1/*
2 * CDDL HEADER START
3 *
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.
7 *
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.
12 *
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]
18 *
19 * CDDL HEADER END
20 */
21
22/*
23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
25 */
26
27#include <sys/zfs_context.h>
28#include <sys/spa_impl.h>
29#include <sys/dmu.h>
30#include <sys/dmu_tx.h>
31#include <sys/zap.h>
32#include <sys/vdev_impl.h>
33#include <sys/metaslab.h>
34#include <sys/metaslab_impl.h>
35#include <sys/uberblock_impl.h>
36#include <sys/txg.h>
37#include <sys/avl.h>
38#include <sys/bpobj.h>
39#include <sys/dsl_pool.h>
40#include <sys/dsl_synctask.h>
41#include <sys/dsl_dir.h>
42#include <sys/arc.h>
43#include <sys/zfeature.h>
44#include <sys/vdev_indirect_births.h>
45#include <sys/vdev_indirect_mapping.h>
46#include <sys/abd.h>
47#include <sys/vdev_initialize.h>
48#include <sys/vdev_trim.h>
49
50/*
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.  (Note that members of a mirror can also be removed
55 * by the detach operation.)
56 *
57 * Log vdevs are removed by evacuating them and then turning the vdev
58 * into a hole vdev while holding spa config locks.
59 *
60 * Top level vdevs are removed and converted into an indirect vdev via
61 * a multi-step process:
62 *
63 *  - Disable allocations from this device (spa_vdev_remove_top).
64 *
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.
70 *
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).
74 *
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).
79 */
80
81typedef struct vdev_copy_arg {
82	metaslab_t	*vca_msp;
83	uint64_t	vca_outstanding_bytes;
84	kcondvar_t	vca_cv;
85	kmutex_t	vca_lock;
86} vdev_copy_arg_t;
87
88/*
89 * The maximum amount of memory we can use for outstanding i/o while
90 * doing a device removal.  This determines how much i/o we can have
91 * in flight concurrently.
92 */
93int zfs_remove_max_copy_bytes = 64 * 1024 * 1024;
94
95/*
96 * The largest contiguous segment that we will attempt to allocate when
97 * removing a device.  This can be no larger than SPA_MAXBLOCKSIZE.  If
98 * there is a performance problem with attempting to allocate large blocks,
99 * consider decreasing this.
100 *
101 * Note: we will issue I/Os of up to this size.  The mpt driver does not
102 * respond well to I/Os larger than 1MB, so we set this to 1MB.  (When
103 * mpt processes an I/O larger than 1MB, it needs to do an allocation of
104 * 2 physically contiguous pages; if this allocation fails, mpt will drop
105 * the I/O and hang the device.)
106 */
107int zfs_remove_max_segment = 1024 * 1024;
108
109/*
110 * Allow a remap segment to span free chunks of at most this size. The main
111 * impact of a larger span is that we will read and write larger, more
112 * contiguous chunks, with more "unnecessary" data -- trading off bandwidth
113 * for iops.  The value here was chosen to align with
114 * zfs_vdev_read_gap_limit, which is a similar concept when doing regular
115 * reads (but there's no reason it has to be the same).
116 *
117 * Additionally, a higher span will have the following relatively minor
118 * effects:
119 *  - the mapping will be smaller, since one entry can cover more allocated
120 *    segments
121 *  - more of the fragmentation in the removing device will be preserved
122 *  - we'll do larger allocations, which may fail and fall back on smaller
123 *    allocations
124 */
125int vdev_removal_max_span = 32 * 1024;
126
127/*
128 * This is used by the test suite so that it can ensure that certain
129 * actions happen while in the middle of a removal.
130 */
131int zfs_removal_suspend_progress = 0;
132
133#define	VDEV_REMOVAL_ZAP_OBJS	"lzap"
134
135static void spa_vdev_remove_thread(void *arg);
136
137static void
138spa_sync_removing_state(spa_t *spa, dmu_tx_t *tx)
139{
140	VERIFY0(zap_update(spa->spa_dsl_pool->dp_meta_objset,
141	    DMU_POOL_DIRECTORY_OBJECT,
142	    DMU_POOL_REMOVING, sizeof (uint64_t),
143	    sizeof (spa->spa_removing_phys) / sizeof (uint64_t),
144	    &spa->spa_removing_phys, tx));
145}
146
147static nvlist_t *
148spa_nvlist_lookup_by_guid(nvlist_t **nvpp, int count, uint64_t target_guid)
149{
150	for (int i = 0; i < count; i++) {
151		uint64_t guid =
152		    fnvlist_lookup_uint64(nvpp[i], ZPOOL_CONFIG_GUID);
153
154		if (guid == target_guid)
155			return (nvpp[i]);
156	}
157
158	return (NULL);
159}
160
161static void
162spa_vdev_remove_aux(nvlist_t *config, char *name, nvlist_t **dev, int count,
163    nvlist_t *dev_to_remove)
164{
165	nvlist_t **newdev = NULL;
166
167	if (count > 1)
168		newdev = kmem_alloc((count - 1) * sizeof (void *), KM_SLEEP);
169
170	for (int i = 0, j = 0; i < count; i++) {
171		if (dev[i] == dev_to_remove)
172			continue;
173		VERIFY(nvlist_dup(dev[i], &newdev[j++], KM_SLEEP) == 0);
174	}
175
176	VERIFY(nvlist_remove(config, name, DATA_TYPE_NVLIST_ARRAY) == 0);
177	VERIFY(nvlist_add_nvlist_array(config, name, newdev, count - 1) == 0);
178
179	for (int i = 0; i < count - 1; i++)
180		nvlist_free(newdev[i]);
181
182	if (count > 1)
183		kmem_free(newdev, (count - 1) * sizeof (void *));
184}
185
186static spa_vdev_removal_t *
187spa_vdev_removal_create(vdev_t *vd)
188{
189	spa_vdev_removal_t *svr = kmem_zalloc(sizeof (*svr), KM_SLEEP);
190	mutex_init(&svr->svr_lock, NULL, MUTEX_DEFAULT, NULL);
191	cv_init(&svr->svr_cv, NULL, CV_DEFAULT, NULL);
192	svr->svr_allocd_segs = range_tree_create(NULL, NULL);
193	svr->svr_vdev_id = vd->vdev_id;
194
195	for (int i = 0; i < TXG_SIZE; i++) {
196		svr->svr_frees[i] = range_tree_create(NULL, NULL);
197		list_create(&svr->svr_new_segments[i],
198		    sizeof (vdev_indirect_mapping_entry_t),
199		    offsetof(vdev_indirect_mapping_entry_t, vime_node));
200	}
201
202	return (svr);
203}
204
205void
206spa_vdev_removal_destroy(spa_vdev_removal_t *svr)
207{
208	for (int i = 0; i < TXG_SIZE; i++) {
209		ASSERT0(svr->svr_bytes_done[i]);
210		ASSERT0(svr->svr_max_offset_to_sync[i]);
211		range_tree_destroy(svr->svr_frees[i]);
212		list_destroy(&svr->svr_new_segments[i]);
213	}
214
215	range_tree_destroy(svr->svr_allocd_segs);
216	mutex_destroy(&svr->svr_lock);
217	cv_destroy(&svr->svr_cv);
218	kmem_free(svr, sizeof (*svr));
219}
220
221/*
222 * This is called as a synctask in the txg in which we will mark this vdev
223 * as removing (in the config stored in the MOS).
224 *
225 * It begins the evacuation of a toplevel vdev by:
226 * - initializing the spa_removing_phys which tracks this removal
227 * - computing the amount of space to remove for accounting purposes
228 * - dirtying all dbufs in the spa_config_object
229 * - creating the spa_vdev_removal
230 * - starting the spa_vdev_remove_thread
231 */
232static void
233vdev_remove_initiate_sync(void *arg, dmu_tx_t *tx)
234{
235	int vdev_id = (uintptr_t)arg;
236	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
237	vdev_t *vd = vdev_lookup_top(spa, vdev_id);
238	vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
239	objset_t *mos = spa->spa_dsl_pool->dp_meta_objset;
240	spa_vdev_removal_t *svr = NULL;
241	uint64_t txg = dmu_tx_get_txg(tx);
242
243	ASSERT3P(vd->vdev_ops, !=, &vdev_raidz_ops);
244	svr = spa_vdev_removal_create(vd);
245
246	ASSERT(vd->vdev_removing);
247	ASSERT3P(vd->vdev_indirect_mapping, ==, NULL);
248
249	spa_feature_incr(spa, SPA_FEATURE_DEVICE_REMOVAL, tx);
250	if (spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS)) {
251		/*
252		 * By activating the OBSOLETE_COUNTS feature, we prevent
253		 * the pool from being downgraded and ensure that the
254		 * refcounts are precise.
255		 */
256		spa_feature_incr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
257		uint64_t one = 1;
258		VERIFY0(zap_add(spa->spa_meta_objset, vd->vdev_top_zap,
259		    VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE, sizeof (one), 1,
260		    &one, tx));
261		ASSERT3U(vdev_obsolete_counts_are_precise(vd), !=, 0);
262	}
263
264	vic->vic_mapping_object = vdev_indirect_mapping_alloc(mos, tx);
265	vd->vdev_indirect_mapping =
266	    vdev_indirect_mapping_open(mos, vic->vic_mapping_object);
267	vic->vic_births_object = vdev_indirect_births_alloc(mos, tx);
268	vd->vdev_indirect_births =
269	    vdev_indirect_births_open(mos, vic->vic_births_object);
270	spa->spa_removing_phys.sr_removing_vdev = vd->vdev_id;
271	spa->spa_removing_phys.sr_start_time = gethrestime_sec();
272	spa->spa_removing_phys.sr_end_time = 0;
273	spa->spa_removing_phys.sr_state = DSS_SCANNING;
274	spa->spa_removing_phys.sr_to_copy = 0;
275	spa->spa_removing_phys.sr_copied = 0;
276
277	/*
278	 * Note: We can't use vdev_stat's vs_alloc for sr_to_copy, because
279	 * there may be space in the defer tree, which is free, but still
280	 * counted in vs_alloc.
281	 */
282	for (uint64_t i = 0; i < vd->vdev_ms_count; i++) {
283		metaslab_t *ms = vd->vdev_ms[i];
284		if (ms->ms_sm == NULL)
285			continue;
286
287		spa->spa_removing_phys.sr_to_copy +=
288		    metaslab_allocated_space(ms);
289
290		/*
291		 * Space which we are freeing this txg does not need to
292		 * be copied.
293		 */
294		spa->spa_removing_phys.sr_to_copy -=
295		    range_tree_space(ms->ms_freeing);
296
297		ASSERT0(range_tree_space(ms->ms_freed));
298		for (int t = 0; t < TXG_SIZE; t++)
299			ASSERT0(range_tree_space(ms->ms_allocating[t]));
300	}
301
302	/*
303	 * Sync tasks are called before metaslab_sync(), so there should
304	 * be no already-synced metaslabs in the TXG_CLEAN list.
