xref: /titanic_50/usr/src/uts/common/fs/zfs/metaslab.c (revision 10d63b7db37a83b39c7f511cf9426c9d03ea0760)
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, 2014 by Delphix. All rights reserved.
24  * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25  */
26 
27 #include <sys/zfs_context.h>
28 #include <sys/dmu.h>
29 #include <sys/dmu_tx.h>
30 #include <sys/space_map.h>
31 #include <sys/metaslab_impl.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/zio.h>
34 #include <sys/spa_impl.h>
35 #include <sys/zfeature.h>
36 
37 /*
38  * Allow allocations to switch to gang blocks quickly. We do this to
39  * avoid having to load lots of space_maps in a given txg. There are,
40  * however, some cases where we want to avoid "fast" ganging and instead
41  * we want to do an exhaustive search of all metaslabs on this device.
42  * Currently we don't allow any gang, slog, or dump device related allocations
43  * to "fast" gang.
44  */
45 #define	CAN_FASTGANG(flags) \
46 	(!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \
47 	METASLAB_GANG_AVOID)))
48 
49 #define	METASLAB_WEIGHT_PRIMARY		(1ULL << 63)
50 #define	METASLAB_WEIGHT_SECONDARY	(1ULL << 62)
51 #define	METASLAB_ACTIVE_MASK		\
52 	(METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
53 
54 uint64_t metaslab_aliquot = 512ULL << 10;
55 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1;	/* force gang blocks */
56 
57 /*
58  * The in-core space map representation is more compact than its on-disk form.
59  * The zfs_condense_pct determines how much more compact the in-core
60  * space_map representation must be before we compact it on-disk.
61  * Values should be greater than or equal to 100.
62  */
63 int zfs_condense_pct = 200;
64 
65 /*
66  * Condensing a metaslab is not guaranteed to actually reduce the amount of
67  * space used on disk. In particular, a space map uses data in increments of
68  * MAX(1 << ashift, space_map_blksize), so a metaslab might use the
69  * same number of blocks after condensing. Since the goal of condensing is to
70  * reduce the number of IOPs required to read the space map, we only want to
71  * condense when we can be sure we will reduce the number of blocks used by the
72  * space map. Unfortunately, we cannot precisely compute whether or not this is
73  * the case in metaslab_should_condense since we are holding ms_lock. Instead,
74  * we apply the following heuristic: do not condense a spacemap unless the
75  * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
76  * blocks.
77  */
78 int zfs_metaslab_condense_block_threshold = 4;
79 
80 /*
81  * The zfs_mg_noalloc_threshold defines which metaslab groups should
82  * be eligible for allocation. The value is defined as a percentage of
83  * free space. Metaslab groups that have more free space than
84  * zfs_mg_noalloc_threshold are always eligible for allocations. Once
85  * a metaslab group's free space is less than or equal to the
86  * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
87  * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
88  * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
89  * groups are allowed to accept allocations. Gang blocks are always
90  * eligible to allocate on any metaslab group. The default value of 0 means
91  * no metaslab group will be excluded based on this criterion.
92  */
93 int zfs_mg_noalloc_threshold = 0;
94 
95 /*
96  * Metaslab groups are considered eligible for allocations if their
97  * fragmenation metric (measured as a percentage) is less than or equal to
98  * zfs_mg_fragmentation_threshold. If a metaslab group exceeds this threshold
99  * then it will be skipped unless all metaslab groups within the metaslab
100  * class have also crossed this threshold.
101  */
102 int zfs_mg_fragmentation_threshold = 85;
103 
104 /*
105  * Allow metaslabs to keep their active state as long as their fragmentation
106  * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
107  * active metaslab that exceeds this threshold will no longer keep its active
108  * status allowing better metaslabs to be selected.
109  */
110 int zfs_metaslab_fragmentation_threshold = 70;
111 
112 /*
113  * When set will load all metaslabs when pool is first opened.
114  */
115 int metaslab_debug_load = 0;
116 
117 /*
118  * When set will prevent metaslabs from being unloaded.
119  */
120 int metaslab_debug_unload = 0;
121 
122 /*
123  * Minimum size which forces the dynamic allocator to change
124  * it's allocation strategy.  Once the space map cannot satisfy
125  * an allocation of this size then it switches to using more
126  * aggressive strategy (i.e search by size rather than offset).
127  */
128 uint64_t metaslab_df_alloc_threshold = SPA_OLD_MAXBLOCKSIZE;
129 
130 /*
131  * The minimum free space, in percent, which must be available
132  * in a space map to continue allocations in a first-fit fashion.
133  * Once the space_map's free space drops below this level we dynamically
134  * switch to using best-fit allocations.
135  */
136 int metaslab_df_free_pct = 4;
137 
138 /*
139  * A metaslab is considered "free" if it contains a contiguous
140  * segment which is greater than metaslab_min_alloc_size.
141  */
142 uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
143 
144 /*
145  * Percentage of all cpus that can be used by the metaslab taskq.
146  */
147 int metaslab_load_pct = 50;
148 
149 /*
150  * Determines how many txgs a metaslab may remain loaded without having any
151  * allocations from it. As long as a metaslab continues to be used we will
152  * keep it loaded.
153  */
154 int metaslab_unload_delay = TXG_SIZE * 2;
155 
156 /*
157  * Max number of metaslabs per group to preload.
158  */
159 int metaslab_preload_limit = SPA_DVAS_PER_BP;
160 
161 /*
162  * Enable/disable preloading of metaslab.
163  */
164 boolean_t metaslab_preload_enabled = B_TRUE;
165 
166 /*
167  * Enable/disable fragmentation weighting on metaslabs.
168  */
169 boolean_t metaslab_fragmentation_factor_enabled = B_TRUE;
170 
171 /*
172  * Enable/disable lba weighting (i.e. outer tracks are given preference).
173  */
174 boolean_t metaslab_lba_weighting_enabled = B_TRUE;
175 
176 /*
177  * Enable/disable metaslab group biasing.
178  */
179 boolean_t metaslab_bias_enabled = B_TRUE;
180 
181 static uint64_t metaslab_fragmentation(metaslab_t *);
182 
183 /*
184  * ==========================================================================
185  * Metaslab classes
186  * ==========================================================================
187  */
188 metaslab_class_t *
189 metaslab_class_create(spa_t *spa, metaslab_ops_t *ops)
190 {
191 	metaslab_class_t *mc;
192 
193 	mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
194 
195 	mc->mc_spa = spa;
196 	mc->mc_rotor = NULL;
197 	mc->mc_ops = ops;
198 
199 	return (mc);
200 }
201 
202 void
203 metaslab_class_destroy(metaslab_class_t *mc)
204 {
205 	ASSERT(mc->mc_rotor == NULL);
206 	ASSERT(mc->mc_alloc == 0);
207 	ASSERT(mc->mc_deferred == 0);
208 	ASSERT(mc->mc_space == 0);
209 	ASSERT(mc->mc_dspace == 0);
210 
211 	kmem_free(mc, sizeof (metaslab_class_t));
212 }
213 
214 int
215 metaslab_class_validate(metaslab_class_t *mc)
216 {
217 	metaslab_group_t *mg;
218 	vdev_t *vd;
219 
220 	/*
221 	 * Must hold one of the spa_config locks.
222 	 */
223 	ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
224 	    spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
225 
226 	if ((mg = mc->mc_rotor) == NULL)
227 		return (0);
228 
229 	do {
230 		vd = mg->mg_vd;
231 		ASSERT(vd->vdev_mg != NULL);
232 		ASSERT3P(vd->vdev_top, ==, vd);
233 		ASSERT3P(mg->mg_class, ==, mc);
234 		ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
235 	} while ((mg = mg->mg_next) != mc->mc_rotor);
236 
237 	return (0);
238 }
239 
240 void
241 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
242     int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
243 {
244 	atomic_add_64(&mc->mc_alloc, alloc_delta);
245 	atomic_add_64(&mc->mc_deferred, defer_delta);
246 	atomic_add_64(&mc->mc_space, space_delta);
247 	atomic_add_64(&mc->mc_dspace, dspace_delta);
248 }
249 
250 uint64_t
251 metaslab_class_get_alloc(metaslab_class_t *mc)
252 {
253 	return (mc->mc_alloc);
254 }
255 
256 uint64_t
257 metaslab_class_get_deferred(metaslab_class_t *mc)
258 {
259 	return (mc->mc_deferred);
260 }
261 
262 uint64_t
263 metaslab_class_get_space(metaslab_class_t *mc)
264 {
265 	return (mc->mc_space);
266 }
267 
268 uint64_t
269 metaslab_class_get_dspace(metaslab_class_t *mc)
270 {
271 	return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
272 }
273 
274 void
275 metaslab_class_histogram_verify(metaslab_class_t *mc)
276 {
277 	vdev_t *rvd = mc->mc_spa->spa_root_vdev;
278 	uint64_t *mc_hist;
279 	int i;
280 
281 	if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
282 		return;
283 
284 	mc_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
285 	    KM_SLEEP);
286 
287 	for (int c = 0; c < rvd->vdev_children; c++) {
288 		vdev_t *tvd = rvd->vdev_child[c];
289 		metaslab_group_t *mg = tvd->vdev_mg;
290 
291 		/*
292 		 * Skip any holes, uninitialized top-levels, or
293 		 * vdevs that are not in this metalab class.
294 		 */
295 		if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
296 		    mg->mg_class != mc) {
297 			continue;
298 		}
299 
300 		for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
301 			mc_hist[i] += mg->mg_histogram[i];
302 	}
303 
304 	for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
305 		VERIFY3U(mc_hist[i], ==, mc->mc_histogram[i]);
306 
307 	kmem_free(mc_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
308 }
309 
310 /*
311  * Calculate the metaslab class's fragmentation metric. The metric
312  * is weighted based on the space contribution of each metaslab group.
313  * The return value will be a number between 0 and 100 (inclusive), or
314  * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
315  * zfs_frag_table for more information about the metric.
316  */
317 uint64_t
318 metaslab_class_fragmentation(metaslab_class_t *mc)
319 {
320 	vdev_t *rvd = mc->mc_spa->spa_root_vdev;
321 	uint64_t fragmentation = 0;
322 
323 	spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
324 
325 	for (int c = 0; c < rvd->vdev_children; c++) {
326 		vdev_t *tvd = rvd->vdev_child[c];
327 		metaslab_group_t *mg = tvd->vdev_mg;
328 
329 		/*
330 		 * Skip any holes, uninitialized top-levels, or
331 		 * vdevs that are not in this metalab class.
332 		 */
333 		if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
334 		    mg->mg_class != mc) {
335 			continue;
336 		}
337 
338 		/*
339 		 * If a metaslab group does not contain a fragmentation
340 		 * metric then just bail out.
341 		 */
342 		if (mg->mg_fragmentation == ZFS_FRAG_INVALID) {
343 			spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
344 			return (ZFS_FRAG_INVALID);
345 		}
346 
347 		/*
348 		 * Determine how much this metaslab_group is contributing
349 		 * to the overall pool fragmentation metric.
350 		 */
351 		fragmentation += mg->mg_fragmentation *
352 		    metaslab_group_get_space(mg);
353 	}
354 	fragmentation /= metaslab_class_get_space(mc);
355 
356 	ASSERT3U(fragmentation, <=, 100);
357 	spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
358 	return (fragmentation);
359 }
360 
361 /*
362  * Calculate the amount of expandable space that is available in
363  * this metaslab class. If a device is expanded then its expandable
364  * space will be the amount of allocatable space that is currently not
365  * part of this metaslab class.
