4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
27 #include <sys/zfs_context.h>
28 #include <sys/spa_impl.h>
30 #include <sys/dmu_tx.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/metaslab.h>
34 #include <sys/metaslab_impl.h>
35 #include <sys/uberblock_impl.h>
38 #include <sys/bpobj.h>
39 #include <sys/dsl_pool.h>
40 #include <sys/dsl_synctask.h>
41 #include <sys/dsl_dir.h>
43 #include <sys/zfeature.h>
44 #include <sys/vdev_indirect_births.h>
45 #include <sys/vdev_indirect_mapping.h>
47 #include <sys/vdev_initialize.h>
48 #include <sys/trace_vdev.h>
51 * This file contains the necessary logic to remove vdevs from a
52 * storage pool. Currently, the only devices that can be removed
53 * are log, cache, and spare devices; and top level vdevs from a pool
54 * w/o raidz or mirrors. (Note that members of a mirror can be removed
55 * by the detach operation.)
57 * Log vdevs are removed by evacuating them and then turning the vdev
58 * into a hole vdev while holding spa config locks.
60 * Top level vdevs are removed and converted into an indirect vdev via
61 * a multi-step process:
63 * - Disable allocations from this device (spa_vdev_remove_top).
65 * - From a new thread (spa_vdev_remove_thread), copy data from
66 * the removing vdev to a different vdev. The copy happens in open
67 * context (spa_vdev_copy_impl) and issues a sync task
68 * (vdev_mapping_sync) so the sync thread can update the partial
69 * indirect mappings in core and on disk.
71 * - If a free happens during a removal, it is freed from the
72 * removing vdev, and if it has already been copied, from the new
73 * location as well (free_from_removing_vdev).
75 * - After the removal is completed, the copy thread converts the vdev
76 * into an indirect vdev (vdev_remove_complete) before instructing
77 * the sync thread to destroy the space maps and finish the removal
78 * (spa_finish_removal).
81 typedef struct vdev_copy_arg
{
83 uint64_t vca_outstanding_bytes
;
84 uint64_t vca_read_error_bytes
;
85 uint64_t vca_write_error_bytes
;
91 * The maximum amount of memory we can use for outstanding i/o while
92 * doing a device removal. This determines how much i/o we can have
93 * in flight concurrently.
95 int zfs_remove_max_copy_bytes
= 64 * 1024 * 1024;
98 * The largest contiguous segment that we will attempt to allocate when
99 * removing a device. This can be no larger than SPA_MAXBLOCKSIZE. If
100 * there is a performance problem with attempting to allocate large blocks,
101 * consider decreasing this.
103 int zfs_remove_max_segment
= SPA_MAXBLOCKSIZE
;
106 * Ignore hard IO errors during device removal. When set if a device
107 * encounters hard IO error during the removal process the removal will
108 * not be cancelled. This can result in a normally recoverable block
109 * becoming permanently damaged and is not recommended.
111 int zfs_removal_ignore_errors
= 0;
114 * Allow a remap segment to span free chunks of at most this size. The main
115 * impact of a larger span is that we will read and write larger, more
116 * contiguous chunks, with more "unnecessary" data -- trading off bandwidth
117 * for iops. The value here was chosen to align with
118 * zfs_vdev_read_gap_limit, which is a similar concept when doing regular
119 * reads (but there's no reason it has to be the same).
121 * Additionally, a higher span will have the following relatively minor
123 * - the mapping will be smaller, since one entry can cover more allocated
125 * - more of the fragmentation in the removing device will be preserved
126 * - we'll do larger allocations, which may fail and fall back on smaller
129 int vdev_removal_max_span
= 32 * 1024;
132 * This is used by the test suite so that it can ensure that certain
133 * actions happen while in the middle of a removal.
135 int zfs_removal_suspend_progress
= 0;
137 #define VDEV_REMOVAL_ZAP_OBJS "lzap"
139 static void spa_vdev_remove_thread(void *arg
);
140 static int spa_vdev_remove_cancel_impl(spa_t
*spa
);
143 spa_sync_removing_state(spa_t
*spa
, dmu_tx_t
*tx
)
145 VERIFY0(zap_update(spa
->spa_dsl_pool
->dp_meta_objset
,
146 DMU_POOL_DIRECTORY_OBJECT
,
147 DMU_POOL_REMOVING
, sizeof (uint64_t),
148 sizeof (spa
->spa_removing_phys
) / sizeof (uint64_t),
149 &spa
->spa_removing_phys
, tx
));
153 spa_nvlist_lookup_by_guid(nvlist_t
**nvpp
, int count
, uint64_t target_guid
)
155 for (int i
= 0; i
< count
; i
++) {
157 fnvlist_lookup_uint64(nvpp
[i
], ZPOOL_CONFIG_GUID
);
159 if (guid
== target_guid
)
167 spa_vdev_remove_aux(nvlist_t
*config
, char *name
, nvlist_t
**dev
, int count
,
168 nvlist_t
*dev_to_remove
)
170 nvlist_t
**newdev
= NULL
;
173 newdev
= kmem_alloc((count
- 1) * sizeof (void *), KM_SLEEP
);
175 for (int i
= 0, j
= 0; i
< count
; i
++) {
176 if (dev
[i
] == dev_to_remove
)
178 VERIFY(nvlist_dup(dev
[i
], &newdev
[j
++], KM_SLEEP
) == 0);
181 VERIFY(nvlist_remove(config
, name
, DATA_TYPE_NVLIST_ARRAY
) == 0);
182 VERIFY(nvlist_add_nvlist_array(config
, name
, newdev
, count
- 1) == 0);
184 for (int i
= 0; i
< count
- 1; i
++)
185 nvlist_free(newdev
[i
]);
188 kmem_free(newdev
, (count
- 1) * sizeof (void *));
191 static spa_vdev_removal_t
*
192 spa_vdev_removal_create(vdev_t
*vd
)
194 spa_vdev_removal_t
*svr
= kmem_zalloc(sizeof (*svr
), KM_SLEEP
);
195 mutex_init(&svr
->svr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
196 cv_init(&svr
->svr_cv
, NULL
, CV_DEFAULT
, NULL
);
197 svr
->svr_allocd_segs
= range_tree_create(NULL
, NULL
);
198 svr
->svr_vdev_id
= vd
->vdev_id
;
200 for (int i
= 0; i
< TXG_SIZE
; i
++) {
201 svr
->svr_frees
[i
] = range_tree_create(NULL
, NULL
);
202 list_create(&svr
->svr_new_segments
[i
],
203 sizeof (vdev_indirect_mapping_entry_t
),
204 offsetof(vdev_indirect_mapping_entry_t
, vime_node
));
211 spa_vdev_removal_destroy(spa_vdev_removal_t
*svr
)
213 for (int i
= 0; i
< TXG_SIZE
; i
++) {
214 ASSERT0(svr
->svr_bytes_done
[i
]);
215 ASSERT0(svr
->svr_max_offset_to_sync
[i
]);
216 range_tree_destroy(svr
->svr_frees
[i
]);
217 list_destroy(&svr
->svr_new_segments
[i
]);
220 range_tree_destroy(svr
->svr_allocd_segs
);
221 mutex_destroy(&svr
->svr_lock
);
222 cv_destroy(&svr
->svr_cv
);
223 kmem_free(svr
, sizeof (*svr
));
227 * This is called as a synctask in the txg in which we will mark this vdev
228 * as removing (in the config stored in the MOS).
230 * It begins the evacuation of a toplevel vdev by:
231 * - initializing the spa_removing_phys which tracks this removal
232 * - computing the amount of space to remove for accounting purposes
233 * - dirtying all dbufs in the spa_config_object
234 * - creating the spa_vdev_removal
235 * - starting the spa_vdev_remove_thread
238 vdev_remove_initiate_sync(void *arg
, dmu_tx_t
*tx
)
240 int vdev_id
= (uintptr_t)arg
;
241 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
242 vdev_t
*vd
= vdev_lookup_top(spa
, vdev_id
);
243 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
244 objset_t
*mos
= spa
->spa_dsl_pool
->dp_meta_objset
;
245 spa_vdev_removal_t
*svr
= NULL
;
246 ASSERTV(uint64_t txg
= dmu_tx_get_txg(tx
));
248 ASSERT3P(vd
->vdev_ops
, !=, &vdev_raidz_ops
);
249 svr
= spa_vdev_removal_create(vd
);
251 ASSERT(vd
->vdev_removing
);
252 ASSERT3P(vd
->vdev_indirect_mapping
, ==, NULL
);
254 spa_feature_incr(spa
, SPA_FEATURE_DEVICE_REMOVAL
, tx
);
255 if (spa_feature_is_enabled(spa
, SPA_FEATURE_OBSOLETE_COUNTS
)) {
257 * By activating the OBSOLETE_COUNTS feature, we prevent
258 * the pool from being downgraded and ensure that the
259 * refcounts are precise.
261 spa_feature_incr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
263 VERIFY0(zap_add(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
264 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE
, sizeof (one
), 1,
266 ASSERTV(boolean_t are_precise
);
267 ASSERT0(vdev_obsolete_counts_are_precise(vd
, &are_precise
));
268 ASSERT3B(are_precise
, ==, B_TRUE
);
271 vic
->vic_mapping_object
= vdev_indirect_mapping_alloc(mos
, tx
);
272 vd
->vdev_indirect_mapping
=
273 vdev_indirect_mapping_open(mos
, vic
->vic_mapping_object
);
274 vic
->vic_births_object
= vdev_indirect_births_alloc(mos
, tx
);
275 vd
->vdev_indirect_births
=
276 vdev_indirect_births_open(mos
, vic
->vic_births_object
);
277 spa
->spa_removing_phys
.sr_removing_vdev
= vd
->vdev_id
;
278 spa
->spa_removing_phys
.sr_start_time
= gethrestime_sec();
279 spa
->spa_removing_phys
.sr_end_time
= 0;
280 spa
->spa_removing_phys
.sr_state
= DSS_SCANNING
;
281 spa
->spa_removing_phys
.sr_to_copy
= 0;
282 spa
->spa_removing_phys
.sr_copied
= 0;
285 * Note: We can't use vdev_stat's vs_alloc for sr_to_copy, because
286 * there may be space in the defer tree, which is free, but still
287 * counted in vs_alloc.