305	 */
306	ASSERT3P(txg_list_head(&vd->vdev_ms_list, TXG_CLEAN(txg)), ==, NULL);
307
308	spa_sync_removing_state(spa, tx);
309
310	/*
311	 * All blocks that we need to read the most recent mapping must be
312	 * stored on concrete vdevs.  Therefore, we must dirty anything that
313	 * is read before spa_remove_init().  Specifically, the
314	 * spa_config_object.  (Note that although we already modified the
315	 * spa_config_object in spa_sync_removing_state, that may not have
316	 * modified all blocks of the object.)
317	 */
318	dmu_object_info_t doi;
319	VERIFY0(dmu_object_info(mos, DMU_POOL_DIRECTORY_OBJECT, &doi));
320	for (uint64_t offset = 0; offset < doi.doi_max_offset; ) {
321		dmu_buf_t *dbuf;
322		VERIFY0(dmu_buf_hold(mos, DMU_POOL_DIRECTORY_OBJECT,
323		    offset, FTAG, &dbuf, 0));
324		dmu_buf_will_dirty(dbuf, tx);
325		offset += dbuf->db_size;
326		dmu_buf_rele(dbuf, FTAG);
327	}
328
329	/*
330	 * Now that we've allocated the im_object, dirty the vdev to ensure
331	 * that the object gets written to the config on disk.
332	 */
333	vdev_config_dirty(vd);
334
335	zfs_dbgmsg("starting removal thread for vdev %llu (%p) in txg %llu "
336	    "im_obj=%llu", vd->vdev_id, vd, dmu_tx_get_txg(tx),
337	    vic->vic_mapping_object);
338
339	spa_history_log_internal(spa, "vdev remove started", tx,
340	    "%s vdev %llu %s", spa_name(spa), vd->vdev_id,
341	    (vd->vdev_path != NULL) ? vd->vdev_path : "-");
342	/*
343	 * Setting spa_vdev_removal causes subsequent frees to call
344	 * free_from_removing_vdev().  Note that we don't need any locking
345	 * because we are the sync thread, and metaslab_free_impl() is only
346	 * called from syncing context (potentially from a zio taskq thread,
347	 * but in any case only when there are outstanding free i/os, which
348	 * there are not).
349	 */
350	ASSERT3P(spa->spa_vdev_removal, ==, NULL);
351	spa->spa_vdev_removal = svr;
352	svr->svr_thread = thread_create(NULL, 0,
353	    spa_vdev_remove_thread, spa, 0, &p0, TS_RUN, minclsyspri);
354}
355
356/*
357 * When we are opening a pool, we must read the mapping for each
358 * indirect vdev in order from most recently removed to least
359 * recently removed.  We do this because the blocks for the mapping
360 * of older indirect vdevs may be stored on more recently removed vdevs.
361 * In order to read each indirect mapping object, we must have
362 * initialized all more recently removed vdevs.
363 */
364int
365spa_remove_init(spa_t *spa)
366{
367	int error;
368
369	error = zap_lookup(spa->spa_dsl_pool->dp_meta_objset,
370	    DMU_POOL_DIRECTORY_OBJECT,
371	    DMU_POOL_REMOVING, sizeof (uint64_t),
372	    sizeof (spa->spa_removing_phys) / sizeof (uint64_t),
373	    &spa->spa_removing_phys);
374
375	if (error == ENOENT) {
376		spa->spa_removing_phys.sr_state = DSS_NONE;
377		spa->spa_removing_phys.sr_removing_vdev = -1;
378		spa->spa_removing_phys.sr_prev_indirect_vdev = -1;
379		spa->spa_indirect_vdevs_loaded = B_TRUE;
380		return (0);
381	} else if (error != 0) {
382		return (error);
383	}
384
385	if (spa->spa_removing_phys.sr_state == DSS_SCANNING) {
386		/*
387		 * We are currently removing a vdev.  Create and
388		 * initialize a spa_vdev_removal_t from the bonus
389		 * buffer of the removing vdevs vdev_im_object, and
390		 * initialize its partial mapping.
391		 */
392		spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
393		vdev_t *vd = vdev_lookup_top(spa,
394		    spa->spa_removing_phys.sr_removing_vdev);
395
396		if (vd == NULL) {
397			spa_config_exit(spa, SCL_STATE, FTAG);
398			return (EINVAL);
399		}
400
401		vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
402
403		ASSERT(vdev_is_concrete(vd));
404		spa_vdev_removal_t *svr = spa_vdev_removal_create(vd);
405		ASSERT3U(svr->svr_vdev_id, ==, vd->vdev_id);
406		ASSERT(vd->vdev_removing);
407
408		vd->vdev_indirect_mapping = vdev_indirect_mapping_open(
409		    spa->spa_meta_objset, vic->vic_mapping_object);
410		vd->vdev_indirect_births = vdev_indirect_births_open(
411		    spa->spa_meta_objset, vic->vic_births_object);
412		spa_config_exit(spa, SCL_STATE, FTAG);
413
414		spa->spa_vdev_removal = svr;
415	}
416
417	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
418	uint64_t indirect_vdev_id =
419	    spa->spa_removing_phys.sr_prev_indirect_vdev;
420	while (indirect_vdev_id != UINT64_MAX) {
421		vdev_t *vd = vdev_lookup_top(spa, indirect_vdev_id);
422		vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
423
424		ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
425		vd->vdev_indirect_mapping = vdev_indirect_mapping_open(
426		    spa->spa_meta_objset, vic->vic_mapping_object);
427		vd->vdev_indirect_births = vdev_indirect_births_open(
428		    spa->spa_meta_objset, vic->vic_births_object);
429
430		indirect_vdev_id = vic->vic_prev_indirect_vdev;
431	}
432	spa_config_exit(spa, SCL_STATE, FTAG);
433
434	/*
435	 * Now that we've loaded all the indirect mappings, we can allow
436	 * reads from other blocks (e.g. via predictive prefetch).
437	 */
438	spa->spa_indirect_vdevs_loaded = B_TRUE;
439	return (0);
440}
441
442void
443spa_restart_removal(spa_t *spa)
444{
445	spa_vdev_removal_t *svr = spa->spa_vdev_removal;
446
447	if (svr == NULL)
448		return;
449
450	/*
451	 * In general when this function is called there is no
452	 * removal thread running. The only scenario where this
453	 * is not true is during spa_import() where this function
454	 * is called twice [once from spa_import_impl() and
455	 * spa_async_resume()]. Thus, in the scenario where we
456	 * import a pool that has an ongoing removal we don't
457	 * want to spawn a second thread.
458	 */
459	if (svr->svr_thread != NULL)
460		return;
461
462	if (!spa_writeable(spa))
463		return;
464
465	zfs_dbgmsg("restarting removal of %llu", svr->svr_vdev_id);
466	svr->svr_thread = thread_create(NULL, 0, spa_vdev_remove_thread, spa,
467	    0, &p0, TS_RUN, minclsyspri);
468}
469
470/*
471 * Process freeing from a device which is in the middle of being removed.
472 * We must handle this carefully so that we attempt to copy freed data,
473 * and we correctly free already-copied data.
474 */
475void
476free_from_removing_vdev(vdev_t *vd, uint64_t offset, uint64_t size)
477{
478	spa_t *spa = vd->vdev_spa;
479	spa_vdev_removal_t *svr = spa->spa_vdev_removal;
480	vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
481	uint64_t txg = spa_syncing_txg(spa);
482	uint64_t max_offset_yet = 0;
483
484	ASSERT(vd->vdev_indirect_config.vic_mapping_object != 0);
485	ASSERT3U(vd->vdev_indirect_config.vic_mapping_object, ==,
486	    vdev_indirect_mapping_object(vim));
487	ASSERT3U(vd->vdev_id, ==, svr->svr_vdev_id);
488
489	mutex_enter(&svr->svr_lock);
490
491	/*
492	 * Remove the segment from the removing vdev's spacemap.  This
493	 * ensures that we will not attempt to copy this space (if the
494	 * removal thread has not yet visited it), and also ensures
495	 * that we know what is actually allocated on the new vdevs
496	 * (needed if we cancel the removal).
497	 *
498	 * Note: we must do the metaslab_free_concrete() with the svr_lock
499	 * held, so that the remove_thread can not load this metaslab and then
500	 * visit this offset between the time that we metaslab_free_concrete()
501	 * and when we check to see if it has been visited.
502	 *
503	 * Note: The checkpoint flag is set to false as having/taking
504	 * a checkpoint and removing a device can't happen at the same
505	 * time.
506	 */
507	ASSERT(!spa_has_checkpoint(spa));
508	metaslab_free_concrete(vd, offset, size, B_FALSE);
509
510	uint64_t synced_size = 0;
511	uint64_t synced_offset = 0;
512	uint64_t max_offset_synced = vdev_indirect_mapping_max_offset(vim);
513	if (offset < max_offset_synced) {
514		/*
515		 * The mapping for this offset is already on disk.
516		 * Free from the new location.
517		 *
518		 * Note that we use svr_max_synced_offset because it is
519		 * updated atomically with respect to the in-core mapping.
520		 * By contrast, vim_max_offset is not.
521		 *
522		 * This block may be split between a synced entry and an
523		 * in-flight or unvisited entry.  Only process the synced
524		 * portion of it here.
525		 */
526		synced_size = MIN(size, max_offset_synced - offset);
527		synced_offset = offset;
528
529		ASSERT3U(max_offset_yet, <=, max_offset_synced);
530		max_offset_yet = max_offset_synced;
531
532		DTRACE_PROBE3(remove__free__synced,
533		    spa_t *, spa,
534		    uint64_t, offset,
535		    uint64_t, synced_size);
536
537		size -= synced_size;
538		offset += synced_size;
539	}
540
541	/*
542	 * Look at all in-flight txgs starting from the currently syncing one
543	 * and see if a section of this free is being copied. By starting from
544	 * this txg and iterating forward, we might find that this region
545	 * was copied in two different txgs and handle it appropriately.
546	 */
547	for (int i = 0; i < TXG_CONCURRENT_STATES; i++) {
548		int txgoff = (txg + i) & TXG_MASK;
549		if (size > 0 && offset < svr->svr_max_offset_to_sync[txgoff]) {
550			/*
551			 * The mapping for this offset is in flight, and
552			 * will be synced in txg+i.
553			 */
554			uint64_t inflight_size = MIN(size,
555			    svr->svr_max_offset_to_sync[txgoff] - offset);
556
557			DTRACE_PROBE4(remove__free__inflight,
558			    spa_t *, spa,
559			    uint64_t, offset,
560			    uint64_t, inflight_size,
561			    uint64_t, txg + i);
562
563			/*
564			 * We copy data in order of increasing offset.
565			 * Therefore the max_offset_to_sync[] must increase
566			 * (or be zero, indicating that nothing is being
567			 * copied in that txg).
568			 */
569			if (svr->svr_max_offset_to_sync[txgoff] != 0) {
570				ASSERT3U(svr->svr_max_offset_to_sync[txgoff],
571				    >=, max_offset_yet);
572				max_offset_yet =
573				    svr->svr_max_offset_to_sync[txgoff];
574			}
575
576			/*
577			 * We've already committed to copying this segment:
578			 * we have allocated space elsewhere in the pool for
579			 * it and have an IO outstanding to copy the data. We
580			 * cannot free the space before the copy has
581			 * completed, or else the copy IO might overwrite any
582			 * new data. To free that space, we record the
583			 * segment in the appropriate svr_frees tree and free
584			 * the mapped space later, in the txg where we have
585			 * completed the copy and synced the mapping (see
586			 * vdev_mapping_sync).
