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
)
295 * Sync tasks happen before metaslab_sync(), therefore
296 * smp_alloc and sm_alloc must be the same.
298 ASSERT3U(space_map_allocated(ms
->ms_sm
), ==,
299 ms
->ms_sm
->sm_phys
->smp_alloc
);
301 spa
->spa_removing_phys
.sr_to_copy
+=
302 space_map_allocated(ms
->ms_sm
);
305 * Space which we are freeing this txg does not need to
308 spa
->spa_removing_phys
.sr_to_copy
-=
309 range_tree_space(ms
->ms_freeing
);
311 ASSERT0(range_tree_space(ms
->ms_freed
));
312 for (int t
= 0; t
< TXG_SIZE
; t
++)
313 ASSERT0(range_tree_space(ms
->ms_allocating
[t
]));
317 * Sync tasks are called before metaslab_sync(), so there should
318 * be no already-synced metaslabs in the TXG_CLEAN list.
320 ASSERT3P(txg_list_head(&vd
->vdev_ms_list
, TXG_CLEAN(txg
)), ==, NULL
);
322 spa_sync_removing_state(spa
, tx
);
325 * All blocks that we need to read the most recent mapping must be
326 * stored on concrete vdevs. Therefore, we must dirty anything that
327 * is read before spa_remove_init(). Specifically, the
328 * spa_config_object. (Note that although we already modified the
329 * spa_config_object in spa_sync_removing_state, that may not have
330 * modified all blocks of the object.)
332 dmu_object_info_t doi
;
333 VERIFY0(dmu_object_info(mos
, DMU_POOL_DIRECTORY_OBJECT
, &doi
));
334 for (uint64_t offset
= 0; offset
< doi
.doi_max_offset
; ) {
336 VERIFY0(dmu_buf_hold(mos
, DMU_POOL_DIRECTORY_OBJECT
,
337 offset
, FTAG
, &dbuf
, 0));
338 dmu_buf_will_dirty(dbuf
, tx
);
339 offset
+= dbuf
->db_size
;
340 dmu_buf_rele(dbuf
, FTAG
);
344 * Now that we've allocated the im_object, dirty the vdev to ensure
345 * that the object gets written to the config on disk.
347 vdev_config_dirty(vd
);
349 zfs_dbgmsg("starting removal thread for vdev %llu (%p) in txg %llu "
350 "im_obj=%llu", vd
->vdev_id
, vd
, dmu_tx_get_txg(tx
),
351 vic
->vic_mapping_object
);
353 spa_history_log_internal(spa
, "vdev remove started", tx
,
354 "%s vdev %llu %s", spa_name(spa
), vd
->vdev_id
,
355 (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
357 * Setting spa_vdev_removal causes subsequent frees to call
358 * free_from_removing_vdev(). Note that we don't need any locking
359 * because we are the sync thread, and metaslab_free_impl() is only
360 * called from syncing context (potentially from a zio taskq thread,
361 * but in any case only when there are outstanding free i/os, which
364 ASSERT3P(spa
->spa_vdev_removal
, ==, NULL
);
365 spa
->spa_vdev_removal
= svr
;
366 svr
->svr_thread
= thread_create(NULL
, 0,
367 spa_vdev_remove_thread
, spa
, 0, &p0
, TS_RUN
, minclsyspri
);
371 * When we are opening a pool, we must read the mapping for each
372 * indirect vdev in order from most recently removed to least
373 * recently removed. We do this because the blocks for the mapping
374 * of older indirect vdevs may be stored on more recently removed vdevs.
375 * In order to read each indirect mapping object, we must have
376 * initialized all more recently removed vdevs.
379 spa_remove_init(spa_t
*spa
)
383 error
= zap_lookup(spa
->spa_dsl_pool
->dp_meta_objset
,
384 DMU_POOL_DIRECTORY_OBJECT
,
385 DMU_POOL_REMOVING
, sizeof (uint64_t),
386 sizeof (spa
->spa_removing_phys
) / sizeof (uint64_t),
387 &spa
->spa_removing_phys
);
389 if (error
== ENOENT
) {
390 spa
->spa_removing_phys
.sr_state
= DSS_NONE
;
391 spa
->spa_removing_phys
.sr_removing_vdev
= -1;
392 spa
->spa_removing_phys
.sr_prev_indirect_vdev
= -1;
393 spa
->spa_indirect_vdevs_loaded
= B_TRUE
;
395 } else if (error
!= 0) {
399 if (spa
->spa_removing_phys
.sr_state
== DSS_SCANNING
) {
401 * We are currently removing a vdev. Create and
402 * initialize a spa_vdev_removal_t from the bonus
403 * buffer of the removing vdevs vdev_im_object, and
404 * initialize its partial mapping.
406 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
407 vdev_t
*vd
= vdev_lookup_top(spa
,
408 spa
->spa_removing_phys
.sr_removing_vdev
);
411 spa_config_exit(spa
, SCL_STATE
, FTAG
);
415 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
417 ASSERT(vdev_is_concrete(vd
));
418 spa_vdev_removal_t
*svr
= spa_vdev_removal_create(vd
);
419 ASSERT3U(svr
->svr_vdev_id
, ==, vd
->vdev_id
);
420 ASSERT(vd
->vdev_removing
);
422 vd
->vdev_indirect_mapping
= vdev_indirect_mapping_open(
423 spa
->spa_meta_objset
, vic
->vic_mapping_object
);
424 vd
->vdev_indirect_births
= vdev_indirect_births_open(
425 spa
->spa_meta_objset
, vic
->vic_births_object
);
426 spa_config_exit(spa
, SCL_STATE
, FTAG
);
428 spa
->spa_vdev_removal
= svr
;
431 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
432 uint64_t indirect_vdev_id
=
433 spa
->spa_removing_phys
.sr_prev_indirect_vdev
;
434 while (indirect_vdev_id
!= UINT64_MAX
) {
435 vdev_t
*vd
= vdev_lookup_top(spa
, indirect_vdev_id
);
436 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
438 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
439 vd
->vdev_indirect_mapping
= vdev_indirect_mapping_open(
440 spa
->spa_meta_objset
, vic
->vic_mapping_object
);
441 vd
->vdev_indirect_births
= vdev_indirect_births_open(
442 spa
->spa_meta_objset
, vic
->vic_births_object
);
444 indirect_vdev_id
= vic
->vic_prev_indirect_vdev
;
446 spa_config_exit(spa
, SCL_STATE
, FTAG
);
449 * Now that we've loaded all the indirect mappings, we can allow
450 * reads from other blocks (e.g. via predictive prefetch).
452 spa
->spa_indirect_vdevs_loaded
= B_TRUE
;
457 spa_restart_removal(spa_t
*spa
)
459 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
465 * In general when this function is called there is no
466 * removal thread running. The only scenario where this
467 * is not true is during spa_import() where this function
468 * is called twice [once from spa_import_impl() and
469 * spa_async_resume()]. Thus, in the scenario where we
470 * import a pool that has an ongoing removal we don't
471 * want to spawn a second thread.
473 if (svr
->svr_thread
!= NULL
)
476 if (!spa_writeable(spa
))
479 zfs_dbgmsg("restarting removal of %llu", svr
->svr_vdev_id
);
480 svr
->svr_thread
= thread_create(NULL
, 0, spa_vdev_remove_thread
, spa
,
481 0, &p0
, TS_RUN
, minclsyspri
);
485 * Process freeing from a device which is in the middle of being removed.
486 * We must handle this carefully so that we attempt to copy freed data,
487 * and we correctly free already-copied data.
490 free_from_removing_vdev(vdev_t
*vd
, uint64_t offset
, uint64_t size
)
492 spa_t
*spa
= vd
->vdev_spa
;
493 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
494 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
495 uint64_t txg
= spa_syncing_txg(spa
);
496 uint64_t max_offset_yet
= 0;
498 ASSERT(vd
->vdev_indirect_config
.vic_mapping_object
!= 0);
499 ASSERT3U(vd
->vdev_indirect_config
.vic_mapping_object
, ==,
500 vdev_indirect_mapping_object(vim
));
501 ASSERT3U(vd
->vdev_id
, ==, svr
->svr_vdev_id
);
503 mutex_enter(&svr
->svr_lock
);
506 * Remove the segment from the removing vdev's spacemap. This
507 * ensures that we will not attempt to copy this space (if the
508 * removal thread has not yet visited it), and also ensures
509 * that we know what is actually allocated on the new vdevs
510 * (needed if we cancel the removal).
512 * Note: we must do the metaslab_free_concrete() with the svr_lock
513 * held, so that the remove_thread can not load this metaslab and then
514 * visit this offset between the time that we metaslab_free_concrete()
515 * and when we check to see if it has been visited.
517 * Note: The checkpoint flag is set to false as having/taking
518 * a checkpoint and removing a device can't happen at the same
521 ASSERT(!spa_has_checkpoint(spa
));
522 metaslab_free_concrete(vd
, offset
, size
, B_FALSE
);
524 uint64_t synced_size
= 0;
525 uint64_t synced_offset
= 0;
526 uint64_t max_offset_synced
= vdev_indirect_mapping_max_offset(vim
);
527 if (offset
< max_offset_synced
) {
529 * The mapping for this offset is already on disk.
530 * Free from the new location.
532 * Note that we use svr_max_synced_offset because it is
533 * updated atomically with respect to the in-core mapping.
534 * By contrast, vim_max_offset is not.
536 * This block may be split between a synced entry and an
537 * in-flight or unvisited entry. Only process the synced
538 * portion of it here.
