xref: /titanic_50/usr/src/uts/common/fs/zfs/metaslab.c (revision 9584cebb1c69707f4c67306b661c2ed47d8676f1)
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) 2013 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 
36 /*
37  * Allow allocations to switch to gang blocks quickly. We do this to
38  * avoid having to load lots of space_maps in a given txg. There are,
39  * however, some cases where we want to avoid "fast" ganging and instead
40  * we want to do an exhaustive search of all metaslabs on this device.
41  * Currently we don't allow any gang, zil, or dump device related allocations
42  * to "fast" gang.
43  */
44 #define	CAN_FASTGANG(flags) \
45 	(!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \
46 	METASLAB_GANG_AVOID)))
47 
48 #define	METASLAB_WEIGHT_PRIMARY		(1ULL << 63)
49 #define	METASLAB_WEIGHT_SECONDARY	(1ULL << 62)
50 #define	METASLAB_ACTIVE_MASK		\
51 	(METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
52 
53 uint64_t metaslab_aliquot = 512ULL << 10;
54 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1;	/* force gang blocks */
55 
56 /*
57  * The in-core space map representation is more compact than its on-disk form.
58  * The zfs_condense_pct determines how much more compact the in-core
59  * space_map representation must be before we compact it on-disk.
60  * Values should be greater than or equal to 100.
61  */
62 int zfs_condense_pct = 200;
63 
64 /*
65  * This value defines the number of allowed allocation failures per vdev.
66  * If a device reaches this threshold in a given txg then we consider skipping
67  * allocations on that device. The value of zfs_mg_alloc_failures is computed
68  * in zio_init() unless it has been overridden in /etc/system.
69  */
70 int zfs_mg_alloc_failures = 0;
71 
72 /*
73  * The zfs_mg_noalloc_threshold defines which metaslab groups should
74  * be eligible for allocation. The value is defined as a percentage of
75  * a free space. Metaslab groups that have more free space than
76  * zfs_mg_noalloc_threshold are always eligible for allocations. Once
77  * a metaslab group's free space is less than or equal to the
78  * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
79  * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
80  * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
81  * groups are allowed to accept allocations. Gang blocks are always
82  * eligible to allocate on any metaslab group. The default value of 0 means
83  * no metaslab group will be excluded based on this criterion.
84  */
85 int zfs_mg_noalloc_threshold = 0;
86 
87 /*
88  * When set will load all metaslabs when pool is first opened.
89  */
90 int metaslab_debug_load = 0;
91 
92 /*
93  * When set will prevent metaslabs from being unloaded.
94  */
95 int metaslab_debug_unload = 0;
96 
97 /*
98  * Minimum size which forces the dynamic allocator to change
99  * it's allocation strategy.  Once the space map cannot satisfy
100  * an allocation of this size then it switches to using more
101  * aggressive strategy (i.e search by size rather than offset).
102  */
103 uint64_t metaslab_df_alloc_threshold = SPA_MAXBLOCKSIZE;
104 
105 /*
106  * The minimum free space, in percent, which must be available
107  * in a space map to continue allocations in a first-fit fashion.
108  * Once the space_map's free space drops below this level we dynamically
109  * switch to using best-fit allocations.
110  */
111 int metaslab_df_free_pct = 4;
112 
113 /*
114  * A metaslab is considered "free" if it contains a contiguous
115  * segment which is greater than metaslab_min_alloc_size.
116  */
117 uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
118 
119 /*
120  * Percentage of all cpus that can be used by the metaslab taskq.
121  */
122 int metaslab_load_pct = 50;
123 
124 /*
125  * Determines how many txgs a metaslab may remain loaded without having any
126  * allocations from it. As long as a metaslab continues to be used we will
127  * keep it loaded.
128  */
129 int metaslab_unload_delay = TXG_SIZE * 2;
130 
131 /*
132  * Should we be willing to write data to degraded vdevs?
133  */
134 boolean_t zfs_write_to_degraded = B_FALSE;
135 
136 /*
137  * Max number of metaslabs per group to preload.
138  */
139 int metaslab_preload_limit = SPA_DVAS_PER_BP;
140 
141 /*
142  * Enable/disable preloading of metaslab.
143  */
144 boolean_t metaslab_preload_enabled = B_TRUE;
145 
146 /*
147  * Enable/disable additional weight factor for each metaslab.
148  */
149 boolean_t metaslab_weight_factor_enable = B_FALSE;
150 
151 
152 /*
153  * ==========================================================================
154  * Metaslab classes
155  * ==========================================================================
156  */
157 metaslab_class_t *
158 metaslab_class_create(spa_t *spa, metaslab_ops_t *ops)
159 {
160 	metaslab_class_t *mc;
161 
162 	mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
163 
164 	mc->mc_spa = spa;
165 	mc->mc_rotor = NULL;
166 	mc->mc_ops = ops;
167 
168 	return (mc);
169 }
170 
171 void
172 metaslab_class_destroy(metaslab_class_t *mc)
173 {
174 	ASSERT(mc->mc_rotor == NULL);
175 	ASSERT(mc->mc_alloc == 0);
176 	ASSERT(mc->mc_deferred == 0);
177 	ASSERT(mc->mc_space == 0);
178 	ASSERT(mc->mc_dspace == 0);
179 
180 	kmem_free(mc, sizeof (metaslab_class_t));
181 }
182 
183 int
184 metaslab_class_validate(metaslab_class_t *mc)
185 {
186 	metaslab_group_t *mg;
187 	vdev_t *vd;
188 
189 	/*
190 	 * Must hold one of the spa_config locks.
191 	 */
192 	ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
193 	    spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
194 
195 	if ((mg = mc->mc_rotor) == NULL)
196 		return (0);
197 
198 	do {
199 		vd = mg->mg_vd;
200 		ASSERT(vd->vdev_mg != NULL);
201 		ASSERT3P(vd->vdev_top, ==, vd);
202 		ASSERT3P(mg->mg_class, ==, mc);
203 		ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
204 	} while ((mg = mg->mg_next) != mc->mc_rotor);
205 
206 	return (0);
207 }
208 
209 void
210 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
211     int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
212 {
213 	atomic_add_64(&mc->mc_alloc, alloc_delta);
214 	atomic_add_64(&mc->mc_deferred, defer_delta);
215 	atomic_add_64(&mc->mc_space, space_delta);
216 	atomic_add_64(&mc->mc_dspace, dspace_delta);
217 }
218 
219 uint64_t
220 metaslab_class_get_alloc(metaslab_class_t *mc)
221 {
222 	return (mc->mc_alloc);
223 }
224 
225 uint64_t
226 metaslab_class_get_deferred(metaslab_class_t *mc)
227 {
228 	return (mc->mc_deferred);
229 }
230 
231 uint64_t
232 metaslab_class_get_space(metaslab_class_t *mc)
233 {
234 	return (mc->mc_space);
235 }
236 
237 uint64_t
238 metaslab_class_get_dspace(metaslab_class_t *mc)
239 {
240 	return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
241 }
242 
243 /*
244  * ==========================================================================
245  * Metaslab groups
246  * ==========================================================================
247  */
248 static int
249 metaslab_compare(const void *x1, const void *x2)
250 {
251 	const metaslab_t *m1 = x1;
252 	const metaslab_t *m2 = x2;
253 
254 	if (m1->ms_weight < m2->ms_weight)
255 		return (1);
256 	if (m1->ms_weight > m2->ms_weight)
257 		return (-1);
258 
259 	/*
260 	 * If the weights are identical, use the offset to force uniqueness.
261 	 */
262 	if (m1->ms_start < m2->ms_start)
263 		return (-1);
264 	if (m1->ms_start > m2->ms_start)
265 		return (1);
266 
267 	ASSERT3P(m1, ==, m2);
268 
269 	return (0);
270 }
271 
272 /*
273  * Update the allocatable flag and the metaslab group's capacity.
274  * The allocatable flag is set to true if the capacity is below
275  * the zfs_mg_noalloc_threshold. If a metaslab group transitions
276  * from allocatable to non-allocatable or vice versa then the metaslab
277  * group's class is updated to reflect the transition.
278  */
279 static void
280 metaslab_group_alloc_update(metaslab_group_t *mg)
281 {
282 	vdev_t *vd = mg->mg_vd;
283 	metaslab_class_t *mc = mg->mg_class;
284 	vdev_stat_t *vs = &vd->vdev_stat;
285 	boolean_t was_allocatable;
286 
287 	ASSERT(vd == vd->vdev_top);
288 
289 	mutex_enter(&mg->mg_lock);
290 	was_allocatable = mg->mg_allocatable;
291 
292 	mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
293 	    (vs->vs_space + 1);
294 
295 	mg->mg_allocatable = (mg->mg_free_capacity > zfs_mg_noalloc_threshold);
296 
297 	/*
298 	 * The mc_alloc_groups maintains a count of the number of
299 	 * groups in this metaslab class that are still above the
300 	 * zfs_mg_noalloc_threshold. This is used by the allocating
301 	 * threads to determine if they should avoid allocations to
302 	 * a given group. The allocator will avoid allocations to a group
303 	 * if that group has reached or is below the zfs_mg_noalloc_threshold
304 	 * and there are still other groups that are above the threshold.
305 	 * When a group transitions from allocatable to non-allocatable or
306 	 * vice versa we update the metaslab class to reflect that change.
307 	 * When the mc_alloc_groups value drops to 0 that means that all
308 	 * groups have reached the zfs_mg_noalloc_threshold making all groups
309 	 * eligible for allocations. This effectively means that all devices
310 	 * are balanced again.
