xref: /titanic_52/usr/src/uts/common/fs/zfs/metaslab.c (revision 9525b14bcdeb5b5f6f95ab27c2f48f18bd2ec829)
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 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
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 uint64_t metaslab_aliquot = 512ULL << 10;
35 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1;	/* force gang blocks */
36 
37 /*
38  * Metaslab debugging: when set, keeps all space maps in core to verify frees.
39  */
40 static int metaslab_debug = 0;
41 
42 /*
43  * Minimum size which forces the dynamic allocator to change
44  * it's allocation strategy. Once the space map cannot satisfy
45  * an allocation of this size then it switches to using more
46  * aggressive strategy (i.e search by size rather than offset).
47  */
48 uint64_t metaslab_df_alloc_threshold = SPA_MAXBLOCKSIZE;
49 
50 /*
51  * The minimum free space, in percent, which must be available
52  * in a space map to continue allocations in a first-fit fashion.
53  * Once the space_map's free space drops below this level we dynamically
54  * switch to using best-fit allocations.
55  */
56 int metaslab_df_free_pct = 30;
57 
58 /*
59  * ==========================================================================
60  * Metaslab classes
61  * ==========================================================================
62  */
63 metaslab_class_t *
64 metaslab_class_create(spa_t *spa, space_map_ops_t *ops)
65 {
66 	metaslab_class_t *mc;
67 
68 	mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
69 
70 	mc->mc_spa = spa;
71 	mc->mc_rotor = NULL;
72 	mc->mc_ops = ops;
73 
74 	return (mc);
75 }
76 
77 void
78 metaslab_class_destroy(metaslab_class_t *mc)
79 {
80 	ASSERT(mc->mc_rotor == NULL);
81 	ASSERT(mc->mc_alloc == 0);
82 	ASSERT(mc->mc_deferred == 0);
83 	ASSERT(mc->mc_space == 0);
84 	ASSERT(mc->mc_dspace == 0);
85 
86 	kmem_free(mc, sizeof (metaslab_class_t));
87 }
88 
89 int
90 metaslab_class_validate(metaslab_class_t *mc)
91 {
92 	metaslab_group_t *mg;
93 	vdev_t *vd;
94 
95 	/*
96 	 * Must hold one of the spa_config locks.
97 	 */
98 	ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
99 	    spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
100 
101 	if ((mg = mc->mc_rotor) == NULL)
102 		return (0);
103 
104 	do {
105 		vd = mg->mg_vd;
106 		ASSERT(vd->vdev_mg != NULL);
107 		ASSERT3P(vd->vdev_top, ==, vd);
108 		ASSERT3P(mg->mg_class, ==, mc);
109 		ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
110 	} while ((mg = mg->mg_next) != mc->mc_rotor);
111 
112 	return (0);
113 }
114 
115 void
116 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
117     int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
118 {
119 	atomic_add_64(&mc->mc_alloc, alloc_delta);
120 	atomic_add_64(&mc->mc_deferred, defer_delta);
121 	atomic_add_64(&mc->mc_space, space_delta);
122 	atomic_add_64(&mc->mc_dspace, dspace_delta);
123 }
124 
125 uint64_t
126 metaslab_class_get_alloc(metaslab_class_t *mc)
127 {
128 	return (mc->mc_alloc);
129 }
130 
131 uint64_t
132 metaslab_class_get_deferred(metaslab_class_t *mc)
133 {
134 	return (mc->mc_deferred);
135 }
136 
137 uint64_t
138 metaslab_class_get_space(metaslab_class_t *mc)
139 {
140 	return (mc->mc_space);
141 }
142 
143 uint64_t
144 metaslab_class_get_dspace(metaslab_class_t *mc)
145 {
146 	return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
147 }
148 
149 /*
150  * ==========================================================================
151  * Metaslab groups
152  * ==========================================================================
153  */
154 static int
155 metaslab_compare(const void *x1, const void *x2)
156 {
157 	const metaslab_t *m1 = x1;
158 	const metaslab_t *m2 = x2;
159 
160 	if (m1->ms_weight < m2->ms_weight)
161 		return (1);
162 	if (m1->ms_weight > m2->ms_weight)
163 		return (-1);
164 
165 	/*
166 	 * If the weights are identical, use the offset to force uniqueness.