289 for (uint64_t i
= 0; i
< vd
->vdev_ms_count
; i
++) {
290 metaslab_t
*ms
= vd
->vdev_ms
[i
];
291 if (ms
->ms_sm
== NULL
)
294 spa
->spa_removing_phys
.sr_to_copy
+=
295 metaslab_allocated_space(ms
);
298 * Space which we are freeing this txg does not need to
301 spa
->spa_removing_phys
.sr_to_copy
-=
302 range_tree_space(ms
->ms_freeing
);
304 ASSERT0(range_tree_space(ms
->ms_freed
));
305 for (int t
= 0; t
< TXG_SIZE
; t
++)
306 ASSERT0(range_tree_space(ms
->ms_allocating
[t
]));
310 * Sync tasks are called before metaslab_sync(), so there should
311 * be no already-synced metaslabs in the TXG_CLEAN list.
313 ASSERT3P(txg_list_head(&vd
->vdev_ms_list
, TXG_CLEAN(txg
)), ==, NULL
);
315 spa_sync_removing_state(spa
, tx
);
318 * All blocks that we need to read the most recent mapping must be
319 * stored on concrete vdevs. Therefore, we must dirty anything that
320 * is read before spa_remove_init(). Specifically, the
321 * spa_config_object. (Note that although we already modified the
322 * spa_config_object in spa_sync_removing_state, that may not have
323 * modified all blocks of the object.)
325 dmu_object_info_t doi
;
326 VERIFY0(dmu_object_info(mos
, DMU_POOL_DIRECTORY_OBJECT
, &doi
));
327 for (uint64_t offset
= 0; offset
< doi
.doi_max_offset
; ) {
329 VERIFY0(dmu_buf_hold(mos
, DMU_POOL_DIRECTORY_OBJECT
,
330 offset
, FTAG
, &dbuf
, 0));
331 dmu_buf_will_dirty(dbuf
, tx
);
332 offset
+= dbuf
->db_size
;
333 dmu_buf_rele(dbuf
, FTAG
);
337 * Now that we've allocated the im_object, dirty the vdev to ensure
338 * that the object gets written to the config on disk.
340 vdev_config_dirty(vd
);
342 zfs_dbgmsg("starting removal thread for vdev %llu (%p) in txg %llu "
343 "im_obj=%llu", vd
->vdev_id
, vd
, dmu_tx_get_txg(tx
),
344 vic
->vic_mapping_object
);
346 spa_history_log_internal(spa
, "vdev remove started", tx
,
347 "%s vdev %llu %s", spa_name(spa
), vd
->vdev_id
,
348 (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
350 * Setting spa_vdev_removal causes subsequent frees to call
351 * free_from_removing_vdev(). Note that we don't need any locking
352 * because we are the sync thread, and metaslab_free_impl() is only
353 * called from syncing context (potentially from a zio taskq thread,
354 * but in any case only when there are outstanding free i/os, which
357 ASSERT3P(spa
->spa_vdev_removal
, ==, NULL
);
358 spa
->spa_vdev_removal
= svr
;
359 svr
->svr_thread
= thread_create(NULL
, 0,
360 spa_vdev_remove_thread
, spa
, 0, &p0
, TS_RUN
, minclsyspri
);
364 * When we are opening a pool, we must read the mapping for each
365 * indirect vdev in order from most recently removed to least
366 * recently removed. We do this because the blocks for the mapping
367 * of older indirect vdevs may be stored on more recently removed vdevs.
368 * In order to read each indirect mapping object, we must have
369 * initialized all more recently removed vdevs.
372 spa_remove_init(spa_t
*spa
)
376 error
= zap_lookup(spa
->spa_dsl_pool
->dp_meta_objset
,
377 DMU_POOL_DIRECTORY_OBJECT
,
378 DMU_POOL_REMOVING
, sizeof (uint64_t),
379 sizeof (spa
->spa_removing_phys
) / sizeof (uint64_t),
380 &spa
->spa_removing_phys
);
382 if (error
== ENOENT
) {
383 spa
->spa_removing_phys
.sr_state
= DSS_NONE
;
384 spa
->spa_removing_phys
.sr_removing_vdev
= -1;
385 spa
->spa_removing_phys
.sr_prev_indirect_vdev
= -1;
386 spa
->spa_indirect_vdevs_loaded
= B_TRUE
;
388 } else if (error
!= 0) {
392 if (spa
->spa_removing_phys
.sr_state
== DSS_SCANNING
) {
394 * We are currently removing a vdev. Create and
395 * initialize a spa_vdev_removal_t from the bonus
396 * buffer of the removing vdevs vdev_im_object, and
397 * initialize its partial mapping.
399 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
400 vdev_t
*vd
= vdev_lookup_top(spa
,
401 spa
->spa_removing_phys
.sr_removing_vdev
);
404 spa_config_exit(spa
, SCL_STATE
, FTAG
);
408 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
410 ASSERT(vdev_is_concrete(vd
));
411 spa_vdev_removal_t
*svr
= spa_vdev_removal_create(vd
);
412 ASSERT3U(svr
->svr_vdev_id
, ==, vd
->vdev_id
);
413 ASSERT(vd
->vdev_removing
);
415 vd
->vdev_indirect_mapping
= vdev_indirect_mapping_open(
416 spa
->spa_meta_objset
, vic
->vic_mapping_object
);
417 vd
->vdev_indirect_births
= vdev_indirect_births_open(
418 spa
->spa_meta_objset
, vic
->vic_births_object
);
419 spa_config_exit(spa
, SCL_STATE
, FTAG
);
421 spa
->spa_vdev_removal
= svr
;
424 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
425 uint64_t indirect_vdev_id
=
426 spa
->spa_removing_phys
.sr_prev_indirect_vdev
;
427 while (indirect_vdev_id
!= UINT64_MAX
) {
428 vdev_t
*vd
= vdev_lookup_top(spa
, indirect_vdev_id
);
429 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
431 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
432 vd
->vdev_indirect_mapping
= vdev_indirect_mapping_open(
433 spa
->spa_meta_objset
, vic
->vic_mapping_object
);
434 vd
->vdev_indirect_births
= vdev_indirect_births_open(
435 spa
->spa_meta_objset
, vic
->vic_births_object
);
437 indirect_vdev_id
= vic
->vic_prev_indirect_vdev
;
439 spa_config_exit(spa
, SCL_STATE
, FTAG
);
442 * Now that we've loaded all the indirect mappings, we can allow
443 * reads from other blocks (e.g. via predictive prefetch).
445 spa
->spa_indirect_vdevs_loaded
= B_TRUE
;
450 spa_restart_removal(spa_t
*spa
)
452 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
458 * In general when this function is called there is no
459 * removal thread running. The only scenario where this
460 * is not true is during spa_import() where this function
461 * is called twice [once from spa_import_impl() and
462 * spa_async_resume()]. Thus, in the scenario where we
463 * import a pool that has an ongoing removal we don't
464 * want to spawn a second thread.
466 if (svr
->svr_thread
!= NULL
)
469 if (!spa_writeable(spa
))
472 zfs_dbgmsg("restarting removal of %llu", svr
->svr_vdev_id
);
473 svr
->svr_thread
= thread_create(NULL
, 0, spa_vdev_remove_thread
, spa
,
474 0, &p0
, TS_RUN
, minclsyspri
);
478 * Process freeing from a device which is in the middle of being removed.
479 * We must handle this carefully so that we attempt to copy freed data,
480 * and we correctly free already-copied data.
483 free_from_removing_vdev(vdev_t
*vd
, uint64_t offset
, uint64_t size
)
485 spa_t
*spa
= vd
->vdev_spa
;
486 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
487 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
488 uint64_t txg
= spa_syncing_txg(spa
);
489 uint64_t max_offset_yet
= 0;
491 ASSERT(vd
->vdev_indirect_config
.vic_mapping_object
!= 0);
492 ASSERT3U(vd
->vdev_indirect_config
.vic_mapping_object
, ==,
493 vdev_indirect_mapping_object(vim
));
494 ASSERT3U(vd
->vdev_id
, ==, svr
->svr_vdev_id
);
496 mutex_enter(&svr
->svr_lock
);
499 * Remove the segment from the removing vdev's spacemap. This
500 * ensures that we will not attempt to copy this space (if the
501 * removal thread has not yet visited it), and also ensures
502 * that we know what is actually allocated on the new vdevs
503 * (needed if we cancel the removal).
505 * Note: we must do the metaslab_free_concrete() with the svr_lock
506 * held, so that the remove_thread can not load this metaslab and then
507 * visit this offset between the time that we metaslab_free_concrete()
508 * and when we check to see if it has been visited.
510 * Note: The checkpoint flag is set to false as having/taking
511 * a checkpoint and removing a device can't happen at the same
514 ASSERT(!spa_has_checkpoint(spa
));
515 metaslab_free_concrete(vd
, offset
, size
, B_FALSE
);
517 uint64_t synced_size
= 0;
518 uint64_t synced_offset
= 0;
519 uint64_t max_offset_synced
= vdev_indirect_mapping_max_offset(vim
);
520 if (offset
< max_offset_synced
) {
522 * The mapping for this offset is already on disk.
523 * Free from the new location.
525 * Note that we use svr_max_synced_offset because it is
526 * updated atomically with respect to the in-core mapping.
527 * By contrast, vim_max_offset is not.
529 * This block may be split between a synced entry and an
530 * in-flight or unvisited entry. Only process the synced
531 * portion of it here.
533 synced_size
= MIN(size
, max_offset_synced
- offset
);
534 synced_offset
= offset
;
536 ASSERT3U(max_offset_yet
, <=, max_offset_synced
);
537 max_offset_yet
= max_offset_synced
;
539 DTRACE_PROBE3(remove__free__synced
,
542 uint64_t, synced_size
);
545 offset
+= synced_size
;
549 * Look at all in-flight txgs starting from the currently syncing one
550 * and see if a section of this free is being copied. By starting from
551 * this txg and iterating forward, we might find that this region
552 * was copied in two different txgs and handle it appropriately.
554 for (int i
= 0; i
< TXG_CONCURRENT_STATES
; i
++) {
555 int txgoff
= (txg
+ i
) & TXG_MASK
;
556 if (size
> 0 && offset
< svr
->svr_max_offset_to_sync
[txgoff
]) {
558 * The mapping for this offset is in flight, and
559 * will be synced in txg+i.
561 uint64_t inflight_size
= MIN(size
,
562 svr
->svr_max_offset_to_sync
[txgoff
] - offset
);
564 DTRACE_PROBE4(remove__free__inflight
,
567 uint64_t, inflight_size
,
571 * We copy data in order of increasing offset.
572 * Therefore the max_offset_to_sync[] must increase
573 * (or be zero, indicating that nothing is being
574 * copied in that txg).
