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 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
26 * Copyright (c) 2017, Intel Corporation.
27 */
28
29#include <sys/zfs_context.h>
30#include <sys/dmu.h>
31#include <sys/dmu_tx.h>
32#include <sys/space_map.h>
33#include <sys/metaslab_impl.h>
34#include <sys/vdev_impl.h>
35#include <sys/zio.h>
36#include <sys/spa_impl.h>
37#include <sys/zfeature.h>
38#include <sys/vdev_indirect_mapping.h>
39#include <sys/zap.h>
40#include <sys/btree.h>
41
42#define	GANG_ALLOCATION(flags) \
43	((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER))
44
45uint64_t metaslab_aliquot = 512ULL << 10;
46uint64_t metaslab_force_ganging = SPA_MAXBLOCKSIZE + 1;	/* force gang blocks */
47
48/*
49 * In pools where the log space map feature is not enabled we touch
50 * multiple metaslabs (and their respective space maps) with each
51 * transaction group. Thus, we benefit from having a small space map
52 * block size since it allows us to issue more I/O operations scattered
53 * around the disk. So a sane default for the space map block size
54 * is 8~16K.
55 */
56int zfs_metaslab_sm_blksz_no_log = (1 << 14);
57
58/*
59 * When the log space map feature is enabled, we accumulate a lot of
60 * changes per metaslab that are flushed once in a while so we benefit
61 * from a bigger block size like 128K for the metaslab space maps.
62 */
63int zfs_metaslab_sm_blksz_with_log = (1 << 17);
64
65/*
66 * The in-core space map representation is more compact than its on-disk form.
67 * The zfs_condense_pct determines how much more compact the in-core
68 * space map representation must be before we compact it on-disk.
69 * Values should be greater than or equal to 100.
70 */
71int zfs_condense_pct = 200;
72
73/*
74 * Condensing a metaslab is not guaranteed to actually reduce the amount of
75 * space used on disk. In particular, a space map uses data in increments of
76 * MAX(1 << ashift, space_map_blksize), so a metaslab might use the
77 * same number of blocks after condensing. Since the goal of condensing is to
78 * reduce the number of IOPs required to read the space map, we only want to
79 * condense when we can be sure we will reduce the number of blocks used by the
80 * space map. Unfortunately, we cannot precisely compute whether or not this is
81 * the case in metaslab_should_condense since we are holding ms_lock. Instead,
82 * we apply the following heuristic: do not condense a spacemap unless the
83 * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
84 * blocks.
85 */
86int zfs_metaslab_condense_block_threshold = 4;
87
88/*
89 * The zfs_mg_noalloc_threshold defines which metaslab groups should
90 * be eligible for allocation. The value is defined as a percentage of
91 * free space. Metaslab groups that have more free space than
92 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
93 * a metaslab group's free space is less than or equal to the
94 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
95 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
96 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
97 * groups are allowed to accept allocations. Gang blocks are always
98 * eligible to allocate on any metaslab group. The default value of 0 means
99 * no metaslab group will be excluded based on this criterion.
100 */
101int zfs_mg_noalloc_threshold = 0;
102
103/*
104 * Metaslab groups are considered eligible for allocations if their
105 * fragmenation metric (measured as a percentage) is less than or
106 * equal to zfs_mg_fragmentation_threshold. If a metaslab group
107 * exceeds this threshold then it will be skipped unless all metaslab
108 * groups within the metaslab class have also crossed this threshold.
109 *
110 * This tunable was introduced to avoid edge cases where we continue
111 * allocating from very fragmented disks in our pool while other, less
112 * fragmented disks, exists. On the other hand, if all disks in the
113 * pool are uniformly approaching the threshold, the threshold can
114 * be a speed bump in performance, where we keep switching the disks
115 * that we allocate from (e.g. we allocate some segments from disk A
116 * making it bypassing the threshold while freeing segments from disk
117 * B getting its fragmentation below the threshold).
118 *
119 * Empirically, we've seen that our vdev selection for allocations is
120 * good enough that fragmentation increases uniformly across all vdevs
121 * the majority of the time. Thus we set the threshold percentage high
122 * enough to avoid hitting the speed bump on pools that are being pushed
123 * to the edge.
124 */
125int zfs_mg_fragmentation_threshold = 95;
126
127/*
128 * Allow metaslabs to keep their active state as long as their fragmentation
129 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
130 * active metaslab that exceeds this threshold will no longer keep its active
131 * status allowing better metaslabs to be selected.
132 */
133int zfs_metaslab_fragmentation_threshold = 70;
134
135/*
136 * When set will load all metaslabs when pool is first opened.
137 */
138int metaslab_debug_load = 0;
139
140/*
141 * When set will prevent metaslabs from being unloaded.
142 */
143int metaslab_debug_unload = 0;
144
145/*
146 * Minimum size which forces the dynamic allocator to change
147 * it's allocation strategy.  Once the space map cannot satisfy
148 * an allocation of this size then it switches to using more
149 * aggressive strategy (i.e search by size rather than offset).
150 */
151uint64_t metaslab_df_alloc_threshold = SPA_OLD_MAXBLOCKSIZE;
152
153/*
154 * The minimum free space, in percent, which must be available
155 * in a space map to continue allocations in a first-fit fashion.
156 * Once the space map's free space drops below this level we dynamically
157 * switch to using best-fit allocations.
158 */
159int metaslab_df_free_pct = 4;
160
161/*
162 * Maximum distance to search forward from the last offset. Without this
163 * limit, fragmented pools can see >100,000 iterations and
164 * metaslab_block_picker() becomes the performance limiting factor on
165 * high-performance storage.
166 *
167 * With the default setting of 16MB, we typically see less than 500
168 * iterations, even with very fragmented, ashift=9 pools. The maximum number
169 * of iterations possible is:
170 *     metaslab_df_max_search / (2 * (1<<ashift))
171 * With the default setting of 16MB this is 16*1024 (with ashift=9) or
172 * 2048 (with ashift=12).
173 */
174int metaslab_df_max_search = 16 * 1024 * 1024;
175
176/*
177 * Forces the metaslab_block_picker function to search for at least this many
178 * segments forwards until giving up on finding a segment that the allocation
179 * will fit into.
180 */
181uint32_t metaslab_min_search_count = 100;
182
183/*
184 * If we are not searching forward (due to metaslab_df_max_search,
185 * metaslab_df_free_pct, or metaslab_df_alloc_threshold), this tunable
186 * controls what segment is used.  If it is set, we will use the largest free
187 * segment.  If it is not set, we will use a segment of exactly the requested
188 * size (or larger).
189 */
190int metaslab_df_use_largest_segment = B_FALSE;
191
192/*
193 * A metaslab is considered "free" if it contains a contiguous
194 * segment which is greater than metaslab_min_alloc_size.
195 */
196uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
197
198/*
199 * Percentage of all cpus that can be used by the metaslab taskq.
200 */
201int metaslab_load_pct = 50;
202
203/*
204 * These tunables control how long a metaslab will remain loaded after the
205 * last allocation from it.  A metaslab can't be unloaded until at least
206 * metaslab_unload_delay TXG's and metaslab_unload_delay_ms milliseconds
207 * have elapsed.  However, zfs_metaslab_mem_limit may cause it to be
208 * unloaded sooner.  These settings are intended to be generous -- to keep
209 * metaslabs loaded for a long time, reducing the rate of metaslab loading.
210 */
211int metaslab_unload_delay = 32;
212int metaslab_unload_delay_ms = 10 * 60 * 1000; /* ten minutes */
213
214/*
215 * Max number of metaslabs per group to preload.
216 */
217int metaslab_preload_limit = 10;
218
219/*
220 * Enable/disable preloading of metaslab.
221 */
222boolean_t metaslab_preload_enabled = B_TRUE;
223
224/*
225 * Enable/disable fragmentation weighting on metaslabs.
226 */
227boolean_t metaslab_fragmentation_factor_enabled = B_TRUE;
228
229/*
230 * Enable/disable lba weighting (i.e. outer tracks are given preference).
231 */
232boolean_t metaslab_lba_weighting_enabled = B_TRUE;
233
234/*
235 * Enable/disable metaslab group biasing.
236 */
237boolean_t metaslab_bias_enabled = B_TRUE;
238
239/*
240 * Enable/disable remapping of indirect DVAs to their concrete vdevs.
241 */
242boolean_t zfs_remap_blkptr_enable = B_TRUE;
243
244/*
245 * Enable/disable segment-based metaslab selection.
246 */
247boolean_t zfs_metaslab_segment_weight_enabled = B_TRUE;
248
249/*
250 * When using segment-based metaslab selection, we will continue
251 * allocating from the active metaslab until we have exhausted
252 * zfs_metaslab_switch_threshold of its buckets.
253 */
254int zfs_metaslab_switch_threshold = 2;
255
256/*
257 * Internal switch to enable/disable the metaslab allocation tracing
258 * facility.
259 */
260boolean_t metaslab_trace_enabled = B_TRUE;
261
262/*
263 * Maximum entries that the metaslab allocation tracing facility will keep
264 * in a given list when running in non-debug mode. We limit the number
265 * of entries in non-debug mode to prevent us from using up too much memory.
266 * The limit should be sufficiently large that we don't expect any allocation
267 * to every exceed this value. In debug mode, the system will panic if this
268 * limit is ever reached allowing for further investigation.
269 */
270uint64_t metaslab_trace_max_entries = 5000;
271
272/*
273 * Maximum number of metaslabs per group that can be disabled
274 * simultaneously.
275 */
276int max_disabled_ms = 3;
277
278/*
279 * Time (in seconds) to respect ms_max_size when the metaslab is not loaded.
280 * To avoid 64-bit overflow, don't set above UINT32_MAX.
281 */
282unsigned long zfs_metaslab_max_size_cache_sec = 3600; /* 1 hour */
283
284/*
285 * Maximum percentage of memory to use on storing loaded metaslabs. If loading
286 * a metaslab would take it over this percentage, the oldest selected metaslab
287 * is automatically unloaded.
288 */
289int zfs_metaslab_mem_limit = 75;
290
291/*
292 * Force the per-metaslab range trees to use 64-bit integers to store
293 * segments. Used for debugging purposes.
294 */
295boolean_t zfs_metaslab_force_large_segs = B_FALSE;
296
297/*
298 * By default we only store segments over a certain size in the size-sorted
299 * metaslab trees (ms_allocatable_by_size and
300 * ms_unflushed_frees_by_size). This dramatically reduces memory usage and
301 * improves load and unload times at the cost of causing us to use slightly
302 * larger segments than we would otherwise in some cases.
303 */
304uint32_t metaslab_by_size_min_shift = 14;
305
306static uint64_t metaslab_weight(metaslab_t *);
307static void metaslab_set_fragmentation(metaslab_t *);
308static void metaslab_free_impl(vdev_t *, uint64_t, uint64_t, boolean_t);
309static void metaslab_check_free_impl(vdev_t *, uint64_t, uint64_t);
310static void metaslab_passivate(metaslab_t *msp, uint64_t weight);
311static uint64_t metaslab_weight_from_range_tree(metaslab_t *msp);
312static void metaslab_flush_update(metaslab_t *, dmu_tx_t *);
313static unsigned int metaslab_idx_func(multilist_t *, void *);
314static void metaslab_evict(metaslab_t *, uint64_t);
315static void metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg);
316
317kmem_cache_t *metaslab_alloc_trace_cache;
318
319typedef struct metaslab_stats {
320	kstat_named_t metaslabstat_trace_over_limit;
321	kstat_named_t metaslabstat_df_find_under_floor;
322	kstat_named_t metaslabstat_reload_tree;
323} metaslab_stats_t;
324
325static metaslab_stats_t metaslab_stats = {
326	{ "trace_over_limit",		KSTAT_DATA_UINT64 },
327	{ "df_find_under_floor",	KSTAT_DATA_UINT64 },
328	{ "reload_tree",		KSTAT_DATA_UINT64 },
329};
330
331#define	METASLABSTAT_BUMP(stat) \
332	atomic_inc_64(&metaslab_stats.stat.value.ui64);
333
334
335kstat_t *metaslab_ksp;
336
337void
338metaslab_stat_init(void)
339{
340	ASSERT(metaslab_alloc_trace_cache == NULL);
341	metaslab_alloc_trace_cache = kmem_cache_create(
342	    "metaslab_alloc_trace_cache", sizeof (metaslab_alloc_trace_t),
343	    0, NULL, NULL, NULL, NULL, NULL, 0);
344	metaslab_ksp = kstat_create("zfs", 0, "metaslab_stats",
345	    "misc", KSTAT_TYPE_NAMED, sizeof (metaslab_stats) /
346	    sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
347	if (metaslab_ksp != NULL) {
348		metaslab_ksp->ks_data = &metaslab_stats;
349		kstat_install(metaslab_ksp);
350	}
351}
352
353void
354metaslab_stat_fini(void)
355{
356	if (metaslab_ksp != NULL) {
357		kstat_delete(metaslab_ksp);
358		metaslab_ksp = NULL;
359	}
360
361	kmem_cache_destroy(metaslab_alloc_trace_cache);
362	metaslab_alloc_trace_cache = NULL;
363}
364
365/*
366 * ==========================================================================
367 * Metaslab classes
368 * ==========================================================================
369 */
370metaslab_class_t *
371metaslab_class_create(spa_t *spa, metaslab_ops_t *ops)
372{
373	metaslab_class_t *mc;
374
375	mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
376
377	mc->mc_spa = spa;
378	mc->mc_rotor = NULL;
379	mc->mc_ops = ops;
380	mutex_init(&mc->mc_lock, NULL, MUTEX_DEFAULT, NULL);
381	mc->mc_metaslab_txg_list = multilist_create(sizeof (metaslab_t),
382	    offsetof(metaslab_t, ms_class_txg_node), metaslab_idx_func);
383	mc->mc_alloc_slots = kmem_zalloc(spa->spa_alloc_count *
384	    sizeof (zfs_refcount_t), KM_SLEEP);
385	mc->mc_alloc_max_slots = kmem_zalloc(spa->spa_alloc_count *
386	    sizeof (uint64_t), KM_SLEEP);
387	for (int i = 0; i < spa->spa_alloc_count; i++)
388		zfs_refcount_create_tracked(&mc->mc_alloc_slots[i]);
389
390	return (mc);
391}
392
393void
394metaslab_class_destroy(metaslab_class_t *mc)
395{
396	ASSERT(mc->mc_rotor == NULL);
397	ASSERT(mc->mc_alloc == 0);
398	ASSERT(mc->mc_deferred == 0);
399	ASSERT(mc->mc_space == 0);
400	ASSERT(mc->mc_dspace == 0);
401
402	for (int i = 0; i < mc->mc_spa->spa_alloc_count; i++)
403		zfs_refcount_destroy(&mc->mc_alloc_slots[i]);
404	kmem_free(mc->mc_alloc_slots, mc->mc_spa->spa_alloc_count *
405	    sizeof (zfs_refcount_t));
406	kmem_free(mc->mc_alloc_max_slots, mc->mc_spa->spa_alloc_count *
407	    sizeof (uint64_t));
408	mutex_destroy(&mc->mc_lock);
409	multilist_destroy(mc->mc_metaslab_txg_list);
410	kmem_free(mc, sizeof (metaslab_class_t));
411}
412
413int
414metaslab_class_validate(metaslab_class_t *mc)
415{
416	metaslab_group_t *mg;
417	vdev_t *vd;
418
419	/*
420	 * Must hold one of the spa_config locks.
421	 */
422	ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
423	    spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
424
425	if ((mg = mc->mc_rotor) == NULL)
426		return (0);
427
428	do {
429		vd = mg->mg_vd;
430		ASSERT(vd->vdev_mg != NULL);
431		ASSERT3P(vd->vdev_top, ==, vd);
432		ASSERT3P(mg->mg_class, ==, mc);
433		ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
434	} while ((mg = mg->mg_next) != mc->mc_rotor);
435
436	return (0);
437}
438
439static void
440metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
441    int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
442{
443	atomic_add_64(&mc->mc_alloc, alloc_delta);
444	atomic_add_64(&mc->mc_deferred, defer_delta);
445	atomic_add_64(&mc->mc_space, space_delta);
446	atomic_add_64(&mc->mc_dspace, dspace_delta);
447}
448
449uint64_t
450metaslab_class_get_alloc(metaslab_class_t *mc)
451{
452	return (mc->mc_alloc);
453}
454
455uint64_t
456metaslab_class_get_deferred(metaslab_class_t *mc)
457{
458	return (mc->mc_deferred);
459}
460
461uint64_t
462metaslab_class_get_space(metaslab_class_t *mc)
463{
464	return (mc->mc_space);
465}
466
467uint64_t
468metaslab_class_get_dspace(metaslab_class_t *mc)
469{
470	return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
471}
472
473void
474metaslab_class_histogram_verify(metaslab_class_t *mc)
475{
476	spa_t *spa = mc->mc_spa;
477	vdev_t *rvd = spa->spa_root_vdev;
478	uint64_t *mc_hist;
479	int i;
480
481	if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
482		return;
483
484	mc_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
485	    KM_SLEEP);
486
487	for (int c = 0; c < rvd->vdev_children; c++) {
488		vdev_t *tvd = rvd->vdev_child[c];
489		metaslab_group_t *mg = tvd->vdev_mg;
490
491		/*
492		 * Skip any holes, uninitialized top-levels, or
493		 * vdevs that are not in this metalab class.
494		 */
495		if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 ||
496		    mg->mg_class != mc) {
497			continue;
498		}
499
500		for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
501			mc_hist[i] += mg->mg_histogram[i];
502	}
503
504	for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
505		VERIFY3U(mc_hist[i], ==, mc->mc_histogram[i]);
506
507	kmem_free(mc_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
508}
509
510/*
511 * Calculate the metaslab class's fragmentation metric. The metric
512 * is weighted based on the space contribution of each metaslab group.
513 * The return value will be a number between 0 and 100 (inclusive), or
514 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
515 * zfs_frag_table for more information about the metric.
516 */
517uint64_t
518metaslab_class_fragmentation(metaslab_class_t *mc)
519{
520	vdev_t *rvd = mc->mc_spa->spa_root_vdev;
521	uint64_t fragmentation = 0;
522
523	spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
524
525	for (int c = 0; c < rvd->vdev_children; c++) {
526		vdev_t *tvd = rvd->vdev_child[c];
527		metaslab_group_t *mg = tvd->vdev_mg;
528
529		/*
530		 * Skip any holes, uninitialized top-levels,
531		 * or vdevs that are not in this metalab class.
532		 */
533		if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 ||
534		    mg->mg_class != mc) {
535			continue;
536		}
537
538		/*
539		 * If a metaslab group does not contain a fragmentation
540		 * metric then just bail out.
541		 */
542		if (mg->mg_fragmentation == ZFS_FRAG_INVALID) {
543			spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
544			return (ZFS_FRAG_INVALID);
545		}
546
547		/*
548		 * Determine how much this metaslab_group is contributing
549		 * to the overall pool fragmentation metric.
550		 */
551		fragmentation += mg->mg_fragmentation *
552		    metaslab_group_get_space(mg);
553	}
554	fragmentation /= metaslab_class_get_space(mc);
555
556	ASSERT3U(fragmentation, <=, 100);
557	spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
558	return (fragmentation);
559}
560
561/*
562 * Calculate the amount of expandable space that is available in
563 * this metaslab class. If a device is expanded then its expandable
564 * space will be the amount of allocatable space that is currently not
565 * part of this metaslab class.
566 */
567uint64_t
568metaslab_class_expandable_space(metaslab_class_t *mc)
569{
570	vdev_t *rvd = mc->mc_spa->spa_root_vdev;
571	uint64_t space = 0;
572
573	spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
574	for (int c = 0; c < rvd->vdev_children; c++) {
575		uint64_t tspace;
576		vdev_t *tvd = rvd->vdev_child[c];
577		metaslab_group_t *mg = tvd->vdev_mg;
578
579		if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 ||
580		    mg->mg_class != mc) {
581			continue;
582		}
583
584		/*
585		 * Calculate if we have enough space to add additional
586		 * metaslabs. We report the expandable space in terms
587		 * of the metaslab size since that's the unit of expansion.
588		 * Adjust by efi system partition size.
589		 */
590		tspace = tvd->vdev_max_asize - tvd->vdev_asize;
591		if (tspace > mc->mc_spa->spa_bootsize) {
592			tspace -= mc->mc_spa->spa_bootsize;
593		}
594		space += P2ALIGN(tspace, 1ULL << tvd->vdev_ms_shift);
595	}
596	spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
597	return (space);
598}
599
600void
601metaslab_class_evict_old(metaslab_class_t *mc, uint64_t txg)
602{
603	multilist_t *ml = mc->mc_metaslab_txg_list;
604	for (int i = 0; i < multilist_get_num_sublists(ml); i++) {
605		multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
606		metaslab_t *msp = multilist_sublist_head(mls);
607		multilist_sublist_unlock(mls);
608		while (msp != NULL) {
609			mutex_enter(&msp->ms_lock);
610
611			/*
612			 * If the metaslab has been removed from the list
613			 * (which could happen if we were at the memory limit
614			 * and it was evicted during this loop), then we can't
615			 * proceed and we should restart the sublist.
616			 */
617			if (!multilist_link_active(&msp->ms_class_txg_node)) {
618				mutex_exit(&msp->ms_lock);
619				i--;
620				break;
621			}
622			mls = multilist_sublist_lock(ml, i);
623			metaslab_t *next_msp = multilist_sublist_next(mls, msp);
624			multilist_sublist_unlock(mls);
625			if (txg >
626			    msp->ms_selected_txg + metaslab_unload_delay &&
627			    gethrtime() > msp->ms_selected_time +
628			    (uint64_t)MSEC2NSEC(metaslab_unload_delay_ms)) {
629				metaslab_evict(msp, txg);
630			} else {
631				/*
632				 * Once we've hit a metaslab selected too
633				 * recently to evict, we're done evicting for
634				 * now.
635				 */
636				mutex_exit(&msp->ms_lock);
637				break;
638			}
639			mutex_exit(&msp->ms_lock);
640			msp = next_msp;
641		}
642	}
643}
644
645static int
646metaslab_compare(const void *x1, const void *x2)
647{
648	const metaslab_t *m1 = (const metaslab_t *)x1;
649	const metaslab_t *m2 = (const metaslab_t *)x2;
650
651	int sort1 = 0;
652	int sort2 = 0;
653	if (m1->ms_allocator != -1 && m1->ms_primary)
654		sort1 = 1;
655	else if (m1->ms_allocator != -1 && !m1->ms_primary)
656		sort1 = 2;
657	if (m2->ms_allocator != -1 && m2->ms_primary)
658		sort2 = 1;
659	else if (m2->ms_allocator != -1 && !m2->ms_primary)
660		sort2 = 2;
661
662	/*
663	 * Sort inactive metaslabs first, then primaries, then secondaries. When
664	 * selecting a metaslab to allocate from, an allocator first tries its
665	 * primary, then secondary active metaslab. If it doesn't have active
666	 * metaslabs, or can't allocate from them, it searches for an inactive
667	 * metaslab to activate. If it can't find a suitable one, it will steal
668	 * a primary or secondary metaslab from another allocator.
669	 */
670	if (sort1 < sort2)
671		return (-1);
672	if (sort1 > sort2)
673		return (1);
674
675	int cmp = TREE_CMP(m2->ms_weight, m1->ms_weight);
676	if (likely(cmp))
677		return (cmp);
678
679	IMPLY(TREE_CMP(m1->ms_start, m2->ms_start) == 0, m1 == m2);
680
681	return (TREE_CMP(m1->ms_start, m2->ms_start));
682}
683
684/*
685 * ==========================================================================
686 * Metaslab groups
687 * ==========================================================================
688 */
689/*
690 * Update the allocatable flag and the metaslab group's capacity.
691 * The allocatable flag is set to true if the capacity is below
692 * the zfs_mg_noalloc_threshold or has a fragmentation value that is
693 * greater than zfs_mg_fragmentation_threshold. If a metaslab group
694 * transitions from allocatable to non-allocatable or vice versa then the
695 * metaslab group's class is updated to reflect the transition.
696 */
697static void
698metaslab_group_alloc_update(metaslab_group_t *mg)
699{
700	vdev_t *vd = mg->mg_vd;
701	metaslab_class_t *mc = mg->mg_class;
702	vdev_stat_t *vs = &vd->vdev_stat;
703	boolean_t was_allocatable;
704	boolean_t was_initialized;
705
706	ASSERT(vd == vd->vdev_top);
707	ASSERT3U(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_READER), ==,
708	    SCL_ALLOC);
709
710	mutex_enter(&mg->mg_lock);
711	was_allocatable = mg->mg_allocatable;
712	was_initialized = mg->mg_initialized;
713
714	mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
715	    (vs->vs_space + 1);
716
717	mutex_enter(&mc->mc_lock);
718
719	/*
720	 * If the metaslab group was just added then it won't
721	 * have any space until we finish syncing out this txg.
722	 * At that point we will consider it initialized and available
723	 * for allocations.  We also don't consider non-activated
724	 * metaslab groups (e.g. vdevs that are in the middle of being removed)
725	 * to be initialized, because they can't be used for allocation.
726	 */
727	mg->mg_initialized = metaslab_group_initialized(mg);
728	if (!was_initialized && mg->mg_initialized) {
729		mc->mc_groups++;
730	} else if (was_initialized && !mg->mg_initialized) {
731		ASSERT3U(mc->mc_groups, >, 0);
732		mc->mc_groups--;
733	}
734	if (mg->mg_initialized)
735		mg->mg_no_free_space = B_FALSE;
736
737	/*
738	 * A metaslab group is considered allocatable if it has plenty
739	 * of free space or is not heavily fragmented. We only take
740	 * fragmentation into account if the metaslab group has a valid
741	 * fragmentation metric (i.e. a value between 0 and 100).
742	 */
743	mg->mg_allocatable = (mg->mg_activation_count > 0 &&
744	    mg->mg_free_capacity > zfs_mg_noalloc_threshold &&
745	    (mg->mg_fragmentation == ZFS_FRAG_INVALID ||
746	    mg->mg_fragmentation <= zfs_mg_fragmentation_threshold));
747
748	/*
749	 * The mc_alloc_groups maintains a count of the number of
750	 * groups in this metaslab class that are still above the
751	 * zfs_mg_noalloc_threshold. This is used by the allocating
752	 * threads to determine if they should avoid allocations to
753	 * a given group. The allocator will avoid allocations to a group
754	 * if that group has reached or is below the zfs_mg_noalloc_threshold
755	 * and there are still other groups that are above the threshold.
756	 * When a group transitions from allocatable to non-allocatable or
757	 * vice versa we update the metaslab class to reflect that change.
758	 * When the mc_alloc_groups value drops to 0 that means that all
759	 * groups have reached the zfs_mg_noalloc_threshold making all groups
760	 * eligible for allocations. This effectively means that all devices
761	 * are balanced again.
762	 */
763	if (was_allocatable && !mg->mg_allocatable)
764		mc->mc_alloc_groups--;
765	else if (!was_allocatable && mg->mg_allocatable)
766		mc->mc_alloc_groups++;
767	mutex_exit(&mc->mc_lock);
768
769	mutex_exit(&mg->mg_lock);
770}
771
772int
773metaslab_sort_by_flushed(const void *va, const void *vb)
774{
775	const metaslab_t *a = va;
776	const metaslab_t *b = vb;
777
778	int cmp = TREE_CMP(a->ms_unflushed_txg, b->ms_unflushed_txg);
779	if (likely(cmp))
780		return (cmp);
781
782	uint64_t a_vdev_id = a->ms_group->mg_vd->vdev_id;
783	uint64_t b_vdev_id = b->ms_group->mg_vd->vdev_id;
784	cmp = TREE_CMP(a_vdev_id, b_vdev_id);
785	if (cmp)
786		return (cmp);
787
788	return (TREE_CMP(a->ms_id, b->ms_id));
789}
790
791metaslab_group_t *
792metaslab_group_create(metaslab_class_t *mc, vdev_t *vd, int allocators)
793{
794	metaslab_group_t *mg;
795
796	mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
797	mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
798	mutex_init(&mg->mg_ms_disabled_lock, NULL, MUTEX_DEFAULT, NULL);
799	cv_init(&mg->mg_ms_disabled_cv, NULL, CV_DEFAULT, NULL);
800	mg->mg_primaries = kmem_zalloc(allocators * sizeof (metaslab_t *),
801	    KM_SLEEP);
802	mg->mg_secondaries = kmem_zalloc(allocators * sizeof (metaslab_t *),
803	    KM_SLEEP);
804	avl_create(&mg->mg_metaslab_tree, metaslab_compare,
805	    sizeof (metaslab_t), offsetof(metaslab_t, ms_group_node));
806	mg->mg_vd = vd;
807	mg->mg_class = mc;
808	mg->mg_activation_count = 0;
809	mg->mg_initialized = B_FALSE;
810	mg->mg_no_free_space = B_TRUE;
811	mg->mg_allocators = allocators;
812
813	mg->mg_alloc_queue_depth = kmem_zalloc(allocators *
814	    sizeof (zfs_refcount_t), KM_SLEEP);
815	mg->mg_cur_max_alloc_queue_depth = kmem_zalloc(allocators *
816	    sizeof (uint64_t), KM_SLEEP);
817	for (int i = 0; i < allocators; i++) {
818		zfs_refcount_create_tracked(&mg->mg_alloc_queue_depth[i]);
819		mg->mg_cur_max_alloc_queue_depth[i] = 0;
820	}
821
822	mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct,
823	    minclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT);
824
825	return (mg);
826}
827
828void
829metaslab_group_destroy(metaslab_group_t *mg)
830{
831	ASSERT(mg->mg_prev == NULL);
832	ASSERT(mg->mg_next == NULL);
833	/*
834	 * We may have gone below zero with the activation count
835	 * either because we never activated in the first place or
836	 * because we're done, and possibly removing the vdev.