366  */
367 uint64_t
368 metaslab_class_expandable_space(metaslab_class_t *mc)
369 {
370 	vdev_t *rvd = mc->mc_spa->spa_root_vdev;
371 	uint64_t space = 0;
372 
373 	spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
374 	for (int c = 0; c < rvd->vdev_children; c++) {
375 		vdev_t *tvd = rvd->vdev_child[c];
376 		metaslab_group_t *mg = tvd->vdev_mg;
377 
378 		if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
379 		    mg->mg_class != mc) {
380 			continue;
381 		}
382 
383 		space += tvd->vdev_max_asize - tvd->vdev_asize;
384 	}
385 	spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
386 	return (space);
387 }
388 
389 /*
390  * ==========================================================================
391  * Metaslab groups
392  * ==========================================================================
393  */
394 static int
395 metaslab_compare(const void *x1, const void *x2)
396 {
397 	const metaslab_t *m1 = x1;
398 	const metaslab_t *m2 = x2;
399 
400 	if (m1->ms_weight < m2->ms_weight)
401 		return (1);
402 	if (m1->ms_weight > m2->ms_weight)
403 		return (-1);
404 
405 	/*
406 	 * If the weights are identical, use the offset to force uniqueness.
407 	 */
408 	if (m1->ms_start < m2->ms_start)
409 		return (-1);
410 	if (m1->ms_start > m2->ms_start)
411 		return (1);
412 
413 	ASSERT3P(m1, ==, m2);
414 
415 	return (0);
416 }
417 
418 /*
419  * Update the allocatable flag and the metaslab group's capacity.
420  * The allocatable flag is set to true if the capacity is below
421  * the zfs_mg_noalloc_threshold. If a metaslab group transitions
422  * from allocatable to non-allocatable or vice versa then the metaslab
423  * group's class is updated to reflect the transition.
424  */
425 static void
426 metaslab_group_alloc_update(metaslab_group_t *mg)
427 {
428 	vdev_t *vd = mg->mg_vd;
429 	metaslab_class_t *mc = mg->mg_class;
430 	vdev_stat_t *vs = &vd->vdev_stat;
431 	boolean_t was_allocatable;
432 
433 	ASSERT(vd == vd->vdev_top);
434 
435 	mutex_enter(&mg->mg_lock);
436 	was_allocatable = mg->mg_allocatable;
437 
438 	mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
439 	    (vs->vs_space + 1);
440 
441 	/*
442 	 * A metaslab group is considered allocatable if it has plenty
443 	 * of free space or is not heavily fragmented. We only take
444 	 * fragmentation into account if the metaslab group has a valid
445 	 * fragmentation metric (i.e. a value between 0 and 100).
446 	 */
447 	mg->mg_allocatable = (mg->mg_free_capacity > zfs_mg_noalloc_threshold &&
448 	    (mg->mg_fragmentation == ZFS_FRAG_INVALID ||
449 	    mg->mg_fragmentation <= zfs_mg_fragmentation_threshold));
450 
451 	/*
452 	 * The mc_alloc_groups maintains a count of the number of
453 	 * groups in this metaslab class that are still above the
454 	 * zfs_mg_noalloc_threshold. This is used by the allocating
455 	 * threads to determine if they should avoid allocations to
456 	 * a given group. The allocator will avoid allocations to a group
457 	 * if that group has reached or is below the zfs_mg_noalloc_threshold
458 	 * and there are still other groups that are above the threshold.
459 	 * When a group transitions from allocatable to non-allocatable or
460 	 * vice versa we update the metaslab class to reflect that change.
461 	 * When the mc_alloc_groups value drops to 0 that means that all
462 	 * groups have reached the zfs_mg_noalloc_threshold making all groups
463 	 * eligible for allocations. This effectively means that all devices
464 	 * are balanced again.
465 	 */
466 	if (was_allocatable && !mg->mg_allocatable)
467 		mc->mc_alloc_groups--;
468 	else if (!was_allocatable && mg->mg_allocatable)
469 		mc->mc_alloc_groups++;
470 
471 	mutex_exit(&mg->mg_lock);
472 }
473 
474 metaslab_group_t *
475 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
476 {
477 	metaslab_group_t *mg;
478 
479 	mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
480 	mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
481 	avl_create(&mg->mg_metaslab_tree, metaslab_compare,
482 	    sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
483 	mg->mg_vd = vd;
484 	mg->mg_class = mc;
485 	mg->mg_activation_count = 0;
486 
487 	mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct,
488 	    minclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT);
489 
490 	return (mg);
491 }
492 
493 void
494 metaslab_group_destroy(metaslab_group_t *mg)
495 {
496 	ASSERT(mg->mg_prev == NULL);
497 	ASSERT(mg->mg_next == NULL);
498 	/*
499 	 * We may have gone below zero with the activation count
500 	 * either because we never activated in the first place or
501 	 * because we're done, and possibly removing the vdev.
502 	 */
503 	ASSERT(mg->mg_activation_count <= 0);
504 
505 	taskq_destroy(mg->mg_taskq);
506 	avl_destroy(&mg->mg_metaslab_tree);
507 	mutex_destroy(&mg->mg_lock);
508 	kmem_free(mg, sizeof (metaslab_group_t));
509 }
510 
511 void
512 metaslab_group_activate(metaslab_group_t *mg)
513 {
514 	metaslab_class_t *mc = mg->mg_class;
515 	metaslab_group_t *mgprev, *mgnext;
516 
517 	ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
518 
519 	ASSERT(mc->mc_rotor != mg);
520 	ASSERT(mg->mg_prev == NULL);
521 	ASSERT(mg->mg_next == NULL);
522 	ASSERT(mg->mg_activation_count <= 0);
523 
524 	if (++mg->mg_activation_count <= 0)
525 		return;
526 
527 	mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
528 	metaslab_group_alloc_update(mg);
529 
530 	if ((mgprev = mc->mc_rotor) == NULL) {
531 		mg->mg_prev = mg;
532 		mg->mg_next = mg;
533 	} else {
534 		mgnext = mgprev->mg_next;
535 		mg->mg_prev = mgprev;
536 		mg->mg_next = mgnext;
537 		mgprev->mg_next = mg;
538 		mgnext->mg_prev = mg;
539 	}
540 	mc->mc_rotor = mg;
541 }
542 
543 void
544 metaslab_group_passivate(metaslab_group_t *mg)
545 {
546 	metaslab_class_t *mc = mg->mg_class;
547 	metaslab_group_t *mgprev, *mgnext;
548 
549 	ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
550 
551 	if (--mg->mg_activation_count != 0) {
552 		ASSERT(mc->mc_rotor != mg);
553 		ASSERT(mg->mg_prev == NULL);
554 		ASSERT(mg->mg_next == NULL);
555 		ASSERT(mg->mg_activation_count < 0);
556 		return;
557 	}
558 
559 	taskq_wait(mg->mg_taskq);
560 	metaslab_group_alloc_update(mg);
561 
562 	mgprev = mg->mg_prev;
563 	mgnext = mg->mg_next;
564 
565 	if (mg == mgnext) {
566 		mc->mc_rotor = NULL;
567 	} else {
568 		mc->mc_rotor = mgnext;
569 		mgprev->mg_next = mgnext;
570 		mgnext->mg_prev = mgprev;
571 	}
572 
573 	mg->mg_prev = NULL;
574 	mg->mg_next = NULL;
575 }
576 
577 uint64_t
578 metaslab_group_get_space(metaslab_group_t *mg)
579 {
580 	return ((1ULL << mg->mg_vd->vdev_ms_shift) * mg->mg_vd->vdev_ms_count);
581 }
582 
583 void
584 metaslab_group_histogram_verify(metaslab_group_t *mg)
585 {
586 	uint64_t *mg_hist;
587 	vdev_t *vd = mg->mg_vd;
588 	uint64_t ashift = vd->vdev_ashift;
589 	int i;
590 
591 	if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
592 		return;
593 
594 	mg_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
595 	    KM_SLEEP);
596 
597 	ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE, >=,
598 	    SPACE_MAP_HISTOGRAM_SIZE + ashift);
599 
600 	for (int m = 0; m < vd->vdev_ms_count; m++) {
601 		metaslab_t *msp = vd->vdev_ms[m];
602 
603 		if (msp->ms_sm == NULL)
604 			continue;
605 
606 		for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++)
607 			mg_hist[i + ashift] +=
608 			    msp->ms_sm->sm_phys->smp_histogram[i];
609 	}
610 
611 	for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i ++)
612 		VERIFY3U(mg_hist[i], ==, mg->mg_histogram[i]);
613 
614 	kmem_free(mg_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
615 }
616 
617 static void
618 metaslab_group_histogram_add(metaslab_group_t *mg, metaslab_t *msp)
619 {
620 	metaslab_class_t *mc = mg->mg_class;
621 	uint64_t ashift = mg->mg_vd->vdev_ashift;
622 
623 	ASSERT(MUTEX_HELD(&msp->ms_lock));
624 	if (msp->ms_sm == NULL)
625 		return;
626 
627 	mutex_enter(&mg->mg_lock);
628 	for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
629 		mg->mg_histogram[i + ashift] +=
630 		    msp->ms_sm->sm_phys->smp_histogram[i];
631 		mc->mc_histogram[i + ashift] +=
632 		    msp->ms_sm->sm_phys->smp_histogram[i];
633 	}
634 	mutex_exit(&mg->mg_lock);
635 }
636 
637 void
638 metaslab_group_histogram_remove(metaslab_group_t *mg, metaslab_t *msp)
639 {
640 	metaslab_class_t *mc = mg->mg_class;
641 	uint64_t ashift = mg->mg_vd->vdev_ashift;
642 
643 	ASSERT(MUTEX_HELD(&msp->ms_lock));
644 	if (msp->ms_sm == NULL)
645 		return;
646 
647 	mutex_enter(&mg->mg_lock);
648 	for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
649 		ASSERT3U(mg->mg_histogram[i + ashift], >=,
650 		    msp->ms_sm->sm_phys->smp_histogram[i]);
651 		ASSERT3U(mc->mc_histogram[i + ashift], >=,
652 		    msp->ms_sm->sm_phys->smp_histogram[i]);
653 
654 		mg->mg_histogram[i + ashift] -=
655 		    msp->ms_sm->sm_phys->smp_histogram[i];
656 		mc->mc_histogram[i + ashift] -=
657 		    msp->ms_sm->sm_phys->smp_histogram[i];
658 	}
659 	mutex_exit(&mg->mg_lock);
660 }
661 
662 static void
663 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
664 {
665 	ASSERT(msp->ms_group == NULL);
666 	mutex_enter(&mg->mg_lock);
667 	msp->ms_group = mg;
668 	msp->ms_weight = 0;
669 	avl_add(&mg->mg_metaslab_tree, msp);
670 	mutex_exit(&mg->mg_lock);
671 
672 	mutex_enter(&msp->ms_lock);
673 	metaslab_group_histogram_add(mg, msp);
674 	mutex_exit(&msp->ms_lock);
675 }
676 
677 static void
678 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
679 {
680 	mutex_enter(&msp->ms_lock);
681 	metaslab_group_histogram_remove(mg, msp);
682 	mutex_exit(&msp->ms_lock);
683 
684 	mutex_enter(&mg->mg_lock);
685 	ASSERT(msp->ms_group == mg);
686 	avl_remove(&mg->mg_metaslab_tree, msp);
687 	msp->ms_group = NULL;
688 	mutex_exit(&mg->mg_lock);
689 }
690 
691 static void
692 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
693 {
694 	/*
695 	 * Although in principle the weight can be any value, in
696 	 * practice we do not use values in the range [1, 511].