311 	 */
312 	if (was_allocatable && !mg->mg_allocatable)
313 		mc->mc_alloc_groups--;
314 	else if (!was_allocatable && mg->mg_allocatable)
315 		mc->mc_alloc_groups++;
316 	mutex_exit(&mg->mg_lock);
317 }
318 
319 metaslab_group_t *
320 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
321 {
322 	metaslab_group_t *mg;
323 
324 	mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
325 	mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
326 	avl_create(&mg->mg_metaslab_tree, metaslab_compare,
327 	    sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
328 	mg->mg_vd = vd;
329 	mg->mg_class = mc;
330 	mg->mg_activation_count = 0;
331 
332 	mg->mg_taskq = taskq_create("metaslab_group_tasksq", metaslab_load_pct,
333 	    minclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT);
334 
335 	return (mg);
336 }
337 
338 void
339 metaslab_group_destroy(metaslab_group_t *mg)
340 {
341 	ASSERT(mg->mg_prev == NULL);
342 	ASSERT(mg->mg_next == NULL);
343 	/*
344 	 * We may have gone below zero with the activation count
345 	 * either because we never activated in the first place or
346 	 * because we're done, and possibly removing the vdev.
347 	 */
348 	ASSERT(mg->mg_activation_count <= 0);
349 
350 	avl_destroy(&mg->mg_metaslab_tree);
351 	mutex_destroy(&mg->mg_lock);
352 	kmem_free(mg, sizeof (metaslab_group_t));
353 }
354 
355 void
356 metaslab_group_activate(metaslab_group_t *mg)
357 {
358 	metaslab_class_t *mc = mg->mg_class;
359 	metaslab_group_t *mgprev, *mgnext;
360 
361 	ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
362 
363 	ASSERT(mc->mc_rotor != mg);
364 	ASSERT(mg->mg_prev == NULL);
365 	ASSERT(mg->mg_next == NULL);
366 	ASSERT(mg->mg_activation_count <= 0);
367 
368 	if (++mg->mg_activation_count <= 0)
369 		return;
370 
371 	mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
372 	metaslab_group_alloc_update(mg);
373 
374 	if ((mgprev = mc->mc_rotor) == NULL) {
375 		mg->mg_prev = mg;
376 		mg->mg_next = mg;
377 	} else {
378 		mgnext = mgprev->mg_next;
379 		mg->mg_prev = mgprev;
380 		mg->mg_next = mgnext;
381 		mgprev->mg_next = mg;
382 		mgnext->mg_prev = mg;
383 	}
384 	mc->mc_rotor = mg;
385 }
386 
387 void
388 metaslab_group_passivate(metaslab_group_t *mg)
389 {
390 	metaslab_class_t *mc = mg->mg_class;
391 	metaslab_group_t *mgprev, *mgnext;
392 
393 	ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
394 
395 	if (--mg->mg_activation_count != 0) {
396 		ASSERT(mc->mc_rotor != mg);
397 		ASSERT(mg->mg_prev == NULL);
398 		ASSERT(mg->mg_next == NULL);
399 		ASSERT(mg->mg_activation_count < 0);
400 		return;
401 	}
402 
403 	taskq_wait(mg->mg_taskq);
404 
405 	mgprev = mg->mg_prev;
406 	mgnext = mg->mg_next;
407 
408 	if (mg == mgnext) {
409 		mc->mc_rotor = NULL;
410 	} else {
411 		mc->mc_rotor = mgnext;
412 		mgprev->mg_next = mgnext;
413 		mgnext->mg_prev = mgprev;
414 	}
415 
416 	mg->mg_prev = NULL;
417 	mg->mg_next = NULL;
418 }
419 
420 static void
421 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
422 {
423 	mutex_enter(&mg->mg_lock);
424 	ASSERT(msp->ms_group == NULL);
425 	msp->ms_group = mg;
426 	msp->ms_weight = 0;
427 	avl_add(&mg->mg_metaslab_tree, msp);
428 	mutex_exit(&mg->mg_lock);
429 }
430 
431 static void
432 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
433 {
434 	mutex_enter(&mg->mg_lock);
435 	ASSERT(msp->ms_group == mg);
436 	avl_remove(&mg->mg_metaslab_tree, msp);
437 	msp->ms_group = NULL;
438 	mutex_exit(&mg->mg_lock);
439 }
440 
441 static void
442 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
443 {
444 	/*
445 	 * Although in principle the weight can be any value, in
446 	 * practice we do not use values in the range [1, 510].
447 	 */
448 	ASSERT(weight >= SPA_MINBLOCKSIZE-1 || weight == 0);
449 	ASSERT(MUTEX_HELD(&msp->ms_lock));
450 
451 	mutex_enter(&mg->mg_lock);
452 	ASSERT(msp->ms_group == mg);
453 	avl_remove(&mg->mg_metaslab_tree, msp);
454 	msp->ms_weight = weight;
455 	avl_add(&mg->mg_metaslab_tree, msp);
456 	mutex_exit(&mg->mg_lock);
457 }
458 
459 /*
460  * Determine if a given metaslab group should skip allocations. A metaslab
461  * group should avoid allocations if its used capacity has crossed the
462  * zfs_mg_noalloc_threshold and there is at least one metaslab group
463  * that can still handle allocations.
464  */
465 static boolean_t
466 metaslab_group_allocatable(metaslab_group_t *mg)
467 {
468 	vdev_t *vd = mg->mg_vd;
469 	spa_t *spa = vd->vdev_spa;
470 	metaslab_class_t *mc = mg->mg_class;
471 
472 	/*
473 	 * A metaslab group is considered allocatable if its free capacity
474 	 * is greater than the set value of zfs_mg_noalloc_threshold, it's
475 	 * associated with a slog, or there are no other metaslab groups
476 	 * with free capacity greater than zfs_mg_noalloc_threshold.
477 	 */
478 	return (mg->mg_free_capacity > zfs_mg_noalloc_threshold ||
479 	    mc != spa_normal_class(spa) || mc->mc_alloc_groups == 0);
480 }
481 
482 /*
483  * ==========================================================================
484  * Range tree callbacks
485  * ==========================================================================
486  */
487 
488 /*
489  * Comparison function for the private size-ordered tree. Tree is sorted
490  * by size, larger sizes at the end of the tree.
491  */
492 static int
493 metaslab_rangesize_compare(const void *x1, const void *x2)
494 {
495 	const range_seg_t *r1 = x1;
496 	const range_seg_t *r2 = x2;
497 	uint64_t rs_size1 = r1->rs_end - r1->rs_start;
498 	uint64_t rs_size2 = r2->rs_end - r2->rs_start;
499 
500 	if (rs_size1 < rs_size2)
501 		return (-1);
502 	if (rs_size1 > rs_size2)
503 		return (1);
504 
505 	if (r1->rs_start < r2->rs_start)
506 		return (-1);
507 
508 	if (r1->rs_start > r2->rs_start)
509 		return (1);
510 
511 	return (0);
512 }
513 
514 /*
515  * Create any block allocator specific components. The current allocators
516  * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
517  */
518 static void
519 metaslab_rt_create(range_tree_t *rt, void *arg)
520 {
521 	metaslab_t *msp = arg;
522 
523 	ASSERT3P(rt->rt_arg, ==, msp);
524 	ASSERT(msp->ms_tree == NULL);
525 
526 	avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
527 	    sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
528 }
529 
530 /*
531  * Destroy the block allocator specific components.
532  */
533 static void
534 metaslab_rt_destroy(range_tree_t *rt, void *arg)
535 {
536 	metaslab_t *msp = arg;
537 
538 	ASSERT3P(rt->rt_arg, ==, msp);
539 	ASSERT3P(msp->ms_tree, ==, rt);
540 	ASSERT0(avl_numnodes(&msp->ms_size_tree));
541 
542 	avl_destroy(&msp->ms_size_tree);
543 }
544 
545 static void
546 metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg)
547 {
548 	metaslab_t *msp = arg;
549 
550 	ASSERT3P(rt->rt_arg, ==, msp);
551 	ASSERT3P(msp->ms_tree, ==, rt);
552 	VERIFY(!msp->ms_condensing);
553 	avl_add(&msp->ms_size_tree, rs);
554 }
555 
556 static void
557 metaslab_rt_remove(range_tree_t *rt, range_seg_t *rs, void *arg)
558 {
559 	metaslab_t *msp = arg;
560 
561 	ASSERT3P(rt->rt_arg, ==, msp);
562 	ASSERT3P(msp->ms_tree, ==, rt);
563 	VERIFY(!msp->ms_condensing);
564 	avl_remove(&msp->ms_size_tree, rs);
565 }
566 
567 static void
568 metaslab_rt_vacate(range_tree_t *rt, void *arg)
569 {
570 	metaslab_t *msp = arg;
571 
572 	ASSERT3P(rt->rt_arg, ==, msp);
573 	ASSERT3P(msp->ms_tree, ==, rt);
574 
575 	/*
576 	 * Normally one would walk the tree freeing nodes along the way.
577 	 * Since the nodes are shared with the range trees we can avoid
578 	 * walking all nodes and just reinitialize the avl tree. The nodes
579 	 * will be freed by the range tree, so we don't want to free them here.
580 	 */
581 	avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
582 	    sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
583 }
584 
585 static range_tree_ops_t metaslab_rt_ops = {
586 	metaslab_rt_create,
587 	metaslab_rt_destroy,
588 	metaslab_rt_add,
589 	metaslab_rt_remove,
590 	metaslab_rt_vacate
591 };
592 
593 /*
594  * ==========================================================================
595  * Metaslab block operations
596  * ==========================================================================
597  */
598 
599 /*
600  * Return the maximum contiguous segment within the metaslab.