587			 */
588			range_tree_add(svr->svr_frees[txgoff],
589			    offset, inflight_size);
590			size -= inflight_size;
591			offset += inflight_size;
592
593			/*
594			 * This space is already accounted for as being
595			 * done, because it is being copied in txg+i.
596			 * However, if i!=0, then it is being copied in
597			 * a future txg.  If we crash after this txg
598			 * syncs but before txg+i syncs, then the space
599			 * will be free.  Therefore we must account
600			 * for the space being done in *this* txg
601			 * (when it is freed) rather than the future txg
602			 * (when it will be copied).
603			 */
604			ASSERT3U(svr->svr_bytes_done[txgoff], >=,
605			    inflight_size);
606			svr->svr_bytes_done[txgoff] -= inflight_size;
607			svr->svr_bytes_done[txg & TXG_MASK] += inflight_size;
608		}
609	}
610	ASSERT0(svr->svr_max_offset_to_sync[TXG_CLEAN(txg) & TXG_MASK]);
611
612	if (size > 0) {
613		/*
614		 * The copy thread has not yet visited this offset.  Ensure
615		 * that it doesn't.
616		 */
617
618		DTRACE_PROBE3(remove__free__unvisited,
619		    spa_t *, spa,
620		    uint64_t, offset,
621		    uint64_t, size);
622
623		if (svr->svr_allocd_segs != NULL)
624			range_tree_clear(svr->svr_allocd_segs, offset, size);
625
626		/*
627		 * Since we now do not need to copy this data, for
628		 * accounting purposes we have done our job and can count
629		 * it as completed.
630		 */
631		svr->svr_bytes_done[txg & TXG_MASK] += size;
632	}
633	mutex_exit(&svr->svr_lock);
634
635	/*
636	 * Now that we have dropped svr_lock, process the synced portion
637	 * of this free.
638	 */
639	if (synced_size > 0) {
640		vdev_indirect_mark_obsolete(vd, synced_offset, synced_size);
641
642		/*
643		 * Note: this can only be called from syncing context,
644		 * and the vdev_indirect_mapping is only changed from the
645		 * sync thread, so we don't need svr_lock while doing
646		 * metaslab_free_impl_cb.
647		 */
648		boolean_t checkpoint = B_FALSE;
649		vdev_indirect_ops.vdev_op_remap(vd, synced_offset, synced_size,
650		    metaslab_free_impl_cb, &checkpoint);
651	}
652}
653
654/*
655 * Stop an active removal and update the spa_removing phys.
656 */
657static void
658spa_finish_removal(spa_t *spa, dsl_scan_state_t state, dmu_tx_t *tx)
659{
660	spa_vdev_removal_t *svr = spa->spa_vdev_removal;
661	ASSERT3U(dmu_tx_get_txg(tx), ==, spa_syncing_txg(spa));
662
663	/* Ensure the removal thread has completed before we free the svr. */
664	spa_vdev_remove_suspend(spa);
665
666	ASSERT(state == DSS_FINISHED || state == DSS_CANCELED);
667
668	if (state == DSS_FINISHED) {
669		spa_removing_phys_t *srp = &spa->spa_removing_phys;
670		vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
671		vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
672
673		if (srp->sr_prev_indirect_vdev != UINT64_MAX) {
674			vdev_t *pvd = vdev_lookup_top(spa,
675			    srp->sr_prev_indirect_vdev);
676			ASSERT3P(pvd->vdev_ops, ==, &vdev_indirect_ops);
677		}
678
679		vic->vic_prev_indirect_vdev = srp->sr_prev_indirect_vdev;
680		srp->sr_prev_indirect_vdev = vd->vdev_id;
681	}
682	spa->spa_removing_phys.sr_state = state;
683	spa->spa_removing_phys.sr_end_time = gethrestime_sec();
684
685	spa->spa_vdev_removal = NULL;
686	spa_vdev_removal_destroy(svr);
687
688	spa_sync_removing_state(spa, tx);
689
690	vdev_config_dirty(spa->spa_root_vdev);
691}
692
693static void
694free_mapped_segment_cb(void *arg, uint64_t offset, uint64_t size)
695{
696	vdev_t *vd = arg;
697	vdev_indirect_mark_obsolete(vd, offset, size);
698	boolean_t checkpoint = B_FALSE;
699	vdev_indirect_ops.vdev_op_remap(vd, offset, size,
700	    metaslab_free_impl_cb, &checkpoint);
701}
702
703/*
704 * On behalf of the removal thread, syncs an incremental bit more of
705 * the indirect mapping to disk and updates the in-memory mapping.
706 * Called as a sync task in every txg that the removal thread makes progress.
707 */
708static void
709vdev_mapping_sync(void *arg, dmu_tx_t *tx)
710{
711	spa_vdev_removal_t *svr = arg;
712	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
713	vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
714	vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
715	uint64_t txg = dmu_tx_get_txg(tx);
716	vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
717
718	ASSERT(vic->vic_mapping_object != 0);
719	ASSERT3U(txg, ==, spa_syncing_txg(spa));
720
721	vdev_indirect_mapping_add_entries(vim,
722	    &svr->svr_new_segments[txg & TXG_MASK], tx);
723	vdev_indirect_births_add_entry(vd->vdev_indirect_births,
724	    vdev_indirect_mapping_max_offset(vim), dmu_tx_get_txg(tx), tx);
725
726	/*
727	 * Free the copied data for anything that was freed while the
728	 * mapping entries were in flight.
729	 */
730	mutex_enter(&svr->svr_lock);
731	range_tree_vacate(svr->svr_frees[txg & TXG_MASK],
732	    free_mapped_segment_cb, vd);
733	ASSERT3U(svr->svr_max_offset_to_sync[txg & TXG_MASK], >=,
734	    vdev_indirect_mapping_max_offset(vim));
735	svr->svr_max_offset_to_sync[txg & TXG_MASK] = 0;
736	mutex_exit(&svr->svr_lock);
737
738	spa_sync_removing_state(spa, tx);
739}
740
741typedef struct vdev_copy_segment_arg {
742	spa_t *vcsa_spa;
743	dva_t *vcsa_dest_dva;
744	uint64_t vcsa_txg;
745	range_tree_t *vcsa_obsolete_segs;
746} vdev_copy_segment_arg_t;
747
748static void
749unalloc_seg(void *arg, uint64_t start, uint64_t size)
750{
751	vdev_copy_segment_arg_t *vcsa = arg;
752	spa_t *spa = vcsa->vcsa_spa;
753	blkptr_t bp = { 0 };
754
755	BP_SET_BIRTH(&bp, TXG_INITIAL, TXG_INITIAL);
756	BP_SET_LSIZE(&bp, size);
757	BP_SET_PSIZE(&bp, size);
758	BP_SET_COMPRESS(&bp, ZIO_COMPRESS_OFF);
759	BP_SET_CHECKSUM(&bp, ZIO_CHECKSUM_OFF);
760	BP_SET_TYPE(&bp, DMU_OT_NONE);
761	BP_SET_LEVEL(&bp, 0);
762	BP_SET_DEDUP(&bp, 0);
763	BP_SET_BYTEORDER(&bp, ZFS_HOST_BYTEORDER);
764
765	DVA_SET_VDEV(&bp.blk_dva[0], DVA_GET_VDEV(vcsa->vcsa_dest_dva));
766	DVA_SET_OFFSET(&bp.blk_dva[0],
767	    DVA_GET_OFFSET(vcsa->vcsa_dest_dva) + start);
768	DVA_SET_ASIZE(&bp.blk_dva[0], size);
769
770	zio_free(spa, vcsa->vcsa_txg, &bp);
771}
772
773/*
774 * All reads and writes associated with a call to spa_vdev_copy_segment()
775 * are done.
776 */
777static void
778spa_vdev_copy_segment_done(zio_t *zio)
779{
780	vdev_copy_segment_arg_t *vcsa = zio->io_private;
781
782	range_tree_vacate(vcsa->vcsa_obsolete_segs,
783	    unalloc_seg, vcsa);
784	range_tree_destroy(vcsa->vcsa_obsolete_segs);
785	kmem_free(vcsa, sizeof (*vcsa));
786
787	spa_config_exit(zio->io_spa, SCL_STATE, zio->io_spa);
788}
789
790/*
791 * The write of the new location is done.
792 */
793static void
794spa_vdev_copy_segment_write_done(zio_t *zio)
795{
796	vdev_copy_arg_t *vca = zio->io_private;
797
798	abd_free(zio->io_abd);
799
800	mutex_enter(&vca->vca_lock);
801	vca->vca_outstanding_bytes -= zio->io_size;
802	cv_signal(&vca->vca_cv);
803	mutex_exit(&vca->vca_lock);
804}
805
806/*
807 * The read of the old location is done.  The parent zio is the write to
808 * the new location.  Allow it to start.
809 */
810static void
811spa_vdev_copy_segment_read_done(zio_t *zio)
812{
813	zio_nowait(zio_unique_parent(zio));
814}
815
816/*
817 * If the old and new vdevs are mirrors, we will read both sides of the old
818 * mirror, and write each copy to the corresponding side of the new mirror.
819 * If the old and new vdevs have a different number of children, we will do
820 * this as best as possible.  Since we aren't verifying checksums, this
821 * ensures that as long as there's a good copy of the data, we'll have a
822 * good copy after the removal, even if there's silent damage to one side
823 * of the mirror. If we're removing a mirror that has some silent damage,
824 * we'll have exactly the same damage in the new location (assuming that
825 * the new location is also a mirror).
826 *
827 * We accomplish this by creating a tree of zio_t's, with as many writes as
828 * there are "children" of the new vdev (a non-redundant vdev counts as one
829 * child, a 2-way mirror has 2 children, etc). Each write has an associated
830 * read from a child of the old vdev. Typically there will be the same
831 * number of children of the old and new vdevs.  However, if there are more
832 * children of the new vdev, some child(ren) of the old vdev will be issued
833 * multiple reads.  If there are more children of the old vdev, some copies
834 * will be dropped.
835 *
836 * For example, the tree of zio_t's for a 2-way mirror is:
837 *
838 *                            null
839 *                           /    \
840 *    write(new vdev, child 0)      write(new vdev, child 1)
841 *      |                             |
842 *    read(old vdev, child 0)       read(old vdev, child 1)
843 *
844 * Child zio's complete before their parents complete.  However, zio's
845 * created with zio_vdev_child_io() may be issued before their children
846 * complete.  In this case we need to make sure that the children (reads)
847 * complete before the parents (writes) are *issued*.  We do this by not
848 * calling zio_nowait() on each write until its corresponding read has
849 * completed.
850 *
851 * The spa_config_lock must be held while zio's created by
852 * zio_vdev_child_io() are in progress, to ensure that the vdev tree does
853 * not change (e.g. due to a concurrent "zpool attach/detach"). The "null"
854 * zio is needed to release the spa_config_lock after all the reads and
855 * writes complete. (Note that we can't grab the config lock for each read,
856 * because it is not reentrant - we could deadlock with a thread waiting
857 * for a write lock.)