167 	 */
168 	if (m1->ms_map.sm_start < m2->ms_map.sm_start)
169 		return (-1);
170 	if (m1->ms_map.sm_start > m2->ms_map.sm_start)
171 		return (1);
172 
173 	ASSERT3P(m1, ==, m2);
174 
175 	return (0);
176 }
177 
178 metaslab_group_t *
179 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
180 {
181 	metaslab_group_t *mg;
182 
183 	mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
184 	mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
185 	avl_create(&mg->mg_metaslab_tree, metaslab_compare,
186 	    sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
187 	mg->mg_vd = vd;
188 	mg->mg_class = mc;
189 	mg->mg_activation_count = 0;
190 
191 	return (mg);
192 }
193 
194 void
195 metaslab_group_destroy(metaslab_group_t *mg)
196 {
197 	ASSERT(mg->mg_prev == NULL);
198 	ASSERT(mg->mg_next == NULL);
199 	/*
200 	 * We may have gone below zero with the activation count
201 	 * either because we never activated in the first place or
202 	 * because we're done, and possibly removing the vdev.
203 	 */
204 	ASSERT(mg->mg_activation_count <= 0);
205 
206 	avl_destroy(&mg->mg_metaslab_tree);
207 	mutex_destroy(&mg->mg_lock);
208 	kmem_free(mg, sizeof (metaslab_group_t));
209 }
210 
211 void
212 metaslab_group_activate(metaslab_group_t *mg)
213 {
214 	metaslab_class_t *mc = mg->mg_class;
215 	metaslab_group_t *mgprev, *mgnext;
216 
217 	ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
218 
219 	ASSERT(mc->mc_rotor != mg);
220 	ASSERT(mg->mg_prev == NULL);
221 	ASSERT(mg->mg_next == NULL);
222 	ASSERT(mg->mg_activation_count <= 0);
223 
224 	if (++mg->mg_activation_count <= 0)
225 		return;
226 
227 	mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
228 
229 	if ((mgprev = mc->mc_rotor) == NULL) {
230 		mg->mg_prev = mg;
231 		mg->mg_next = mg;
232 	} else {
233 		mgnext = mgprev->mg_next;
234 		mg->mg_prev = mgprev;
235 		mg->mg_next = mgnext;
236 		mgprev->mg_next = mg;
237 		mgnext->mg_prev = mg;
238 	}
239 	mc->mc_rotor = mg;
240 }
241 
242 void
243 metaslab_group_passivate(metaslab_group_t *mg)
244 {
245 	metaslab_class_t *mc = mg->mg_class;
246 	metaslab_group_t *mgprev, *mgnext;
247 
248 	ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
249 
250 	if (--mg->mg_activation_count != 0) {
251 		ASSERT(mc->mc_rotor != mg);
252 		ASSERT(mg->mg_prev == NULL);
253 		ASSERT(mg->mg_next == NULL);
254 		ASSERT(mg->mg_activation_count < 0);
255 		return;
256 	}
257 
258 	mgprev = mg->mg_prev;
259 	mgnext = mg->mg_next;
260 
261 	if (mg == mgnext) {
262 		mc->mc_rotor = NULL;
263 	} else {
264 		mc->mc_rotor = mgnext;
265 		mgprev->mg_next = mgnext;
266 		mgnext->mg_prev = mgprev;
267 	}
268 
269 	mg->mg_prev = NULL;
270 	mg->mg_next = NULL;
271 }
272 
273 static void
274 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
275 {
276 	mutex_enter(&mg->mg_lock);
277 	ASSERT(msp->ms_group == NULL);
278 	msp->ms_group = mg;
279 	msp->ms_weight = 0;
280 	avl_add(&mg->mg_metaslab_tree, msp);
281 	mutex_exit(&mg->mg_lock);
282 }
283 
284 static void
285 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
286 {
287 	mutex_enter(&mg->mg_lock);
288 	ASSERT(msp->ms_group == mg);
289 	avl_remove(&mg->mg_metaslab_tree, msp);
290 	msp->ms_group = NULL;
291 	mutex_exit(&mg->mg_lock);
292 }
293 
294 static void
295 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
296 {
297 	/*
298 	 * Although in principle the weight can be any value, in
299 	 * practice we do not use values in the range [1, 510].
