xref: /titanic_44/usr/src/uts/common/fs/zfs/metaslab.c (revision b17f03d7d89b75b69b9b7db22f2316b700e3a5a8)
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) 2012 by Delphix. All rights reserved.
24  */
25 
26 #include <sys/zfs_context.h>
27 #include <sys/dmu.h>
28 #include <sys/dmu_tx.h>
29 #include <sys/space_map.h>
30 #include <sys/metaslab_impl.h>
31 #include <sys/vdev_impl.h>
32 #include <sys/zio.h>
33 
34 /*
35  * Allow allocations to switch to gang blocks quickly. We do this to
36  * avoid having to load lots of space_maps in a given txg. There are,
37  * however, some cases where we want to avoid "fast" ganging and instead
38  * we want to do an exhaustive search of all metaslabs on this device.
39  * Currently we don't allow any gang, zil, or dump device related allocations
40  * to "fast" gang.
41  */
42 #define	CAN_FASTGANG(flags) \
43 	(!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \
44 	METASLAB_GANG_AVOID)))
45 
46 uint64_t metaslab_aliquot = 512ULL << 10;
47 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1;	/* force gang blocks */
48 
49 /*
50  * This value defines the number of allowed allocation failures per vdev.
51  * If a device reaches this threshold in a given txg then we consider skipping
52  * allocations on that device.
53  */
54 int zfs_mg_alloc_failures;
55 
56 /*
57  * Metaslab debugging: when set, keeps all space maps in core to verify frees.
58  */
59 static int metaslab_debug = 0;
60 
61 /*
62  * Minimum size which forces the dynamic allocator to change
63  * it's allocation strategy.  Once the space map cannot satisfy
64  * an allocation of this size then it switches to using more
65  * aggressive strategy (i.e search by size rather than offset).
66  */
67 uint64_t metaslab_df_alloc_threshold = SPA_MAXBLOCKSIZE;
68 
69 /*
70  * The minimum free space, in percent, which must be available
71  * in a space map to continue allocations in a first-fit fashion.
72  * Once the space_map's free space drops below this level we dynamically
73  * switch to using best-fit allocations.
74  */
75 int metaslab_df_free_pct = 4;
76 
77 /*
78  * A metaslab is considered "free" if it contains a contiguous
79  * segment which is greater than metaslab_min_alloc_size.
80  */
81 uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
82 
83 /*
84  * Max number of space_maps to prefetch.
85  */
86 int metaslab_prefetch_limit = SPA_DVAS_PER_BP;
87 
88 /*
89  * Percentage bonus multiplier for metaslabs that are in the bonus area.
90  */
91 int metaslab_smo_bonus_pct = 150;
92 
93 /*
94  * ==========================================================================
95  * Metaslab classes
96  * ==========================================================================
97  */
98 metaslab_class_t *
99 metaslab_class_create(spa_t *spa, space_map_ops_t *ops)
100 {
101 	metaslab_class_t *mc;
102 
103 	mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
104 
105 	mc->mc_spa = spa;
106 	mc->mc_rotor = NULL;
107 	mc->mc_ops = ops;
108 
109 	return (mc);
110 }
111 
112 void
113 metaslab_class_destroy(metaslab_class_t *mc)
114 {
115 	ASSERT(mc->mc_rotor == NULL);
116 	ASSERT(mc->mc_alloc == 0);
117 	ASSERT(mc->mc_deferred == 0);
118 	ASSERT(mc->mc_space == 0);
119 	ASSERT(mc->mc_dspace == 0);
120 
121 	kmem_free(mc, sizeof (metaslab_class_t));
122 }
123 
124 int
125 metaslab_class_validate(metaslab_class_t *mc)
126 {
127 	metaslab_group_t *mg;
128 	vdev_t *vd;
129 
130 	/*
131 	 * Must hold one of the spa_config locks.
132 	 */
133 	ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
134 	    spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
135 
136 	if ((mg = mc->mc_rotor) == NULL)
137 		return (0);
138 
139 	do {
140 		vd = mg->mg_vd;
141 		ASSERT(vd->vdev_mg != NULL);
142 		ASSERT3P(vd->vdev_top, ==, vd);
143 		ASSERT3P(mg->mg_class, ==, mc);
144 		ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
145 	} while ((mg = mg->mg_next) != mc->mc_rotor);
146 
147 	return (0);
148 }
149 
150 void
151 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
152     int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
153 {
154 	atomic_add_64(&mc->mc_alloc, alloc_delta);
155 	atomic_add_64(&mc->mc_deferred, defer_delta);
156 	atomic_add_64(&mc->mc_space, space_delta);
157 	atomic_add_64(&mc->mc_dspace, dspace_delta);
158 }
159 
160 uint64_t
161 metaslab_class_get_alloc(metaslab_class_t *mc)
162 {
163 	return (mc->mc_alloc);
164 }
165 
166 uint64_t
167 metaslab_class_get_deferred(metaslab_class_t *mc)
168 {
169 	return (mc->mc_deferred);
170 }
171 
172 uint64_t
173 metaslab_class_get_space(metaslab_class_t *mc)
174 {
175 	return (mc->mc_space);
176 }
177 
178 uint64_t
179 metaslab_class_get_dspace(metaslab_class_t *mc)
180 {
181 	return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
182 }
183 
184 /*
185  * ==========================================================================
186  * Metaslab groups
187  * ==========================================================================
188  */
189 static int
190 metaslab_compare(const void *x1, const void *x2)
191 {
192 	const metaslab_t *m1 = x1;
193 	const metaslab_t *m2 = x2;
194 
195 	if (m1->ms_weight < m2->ms_weight)
196 		return (1);
197 	if (m1->ms_weight > m2->ms_weight)
198 		return (-1);
199 
200 	/*
201 	 * If the weights are identical, use the offset to force uniqueness.
202 	 */
203 	if (m1->ms_map.sm_start < m2->ms_map.sm_start)
204 		return (-1);
205 	if (m1->ms_map.sm_start > m2->ms_map.sm_start)
206 		return (1);
207 
208 	ASSERT3P(m1, ==, m2);
209 
210 	return (0);
211 }
212 
213 metaslab_group_t *
214 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
215 {
216 	metaslab_group_t *mg;
217 
218 	mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
219 	mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
220 	avl_create(&mg->mg_metaslab_tree, metaslab_compare,
221 	    sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
222 	mg->mg_vd = vd;
223 	mg->mg_class = mc;
224 	mg->mg_activation_count = 0;
225 
226 	return (mg);
227 }
228 
229 void
230 metaslab_group_destroy(metaslab_group_t *mg)
231 {
232 	ASSERT(mg->mg_prev == NULL);
233 	ASSERT(mg->mg_next == NULL);
234 	/*
235 	 * We may have gone below zero with the activation count
236 	 * either because we never activated in the first place or
237 	 * because we're done, and possibly removing the vdev.