576 if (svr
->svr_max_offset_to_sync
[txgoff
] != 0) {
577 ASSERT3U(svr
->svr_max_offset_to_sync
[txgoff
],
580 svr
->svr_max_offset_to_sync
[txgoff
];
584 * We've already committed to copying this segment:
585 * we have allocated space elsewhere in the pool for
586 * it and have an IO outstanding to copy the data. We
587 * cannot free the space before the copy has
588 * completed, or else the copy IO might overwrite any
589 * new data. To free that space, we record the
590 * segment in the appropriate svr_frees tree and free
591 * the mapped space later, in the txg where we have
592 * completed the copy and synced the mapping (see
593 * vdev_mapping_sync).
595 range_tree_add(svr
->svr_frees
[txgoff
],
596 offset
, inflight_size
);
597 size
-= inflight_size
;
598 offset
+= inflight_size
;
601 * This space is already accounted for as being
602 * done, because it is being copied in txg+i.
603 * However, if i!=0, then it is being copied in
604 * a future txg. If we crash after this txg
605 * syncs but before txg+i syncs, then the space
606 * will be free. Therefore we must account
607 * for the space being done in *this* txg
608 * (when it is freed) rather than the future txg
609 * (when it will be copied).
611 ASSERT3U(svr
->svr_bytes_done
[txgoff
], >=,
613 svr
->svr_bytes_done
[txgoff
] -= inflight_size
;
614 svr
->svr_bytes_done
[txg
& TXG_MASK
] += inflight_size
;
617 ASSERT0(svr
->svr_max_offset_to_sync
[TXG_CLEAN(txg
) & TXG_MASK
]);
621 * The copy thread has not yet visited this offset. Ensure
625 DTRACE_PROBE3(remove__free__unvisited
,
630 if (svr
->svr_allocd_segs
!= NULL
)
631 range_tree_clear(svr
->svr_allocd_segs
, offset
, size
);
634 * Since we now do not need to copy this data, for
635 * accounting purposes we have done our job and can count
638 svr
->svr_bytes_done
[txg
& TXG_MASK
] += size
;
640 mutex_exit(&svr
->svr_lock
);
643 * Now that we have dropped svr_lock, process the synced portion
646 if (synced_size
> 0) {
647 vdev_indirect_mark_obsolete(vd
, synced_offset
, synced_size
);
650 * Note: this can only be called from syncing context,
651 * and the vdev_indirect_mapping is only changed from the
652 * sync thread, so we don't need svr_lock while doing
653 * metaslab_free_impl_cb.
655 boolean_t checkpoint
= B_FALSE
;
656 vdev_indirect_ops
.vdev_op_remap(vd
, synced_offset
, synced_size
,
657 metaslab_free_impl_cb
, &checkpoint
);
662 * Stop an active removal and update the spa_removing phys.
665 spa_finish_removal(spa_t
*spa
, dsl_scan_state_t state
, dmu_tx_t
*tx
)
667 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
668 ASSERT3U(dmu_tx_get_txg(tx
), ==, spa_syncing_txg(spa
));
670 /* Ensure the removal thread has completed before we free the svr. */
671 spa_vdev_remove_suspend(spa
);
673 ASSERT(state
== DSS_FINISHED
|| state
== DSS_CANCELED
);
675 if (state
== DSS_FINISHED
) {
676 spa_removing_phys_t
*srp
= &spa
->spa_removing_phys
;
677 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
678 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
680 if (srp
->sr_prev_indirect_vdev
!= -1) {
682 pvd
= vdev_lookup_top(spa
,
683 srp
->sr_prev_indirect_vdev
);
684 ASSERT3P(pvd
->vdev_ops
, ==, &vdev_indirect_ops
);
687 vic
->vic_prev_indirect_vdev
= srp
->sr_prev_indirect_vdev
;
688 srp
->sr_prev_indirect_vdev
= vd
->vdev_id
;
690 spa
->spa_removing_phys
.sr_state
= state
;
691 spa
->spa_removing_phys
.sr_end_time
= gethrestime_sec();
693 spa
->spa_vdev_removal
= NULL
;
694 spa_vdev_removal_destroy(svr
);
696 spa_sync_removing_state(spa
, tx
);
698 vdev_config_dirty(spa
->spa_root_vdev
);
702 free_mapped_segment_cb(void *arg
, uint64_t offset
, uint64_t size
)
705 vdev_indirect_mark_obsolete(vd
, offset
, size
);
706 boolean_t checkpoint
= B_FALSE
;
707 vdev_indirect_ops
.vdev_op_remap(vd
, offset
, size
,
708 metaslab_free_impl_cb
, &checkpoint
);
712 * On behalf of the removal thread, syncs an incremental bit more of
713 * the indirect mapping to disk and updates the in-memory mapping.
714 * Called as a sync task in every txg that the removal thread makes progress.
717 vdev_mapping_sync(void *arg
, dmu_tx_t
*tx
)
719 spa_vdev_removal_t
*svr
= arg
;
720 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
721 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
722 ASSERTV(vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
);
723 uint64_t txg
= dmu_tx_get_txg(tx
);
724 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
726 ASSERT(vic
->vic_mapping_object
!= 0);
727 ASSERT3U(txg
, ==, spa_syncing_txg(spa
));
729 vdev_indirect_mapping_add_entries(vim
,
730 &svr
->svr_new_segments
[txg
& TXG_MASK
], tx
);
731 vdev_indirect_births_add_entry(vd
->vdev_indirect_births
,
732 vdev_indirect_mapping_max_offset(vim
), dmu_tx_get_txg(tx
), tx
);
735 * Free the copied data for anything that was freed while the
736 * mapping entries were in flight.
738 mutex_enter(&svr
->svr_lock
);
739 range_tree_vacate(svr
->svr_frees
[txg
& TXG_MASK
],
740 free_mapped_segment_cb
, vd
);
741 ASSERT3U(svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
], >=,
742 vdev_indirect_mapping_max_offset(vim
));
743 svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] = 0;
744 mutex_exit(&svr
->svr_lock
);
746 spa_sync_removing_state(spa
, tx
);
749 typedef struct vdev_copy_segment_arg
{
751 dva_t
*vcsa_dest_dva
;
753 range_tree_t
*vcsa_obsolete_segs
;
754 } vdev_copy_segment_arg_t
;
757 unalloc_seg(void *arg
, uint64_t start
, uint64_t size
)
759 vdev_copy_segment_arg_t
*vcsa
= arg
;
760 spa_t
*spa
= vcsa
->vcsa_spa
;
761 blkptr_t bp
= { { { {0} } } };
763 BP_SET_BIRTH(&bp
, TXG_INITIAL
, TXG_INITIAL
);
764 BP_SET_LSIZE(&bp
, size
);
765 BP_SET_PSIZE(&bp
, size
);
766 BP_SET_COMPRESS(&bp
, ZIO_COMPRESS_OFF
);
767 BP_SET_CHECKSUM(&bp
, ZIO_CHECKSUM_OFF
);
768 BP_SET_TYPE(&bp
, DMU_OT_NONE
);
769 BP_SET_LEVEL(&bp
, 0);
770 BP_SET_DEDUP(&bp
, 0);
771 BP_SET_BYTEORDER(&bp
, ZFS_HOST_BYTEORDER
);
773 DVA_SET_VDEV(&bp
.blk_dva
[0], DVA_GET_VDEV(vcsa
->vcsa_dest_dva
));
774 DVA_SET_OFFSET(&bp
.blk_dva
[0],
775 DVA_GET_OFFSET(vcsa
->vcsa_dest_dva
) + start
);
776 DVA_SET_ASIZE(&bp
.blk_dva
[0], size
);
778 zio_free(spa
, vcsa
->vcsa_txg
, &bp
);
782 * All reads and writes associated with a call to spa_vdev_copy_segment()
786 spa_vdev_copy_segment_done(zio_t
*zio
)
788 vdev_copy_segment_arg_t
*vcsa
= zio
->io_private
;
790 range_tree_vacate(vcsa
->vcsa_obsolete_segs
,
792 range_tree_destroy(vcsa
->vcsa_obsolete_segs
);
793 kmem_free(vcsa
, sizeof (*vcsa
));
795 spa_config_exit(zio
->io_spa
, SCL_STATE
, zio
->io_spa
);
799 * The write of the new location is done.
802 spa_vdev_copy_segment_write_done(zio_t
*zio
)
804 vdev_copy_arg_t
*vca
= zio
->io_private
;
806 abd_free(zio
->io_abd
);
808 mutex_enter(&vca
->vca_lock
);
809 vca
->vca_outstanding_bytes
-= zio
->io_size
;
811 if (zio
->io_error
!= 0)
812 vca
->vca_write_error_bytes
+= zio
->io_size
;
814 cv_signal(&vca
->vca_cv
);
815 mutex_exit(&vca
->vca_lock
);
819 * The read of the old location is done. The parent zio is the write to
820 * the new location. Allow it to start.
823 spa_vdev_copy_segment_read_done(zio_t
*zio
)
825 vdev_copy_arg_t
*vca
= zio
->io_private
;
827 if (zio
->io_error
!= 0) {
828 mutex_enter(&vca
->vca_lock
);
829 vca
->vca_read_error_bytes
+= zio
->io_size
;
830 mutex_exit(&vca
->vca_lock
);
833 zio_nowait(zio_unique_parent(zio
));
837 * If the old and new vdevs are mirrors, we will read both sides of the old
838 * mirror, and write each copy to the corresponding side of the new mirror.
839 * If the old and new vdevs have a different number of children, we will do
840 * this as best as possible. Since we aren't verifying checksums, this
841 * ensures that as long as there's a good copy of the data, we'll have a
842 * good copy after the removal, even if there's silent damage to one side
843 * of the mirror. If we're removing a mirror that has some silent damage,
844 * we'll have exactly the same damage in the new location (assuming that
845 * the new location is also a mirror).
847 * We accomplish this by creating a tree of zio_t's, with as many writes as
848 * there are "children" of the new vdev (a non-redundant vdev counts as one
849 * child, a 2-way mirror has 2 children, etc). Each write has an associated
850 * read from a child of the old vdev. Typically there will be the same
851 * number of children of the old and new vdevs. However, if there are more
852 * children of the new vdev, some child(ren) of the old vdev will be issued
853 * multiple reads. If there are more children of the old vdev, some copies
856 * For example, the tree of zio_t's for a 2-way mirror is:
860 * write(new vdev, child 0) write(new vdev, child 1)
862 * read(old vdev, child 0) read(old vdev, child 1)
864 * Child zio's complete before their parents complete. However, zio's
865 * created with zio_vdev_child_io() may be issued before their children
866 * complete. In this case we need to make sure that the children (reads)
867 * complete before the parents (writes) are *issued*. We do this by not
868 * calling zio_nowait() on each write until its corresponding read has
871 * The spa_config_lock must be held while zio's created by
872 * zio_vdev_child_io() are in progress, to ensure that the vdev tree does
873 * not change (e.g. due to a concurrent "zpool attach/detach"). The "null"
874 * zio is needed to release the spa_config_lock after all the reads and
875 * writes complete. (Note that we can't grab the config lock for each read,
876 * because it is not reentrant - we could deadlock with a thread waiting
880 spa_vdev_copy_one_child(vdev_copy_arg_t
*vca
, zio_t
*nzio
,
881 vdev_t
*source_vd
, uint64_t source_offset
,
882 vdev_t
*dest_child_vd
, uint64_t dest_offset
, int dest_id
, uint64_t size
)
884 ASSERT3U(spa_config_held(nzio
->io_spa
, SCL_ALL
, RW_READER
), !=, 0);
887 * If the destination child in unwritable then there is no point
888 * in issuing the source reads which cannot be written.