858 */
859static void
860spa_vdev_copy_one_child(vdev_copy_arg_t *vca, zio_t *nzio,
861    vdev_t *source_vd, uint64_t source_offset,
862    vdev_t *dest_child_vd, uint64_t dest_offset, int dest_id, uint64_t size)
863{
864	ASSERT3U(spa_config_held(nzio->io_spa, SCL_ALL, RW_READER), !=, 0);
865
866	mutex_enter(&vca->vca_lock);
867	vca->vca_outstanding_bytes += size;
868	mutex_exit(&vca->vca_lock);
869
870	abd_t *abd = abd_alloc_for_io(size, B_FALSE);
871
872	vdev_t *source_child_vd;
873	if (source_vd->vdev_ops == &vdev_mirror_ops && dest_id != -1) {
874		/*
875		 * Source and dest are both mirrors.  Copy from the same
876		 * child id as we are copying to (wrapping around if there
877		 * are more dest children than source children).
878		 */
879		source_child_vd =
880		    source_vd->vdev_child[dest_id % source_vd->vdev_children];
881	} else {
882		source_child_vd = source_vd;
883	}
884
885	zio_t *write_zio = zio_vdev_child_io(nzio, NULL,
886	    dest_child_vd, dest_offset, abd, size,
887	    ZIO_TYPE_WRITE, ZIO_PRIORITY_REMOVAL,
888	    ZIO_FLAG_CANFAIL,
889	    spa_vdev_copy_segment_write_done, vca);
890
891	zio_nowait(zio_vdev_child_io(write_zio, NULL,
892	    source_child_vd, source_offset, abd, size,
893	    ZIO_TYPE_READ, ZIO_PRIORITY_REMOVAL,
894	    ZIO_FLAG_CANFAIL,
895	    spa_vdev_copy_segment_read_done, vca));
896}
897
898/*
899 * Allocate a new location for this segment, and create the zio_t's to
900 * read from the old location and write to the new location.
901 */
902static int
903spa_vdev_copy_segment(vdev_t *vd, range_tree_t *segs,
904    uint64_t maxalloc, uint64_t txg,
905    vdev_copy_arg_t *vca, zio_alloc_list_t *zal)
906{
907	metaslab_group_t *mg = vd->vdev_mg;
908	spa_t *spa = vd->vdev_spa;
909	spa_vdev_removal_t *svr = spa->spa_vdev_removal;
910	vdev_indirect_mapping_entry_t *entry;
911	dva_t dst = { 0 };
912	uint64_t start = range_tree_min(segs);
913
914	ASSERT3U(maxalloc, <=, SPA_MAXBLOCKSIZE);
915
916	uint64_t size = range_tree_span(segs);
917	if (range_tree_span(segs) > maxalloc) {
918		/*
919		 * We can't allocate all the segments.  Prefer to end
920		 * the allocation at the end of a segment, thus avoiding
921		 * additional split blocks.
922		 */
923		range_seg_t search;
924		avl_index_t where;
925		search.rs_start = start + maxalloc;
926		search.rs_end = search.rs_start;
927		range_seg_t *rs = avl_find(&segs->rt_root, &search, &where);
928		if (rs == NULL) {
929			rs = avl_nearest(&segs->rt_root, where, AVL_BEFORE);
930		} else {
931			rs = AVL_PREV(&segs->rt_root, rs);
932		}
933		if (rs != NULL) {
934			size = rs->rs_end - start;
935		} else {
936			/*
937			 * There are no segments that end before maxalloc.
938			 * I.e. the first segment is larger than maxalloc,
939			 * so we must split it.
940			 */
941			size = maxalloc;
942		}
943	}
944	ASSERT3U(size, <=, maxalloc);
945
946	/*
947	 * An allocation class might not have any remaining vdevs or space
948	 */
949	metaslab_class_t *mc = mg->mg_class;
950	if (mc != spa_normal_class(spa) && mc->mc_groups <= 1)
951		mc = spa_normal_class(spa);
952	int error = metaslab_alloc_dva(spa, mc, size, &dst, 0, NULL, txg, 0,
953	    zal, 0);
954	if (error == ENOSPC && mc != spa_normal_class(spa)) {
955		error = metaslab_alloc_dva(spa, spa_normal_class(spa), size,
956		    &dst, 0, NULL, txg, 0, zal, 0);
957	}
958	if (error != 0)
959		return (error);
960
961	/*
962	 * Determine the ranges that are not actually needed.  Offsets are
963	 * relative to the start of the range to be copied (i.e. relative to the
964	 * local variable "start").
965	 */
966	range_tree_t *obsolete_segs = range_tree_create(NULL, NULL);
967
968	range_seg_t *rs = avl_first(&segs->rt_root);
969	ASSERT3U(rs->rs_start, ==, start);
970	uint64_t prev_seg_end = rs->rs_end;
971	while ((rs = AVL_NEXT(&segs->rt_root, rs)) != NULL) {
972		if (rs->rs_start >= start + size) {
973			break;
974		} else {
975			range_tree_add(obsolete_segs,
976			    prev_seg_end - start,
977			    rs->rs_start - prev_seg_end);
978		}
979		prev_seg_end = rs->rs_end;
980	}
981	/* We don't end in the middle of an obsolete range */
982	ASSERT3U(start + size, <=, prev_seg_end);
983
984	range_tree_clear(segs, start, size);
985
986	/*
987	 * We can't have any padding of the allocated size, otherwise we will
988	 * misunderstand what's allocated, and the size of the mapping.
989	 * The caller ensures this will be true by passing in a size that is
990	 * aligned to the worst (highest) ashift in the pool.
991	 */
992	ASSERT3U(DVA_GET_ASIZE(&dst), ==, size);
993
994	entry = kmem_zalloc(sizeof (vdev_indirect_mapping_entry_t), KM_SLEEP);
995	DVA_MAPPING_SET_SRC_OFFSET(&entry->vime_mapping, start);
996	entry->vime_mapping.vimep_dst = dst;
997	if (spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS)) {
998		entry->vime_obsolete_count = range_tree_space(obsolete_segs);
999	}
1000
1001	vdev_copy_segment_arg_t *vcsa = kmem_zalloc(sizeof (*vcsa), KM_SLEEP);
1002	vcsa->vcsa_dest_dva = &entry->vime_mapping.vimep_dst;
1003	vcsa->vcsa_obsolete_segs = obsolete_segs;
1004	vcsa->vcsa_spa = spa;
1005	vcsa->vcsa_txg = txg;
1006
1007	/*
1008	 * See comment before spa_vdev_copy_one_child().
1009	 */
1010	spa_config_enter(spa, SCL_STATE, spa, RW_READER);
1011	zio_t *nzio = zio_null(spa->spa_txg_zio[txg & TXG_MASK], spa, NULL,
1012	    spa_vdev_copy_segment_done, vcsa, 0);
1013	vdev_t *dest_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dst));
1014	if (dest_vd->vdev_ops == &vdev_mirror_ops) {
1015		for (int i = 0; i < dest_vd->vdev_children; i++) {
1016			vdev_t *child = dest_vd->vdev_child[i];
1017			spa_vdev_copy_one_child(vca, nzio, vd, start,
1018			    child, DVA_GET_OFFSET(&dst), i, size);
1019		}
1020	} else {
1021		spa_vdev_copy_one_child(vca, nzio, vd, start,
1022		    dest_vd, DVA_GET_OFFSET(&dst), -1, size);
1023	}
1024	zio_nowait(nzio);
1025
1026	list_insert_tail(&svr->svr_new_segments[txg & TXG_MASK], entry);
1027	ASSERT3U(start + size, <=, vd->vdev_ms_count << vd->vdev_ms_shift);
1028	vdev_dirty(vd, 0, NULL, txg);
1029
1030	return (0);
1031}
1032
1033/*
1034 * Complete the removal of a toplevel vdev. This is called as a
1035 * synctask in the same txg that we will sync out the new config (to the
1036 * MOS object) which indicates that this vdev is indirect.
1037 */
1038static void
1039vdev_remove_complete_sync(void *arg, dmu_tx_t *tx)
1040{
1041	spa_vdev_removal_t *svr = arg;
1042	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
1043	vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
1044
1045	ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
1046
1047	for (int i = 0; i < TXG_SIZE; i++) {
1048		ASSERT0(svr->svr_bytes_done[i]);
1049	}
1050
1051	ASSERT3U(spa->spa_removing_phys.sr_copied, ==,
1052	    spa->spa_removing_phys.sr_to_copy);
1053
1054	vdev_destroy_spacemaps(vd, tx);
1055
1056	/* destroy leaf zaps, if any */
1057	ASSERT3P(svr->svr_zaplist, !=, NULL);
1058	for (nvpair_t *pair = nvlist_next_nvpair(svr->svr_zaplist, NULL);
1059	    pair != NULL;
1060	    pair = nvlist_next_nvpair(svr->svr_zaplist, pair)) {
1061		vdev_destroy_unlink_zap(vd, fnvpair_value_uint64(pair), tx);
1062	}
1063	fnvlist_free(svr->svr_zaplist);
1064
1065	spa_finish_removal(dmu_tx_pool(tx)->dp_spa, DSS_FINISHED, tx);
1066	/* vd->vdev_path is not available here */
1067	spa_history_log_internal(spa, "vdev remove completed",  tx,
1068	    "%s vdev %llu", spa_name(spa), vd->vdev_id);
1069}
1070
1071static void
1072vdev_remove_enlist_zaps(vdev_t *vd, nvlist_t *zlist)
1073{
1074	ASSERT3P(zlist, !=, NULL);
1075	ASSERT3P(vd->vdev_ops, !=, &vdev_raidz_ops);
1076
1077	if (vd->vdev_leaf_zap != 0) {
1078		char zkey[32];
1079		(void) snprintf(zkey, sizeof (zkey), "%s-%"PRIu64,
1080		    VDEV_REMOVAL_ZAP_OBJS, vd->vdev_leaf_zap);
1081		fnvlist_add_uint64(zlist, zkey, vd->vdev_leaf_zap);
1082	}
1083
1084	for (uint64_t id = 0; id < vd->vdev_children; id++) {
1085		vdev_remove_enlist_zaps(vd->vdev_child[id], zlist);
1086	}
1087}
1088
1089static void
1090vdev_remove_replace_with_indirect(vdev_t *vd, uint64_t txg)
1091{
1092	vdev_t *ivd;
1093	dmu_tx_t *tx;
1094	spa_t *spa = vd->vdev_spa;
1095	spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1096
1097	/*
1098	 * First, build a list of leaf zaps to be destroyed.
1099	 * This is passed to the sync context thread,
1100	 * which does the actual unlinking.
1101	 */
1102	svr->svr_zaplist = fnvlist_alloc();
1103	vdev_remove_enlist_zaps(vd, svr->svr_zaplist);
1104
1105	ivd = vdev_add_parent(vd, &vdev_indirect_ops);
1106	ivd->vdev_removing = 0;
1107
1108	vd->vdev_leaf_zap = 0;
1109
1110	vdev_remove_child(ivd, vd);
1111	vdev_compact_children(ivd);
1112
1113	ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
1114
1115	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1116	dsl_sync_task_nowait(spa->spa_dsl_pool, vdev_remove_complete_sync, svr,
1117	    0, ZFS_SPACE_CHECK_NONE, tx);
1118	dmu_tx_commit(tx);
1119
1120	/*
1121	 * Indicate that this thread has exited.