540 synced_size
= MIN(size
, max_offset_synced
- offset
);
541 synced_offset
= offset
;
543 ASSERT3U(max_offset_yet
, <=, max_offset_synced
);
544 max_offset_yet
= max_offset_synced
;
546 DTRACE_PROBE3(remove__free__synced
,
549 uint64_t, synced_size
);
552 offset
+= synced_size
;
556 * Look at all in-flight txgs starting from the currently syncing one
557 * and see if a section of this free is being copied. By starting from
558 * this txg and iterating forward, we might find that this region
559 * was copied in two different txgs and handle it appropriately.
561 for (int i
= 0; i
< TXG_CONCURRENT_STATES
; i
++) {
562 int txgoff
= (txg
+ i
) & TXG_MASK
;
563 if (size
> 0 && offset
< svr
->svr_max_offset_to_sync
[txgoff
]) {
565 * The mapping for this offset is in flight, and
566 * will be synced in txg+i.
568 uint64_t inflight_size
= MIN(size
,
569 svr
->svr_max_offset_to_sync
[txgoff
] - offset
);
571 DTRACE_PROBE4(remove__free__inflight
,
574 uint64_t, inflight_size
,
578 * We copy data in order of increasing offset.
579 * Therefore the max_offset_to_sync[] must increase
580 * (or be zero, indicating that nothing is being
581 * copied in that txg).
583 if (svr
->svr_max_offset_to_sync
[txgoff
] != 0) {
584 ASSERT3U(svr
->svr_max_offset_to_sync
[txgoff
],
587 svr
->svr_max_offset_to_sync
[txgoff
];
591 * We've already committed to copying this segment:
592 * we have allocated space elsewhere in the pool for
593 * it and have an IO outstanding to copy the data. We
594 * cannot free the space before the copy has
595 * completed, or else the copy IO might overwrite any
596 * new data. To free that space, we record the
597 * segment in the appropriate svr_frees tree and free
598 * the mapped space later, in the txg where we have
599 * completed the copy and synced the mapping (see
600 * vdev_mapping_sync).
602 range_tree_add(svr
->svr_frees
[txgoff
],
603 offset
, inflight_size
);
604 size
-= inflight_size
;
605 offset
+= inflight_size
;
608 * This space is already accounted for as being
609 * done, because it is being copied in txg+i.
610 * However, if i!=0, then it is being copied in
611 * a future txg. If we crash after this txg
612 * syncs but before txg+i syncs, then the space
613 * will be free. Therefore we must account
614 * for the space being done in *this* txg
615 * (when it is freed) rather than the future txg
616 * (when it will be copied).
618 ASSERT3U(svr
->svr_bytes_done
[txgoff
], >=,
620 svr
->svr_bytes_done
[txgoff
] -= inflight_size
;
621 svr
->svr_bytes_done
[txg
& TXG_MASK
] += inflight_size
;
624 ASSERT0(svr
->svr_max_offset_to_sync
[TXG_CLEAN(txg
) & TXG_MASK
]);
628 * The copy thread has not yet visited this offset. Ensure
632 DTRACE_PROBE3(remove__free__unvisited
,
637 if (svr
->svr_allocd_segs
!= NULL
)
638 range_tree_clear(svr
->svr_allocd_segs
, offset
, size
);
641 * Since we now do not need to copy this data, for
642 * accounting purposes we have done our job and can count
645 svr
->svr_bytes_done
[txg
& TXG_MASK
] += size
;
647 mutex_exit(&svr
->svr_lock
);
650 * Now that we have dropped svr_lock, process the synced portion
653 if (synced_size
> 0) {
654 vdev_indirect_mark_obsolete(vd
, synced_offset
, synced_size
);
657 * Note: this can only be called from syncing context,
658 * and the vdev_indirect_mapping is only changed from the
659 * sync thread, so we don't need svr_lock while doing
660 * metaslab_free_impl_cb.
662 boolean_t checkpoint
= B_FALSE
;
663 vdev_indirect_ops
.vdev_op_remap(vd
, synced_offset
, synced_size
,
664 metaslab_free_impl_cb
, &checkpoint
);
669 * Stop an active removal and update the spa_removing phys.
672 spa_finish_removal(spa_t
*spa
, dsl_scan_state_t state
, dmu_tx_t
*tx
)
674 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
675 ASSERT3U(dmu_tx_get_txg(tx
), ==, spa_syncing_txg(spa
));
677 /* Ensure the removal thread has completed before we free the svr. */
678 spa_vdev_remove_suspend(spa
);
680 ASSERT(state
== DSS_FINISHED
|| state
== DSS_CANCELED
);
682 if (state
== DSS_FINISHED
) {
683 spa_removing_phys_t
*srp
= &spa
->spa_removing_phys
;
684 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
685 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
687 if (srp
->sr_prev_indirect_vdev
!= -1) {
689 pvd
= vdev_lookup_top(spa
,
690 srp
->sr_prev_indirect_vdev
);
691 ASSERT3P(pvd
->vdev_ops
, ==, &vdev_indirect_ops
);
694 vic
->vic_prev_indirect_vdev
= srp
->sr_prev_indirect_vdev
;
695 srp
->sr_prev_indirect_vdev
= vd
->vdev_id
;
697 spa
->spa_removing_phys
.sr_state
= state
;
698 spa
->spa_removing_phys
.sr_end_time
= gethrestime_sec();
700 spa
->spa_vdev_removal
= NULL
;
701 spa_vdev_removal_destroy(svr
);
703 spa_sync_removing_state(spa
, tx
);
705 vdev_config_dirty(spa
->spa_root_vdev
);
709 free_mapped_segment_cb(void *arg
, uint64_t offset
, uint64_t size
)
712 vdev_indirect_mark_obsolete(vd
, offset
, size
);
713 boolean_t checkpoint
= B_FALSE
;
714 vdev_indirect_ops
.vdev_op_remap(vd
, offset
, size
,
715 metaslab_free_impl_cb
, &checkpoint
);
719 * On behalf of the removal thread, syncs an incremental bit more of
720 * the indirect mapping to disk and updates the in-memory mapping.
721 * Called as a sync task in every txg that the removal thread makes progress.
724 vdev_mapping_sync(void *arg
, dmu_tx_t
*tx
)
726 spa_vdev_removal_t
*svr
= arg
;
727 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
728 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
729 ASSERTV(vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
);
730 uint64_t txg
= dmu_tx_get_txg(tx
);
731 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
733 ASSERT(vic
->vic_mapping_object
!= 0);
734 ASSERT3U(txg
, ==, spa_syncing_txg(spa
));
736 vdev_indirect_mapping_add_entries(vim
,
737 &svr
->svr_new_segments
[txg
& TXG_MASK
], tx
);
738 vdev_indirect_births_add_entry(vd
->vdev_indirect_births
,
739 vdev_indirect_mapping_max_offset(vim
), dmu_tx_get_txg(tx
), tx
);
742 * Free the copied data for anything that was freed while the
743 * mapping entries were in flight.
745 mutex_enter(&svr
->svr_lock
);
746 range_tree_vacate(svr
->svr_frees
[txg
& TXG_MASK
],
747 free_mapped_segment_cb
, vd
);
748 ASSERT3U(svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
], >=,
749 vdev_indirect_mapping_max_offset(vim
));
750 svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] = 0;
751 mutex_exit(&svr
->svr_lock
);
753 spa_sync_removing_state(spa
, tx
);
756 typedef struct vdev_copy_segment_arg
{
758 dva_t
*vcsa_dest_dva
;
760 range_tree_t
*vcsa_obsolete_segs
;
761 } vdev_copy_segment_arg_t
;
764 unalloc_seg(void *arg
, uint64_t start
, uint64_t size
)
766 vdev_copy_segment_arg_t
*vcsa
= arg
;
767 spa_t
*spa
= vcsa
->vcsa_spa
;
768 blkptr_t bp
= { { { {0} } } };
770 BP_SET_BIRTH(&bp
, TXG_INITIAL
, TXG_INITIAL
);
771 BP_SET_LSIZE(&bp
, size
);
772 BP_SET_PSIZE(&bp
, size
);
773 BP_SET_COMPRESS(&bp
, ZIO_COMPRESS_OFF
);
774 BP_SET_CHECKSUM(&bp
, ZIO_CHECKSUM_OFF
);
775 BP_SET_TYPE(&bp
, DMU_OT_NONE
);
776 BP_SET_LEVEL(&bp
, 0);
777 BP_SET_DEDUP(&bp
, 0);
778 BP_SET_BYTEORDER(&bp
, ZFS_HOST_BYTEORDER
);
780 DVA_SET_VDEV(&bp
.blk_dva
[0], DVA_GET_VDEV(vcsa
->vcsa_dest_dva
));
781 DVA_SET_OFFSET(&bp
.blk_dva
[0],
782 DVA_GET_OFFSET(vcsa
->vcsa_dest_dva
) + start
);
783 DVA_SET_ASIZE(&bp
.blk_dva
[0], size
);
785 zio_free(spa
, vcsa
->vcsa_txg
, &bp
);
789 * All reads and writes associated with a call to spa_vdev_copy_segment()
793 spa_vdev_copy_segment_done(zio_t
*zio
)
795 vdev_copy_segment_arg_t
*vcsa
= zio
->io_private
;
797 range_tree_vacate(vcsa
->vcsa_obsolete_segs
,
799 range_tree_destroy(vcsa
->vcsa_obsolete_segs
);
800 kmem_free(vcsa
, sizeof (*vcsa
));
802 spa_config_exit(zio
->io_spa
, SCL_STATE
, zio
->io_spa
);
806 * The write of the new location is done.