837	 */
838	ASSERT(mg->mg_activation_count <= 0);
839
840	taskq_destroy(mg->mg_taskq);
841	avl_destroy(&mg->mg_metaslab_tree);
842	kmem_free(mg->mg_primaries, mg->mg_allocators * sizeof (metaslab_t *));
843	kmem_free(mg->mg_secondaries, mg->mg_allocators *
844	    sizeof (metaslab_t *));
845	mutex_destroy(&mg->mg_lock);
846	mutex_destroy(&mg->mg_ms_disabled_lock);
847	cv_destroy(&mg->mg_ms_disabled_cv);
848
849	for (int i = 0; i < mg->mg_allocators; i++) {
850		zfs_refcount_destroy(&mg->mg_alloc_queue_depth[i]);
851		mg->mg_cur_max_alloc_queue_depth[i] = 0;
852	}
853	kmem_free(mg->mg_alloc_queue_depth, mg->mg_allocators *
854	    sizeof (zfs_refcount_t));
855	kmem_free(mg->mg_cur_max_alloc_queue_depth, mg->mg_allocators *
856	    sizeof (uint64_t));
857
858	kmem_free(mg, sizeof (metaslab_group_t));
859}
860
861void
862metaslab_group_activate(metaslab_group_t *mg)
863{
864	metaslab_class_t *mc = mg->mg_class;
865	metaslab_group_t *mgprev, *mgnext;
866
867	ASSERT3U(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER), !=, 0);
868
869	ASSERT(mc->mc_rotor != mg);
870	ASSERT(mg->mg_prev == NULL);
871	ASSERT(mg->mg_next == NULL);
872	ASSERT(mg->mg_activation_count <= 0);
873
874	if (++mg->mg_activation_count <= 0)
875		return;
876
877	mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
878	metaslab_group_alloc_update(mg);
879
880	if ((mgprev = mc->mc_rotor) == NULL) {
881		mg->mg_prev = mg;
882		mg->mg_next = mg;
883	} else {
884		mgnext = mgprev->mg_next;
885		mg->mg_prev = mgprev;
886		mg->mg_next = mgnext;
887		mgprev->mg_next = mg;
888		mgnext->mg_prev = mg;
889	}
890	mc->mc_rotor = mg;
891}
892
893/*
894 * Passivate a metaslab group and remove it from the allocation rotor.
895 * Callers must hold both the SCL_ALLOC and SCL_ZIO lock prior to passivating
896 * a metaslab group. This function will momentarily drop spa_config_locks
897 * that are lower than the SCL_ALLOC lock (see comment below).
898 */
899void
900metaslab_group_passivate(metaslab_group_t *mg)
901{
902	metaslab_class_t *mc = mg->mg_class;
903	spa_t *spa = mc->mc_spa;
904	metaslab_group_t *mgprev, *mgnext;
905	int locks = spa_config_held(spa, SCL_ALL, RW_WRITER);
906
907	ASSERT3U(spa_config_held(spa, SCL_ALLOC | SCL_ZIO, RW_WRITER), ==,
908	    (SCL_ALLOC | SCL_ZIO));
909
910	if (--mg->mg_activation_count != 0) {
911		ASSERT(mc->mc_rotor != mg);
912		ASSERT(mg->mg_prev == NULL);
913		ASSERT(mg->mg_next == NULL);
914		ASSERT(mg->mg_activation_count < 0);
915		return;
916	}
917
918	/*
919	 * The spa_config_lock is an array of rwlocks, ordered as
920	 * follows (from highest to lowest):
921	 *	SCL_CONFIG > SCL_STATE > SCL_L2ARC > SCL_ALLOC >
922	 *	SCL_ZIO > SCL_FREE > SCL_VDEV
923	 * (For more information about the spa_config_lock see spa_misc.c)
924	 * The higher the lock, the broader its coverage. When we passivate
925	 * a metaslab group, we must hold both the SCL_ALLOC and the SCL_ZIO
926	 * config locks. However, the metaslab group's taskq might be trying
927	 * to preload metaslabs so we must drop the SCL_ZIO lock and any
928	 * lower locks to allow the I/O to complete. At a minimum,
929	 * we continue to hold the SCL_ALLOC lock, which prevents any future
930	 * allocations from taking place and any changes to the vdev tree.
931	 */
932	spa_config_exit(spa, locks & ~(SCL_ZIO - 1), spa);
933	taskq_wait(mg->mg_taskq);
934	spa_config_enter(spa, locks & ~(SCL_ZIO - 1), spa, RW_WRITER);
935	metaslab_group_alloc_update(mg);
936	for (int i = 0; i < mg->mg_allocators; i++) {
937		metaslab_t *msp = mg->mg_primaries[i];
938		if (msp != NULL) {
939			mutex_enter(&msp->ms_lock);
940			metaslab_passivate(msp,
941			    metaslab_weight_from_range_tree(msp));
942			mutex_exit(&msp->ms_lock);
943		}
944		msp = mg->mg_secondaries[i];
945		if (msp != NULL) {
946			mutex_enter(&msp->ms_lock);
947			metaslab_passivate(msp,
948			    metaslab_weight_from_range_tree(msp));
949			mutex_exit(&msp->ms_lock);
950		}
951	}
952
953	mgprev = mg->mg_prev;
954	mgnext = mg->mg_next;
955
956	if (mg == mgnext) {
957		mc->mc_rotor = NULL;
958	} else {
959		mc->mc_rotor = mgnext;
960		mgprev->mg_next = mgnext;
961		mgnext->mg_prev = mgprev;
962	}
963
964	mg->mg_prev = NULL;
965	mg->mg_next = NULL;
966}
967
968boolean_t
969metaslab_group_initialized(metaslab_group_t *mg)
970{
971	vdev_t *vd = mg->mg_vd;
972	vdev_stat_t *vs = &vd->vdev_stat;
973
974	return (vs->vs_space != 0 && mg->mg_activation_count > 0);
975}
976
977uint64_t
978metaslab_group_get_space(metaslab_group_t *mg)
979{
980	return ((1ULL << mg->mg_vd->vdev_ms_shift) * mg->mg_vd->vdev_ms_count);
981}
982
983void
984metaslab_group_histogram_verify(metaslab_group_t *mg)
985{
986	uint64_t *mg_hist;
987	vdev_t *vd = mg->mg_vd;
988	uint64_t ashift = vd->vdev_ashift;
989	int i;
990
991	if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
992		return;
993
994	mg_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
995	    KM_SLEEP);
996
997	ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE, >=,
998	    SPACE_MAP_HISTOGRAM_SIZE + ashift);
999
1000	for (int m = 0; m < vd->vdev_ms_count; m++) {
1001		metaslab_t *msp = vd->vdev_ms[m];
1002
1003		/* skip if not active or not a member */
1004		if (msp->ms_sm == NULL || msp->ms_group != mg)
1005			continue;
1006
1007		for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++)
1008			mg_hist[i + ashift] +=
1009			    msp->ms_sm->sm_phys->smp_histogram[i];
1010	}
1011
1012	for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i ++)
1013		VERIFY3U(mg_hist[i], ==, mg->mg_histogram[i]);
1014
1015	kmem_free(mg_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
1016}
1017
1018static void
1019metaslab_group_histogram_add(metaslab_group_t *mg, metaslab_t *msp)
1020{
1021	metaslab_class_t *mc = mg->mg_class;
1022	uint64_t ashift = mg->mg_vd->vdev_ashift;
1023
1024	ASSERT(MUTEX_HELD(&msp->ms_lock));
1025	if (msp->ms_sm == NULL)
1026		return;
1027
1028	mutex_enter(&mg->mg_lock);
1029	for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
1030		mg->mg_histogram[i + ashift] +=
1031		    msp->ms_sm->sm_phys->smp_histogram[i];
1032		mc->mc_histogram[i + ashift] +=
1033		    msp->ms_sm->sm_phys->smp_histogram[i];
1034	}
1035	mutex_exit(&mg->mg_lock);
1036}
1037
1038void
1039metaslab_group_histogram_remove(metaslab_group_t *mg, metaslab_t *msp)
1040{
1041	metaslab_class_t *mc = mg->mg_class;
1042	uint64_t ashift = mg->mg_vd->vdev_ashift;
1043
1044	ASSERT(MUTEX_HELD(&msp->ms_lock));
1045	if (msp->ms_sm == NULL)
1046		return;
1047
1048	mutex_enter(&mg->mg_lock);
1049	for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
1050		ASSERT3U(mg->mg_histogram[i + ashift], >=,
1051		    msp->ms_sm->sm_phys->smp_histogram[i]);
1052		ASSERT3U(mc->mc_histogram[i + ashift], >=,
1053		    msp->ms_sm->sm_phys->smp_histogram[i]);
1054
1055		mg->mg_histogram[i + ashift] -=
1056		    msp->ms_sm->sm_phys->smp_histogram[i];
1057		mc->mc_histogram[i + ashift] -=
1058		    msp->ms_sm->sm_phys->smp_histogram[i];
1059	}
1060	mutex_exit(&mg->mg_lock);
1061}
1062
1063static void
1064metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
1065{
1066	ASSERT(msp->ms_group == NULL);
1067	mutex_enter(&mg->mg_lock);
1068	msp->ms_group = mg;
1069	msp->ms_weight = 0;
1070	avl_add(&mg->mg_metaslab_tree, msp);
1071	mutex_exit(&mg->mg_lock);
1072
1073	mutex_enter(&msp->ms_lock);
1074	metaslab_group_histogram_add(mg, msp);
1075	mutex_exit(&msp->ms_lock);
1076}
1077
1078static void
1079metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
1080{
1081	mutex_enter(&msp->ms_lock);
1082	metaslab_group_histogram_remove(mg, msp);
1083	mutex_exit(&msp->ms_lock);
1084
1085	mutex_enter(&mg->mg_lock);
1086	ASSERT(msp->ms_group == mg);
1087	avl_remove(&mg->mg_metaslab_tree, msp);
1088
1089	metaslab_class_t *mc = msp->ms_group->mg_class;
1090	multilist_sublist_t *mls =
1091	    multilist_sublist_lock_obj(mc->mc_metaslab_txg_list, msp);
1092	if (multilist_link_active(&msp->ms_class_txg_node))
1093		multilist_sublist_remove(mls, msp);
1094	multilist_sublist_unlock(mls);
1095
1096	msp->ms_group = NULL;
1097	mutex_exit(&mg->mg_lock);
1098}
1099
1100static void
1101metaslab_group_sort_impl(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
1102{
1103	ASSERT(MUTEX_HELD(&msp->ms_lock));
1104	ASSERT(MUTEX_HELD(&mg->mg_lock));
1105	ASSERT(msp->ms_group == mg);
1106
1107	avl_remove(&mg->mg_metaslab_tree, msp);
1108	msp->ms_weight = weight;
1109	avl_add(&mg->mg_metaslab_tree, msp);
1110
1111}
1112
1113static void
1114metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
1115{
1116	/*
1117	 * Although in principle the weight can be any value, in
1118	 * practice we do not use values in the range [1, 511].
1119	 */
1120	ASSERT(weight >= SPA_MINBLOCKSIZE || weight == 0);
1121	ASSERT(MUTEX_HELD(&msp->ms_lock));
1122
1123	mutex_enter(&mg->mg_lock);
1124	metaslab_group_sort_impl(mg, msp, weight);
1125	mutex_exit(&mg->mg_lock);
1126}
1127
1128/*
1129 * Calculate the fragmentation for a given metaslab group. We can use
1130 * a simple average here since all metaslabs within the group must have
1131 * the same size. The return value will be a value between 0 and 100
1132 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
1133 * group have a fragmentation metric.
1134 */
1135uint64_t
1136metaslab_group_fragmentation(metaslab_group_t *mg)
1137{
1138	vdev_t *vd = mg->mg_vd;
1139	uint64_t fragmentation = 0;
1140	uint64_t valid_ms = 0;
1141
1142	for (int m = 0; m < vd->vdev_ms_count; m++) {
1143		metaslab_t *msp = vd->vdev_ms[m];
1144
1145		if (msp->ms_fragmentation == ZFS_FRAG_INVALID)
1146			continue;
1147		if (msp->ms_group != mg)
1148			continue;
1149
1150		valid_ms++;
1151		fragmentation += msp->ms_fragmentation;
1152	}
1153
1154	if (valid_ms <= mg->mg_vd->vdev_ms_count / 2)
1155		return (ZFS_FRAG_INVALID);
1156
1157	fragmentation /= valid_ms;
1158	ASSERT3U(fragmentation, <=, 100);
1159	return (fragmentation);
1160}
1161
1162/*
1163 * Determine if a given metaslab group should skip allocations. A metaslab
1164 * group should avoid allocations if its free capacity is less than the
1165 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
1166 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
1167 * that can still handle allocations. If the allocation throttle is enabled
1168 * then we skip allocations to devices that have reached their maximum
1169 * allocation queue depth unless the selected metaslab group is the only
1170 * eligible group remaining.
1171 */
1172static boolean_t
1173metaslab_group_allocatable(metaslab_group_t *mg, metaslab_group_t *rotor,
1174    uint64_t psize, int allocator, int d)
1175{
1176	spa_t *spa = mg->mg_vd->vdev_spa;
1177	metaslab_class_t *mc = mg->mg_class;
1178
1179	/*
1180	 * We can only consider skipping this metaslab group if it's
1181	 * in the normal metaslab class and there are other metaslab
1182	 * groups to select from. Otherwise, we always consider it eligible
1183	 * for allocations.
1184	 */
1185	if ((mc != spa_normal_class(spa) &&
1186	    mc != spa_special_class(spa) &&
1187	    mc != spa_dedup_class(spa)) ||
1188	    mc->mc_groups <= 1)
1189		return (B_TRUE);
1190
1191	/*
1192	 * If the metaslab group's mg_allocatable flag is set (see comments
1193	 * in metaslab_group_alloc_update() for more information) and
1194	 * the allocation throttle is disabled then allow allocations to this
1195	 * device. However, if the allocation throttle is enabled then
1196	 * check if we have reached our allocation limit (mg_alloc_queue_depth)
1197	 * to determine if we should allow allocations to this metaslab group.
1198	 * If all metaslab groups are no longer considered allocatable
1199	 * (mc_alloc_groups == 0) or we're trying to allocate the smallest
1200	 * gang block size then we allow allocations on this metaslab group
1201	 * regardless of the mg_allocatable or throttle settings.
1202	 */
1203	if (mg->mg_allocatable) {
1204		metaslab_group_t *mgp;
1205		int64_t qdepth;
1206		uint64_t qmax = mg->mg_cur_max_alloc_queue_depth[allocator];
1207
1208		if (!mc->mc_alloc_throttle_enabled)
1209			return (B_TRUE);
1210
1211		/*
1212		 * If this metaslab group does not have any free space, then
1213		 * there is no point in looking further.
1214		 */
1215		if (mg->mg_no_free_space)
1216			return (B_FALSE);
1217
1218		/*
1219		 * Relax allocation throttling for ditto blocks.  Due to
1220		 * random imbalances in allocation it tends to push copies
1221		 * to one vdev, that looks a bit better at the moment.
1222		 */
1223		qmax = qmax * (4 + d) / 4;
1224
1225		qdepth = zfs_refcount_count(
1226		    &mg->mg_alloc_queue_depth[allocator]);
1227
1228		/*
1229		 * If this metaslab group is below its qmax or it's
1230		 * the only allocatable metasable group, then attempt
1231		 * to allocate from it.
1232		 */
1233		if (qdepth < qmax || mc->mc_alloc_groups == 1)
1234			return (B_TRUE);
1235		ASSERT3U(mc->mc_alloc_groups, >, 1);
1236
1237		/*
1238		 * Since this metaslab group is at or over its qmax, we
1239		 * need to determine if there are metaslab groups after this
1240		 * one that might be able to handle this allocation. This is
1241		 * racy since we can't hold the locks for all metaslab
1242		 * groups at the same time when we make this check.
1243		 */
1244		for (mgp = mg->mg_next; mgp != rotor; mgp = mgp->mg_next) {
1245			qmax = mgp->mg_cur_max_alloc_queue_depth[allocator];
1246			qmax = qmax * (4 + d) / 4;
1247			qdepth = zfs_refcount_count(
1248			    &mgp->mg_alloc_queue_depth[allocator]);
1249
1250			/*
1251			 * If there is another metaslab group that
1252			 * might be able to handle the allocation, then
1253			 * we return false so that we skip this group.
1254			 */
1255			if (qdepth < qmax && !mgp->mg_no_free_space)
1256				return (B_FALSE);
1257		}
1258
1259		/*
1260		 * We didn't find another group to handle the allocation
1261		 * so we can't skip this metaslab group even though
1262		 * we are at or over our qmax.
1263		 */
1264		return (B_TRUE);
1265
1266	} else if (mc->mc_alloc_groups == 0 || psize == SPA_MINBLOCKSIZE) {
1267		return (B_TRUE);
1268	}
1269	return (B_FALSE);
1270}
1271
1272/*
1273 * ==========================================================================
1274 * Range tree callbacks
1275 * ==========================================================================
1276 */
1277
1278/*
1279 * Comparison function for the private size-ordered tree using 32-bit
1280 * ranges. Tree is sorted by size, larger sizes at the end of the tree.
1281 */
1282static int
1283metaslab_rangesize32_compare(const void *x1, const void *x2)
1284{
1285	const range_seg32_t *r1 = x1;
1286	const range_seg32_t *r2 = x2;
1287
1288	uint64_t rs_size1 = r1->rs_end - r1->rs_start;
1289	uint64_t rs_size2 = r2->rs_end - r2->rs_start;
1290
1291	int cmp = TREE_CMP(rs_size1, rs_size2);
1292	if (likely(cmp))
1293		return (cmp);
1294
1295	return (TREE_CMP(r1->rs_start, r2->rs_start));
1296}
1297
1298/*
1299 * Comparison function for the private size-ordered tree using 64-bit
1300 * ranges. Tree is sorted by size, larger sizes at the end of the tree.
1301 */
1302static int
1303metaslab_rangesize64_compare(const void *x1, const void *x2)
1304{
1305	const range_seg64_t *r1 = x1;
1306	const range_seg64_t *r2 = x2;
1307
1308	uint64_t rs_size1 = r1->rs_end - r1->rs_start;
1309	uint64_t rs_size2 = r2->rs_end - r2->rs_start;
1310
1311	int cmp = TREE_CMP(rs_size1, rs_size2);
1312	if (likely(cmp))
1313		return (cmp);
1314
1315	return (TREE_CMP(r1->rs_start, r2->rs_start));
1316}
1317typedef struct metaslab_rt_arg {
1318	zfs_btree_t *mra_bt;
1319	uint32_t mra_floor_shift;
1320} metaslab_rt_arg_t;
1321
1322struct mssa_arg {
1323	range_tree_t *rt;
1324	metaslab_rt_arg_t *mra;
1325};
1326
1327static void
1328metaslab_size_sorted_add(void *arg, uint64_t start, uint64_t size)
1329{
1330	struct mssa_arg *mssap = arg;
1331	range_tree_t *rt = mssap->rt;
1332	metaslab_rt_arg_t *mrap = mssap->mra;
1333	range_seg_max_t seg = {0};
1334	rs_set_start(&seg, rt, start);
1335	rs_set_end(&seg, rt, start + size);
1336	metaslab_rt_add(rt, &seg, mrap);
1337}
1338
1339static void
1340metaslab_size_tree_full_load(range_tree_t *rt)
1341{
1342	metaslab_rt_arg_t *mrap = rt->rt_arg;
1343#ifdef _METASLAB_TRACING
1344	METASLABSTAT_BUMP(metaslabstat_reload_tree);
1345#endif
1346	ASSERT0(zfs_btree_numnodes(mrap->mra_bt));
1347	mrap->mra_floor_shift = 0;
1348	struct mssa_arg arg = {0};
1349	arg.rt = rt;
1350	arg.mra = mrap;
1351	range_tree_walk(rt, metaslab_size_sorted_add, &arg);
1352}
1353
1354/*
1355 * Create any block allocator specific components. The current allocators
1356 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
1357 */
1358/* ARGSUSED */
1359static void
1360metaslab_rt_create(range_tree_t *rt, void *arg)
1361{
1362	metaslab_rt_arg_t *mrap = arg;
1363	zfs_btree_t *size_tree = mrap->mra_bt;
1364
1365	size_t size;
1366	int (*compare) (const void *, const void *);
1367	switch (rt->rt_type) {
1368	case RANGE_SEG32:
1369		size = sizeof (range_seg32_t);
1370		compare = metaslab_rangesize32_compare;
1371		break;
1372	case RANGE_SEG64:
1373		size = sizeof (range_seg64_t);
1374		compare = metaslab_rangesize64_compare;
1375		break;
1376	default:
1377		panic("Invalid range seg type %d", rt->rt_type);
1378	}
1379	zfs_btree_create(size_tree, compare, size);
1380	mrap->mra_floor_shift = metaslab_by_size_min_shift;
1381}
1382
1383/* ARGSUSED */
1384static void
1385metaslab_rt_destroy(range_tree_t *rt, void *arg)
1386{
1387	metaslab_rt_arg_t *mrap = arg;
1388	zfs_btree_t *size_tree = mrap->mra_bt;
1389
1390	zfs_btree_destroy(size_tree);
1391	kmem_free(mrap, sizeof (*mrap));
1392}
1393
1394/* ARGSUSED */
1395static void
1396metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg)
1397{
1398	metaslab_rt_arg_t *mrap = arg;
1399	zfs_btree_t *size_tree = mrap->mra_bt;
1400
1401	if (rs_get_end(rs, rt) - rs_get_start(rs, rt) <
1402	    (1 << mrap->mra_floor_shift))
1403		return;
1404
1405	zfs_btree_add(size_tree, rs);
1406}
1407
1408/* ARGSUSED */
1409static void
1410metaslab_rt_remove(range_tree_t *rt, range_seg_t *rs, void *arg)
1411{
1412	metaslab_rt_arg_t *mrap = arg;
1413	zfs_btree_t *size_tree = mrap->mra_bt;
1414
1415	if (rs_get_end(rs, rt) - rs_get_start(rs, rt) < (1 <<
1416	    mrap->mra_floor_shift))
1417		return;
1418
1419	zfs_btree_remove(size_tree, rs);
1420}
1421
1422/* ARGSUSED */
1423static void
1424metaslab_rt_vacate(range_tree_t *rt, void *arg)
1425{
1426	metaslab_rt_arg_t *mrap = arg;
1427	zfs_btree_t *size_tree = mrap->mra_bt;
1428	zfs_btree_clear(size_tree);
1429	zfs_btree_destroy(size_tree);
1430
1431	metaslab_rt_create(rt, arg);
1432}
1433
1434static range_tree_ops_t metaslab_rt_ops = {
1435	.rtop_create = metaslab_rt_create,
1436	.rtop_destroy = metaslab_rt_destroy,
1437	.rtop_add = metaslab_rt_add,
1438	.rtop_remove = metaslab_rt_remove,
1439	.rtop_vacate = metaslab_rt_vacate
1440};
1441
1442/*
1443 * ==========================================================================
1444 * Common allocator routines
1445 * ==========================================================================
1446 */
1447
1448/*
1449 * Return the maximum contiguous segment within the metaslab.
1450 */
1451uint64_t
1452metaslab_largest_allocatable(metaslab_t *msp)
1453{
1454	zfs_btree_t *t = &msp->ms_allocatable_by_size;
1455	range_seg_t *rs;
1456
1457	if (t == NULL)
1458		return (0);
1459	if (zfs_btree_numnodes(t) == 0)
1460		metaslab_size_tree_full_load(msp->ms_allocatable);
1461
1462	rs = zfs_btree_last(t, NULL);
1463	if (rs == NULL)
1464		return (0);
1465
1466	return (rs_get_end(rs, msp->ms_allocatable) - rs_get_start(rs,
1467	    msp->ms_allocatable));
1468}
1469
1470/*
1471 * Return the maximum contiguous segment within the unflushed frees of this
1472 * metaslab.
1473 */
1474uint64_t
1475metaslab_largest_unflushed_free(metaslab_t *msp)
1476{
1477	ASSERT(MUTEX_HELD(&msp->ms_lock));
1478
1479	if (msp->ms_unflushed_frees == NULL)
1480		return (0);
1481
1482	if (zfs_btree_numnodes(&msp->ms_unflushed_frees_by_size) == 0)
1483		metaslab_size_tree_full_load(msp->ms_unflushed_frees);
1484	range_seg_t *rs = zfs_btree_last(&msp->ms_unflushed_frees_by_size,
1485	    NULL);
1486	if (rs == NULL)
1487		return (0);
1488
1489	/*
1490	 * When a range is freed from the metaslab, that range is added to
1491	 * both the unflushed frees and the deferred frees. While the block
1492	 * will eventually be usable, if the metaslab were loaded the range
1493	 * would not be added to the ms_allocatable tree until TXG_DEFER_SIZE
1494	 * txgs had passed.  As a result, when attempting to estimate an upper
1495	 * bound for the largest currently-usable free segment in the
1496	 * metaslab, we need to not consider any ranges currently in the defer
1497	 * trees. This algorithm approximates the largest available chunk in
1498	 * the largest range in the unflushed_frees tree by taking the first
1499	 * chunk.  While this may be a poor estimate, it should only remain so
1500	 * briefly and should eventually self-correct as frees are no longer
1501	 * deferred. Similar logic applies to the ms_freed tree. See
1502	 * metaslab_load() for more details.
1503	 *
1504	 * There are two primary sources of innacuracy in this estimate. Both
1505	 * are tolerated for performance reasons. The first source is that we
1506	 * only check the largest segment for overlaps. Smaller segments may
1507	 * have more favorable overlaps with the other trees, resulting in
1508	 * larger usable chunks.  Second, we only look at the first chunk in
1509	 * the largest segment; there may be other usable chunks in the
1510	 * largest segment, but we ignore them.
1511	 */
1512	uint64_t rstart = rs_get_start(rs, msp->ms_unflushed_frees);
1513	uint64_t rsize = rs_get_end(rs, msp->ms_unflushed_frees) - rstart;
1514	for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1515		uint64_t start = 0;
1516		uint64_t size = 0;
1517		boolean_t found = range_tree_find_in(msp->ms_defer[t], rstart,
1518		    rsize, &start, &size);
1519		if (found) {
1520			if (rstart == start)
1521				return (0);
1522			rsize = start - rstart;
1523		}
1524	}
1525
1526	uint64_t start = 0;
1527	uint64_t size = 0;
1528	boolean_t found = range_tree_find_in(msp->ms_freed, rstart,
1529	    rsize, &start, &size);
1530	if (found)
1531		rsize = start - rstart;
1532
1533	return (rsize);
1534}
1535
1536static range_seg_t *
1537metaslab_block_find(zfs_btree_t *t, range_tree_t *rt, uint64_t start,
1538    uint64_t size, zfs_btree_index_t *where)
1539{
1540	range_seg_t *rs;
1541	range_seg_max_t rsearch;
1542
1543	rs_set_start(&rsearch, rt, start);
1544	rs_set_end(&rsearch, rt, start + size);
1545
1546	rs = zfs_btree_find(t, &rsearch, where);
1547	if (rs == NULL) {
1548		rs = zfs_btree_next(t, where, where);
1549	}
1550
1551	return (rs);
1552}
1553
1554/*
1555 * This is a helper function that can be used by the allocator to find a
1556 * suitable block to allocate. This will search the specified B-tree looking
1557 * for a block that matches the specified criteria.
1558 */
1559static uint64_t
1560metaslab_block_picker(range_tree_t *rt, uint64_t *cursor, uint64_t size,
1561    uint64_t max_search)
1562{
1563	if (*cursor == 0)
1564		*cursor = rt->rt_start;
1565	zfs_btree_t *bt = &rt->rt_root;
1566	zfs_btree_index_t where;
1567	range_seg_t *rs = metaslab_block_find(bt, rt, *cursor, size, &where);
1568	uint64_t first_found;
1569	int count_searched = 0;
1570
1571	if (rs != NULL)
1572		first_found = rs_get_start(rs, rt);
1573
1574	while (rs != NULL && (rs_get_start(rs, rt) - first_found <=
1575	    max_search || count_searched < metaslab_min_search_count)) {
1576		uint64_t offset = rs_get_start(rs, rt);
1577		if (offset + size <= rs_get_end(rs, rt)) {
1578			*cursor = offset + size;
1579			return (offset);
1580		}
1581		rs = zfs_btree_next(bt, &where, &where);
1582		count_searched++;
1583	}
1584
1585	*cursor = 0;
1586	return (-1ULL);
1587}
1588
1589/*
1590 * ==========================================================================
1591 * Dynamic Fit (df) block allocator
1592 *
1593 * Search for a free chunk of at least this size, starting from the last
1594 * offset (for this alignment of block) looking for up to
1595 * metaslab_df_max_search bytes (16MB).  If a large enough free chunk is not
1596 * found within 16MB, then return a free chunk of exactly the requested size (or
1597 * larger).
1598 *
1599 * If it seems like searching from the last offset will be unproductive, skip
1600 * that and just return a free chunk of exactly the requested size (or larger).
1601 * This is based on metaslab_df_alloc_threshold and metaslab_df_free_pct.  This
1602 * mechanism is probably not very useful and may be removed in the future.
1603 *
1604 * The behavior when not searching can be changed to return the largest free
1605 * chunk, instead of a free chunk of exactly the requested size, by setting
1606 * metaslab_df_use_largest_segment.