697 	 */
698 	ASSERT(weight >= SPA_MINBLOCKSIZE || weight == 0);
699 	ASSERT(MUTEX_HELD(&msp->ms_lock));
700 
701 	mutex_enter(&mg->mg_lock);
702 	ASSERT(msp->ms_group == mg);
703 	avl_remove(&mg->mg_metaslab_tree, msp);
704 	msp->ms_weight = weight;
705 	avl_add(&mg->mg_metaslab_tree, msp);
706 	mutex_exit(&mg->mg_lock);
707 }
708 
709 /*
710  * Calculate the fragmentation for a given metaslab group. We can use
711  * a simple average here since all metaslabs within the group must have
712  * the same size. The return value will be a value between 0 and 100
713  * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
714  * group have a fragmentation metric.
715  */
716 uint64_t
717 metaslab_group_fragmentation(metaslab_group_t *mg)
718 {
719 	vdev_t *vd = mg->mg_vd;
720 	uint64_t fragmentation = 0;
721 	uint64_t valid_ms = 0;
722 
723 	for (int m = 0; m < vd->vdev_ms_count; m++) {
724 		metaslab_t *msp = vd->vdev_ms[m];
725 
726 		if (msp->ms_fragmentation == ZFS_FRAG_INVALID)
727 			continue;
728 
729 		valid_ms++;
730 		fragmentation += msp->ms_fragmentation;
731 	}
732 
733 	if (valid_ms <= vd->vdev_ms_count / 2)
734 		return (ZFS_FRAG_INVALID);
735 
736 	fragmentation /= valid_ms;
737 	ASSERT3U(fragmentation, <=, 100);
738 	return (fragmentation);
739 }
740 
741 /*
742  * Determine if a given metaslab group should skip allocations. A metaslab
743  * group should avoid allocations if its free capacity is less than the
744  * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
745  * zfs_mg_fragmentation_threshold and there is at least one metaslab group
746  * that can still handle allocations.
747  */
748 static boolean_t
749 metaslab_group_allocatable(metaslab_group_t *mg)
750 {
751 	vdev_t *vd = mg->mg_vd;
752 	spa_t *spa = vd->vdev_spa;
753 	metaslab_class_t *mc = mg->mg_class;
754 
755 	/*
756 	 * We use two key metrics to determine if a metaslab group is
757 	 * considered allocatable -- free space and fragmentation. If
758 	 * the free space is greater than the free space threshold and
759 	 * the fragmentation is less than the fragmentation threshold then
760 	 * consider the group allocatable. There are two case when we will
761 	 * not consider these key metrics. The first is if the group is
762 	 * associated with a slog device and the second is if all groups
763 	 * in this metaslab class have already been consider ineligible
764 	 * for allocations.
765 	 */
766 	return ((mg->mg_free_capacity > zfs_mg_noalloc_threshold &&
767 	    (mg->mg_fragmentation == ZFS_FRAG_INVALID ||
768 	    mg->mg_fragmentation <= zfs_mg_fragmentation_threshold)) ||
769 	    mc != spa_normal_class(spa) || mc->mc_alloc_groups == 0);
770 }
771 
772 /*
773  * ==========================================================================
774  * Range tree callbacks
775  * ==========================================================================
776  */
777 
778 /*
779  * Comparison function for the private size-ordered tree. Tree is sorted
780  * by size, larger sizes at the end of the tree.
781  */
782 static int
783 metaslab_rangesize_compare(const void *x1, const void *x2)
784 {
785 	const range_seg_t *r1 = x1;
786 	const range_seg_t *r2 = x2;
787 	uint64_t rs_size1 = r1->rs_end - r1->rs_start;
788 	uint64_t rs_size2 = r2->rs_end - r2->rs_start;
789 
790 	if (rs_size1 < rs_size2)
791 		return (-1);
792 	if (rs_size1 > rs_size2)
793 		return (1);
794 
795 	if (r1->rs_start < r2->rs_start)
796 		return (-1);
797 
798 	if (r1->rs_start > r2->rs_start)
799 		return (1);
800 
801 	return (0);
802 }
803 
804 /*
805  * Create any block allocator specific components. The current allocators
806  * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
807  */
808 static void
809 metaslab_rt_create(range_tree_t *rt, void *arg)
810 {
811 	metaslab_t *msp = arg;
812 
813 	ASSERT3P(rt->rt_arg, ==, msp);
814 	ASSERT(msp->ms_tree == NULL);
815 
816 	avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
817 	    sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
818 }
819 
820 /*
821  * Destroy the block allocator specific components.
822  */
823 static void
824 metaslab_rt_destroy(range_tree_t *rt, void *arg)
825 {
826 	metaslab_t *msp = arg;
827 
828 	ASSERT3P(rt->rt_arg, ==, msp);
829 	ASSERT3P(msp->ms_tree, ==, rt);
830 	ASSERT0(avl_numnodes(&msp->ms_size_tree));
831 
832 	avl_destroy(&msp->ms_size_tree);
833 }
834 
835 static void
836 metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg)
837 {
838 	metaslab_t *msp = arg;
839 
840 	ASSERT3P(rt->rt_arg, ==, msp);
841 	ASSERT3P(msp->ms_tree, ==, rt);
842 	VERIFY(!msp->ms_condensing);
843 	avl_add(&msp->ms_size_tree, rs);
844 }
845 
846 static void
847 metaslab_rt_remove(range_tree_t *rt, range_seg_t *rs, void *arg)
848 {
849 	metaslab_t *msp = arg;
850 
851 	ASSERT3P(rt->rt_arg, ==, msp);
852 	ASSERT3P(msp->ms_tree, ==, rt);
853 	VERIFY(!msp->ms_condensing);
854 	avl_remove(&msp->ms_size_tree, rs);
855 }
856 
857 static void
858 metaslab_rt_vacate(range_tree_t *rt, void *arg)
859 {
860 	metaslab_t *msp = arg;
861 
862 	ASSERT3P(rt->rt_arg, ==, msp);
863 	ASSERT3P(msp->ms_tree, ==, rt);
864 
865 	/*
866 	 * Normally one would walk the tree freeing nodes along the way.
867 	 * Since the nodes are shared with the range trees we can avoid
868 	 * walking all nodes and just reinitialize the avl tree. The nodes
869 	 * will be freed by the range tree, so we don't want to free them here.
870 	 */
871 	avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
872 	    sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
873 }
874 
875 static range_tree_ops_t metaslab_rt_ops = {
876 	metaslab_rt_create,
877 	metaslab_rt_destroy,
878 	metaslab_rt_add,
879 	metaslab_rt_remove,
880 	metaslab_rt_vacate
881 };
882 
883 /*
884  * ==========================================================================
885  * Metaslab block operations
886  * ==========================================================================
887  */
888 
889 /*
890  * Return the maximum contiguous segment within the metaslab.
891  */
892 uint64_t
893 metaslab_block_maxsize(metaslab_t *msp)
894 {
895 	avl_tree_t *t = &msp->ms_size_tree;
896 	range_seg_t *rs;
897 
898 	if (t == NULL || (rs = avl_last(t)) == NULL)
899 		return (0ULL);
900 
901 	return (rs->rs_end - rs->rs_start);
902 }
903 
904 uint64_t
905 metaslab_block_alloc(metaslab_t *msp, uint64_t size)
906 {
907 	uint64_t start;
908 	range_tree_t *rt = msp->ms_tree;
909 
910 	VERIFY(!msp->ms_condensing);
911 
912 	start = msp->ms_ops->msop_alloc(msp, size);
913 	if (start != -1ULL) {
914 		vdev_t *vd = msp->ms_group->mg_vd;
915 
916 		VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
917 		VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
918 		VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
919 		range_tree_remove(rt, start, size);
920 	}
921 	return (start);
922 }
923 
924 /*
925  * ==========================================================================
926  * Common allocator routines
927  * ==========================================================================
928  */
929 
930 /*
931  * This is a helper function that can be used by the allocator to find
932  * a suitable block to allocate. This will search the specified AVL
933  * tree looking for a block that matches the specified criteria.
934  */
935 static uint64_t
936 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
937     uint64_t align)
938 {
939 	range_seg_t *rs, rsearch;
940 	avl_index_t where;
941 
942 	rsearch.rs_start = *cursor;
943 	rsearch.rs_end = *cursor + size;
944 
945 	rs = avl_find(t, &rsearch, &where);
946 	if (rs == NULL)
947 		rs = avl_nearest(t, where, AVL_AFTER);
948 
949 	while (rs != NULL) {
950 		uint64_t offset = P2ROUNDUP(rs->rs_start, align);
951 
952 		if (offset + size <= rs->rs_end) {
953 			*cursor = offset + size;
954 			return (offset);
955 		}
956 		rs = AVL_NEXT(t, rs);
957 	}
958 
959 	/*
960 	 * If we know we've searched the whole map (*cursor == 0), give up.
961 	 * Otherwise, reset the cursor to the beginning and try again.
962 	 */
963 	if (*cursor == 0)
964 		return (-1ULL);
965 
966 	*cursor = 0;
967 	return (metaslab_block_picker(t, cursor, size, align));
968 }
969 
970 /*
971  * ==========================================================================
972  * The first-fit block allocator
973  * ==========================================================================
974  */
975 static uint64_t
976 metaslab_ff_alloc(metaslab_t *msp, uint64_t size)
977 {
978 	/*
979 	 * Find the largest power of 2 block size that evenly divides the
980 	 * requested size. This is used to try to allocate blocks with similar
981 	 * alignment from the same area of the metaslab (i.e. same cursor
982 	 * bucket) but it does not guarantee that other allocations sizes
983 	 * may exist in the same region.
984 	 */
985 	uint64_t align = size & -size;
986 	uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
987 	avl_tree_t *t = &msp->ms_tree->rt_root;
988 
989 	return (metaslab_block_picker(t, cursor, size, align));
990 }
991 
992 static metaslab_ops_t metaslab_ff_ops = {
993 	metaslab_ff_alloc
994 };
995 
996 /*
997  * ==========================================================================
998  * Dynamic block allocator -
999  * Uses the first fit allocation scheme until space get low and then
1000  * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1001  * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1002  * ==========================================================================
1003  */
1004 static uint64_t
1005 metaslab_df_alloc(metaslab_t *msp, uint64_t size)
1006 {
1007 	/*
1008 	 * Find the largest power of 2 block size that evenly divides the
1009 	 * requested size. This is used to try to allocate blocks with similar
1010 	 * alignment from the same area of the metaslab (i.e. same cursor
1011 	 * bucket) but it does not guarantee that other allocations sizes
1012 	 * may exist in the same region.
1013 	 */
1014 	uint64_t align = size & -size;
1015 	uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1016 	range_tree_t *rt = msp->ms_tree;
1017 	avl_tree_t *t = &rt->rt_root;
1018 	uint64_t max_size = metaslab_block_maxsize(msp);
1019 	int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
1020 
1021 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1022 	ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1023 
1024 	if (max_size < size)
1025 		return (-1ULL);
1026 
1027 	/*
1028 	 * If we're running low on space switch to using the size
1029 	 * sorted AVL tree (best-fit).
1030 	 */
1031 	if (max_size < metaslab_df_alloc_threshold ||
1032 	    free_pct < metaslab_df_free_pct) {
1033 		t = &msp->ms_size_tree;
1034 		*cursor = 0;
1035 	}
1036 
1037 	return (metaslab_block_picker(t, cursor, size, 1ULL));
1038 }
1039 
1040 static metaslab_ops_t metaslab_df_ops = {
1041 	metaslab_df_alloc
1042 };
1043 
1044 /*
1045  * ==========================================================================
1046  * Cursor fit block allocator -
1047  * Select the largest region in the metaslab, set the cursor to the beginning
1048  * of the range and the cursor_end to the end of the range. As allocations
1049  * are made advance the cursor. Continue allocating from the cursor until
1050  * the range is exhausted and then find a new range.