809 spa_vdev_copy_segment_write_done(zio_t
*zio
)
811 vdev_copy_arg_t
*vca
= zio
->io_private
;
813 abd_free(zio
->io_abd
);
815 mutex_enter(&vca
->vca_lock
);
816 vca
->vca_outstanding_bytes
-= zio
->io_size
;
818 if (zio
->io_error
!= 0)
819 vca
->vca_write_error_bytes
+= zio
->io_size
;
821 cv_signal(&vca
->vca_cv
);
822 mutex_exit(&vca
->vca_lock
);
826 * The read of the old location is done. The parent zio is the write to
827 * the new location. Allow it to start.
830 spa_vdev_copy_segment_read_done(zio_t
*zio
)
832 vdev_copy_arg_t
*vca
= zio
->io_private
;
834 if (zio
->io_error
!= 0) {
835 mutex_enter(&vca
->vca_lock
);
836 vca
->vca_read_error_bytes
+= zio
->io_size
;
837 mutex_exit(&vca
->vca_lock
);
840 zio_nowait(zio_unique_parent(zio
));
844 * If the old and new vdevs are mirrors, we will read both sides of the old
845 * mirror, and write each copy to the corresponding side of the new mirror.
846 * If the old and new vdevs have a different number of children, we will do
847 * this as best as possible. Since we aren't verifying checksums, this
848 * ensures that as long as there's a good copy of the data, we'll have a
849 * good copy after the removal, even if there's silent damage to one side
850 * of the mirror. If we're removing a mirror that has some silent damage,
851 * we'll have exactly the same damage in the new location (assuming that
852 * the new location is also a mirror).
854 * We accomplish this by creating a tree of zio_t's, with as many writes as
855 * there are "children" of the new vdev (a non-redundant vdev counts as one
856 * child, a 2-way mirror has 2 children, etc). Each write has an associated
857 * read from a child of the old vdev. Typically there will be the same
858 * number of children of the old and new vdevs. However, if there are more
859 * children of the new vdev, some child(ren) of the old vdev will be issued
860 * multiple reads. If there are more children of the old vdev, some copies
863 * For example, the tree of zio_t's for a 2-way mirror is:
867 * write(new vdev, child 0) write(new vdev, child 1)
869 * read(old vdev, child 0) read(old vdev, child 1)
871 * Child zio's complete before their parents complete. However, zio's
872 * created with zio_vdev_child_io() may be issued before their children
873 * complete. In this case we need to make sure that the children (reads)
874 * complete before the parents (writes) are *issued*. We do this by not
875 * calling zio_nowait() on each write until its corresponding read has
878 * The spa_config_lock must be held while zio's created by
879 * zio_vdev_child_io() are in progress, to ensure that the vdev tree does
880 * not change (e.g. due to a concurrent "zpool attach/detach"). The "null"
881 * zio is needed to release the spa_config_lock after all the reads and
882 * writes complete. (Note that we can't grab the config lock for each read,
883 * because it is not reentrant - we could deadlock with a thread waiting
887 spa_vdev_copy_one_child(vdev_copy_arg_t
*vca
, zio_t
*nzio
,
888 vdev_t
*source_vd
, uint64_t source_offset
,
889 vdev_t
*dest_child_vd
, uint64_t dest_offset
, int dest_id
, uint64_t size
)
891 ASSERT3U(spa_config_held(nzio
->io_spa
, SCL_ALL
, RW_READER
), !=, 0);
894 * If the destination child in unwritable then there is no point
895 * in issuing the source reads which cannot be written.
897 if (!vdev_writeable(dest_child_vd
))
900 mutex_enter(&vca
->vca_lock
);
901 vca
->vca_outstanding_bytes
+= size
;
902 mutex_exit(&vca
->vca_lock
);
904 abd_t
*abd
= abd_alloc_for_io(size
, B_FALSE
);
906 vdev_t
*source_child_vd
= NULL
;
907 if (source_vd
->vdev_ops
== &vdev_mirror_ops
&& dest_id
!= -1) {
909 * Source and dest are both mirrors. Copy from the same
910 * child id as we are copying to (wrapping around if there
911 * are more dest children than source children). If the
912 * preferred source child is unreadable select another.
914 for (int i
= 0; i
< source_vd
->vdev_children
; i
++) {
915 source_child_vd
= source_vd
->vdev_child
[
916 (dest_id
+ i
) % source_vd
->vdev_children
];
917 if (vdev_readable(source_child_vd
))
921 source_child_vd
= source_vd
;
925 * There should always be at least one readable source child or
926 * the pool would be in a suspended state. Somehow selecting an
927 * unreadable child would result in IO errors, the removal process
928 * being cancelled, and the pool reverting to its pre-removal state.
930 ASSERT3P(source_child_vd
, !=, NULL
);
932 zio_t
*write_zio
= zio_vdev_child_io(nzio
, NULL
,
933 dest_child_vd
, dest_offset
, abd
, size
,
934 ZIO_TYPE_WRITE
, ZIO_PRIORITY_REMOVAL
,
936 spa_vdev_copy_segment_write_done
, vca
);
938 zio_nowait(zio_vdev_child_io(write_zio
, NULL
,
939 source_child_vd
, source_offset
, abd
, size
,
940 ZIO_TYPE_READ
, ZIO_PRIORITY_REMOVAL
,
942 spa_vdev_copy_segment_read_done
, vca
));
946 * Allocate a new location for this segment, and create the zio_t's to
947 * read from the old location and write to the new location.
950 spa_vdev_copy_segment(vdev_t
*vd
, range_tree_t
*segs
,
951 uint64_t maxalloc
, uint64_t txg
,
952 vdev_copy_arg_t
*vca
, zio_alloc_list_t
*zal
)
954 metaslab_group_t
*mg
= vd
->vdev_mg
;
955 spa_t
*spa
= vd
->vdev_spa
;
956 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
957 vdev_indirect_mapping_entry_t
*entry
;
959 uint64_t start
= range_tree_min(segs
);
961 ASSERT3U(maxalloc
, <=, SPA_MAXBLOCKSIZE
);
963 uint64_t size
= range_tree_span(segs
);
964 if (range_tree_span(segs
) > maxalloc
) {
966 * We can't allocate all the segments. Prefer to end
967 * the allocation at the end of a segment, thus avoiding
968 * additional split blocks.
972 search
.rs_start
= start
+ maxalloc
;
973 search
.rs_end
= search
.rs_start
;
974 range_seg_t
*rs
= avl_find(&segs
->rt_root
, &search
, &where
);
976 rs
= avl_nearest(&segs
->rt_root
, where
, AVL_BEFORE
);
978 rs
= AVL_PREV(&segs
->rt_root
, rs
);
981 size
= rs
->rs_end
- start
;
984 * There are no segments that end before maxalloc.
985 * I.e. the first segment is larger than maxalloc,
986 * so we must split it.
991 ASSERT3U(size
, <=, maxalloc
);
994 * An allocation class might not have any remaining vdevs or space
996 metaslab_class_t
*mc
= mg
->mg_class
;
997 if (mc
!= spa_normal_class(spa
) && mc
->mc_groups
<= 1)
998 mc
= spa_normal_class(spa
);
999 int error
= metaslab_alloc_dva(spa
, mc
, size
, &dst
, 0, NULL
, txg
, 0,
1001 if (error
== ENOSPC
&& mc
!= spa_normal_class(spa
)) {
1002 error
= metaslab_alloc_dva(spa
, spa_normal_class(spa
), size
,
1003 &dst
, 0, NULL
, txg
, 0, zal
, 0);
1009 * Determine the ranges that are not actually needed. Offsets are
1010 * relative to the start of the range to be copied (i.e. relative to the
1011 * local variable "start").
1013 range_tree_t
*obsolete_segs
= range_tree_create(NULL
, NULL
);
1015 range_seg_t
*rs
= avl_first(&segs
->rt_root
);
1016 ASSERT3U(rs
->rs_start
, ==, start
);
1017 uint64_t prev_seg_end
= rs
->rs_end
;
1018 while ((rs
= AVL_NEXT(&segs
->rt_root
, rs
)) != NULL
) {
1019 if (rs
->rs_start
>= start
+ size
) {
1022 range_tree_add(obsolete_segs
,
1023 prev_seg_end
- start
,
1024 rs
->rs_start
- prev_seg_end
);
1026 prev_seg_end
= rs
->rs_end
;
1028 /* We don't end in the middle of an obsolete range */
1029 ASSERT3U(start
+ size
, <=, prev_seg_end
);
1031 range_tree_clear(segs
, start
, size
);
1034 * We can't have any padding of the allocated size, otherwise we will
1035 * misunderstand what's allocated, and the size of the mapping.
1036 * The caller ensures this will be true by passing in a size that is
1037 * aligned to the worst (highest) ashift in the pool.
1039 ASSERT3U(DVA_GET_ASIZE(&dst
), ==, size
);
1041 entry
= kmem_zalloc(sizeof (vdev_indirect_mapping_entry_t
), KM_SLEEP
);
1042 DVA_MAPPING_SET_SRC_OFFSET(&entry
->vime_mapping
, start
);
1043 entry
->vime_mapping
.vimep_dst
= dst
;
1044 if (spa_feature_is_enabled(spa
, SPA_FEATURE_OBSOLETE_COUNTS
)) {
1045 entry
->vime_obsolete_count
= range_tree_space(obsolete_segs
);
1048 vdev_copy_segment_arg_t
*vcsa
= kmem_zalloc(sizeof (*vcsa
), KM_SLEEP
);
1049 vcsa
->vcsa_dest_dva
= &entry
->vime_mapping
.vimep_dst
;
1050 vcsa
->vcsa_obsolete_segs
= obsolete_segs
;
1051 vcsa
->vcsa_spa
= spa
;
1052 vcsa
->vcsa_txg
= txg
;
1055 * See comment before spa_vdev_copy_one_child().