601  */
602 uint64_t
603 metaslab_block_maxsize(metaslab_t *msp)
604 {
605 	avl_tree_t *t = &msp->ms_size_tree;
606 	range_seg_t *rs;
607 
608 	if (t == NULL || (rs = avl_last(t)) == NULL)
609 		return (0ULL);
610 
611 	return (rs->rs_end - rs->rs_start);
612 }
613 
614 uint64_t
615 metaslab_block_alloc(metaslab_t *msp, uint64_t size)
616 {
617 	uint64_t start;
618 	range_tree_t *rt = msp->ms_tree;
619 
620 	VERIFY(!msp->ms_condensing);
621 
622 	start = msp->ms_ops->msop_alloc(msp, size);
623 	if (start != -1ULL) {
624 		vdev_t *vd = msp->ms_group->mg_vd;
625 
626 		VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
627 		VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
628 		VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
629 		range_tree_remove(rt, start, size);
630 	}
631 	return (start);
632 }
633 
634 /*
635  * ==========================================================================
636  * Common allocator routines
637  * ==========================================================================
638  */
639 
640 /*
641  * This is a helper function that can be used by the allocator to find
642  * a suitable block to allocate. This will search the specified AVL
643  * tree looking for a block that matches the specified criteria.
644  */
645 static uint64_t
646 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
647     uint64_t align)
648 {
649 	range_seg_t *rs, rsearch;
650 	avl_index_t where;
651 
652 	rsearch.rs_start = *cursor;
653 	rsearch.rs_end = *cursor + size;
654 
655 	rs = avl_find(t, &rsearch, &where);
656 	if (rs == NULL)
657 		rs = avl_nearest(t, where, AVL_AFTER);
658 
659 	while (rs != NULL) {
660 		uint64_t offset = P2ROUNDUP(rs->rs_start, align);
661 
662 		if (offset + size <= rs->rs_end) {
663 			*cursor = offset + size;
664 			return (offset);
665 		}
666 		rs = AVL_NEXT(t, rs);
667 	}
668 
669 	/*
670 	 * If we know we've searched the whole map (*cursor == 0), give up.
671 	 * Otherwise, reset the cursor to the beginning and try again.
672 	 */
673 	if (*cursor == 0)
674 		return (-1ULL);
675 
676 	*cursor = 0;
677 	return (metaslab_block_picker(t, cursor, size, align));
678 }
679 
680 /*
681  * ==========================================================================
682  * The first-fit block allocator
683  * ==========================================================================
684  */
685 static uint64_t
686 metaslab_ff_alloc(metaslab_t *msp, uint64_t size)
687 {
688 	/*
689 	 * Find the largest power of 2 block size that evenly divides the
690 	 * requested size. This is used to try to allocate blocks with similar
691 	 * alignment from the same area of the metaslab (i.e. same cursor
692 	 * bucket) but it does not guarantee that other allocations sizes
693 	 * may exist in the same region.
694 	 */
695 	uint64_t align = size & -size;
696 	uint64_t *cursor = &msp->ms_lbas[highbit(align) - 1];
697 	avl_tree_t *t = &msp->ms_tree->rt_root;
698 
699 	return (metaslab_block_picker(t, cursor, size, align));
700 }
701 
702 /* ARGSUSED */
703 static boolean_t
704 metaslab_ff_fragmented(metaslab_t *msp)
705 {
706 	return (B_TRUE);
707 }
708 
709 static metaslab_ops_t metaslab_ff_ops = {
710 	metaslab_ff_alloc,
711 	metaslab_ff_fragmented
712 };
713 
714 /*
715  * ==========================================================================
716  * Dynamic block allocator -
717  * Uses the first fit allocation scheme until space get low and then
718  * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
719  * and metaslab_df_free_pct to determine when to switch the allocation scheme.
720  * ==========================================================================
721  */
722 static uint64_t
723 metaslab_df_alloc(metaslab_t *msp, uint64_t size)
724 {
725 	/*
726 	 * Find the largest power of 2 block size that evenly divides the
727 	 * requested size. This is used to try to allocate blocks with similar
728 	 * alignment from the same area of the metaslab (i.e. same cursor
729 	 * bucket) but it does not guarantee that other allocations sizes
730 	 * may exist in the same region.
731 	 */
732 	uint64_t align = size & -size;
733 	uint64_t *cursor = &msp->ms_lbas[highbit(align) - 1];
734 	range_tree_t *rt = msp->ms_tree;
735 	avl_tree_t *t = &rt->rt_root;
736 	uint64_t max_size = metaslab_block_maxsize(msp);
737 	int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
738 
739 	ASSERT(MUTEX_HELD(&msp->ms_lock));
740 	ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
741 
742 	if (max_size < size)
743 		return (-1ULL);
744 
745 	/*
746 	 * If we're running low on space switch to using the size
747 	 * sorted AVL tree (best-fit).
748 	 */
749 	if (max_size < metaslab_df_alloc_threshold ||
750 	    free_pct < metaslab_df_free_pct) {
751 		t = &msp->ms_size_tree;
752 		*cursor = 0;
753 	}
754 
755 	return (metaslab_block_picker(t, cursor, size, 1ULL));
756 }
757 
758 static boolean_t
759 metaslab_df_fragmented(metaslab_t *msp)
760 {
761 	range_tree_t *rt = msp->ms_tree;
762 	uint64_t max_size = metaslab_block_maxsize(msp);
763 	int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
764 
765 	if (max_size >= metaslab_df_alloc_threshold &&
766 	    free_pct >= metaslab_df_free_pct)
767 		return (B_FALSE);
768 
769 	return (B_TRUE);
770 }
771 
772 static metaslab_ops_t metaslab_df_ops = {
773 	metaslab_df_alloc,
774 	metaslab_df_fragmented
775 };
776 
777 /*
778  * ==========================================================================
779  * Cursor fit block allocator -
780  * Select the largest region in the metaslab, set the cursor to the beginning
781  * of the range and the cursor_end to the end of the range. As allocations
782  * are made advance the cursor. Continue allocating from the cursor until
783  * the range is exhausted and then find a new range.
784  * ==========================================================================
785  */
786 static uint64_t
787 metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
788 {
789 	range_tree_t *rt = msp->ms_tree;
790 	avl_tree_t *t = &msp->ms_size_tree;
791 	uint64_t *cursor = &msp->ms_lbas[0];
792 	uint64_t *cursor_end = &msp->ms_lbas[1];
793 	uint64_t offset = 0;
794 
795 	ASSERT(MUTEX_HELD(&msp->ms_lock));
796 	ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&rt->rt_root));
797 
798 	ASSERT3U(*cursor_end, >=, *cursor);
799 
800 	if ((*cursor + size) > *cursor_end) {
801 		range_seg_t *rs;
802 
803 		rs = avl_last(&msp->ms_size_tree);
804 		if (rs == NULL || (rs->rs_end - rs->rs_start) < size)
805 			return (-1ULL);
806 
807 		*cursor = rs->rs_start;
808 		*cursor_end = rs->rs_end;
809 	}
810 
811 	offset = *cursor;
812 	*cursor += size;
813 
814 	return (offset);
815 }
816 
817 static boolean_t
818 metaslab_cf_fragmented(metaslab_t *msp)
819 {
820 	return (metaslab_block_maxsize(msp) < metaslab_min_alloc_size);
821 }
822 
823 static metaslab_ops_t metaslab_cf_ops = {
824 	metaslab_cf_alloc,
825 	metaslab_cf_fragmented
826 };
827 
828 /*
829  * ==========================================================================
830  * New dynamic fit allocator -
831  * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
832  * contiguous blocks. If no region is found then just use the largest segment
833  * that remains.
834  * ==========================================================================
835  */
836 
837 /*
838  * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
839  * to request from the allocator.
840  */
841 uint64_t metaslab_ndf_clump_shift = 4;
842 
843 static uint64_t
844 metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
845 {
846 	avl_tree_t *t = &msp->ms_tree->rt_root;
847 	avl_index_t where;
848 	range_seg_t *rs, rsearch;
849 	uint64_t hbit = highbit(size);
850 	uint64_t *cursor = &msp->ms_lbas[hbit - 1];
851 	uint64_t max_size = metaslab_block_maxsize(msp);
852 
853 	ASSERT(MUTEX_HELD(&msp->ms_lock));
854 	ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
855 
856 	if (max_size < size)
857 		return (-1ULL);
858 
859 	rsearch.rs_start = *cursor;
860 	rsearch.rs_end = *cursor + size;
861 
862 	rs = avl_find(t, &rsearch, &where);
863 	if (rs == NULL || (rs->rs_end - rs->rs_start) < size) {
864 		t = &msp->ms_size_tree;
865 
866 		rsearch.rs_start = 0;
867 		rsearch.rs_end = MIN(max_size,
868 		    1ULL << (hbit + metaslab_ndf_clump_shift));
869 		rs = avl_find(t, &rsearch, &where);
870 		if (rs == NULL)
871 			rs = avl_nearest(t, where, AVL_AFTER);
872 		ASSERT(rs != NULL);
873 	}
874 
875 	if ((rs->rs_end - rs->rs_start) >= size) {
876 		*cursor = rs->rs_start + size;
877 		return (rs->rs_start);
878 	}
879 	return (-1ULL);
880 }
881 
882 static boolean_t
883 metaslab_ndf_fragmented(metaslab_t *msp)
884 {
885 	return (metaslab_block_maxsize(msp) <=
886 	    (metaslab_min_alloc_size << metaslab_ndf_clump_shift));
887 }
888 
889 static metaslab_ops_t metaslab_ndf_ops = {
890 	metaslab_ndf_alloc,
891 	metaslab_ndf_fragmented
892 };
893 
894 metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
895 
896 /*
897  * ==========================================================================
898  * Metaslabs
899  * ==========================================================================
900  */
901 
902 /*
903  * Wait for any in-progress metaslab loads to complete.