1122	 * After this, we can not use svr.
1123	 */
1124	mutex_enter(&svr->svr_lock);
1125	svr->svr_thread = NULL;
1126	cv_broadcast(&svr->svr_cv);
1127	mutex_exit(&svr->svr_lock);
1128}
1129
1130/*
1131 * Complete the removal of a toplevel vdev. This is called in open
1132 * context by the removal thread after we have copied all vdev's data.
1133 */
1134static void
1135vdev_remove_complete(spa_t *spa)
1136{
1137	uint64_t txg;
1138
1139	/*
1140	 * Wait for any deferred frees to be synced before we call
1141	 * vdev_metaslab_fini()
1142	 */
1143	txg_wait_synced(spa->spa_dsl_pool, 0);
1144	txg = spa_vdev_enter(spa);
1145	vdev_t *vd = vdev_lookup_top(spa, spa->spa_vdev_removal->svr_vdev_id);
1146	ASSERT3P(vd->vdev_initialize_thread, ==, NULL);
1147	ASSERT3P(vd->vdev_trim_thread, ==, NULL);
1148	ASSERT3P(vd->vdev_autotrim_thread, ==, NULL);
1149
1150	sysevent_t *ev = spa_event_create(spa, vd, NULL,
1151	    ESC_ZFS_VDEV_REMOVE_DEV);
1152
1153	zfs_dbgmsg("finishing device removal for vdev %llu in txg %llu",
1154	    vd->vdev_id, txg);
1155
1156	/*
1157	 * Discard allocation state.
1158	 */
1159	if (vd->vdev_mg != NULL) {
1160		vdev_metaslab_fini(vd);
1161		metaslab_group_destroy(vd->vdev_mg);
1162		vd->vdev_mg = NULL;
1163		spa_log_sm_set_blocklimit(spa);
1164	}
1165	ASSERT0(vd->vdev_stat.vs_space);
1166	ASSERT0(vd->vdev_stat.vs_dspace);
1167
1168	vdev_remove_replace_with_indirect(vd, txg);
1169
1170	/*
1171	 * We now release the locks, allowing spa_sync to run and finish the
1172	 * removal via vdev_remove_complete_sync in syncing context.
1173	 *
1174	 * Note that we hold on to the vdev_t that has been replaced.  Since
1175	 * it isn't part of the vdev tree any longer, it can't be concurrently
1176	 * manipulated, even while we don't have the config lock.
1177	 */
1178	(void) spa_vdev_exit(spa, NULL, txg, 0);
1179
1180	/*
1181	 * Top ZAP should have been transferred to the indirect vdev in
1182	 * vdev_remove_replace_with_indirect.
1183	 */
1184	ASSERT0(vd->vdev_top_zap);
1185
1186	/*
1187	 * Leaf ZAP should have been moved in vdev_remove_replace_with_indirect.
1188	 */
1189	ASSERT0(vd->vdev_leaf_zap);
1190
1191	txg = spa_vdev_enter(spa);
1192	(void) vdev_label_init(vd, 0, VDEV_LABEL_REMOVE);
1193	/*
1194	 * Request to update the config and the config cachefile.
1195	 */
1196	vdev_config_dirty(spa->spa_root_vdev);
1197	(void) spa_vdev_exit(spa, vd, txg, 0);
1198
1199	spa_event_post(ev);
1200}
1201
1202/*
1203 * Evacuates a segment of size at most max_alloc from the vdev
1204 * via repeated calls to spa_vdev_copy_segment. If an allocation
1205 * fails, the pool is probably too fragmented to handle such a
1206 * large size, so decrease max_alloc so that the caller will not try
1207 * this size again this txg.
1208 */
1209static void
1210spa_vdev_copy_impl(vdev_t *vd, spa_vdev_removal_t *svr, vdev_copy_arg_t *vca,
1211    uint64_t *max_alloc, dmu_tx_t *tx)
1212{
1213	uint64_t txg = dmu_tx_get_txg(tx);
1214	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
1215
1216	mutex_enter(&svr->svr_lock);
1217
1218	/*
1219	 * Determine how big of a chunk to copy.  We can allocate up
1220	 * to max_alloc bytes, and we can span up to vdev_removal_max_span
1221	 * bytes of unallocated space at a time.  "segs" will track the
1222	 * allocated segments that we are copying.  We may also be copying
1223	 * free segments (of up to vdev_removal_max_span bytes).
1224	 */
1225	range_tree_t *segs = range_tree_create(NULL, NULL);
1226	for (;;) {
1227		range_seg_t *rs = avl_first(&svr->svr_allocd_segs->rt_root);
1228		if (rs == NULL)
1229			break;
1230
1231		uint64_t seg_length;
1232
1233		if (range_tree_is_empty(segs)) {
1234			/* need to truncate the first seg based on max_alloc */
1235			seg_length =
1236			    MIN(rs->rs_end - rs->rs_start, *max_alloc);
1237		} else {
1238			if (rs->rs_start - range_tree_max(segs) >
1239			    vdev_removal_max_span) {
1240				/*
1241				 * Including this segment would cause us to
1242				 * copy a larger unneeded chunk than is allowed.
1243				 */
1244				break;
1245			} else if (rs->rs_end - range_tree_min(segs) >
1246			    *max_alloc) {
1247				/*
1248				 * This additional segment would extend past
1249				 * max_alloc. Rather than splitting this
1250				 * segment, leave it for the next mapping.
1251				 */
1252				break;
1253			} else {
1254				seg_length = rs->rs_end - rs->rs_start;
1255			}
1256		}
1257
1258		range_tree_add(segs, rs->rs_start, seg_length);
1259		range_tree_remove(svr->svr_allocd_segs,
1260		    rs->rs_start, seg_length);
1261	}
1262
1263	if (range_tree_is_empty(segs)) {
1264		mutex_exit(&svr->svr_lock);
1265		range_tree_destroy(segs);
1266		return;
1267	}
1268
1269	if (svr->svr_max_offset_to_sync[txg & TXG_MASK] == 0) {
1270		dsl_sync_task_nowait(dmu_tx_pool(tx), vdev_mapping_sync,
1271		    svr, 0, ZFS_SPACE_CHECK_NONE, tx);
1272	}
1273
1274	svr->svr_max_offset_to_sync[txg & TXG_MASK] = range_tree_max(segs);
1275
1276	/*
1277	 * Note: this is the amount of *allocated* space
1278	 * that we are taking care of each txg.
1279	 */
1280	svr->svr_bytes_done[txg & TXG_MASK] += range_tree_space(segs);
1281
1282	mutex_exit(&svr->svr_lock);
1283
1284	zio_alloc_list_t zal;
1285	metaslab_trace_init(&zal);
1286	uint64_t thismax = SPA_MAXBLOCKSIZE;
1287	while (!range_tree_is_empty(segs)) {
1288		int error = spa_vdev_copy_segment(vd,
1289		    segs, thismax, txg, vca, &zal);
1290
1291		if (error == ENOSPC) {
1292			/*
1293			 * Cut our segment in half, and don't try this
1294			 * segment size again this txg.  Note that the
1295			 * allocation size must be aligned to the highest
1296			 * ashift in the pool, so that the allocation will
1297			 * not be padded out to a multiple of the ashift,
1298			 * which could cause us to think that this mapping
1299			 * is larger than we intended.
1300			 */
1301			ASSERT3U(spa->spa_max_ashift, >=, SPA_MINBLOCKSHIFT);
1302			ASSERT3U(spa->spa_max_ashift, ==, spa->spa_min_ashift);
1303			uint64_t attempted =
1304			    MIN(range_tree_span(segs), thismax);
1305			thismax = P2ROUNDUP(attempted / 2,
1306			    1 << spa->spa_max_ashift);
1307			/*
1308			 * The minimum-size allocation can not fail.
1309			 */
1310			ASSERT3U(attempted, >, 1 << spa->spa_max_ashift);
1311			*max_alloc = attempted - (1 << spa->spa_max_ashift);
1312		} else {
1313			ASSERT0(error);
1314
1315			/*
1316			 * We've performed an allocation, so reset the
1317			 * alloc trace list.
1318			 */
1319			metaslab_trace_fini(&zal);
1320			metaslab_trace_init(&zal);
1321		}
1322	}
1323	metaslab_trace_fini(&zal);
1324	range_tree_destroy(segs);
1325}
1326
1327/*
1328 * The removal thread operates in open context.  It iterates over all
1329 * allocated space in the vdev, by loading each metaslab's spacemap.
1330 * For each contiguous segment of allocated space (capping the segment
1331 * size at SPA_MAXBLOCKSIZE), we:
1332 *    - Allocate space for it on another vdev.
1333 *    - Create a new mapping from the old location to the new location
1334 *      (as a record in svr_new_segments).
1335 *    - Initiate a logical read zio to get the data off the removing disk.
1336 *    - In the read zio's done callback, initiate a logical write zio to
1337 *      write it to the new vdev.
1338 * Note that all of this will take effect when a particular TXG syncs.
1339 * The sync thread ensures that all the phys reads and writes for the syncing
1340 * TXG have completed (see spa_txg_zio) and writes the new mappings to disk
1341 * (see vdev_mapping_sync()).
1342 */
1343static void
1344spa_vdev_remove_thread(void *arg)
1345{
1346	spa_t *spa = arg;
1347	spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1348	vdev_copy_arg_t vca;
1349	uint64_t max_alloc = zfs_remove_max_segment;
1350	uint64_t last_txg = 0;
1351
1352	spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
1353	vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
1354	vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
1355	uint64_t start_offset = vdev_indirect_mapping_max_offset(vim);
1356
1357	ASSERT3P(vd->vdev_ops, !=, &vdev_indirect_ops);
1358	ASSERT(vdev_is_concrete(vd));
1359	ASSERT(vd->vdev_removing);
1360	ASSERT(vd->vdev_indirect_config.vic_mapping_object != 0);
1361	ASSERT(vim != NULL);
1362
1363	mutex_init(&vca.vca_lock, NULL, MUTEX_DEFAULT, NULL);
1364	cv_init(&vca.vca_cv, NULL, CV_DEFAULT, NULL);
1365	vca.vca_outstanding_bytes = 0;
1366
1367	mutex_enter(&svr->svr_lock);
1368
1369	/*
1370	 * Start from vim_max_offset so we pick up where we left off
1371	 * if we are restarting the removal after opening the pool.
1372	 */
1373	uint64_t msi;
1374	for (msi = start_offset >> vd->vdev_ms_shift;
1375	    msi < vd->vdev_ms_count && !svr->svr_thread_exit; msi++) {
1376		metaslab_t *msp = vd->vdev_ms[msi];
1377		ASSERT3U(msi, <=, vd->vdev_ms_count);
1378
1379		ASSERT0(range_tree_space(svr->svr_allocd_segs));
1380
1381		mutex_enter(&msp->ms_sync_lock);
1382		mutex_enter(&msp->ms_lock);
1383
1384		/*
1385		 * Assert nothing in flight -- ms_*tree is empty.