1057 spa_config_enter(spa
, SCL_STATE
, spa
, RW_READER
);
1058 zio_t
*nzio
= zio_null(spa
->spa_txg_zio
[txg
& TXG_MASK
], spa
, NULL
,
1059 spa_vdev_copy_segment_done
, vcsa
, 0);
1060 vdev_t
*dest_vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dst
));
1061 if (dest_vd
->vdev_ops
== &vdev_mirror_ops
) {
1062 for (int i
= 0; i
< dest_vd
->vdev_children
; i
++) {
1063 vdev_t
*child
= dest_vd
->vdev_child
[i
];
1064 spa_vdev_copy_one_child(vca
, nzio
, vd
, start
,
1065 child
, DVA_GET_OFFSET(&dst
), i
, size
);
1068 spa_vdev_copy_one_child(vca
, nzio
, vd
, start
,
1069 dest_vd
, DVA_GET_OFFSET(&dst
), -1, size
);
1073 list_insert_tail(&svr
->svr_new_segments
[txg
& TXG_MASK
], entry
);
1074 ASSERT3U(start
+ size
, <=, vd
->vdev_ms_count
<< vd
->vdev_ms_shift
);
1075 vdev_dirty(vd
, 0, NULL
, txg
);
1081 * Complete the removal of a toplevel vdev. This is called as a
1082 * synctask in the same txg that we will sync out the new config (to the
1083 * MOS object) which indicates that this vdev is indirect.
1086 vdev_remove_complete_sync(void *arg
, dmu_tx_t
*tx
)
1088 spa_vdev_removal_t
*svr
= arg
;
1089 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1090 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1092 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
1094 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1095 ASSERT0(svr
->svr_bytes_done
[i
]);
1098 ASSERT3U(spa
->spa_removing_phys
.sr_copied
, ==,
1099 spa
->spa_removing_phys
.sr_to_copy
);
1101 vdev_destroy_spacemaps(vd
, tx
);
1103 /* destroy leaf zaps, if any */
1104 ASSERT3P(svr
->svr_zaplist
, !=, NULL
);
1105 for (nvpair_t
*pair
= nvlist_next_nvpair(svr
->svr_zaplist
, NULL
);
1107 pair
= nvlist_next_nvpair(svr
->svr_zaplist
, pair
)) {
1108 vdev_destroy_unlink_zap(vd
, fnvpair_value_uint64(pair
), tx
);
1110 fnvlist_free(svr
->svr_zaplist
);
1112 spa_finish_removal(dmu_tx_pool(tx
)->dp_spa
, DSS_FINISHED
, tx
);
1113 /* vd->vdev_path is not available here */
1114 spa_history_log_internal(spa
, "vdev remove completed", tx
,
1115 "%s vdev %llu", spa_name(spa
), vd
->vdev_id
);
1119 vdev_remove_enlist_zaps(vdev_t
*vd
, nvlist_t
*zlist
)
1121 ASSERT3P(zlist
, !=, NULL
);
1122 ASSERT3P(vd
->vdev_ops
, !=, &vdev_raidz_ops
);
1124 if (vd
->vdev_leaf_zap
!= 0) {
1126 (void) snprintf(zkey
, sizeof (zkey
), "%s-%llu",
1127 VDEV_REMOVAL_ZAP_OBJS
, (u_longlong_t
)vd
->vdev_leaf_zap
);
1128 fnvlist_add_uint64(zlist
, zkey
, vd
->vdev_leaf_zap
);
1131 for (uint64_t id
= 0; id
< vd
->vdev_children
; id
++) {
1132 vdev_remove_enlist_zaps(vd
->vdev_child
[id
], zlist
);
1137 vdev_remove_replace_with_indirect(vdev_t
*vd
, uint64_t txg
)
1141 spa_t
*spa
= vd
->vdev_spa
;
1142 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1145 * First, build a list of leaf zaps to be destroyed.
1146 * This is passed to the sync context thread,
1147 * which does the actual unlinking.
1149 svr
->svr_zaplist
= fnvlist_alloc();
1150 vdev_remove_enlist_zaps(vd
, svr
->svr_zaplist
);
1152 ivd
= vdev_add_parent(vd
, &vdev_indirect_ops
);
1153 ivd
->vdev_removing
= 0;
1155 vd
->vdev_leaf_zap
= 0;
1157 vdev_remove_child(ivd
, vd
);
1158 vdev_compact_children(ivd
);
1160 ASSERT(!list_link_active(&vd
->vdev_state_dirty_node
));
1162 mutex_enter(&svr
->svr_lock
);
1163 svr
->svr_thread
= NULL
;
1164 cv_broadcast(&svr
->svr_cv
);
1165 mutex_exit(&svr
->svr_lock
);
1167 /* After this, we can not use svr. */
1168 tx
= dmu_tx_create_assigned(spa
->spa_dsl_pool
, txg
);
1169 dsl_sync_task_nowait(spa
->spa_dsl_pool
, vdev_remove_complete_sync
, svr
,
1170 0, ZFS_SPACE_CHECK_NONE
, tx
);
1175 * Complete the removal of a toplevel vdev. This is called in open
1176 * context by the removal thread after we have copied all vdev's data.
1179 vdev_remove_complete(spa_t
*spa
)
1184 * Wait for any deferred frees to be synced before we call
1185 * vdev_metaslab_fini()
1187 txg_wait_synced(spa
->spa_dsl_pool
, 0);
1188 txg
= spa_vdev_enter(spa
);
1189 vdev_t
*vd
= vdev_lookup_top(spa
, spa
->spa_vdev_removal
->svr_vdev_id
);
1190 ASSERT3P(vd
->vdev_initialize_thread
, ==, NULL
);
1192 sysevent_t
*ev
= spa_event_create(spa
, vd
, NULL
,
1193 ESC_ZFS_VDEV_REMOVE_DEV
);
1195 zfs_dbgmsg("finishing device removal for vdev %llu in txg %llu",
1199 * Discard allocation state.
1201 if (vd
->vdev_mg
!= NULL
) {
1202 vdev_metaslab_fini(vd
);
1203 metaslab_group_destroy(vd
->vdev_mg
);
1206 ASSERT0(vd
->vdev_stat
.vs_space
);
1207 ASSERT0(vd
->vdev_stat
.vs_dspace
);
1209 vdev_remove_replace_with_indirect(vd
, txg
);
1212 * We now release the locks, allowing spa_sync to run and finish the
1213 * removal via vdev_remove_complete_sync in syncing context.
1215 * Note that we hold on to the vdev_t that has been replaced. Since
1216 * it isn't part of the vdev tree any longer, it can't be concurrently
1217 * manipulated, even while we don't have the config lock.
1219 (void) spa_vdev_exit(spa
, NULL
, txg
, 0);
1222 * Top ZAP should have been transferred to the indirect vdev in
1223 * vdev_remove_replace_with_indirect.
1225 ASSERT0(vd
->vdev_top_zap
);
1228 * Leaf ZAP should have been moved in vdev_remove_replace_with_indirect.
1230 ASSERT0(vd
->vdev_leaf_zap
);
1232 txg
= spa_vdev_enter(spa
);
1233 (void) vdev_label_init(vd
, 0, VDEV_LABEL_REMOVE
);
1235 * Request to update the config and the config cachefile.
1237 vdev_config_dirty(spa
->spa_root_vdev
);
1238 (void) spa_vdev_exit(spa
, vd
, txg
, 0);
1245 * Evacuates a segment of size at most max_alloc from the vdev
1246 * via repeated calls to spa_vdev_copy_segment. If an allocation
1247 * fails, the pool is probably too fragmented to handle such a
1248 * large size, so decrease max_alloc so that the caller will not try
1249 * this size again this txg.
1252 spa_vdev_copy_impl(vdev_t
*vd
, spa_vdev_removal_t
*svr
, vdev_copy_arg_t
*vca
,
1253 uint64_t *max_alloc
, dmu_tx_t
*tx
)
1255 uint64_t txg
= dmu_tx_get_txg(tx
);
1256 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1258 mutex_enter(&svr
->svr_lock
);
1261 * Determine how big of a chunk to copy. We can allocate up
1262 * to max_alloc bytes, and we can span up to vdev_removal_max_span
1263 * bytes of unallocated space at a time. "segs" will track the
1264 * allocated segments that we are copying. We may also be copying
1265 * free segments (of up to vdev_removal_max_span bytes).
1267 range_tree_t
*segs
= range_tree_create(NULL
, NULL
);
1269 range_seg_t
*rs
= range_tree_first(svr
->svr_allocd_segs
);
1274 uint64_t seg_length
;
1276 if (range_tree_is_empty(segs
)) {
1277 /* need to truncate the first seg based on max_alloc */
1279 MIN(rs
->rs_end
- rs
->rs_start
, *max_alloc
);
1281 if (rs
->rs_start
- range_tree_max(segs
) >
1282 vdev_removal_max_span
) {
1284 * Including this segment would cause us to
1285 * copy a larger unneeded chunk than is allowed.