1051  * ==========================================================================
1052  */
1053 static uint64_t
1054 metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
1055 {
1056 	range_tree_t *rt = msp->ms_tree;
1057 	avl_tree_t *t = &msp->ms_size_tree;
1058 	uint64_t *cursor = &msp->ms_lbas[0];
1059 	uint64_t *cursor_end = &msp->ms_lbas[1];
1060 	uint64_t offset = 0;
1061 
1062 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1063 	ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&rt->rt_root));
1064 
1065 	ASSERT3U(*cursor_end, >=, *cursor);
1066 
1067 	if ((*cursor + size) > *cursor_end) {
1068 		range_seg_t *rs;
1069 
1070 		rs = avl_last(&msp->ms_size_tree);
1071 		if (rs == NULL || (rs->rs_end - rs->rs_start) < size)
1072 			return (-1ULL);
1073 
1074 		*cursor = rs->rs_start;
1075 		*cursor_end = rs->rs_end;
1076 	}
1077 
1078 	offset = *cursor;
1079 	*cursor += size;
1080 
1081 	return (offset);
1082 }
1083 
1084 static metaslab_ops_t metaslab_cf_ops = {
1085 	metaslab_cf_alloc
1086 };
1087 
1088 /*
1089  * ==========================================================================
1090  * New dynamic fit allocator -
1091  * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1092  * contiguous blocks. If no region is found then just use the largest segment
1093  * that remains.
1094  * ==========================================================================
1095  */
1096 
1097 /*
1098  * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1099  * to request from the allocator.
1100  */
1101 uint64_t metaslab_ndf_clump_shift = 4;
1102 
1103 static uint64_t
1104 metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
1105 {
1106 	avl_tree_t *t = &msp->ms_tree->rt_root;
1107 	avl_index_t where;
1108 	range_seg_t *rs, rsearch;
1109 	uint64_t hbit = highbit64(size);
1110 	uint64_t *cursor = &msp->ms_lbas[hbit - 1];
1111 	uint64_t max_size = metaslab_block_maxsize(msp);
1112 
1113 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1114 	ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1115 
1116 	if (max_size < size)
1117 		return (-1ULL);
1118 
1119 	rsearch.rs_start = *cursor;
1120 	rsearch.rs_end = *cursor + size;
1121 
1122 	rs = avl_find(t, &rsearch, &where);
1123 	if (rs == NULL || (rs->rs_end - rs->rs_start) < size) {
1124 		t = &msp->ms_size_tree;
1125 
1126 		rsearch.rs_start = 0;
1127 		rsearch.rs_end = MIN(max_size,
1128 		    1ULL << (hbit + metaslab_ndf_clump_shift));
1129 		rs = avl_find(t, &rsearch, &where);
1130 		if (rs == NULL)
1131 			rs = avl_nearest(t, where, AVL_AFTER);
1132 		ASSERT(rs != NULL);
1133 	}
1134 
1135 	if ((rs->rs_end - rs->rs_start) >= size) {
1136 		*cursor = rs->rs_start + size;
1137 		return (rs->rs_start);
1138 	}
1139 	return (-1ULL);
1140 }
1141 
1142 static metaslab_ops_t metaslab_ndf_ops = {
1143 	metaslab_ndf_alloc
1144 };
1145 
1146 metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
1147 
1148 /*
1149  * ==========================================================================
1150  * Metaslabs
1151  * ==========================================================================
1152  */
1153 
1154 /*
1155  * Wait for any in-progress metaslab loads to complete.
1156  */
1157 void
1158 metaslab_load_wait(metaslab_t *msp)
1159 {
1160 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1161 
1162 	while (msp->ms_loading) {
1163 		ASSERT(!msp->ms_loaded);
1164 		cv_wait(&msp->ms_load_cv, &msp->ms_lock);
1165 	}
1166 }
1167 
1168 int
1169 metaslab_load(metaslab_t *msp)
1170 {
1171 	int error = 0;
1172 
1173 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1174 	ASSERT(!msp->ms_loaded);
1175 	ASSERT(!msp->ms_loading);
1176 
1177 	msp->ms_loading = B_TRUE;
1178 
1179 	/*
1180 	 * If the space map has not been allocated yet, then treat
1181 	 * all the space in the metaslab as free and add it to the
1182 	 * ms_tree.
1183 	 */
1184 	if (msp->ms_sm != NULL)
1185 		error = space_map_load(msp->ms_sm, msp->ms_tree, SM_FREE);
1186 	else
1187 		range_tree_add(msp->ms_tree, msp->ms_start, msp->ms_size);
1188 
1189 	msp->ms_loaded = (error == 0);
1190 	msp->ms_loading = B_FALSE;
1191 
1192 	if (msp->ms_loaded) {
1193 		for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1194 			range_tree_walk(msp->ms_defertree[t],
1195 			    range_tree_remove, msp->ms_tree);
1196 		}
1197 	}
1198 	cv_broadcast(&msp->ms_load_cv);
1199 	return (error);
1200 }
1201 
1202 void
1203 metaslab_unload(metaslab_t *msp)
1204 {
1205 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1206 	range_tree_vacate(msp->ms_tree, NULL, NULL);
1207 	msp->ms_loaded = B_FALSE;
1208 	msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
1209 }
1210 
1211 int
1212 metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, uint64_t txg,
1213     metaslab_t **msp)
1214 {
1215 	vdev_t *vd = mg->mg_vd;
1216 	objset_t *mos = vd->vdev_spa->spa_meta_objset;
1217 	metaslab_t *ms;
1218 	int error;
1219 
1220 	ms = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
1221 	mutex_init(&ms->ms_lock, NULL, MUTEX_DEFAULT, NULL);
1222 	cv_init(&ms->ms_load_cv, NULL, CV_DEFAULT, NULL);
1223 	ms->ms_id = id;
1224 	ms->ms_start = id << vd->vdev_ms_shift;
1225 	ms->ms_size = 1ULL << vd->vdev_ms_shift;
1226 
1227 	/*
1228 	 * We only open space map objects that already exist. All others
1229 	 * will be opened when we finally allocate an object for it.
1230 	 */
1231 	if (object != 0) {
1232 		error = space_map_open(&ms->ms_sm, mos, object, ms->ms_start,
1233 		    ms->ms_size, vd->vdev_ashift, &ms->ms_lock);
1234 
1235 		if (error != 0) {
1236 			kmem_free(ms, sizeof (metaslab_t));
1237 			return (error);
1238 		}
1239 
1240 		ASSERT(ms->ms_sm != NULL);
1241 	}
1242 
1243 	/*
1244 	 * We create the main range tree here, but we don't create the
1245 	 * alloctree and freetree until metaslab_sync_done().  This serves
1246 	 * two purposes: it allows metaslab_sync_done() to detect the
1247 	 * addition of new space; and for debugging, it ensures that we'd
1248 	 * data fault on any attempt to use this metaslab before it's ready.
1249 	 */
1250 	ms->ms_tree = range_tree_create(&metaslab_rt_ops, ms, &ms->ms_lock);
1251 	metaslab_group_add(mg, ms);
1252 
1253 	ms->ms_fragmentation = metaslab_fragmentation(ms);
1254 	ms->ms_ops = mg->mg_class->mc_ops;
1255 
1256 	/*
1257 	 * If we're opening an existing pool (txg == 0) or creating
1258 	 * a new one (txg == TXG_INITIAL), all space is available now.
1259 	 * If we're adding space to an existing pool, the new space
1260 	 * does not become available until after this txg has synced.
1261 	 */
1262 	if (txg <= TXG_INITIAL)
1263 		metaslab_sync_done(ms, 0);
1264 
1265 	/*
1266 	 * If metaslab_debug_load is set and we're initializing a metaslab
1267 	 * that has an allocated space_map object then load the its space
1268 	 * map so that can verify frees.
1269 	 */
1270 	if (metaslab_debug_load && ms->ms_sm != NULL) {
1271 		mutex_enter(&ms->ms_lock);
1272 		VERIFY0(metaslab_load(ms));
1273 		mutex_exit(&ms->ms_lock);
1274 	}
1275 
1276 	if (txg != 0) {
1277 		vdev_dirty(vd, 0, NULL, txg);
1278 		vdev_dirty(vd, VDD_METASLAB, ms, txg);
1279 	}
1280 
1281 	*msp = ms;
1282 
1283 	return (0);
1284 }
1285 
1286 void
1287 metaslab_fini(metaslab_t *msp)
1288 {
1289 	metaslab_group_t *mg = msp->ms_group;
1290 
1291 	metaslab_group_remove(mg, msp);
1292 
1293 	mutex_enter(&msp->ms_lock);
1294 
1295 	VERIFY(msp->ms_group == NULL);
1296 	vdev_space_update(mg->mg_vd, -space_map_allocated(msp->ms_sm),
1297 	    0, -msp->ms_size);
1298 	space_map_close(msp->ms_sm);
1299 
1300 	metaslab_unload(msp);
1301 	range_tree_destroy(msp->ms_tree);
1302 
1303 	for (int t = 0; t < TXG_SIZE; t++) {
1304 		range_tree_destroy(msp->ms_alloctree[t]);
1305 		range_tree_destroy(msp->ms_freetree[t]);
1306 	}
1307 
1308 	for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1309 		range_tree_destroy(msp->ms_defertree[t]);
1310 	}
1311 
1312 	ASSERT0(msp->ms_deferspace);
1313 
1314 	mutex_exit(&msp->ms_lock);
1315 	cv_destroy(&msp->ms_load_cv);
1316 	mutex_destroy(&msp->ms_lock);
1317 
1318 	kmem_free(msp, sizeof (metaslab_t));
1319 }
1320 
1321 #define	FRAGMENTATION_TABLE_SIZE	17
1322 
1323 /*
1324  * This table defines a segment size based fragmentation metric that will
1325  * allow each metaslab to derive its own fragmentation value. This is done
1326  * by calculating the space in each bucket of the spacemap histogram and
1327  * multiplying that by the fragmetation metric in this table. Doing
1328  * this for all buckets and dividing it by the total amount of free
1329  * space in this metaslab (i.e. the total free space in all buckets) gives
1330  * us the fragmentation metric. This means that a high fragmentation metric
1331  * equates to most of the free space being comprised of small segments.
1332  * Conversely, if the metric is low, then most of the free space is in
1333  * large segments. A 10% change in fragmentation equates to approximately
1334  * double the number of segments.
1335  *
1336  * This table defines 0% fragmented space using 16MB segments. Testing has
1337  * shown that segments that are greater than or equal to 16MB do not suffer
1338  * from drastic performance problems. Using this value, we derive the rest
1339  * of the table. Since the fragmentation value is never stored on disk, it
1340  * is possible to change these calculations in the future.
1341  */
1342 int zfs_frag_table[FRAGMENTATION_TABLE_SIZE] = {
1343 	100,	/* 512B	*/
1344 	100,	/* 1K	*/
1345 	98,	/* 2K	*/
1346 	95,	/* 4K	*/
1347 	90,	/* 8K	*/
1348 	80,	/* 16K	*/
1349 	70,	/* 32K	*/
1350 	60,	/* 64K	*/
1351 	50,	/* 128K	*/
1352 	40,	/* 256K	*/
1353 	30,	/* 512K	*/
1354 	20,	/* 1M	*/
1355 	15,	/* 2M	*/
1356 	10,	/* 4M	*/
1357 	5,	/* 8M	*/
1358 	0	/* 16M	*/
1359 };
1360 
1361 /*
1362  * Calclate the metaslab's fragmentation metric. A return value
1363  * of ZFS_FRAG_INVALID means that the metaslab has not been upgraded and does
1364  * not support this metric. Otherwise, the return value should be in the
1365  * range [0, 100].