300 	 */
301 	ASSERT(weight >= SPA_MINBLOCKSIZE-1 || weight == 0);
302 	ASSERT(MUTEX_HELD(&msp->ms_lock));
303 
304 	mutex_enter(&mg->mg_lock);
305 	ASSERT(msp->ms_group == mg);
306 	avl_remove(&mg->mg_metaslab_tree, msp);
307 	msp->ms_weight = weight;
308 	avl_add(&mg->mg_metaslab_tree, msp);
309 	mutex_exit(&mg->mg_lock);
310 }
311 
312 /*
313  * This is a helper function that can be used by the allocator to find
314  * a suitable block to allocate. This will search the specified AVL
315  * tree looking for a block that matches the specified criteria.
316  */
317 static uint64_t
318 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
319     uint64_t align)
320 {
321 	space_seg_t *ss, ssearch;
322 	avl_index_t where;
323 
324 	ssearch.ss_start = *cursor;
325 	ssearch.ss_end = *cursor + size;
326 
327 	ss = avl_find(t, &ssearch, &where);
328 	if (ss == NULL)
329 		ss = avl_nearest(t, where, AVL_AFTER);
330 
331 	while (ss != NULL) {
332 		uint64_t offset = P2ROUNDUP(ss->ss_start, align);
333 
334 		if (offset + size <= ss->ss_end) {
335 			*cursor = offset + size;
336 			return (offset);
337 		}
338 		ss = AVL_NEXT(t, ss);
339 	}
340 
341 	/*
342 	 * If we know we've searched the whole map (*cursor == 0), give up.
343 	 * Otherwise, reset the cursor to the beginning and try again.
344 	 */
345 	if (*cursor == 0)
346 		return (-1ULL);
347 
348 	*cursor = 0;
349 	return (metaslab_block_picker(t, cursor, size, align));
350 }
351 
352 /*
353  * ==========================================================================
354  * The first-fit block allocator
355  * ==========================================================================
356  */
357 static void
358 metaslab_ff_load(space_map_t *sm)
359 {
360 	ASSERT(sm->sm_ppd == NULL);
361 	sm->sm_ppd = kmem_zalloc(64 * sizeof (uint64_t), KM_SLEEP);
362 	sm->sm_pp_root = NULL;
363 }
364 
365 static void
366 metaslab_ff_unload(space_map_t *sm)
367 {
368 	kmem_free(sm->sm_ppd, 64 * sizeof (uint64_t));
369 	sm->sm_ppd = NULL;
370 }
371 
372 static uint64_t
373 metaslab_ff_alloc(space_map_t *sm, uint64_t size)
374 {
375 	avl_tree_t *t = &sm->sm_root;
376 	uint64_t align = size & -size;
377 	uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
378 
379 	return (metaslab_block_picker(t, cursor, size, align));
380 }
381 
382 /* ARGSUSED */
383 static void
384 metaslab_ff_claim(space_map_t *sm, uint64_t start, uint64_t size)
385 {
386 	/* No need to update cursor */
387 }
388 
389 /* ARGSUSED */
390 static void
391 metaslab_ff_free(space_map_t *sm, uint64_t start, uint64_t size)
392 {
393 	/* No need to update cursor */
394 }
395 
396 static space_map_ops_t metaslab_ff_ops = {
397 	metaslab_ff_load,
398 	metaslab_ff_unload,
399 	metaslab_ff_alloc,
400 	metaslab_ff_claim,
401 	metaslab_ff_free,
402 	NULL	/* maxsize */
403 };
404 
405 /*
406  * Dynamic block allocator -
407  * Uses the first fit allocation scheme until space get low and then
408  * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
409  * and metaslab_df_free_pct to determine when to switch the allocation scheme.