1607 * ==========================================================================
1608 */
1609static uint64_t
1610metaslab_df_alloc(metaslab_t *msp, uint64_t size)
1611{
1612	/*
1613	 * Find the largest power of 2 block size that evenly divides the
1614	 * requested size. This is used to try to allocate blocks with similar
1615	 * alignment from the same area of the metaslab (i.e. same cursor
1616	 * bucket) but it does not guarantee that other allocations sizes
1617	 * may exist in the same region.
1618	 */
1619	uint64_t align = size & -size;
1620	uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1621	range_tree_t *rt = msp->ms_allocatable;
1622	int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
1623	uint64_t offset;
1624
1625	ASSERT(MUTEX_HELD(&msp->ms_lock));
1626
1627	/*
1628	 * If we're running low on space, find a segment based on size,
1629	 * rather than iterating based on offset.
1630	 */
1631	if (metaslab_largest_allocatable(msp) < metaslab_df_alloc_threshold ||
1632	    free_pct < metaslab_df_free_pct) {
1633		offset = -1;
1634	} else {
1635		offset = metaslab_block_picker(rt,
1636		    cursor, size, metaslab_df_max_search);
1637	}
1638
1639	if (offset == -1) {
1640		range_seg_t *rs;
1641		if (zfs_btree_numnodes(&msp->ms_allocatable_by_size) == 0)
1642			metaslab_size_tree_full_load(msp->ms_allocatable);
1643		if (metaslab_df_use_largest_segment) {
1644			/* use largest free segment */
1645			rs = zfs_btree_last(&msp->ms_allocatable_by_size, NULL);
1646		} else {
1647			zfs_btree_index_t where;
1648			/* use segment of this size, or next largest */
1649#ifdef _METASLAB_TRACING
1650			metaslab_rt_arg_t *mrap = msp->ms_allocatable->rt_arg;
1651			if (size < (1 << mrap->mra_floor_shift)) {
1652				METASLABSTAT_BUMP(
1653				    metaslabstat_df_find_under_floor);
1654			}
1655#endif
1656			rs = metaslab_block_find(&msp->ms_allocatable_by_size,
1657			    rt, msp->ms_start, size, &where);
1658		}
1659		if (rs != NULL && rs_get_start(rs, rt) + size <= rs_get_end(rs,
1660		    rt)) {
1661			offset = rs_get_start(rs, rt);
1662			*cursor = offset + size;
1663		}
1664	}
1665
1666	return (offset);
1667}
1668
1669static metaslab_ops_t metaslab_df_ops = {
1670	metaslab_df_alloc
1671};
1672
1673/*
1674 * ==========================================================================
1675 * Cursor fit block allocator -
1676 * Select the largest region in the metaslab, set the cursor to the beginning
1677 * of the range and the cursor_end to the end of the range. As allocations
1678 * are made advance the cursor. Continue allocating from the cursor until
1679 * the range is exhausted and then find a new range.
1680 * ==========================================================================
1681 */
1682static uint64_t
1683metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
1684{
1685	range_tree_t *rt = msp->ms_allocatable;
1686	zfs_btree_t *t = &msp->ms_allocatable_by_size;
1687	uint64_t *cursor = &msp->ms_lbas[0];
1688	uint64_t *cursor_end = &msp->ms_lbas[1];
1689	uint64_t offset = 0;
1690
1691	ASSERT(MUTEX_HELD(&msp->ms_lock));
1692
1693	ASSERT3U(*cursor_end, >=, *cursor);
1694
1695	if ((*cursor + size) > *cursor_end) {
1696		range_seg_t *rs;
1697
1698		if (zfs_btree_numnodes(t) == 0)
1699			metaslab_size_tree_full_load(msp->ms_allocatable);
1700		rs = zfs_btree_last(t, NULL);
1701		if (rs == NULL || (rs_get_end(rs, rt) - rs_get_start(rs, rt)) <
1702		    size)
1703			return (-1ULL);
1704
1705		*cursor = rs_get_start(rs, rt);
1706		*cursor_end = rs_get_end(rs, rt);
1707	}
1708
1709	offset = *cursor;
1710	*cursor += size;
1711
1712	return (offset);
1713}
1714
1715static metaslab_ops_t metaslab_cf_ops = {
1716	metaslab_cf_alloc
1717};
1718
1719/*
1720 * ==========================================================================
1721 * New dynamic fit allocator -
1722 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1723 * contiguous blocks. If no region is found then just use the largest segment
1724 * that remains.
1725 * ==========================================================================
1726 */
1727
1728/*
1729 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1730 * to request from the allocator.
1731 */
1732uint64_t metaslab_ndf_clump_shift = 4;
1733
1734static uint64_t
1735metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
1736{
1737	zfs_btree_t *t = &msp->ms_allocatable->rt_root;
1738	range_tree_t *rt = msp->ms_allocatable;
1739	zfs_btree_index_t where;
1740	range_seg_t *rs;
1741	range_seg_max_t rsearch;
1742	uint64_t hbit = highbit64(size);
1743	uint64_t *cursor = &msp->ms_lbas[hbit - 1];
1744	uint64_t max_size = metaslab_largest_allocatable(msp);
1745
1746	ASSERT(MUTEX_HELD(&msp->ms_lock));
1747
1748	if (max_size < size)
1749		return (-1ULL);
1750
1751	rs_set_start(&rsearch, rt, *cursor);
1752	rs_set_end(&rsearch, rt, *cursor + size);
1753
1754	rs = zfs_btree_find(t, &rsearch, &where);
1755	if (rs == NULL || (rs_get_end(rs, rt) - rs_get_start(rs, rt)) < size) {
1756		t = &msp->ms_allocatable_by_size;
1757
1758		rs_set_start(&rsearch, rt, 0);
1759		rs_set_end(&rsearch, rt, MIN(max_size, 1ULL << (hbit +
1760		    metaslab_ndf_clump_shift)));
1761
1762		rs = zfs_btree_find(t, &rsearch, &where);
1763		if (rs == NULL)
1764			rs = zfs_btree_next(t, &where, &where);
1765		ASSERT(rs != NULL);
1766	}
1767
1768	if ((rs_get_end(rs, rt) - rs_get_start(rs, rt)) >= size) {
1769		*cursor = rs_get_start(rs, rt) + size;
1770		return (rs_get_start(rs, rt));
1771	}
1772	return (-1ULL);
1773}
1774
1775static metaslab_ops_t metaslab_ndf_ops = {
1776	metaslab_ndf_alloc
1777};
1778
1779metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
1780
1781/*
1782 * ==========================================================================
1783 * Metaslabs
1784 * ==========================================================================
1785 */
1786
1787/*
1788 * Wait for any in-progress metaslab loads to complete.
1789 */
1790void
1791metaslab_load_wait(metaslab_t *msp)
1792{
1793	ASSERT(MUTEX_HELD(&msp->ms_lock));
1794
1795	while (msp->ms_loading) {
1796		ASSERT(!msp->ms_loaded);
1797		cv_wait(&msp->ms_load_cv, &msp->ms_lock);
1798	}
1799}
1800
1801/*
1802 * Wait for any in-progress flushing to complete.
1803 */
1804void
1805metaslab_flush_wait(metaslab_t *msp)
1806{
1807	ASSERT(MUTEX_HELD(&msp->ms_lock));
1808
1809	while (msp->ms_flushing)
1810		cv_wait(&msp->ms_flush_cv, &msp->ms_lock);
1811}
1812
1813static unsigned int
1814metaslab_idx_func(multilist_t *ml, void *arg)
1815{
1816	metaslab_t *msp = arg;
1817	return (msp->ms_id % multilist_get_num_sublists(ml));
1818}
1819
1820uint64_t
1821metaslab_allocated_space(metaslab_t *msp)
1822{
1823	return (msp->ms_allocated_space);
1824}
1825
1826/*
1827 * Verify that the space accounting on disk matches the in-core range_trees.
1828 */
1829static void
1830metaslab_verify_space(metaslab_t *msp, uint64_t txg)
1831{
1832	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1833	uint64_t allocating = 0;
1834	uint64_t sm_free_space, msp_free_space;
1835
1836	ASSERT(MUTEX_HELD(&msp->ms_lock));
1837	ASSERT(!msp->ms_condensing);
1838
1839	if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0)
1840		return;
1841
1842	/*
1843	 * We can only verify the metaslab space when we're called
1844	 * from syncing context with a loaded metaslab that has an
1845	 * allocated space map. Calling this in non-syncing context
1846	 * does not provide a consistent view of the metaslab since
1847	 * we're performing allocations in the future.
1848	 */
1849	if (txg != spa_syncing_txg(spa) || msp->ms_sm == NULL ||
1850	    !msp->ms_loaded)
1851		return;
1852
1853	/*
1854	 * Even though the smp_alloc field can get negative,
1855	 * when it comes to a metaslab's space map, that should
1856	 * never be the case.
1857	 */
1858	ASSERT3S(space_map_allocated(msp->ms_sm), >=, 0);
1859
1860	ASSERT3U(space_map_allocated(msp->ms_sm), >=,
1861	    range_tree_space(msp->ms_unflushed_frees));
1862
1863	ASSERT3U(metaslab_allocated_space(msp), ==,
1864	    space_map_allocated(msp->ms_sm) +
1865	    range_tree_space(msp->ms_unflushed_allocs) -
1866	    range_tree_space(msp->ms_unflushed_frees));
1867
1868	sm_free_space = msp->ms_size - metaslab_allocated_space(msp);
1869
1870	/*
1871	 * Account for future allocations since we would have
1872	 * already deducted that space from the ms_allocatable.
1873	 */
1874	for (int t = 0; t < TXG_CONCURRENT_STATES; t++) {
1875		allocating +=
1876		    range_tree_space(msp->ms_allocating[(txg + t) & TXG_MASK]);
1877	}
1878	ASSERT3U(allocating + msp->ms_allocated_this_txg, ==,
1879	    msp->ms_allocating_total);
1880
1881	ASSERT3U(msp->ms_deferspace, ==,
1882	    range_tree_space(msp->ms_defer[0]) +
1883	    range_tree_space(msp->ms_defer[1]));
1884
1885	msp_free_space = range_tree_space(msp->ms_allocatable) + allocating +
1886	    msp->ms_deferspace + range_tree_space(msp->ms_freed);
1887
1888	VERIFY3U(sm_free_space, ==, msp_free_space);
1889}
1890
1891static void
1892metaslab_aux_histograms_clear(metaslab_t *msp)
1893{
1894	/*
1895	 * Auxiliary histograms are only cleared when resetting them,
1896	 * which can only happen while the metaslab is loaded.
1897	 */
1898	ASSERT(msp->ms_loaded);
1899
1900	bzero(msp->ms_synchist, sizeof (msp->ms_synchist));
1901	for (int t = 0; t < TXG_DEFER_SIZE; t++)
1902		bzero(msp->ms_deferhist[t], sizeof (msp->ms_deferhist[t]));
1903}
1904
1905static void
1906metaslab_aux_histogram_add(uint64_t *histogram, uint64_t shift,
1907    range_tree_t *rt)
1908{
1909	/*
1910	 * This is modeled after space_map_histogram_add(), so refer to that
1911	 * function for implementation details. We want this to work like
1912	 * the space map histogram, and not the range tree histogram, as we
1913	 * are essentially constructing a delta that will be later subtracted
1914	 * from the space map histogram.
1915	 */
1916	int idx = 0;
1917	for (int i = shift; i < RANGE_TREE_HISTOGRAM_SIZE; i++) {
1918		ASSERT3U(i, >=, idx + shift);
1919		histogram[idx] += rt->rt_histogram[i] << (i - idx - shift);
1920
1921		if (idx < SPACE_MAP_HISTOGRAM_SIZE - 1) {
1922			ASSERT3U(idx + shift, ==, i);
1923			idx++;
1924			ASSERT3U(idx, <, SPACE_MAP_HISTOGRAM_SIZE);
1925		}
1926	}
1927}
1928
1929/*
1930 * Called at every sync pass that the metaslab gets synced.
1931 *
1932 * The reason is that we want our auxiliary histograms to be updated
1933 * wherever the metaslab's space map histogram is updated. This way
1934 * we stay consistent on which parts of the metaslab space map's
1935 * histogram are currently not available for allocations (e.g because
1936 * they are in the defer, freed, and freeing trees).
1937 */
1938static void
1939metaslab_aux_histograms_update(metaslab_t *msp)
1940{
1941	space_map_t *sm = msp->ms_sm;
1942	ASSERT(sm != NULL);
1943
1944	/*
1945	 * This is similar to the metaslab's space map histogram updates
1946	 * that take place in metaslab_sync(). The only difference is that
1947	 * we only care about segments that haven't made it into the
1948	 * ms_allocatable tree yet.
1949	 */
1950	if (msp->ms_loaded) {
1951		metaslab_aux_histograms_clear(msp);
1952
1953		metaslab_aux_histogram_add(msp->ms_synchist,
1954		    sm->sm_shift, msp->ms_freed);
1955
1956		for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1957			metaslab_aux_histogram_add(msp->ms_deferhist[t],
1958			    sm->sm_shift, msp->ms_defer[t]);
1959		}
1960	}
1961
1962	metaslab_aux_histogram_add(msp->ms_synchist,
1963	    sm->sm_shift, msp->ms_freeing);
1964}
1965
1966/*
1967 * Called every time we are done syncing (writing to) the metaslab,
1968 * i.e. at the end of each sync pass.
1969 * [see the comment in metaslab_impl.h for ms_synchist, ms_deferhist]
1970 */
1971static void
1972metaslab_aux_histograms_update_done(metaslab_t *msp, boolean_t defer_allowed)
1973{
1974	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1975	space_map_t *sm = msp->ms_sm;
1976
1977	if (sm == NULL) {
1978		/*
1979		 * We came here from metaslab_init() when creating/opening a
1980		 * pool, looking at a metaslab that hasn't had any allocations
1981		 * yet.
1982		 */
1983		return;
1984	}
1985
1986	/*
1987	 * This is similar to the actions that we take for the ms_freed
1988	 * and ms_defer trees in metaslab_sync_done().
1989	 */
1990	uint64_t hist_index = spa_syncing_txg(spa) % TXG_DEFER_SIZE;
1991	if (defer_allowed) {
1992		bcopy(msp->ms_synchist, msp->ms_deferhist[hist_index],
1993		    sizeof (msp->ms_synchist));
1994	} else {
1995		bzero(msp->ms_deferhist[hist_index],
1996		    sizeof (msp->ms_deferhist[hist_index]));
1997	}
1998	bzero(msp->ms_synchist, sizeof (msp->ms_synchist));
1999}
2000
2001/*
2002 * Ensure that the metaslab's weight and fragmentation are consistent
2003 * with the contents of the histogram (either the range tree's histogram
2004 * or the space map's depending whether the metaslab is loaded).
2005 */
2006static void
2007metaslab_verify_weight_and_frag(metaslab_t *msp)
2008{
2009	ASSERT(MUTEX_HELD(&msp->ms_lock));
2010
2011	if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0)
2012		return;
2013
2014	/*
2015	 * We can end up here from vdev_remove_complete(), in which case we
2016	 * cannot do these assertions because we hold spa config locks and
2017	 * thus we are not allowed to read from the DMU.
2018	 *
2019	 * We check if the metaslab group has been removed and if that's
2020	 * the case we return immediately as that would mean that we are
2021	 * here from the aforementioned code path.
2022	 */
2023	if (msp->ms_group == NULL)
2024		return;
2025
2026	/*
2027	 * Devices being removed always return a weight of 0 and leave
2028	 * fragmentation and ms_max_size as is - there is nothing for
2029	 * us to verify here.
2030	 */
2031	vdev_t *vd = msp->ms_group->mg_vd;
2032	if (vd->vdev_removing)
2033		return;
2034
2035	/*
2036	 * If the metaslab is dirty it probably means that we've done
2037	 * some allocations or frees that have changed our histograms
2038	 * and thus the weight.
2039	 */
2040	for (int t = 0; t < TXG_SIZE; t++) {
2041		if (txg_list_member(&vd->vdev_ms_list, msp, t))
2042			return;
2043	}
2044
2045	/*
2046	 * This verification checks that our in-memory state is consistent
2047	 * with what's on disk. If the pool is read-only then there aren't
2048	 * any changes and we just have the initially-loaded state.
2049	 */
2050	if (!spa_writeable(msp->ms_group->mg_vd->vdev_spa))
2051		return;
2052
2053	/* some extra verification for in-core tree if you can */
2054	if (msp->ms_loaded) {
2055		range_tree_stat_verify(msp->ms_allocatable);
2056		VERIFY(space_map_histogram_verify(msp->ms_sm,
2057		    msp->ms_allocatable));
2058	}
2059
2060	uint64_t weight = msp->ms_weight;
2061	uint64_t was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
2062	boolean_t space_based = WEIGHT_IS_SPACEBASED(msp->ms_weight);
2063	uint64_t frag = msp->ms_fragmentation;
2064	uint64_t max_segsize = msp->ms_max_size;
2065
2066	msp->ms_weight = 0;
2067	msp->ms_fragmentation = 0;
2068
2069	/*
2070	 * This function is used for verification purposes. Regardless of
2071	 * whether metaslab_weight() thinks this metaslab should be active or
2072	 * not, we want to ensure that the actual weight (and therefore the
2073	 * value of ms_weight) would be the same if it was to be recalculated
2074	 * at this point.
2075	 */
2076	msp->ms_weight = metaslab_weight(msp) | was_active;
2077
2078	VERIFY3U(max_segsize, ==, msp->ms_max_size);
2079
2080	/*
2081	 * If the weight type changed then there is no point in doing
2082	 * verification. Revert fields to their original values.
2083	 */
2084	if ((space_based && !WEIGHT_IS_SPACEBASED(msp->ms_weight)) ||
2085	    (!space_based && WEIGHT_IS_SPACEBASED(msp->ms_weight))) {
2086		msp->ms_fragmentation = frag;
2087		msp->ms_weight = weight;
2088		return;
2089	}
2090
2091	VERIFY3U(msp->ms_fragmentation, ==, frag);
2092	VERIFY3U(msp->ms_weight, ==, weight);
2093}
2094
2095/*
2096 * If we're over the zfs_metaslab_mem_limit, select the loaded metaslab from
2097 * this class that was used longest ago, and attempt to unload it.  We don't
2098 * want to spend too much time in this loop to prevent performance
2099 * degredation, and we expect that most of the time this operation will
2100 * succeed. Between that and the normal unloading processing during txg sync,
2101 * we expect this to keep the metaslab memory usage under control.
2102 */
2103static void
2104metaslab_potentially_evict(metaslab_class_t *mc)
2105{
2106#ifdef _KERNEL
2107	uint64_t allmem = arc_all_memory();
2108	extern kmem_cache_t *zfs_btree_leaf_cache;
2109	uint64_t inuse = kmem_cache_stat(zfs_btree_leaf_cache, "buf_inuse");
2110	uint64_t size =  kmem_cache_stat(zfs_btree_leaf_cache, "buf_size");
2111	int tries = 0;
2112	for (; allmem * zfs_metaslab_mem_limit / 100 < inuse * size &&
2113	    tries < multilist_get_num_sublists(mc->mc_metaslab_txg_list) * 2;
2114	    tries++) {
2115		unsigned int idx = multilist_get_random_index(
2116		    mc->mc_metaslab_txg_list);
2117		multilist_sublist_t *mls =
2118		    multilist_sublist_lock(mc->mc_metaslab_txg_list, idx);
2119		metaslab_t *msp = multilist_sublist_head(mls);
2120		multilist_sublist_unlock(mls);
2121		while (msp != NULL && allmem * zfs_metaslab_mem_limit / 100 <
2122		    inuse * size) {
2123			VERIFY3P(mls, ==, multilist_sublist_lock(
2124			    mc->mc_metaslab_txg_list, idx));
2125			ASSERT3U(idx, ==,
2126			    metaslab_idx_func(mc->mc_metaslab_txg_list, msp));
2127
2128			if (!multilist_link_active(&msp->ms_class_txg_node)) {
2129				multilist_sublist_unlock(mls);
2130				break;
2131			}
2132			metaslab_t *next_msp = multilist_sublist_next(mls, msp);
2133			multilist_sublist_unlock(mls);
2134			/*
2135			 * If the metaslab is currently loading there are two
2136			 * cases. If it's the metaslab we're evicting, we
2137			 * can't continue on or we'll panic when we attempt to
2138			 * recursively lock the mutex. If it's another
2139			 * metaslab that's loading, it can be safely skipped,
2140			 * since we know it's very new and therefore not a
2141			 * good eviction candidate. We check later once the
2142			 * lock is held that the metaslab is fully loaded
2143			 * before actually unloading it.
2144			 */
2145			if (msp->ms_loading) {
2146				msp = next_msp;
2147				inuse = kmem_cache_stat(zfs_btree_leaf_cache,
2148				    "buf_inuse");
2149				continue;
2150			}
2151			/*
2152			 * We can't unload metaslabs with no spacemap because
2153			 * they're not ready to be unloaded yet. We can't
2154			 * unload metaslabs with outstanding allocations
2155			 * because doing so could cause the metaslab's weight
2156			 * to decrease while it's unloaded, which violates an
2157			 * invariant that we use to prevent unnecessary
2158			 * loading. We also don't unload metaslabs that are
2159			 * currently active because they are high-weight
2160			 * metaslabs that are likely to be used in the near
2161			 * future.
2162			 */
2163			mutex_enter(&msp->ms_lock);
2164			if (msp->ms_allocator == -1 && msp->ms_sm != NULL &&
2165			    msp->ms_allocating_total == 0) {
2166				metaslab_unload(msp);
2167			}
2168			mutex_exit(&msp->ms_lock);
2169			msp = next_msp;
2170			inuse = kmem_cache_stat(zfs_btree_leaf_cache,
2171			    "buf_inuse");
2172		}
2173	}
2174#endif
2175}
2176
2177static int
2178metaslab_load_impl(metaslab_t *msp)
2179{
2180	int error = 0;
2181
2182	ASSERT(MUTEX_HELD(&msp->ms_lock));
2183	ASSERT(msp->ms_loading);
2184	ASSERT(!msp->ms_condensing);
2185
2186	/*
2187	 * We temporarily drop the lock to unblock other operations while we
2188	 * are reading the space map. Therefore, metaslab_sync() and
2189	 * metaslab_sync_done() can run at the same time as we do.
2190	 *
2191	 * If we are using the log space maps, metaslab_sync() can't write to
2192	 * the metaslab's space map while we are loading as we only write to
2193	 * it when we are flushing the metaslab, and that can't happen while
2194	 * we are loading it.
2195	 *
2196	 * If we are not using log space maps though, metaslab_sync() can
2197	 * append to the space map while we are loading. Therefore we load
2198	 * only entries that existed when we started the load. Additionally,
2199	 * metaslab_sync_done() has to wait for the load to complete because
2200	 * there are potential races like metaslab_load() loading parts of the
2201	 * space map that are currently being appended by metaslab_sync(). If
2202	 * we didn't, the ms_allocatable would have entries that
2203	 * metaslab_sync_done() would try to re-add later.
2204	 *
2205	 * That's why before dropping the lock we remember the synced length
2206	 * of the metaslab and read up to that point of the space map,
2207	 * ignoring entries appended by metaslab_sync() that happen after we
2208	 * drop the lock.
2209	 */
2210	uint64_t length = msp->ms_synced_length;
2211	mutex_exit(&msp->ms_lock);
2212
2213	hrtime_t load_start = gethrtime();
2214	metaslab_rt_arg_t *mrap;
2215	if (msp->ms_allocatable->rt_arg == NULL) {
2216		mrap = kmem_zalloc(sizeof (*mrap), KM_SLEEP);
2217	} else {
2218		mrap = msp->ms_allocatable->rt_arg;
2219		msp->ms_allocatable->rt_ops = NULL;
2220		msp->ms_allocatable->rt_arg = NULL;
2221	}
2222	mrap->mra_bt = &msp->ms_allocatable_by_size;
2223	mrap->mra_floor_shift = metaslab_by_size_min_shift;
2224
2225	if (msp->ms_sm != NULL) {
2226		error = space_map_load_length(msp->ms_sm, msp->ms_allocatable,
2227		    SM_FREE, length);
2228
2229		/* Now, populate the size-sorted tree. */
2230		metaslab_rt_create(msp->ms_allocatable, mrap);
2231		msp->ms_allocatable->rt_ops = &metaslab_rt_ops;
2232		msp->ms_allocatable->rt_arg = mrap;
2233
2234		struct mssa_arg arg = {0};
2235		arg.rt = msp->ms_allocatable;
2236		arg.mra = mrap;
2237		range_tree_walk(msp->ms_allocatable, metaslab_size_sorted_add,
2238		    &arg);
2239	} else {
2240		/*
2241		 * Add the size-sorted tree first, since we don't need to load
2242		 * the metaslab from the spacemap.
2243		 */
2244		metaslab_rt_create(msp->ms_allocatable, mrap);
2245		msp->ms_allocatable->rt_ops = &metaslab_rt_ops;
2246		msp->ms_allocatable->rt_arg = mrap;
2247		/*
2248		 * The space map has not been allocated yet, so treat
2249		 * all the space in the metaslab as free and add it to the
2250		 * ms_allocatable tree.
2251		 */
2252		range_tree_add(msp->ms_allocatable,
2253		    msp->ms_start, msp->ms_size);
2254
2255		if (msp->ms_freed != NULL) {
2256			/*
2257			 * If the ms_sm doesn't exist, this means that this
2258			 * metaslab hasn't gone through metaslab_sync() and
2259			 * thus has never been dirtied. So we shouldn't
2260			 * expect any unflushed allocs or frees from previous
2261			 * TXGs.
2262			 *
2263			 * Note: ms_freed and all the other trees except for
2264			 * the ms_allocatable, can be NULL at this point only
2265			 * if this is a new metaslab of a vdev that just got
2266			 * expanded.
2267			 */
2268			ASSERT(range_tree_is_empty(msp->ms_unflushed_allocs));
2269			ASSERT(range_tree_is_empty(msp->ms_unflushed_frees));
2270		}
2271	}
2272
2273	/*
2274	 * We need to grab the ms_sync_lock to prevent metaslab_sync() from
2275	 * changing the ms_sm (or log_sm) and the metaslab's range trees
2276	 * while we are about to use them and populate the ms_allocatable.
2277	 * The ms_lock is insufficient for this because metaslab_sync() doesn't
2278	 * hold the ms_lock while writing the ms_checkpointing tree to disk.
2279	 */
2280	mutex_enter(&msp->ms_sync_lock);
2281	mutex_enter(&msp->ms_lock);
2282
2283	ASSERT(!msp->ms_condensing);
2284	ASSERT(!msp->ms_flushing);
2285
2286	if (error != 0) {
2287		mutex_exit(&msp->ms_sync_lock);
2288		return (error);
2289	}
2290
2291	ASSERT3P(msp->ms_group, !=, NULL);
2292	msp->ms_loaded = B_TRUE;
2293
2294	/*
2295	 * Apply all the unflushed changes to ms_allocatable right
2296	 * away so any manipulations we do below have a clear view
2297	 * of what is allocated and what is free.
2298	 */
2299	range_tree_walk(msp->ms_unflushed_allocs,
2300	    range_tree_remove, msp->ms_allocatable);
2301	range_tree_walk(msp->ms_unflushed_frees,
2302	    range_tree_add, msp->ms_allocatable);
2303
2304	msp->ms_loaded = B_TRUE;
2305
2306	ASSERT3P(msp->ms_group, !=, NULL);
2307	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2308	if (spa_syncing_log_sm(spa) != NULL) {
2309		ASSERT(spa_feature_is_enabled(spa,
2310		    SPA_FEATURE_LOG_SPACEMAP));
2311
2312		/*
2313		 * If we use a log space map we add all the segments
2314		 * that are in ms_unflushed_frees so they are available
2315		 * for allocation.
2316		 *
2317		 * ms_allocatable needs to contain all free segments
2318		 * that are ready for allocations (thus not segments
2319		 * from ms_freeing, ms_freed, and the ms_defer trees).
2320		 * But if we grab the lock in this code path at a sync
2321		 * pass later that 1, then it also contains the
2322		 * segments of ms_freed (they were added to it earlier
2323		 * in this path through ms_unflushed_frees). So we
2324		 * need to remove all the segments that exist in
2325		 * ms_freed from ms_allocatable as they will be added
2326		 * later in metaslab_sync_done().
2327		 *
2328		 * When there's no log space map, the ms_allocatable
2329		 * correctly doesn't contain any segments that exist
2330		 * in ms_freed [see ms_synced_length].
2331		 */
2332		range_tree_walk(msp->ms_freed,
2333		    range_tree_remove, msp->ms_allocatable);
2334	}
2335
2336	/*
2337	 * If we are not using the log space map, ms_allocatable
2338	 * contains the segments that exist in the ms_defer trees
2339	 * [see ms_synced_length]. Thus we need to remove them
2340	 * from ms_allocatable as they will be added again in
2341	 * metaslab_sync_done().
2342	 *
2343	 * If we are using the log space map, ms_allocatable still
2344	 * contains the segments that exist in the ms_defer trees.
2345	 * Not because it read them through the ms_sm though. But
2346	 * because these segments are part of ms_unflushed_frees
2347	 * whose segments we add to ms_allocatable earlier in this
2348	 * code path.
2349	 */
2350	for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2351		range_tree_walk(msp->ms_defer[t],
2352		    range_tree_remove, msp->ms_allocatable);
2353	}
2354
2355	/*
2356	 * Call metaslab_recalculate_weight_and_sort() now that the
2357	 * metaslab is loaded so we get the metaslab's real weight.