1288 } else if (rs
->rs_end
- range_tree_min(segs
) >
1291 * This additional segment would extend past
1292 * max_alloc. Rather than splitting this
1293 * segment, leave it for the next mapping.
1297 seg_length
= rs
->rs_end
- rs
->rs_start
;
1301 range_tree_add(segs
, rs
->rs_start
, seg_length
);
1302 range_tree_remove(svr
->svr_allocd_segs
,
1303 rs
->rs_start
, seg_length
);
1306 if (range_tree_is_empty(segs
)) {
1307 mutex_exit(&svr
->svr_lock
);
1308 range_tree_destroy(segs
);
1312 if (svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] == 0) {
1313 dsl_sync_task_nowait(dmu_tx_pool(tx
), vdev_mapping_sync
,
1314 svr
, 0, ZFS_SPACE_CHECK_NONE
, tx
);
1317 svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] = range_tree_max(segs
);
1320 * Note: this is the amount of *allocated* space
1321 * that we are taking care of each txg.
1323 svr
->svr_bytes_done
[txg
& TXG_MASK
] += range_tree_space(segs
);
1325 mutex_exit(&svr
->svr_lock
);
1327 zio_alloc_list_t zal
;
1328 metaslab_trace_init(&zal
);
1329 uint64_t thismax
= SPA_MAXBLOCKSIZE
;
1330 while (!range_tree_is_empty(segs
)) {
1331 int error
= spa_vdev_copy_segment(vd
,
1332 segs
, thismax
, txg
, vca
, &zal
);
1334 if (error
== ENOSPC
) {
1336 * Cut our segment in half, and don't try this
1337 * segment size again this txg. Note that the
1338 * allocation size must be aligned to the highest
1339 * ashift in the pool, so that the allocation will
1340 * not be padded out to a multiple of the ashift,
1341 * which could cause us to think that this mapping
1342 * is larger than we intended.
1344 ASSERT3U(spa
->spa_max_ashift
, >=, SPA_MINBLOCKSHIFT
);
1345 ASSERT3U(spa
->spa_max_ashift
, ==, spa
->spa_min_ashift
);
1346 uint64_t attempted
=
1347 MIN(range_tree_span(segs
), thismax
);
1348 thismax
= P2ROUNDUP(attempted
/ 2,
1349 1 << spa
->spa_max_ashift
);
1351 * The minimum-size allocation can not fail.
1353 ASSERT3U(attempted
, >, 1 << spa
->spa_max_ashift
);
1354 *max_alloc
= attempted
- (1 << spa
->spa_max_ashift
);
1359 * We've performed an allocation, so reset the
1362 metaslab_trace_fini(&zal
);
1363 metaslab_trace_init(&zal
);
1366 metaslab_trace_fini(&zal
);
1367 range_tree_destroy(segs
);
1371 * The removal thread operates in open context. It iterates over all
1372 * allocated space in the vdev, by loading each metaslab's spacemap.
1373 * For each contiguous segment of allocated space (capping the segment
1374 * size at SPA_MAXBLOCKSIZE), we:
1375 * - Allocate space for it on another vdev.
1376 * - Create a new mapping from the old location to the new location
1377 * (as a record in svr_new_segments).
1378 * - Initiate a physical read zio to get the data off the removing disk.
1379 * - In the read zio's done callback, initiate a physical write zio to
1380 * write it to the new vdev.
1381 * Note that all of this will take effect when a particular TXG syncs.
1382 * The sync thread ensures that all the phys reads and writes for the syncing
1383 * TXG have completed (see spa_txg_zio) and writes the new mappings to disk
1384 * (see vdev_mapping_sync()).
1387 spa_vdev_remove_thread(void *arg
)
1390 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1391 vdev_copy_arg_t vca
;
1392 uint64_t max_alloc
= zfs_remove_max_segment
;
1393 uint64_t last_txg
= 0;
1395 spa_config_enter(spa
, SCL_CONFIG
, FTAG
, RW_READER
);
1396 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1397 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
1398 uint64_t start_offset
= vdev_indirect_mapping_max_offset(vim
);
1400 ASSERT3P(vd
->vdev_ops
, !=, &vdev_indirect_ops
);
1401 ASSERT(vdev_is_concrete(vd
));
1402 ASSERT(vd
->vdev_removing
);
1403 ASSERT(vd
->vdev_indirect_config
.vic_mapping_object
!= 0);
1404 ASSERT(vim
!= NULL
);
1406 mutex_init(&vca
.vca_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1407 cv_init(&vca
.vca_cv
, NULL
, CV_DEFAULT
, NULL
);
1408 vca
.vca_outstanding_bytes
= 0;
1409 vca
.vca_read_error_bytes
= 0;
1410 vca
.vca_write_error_bytes
= 0;
1412 mutex_enter(&svr
->svr_lock
);
1415 * Start from vim_max_offset so we pick up where we left off
1416 * if we are restarting the removal after opening the pool.
1419 for (msi
= start_offset
>> vd
->vdev_ms_shift
;
1420 msi
< vd
->vdev_ms_count
&& !svr
->svr_thread_exit
; msi
++) {
1421 metaslab_t
*msp
= vd
->vdev_ms
[msi
];
1422 ASSERT3U(msi
, <=, vd
->vdev_ms_count
);
1424 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1426 mutex_enter(&msp
->ms_sync_lock
);
1427 mutex_enter(&msp
->ms_lock
);
1430 * Assert nothing in flight -- ms_*tree is empty.
1432 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1433 ASSERT0(range_tree_space(msp
->ms_allocating
[i
]));
1437 * If the metaslab has ever been allocated from (ms_sm!=NULL),
1438 * read the allocated segments from the space map object
1439 * into svr_allocd_segs. Since we do this while holding
1440 * svr_lock and ms_sync_lock, concurrent frees (which
1441 * would have modified the space map) will wait for us
1442 * to finish loading the spacemap, and then take the
1443 * appropriate action (see free_from_removing_vdev()).
1445 if (msp
->ms_sm
!= NULL
) {
1446 space_map_t
*sm
= NULL
;
1449 * We have to open a new space map here, because
1450 * ms_sm's sm_length and sm_alloc may not reflect
1451 * what's in the object contents, if we are in between
1452 * metaslab_sync() and metaslab_sync_done().
1454 VERIFY0(space_map_open(&sm
,
1455 spa
->spa_dsl_pool
->dp_meta_objset
,
1456 msp
->ms_sm
->sm_object
, msp
->ms_sm
->sm_start
,
1457 msp
->ms_sm
->sm_size
, msp
->ms_sm
->sm_shift
));
1458 space_map_update(sm
);
1459 VERIFY0(space_map_load(sm
, svr
->svr_allocd_segs
,
1461 space_map_close(sm
);
1463 range_tree_walk(msp
->ms_freeing
,
1464 range_tree_remove
, svr
->svr_allocd_segs
);
1467 * When we are resuming from a paused removal (i.e.
1468 * when importing a pool with a removal in progress),
1469 * discard any state that we have already processed.
1471 range_tree_clear(svr
->svr_allocd_segs
, 0, start_offset
);
1473 mutex_exit(&msp
->ms_lock
);
1474 mutex_exit(&msp
->ms_sync_lock
);
1477 zfs_dbgmsg("copying %llu segments for metaslab %llu",
1478 avl_numnodes(&svr
->svr_allocd_segs
->rt_root
),
1481 while (!svr
->svr_thread_exit
&&
1482 !range_tree_is_empty(svr
->svr_allocd_segs
)) {
1484 mutex_exit(&svr
->svr_lock
);
1487 * We need to periodically drop the config lock so that
1488 * writers can get in. Additionally, we can't wait
1489 * for a txg to sync while holding a config lock
1490 * (since a waiting writer could cause a 3-way deadlock
1491 * with the sync thread, which also gets a config
1492 * lock for reader). So we can't hold the config lock
1493 * while calling dmu_tx_assign().
1495 spa_config_exit(spa
, SCL_CONFIG
, FTAG
);
1498 * This delay will pause the removal around the point
1499 * specified by zfs_removal_suspend_progress. We do this
1500 * solely from the test suite or during debugging.
1502 uint64_t bytes_copied
=
1503 spa
->spa_removing_phys
.sr_copied
;
1504 for (int i
= 0; i
< TXG_SIZE
; i
++)
1505 bytes_copied
+= svr
->svr_bytes_done
[i
];
1506 while (zfs_removal_suspend_progress
&&
1507 !svr
->svr_thread_exit
)
1510 mutex_enter(&vca
.vca_lock
);
1511 while (vca
.vca_outstanding_bytes
>
1512 zfs_remove_max_copy_bytes
) {
1513 cv_wait(&vca
.vca_cv
, &vca
.vca_lock
);
1515 mutex_exit(&vca
.vca_lock
);
1518 dmu_tx_create_dd(spa_get_dsl(spa
)->dp_mos_dir
);
1519 dmu_tx_hold_space(tx
, SPA_MAXBLOCKSIZE
);
1520 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
1521 uint64_t txg
= dmu_tx_get_txg(tx
);
1524 * Reacquire the vdev_config lock. The vdev_t
1525 * that we're removing may have changed, e.g. due
1526 * to a vdev_attach or vdev_detach.