238 	 */
239 	ASSERT(mg->mg_activation_count <= 0);
240 
241 	avl_destroy(&mg->mg_metaslab_tree);
242 	mutex_destroy(&mg->mg_lock);
243 	kmem_free(mg, sizeof (metaslab_group_t));
244 }
245 
246 void
247 metaslab_group_activate(metaslab_group_t *mg)
248 {
249 	metaslab_class_t *mc = mg->mg_class;
250 	metaslab_group_t *mgprev, *mgnext;
251 
252 	ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
253 
254 	ASSERT(mc->mc_rotor != mg);
255 	ASSERT(mg->mg_prev == NULL);
256 	ASSERT(mg->mg_next == NULL);
257 	ASSERT(mg->mg_activation_count <= 0);
258 
259 	if (++mg->mg_activation_count <= 0)
260 		return;
261 
262 	mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
263 
264 	if ((mgprev = mc->mc_rotor) == NULL) {
265 		mg->mg_prev = mg;
266 		mg->mg_next = mg;
267 	} else {
268 		mgnext = mgprev->mg_next;
269 		mg->mg_prev = mgprev;
270 		mg->mg_next = mgnext;
271 		mgprev->mg_next = mg;
272 		mgnext->mg_prev = mg;
273 	}
274 	mc->mc_rotor = mg;
275 }
276 
277 void
278 metaslab_group_passivate(metaslab_group_t *mg)
279 {
280 	metaslab_class_t *mc = mg->mg_class;
281 	metaslab_group_t *mgprev, *mgnext;
282 
283 	ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
284 
285 	if (--mg->mg_activation_count != 0) {
286 		ASSERT(mc->mc_rotor != mg);
287 		ASSERT(mg->mg_prev == NULL);
288 		ASSERT(mg->mg_next == NULL);
289 		ASSERT(mg->mg_activation_count < 0);
290 		return;
291 	}
292 
293 	mgprev = mg->mg_prev;
294 	mgnext = mg->mg_next;
295 
296 	if (mg == mgnext) {
297 		mc->mc_rotor = NULL;
298 	} else {
299 		mc->mc_rotor = mgnext;
300 		mgprev->mg_next = mgnext;
301 		mgnext->mg_prev = mgprev;
302 	}
303 
304 	mg->mg_prev = NULL;
305 	mg->mg_next = NULL;
306 }
307 
308 static void
309 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
310 {
311 	mutex_enter(&mg->mg_lock);
312 	ASSERT(msp->ms_group == NULL);
313 	msp->ms_group = mg;
314 	msp->ms_weight = 0;
315 	avl_add(&mg->mg_metaslab_tree, msp);
316 	mutex_exit(&mg->mg_lock);
317 }
318 
319 static void
320 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
321 {
322 	mutex_enter(&mg->mg_lock);
323 	ASSERT(msp->ms_group == mg);
324 	avl_remove(&mg->mg_metaslab_tree, msp);
325 	msp->ms_group = NULL;
326 	mutex_exit(&mg->mg_lock);
327 }
328 
329 static void
330 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
331 {
332 	/*
333 	 * Although in principle the weight can be any value, in
334 	 * practice we do not use values in the range [1, 510].
335 	 */
336 	ASSERT(weight >= SPA_MINBLOCKSIZE-1 || weight == 0);
337 	ASSERT(MUTEX_HELD(&msp->ms_lock));
338 
339 	mutex_enter(&mg->mg_lock);
340 	ASSERT(msp->ms_group == mg);
341 	avl_remove(&mg->mg_metaslab_tree, msp);
342 	msp->ms_weight = weight;
343 	avl_add(&mg->mg_metaslab_tree, msp);
344 	mutex_exit(&mg->mg_lock);
345 }
346 
347 /*
348  * ==========================================================================
349  * Common allocator routines
350  * ==========================================================================
351  */
352 static int
353 metaslab_segsize_compare(const void *x1, const void *x2)
354 {
355 	const space_seg_t *s1 = x1;
356 	const space_seg_t *s2 = x2;
357 	uint64_t ss_size1 = s1->ss_end - s1->ss_start;
358 	uint64_t ss_size2 = s2->ss_end - s2->ss_start;
359 
360 	if (ss_size1 < ss_size2)
361 		return (-1);
362 	if (ss_size1 > ss_size2)
363 		return (1);
364 
365 	if (s1->ss_start < s2->ss_start)
366 		return (-1);
367 	if (s1->ss_start > s2->ss_start)
368 		return (1);
369 
370 	return (0);
371 }
372 
373 /*
374  * This is a helper function that can be used by the allocator to find
375  * a suitable block to allocate. This will search the specified AVL
376  * tree looking for a block that matches the specified criteria.
377  */
378 static uint64_t
379 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
380     uint64_t align)
381 {
382 	space_seg_t *ss, ssearch;
383 	avl_index_t where;
384 
385 	ssearch.ss_start = *cursor;
386 	ssearch.ss_end = *cursor + size;
387 
388 	ss = avl_find(t, &ssearch, &where);
389 	if (ss == NULL)
390 		ss = avl_nearest(t, where, AVL_AFTER);
391 
392 	while (ss != NULL) {
393 		uint64_t offset = P2ROUNDUP(ss->ss_start, align);
394 
395 		if (offset + size <= ss->ss_end) {
396 			*cursor = offset + size;
397 			return (offset);
398 		}
399 		ss = AVL_NEXT(t, ss);
400 	}
401 
402 	/*
403 	 * If we know we've searched the whole map (*cursor == 0), give up.
404 	 * Otherwise, reset the cursor to the beginning and try again.
405 	 */
406 	if (*cursor == 0)
407 		return (-1ULL);
408 
409 	*cursor = 0;
410 	return (metaslab_block_picker(t, cursor, size, align));
411 }
412 
413 static void
414 metaslab_pp_load(space_map_t *sm)
415 {
416 	space_seg_t *ss;
417 
418 	ASSERT(sm->sm_ppd == NULL);
419 	sm->sm_ppd = kmem_zalloc(64 * sizeof (uint64_t), KM_SLEEP);
420 
421 	sm->sm_pp_root = kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
422 	avl_create(sm->sm_pp_root, metaslab_segsize_compare,
423 	    sizeof (space_seg_t), offsetof(struct space_seg, ss_pp_node));
424 
425 	for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
426 		avl_add(sm->sm_pp_root, ss);
427 }
428 
429 static void
430 metaslab_pp_unload(space_map_t *sm)
431 {
432 	void *cookie = NULL;
433 
434 	kmem_free(sm->sm_ppd, 64 * sizeof (uint64_t));
435 	sm->sm_ppd = NULL;
436 
437 	while (avl_destroy_nodes(sm->sm_pp_root, &cookie) != NULL) {
438 		/* tear down the tree */
439 	}
440 
441 	avl_destroy(sm->sm_pp_root);
442 	kmem_free(sm->sm_pp_root, sizeof (avl_tree_t));
443 	sm->sm_pp_root = NULL;
444 }
445 
446 /* ARGSUSED */
447 static void
448 metaslab_pp_claim(space_map_t *sm, uint64_t start, uint64_t size)
449 {
450 	/* No need to update cursor */
451 }
452 
453 /* ARGSUSED */
454 static void
455 metaslab_pp_free(space_map_t *sm, uint64_t start, uint64_t size)
456 {
457 	/* No need to update cursor */
458 }
459 
460 /*
461  * Return the maximum contiguous segment within the metaslab.