2358	 *
2359	 * Unless this metaslab was created with older software and
2360	 * has not yet been converted to use segment-based weight, we
2361	 * expect the new weight to be better or equal to the weight
2362	 * that the metaslab had while it was not loaded. This is
2363	 * because the old weight does not take into account the
2364	 * consolidation of adjacent segments between TXGs. [see
2365	 * comment for ms_synchist and ms_deferhist[] for more info]
2366	 */
2367	uint64_t weight = msp->ms_weight;
2368	uint64_t max_size = msp->ms_max_size;
2369	metaslab_recalculate_weight_and_sort(msp);
2370	if (!WEIGHT_IS_SPACEBASED(weight))
2371		ASSERT3U(weight, <=, msp->ms_weight);
2372	msp->ms_max_size = metaslab_largest_allocatable(msp);
2373	ASSERT3U(max_size, <=, msp->ms_max_size);
2374	hrtime_t load_end = gethrtime();
2375		msp->ms_load_time = load_end;
2376	if (zfs_flags & ZFS_DEBUG_LOG_SPACEMAP) {
2377		zfs_dbgmsg("loading: txg %llu, spa %s, vdev_id %llu, "
2378		    "ms_id %llu, smp_length %llu, "
2379		    "unflushed_allocs %llu, unflushed_frees %llu, "
2380		    "freed %llu, defer %llu + %llu, "
2381		    "loading_time %lld ms, ms_max_size %llu, "
2382		    "max size error %llu",
2383		    spa_syncing_txg(spa), spa_name(spa),
2384		    msp->ms_group->mg_vd->vdev_id, msp->ms_id,
2385		    space_map_length(msp->ms_sm),
2386		    range_tree_space(msp->ms_unflushed_allocs),
2387		    range_tree_space(msp->ms_unflushed_frees),
2388		    range_tree_space(msp->ms_freed),
2389		    range_tree_space(msp->ms_defer[0]),
2390		    range_tree_space(msp->ms_defer[1]),
2391		    (longlong_t)((load_end - load_start) / 1000000),
2392		    msp->ms_max_size, msp->ms_max_size - max_size);
2393	}
2394
2395	metaslab_verify_space(msp, spa_syncing_txg(spa));
2396	mutex_exit(&msp->ms_sync_lock);
2397	return (0);
2398}
2399
2400int
2401metaslab_load(metaslab_t *msp)
2402{
2403	ASSERT(MUTEX_HELD(&msp->ms_lock));
2404
2405	/*
2406	 * There may be another thread loading the same metaslab, if that's
2407	 * the case just wait until the other thread is done and return.
2408	 */
2409	metaslab_load_wait(msp);
2410	if (msp->ms_loaded)
2411		return (0);
2412	VERIFY(!msp->ms_loading);
2413	ASSERT(!msp->ms_condensing);
2414
2415	/*
2416	 * We set the loading flag BEFORE potentially dropping the lock to
2417	 * wait for an ongoing flush (see ms_flushing below). This way other
2418	 * threads know that there is already a thread that is loading this
2419	 * metaslab.
2420	 */
2421	msp->ms_loading = B_TRUE;
2422
2423	/*
2424	 * Wait for any in-progress flushing to finish as we drop the ms_lock
2425	 * both here (during space_map_load()) and in metaslab_flush() (when
2426	 * we flush our changes to the ms_sm).
2427	 */
2428	if (msp->ms_flushing)
2429		metaslab_flush_wait(msp);
2430
2431	/*
2432	 * In the possibility that we were waiting for the metaslab to be
2433	 * flushed (where we temporarily dropped the ms_lock), ensure that
2434	 * no one else loaded the metaslab somehow.
2435	 */
2436	ASSERT(!msp->ms_loaded);
2437
2438	/*
2439	 * If we're loading a metaslab in the normal class, consider evicting
2440	 * another one to keep our memory usage under the limit defined by the
2441	 * zfs_metaslab_mem_limit tunable.
2442	 */
2443	if (spa_normal_class(msp->ms_group->mg_class->mc_spa) ==
2444	    msp->ms_group->mg_class) {
2445		metaslab_potentially_evict(msp->ms_group->mg_class);
2446	}
2447
2448	int error = metaslab_load_impl(msp);
2449
2450	ASSERT(MUTEX_HELD(&msp->ms_lock));
2451	msp->ms_loading = B_FALSE;
2452	cv_broadcast(&msp->ms_load_cv);
2453
2454	return (error);
2455}
2456
2457void
2458metaslab_unload(metaslab_t *msp)
2459{
2460	ASSERT(MUTEX_HELD(&msp->ms_lock));
2461
2462	/*
2463	 * This can happen if a metaslab is selected for eviction (in
2464	 * metaslab_potentially_evict) and then unloaded during spa_sync (via
2465	 * metaslab_class_evict_old).
2466	 */
2467	if (!msp->ms_loaded)
2468		return;
2469
2470	range_tree_vacate(msp->ms_allocatable, NULL, NULL);
2471	msp->ms_loaded = B_FALSE;
2472	msp->ms_unload_time = gethrtime();
2473
2474	msp->ms_activation_weight = 0;
2475	msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
2476
2477	if (msp->ms_group != NULL) {
2478		metaslab_class_t *mc = msp->ms_group->mg_class;
2479		multilist_sublist_t *mls =
2480		    multilist_sublist_lock_obj(mc->mc_metaslab_txg_list, msp);
2481		if (multilist_link_active(&msp->ms_class_txg_node))
2482			multilist_sublist_remove(mls, msp);
2483		multilist_sublist_unlock(mls);
2484	}
2485
2486	/*
2487	 * We explicitly recalculate the metaslab's weight based on its space
2488	 * map (as it is now not loaded). We want unload metaslabs to always
2489	 * have their weights calculated from the space map histograms, while
2490	 * loaded ones have it calculated from their in-core range tree
2491	 * [see metaslab_load()]. This way, the weight reflects the information
2492	 * available in-core, whether it is loaded or not.
2493	 *
2494	 * If ms_group == NULL means that we came here from metaslab_fini(),
2495	 * at which point it doesn't make sense for us to do the recalculation
2496	 * and the sorting.
2497	 */
2498	if (msp->ms_group != NULL)
2499		metaslab_recalculate_weight_and_sort(msp);
2500}
2501
2502/*
2503 * We want to optimize the memory use of the per-metaslab range
2504 * trees. To do this, we store the segments in the range trees in
2505 * units of sectors, zero-indexing from the start of the metaslab. If
2506 * the vdev_ms_shift - the vdev_ashift is less than 32, we can store
2507 * the ranges using two uint32_ts, rather than two uint64_ts.
2508 */
2509static range_seg_type_t
2510metaslab_calculate_range_tree_type(vdev_t *vdev, metaslab_t *msp,
2511    uint64_t *start, uint64_t *shift)
2512{
2513	if (vdev->vdev_ms_shift - vdev->vdev_ashift < 32 &&
2514	    !zfs_metaslab_force_large_segs) {
2515		*shift = vdev->vdev_ashift;
2516		*start = msp->ms_start;
2517		return (RANGE_SEG32);
2518	} else {
2519		*shift = 0;
2520		*start = 0;
2521		return (RANGE_SEG64);
2522	}
2523}
2524
2525void
2526metaslab_set_selected_txg(metaslab_t *msp, uint64_t txg)
2527{
2528	ASSERT(MUTEX_HELD(&msp->ms_lock));
2529	metaslab_class_t *mc = msp->ms_group->mg_class;
2530	multilist_sublist_t *mls =
2531	    multilist_sublist_lock_obj(mc->mc_metaslab_txg_list, msp);
2532	if (multilist_link_active(&msp->ms_class_txg_node))
2533		multilist_sublist_remove(mls, msp);
2534	msp->ms_selected_txg = txg;
2535	msp->ms_selected_time = gethrtime();
2536	multilist_sublist_insert_tail(mls, msp);
2537	multilist_sublist_unlock(mls);
2538}
2539
2540void
2541metaslab_space_update(vdev_t *vd, metaslab_class_t *mc, int64_t alloc_delta,
2542    int64_t defer_delta, int64_t space_delta)
2543{
2544	vdev_space_update(vd, alloc_delta, defer_delta, space_delta);
2545
2546	ASSERT3P(vd->vdev_spa->spa_root_vdev, ==, vd->vdev_parent);
2547	ASSERT(vd->vdev_ms_count != 0);
2548
2549	metaslab_class_space_update(mc, alloc_delta, defer_delta, space_delta,
2550	    vdev_deflated_space(vd, space_delta));
2551}
2552
2553int
2554metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object,
2555    uint64_t txg, metaslab_t **msp)
2556{
2557	vdev_t *vd = mg->mg_vd;
2558	spa_t *spa = vd->vdev_spa;
2559	objset_t *mos = spa->spa_meta_objset;
2560	metaslab_t *ms;
2561	int error;
2562
2563	ms = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
2564	mutex_init(&ms->ms_lock, NULL, MUTEX_DEFAULT, NULL);
2565	mutex_init(&ms->ms_sync_lock, NULL, MUTEX_DEFAULT, NULL);
2566	cv_init(&ms->ms_load_cv, NULL, CV_DEFAULT, NULL);
2567	cv_init(&ms->ms_flush_cv, NULL, CV_DEFAULT, NULL);
2568	multilist_link_init(&ms->ms_class_txg_node);
2569
2570	ms->ms_id = id;
2571	ms->ms_start = id << vd->vdev_ms_shift;
2572	ms->ms_size = 1ULL << vd->vdev_ms_shift;
2573	ms->ms_allocator = -1;
2574	ms->ms_new = B_TRUE;
2575
2576	/*
2577	 * We only open space map objects that already exist. All others
2578	 * will be opened when we finally allocate an object for it.
2579	 *
2580	 * Note:
2581	 * When called from vdev_expand(), we can't call into the DMU as
2582	 * we are holding the spa_config_lock as a writer and we would
2583	 * deadlock [see relevant comment in vdev_metaslab_init()]. in
2584	 * that case, the object parameter is zero though, so we won't
2585	 * call into the DMU.
2586	 */
2587	if (object != 0) {
2588		error = space_map_open(&ms->ms_sm, mos, object, ms->ms_start,
2589		    ms->ms_size, vd->vdev_ashift);
2590
2591		if (error != 0) {
2592			kmem_free(ms, sizeof (metaslab_t));
2593			return (error);
2594		}
2595
2596		ASSERT(ms->ms_sm != NULL);
2597		ASSERT3S(space_map_allocated(ms->ms_sm), >=, 0);
2598		ms->ms_allocated_space = space_map_allocated(ms->ms_sm);
2599	}
2600
2601	range_seg_type_t type;
2602	uint64_t shift, start;
2603	type = metaslab_calculate_range_tree_type(vd, ms, &start, &shift);
2604
2605	/*
2606	 * We create the ms_allocatable here, but we don't create the
2607	 * other range trees until metaslab_sync_done().  This serves
2608	 * two purposes: it allows metaslab_sync_done() to detect the
2609	 * addition of new space; and for debugging, it ensures that
2610	 * we'd data fault on any attempt to use this metaslab before
2611	 * it's ready.
2612	 */
2613	ms->ms_allocatable = range_tree_create(NULL, type, NULL, start, shift);
2614
2615	ms->ms_trim = range_tree_create(NULL, type, NULL, start, shift);
2616
2617	metaslab_group_add(mg, ms);
2618	metaslab_set_fragmentation(ms);
2619
2620	/*
2621	 * If we're opening an existing pool (txg == 0) or creating
2622	 * a new one (txg == TXG_INITIAL), all space is available now.
2623	 * If we're adding space to an existing pool, the new space
2624	 * does not become available until after this txg has synced.
2625	 * The metaslab's weight will also be initialized when we sync
2626	 * out this txg. This ensures that we don't attempt to allocate
2627	 * from it before we have initialized it completely.
2628	 */
2629	if (txg <= TXG_INITIAL) {
2630		metaslab_sync_done(ms, 0);
2631		metaslab_space_update(vd, mg->mg_class,
2632		    metaslab_allocated_space(ms), 0, 0);
2633	}
2634
2635	if (txg != 0) {
2636		vdev_dirty(vd, 0, NULL, txg);
2637		vdev_dirty(vd, VDD_METASLAB, ms, txg);
2638	}
2639
2640	*msp = ms;
2641
2642	return (0);
2643}
2644
2645static void
2646metaslab_fini_flush_data(metaslab_t *msp)
2647{
2648	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2649
2650	if (metaslab_unflushed_txg(msp) == 0) {
2651		ASSERT3P(avl_find(&spa->spa_metaslabs_by_flushed, msp, NULL),
2652		    ==, NULL);
2653		return;
2654	}
2655	ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP));
2656
2657	mutex_enter(&spa->spa_flushed_ms_lock);
2658	avl_remove(&spa->spa_metaslabs_by_flushed, msp);
2659	mutex_exit(&spa->spa_flushed_ms_lock);
2660
2661	spa_log_sm_decrement_mscount(spa, metaslab_unflushed_txg(msp));
2662	spa_log_summary_decrement_mscount(spa, metaslab_unflushed_txg(msp));
2663}
2664
2665uint64_t
2666metaslab_unflushed_changes_memused(metaslab_t *ms)
2667{
2668	return ((range_tree_numsegs(ms->ms_unflushed_allocs) +
2669	    range_tree_numsegs(ms->ms_unflushed_frees)) *
2670	    ms->ms_unflushed_allocs->rt_root.bt_elem_size);
2671}
2672
2673void
2674metaslab_fini(metaslab_t *msp)
2675{
2676	metaslab_group_t *mg = msp->ms_group;
2677	vdev_t *vd = mg->mg_vd;
2678	spa_t *spa = vd->vdev_spa;
2679
2680	metaslab_fini_flush_data(msp);
2681
2682	metaslab_group_remove(mg, msp);
2683
2684	mutex_enter(&msp->ms_lock);
2685	VERIFY(msp->ms_group == NULL);
2686	metaslab_space_update(vd, mg->mg_class,
2687	    -metaslab_allocated_space(msp), 0, -msp->ms_size);
2688
2689	space_map_close(msp->ms_sm);
2690	msp->ms_sm = NULL;
2691
2692	metaslab_unload(msp);
2693	range_tree_destroy(msp->ms_allocatable);
2694	range_tree_destroy(msp->ms_freeing);
2695	range_tree_destroy(msp->ms_freed);
2696
2697	ASSERT3U(spa->spa_unflushed_stats.sus_memused, >=,
2698	    metaslab_unflushed_changes_memused(msp));
2699	spa->spa_unflushed_stats.sus_memused -=
2700	    metaslab_unflushed_changes_memused(msp);
2701	range_tree_vacate(msp->ms_unflushed_allocs, NULL, NULL);
2702	range_tree_destroy(msp->ms_unflushed_allocs);
2703	range_tree_vacate(msp->ms_unflushed_frees, NULL, NULL);
2704	range_tree_destroy(msp->ms_unflushed_frees);
2705
2706	for (int t = 0; t < TXG_SIZE; t++) {
2707		range_tree_destroy(msp->ms_allocating[t]);
2708	}
2709
2710	for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2711		range_tree_destroy(msp->ms_defer[t]);
2712	}
2713	ASSERT0(msp->ms_deferspace);
2714
2715	range_tree_destroy(msp->ms_checkpointing);
2716
2717	for (int t = 0; t < TXG_SIZE; t++)
2718		ASSERT(!txg_list_member(&vd->vdev_ms_list, msp, t));
2719
2720	range_tree_vacate(msp->ms_trim, NULL, NULL);
2721	range_tree_destroy(msp->ms_trim);
2722
2723	mutex_exit(&msp->ms_lock);
2724	cv_destroy(&msp->ms_load_cv);
2725	cv_destroy(&msp->ms_flush_cv);
2726	mutex_destroy(&msp->ms_lock);
2727	mutex_destroy(&msp->ms_sync_lock);
2728	ASSERT3U(msp->ms_allocator, ==, -1);
2729
2730	kmem_free(msp, sizeof (metaslab_t));
2731}
2732
2733#define	FRAGMENTATION_TABLE_SIZE	17
2734
2735/*
2736 * This table defines a segment size based fragmentation metric that will
2737 * allow each metaslab to derive its own fragmentation value. This is done
2738 * by calculating the space in each bucket of the spacemap histogram and
2739 * multiplying that by the fragmentation metric in this table. Doing
2740 * this for all buckets and dividing it by the total amount of free
2741 * space in this metaslab (i.e. the total free space in all buckets) gives
2742 * us the fragmentation metric. This means that a high fragmentation metric
2743 * equates to most of the free space being comprised of small segments.
2744 * Conversely, if the metric is low, then most of the free space is in
2745 * large segments. A 10% change in fragmentation equates to approximately
2746 * double the number of segments.
2747 *
2748 * This table defines 0% fragmented space using 16MB segments. Testing has
2749 * shown that segments that are greater than or equal to 16MB do not suffer
2750 * from drastic performance problems. Using this value, we derive the rest
2751 * of the table. Since the fragmentation value is never stored on disk, it
2752 * is possible to change these calculations in the future.
2753 */
2754int zfs_frag_table[FRAGMENTATION_TABLE_SIZE] = {
2755	100,	/* 512B	*/
2756	100,	/* 1K	*/
2757	98,	/* 2K	*/
2758	95,	/* 4K	*/
2759	90,	/* 8K	*/
2760	80,	/* 16K	*/
2761	70,	/* 32K	*/
2762	60,	/* 64K	*/
2763	50,	/* 128K	*/
2764	40,	/* 256K	*/
2765	30,	/* 512K	*/
2766	20,	/* 1M	*/
2767	15,	/* 2M	*/
2768	10,	/* 4M	*/
2769	5,	/* 8M	*/
2770	0	/* 16M	*/
2771};
2772
2773/*
2774 * Calculate the metaslab's fragmentation metric and set ms_fragmentation.
2775 * Setting this value to ZFS_FRAG_INVALID means that the metaslab has not
2776 * been upgraded and does not support this metric. Otherwise, the return
2777 * value should be in the range [0, 100].
2778 */
2779static void
2780metaslab_set_fragmentation(metaslab_t *msp)
2781{
2782	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2783	uint64_t fragmentation = 0;
2784	uint64_t total = 0;
2785	boolean_t feature_enabled = spa_feature_is_enabled(spa,
2786	    SPA_FEATURE_SPACEMAP_HISTOGRAM);
2787
2788	if (!feature_enabled) {
2789		msp->ms_fragmentation = ZFS_FRAG_INVALID;
2790		return;
2791	}
2792
2793	/*
2794	 * A null space map means that the entire metaslab is free
2795	 * and thus is not fragmented.
2796	 */
2797	if (msp->ms_sm == NULL) {
2798		msp->ms_fragmentation = 0;
2799		return;
2800	}
2801
2802	/*
2803	 * If this metaslab's space map has not been upgraded, flag it
2804	 * so that we upgrade next time we encounter it.
2805	 */
2806	if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t)) {
2807		uint64_t txg = spa_syncing_txg(spa);
2808		vdev_t *vd = msp->ms_group->mg_vd;
2809
2810		/*
2811		 * If we've reached the final dirty txg, then we must
2812		 * be shutting down the pool. We don't want to dirty
2813		 * any data past this point so skip setting the condense
2814		 * flag. We can retry this action the next time the pool
2815		 * is imported.
2816		 */
2817		if (spa_writeable(spa) && txg < spa_final_dirty_txg(spa)) {
2818			msp->ms_condense_wanted = B_TRUE;
2819			vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
2820			zfs_dbgmsg("txg %llu, requesting force condense: "
2821			    "ms_id %llu, vdev_id %llu", txg, msp->ms_id,
2822			    vd->vdev_id);
2823		}
2824		msp->ms_fragmentation = ZFS_FRAG_INVALID;
2825		return;
2826	}
2827
2828	for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
2829		uint64_t space = 0;
2830		uint8_t shift = msp->ms_sm->sm_shift;
2831
2832		int idx = MIN(shift - SPA_MINBLOCKSHIFT + i,
2833		    FRAGMENTATION_TABLE_SIZE - 1);
2834
2835		if (msp->ms_sm->sm_phys->smp_histogram[i] == 0)
2836			continue;
2837
2838		space = msp->ms_sm->sm_phys->smp_histogram[i] << (i + shift);
2839		total += space;
2840
2841		ASSERT3U(idx, <, FRAGMENTATION_TABLE_SIZE);
2842		fragmentation += space * zfs_frag_table[idx];
2843	}
2844
2845	if (total > 0)
2846		fragmentation /= total;
2847	ASSERT3U(fragmentation, <=, 100);
2848
2849	msp->ms_fragmentation = fragmentation;
2850}
2851
2852/*
2853 * Compute a weight -- a selection preference value -- for the given metaslab.
2854 * This is based on the amount of free space, the level of fragmentation,
2855 * the LBA range, and whether the metaslab is loaded.
2856 */
2857static uint64_t
2858metaslab_space_weight(metaslab_t *msp)
2859{
2860	metaslab_group_t *mg = msp->ms_group;
2861	vdev_t *vd = mg->mg_vd;
2862	uint64_t weight, space;
2863
2864	ASSERT(MUTEX_HELD(&msp->ms_lock));
2865
2866	/*
2867	 * The baseline weight is the metaslab's free space.
2868	 */
2869	space = msp->ms_size - metaslab_allocated_space(msp);
2870
2871	if (metaslab_fragmentation_factor_enabled &&
2872	    msp->ms_fragmentation != ZFS_FRAG_INVALID) {
2873		/*
2874		 * Use the fragmentation information to inversely scale
2875		 * down the baseline weight. We need to ensure that we
2876		 * don't exclude this metaslab completely when it's 100%
2877		 * fragmented. To avoid this we reduce the fragmented value
2878		 * by 1.
2879		 */
2880		space = (space * (100 - (msp->ms_fragmentation - 1))) / 100;
2881
2882		/*
2883		 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
2884		 * this metaslab again. The fragmentation metric may have
2885		 * decreased the space to something smaller than
2886		 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
2887		 * so that we can consume any remaining space.
2888		 */
2889		if (space > 0 && space < SPA_MINBLOCKSIZE)
2890			space = SPA_MINBLOCKSIZE;
2891	}
2892	weight = space;
2893
2894	/*
2895	 * Modern disks have uniform bit density and constant angular velocity.
2896	 * Therefore, the outer recording zones are faster (higher bandwidth)
2897	 * than the inner zones by the ratio of outer to inner track diameter,
2898	 * which is typically around 2:1.  We account for this by assigning
2899	 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
2900	 * In effect, this means that we'll select the metaslab with the most
2901	 * free bandwidth rather than simply the one with the most free space.
2902	 */
2903	if (!vd->vdev_nonrot && metaslab_lba_weighting_enabled) {
2904		weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count;
2905		ASSERT(weight >= space && weight <= 2 * space);
2906	}
2907
2908	/*
2909	 * If this metaslab is one we're actively using, adjust its
2910	 * weight to make it preferable to any inactive metaslab so
2911	 * we'll polish it off. If the fragmentation on this metaslab
2912	 * has exceed our threshold, then don't mark it active.
2913	 */
2914	if (msp->ms_loaded && msp->ms_fragmentation != ZFS_FRAG_INVALID &&
2915	    msp->ms_fragmentation <= zfs_metaslab_fragmentation_threshold) {
2916		weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
2917	}
2918
2919	WEIGHT_SET_SPACEBASED(weight);
2920	return (weight);
2921}
2922
2923/*
2924 * Return the weight of the specified metaslab, according to the segment-based
2925 * weighting algorithm. The metaslab must be loaded. This function can
2926 * be called within a sync pass since it relies only on the metaslab's
2927 * range tree which is always accurate when the metaslab is loaded.
2928 */
2929static uint64_t
2930metaslab_weight_from_range_tree(metaslab_t *msp)
2931{
2932	uint64_t weight = 0;
2933	uint32_t segments = 0;
2934
2935	ASSERT(msp->ms_loaded);
2936
2937	for (int i = RANGE_TREE_HISTOGRAM_SIZE - 1; i >= SPA_MINBLOCKSHIFT;
2938	    i--) {
2939		uint8_t shift = msp->ms_group->mg_vd->vdev_ashift;
2940		int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1;
2941
2942		segments <<= 1;
2943		segments += msp->ms_allocatable->rt_histogram[i];
2944
2945		/*
2946		 * The range tree provides more precision than the space map
2947		 * and must be downgraded so that all values fit within the
2948		 * space map's histogram. This allows us to compare loaded
2949		 * vs. unloaded metaslabs to determine which metaslab is
2950		 * considered "best".
2951		 */
2952		if (i > max_idx)
2953			continue;
2954
2955		if (segments != 0) {
2956			WEIGHT_SET_COUNT(weight, segments);
2957			WEIGHT_SET_INDEX(weight, i);
2958			WEIGHT_SET_ACTIVE(weight, 0);
2959			break;
2960		}
2961	}
2962	return (weight);
2963}
2964
2965/*
2966 * Calculate the weight based on the on-disk histogram. Should be applied
2967 * only to unloaded metaslabs  (i.e no incoming allocations) in-order to
2968 * give results consistent with the on-disk state
2969 */
2970static uint64_t
2971metaslab_weight_from_spacemap(metaslab_t *msp)
2972{
2973	space_map_t *sm = msp->ms_sm;
2974	ASSERT(!msp->ms_loaded);
2975	ASSERT(sm != NULL);
2976	ASSERT3U(space_map_object(sm), !=, 0);
2977	ASSERT3U(sm->sm_dbuf->db_size, ==, sizeof (space_map_phys_t));
2978
2979	/*
2980	 * Create a joint histogram from all the segments that have made
2981	 * it to the metaslab's space map histogram, that are not yet
2982	 * available for allocation because they are still in the freeing
2983	 * pipeline (e.g. freeing, freed, and defer trees). Then subtract
2984	 * these segments from the space map's histogram to get a more
2985	 * accurate weight.
2986	 */
2987	uint64_t deferspace_histogram[SPACE_MAP_HISTOGRAM_SIZE] = {0};
2988	for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++)
2989		deferspace_histogram[i] += msp->ms_synchist[i];
2990	for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2991		for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
2992			deferspace_histogram[i] += msp->ms_deferhist[t][i];
2993		}
2994	}
2995
2996	uint64_t weight = 0;
2997	for (int i = SPACE_MAP_HISTOGRAM_SIZE - 1; i >= 0; i--) {
2998		ASSERT3U(sm->sm_phys->smp_histogram[i], >=,
2999		    deferspace_histogram[i]);
3000		uint64_t count =
3001		    sm->sm_phys->smp_histogram[i] - deferspace_histogram[i];
3002		if (count != 0) {
3003			WEIGHT_SET_COUNT(weight, count);
3004			WEIGHT_SET_INDEX(weight, i + sm->sm_shift);
3005			WEIGHT_SET_ACTIVE(weight, 0);
3006			break;
3007		}
3008	}
3009	return (weight);
3010}
3011
3012/*
3013 * Compute a segment-based weight for the specified metaslab. The weight
3014 * is determined by highest bucket in the histogram. The information
3015 * for the highest bucket is encoded into the weight value.
3016 */
3017static uint64_t
3018metaslab_segment_weight(metaslab_t *msp)
3019{
3020	metaslab_group_t *mg = msp->ms_group;
3021	uint64_t weight = 0;
3022	uint8_t shift = mg->mg_vd->vdev_ashift;
3023
3024	ASSERT(MUTEX_HELD(&msp->ms_lock));
3025
3026	/*
3027	 * The metaslab is completely free.
3028	 */
3029	if (metaslab_allocated_space(msp) == 0) {
3030		int idx = highbit64(msp->ms_size) - 1;
3031		int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1;
3032
3033		if (idx < max_idx) {
3034			WEIGHT_SET_COUNT(weight, 1ULL);
3035			WEIGHT_SET_INDEX(weight, idx);
3036		} else {
3037			WEIGHT_SET_COUNT(weight, 1ULL << (idx - max_idx));
3038			WEIGHT_SET_INDEX(weight, max_idx);
3039		}
3040		WEIGHT_SET_ACTIVE(weight, 0);
3041		ASSERT(!WEIGHT_IS_SPACEBASED(weight));
3042		return (weight);
3043	}
3044
3045	ASSERT3U(msp->ms_sm->sm_dbuf->db_size, ==, sizeof (space_map_phys_t));
3046
3047	/*
3048	 * If the metaslab is fully allocated then just make the weight 0.
3049	 */
3050	if (metaslab_allocated_space(msp) == msp->ms_size)
3051		return (0);
3052	/*
3053	 * If the metaslab is already loaded, then use the range tree to
3054	 * determine the weight. Otherwise, we rely on the space map information
3055	 * to generate the weight.
3056	 */
3057	if (msp->ms_loaded) {
3058		weight = metaslab_weight_from_range_tree(msp);
3059	} else {
3060		weight = metaslab_weight_from_spacemap(msp);
3061	}
3062
3063	/*
3064	 * If the metaslab was active the last time we calculated its weight
3065	 * then keep it active. We want to consume the entire region that
3066	 * is associated with this weight.
3067	 */
3068	if (msp->ms_activation_weight != 0 && weight != 0)
3069		WEIGHT_SET_ACTIVE(weight, WEIGHT_GET_ACTIVE(msp->ms_weight));
3070	return (weight);
3071}
3072
3073/*
3074 * Determine if we should attempt to allocate from this metaslab. If the
3075 * metaslab is loaded, then we can determine if the desired allocation
3076 * can be satisfied by looking at the size of the maximum free segment
3077 * on that metaslab. Otherwise, we make our decision based on the metaslab's
3078 * weight. For segment-based weighting we can determine the maximum
3079 * allocation based on the index encoded in its value. For space-based
3080 * weights we rely on the entire weight (excluding the weight-type bit).