1528 spa_config_enter(spa
, SCL_CONFIG
, FTAG
, RW_READER
);
1529 vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1531 if (txg
!= last_txg
)
1532 max_alloc
= zfs_remove_max_segment
;
1535 spa_vdev_copy_impl(vd
, svr
, &vca
, &max_alloc
, tx
);
1538 mutex_enter(&svr
->svr_lock
);
1541 mutex_enter(&vca
.vca_lock
);
1542 if (zfs_removal_ignore_errors
== 0 &&
1543 (vca
.vca_read_error_bytes
> 0 ||
1544 vca
.vca_write_error_bytes
> 0)) {
1545 svr
->svr_thread_exit
= B_TRUE
;
1547 mutex_exit(&vca
.vca_lock
);
1550 mutex_exit(&svr
->svr_lock
);
1552 spa_config_exit(spa
, SCL_CONFIG
, FTAG
);
1555 * Wait for all copies to finish before cleaning up the vca.
1557 txg_wait_synced(spa
->spa_dsl_pool
, 0);
1558 ASSERT0(vca
.vca_outstanding_bytes
);
1560 mutex_destroy(&vca
.vca_lock
);
1561 cv_destroy(&vca
.vca_cv
);
1563 if (svr
->svr_thread_exit
) {
1564 mutex_enter(&svr
->svr_lock
);
1565 range_tree_vacate(svr
->svr_allocd_segs
, NULL
, NULL
);
1566 svr
->svr_thread
= NULL
;
1567 cv_broadcast(&svr
->svr_cv
);
1568 mutex_exit(&svr
->svr_lock
);
1571 * During the removal process an unrecoverable read or write
1572 * error was encountered. The removal process must be
1573 * cancelled or this damage may become permanent.
1575 if (zfs_removal_ignore_errors
== 0 &&
1576 (vca
.vca_read_error_bytes
> 0 ||
1577 vca
.vca_write_error_bytes
> 0)) {
1578 zfs_dbgmsg("canceling removal due to IO errors: "
1579 "[read_error_bytes=%llu] [write_error_bytes=%llu]",
1580 vca
.vca_read_error_bytes
,
1581 vca
.vca_write_error_bytes
);
1582 spa_vdev_remove_cancel_impl(spa
);
1585 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1586 vdev_remove_complete(spa
);
1591 spa_vdev_remove_suspend(spa_t
*spa
)
1593 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1598 mutex_enter(&svr
->svr_lock
);
1599 svr
->svr_thread_exit
= B_TRUE
;
1600 while (svr
->svr_thread
!= NULL
)
1601 cv_wait(&svr
->svr_cv
, &svr
->svr_lock
);
1602 svr
->svr_thread_exit
= B_FALSE
;
1603 mutex_exit(&svr
->svr_lock
);
1608 spa_vdev_remove_cancel_check(void *arg
, dmu_tx_t
*tx
)
1610 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1612 if (spa
->spa_vdev_removal
== NULL
)
1613 return (ENOTACTIVE
);
1618 * Cancel a removal by freeing all entries from the partial mapping
1619 * and marking the vdev as no longer being removing.
1623 spa_vdev_remove_cancel_sync(void *arg
, dmu_tx_t
*tx
)
1625 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1626 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1627 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1628 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
1629 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
1630 objset_t
*mos
= spa
->spa_meta_objset
;
1632 ASSERT3P(svr
->svr_thread
, ==, NULL
);
1634 spa_feature_decr(spa
, SPA_FEATURE_DEVICE_REMOVAL
, tx
);
1636 boolean_t are_precise
;
1637 VERIFY0(vdev_obsolete_counts_are_precise(vd
, &are_precise
));
1639 spa_feature_decr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
1640 VERIFY0(zap_remove(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
1641 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE
, tx
));
1644 uint64_t obsolete_sm_object
;
1645 VERIFY0(vdev_obsolete_sm_object(vd
, &obsolete_sm_object
));
1646 if (obsolete_sm_object
!= 0) {
1647 ASSERT(vd
->vdev_obsolete_sm
!= NULL
);
1648 ASSERT3U(obsolete_sm_object
, ==,
1649 space_map_object(vd
->vdev_obsolete_sm
));
1651 space_map_free(vd
->vdev_obsolete_sm
, tx
);
1652 VERIFY0(zap_remove(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
1653 VDEV_TOP_ZAP_INDIRECT_OBSOLETE_SM
, tx
));
1654 space_map_close(vd
->vdev_obsolete_sm
);
1655 vd
->vdev_obsolete_sm
= NULL
;
1656 spa_feature_decr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
1658 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1659 ASSERT(list_is_empty(&svr
->svr_new_segments
[i
]));
1660 ASSERT3U(svr
->svr_max_offset_to_sync
[i
], <=,
1661 vdev_indirect_mapping_max_offset(vim
));
1664 for (uint64_t msi
= 0; msi
< vd
->vdev_ms_count
; msi
++) {
1665 metaslab_t
*msp
= vd
->vdev_ms
[msi
];
1667 if (msp
->ms_start
>= vdev_indirect_mapping_max_offset(vim
))
1670 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1672 mutex_enter(&msp
->ms_lock
);
1675 * Assert nothing in flight -- ms_*tree is empty.
1677 for (int i
= 0; i
< TXG_SIZE
; i
++)
1678 ASSERT0(range_tree_space(msp
->ms_allocating
[i
]));
1679 for (int i
= 0; i
< TXG_DEFER_SIZE
; i
++)
1680 ASSERT0(range_tree_space(msp
->ms_defer
[i
]));
1681 ASSERT0(range_tree_space(msp
->ms_freed
));
1683 if (msp
->ms_sm
!= NULL
) {
1685 * Assert that the in-core spacemap has the same
1686 * length as the on-disk one, so we can use the
1687 * existing in-core spacemap to load it from disk.
1689 ASSERT3U(msp
->ms_sm
->sm_alloc
, ==,
1690 msp
->ms_sm
->sm_phys
->smp_alloc
);
1691 ASSERT3U(msp
->ms_sm
->sm_length
, ==,
1692 msp
->ms_sm
->sm_phys
->smp_objsize
);
1694 mutex_enter(&svr
->svr_lock
);
1695 VERIFY0(space_map_load(msp
->ms_sm
,
1696 svr
->svr_allocd_segs
, SM_ALLOC
));
1697 range_tree_walk(msp
->ms_freeing
,
1698 range_tree_remove
, svr
->svr_allocd_segs
);
1701 * Clear everything past what has been synced,
1702 * because we have not allocated mappings for it yet.
1704 uint64_t syncd
= vdev_indirect_mapping_max_offset(vim
);
1705 uint64_t sm_end
= msp
->ms_sm
->sm_start
+
1706 msp
->ms_sm
->sm_size
;
1708 range_tree_clear(svr
->svr_allocd_segs
,
1709 syncd
, sm_end
- syncd
);
1711 mutex_exit(&svr
->svr_lock
);
1713 mutex_exit(&msp
->ms_lock
);
1715 mutex_enter(&svr
->svr_lock
);
1716 range_tree_vacate(svr
->svr_allocd_segs
,
1717 free_mapped_segment_cb
, vd
);
1718 mutex_exit(&svr
->svr_lock
);
1722 * Note: this must happen after we invoke free_mapped_segment_cb,
1723 * because it adds to the obsolete_segments.
1725 range_tree_vacate(vd
->vdev_obsolete_segments
, NULL
, NULL
);
1727 ASSERT3U(vic
->vic_mapping_object
, ==,
1728 vdev_indirect_mapping_object(vd
->vdev_indirect_mapping
));
1729 vdev_indirect_mapping_close(vd
->vdev_indirect_mapping
);
1730 vd
->vdev_indirect_mapping
= NULL
;
1731 vdev_indirect_mapping_free(mos
, vic
->vic_mapping_object
, tx
);
1732 vic
->vic_mapping_object
= 0;
1734 ASSERT3U(vic
->vic_births_object
, ==,
1735 vdev_indirect_births_object(vd
->vdev_indirect_births
));
1736 vdev_indirect_births_close(vd
->vdev_indirect_births
);
1737 vd
->vdev_indirect_births
= NULL
;
1738 vdev_indirect_births_free(mos
, vic
->vic_births_object
, tx
);
1739 vic
->vic_births_object
= 0;
1742 * We may have processed some frees from the removing vdev in this
1743 * txg, thus increasing svr_bytes_done; discard that here to
1744 * satisfy the assertions in spa_vdev_removal_destroy().
1745 * Note that future txg's can not have any bytes_done, because
1746 * future TXG's are only modified from open context, and we have
1747 * already shut down the copying thread.
1749 svr
->svr_bytes_done
[dmu_tx_get_txg(tx
) & TXG_MASK
] = 0;
1750 spa_finish_removal(spa
, DSS_CANCELED
, tx
);
1752 vd
->vdev_removing
= B_FALSE
;
1753 vdev_config_dirty(vd
);
1755 zfs_dbgmsg("canceled device removal for vdev %llu in %llu",
1756 vd
->vdev_id
, dmu_tx_get_txg(tx
));
1757 spa_history_log_internal(spa
, "vdev remove canceled", tx
,
1758 "%s vdev %llu %s", spa_name(spa
),
1759 vd
->vdev_id
, (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
1763 spa_vdev_remove_cancel_impl(spa_t
*spa
)
1765 uint64_t vdid
= spa
->spa_vdev_removal
->svr_vdev_id
;
1767 int error
= dsl_sync_task(spa
->spa_name
, spa_vdev_remove_cancel_check
,
1768 spa_vdev_remove_cancel_sync
, NULL
, 0,
1769 ZFS_SPACE_CHECK_EXTRA_RESERVED
);
1772 spa_config_enter(spa
, SCL_ALLOC
| SCL_VDEV
, FTAG
, RW_WRITER
);
1773 vdev_t
*vd
= vdev_lookup_top(spa
, vdid
);
1774 metaslab_group_activate(vd
->vdev_mg
);
1775 spa_config_exit(spa
, SCL_ALLOC
| SCL_VDEV
, FTAG
);
1782 spa_vdev_remove_cancel(spa_t
*spa
)
1784 spa_vdev_remove_suspend(spa
);
1786 if (spa
->spa_vdev_removal
== NULL
)
1787 return (ENOTACTIVE
);
1789 return (spa_vdev_remove_cancel_impl(spa
));
1793 * Called every sync pass of every txg if there's a svr.