462  */
463 uint64_t
464 metaslab_pp_maxsize(space_map_t *sm)
465 {
466 	avl_tree_t *t = sm->sm_pp_root;
467 	space_seg_t *ss;
468 
469 	if (t == NULL || (ss = avl_last(t)) == NULL)
470 		return (0ULL);
471 
472 	return (ss->ss_end - ss->ss_start);
473 }
474 
475 /*
476  * ==========================================================================
477  * The first-fit block allocator
478  * ==========================================================================
479  */
480 static uint64_t
481 metaslab_ff_alloc(space_map_t *sm, uint64_t size)
482 {
483 	avl_tree_t *t = &sm->sm_root;
484 	uint64_t align = size & -size;
485 	uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
486 
487 	return (metaslab_block_picker(t, cursor, size, align));
488 }
489 
490 /* ARGSUSED */
491 boolean_t
492 metaslab_ff_fragmented(space_map_t *sm)
493 {
494 	return (B_TRUE);
495 }
496 
497 static space_map_ops_t metaslab_ff_ops = {
498 	metaslab_pp_load,
499 	metaslab_pp_unload,
500 	metaslab_ff_alloc,
501 	metaslab_pp_claim,
502 	metaslab_pp_free,
503 	metaslab_pp_maxsize,
504 	metaslab_ff_fragmented
505 };
506 
507 /*
508  * ==========================================================================
509  * Dynamic block allocator -
510  * Uses the first fit allocation scheme until space get low and then
511  * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
512  * and metaslab_df_free_pct to determine when to switch the allocation scheme.
513  * ==========================================================================
514  */
515 static uint64_t
516 metaslab_df_alloc(space_map_t *sm, uint64_t size)
517 {
518 	avl_tree_t *t = &sm->sm_root;
519 	uint64_t align = size & -size;
520 	uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
521 	uint64_t max_size = metaslab_pp_maxsize(sm);
522 	int free_pct = sm->sm_space * 100 / sm->sm_size;
523 
524 	ASSERT(MUTEX_HELD(sm->sm_lock));
525 	ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
526 
527 	if (max_size < size)
528 		return (-1ULL);
529 
530 	/*
531 	 * If we're running low on space switch to using the size
532 	 * sorted AVL tree (best-fit).
533 	 */
534 	if (max_size < metaslab_df_alloc_threshold ||
535 	    free_pct < metaslab_df_free_pct) {
536 		t = sm->sm_pp_root;
537 		*cursor = 0;
538 	}
539 
540 	return (metaslab_block_picker(t, cursor, size, 1ULL));
541 }
542 
543 static boolean_t
544 metaslab_df_fragmented(space_map_t *sm)
545 {
546 	uint64_t max_size = metaslab_pp_maxsize(sm);
547 	int free_pct = sm->sm_space * 100 / sm->sm_size;
548 
549 	if (max_size >= metaslab_df_alloc_threshold &&
550 	    free_pct >= metaslab_df_free_pct)
551 		return (B_FALSE);
552 
553 	return (B_TRUE);
554 }
555 
556 static space_map_ops_t metaslab_df_ops = {
557 	metaslab_pp_load,
558 	metaslab_pp_unload,
559 	metaslab_df_alloc,
560 	metaslab_pp_claim,
561 	metaslab_pp_free,
562 	metaslab_pp_maxsize,
563 	metaslab_df_fragmented
564 };
565 
566 /*
567  * ==========================================================================
568  * Other experimental allocators
569  * ==========================================================================
570  */
571 static uint64_t
572 metaslab_cdf_alloc(space_map_t *sm, uint64_t size)
573 {
574 	avl_tree_t *t = &sm->sm_root;
575 	uint64_t *cursor = (uint64_t *)sm->sm_ppd;
576 	uint64_t *extent_end = (uint64_t *)sm->sm_ppd + 1;
577 	uint64_t max_size = metaslab_pp_maxsize(sm);
578 	uint64_t rsize = size;
579 	uint64_t offset = 0;
580 
581 	ASSERT(MUTEX_HELD(sm->sm_lock));
582 	ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
583 
584 	if (max_size < size)
585 		return (-1ULL);
586 
587 	ASSERT3U(*extent_end, >=, *cursor);
588 
589 	/*
590 	 * If we're running low on space switch to using the size
591 	 * sorted AVL tree (best-fit).
592 	 */
593 	if ((*cursor + size) > *extent_end) {
594 
595 		t = sm->sm_pp_root;
596 		*cursor = *extent_end = 0;
597 
598 		if (max_size > 2 * SPA_MAXBLOCKSIZE)
599 			rsize = MIN(metaslab_min_alloc_size, max_size);
600 		offset = metaslab_block_picker(t, extent_end, rsize, 1ULL);
601 		if (offset != -1)
602 			*cursor = offset + size;
603 	} else {
604 		offset = metaslab_block_picker(t, cursor, rsize, 1ULL);
605 	}
606 	ASSERT3U(*cursor, <=, *extent_end);
607 	return (offset);
608 }
609 
610 static boolean_t
611 metaslab_cdf_fragmented(space_map_t *sm)
612 {
613 	uint64_t max_size = metaslab_pp_maxsize(sm);
614 
615 	if (max_size > (metaslab_min_alloc_size * 10))
616 		return (B_FALSE);
617 	return (B_TRUE);
618 }
619 
620 static space_map_ops_t metaslab_cdf_ops = {
621 	metaslab_pp_load,
622 	metaslab_pp_unload,
623 	metaslab_cdf_alloc,
624 	metaslab_pp_claim,
625 	metaslab_pp_free,
626 	metaslab_pp_maxsize,
627 	metaslab_cdf_fragmented
628 };
629 
630 uint64_t metaslab_ndf_clump_shift = 4;
631 
632 static uint64_t
633 metaslab_ndf_alloc(space_map_t *sm, uint64_t size)
634 {
635 	avl_tree_t *t = &sm->sm_root;
636 	avl_index_t where;
637 	space_seg_t *ss, ssearch;
638 	uint64_t hbit = highbit(size);
639 	uint64_t *cursor = (uint64_t *)sm->sm_ppd + hbit - 1;
640 	uint64_t max_size = metaslab_pp_maxsize(sm);
641 
642 	ASSERT(MUTEX_HELD(sm->sm_lock));
643 	ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
644 
645 	if (max_size < size)
646 		return (-1ULL);
647 
648 	ssearch.ss_start = *cursor;
649 	ssearch.ss_end = *cursor + size;
650 
651 	ss = avl_find(t, &ssearch, &where);
652 	if (ss == NULL || (ss->ss_start + size > ss->ss_end)) {
653 		t = sm->sm_pp_root;
654 
655 		ssearch.ss_start = 0;
656 		ssearch.ss_end = MIN(max_size,
657 		    1ULL << (hbit + metaslab_ndf_clump_shift));
658 		ss = avl_find(t, &ssearch, &where);
659 		if (ss == NULL)
660 			ss = avl_nearest(t, where, AVL_AFTER);
661 		ASSERT(ss != NULL);
662 	}
663 
664 	if (ss != NULL) {
665 		if (ss->ss_start + size <= ss->ss_end) {
666 			*cursor = ss->ss_start + size;
667 			return (ss->ss_start);
668 		}
669 	}
670 	return (-1ULL);
671 }
672 
673 static boolean_t
674 metaslab_ndf_fragmented(space_map_t *sm)
675 {
676 	uint64_t max_size = metaslab_pp_maxsize(sm);
677 
678 	if (max_size > (metaslab_min_alloc_size << metaslab_ndf_clump_shift))
679 		return (B_FALSE);
680 	return (B_TRUE);
681 }
682 
683 
684 static space_map_ops_t metaslab_ndf_ops = {
685 	metaslab_pp_load,
686 	metaslab_pp_unload,
687 	metaslab_ndf_alloc,
688 	metaslab_pp_claim,
689 	metaslab_pp_free,
690 	metaslab_pp_maxsize,
691 	metaslab_ndf_fragmented
692 };
693 
694 space_map_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
695 
696 /*
697  * ==========================================================================
698  * Metaslabs
699  * ==========================================================================
700  */
701 metaslab_t *
702 metaslab_init(metaslab_group_t *mg, space_map_obj_t *smo,
703 	uint64_t start, uint64_t size, uint64_t txg)
704 {
705 	vdev_t *vd = mg->mg_vd;
706 	metaslab_t *msp;
707 
708 	msp = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
709 	mutex_init(&msp->ms_lock, NULL, MUTEX_DEFAULT, NULL);
710 
711 	msp->ms_smo_syncing = *smo;
712 
713 	/*
714 	 * We create the main space map here, but we don't create the
715 	 * allocmaps and freemaps until metaslab_sync_done().  This serves
716 	 * two purposes: it allows metaslab_sync_done() to detect the
717 	 * addition of new space; and for debugging, it ensures that we'd
718 	 * data fault on any attempt to use this metaslab before it's ready.