890 if (!vdev_writeable(dest_child_vd
))
893 mutex_enter(&vca
->vca_lock
);
894 vca
->vca_outstanding_bytes
+= size
;
895 mutex_exit(&vca
->vca_lock
);
897 abd_t
*abd
= abd_alloc_for_io(size
, B_FALSE
);
899 vdev_t
*source_child_vd
= NULL
;
900 if (source_vd
->vdev_ops
== &vdev_mirror_ops
&& dest_id
!= -1) {
902 * Source and dest are both mirrors. Copy from the same
903 * child id as we are copying to (wrapping around if there
904 * are more dest children than source children). If the
905 * preferred source child is unreadable select another.
907 for (int i
= 0; i
< source_vd
->vdev_children
; i
++) {
908 source_child_vd
= source_vd
->vdev_child
[
909 (dest_id
+ i
) % source_vd
->vdev_children
];
910 if (vdev_readable(source_child_vd
))
914 source_child_vd
= source_vd
;
918 * There should always be at least one readable source child or
919 * the pool would be in a suspended state. Somehow selecting an
920 * unreadable child would result in IO errors, the removal process
921 * being cancelled, and the pool reverting to its pre-removal state.
923 ASSERT3P(source_child_vd
, !=, NULL
);
925 zio_t
*write_zio
= zio_vdev_child_io(nzio
, NULL
,
926 dest_child_vd
, dest_offset
, abd
, size
,
927 ZIO_TYPE_WRITE
, ZIO_PRIORITY_REMOVAL
,
929 spa_vdev_copy_segment_write_done
, vca
);
931 zio_nowait(zio_vdev_child_io(write_zio
, NULL
,
932 source_child_vd
, source_offset
, abd
, size
,
933 ZIO_TYPE_READ
, ZIO_PRIORITY_REMOVAL
,
935 spa_vdev_copy_segment_read_done
, vca
));
939 * Allocate a new location for this segment, and create the zio_t's to
940 * read from the old location and write to the new location.
943 spa_vdev_copy_segment(vdev_t
*vd
, range_tree_t
*segs
,
944 uint64_t maxalloc
, uint64_t txg
,
945 vdev_copy_arg_t
*vca
, zio_alloc_list_t
*zal
)
947 metaslab_group_t
*mg
= vd
->vdev_mg
;
948 spa_t
*spa
= vd
->vdev_spa
;
949 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
950 vdev_indirect_mapping_entry_t
*entry
;
952 uint64_t start
= range_tree_min(segs
);
954 ASSERT3U(maxalloc
, <=, SPA_MAXBLOCKSIZE
);
956 uint64_t size
= range_tree_span(segs
);
957 if (range_tree_span(segs
) > maxalloc
) {
959 * We can't allocate all the segments. Prefer to end
960 * the allocation at the end of a segment, thus avoiding
961 * additional split blocks.
965 search
.rs_start
= start
+ maxalloc
;
966 search
.rs_end
= search
.rs_start
;
967 range_seg_t
*rs
= avl_find(&segs
->rt_root
, &search
, &where
);
969 rs
= avl_nearest(&segs
->rt_root
, where
, AVL_BEFORE
);
971 rs
= AVL_PREV(&segs
->rt_root
, rs
);
974 size
= rs
->rs_end
- start
;
977 * There are no segments that end before maxalloc.
978 * I.e. the first segment is larger than maxalloc,
979 * so we must split it.
984 ASSERT3U(size
, <=, maxalloc
);
987 * An allocation class might not have any remaining vdevs or space
989 metaslab_class_t
*mc
= mg
->mg_class
;
990 if (mc
!= spa_normal_class(spa
) && mc
->mc_groups
<= 1)
991 mc
= spa_normal_class(spa
);
992 int error
= metaslab_alloc_dva(spa
, mc
, size
, &dst
, 0, NULL
, txg
, 0,
994 if (error
== ENOSPC
&& mc
!= spa_normal_class(spa
)) {
995 error
= metaslab_alloc_dva(spa
, spa_normal_class(spa
), size
,
996 &dst
, 0, NULL
, txg
, 0, zal
, 0);
1002 * Determine the ranges that are not actually needed. Offsets are
1003 * relative to the start of the range to be copied (i.e. relative to the
1004 * local variable "start").
1006 range_tree_t
*obsolete_segs
= range_tree_create(NULL
, NULL
);
1008 range_seg_t
*rs
= avl_first(&segs
->rt_root
);
1009 ASSERT3U(rs
->rs_start
, ==, start
);
1010 uint64_t prev_seg_end
= rs
->rs_end
;
1011 while ((rs
= AVL_NEXT(&segs
->rt_root
, rs
)) != NULL
) {
1012 if (rs
->rs_start
>= start
+ size
) {
1015 range_tree_add(obsolete_segs
,
1016 prev_seg_end
- start
,
1017 rs
->rs_start
- prev_seg_end
);
1019 prev_seg_end
= rs
->rs_end
;
1021 /* We don't end in the middle of an obsolete range */
1022 ASSERT3U(start
+ size
, <=, prev_seg_end
);
1024 range_tree_clear(segs
, start
, size
);
1027 * We can't have any padding of the allocated size, otherwise we will
1028 * misunderstand what's allocated, and the size of the mapping.
1029 * The caller ensures this will be true by passing in a size that is
1030 * aligned to the worst (highest) ashift in the pool.
1032 ASSERT3U(DVA_GET_ASIZE(&dst
), ==, size
);
1034 entry
= kmem_zalloc(sizeof (vdev_indirect_mapping_entry_t
), KM_SLEEP
);
1035 DVA_MAPPING_SET_SRC_OFFSET(&entry
->vime_mapping
, start
);
1036 entry
->vime_mapping
.vimep_dst
= dst
;
1037 if (spa_feature_is_enabled(spa
, SPA_FEATURE_OBSOLETE_COUNTS
)) {
1038 entry
->vime_obsolete_count
= range_tree_space(obsolete_segs
);
1041 vdev_copy_segment_arg_t
*vcsa
= kmem_zalloc(sizeof (*vcsa
), KM_SLEEP
);
1042 vcsa
->vcsa_dest_dva
= &entry
->vime_mapping
.vimep_dst
;
1043 vcsa
->vcsa_obsolete_segs
= obsolete_segs
;
1044 vcsa
->vcsa_spa
= spa
;
1045 vcsa
->vcsa_txg
= txg
;
1048 * See comment before spa_vdev_copy_one_child().
1050 spa_config_enter(spa
, SCL_STATE
, spa
, RW_READER
);
1051 zio_t
*nzio
= zio_null(spa
->spa_txg_zio
[txg
& TXG_MASK
], spa
, NULL
,
1052 spa_vdev_copy_segment_done
, vcsa
, 0);
1053 vdev_t
*dest_vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dst
));
1054 if (dest_vd
->vdev_ops
== &vdev_mirror_ops
) {
1055 for (int i
= 0; i
< dest_vd
->vdev_children
; i
++) {
1056 vdev_t
*child
= dest_vd
->vdev_child
[i
];
1057 spa_vdev_copy_one_child(vca
, nzio
, vd
, start
,
1058 child
, DVA_GET_OFFSET(&dst
), i
, size
);
1061 spa_vdev_copy_one_child(vca
, nzio
, vd
, start
,
1062 dest_vd
, DVA_GET_OFFSET(&dst
), -1, size
);
1066 list_insert_tail(&svr
->svr_new_segments
[txg
& TXG_MASK
], entry
);
1067 ASSERT3U(start
+ size
, <=, vd
->vdev_ms_count
<< vd
->vdev_ms_shift
);
1068 vdev_dirty(vd
, 0, NULL
, txg
);
1074 * Complete the removal of a toplevel vdev. This is called as a
1075 * synctask in the same txg that we will sync out the new config (to the
1076 * MOS object) which indicates that this vdev is indirect.
1079 vdev_remove_complete_sync(void *arg
, dmu_tx_t
*tx
)
1081 spa_vdev_removal_t
*svr
= arg
;
1082 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1083 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1085 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
1087 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1088 ASSERT0(svr
->svr_bytes_done
[i
]);
1091 ASSERT3U(spa
->spa_removing_phys
.sr_copied
, ==,
1092 spa
->spa_removing_phys
.sr_to_copy
);
1094 vdev_destroy_spacemaps(vd
, tx
);
1096 /* destroy leaf zaps, if any */
1097 ASSERT3P(svr
->svr_zaplist
, !=, NULL
);
1098 for (nvpair_t
*pair
= nvlist_next_nvpair(svr
->svr_zaplist
, NULL
);
1100 pair
= nvlist_next_nvpair(svr
->svr_zaplist
, pair
)) {
1101 vdev_destroy_unlink_zap(vd
, fnvpair_value_uint64(pair
), tx
);
1103 fnvlist_free(svr
->svr_zaplist
);
1105 spa_finish_removal(dmu_tx_pool(tx
)->dp_spa
, DSS_FINISHED
, tx
);
1106 /* vd->vdev_path is not available here */
1107 spa_history_log_internal(spa
, "vdev remove completed", tx
,
1108 "%s vdev %llu", spa_name(spa
), vd
->vdev_id
);
1112 vdev_remove_enlist_zaps(vdev_t
*vd
, nvlist_t
*zlist
)
1114 ASSERT3P(zlist
, !=, NULL
);
1115 ASSERT3P(vd
->vdev_ops
, !=, &vdev_raidz_ops
);
1117 if (vd
->vdev_leaf_zap
!= 0) {
1119 (void) snprintf(zkey
, sizeof (zkey
), "%s-%llu",
1120 VDEV_REMOVAL_ZAP_OBJS
, (u_longlong_t
)vd
->vdev_leaf_zap
);
1121 fnvlist_add_uint64(zlist
, zkey
, vd
->vdev_leaf_zap
);
1124 for (uint64_t id
= 0; id
< vd
->vdev_children
; id
++) {
1125 vdev_remove_enlist_zaps(vd
->vdev_child
[id
], zlist
);
1130 vdev_remove_replace_with_indirect(vdev_t
*vd
, uint64_t txg
)
1134 spa_t
*spa
= vd
->vdev_spa
;
1135 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1138 * First, build a list of leaf zaps to be destroyed.