410  */
411 
412 uint64_t
413 metaslab_df_maxsize(space_map_t *sm)
414 {
415 	avl_tree_t *t = sm->sm_pp_root;
416 	space_seg_t *ss;
417 
418 	if (t == NULL || (ss = avl_last(t)) == NULL)
419 		return (0ULL);
420 
421 	return (ss->ss_end - ss->ss_start);
422 }
423 
424 static int
425 metaslab_df_seg_compare(const void *x1, const void *x2)
426 {
427 	const space_seg_t *s1 = x1;
428 	const space_seg_t *s2 = x2;
429 	uint64_t ss_size1 = s1->ss_end - s1->ss_start;
430 	uint64_t ss_size2 = s2->ss_end - s2->ss_start;
431 
432 	if (ss_size1 < ss_size2)
433 		return (-1);
434 	if (ss_size1 > ss_size2)
435 		return (1);
436 
437 	if (s1->ss_start < s2->ss_start)
438 		return (-1);
439 	if (s1->ss_start > s2->ss_start)
440 		return (1);
441 
442 	return (0);
443 }
444 
445 static void
446 metaslab_df_load(space_map_t *sm)
447 {
448 	space_seg_t *ss;
449 
450 	ASSERT(sm->sm_ppd == NULL);
451 	sm->sm_ppd = kmem_zalloc(64 * sizeof (uint64_t), KM_SLEEP);
452 
453 	sm->sm_pp_root = kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
454 	avl_create(sm->sm_pp_root, metaslab_df_seg_compare,
455 	    sizeof (space_seg_t), offsetof(struct space_seg, ss_pp_node));
456 
457 	for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
458 		avl_add(sm->sm_pp_root, ss);
459 }
460 
461 static void
462 metaslab_df_unload(space_map_t *sm)
463 {
464 	void *cookie = NULL;
465 
466 	kmem_free(sm->sm_ppd, 64 * sizeof (uint64_t));
467 	sm->sm_ppd = NULL;
468 
469 	while (avl_destroy_nodes(sm->sm_pp_root, &cookie) != NULL) {
470 		/* tear down the tree */
471 	}
472 
473 	avl_destroy(sm->sm_pp_root);
474 	kmem_free(sm->sm_pp_root, sizeof (avl_tree_t));
475 	sm->sm_pp_root = NULL;
476 }
477 
478 static uint64_t
479 metaslab_df_alloc(space_map_t *sm, uint64_t size)
480 {
481 	avl_tree_t *t = &sm->sm_root;
482 	uint64_t align = size & -size;
483 	uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
484 	uint64_t max_size = metaslab_df_maxsize(sm);
485 	int free_pct = sm->sm_space * 100 / sm->sm_size;
486 
487 	ASSERT(MUTEX_HELD(sm->sm_lock));
488 	ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
489 
490 	if (max_size < size)
491 		return (-1ULL);
492 
493 	/*
494 	 * If we're running low on space switch to using the size
495 	 * sorted AVL tree (best-fit).
496 	 */
497 	if (max_size < metaslab_df_alloc_threshold ||
498 	    free_pct < metaslab_df_free_pct) {
499 		t = sm->sm_pp_root;
500 		*cursor = 0;
501 	}
502 
503 	return (metaslab_block_picker(t, cursor, size, 1ULL));
504 }
505 
506 /* ARGSUSED */
507 static void
508 metaslab_df_claim(space_map_t *sm, uint64_t start, uint64_t size)
509 {
510 	/* No need to update cursor */
511 }
512 
513 /* ARGSUSED */
514 static void
515 metaslab_df_free(space_map_t *sm, uint64_t start, uint64_t size)
516 {
517 	/* No need to update cursor */
518 }
519 
520 static space_map_ops_t metaslab_df_ops = {
521 	metaslab_df_load,
522 	metaslab_df_unload,
523 	metaslab_df_alloc,
524 	metaslab_df_claim,
525 	metaslab_df_free,
526 	metaslab_df_maxsize
527 };
528 
529 space_map_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
530 
531 /*
532  * ==========================================================================
533  * Metaslabs
534  * ==========================================================================
535  */
536 metaslab_t *
537 metaslab_init(metaslab_group_t *mg, space_map_obj_t *smo,
538 	uint64_t start, uint64_t size, uint64_t txg)
539 {
540 	vdev_t *vd = mg->mg_vd;
541 	metaslab_t *msp;
542 
543 	msp = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
544 	mutex_init(&msp->ms_lock, NULL, MUTEX_DEFAULT, NULL);
545 
546 	msp->ms_smo_syncing = *smo;
547 
548 	/*
549 	 * We create the main space map here, but we don't create the
550 	 * allocmaps and freemaps until metaslab_sync_done().  This serves
551 	 * two purposes: it allows metaslab_sync_done() to detect the
552 	 * addition of new space; and for debugging, it ensures that we'd
553 	 * data fault on any attempt to use this metaslab before it's ready.