3081 */
3082boolean_t
3083metaslab_should_allocate(metaslab_t *msp, uint64_t asize, boolean_t try_hard)
3084{
3085	/*
3086	 * If the metaslab is loaded, ms_max_size is definitive and we can use
3087	 * the fast check. If it's not, the ms_max_size is a lower bound (once
3088	 * set), and we should use the fast check as long as we're not in
3089	 * try_hard and it's been less than zfs_metaslab_max_size_cache_sec
3090	 * seconds since the metaslab was unloaded.
3091	 */
3092	if (msp->ms_loaded ||
3093	    (msp->ms_max_size != 0 && !try_hard && gethrtime() <
3094	    msp->ms_unload_time + SEC2NSEC(zfs_metaslab_max_size_cache_sec)))
3095		return (msp->ms_max_size >= asize);
3096
3097	boolean_t should_allocate;
3098	if (!WEIGHT_IS_SPACEBASED(msp->ms_weight)) {
3099		/*
3100		 * The metaslab segment weight indicates segments in the
3101		 * range [2^i, 2^(i+1)), where i is the index in the weight.
3102		 * Since the asize might be in the middle of the range, we
3103		 * should attempt the allocation if asize < 2^(i+1).
3104		 */
3105		should_allocate = (asize <
3106		    1ULL << (WEIGHT_GET_INDEX(msp->ms_weight) + 1));
3107	} else {
3108		should_allocate = (asize <=
3109		    (msp->ms_weight & ~METASLAB_WEIGHT_TYPE));
3110	}
3111
3112	return (should_allocate);
3113}
3114
3115static uint64_t
3116metaslab_weight(metaslab_t *msp)
3117{
3118	vdev_t *vd = msp->ms_group->mg_vd;
3119	spa_t *spa = vd->vdev_spa;
3120	uint64_t weight;
3121
3122	ASSERT(MUTEX_HELD(&msp->ms_lock));
3123
3124	metaslab_set_fragmentation(msp);
3125
3126	/*
3127	 * Update the maximum size. If the metaslab is loaded, this will
3128	 * ensure that we get an accurate maximum size if newly freed space
3129	 * has been added back into the free tree. If the metaslab is
3130	 * unloaded, we check if there's a larger free segment in the
3131	 * unflushed frees. This is a lower bound on the largest allocatable
3132	 * segment size. Coalescing of adjacent entries may reveal larger
3133	 * allocatable segments, but we aren't aware of those until loading
3134	 * the space map into a range tree.
3135	 */
3136	if (msp->ms_loaded) {
3137		msp->ms_max_size = metaslab_largest_allocatable(msp);
3138	} else {
3139		msp->ms_max_size = MAX(msp->ms_max_size,
3140		    metaslab_largest_unflushed_free(msp));
3141	}
3142
3143	/*
3144	 * Segment-based weighting requires space map histogram support.
3145	 */
3146	if (zfs_metaslab_segment_weight_enabled &&
3147	    spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
3148	    (msp->ms_sm == NULL || msp->ms_sm->sm_dbuf->db_size ==
3149	    sizeof (space_map_phys_t))) {
3150		weight = metaslab_segment_weight(msp);
3151	} else {
3152		weight = metaslab_space_weight(msp);
3153	}
3154	return (weight);
3155}
3156
3157void
3158metaslab_recalculate_weight_and_sort(metaslab_t *msp)
3159{
3160	ASSERT(MUTEX_HELD(&msp->ms_lock));
3161
3162	/* note: we preserve the mask (e.g. indication of primary, etc..) */
3163	uint64_t was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
3164	metaslab_group_sort(msp->ms_group, msp,
3165	    metaslab_weight(msp) | was_active);
3166}
3167
3168static int
3169metaslab_activate_allocator(metaslab_group_t *mg, metaslab_t *msp,
3170    int allocator, uint64_t activation_weight)
3171{
3172	ASSERT(MUTEX_HELD(&msp->ms_lock));
3173
3174	/*
3175	 * If we're activating for the claim code, we don't want to actually
3176	 * set the metaslab up for a specific allocator.
3177	 */
3178	if (activation_weight == METASLAB_WEIGHT_CLAIM) {
3179		ASSERT0(msp->ms_activation_weight);
3180		msp->ms_activation_weight = msp->ms_weight;
3181		metaslab_group_sort(mg, msp, msp->ms_weight |
3182		    activation_weight);
3183		return (0);
3184	}
3185
3186	metaslab_t **arr = (activation_weight == METASLAB_WEIGHT_PRIMARY ?
3187	    mg->mg_primaries : mg->mg_secondaries);
3188
3189	mutex_enter(&mg->mg_lock);
3190	if (arr[allocator] != NULL) {
3191		mutex_exit(&mg->mg_lock);
3192		return (EEXIST);
3193	}
3194
3195	arr[allocator] = msp;
3196	ASSERT3S(msp->ms_allocator, ==, -1);
3197	msp->ms_allocator = allocator;
3198	msp->ms_primary = (activation_weight == METASLAB_WEIGHT_PRIMARY);
3199
3200	ASSERT0(msp->ms_activation_weight);
3201	msp->ms_activation_weight = msp->ms_weight;
3202	metaslab_group_sort_impl(mg, msp,
3203	    msp->ms_weight | activation_weight);
3204
3205	mutex_exit(&mg->mg_lock);
3206
3207	return (0);
3208}
3209
3210static int
3211metaslab_activate(metaslab_t *msp, int allocator, uint64_t activation_weight)
3212{
3213	ASSERT(MUTEX_HELD(&msp->ms_lock));
3214
3215	/*
3216	 * The current metaslab is already activated for us so there
3217	 * is nothing to do. Already activated though, doesn't mean
3218	 * that this metaslab is activated for our allocator nor our
3219	 * requested activation weight. The metaslab could have started
3220	 * as an active one for our allocator but changed allocators
3221	 * while we were waiting to grab its ms_lock or we stole it
3222	 * [see find_valid_metaslab()]. This means that there is a
3223	 * possibility of passivating a metaslab of another allocator
3224	 * or from a different activation mask, from this thread.
3225	 */
3226	if ((msp->ms_weight & METASLAB_ACTIVE_MASK) != 0) {
3227		ASSERT(msp->ms_loaded);
3228		return (0);
3229	}
3230
3231	int error = metaslab_load(msp);
3232	if (error != 0) {
3233		metaslab_group_sort(msp->ms_group, msp, 0);
3234		return (error);
3235	}
3236
3237	/*
3238	 * When entering metaslab_load() we may have dropped the
3239	 * ms_lock because we were loading this metaslab, or we
3240	 * were waiting for another thread to load it for us. In
3241	 * that scenario, we recheck the weight of the metaslab
3242	 * to see if it was activated by another thread.
3243	 *
3244	 * If the metaslab was activated for another allocator or
3245	 * it was activated with a different activation weight (e.g.
3246	 * we wanted to make it a primary but it was activated as
3247	 * secondary) we return error (EBUSY).
3248	 *
3249	 * If the metaslab was activated for the same allocator
3250	 * and requested activation mask, skip activating it.
3251	 */
3252	if ((msp->ms_weight & METASLAB_ACTIVE_MASK) != 0) {
3253		if (msp->ms_allocator != allocator)
3254			return (EBUSY);
3255
3256		if ((msp->ms_weight & activation_weight) == 0)
3257			return (EBUSY);
3258
3259		EQUIV((activation_weight == METASLAB_WEIGHT_PRIMARY),
3260		    msp->ms_primary);
3261		return (0);
3262	}
3263
3264	/*
3265	 * If the metaslab has literally 0 space, it will have weight 0. In
3266	 * that case, don't bother activating it. This can happen if the
3267	 * metaslab had space during find_valid_metaslab, but another thread
3268	 * loaded it and used all that space while we were waiting to grab the
3269	 * lock.
3270	 */
3271	if (msp->ms_weight == 0) {
3272		ASSERT0(range_tree_space(msp->ms_allocatable));
3273		return (SET_ERROR(ENOSPC));
3274	}
3275
3276	if ((error = metaslab_activate_allocator(msp->ms_group, msp,
3277	    allocator, activation_weight)) != 0) {
3278		return (error);
3279	}
3280
3281	ASSERT(msp->ms_loaded);
3282	ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
3283
3284	return (0);
3285}
3286
3287static void
3288metaslab_passivate_allocator(metaslab_group_t *mg, metaslab_t *msp,
3289    uint64_t weight)
3290{
3291	ASSERT(MUTEX_HELD(&msp->ms_lock));
3292	ASSERT(msp->ms_loaded);
3293
3294	if (msp->ms_weight & METASLAB_WEIGHT_CLAIM) {
3295		metaslab_group_sort(mg, msp, weight);
3296		return;
3297	}
3298
3299	mutex_enter(&mg->mg_lock);
3300	ASSERT3P(msp->ms_group, ==, mg);
3301	ASSERT3S(0, <=, msp->ms_allocator);
3302	ASSERT3U(msp->ms_allocator, <, mg->mg_allocators);
3303
3304	if (msp->ms_primary) {
3305		ASSERT3P(mg->mg_primaries[msp->ms_allocator], ==, msp);
3306		ASSERT(msp->ms_weight & METASLAB_WEIGHT_PRIMARY);
3307		mg->mg_primaries[msp->ms_allocator] = NULL;
3308	} else {
3309		ASSERT3P(mg->mg_secondaries[msp->ms_allocator], ==, msp);
3310		ASSERT(msp->ms_weight & METASLAB_WEIGHT_SECONDARY);
3311		mg->mg_secondaries[msp->ms_allocator] = NULL;
3312	}
3313	msp->ms_allocator = -1;
3314	metaslab_group_sort_impl(mg, msp, weight);
3315	mutex_exit(&mg->mg_lock);
3316}
3317
3318static void
3319metaslab_passivate(metaslab_t *msp, uint64_t weight)
3320{
3321	uint64_t size = weight & ~METASLAB_WEIGHT_TYPE;
3322
3323	/*
3324	 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
3325	 * this metaslab again.  In that case, it had better be empty,
3326	 * or we would be leaving space on the table.
3327	 */
3328	ASSERT(size >= SPA_MINBLOCKSIZE ||
3329	    range_tree_is_empty(msp->ms_allocatable));
3330	ASSERT0(weight & METASLAB_ACTIVE_MASK);
3331
3332	ASSERT(msp->ms_activation_weight != 0);
3333	msp->ms_activation_weight = 0;
3334	metaslab_passivate_allocator(msp->ms_group, msp, weight);
3335	ASSERT0(msp->ms_weight & METASLAB_ACTIVE_MASK);
3336}
3337
3338/*
3339 * Segment-based metaslabs are activated once and remain active until
3340 * we either fail an allocation attempt (similar to space-based metaslabs)
3341 * or have exhausted the free space in zfs_metaslab_switch_threshold
3342 * buckets since the metaslab was activated. This function checks to see
3343 * if we've exhaused the zfs_metaslab_switch_threshold buckets in the
3344 * metaslab and passivates it proactively. This will allow us to select a
3345 * metaslabs with larger contiguous region if any remaining within this
3346 * metaslab group. If we're in sync pass > 1, then we continue using this
3347 * metaslab so that we don't dirty more block and cause more sync passes.
3348 */
3349void
3350metaslab_segment_may_passivate(metaslab_t *msp)
3351{
3352	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
3353
3354	if (WEIGHT_IS_SPACEBASED(msp->ms_weight) || spa_sync_pass(spa) > 1)
3355		return;
3356
3357	/*
3358	 * Since we are in the middle of a sync pass, the most accurate
3359	 * information that is accessible to us is the in-core range tree
3360	 * histogram; calculate the new weight based on that information.
3361	 */
3362	uint64_t weight = metaslab_weight_from_range_tree(msp);
3363	int activation_idx = WEIGHT_GET_INDEX(msp->ms_activation_weight);
3364	int current_idx = WEIGHT_GET_INDEX(weight);
3365
3366	if (current_idx <= activation_idx - zfs_metaslab_switch_threshold)
3367		metaslab_passivate(msp, weight);
3368}
3369
3370static void
3371metaslab_preload(void *arg)
3372{
3373	metaslab_t *msp = arg;
3374	metaslab_class_t *mc = msp->ms_group->mg_class;
3375	spa_t *spa = mc->mc_spa;
3376
3377	ASSERT(!MUTEX_HELD(&msp->ms_group->mg_lock));
3378
3379	mutex_enter(&msp->ms_lock);
3380	(void) metaslab_load(msp);
3381	metaslab_set_selected_txg(msp, spa_syncing_txg(spa));
3382	mutex_exit(&msp->ms_lock);
3383}
3384
3385static void
3386metaslab_group_preload(metaslab_group_t *mg)
3387{
3388	spa_t *spa = mg->mg_vd->vdev_spa;
3389	metaslab_t *msp;
3390	avl_tree_t *t = &mg->mg_metaslab_tree;
3391	int m = 0;
3392
3393	if (spa_shutting_down(spa) || !metaslab_preload_enabled) {
3394		taskq_wait(mg->mg_taskq);
3395		return;
3396	}
3397
3398	mutex_enter(&mg->mg_lock);
3399
3400	/*
3401	 * Load the next potential metaslabs
3402	 */
3403	for (msp = avl_first(t); msp != NULL; msp = AVL_NEXT(t, msp)) {
3404		ASSERT3P(msp->ms_group, ==, mg);
3405
3406		/*
3407		 * We preload only the maximum number of metaslabs specified
3408		 * by metaslab_preload_limit. If a metaslab is being forced
3409		 * to condense then we preload it too. This will ensure
3410		 * that force condensing happens in the next txg.
3411		 */
3412		if (++m > metaslab_preload_limit && !msp->ms_condense_wanted) {
3413			continue;
3414		}
3415
3416		VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
3417		    msp, TQ_SLEEP) != TASKQID_INVALID);
3418	}
3419	mutex_exit(&mg->mg_lock);
3420}
3421
3422/*
3423 * Determine if the space map's on-disk footprint is past our tolerance for
3424 * inefficiency. We would like to use the following criteria to make our
3425 * decision:
3426 *
3427 * 1. Do not condense if the size of the space map object would dramatically
3428 *    increase as a result of writing out the free space range tree.
3429 *
3430 * 2. Condense if the on on-disk space map representation is at least
3431 *    zfs_condense_pct/100 times the size of the optimal representation
3432 *    (i.e. zfs_condense_pct = 110 and in-core = 1MB, optimal = 1.1MB).
3433 *
3434 * 3. Do not condense if the on-disk size of the space map does not actually
3435 *    decrease.
3436 *
3437 * Unfortunately, we cannot compute the on-disk size of the space map in this
3438 * context because we cannot accurately compute the effects of compression, etc.
3439 * Instead, we apply the heuristic described in the block comment for
3440 * zfs_metaslab_condense_block_threshold - we only condense if the space used
3441 * is greater than a threshold number of blocks.
3442 */
3443static boolean_t
3444metaslab_should_condense(metaslab_t *msp)
3445{
3446	space_map_t *sm = msp->ms_sm;
3447	vdev_t *vd = msp->ms_group->mg_vd;
3448	uint64_t vdev_blocksize = 1 << vd->vdev_ashift;
3449
3450	ASSERT(MUTEX_HELD(&msp->ms_lock));
3451	ASSERT(msp->ms_loaded);
3452	ASSERT(sm != NULL);
3453	ASSERT3U(spa_sync_pass(vd->vdev_spa), ==, 1);
3454
3455	/*
3456	 * We always condense metaslabs that are empty and metaslabs for
3457	 * which a condense request has been made.
3458	 */
3459	if (range_tree_numsegs(msp->ms_allocatable) == 0 ||
3460	    msp->ms_condense_wanted)
3461		return (B_TRUE);
3462
3463	uint64_t record_size = MAX(sm->sm_blksz, vdev_blocksize);
3464	uint64_t object_size = space_map_length(sm);
3465	uint64_t optimal_size = space_map_estimate_optimal_size(sm,
3466	    msp->ms_allocatable, SM_NO_VDEVID);
3467
3468	return (object_size >= (optimal_size * zfs_condense_pct / 100) &&
3469	    object_size > zfs_metaslab_condense_block_threshold * record_size);
3470}
3471
3472/*
3473 * Condense the on-disk space map representation to its minimized form.
3474 * The minimized form consists of a small number of allocations followed
3475 * by the entries of the free range tree (ms_allocatable). The condensed
3476 * spacemap contains all the entries of previous TXGs (including those in
3477 * the pool-wide log spacemaps; thus this is effectively a superset of
3478 * metaslab_flush()), but this TXG's entries still need to be written.
3479 */
3480static void
3481metaslab_condense(metaslab_t *msp, dmu_tx_t *tx)
3482{
3483	range_tree_t *condense_tree;
3484	space_map_t *sm = msp->ms_sm;
3485	uint64_t txg = dmu_tx_get_txg(tx);
3486	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
3487
3488	ASSERT(MUTEX_HELD(&msp->ms_lock));
3489	ASSERT(msp->ms_loaded);
3490	ASSERT(msp->ms_sm != NULL);
3491
3492	/*
3493	 * In order to condense the space map, we need to change it so it
3494	 * only describes which segments are currently allocated and free.
3495	 *
3496	 * All the current free space resides in the ms_allocatable, all
3497	 * the ms_defer trees, and all the ms_allocating trees. We ignore
3498	 * ms_freed because it is empty because we're in sync pass 1. We
3499	 * ignore ms_freeing because these changes are not yet reflected
3500	 * in the spacemap (they will be written later this txg).
3501	 *
3502	 * So to truncate the space map to represent all the entries of
3503	 * previous TXGs we do the following:
3504	 *
3505	 * 1] We create a range tree (condense tree) that is 100% empty.
3506	 * 2] We add to it all segments found in the ms_defer trees
3507	 *    as those segments are marked as free in the original space
3508	 *    map. We do the same with the ms_allocating trees for the same
3509	 *    reason. Adding these segments should be a relatively
3510	 *    inexpensive operation since we expect these trees to have a
3511	 *    small number of nodes.
3512	 * 3] We vacate any unflushed allocs, since they are not frees we
3513	 *    need to add to the condense tree. Then we vacate any
3514	 *    unflushed frees as they should already be part of ms_allocatable.
3515	 * 4] At this point, we would ideally like to add all segments
3516	 *    in the ms_allocatable tree from the condense tree. This way
3517	 *    we would write all the entries of the condense tree as the
3518	 *    condensed space map, which would only contain freeed
3519	 *    segments with everything else assumed to be allocated.
3520	 *
3521	 *    Doing so can be prohibitively expensive as ms_allocatable can
3522	 *    be large, and therefore computationally expensive to add to
3523	 *    the condense_tree. Instead we first sync out an entry marking
3524	 *    everything as allocated, then the condense_tree and then the
3525	 *    ms_allocatable, in the condensed space map. While this is not
3526	 *    optimal, it is typically close to optimal and more importantly
3527	 *    much cheaper to compute.
3528	 *
3529	 * 5] Finally, as both of the unflushed trees were written to our
3530	 *    new and condensed metaslab space map, we basically flushed
3531	 *    all the unflushed changes to disk, thus we call
3532	 *    metaslab_flush_update().
3533	 */
3534	ASSERT3U(spa_sync_pass(spa), ==, 1);
3535	ASSERT(range_tree_is_empty(msp->ms_freed)); /* since it is pass 1 */
3536
3537	zfs_dbgmsg("condensing: txg %llu, msp[%llu] %p, vdev id %llu, "
3538	    "spa %s, smp size %llu, segments %lu, forcing condense=%s", txg,
3539	    msp->ms_id, msp, msp->ms_group->mg_vd->vdev_id,
3540	    spa->spa_name, space_map_length(msp->ms_sm),
3541	    range_tree_numsegs(msp->ms_allocatable),
3542	    msp->ms_condense_wanted ? "TRUE" : "FALSE");
3543
3544	msp->ms_condense_wanted = B_FALSE;
3545
3546	range_seg_type_t type;
3547	uint64_t shift, start;
3548	type = metaslab_calculate_range_tree_type(msp->ms_group->mg_vd, msp,
3549	    &start, &shift);
3550
3551	condense_tree = range_tree_create(NULL, type, NULL, start, shift);
3552
3553	for (int t = 0; t < TXG_DEFER_SIZE; t++) {
3554		range_tree_walk(msp->ms_defer[t],
3555		    range_tree_add, condense_tree);
3556	}
3557
3558	for (int t = 0; t < TXG_CONCURRENT_STATES; t++) {
3559		range_tree_walk(msp->ms_allocating[(txg + t) & TXG_MASK],
3560		    range_tree_add, condense_tree);
3561	}
3562
3563	ASSERT3U(spa->spa_unflushed_stats.sus_memused, >=,
3564	    metaslab_unflushed_changes_memused(msp));
3565	spa->spa_unflushed_stats.sus_memused -=
3566	    metaslab_unflushed_changes_memused(msp);
3567	range_tree_vacate(msp->ms_unflushed_allocs, NULL, NULL);
3568	range_tree_vacate(msp->ms_unflushed_frees, NULL, NULL);
3569
3570	/*
3571	 * We're about to drop the metaslab's lock thus allowing other
3572	 * consumers to change its content. Set the metaslab's ms_condensing
3573	 * flag to ensure that allocations on this metaslab do not occur
3574	 * while we're in the middle of committing it to disk. This is only
3575	 * critical for ms_allocatable as all other range trees use per TXG
3576	 * views of their content.
3577	 */
3578	msp->ms_condensing = B_TRUE;
3579
3580	mutex_exit(&msp->ms_lock);
3581	uint64_t object = space_map_object(msp->ms_sm);
3582	space_map_truncate(sm,
3583	    spa_feature_is_enabled(spa, SPA_FEATURE_LOG_SPACEMAP) ?
3584	    zfs_metaslab_sm_blksz_with_log : zfs_metaslab_sm_blksz_no_log, tx);
3585
3586	/*
3587	 * space_map_truncate() may have reallocated the spacemap object.
3588	 * If so, update the vdev_ms_array.
3589	 */
3590	if (space_map_object(msp->ms_sm) != object) {
3591		object = space_map_object(msp->ms_sm);
3592		dmu_write(spa->spa_meta_objset,
3593		    msp->ms_group->mg_vd->vdev_ms_array, sizeof (uint64_t) *
3594		    msp->ms_id, sizeof (uint64_t), &object, tx);
3595	}
3596
3597	/*
3598	 * Note:
3599	 * When the log space map feature is enabled, each space map will
3600	 * always have ALLOCS followed by FREES for each sync pass. This is
3601	 * typically true even when the log space map feature is disabled,
3602	 * except from the case where a metaslab goes through metaslab_sync()
3603	 * and gets condensed. In that case the metaslab's space map will have
3604	 * ALLOCS followed by FREES (due to condensing) followed by ALLOCS
3605	 * followed by FREES (due to space_map_write() in metaslab_sync()) for
3606	 * sync pass 1.
3607	 */
3608	range_tree_t *tmp_tree = range_tree_create(NULL, type, NULL, start,
3609	    shift);
3610	range_tree_add(tmp_tree, msp->ms_start, msp->ms_size);
3611	space_map_write(sm, tmp_tree, SM_ALLOC, SM_NO_VDEVID, tx);
3612	space_map_write(sm, msp->ms_allocatable, SM_FREE, SM_NO_VDEVID, tx);
3613	space_map_write(sm, condense_tree, SM_FREE, SM_NO_VDEVID, tx);
3614
3615	range_tree_vacate(condense_tree, NULL, NULL);
3616	range_tree_destroy(condense_tree);
3617	range_tree_vacate(tmp_tree, NULL, NULL);
3618	range_tree_destroy(tmp_tree);
3619	mutex_enter(&msp->ms_lock);
3620
3621	msp->ms_condensing = B_FALSE;
3622	metaslab_flush_update(msp, tx);
3623}
3624
3625/*
3626 * Called when the metaslab has been flushed (its own spacemap now reflects
3627 * all the contents of the pool-wide spacemap log). Updates the metaslab's
3628 * metadata and any pool-wide related log space map data (e.g. summary,
3629 * obsolete logs, etc.) to reflect that.
3630 */
3631static void
3632metaslab_flush_update(metaslab_t *msp, dmu_tx_t *tx)
3633{
3634	metaslab_group_t *mg = msp->ms_group;
3635	spa_t *spa = mg->mg_vd->vdev_spa;
3636
3637	ASSERT(MUTEX_HELD(&msp->ms_lock));
3638
3639	ASSERT3U(spa_sync_pass(spa), ==, 1);
3640	ASSERT(range_tree_is_empty(msp->ms_unflushed_allocs));
3641	ASSERT(range_tree_is_empty(msp->ms_unflushed_frees));
3642
3643	/*
3644	 * Just because a metaslab got flushed, that doesn't mean that
3645	 * it will pass through metaslab_sync_done(). Thus, make sure to
3646	 * update ms_synced_length here in case it doesn't.
3647	 */
3648	msp->ms_synced_length = space_map_length(msp->ms_sm);
3649
3650	/*
3651	 * We may end up here from metaslab_condense() without the
3652	 * feature being active. In that case this is a no-op.
3653	 */
3654	if (!spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP))
3655		return;
3656
3657	ASSERT(spa_syncing_log_sm(spa) != NULL);
3658	ASSERT(msp->ms_sm != NULL);
3659	ASSERT(metaslab_unflushed_txg(msp) != 0);
3660	ASSERT3P(avl_find(&spa->spa_metaslabs_by_flushed, msp, NULL), ==, msp);
3661
3662	VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(spa));
3663
3664	/* update metaslab's position in our flushing tree */
3665	uint64_t ms_prev_flushed_txg = metaslab_unflushed_txg(msp);
3666	mutex_enter(&spa->spa_flushed_ms_lock);
3667	avl_remove(&spa->spa_metaslabs_by_flushed, msp);
3668	metaslab_set_unflushed_txg(msp, spa_syncing_txg(spa), tx);
3669	avl_add(&spa->spa_metaslabs_by_flushed, msp);
3670	mutex_exit(&spa->spa_flushed_ms_lock);
3671
3672	/* update metaslab counts of spa_log_sm_t nodes */
3673	spa_log_sm_decrement_mscount(spa, ms_prev_flushed_txg);
3674	spa_log_sm_increment_current_mscount(spa);
3675
3676	/* cleanup obsolete logs if any */
3677	uint64_t log_blocks_before = spa_log_sm_nblocks(spa);
3678	spa_cleanup_old_sm_logs(spa, tx);
3679	uint64_t log_blocks_after = spa_log_sm_nblocks(spa);
3680	VERIFY3U(log_blocks_after, <=, log_blocks_before);
3681
3682	/* update log space map summary */
3683	uint64_t blocks_gone = log_blocks_before - log_blocks_after;
3684	spa_log_summary_add_flushed_metaslab(spa);
3685	spa_log_summary_decrement_mscount(spa, ms_prev_flushed_txg);
3686	spa_log_summary_decrement_blkcount(spa, blocks_gone);
3687}
3688
3689boolean_t
3690metaslab_flush(metaslab_t *msp, dmu_tx_t *tx)
3691{
3692	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
3693
3694	ASSERT(MUTEX_HELD(&msp->ms_lock));
3695	ASSERT3U(spa_sync_pass(spa), ==, 1);
3696	ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP));
3697
3698	ASSERT(msp->ms_sm != NULL);
3699	ASSERT(metaslab_unflushed_txg(msp) != 0);
3700	ASSERT(avl_find(&spa->spa_metaslabs_by_flushed, msp, NULL) != NULL);
3701
3702	/*
3703	 * There is nothing wrong with flushing the same metaslab twice, as
3704	 * this codepath should work on that case. However, the current
3705	 * flushing scheme makes sure to avoid this situation as we would be
3706	 * making all these calls without having anything meaningful to write
3707	 * to disk. We assert this behavior here.
3708	 */
3709	ASSERT3U(metaslab_unflushed_txg(msp), <, dmu_tx_get_txg(tx));
3710
3711	/*
3712	 * We can not flush while loading, because then we would
3713	 * not load the ms_unflushed_{allocs,frees}.
3714	 */
3715	if (msp->ms_loading)
3716		return (B_FALSE);
3717
3718	metaslab_verify_space(msp, dmu_tx_get_txg(tx));
3719	metaslab_verify_weight_and_frag(msp);
3720
3721	/*
3722	 * Metaslab condensing is effectively flushing. Therefore if the
3723	 * metaslab can be condensed we can just condense it instead of
3724	 * flushing it.
3725	 *
3726	 * Note that metaslab_condense() does call metaslab_flush_update()
3727	 * so we can just return immediately after condensing. We also
3728	 * don't need to care about setting ms_flushing or broadcasting
3729	 * ms_flush_cv, even if we temporarily drop the ms_lock in
3730	 * metaslab_condense(), as the metaslab is already loaded.
3731	 */
3732	if (msp->ms_loaded && metaslab_should_condense(msp)) {
3733		metaslab_group_t *mg = msp->ms_group;
3734
3735		/*
3736		 * For all histogram operations below refer to the
3737		 * comments of metaslab_sync() where we follow a
3738		 * similar procedure.
3739		 */
3740		metaslab_group_histogram_verify(mg);
3741		metaslab_class_histogram_verify(mg->mg_class);
3742		metaslab_group_histogram_remove(mg, msp);
3743
3744		metaslab_condense(msp, tx);
3745
3746		space_map_histogram_clear(msp->ms_sm);
3747		space_map_histogram_add(msp->ms_sm, msp->ms_allocatable, tx);
3748		ASSERT(range_tree_is_empty(msp->ms_freed));
3749		for (int t = 0; t < TXG_DEFER_SIZE; t++) {
3750			space_map_histogram_add(msp->ms_sm,
3751			    msp->ms_defer[t], tx);
3752		}
3753		metaslab_aux_histograms_update(msp);
3754
3755		metaslab_group_histogram_add(mg, msp);
3756		metaslab_group_histogram_verify(mg);
3757		metaslab_class_histogram_verify(mg->mg_class);
3758
3759		metaslab_verify_space(msp, dmu_tx_get_txg(tx));
3760
3761		/*
3762		 * Since we recreated the histogram (and potentially
3763		 * the ms_sm too while condensing) ensure that the
3764		 * weight is updated too because we are not guaranteed
3765		 * that this metaslab is dirty and will go through
3766		 * metaslab_sync_done().