1386		 */
1387		for (int i = 0; i < TXG_SIZE; i++) {
1388			ASSERT0(range_tree_space(msp->ms_allocating[i]));
1389		}
1390
1391		/*
1392		 * If the metaslab has ever been allocated from (ms_sm!=NULL),
1393		 * read the allocated segments from the space map object
1394		 * into svr_allocd_segs. Since we do this while holding
1395		 * svr_lock and ms_sync_lock, concurrent frees (which
1396		 * would have modified the space map) will wait for us
1397		 * to finish loading the spacemap, and then take the
1398		 * appropriate action (see free_from_removing_vdev()).
1399		 */
1400		if (msp->ms_sm != NULL) {
1401			VERIFY0(space_map_load(msp->ms_sm,
1402			    svr->svr_allocd_segs, SM_ALLOC));
1403
1404			range_tree_walk(msp->ms_unflushed_allocs,
1405			    range_tree_add, svr->svr_allocd_segs);
1406			range_tree_walk(msp->ms_unflushed_frees,
1407			    range_tree_remove, svr->svr_allocd_segs);
1408			range_tree_walk(msp->ms_freeing,
1409			    range_tree_remove, svr->svr_allocd_segs);
1410
1411			/*
1412			 * When we are resuming from a paused removal (i.e.
1413			 * when importing a pool with a removal in progress),
1414			 * discard any state that we have already processed.
1415			 */
1416			range_tree_clear(svr->svr_allocd_segs, 0, start_offset);
1417		}
1418		mutex_exit(&msp->ms_lock);
1419		mutex_exit(&msp->ms_sync_lock);
1420
1421		vca.vca_msp = msp;
1422		zfs_dbgmsg("copying %llu segments for metaslab %llu",
1423		    avl_numnodes(&svr->svr_allocd_segs->rt_root),
1424		    msp->ms_id);
1425
1426		while (!svr->svr_thread_exit &&
1427		    !range_tree_is_empty(svr->svr_allocd_segs)) {
1428
1429			mutex_exit(&svr->svr_lock);
1430
1431			/*
1432			 * We need to periodically drop the config lock so that
1433			 * writers can get in.  Additionally, we can't wait
1434			 * for a txg to sync while holding a config lock
1435			 * (since a waiting writer could cause a 3-way deadlock
1436			 * with the sync thread, which also gets a config
1437			 * lock for reader).  So we can't hold the config lock
1438			 * while calling dmu_tx_assign().
1439			 */
1440			spa_config_exit(spa, SCL_CONFIG, FTAG);
1441
1442			/*
1443			 * This delay will pause the removal around the point
1444			 * specified by zfs_removal_suspend_progress. We do this
1445			 * solely from the test suite or during debugging.
1446			 */
1447			uint64_t bytes_copied =
1448			    spa->spa_removing_phys.sr_copied;
1449			for (int i = 0; i < TXG_SIZE; i++)
1450				bytes_copied += svr->svr_bytes_done[i];
1451			while (zfs_removal_suspend_progress &&
1452			    !svr->svr_thread_exit)
1453				delay(hz);
1454
1455			mutex_enter(&vca.vca_lock);
1456			while (vca.vca_outstanding_bytes >
1457			    zfs_remove_max_copy_bytes) {
1458				cv_wait(&vca.vca_cv, &vca.vca_lock);
1459			}
1460			mutex_exit(&vca.vca_lock);
1461
1462			dmu_tx_t *tx =
1463			    dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
1464
1465			VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
1466			uint64_t txg = dmu_tx_get_txg(tx);
1467
1468			/*
1469			 * Reacquire the vdev_config lock.  The vdev_t
1470			 * that we're removing may have changed, e.g. due
1471			 * to a vdev_attach or vdev_detach.
1472			 */
1473			spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
1474			vd = vdev_lookup_top(spa, svr->svr_vdev_id);
1475
1476			if (txg != last_txg)
1477				max_alloc = zfs_remove_max_segment;
1478			last_txg = txg;
1479
1480			spa_vdev_copy_impl(vd, svr, &vca, &max_alloc, tx);
1481
1482			dmu_tx_commit(tx);
1483			mutex_enter(&svr->svr_lock);
1484		}
1485	}
1486
1487	mutex_exit(&svr->svr_lock);
1488
1489	spa_config_exit(spa, SCL_CONFIG, FTAG);
1490
1491	/*
1492	 * Wait for all copies to finish before cleaning up the vca.
1493	 */
1494	txg_wait_synced(spa->spa_dsl_pool, 0);
1495	ASSERT0(vca.vca_outstanding_bytes);
1496
1497	mutex_destroy(&vca.vca_lock);
1498	cv_destroy(&vca.vca_cv);
1499
1500	if (svr->svr_thread_exit) {
1501		mutex_enter(&svr->svr_lock);
1502		range_tree_vacate(svr->svr_allocd_segs, NULL, NULL);
1503		svr->svr_thread = NULL;
1504		cv_broadcast(&svr->svr_cv);
1505		mutex_exit(&svr->svr_lock);
1506	} else {
1507		ASSERT0(range_tree_space(svr->svr_allocd_segs));
1508		vdev_remove_complete(spa);
1509	}
1510}
1511
1512void
1513spa_vdev_remove_suspend(spa_t *spa)
1514{
1515	spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1516
1517	if (svr == NULL)
1518		return;
1519
1520	mutex_enter(&svr->svr_lock);
1521	svr->svr_thread_exit = B_TRUE;
1522	while (svr->svr_thread != NULL)
1523		cv_wait(&svr->svr_cv, &svr->svr_lock);
1524	svr->svr_thread_exit = B_FALSE;
1525	mutex_exit(&svr->svr_lock);
1526}
1527
1528/* ARGSUSED */
1529static int
1530spa_vdev_remove_cancel_check(void *arg, dmu_tx_t *tx)
1531{
1532	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
1533
1534	if (spa->spa_vdev_removal == NULL)
1535		return (ENOTACTIVE);
1536	return (0);
1537}
1538
1539/*
1540 * Cancel a removal by freeing all entries from the partial mapping
1541 * and marking the vdev as no longer being removing.
1542 */
1543/* ARGSUSED */
1544static void
1545spa_vdev_remove_cancel_sync(void *arg, dmu_tx_t *tx)
1546{
1547	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
1548	spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1549	vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
1550	vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
1551	vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
1552	objset_t *mos = spa->spa_meta_objset;
1553
1554	ASSERT3P(svr->svr_thread, ==, NULL);
1555
1556	spa_feature_decr(spa, SPA_FEATURE_DEVICE_REMOVAL, tx);
1557	if (vdev_obsolete_counts_are_precise(vd)) {
1558		spa_feature_decr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
1559		VERIFY0(zap_remove(spa->spa_meta_objset, vd->vdev_top_zap,
1560		    VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE, tx));
1561	}
1562
1563	if (vdev_obsolete_sm_object(vd) != 0) {
1564		ASSERT(vd->vdev_obsolete_sm != NULL);
1565		ASSERT3U(vdev_obsolete_sm_object(vd), ==,
1566		    space_map_object(vd->vdev_obsolete_sm));
1567
1568		space_map_free(vd->vdev_obsolete_sm, tx);
1569		VERIFY0(zap_remove(spa->spa_meta_objset, vd->vdev_top_zap,
1570		    VDEV_TOP_ZAP_INDIRECT_OBSOLETE_SM, tx));
1571		space_map_close(vd->vdev_obsolete_sm);
1572		vd->vdev_obsolete_sm = NULL;
1573		spa_feature_decr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
1574	}
1575	for (int i = 0; i < TXG_SIZE; i++) {
1576		ASSERT(list_is_empty(&svr->svr_new_segments[i]));
1577		ASSERT3U(svr->svr_max_offset_to_sync[i], <=,
1578		    vdev_indirect_mapping_max_offset(vim));
1579	}
1580
1581	for (uint64_t msi = 0; msi < vd->vdev_ms_count; msi++) {
1582		metaslab_t *msp = vd->vdev_ms[msi];
1583
1584		if (msp->ms_start >= vdev_indirect_mapping_max_offset(vim))
1585			break;
1586
1587		ASSERT0(range_tree_space(svr->svr_allocd_segs));
1588
1589		mutex_enter(&msp->ms_lock);
1590
1591		/*
1592		 * Assert nothing in flight -- ms_*tree is empty.
1593		 */
1594		for (int i = 0; i < TXG_SIZE; i++)
1595			ASSERT0(range_tree_space(msp->ms_allocating[i]));
1596		for (int i = 0; i < TXG_DEFER_SIZE; i++)
1597			ASSERT0(range_tree_space(msp->ms_defer[i]));
1598		ASSERT0(range_tree_space(msp->ms_freed));
1599
1600		if (msp->ms_sm != NULL) {
1601			mutex_enter(&svr->svr_lock);
1602			VERIFY0(space_map_load(msp->ms_sm,
1603			    svr->svr_allocd_segs, SM_ALLOC));
1604
1605			range_tree_walk(msp->ms_unflushed_allocs,
1606			    range_tree_add, svr->svr_allocd_segs);
1607			range_tree_walk(msp->ms_unflushed_frees,
1608			    range_tree_remove, svr->svr_allocd_segs);
1609			range_tree_walk(msp->ms_freeing,
1610			    range_tree_remove, svr->svr_allocd_segs);
1611
1612			/*
1613			 * Clear everything past what has been synced,
1614			 * because we have not allocated mappings for it yet.
1615			 */
1616			uint64_t syncd = vdev_indirect_mapping_max_offset(vim);
1617			uint64_t sm_end = msp->ms_sm->sm_start +
1618			    msp->ms_sm->sm_size;
1619			if (sm_end > syncd)
1620				range_tree_clear(svr->svr_allocd_segs,
1621				    syncd, sm_end - syncd);
1622
1623			mutex_exit(&svr->svr_lock);
1624		}
1625		mutex_exit(&msp->ms_lock);
1626
1627		mutex_enter(&svr->svr_lock);
1628		range_tree_vacate(svr->svr_allocd_segs,
1629		    free_mapped_segment_cb, vd);
1630		mutex_exit(&svr->svr_lock);
1631	}
1632
1633	/*
1634	 * Note: this must happen after we invoke free_mapped_segment_cb,
1635	 * because it adds to the obsolete_segments.
1636	 */
1637	range_tree_vacate(vd->vdev_obsolete_segments, NULL, NULL);
1638
1639	ASSERT3U(vic->vic_mapping_object, ==,
1640	    vdev_indirect_mapping_object(vd->vdev_indirect_mapping));
1641	vdev_indirect_mapping_close(vd->vdev_indirect_mapping);
1642	vd->vdev_indirect_mapping = NULL;
1643	vdev_indirect_mapping_free(mos, vic->vic_mapping_object, tx);
1644	vic->vic_mapping_object = 0;
1645
1646	ASSERT3U(vic->vic_births_object, ==,
1647	    vdev_indirect_births_object(vd->vdev_indirect_births));
1648	vdev_indirect_births_close(vd->vdev_indirect_births);
1649	vd->vdev_indirect_births = NULL;
1650	vdev_indirect_births_free(mos, vic->vic_births_object, tx);
1651	vic->vic_births_object = 0;
1652
1653	/*
1654	 * We may have processed some frees from the removing vdev in this
1655	 * txg, thus increasing svr_bytes_done; discard that here to
1656	 * satisfy the assertions in spa_vdev_removal_destroy().