719 	 */
720 	space_map_create(&msp->ms_map, start, size,
721 	    vd->vdev_ashift, &msp->ms_lock);
722 
723 	metaslab_group_add(mg, msp);
724 
725 	if (metaslab_debug && smo->smo_object != 0) {
726 		mutex_enter(&msp->ms_lock);
727 		VERIFY(space_map_load(&msp->ms_map, mg->mg_class->mc_ops,
728 		    SM_FREE, smo, spa_meta_objset(vd->vdev_spa)) == 0);
729 		mutex_exit(&msp->ms_lock);
730 	}
731 
732 	/*
733 	 * If we're opening an existing pool (txg == 0) or creating
734 	 * a new one (txg == TXG_INITIAL), all space is available now.
735 	 * If we're adding space to an existing pool, the new space
736 	 * does not become available until after this txg has synced.
737 	 */
738 	if (txg <= TXG_INITIAL)
739 		metaslab_sync_done(msp, 0);
740 
741 	if (txg != 0) {
742 		vdev_dirty(vd, 0, NULL, txg);
743 		vdev_dirty(vd, VDD_METASLAB, msp, txg);
744 	}
745 
746 	return (msp);
747 }
748 
749 void
750 metaslab_fini(metaslab_t *msp)
751 {
752 	metaslab_group_t *mg = msp->ms_group;
753 
754 	vdev_space_update(mg->mg_vd,
755 	    -msp->ms_smo.smo_alloc, 0, -msp->ms_map.sm_size);
756 
757 	metaslab_group_remove(mg, msp);
758 
759 	mutex_enter(&msp->ms_lock);
760 
761 	space_map_unload(&msp->ms_map);
762 	space_map_destroy(&msp->ms_map);
763 
764 	for (int t = 0; t < TXG_SIZE; t++) {
765 		space_map_destroy(&msp->ms_allocmap[t]);
766 		space_map_destroy(&msp->ms_freemap[t]);
767 	}
768 
769 	for (int t = 0; t < TXG_DEFER_SIZE; t++)
770 		space_map_destroy(&msp->ms_defermap[t]);
771 
772 	ASSERT0(msp->ms_deferspace);
773 
774 	mutex_exit(&msp->ms_lock);
775 	mutex_destroy(&msp->ms_lock);
776 
777 	kmem_free(msp, sizeof (metaslab_t));
778 }
779 
780 #define	METASLAB_WEIGHT_PRIMARY		(1ULL << 63)
781 #define	METASLAB_WEIGHT_SECONDARY	(1ULL << 62)
782 #define	METASLAB_ACTIVE_MASK		\
783 	(METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
784 
785 static uint64_t
786 metaslab_weight(metaslab_t *msp)
787 {
788 	metaslab_group_t *mg = msp->ms_group;
789 	space_map_t *sm = &msp->ms_map;
790 	space_map_obj_t *smo = &msp->ms_smo;
791 	vdev_t *vd = mg->mg_vd;
792 	uint64_t weight, space;
793 
794 	ASSERT(MUTEX_HELD(&msp->ms_lock));
795 
796 	/*
797 	 * The baseline weight is the metaslab's free space.
798 	 */
799 	space = sm->sm_size - smo->smo_alloc;
800 	weight = space;
801 
802 	/*
803 	 * Modern disks have uniform bit density and constant angular velocity.
804 	 * Therefore, the outer recording zones are faster (higher bandwidth)
805 	 * than the inner zones by the ratio of outer to inner track diameter,
806 	 * which is typically around 2:1.  We account for this by assigning
807 	 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
808 	 * In effect, this means that we'll select the metaslab with the most
809 	 * free bandwidth rather than simply the one with the most free space.
810 	 */
811 	weight = 2 * weight -
812 	    ((sm->sm_start >> vd->vdev_ms_shift) * weight) / vd->vdev_ms_count;
813 	ASSERT(weight >= space && weight <= 2 * space);
814 
815 	/*
816 	 * For locality, assign higher weight to metaslabs which have
817 	 * a lower offset than what we've already activated.
818 	 */
819 	if (sm->sm_start <= mg->mg_bonus_area)
820 		weight *= (metaslab_smo_bonus_pct / 100);
821 	ASSERT(weight >= space &&
822 	    weight <= 2 * (metaslab_smo_bonus_pct / 100) * space);
823 
824 	if (sm->sm_loaded && !sm->sm_ops->smop_fragmented(sm)) {
825 		/*
826 		 * If this metaslab is one we're actively using, adjust its
827 		 * weight to make it preferable to any inactive metaslab so
828 		 * we'll polish it off.
829 		 */
830 		weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
831 	}
832 	return (weight);
833 }
834 
835 static void
836 metaslab_prefetch(metaslab_group_t *mg)
837 {
838 	spa_t *spa = mg->mg_vd->vdev_spa;
839 	metaslab_t *msp;
840 	avl_tree_t *t = &mg->mg_metaslab_tree;
841 	int m;
842 
843 	mutex_enter(&mg->mg_lock);
844 
845 	/*
846 	 * Prefetch the next potential metaslabs
847 	 */
848 	for (msp = avl_first(t), m = 0; msp; msp = AVL_NEXT(t, msp), m++) {
849 		space_map_t *sm = &msp->ms_map;
850 		space_map_obj_t *smo = &msp->ms_smo;
851 
852 		/* If we have reached our prefetch limit then we're done */
853 		if (m >= metaslab_prefetch_limit)
854 			break;
855 
856 		if (!sm->sm_loaded && smo->smo_object != 0) {
857 			mutex_exit(&mg->mg_lock);
858 			dmu_prefetch(spa_meta_objset(spa), smo->smo_object,
859 			    0ULL, smo->smo_objsize);
860 			mutex_enter(&mg->mg_lock);
861 		}
862 	}
863 	mutex_exit(&mg->mg_lock);
864 }
865 
866 static int
867 metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
868 {
869 	metaslab_group_t *mg = msp->ms_group;
870 	space_map_t *sm = &msp->ms_map;
871 	space_map_ops_t *sm_ops = msp->ms_group->mg_class->mc_ops;
872 
873 	ASSERT(MUTEX_HELD(&msp->ms_lock));
874 
875 	if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
876 		space_map_load_wait(sm);
877 		if (!sm->sm_loaded) {
878 			space_map_obj_t *smo = &msp->ms_smo;
879 
880 			int error = space_map_load(sm, sm_ops, SM_FREE, smo,
881 			    spa_meta_objset(msp->ms_group->mg_vd->vdev_spa));
882 			if (error)  {
883 				metaslab_group_sort(msp->ms_group, msp, 0);
884 				return (error);
885 			}
886 			for (int t = 0; t < TXG_DEFER_SIZE; t++)
887 				space_map_walk(&msp->ms_defermap[t],
888 				    space_map_claim, sm);
889 
890 		}
891 
892 		/*
893 		 * Track the bonus area as we activate new metaslabs.