1366  */
1367 static uint64_t
1368 metaslab_fragmentation(metaslab_t *msp)
1369 {
1370 	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1371 	uint64_t fragmentation = 0;
1372 	uint64_t total = 0;
1373 	boolean_t feature_enabled = spa_feature_is_enabled(spa,
1374 	    SPA_FEATURE_SPACEMAP_HISTOGRAM);
1375 
1376 	if (!feature_enabled)
1377 		return (ZFS_FRAG_INVALID);
1378 
1379 	/*
1380 	 * A null space map means that the entire metaslab is free
1381 	 * and thus is not fragmented.
1382 	 */
1383 	if (msp->ms_sm == NULL)
1384 		return (0);
1385 
1386 	/*
1387 	 * If this metaslab's space_map has not been upgraded, flag it
1388 	 * so that we upgrade next time we encounter it.
1389 	 */
1390 	if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t)) {
1391 		uint64_t txg = spa_syncing_txg(spa);
1392 		vdev_t *vd = msp->ms_group->mg_vd;
1393 
1394 		if (spa_writeable(spa)) {
1395 			msp->ms_condense_wanted = B_TRUE;
1396 			vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1397 			spa_dbgmsg(spa, "txg %llu, requesting force condense: "
1398 			    "msp %p, vd %p", txg, msp, vd);
1399 		}
1400 		return (ZFS_FRAG_INVALID);
1401 	}
1402 
1403 	for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
1404 		uint64_t space = 0;
1405 		uint8_t shift = msp->ms_sm->sm_shift;
1406 		int idx = MIN(shift - SPA_MINBLOCKSHIFT + i,
1407 		    FRAGMENTATION_TABLE_SIZE - 1);
1408 
1409 		if (msp->ms_sm->sm_phys->smp_histogram[i] == 0)
1410 			continue;
1411 
1412 		space = msp->ms_sm->sm_phys->smp_histogram[i] << (i + shift);
1413 		total += space;
1414 
1415 		ASSERT3U(idx, <, FRAGMENTATION_TABLE_SIZE);
1416 		fragmentation += space * zfs_frag_table[idx];
1417 	}
1418 
1419 	if (total > 0)
1420 		fragmentation /= total;
1421 	ASSERT3U(fragmentation, <=, 100);
1422 	return (fragmentation);
1423 }
1424 
1425 /*
1426  * Compute a weight -- a selection preference value -- for the given metaslab.
1427  * This is based on the amount of free space, the level of fragmentation,
1428  * the LBA range, and whether the metaslab is loaded.
1429  */
1430 static uint64_t
1431 metaslab_weight(metaslab_t *msp)
1432 {
1433 	metaslab_group_t *mg = msp->ms_group;
1434 	vdev_t *vd = mg->mg_vd;
1435 	uint64_t weight, space;
1436 
1437 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1438 
1439 	/*
1440 	 * This vdev is in the process of being removed so there is nothing
1441 	 * for us to do here.
1442 	 */
1443 	if (vd->vdev_removing) {
1444 		ASSERT0(space_map_allocated(msp->ms_sm));
1445 		ASSERT0(vd->vdev_ms_shift);
1446 		return (0);
1447 	}
1448 
1449 	/*
1450 	 * The baseline weight is the metaslab's free space.
1451 	 */
1452 	space = msp->ms_size - space_map_allocated(msp->ms_sm);
1453 
1454 	msp->ms_fragmentation = metaslab_fragmentation(msp);
1455 	if (metaslab_fragmentation_factor_enabled &&
1456 	    msp->ms_fragmentation != ZFS_FRAG_INVALID) {
1457 		/*
1458 		 * Use the fragmentation information to inversely scale
1459 		 * down the baseline weight. We need to ensure that we
1460 		 * don't exclude this metaslab completely when it's 100%
1461 		 * fragmented. To avoid this we reduce the fragmented value
1462 		 * by 1.
1463 		 */
1464 		space = (space * (100 - (msp->ms_fragmentation - 1))) / 100;
1465 
1466 		/*
1467 		 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
1468 		 * this metaslab again. The fragmentation metric may have
1469 		 * decreased the space to something smaller than
1470 		 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
1471 		 * so that we can consume any remaining space.
1472 		 */
1473 		if (space > 0 && space < SPA_MINBLOCKSIZE)
1474 			space = SPA_MINBLOCKSIZE;
1475 	}
1476 	weight = space;
1477 
1478 	/*
1479 	 * Modern disks have uniform bit density and constant angular velocity.
1480 	 * Therefore, the outer recording zones are faster (higher bandwidth)
1481 	 * than the inner zones by the ratio of outer to inner track diameter,
1482 	 * which is typically around 2:1.  We account for this by assigning
1483 	 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1484 	 * In effect, this means that we'll select the metaslab with the most
1485 	 * free bandwidth rather than simply the one with the most free space.
1486 	 */
1487 	if (metaslab_lba_weighting_enabled) {
1488 		weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count;
1489 		ASSERT(weight >= space && weight <= 2 * space);
1490 	}
1491 
1492 	/*
1493 	 * If this metaslab is one we're actively using, adjust its
1494 	 * weight to make it preferable to any inactive metaslab so
1495 	 * we'll polish it off. If the fragmentation on this metaslab
1496 	 * has exceed our threshold, then don't mark it active.
1497 	 */
1498 	if (msp->ms_loaded && msp->ms_fragmentation != ZFS_FRAG_INVALID &&
1499 	    msp->ms_fragmentation <= zfs_metaslab_fragmentation_threshold) {
1500 		weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
1501 	}
1502 
1503 	return (weight);
1504 }
1505 
1506 static int
1507 metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
1508 {
1509 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1510 
1511 	if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1512 		metaslab_load_wait(msp);
1513 		if (!msp->ms_loaded) {
1514 			int error = metaslab_load(msp);
1515 			if (error) {
1516 				metaslab_group_sort(msp->ms_group, msp, 0);
1517 				return (error);
1518 			}
1519 		}
1520 
1521 		metaslab_group_sort(msp->ms_group, msp,
1522 		    msp->ms_weight | activation_weight);
1523 	}
1524 	ASSERT(msp->ms_loaded);
1525 	ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
1526 
1527 	return (0);
1528 }
1529 
1530 static void
1531 metaslab_passivate(metaslab_t *msp, uint64_t size)
1532 {
1533 	/*
1534 	 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1535 	 * this metaslab again.  In that case, it had better be empty,
1536 	 * or we would be leaving space on the table.
1537 	 */
1538 	ASSERT(size >= SPA_MINBLOCKSIZE || range_tree_space(msp->ms_tree) == 0);
1539 	metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
1540 	ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
1541 }
1542 
1543 static void
1544 metaslab_preload(void *arg)
1545 {
1546 	metaslab_t *msp = arg;
1547 	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1548 
1549 	ASSERT(!MUTEX_HELD(&msp->ms_group->mg_lock));
1550 
1551 	mutex_enter(&msp->ms_lock);
1552 	metaslab_load_wait(msp);
1553 	if (!msp->ms_loaded)
1554 		(void) metaslab_load(msp);
1555 
1556 	/*
1557 	 * Set the ms_access_txg value so that we don't unload it right away.
1558 	 */
1559 	msp->ms_access_txg = spa_syncing_txg(spa) + metaslab_unload_delay + 1;
1560 	mutex_exit(&msp->ms_lock);
1561 }
1562 
1563 static void
1564 metaslab_group_preload(metaslab_group_t *mg)
1565 {
1566 	spa_t *spa = mg->mg_vd->vdev_spa;
1567 	metaslab_t *msp;
1568 	avl_tree_t *t = &mg->mg_metaslab_tree;
1569 	int m = 0;
1570 
1571 	if (spa_shutting_down(spa) || !metaslab_preload_enabled) {
1572 		taskq_wait(mg->mg_taskq);
1573 		return;
1574 	}
1575 
1576 	mutex_enter(&mg->mg_lock);
1577 	/*
1578 	 * Load the next potential metaslabs
1579 	 */
1580 	msp = avl_first(t);
1581 	while (msp != NULL) {
1582 		metaslab_t *msp_next = AVL_NEXT(t, msp);
1583 
1584 		/*
1585 		 * We preload only the maximum number of metaslabs specified
1586 		 * by metaslab_preload_limit. If a metaslab is being forced
1587 		 * to condense then we preload it too. This will ensure
1588 		 * that force condensing happens in the next txg.
1589 		 */
1590 		if (++m > metaslab_preload_limit && !msp->ms_condense_wanted) {
1591 			msp = msp_next;
1592 			continue;
1593 		}
1594 
1595 		/*
1596 		 * We must drop the metaslab group lock here to preserve
1597 		 * lock ordering with the ms_lock (when grabbing both
1598 		 * the mg_lock and the ms_lock, the ms_lock must be taken
1599 		 * first).  As a result, it is possible that the ordering
1600 		 * of the metaslabs within the avl tree may change before
1601 		 * we reacquire the lock. The metaslab cannot be removed from
1602 		 * the tree while we're in syncing context so it is safe to
1603 		 * drop the mg_lock here. If the metaslabs are reordered
1604 		 * nothing will break -- we just may end up loading a
1605 		 * less than optimal one.
1606 		 */
1607 		mutex_exit(&mg->mg_lock);
1608 		VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
1609 		    msp, TQ_SLEEP) != NULL);
1610 		mutex_enter(&mg->mg_lock);
1611 		msp = msp_next;
1612 	}
1613 	mutex_exit(&mg->mg_lock);
1614 }
1615 
1616 /*
1617  * Determine if the space map's on-disk footprint is past our tolerance
1618  * for inefficiency. We would like to use the following criteria to make
1619  * our decision:
1620  *
1621  * 1. The size of the space map object should not dramatically increase as a
1622  * result of writing out the free space range tree.
1623  *
1624  * 2. The minimal on-disk space map representation is zfs_condense_pct/100
1625  * times the size than the free space range tree representation
1626  * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
1627  *
1628  * 3. The on-disk size of the space map should actually decrease.
1629  *
1630  * Checking the first condition is tricky since we don't want to walk
1631  * the entire AVL tree calculating the estimated on-disk size. Instead we
1632  * use the size-ordered range tree in the metaslab and calculate the
1633  * size required to write out the largest segment in our free tree. If the
1634  * size required to represent that segment on disk is larger than the space
1635  * map object then we avoid condensing this map.
1636  *
1637  * To determine the second criterion we use a best-case estimate and assume
1638  * each segment can be represented on-disk as a single 64-bit entry. We refer
1639  * to this best-case estimate as the space map's minimal form.
1640  *
1641  * Unfortunately, we cannot compute the on-disk size of the space map in this
1642  * context because we cannot accurately compute the effects of compression, etc.
1643  * Instead, we apply the heuristic described in the block comment for
1644  * zfs_metaslab_condense_block_threshold - we only condense if the space used
1645  * is greater than a threshold number of blocks.