1657	 * Note that future txg's can not have any bytes_done, because
1658	 * future TXG's are only modified from open context, and we have
1659	 * already shut down the copying thread.
1660	 */
1661	svr->svr_bytes_done[dmu_tx_get_txg(tx) & TXG_MASK] = 0;
1662	spa_finish_removal(spa, DSS_CANCELED, tx);
1663
1664	vd->vdev_removing = B_FALSE;
1665	vdev_config_dirty(vd);
1666
1667	zfs_dbgmsg("canceled device removal for vdev %llu in %llu",
1668	    vd->vdev_id, dmu_tx_get_txg(tx));
1669	spa_history_log_internal(spa, "vdev remove canceled", tx,
1670	    "%s vdev %llu %s", spa_name(spa),
1671	    vd->vdev_id, (vd->vdev_path != NULL) ? vd->vdev_path : "-");
1672}
1673
1674int
1675spa_vdev_remove_cancel(spa_t *spa)
1676{
1677	spa_vdev_remove_suspend(spa);
1678
1679	if (spa->spa_vdev_removal == NULL)
1680		return (ENOTACTIVE);
1681
1682	uint64_t vdid = spa->spa_vdev_removal->svr_vdev_id;
1683
1684	int error = dsl_sync_task(spa->spa_name, spa_vdev_remove_cancel_check,
1685	    spa_vdev_remove_cancel_sync, NULL, 0,
1686	    ZFS_SPACE_CHECK_EXTRA_RESERVED);
1687
1688	if (error == 0) {
1689		spa_config_enter(spa, SCL_ALLOC | SCL_VDEV, FTAG, RW_WRITER);
1690		vdev_t *vd = vdev_lookup_top(spa, vdid);
1691		metaslab_group_activate(vd->vdev_mg);
1692		spa_config_exit(spa, SCL_ALLOC | SCL_VDEV, FTAG);
1693	}
1694
1695	return (error);
1696}
1697
1698void
1699svr_sync(spa_t *spa, dmu_tx_t *tx)
1700{
1701	spa_vdev_removal_t *svr = spa->spa_vdev_removal;
1702	int txgoff = dmu_tx_get_txg(tx) & TXG_MASK;
1703
1704	if (svr == NULL)
1705		return;
1706
1707	/*
1708	 * This check is necessary so that we do not dirty the
1709	 * DIRECTORY_OBJECT via spa_sync_removing_state() when there
1710	 * is nothing to do.  Dirtying it every time would prevent us
1711	 * from syncing-to-convergence.
1712	 */
1713	if (svr->svr_bytes_done[txgoff] == 0)
1714		return;
1715
1716	/*
1717	 * Update progress accounting.
1718	 */
1719	spa->spa_removing_phys.sr_copied += svr->svr_bytes_done[txgoff];
1720	svr->svr_bytes_done[txgoff] = 0;
1721
1722	spa_sync_removing_state(spa, tx);
1723}
1724
1725static void
1726vdev_remove_make_hole_and_free(vdev_t *vd)
1727{
1728	uint64_t id = vd->vdev_id;
1729	spa_t *spa = vd->vdev_spa;
1730	vdev_t *rvd = spa->spa_root_vdev;
1731
1732	ASSERT(MUTEX_HELD(&spa_namespace_lock));
1733	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1734
1735	vdev_free(vd);
1736
1737	vd = vdev_alloc_common(spa, id, 0, &vdev_hole_ops);
1738	vdev_add_child(rvd, vd);
1739	vdev_config_dirty(rvd);
1740
1741	/*
1742	 * Reassess the health of our root vdev.
1743	 */
1744	vdev_reopen(rvd);
1745}
1746
1747/*
1748 * Remove a log device.  The config lock is held for the specified TXG.
1749 */
1750static int
1751spa_vdev_remove_log(vdev_t *vd, uint64_t *txg)
1752{
1753	metaslab_group_t *mg = vd->vdev_mg;
1754	spa_t *spa = vd->vdev_spa;
1755	int error = 0;
1756
1757	ASSERT(vd->vdev_islog);
1758	ASSERT(vd == vd->vdev_top);
1759	ASSERT(MUTEX_HELD(&spa_namespace_lock));
1760
1761	/*
1762	 * Stop allocating from this vdev.
1763	 */
1764	metaslab_group_passivate(mg);
1765
1766	/*
1767	 * Wait for the youngest allocations and frees to sync,
1768	 * and then wait for the deferral of those frees to finish.
1769	 */
1770	spa_vdev_config_exit(spa, NULL,
1771	    *txg + TXG_CONCURRENT_STATES + TXG_DEFER_SIZE, 0, FTAG);
1772
1773	/*
1774	 * Evacuate the device.  We don't hold the config lock as
1775	 * writer since we need to do I/O but we do keep the
1776	 * spa_namespace_lock held.  Once this completes the device
1777	 * should no longer have any blocks allocated on it.
1778	 */
1779	ASSERT(MUTEX_HELD(&spa_namespace_lock));
1780	if (vd->vdev_stat.vs_alloc != 0)
1781		error = spa_reset_logs(spa);
1782
1783	*txg = spa_vdev_config_enter(spa);
1784
1785	if (error != 0) {
1786		metaslab_group_activate(mg);
1787		return (error);
1788	}
1789	ASSERT0(vd->vdev_stat.vs_alloc);
1790
1791	/*
1792	 * The evacuation succeeded.  Remove any remaining MOS metadata
1793	 * associated with this vdev, and wait for these changes to sync.
1794	 */
1795	vd->vdev_removing = B_TRUE;
1796
1797	vdev_dirty_leaves(vd, VDD_DTL, *txg);
1798	vdev_config_dirty(vd);
1799
1800	/*
1801	 * When the log space map feature is enabled we look at
1802	 * the vdev's top_zap to find the on-disk flush data of
1803	 * the metaslab we just flushed. Thus, while removing a
1804	 * log vdev we make sure to call vdev_metaslab_fini()
1805	 * first, which removes all metaslabs of this vdev from
1806	 * spa_metaslabs_by_flushed before vdev_remove_empty()
1807	 * destroys the top_zap of this log vdev.
1808	 *
1809	 * This avoids the scenario where we flush a metaslab
1810	 * from the log vdev being removed that doesn't have a
1811	 * top_zap and end up failing to lookup its on-disk flush
1812	 * data.
1813	 *
1814	 * We don't call metaslab_group_destroy() right away
1815	 * though (it will be called in vdev_free() later) as
1816	 * during metaslab_sync() of metaslabs from other vdevs
1817	 * we may touch the metaslab group of this vdev through
1818	 * metaslab_class_histogram_verify()
1819	 */
1820	vdev_metaslab_fini(vd);
1821	spa_log_sm_set_blocklimit(spa);
1822
1823	spa_history_log_internal(spa, "vdev remove", NULL,
1824	    "%s vdev %llu (log) %s", spa_name(spa), vd->vdev_id,
1825	    (vd->vdev_path != NULL) ? vd->vdev_path : "-");
1826
1827	/* Make sure these changes are sync'ed */
1828	spa_vdev_config_exit(spa, NULL, *txg, 0, FTAG);
1829
1830	/* Stop initializing and TRIM */
1831	vdev_initialize_stop_all(vd, VDEV_INITIALIZE_CANCELED);
1832	vdev_trim_stop_all(vd, VDEV_TRIM_CANCELED);
1833	vdev_autotrim_stop_wait(vd);
1834
1835	*txg = spa_vdev_config_enter(spa);
1836
1837	sysevent_t *ev = spa_event_create(spa, vd, NULL,
1838	    ESC_ZFS_VDEV_REMOVE_DEV);
1839	ASSERT(MUTEX_HELD(&spa_namespace_lock));
1840	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1841
1842	/* The top ZAP should have been destroyed by vdev_remove_empty. */
1843	ASSERT0(vd->vdev_top_zap);
1844	/* The leaf ZAP should have been destroyed by vdev_dtl_sync. */
1845	ASSERT0(vd->vdev_leaf_zap);
1846
1847	(void) vdev_label_init(vd, 0, VDEV_LABEL_REMOVE);
1848
1849	if (list_link_active(&vd->vdev_state_dirty_node))
1850		vdev_state_clean(vd);
1851	if (list_link_active(&vd->vdev_config_dirty_node))
1852		vdev_config_clean(vd);
1853
1854	ASSERT0(vd->vdev_stat.vs_alloc);
1855
1856	/*
1857	 * Clean up the vdev namespace.
1858	 */
1859	vdev_remove_make_hole_and_free(vd);
1860
1861	if (ev != NULL)
1862		spa_event_post(ev);
1863
1864	return (0);
1865}
1866
1867static int
1868spa_vdev_remove_top_check(vdev_t *vd)
1869{
1870	spa_t *spa = vd->vdev_spa;
1871
1872	if (vd != vd->vdev_top)
1873		return (SET_ERROR(ENOTSUP));
1874
1875	if (!spa_feature_is_enabled(spa, SPA_FEATURE_DEVICE_REMOVAL))
1876		return (SET_ERROR(ENOTSUP));
1877
1878	/* available space in the pool's normal class */
1879	uint64_t available = dsl_dir_space_available(
1880	    spa->spa_dsl_pool->dp_root_dir, NULL, 0, B_TRUE);
1881
1882	metaslab_class_t *mc = vd->vdev_mg->mg_class;
1883
1884	/*
1885	 * When removing a vdev from an allocation class that has
1886	 * remaining vdevs, include available space from the class.
1887	 */
1888	if (mc != spa_normal_class(spa) && mc->mc_groups > 1) {
1889		uint64_t class_avail = metaslab_class_get_space(mc) -
1890		    metaslab_class_get_alloc(mc);
1891
1892		/* add class space, adjusted for overhead */
1893		available += (class_avail * 94) / 100;
1894	}
1895
1896	/*
1897	 * There has to be enough free space to remove the
1898	 * device and leave double the "slop" space (i.e. we
1899	 * must leave at least 3% of the pool free, in addition to
1900	 * the normal slop space).
1901	 */
1902	if (available < vd->vdev_stat.vs_dspace + spa_get_slop_space(spa)) {
1903		return (SET_ERROR(ENOSPC));
1904	}
1905
1906	/*
1907	 * There can not be a removal in progress.
1908	 */
1909	if (spa->spa_removing_phys.sr_state == DSS_SCANNING)
1910		return (SET_ERROR(EBUSY));
1911
1912	/*
1913	 * The device must have all its data.
1914	 */
1915	if (!vdev_dtl_empty(vd, DTL_MISSING) ||
1916	    !vdev_dtl_empty(vd, DTL_OUTAGE))
1917		return (SET_ERROR(EBUSY));
1918
1919	/*
1920	 * The device must be healthy.
1921	 */
1922	if (!vdev_readable(vd))
1923		return (SET_ERROR(EIO));
1924
1925	/*
1926	 * All vdevs in normal class must have the same ashift.
1927	 */
1928	if (spa->spa_max_ashift != spa->spa_min_ashift) {
1929		return (SET_ERROR(EINVAL));
1930	}
1931
1932	/*
1933	 * All vdevs in normal class must have the same ashift
1934	 * and not be raidz.