1139 * This is passed to the sync context thread,
1140 * which does the actual unlinking.
1142 svr
->svr_zaplist
= fnvlist_alloc();
1143 vdev_remove_enlist_zaps(vd
, svr
->svr_zaplist
);
1145 ivd
= vdev_add_parent(vd
, &vdev_indirect_ops
);
1146 ivd
->vdev_removing
= 0;
1148 vd
->vdev_leaf_zap
= 0;
1150 vdev_remove_child(ivd
, vd
);
1151 vdev_compact_children(ivd
);
1153 ASSERT(!list_link_active(&vd
->vdev_state_dirty_node
));
1155 mutex_enter(&svr
->svr_lock
);
1156 svr
->svr_thread
= NULL
;
1157 cv_broadcast(&svr
->svr_cv
);
1158 mutex_exit(&svr
->svr_lock
);
1160 /* After this, we can not use svr. */
1161 tx
= dmu_tx_create_assigned(spa
->spa_dsl_pool
, txg
);
1162 dsl_sync_task_nowait(spa
->spa_dsl_pool
, vdev_remove_complete_sync
, svr
,
1163 0, ZFS_SPACE_CHECK_NONE
, tx
);
1168 * Complete the removal of a toplevel vdev. This is called in open
1169 * context by the removal thread after we have copied all vdev's data.
1172 vdev_remove_complete(spa_t
*spa
)
1177 * Wait for any deferred frees to be synced before we call
1178 * vdev_metaslab_fini()
1180 txg_wait_synced(spa
->spa_dsl_pool
, 0);
1181 txg
= spa_vdev_enter(spa
);
1182 vdev_t
*vd
= vdev_lookup_top(spa
, spa
->spa_vdev_removal
->svr_vdev_id
);
1183 ASSERT3P(vd
->vdev_initialize_thread
, ==, NULL
);
1185 sysevent_t
*ev
= spa_event_create(spa
, vd
, NULL
,
1186 ESC_ZFS_VDEV_REMOVE_DEV
);
1188 zfs_dbgmsg("finishing device removal for vdev %llu in txg %llu",
1192 * Discard allocation state.
1194 if (vd
->vdev_mg
!= NULL
) {
1195 vdev_metaslab_fini(vd
);
1196 metaslab_group_destroy(vd
->vdev_mg
);
1199 ASSERT0(vd
->vdev_stat
.vs_space
);
1200 ASSERT0(vd
->vdev_stat
.vs_dspace
);
1202 vdev_remove_replace_with_indirect(vd
, txg
);
1205 * We now release the locks, allowing spa_sync to run and finish the
1206 * removal via vdev_remove_complete_sync in syncing context.
1208 * Note that we hold on to the vdev_t that has been replaced. Since
1209 * it isn't part of the vdev tree any longer, it can't be concurrently
1210 * manipulated, even while we don't have the config lock.
1212 (void) spa_vdev_exit(spa
, NULL
, txg
, 0);
1215 * Top ZAP should have been transferred to the indirect vdev in
1216 * vdev_remove_replace_with_indirect.
1218 ASSERT0(vd
->vdev_top_zap
);
1221 * Leaf ZAP should have been moved in vdev_remove_replace_with_indirect.
1223 ASSERT0(vd
->vdev_leaf_zap
);
1225 txg
= spa_vdev_enter(spa
);
1226 (void) vdev_label_init(vd
, 0, VDEV_LABEL_REMOVE
);
1228 * Request to update the config and the config cachefile.
1230 vdev_config_dirty(spa
->spa_root_vdev
);
1231 (void) spa_vdev_exit(spa
, vd
, txg
, 0);
1238 * Evacuates a segment of size at most max_alloc from the vdev
1239 * via repeated calls to spa_vdev_copy_segment. If an allocation
1240 * fails, the pool is probably too fragmented to handle such a
1241 * large size, so decrease max_alloc so that the caller will not try
1242 * this size again this txg.
1245 spa_vdev_copy_impl(vdev_t
*vd
, spa_vdev_removal_t
*svr
, vdev_copy_arg_t
*vca
,
1246 uint64_t *max_alloc
, dmu_tx_t
*tx
)
1248 uint64_t txg
= dmu_tx_get_txg(tx
);
1249 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1251 mutex_enter(&svr
->svr_lock
);
1254 * Determine how big of a chunk to copy. We can allocate up
1255 * to max_alloc bytes, and we can span up to vdev_removal_max_span
1256 * bytes of unallocated space at a time. "segs" will track the
1257 * allocated segments that we are copying. We may also be copying
1258 * free segments (of up to vdev_removal_max_span bytes).
1260 range_tree_t
*segs
= range_tree_create(NULL
, NULL
);
1262 range_seg_t
*rs
= range_tree_first(svr
->svr_allocd_segs
);
1267 uint64_t seg_length
;
1269 if (range_tree_is_empty(segs
)) {
1270 /* need to truncate the first seg based on max_alloc */
1272 MIN(rs
->rs_end
- rs
->rs_start
, *max_alloc
);
1274 if (rs
->rs_start
- range_tree_max(segs
) >
1275 vdev_removal_max_span
) {
1277 * Including this segment would cause us to
1278 * copy a larger unneeded chunk than is allowed.
1281 } else if (rs
->rs_end
- range_tree_min(segs
) >
1284 * This additional segment would extend past
1285 * max_alloc. Rather than splitting this
1286 * segment, leave it for the next mapping.
1290 seg_length
= rs
->rs_end
- rs
->rs_start
;
1294 range_tree_add(segs
, rs
->rs_start
, seg_length
);
1295 range_tree_remove(svr
->svr_allocd_segs
,
1296 rs
->rs_start
, seg_length
);
1299 if (range_tree_is_empty(segs
)) {
1300 mutex_exit(&svr
->svr_lock
);
1301 range_tree_destroy(segs
);
1305 if (svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] == 0) {
1306 dsl_sync_task_nowait(dmu_tx_pool(tx
), vdev_mapping_sync
,
1307 svr
, 0, ZFS_SPACE_CHECK_NONE
, tx
);
1310 svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] = range_tree_max(segs
);
1313 * Note: this is the amount of *allocated* space
1314 * that we are taking care of each txg.
1316 svr
->svr_bytes_done
[txg
& TXG_MASK
] += range_tree_space(segs
);
1318 mutex_exit(&svr
->svr_lock
);
1320 zio_alloc_list_t zal
;
1321 metaslab_trace_init(&zal
);
1322 uint64_t thismax
= SPA_MAXBLOCKSIZE
;
1323 while (!range_tree_is_empty(segs
)) {
1324 int error
= spa_vdev_copy_segment(vd
,
1325 segs
, thismax
, txg
, vca
, &zal
);
1327 if (error
== ENOSPC
) {
1329 * Cut our segment in half, and don't try this
1330 * segment size again this txg. Note that the
1331 * allocation size must be aligned to the highest
1332 * ashift in the pool, so that the allocation will
1333 * not be padded out to a multiple of the ashift,
1334 * which could cause us to think that this mapping
1335 * is larger than we intended.
1337 ASSERT3U(spa
->spa_max_ashift
, >=, SPA_MINBLOCKSHIFT
);
1338 ASSERT3U(spa
->spa_max_ashift
, ==, spa
->spa_min_ashift
);
1339 uint64_t attempted
=
1340 MIN(range_tree_span(segs
), thismax
);
1341 thismax
= P2ROUNDUP(attempted
/ 2,
1342 1 << spa
->spa_max_ashift
);
1344 * The minimum-size allocation can not fail.
1346 ASSERT3U(attempted
, >, 1 << spa
->spa_max_ashift
);
1347 *max_alloc
= attempted
- (1 << spa
->spa_max_ashift
);
1352 * We've performed an allocation, so reset the
1355 metaslab_trace_fini(&zal
);
1356 metaslab_trace_init(&zal
);
1359 metaslab_trace_fini(&zal
);
1360 range_tree_destroy(segs
);
1364 * The removal thread operates in open context. It iterates over all
1365 * allocated space in the vdev, by loading each metaslab's spacemap.
1366 * For each contiguous segment of allocated space (capping the segment
1367 * size at SPA_MAXBLOCKSIZE), we:
1368 * - Allocate space for it on another vdev.
1369 * - Create a new mapping from the old location to the new location
1370 * (as a record in svr_new_segments).
1371 * - Initiate a physical read zio to get the data off the removing disk.
1372 * - In the read zio's done callback, initiate a physical write zio to
1373 * write it to the new vdev.
1374 * Note that all of this will take effect when a particular TXG syncs.
1375 * The sync thread ensures that all the phys reads and writes for the syncing
1376 * TXG have completed (see spa_txg_zio) and writes the new mappings to disk
1377 * (see vdev_mapping_sync()).
1380 spa_vdev_remove_thread(void *arg
)
1383 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1384 vdev_copy_arg_t vca
;
1385 uint64_t max_alloc
= zfs_remove_max_segment
;
1386 uint64_t last_txg
= 0;
1388 spa_config_enter(spa
, SCL_CONFIG
, FTAG
, RW_READER
);
1389 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1390 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
1391 uint64_t start_offset
= vdev_indirect_mapping_max_offset(vim
);
1393 ASSERT3P(vd
->vdev_ops
, !=, &vdev_indirect_ops
);
1394 ASSERT(vdev_is_concrete(vd
));
1395 ASSERT(vd
->vdev_removing
);
1396 ASSERT(vd
->vdev_indirect_config
.vic_mapping_object
!= 0);
1397 ASSERT(vim
!= NULL
);
1399 mutex_init(&vca
.vca_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1400 cv_init(&vca
.vca_cv
, NULL
, CV_DEFAULT
, NULL
);
1401 vca
.vca_outstanding_bytes
= 0;
1402 vca
.vca_read_error_bytes
= 0;
1403 vca
.vca_write_error_bytes
= 0;
1405 mutex_enter(&svr
->svr_lock
);
1408 * Start from vim_max_offset so we pick up where we left off
1409 * if we are restarting the removal after opening the pool.