554 	 */
555 	space_map_create(&msp->ms_map, start, size,
556 	    vd->vdev_ashift, &msp->ms_lock);
557 
558 	metaslab_group_add(mg, msp);
559 
560 	if (metaslab_debug && smo->smo_object != 0) {
561 		mutex_enter(&msp->ms_lock);
562 		VERIFY(space_map_load(&msp->ms_map, mg->mg_class->mc_ops,
563 		    SM_FREE, smo, spa_meta_objset(vd->vdev_spa)) == 0);
564 		mutex_exit(&msp->ms_lock);
565 	}
566 
567 	/*
568 	 * If we're opening an existing pool (txg == 0) or creating
569 	 * a new one (txg == TXG_INITIAL), all space is available now.
570 	 * If we're adding space to an existing pool, the new space
571 	 * does not become available until after this txg has synced.
572 	 */
573 	if (txg <= TXG_INITIAL)
574 		metaslab_sync_done(msp, 0);
575 
576 	if (txg != 0) {
577 		vdev_dirty(vd, 0, NULL, txg);
578 		vdev_dirty(vd, VDD_METASLAB, msp, txg);
579 	}
580 
581 	return (msp);
582 }
583 
584 void
585 metaslab_fini(metaslab_t *msp)
586 {
587 	metaslab_group_t *mg = msp->ms_group;
588 
589 	vdev_space_update(mg->mg_vd,
590 	    -msp->ms_smo.smo_alloc, 0, -msp->ms_map.sm_size);
591 
592 	metaslab_group_remove(mg, msp);
593 
594 	mutex_enter(&msp->ms_lock);
595 
596 	space_map_unload(&msp->ms_map);
597 	space_map_destroy(&msp->ms_map);
598 
599 	for (int t = 0; t < TXG_SIZE; t++) {
600 		space_map_destroy(&msp->ms_allocmap[t]);
601 		space_map_destroy(&msp->ms_freemap[t]);
602 	}
603 
604 	for (int t = 0; t < TXG_DEFER_SIZE; t++)
605 		space_map_destroy(&msp->ms_defermap[t]);
606 
607 	ASSERT3S(msp->ms_deferspace, ==, 0);
608 
609 	mutex_exit(&msp->ms_lock);
610 	mutex_destroy(&msp->ms_lock);
611 
612 	kmem_free(msp, sizeof (metaslab_t));
613 }
614 
615 #define	METASLAB_WEIGHT_PRIMARY		(1ULL << 63)
616 #define	METASLAB_WEIGHT_SECONDARY	(1ULL << 62)
617 #define	METASLAB_ACTIVE_MASK		\
618 	(METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
619 #define	METASLAB_SMO_BONUS_MULTIPLIER	2
620 
621 static uint64_t
622 metaslab_weight(metaslab_t *msp)
623 {
624 	metaslab_group_t *mg = msp->ms_group;
625 	space_map_t *sm = &msp->ms_map;
626 	space_map_obj_t *smo = &msp->ms_smo;
627 	vdev_t *vd = mg->mg_vd;
628 	uint64_t weight, space;
629 
630 	ASSERT(MUTEX_HELD(&msp->ms_lock));
631 
632 	/*
633 	 * The baseline weight is the metaslab's free space.
634 	 */
635 	space = sm->sm_size - smo->smo_alloc;
636 	weight = space;
637 
638 	/*
639 	 * Modern disks have uniform bit density and constant angular velocity.
640 	 * Therefore, the outer recording zones are faster (higher bandwidth)
641 	 * than the inner zones by the ratio of outer to inner track diameter,
642 	 * which is typically around 2:1.  We account for this by assigning
643 	 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
644 	 * In effect, this means that we'll select the metaslab with the most
645 	 * free bandwidth rather than simply the one with the most free space.