904  */
905 void
906 metaslab_load_wait(metaslab_t *msp)
907 {
908 	ASSERT(MUTEX_HELD(&msp->ms_lock));
909 
910 	while (msp->ms_loading) {
911 		ASSERT(!msp->ms_loaded);
912 		cv_wait(&msp->ms_load_cv, &msp->ms_lock);
913 	}
914 }
915 
916 int
917 metaslab_load(metaslab_t *msp)
918 {
919 	int error = 0;
920 
921 	ASSERT(MUTEX_HELD(&msp->ms_lock));
922 	ASSERT(!msp->ms_loaded);
923 	ASSERT(!msp->ms_loading);
924 
925 	msp->ms_loading = B_TRUE;
926 
927 	/*
928 	 * If the space map has not been allocated yet, then treat
929 	 * all the space in the metaslab as free and add it to the
930 	 * ms_tree.
931 	 */
932 	if (msp->ms_sm != NULL)
933 		error = space_map_load(msp->ms_sm, msp->ms_tree, SM_FREE);
934 	else
935 		range_tree_add(msp->ms_tree, msp->ms_start, msp->ms_size);
936 
937 	msp->ms_loaded = (error == 0);
938 	msp->ms_loading = B_FALSE;
939 
940 	if (msp->ms_loaded) {
941 		for (int t = 0; t < TXG_DEFER_SIZE; t++) {
942 			range_tree_walk(msp->ms_defertree[t],
943 			    range_tree_remove, msp->ms_tree);
944 		}
945 	}
946 	cv_broadcast(&msp->ms_load_cv);
947 	return (error);
948 }
949 
950 void
951 metaslab_unload(metaslab_t *msp)
952 {
953 	ASSERT(MUTEX_HELD(&msp->ms_lock));
954 	range_tree_vacate(msp->ms_tree, NULL, NULL);
955 	msp->ms_loaded = B_FALSE;
956 	msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
957 }
958 
959 metaslab_t *
960 metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, uint64_t txg)
961 {
962 	vdev_t *vd = mg->mg_vd;
963 	objset_t *mos = vd->vdev_spa->spa_meta_objset;
964 	metaslab_t *msp;
965 
966 	msp = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
967 	mutex_init(&msp->ms_lock, NULL, MUTEX_DEFAULT, NULL);
968 	cv_init(&msp->ms_load_cv, NULL, CV_DEFAULT, NULL);
969 	msp->ms_id = id;
970 	msp->ms_start = id << vd->vdev_ms_shift;
971 	msp->ms_size = 1ULL << vd->vdev_ms_shift;
972 
973 	/*
974 	 * We only open space map objects that already exist. All others
975 	 * will be opened when we finally allocate an object for it.
976 	 */
977 	if (object != 0) {
978 		VERIFY0(space_map_open(&msp->ms_sm, mos, object, msp->ms_start,
979 		    msp->ms_size, vd->vdev_ashift, &msp->ms_lock));
980 		ASSERT(msp->ms_sm != NULL);
981 	}
982 
983 	/*
984 	 * We create the main range tree here, but we don't create the
985 	 * alloctree and freetree until metaslab_sync_done().  This serves
986 	 * two purposes: it allows metaslab_sync_done() to detect the
987 	 * addition of new space; and for debugging, it ensures that we'd
988 	 * data fault on any attempt to use this metaslab before it's ready.
989 	 */
990 	msp->ms_tree = range_tree_create(&metaslab_rt_ops, msp, &msp->ms_lock);
991 	metaslab_group_add(mg, msp);
992 
993 	msp->ms_ops = mg->mg_class->mc_ops;
994 
995 	/*
996 	 * If we're opening an existing pool (txg == 0) or creating
997 	 * a new one (txg == TXG_INITIAL), all space is available now.
998 	 * If we're adding space to an existing pool, the new space
999 	 * does not become available until after this txg has synced.
1000 	 */
1001 	if (txg <= TXG_INITIAL)
1002 		metaslab_sync_done(msp, 0);
1003 
1004 	/*
1005 	 * If metaslab_debug_load is set and we're initializing a metaslab
1006 	 * that has an allocated space_map object then load the its space
1007 	 * map so that can verify frees.
1008 	 */
1009 	if (metaslab_debug_load && msp->ms_sm != NULL) {
1010 		mutex_enter(&msp->ms_lock);
1011 		VERIFY0(metaslab_load(msp));
1012 		mutex_exit(&msp->ms_lock);
1013 	}
1014 
1015 	if (txg != 0) {
1016 		vdev_dirty(vd, 0, NULL, txg);
1017 		vdev_dirty(vd, VDD_METASLAB, msp, txg);
1018 	}
1019 
1020 	return (msp);
1021 }
1022 
1023 void
1024 metaslab_fini(metaslab_t *msp)
1025 {
1026 	metaslab_group_t *mg = msp->ms_group;
1027 
1028 	metaslab_group_remove(mg, msp);
1029 
1030 	mutex_enter(&msp->ms_lock);
1031 
1032 	VERIFY(msp->ms_group == NULL);
1033 	vdev_space_update(mg->mg_vd, -space_map_allocated(msp->ms_sm),
1034 	    0, -msp->ms_size);
1035 	space_map_close(msp->ms_sm);
1036 
1037 	metaslab_unload(msp);
1038 	range_tree_destroy(msp->ms_tree);
1039 
1040 	for (int t = 0; t < TXG_SIZE; t++) {
1041 		range_tree_destroy(msp->ms_alloctree[t]);
1042 		range_tree_destroy(msp->ms_freetree[t]);
1043 	}
1044 
1045 	for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1046 		range_tree_destroy(msp->ms_defertree[t]);
1047 	}
1048 
1049 	ASSERT0(msp->ms_deferspace);
1050 
1051 	mutex_exit(&msp->ms_lock);
1052 	cv_destroy(&msp->ms_load_cv);
1053 	mutex_destroy(&msp->ms_lock);
1054 
1055 	kmem_free(msp, sizeof (metaslab_t));
1056 }
1057 
1058 /*
1059  * Apply a weighting factor based on the histogram information for this
1060  * metaslab. The current weighting factor is somewhat arbitrary and requires
1061  * additional investigation. The implementation provides a measure of
1062  * "weighted" free space and gives a higher weighting for larger contiguous
1063  * regions. The weighting factor is determined by counting the number of
1064  * sm_shift sectors that exist in each region represented by the histogram.
1065  * That value is then multiplied by the power of 2 exponent and the sm_shift
1066  * value.
1067  *
1068  * For example, assume the 2^21 histogram bucket has 4 2MB regions and the
1069  * metaslab has an sm_shift value of 9 (512B):
1070  *
1071  * 1) calculate the number of sm_shift sectors in the region:
1072  *	2^21 / 2^9 = 2^12 = 4096 * 4 (number of regions) = 16384
1073  * 2) multiply by the power of 2 exponent and the sm_shift value:
1074  *	16384 * 21 * 9 = 3096576
1075  * This value will be added to the weighting of the metaslab.
1076  */
1077 static uint64_t
1078 metaslab_weight_factor(metaslab_t *msp)
1079 {
1080 	uint64_t factor = 0;
1081 	uint64_t sectors;
1082 	int i;
1083 
1084 	/*
1085 	 * A null space map means that the entire metaslab is free,
1086 	 * calculate a weight factor that spans the entire size of the
1087 	 * metaslab.
1088 	 */
1089 	if (msp->ms_sm == NULL) {
1090 		vdev_t *vd = msp->ms_group->mg_vd;
1091 
1092 		i = highbit(msp->ms_size) - 1;
1093 		sectors = msp->ms_size >> vd->vdev_ashift;
1094 		return (sectors * i * vd->vdev_ashift);
1095 	}
1096 
1097 	if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
1098 		return (0);
1099 
1100 	for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE(msp->ms_sm); i++) {
1101 		if (msp->ms_sm->sm_phys->smp_histogram[i] == 0)
1102 			continue;
1103 
1104 		/*
1105 		 * Determine the number of sm_shift sectors in the region
1106 		 * indicated by the histogram. For example, given an
1107 		 * sm_shift value of 9 (512 bytes) and i = 4 then we know
1108 		 * that we're looking at an 8K region in the histogram
1109 		 * (i.e. 9 + 4 = 13, 2^13 = 8192). To figure out the
1110 		 * number of sm_shift sectors (512 bytes in this example),
1111 		 * we would take 8192 / 512 = 16. Since the histogram
1112 		 * is offset by sm_shift we can simply use the value of
1113 		 * of i to calculate this (i.e. 2^i = 16 where i = 4).
1114 		 */
1115 		sectors = msp->ms_sm->sm_phys->smp_histogram[i] << i;
1116 		factor += (i + msp->ms_sm->sm_shift) * sectors;
1117 	}
1118 	return (factor * msp->ms_sm->sm_shift);
1119 }
1120 
1121 static uint64_t
1122 metaslab_weight(metaslab_t *msp)
1123 {
1124 	metaslab_group_t *mg = msp->ms_group;
1125 	vdev_t *vd = mg->mg_vd;
1126 	uint64_t weight, space;
1127 
1128 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1129 
1130 	/*
1131 	 * This vdev is in the process of being removed so there is nothing
1132 	 * for us to do here.
1133 	 */
1134 	if (vd->vdev_removing) {
1135 		ASSERT0(space_map_allocated(msp->ms_sm));
1136 		ASSERT0(vd->vdev_ms_shift);
1137 		return (0);
1138 	}
1139 
1140 	/*
1141 	 * The baseline weight is the metaslab's free space.
1142 	 */
1143 	space = msp->ms_size - space_map_allocated(msp->ms_sm);
1144 	weight = space;
1145 
1146 	/*
1147 	 * Modern disks have uniform bit density and constant angular velocity.
1148 	 * Therefore, the outer recording zones are faster (higher bandwidth)
1149 	 * than the inner zones by the ratio of outer to inner track diameter,
1150 	 * which is typically around 2:1.  We account for this by assigning
1151 	 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1152 	 * In effect, this means that we'll select the metaslab with the most
1153 	 * free bandwidth rather than simply the one with the most free space.