1412 for (msi
= start_offset
>> vd
->vdev_ms_shift
;
1413 msi
< vd
->vdev_ms_count
&& !svr
->svr_thread_exit
; msi
++) {
1414 metaslab_t
*msp
= vd
->vdev_ms
[msi
];
1415 ASSERT3U(msi
, <=, vd
->vdev_ms_count
);
1417 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1419 mutex_enter(&msp
->ms_sync_lock
);
1420 mutex_enter(&msp
->ms_lock
);
1423 * Assert nothing in flight -- ms_*tree is empty.
1425 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1426 ASSERT0(range_tree_space(msp
->ms_allocating
[i
]));
1430 * If the metaslab has ever been allocated from (ms_sm!=NULL),
1431 * read the allocated segments from the space map object
1432 * into svr_allocd_segs. Since we do this while holding
1433 * svr_lock and ms_sync_lock, concurrent frees (which
1434 * would have modified the space map) will wait for us
1435 * to finish loading the spacemap, and then take the
1436 * appropriate action (see free_from_removing_vdev()).
1438 if (msp
->ms_sm
!= NULL
) {
1439 VERIFY0(space_map_load(msp
->ms_sm
,
1440 svr
->svr_allocd_segs
, SM_ALLOC
));
1442 range_tree_walk(msp
->ms_freeing
,
1443 range_tree_remove
, svr
->svr_allocd_segs
);
1446 * When we are resuming from a paused removal (i.e.
1447 * when importing a pool with a removal in progress),
1448 * discard any state that we have already processed.
1450 range_tree_clear(svr
->svr_allocd_segs
, 0, start_offset
);
1452 mutex_exit(&msp
->ms_lock
);
1453 mutex_exit(&msp
->ms_sync_lock
);
1456 zfs_dbgmsg("copying %llu segments for metaslab %llu",
1457 avl_numnodes(&svr
->svr_allocd_segs
->rt_root
),
1460 while (!svr
->svr_thread_exit
&&
1461 !range_tree_is_empty(svr
->svr_allocd_segs
)) {
1463 mutex_exit(&svr
->svr_lock
);
1466 * We need to periodically drop the config lock so that
1467 * writers can get in. Additionally, we can't wait
1468 * for a txg to sync while holding a config lock
1469 * (since a waiting writer could cause a 3-way deadlock
1470 * with the sync thread, which also gets a config
1471 * lock for reader). So we can't hold the config lock
1472 * while calling dmu_tx_assign().
1474 spa_config_exit(spa
, SCL_CONFIG
, FTAG
);
1477 * This delay will pause the removal around the point
1478 * specified by zfs_removal_suspend_progress. We do this
1479 * solely from the test suite or during debugging.
1481 uint64_t bytes_copied
=
1482 spa
->spa_removing_phys
.sr_copied
;
1483 for (int i
= 0; i
< TXG_SIZE
; i
++)
1484 bytes_copied
+= svr
->svr_bytes_done
[i
];
1485 while (zfs_removal_suspend_progress
&&
1486 !svr
->svr_thread_exit
)
1489 mutex_enter(&vca
.vca_lock
);
1490 while (vca
.vca_outstanding_bytes
>
1491 zfs_remove_max_copy_bytes
) {
1492 cv_wait(&vca
.vca_cv
, &vca
.vca_lock
);
1494 mutex_exit(&vca
.vca_lock
);
1497 dmu_tx_create_dd(spa_get_dsl(spa
)->dp_mos_dir
);
1498 dmu_tx_hold_space(tx
, SPA_MAXBLOCKSIZE
);
1499 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
1500 uint64_t txg
= dmu_tx_get_txg(tx
);
1503 * Reacquire the vdev_config lock. The vdev_t
1504 * that we're removing may have changed, e.g. due
1505 * to a vdev_attach or vdev_detach.
1507 spa_config_enter(spa
, SCL_CONFIG
, FTAG
, RW_READER
);
1508 vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1510 if (txg
!= last_txg
)
1511 max_alloc
= zfs_remove_max_segment
;
1514 spa_vdev_copy_impl(vd
, svr
, &vca
, &max_alloc
, tx
);
1517 mutex_enter(&svr
->svr_lock
);
1520 mutex_enter(&vca
.vca_lock
);
1521 if (zfs_removal_ignore_errors
== 0 &&
1522 (vca
.vca_read_error_bytes
> 0 ||
1523 vca
.vca_write_error_bytes
> 0)) {
1524 svr
->svr_thread_exit
= B_TRUE
;
1526 mutex_exit(&vca
.vca_lock
);
1529 mutex_exit(&svr
->svr_lock
);
1531 spa_config_exit(spa
, SCL_CONFIG
, FTAG
);
1534 * Wait for all copies to finish before cleaning up the vca.
1536 txg_wait_synced(spa
->spa_dsl_pool
, 0);
1537 ASSERT0(vca
.vca_outstanding_bytes
);
1539 mutex_destroy(&vca
.vca_lock
);
1540 cv_destroy(&vca
.vca_cv
);
1542 if (svr
->svr_thread_exit
) {
1543 mutex_enter(&svr
->svr_lock
);
1544 range_tree_vacate(svr
->svr_allocd_segs
, NULL
, NULL
);
1545 svr
->svr_thread
= NULL
;
1546 cv_broadcast(&svr
->svr_cv
);
1547 mutex_exit(&svr
->svr_lock
);
1550 * During the removal process an unrecoverable read or write
1551 * error was encountered. The removal process must be
1552 * cancelled or this damage may become permanent.
1554 if (zfs_removal_ignore_errors
== 0 &&
1555 (vca
.vca_read_error_bytes
> 0 ||
1556 vca
.vca_write_error_bytes
> 0)) {
1557 zfs_dbgmsg("canceling removal due to IO errors: "
1558 "[read_error_bytes=%llu] [write_error_bytes=%llu]",
1559 vca
.vca_read_error_bytes
,
1560 vca
.vca_write_error_bytes
);
1561 spa_vdev_remove_cancel_impl(spa
);
1564 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1565 vdev_remove_complete(spa
);
1570 spa_vdev_remove_suspend(spa_t
*spa
)
1572 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1577 mutex_enter(&svr
->svr_lock
);
1578 svr
->svr_thread_exit
= B_TRUE
;
1579 while (svr
->svr_thread
!= NULL
)
1580 cv_wait(&svr
->svr_cv
, &svr
->svr_lock
);
1581 svr
->svr_thread_exit
= B_FALSE
;
1582 mutex_exit(&svr
->svr_lock
);
1587 spa_vdev_remove_cancel_check(void *arg
, dmu_tx_t
*tx
)
1589 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1591 if (spa
->spa_vdev_removal
== NULL
)
1592 return (ENOTACTIVE
);
1597 * Cancel a removal by freeing all entries from the partial mapping
1598 * and marking the vdev as no longer being removing.
1602 spa_vdev_remove_cancel_sync(void *arg
, dmu_tx_t
*tx
)
1604 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1605 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1606 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1607 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
1608 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
1609 objset_t
*mos
= spa
->spa_meta_objset
;
1611 ASSERT3P(svr
->svr_thread
, ==, NULL
);
1613 spa_feature_decr(spa
, SPA_FEATURE_DEVICE_REMOVAL
, tx
);
1615 boolean_t are_precise
;
1616 VERIFY0(vdev_obsolete_counts_are_precise(vd
, &are_precise
));
1618 spa_feature_decr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
1619 VERIFY0(zap_remove(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
1620 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE
, tx
));
1623 uint64_t obsolete_sm_object
;
1624 VERIFY0(vdev_obsolete_sm_object(vd
, &obsolete_sm_object
));
1625 if (obsolete_sm_object
!= 0) {
1626 ASSERT(vd
->vdev_obsolete_sm
!= NULL
);
1627 ASSERT3U(obsolete_sm_object
, ==,
1628 space_map_object(vd
->vdev_obsolete_sm
));
1630 space_map_free(vd
->vdev_obsolete_sm
, tx
);
1631 VERIFY0(zap_remove(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
1632 VDEV_TOP_ZAP_INDIRECT_OBSOLETE_SM
, tx
));
1633 space_map_close(vd
->vdev_obsolete_sm
);
1634 vd
->vdev_obsolete_sm
= NULL
;
1635 spa_feature_decr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
1637 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1638 ASSERT(list_is_empty(&svr
->svr_new_segments
[i
]));
1639 ASSERT3U(svr
->svr_max_offset_to_sync
[i
], <=,
1640 vdev_indirect_mapping_max_offset(vim
));
1643 for (uint64_t msi
= 0; msi
< vd
->vdev_ms_count
; msi
++) {
1644 metaslab_t
*msp
= vd
->vdev_ms
[msi
];
1646 if (msp
->ms_start
>= vdev_indirect_mapping_max_offset(vim
))
1649 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1651 mutex_enter(&msp
->ms_lock
);
1654 * Assert nothing in flight -- ms_*tree is empty.
1656 for (int i
= 0; i
< TXG_SIZE
; i
++)
1657 ASSERT0(range_tree_space(msp
->ms_allocating
[i
]));
1658 for (int i
= 0; i
< TXG_DEFER_SIZE
; i
++)
1659 ASSERT0(range_tree_space(msp
->ms_defer
[i
]));
1660 ASSERT0(range_tree_space(msp
->ms_freed
));
1662 if (msp
->ms_sm
!= NULL
) {
1663 mutex_enter(&svr
->svr_lock
);
1664 VERIFY0(space_map_load(msp
->ms_sm
,
1665 svr
->svr_allocd_segs
, SM_ALLOC
));
1666 range_tree_walk(msp
->ms_freeing
,
1667 range_tree_remove
, svr
->svr_allocd_segs
);
1670 * Clear everything past what has been synced,
1671 * because we have not allocated mappings for it yet.
1673 uint64_t syncd
= vdev_indirect_mapping_max_offset(vim
);
1674 uint64_t sm_end
= msp
->ms_sm
->sm_start
+
1675 msp
->ms_sm
->sm_size
;
1677 range_tree_clear(svr
->svr_allocd_segs
,
1678 syncd
, sm_end
- syncd
);
1680 mutex_exit(&svr
->svr_lock
);
1682 mutex_exit(&msp
->ms_lock
);
1684 mutex_enter(&svr
->svr_lock
);
1685 range_tree_vacate(svr
->svr_allocd_segs
,
1686 free_mapped_segment_cb
, vd
);
1687 mutex_exit(&svr
->svr_lock
);
1691 * Note: this must happen after we invoke free_mapped_segment_cb,
1692 * because it adds to the obsolete_segments.