1796 svr_sync(spa_t
*spa
, dmu_tx_t
*tx
)
1798 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1799 int txgoff
= dmu_tx_get_txg(tx
) & TXG_MASK
;
1805 * This check is necessary so that we do not dirty the
1806 * DIRECTORY_OBJECT via spa_sync_removing_state() when there
1807 * is nothing to do. Dirtying it every time would prevent us
1808 * from syncing-to-convergence.
1810 if (svr
->svr_bytes_done
[txgoff
] == 0)
1814 * Update progress accounting.
1816 spa
->spa_removing_phys
.sr_copied
+= svr
->svr_bytes_done
[txgoff
];
1817 svr
->svr_bytes_done
[txgoff
] = 0;
1819 spa_sync_removing_state(spa
, tx
);
1823 vdev_remove_make_hole_and_free(vdev_t
*vd
)
1825 uint64_t id
= vd
->vdev_id
;
1826 spa_t
*spa
= vd
->vdev_spa
;
1827 vdev_t
*rvd
= spa
->spa_root_vdev
;
1828 boolean_t last_vdev
= (id
== (rvd
->vdev_children
- 1));
1830 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
1831 ASSERT(spa_config_held(spa
, SCL_ALL
, RW_WRITER
) == SCL_ALL
);
1836 vdev_compact_children(rvd
);
1838 vd
= vdev_alloc_common(spa
, id
, 0, &vdev_hole_ops
);
1839 vdev_add_child(rvd
, vd
);
1841 vdev_config_dirty(rvd
);
1844 * Reassess the health of our root vdev.
1850 * Remove a log device. The config lock is held for the specified TXG.
1853 spa_vdev_remove_log(vdev_t
*vd
, uint64_t *txg
)
1855 metaslab_group_t
*mg
= vd
->vdev_mg
;
1856 spa_t
*spa
= vd
->vdev_spa
;
1859 ASSERT(vd
->vdev_islog
);
1860 ASSERT(vd
== vd
->vdev_top
);
1863 * Stop allocating from this vdev.
1865 metaslab_group_passivate(mg
);
1868 * Wait for the youngest allocations and frees to sync,
1869 * and then wait for the deferral of those frees to finish.
1871 spa_vdev_config_exit(spa
, NULL
,
1872 *txg
+ TXG_CONCURRENT_STATES
+ TXG_DEFER_SIZE
, 0, FTAG
);
1875 * Evacuate the device. We don't hold the config lock as writer
1876 * since we need to do I/O but we do keep the
1877 * spa_namespace_lock held. Once this completes the device
1878 * should no longer have any blocks allocated on it.
1880 if (vd
->vdev_islog
) {
1881 if (vd
->vdev_stat
.vs_alloc
!= 0)
1882 error
= spa_reset_logs(spa
);
1885 *txg
= spa_vdev_config_enter(spa
);
1888 metaslab_group_activate(mg
);
1891 ASSERT0(vd
->vdev_stat
.vs_alloc
);
1894 * The evacuation succeeded. Remove any remaining MOS metadata
1895 * associated with this vdev, and wait for these changes to sync.
1897 vd
->vdev_removing
= B_TRUE
;
1899 vdev_dirty_leaves(vd
, VDD_DTL
, *txg
);
1900 vdev_config_dirty(vd
);
1902 spa_vdev_config_exit(spa
, NULL
, *txg
, 0, FTAG
);
1904 /* Stop initializing */
1905 vdev_initialize_stop_all(vd
, VDEV_INITIALIZE_CANCELED
);
1907 *txg
= spa_vdev_config_enter(spa
);
1909 sysevent_t
*ev
= spa_event_create(spa
, vd
, NULL
,
1910 ESC_ZFS_VDEV_REMOVE_DEV
);
1911 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
1912 ASSERT(spa_config_held(spa
, SCL_ALL
, RW_WRITER
) == SCL_ALL
);
1914 /* The top ZAP should have been destroyed by vdev_remove_empty. */
1915 ASSERT0(vd
->vdev_top_zap
);
1916 /* The leaf ZAP should have been destroyed by vdev_dtl_sync. */
1917 ASSERT0(vd
->vdev_leaf_zap
);
1919 (void) vdev_label_init(vd
, 0, VDEV_LABEL_REMOVE
);
1921 if (list_link_active(&vd
->vdev_state_dirty_node
))
1922 vdev_state_clean(vd
);
1923 if (list_link_active(&vd
->vdev_config_dirty_node
))
1924 vdev_config_clean(vd
);
1927 * Clean up the vdev namespace.
1929 vdev_remove_make_hole_and_free(vd
);
1938 spa_vdev_remove_top_check(vdev_t
*vd
)
1940 spa_t
*spa
= vd
->vdev_spa
;
1942 if (vd
!= vd
->vdev_top
)
1943 return (SET_ERROR(ENOTSUP
));
1945 if (!spa_feature_is_enabled(spa
, SPA_FEATURE_DEVICE_REMOVAL
))
1946 return (SET_ERROR(ENOTSUP
));
1948 /* available space in the pool's normal class */
1949 uint64_t available
= dsl_dir_space_available(
1950 spa
->spa_dsl_pool
->dp_root_dir
, NULL
, 0, B_TRUE
);
1952 metaslab_class_t
*mc
= vd
->vdev_mg
->mg_class
;
1955 * When removing a vdev from an allocation class that has
1956 * remaining vdevs, include available space from the class.
1958 if (mc
!= spa_normal_class(spa
) && mc
->mc_groups
> 1) {
1959 uint64_t class_avail
= metaslab_class_get_space(mc
) -
1960 metaslab_class_get_alloc(mc
);
1962 /* add class space, adjusted for overhead */
1963 available
+= (class_avail
* 94) / 100;
1967 * There has to be enough free space to remove the
1968 * device and leave double the "slop" space (i.e. we
1969 * must leave at least 3% of the pool free, in addition to
1970 * the normal slop space).
1972 if (available
< vd
->vdev_stat
.vs_dspace
+ spa_get_slop_space(spa
)) {
1973 return (SET_ERROR(ENOSPC
));
1977 * There can not be a removal in progress.
1979 if (spa
->spa_removing_phys
.sr_state
== DSS_SCANNING
)
1980 return (SET_ERROR(EBUSY
));
1983 * The device must have all its data.
1985 if (!vdev_dtl_empty(vd
, DTL_MISSING
) ||
1986 !vdev_dtl_empty(vd
, DTL_OUTAGE
))
1987 return (SET_ERROR(EBUSY
));
1990 * The device must be healthy.
1992 if (!vdev_readable(vd
))
1993 return (SET_ERROR(EIO
));
1996 * All vdevs in normal class must have the same ashift.
1998 if (spa
->spa_max_ashift
!= spa
->spa_min_ashift
) {
1999 return (SET_ERROR(EINVAL
));
2003 * All vdevs in normal class must have the same ashift
2006 vdev_t
*rvd
= spa
->spa_root_vdev
;
2007 int num_indirect
= 0;
2008 for (uint64_t id
= 0; id
< rvd
->vdev_children
; id
++) {
2009 vdev_t
*cvd
= rvd
->vdev_child
[id
];
2010 if (cvd
->vdev_ashift
!= 0 && !cvd
->vdev_islog
)
2011 ASSERT3U(cvd
->vdev_ashift
, ==, spa
->spa_max_ashift
);
2012 if (cvd
->vdev_ops
== &vdev_indirect_ops
)
2014 if (!vdev_is_concrete(cvd
))
2016 if (cvd
->vdev_ops
== &vdev_raidz_ops
)
2017 return (SET_ERROR(EINVAL
));
2019 * Need the mirror to be mirror of leaf vdevs only
2021 if (cvd
->vdev_ops
== &vdev_mirror_ops
) {
2022 for (uint64_t cid
= 0;
2023 cid
< cvd
->vdev_children
; cid
++) {
2024 if (!cvd
->vdev_child
[cid
]->vdev_ops
->
2026 return (SET_ERROR(EINVAL
));
2035 * Initiate removal of a top-level vdev, reducing the total space in the pool.
2036 * The config lock is held for the specified TXG. Once initiated,
2037 * evacuation of all allocated space (copying it to other vdevs) happens
2038 * in the background (see spa_vdev_remove_thread()), and can be canceled
2039 * (see spa_vdev_remove_cancel()). If successful, the vdev will
2040 * be transformed to an indirect vdev (see spa_vdev_remove_complete()).
2043 spa_vdev_remove_top(vdev_t
*vd
, uint64_t *txg
)
2045 spa_t
*spa
= vd
->vdev_spa
;
2049 * Check for errors up-front, so that we don't waste time
2050 * passivating the metaslab group and clearing the ZIL if there
2053 error
= spa_vdev_remove_top_check(vd
);
2058 * Stop allocating from this vdev. Note that we must check
2059 * that this is not the only device in the pool before
2060 * passivating, otherwise we will not be able to make
2061 * progress because we can't allocate from any vdevs.