1154 	 */
1155 	weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count;
1156 	ASSERT(weight >= space && weight <= 2 * space);
1157 
1158 	msp->ms_factor = metaslab_weight_factor(msp);
1159 	if (metaslab_weight_factor_enable)
1160 		weight += msp->ms_factor;
1161 
1162 	if (msp->ms_loaded && !msp->ms_ops->msop_fragmented(msp)) {
1163 		/*
1164 		 * If this metaslab is one we're actively using, adjust its
1165 		 * weight to make it preferable to any inactive metaslab so
1166 		 * we'll polish it off.
1167 		 */
1168 		weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
1169 	}
1170 
1171 	return (weight);
1172 }
1173 
1174 static int
1175 metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
1176 {
1177 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1178 
1179 	if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1180 		metaslab_load_wait(msp);
1181 		if (!msp->ms_loaded) {
1182 			int error = metaslab_load(msp);
1183 			if (error) {
1184 				metaslab_group_sort(msp->ms_group, msp, 0);
1185 				return (error);
1186 			}
1187 		}
1188 
1189 		metaslab_group_sort(msp->ms_group, msp,
1190 		    msp->ms_weight | activation_weight);
1191 	}
1192 	ASSERT(msp->ms_loaded);
1193 	ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
1194 
1195 	return (0);
1196 }
1197 
1198 static void
1199 metaslab_passivate(metaslab_t *msp, uint64_t size)
1200 {
1201 	/*
1202 	 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1203 	 * this metaslab again.  In that case, it had better be empty,
1204 	 * or we would be leaving space on the table.
1205 	 */
1206 	ASSERT(size >= SPA_MINBLOCKSIZE || range_tree_space(msp->ms_tree) == 0);
1207 	metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
1208 	ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
1209 }
1210 
1211 static void
1212 metaslab_preload(void *arg)
1213 {
1214 	metaslab_t *msp = arg;
1215 	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1216 
1217 	mutex_enter(&msp->ms_lock);
1218 	metaslab_load_wait(msp);
1219 	if (!msp->ms_loaded)
1220 		(void) metaslab_load(msp);
1221 
1222 	/*
1223 	 * Set the ms_access_txg value so that we don't unload it right away.
1224 	 */
1225 	msp->ms_access_txg = spa_syncing_txg(spa) + metaslab_unload_delay + 1;
1226 	mutex_exit(&msp->ms_lock);
1227 }
1228 
1229 static void
1230 metaslab_group_preload(metaslab_group_t *mg)
1231 {
1232 	spa_t *spa = mg->mg_vd->vdev_spa;
1233 	metaslab_t *msp;
1234 	avl_tree_t *t = &mg->mg_metaslab_tree;
1235 	int m = 0;
1236 
1237 	if (spa_shutting_down(spa) || !metaslab_preload_enabled) {
1238 		taskq_wait(mg->mg_taskq);
1239 		return;
1240 	}
1241 	mutex_enter(&mg->mg_lock);
1242 
1243 	/*
1244 	 * Prefetch the next potential metaslabs
1245 	 */
1246 	for (msp = avl_first(t); msp != NULL; msp = AVL_NEXT(t, msp)) {
1247 
1248 		/* If we have reached our preload limit then we're done */
1249 		if (++m > metaslab_preload_limit)
1250 			break;
1251 
1252 		VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
1253 		    msp, TQ_SLEEP) != NULL);
1254 	}
1255 	mutex_exit(&mg->mg_lock);
1256 }
1257 
1258 /*
1259  * Determine if the space map's on-disk footprint is past our tolerance
1260  * for inefficiency. We would like to use the following criteria to make
1261  * our decision:
1262  *
1263  * 1. The size of the space map object should not dramatically increase as a
1264  * result of writing out the free space range tree.
1265  *
1266  * 2. The minimal on-disk space map representation is zfs_condense_pct/100
1267  * times the size than the free space range tree representation
1268  * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
1269  *
1270  * Checking the first condition is tricky since we don't want to walk
1271  * the entire AVL tree calculating the estimated on-disk size. Instead we
1272  * use the size-ordered range tree in the metaslab and calculate the
1273  * size required to write out the largest segment in our free tree. If the
1274  * size required to represent that segment on disk is larger than the space
1275  * map object then we avoid condensing this map.
1276  *
1277  * To determine the second criterion we use a best-case estimate and assume
1278  * each segment can be represented on-disk as a single 64-bit entry. We refer
1279  * to this best-case estimate as the space map's minimal form.
1280  */
1281 static boolean_t
1282 metaslab_should_condense(metaslab_t *msp)
1283 {
1284 	space_map_t *sm = msp->ms_sm;
1285 	range_seg_t *rs;
1286 	uint64_t size, entries, segsz;
1287 
1288 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1289 	ASSERT(msp->ms_loaded);
1290 
1291 	/*
1292 	 * Use the ms_size_tree range tree, which is ordered by size, to
1293 	 * obtain the largest segment in the free tree. If the tree is empty
1294 	 * then we should condense the map.
1295 	 */
1296 	rs = avl_last(&msp->ms_size_tree);
1297 	if (rs == NULL)
1298 		return (B_TRUE);
1299 
1300 	/*
1301 	 * Calculate the number of 64-bit entries this segment would
1302 	 * require when written to disk. If this single segment would be
1303 	 * larger on-disk than the entire current on-disk structure, then
1304 	 * clearly condensing will increase the on-disk structure size.
1305 	 */
1306 	size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
1307 	entries = size / (MIN(size, SM_RUN_MAX));
1308 	segsz = entries * sizeof (uint64_t);
1309 
1310 	return (segsz <= space_map_length(msp->ms_sm) &&
1311 	    space_map_length(msp->ms_sm) >= (zfs_condense_pct *
1312 	    sizeof (uint64_t) * avl_numnodes(&msp->ms_tree->rt_root)) / 100);
1313 }
1314 
1315 /*
1316  * Condense the on-disk space map representation to its minimized form.
1317  * The minimized form consists of a small number of allocations followed by
1318  * the entries of the free range tree.
1319  */
1320 static void
1321 metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
1322 {
1323 	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1324 	range_tree_t *freetree = msp->ms_freetree[txg & TXG_MASK];
1325 	range_tree_t *condense_tree;
1326 	space_map_t *sm = msp->ms_sm;
1327 
1328 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1329 	ASSERT3U(spa_sync_pass(spa), ==, 1);
1330 	ASSERT(msp->ms_loaded);
1331 
1332 	spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, "
1333 	    "smp size %llu, segments %lu", txg, msp->ms_id, msp,
1334 	    space_map_length(msp->ms_sm), avl_numnodes(&msp->ms_tree->rt_root));
1335 
1336 	/*
1337 	 * Create an range tree that is 100% allocated. We remove segments
1338 	 * that have been freed in this txg, any deferred frees that exist,
1339 	 * and any allocation in the future. Removing segments should be
1340 	 * a relatively inexpensive operation since we expect these trees to
1341 	 * have a small number of nodes.
1342 	 */
1343 	condense_tree = range_tree_create(NULL, NULL, &msp->ms_lock);
1344 	range_tree_add(condense_tree, msp->ms_start, msp->ms_size);
1345 
1346 	/*
1347 	 * Remove what's been freed in this txg from the condense_tree.
1348 	 * Since we're in sync_pass 1, we know that all the frees from
1349 	 * this txg are in the freetree.
1350 	 */
1351 	range_tree_walk(freetree, range_tree_remove, condense_tree);
1352 
1353 	for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1354 		range_tree_walk(msp->ms_defertree[t],
1355 		    range_tree_remove, condense_tree);
1356 	}
1357 
1358 	for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
1359 		range_tree_walk(msp->ms_alloctree[(txg + t) & TXG_MASK],
1360 		    range_tree_remove, condense_tree);
1361 	}
1362 
1363 	/*
1364 	 * We're about to drop the metaslab's lock thus allowing
1365 	 * other consumers to change it's content. Set the
1366 	 * metaslab's ms_condensing flag to ensure that
1367 	 * allocations on this metaslab do not occur while we're
1368 	 * in the middle of committing it to disk. This is only critical
1369 	 * for the ms_tree as all other range trees use per txg
1370 	 * views of their content.
1371 	 */
1372 	msp->ms_condensing = B_TRUE;
1373 
1374 	mutex_exit(&msp->ms_lock);
1375 	space_map_truncate(sm, tx);
1376 	mutex_enter(&msp->ms_lock);
1377 
1378 	/*
1379 	 * While we would ideally like to create a space_map representation
1380 	 * that consists only of allocation records, doing so can be
1381 	 * prohibitively expensive because the in-core free tree can be
1382 	 * large, and therefore computationally expensive to subtract
1383 	 * from the condense_tree. Instead we sync out two trees, a cheap
1384 	 * allocation only tree followed by the in-core free tree. While not
1385 	 * optimal, this is typically close to optimal, and much cheaper to
1386 	 * compute.
1387 	 */
1388 	space_map_write(sm, condense_tree, SM_ALLOC, tx);
1389 	range_tree_vacate(condense_tree, NULL, NULL);
1390 	range_tree_destroy(condense_tree);
1391 
1392 	space_map_write(sm, msp->ms_tree, SM_FREE, tx);
1393 	msp->ms_condensing = B_FALSE;
1394 }
1395 
1396 /*
1397  * Write a metaslab to disk in the context of the specified transaction group.
1398  */
1399 void
1400 metaslab_sync(metaslab_t *msp, uint64_t txg)
1401 {
1402 	metaslab_group_t *mg = msp->ms_group;
1403 	vdev_t *vd = mg->mg_vd;
1404 	spa_t *spa = vd->vdev_spa;
1405 	objset_t *mos = spa_meta_objset(spa);
1406 	range_tree_t *alloctree = msp->ms_alloctree[txg & TXG_MASK];
1407 	range_tree_t **freetree = &msp->ms_freetree[txg & TXG_MASK];
1408 	range_tree_t **freed_tree =
1409 	    &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
1410 	dmu_tx_t *tx;
1411 	uint64_t object = space_map_object(msp->ms_sm);
1412 
1413 	ASSERT(!vd->vdev_ishole);
1414 
1415 	/*
1416 	 * This metaslab has just been added so there's no work to do now.