894 		 */
895 		if (sm->sm_start > mg->mg_bonus_area) {
896 			mutex_enter(&mg->mg_lock);
897 			mg->mg_bonus_area = sm->sm_start;
898 			mutex_exit(&mg->mg_lock);
899 		}
900 
901 		metaslab_group_sort(msp->ms_group, msp,
902 		    msp->ms_weight | activation_weight);
903 	}
904 	ASSERT(sm->sm_loaded);
905 	ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
906 
907 	return (0);
908 }
909 
910 static void
911 metaslab_passivate(metaslab_t *msp, uint64_t size)
912 {
913 	/*
914 	 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
915 	 * this metaslab again.  In that case, it had better be empty,
916 	 * or we would be leaving space on the table.
917 	 */
918 	ASSERT(size >= SPA_MINBLOCKSIZE || msp->ms_map.sm_space == 0);
919 	metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
920 	ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
921 }
922 
923 /*
924  * Write a metaslab to disk in the context of the specified transaction group.
925  */
926 void
927 metaslab_sync(metaslab_t *msp, uint64_t txg)
928 {
929 	vdev_t *vd = msp->ms_group->mg_vd;
930 	spa_t *spa = vd->vdev_spa;
931 	objset_t *mos = spa_meta_objset(spa);
932 	space_map_t *allocmap = &msp->ms_allocmap[txg & TXG_MASK];
933 	space_map_t *freemap = &msp->ms_freemap[txg & TXG_MASK];
934 	space_map_t *freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
935 	space_map_t *sm = &msp->ms_map;
936 	space_map_obj_t *smo = &msp->ms_smo_syncing;
937 	dmu_buf_t *db;
938 	dmu_tx_t *tx;
939 
940 	ASSERT(!vd->vdev_ishole);
941 
942 	if (allocmap->sm_space == 0 && freemap->sm_space == 0)
943 		return;
944 
945 	/*
946 	 * The only state that can actually be changing concurrently with
947 	 * metaslab_sync() is the metaslab's ms_map.  No other thread can
948 	 * be modifying this txg's allocmap, freemap, freed_map, or smo.
949 	 * Therefore, we only hold ms_lock to satify space_map ASSERTs.
950 	 * We drop it whenever we call into the DMU, because the DMU
951 	 * can call down to us (e.g. via zio_free()) at any time.
952 	 */
953 
954 	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
955 
956 	if (smo->smo_object == 0) {
957 		ASSERT(smo->smo_objsize == 0);
958 		ASSERT(smo->smo_alloc == 0);
959 		smo->smo_object = dmu_object_alloc(mos,
960 		    DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
961 		    DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
962 		ASSERT(smo->smo_object != 0);
963 		dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
964 		    (sm->sm_start >> vd->vdev_ms_shift),
965 		    sizeof (uint64_t), &smo->smo_object, tx);
966 	}
967 
968 	mutex_enter(&msp->ms_lock);
969 
970 	space_map_walk(freemap, space_map_add, freed_map);
971 
972 	if (sm->sm_loaded && spa_sync_pass(spa) == 1 && smo->smo_objsize >=
973 	    2 * sizeof (uint64_t) * avl_numnodes(&sm->sm_root)) {
974 		/*
975 		 * The in-core space map representation is twice as compact
976 		 * as the on-disk one, so it's time to condense the latter
977 		 * by generating a pure allocmap from first principles.
978 		 *
979 		 * This metaslab is 100% allocated,
980 		 * minus the content of the in-core map (sm),
981 		 * minus what's been freed this txg (freed_map),
982 		 * minus deferred frees (ms_defermap[]),
983 		 * minus allocations from txgs in the future
984 		 * (because they haven't been committed yet).
985 		 */
986 		space_map_vacate(allocmap, NULL, NULL);
987 		space_map_vacate(freemap, NULL, NULL);
988 
989 		space_map_add(allocmap, allocmap->sm_start, allocmap->sm_size);
990 
991 		space_map_walk(sm, space_map_remove, allocmap);
992 		space_map_walk(freed_map, space_map_remove, allocmap);
993 
994 		for (int t = 0; t < TXG_DEFER_SIZE; t++)
995 			space_map_walk(&msp->ms_defermap[t],
996 			    space_map_remove, allocmap);
997 
998 		for (int t = 1; t < TXG_CONCURRENT_STATES; t++)
999 			space_map_walk(&msp->ms_allocmap[(txg + t) & TXG_MASK],
1000 			    space_map_remove, allocmap);
1001 
1002 		mutex_exit(&msp->ms_lock);
1003 		space_map_truncate(smo, mos, tx);
1004 		mutex_enter(&msp->ms_lock);
1005 	}
1006 
1007 	space_map_sync(allocmap, SM_ALLOC, smo, mos, tx);
1008 	space_map_sync(freemap, SM_FREE, smo, mos, tx);
1009 
1010 	mutex_exit(&msp->ms_lock);
1011 
1012 	VERIFY(0 == dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
1013 	dmu_buf_will_dirty(db, tx);
1014 	ASSERT3U(db->db_size, >=, sizeof (*smo));
1015 	bcopy(smo, db->db_data, sizeof (*smo));
1016 	dmu_buf_rele(db, FTAG);
1017 
1018 	dmu_tx_commit(tx);
1019 }
1020 
1021 /*
1022  * Called after a transaction group has completely synced to mark
1023  * all of the metaslab's free space as usable.
1024  */
1025 void
1026 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
1027 {
1028 	space_map_obj_t *smo = &msp->ms_smo;
1029 	space_map_obj_t *smosync = &msp->ms_smo_syncing;
1030 	space_map_t *sm = &msp->ms_map;
1031 	space_map_t *freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
1032 	space_map_t *defer_map = &msp->ms_defermap[txg % TXG_DEFER_SIZE];
1033 	metaslab_group_t *mg = msp->ms_group;
1034 	vdev_t *vd = mg->mg_vd;
1035 	int64_t alloc_delta, defer_delta;
1036 
1037 	ASSERT(!vd->vdev_ishole);
1038 
1039 	mutex_enter(&msp->ms_lock);
1040 
1041 	/*
1042 	 * If this metaslab is just becoming available, initialize its
1043 	 * allocmaps and freemaps and add its capacity to the vdev.