3767		 */
3768		metaslab_recalculate_weight_and_sort(msp);
3769		return (B_TRUE);
3770	}
3771
3772	msp->ms_flushing = B_TRUE;
3773	uint64_t sm_len_before = space_map_length(msp->ms_sm);
3774
3775	mutex_exit(&msp->ms_lock);
3776	space_map_write(msp->ms_sm, msp->ms_unflushed_allocs, SM_ALLOC,
3777	    SM_NO_VDEVID, tx);
3778	space_map_write(msp->ms_sm, msp->ms_unflushed_frees, SM_FREE,
3779	    SM_NO_VDEVID, tx);
3780	mutex_enter(&msp->ms_lock);
3781
3782	uint64_t sm_len_after = space_map_length(msp->ms_sm);
3783	if (zfs_flags & ZFS_DEBUG_LOG_SPACEMAP) {
3784		zfs_dbgmsg("flushing: txg %llu, spa %s, vdev_id %llu, "
3785		    "ms_id %llu, unflushed_allocs %llu, unflushed_frees %llu, "
3786		    "appended %llu bytes", dmu_tx_get_txg(tx), spa_name(spa),
3787		    msp->ms_group->mg_vd->vdev_id, msp->ms_id,
3788		    range_tree_space(msp->ms_unflushed_allocs),
3789		    range_tree_space(msp->ms_unflushed_frees),
3790		    (sm_len_after - sm_len_before));
3791	}
3792
3793	ASSERT3U(spa->spa_unflushed_stats.sus_memused, >=,
3794	    metaslab_unflushed_changes_memused(msp));
3795	spa->spa_unflushed_stats.sus_memused -=
3796	    metaslab_unflushed_changes_memused(msp);
3797	range_tree_vacate(msp->ms_unflushed_allocs, NULL, NULL);
3798	range_tree_vacate(msp->ms_unflushed_frees, NULL, NULL);
3799
3800	metaslab_verify_space(msp, dmu_tx_get_txg(tx));
3801	metaslab_verify_weight_and_frag(msp);
3802
3803	metaslab_flush_update(msp, tx);
3804
3805	metaslab_verify_space(msp, dmu_tx_get_txg(tx));
3806	metaslab_verify_weight_and_frag(msp);
3807
3808	msp->ms_flushing = B_FALSE;
3809	cv_broadcast(&msp->ms_flush_cv);
3810	return (B_TRUE);
3811}
3812
3813/*
3814 * Write a metaslab to disk in the context of the specified transaction group.
3815 */
3816void
3817metaslab_sync(metaslab_t *msp, uint64_t txg)
3818{
3819	metaslab_group_t *mg = msp->ms_group;
3820	vdev_t *vd = mg->mg_vd;
3821	spa_t *spa = vd->vdev_spa;
3822	objset_t *mos = spa_meta_objset(spa);
3823	range_tree_t *alloctree = msp->ms_allocating[txg & TXG_MASK];
3824	dmu_tx_t *tx;
3825
3826	ASSERT(!vd->vdev_ishole);
3827
3828	/*
3829	 * This metaslab has just been added so there's no work to do now.
3830	 */
3831	if (msp->ms_freeing == NULL) {
3832		ASSERT3P(alloctree, ==, NULL);
3833		return;
3834	}
3835
3836	ASSERT3P(alloctree, !=, NULL);
3837	ASSERT3P(msp->ms_freeing, !=, NULL);
3838	ASSERT3P(msp->ms_freed, !=, NULL);
3839	ASSERT3P(msp->ms_checkpointing, !=, NULL);
3840	ASSERT3P(msp->ms_trim, !=, NULL);
3841
3842	/*
3843	 * Normally, we don't want to process a metaslab if there are no
3844	 * allocations or frees to perform. However, if the metaslab is being
3845	 * forced to condense, it's loaded and we're not beyond the final
3846	 * dirty txg, we need to let it through. Not condensing beyond the
3847	 * final dirty txg prevents an issue where metaslabs that need to be
3848	 * condensed but were loaded for other reasons could cause a panic
3849	 * here. By only checking the txg in that branch of the conditional,
3850	 * we preserve the utility of the VERIFY statements in all other
3851	 * cases.
3852	 */
3853	if (range_tree_is_empty(alloctree) &&
3854	    range_tree_is_empty(msp->ms_freeing) &&
3855	    range_tree_is_empty(msp->ms_checkpointing) &&
3856	    !(msp->ms_loaded && msp->ms_condense_wanted &&
3857	    txg <= spa_final_dirty_txg(spa)))
3858		return;
3859
3860
3861	VERIFY3U(txg, <=, spa_final_dirty_txg(spa));
3862
3863	/*
3864	 * The only state that can actually be changing concurrently
3865	 * with metaslab_sync() is the metaslab's ms_allocatable. No
3866	 * other thread can be modifying this txg's alloc, freeing,
3867	 * freed, or space_map_phys_t.  We drop ms_lock whenever we
3868	 * could call into the DMU, because the DMU can call down to
3869	 * us (e.g. via zio_free()) at any time.
3870	 *
3871	 * The spa_vdev_remove_thread() can be reading metaslab state
3872	 * concurrently, and it is locked out by the ms_sync_lock.
3873	 * Note that the ms_lock is insufficient for this, because it
3874	 * is dropped by space_map_write().
3875	 */
3876	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
3877
3878	/*
3879	 * Generate a log space map if one doesn't exist already.
3880	 */
3881	spa_generate_syncing_log_sm(spa, tx);
3882
3883	if (msp->ms_sm == NULL) {
3884		uint64_t new_object = space_map_alloc(mos,
3885		    spa_feature_is_enabled(spa, SPA_FEATURE_LOG_SPACEMAP) ?
3886		    zfs_metaslab_sm_blksz_with_log :
3887		    zfs_metaslab_sm_blksz_no_log, tx);
3888		VERIFY3U(new_object, !=, 0);
3889
3890		dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
3891		    msp->ms_id, sizeof (uint64_t), &new_object, tx);
3892
3893		VERIFY0(space_map_open(&msp->ms_sm, mos, new_object,
3894		    msp->ms_start, msp->ms_size, vd->vdev_ashift));
3895		ASSERT(msp->ms_sm != NULL);
3896
3897		ASSERT(range_tree_is_empty(msp->ms_unflushed_allocs));
3898		ASSERT(range_tree_is_empty(msp->ms_unflushed_frees));
3899		ASSERT0(metaslab_allocated_space(msp));
3900	}
3901
3902	if (metaslab_unflushed_txg(msp) == 0 &&
3903	    spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) {
3904		ASSERT(spa_syncing_log_sm(spa) != NULL);
3905
3906		metaslab_set_unflushed_txg(msp, spa_syncing_txg(spa), tx);
3907		spa_log_sm_increment_current_mscount(spa);
3908		spa_log_summary_add_flushed_metaslab(spa);
3909
3910		ASSERT(msp->ms_sm != NULL);
3911		mutex_enter(&spa->spa_flushed_ms_lock);
3912		avl_add(&spa->spa_metaslabs_by_flushed, msp);
3913		mutex_exit(&spa->spa_flushed_ms_lock);
3914
3915		ASSERT(range_tree_is_empty(msp->ms_unflushed_allocs));
3916		ASSERT(range_tree_is_empty(msp->ms_unflushed_frees));
3917	}
3918
3919	if (!range_tree_is_empty(msp->ms_checkpointing) &&
3920	    vd->vdev_checkpoint_sm == NULL) {
3921		ASSERT(spa_has_checkpoint(spa));
3922
3923		uint64_t new_object = space_map_alloc(mos,
3924		    zfs_vdev_standard_sm_blksz, tx);
3925		VERIFY3U(new_object, !=, 0);
3926
3927		VERIFY0(space_map_open(&vd->vdev_checkpoint_sm,
3928		    mos, new_object, 0, vd->vdev_asize, vd->vdev_ashift));
3929		ASSERT3P(vd->vdev_checkpoint_sm, !=, NULL);
3930
3931		/*
3932		 * We save the space map object as an entry in vdev_top_zap
3933		 * so it can be retrieved when the pool is reopened after an
3934		 * export or through zdb.
3935		 */
3936		VERIFY0(zap_add(vd->vdev_spa->spa_meta_objset,
3937		    vd->vdev_top_zap, VDEV_TOP_ZAP_POOL_CHECKPOINT_SM,
3938		    sizeof (new_object), 1, &new_object, tx));
3939	}
3940
3941	mutex_enter(&msp->ms_sync_lock);
3942	mutex_enter(&msp->ms_lock);
3943
3944	/*
3945	 * Note: metaslab_condense() clears the space map's histogram.
3946	 * Therefore we must verify and remove this histogram before
3947	 * condensing.
3948	 */
3949	metaslab_group_histogram_verify(mg);
3950	metaslab_class_histogram_verify(mg->mg_class);
3951	metaslab_group_histogram_remove(mg, msp);
3952
3953	if (spa->spa_sync_pass == 1 && msp->ms_loaded &&
3954	    metaslab_should_condense(msp))
3955		metaslab_condense(msp, tx);
3956
3957	/*
3958	 * We'll be going to disk to sync our space accounting, thus we
3959	 * drop the ms_lock during that time so allocations coming from
3960	 * open-context (ZIL) for future TXGs do not block.
3961	 */
3962	mutex_exit(&msp->ms_lock);
3963	space_map_t *log_sm = spa_syncing_log_sm(spa);
3964	if (log_sm != NULL) {
3965		ASSERT(spa_feature_is_enabled(spa, SPA_FEATURE_LOG_SPACEMAP));
3966
3967		space_map_write(log_sm, alloctree, SM_ALLOC,
3968		    vd->vdev_id, tx);
3969		space_map_write(log_sm, msp->ms_freeing, SM_FREE,
3970		    vd->vdev_id, tx);
3971		mutex_enter(&msp->ms_lock);
3972
3973		ASSERT3U(spa->spa_unflushed_stats.sus_memused, >=,
3974		    metaslab_unflushed_changes_memused(msp));
3975		spa->spa_unflushed_stats.sus_memused -=
3976		    metaslab_unflushed_changes_memused(msp);
3977		range_tree_remove_xor_add(alloctree,
3978		    msp->ms_unflushed_frees, msp->ms_unflushed_allocs);
3979		range_tree_remove_xor_add(msp->ms_freeing,
3980		    msp->ms_unflushed_allocs, msp->ms_unflushed_frees);
3981		spa->spa_unflushed_stats.sus_memused +=
3982		    metaslab_unflushed_changes_memused(msp);
3983	} else {
3984		ASSERT(!spa_feature_is_enabled(spa, SPA_FEATURE_LOG_SPACEMAP));
3985
3986		space_map_write(msp->ms_sm, alloctree, SM_ALLOC,
3987		    SM_NO_VDEVID, tx);
3988		space_map_write(msp->ms_sm, msp->ms_freeing, SM_FREE,
3989		    SM_NO_VDEVID, tx);
3990		mutex_enter(&msp->ms_lock);
3991	}
3992
3993	msp->ms_allocated_space += range_tree_space(alloctree);
3994	ASSERT3U(msp->ms_allocated_space, >=,
3995	    range_tree_space(msp->ms_freeing));
3996	msp->ms_allocated_space -= range_tree_space(msp->ms_freeing);
3997
3998	if (!range_tree_is_empty(msp->ms_checkpointing)) {
3999		ASSERT(spa_has_checkpoint(spa));
4000		ASSERT3P(vd->vdev_checkpoint_sm, !=, NULL);
4001
4002		/*
4003		 * Since we are doing writes to disk and the ms_checkpointing
4004		 * tree won't be changing during that time, we drop the
4005		 * ms_lock while writing to the checkpoint space map, for the
4006		 * same reason mentioned above.
4007		 */
4008		mutex_exit(&msp->ms_lock);
4009		space_map_write(vd->vdev_checkpoint_sm,
4010		    msp->ms_checkpointing, SM_FREE, SM_NO_VDEVID, tx);
4011		mutex_enter(&msp->ms_lock);
4012
4013		spa->spa_checkpoint_info.sci_dspace +=
4014		    range_tree_space(msp->ms_checkpointing);
4015		vd->vdev_stat.vs_checkpoint_space +=
4016		    range_tree_space(msp->ms_checkpointing);
4017		ASSERT3U(vd->vdev_stat.vs_checkpoint_space, ==,
4018		    -space_map_allocated(vd->vdev_checkpoint_sm));
4019
4020		range_tree_vacate(msp->ms_checkpointing, NULL, NULL);
4021	}
4022
4023	if (msp->ms_loaded) {
4024		/*
4025		 * When the space map is loaded, we have an accurate
4026		 * histogram in the range tree. This gives us an opportunity
4027		 * to bring the space map's histogram up-to-date so we clear
4028		 * it first before updating it.
4029		 */
4030		space_map_histogram_clear(msp->ms_sm);
4031		space_map_histogram_add(msp->ms_sm, msp->ms_allocatable, tx);
4032
4033		/*
4034		 * Since we've cleared the histogram we need to add back
4035		 * any free space that has already been processed, plus
4036		 * any deferred space. This allows the on-disk histogram
4037		 * to accurately reflect all free space even if some space
4038		 * is not yet available for allocation (i.e. deferred).
4039		 */
4040		space_map_histogram_add(msp->ms_sm, msp->ms_freed, tx);
4041
4042		/*
4043		 * Add back any deferred free space that has not been
4044		 * added back into the in-core free tree yet. This will
4045		 * ensure that we don't end up with a space map histogram
4046		 * that is completely empty unless the metaslab is fully
4047		 * allocated.
4048		 */
4049		for (int t = 0; t < TXG_DEFER_SIZE; t++) {
4050			space_map_histogram_add(msp->ms_sm,
4051			    msp->ms_defer[t], tx);
4052		}
4053	}
4054
4055	/*
4056	 * Always add the free space from this sync pass to the space
4057	 * map histogram. We want to make sure that the on-disk histogram
4058	 * accounts for all free space. If the space map is not loaded,
4059	 * then we will lose some accuracy but will correct it the next
4060	 * time we load the space map.
4061	 */
4062	space_map_histogram_add(msp->ms_sm, msp->ms_freeing, tx);
4063	metaslab_aux_histograms_update(msp);
4064
4065	metaslab_group_histogram_add(mg, msp);
4066	metaslab_group_histogram_verify(mg);
4067	metaslab_class_histogram_verify(mg->mg_class);
4068
4069	/*
4070	 * For sync pass 1, we avoid traversing this txg's free range tree
4071	 * and instead will just swap the pointers for freeing and freed.
4072	 * We can safely do this since the freed_tree is guaranteed to be
4073	 * empty on the initial pass.
4074	 *
4075	 * Keep in mind that even if we are currently using a log spacemap
4076	 * we want current frees to end up in the ms_allocatable (but not
4077	 * get appended to the ms_sm) so their ranges can be reused as usual.
4078	 */
4079	if (spa_sync_pass(spa) == 1) {
4080		range_tree_swap(&msp->ms_freeing, &msp->ms_freed);
4081		ASSERT0(msp->ms_allocated_this_txg);
4082	} else {
4083		range_tree_vacate(msp->ms_freeing,
4084		    range_tree_add, msp->ms_freed);
4085	}
4086	msp->ms_allocated_this_txg += range_tree_space(alloctree);
4087	range_tree_vacate(alloctree, NULL, NULL);
4088
4089	ASSERT0(range_tree_space(msp->ms_allocating[txg & TXG_MASK]));
4090	ASSERT0(range_tree_space(msp->ms_allocating[TXG_CLEAN(txg)
4091	    & TXG_MASK]));
4092	ASSERT0(range_tree_space(msp->ms_freeing));
4093	ASSERT0(range_tree_space(msp->ms_checkpointing));
4094
4095	mutex_exit(&msp->ms_lock);
4096
4097	/*
4098	 * Verify that the space map object ID has been recorded in the
4099	 * vdev_ms_array.
4100	 */
4101	uint64_t object;
4102	VERIFY0(dmu_read(mos, vd->vdev_ms_array,
4103	    msp->ms_id * sizeof (uint64_t), sizeof (uint64_t), &object, 0));
4104	VERIFY3U(object, ==, space_map_object(msp->ms_sm));
4105
4106	mutex_exit(&msp->ms_sync_lock);
4107	dmu_tx_commit(tx);
4108}
4109
4110static void
4111metaslab_evict(metaslab_t *msp, uint64_t txg)
4112{
4113	if (!msp->ms_loaded || msp->ms_disabled != 0)
4114		return;
4115
4116	for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
4117		VERIFY0(range_tree_space(
4118		    msp->ms_allocating[(txg + t) & TXG_MASK]));
4119	}
4120	if (msp->ms_allocator != -1)
4121		metaslab_passivate(msp, msp->ms_weight & ~METASLAB_ACTIVE_MASK);
4122
4123	if (!metaslab_debug_unload)
4124		metaslab_unload(msp);
4125}
4126
4127/*
4128 * Called after a transaction group has completely synced to mark
4129 * all of the metaslab's free space as usable.
4130 */
4131void
4132metaslab_sync_done(metaslab_t *msp, uint64_t txg)
4133{
4134	metaslab_group_t *mg = msp->ms_group;
4135	vdev_t *vd = mg->mg_vd;
4136	spa_t *spa = vd->vdev_spa;
4137	range_tree_t **defer_tree;
4138	int64_t alloc_delta, defer_delta;
4139	boolean_t defer_allowed = B_TRUE;
4140
4141	ASSERT(!vd->vdev_ishole);
4142
4143	mutex_enter(&msp->ms_lock);
4144
4145	/*
4146	 * If this metaslab is just becoming available, initialize its
4147	 * range trees and add its capacity to the vdev.
4148	 */
4149	if (msp->ms_freed == NULL) {
4150		range_seg_type_t type;
4151		uint64_t shift, start;
4152		type = metaslab_calculate_range_tree_type(vd, msp, &start,
4153		    &shift);
4154
4155		for (int t = 0; t < TXG_SIZE; t++) {
4156			ASSERT(msp->ms_allocating[t] == NULL);
4157
4158			msp->ms_allocating[t] = range_tree_create(NULL, type,
4159			    NULL, start, shift);
4160		}
4161
4162		ASSERT3P(msp->ms_freeing, ==, NULL);
4163		msp->ms_freeing = range_tree_create(NULL, type, NULL, start,
4164		    shift);
4165
4166		ASSERT3P(msp->ms_freed, ==, NULL);
4167		msp->ms_freed = range_tree_create(NULL, type, NULL, start,
4168		    shift);
4169
4170		for (int t = 0; t < TXG_DEFER_SIZE; t++) {
4171			ASSERT3P(msp->ms_defer[t], ==, NULL);
4172			msp->ms_defer[t] = range_tree_create(NULL, type, NULL,
4173			    start, shift);
4174		}
4175
4176		ASSERT3P(msp->ms_checkpointing, ==, NULL);
4177		msp->ms_checkpointing = range_tree_create(NULL, type, NULL,
4178		    start, shift);
4179
4180		ASSERT3P(msp->ms_unflushed_allocs, ==, NULL);
4181		msp->ms_unflushed_allocs = range_tree_create(NULL, type, NULL,
4182		    start, shift);
4183
4184		metaslab_rt_arg_t *mrap = kmem_zalloc(sizeof (*mrap), KM_SLEEP);
4185		mrap->mra_bt = &msp->ms_unflushed_frees_by_size;
4186		mrap->mra_floor_shift = metaslab_by_size_min_shift;
4187		ASSERT3P(msp->ms_unflushed_frees, ==, NULL);
4188		msp->ms_unflushed_frees = range_tree_create(&metaslab_rt_ops,
4189		    type, mrap, start, shift);
4190
4191		metaslab_space_update(vd, mg->mg_class, 0, 0, msp->ms_size);
4192	}
4193	ASSERT0(range_tree_space(msp->ms_freeing));
4194	ASSERT0(range_tree_space(msp->ms_checkpointing));
4195
4196	defer_tree = &msp->ms_defer[txg % TXG_DEFER_SIZE];
4197
4198	uint64_t free_space = metaslab_class_get_space(spa_normal_class(spa)) -
4199	    metaslab_class_get_alloc(spa_normal_class(spa));
4200	if (free_space <= spa_get_slop_space(spa) || vd->vdev_removing) {
4201		defer_allowed = B_FALSE;
4202	}
4203
4204	defer_delta = 0;
4205	alloc_delta = msp->ms_allocated_this_txg -
4206	    range_tree_space(msp->ms_freed);
4207
4208	if (defer_allowed) {
4209		defer_delta = range_tree_space(msp->ms_freed) -
4210		    range_tree_space(*defer_tree);
4211	} else {
4212		defer_delta -= range_tree_space(*defer_tree);
4213	}
4214	metaslab_space_update(vd, mg->mg_class, alloc_delta + defer_delta,
4215	    defer_delta, 0);
4216
4217	if (spa_syncing_log_sm(spa) == NULL) {
4218		/*
4219		 * If there's a metaslab_load() in progress and we don't have
4220		 * a log space map, it means that we probably wrote to the
4221		 * metaslab's space map. If this is the case, we need to
4222		 * make sure that we wait for the load to complete so that we
4223		 * have a consistent view at the in-core side of the metaslab.
4224		 */
4225		metaslab_load_wait(msp);
4226	} else {
4227		ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP));
4228	}
4229
4230	/*
4231	 * When auto-trimming is enabled, free ranges which are added to
4232	 * ms_allocatable are also be added to ms_trim.  The ms_trim tree is
4233	 * periodically consumed by the vdev_autotrim_thread() which issues
4234	 * trims for all ranges and then vacates the tree.  The ms_trim tree
4235	 * can be discarded at any time with the sole consequence of recent
4236	 * frees not being trimmed.
4237	 */
4238	if (spa_get_autotrim(spa) == SPA_AUTOTRIM_ON) {
4239		range_tree_walk(*defer_tree, range_tree_add, msp->ms_trim);
4240		if (!defer_allowed) {
4241			range_tree_walk(msp->ms_freed, range_tree_add,
4242			    msp->ms_trim);
4243		}
4244	} else {
4245		range_tree_vacate(msp->ms_trim, NULL, NULL);
4246	}
4247
4248	/*
4249	 * Move the frees from the defer_tree back to the free
4250	 * range tree (if it's loaded). Swap the freed_tree and
4251	 * the defer_tree -- this is safe to do because we've
4252	 * just emptied out the defer_tree.
4253	 */
4254	range_tree_vacate(*defer_tree,
4255	    msp->ms_loaded ? range_tree_add : NULL, msp->ms_allocatable);
4256	if (defer_allowed) {
4257		range_tree_swap(&msp->ms_freed, defer_tree);
4258	} else {
4259		range_tree_vacate(msp->ms_freed,
4260		    msp->ms_loaded ? range_tree_add : NULL,
4261		    msp->ms_allocatable);
4262	}
4263
4264	msp->ms_synced_length = space_map_length(msp->ms_sm);
4265
4266	msp->ms_deferspace += defer_delta;
4267	ASSERT3S(msp->ms_deferspace, >=, 0);
4268	ASSERT3S(msp->ms_deferspace, <=, msp->ms_size);
4269	if (msp->ms_deferspace != 0) {
4270		/*
4271		 * Keep syncing this metaslab until all deferred frees
4272		 * are back in circulation.
4273		 */
4274		vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
4275	}
4276	metaslab_aux_histograms_update_done(msp, defer_allowed);
4277
4278	if (msp->ms_new) {
4279		msp->ms_new = B_FALSE;
4280		mutex_enter(&mg->mg_lock);
4281		mg->mg_ms_ready++;
4282		mutex_exit(&mg->mg_lock);
4283	}
4284
4285	/*
4286	 * Re-sort metaslab within its group now that we've adjusted
4287	 * its allocatable space.
4288	 */
4289	metaslab_recalculate_weight_and_sort(msp);
4290
4291	/*
4292	 * If the metaslab is loaded and we've not tried to load or allocate
4293	 * from it in 'metaslab_unload_delay' txgs, then unload it.
4294	 */
4295	if (msp->ms_loaded &&
4296	    msp->ms_disabled == 0 &&
4297	    msp->ms_selected_txg + metaslab_unload_delay < txg) {
4298
4299		for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
4300			VERIFY0(range_tree_space(
4301			    msp->ms_allocating[(txg + t) & TXG_MASK]));
4302		}
4303		if (msp->ms_allocator != -1) {
4304			metaslab_passivate(msp, msp->ms_weight &
4305			    ~METASLAB_ACTIVE_MASK);
4306		}
4307
4308		if (!metaslab_debug_unload)
4309			metaslab_unload(msp);
4310	}
4311
4312	ASSERT0(range_tree_space(msp->ms_allocating[txg & TXG_MASK]));
4313	ASSERT0(range_tree_space(msp->ms_freeing));
4314	ASSERT0(range_tree_space(msp->ms_freed));
4315	ASSERT0(range_tree_space(msp->ms_checkpointing));
4316	msp->ms_allocating_total -= msp->ms_allocated_this_txg;
4317	msp->ms_allocated_this_txg = 0;
4318	mutex_exit(&msp->ms_lock);
4319}
4320
4321void
4322metaslab_sync_reassess(metaslab_group_t *mg)
4323{
4324	spa_t *spa = mg->mg_class->mc_spa;
4325
4326	spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
4327	metaslab_group_alloc_update(mg);
4328	mg->mg_fragmentation = metaslab_group_fragmentation(mg);
4329
4330	/*
4331	 * Preload the next potential metaslabs but only on active
4332	 * metaslab groups. We can get into a state where the metaslab
4333	 * is no longer active since we dirty metaslabs as we remove a
4334	 * a device, thus potentially making the metaslab group eligible
4335	 * for preloading.
4336	 */
4337	if (mg->mg_activation_count > 0) {
4338		metaslab_group_preload(mg);
4339	}
4340	spa_config_exit(spa, SCL_ALLOC, FTAG);
4341}
4342
4343/*
4344 * When writing a ditto block (i.e. more than one DVA for a given BP) on
4345 * the same vdev as an existing DVA of this BP, then try to allocate it
4346 * on a different metaslab than existing DVAs (i.e. a unique metaslab).
4347 */
4348static boolean_t
4349metaslab_is_unique(metaslab_t *msp, dva_t *dva)
4350{
4351	uint64_t dva_ms_id;
4352
4353	if (DVA_GET_ASIZE(dva) == 0)
4354		return (B_TRUE);
4355
4356	if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
4357		return (B_TRUE);
4358
4359	dva_ms_id = DVA_GET_OFFSET(dva) >> msp->ms_group->mg_vd->vdev_ms_shift;
4360
4361	return (msp->ms_id != dva_ms_id);
4362}
4363
4364/*
4365 * ==========================================================================
4366 * Metaslab allocation tracing facility
4367 * ==========================================================================
4368 */
4369
4370/*
4371 * Add an allocation trace element to the allocation tracing list.
4372 */
4373static void
4374metaslab_trace_add(zio_alloc_list_t *zal, metaslab_group_t *mg,
4375    metaslab_t *msp, uint64_t psize, uint32_t dva_id, uint64_t offset,
4376    int allocator)
4377{
4378	if (!metaslab_trace_enabled)
4379		return;
4380
4381	/*
4382	 * When the tracing list reaches its maximum we remove
4383	 * the second element in the list before adding a new one.
4384	 * By removing the second element we preserve the original
4385	 * entry as a clue to what allocations steps have already been
4386	 * performed.
4387	 */
4388	if (zal->zal_size == metaslab_trace_max_entries) {
4389		metaslab_alloc_trace_t *mat_next;
4390#ifdef DEBUG
4391		panic("too many entries in allocation list");
4392#endif
4393		METASLABSTAT_BUMP(metaslabstat_trace_over_limit);
4394		zal->zal_size--;
4395		mat_next = list_next(&zal->zal_list, list_head(&zal->zal_list));
4396		list_remove(&zal->zal_list, mat_next);
4397		kmem_cache_free(metaslab_alloc_trace_cache, mat_next);
4398	}
4399
4400	metaslab_alloc_trace_t *mat =
4401	    kmem_cache_alloc(metaslab_alloc_trace_cache, KM_SLEEP);
4402	list_link_init(&mat->mat_list_node);
4403	mat->mat_mg = mg;
4404	mat->mat_msp = msp;
4405	mat->mat_size = psize;
4406	mat->mat_dva_id = dva_id;
4407	mat->mat_offset = offset;
4408	mat->mat_weight = 0;
4409	mat->mat_allocator = allocator;
4410
4411	if (msp != NULL)
4412		mat->mat_weight = msp->ms_weight;
4413
4414	/*
4415	 * The list is part of the zio so locking is not required. Only
4416	 * a single thread will perform allocations for a given zio.