1417 	 */
1418 	if (*freetree == NULL) {
1419 		ASSERT3P(alloctree, ==, NULL);
1420 		return;
1421 	}
1422 
1423 	ASSERT3P(alloctree, !=, NULL);
1424 	ASSERT3P(*freetree, !=, NULL);
1425 	ASSERT3P(*freed_tree, !=, NULL);
1426 
1427 	if (range_tree_space(alloctree) == 0 &&
1428 	    range_tree_space(*freetree) == 0)
1429 		return;
1430 
1431 	/*
1432 	 * The only state that can actually be changing concurrently with
1433 	 * metaslab_sync() is the metaslab's ms_tree.  No other thread can
1434 	 * be modifying this txg's alloctree, freetree, freed_tree, or
1435 	 * space_map_phys_t. Therefore, we only hold ms_lock to satify
1436 	 * space_map ASSERTs. We drop it whenever we call into the DMU,
1437 	 * because the DMU can call down to us (e.g. via zio_free()) at
1438 	 * any time.
1439 	 */
1440 
1441 	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
1442 
1443 	if (msp->ms_sm == NULL) {
1444 		uint64_t new_object;
1445 
1446 		new_object = space_map_alloc(mos, tx);
1447 		VERIFY3U(new_object, !=, 0);
1448 
1449 		VERIFY0(space_map_open(&msp->ms_sm, mos, new_object,
1450 		    msp->ms_start, msp->ms_size, vd->vdev_ashift,
1451 		    &msp->ms_lock));
1452 		ASSERT(msp->ms_sm != NULL);
1453 	}
1454 
1455 	mutex_enter(&msp->ms_lock);
1456 
1457 	if (msp->ms_loaded && spa_sync_pass(spa) == 1 &&
1458 	    metaslab_should_condense(msp)) {
1459 		metaslab_condense(msp, txg, tx);
1460 	} else {
1461 		space_map_write(msp->ms_sm, alloctree, SM_ALLOC, tx);
1462 		space_map_write(msp->ms_sm, *freetree, SM_FREE, tx);
1463 	}
1464 
1465 	range_tree_vacate(alloctree, NULL, NULL);
1466 
1467 	if (msp->ms_loaded) {
1468 		/*
1469 		 * When the space map is loaded, we have an accruate
1470 		 * histogram in the range tree. This gives us an opportunity
1471 		 * to bring the space map's histogram up-to-date so we clear
1472 		 * it first before updating it.
1473 		 */
1474 		space_map_histogram_clear(msp->ms_sm);
1475 		space_map_histogram_add(msp->ms_sm, msp->ms_tree, tx);
1476 	} else {
1477 		/*
1478 		 * Since the space map is not loaded we simply update the
1479 		 * exisiting histogram with what was freed in this txg. This
1480 		 * means that the on-disk histogram may not have an accurate
1481 		 * view of the free space but it's close enough to allow
1482 		 * us to make allocation decisions.
1483 		 */
1484 		space_map_histogram_add(msp->ms_sm, *freetree, tx);
1485 	}
1486 
1487 	/*
1488 	 * For sync pass 1, we avoid traversing this txg's free range tree
1489 	 * and instead will just swap the pointers for freetree and
1490 	 * freed_tree. We can safely do this since the freed_tree is
1491 	 * guaranteed to be empty on the initial pass.
1492 	 */
1493 	if (spa_sync_pass(spa) == 1) {
1494 		range_tree_swap(freetree, freed_tree);
1495 	} else {
1496 		range_tree_vacate(*freetree, range_tree_add, *freed_tree);
1497 	}
1498 
1499 	ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
1500 	ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
1501 
1502 	mutex_exit(&msp->ms_lock);
1503 
1504 	if (object != space_map_object(msp->ms_sm)) {
1505 		object = space_map_object(msp->ms_sm);
1506 		dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
1507 		    msp->ms_id, sizeof (uint64_t), &object, tx);
1508 	}
1509 	dmu_tx_commit(tx);
1510 }
1511 
1512 /*
1513  * Called after a transaction group has completely synced to mark
1514  * all of the metaslab's free space as usable.
1515  */
1516 void
1517 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
1518 {
1519 	metaslab_group_t *mg = msp->ms_group;
1520 	vdev_t *vd = mg->mg_vd;
1521 	range_tree_t **freed_tree;
1522 	range_tree_t **defer_tree;
1523 	int64_t alloc_delta, defer_delta;
1524 
1525 	ASSERT(!vd->vdev_ishole);
1526 
1527 	mutex_enter(&msp->ms_lock);
1528 
1529 	/*
1530 	 * If this metaslab is just becoming available, initialize its
1531 	 * alloctrees, freetrees, and defertree and add its capacity to
1532 	 * the vdev.
1533 	 */
1534 	if (msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK] == NULL) {
1535 		for (int t = 0; t < TXG_SIZE; t++) {
1536 			ASSERT(msp->ms_alloctree[t] == NULL);
1537 			ASSERT(msp->ms_freetree[t] == NULL);
1538 
1539 			msp->ms_alloctree[t] = range_tree_create(NULL, msp,
1540 			    &msp->ms_lock);
1541 			msp->ms_freetree[t] = range_tree_create(NULL, msp,
1542 			    &msp->ms_lock);
1543 		}
1544 
1545 		for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1546 			ASSERT(msp->ms_defertree[t] == NULL);
1547 
1548 			msp->ms_defertree[t] = range_tree_create(NULL, msp,
1549 			    &msp->ms_lock);
1550 		}
1551 
1552 		vdev_space_update(vd, 0, 0, msp->ms_size);
1553 	}
1554 
1555 	freed_tree = &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
1556 	defer_tree = &msp->ms_defertree[txg % TXG_DEFER_SIZE];
1557 
1558 	alloc_delta = space_map_alloc_delta(msp->ms_sm);
1559 	defer_delta = range_tree_space(*freed_tree) -
1560 	    range_tree_space(*defer_tree);
1561 
1562 	vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
1563 
1564 	ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
1565 	ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
1566 
1567 	/*
1568 	 * If there's a metaslab_load() in progress, wait for it to complete
1569 	 * so that we have a consistent view of the in-core space map.
1570 	 */
1571 	metaslab_load_wait(msp);
1572 
1573 	/*
1574 	 * Move the frees from the defer_tree back to the free
1575 	 * range tree (if it's loaded). Swap the freed_tree and the
1576 	 * defer_tree -- this is safe to do because we've just emptied out
1577 	 * the defer_tree.
1578 	 */
1579 	range_tree_vacate(*defer_tree,
1580 	    msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
1581 	range_tree_swap(freed_tree, defer_tree);
1582 
1583 	space_map_update(msp->ms_sm);
1584 
1585 	msp->ms_deferspace += defer_delta;
1586 	ASSERT3S(msp->ms_deferspace, >=, 0);
1587 	ASSERT3S(msp->ms_deferspace, <=, msp->ms_size);
1588 	if (msp->ms_deferspace != 0) {
1589 		/*
1590 		 * Keep syncing this metaslab until all deferred frees
1591 		 * are back in circulation.
1592 		 */
1593 		vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1594 	}
1595 
1596 	if (msp->ms_loaded && msp->ms_access_txg < txg) {
1597 		for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
1598 			VERIFY0(range_tree_space(
1599 			    msp->ms_alloctree[(txg + t) & TXG_MASK]));
1600 		}
1601 
1602 		if (!metaslab_debug_unload)
1603 			metaslab_unload(msp);
1604 	}
1605 
1606 	metaslab_group_sort(mg, msp, metaslab_weight(msp));
1607 	mutex_exit(&msp->ms_lock);
1608 
1609 }
1610 
1611 void
1612 metaslab_sync_reassess(metaslab_group_t *mg)
1613 {
1614 	int64_t failures = mg->mg_alloc_failures;
1615 
1616 	metaslab_group_alloc_update(mg);
1617 	atomic_add_64(&mg->mg_alloc_failures, -failures);
1618 
1619 	/*
1620 	 * Preload the next potential metaslabs
1621 	 */
1622 	metaslab_group_preload(mg);
1623 }
1624 
1625 static uint64_t
1626 metaslab_distance(metaslab_t *msp, dva_t *dva)
1627 {
1628 	uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
1629 	uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
1630 	uint64_t start = msp->ms_id;
1631 
1632 	if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
1633 		return (1ULL << 63);
1634 
1635 	if (offset < start)
1636 		return ((start - offset) << ms_shift);
1637 	if (offset > start)
1638 		return ((offset - start) << ms_shift);
1639 	return (0);
1640 }
1641 
1642 static uint64_t
1643 metaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize,
1644     uint64_t txg, uint64_t min_distance, dva_t *dva, int d, int flags)
1645 {
1646 	spa_t *spa = mg->mg_vd->vdev_spa;
1647 	metaslab_t *msp = NULL;
1648 	uint64_t offset = -1ULL;
1649 	avl_tree_t *t = &mg->mg_metaslab_tree;
1650 	uint64_t activation_weight;
1651 	uint64_t target_distance;
1652 	int i;
1653 
1654 	activation_weight = METASLAB_WEIGHT_PRIMARY;
1655 	for (i = 0; i < d; i++) {
1656 		if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
1657 			activation_weight = METASLAB_WEIGHT_SECONDARY;
1658 			break;
1659 		}
1660 	}
1661 
1662 	for (;;) {
1663 		boolean_t was_active;
1664 
1665 		mutex_enter(&mg->mg_lock);
1666 		for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
1667 			if (msp->ms_weight < asize) {
1668 				spa_dbgmsg(spa, "%s: failed to meet weight "
1669 				    "requirement: vdev %llu, txg %llu, mg %p, "
1670 				    "msp %p, psize %llu, asize %llu, "
1671 				    "failures %llu, weight %llu",
1672 				    spa_name(spa), mg->mg_vd->vdev_id, txg,
1673 				    mg, msp, psize, asize,
1674 				    mg->mg_alloc_failures, msp->ms_weight);
1675 				mutex_exit(&mg->mg_lock);
1676 				return (-1ULL);
1677 			}
1678 
1679 			/*
1680 			 * If the selected metaslab is condensing, skip it.