1044 	 */
1045 	if (freed_map->sm_size == 0) {
1046 		for (int t = 0; t < TXG_SIZE; t++) {
1047 			space_map_create(&msp->ms_allocmap[t], sm->sm_start,
1048 			    sm->sm_size, sm->sm_shift, sm->sm_lock);
1049 			space_map_create(&msp->ms_freemap[t], sm->sm_start,
1050 			    sm->sm_size, sm->sm_shift, sm->sm_lock);
1051 		}
1052 
1053 		for (int t = 0; t < TXG_DEFER_SIZE; t++)
1054 			space_map_create(&msp->ms_defermap[t], sm->sm_start,
1055 			    sm->sm_size, sm->sm_shift, sm->sm_lock);
1056 
1057 		vdev_space_update(vd, 0, 0, sm->sm_size);
1058 	}
1059 
1060 	alloc_delta = smosync->smo_alloc - smo->smo_alloc;
1061 	defer_delta = freed_map->sm_space - defer_map->sm_space;
1062 
1063 	vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
1064 
1065 	ASSERT(msp->ms_allocmap[txg & TXG_MASK].sm_space == 0);
1066 	ASSERT(msp->ms_freemap[txg & TXG_MASK].sm_space == 0);
1067 
1068 	/*
1069 	 * If there's a space_map_load() in progress, wait for it to complete
1070 	 * so that we have a consistent view of the in-core space map.
1071 	 * Then, add defer_map (oldest deferred frees) to this map and
1072 	 * transfer freed_map (this txg's frees) to defer_map.
1073 	 */
1074 	space_map_load_wait(sm);
1075 	space_map_vacate(defer_map, sm->sm_loaded ? space_map_free : NULL, sm);
1076 	space_map_vacate(freed_map, space_map_add, defer_map);
1077 
1078 	*smo = *smosync;
1079 
1080 	msp->ms_deferspace += defer_delta;
1081 	ASSERT3S(msp->ms_deferspace, >=, 0);
1082 	ASSERT3S(msp->ms_deferspace, <=, sm->sm_size);
1083 	if (msp->ms_deferspace != 0) {
1084 		/*
1085 		 * Keep syncing this metaslab until all deferred frees
1086 		 * are back in circulation.
1087 		 */
1088 		vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1089 	}
1090 
1091 	/*
1092 	 * If the map is loaded but no longer active, evict it as soon as all
1093 	 * future allocations have synced.  (If we unloaded it now and then
1094 	 * loaded a moment later, the map wouldn't reflect those allocations.)
1095 	 */
1096 	if (sm->sm_loaded && (msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1097 		int evictable = 1;
1098 
1099 		for (int t = 1; t < TXG_CONCURRENT_STATES; t++)
1100 			if (msp->ms_allocmap[(txg + t) & TXG_MASK].sm_space)
1101 				evictable = 0;
1102 
1103 		if (evictable && !metaslab_debug)
1104 			space_map_unload(sm);
1105 	}
1106 
1107 	metaslab_group_sort(mg, msp, metaslab_weight(msp));
1108 
1109 	mutex_exit(&msp->ms_lock);
1110 }
1111 
1112 void
1113 metaslab_sync_reassess(metaslab_group_t *mg)
1114 {
1115 	vdev_t *vd = mg->mg_vd;
1116 	int64_t failures = mg->mg_alloc_failures;
1117 
1118 	/*
1119 	 * Re-evaluate all metaslabs which have lower offsets than the
1120 	 * bonus area.
1121 	 */
1122 	for (int m = 0; m < vd->vdev_ms_count; m++) {
1123 		metaslab_t *msp = vd->vdev_ms[m];
1124 
1125 		if (msp->ms_map.sm_start > mg->mg_bonus_area)
1126 			break;
1127 
1128 		mutex_enter(&msp->ms_lock);
1129 		metaslab_group_sort(mg, msp, metaslab_weight(msp));
1130 		mutex_exit(&msp->ms_lock);
1131 	}
1132 
1133 	atomic_add_64(&mg->mg_alloc_failures, -failures);
1134 
1135 	/*
1136 	 * Prefetch the next potential metaslabs
1137 	 */
1138 	metaslab_prefetch(mg);
1139 }
1140 
1141 static uint64_t
1142 metaslab_distance(metaslab_t *msp, dva_t *dva)
1143 {
1144 	uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
1145 	uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
1146 	uint64_t start = msp->ms_map.sm_start >> ms_shift;
1147 
1148 	if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
1149 		return (1ULL << 63);
1150 
1151 	if (offset < start)
1152 		return ((start - offset) << ms_shift);
1153 	if (offset > start)
1154 		return ((offset - start) << ms_shift);
1155 	return (0);
1156 }
1157 
1158 static uint64_t
1159 metaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize,
1160     uint64_t txg, uint64_t min_distance, dva_t *dva, int d, int flags)
1161 {
1162 	spa_t *spa = mg->mg_vd->vdev_spa;
1163 	metaslab_t *msp = NULL;
1164 	uint64_t offset = -1ULL;
1165 	avl_tree_t *t = &mg->mg_metaslab_tree;
1166 	uint64_t activation_weight;
1167 	uint64_t target_distance;
1168 	int i;
1169 
1170 	activation_weight = METASLAB_WEIGHT_PRIMARY;
1171 	for (i = 0; i < d; i++) {
1172 		if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
1173 			activation_weight = METASLAB_WEIGHT_SECONDARY;
1174 			break;
1175 		}
1176 	}
1177 
1178 	for (;;) {
1179 		boolean_t was_active;
1180 
1181 		mutex_enter(&mg->mg_lock);
1182 		for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
1183 			if (msp->ms_weight < asize) {
1184 				spa_dbgmsg(spa, "%s: failed to meet weight "
1185 				    "requirement: vdev %llu, txg %llu, mg %p, "
1186 				    "msp %p, psize %llu, asize %llu, "
1187 				    "failures %llu, weight %llu",
1188 				    spa_name(spa), mg->mg_vd->vdev_id, txg,
1189 				    mg, msp, psize, asize,
1190 				    mg->mg_alloc_failures, msp->ms_weight);
1191 				mutex_exit(&mg->mg_lock);
1192 				return (-1ULL);
1193 			}
1194 			was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
1195 			if (activation_weight == METASLAB_WEIGHT_PRIMARY)
1196 				break;
1197 
1198 			target_distance = min_distance +
1199 			    (msp->ms_smo.smo_alloc ? 0 : min_distance >> 1);
1200 
1201 			for (i = 0; i < d; i++)
1202 				if (metaslab_distance(msp, &dva[i]) <
1203 				    target_distance)
1204 					break;
1205 			if (i == d)
1206 				break;
1207 		}
1208 		mutex_exit(&mg->mg_lock);
1209 		if (msp == NULL)
1210 			return (-1ULL);
1211 
1212 		/*
1213 		 * If we've already reached the allowable number of failed
1214 		 * allocation attempts on this metaslab group then we
1215 		 * consider skipping it. We skip it only if we're allowed
1216 		 * to "fast" gang, the physical size is larger than
1217 		 * a gang block, and we're attempting to allocate from
1218 		 * the primary metaslab.
1219 		 */
1220 		if (mg->mg_alloc_failures > zfs_mg_alloc_failures &&
1221 		    CAN_FASTGANG(flags) && psize > SPA_GANGBLOCKSIZE &&
1222 		    activation_weight == METASLAB_WEIGHT_PRIMARY) {
1223 			spa_dbgmsg(spa, "%s: skipping metaslab group: "
1224 			    "vdev %llu, txg %llu, mg %p, psize %llu, "
1225 			    "asize %llu, failures %llu", spa_name(spa),
1226 			    mg->mg_vd->vdev_id, txg, mg, psize, asize,
1227 			    mg->mg_alloc_failures);
1228 			return (-1ULL);
1229 		}
1230 
1231 		mutex_enter(&msp->ms_lock);
1232 
1233 		/*
1234 		 * Ensure that the metaslab we have selected is still
1235 		 * capable of handling our request. It's possible that
1236 		 * another thread may have changed the weight while we
1237 		 * were blocked on the metaslab lock.