1646  */
1647 static boolean_t
1648 metaslab_should_condense(metaslab_t *msp)
1649 {
1650 	space_map_t *sm = msp->ms_sm;
1651 	range_seg_t *rs;
1652 	uint64_t size, entries, segsz, object_size, optimal_size, record_size;
1653 	dmu_object_info_t doi;
1654 	uint64_t vdev_blocksize = 1 << msp->ms_group->mg_vd->vdev_ashift;
1655 
1656 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1657 	ASSERT(msp->ms_loaded);
1658 
1659 	/*
1660 	 * Use the ms_size_tree range tree, which is ordered by size, to
1661 	 * obtain the largest segment in the free tree. We always condense
1662 	 * metaslabs that are empty and metaslabs for which a condense
1663 	 * request has been made.
1664 	 */
1665 	rs = avl_last(&msp->ms_size_tree);
1666 	if (rs == NULL || msp->ms_condense_wanted)
1667 		return (B_TRUE);
1668 
1669 	/*
1670 	 * Calculate the number of 64-bit entries this segment would
1671 	 * require when written to disk. If this single segment would be
1672 	 * larger on-disk than the entire current on-disk structure, then
1673 	 * clearly condensing will increase the on-disk structure size.
1674 	 */
1675 	size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
1676 	entries = size / (MIN(size, SM_RUN_MAX));
1677 	segsz = entries * sizeof (uint64_t);
1678 
1679 	optimal_size = sizeof (uint64_t) * avl_numnodes(&msp->ms_tree->rt_root);
1680 	object_size = space_map_length(msp->ms_sm);
1681 
1682 	dmu_object_info_from_db(sm->sm_dbuf, &doi);
1683 	record_size = MAX(doi.doi_data_block_size, vdev_blocksize);
1684 
1685 	return (segsz <= object_size &&
1686 	    object_size >= (optimal_size * zfs_condense_pct / 100) &&
1687 	    object_size > zfs_metaslab_condense_block_threshold * record_size);
1688 }
1689 
1690 /*
1691  * Condense the on-disk space map representation to its minimized form.
1692  * The minimized form consists of a small number of allocations followed by
1693  * the entries of the free range tree.
1694  */
1695 static void
1696 metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
1697 {
1698 	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1699 	range_tree_t *freetree = msp->ms_freetree[txg & TXG_MASK];
1700 	range_tree_t *condense_tree;
1701 	space_map_t *sm = msp->ms_sm;
1702 
1703 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1704 	ASSERT3U(spa_sync_pass(spa), ==, 1);
1705 	ASSERT(msp->ms_loaded);
1706 
1707 
1708 	spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, "
1709 	    "smp size %llu, segments %lu, forcing condense=%s", txg,
1710 	    msp->ms_id, msp, space_map_length(msp->ms_sm),
1711 	    avl_numnodes(&msp->ms_tree->rt_root),
1712 	    msp->ms_condense_wanted ? "TRUE" : "FALSE");
1713 
1714 	msp->ms_condense_wanted = B_FALSE;
1715 
1716 	/*
1717 	 * Create an range tree that is 100% allocated. We remove segments
1718 	 * that have been freed in this txg, any deferred frees that exist,
1719 	 * and any allocation in the future. Removing segments should be
1720 	 * a relatively inexpensive operation since we expect these trees to
1721 	 * have a small number of nodes.
1722 	 */
1723 	condense_tree = range_tree_create(NULL, NULL, &msp->ms_lock);
1724 	range_tree_add(condense_tree, msp->ms_start, msp->ms_size);
1725 
1726 	/*
1727 	 * Remove what's been freed in this txg from the condense_tree.
1728 	 * Since we're in sync_pass 1, we know that all the frees from
1729 	 * this txg are in the freetree.
1730 	 */
1731 	range_tree_walk(freetree, range_tree_remove, condense_tree);
1732 
1733 	for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1734 		range_tree_walk(msp->ms_defertree[t],
1735 		    range_tree_remove, condense_tree);
1736 	}
1737 
1738 	for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
1739 		range_tree_walk(msp->ms_alloctree[(txg + t) & TXG_MASK],
1740 		    range_tree_remove, condense_tree);
1741 	}
1742 
1743 	/*
1744 	 * We're about to drop the metaslab's lock thus allowing
1745 	 * other consumers to change it's content. Set the
1746 	 * metaslab's ms_condensing flag to ensure that
1747 	 * allocations on this metaslab do not occur while we're
1748 	 * in the middle of committing it to disk. This is only critical
1749 	 * for the ms_tree as all other range trees use per txg
1750 	 * views of their content.
1751 	 */
1752 	msp->ms_condensing = B_TRUE;
1753 
1754 	mutex_exit(&msp->ms_lock);
1755 	space_map_truncate(sm, tx);
1756 	mutex_enter(&msp->ms_lock);
1757 
1758 	/*
1759 	 * While we would ideally like to create a space_map representation
1760 	 * that consists only of allocation records, doing so can be
1761 	 * prohibitively expensive because the in-core free tree can be
1762 	 * large, and therefore computationally expensive to subtract
1763 	 * from the condense_tree. Instead we sync out two trees, a cheap
1764 	 * allocation only tree followed by the in-core free tree. While not
1765 	 * optimal, this is typically close to optimal, and much cheaper to
1766 	 * compute.
1767 	 */
1768 	space_map_write(sm, condense_tree, SM_ALLOC, tx);
1769 	range_tree_vacate(condense_tree, NULL, NULL);
1770 	range_tree_destroy(condense_tree);
1771 
1772 	space_map_write(sm, msp->ms_tree, SM_FREE, tx);
1773 	msp->ms_condensing = B_FALSE;
1774 }
1775 
1776 /*
1777  * Write a metaslab to disk in the context of the specified transaction group.
1778  */
1779 void
1780 metaslab_sync(metaslab_t *msp, uint64_t txg)
1781 {
1782 	metaslab_group_t *mg = msp->ms_group;
1783 	vdev_t *vd = mg->mg_vd;
1784 	spa_t *spa = vd->vdev_spa;
1785 	objset_t *mos = spa_meta_objset(spa);
1786 	range_tree_t *alloctree = msp->ms_alloctree[txg & TXG_MASK];
1787 	range_tree_t **freetree = &msp->ms_freetree[txg & TXG_MASK];
1788 	range_tree_t **freed_tree =
1789 	    &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
1790 	dmu_tx_t *tx;
1791 	uint64_t object = space_map_object(msp->ms_sm);
1792 
1793 	ASSERT(!vd->vdev_ishole);
1794 
1795 	/*
1796 	 * This metaslab has just been added so there's no work to do now.
1797 	 */
1798 	if (*freetree == NULL) {
1799 		ASSERT3P(alloctree, ==, NULL);
1800 		return;
1801 	}
1802 
1803 	ASSERT3P(alloctree, !=, NULL);
1804 	ASSERT3P(*freetree, !=, NULL);
1805 	ASSERT3P(*freed_tree, !=, NULL);
1806 
1807 	/*
1808 	 * Normally, we don't want to process a metaslab if there
1809 	 * are no allocations or frees to perform. However, if the metaslab
1810 	 * is being forced to condense we need to let it through.
1811 	 */
1812 	if (range_tree_space(alloctree) == 0 &&
1813 	    range_tree_space(*freetree) == 0 &&
1814 	    !msp->ms_condense_wanted)
1815 		return;
1816 
1817 	/*
1818 	 * The only state that can actually be changing concurrently with
1819 	 * metaslab_sync() is the metaslab's ms_tree.  No other thread can
1820 	 * be modifying this txg's alloctree, freetree, freed_tree, or
1821 	 * space_map_phys_t. Therefore, we only hold ms_lock to satify
1822 	 * space_map ASSERTs. We drop it whenever we call into the DMU,
1823 	 * because the DMU can call down to us (e.g. via zio_free()) at
1824 	 * any time.
1825 	 */
1826 
1827 	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
1828 
1829 	if (msp->ms_sm == NULL) {
1830 		uint64_t new_object;
1831 
1832 		new_object = space_map_alloc(mos, tx);
1833 		VERIFY3U(new_object, !=, 0);
1834 
1835 		VERIFY0(space_map_open(&msp->ms_sm, mos, new_object,
1836 		    msp->ms_start, msp->ms_size, vd->vdev_ashift,
1837 		    &msp->ms_lock));
1838 		ASSERT(msp->ms_sm != NULL);
1839 	}
1840 
1841 	mutex_enter(&msp->ms_lock);
1842 
1843 	/*
1844 	 * Note: metaslab_condense() clears the space_map's histogram.
1845 	 * Therefore we must verify and remove this histogram before
1846 	 * condensing.
1847 	 */
1848 	metaslab_group_histogram_verify(mg);
1849 	metaslab_class_histogram_verify(mg->mg_class);
1850 	metaslab_group_histogram_remove(mg, msp);
1851 
1852 	if (msp->ms_loaded && spa_sync_pass(spa) == 1 &&
1853 	    metaslab_should_condense(msp)) {
1854 		metaslab_condense(msp, txg, tx);
1855 	} else {
1856 		space_map_write(msp->ms_sm, alloctree, SM_ALLOC, tx);
1857 		space_map_write(msp->ms_sm, *freetree, SM_FREE, tx);
1858 	}
1859 
1860 	if (msp->ms_loaded) {
1861 		/*
1862 		 * When the space map is loaded, we have an accruate
1863 		 * histogram in the range tree. This gives us an opportunity
1864 		 * to bring the space map's histogram up-to-date so we clear
1865 		 * it first before updating it.
1866 		 */
1867 		space_map_histogram_clear(msp->ms_sm);
1868 		space_map_histogram_add(msp->ms_sm, msp->ms_tree, tx);
1869 	} else {
1870 		/*
1871 		 * Since the space map is not loaded we simply update the
1872 		 * exisiting histogram with what was freed in this txg. This
1873 		 * means that the on-disk histogram may not have an accurate
1874 		 * view of the free space but it's close enough to allow
1875 		 * us to make allocation decisions.
1876 		 */
1877 		space_map_histogram_add(msp->ms_sm, *freetree, tx);
1878 	}
1879 	metaslab_group_histogram_add(mg, msp);
1880 	metaslab_group_histogram_verify(mg);
1881 	metaslab_class_histogram_verify(mg->mg_class);
1882 
1883 	/*
1884 	 * For sync pass 1, we avoid traversing this txg's free range tree
1885 	 * and instead will just swap the pointers for freetree and
1886 	 * freed_tree. We can safely do this since the freed_tree is
1887 	 * guaranteed to be empty on the initial pass.
1888 	 */
1889 	if (spa_sync_pass(spa) == 1) {
1890 		range_tree_swap(freetree, freed_tree);
1891 	} else {
1892 		range_tree_vacate(*freetree, range_tree_add, *freed_tree);
1893 	}
1894 	range_tree_vacate(alloctree, NULL, NULL);
1895 
1896 	ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
1897 	ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
1898 
1899 	mutex_exit(&msp->ms_lock);
1900 
1901 	if (object != space_map_object(msp->ms_sm)) {
1902 		object = space_map_object(msp->ms_sm);
1903 		dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
1904 		    msp->ms_id, sizeof (uint64_t), &object, tx);
1905 	}
1906 	dmu_tx_commit(tx);
1907 }
1908 
1909 /*
1910  * Called after a transaction group has completely synced to mark
1911  * all of the metaslab's free space as usable.
1912  */
1913 void
1914 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
1915 {
1916 	metaslab_group_t *mg = msp->ms_group;
1917 	vdev_t *vd = mg->mg_vd;
1918 	range_tree_t **freed_tree;
1919 	range_tree_t **defer_tree;
1920 	int64_t alloc_delta, defer_delta;
1921 
1922 	ASSERT(!vd->vdev_ishole);
1923 
1924 	mutex_enter(&msp->ms_lock);
1925 
1926 	/*
1927 	 * If this metaslab is just becoming available, initialize its
1928 	 * alloctrees, freetrees, and defertree and add its capacity to
1929 	 * the vdev.