4417	 */
4418	list_insert_tail(&zal->zal_list, mat);
4419	zal->zal_size++;
4420
4421	ASSERT3U(zal->zal_size, <=, metaslab_trace_max_entries);
4422}
4423
4424void
4425metaslab_trace_init(zio_alloc_list_t *zal)
4426{
4427	list_create(&zal->zal_list, sizeof (metaslab_alloc_trace_t),
4428	    offsetof(metaslab_alloc_trace_t, mat_list_node));
4429	zal->zal_size = 0;
4430}
4431
4432void
4433metaslab_trace_fini(zio_alloc_list_t *zal)
4434{
4435	metaslab_alloc_trace_t *mat;
4436
4437	while ((mat = list_remove_head(&zal->zal_list)) != NULL)
4438		kmem_cache_free(metaslab_alloc_trace_cache, mat);
4439	list_destroy(&zal->zal_list);
4440	zal->zal_size = 0;
4441}
4442
4443/*
4444 * ==========================================================================
4445 * Metaslab block operations
4446 * ==========================================================================
4447 */
4448
4449static void
4450metaslab_group_alloc_increment(spa_t *spa, uint64_t vdev, void *tag, int flags,
4451    int allocator)
4452{
4453	if (!(flags & METASLAB_ASYNC_ALLOC) ||
4454	    (flags & METASLAB_DONT_THROTTLE))
4455		return;
4456
4457	metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
4458	if (!mg->mg_class->mc_alloc_throttle_enabled)
4459		return;
4460
4461	(void) zfs_refcount_add(&mg->mg_alloc_queue_depth[allocator], tag);
4462}
4463
4464static void
4465metaslab_group_increment_qdepth(metaslab_group_t *mg, int allocator)
4466{
4467	uint64_t max = mg->mg_max_alloc_queue_depth;
4468	uint64_t cur = mg->mg_cur_max_alloc_queue_depth[allocator];
4469	while (cur < max) {
4470		if (atomic_cas_64(&mg->mg_cur_max_alloc_queue_depth[allocator],
4471		    cur, cur + 1) == cur) {
4472			atomic_inc_64(
4473			    &mg->mg_class->mc_alloc_max_slots[allocator]);
4474			return;
4475		}
4476		cur = mg->mg_cur_max_alloc_queue_depth[allocator];
4477	}
4478}
4479
4480void
4481metaslab_group_alloc_decrement(spa_t *spa, uint64_t vdev, void *tag, int flags,
4482    int allocator, boolean_t io_complete)
4483{
4484	if (!(flags & METASLAB_ASYNC_ALLOC) ||
4485	    (flags & METASLAB_DONT_THROTTLE))
4486		return;
4487
4488	metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
4489	if (!mg->mg_class->mc_alloc_throttle_enabled)
4490		return;
4491
4492	(void) zfs_refcount_remove(&mg->mg_alloc_queue_depth[allocator], tag);
4493	if (io_complete)
4494		metaslab_group_increment_qdepth(mg, allocator);
4495}
4496
4497void
4498metaslab_group_alloc_verify(spa_t *spa, const blkptr_t *bp, void *tag,
4499    int allocator)
4500{
4501#ifdef ZFS_DEBUG
4502	const dva_t *dva = bp->blk_dva;
4503	int ndvas = BP_GET_NDVAS(bp);
4504
4505	for (int d = 0; d < ndvas; d++) {
4506		uint64_t vdev = DVA_GET_VDEV(&dva[d]);
4507		metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
4508		VERIFY(zfs_refcount_not_held(
4509		    &mg->mg_alloc_queue_depth[allocator], tag));
4510	}
4511#endif
4512}
4513
4514static uint64_t
4515metaslab_block_alloc(metaslab_t *msp, uint64_t size, uint64_t txg)
4516{
4517	uint64_t start;
4518	range_tree_t *rt = msp->ms_allocatable;
4519	metaslab_class_t *mc = msp->ms_group->mg_class;
4520
4521	ASSERT(MUTEX_HELD(&msp->ms_lock));
4522	VERIFY(!msp->ms_condensing);
4523	VERIFY0(msp->ms_disabled);
4524
4525	start = mc->mc_ops->msop_alloc(msp, size);
4526	if (start != -1ULL) {
4527		metaslab_group_t *mg = msp->ms_group;
4528		vdev_t *vd = mg->mg_vd;
4529
4530		VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
4531		VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
4532		VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
4533		range_tree_remove(rt, start, size);
4534		range_tree_clear(msp->ms_trim, start, size);
4535
4536		if (range_tree_is_empty(msp->ms_allocating[txg & TXG_MASK]))
4537			vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
4538
4539		range_tree_add(msp->ms_allocating[txg & TXG_MASK], start, size);
4540		msp->ms_allocating_total += size;
4541
4542		/* Track the last successful allocation */
4543		msp->ms_alloc_txg = txg;
4544		metaslab_verify_space(msp, txg);
4545	}
4546
4547	/*
4548	 * Now that we've attempted the allocation we need to update the
4549	 * metaslab's maximum block size since it may have changed.
4550	 */
4551	msp->ms_max_size = metaslab_largest_allocatable(msp);
4552	return (start);
4553}
4554
4555/*
4556 * Find the metaslab with the highest weight that is less than what we've
4557 * already tried.  In the common case, this means that we will examine each
4558 * metaslab at most once. Note that concurrent callers could reorder metaslabs
4559 * by activation/passivation once we have dropped the mg_lock. If a metaslab is
4560 * activated by another thread, and we fail to allocate from the metaslab we
4561 * have selected, we may not try the newly-activated metaslab, and instead
4562 * activate another metaslab.  This is not optimal, but generally does not cause
4563 * any problems (a possible exception being if every metaslab is completely full
4564 * except for the the newly-activated metaslab which we fail to examine).
4565 */
4566static metaslab_t *
4567find_valid_metaslab(metaslab_group_t *mg, uint64_t activation_weight,
4568    dva_t *dva, int d, boolean_t want_unique, uint64_t asize, int allocator,
4569    boolean_t try_hard, zio_alloc_list_t *zal, metaslab_t *search,
4570    boolean_t *was_active)
4571{
4572	avl_index_t idx;
4573	avl_tree_t *t = &mg->mg_metaslab_tree;
4574	metaslab_t *msp = avl_find(t, search, &idx);
4575	if (msp == NULL)
4576		msp = avl_nearest(t, idx, AVL_AFTER);
4577
4578	for (; msp != NULL; msp = AVL_NEXT(t, msp)) {
4579		int i;
4580		if (!metaslab_should_allocate(msp, asize, try_hard)) {
4581			metaslab_trace_add(zal, mg, msp, asize, d,
4582			    TRACE_TOO_SMALL, allocator);
4583			continue;
4584		}
4585
4586		/*
4587		 * If the selected metaslab is condensing or disabled,
4588		 * skip it.
4589		 */
4590		if (msp->ms_condensing || msp->ms_disabled > 0)
4591			continue;
4592
4593		*was_active = msp->ms_allocator != -1;
4594		/*
4595		 * If we're activating as primary, this is our first allocation
4596		 * from this disk, so we don't need to check how close we are.
4597		 * If the metaslab under consideration was already active,
4598		 * we're getting desperate enough to steal another allocator's
4599		 * metaslab, so we still don't care about distances.
4600		 */
4601		if (activation_weight == METASLAB_WEIGHT_PRIMARY || *was_active)
4602			break;
4603
4604		for (i = 0; i < d; i++) {
4605			if (want_unique &&
4606			    !metaslab_is_unique(msp, &dva[i]))
4607				break;  /* try another metaslab */
4608		}
4609		if (i == d)
4610			break;
4611	}
4612
4613	if (msp != NULL) {
4614		search->ms_weight = msp->ms_weight;
4615		search->ms_start = msp->ms_start + 1;
4616		search->ms_allocator = msp->ms_allocator;
4617		search->ms_primary = msp->ms_primary;
4618	}
4619	return (msp);
4620}
4621
4622void
4623metaslab_active_mask_verify(metaslab_t *msp)
4624{
4625	ASSERT(MUTEX_HELD(&msp->ms_lock));
4626
4627	if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0)
4628		return;
4629
4630	if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0)
4631		return;
4632
4633	if (msp->ms_weight & METASLAB_WEIGHT_PRIMARY) {
4634		VERIFY0(msp->ms_weight & METASLAB_WEIGHT_SECONDARY);
4635		VERIFY0(msp->ms_weight & METASLAB_WEIGHT_CLAIM);
4636		VERIFY3S(msp->ms_allocator, !=, -1);
4637		VERIFY(msp->ms_primary);
4638		return;
4639	}
4640
4641	if (msp->ms_weight & METASLAB_WEIGHT_SECONDARY) {
4642		VERIFY0(msp->ms_weight & METASLAB_WEIGHT_PRIMARY);
4643		VERIFY0(msp->ms_weight & METASLAB_WEIGHT_CLAIM);
4644		VERIFY3S(msp->ms_allocator, !=, -1);
4645		VERIFY(!msp->ms_primary);
4646		return;
4647	}
4648
4649	if (msp->ms_weight & METASLAB_WEIGHT_CLAIM) {
4650		VERIFY0(msp->ms_weight & METASLAB_WEIGHT_PRIMARY);
4651		VERIFY0(msp->ms_weight & METASLAB_WEIGHT_SECONDARY);
4652		VERIFY3S(msp->ms_allocator, ==, -1);
4653		return;
4654	}
4655}
4656
4657/* ARGSUSED */
4658static uint64_t
4659metaslab_group_alloc_normal(metaslab_group_t *mg, zio_alloc_list_t *zal,
4660    uint64_t asize, uint64_t txg, boolean_t want_unique, dva_t *dva, int d,
4661    int allocator, boolean_t try_hard)
4662{
4663	metaslab_t *msp = NULL;
4664	uint64_t offset = -1ULL;
4665
4666	uint64_t activation_weight = METASLAB_WEIGHT_PRIMARY;
4667	for (int i = 0; i < d; i++) {
4668		if (activation_weight == METASLAB_WEIGHT_PRIMARY &&
4669		    DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
4670			activation_weight = METASLAB_WEIGHT_SECONDARY;
4671		} else if (activation_weight == METASLAB_WEIGHT_SECONDARY &&
4672		    DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
4673			activation_weight = METASLAB_WEIGHT_CLAIM;
4674			break;
4675		}
4676	}
4677
4678	/*
4679	 * If we don't have enough metaslabs active to fill the entire array, we
4680	 * just use the 0th slot.
4681	 */
4682	if (mg->mg_ms_ready < mg->mg_allocators * 3)
4683		allocator = 0;
4684
4685	ASSERT3U(mg->mg_vd->vdev_ms_count, >=, 2);
4686
4687	metaslab_t *search = kmem_alloc(sizeof (*search), KM_SLEEP);
4688	search->ms_weight = UINT64_MAX;
4689	search->ms_start = 0;
4690	/*
4691	 * At the end of the metaslab tree are the already-active metaslabs,
4692	 * first the primaries, then the secondaries. When we resume searching
4693	 * through the tree, we need to consider ms_allocator and ms_primary so
4694	 * we start in the location right after where we left off, and don't
4695	 * accidentally loop forever considering the same metaslabs.
4696	 */
4697	search->ms_allocator = -1;
4698	search->ms_primary = B_TRUE;
4699	for (;;) {
4700		boolean_t was_active = B_FALSE;
4701
4702		mutex_enter(&mg->mg_lock);
4703
4704		if (activation_weight == METASLAB_WEIGHT_PRIMARY &&
4705		    mg->mg_primaries[allocator] != NULL) {
4706			msp = mg->mg_primaries[allocator];
4707
4708			/*
4709			 * Even though we don't hold the ms_lock for the
4710			 * primary metaslab, those fields should not
4711			 * change while we hold the mg_lock. Thus is is
4712			 * safe to make assertions on them.
4713			 */
4714			ASSERT(msp->ms_primary);
4715			ASSERT3S(msp->ms_allocator, ==, allocator);
4716			ASSERT(msp->ms_loaded);
4717
4718			was_active = B_TRUE;
4719			ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
4720		} else if (activation_weight == METASLAB_WEIGHT_SECONDARY &&
4721		    mg->mg_secondaries[allocator] != NULL) {
4722			msp = mg->mg_secondaries[allocator];
4723
4724			/*
4725			 * See comment above about the similar assertions
4726			 * for the primary metaslab.
4727			 */
4728			ASSERT(!msp->ms_primary);
4729			ASSERT3S(msp->ms_allocator, ==, allocator);
4730			ASSERT(msp->ms_loaded);
4731
4732			was_active = B_TRUE;
4733			ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
4734		} else {
4735			msp = find_valid_metaslab(mg, activation_weight, dva, d,
4736			    want_unique, asize, allocator, try_hard, zal,
4737			    search, &was_active);
4738		}
4739
4740		mutex_exit(&mg->mg_lock);
4741		if (msp == NULL) {
4742			kmem_free(search, sizeof (*search));
4743			return (-1ULL);
4744		}
4745		mutex_enter(&msp->ms_lock);
4746
4747		metaslab_active_mask_verify(msp);
4748
4749		/*
4750		 * This code is disabled out because of issues with
4751		 * tracepoints in non-gpl kernel modules.
4752		 */
4753#if 0
4754		DTRACE_PROBE3(ms__activation__attempt,
4755		    metaslab_t *, msp, uint64_t, activation_weight,
4756		    boolean_t, was_active);
4757#endif
4758
4759		/*
4760		 * Ensure that the metaslab we have selected is still
4761		 * capable of handling our request. It's possible that
4762		 * another thread may have changed the weight while we
4763		 * were blocked on the metaslab lock. We check the
4764		 * active status first to see if we need to set_selected_txg
4765		 * a new metaslab.
4766		 */
4767		if (was_active && !(msp->ms_weight & METASLAB_ACTIVE_MASK)) {
4768			ASSERT3S(msp->ms_allocator, ==, -1);
4769			mutex_exit(&msp->ms_lock);
4770			continue;
4771		}
4772
4773		/*
4774		 * If the metaslab was activated for another allocator
4775		 * while we were waiting in the ms_lock above, or it's
4776		 * a primary and we're seeking a secondary (or vice versa),
4777		 * we go back and select a new metaslab.
4778		 */
4779		if (!was_active && (msp->ms_weight & METASLAB_ACTIVE_MASK) &&
4780		    (msp->ms_allocator != -1) &&
4781		    (msp->ms_allocator != allocator || ((activation_weight ==
4782		    METASLAB_WEIGHT_PRIMARY) != msp->ms_primary))) {
4783			ASSERT(msp->ms_loaded);
4784			ASSERT((msp->ms_weight & METASLAB_WEIGHT_CLAIM) ||
4785			    msp->ms_allocator != -1);
4786			mutex_exit(&msp->ms_lock);
4787			continue;
4788		}
4789
4790		/*
4791		 * This metaslab was used for claiming regions allocated
4792		 * by the ZIL during pool import. Once these regions are
4793		 * claimed we don't need to keep the CLAIM bit set
4794		 * anymore. Passivate this metaslab to zero its activation
4795		 * mask.
4796		 */
4797		if (msp->ms_weight & METASLAB_WEIGHT_CLAIM &&
4798		    activation_weight != METASLAB_WEIGHT_CLAIM) {
4799			ASSERT(msp->ms_loaded);
4800			ASSERT3S(msp->ms_allocator, ==, -1);
4801			metaslab_passivate(msp, msp->ms_weight &
4802			    ~METASLAB_WEIGHT_CLAIM);
4803			mutex_exit(&msp->ms_lock);
4804			continue;
4805		}
4806
4807		metaslab_set_selected_txg(msp, txg);
4808
4809		int activation_error =
4810		    metaslab_activate(msp, allocator, activation_weight);
4811		metaslab_active_mask_verify(msp);
4812
4813		/*
4814		 * If the metaslab was activated by another thread for
4815		 * another allocator or activation_weight (EBUSY), or it
4816		 * failed because another metaslab was assigned as primary
4817		 * for this allocator (EEXIST) we continue using this
4818		 * metaslab for our allocation, rather than going on to a
4819		 * worse metaslab (we waited for that metaslab to be loaded
4820		 * after all).
4821		 *
4822		 * If the activation failed due to an I/O error or ENOSPC we
4823		 * skip to the next metaslab.
4824		 */
4825		boolean_t activated;
4826		if (activation_error == 0) {
4827			activated = B_TRUE;
4828		} else if (activation_error == EBUSY ||
4829		    activation_error == EEXIST) {
4830			activated = B_FALSE;
4831		} else {
4832			mutex_exit(&msp->ms_lock);
4833			continue;
4834		}
4835		ASSERT(msp->ms_loaded);
4836
4837		/*
4838		 * Now that we have the lock, recheck to see if we should
4839		 * continue to use this metaslab for this allocation. The
4840		 * the metaslab is now loaded so metaslab_should_allocate()
4841		 * can accurately determine if the allocation attempt should
4842		 * proceed.
4843		 */
4844		if (!metaslab_should_allocate(msp, asize, try_hard)) {
4845			/* Passivate this metaslab and select a new one. */
4846			metaslab_trace_add(zal, mg, msp, asize, d,
4847			    TRACE_TOO_SMALL, allocator);
4848			goto next;
4849		}
4850
4851		/*
4852		 * If this metaslab is currently condensing then pick again
4853		 * as we can't manipulate this metaslab until it's committed
4854		 * to disk. If this metaslab is being initialized, we shouldn't
4855		 * allocate from it since the allocated region might be
4856		 * overwritten after allocation.
4857		 */
4858		if (msp->ms_condensing) {
4859			metaslab_trace_add(zal, mg, msp, asize, d,
4860			    TRACE_CONDENSING, allocator);
4861			if (activated) {
4862				metaslab_passivate(msp, msp->ms_weight &
4863				    ~METASLAB_ACTIVE_MASK);
4864			}
4865			mutex_exit(&msp->ms_lock);
4866			continue;
4867		} else if (msp->ms_disabled > 0) {
4868			metaslab_trace_add(zal, mg, msp, asize, d,
4869			    TRACE_DISABLED, allocator);
4870			if (activated) {
4871				metaslab_passivate(msp, msp->ms_weight &
4872				    ~METASLAB_ACTIVE_MASK);
4873			}
4874			mutex_exit(&msp->ms_lock);
4875			continue;
4876		}
4877
4878		offset = metaslab_block_alloc(msp, asize, txg);
4879		metaslab_trace_add(zal, mg, msp, asize, d, offset, allocator);
4880
4881		if (offset != -1ULL) {
4882			/* Proactively passivate the metaslab, if needed */
4883			if (activated)
4884				metaslab_segment_may_passivate(msp);
4885			break;
4886		}
4887next:
4888		ASSERT(msp->ms_loaded);
4889
4890		/*
4891		 * This code is disabled out because of issues with
4892		 * tracepoints in non-gpl kernel modules.
4893		 */
4894#if 0
4895		DTRACE_PROBE2(ms__alloc__failure, metaslab_t *, msp,
4896		    uint64_t, asize);
4897#endif
4898
4899		/*
4900		 * We were unable to allocate from this metaslab so determine
4901		 * a new weight for this metaslab. Now that we have loaded
4902		 * the metaslab we can provide a better hint to the metaslab
4903		 * selector.
4904		 *
4905		 * For space-based metaslabs, we use the maximum block size.
4906		 * This information is only available when the metaslab
4907		 * is loaded and is more accurate than the generic free
4908		 * space weight that was calculated by metaslab_weight().
4909		 * This information allows us to quickly compare the maximum
4910		 * available allocation in the metaslab to the allocation
4911		 * size being requested.
4912		 *
4913		 * For segment-based metaslabs, determine the new weight
4914		 * based on the highest bucket in the range tree. We
4915		 * explicitly use the loaded segment weight (i.e. the range
4916		 * tree histogram) since it contains the space that is
4917		 * currently available for allocation and is accurate
4918		 * even within a sync pass.
4919		 */
4920		uint64_t weight;
4921		if (WEIGHT_IS_SPACEBASED(msp->ms_weight)) {
4922			weight = metaslab_largest_allocatable(msp);
4923			WEIGHT_SET_SPACEBASED(weight);
4924		} else {
4925			weight = metaslab_weight_from_range_tree(msp);
4926		}
4927
4928		if (activated) {
4929			metaslab_passivate(msp, weight);
4930		} else {
4931			/*
4932			 * For the case where we use the metaslab that is
4933			 * active for another allocator we want to make
4934			 * sure that we retain the activation mask.
4935			 *
4936			 * Note that we could attempt to use something like
4937			 * metaslab_recalculate_weight_and_sort() that
4938			 * retains the activation mask here. That function
4939			 * uses metaslab_weight() to set the weight though
4940			 * which is not as accurate as the calculations
4941			 * above.
4942			 */
4943			weight |= msp->ms_weight & METASLAB_ACTIVE_MASK;
4944			metaslab_group_sort(mg, msp, weight);
4945		}
4946		metaslab_active_mask_verify(msp);
4947
4948		/*
4949		 * We have just failed an allocation attempt, check
4950		 * that metaslab_should_allocate() agrees. Otherwise,
4951		 * we may end up in an infinite loop retrying the same
4952		 * metaslab.
4953		 */
4954		ASSERT(!metaslab_should_allocate(msp, asize, try_hard));
4955
4956		mutex_exit(&msp->ms_lock);
4957	}
4958	mutex_exit(&msp->ms_lock);
4959	kmem_free(search, sizeof (*search));
4960	return (offset);
4961}
4962
4963static uint64_t
4964metaslab_group_alloc(metaslab_group_t *mg, zio_alloc_list_t *zal,
4965    uint64_t asize, uint64_t txg, boolean_t want_unique, dva_t *dva, int d,
4966    int allocator, boolean_t try_hard)
4967{
4968	uint64_t offset;
4969	ASSERT(mg->mg_initialized);
4970
4971	offset = metaslab_group_alloc_normal(mg, zal, asize, txg, want_unique,
4972	    dva, d, allocator, try_hard);
4973
4974	mutex_enter(&mg->mg_lock);
4975	if (offset == -1ULL) {
4976		mg->mg_failed_allocations++;
4977		metaslab_trace_add(zal, mg, NULL, asize, d,
4978		    TRACE_GROUP_FAILURE, allocator);
4979		if (asize == SPA_GANGBLOCKSIZE) {
4980			/*
4981			 * This metaslab group was unable to allocate
4982			 * the minimum gang block size so it must be out of
4983			 * space. We must notify the allocation throttle
4984			 * to start skipping allocation attempts to this
4985			 * metaslab group until more space becomes available.
4986			 * Note: this failure cannot be caused by the
4987			 * allocation throttle since the allocation throttle
4988			 * is only responsible for skipping devices and
4989			 * not failing block allocations.
4990			 */
4991			mg->mg_no_free_space = B_TRUE;
4992		}
4993	}
4994	mg->mg_allocations++;
4995	mutex_exit(&mg->mg_lock);
4996	return (offset);
4997}
4998
4999/*
5000 * Allocate a block for the specified i/o.
5001 */
5002int
5003metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
5004    dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags,
5005    zio_alloc_list_t *zal, int allocator)
5006{
5007	metaslab_group_t *mg, *rotor;
5008	vdev_t *vd;
5009	boolean_t try_hard = B_FALSE;
5010
5011	ASSERT(!DVA_IS_VALID(&dva[d]));
5012
5013	/*
5014	 * For testing, make some blocks above a certain size be gang blocks.
5015	 * This will also test spilling from special to normal.
5016	 */
5017	if (psize >= metaslab_force_ganging && (ddi_get_lbolt() & 3) == 0) {
5018		metaslab_trace_add(zal, NULL, NULL, psize, d, TRACE_FORCE_GANG,
5019		    allocator);
5020		return (SET_ERROR(ENOSPC));
5021	}
5022
5023	/*
5024	 * Start at the rotor and loop through all mgs until we find something.
5025	 * Note that there's no locking on mc_rotor or mc_aliquot because
5026	 * nothing actually breaks if we miss a few updates -- we just won't
5027	 * allocate quite as evenly.  It all balances out over time.
5028	 *
5029	 * If we are doing ditto or log blocks, try to spread them across
5030	 * consecutive vdevs.  If we're forced to reuse a vdev before we've
5031	 * allocated all of our ditto blocks, then try and spread them out on
5032	 * that vdev as much as possible.  If it turns out to not be possible,
5033	 * gradually lower our standards until anything becomes acceptable.
5034	 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
5035	 * gives us hope of containing our fault domains to something we're
5036	 * able to reason about.  Otherwise, any two top-level vdev failures
5037	 * will guarantee the loss of data.  With consecutive allocation,
5038	 * only two adjacent top-level vdev failures will result in data loss.
5039	 *
5040	 * If we are doing gang blocks (hintdva is non-NULL), try to keep
5041	 * ourselves on the same vdev as our gang block header.  That
5042	 * way, we can hope for locality in vdev_cache, plus it makes our
5043	 * fault domains something tractable.
5044	 */
5045	if (hintdva) {
5046		vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
5047
5048		/*
5049		 * It's possible the vdev we're using as the hint no
5050		 * longer exists or its mg has been closed (e.g. by
5051		 * device removal).  Consult the rotor when
5052		 * all else fails.
5053		 */
5054		if (vd != NULL && vd->vdev_mg != NULL) {
5055			mg = vd->vdev_mg;
5056
5057			if (flags & METASLAB_HINTBP_AVOID &&
5058			    mg->mg_next != NULL)
5059				mg = mg->mg_next;
5060		} else {
5061			mg = mc->mc_rotor;
5062		}
5063	} else if (d != 0) {
5064		vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
5065		mg = vd->vdev_mg->mg_next;
5066	} else {
5067		ASSERT(mc->mc_rotor != NULL);
5068		mg = mc->mc_rotor;
5069	}
5070
5071	/*
5072	 * If the hint put us into the wrong metaslab class, or into a
5073	 * metaslab group that has been passivated, just follow the rotor.
5074	 */
5075	if (mg->mg_class != mc || mg->mg_activation_count <= 0)
5076		mg = mc->mc_rotor;
5077
5078	rotor = mg;
5079top:
5080	do {
5081		boolean_t allocatable;
5082
5083		ASSERT(mg->mg_activation_count == 1);
5084		vd = mg->mg_vd;
5085
5086		/*
5087		 * Don't allocate from faulted devices.
5088		 */
5089		if (try_hard) {
5090			spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
5091			allocatable = vdev_allocatable(vd);
5092			spa_config_exit(spa, SCL_ZIO, FTAG);
5093		} else {
5094			allocatable = vdev_allocatable(vd);
5095		}
5096
5097		/*
5098		 * Determine if the selected metaslab group is eligible
5099		 * for allocations. If we're ganging then don't allow
5100		 * this metaslab group to skip allocations since that would
5101		 * inadvertently return ENOSPC and suspend the pool
5102		 * even though space is still available.
5103		 */
5104		if (allocatable && !GANG_ALLOCATION(flags) && !try_hard) {
5105			allocatable = metaslab_group_allocatable(mg, rotor,
5106			    psize, allocator, d);
5107		}
5108
5109		if (!allocatable) {
5110			metaslab_trace_add(zal, mg, NULL, psize, d,
5111			    TRACE_NOT_ALLOCATABLE, allocator);
5112			goto next;
5113		}
5114
5115		ASSERT(mg->mg_initialized);
5116
5117		/*
5118		 * Avoid writing single-copy data to a failing,
5119		 * non-redundant vdev, unless we've already tried all
5120		 * other vdevs.
5121		 */
5122		if ((vd->vdev_stat.vs_write_errors > 0 ||
5123		    vd->vdev_state < VDEV_STATE_HEALTHY) &&
5124		    d == 0 && !try_hard && vd->vdev_children == 0) {
5125			metaslab_trace_add(zal, mg, NULL, psize, d,
5126			    TRACE_VDEV_ERROR, allocator);
5127			goto next;
5128		}
5129
5130		ASSERT(mg->mg_class == mc);
5131
5132		uint64_t asize = vdev_psize_to_asize(vd, psize);
5133		ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
5134
5135		/*
5136		 * If we don't need to try hard, then require that the
5137		 * block be on an different metaslab from any other DVAs
5138		 * in this BP (unique=true).  If we are trying hard, then
5139		 * allow any metaslab to be used (unique=false).
5140		 */
5141		uint64_t offset = metaslab_group_alloc(mg, zal, asize, txg,
5142		    !try_hard, dva, d, allocator, try_hard);
5143
5144		if (offset != -1ULL) {
5145			/*
5146			 * If we've just selected this metaslab group,
5147			 * figure out whether the corresponding vdev is
5148			 * over- or under-used relative to the pool,
5149			 * and set an allocation bias to even it out.
5150			 */
5151			if (mc->mc_aliquot == 0 && metaslab_bias_enabled) {
5152				vdev_stat_t *vs = &vd->vdev_stat;
5153				int64_t vu, cu;
5154
5155				vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
5156				cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
5157
5158				/*
5159				 * Calculate how much more or less we should
5160				 * try to allocate from this device during
5161				 * this iteration around the rotor.
5162				 * For example, if a device is 80% full
5163				 * and the pool is 20% full then we should
5164				 * reduce allocations by 60% on this device.
5165				 *
5166				 * mg_bias = (20 - 80) * 512K / 100 = -307K
5167				 *
5168				 * This reduces allocations by 307K for this
5169				 * iteration.
5170				 */
5171				mg->mg_bias = ((cu - vu) *
5172				    (int64_t)mg->mg_aliquot) / 100;
5173			} else if (!metaslab_bias_enabled) {
5174				mg->mg_bias = 0;
5175			}
5176
5177			if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
5178			    mg->mg_aliquot + mg->mg_bias) {
5179				mc->mc_rotor = mg->mg_next;
5180				mc->mc_aliquot = 0;
5181			}
5182
5183			DVA_SET_VDEV(&dva[d], vd->vdev_id);
5184			DVA_SET_OFFSET(&dva[d], offset);
5185			DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
5186			DVA_SET_ASIZE(&dva[d], asize);
5187
5188			return (0);
5189		}
5190next:
5191		mc->mc_rotor = mg->mg_next;
5192		mc->mc_aliquot = 0;
5193	} while ((mg = mg->mg_next) != rotor);
5194
5195	/*
5196	 * If we haven't tried hard, do so now.