1238 		 */
1239 		if (msp->ms_weight < asize || (was_active &&
1240 		    !(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
1241 		    activation_weight == METASLAB_WEIGHT_PRIMARY)) {
1242 			mutex_exit(&msp->ms_lock);
1243 			continue;
1244 		}
1245 
1246 		if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
1247 		    activation_weight == METASLAB_WEIGHT_PRIMARY) {
1248 			metaslab_passivate(msp,
1249 			    msp->ms_weight & ~METASLAB_ACTIVE_MASK);
1250 			mutex_exit(&msp->ms_lock);
1251 			continue;
1252 		}
1253 
1254 		if (metaslab_activate(msp, activation_weight) != 0) {
1255 			mutex_exit(&msp->ms_lock);
1256 			continue;
1257 		}
1258 
1259 		if ((offset = space_map_alloc(&msp->ms_map, asize)) != -1ULL)
1260 			break;
1261 
1262 		atomic_inc_64(&mg->mg_alloc_failures);
1263 
1264 		metaslab_passivate(msp, space_map_maxsize(&msp->ms_map));
1265 
1266 		mutex_exit(&msp->ms_lock);
1267 	}
1268 
1269 	if (msp->ms_allocmap[txg & TXG_MASK].sm_space == 0)
1270 		vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
1271 
1272 	space_map_add(&msp->ms_allocmap[txg & TXG_MASK], offset, asize);
1273 
1274 	mutex_exit(&msp->ms_lock);
1275 
1276 	return (offset);
1277 }
1278 
1279 /*
1280  * Allocate a block for the specified i/o.
1281  */
1282 static int
1283 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
1284     dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
1285 {
1286 	metaslab_group_t *mg, *rotor;
1287 	vdev_t *vd;
1288 	int dshift = 3;
1289 	int all_zero;
1290 	int zio_lock = B_FALSE;
1291 	boolean_t allocatable;
1292 	uint64_t offset = -1ULL;
1293 	uint64_t asize;
1294 	uint64_t distance;
1295 
1296 	ASSERT(!DVA_IS_VALID(&dva[d]));
1297 
1298 	/*
1299 	 * For testing, make some blocks above a certain size be gang blocks.
1300 	 */
1301 	if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0)
1302 		return (ENOSPC);
1303 
1304 	/*
1305 	 * Start at the rotor and loop through all mgs until we find something.
1306 	 * Note that there's no locking on mc_rotor or mc_aliquot because
1307 	 * nothing actually breaks if we miss a few updates -- we just won't
1308 	 * allocate quite as evenly.  It all balances out over time.
1309 	 *
1310 	 * If we are doing ditto or log blocks, try to spread them across
1311 	 * consecutive vdevs.  If we're forced to reuse a vdev before we've
1312 	 * allocated all of our ditto blocks, then try and spread them out on
1313 	 * that vdev as much as possible.  If it turns out to not be possible,
1314 	 * gradually lower our standards until anything becomes acceptable.
1315 	 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
1316 	 * gives us hope of containing our fault domains to something we're
1317 	 * able to reason about.  Otherwise, any two top-level vdev failures
1318 	 * will guarantee the loss of data.  With consecutive allocation,
1319 	 * only two adjacent top-level vdev failures will result in data loss.
1320 	 *
1321 	 * If we are doing gang blocks (hintdva is non-NULL), try to keep
1322 	 * ourselves on the same vdev as our gang block header.  That
1323 	 * way, we can hope for locality in vdev_cache, plus it makes our
1324 	 * fault domains something tractable.
1325 	 */
1326 	if (hintdva) {
1327 		vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
1328 
1329 		/*
1330 		 * It's possible the vdev we're using as the hint no
1331 		 * longer exists (i.e. removed). Consult the rotor when
1332 		 * all else fails.
1333 		 */
1334 		if (vd != NULL) {
1335 			mg = vd->vdev_mg;
1336 
1337 			if (flags & METASLAB_HINTBP_AVOID &&
1338 			    mg->mg_next != NULL)
1339 				mg = mg->mg_next;
1340 		} else {
1341 			mg = mc->mc_rotor;
1342 		}
1343 	} else if (d != 0) {
1344 		vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
1345 		mg = vd->vdev_mg->mg_next;
1346 	} else {
1347 		mg = mc->mc_rotor;
1348 	}
1349 
1350 	/*
1351 	 * If the hint put us into the wrong metaslab class, or into a
1352 	 * metaslab group that has been passivated, just follow the rotor.
1353 	 */
1354 	if (mg->mg_class != mc || mg->mg_activation_count <= 0)
1355 		mg = mc->mc_rotor;
1356 
1357 	rotor = mg;
1358 top:
1359 	all_zero = B_TRUE;
1360 	do {
1361 		ASSERT(mg->mg_activation_count == 1);
1362 
1363 		vd = mg->mg_vd;
1364 
1365 		/*
1366 		 * Don't allocate from faulted devices.
1367 		 */
1368 		if (zio_lock) {
1369 			spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
1370 			allocatable = vdev_allocatable(vd);
1371 			spa_config_exit(spa, SCL_ZIO, FTAG);
1372 		} else {
1373 			allocatable = vdev_allocatable(vd);
1374 		}
1375 		if (!allocatable)
1376 			goto next;
1377 
1378 		/*
1379 		 * Avoid writing single-copy data to a failing vdev
1380 		 */
1381 		if ((vd->vdev_stat.vs_write_errors > 0 ||
1382 		    vd->vdev_state < VDEV_STATE_HEALTHY) &&
1383 		    d == 0 && dshift == 3) {
1384 			all_zero = B_FALSE;
1385 			goto next;
1386 		}
1387 
1388 		ASSERT(mg->mg_class == mc);
1389 
1390 		distance = vd->vdev_asize >> dshift;
1391 		if (distance <= (1ULL << vd->vdev_ms_shift))
1392 			distance = 0;
1393 		else
1394 			all_zero = B_FALSE;
1395 
1396 		asize = vdev_psize_to_asize(vd, psize);
1397 		ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
1398 
1399 		offset = metaslab_group_alloc(mg, psize, asize, txg, distance,
1400 		    dva, d, flags);
1401 		if (offset != -1ULL) {
1402 			/*
1403 			 * If we've just selected this metaslab group,
1404 			 * figure out whether the corresponding vdev is
1405 			 * over- or under-used relative to the pool,
1406 			 * and set an allocation bias to even it out.
1407 			 */
1408 			if (mc->mc_aliquot == 0) {
1409 				vdev_stat_t *vs = &vd->vdev_stat;
1410 				int64_t vu, cu;
1411 
1412 				vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
1413 				cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
1414 
1415 				/*
1416 				 * Calculate how much more or less we should
1417 				 * try to allocate from this device during
1418 				 * this iteration around the rotor.
1419 				 * For example, if a device is 80% full
1420 				 * and the pool is 20% full then we should
1421 				 * reduce allocations by 60% on this device.
1422 				 *
1423 				 * mg_bias = (20 - 80) * 512K / 100 = -307K
1424 				 *
1425 				 * This reduces allocations by 307K for this
1426 				 * iteration.
