metadata
1890 * associated with this vdev, and wait for these changes to sync.
1892 vd
->vdev_removing
= B_TRUE
;
1894 vdev_dirty_leaves(vd
, VDD_DTL
, *txg
);
1895 vdev_config_dirty(vd
);
1897 spa_vdev_config_exit(spa
, NULL
, *txg
, 0, FTAG
);
1899 *txg
= spa_vdev_config_enter(spa
);
1901 sysevent_t
*ev
= spa_event_create(spa
, vd
, NULL
,
1902 ESC_ZFS_VDEV_REMOVE_DEV
);
1903 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
1904 ASSERT(spa_config_held(spa
, SCL_ALL
, RW_WRITER
) == SCL_ALL
);
1906 /* The top ZAP should have been destroyed by vdev_remove_empty. */
1907 ASSERT0(vd
->vdev_top_zap
);
1908 /* The leaf ZAP should have been destroyed by vdev_dtl_sync. */
1909 ASSERT0(vd
->vdev_leaf_zap
);
1911 (void) vdev_label_init(vd
, 0, VDEV_LABEL_REMOVE
);
1913 if (list_link_active(&vd
->vdev_state_dirty_node
))
1914 vdev_state_clean(vd
);
1915 if (list_link_active(&vd
->vdev_config_dirty_node
))
1916 vdev_config_clean(vd
);
1919 * Clean up the vdev namespace.
1921 vdev_remove_make_hole_and_free(vd
);
1930 spa_vdev_remove_top_check(vdev_t
*vd
)
1932 spa_t
*spa
= vd
->vdev_spa
;
1934 if (vd
!= vd
->vdev_top
)
1935 return (SET_ERROR(ENOTSUP
));
1937 if (!spa_feature_is_enabled(spa
, SPA_FEATURE_DEVICE_REMOVAL
))
1938 return (SET_ERROR(ENOTSUP
));
1940 /* available space in the pool's normal class */
1941 uint64_t available
= dsl_dir_space_available(
1942 spa
->spa_dsl_pool
->dp_root_dir
, NULL
, 0, B_TRUE
);
1944 metaslab_class_t
*mc
= vd
->vdev_mg
->mg_class
;
1947 * When removing a vdev from an allocation class that has
1948 * remaining vdevs, include available space from the class.
1950 if (mc
!= spa_normal_class(spa
) && mc
->mc_groups
> 1) {
1951 uint64_t class_avail
= metaslab_class_get_space(mc
) -
1952 metaslab_class_get_alloc(mc
);
1954 /* add class space, adjusted for overhead */
1955 available
+= (class_avail
* 94) / 100;
1959 * There has to be enough free space to remove the
1960 * device and leave double the "slop" space (i.e. we
1961 * must leave at least 3% of the pool free, in addition to
1962 * the normal slop space).
1964 if (available
< vd
->vdev_stat
.vs_dspace
+ spa_get_slop_space(spa
)) {
1965 return (SET_ERROR(ENOSPC
));
1969 * There can not be a removal in progress.
1971 if (spa
->spa_removing_phys
.sr_state
== DSS_SCANNING
)
1972 return (SET_ERROR(EBUSY
));
1975 * The device must have all its data.
1977 if (!vdev_dtl_empty(vd
, DTL_MISSING
) ||
1978 !vdev_dtl_empty(vd
, DTL_OUTAGE
))
1979 return (SET_ERROR(EBUSY
));
1982 * The device must be healthy.
1984 if (!vdev_readable(vd
))
1985 return (SET_ERROR(EIO
));
1988 * All vdevs in normal class must have the same ashift.
1990 if (spa
->spa_max_ashift
!= spa
->spa_min_ashift
) {
1991 return (SET_ERROR(EINVAL
));
1995 * All vdevs in normal class must have the same ashift
1998 vdev_t
*rvd
= spa
->spa_root_vdev
;
1999 int num_indirect
= 0;
2000 for (uint64_t id
= 0; id
< rvd
->vdev_children
; id
++) {
2001 vdev_t
*cvd
= rvd
->vdev_child
[id
];
2002 if (cvd
->vdev_ashift
!= 0 && !cvd
->vdev_islog
)
2003 ASSERT3U(cvd
->vdev_ashift
, ==, spa
->spa_max_ashift
);
2004 if (cvd
->vdev_ops
== &vdev_indirect_ops
)
2006 if (!vdev_is_concrete(cvd
))
2008 if (cvd
->vdev_ops
== &vdev_raidz_ops
)
2009 return (SET_ERROR(EINVAL
));
2011 * Need the mirror to be mirror of leaf vdevs only
2013 if (cvd
->vdev_ops
== &vdev_mirror_ops
) {
2014 for (uint64_t cid
= 0;
2015 cid
< cvd
->vdev_children
; cid
++) {
2016 if (!cvd
->vdev_child
[cid
]->vdev_ops
->
2018 return (SET_ERROR(EINVAL
));
2027 * Initiate removal of a top-level vdev, reducing the total space in the pool.