1681 			 */
1682 			if (msp->ms_condensing)
1683 				continue;
1684 
1685 			was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
1686 			if (activation_weight == METASLAB_WEIGHT_PRIMARY)
1687 				break;
1688 
1689 			target_distance = min_distance +
1690 			    (space_map_allocated(msp->ms_sm) != 0 ? 0 :
1691 			    min_distance >> 1);
1692 
1693 			for (i = 0; i < d; i++)
1694 				if (metaslab_distance(msp, &dva[i]) <
1695 				    target_distance)
1696 					break;
1697 			if (i == d)
1698 				break;
1699 		}
1700 		mutex_exit(&mg->mg_lock);
1701 		if (msp == NULL)
1702 			return (-1ULL);
1703 
1704 		mutex_enter(&msp->ms_lock);
1705 
1706 		/*
1707 		 * If we've already reached the allowable number of failed
1708 		 * allocation attempts on this metaslab group then we
1709 		 * consider skipping it. We skip it only if we're allowed
1710 		 * to "fast" gang, the physical size is larger than
1711 		 * a gang block, and we're attempting to allocate from
1712 		 * the primary metaslab.
1713 		 */
1714 		if (mg->mg_alloc_failures > zfs_mg_alloc_failures &&
1715 		    CAN_FASTGANG(flags) && psize > SPA_GANGBLOCKSIZE &&
1716 		    activation_weight == METASLAB_WEIGHT_PRIMARY) {
1717 			spa_dbgmsg(spa, "%s: skipping metaslab group: "
1718 			    "vdev %llu, txg %llu, mg %p, msp[%llu] %p, "
1719 			    "psize %llu, asize %llu, failures %llu",
1720 			    spa_name(spa), mg->mg_vd->vdev_id, txg, mg,
1721 			    msp->ms_id, msp, psize, asize,
1722 			    mg->mg_alloc_failures);
1723 			mutex_exit(&msp->ms_lock);
1724 			return (-1ULL);
1725 		}
1726 
1727 		/*
1728 		 * Ensure that the metaslab we have selected is still
1729 		 * capable of handling our request. It's possible that
1730 		 * another thread may have changed the weight while we
1731 		 * were blocked on the metaslab lock.
1732 		 */
1733 		if (msp->ms_weight < asize || (was_active &&
1734 		    !(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
1735 		    activation_weight == METASLAB_WEIGHT_PRIMARY)) {
1736 			mutex_exit(&msp->ms_lock);
1737 			continue;
1738 		}
1739 
1740 		if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
1741 		    activation_weight == METASLAB_WEIGHT_PRIMARY) {
1742 			metaslab_passivate(msp,
1743 			    msp->ms_weight & ~METASLAB_ACTIVE_MASK);
1744 			mutex_exit(&msp->ms_lock);
1745 			continue;
1746 		}
1747 
1748 		if (metaslab_activate(msp, activation_weight) != 0) {
1749 			mutex_exit(&msp->ms_lock);
1750 			continue;
1751 		}
1752 
1753 		/*
1754 		 * If this metaslab is currently condensing then pick again as
1755 		 * we can't manipulate this metaslab until it's committed
1756 		 * to disk.
1757 		 */
1758 		if (msp->ms_condensing) {
1759 			mutex_exit(&msp->ms_lock);
1760 			continue;
1761 		}
1762 
1763 		if ((offset = metaslab_block_alloc(msp, asize)) != -1ULL)
1764 			break;
1765 
1766 		atomic_inc_64(&mg->mg_alloc_failures);
1767 
1768 		metaslab_passivate(msp, metaslab_block_maxsize(msp));
1769 		mutex_exit(&msp->ms_lock);
1770 	}
1771 
1772 	if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
1773 		vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
1774 
1775 	range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, asize);
1776 	msp->ms_access_txg = txg + metaslab_unload_delay;
1777 
1778 	mutex_exit(&msp->ms_lock);
1779 
1780 	return (offset);
1781 }
1782 
1783 /*
1784  * Allocate a block for the specified i/o.
1785  */
1786 static int
1787 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
1788     dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
1789 {
1790 	metaslab_group_t *mg, *rotor;
1791 	vdev_t *vd;
1792 	int dshift = 3;
1793 	int all_zero;
1794 	int zio_lock = B_FALSE;
1795 	boolean_t allocatable;
1796 	uint64_t offset = -1ULL;
1797 	uint64_t asize;
1798 	uint64_t distance;
1799 
1800 	ASSERT(!DVA_IS_VALID(&dva[d]));
1801 
1802 	/*
1803 	 * For testing, make some blocks above a certain size be gang blocks.
1804 	 */
1805 	if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0)
1806 		return (SET_ERROR(ENOSPC));
1807 
1808 	/*
1809 	 * Start at the rotor and loop through all mgs until we find something.
1810 	 * Note that there's no locking on mc_rotor or mc_aliquot because
1811 	 * nothing actually breaks if we miss a few updates -- we just won't
1812 	 * allocate quite as evenly.  It all balances out over time.
1813 	 *
1814 	 * If we are doing ditto or log blocks, try to spread them across
1815 	 * consecutive vdevs.  If we're forced to reuse a vdev before we've
1816 	 * allocated all of our ditto blocks, then try and spread them out on
1817 	 * that vdev as much as possible.  If it turns out to not be possible,
1818 	 * gradually lower our standards until anything becomes acceptable.
1819 	 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
1820 	 * gives us hope of containing our fault domains to something we're
1821 	 * able to reason about.  Otherwise, any two top-level vdev failures
1822 	 * will guarantee the loss of data.  With consecutive allocation,
1823 	 * only two adjacent top-level vdev failures will result in data loss.
1824 	 *
1825 	 * If we are doing gang blocks (hintdva is non-NULL), try to keep
1826 	 * ourselves on the same vdev as our gang block header.  That
1827 	 * way, we can hope for locality in vdev_cache, plus it makes our
1828 	 * fault domains something tractable.
1829 	 */
1830 	if (hintdva) {
1831 		vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
1832 
1833 		/*
1834 		 * It's possible the vdev we're using as the hint no
1835 		 * longer exists (i.e. removed). Consult the rotor when
1836 		 * all else fails.
1837 		 */
1838 		if (vd != NULL) {
1839 			mg = vd->vdev_mg;
1840 
1841 			if (flags & METASLAB_HINTBP_AVOID &&
1842 			    mg->mg_next != NULL)
1843 				mg = mg->mg_next;
1844 		} else {
1845 			mg = mc->mc_rotor;
1846 		}
1847 	} else if (d != 0) {
1848 		vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
1849 		mg = vd->vdev_mg->mg_next;
1850 	} else {
1851 		mg = mc->mc_rotor;
1852 	}
1853 
1854 	/*
1855 	 * If the hint put us into the wrong metaslab class, or into a
1856 	 * metaslab group that has been passivated, just follow the rotor.
1857 	 */
1858 	if (mg->mg_class != mc || mg->mg_activation_count <= 0)
1859 		mg = mc->mc_rotor;
1860 
1861 	rotor = mg;
1862 top:
1863 	all_zero = B_TRUE;
1864 	do {
1865 		ASSERT(mg->mg_activation_count == 1);
1866 
1867 		vd = mg->mg_vd;
1868 
1869 		/*
1870 		 * Don't allocate from faulted devices.
1871 		 */
1872 		if (zio_lock) {
1873 			spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
1874 			allocatable = vdev_allocatable(vd);
1875 			spa_config_exit(spa, SCL_ZIO, FTAG);
1876 		} else {
1877 			allocatable = vdev_allocatable(vd);
1878 		}
1879 
1880 		/*
1881 		 * Determine if the selected metaslab group is eligible
1882 		 * for allocations. If we're ganging or have requested
1883 		 * an allocation for the smallest gang block size
1884 		 * then we don't want to avoid allocating to the this
1885 		 * metaslab group. If we're in this condition we should
1886 		 * try to allocate from any device possible so that we
1887 		 * don't inadvertently return ENOSPC and suspend the pool
1888 		 * even though space is still available.
1889 		 */
1890 		if (allocatable && CAN_FASTGANG(flags) &&
1891 		    psize > SPA_GANGBLOCKSIZE)
1892 			allocatable = metaslab_group_allocatable(mg);
1893 
1894 		if (!allocatable)
1895 			goto next;
1896 
1897 		/*
1898 		 * Avoid writing single-copy data to a failing vdev
1899 		 * unless the user instructs us that it is okay.
1900 		 */
1901 		if ((vd->vdev_stat.vs_write_errors > 0 ||
1902 		    vd->vdev_state < VDEV_STATE_HEALTHY) &&
1903 		    d == 0 && dshift == 3 &&
1904 		    !(zfs_write_to_degraded && vd->vdev_state ==
1905 		    VDEV_STATE_DEGRADED)) {
1906 			all_zero = B_FALSE;
1907 			goto next;
1908 		}
1909 
1910 		ASSERT(mg->mg_class == mc);
1911 
1912 		distance = vd->vdev_asize >> dshift;
1913 		if (distance <= (1ULL << vd->vdev_ms_shift))
1914 			distance = 0;
1915 		else
1916 			all_zero = B_FALSE;
1917 
1918 		asize = vdev_psize_to_asize(vd, psize);
1919 		ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
1920 
1921 		offset = metaslab_group_alloc(mg, psize, asize, txg, distance,
1922 		    dva, d, flags);
1923 		if (offset != -1ULL) {
1924 			/*
1925 			 * If we've just selected this metaslab group,
1926 			 * figure out whether the corresponding vdev is
1927 			 * over- or under-used relative to the pool,
1928 			 * and set an allocation bias to even it out.