646 	 */
647 	weight = 2 * weight -
648 	    ((sm->sm_start >> vd->vdev_ms_shift) * weight) / vd->vdev_ms_count;
649 	ASSERT(weight >= space && weight <= 2 * space);
650 
651 	/*
652 	 * For locality, assign higher weight to metaslabs we've used before.
653 	 */
654 	if (smo->smo_object != 0)
655 		weight *= METASLAB_SMO_BONUS_MULTIPLIER;
656 	ASSERT(weight >= space &&
657 	    weight <= 2 * METASLAB_SMO_BONUS_MULTIPLIER * space);
658 
659 	/*
660 	 * If this metaslab is one we're actively using, adjust its weight to
661 	 * make it preferable to any inactive metaslab so we'll polish it off.
662 	 */
663 	weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
664 
665 	return (weight);
666 }
667 
668 static int
669 metaslab_activate(metaslab_t *msp, uint64_t activation_weight, uint64_t size)
670 {
671 	space_map_t *sm = &msp->ms_map;
672 	space_map_ops_t *sm_ops = msp->ms_group->mg_class->mc_ops;
673 
674 	ASSERT(MUTEX_HELD(&msp->ms_lock));
675 
676 	if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
677 		space_map_load_wait(sm);
678 		if (!sm->sm_loaded) {
679 			int error = space_map_load(sm, sm_ops, SM_FREE,
680 			    &msp->ms_smo,
681 			    spa_meta_objset(msp->ms_group->mg_vd->vdev_spa));
682 			if (error) {
683 				metaslab_group_sort(msp->ms_group, msp, 0);
684 				return (error);
685 			}
686 			for (int t = 0; t < TXG_DEFER_SIZE; t++)
687 				space_map_walk(&msp->ms_defermap[t],
688 				    space_map_claim, sm);
689 		}
690 
691 		/*
692 		 * If we were able to load the map then make sure
693 		 * that this map is still able to satisfy our request.
694 		 */
695 		if (msp->ms_weight < size)
696 			return (ENOSPC);
697 
698 		metaslab_group_sort(msp->ms_group, msp,
699 		    msp->ms_weight | activation_weight);
700 	}
701 	ASSERT(sm->sm_loaded);
702 	ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
703 
704 	return (0);
705 }
706 
707 static void
708 metaslab_passivate(metaslab_t *msp, uint64_t size)
709 {
710 	/*
711 	 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
712 	 * this metaslab again.  In that case, it had better be empty,
713 	 * or we would be leaving space on the table.
714 	 */
715 #if 0
716 	ASSERT(size >= SPA_MINBLOCKSIZE || msp->ms_map.sm_space == 0);
717 #endif
718 	metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
719 	ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
720 }
721 
722 /*
723  * Write a metaslab to disk in the context of the specified transaction group.
724  */
725 void
726 metaslab_sync(metaslab_t *msp, uint64_t txg)
727 {
728 	vdev_t *vd = msp->ms_group->mg_vd;
729 	spa_t *spa = vd->vdev_spa;
730 	objset_t *mos = spa_meta_objset(spa);
731 	space_map_t *allocmap = &msp->ms_allocmap[txg & TXG_MASK];
732 	space_map_t *freemap = &msp->ms_freemap[txg & TXG_MASK];
733 	space_map_t *freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
734 	space_map_t *sm = &msp->ms_map;
735 	space_map_obj_t *smo = &msp->ms_smo_syncing;
736 	dmu_buf_t *db;
737 	dmu_tx_t *tx;
738 
739 	ASSERT(!vd->vdev_ishole);
740 
741 	if (allocmap->sm_space == 0 && freemap->sm_space == 0)
742 		return;
743 
744 	/*
745 	 * The only state that can actually be changing concurrently with
746 	 * metaslab_sync() is the metaslab's ms_map.  No other thread can
747 	 * be modifying this txg's allocmap, freemap, freed_map, or smo.
748 	 * Therefore, we only hold ms_lock to satify space_map ASSERTs.