2028 * The config lock is held for the specified TXG. Once initiated,
2029 * evacuation of all allocated space (copying it to other vdevs) happens
2030 * in the background (see spa_vdev_remove_thread()), and can be canceled
2031 * (see spa_vdev_remove_cancel()). If successful, the vdev will
2032 * be transformed to an indirect vdev (see spa_vdev_remove_complete()).
2035 spa_vdev_remove_top(vdev_t
*vd
, uint64_t *txg
)
2037 spa_t
*spa
= vd
->vdev_spa
;
2041 * Check for errors up-front, so that we don't waste time
2042 * passivating the metaslab group and clearing the ZIL if there
2045 error
= spa_vdev_remove_top_check(vd
);
2050 * Stop allocating from this vdev. Note that we must check
2051 * that this is not the only device in the pool before
2052 * passivating, otherwise we will not be able to make
2053 * progress because we can't allocate from any vdevs.
2054 * The above check for sufficient free space serves this
2057 metaslab_group_t
*mg
= vd
->vdev_mg
;
2058 metaslab_group_passivate(mg
);
2061 * Wait for the youngest allocations and frees to sync,
2062 * and then wait for the deferral of those frees to finish.
2064 spa_vdev_config_exit(spa
, NULL
,
2065 *txg
+ TXG_CONCURRENT_STATES
+ TXG_DEFER_SIZE
, 0, FTAG
);
2068 * We must ensure that no "stubby" log blocks are allocated
2069 * on the device to be removed. These blocks could be
2070 * written at any time, including while we are in the middle
2073 error
= spa_reset_logs(spa
);
2075 *txg
= spa_vdev_config_enter(spa
);
2078 * Things might have changed while the config lock was dropped
2079 * (e.g. space usage). Check for errors again.
2082 error
= spa_vdev_remove_top_check(vd
);
2085 metaslab_group_activate(mg
);
2089 vd
->vdev_removing
= B_TRUE
;
2091 vdev_dirty_leaves(vd
, VDD_DTL
, *txg
);
2092 vdev_config_dirty(vd
);
2093 dmu_tx_t
*tx
= dmu_tx_create_assigned(spa
->spa_dsl_pool
, *txg
);
2094 dsl_sync_task_nowait(spa
->spa_dsl_pool
,
2095 vdev_remove_initiate_sync
,
2096 (void *)(uintptr_t)vd
->vdev_id
, 0, ZFS_SPACE_CHECK_NONE
, tx
);
2103 * Remove a device from the pool.
2105 * Removing a device from the vdev namespace requires several steps
2106 * and can take a significant amount of time. As a result we use
2107 * the spa_vdev_config_[enter/exit] functions which allow us to
2108 * grab and release the spa_config_lock while still holding the namespace
2109 * lock. During each step the configuration is synced out.
2112 spa_vdev_remove(spa_t
*spa
, uint64_t guid
, boolean_t unspare
)
2115 nvlist_t
**spares
, **l2cache
, *nv
;
2117 uint_t nspares
, nl2cache
;
2119 boolean_t locked
= MUTEX_HELD(&spa_namespace_lock
);
2120 sysevent_t
*ev
= NULL
;
2121 char *vd_type
= NULL
, *vd_path
= NULL
;
2123 ASSERT(spa_writeable(spa
));
2126 txg
= spa_vdev_enter(spa
);
2128 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
2129 if (spa_feature_is_active(spa
, SPA_FEATURE_POOL_CHECKPOINT
)) {
2130 error
= (spa_has_checkpoint(spa
)) ?
2131 ZFS_ERR_CHECKPOINT_EXISTS
: ZFS_ERR_DISCARDING_CHECKPOINT
;
2134 return (spa_vdev_exit(spa
, NULL
, txg
, error
));
2139 vd
= spa_lookup_by_guid(spa
, guid
, B_FALSE
);
2141 if (spa
->spa_spares
.sav_vdevs
!= NULL
&&
2142 nvlist_lookup_nvlist_array(spa
->spa_spares
.sav_config
,
2143 ZPOOL_CONFIG_SPARES
, &spares
, &nspares
) == 0 &&
2144 (nv
= spa_nvlist_lookup_by_guid(spares
, nspares
, guid
)) != NULL
) {
2146 * Only remove the hot spare if it's not currently in use
2149 if (vd
== NULL
|| unspare
) {
2151 vd
= spa_lookup_by_guid(spa
, guid
, B_TRUE
);
2152 ev
= spa_event_create(spa
, vd
, NULL
,
2153 ESC_ZFS_VDEV_REMOVE_AUX
);
2155 vd_type
= VDEV_TYPE_SPARE
;
2156 vd_path
= fnvlist_lookup_string(nv
, ZPOOL_CONFIG_PATH
);
2157 spa_vdev_remove_aux(spa
->spa_spares
.sav_config
,
2158 ZPOOL_CONFIG_SPARES
, spares
, nspares
, nv
);
2159 spa_load_spares(spa
);
2160 spa
->spa_spares
.sav_sync
= B_TRUE
;
2162 error
= SET_ERROR(EBUSY
);
2164 } else if (spa
->spa_l2cache
.sav_vdevs
!= NULL
&&
2165 nvlist_lookup_nvlist_array(spa
->spa_l2cache
.sav_config
,
2166 ZPOOL_CONFIG_L2CACHE
, &l2cache
, &nl2cache
) == 0 &&
2167 (nv
= spa_nvlist_lookup_by_guid(l2cache
, nl2cache
, guid
)) != NULL
) {
2168 vd_type
= VDEV_TYPE_L2CACHE
;
2169 vd_path
= fnvlist_lookup_string(nv
, ZPOOL_CONFIG_PATH
);
2171 * Cache devices can always be removed.