1930 	 */
1931 	if (msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK] == NULL) {
1932 		for (int t = 0; t < TXG_SIZE; t++) {
1933 			ASSERT(msp->ms_alloctree[t] == NULL);
1934 			ASSERT(msp->ms_freetree[t] == NULL);
1935 
1936 			msp->ms_alloctree[t] = range_tree_create(NULL, msp,
1937 			    &msp->ms_lock);
1938 			msp->ms_freetree[t] = range_tree_create(NULL, msp,
1939 			    &msp->ms_lock);
1940 		}
1941 
1942 		for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1943 			ASSERT(msp->ms_defertree[t] == NULL);
1944 
1945 			msp->ms_defertree[t] = range_tree_create(NULL, msp,
1946 			    &msp->ms_lock);
1947 		}
1948 
1949 		vdev_space_update(vd, 0, 0, msp->ms_size);
1950 	}
1951 
1952 	freed_tree = &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
1953 	defer_tree = &msp->ms_defertree[txg % TXG_DEFER_SIZE];
1954 
1955 	alloc_delta = space_map_alloc_delta(msp->ms_sm);
1956 	defer_delta = range_tree_space(*freed_tree) -
1957 	    range_tree_space(*defer_tree);
1958 
1959 	vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
1960 
1961 	ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
1962 	ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
1963 
1964 	/*
1965 	 * If there's a metaslab_load() in progress, wait for it to complete
1966 	 * so that we have a consistent view of the in-core space map.
1967 	 */
1968 	metaslab_load_wait(msp);
1969 
1970 	/*
1971 	 * Move the frees from the defer_tree back to the free
1972 	 * range tree (if it's loaded). Swap the freed_tree and the
1973 	 * defer_tree -- this is safe to do because we've just emptied out
1974 	 * the defer_tree.
1975 	 */
1976 	range_tree_vacate(*defer_tree,
1977 	    msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
1978 	range_tree_swap(freed_tree, defer_tree);
1979 
1980 	space_map_update(msp->ms_sm);
1981 
1982 	msp->ms_deferspace += defer_delta;
1983 	ASSERT3S(msp->ms_deferspace, >=, 0);
1984 	ASSERT3S(msp->ms_deferspace, <=, msp->ms_size);
1985 	if (msp->ms_deferspace != 0) {
1986 		/*
1987 		 * Keep syncing this metaslab until all deferred frees
1988 		 * are back in circulation.
1989 		 */
1990 		vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1991 	}
1992 
1993 	if (msp->ms_loaded && msp->ms_access_txg < txg) {
1994 		for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
1995 			VERIFY0(range_tree_space(
1996 			    msp->ms_alloctree[(txg + t) & TXG_MASK]));
1997 		}
1998 
1999 		if (!metaslab_debug_unload)
2000 			metaslab_unload(msp);
2001 	}
2002 
2003 	metaslab_group_sort(mg, msp, metaslab_weight(msp));
2004 	mutex_exit(&msp->ms_lock);
2005 }
2006 
2007 void
2008 metaslab_sync_reassess(metaslab_group_t *mg)
2009 {
2010 	metaslab_group_alloc_update(mg);
2011 	mg->mg_fragmentation = metaslab_group_fragmentation(mg);
2012 
2013 	/*
2014 	 * Preload the next potential metaslabs
2015 	 */
2016 	metaslab_group_preload(mg);
2017 }
2018 
2019 static uint64_t
2020 metaslab_distance(metaslab_t *msp, dva_t *dva)
2021 {
2022 	uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
2023 	uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
2024 	uint64_t start = msp->ms_id;
2025 
2026 	if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
2027 		return (1ULL << 63);
2028 
2029 	if (offset < start)
2030 		return ((start - offset) << ms_shift);
2031 	if (offset > start)
2032 		return ((offset - start) << ms_shift);
2033 	return (0);
2034 }
2035 
2036 static uint64_t
2037 metaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize,
2038     uint64_t txg, uint64_t min_distance, dva_t *dva, int d)
2039 {
2040 	spa_t *spa = mg->mg_vd->vdev_spa;
2041 	metaslab_t *msp = NULL;
2042 	uint64_t offset = -1ULL;
2043 	avl_tree_t *t = &mg->mg_metaslab_tree;
2044 	uint64_t activation_weight;
2045 	uint64_t target_distance;
2046 	int i;
2047 
2048 	activation_weight = METASLAB_WEIGHT_PRIMARY;
2049 	for (i = 0; i < d; i++) {
2050 		if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
2051 			activation_weight = METASLAB_WEIGHT_SECONDARY;
2052 			break;
2053 		}
2054 	}
2055 
2056 	for (;;) {
2057 		boolean_t was_active;
2058 
2059 		mutex_enter(&mg->mg_lock);
2060 		for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
2061 			if (msp->ms_weight < asize) {
2062 				spa_dbgmsg(spa, "%s: failed to meet weight "
2063 				    "requirement: vdev %llu, txg %llu, mg %p, "
2064 				    "msp %p, psize %llu, asize %llu, "
2065 				    "weight %llu", spa_name(spa),
2066 				    mg->mg_vd->vdev_id, txg,
2067 				    mg, msp, psize, asize, msp->ms_weight);
2068 				mutex_exit(&mg->mg_lock);
2069 				return (-1ULL);
2070 			}
2071 
2072 			/*
2073 			 * If the selected metaslab is condensing, skip it.
2074 			 */
2075 			if (msp->ms_condensing)
2076 				continue;
2077 
2078 			was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
2079 			if (activation_weight == METASLAB_WEIGHT_PRIMARY)
2080 				break;
2081 
2082 			target_distance = min_distance +
2083 			    (space_map_allocated(msp->ms_sm) != 0 ? 0 :
2084 			    min_distance >> 1);
2085 
2086 			for (i = 0; i < d; i++)
2087 				if (metaslab_distance(msp, &dva[i]) <
2088 				    target_distance)
2089 					break;
2090 			if (i == d)
2091 				break;
2092 		}
2093 		mutex_exit(&mg->mg_lock);
2094 		if (msp == NULL)
2095 			return (-1ULL);
2096 
2097 		mutex_enter(&msp->ms_lock);
2098 
2099 		/*
2100 		 * Ensure that the metaslab we have selected is still
2101 		 * capable of handling our request. It's possible that
2102 		 * another thread may have changed the weight while we
2103 		 * were blocked on the metaslab lock.
2104 		 */
2105 		if (msp->ms_weight < asize || (was_active &&
2106 		    !(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
2107 		    activation_weight == METASLAB_WEIGHT_PRIMARY)) {
2108 			mutex_exit(&msp->ms_lock);
2109 			continue;
2110 		}
2111 
2112 		if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
2113 		    activation_weight == METASLAB_WEIGHT_PRIMARY) {
2114 			metaslab_passivate(msp,
2115 			    msp->ms_weight & ~METASLAB_ACTIVE_MASK);
2116 			mutex_exit(&msp->ms_lock);
2117 			continue;
2118 		}
2119 
2120 		if (metaslab_activate(msp, activation_weight) != 0) {
2121 			mutex_exit(&msp->ms_lock);
2122 			continue;
2123 		}
2124 
2125 		/*
2126 		 * If this metaslab is currently condensing then pick again as
2127 		 * we can't manipulate this metaslab until it's committed
2128 		 * to disk.
2129 		 */
2130 		if (msp->ms_condensing) {
2131 			mutex_exit(&msp->ms_lock);
2132 			continue;
2133 		}
2134 
2135 		if ((offset = metaslab_block_alloc(msp, asize)) != -1ULL)
2136 			break;
2137 
2138 		metaslab_passivate(msp, metaslab_block_maxsize(msp));
2139 		mutex_exit(&msp->ms_lock);
2140 	}
2141 
2142 	if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
2143 		vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
2144 
2145 	range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, asize);
2146 	msp->ms_access_txg = txg + metaslab_unload_delay;
2147 
2148 	mutex_exit(&msp->ms_lock);
2149 
2150 	return (offset);
2151 }
2152 
2153 /*
2154  * Allocate a block for the specified i/o.
2155  */
2156 static int
2157 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
2158     dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
2159 {
2160 	metaslab_group_t *mg, *rotor;
2161 	vdev_t *vd;
2162 	int dshift = 3;
2163 	int all_zero;
2164 	int zio_lock = B_FALSE;
2165 	boolean_t allocatable;
2166 	uint64_t offset = -1ULL;
2167 	uint64_t asize;
2168 	uint64_t distance;
2169 
2170 	ASSERT(!DVA_IS_VALID(&dva[d]));
2171 
2172 	/*
2173 	 * For testing, make some blocks above a certain size be gang blocks.
2174 	 */
2175 	if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0)
2176 		return (SET_ERROR(ENOSPC));
2177 
2178 	/*
2179 	 * Start at the rotor and loop through all mgs until we find something.
2180 	 * Note that there's no locking on mc_rotor or mc_aliquot because
2181 	 * nothing actually breaks if we miss a few updates -- we just won't
2182 	 * allocate quite as evenly.  It all balances out over time.
2183 	 *
2184 	 * If we are doing ditto or log blocks, try to spread them across
2185 	 * consecutive vdevs.  If we're forced to reuse a vdev before we've
2186 	 * allocated all of our ditto blocks, then try and spread them out on
2187 	 * that vdev as much as possible.  If it turns out to not be possible,
2188 	 * gradually lower our standards until anything becomes acceptable.
2189 	 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
2190 	 * gives us hope of containing our fault domains to something we're
2191 	 * able to reason about.  Otherwise, any two top-level vdev failures
2192 	 * will guarantee the loss of data.  With consecutive allocation,
2193 	 * only two adjacent top-level vdev failures will result in data loss.
2194 	 *
2195 	 * If we are doing gang blocks (hintdva is non-NULL), try to keep
2196 	 * ourselves on the same vdev as our gang block header.  That
2197 	 * way, we can hope for locality in vdev_cache, plus it makes our
2198 	 * fault domains something tractable.
2199 	 */
2200 	if (hintdva) {
2201 		vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
2202 
2203 		/*
2204 		 * It's possible the vdev we're using as the hint no
2205 		 * longer exists (i.e. removed). Consult the rotor when
2206 		 * all else fails.
2207 		 */
2208 		if (vd != NULL) {
2209 			mg = vd->vdev_mg;
2210 
2211 			if (flags & METASLAB_HINTBP_AVOID &&
2212 			    mg->mg_next != NULL)
2213 				mg = mg->mg_next;
2214 		} else {
2215 			mg = mc->mc_rotor;
2216 		}
2217 	} else if (d != 0) {
2218 		vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
2219 		mg = vd->vdev_mg->mg_next;
2220 	} else {
2221 		mg = mc->mc_rotor;
2222 	}
2223 
2224 	/*
2225 	 * If the hint put us into the wrong metaslab class, or into a
2226 	 * metaslab group that has been passivated, just follow the rotor.
2227 	 */
2228 	if (mg->mg_class != mc || mg->mg_activation_count <= 0)
2229 		mg = mc->mc_rotor;
2230 
2231 	rotor = mg;
2232 top:
2233 	all_zero = B_TRUE;
2234 	do {
2235 		ASSERT(mg->mg_activation_count == 1);
2236 
2237 		vd = mg->mg_vd;
2238 
2239 		/*
2240 		 * Don't allocate from faulted devices.