1427 				 */
1428 				mg->mg_bias = ((cu - vu) *
1429 				    (int64_t)mg->mg_aliquot) / 100;
1430 			}
1431 
1432 			if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
1433 			    mg->mg_aliquot + mg->mg_bias) {
1434 				mc->mc_rotor = mg->mg_next;
1435 				mc->mc_aliquot = 0;
1436 			}
1437 
1438 			DVA_SET_VDEV(&dva[d], vd->vdev_id);
1439 			DVA_SET_OFFSET(&dva[d], offset);
1440 			DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
1441 			DVA_SET_ASIZE(&dva[d], asize);
1442 
1443 			return (0);
1444 		}
1445 next:
1446 		mc->mc_rotor = mg->mg_next;
1447 		mc->mc_aliquot = 0;
1448 	} while ((mg = mg->mg_next) != rotor);
1449 
1450 	if (!all_zero) {
1451 		dshift++;
1452 		ASSERT(dshift < 64);
1453 		goto top;
1454 	}
1455 
1456 	if (!allocatable && !zio_lock) {
1457 		dshift = 3;
1458 		zio_lock = B_TRUE;
1459 		goto top;
1460 	}
1461 
1462 	bzero(&dva[d], sizeof (dva_t));
1463 
1464 	return (ENOSPC);
1465 }
1466 
1467 /*
1468  * Free the block represented by DVA in the context of the specified
1469  * transaction group.
1470  */
1471 static void
1472 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
1473 {
1474 	uint64_t vdev = DVA_GET_VDEV(dva);
1475 	uint64_t offset = DVA_GET_OFFSET(dva);
1476 	uint64_t size = DVA_GET_ASIZE(dva);
1477 	vdev_t *vd;
1478 	metaslab_t *msp;
1479 
1480 	ASSERT(DVA_IS_VALID(dva));
1481 
1482 	if (txg > spa_freeze_txg(spa))
1483 		return;
1484 
1485 	if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
1486 	    (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
1487 		cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
1488 		    (u_longlong_t)vdev, (u_longlong_t)offset);
1489 		ASSERT(0);
1490 		return;
1491 	}
1492 
1493 	msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1494 
1495 	if (DVA_GET_GANG(dva))
1496 		size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1497 
1498 	mutex_enter(&msp->ms_lock);
1499 
1500 	if (now) {
1501 		space_map_remove(&msp->ms_allocmap[txg & TXG_MASK],
1502 		    offset, size);
1503 		space_map_free(&msp->ms_map, offset, size);
1504 	} else {
1505 		if (msp->ms_freemap[txg & TXG_MASK].sm_space == 0)
1506 			vdev_dirty(vd, VDD_METASLAB, msp, txg);
1507 		space_map_add(&msp->ms_freemap[txg & TXG_MASK], offset, size);
1508 	}
1509 
1510 	mutex_exit(&msp->ms_lock);
1511 }
1512 
1513 /*
1514  * Intent log support: upon opening the pool after a crash, notify the SPA
1515  * of blocks that the intent log has allocated for immediate write, but
1516  * which are still considered free by the SPA because the last transaction
1517  * group didn't commit yet.
1518  */
1519 static int
1520 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
1521 {
1522 	uint64_t vdev = DVA_GET_VDEV(dva);
1523 	uint64_t offset = DVA_GET_OFFSET(dva);
1524 	uint64_t size = DVA_GET_ASIZE(dva);
1525 	vdev_t *vd;
1526 	metaslab_t *msp;
1527 	int error = 0;
1528 
1529 	ASSERT(DVA_IS_VALID(dva));
1530 
1531 	if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
1532 	    (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
1533 		return (ENXIO);
1534 
1535 	msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1536 
1537 	if (DVA_GET_GANG(dva))
1538 		size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1539 
1540 	mutex_enter(&msp->ms_lock);
1541 
1542 	if ((txg != 0 && spa_writeable(spa)) || !msp->ms_map.sm_loaded)
1543 		error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
1544 
1545 	if (error == 0 && !space_map_contains(&msp->ms_map, offset, size))
1546 		error = ENOENT;
1547 
1548 	if (error || txg == 0) {	/* txg == 0 indicates dry run */
1549 		mutex_exit(&msp->ms_lock);
1550 		return (error);
1551 	}
1552 
1553 	space_map_claim(&msp->ms_map, offset, size);
1554 
1555 	if (spa_writeable(spa)) {	/* don't dirty if we're zdb(1M) */
1556 		if (msp->ms_allocmap[txg & TXG_MASK].sm_space == 0)
1557 			vdev_dirty(vd, VDD_METASLAB, msp, txg);
1558 		space_map_add(&msp->ms_allocmap[txg & TXG_MASK], offset, size);
1559 	}
1560 
1561 	mutex_exit(&msp->ms_lock);
1562 
1563 	return (0);
1564 }
1565 
1566 int
1567 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
1568     int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
1569 {
1570 	dva_t *dva = bp->blk_dva;
1571 	dva_t *hintdva = hintbp->blk_dva;
1572 	int error = 0;
1573 
1574 	ASSERT(bp->blk_birth == 0);
1575 	ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
1576 
1577 	spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
1578 
1579 	if (mc->mc_rotor == NULL) {	/* no vdevs in this class */
1580 		spa_config_exit(spa, SCL_ALLOC, FTAG);
1581 		return (ENOSPC);
1582 	}
1583 
1584 	ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
1585 	ASSERT(BP_GET_NDVAS(bp) == 0);
1586 	ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
1587 
1588 	for (int d = 0; d < ndvas; d++) {
1589 		error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
1590 		    txg, flags);
1591 		if (error) {
1592 			for (d--; d >= 0; d--) {
1593 				metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
1594 				bzero(&dva[d], sizeof (dva_t));
1595 			}
1596 			spa_config_exit(spa, SCL_ALLOC, FTAG);
1597 			return (error);
1598 		}
1599 	}
1600 	ASSERT(error == 0);
1601 	ASSERT(BP_GET_NDVAS(bp) == ndvas);
1602 
1603 	spa_config_exit(spa, SCL_ALLOC, FTAG);
1604 
1605 	BP_SET_BIRTH(bp, txg, txg);
1606 
1607 	return (0);
1608 }
1609 
1610 void
1611 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
1612 {
1613 	const dva_t *dva = bp->blk_dva;
1614 	int ndvas = BP_GET_NDVAS(bp);
1615 
1616 	ASSERT(!BP_IS_HOLE(bp));
1617 	ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
1618 
1619 	spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
1620 
1621 	for (int d = 0; d < ndvas; d++)
1622 		metaslab_free_dva(spa, &dva[d], txg, now);
1623 
1624 	spa_config_exit(spa, SCL_FREE, FTAG);
1625 }
1626 
1627 int
1628 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
1629 {
1630 	const dva_t *dva = bp->blk_dva;
1631 	int ndvas = BP_GET_NDVAS(bp);
1632 	int error = 0;
1633 
1634 	ASSERT(!BP_IS_HOLE(bp));
1635 
1636 	if (txg != 0) {
1637 		/*
1638 		 * First do a dry run to make sure all DVAs are claimable,
1639 		 * so we don't have to unwind from partial failures below.
1640 		 */
1641 		if ((error = metaslab_claim(spa, bp, 0)) != 0)
1642 			return (error);
1643 	}
1644 
1645 	spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
1646 
1647 	for (int d = 0; d < ndvas; d++)
1648 		if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
1649 			break;
1650 
1651 	spa_config_exit(spa, SCL_ALLOC, FTAG);
1652 
1653 	ASSERT(error == 0 || txg == 0);
1654 
1655 	return (error);
1656 }
1657