2173 vd
= spa_lookup_by_guid(spa
, guid
, B_TRUE
);
2174 ev
= spa_event_create(spa
, vd
, NULL
, ESC_ZFS_VDEV_REMOVE_AUX
);
2175 spa_vdev_remove_aux(spa
->spa_l2cache
.sav_config
,
2176 ZPOOL_CONFIG_L2CACHE
, l2cache
, nl2cache
, nv
);
2177 spa_load_l2cache(spa
);
2178 spa
->spa_l2cache
.sav_sync
= B_TRUE
;
2179 } else if (vd
!= NULL
&& vd
->vdev_islog
) {
2182 vd_path
= (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-";
2183 error
= spa_vdev_remove_log(vd
, &txg
);
2184 } else if (vd
!= NULL
) {
2186 error
= spa_vdev_remove_top(vd
, &txg
);
2189 * There is no vdev of any kind with the specified guid.
2191 error
= SET_ERROR(ENOENT
);
2195 error
= spa_vdev_exit(spa
, NULL
, txg
, error
);
2198 * Logging must be done outside the spa config lock. Otherwise,
2199 * this code path could end up holding the spa config lock while
2200 * waiting for a txg_sync so it can write to the internal log.
2201 * Doing that would prevent the txg sync from actually happening,
2202 * causing a deadlock.
2204 if (error
== 0 && vd_type
!= NULL
&& vd_path
!= NULL
) {
2205 spa_history_log_internal(spa
, "vdev remove", NULL
,
2206 "%s vdev (%s) %s", spa_name(spa
), vd_type
, vd_path
);
2216 spa_removal_get_stats(spa_t
*spa
, pool_removal_stat_t
*prs
)
2218 prs
->prs_state
= spa
->spa_removing_phys
.sr_state
;
2220 if (prs
->prs_state
== DSS_NONE
)
2221 return (SET_ERROR(ENOENT
));
2223 prs
->prs_removing_vdev
= spa
->spa_removing_phys
.sr_removing_vdev
;
2224 prs
->prs_start_time
= spa
->spa_removing_phys
.sr_start_time
;
2225 prs
->prs_end_time
= spa
->spa_removing_phys
.sr_end_time
;
2226 prs
->prs_to_copy
= spa
->spa_removing_phys
.sr_to_copy
;
2227 prs
->prs_copied
= spa
->spa_removing_phys
.sr_copied
;
2229 prs
->prs_mapping_memory
= 0;
2230 uint64_t indirect_vdev_id
=
2231 spa
->spa_removing_phys
.sr_prev_indirect_vdev
;
2232 while (indirect_vdev_id
!= -1) {
2233 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[indirect_vdev_id
];
2234 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
2235 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
2237 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
2238 prs
->prs_mapping_memory
+= vdev_indirect_mapping_size(vim
);
2239 indirect_vdev_id
= vic
->vic_prev_indirect_vdev
;
2245 #if defined(_KERNEL)
2246 module_param(zfs_removal_ignore_errors
, int, 0644);
2247 MODULE_PARM_DESC(zfs_removal_ignore_errors
,
2248 "Ignore hard IO errors when removing device");
2250 module_param(zfs_remove_max_segment
, int, 0644);
2251 MODULE_PARM_DESC(zfs_remove_max_segment
,
2252 "Largest contiguous segment to allocate when removing device");
2254 module_param(vdev_removal_max_span
, int, 0644);
2255 MODULE_PARM_DESC(vdev_removal_max_span
,
2256 "Largest span of free chunks a remap segment can span");
2259 module_param(zfs_removal_suspend_progress
, int, 0644);
2260 MODULE_PARM_DESC(zfs_removal_suspend_progress
,
2261 "Pause device removal after this many bytes are copied "
2262 "(debug use only - causes removal to hang)");
2265 EXPORT_SYMBOL(free_from_removing_vdev
);
2266 EXPORT_SYMBOL(spa_removal_get_stats
);
2267 EXPORT_SYMBOL(spa_remove_init
);
2268 EXPORT_SYMBOL(spa_restart_removal
);
2269 EXPORT_SYMBOL(spa_vdev_removal_destroy
);
2270 EXPORT_SYMBOL(spa_vdev_remove
);
2271 EXPORT_SYMBOL(spa_vdev_remove_cancel
);
2272 EXPORT_SYMBOL(spa_vdev_remove_suspend
);
2273 EXPORT_SYMBOL(svr_sync
);