2241 		 */
2242 		if (zio_lock) {
2243 			spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
2244 			allocatable = vdev_allocatable(vd);
2245 			spa_config_exit(spa, SCL_ZIO, FTAG);
2246 		} else {
2247 			allocatable = vdev_allocatable(vd);
2248 		}
2249 
2250 		/*
2251 		 * Determine if the selected metaslab group is eligible
2252 		 * for allocations. If we're ganging or have requested
2253 		 * an allocation for the smallest gang block size
2254 		 * then we don't want to avoid allocating to the this
2255 		 * metaslab group. If we're in this condition we should
2256 		 * try to allocate from any device possible so that we
2257 		 * don't inadvertently return ENOSPC and suspend the pool
2258 		 * even though space is still available.
2259 		 */
2260 		if (allocatable && CAN_FASTGANG(flags) &&
2261 		    psize > SPA_GANGBLOCKSIZE)
2262 			allocatable = metaslab_group_allocatable(mg);
2263 
2264 		if (!allocatable)
2265 			goto next;
2266 
2267 		/*
2268 		 * Avoid writing single-copy data to a failing vdev
2269 		 * unless the user instructs us that it is okay.
2270 		 */
2271 		if ((vd->vdev_stat.vs_write_errors > 0 ||
2272 		    vd->vdev_state < VDEV_STATE_HEALTHY) &&
2273 		    d == 0 && dshift == 3 && vd->vdev_children == 0) {
2274 			all_zero = B_FALSE;
2275 			goto next;
2276 		}
2277 
2278 		ASSERT(mg->mg_class == mc);
2279 
2280 		distance = vd->vdev_asize >> dshift;
2281 		if (distance <= (1ULL << vd->vdev_ms_shift))
2282 			distance = 0;
2283 		else
2284 			all_zero = B_FALSE;
2285 
2286 		asize = vdev_psize_to_asize(vd, psize);
2287 		ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
2288 
2289 		offset = metaslab_group_alloc(mg, psize, asize, txg, distance,
2290 		    dva, d);
2291 		if (offset != -1ULL) {
2292 			/*
2293 			 * If we've just selected this metaslab group,
2294 			 * figure out whether the corresponding vdev is
2295 			 * over- or under-used relative to the pool,
2296 			 * and set an allocation bias to even it out.
2297 			 */
2298 			if (mc->mc_aliquot == 0 && metaslab_bias_enabled) {
2299 				vdev_stat_t *vs = &vd->vdev_stat;
2300 				int64_t vu, cu;
2301 
2302 				vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
2303 				cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
2304 
2305 				/*
2306 				 * Calculate how much more or less we should
2307 				 * try to allocate from this device during
2308 				 * this iteration around the rotor.
2309 				 * For example, if a device is 80% full
2310 				 * and the pool is 20% full then we should
2311 				 * reduce allocations by 60% on this device.
2312 				 *
2313 				 * mg_bias = (20 - 80) * 512K / 100 = -307K
2314 				 *
2315 				 * This reduces allocations by 307K for this
2316 				 * iteration.
2317 				 */
2318 				mg->mg_bias = ((cu - vu) *
2319 				    (int64_t)mg->mg_aliquot) / 100;
2320 			} else if (!metaslab_bias_enabled) {
2321 				mg->mg_bias = 0;
2322 			}
2323 
2324 			if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
2325 			    mg->mg_aliquot + mg->mg_bias) {
2326 				mc->mc_rotor = mg->mg_next;
2327 				mc->mc_aliquot = 0;
2328 			}
2329 
2330 			DVA_SET_VDEV(&dva[d], vd->vdev_id);
2331 			DVA_SET_OFFSET(&dva[d], offset);
2332 			DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
2333 			DVA_SET_ASIZE(&dva[d], asize);
2334 
2335 			return (0);
2336 		}
2337 next:
2338 		mc->mc_rotor = mg->mg_next;
2339 		mc->mc_aliquot = 0;
2340 	} while ((mg = mg->mg_next) != rotor);
2341 
2342 	if (!all_zero) {
2343 		dshift++;
2344 		ASSERT(dshift < 64);
2345 		goto top;
2346 	}
2347 
2348 	if (!allocatable && !zio_lock) {
2349 		dshift = 3;
2350 		zio_lock = B_TRUE;
2351 		goto top;
2352 	}
2353 
2354 	bzero(&dva[d], sizeof (dva_t));
2355 
2356 	return (SET_ERROR(ENOSPC));
2357 }
2358 
2359 /*
2360  * Free the block represented by DVA in the context of the specified
2361  * transaction group.
2362  */
2363 static void
2364 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
2365 {
2366 	uint64_t vdev = DVA_GET_VDEV(dva);
2367 	uint64_t offset = DVA_GET_OFFSET(dva);
2368 	uint64_t size = DVA_GET_ASIZE(dva);
2369 	vdev_t *vd;
2370 	metaslab_t *msp;
2371 
2372 	ASSERT(DVA_IS_VALID(dva));
2373 
2374 	if (txg > spa_freeze_txg(spa))
2375 		return;
2376 
2377 	if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
2378 	    (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
2379 		cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
2380 		    (u_longlong_t)vdev, (u_longlong_t)offset);
2381 		ASSERT(0);
2382 		return;
2383 	}
2384 
2385 	msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2386 
2387 	if (DVA_GET_GANG(dva))
2388 		size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
2389 
2390 	mutex_enter(&msp->ms_lock);
2391 
2392 	if (now) {
2393 		range_tree_remove(msp->ms_alloctree[txg & TXG_MASK],
2394 		    offset, size);
2395 
2396 		VERIFY(!msp->ms_condensing);
2397 		VERIFY3U(offset, >=, msp->ms_start);
2398 		VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size);
2399 		VERIFY3U(range_tree_space(msp->ms_tree) + size, <=,
2400 		    msp->ms_size);
2401 		VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
2402 		VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2403 		range_tree_add(msp->ms_tree, offset, size);
2404 	} else {
2405 		if (range_tree_space(msp->ms_freetree[txg & TXG_MASK]) == 0)
2406 			vdev_dirty(vd, VDD_METASLAB, msp, txg);
2407 		range_tree_add(msp->ms_freetree[txg & TXG_MASK],
2408 		    offset, size);
2409 	}
2410 
2411 	mutex_exit(&msp->ms_lock);
2412 }
2413 
2414 /*
2415  * Intent log support: upon opening the pool after a crash, notify the SPA
2416  * of blocks that the intent log has allocated for immediate write, but
2417  * which are still considered free by the SPA because the last transaction
2418  * group didn't commit yet.
2419  */
2420 static int
2421 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
2422 {
2423 	uint64_t vdev = DVA_GET_VDEV(dva);
2424 	uint64_t offset = DVA_GET_OFFSET(dva);
2425 	uint64_t size = DVA_GET_ASIZE(dva);
2426 	vdev_t *vd;
2427 	metaslab_t *msp;
2428 	int error = 0;
2429 
2430 	ASSERT(DVA_IS_VALID(dva));
2431 
2432 	if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
2433 	    (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
2434 		return (SET_ERROR(ENXIO));
2435 
2436 	msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2437 
2438 	if (DVA_GET_GANG(dva))
2439 		size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
2440 
2441 	mutex_enter(&msp->ms_lock);
2442 
2443 	if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded)
2444 		error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
2445 
2446 	if (error == 0 && !range_tree_contains(msp->ms_tree, offset, size))
2447 		error = SET_ERROR(ENOENT);
2448 
2449 	if (error || txg == 0) {	/* txg == 0 indicates dry run */
2450 		mutex_exit(&msp->ms_lock);
2451 		return (error);
2452 	}
2453 
2454 	VERIFY(!msp->ms_condensing);
2455 	VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
2456 	VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2457 	VERIFY3U(range_tree_space(msp->ms_tree) - size, <=, msp->ms_size);
2458 	range_tree_remove(msp->ms_tree, offset, size);
2459 
2460 	if (spa_writeable(spa)) {	/* don't dirty if we're zdb(1M) */
2461 		if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
2462 			vdev_dirty(vd, VDD_METASLAB, msp, txg);
2463 		range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, size);
2464 	}
2465 
2466 	mutex_exit(&msp->ms_lock);
2467 
2468 	return (0);
2469 }
2470 
2471 int
2472 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
2473     int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
2474 {
2475 	dva_t *dva = bp->blk_dva;
2476 	dva_t *hintdva = hintbp->blk_dva;
2477 	int error = 0;
2478 
2479 	ASSERT(bp->blk_birth == 0);
2480 	ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
2481 
2482 	spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
2483 
2484 	if (mc->mc_rotor == NULL) {	/* no vdevs in this class */
2485 		spa_config_exit(spa, SCL_ALLOC, FTAG);
2486 		return (SET_ERROR(ENOSPC));
2487 	}
2488 
2489 	ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
2490 	ASSERT(BP_GET_NDVAS(bp) == 0);
2491 	ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
2492 
2493 	for (int d = 0; d < ndvas; d++) {
2494 		error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
2495 		    txg, flags);
2496 		if (error != 0) {
2497 			for (d--; d >= 0; d--) {
2498 				metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
2499 				bzero(&dva[d], sizeof (dva_t));
2500 			}
2501 			spa_config_exit(spa, SCL_ALLOC, FTAG);
2502 			return (error);
2503 		}
2504 	}
2505 	ASSERT(error == 0);
2506 	ASSERT(BP_GET_NDVAS(bp) == ndvas);
2507 
2508 	spa_config_exit(spa, SCL_ALLOC, FTAG);
2509 
2510 	BP_SET_BIRTH(bp, txg, txg);
2511 
2512 	return (0);
2513 }
2514 
2515 void
2516 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
2517 {
2518 	const dva_t *dva = bp->blk_dva;
2519 	int ndvas = BP_GET_NDVAS(bp);
2520 
2521 	ASSERT(!BP_IS_HOLE(bp));
2522 	ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
2523 
2524 	spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
2525 
2526 	for (int d = 0; d < ndvas; d++)
2527 		metaslab_free_dva(spa, &dva[d], txg, now);
2528 
2529 	spa_config_exit(spa, SCL_FREE, FTAG);
2530 }
2531 
2532 int
2533 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
2534 {
2535 	const dva_t *dva = bp->blk_dva;
2536 	int ndvas = BP_GET_NDVAS(bp);
2537 	int error = 0;
2538 
2539 	ASSERT(!BP_IS_HOLE(bp));
2540 
2541 	if (txg != 0) {
2542 		/*
2543 		 * First do a dry run to make sure all DVAs are claimable,
2544 		 * so we don't have to unwind from partial failures below.
2545 		 */
2546 		if ((error = metaslab_claim(spa, bp, 0)) != 0)
2547 			return (error);
2548 	}
2549 
2550 	spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
2551 
2552 	for (int d = 0; d < ndvas; d++)
2553 		if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
2554 			break;
2555 
2556 	spa_config_exit(spa, SCL_ALLOC, FTAG);
2557 
2558 	ASSERT(error == 0 || txg == 0);
2559 
2560 	return (error);
2561 }
2562 
2563 void
2564 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
2565 {
2566 	if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
2567 		return;
2568 
2569 	spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
2570 	for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
2571 		uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
2572 		vdev_t *vd = vdev_lookup_top(spa, vdev);
2573 		uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
2574 		uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
2575 		metaslab_t *msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2576 
2577 		if (msp->ms_loaded)
2578 			range_tree_verify(msp->ms_tree, offset, size);
2579 
2580 		for (int j = 0; j < TXG_SIZE; j++)
2581 			range_tree_verify(msp->ms_freetree[j], offset, size);
2582 		for (int j = 0; j < TXG_DEFER_SIZE; j++)
2583 			range_tree_verify(msp->ms_defertree[j], offset, size);
2584 	}
2585 	spa_config_exit(spa, SCL_VDEV, FTAG);
2586 }
2587