2062 * The above check for sufficient free space serves this
2065 metaslab_group_t
*mg
= vd
->vdev_mg
;
2066 metaslab_group_passivate(mg
);
2069 * Wait for the youngest allocations and frees to sync,
2070 * and then wait for the deferral of those frees to finish.
2072 spa_vdev_config_exit(spa
, NULL
,
2073 *txg
+ TXG_CONCURRENT_STATES
+ TXG_DEFER_SIZE
, 0, FTAG
);
2076 * We must ensure that no "stubby" log blocks are allocated
2077 * on the device to be removed. These blocks could be
2078 * written at any time, including while we are in the middle
2081 error
= spa_reset_logs(spa
);
2084 * We stop any initializing that is currently in progress but leave
2085 * the state as "active". This will allow the initializing to resume
2086 * if the removal is canceled sometime later.
2088 vdev_initialize_stop_all(vd
, VDEV_INITIALIZE_ACTIVE
);
2090 *txg
= spa_vdev_config_enter(spa
);
2093 * Things might have changed while the config lock was dropped
2094 * (e.g. space usage). Check for errors again.
2097 error
= spa_vdev_remove_top_check(vd
);
2100 metaslab_group_activate(mg
);
2101 spa_async_request(spa
, SPA_ASYNC_INITIALIZE_RESTART
);
2105 vd
->vdev_removing
= B_TRUE
;
2107 vdev_dirty_leaves(vd
, VDD_DTL
, *txg
);
2108 vdev_config_dirty(vd
);
2109 dmu_tx_t
*tx
= dmu_tx_create_assigned(spa
->spa_dsl_pool
, *txg
);
2110 dsl_sync_task_nowait(spa
->spa_dsl_pool
,
2111 vdev_remove_initiate_sync
,
2112 (void *)(uintptr_t)vd
->vdev_id
, 0, ZFS_SPACE_CHECK_NONE
, tx
);
2119 * Remove a device from the pool.
2121 * Removing a device from the vdev namespace requires several steps
2122 * and can take a significant amount of time. As a result we use
2123 * the spa_vdev_config_[enter/exit] functions which allow us to
2124 * grab and release the spa_config_lock while still holding the namespace
2125 * lock. During each step the configuration is synced out.
2128 spa_vdev_remove(spa_t
*spa
, uint64_t guid
, boolean_t unspare
)
2131 nvlist_t
**spares
, **l2cache
, *nv
;
2133 uint_t nspares
, nl2cache
;
2135 boolean_t locked
= MUTEX_HELD(&spa_namespace_lock
);
2136 sysevent_t
*ev
= NULL
;
2137 char *vd_type
= NULL
, *vd_path
= NULL
;
2139 ASSERT(spa_writeable(spa
));
2142 txg
= spa_vdev_enter(spa
);
2144 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
2145 if (spa_feature_is_active(spa
, SPA_FEATURE_POOL_CHECKPOINT
)) {
2146 error
= (spa_has_checkpoint(spa
)) ?
2147 ZFS_ERR_CHECKPOINT_EXISTS
: ZFS_ERR_DISCARDING_CHECKPOINT
;
2150 return (spa_vdev_exit(spa
, NULL
, txg
, error
));
2155 vd
= spa_lookup_by_guid(spa
, guid
, B_FALSE
);
2157 if (spa
->spa_spares
.sav_vdevs
!= NULL
&&
2158 nvlist_lookup_nvlist_array(spa
->spa_spares
.sav_config
,
2159 ZPOOL_CONFIG_SPARES
, &spares
, &nspares
) == 0 &&
2160 (nv
= spa_nvlist_lookup_by_guid(spares
, nspares
, guid
)) != NULL
) {
2162 * Only remove the hot spare if it's not currently in use
2165 if (vd
== NULL
|| unspare
) {
2167 vd
= spa_lookup_by_guid(spa
, guid
, B_TRUE
);
2168 ev
= spa_event_create(spa
, vd
, NULL
,
2169 ESC_ZFS_VDEV_REMOVE_AUX
);
2171 vd_type
= VDEV_TYPE_SPARE
;
2172 vd_path
= fnvlist_lookup_string(nv
, ZPOOL_CONFIG_PATH
);
2173 spa_vdev_remove_aux(spa
->spa_spares
.sav_config
,
2174 ZPOOL_CONFIG_SPARES
, spares
, nspares
, nv
);
2175 spa_load_spares(spa
);
2176 spa
->spa_spares
.sav_sync
= B_TRUE
;
2178 error
= SET_ERROR(EBUSY
);
2180 } else if (spa
->spa_l2cache
.sav_vdevs
!= NULL
&&
2181 nvlist_lookup_nvlist_array(spa
->spa_l2cache
.sav_config
,
2182 ZPOOL_CONFIG_L2CACHE
, &l2cache
, &nl2cache
) == 0 &&
2183 (nv
= spa_nvlist_lookup_by_guid(l2cache
, nl2cache
, guid
)) != NULL
) {
2184 vd_type
= VDEV_TYPE_L2CACHE
;
2185 vd_path
= fnvlist_lookup_string(nv
, ZPOOL_CONFIG_PATH
);
2187 * Cache devices can always be removed.
2189 vd
= spa_lookup_by_guid(spa
, guid
, B_TRUE
);
2190 ev
= spa_event_create(spa
, vd
, NULL
, ESC_ZFS_VDEV_REMOVE_AUX
);
2191 spa_vdev_remove_aux(spa
->spa_l2cache
.sav_config
,
2192 ZPOOL_CONFIG_L2CACHE
, l2cache
, nl2cache
, nv
);
2193 spa_load_l2cache(spa
);
2194 spa
->spa_l2cache
.sav_sync
= B_TRUE
;
2195 } else if (vd
!= NULL
&& vd
->vdev_islog
) {
2198 vd_path
= (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-";
2199 error
= spa_vdev_remove_log(vd
, &txg
);
2200 } else if (vd
!= NULL
) {
2202 error
= spa_vdev_remove_top(vd
, &txg
);
2205 * There is no vdev of any kind with the specified guid.
2207 error
= SET_ERROR(ENOENT
);
2211 error
= spa_vdev_exit(spa
, NULL
, txg
, error
);
2214 * Logging must be done outside the spa config lock. Otherwise,
2215 * this code path could end up holding the spa config lock while
2216 * waiting for a txg_sync so it can write to the internal log.
2217 * Doing that would prevent the txg sync from actually happening,
2218 * causing a deadlock.
2220 if (error
== 0 && vd_type
!= NULL
&& vd_path
!= NULL
) {
2221 spa_history_log_internal(spa
, "vdev remove", NULL
,
2222 "%s vdev (%s) %s", spa_name(spa
), vd_type
, vd_path
);
2232 spa_removal_get_stats(spa_t
*spa
, pool_removal_stat_t
*prs
)
2234 prs
->prs_state
= spa
->spa_removing_phys
.sr_state
;
2236 if (prs
->prs_state
== DSS_NONE
)
2237 return (SET_ERROR(ENOENT
));
2239 prs
->prs_removing_vdev
= spa
->spa_removing_phys
.sr_removing_vdev
;
2240 prs
->prs_start_time
= spa
->spa_removing_phys
.sr_start_time
;
2241 prs
->prs_end_time
= spa
->spa_removing_phys
.sr_end_time
;
2242 prs
->prs_to_copy
= spa
->spa_removing_phys
.sr_to_copy
;
2243 prs
->prs_copied
= spa
->spa_removing_phys
.sr_copied
;
2245 prs
->prs_mapping_memory
= 0;
2246 uint64_t indirect_vdev_id
=
2247 spa
->spa_removing_phys
.sr_prev_indirect_vdev
;
2248 while (indirect_vdev_id
!= -1) {
2249 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[indirect_vdev_id
];
2250 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
2251 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
2253 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
2254 prs
->prs_mapping_memory
+= vdev_indirect_mapping_size(vim
);
2255 indirect_vdev_id
= vic
->vic_prev_indirect_vdev
;
2261 #if defined(_KERNEL)
2262 module_param(zfs_removal_ignore_errors
, int, 0644);
2263 MODULE_PARM_DESC(zfs_removal_ignore_errors
,
2264 "Ignore hard IO errors when removing device");
2266 module_param(zfs_remove_max_segment
, int, 0644);
2267 MODULE_PARM_DESC(zfs_remove_max_segment
,
2268 "Largest contiguous segment to allocate when removing device");
2270 module_param(vdev_removal_max_span
, int, 0644);
2271 MODULE_PARM_DESC(vdev_removal_max_span
,
2272 "Largest span of free chunks a remap segment can span");
2275 module_param(zfs_removal_suspend_progress
, int, 0644);
2276 MODULE_PARM_DESC(zfs_removal_suspend_progress
,
2277 "Pause device removal after this many bytes are copied "
2278 "(debug use only - causes removal to hang)");
2281 EXPORT_SYMBOL(free_from_removing_vdev
);
2282 EXPORT_SYMBOL(spa_removal_get_stats
);
2283 EXPORT_SYMBOL(spa_remove_init
);
2284 EXPORT_SYMBOL(spa_restart_removal
);
2285 EXPORT_SYMBOL(spa_vdev_removal_destroy
);
2286 EXPORT_SYMBOL(spa_vdev_remove
);
2287 EXPORT_SYMBOL(spa_vdev_remove_cancel
);
2288 EXPORT_SYMBOL(spa_vdev_remove_suspend
);
2289 EXPORT_SYMBOL(svr_sync
);