5197	 */
5198	if (!try_hard) {
5199		try_hard = B_TRUE;
5200		goto top;
5201	}
5202
5203	bzero(&dva[d], sizeof (dva_t));
5204
5205	metaslab_trace_add(zal, rotor, NULL, psize, d, TRACE_ENOSPC, allocator);
5206	return (SET_ERROR(ENOSPC));
5207}
5208
5209void
5210metaslab_free_concrete(vdev_t *vd, uint64_t offset, uint64_t asize,
5211    boolean_t checkpoint)
5212{
5213	metaslab_t *msp;
5214	spa_t *spa = vd->vdev_spa;
5215
5216	ASSERT(vdev_is_concrete(vd));
5217	ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
5218	ASSERT3U(offset >> vd->vdev_ms_shift, <, vd->vdev_ms_count);
5219
5220	msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
5221
5222	VERIFY(!msp->ms_condensing);
5223	VERIFY3U(offset, >=, msp->ms_start);
5224	VERIFY3U(offset + asize, <=, msp->ms_start + msp->ms_size);
5225	VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
5226	VERIFY0(P2PHASE(asize, 1ULL << vd->vdev_ashift));
5227
5228	metaslab_check_free_impl(vd, offset, asize);
5229
5230	mutex_enter(&msp->ms_lock);
5231	if (range_tree_is_empty(msp->ms_freeing) &&
5232	    range_tree_is_empty(msp->ms_checkpointing)) {
5233		vdev_dirty(vd, VDD_METASLAB, msp, spa_syncing_txg(spa));
5234	}
5235
5236	if (checkpoint) {
5237		ASSERT(spa_has_checkpoint(spa));
5238		range_tree_add(msp->ms_checkpointing, offset, asize);
5239	} else {
5240		range_tree_add(msp->ms_freeing, offset, asize);
5241	}
5242	mutex_exit(&msp->ms_lock);
5243}
5244
5245/* ARGSUSED */
5246void
5247metaslab_free_impl_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset,
5248    uint64_t size, void *arg)
5249{
5250	boolean_t *checkpoint = arg;
5251
5252	ASSERT3P(checkpoint, !=, NULL);
5253
5254	if (vd->vdev_ops->vdev_op_remap != NULL)
5255		vdev_indirect_mark_obsolete(vd, offset, size);
5256	else
5257		metaslab_free_impl(vd, offset, size, *checkpoint);
5258}
5259
5260static void
5261metaslab_free_impl(vdev_t *vd, uint64_t offset, uint64_t size,
5262    boolean_t checkpoint)
5263{
5264	spa_t *spa = vd->vdev_spa;
5265
5266	ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
5267
5268	if (spa_syncing_txg(spa) > spa_freeze_txg(spa))
5269		return;
5270
5271	if (spa->spa_vdev_removal != NULL &&
5272	    spa->spa_vdev_removal->svr_vdev_id == vd->vdev_id &&
5273	    vdev_is_concrete(vd)) {
5274		/*
5275		 * Note: we check if the vdev is concrete because when
5276		 * we complete the removal, we first change the vdev to be
5277		 * an indirect vdev (in open context), and then (in syncing
5278		 * context) clear spa_vdev_removal.
5279		 */
5280		free_from_removing_vdev(vd, offset, size);
5281	} else if (vd->vdev_ops->vdev_op_remap != NULL) {
5282		vdev_indirect_mark_obsolete(vd, offset, size);
5283		vd->vdev_ops->vdev_op_remap(vd, offset, size,
5284		    metaslab_free_impl_cb, &checkpoint);
5285	} else {
5286		metaslab_free_concrete(vd, offset, size, checkpoint);
5287	}
5288}
5289
5290typedef struct remap_blkptr_cb_arg {
5291	blkptr_t *rbca_bp;
5292	spa_remap_cb_t rbca_cb;
5293	vdev_t *rbca_remap_vd;
5294	uint64_t rbca_remap_offset;
5295	void *rbca_cb_arg;
5296} remap_blkptr_cb_arg_t;
5297
5298void
5299remap_blkptr_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset,
5300    uint64_t size, void *arg)
5301{
5302	remap_blkptr_cb_arg_t *rbca = arg;
5303	blkptr_t *bp = rbca->rbca_bp;
5304
5305	/* We can not remap split blocks. */
5306	if (size != DVA_GET_ASIZE(&bp->blk_dva[0]))
5307		return;
5308	ASSERT0(inner_offset);
5309
5310	if (rbca->rbca_cb != NULL) {
5311		/*
5312		 * At this point we know that we are not handling split
5313		 * blocks and we invoke the callback on the previous
5314		 * vdev which must be indirect.
5315		 */
5316		ASSERT3P(rbca->rbca_remap_vd->vdev_ops, ==, &vdev_indirect_ops);
5317
5318		rbca->rbca_cb(rbca->rbca_remap_vd->vdev_id,
5319		    rbca->rbca_remap_offset, size, rbca->rbca_cb_arg);
5320
5321		/* set up remap_blkptr_cb_arg for the next call */
5322		rbca->rbca_remap_vd = vd;
5323		rbca->rbca_remap_offset = offset;
5324	}
5325
5326	/*
5327	 * The phys birth time is that of dva[0].  This ensures that we know
5328	 * when each dva was written, so that resilver can determine which
5329	 * blocks need to be scrubbed (i.e. those written during the time
5330	 * the vdev was offline).  It also ensures that the key used in
5331	 * the ARC hash table is unique (i.e. dva[0] + phys_birth).  If
5332	 * we didn't change the phys_birth, a lookup in the ARC for a
5333	 * remapped BP could find the data that was previously stored at
5334	 * this vdev + offset.
5335	 */
5336	vdev_t *oldvd = vdev_lookup_top(vd->vdev_spa,
5337	    DVA_GET_VDEV(&bp->blk_dva[0]));
5338	vdev_indirect_births_t *vib = oldvd->vdev_indirect_births;
5339	bp->blk_phys_birth = vdev_indirect_births_physbirth(vib,
5340	    DVA_GET_OFFSET(&bp->blk_dva[0]), DVA_GET_ASIZE(&bp->blk_dva[0]));
5341
5342	DVA_SET_VDEV(&bp->blk_dva[0], vd->vdev_id);
5343	DVA_SET_OFFSET(&bp->blk_dva[0], offset);
5344}
5345
5346/*
5347 * If the block pointer contains any indirect DVAs, modify them to refer to
5348 * concrete DVAs.  Note that this will sometimes not be possible, leaving
5349 * the indirect DVA in place.  This happens if the indirect DVA spans multiple
5350 * segments in the mapping (i.e. it is a "split block").
5351 *
5352 * If the BP was remapped, calls the callback on the original dva (note the
5353 * callback can be called multiple times if the original indirect DVA refers
5354 * to another indirect DVA, etc).
5355 *
5356 * Returns TRUE if the BP was remapped.
5357 */
5358boolean_t
5359spa_remap_blkptr(spa_t *spa, blkptr_t *bp, spa_remap_cb_t callback, void *arg)
5360{
5361	remap_blkptr_cb_arg_t rbca;
5362
5363	if (!zfs_remap_blkptr_enable)
5364		return (B_FALSE);
5365
5366	if (!spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS))
5367		return (B_FALSE);
5368
5369	/*
5370	 * Dedup BP's can not be remapped, because ddt_phys_select() depends
5371	 * on DVA[0] being the same in the BP as in the DDT (dedup table).
5372	 */
5373	if (BP_GET_DEDUP(bp))
5374		return (B_FALSE);
5375
5376	/*
5377	 * Gang blocks can not be remapped, because
5378	 * zio_checksum_gang_verifier() depends on the DVA[0] that's in
5379	 * the BP used to read the gang block header (GBH) being the same
5380	 * as the DVA[0] that we allocated for the GBH.
5381	 */
5382	if (BP_IS_GANG(bp))
5383		return (B_FALSE);
5384
5385	/*
5386	 * Embedded BP's have no DVA to remap.
5387	 */
5388	if (BP_GET_NDVAS(bp) < 1)
5389		return (B_FALSE);
5390
5391	/*
5392	 * Note: we only remap dva[0].  If we remapped other dvas, we
5393	 * would no longer know what their phys birth txg is.
5394	 */
5395	dva_t *dva = &bp->blk_dva[0];
5396
5397	uint64_t offset = DVA_GET_OFFSET(dva);
5398	uint64_t size = DVA_GET_ASIZE(dva);
5399	vdev_t *vd = vdev_lookup_top(spa, DVA_GET_VDEV(dva));
5400
5401	if (vd->vdev_ops->vdev_op_remap == NULL)
5402		return (B_FALSE);
5403
5404	rbca.rbca_bp = bp;
5405	rbca.rbca_cb = callback;
5406	rbca.rbca_remap_vd = vd;
5407	rbca.rbca_remap_offset = offset;
5408	rbca.rbca_cb_arg = arg;
5409
5410	/*
5411	 * remap_blkptr_cb() will be called in order for each level of
5412	 * indirection, until a concrete vdev is reached or a split block is
5413	 * encountered. old_vd and old_offset are updated within the callback
5414	 * as we go from the one indirect vdev to the next one (either concrete
5415	 * or indirect again) in that order.
5416	 */
5417	vd->vdev_ops->vdev_op_remap(vd, offset, size, remap_blkptr_cb, &rbca);
5418
5419	/* Check if the DVA wasn't remapped because it is a split block */
5420	if (DVA_GET_VDEV(&rbca.rbca_bp->blk_dva[0]) == vd->vdev_id)
5421		return (B_FALSE);
5422
5423	return (B_TRUE);
5424}
5425
5426/*
5427 * Undo the allocation of a DVA which happened in the given transaction group.
5428 */
5429void
5430metaslab_unalloc_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
5431{
5432	metaslab_t *msp;
5433	vdev_t *vd;
5434	uint64_t vdev = DVA_GET_VDEV(dva);
5435	uint64_t offset = DVA_GET_OFFSET(dva);
5436	uint64_t size = DVA_GET_ASIZE(dva);
5437
5438	ASSERT(DVA_IS_VALID(dva));
5439	ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
5440
5441	if (txg > spa_freeze_txg(spa))
5442		return;
5443
5444	if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
5445	    (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
5446		cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
5447		    (u_longlong_t)vdev, (u_longlong_t)offset);
5448		ASSERT(0);
5449		return;
5450	}
5451
5452	ASSERT(!vd->vdev_removing);
5453	ASSERT(vdev_is_concrete(vd));
5454	ASSERT0(vd->vdev_indirect_config.vic_mapping_object);
5455	ASSERT3P(vd->vdev_indirect_mapping, ==, NULL);
5456
5457	if (DVA_GET_GANG(dva))
5458		size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
5459
5460	msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
5461
5462	mutex_enter(&msp->ms_lock);
5463	range_tree_remove(msp->ms_allocating[txg & TXG_MASK],
5464	    offset, size);
5465	msp->ms_allocating_total -= size;
5466
5467	VERIFY(!msp->ms_condensing);
5468	VERIFY3U(offset, >=, msp->ms_start);
5469	VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size);
5470	VERIFY3U(range_tree_space(msp->ms_allocatable) + size, <=,
5471	    msp->ms_size);
5472	VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
5473	VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
5474	range_tree_add(msp->ms_allocatable, offset, size);
5475	mutex_exit(&msp->ms_lock);
5476}
5477
5478/*
5479 * Free the block represented by the given DVA.
5480 */
5481void
5482metaslab_free_dva(spa_t *spa, const dva_t *dva, boolean_t checkpoint)
5483{
5484	uint64_t vdev = DVA_GET_VDEV(dva);
5485	uint64_t offset = DVA_GET_OFFSET(dva);
5486	uint64_t size = DVA_GET_ASIZE(dva);
5487	vdev_t *vd = vdev_lookup_top(spa, vdev);
5488
5489	ASSERT(DVA_IS_VALID(dva));
5490	ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
5491
5492	if (DVA_GET_GANG(dva)) {
5493		size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
5494	}
5495
5496	metaslab_free_impl(vd, offset, size, checkpoint);
5497}
5498
5499/*
5500 * Reserve some allocation slots. The reservation system must be called
5501 * before we call into the allocator. If there aren't any available slots
5502 * then the I/O will be throttled until an I/O completes and its slots are
5503 * freed up. The function returns true if it was successful in placing
5504 * the reservation.
5505 */
5506boolean_t
5507metaslab_class_throttle_reserve(metaslab_class_t *mc, int slots, int allocator,
5508    zio_t *zio, int flags)
5509{
5510	uint64_t available_slots = 0;
5511	boolean_t slot_reserved = B_FALSE;
5512	uint64_t max = mc->mc_alloc_max_slots[allocator];
5513
5514	ASSERT(mc->mc_alloc_throttle_enabled);
5515	mutex_enter(&mc->mc_lock);
5516
5517	uint64_t reserved_slots =
5518	    zfs_refcount_count(&mc->mc_alloc_slots[allocator]);
5519	if (reserved_slots < max)
5520		available_slots = max - reserved_slots;
5521
5522	if (slots <= available_slots || GANG_ALLOCATION(flags) ||
5523	    flags & METASLAB_MUST_RESERVE) {
5524		/*
5525		 * We reserve the slots individually so that we can unreserve
5526		 * them individually when an I/O completes.
5527		 */
5528		for (int d = 0; d < slots; d++) {
5529			reserved_slots =
5530			    zfs_refcount_add(&mc->mc_alloc_slots[allocator],
5531			    zio);
5532		}
5533		zio->io_flags |= ZIO_FLAG_IO_ALLOCATING;
5534		slot_reserved = B_TRUE;
5535	}
5536
5537	mutex_exit(&mc->mc_lock);
5538	return (slot_reserved);
5539}
5540
5541void
5542metaslab_class_throttle_unreserve(metaslab_class_t *mc, int slots,
5543    int allocator, zio_t *zio)
5544{
5545	ASSERT(mc->mc_alloc_throttle_enabled);
5546	mutex_enter(&mc->mc_lock);
5547	for (int d = 0; d < slots; d++) {
5548		(void) zfs_refcount_remove(&mc->mc_alloc_slots[allocator],
5549		    zio);
5550	}
5551	mutex_exit(&mc->mc_lock);
5552}
5553
5554static int
5555metaslab_claim_concrete(vdev_t *vd, uint64_t offset, uint64_t size,
5556    uint64_t txg)
5557{
5558	metaslab_t *msp;
5559	spa_t *spa = vd->vdev_spa;
5560	int error = 0;
5561
5562	if (offset >> vd->vdev_ms_shift >= vd->vdev_ms_count)
5563		return (ENXIO);
5564
5565	ASSERT3P(vd->vdev_ms, !=, NULL);
5566	msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
5567
5568	mutex_enter(&msp->ms_lock);
5569
5570	if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded)
5571		error = metaslab_activate(msp, 0, METASLAB_WEIGHT_CLAIM);
5572	/*
5573	 * No need to fail in that case; someone else has activated the
5574	 * metaslab, but that doesn't preclude us from using it.
5575	 */
5576	if (error == EBUSY)
5577		error = 0;
5578
5579	if (error == 0 &&
5580	    !range_tree_contains(msp->ms_allocatable, offset, size))
5581		error = SET_ERROR(ENOENT);
5582
5583	if (error || txg == 0) {	/* txg == 0 indicates dry run */
5584		mutex_exit(&msp->ms_lock);
5585		return (error);
5586	}
5587
5588	VERIFY(!msp->ms_condensing);
5589	VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
5590	VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
5591	VERIFY3U(range_tree_space(msp->ms_allocatable) - size, <=,
5592	    msp->ms_size);
5593	range_tree_remove(msp->ms_allocatable, offset, size);
5594	range_tree_clear(msp->ms_trim, offset, size);
5595
5596	if (spa_writeable(spa)) {	/* don't dirty if we're zdb(1M) */
5597		metaslab_class_t *mc = msp->ms_group->mg_class;
5598		multilist_sublist_t *mls =
5599		    multilist_sublist_lock_obj(mc->mc_metaslab_txg_list, msp);
5600		if (!multilist_link_active(&msp->ms_class_txg_node)) {
5601			msp->ms_selected_txg = txg;
5602			multilist_sublist_insert_head(mls, msp);
5603		}
5604		multilist_sublist_unlock(mls);
5605
5606		if (range_tree_is_empty(msp->ms_allocating[txg & TXG_MASK]))
5607			vdev_dirty(vd, VDD_METASLAB, msp, txg);
5608		range_tree_add(msp->ms_allocating[txg & TXG_MASK],
5609		    offset, size);
5610		msp->ms_allocating_total += size;
5611	}
5612
5613	mutex_exit(&msp->ms_lock);
5614
5615	return (0);
5616}
5617
5618typedef struct metaslab_claim_cb_arg_t {
5619	uint64_t	mcca_txg;
5620	int		mcca_error;
5621} metaslab_claim_cb_arg_t;
5622
5623/* ARGSUSED */
5624static void
5625metaslab_claim_impl_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset,
5626    uint64_t size, void *arg)
5627{
5628	metaslab_claim_cb_arg_t *mcca_arg = arg;
5629
5630	if (mcca_arg->mcca_error == 0) {
5631		mcca_arg->mcca_error = metaslab_claim_concrete(vd, offset,
5632		    size, mcca_arg->mcca_txg);
5633	}
5634}
5635
5636int
5637metaslab_claim_impl(vdev_t *vd, uint64_t offset, uint64_t size, uint64_t txg)
5638{
5639	if (vd->vdev_ops->vdev_op_remap != NULL) {
5640		metaslab_claim_cb_arg_t arg;
5641
5642		/*
5643		 * Only zdb(1M) can claim on indirect vdevs.  This is used
5644		 * to detect leaks of mapped space (that are not accounted
5645		 * for in the obsolete counts, spacemap, or bpobj).
5646		 */
5647		ASSERT(!spa_writeable(vd->vdev_spa));
5648		arg.mcca_error = 0;
5649		arg.mcca_txg = txg;
5650
5651		vd->vdev_ops->vdev_op_remap(vd, offset, size,
5652		    metaslab_claim_impl_cb, &arg);
5653
5654		if (arg.mcca_error == 0) {
5655			arg.mcca_error = metaslab_claim_concrete(vd,
5656			    offset, size, txg);
5657		}
5658		return (arg.mcca_error);
5659	} else {
5660		return (metaslab_claim_concrete(vd, offset, size, txg));
5661	}
5662}
5663
5664/*
5665 * Intent log support: upon opening the pool after a crash, notify the SPA
5666 * of blocks that the intent log has allocated for immediate write, but
5667 * which are still considered free by the SPA because the last transaction
5668 * group didn't commit yet.
5669 */
5670static int
5671metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
5672{
5673	uint64_t vdev = DVA_GET_VDEV(dva);
5674	uint64_t offset = DVA_GET_OFFSET(dva);
5675	uint64_t size = DVA_GET_ASIZE(dva);
5676	vdev_t *vd;
5677
5678	if ((vd = vdev_lookup_top(spa, vdev)) == NULL) {
5679		return (SET_ERROR(ENXIO));
5680	}
5681
5682	ASSERT(DVA_IS_VALID(dva));
5683
5684	if (DVA_GET_GANG(dva))
5685		size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
5686
5687	return (metaslab_claim_impl(vd, offset, size, txg));
5688}
5689
5690int
5691metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
5692    int ndvas, uint64_t txg, blkptr_t *hintbp, int flags,
5693    zio_alloc_list_t *zal, zio_t *zio, int allocator)
5694{
5695	dva_t *dva = bp->blk_dva;
5696	dva_t *hintdva = (hintbp != NULL) ? hintbp->blk_dva : NULL;
5697	int error = 0;
5698
5699	ASSERT(bp->blk_birth == 0);
5700	ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
5701
5702	spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
5703
5704	if (mc->mc_rotor == NULL) {	/* no vdevs in this class */
5705		spa_config_exit(spa, SCL_ALLOC, FTAG);
5706		return (SET_ERROR(ENOSPC));
5707	}
5708
5709	ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
5710	ASSERT(BP_GET_NDVAS(bp) == 0);
5711	ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
5712	ASSERT3P(zal, !=, NULL);
5713
5714	for (int d = 0; d < ndvas; d++) {
5715		error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
5716		    txg, flags, zal, allocator);
5717		if (error != 0) {
5718			for (d--; d >= 0; d--) {
5719				metaslab_unalloc_dva(spa, &dva[d], txg);
5720				metaslab_group_alloc_decrement(spa,
5721				    DVA_GET_VDEV(&dva[d]), zio, flags,
5722				    allocator, B_FALSE);
5723				bzero(&dva[d], sizeof (dva_t));
5724			}
5725			spa_config_exit(spa, SCL_ALLOC, FTAG);
5726			return (error);
5727		} else {
5728			/*
5729			 * Update the metaslab group's queue depth
5730			 * based on the newly allocated dva.
5731			 */
5732			metaslab_group_alloc_increment(spa,
5733			    DVA_GET_VDEV(&dva[d]), zio, flags, allocator);
5734		}
5735
5736	}
5737	ASSERT(error == 0);
5738	ASSERT(BP_GET_NDVAS(bp) == ndvas);
5739
5740	spa_config_exit(spa, SCL_ALLOC, FTAG);
5741
5742	BP_SET_BIRTH(bp, txg, txg);
5743
5744	return (0);
5745}
5746
5747void
5748metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
5749{
5750	const dva_t *dva = bp->blk_dva;
5751	int ndvas = BP_GET_NDVAS(bp);
5752
5753	ASSERT(!BP_IS_HOLE(bp));
5754	ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
5755
5756	/*
5757	 * If we have a checkpoint for the pool we need to make sure that
5758	 * the blocks that we free that are part of the checkpoint won't be
5759	 * reused until the checkpoint is discarded or we revert to it.
5760	 *
5761	 * The checkpoint flag is passed down the metaslab_free code path
5762	 * and is set whenever we want to add a block to the checkpoint's
5763	 * accounting. That is, we "checkpoint" blocks that existed at the
5764	 * time the checkpoint was created and are therefore referenced by
5765	 * the checkpointed uberblock.
5766	 *
5767	 * Note that, we don't checkpoint any blocks if the current
5768	 * syncing txg <= spa_checkpoint_txg. We want these frees to sync
5769	 * normally as they will be referenced by the checkpointed uberblock.
5770	 */
5771	boolean_t checkpoint = B_FALSE;
5772	if (bp->blk_birth <= spa->spa_checkpoint_txg &&
5773	    spa_syncing_txg(spa) > spa->spa_checkpoint_txg) {
5774		/*
5775		 * At this point, if the block is part of the checkpoint
5776		 * there is no way it was created in the current txg.
5777		 */
5778		ASSERT(!now);
5779		ASSERT3U(spa_syncing_txg(spa), ==, txg);
5780		checkpoint = B_TRUE;
5781	}
5782
5783	spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
5784
5785	for (int d = 0; d < ndvas; d++) {
5786		if (now) {
5787			metaslab_unalloc_dva(spa, &dva[d], txg);
5788		} else {
5789			ASSERT3U(txg, ==, spa_syncing_txg(spa));
5790			metaslab_free_dva(spa, &dva[d], checkpoint);
5791		}
5792	}
5793
5794	spa_config_exit(spa, SCL_FREE, FTAG);
5795}
5796
5797int
5798metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
5799{
5800	const dva_t *dva = bp->blk_dva;
5801	int ndvas = BP_GET_NDVAS(bp);
5802	int error = 0;
5803
5804	ASSERT(!BP_IS_HOLE(bp));
5805
5806	if (txg != 0) {
5807		/*
5808		 * First do a dry run to make sure all DVAs are claimable,
5809		 * so we don't have to unwind from partial failures below.
5810		 */
5811		if ((error = metaslab_claim(spa, bp, 0)) != 0)
5812			return (error);
5813	}
5814
5815	spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
5816
5817	for (int d = 0; d < ndvas; d++) {
5818		error = metaslab_claim_dva(spa, &dva[d], txg);
5819		if (error != 0)
5820			break;
5821	}
5822
5823	spa_config_exit(spa, SCL_ALLOC, FTAG);
5824
5825	ASSERT(error == 0 || txg == 0);
5826
5827	return (error);
5828}
5829
5830/* ARGSUSED */
5831static void
5832metaslab_check_free_impl_cb(uint64_t inner, vdev_t *vd, uint64_t offset,
5833    uint64_t size, void *arg)
5834{
5835	if (vd->vdev_ops == &vdev_indirect_ops)
5836		return;
5837
5838	metaslab_check_free_impl(vd, offset, size);
5839}
5840
5841static void
5842metaslab_check_free_impl(vdev_t *vd, uint64_t offset, uint64_t size)
5843{
5844	metaslab_t *msp;
5845	spa_t *spa = vd->vdev_spa;
5846
5847	if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
5848		return;
5849
5850	if (vd->vdev_ops->vdev_op_remap != NULL) {
5851		vd->vdev_ops->vdev_op_remap(vd, offset, size,
5852		    metaslab_check_free_impl_cb, NULL);
5853		return;
5854	}
5855
5856	ASSERT(vdev_is_concrete(vd));
5857	ASSERT3U(offset >> vd->vdev_ms_shift, <, vd->vdev_ms_count);
5858	ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
5859
5860	msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
5861
5862	mutex_enter(&msp->ms_lock);
5863	if (msp->ms_loaded) {
5864		range_tree_verify_not_present(msp->ms_allocatable,
5865		    offset, size);
5866	}
5867
5868	/*
5869	 * Check all segments that currently exist in the freeing pipeline.
5870	 *
5871	 * It would intuitively make sense to also check the current allocating
5872	 * tree since metaslab_unalloc_dva() exists for extents that are
5873	 * allocated and freed in the same sync pass withing the same txg.
5874	 * Unfortunately there are places (e.g. the ZIL) where we allocate a
5875	 * segment but then we free part of it within the same txg
5876	 * [see zil_sync()]. Thus, we don't call range_tree_verify() in the
5877	 * current allocating tree.
5878	 */
5879	range_tree_verify_not_present(msp->ms_freeing, offset, size);
5880	range_tree_verify_not_present(msp->ms_checkpointing, offset, size);
5881	range_tree_verify_not_present(msp->ms_freed, offset, size);
5882	for (int j = 0; j < TXG_DEFER_SIZE; j++)
5883		range_tree_verify_not_present(msp->ms_defer[j], offset, size);
5884	range_tree_verify_not_present(msp->ms_trim, offset, size);
5885	mutex_exit(&msp->ms_lock);
5886}
5887
5888void
5889metaslab_check_free(spa_t *spa, const blkptr_t *bp)
5890{
5891	if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
5892		return;
5893
5894	spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
5895	for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
5896		uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
5897		vdev_t *vd = vdev_lookup_top(spa, vdev);
5898		uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
5899		uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
5900
5901		if (DVA_GET_GANG(&bp->blk_dva[i]))
5902			size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
5903
5904		ASSERT3P(vd, !=, NULL);
5905
5906		metaslab_check_free_impl(vd, offset, size);
5907	}
5908	spa_config_exit(spa, SCL_VDEV, FTAG);
5909}
5910
5911static void
5912metaslab_group_disable_wait(metaslab_group_t *mg)
5913{
5914	ASSERT(MUTEX_HELD(&mg->mg_ms_disabled_lock));
5915	while (mg->mg_disabled_updating) {
5916		cv_wait(&mg->mg_ms_disabled_cv, &mg->mg_ms_disabled_lock);
5917	}
5918}
5919
5920static void
5921metaslab_group_disabled_increment(metaslab_group_t *mg)
5922{
5923	ASSERT(MUTEX_HELD(&mg->mg_ms_disabled_lock));
5924	ASSERT(mg->mg_disabled_updating);
5925
5926	while (mg->mg_ms_disabled >= max_disabled_ms) {
5927		cv_wait(&mg->mg_ms_disabled_cv, &mg->mg_ms_disabled_lock);
5928	}
5929	mg->mg_ms_disabled++;
5930	ASSERT3U(mg->mg_ms_disabled, <=, max_disabled_ms);
5931}
5932
5933/*
5934 * Mark the metaslab as disabled to prevent any allocations on this metaslab.
5935 * We must also track how many metaslabs are currently disabled within a
5936 * metaslab group and limit them to prevent allocation failures from
5937 * occurring because all metaslabs are disabled.
5938 */
5939void
5940metaslab_disable(metaslab_t *msp)
5941{
5942	ASSERT(!MUTEX_HELD(&msp->ms_lock));
5943	metaslab_group_t *mg = msp->ms_group;
5944
5945	mutex_enter(&mg->mg_ms_disabled_lock);
5946
5947	/*
5948	 * To keep an accurate count of how many threads have disabled
5949	 * a specific metaslab group, we only allow one thread to mark
5950	 * the metaslab group at a time. This ensures that the value of
5951	 * ms_disabled will be accurate when we decide to mark a metaslab
5952	 * group as disabled. To do this we force all other threads
5953	 * to wait till the metaslab's mg_disabled_updating flag is no
5954	 * longer set.
5955	 */
5956	metaslab_group_disable_wait(mg);
5957	mg->mg_disabled_updating = B_TRUE;
5958	if (msp->ms_disabled == 0) {
5959		metaslab_group_disabled_increment(mg);
5960	}
5961	mutex_enter(&msp->ms_lock);
5962	msp->ms_disabled++;
5963	mutex_exit(&msp->ms_lock);
5964
5965	mg->mg_disabled_updating = B_FALSE;
5966	cv_broadcast(&mg->mg_ms_disabled_cv);
5967	mutex_exit(&mg->mg_ms_disabled_lock);
5968}
5969
5970void
5971metaslab_enable(metaslab_t *msp, boolean_t sync,