1935	 */
1936	vdev_t *rvd = spa->spa_root_vdev;
1937	int num_indirect = 0;
1938	for (uint64_t id = 0; id < rvd->vdev_children; id++) {
1939		vdev_t *cvd = rvd->vdev_child[id];
1940		if (cvd->vdev_ashift != 0 && !cvd->vdev_islog)
1941			ASSERT3U(cvd->vdev_ashift, ==, spa->spa_max_ashift);
1942		if (cvd->vdev_ops == &vdev_indirect_ops)
1943			num_indirect++;
1944		if (!vdev_is_concrete(cvd))
1945			continue;
1946		if (cvd->vdev_ops == &vdev_raidz_ops)
1947			return (SET_ERROR(EINVAL));
1948		/*
1949		 * Need the mirror to be mirror of leaf vdevs only
1950		 */
1951		if (cvd->vdev_ops == &vdev_mirror_ops) {
1952			for (uint64_t cid = 0;
1953			    cid < cvd->vdev_children; cid++) {
1954				vdev_t *tmp = cvd->vdev_child[cid];
1955				if (!tmp->vdev_ops->vdev_op_leaf)
1956					return (SET_ERROR(EINVAL));
1957			}
1958		}
1959	}
1960
1961	return (0);
1962}
1963
1964/*
1965 * Initiate removal of a top-level vdev, reducing the total space in the pool.
1966 * The config lock is held for the specified TXG.  Once initiated,
1967 * evacuation of all allocated space (copying it to other vdevs) happens
1968 * in the background (see spa_vdev_remove_thread()), and can be canceled
1969 * (see spa_vdev_remove_cancel()).  If successful, the vdev will
1970 * be transformed to an indirect vdev (see spa_vdev_remove_complete()).
1971 */
1972static int
1973spa_vdev_remove_top(vdev_t *vd, uint64_t *txg)
1974{
1975	spa_t *spa = vd->vdev_spa;
1976	int error;
1977
1978	/*
1979	 * Check for errors up-front, so that we don't waste time
1980	 * passivating the metaslab group and clearing the ZIL if there
1981	 * are errors.
1982	 */
1983	error = spa_vdev_remove_top_check(vd);
1984	if (error != 0)
1985		return (error);
1986
1987	/*
1988	 * Stop allocating from this vdev.  Note that we must check
1989	 * that this is not the only device in the pool before
1990	 * passivating, otherwise we will not be able to make
1991	 * progress because we can't allocate from any vdevs.
1992	 * The above check for sufficient free space serves this
1993	 * purpose.
1994	 */
1995	metaslab_group_t *mg = vd->vdev_mg;
1996	metaslab_group_passivate(mg);
1997
1998	/*
1999	 * Wait for the youngest allocations and frees to sync,
2000	 * and then wait for the deferral of those frees to finish.
2001	 */
2002	spa_vdev_config_exit(spa, NULL,
2003	    *txg + TXG_CONCURRENT_STATES + TXG_DEFER_SIZE, 0, FTAG);
2004
2005	/*
2006	 * We must ensure that no "stubby" log blocks are allocated
2007	 * on the device to be removed.  These blocks could be
2008	 * written at any time, including while we are in the middle
2009	 * of copying them.
2010	 */
2011	error = spa_reset_logs(spa);
2012
2013	/*
2014	 * We stop any initializing and TRIM that is currently in progress
2015	 * but leave the state as "active". This will allow the process to
2016	 * resume if the removal is canceled sometime later.
2017	 */
2018	vdev_initialize_stop_all(vd, VDEV_INITIALIZE_ACTIVE);
2019	vdev_trim_stop_all(vd, VDEV_TRIM_ACTIVE);
2020	vdev_autotrim_stop_wait(vd);
2021
2022	*txg = spa_vdev_config_enter(spa);
2023
2024	/*
2025	 * Things might have changed while the config lock was dropped
2026	 * (e.g. space usage).  Check for errors again.
2027	 */
2028	if (error == 0)
2029		error = spa_vdev_remove_top_check(vd);
2030
2031	if (error != 0) {
2032		metaslab_group_activate(mg);
2033		spa_async_request(spa, SPA_ASYNC_INITIALIZE_RESTART);
2034		spa_async_request(spa, SPA_ASYNC_TRIM_RESTART);
2035		spa_async_request(spa, SPA_ASYNC_AUTOTRIM_RESTART);
2036		return (error);
2037	}
2038
2039	vd->vdev_removing = B_TRUE;
2040
2041	vdev_dirty_leaves(vd, VDD_DTL, *txg);
2042	vdev_config_dirty(vd);
2043	dmu_tx_t *tx = dmu_tx_create_assigned(spa->spa_dsl_pool, *txg);
2044	dsl_sync_task_nowait(spa->spa_dsl_pool,
2045	    vdev_remove_initiate_sync,
2046	    (void *)(uintptr_t)vd->vdev_id, 0, ZFS_SPACE_CHECK_NONE, tx);
2047	dmu_tx_commit(tx);
2048
2049	return (0);
2050}
2051
2052/*
2053 * Remove a device from the pool.
2054 *
2055 * Removing a device from the vdev namespace requires several steps
2056 * and can take a significant amount of time.  As a result we use
2057 * the spa_vdev_config_[enter/exit] functions which allow us to
2058 * grab and release the spa_config_lock while still holding the namespace
2059 * lock.  During each step the configuration is synced out.
2060 */
2061int
2062spa_vdev_remove(spa_t *spa, uint64_t guid, boolean_t unspare)
2063{
2064	vdev_t *vd;
2065	nvlist_t **spares, **l2cache, *nv;
2066	uint64_t txg = 0;
2067	uint_t nspares, nl2cache;
2068	int error = 0;
2069	boolean_t locked = MUTEX_HELD(&spa_namespace_lock);
2070	sysevent_t *ev = NULL;
2071
2072	ASSERT(spa_writeable(spa));
2073
2074	if (!locked)
2075		txg = spa_vdev_enter(spa);
2076
2077	ASSERT(MUTEX_HELD(&spa_namespace_lock));
2078	if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
2079		error = (spa_has_checkpoint(spa)) ?
2080		    ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT;
2081
2082		if (!locked)
2083			return (spa_vdev_exit(spa, NULL, txg, error));
2084
2085		return (error);
2086	}
2087
2088	vd = spa_lookup_by_guid(spa, guid, B_FALSE);
2089
2090	if (spa->spa_spares.sav_vdevs != NULL &&
2091	    nvlist_lookup_nvlist_array(spa->spa_spares.sav_config,
2092	    ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0 &&
2093	    (nv = spa_nvlist_lookup_by_guid(spares, nspares, guid)) != NULL) {
2094		/*
2095		 * Only remove the hot spare if it's not currently in use
2096		 * in this pool.
2097		 */
2098		if (vd == NULL || unspare) {
2099			char *nvstr = fnvlist_lookup_string(nv,
2100			    ZPOOL_CONFIG_PATH);
2101			spa_history_log_internal(spa, "vdev remove", NULL,
2102			    "%s vdev (%s) %s", spa_name(spa),
2103			    VDEV_TYPE_SPARE, nvstr);
2104			if (vd == NULL)
2105				vd = spa_lookup_by_guid(spa, guid, B_TRUE);
2106			ev = spa_event_create(spa, vd, NULL,
2107			    ESC_ZFS_VDEV_REMOVE_AUX);
2108			spa_vdev_remove_aux(spa->spa_spares.sav_config,
2109			    ZPOOL_CONFIG_SPARES, spares, nspares, nv);
2110			spa_load_spares(spa);
2111			spa->spa_spares.sav_sync = B_TRUE;
2112		} else {
2113			error = SET_ERROR(EBUSY);
2114		}
2115	} else if (spa->spa_l2cache.sav_vdevs != NULL &&
2116	    nvlist_lookup_nvlist_array(spa->spa_l2cache.sav_config,
2117	    ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0 &&
2118	    (nv = spa_nvlist_lookup_by_guid(l2cache, nl2cache, guid)) != NULL) {
2119		char *nvstr = fnvlist_lookup_string(nv, ZPOOL_CONFIG_PATH);
2120		spa_history_log_internal(spa, "vdev remove", NULL,
2121		    "%s vdev (%s) %s", spa_name(spa), VDEV_TYPE_L2CACHE, nvstr);
2122		/*
2123		 * Cache devices can always be removed.
2124		 */
2125		vd = spa_lookup_by_guid(spa, guid, B_TRUE);
2126		ev = spa_event_create(spa, vd, NULL, ESC_ZFS_VDEV_REMOVE_AUX);
2127		spa_vdev_remove_aux(spa->spa_l2cache.sav_config,
2128		    ZPOOL_CONFIG_L2CACHE, l2cache, nl2cache, nv);
2129		spa_load_l2cache(spa);
2130		spa->spa_l2cache.sav_sync = B_TRUE;
2131	} else if (vd != NULL && vd->vdev_islog) {
2132		ASSERT(!locked);
2133		error = spa_vdev_remove_log(vd, &txg);
2134	} else if (vd != NULL) {
2135		ASSERT(!locked);
2136		error = spa_vdev_remove_top(vd, &txg);
2137	} else {
2138		/*
2139		 * There is no vdev of any kind with the specified guid.
2140		 */
2141		error = SET_ERROR(ENOENT);
2142	}
2143
2144	if (!locked)
2145		error = spa_vdev_exit(spa, NULL, txg, error);
2146
2147	if (ev != NULL) {
2148		if (error != 0) {
2149			spa_event_discard(ev);
2150		} else {
2151			spa_event_post(ev);
2152		}
2153	}
2154
2155	return (error);
2156}
2157
2158int
2159spa_removal_get_stats(spa_t *spa, pool_removal_stat_t *prs)
2160{
2161	prs->prs_state = spa->spa_removing_phys.sr_state;
2162
2163	if (prs->prs_state == DSS_NONE)
2164		return (SET_ERROR(ENOENT));
2165
2166	prs->prs_removing_vdev = spa->spa_removing_phys.sr_removing_vdev;
2167	prs->prs_start_time = spa->spa_removing_phys.sr_start_time;
2168	prs->prs_end_time = spa->spa_removing_phys.sr_end_time;
2169	prs->prs_to_copy = spa->spa_removing_phys.sr_to_copy;
2170	prs->prs_copied = spa->spa_removing_phys.sr_copied;
2171
2172	if (spa->spa_vdev_removal != NULL) {
2173		for (int i = 0; i < TXG_SIZE; i++) {
2174			prs->prs_copied +=
2175			    spa->spa_vdev_removal->svr_bytes_done[i];
2176		}
2177	}
2178
2179	prs->prs_mapping_memory = 0;
2180	uint64_t indirect_vdev_id =
2181	    spa->spa_removing_phys.sr_prev_indirect_vdev;
2182	while (indirect_vdev_id != -1) {
2183		vdev_t *vd = spa->spa_root_vdev->vdev_child[indirect_vdev_id];
2184		vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
2185		vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
2186
2187		ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
2188		prs->prs_mapping_memory += vdev_indirect_mapping_size(vim);
2189		indirect_vdev_id = vic->vic_prev_indirect_vdev;
2190	}
2191
2192	return (0);
2193}
2194