1694 range_tree_vacate(vd
->vdev_obsolete_segments
, NULL
, NULL
);
1696 ASSERT3U(vic
->vic_mapping_object
, ==,
1697 vdev_indirect_mapping_object(vd
->vdev_indirect_mapping
));
1698 vdev_indirect_mapping_close(vd
->vdev_indirect_mapping
);
1699 vd
->vdev_indirect_mapping
= NULL
;
1700 vdev_indirect_mapping_free(mos
, vic
->vic_mapping_object
, tx
);
1701 vic
->vic_mapping_object
= 0;
1703 ASSERT3U(vic
->vic_births_object
, ==,
1704 vdev_indirect_births_object(vd
->vdev_indirect_births
));
1705 vdev_indirect_births_close(vd
->vdev_indirect_births
);
1706 vd
->vdev_indirect_births
= NULL
;
1707 vdev_indirect_births_free(mos
, vic
->vic_births_object
, tx
);
1708 vic
->vic_births_object
= 0;
1711 * We may have processed some frees from the removing vdev in this
1712 * txg, thus increasing svr_bytes_done; discard that here to
1713 * satisfy the assertions in spa_vdev_removal_destroy().
1714 * Note that future txg's can not have any bytes_done, because
1715 * future TXG's are only modified from open context, and we have
1716 * already shut down the copying thread.
1718 svr
->svr_bytes_done
[dmu_tx_get_txg(tx
) & TXG_MASK
] = 0;
1719 spa_finish_removal(spa
, DSS_CANCELED
, tx
);
1721 vd
->vdev_removing
= B_FALSE
;
1722 vdev_config_dirty(vd
);
1724 zfs_dbgmsg("canceled device removal for vdev %llu in %llu",
1725 vd
->vdev_id
, dmu_tx_get_txg(tx
));
1726 spa_history_log_internal(spa
, "vdev remove canceled", tx
,
1727 "%s vdev %llu %s", spa_name(spa
),
1728 vd
->vdev_id
, (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
1732 spa_vdev_remove_cancel_impl(spa_t
*spa
)
1734 uint64_t vdid
= spa
->spa_vdev_removal
->svr_vdev_id
;
1736 int error
= dsl_sync_task(spa
->spa_name
, spa_vdev_remove_cancel_check
,
1737 spa_vdev_remove_cancel_sync
, NULL
, 0,
1738 ZFS_SPACE_CHECK_EXTRA_RESERVED
);
1741 spa_config_enter(spa
, SCL_ALLOC
| SCL_VDEV
, FTAG
, RW_WRITER
);
1742 vdev_t
*vd
= vdev_lookup_top(spa
, vdid
);
1743 metaslab_group_activate(vd
->vdev_mg
);
1744 spa_config_exit(spa
, SCL_ALLOC
| SCL_VDEV
, FTAG
);
1751 spa_vdev_remove_cancel(spa_t
*spa
)
1753 spa_vdev_remove_suspend(spa
);
1755 if (spa
->spa_vdev_removal
== NULL
)
1756 return (ENOTACTIVE
);
1758 return (spa_vdev_remove_cancel_impl(spa
));
1762 svr_sync(spa_t
*spa
, dmu_tx_t
*tx
)
1764 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1765 int txgoff
= dmu_tx_get_txg(tx
) & TXG_MASK
;
1771 * This check is necessary so that we do not dirty the
1772 * DIRECTORY_OBJECT via spa_sync_removing_state() when there
1773 * is nothing to do. Dirtying it every time would prevent us
1774 * from syncing-to-convergence.
1776 if (svr
->svr_bytes_done
[txgoff
] == 0)
1780 * Update progress accounting.
1782 spa
->spa_removing_phys
.sr_copied
+= svr
->svr_bytes_done
[txgoff
];
1783 svr
->svr_bytes_done
[txgoff
] = 0;
1785 spa_sync_removing_state(spa
, tx
);
1789 vdev_remove_make_hole_and_free(vdev_t
*vd
)
1791 uint64_t id
= vd
->vdev_id
;
1792 spa_t
*spa
= vd
->vdev_spa
;
1793 vdev_t
*rvd
= spa
->spa_root_vdev
;
1794 boolean_t last_vdev
= (id
== (rvd
->vdev_children
- 1));
1796 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
1797 ASSERT(spa_config_held(spa
, SCL_ALL
, RW_WRITER
) == SCL_ALL
);
1802 vdev_compact_children(rvd
);
1804 vd
= vdev_alloc_common(spa
, id
, 0, &vdev_hole_ops
);
1805 vdev_add_child(rvd
, vd
);
1807 vdev_config_dirty(rvd
);
1810 * Reassess the health of our root vdev.
1816 * Remove a log device. The config lock is held for the specified TXG.
1819 spa_vdev_remove_log(vdev_t
*vd
, uint64_t *txg
)
1821 metaslab_group_t
*mg
= vd
->vdev_mg
;
1822 spa_t
*spa
= vd
->vdev_spa
;
1825 ASSERT(vd
->vdev_islog
);
1826 ASSERT(vd
== vd
->vdev_top
);
1827 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
1830 * Stop allocating from this vdev.
1832 metaslab_group_passivate(mg
);
1835 * Wait for the youngest allocations and frees to sync,
1836 * and then wait for the deferral of those frees to finish.
1838 spa_vdev_config_exit(spa
, NULL
,
1839 *txg
+ TXG_CONCURRENT_STATES
+ TXG_DEFER_SIZE
, 0, FTAG
);
1842 * Evacuate the device. We don't hold the config lock as
1843 * writer since we need to do I/O but we do keep the
1844 * spa_namespace_lock held. Once this completes the device
1845 * should no longer have any blocks allocated on it.
1847 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
1848 if (vd
->vdev_stat
.vs_alloc
!= 0)
1849 error
= spa_reset_logs(spa
);
1851 *txg
= spa_vdev_config_enter(spa
);
1854 metaslab_group_activate(mg
);
1857 ASSERT0(vd
->vdev_stat
.vs_alloc
);
1860 * The evacuation succeeded. Remove any remaining MOS metadata
1861 * associated with this vdev, and wait for these changes to sync.
1863 vd
->vdev_removing
= B_TRUE
;
1865 vdev_dirty_leaves(vd
, VDD_DTL
, *txg
);
1866 vdev_config_dirty(vd
);
1868 vdev_metaslab_fini(vd
);
1870 spa_vdev_config_exit(spa
, NULL
, *txg
, 0, FTAG
);
1872 /* Stop initializing */
1873 vdev_initialize_stop_all(vd
, VDEV_INITIALIZE_CANCELED
);
1875 *txg
= spa_vdev_config_enter(spa
);
1877 sysevent_t
*ev
= spa_event_create(spa
, vd
, NULL
,
1878 ESC_ZFS_VDEV_REMOVE_DEV
);
1879 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
1880 ASSERT(spa_config_held(spa
, SCL_ALL
, RW_WRITER
) == SCL_ALL
);
1882 /* The top ZAP should have been destroyed by vdev_remove_empty. */
1883 ASSERT0(vd
->vdev_top_zap
);
1884 /* The leaf ZAP should have been destroyed by vdev_dtl_sync. */
1885 ASSERT0(vd
->vdev_leaf_zap
);
1887 (void) vdev_label_init(vd
, 0, VDEV_LABEL_REMOVE
);
1889 if (list_link_active(&vd
->vdev_state_dirty_node
))
1890 vdev_state_clean(vd
);
1891 if (list_link_active(&vd
->vdev_config_dirty_node
))
1892 vdev_config_clean(vd
);
1894 ASSERT0(vd
->vdev_stat
.vs_alloc
);
1897 * Clean up the vdev namespace.
1899 vdev_remove_make_hole_and_free(vd
);
1908 spa_vdev_remove_top_check(vdev_t
*vd
)
1910 spa_t
*spa
= vd
->vdev_spa
;
1912 if (vd
!= vd
->vdev_top
)
1913 return (SET_ERROR(ENOTSUP
));
1915 if (!spa_feature_is_enabled(spa
, SPA_FEATURE_DEVICE_REMOVAL
))
1916 return (SET_ERROR(ENOTSUP
));
1918 /* available space in the pool's normal class */
1919 uint64_t available
= dsl_dir_space_available(
1920 spa
->spa_dsl_pool
->dp_root_dir
, NULL
, 0, B_TRUE
);
1922 metaslab_class_t
*mc
= vd
->vdev_mg
->mg_class
;
1925 * When removing a vdev from an allocation class that has
1926 * remaining vdevs, include available space from the class.
1928 if (mc
!= spa_normal_class(spa
) && mc
->mc_groups
> 1) {
1929 uint64_t class_avail
= metaslab_class_get_space(mc
) -
1930 metaslab_class_get_alloc(mc
);
1932 /* add class space, adjusted for overhead */
1933 available
+= (class_avail
* 94) / 100;
1937 * There has to be enough free space to remove the
1938 * device and leave double the "slop" space (i.e. we
1939 * must leave at least 3% of the pool free, in addition to
1940 * the normal slop space).
1942 if (available
< vd
->vdev_stat
.vs_dspace
+ spa_get_slop_space(spa
)) {
1943 return (SET_ERROR(ENOSPC
));
1947 * There can not be a removal in progress.
1949 if (spa
->spa_removing_phys
.sr_state
== DSS_SCANNING
)
1950 return (SET_ERROR(EBUSY
));
1953 * The device must have all its data.
1955 if (!vdev_dtl_empty(vd
, DTL_MISSING
) ||
1956 !vdev_dtl_empty(vd
, DTL_OUTAGE
))
1957 return (SET_ERROR(EBUSY
));
1960 * The device must be healthy.
1962 if (!vdev_readable(vd
))
1963 return (SET_ERROR(EIO
));
1966 * All vdevs in normal class must have the same ashift.
1968 if (spa
->spa_max_ashift
!= spa
->spa_min_ashift
) {
1969 return (SET_ERROR(EINVAL
));
1973 * All vdevs in normal class must have the same ashift
1976 vdev_t
*rvd
= spa
->spa_root_vdev
;
1977 int num_indirect
= 0;
1978 for (uint64_t id
= 0; id
< rvd
->vdev_children
; id
++) {
1979 vdev_t
*cvd
= rvd
->vdev_child
[id
];
1980 if (cvd
->vdev_ashift
!= 0 && !cvd
->vdev_islog
)
1981 ASSERT3U(cvd
->vdev_ashift
, ==, spa
->spa_max_ashift
);
1982 if (cvd
->vdev_ops
== &vdev_indirect_ops
)
1984 if (!vdev_is_concrete(cvd
))
1986 if (cvd
->vdev_ops
== &vdev_raidz_ops
)
1987 return (SET_ERROR(EINVAL
));
1989 * Need the mirror to be mirror of leaf vdevs only
1991 if (cvd
->vdev_ops
== &vdev_mirror_ops
) {
1992 for (uint64_t cid
= 0;
1993 cid
< cvd
->vdev_children
; cid
++) {
1994 if (!cvd
->vdev_child
[cid
]->vdev_ops
->
1996 return (SET_ERROR(EINVAL
));
2005 * Initiate removal of a top-level vdev, reducing the total space in the pool.