1929 			 */
1930 			if (mc->mc_aliquot == 0) {
1931 				vdev_stat_t *vs = &vd->vdev_stat;
1932 				int64_t vu, cu;
1933 
1934 				vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
1935 				cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
1936 
1937 				/*
1938 				 * Calculate how much more or less we should
1939 				 * try to allocate from this device during
1940 				 * this iteration around the rotor.
1941 				 * For example, if a device is 80% full
1942 				 * and the pool is 20% full then we should
1943 				 * reduce allocations by 60% on this device.
1944 				 *
1945 				 * mg_bias = (20 - 80) * 512K / 100 = -307K
1946 				 *
1947 				 * This reduces allocations by 307K for this
1948 				 * iteration.
1949 				 */
1950 				mg->mg_bias = ((cu - vu) *
1951 				    (int64_t)mg->mg_aliquot) / 100;
1952 			}
1953 
1954 			if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
1955 			    mg->mg_aliquot + mg->mg_bias) {
1956 				mc->mc_rotor = mg->mg_next;
1957 				mc->mc_aliquot = 0;
1958 			}
1959 
1960 			DVA_SET_VDEV(&dva[d], vd->vdev_id);
1961 			DVA_SET_OFFSET(&dva[d], offset);
1962 			DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
1963 			DVA_SET_ASIZE(&dva[d], asize);
1964 
1965 			return (0);
1966 		}
1967 next:
1968 		mc->mc_rotor = mg->mg_next;
1969 		mc->mc_aliquot = 0;
1970 	} while ((mg = mg->mg_next) != rotor);
1971 
1972 	if (!all_zero) {
1973 		dshift++;
1974 		ASSERT(dshift < 64);
1975 		goto top;
1976 	}
1977 
1978 	if (!allocatable && !zio_lock) {
1979 		dshift = 3;
1980 		zio_lock = B_TRUE;
1981 		goto top;
1982 	}
1983 
1984 	bzero(&dva[d], sizeof (dva_t));
1985 
1986 	return (SET_ERROR(ENOSPC));
1987 }
1988 
1989 /*
1990  * Free the block represented by DVA in the context of the specified
1991  * transaction group.
1992  */
1993 static void
1994 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
1995 {
1996 	uint64_t vdev = DVA_GET_VDEV(dva);
1997 	uint64_t offset = DVA_GET_OFFSET(dva);
1998 	uint64_t size = DVA_GET_ASIZE(dva);
1999 	vdev_t *vd;
2000 	metaslab_t *msp;
2001 
2002 	ASSERT(DVA_IS_VALID(dva));
2003 
2004 	if (txg > spa_freeze_txg(spa))
2005 		return;
2006 
2007 	if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
2008 	    (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
2009 		cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
2010 		    (u_longlong_t)vdev, (u_longlong_t)offset);
2011 		ASSERT(0);
2012 		return;
2013 	}
2014 
2015 	msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2016 
2017 	if (DVA_GET_GANG(dva))
2018 		size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
2019 
2020 	mutex_enter(&msp->ms_lock);
2021 
2022 	if (now) {
2023 		range_tree_remove(msp->ms_alloctree[txg & TXG_MASK],
2024 		    offset, size);
2025 
2026 		VERIFY(!msp->ms_condensing);
2027 		VERIFY3U(offset, >=, msp->ms_start);
2028 		VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size);
2029 		VERIFY3U(range_tree_space(msp->ms_tree) + size, <=,
2030 		    msp->ms_size);
2031 		VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
2032 		VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2033 		range_tree_add(msp->ms_tree, offset, size);
2034 	} else {
2035 		if (range_tree_space(msp->ms_freetree[txg & TXG_MASK]) == 0)
2036 			vdev_dirty(vd, VDD_METASLAB, msp, txg);
2037 		range_tree_add(msp->ms_freetree[txg & TXG_MASK],
2038 		    offset, size);
2039 	}
2040 
2041 	mutex_exit(&msp->ms_lock);
2042 }
2043 
2044 /*
2045  * Intent log support: upon opening the pool after a crash, notify the SPA
2046  * of blocks that the intent log has allocated for immediate write, but
2047  * which are still considered free by the SPA because the last transaction
2048  * group didn't commit yet.
2049  */
2050 static int
2051 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
2052 {
2053 	uint64_t vdev = DVA_GET_VDEV(dva);
2054 	uint64_t offset = DVA_GET_OFFSET(dva);
2055 	uint64_t size = DVA_GET_ASIZE(dva);
2056 	vdev_t *vd;
2057 	metaslab_t *msp;
2058 	int error = 0;
2059 
2060 	ASSERT(DVA_IS_VALID(dva));
2061 
2062 	if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
2063 	    (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
2064 		return (SET_ERROR(ENXIO));
2065 
2066 	msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2067 
2068 	if (DVA_GET_GANG(dva))
2069 		size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
2070 
2071 	mutex_enter(&msp->ms_lock);
2072 
2073 	if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded)
2074 		error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
2075 
2076 	if (error == 0 && !range_tree_contains(msp->ms_tree, offset, size))
2077 		error = SET_ERROR(ENOENT);
2078 
2079 	if (error || txg == 0) {	/* txg == 0 indicates dry run */
2080 		mutex_exit(&msp->ms_lock);
2081 		return (error);
2082 	}
2083 
2084 	VERIFY(!msp->ms_condensing);
2085 	VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
2086 	VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2087 	VERIFY3U(range_tree_space(msp->ms_tree) - size, <=, msp->ms_size);
2088 	range_tree_remove(msp->ms_tree, offset, size);
2089 
2090 	if (spa_writeable(spa)) {	/* don't dirty if we're zdb(1M) */
2091 		if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
2092 			vdev_dirty(vd, VDD_METASLAB, msp, txg);
2093 		range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, size);
2094 	}
2095 
2096 	mutex_exit(&msp->ms_lock);
2097 
2098 	return (0);
2099 }
2100 
2101 int
2102 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
2103     int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
2104 {
2105 	dva_t *dva = bp->blk_dva;
2106 	dva_t *hintdva = hintbp->blk_dva;
2107 	int error = 0;
2108 
2109 	ASSERT(bp->blk_birth == 0);
2110 	ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
2111 
2112 	spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
2113 
2114 	if (mc->mc_rotor == NULL) {	/* no vdevs in this class */
2115 		spa_config_exit(spa, SCL_ALLOC, FTAG);
2116 		return (SET_ERROR(ENOSPC));
2117 	}
2118 
2119 	ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
2120 	ASSERT(BP_GET_NDVAS(bp) == 0);
2121 	ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
2122 
2123 	for (int d = 0; d < ndvas; d++) {
2124 		error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
2125 		    txg, flags);
2126 		if (error != 0) {
2127 			for (d--; d >= 0; d--) {
2128 				metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
2129 				bzero(&dva[d], sizeof (dva_t));
2130 			}
2131 			spa_config_exit(spa, SCL_ALLOC, FTAG);
2132 			return (error);
2133 		}
2134 	}
2135 	ASSERT(error == 0);
2136 	ASSERT(BP_GET_NDVAS(bp) == ndvas);
2137 
2138 	spa_config_exit(spa, SCL_ALLOC, FTAG);
2139 
2140 	BP_SET_BIRTH(bp, txg, txg);
2141 
2142 	return (0);
2143 }
2144 
2145 void
2146 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
2147 {
2148 	const dva_t *dva = bp->blk_dva;
2149 	int ndvas = BP_GET_NDVAS(bp);
2150 
2151 	ASSERT(!BP_IS_HOLE(bp));
2152 	ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
2153 
2154 	spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
2155 
2156 	for (int d = 0; d < ndvas; d++)
2157 		metaslab_free_dva(spa, &dva[d], txg, now);
2158 
2159 	spa_config_exit(spa, SCL_FREE, FTAG);
2160 }
2161 
2162 int
2163 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
2164 {
2165 	const dva_t *dva = bp->blk_dva;
2166 	int ndvas = BP_GET_NDVAS(bp);
2167 	int error = 0;
2168 
2169 	ASSERT(!BP_IS_HOLE(bp));
2170 
2171 	if (txg != 0) {
2172 		/*
2173 		 * First do a dry run to make sure all DVAs are claimable,
2174 		 * so we don't have to unwind from partial failures below.
2175 		 */
2176 		if ((error = metaslab_claim(spa, bp, 0)) != 0)
2177 			return (error);
2178 	}
2179 
2180 	spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
2181 
2182 	for (int d = 0; d < ndvas; d++)
2183 		if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
2184 			break;
2185 
2186 	spa_config_exit(spa, SCL_ALLOC, FTAG);
2187 
2188 	ASSERT(error == 0 || txg == 0);
2189 
2190 	return (error);
2191 }
2192 
2193 void
2194 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
2195 {
2196 	if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
2197 		return;
2198 
2199 	spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
2200 	for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
2201 		uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
2202 		vdev_t *vd = vdev_lookup_top(spa, vdev);
2203 		uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
2204 		uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
2205 		metaslab_t *msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2206 
2207 		if (msp->ms_loaded)
2208 			range_tree_verify(msp->ms_tree, offset, size);
2209 
2210 		for (int j = 0; j < TXG_SIZE; j++)
2211 			range_tree_verify(msp->ms_freetree[j], offset, size);
2212 		for (int j = 0; j < TXG_DEFER_SIZE; j++)
2213 			range_tree_verify(msp->ms_defertree[j], offset, size);
2214 	}
2215 	spa_config_exit(spa, SCL_VDEV, FTAG);
2216 }
2217