2006 * The config lock is held for the specified TXG. Once initiated,
2007 * evacuation of all allocated space (copying it to other vdevs) happens
2008 * in the background (see spa_vdev_remove_thread()), and can be canceled
2009 * (see spa_vdev_remove_cancel()). If successful, the vdev will
2010 * be transformed to an indirect vdev (see spa_vdev_remove_complete()).
2013 spa_vdev_remove_top(vdev_t
*vd
, uint64_t *txg
)
2015 spa_t
*spa
= vd
->vdev_spa
;
2019 * Check for errors up-front, so that we don't waste time
2020 * passivating the metaslab group and clearing the ZIL if there
2023 error
= spa_vdev_remove_top_check(vd
);
2028 * Stop allocating from this vdev. Note that we must check
2029 * that this is not the only device in the pool before
2030 * passivating, otherwise we will not be able to make
2031 * progress because we can't allocate from any vdevs.
2032 * The above check for sufficient free space serves this
2035 metaslab_group_t
*mg
= vd
->vdev_mg
;
2036 metaslab_group_passivate(mg
);
2039 * Wait for the youngest allocations and frees to sync,
2040 * and then wait for the deferral of those frees to finish.
2042 spa_vdev_config_exit(spa
, NULL
,
2043 *txg
+ TXG_CONCURRENT_STATES
+ TXG_DEFER_SIZE
, 0, FTAG
);
2046 * We must ensure that no "stubby" log blocks are allocated
2047 * on the device to be removed. These blocks could be
2048 * written at any time, including while we are in the middle
2051 error
= spa_reset_logs(spa
);
2054 * We stop any initializing that is currently in progress but leave
2055 * the state as "active". This will allow the initializing to resume
2056 * if the removal is canceled sometime later.
2058 vdev_initialize_stop_all(vd
, VDEV_INITIALIZE_ACTIVE
);
2060 *txg
= spa_vdev_config_enter(spa
);
2063 * Things might have changed while the config lock was dropped
2064 * (e.g. space usage). Check for errors again.
2067 error
= spa_vdev_remove_top_check(vd
);
2070 metaslab_group_activate(mg
);
2071 spa_async_request(spa
, SPA_ASYNC_INITIALIZE_RESTART
);
2075 vd
->vdev_removing
= B_TRUE
;
2077 vdev_dirty_leaves(vd
, VDD_DTL
, *txg
);
2078 vdev_config_dirty(vd
);
2079 dmu_tx_t
*tx
= dmu_tx_create_assigned(spa
->spa_dsl_pool
, *txg
);
2080 dsl_sync_task_nowait(spa
->spa_dsl_pool
,
2081 vdev_remove_initiate_sync
,
2082 (void *)(uintptr_t)vd
->vdev_id
, 0, ZFS_SPACE_CHECK_NONE
, tx
);
2089 * Remove a device from the pool.
2091 * Removing a device from the vdev namespace requires several steps
2092 * and can take a significant amount of time. As a result we use
2093 * the spa_vdev_config_[enter/exit] functions which allow us to
2094 * grab and release the spa_config_lock while still holding the namespace
2095 * lock. During each step the configuration is synced out.
2098 spa_vdev_remove(spa_t
*spa
, uint64_t guid
, boolean_t unspare
)
2101 nvlist_t
**spares
, **l2cache
, *nv
;
2103 uint_t nspares
, nl2cache
;
2105 boolean_t locked
= MUTEX_HELD(&spa_namespace_lock
);
2106 sysevent_t
*ev
= NULL
;
2107 char *vd_type
= NULL
, *vd_path
= NULL
;
2109 ASSERT(spa_writeable(spa
));
2112 txg
= spa_vdev_enter(spa
);
2114 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
2115 if (spa_feature_is_active(spa
, SPA_FEATURE_POOL_CHECKPOINT
)) {
2116 error
= (spa_has_checkpoint(spa
)) ?
2117 ZFS_ERR_CHECKPOINT_EXISTS
: ZFS_ERR_DISCARDING_CHECKPOINT
;
2120 return (spa_vdev_exit(spa
, NULL
, txg
, error
));
2125 vd
= spa_lookup_by_guid(spa
, guid
, B_FALSE
);
2127 if (spa
->spa_spares
.sav_vdevs
!= NULL
&&
2128 nvlist_lookup_nvlist_array(spa
->spa_spares
.sav_config
,
2129 ZPOOL_CONFIG_SPARES
, &spares
, &nspares
) == 0 &&
2130 (nv
= spa_nvlist_lookup_by_guid(spares
, nspares
, guid
)) != NULL
) {
2132 * Only remove the hot spare if it's not currently in use
2135 if (vd
== NULL
|| unspare
) {
2137 vd
= spa_lookup_by_guid(spa
, guid
, B_TRUE
);
2138 ev
= spa_event_create(spa
, vd
, NULL
,
2139 ESC_ZFS_VDEV_REMOVE_AUX
);
2141 vd_type
= VDEV_TYPE_SPARE
;
2142 vd_path
= fnvlist_lookup_string(nv
, ZPOOL_CONFIG_PATH
);
2143 spa_vdev_remove_aux(spa
->spa_spares
.sav_config
,
2144 ZPOOL_CONFIG_SPARES
, spares
, nspares
, nv
);
2145 spa_load_spares(spa
);
2146 spa
->spa_spares
.sav_sync
= B_TRUE
;
2148 error
= SET_ERROR(EBUSY
);
2150 } else if (spa
->spa_l2cache
.sav_vdevs
!= NULL
&&
2151 nvlist_lookup_nvlist_array(spa
->spa_l2cache
.sav_config
,
2152 ZPOOL_CONFIG_L2CACHE
, &l2cache
, &nl2cache
) == 0 &&
2153 (nv
= spa_nvlist_lookup_by_guid(l2cache
, nl2cache
, guid
)) != NULL
) {
2154 vd_type
= VDEV_TYPE_L2CACHE
;
2155 vd_path
= fnvlist_lookup_string(nv
, ZPOOL_CONFIG_PATH
);
2157 * Cache devices can always be removed.
2159 vd
= spa_lookup_by_guid(spa
, guid
, B_TRUE
);
2160 ev
= spa_event_create(spa
, vd
, NULL
, ESC_ZFS_VDEV_REMOVE_AUX
);
2161 spa_vdev_remove_aux(spa
->spa_l2cache
.sav_config
,
2162 ZPOOL_CONFIG_L2CACHE
, l2cache
, nl2cache
, nv
);
2163 spa_load_l2cache(spa
);
2164 spa
->spa_l2cache
.sav_sync
= B_TRUE
;
2165 } else if (vd
!= NULL
&& vd
->vdev_islog
) {
2168 vd_path
= (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-";
2169 error
= spa_vdev_remove_log(vd
, &txg
);
2170 } else if (vd
!= NULL
) {
2172 error
= spa_vdev_remove_top(vd
, &txg
);
2175 * There is no vdev of any kind with the specified guid.
2177 error
= SET_ERROR(ENOENT
);
2181 error
= spa_vdev_exit(spa
, NULL
, txg
, error
);
2184 * Logging must be done outside the spa config lock. Otherwise,
2185 * this code path could end up holding the spa config lock while
2186 * waiting for a txg_sync so it can write to the internal log.
2187 * Doing that would prevent the txg sync from actually happening,
2188 * causing a deadlock.
2190 if (error
== 0 && vd_type
!= NULL
&& vd_path
!= NULL
) {
2191 spa_history_log_internal(spa
, "vdev remove", NULL
,
2192 "%s vdev (%s) %s", spa_name(spa
), vd_type
, vd_path
);
2202 spa_removal_get_stats(spa_t
*spa
, pool_removal_stat_t
*prs
)
2204 prs
->prs_state
= spa
->spa_removing_phys
.sr_state
;
2206 if (prs
->prs_state
== DSS_NONE
)
2207 return (SET_ERROR(ENOENT
));
2209 prs
->prs_removing_vdev
= spa
->spa_removing_phys
.sr_removing_vdev
;
2210 prs
->prs_start_time
= spa
->spa_removing_phys
.sr_start_time
;
2211 prs
->prs_end_time
= spa
->spa_removing_phys
.sr_end_time
;
2212 prs
->prs_to_copy
= spa
->spa_removing_phys
.sr_to_copy
;
2213 prs
->prs_copied
= spa
->spa_removing_phys
.sr_copied
;
2215 prs
->prs_mapping_memory
= 0;
2216 uint64_t indirect_vdev_id
=
2217 spa
->spa_removing_phys
.sr_prev_indirect_vdev
;
2218 while (indirect_vdev_id
!= -1) {
2219 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[indirect_vdev_id
];
2220 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
2221 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
2223 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
2224 prs
->prs_mapping_memory
+= vdev_indirect_mapping_size(vim
);
2225 indirect_vdev_id
= vic
->vic_prev_indirect_vdev
;
2231 #if defined(_KERNEL)
2232 module_param(zfs_removal_ignore_errors
, int, 0644);
2233 MODULE_PARM_DESC(zfs_removal_ignore_errors
,
2234 "Ignore hard IO errors when removing device");
2236 module_param(zfs_remove_max_segment
, int, 0644);
2237 MODULE_PARM_DESC(zfs_remove_max_segment
,
2238 "Largest contiguous segment to allocate when removing device");
2240 module_param(vdev_removal_max_span
, int, 0644);
2241 MODULE_PARM_DESC(vdev_removal_max_span
,
2242 "Largest span of free chunks a remap segment can span");
2245 module_param(zfs_removal_suspend_progress
, int, 0644);
2246 MODULE_PARM_DESC(zfs_removal_suspend_progress
,
2247 "Pause device removal after this many bytes are copied "
2248 "(debug use only - causes removal to hang)");
2251 EXPORT_SYMBOL(free_from_removing_vdev
);
2252 EXPORT_SYMBOL(spa_removal_get_stats
);
2253 EXPORT_SYMBOL(spa_remove_init
);
2254 EXPORT_SYMBOL(spa_restart_removal
);
2255 EXPORT_SYMBOL(spa_vdev_removal_destroy
);
2256 EXPORT_SYMBOL(spa_vdev_remove
);
2257 EXPORT_SYMBOL(spa_vdev_remove_cancel
);
2258 EXPORT_SYMBOL(spa_vdev_remove_suspend
);
2259 EXPORT_SYMBOL(svr_sync
);