749 	 * We drop it whenever we call into the DMU, because the DMU
750 	 * can call down to us (e.g. via zio_free()) at any time.
751 	 */
752 
753 	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
754 
755 	if (smo->smo_object == 0) {
756 		ASSERT(smo->smo_objsize == 0);
757 		ASSERT(smo->smo_alloc == 0);
758 		smo->smo_object = dmu_object_alloc(mos,
759 		    DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
760 		    DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
761 		ASSERT(smo->smo_object != 0);
762 		dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
763 		    (sm->sm_start >> vd->vdev_ms_shift),
764 		    sizeof (uint64_t), &smo->smo_object, tx);
765 	}
766 
767 	mutex_enter(&msp->ms_lock);
768 
769 	space_map_walk(freemap, space_map_add, freed_map);
770 
771 	if (sm->sm_loaded && spa_sync_pass(spa) == 1 && smo->smo_objsize >=
772 	    2 * sizeof (uint64_t) * avl_numnodes(&sm->sm_root)) {
773 		/*
774 		 * The in-core space map representation is twice as compact
775 		 * as the on-disk one, so it's time to condense the latter
776 		 * by generating a pure allocmap from first principles.
777 		 *
778 		 * This metaslab is 100% allocated,
779 		 * minus the content of the in-core map (sm),
780 		 * minus what's been freed this txg (freed_map),
781 		 * minus deferred frees (ms_defermap[]),
782 		 * minus allocations from txgs in the future
783 		 * (because they haven't been committed yet).
784 		 */
785 		space_map_vacate(allocmap, NULL, NULL);
786 		space_map_vacate(freemap, NULL, NULL);
787 
788 		space_map_add(allocmap, allocmap->sm_start, allocmap->sm_size);
789 
790 		space_map_walk(sm, space_map_remove, allocmap);
791 		space_map_walk(freed_map, space_map_remove, allocmap);
792 
793 		for (int t = 0; t < TXG_DEFER_SIZE; t++)
794 			space_map_walk(&msp->ms_defermap[t],
795 			    space_map_remove, allocmap);
796 
797 		for (int t = 1; t < TXG_CONCURRENT_STATES; t++)
798 			space_map_walk(&msp->ms_allocmap[(txg + t) & TXG_MASK],
799 			    space_map_remove, allocmap);
800 
801 		mutex_exit(&msp->ms_lock);
802 		space_map_truncate(smo, mos, tx);
803 		mutex_enter(&msp->ms_lock);
804 	}
805 
806 	space_map_sync(allocmap, SM_ALLOC, smo, mos, tx);
807 	space_map_sync(freemap, SM_FREE, smo, mos, tx);
808 
809 	mutex_exit(&msp->ms_lock);
810 
811 	VERIFY(0 == dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
812 	dmu_buf_will_dirty(db, tx);
813 	ASSERT3U(db->db_size, >=, sizeof (*smo));
814 	bcopy(smo, db->db_data, sizeof (*smo));
815 	dmu_buf_rele(db, FTAG);
816 
817 	dmu_tx_commit(tx);
818 }
819 
820 /*
821  * Called after a transaction group has completely synced to mark
822  * all of the metaslab's free space as usable.
823  */
824 void
825 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
826 {
827 	space_map_obj_t *smo = &msp->ms_smo;
828 	space_map_obj_t *smosync = &msp->ms_smo_syncing;
829 	space_map_t *sm = &msp->ms_map;
830 	space_map_t *freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
831 	space_map_t *defer_map = &msp->ms_defermap[txg % TXG_DEFER_SIZE];
832 	metaslab_group_t *mg = msp->ms_group;
833 	vdev_t *vd = mg->mg_vd;
834 	int64_t alloc_delta, defer_delta;
835 
836 	ASSERT(!vd->vdev_ishole);
837 
838 	mutex_enter(&msp->ms_lock);
839 
840 	/*
841 	 * If this metaslab is just becoming available, initialize its
842 	 * allocmaps and freemaps and add its capacity to the vdev.
843 	 */
844 	if (freed_map->sm_size == 0) {
845 		for (int t = 0; t < TXG_SIZE; t++) {
846 			space_map_create(&msp->ms_allocmap[t], sm->sm_start,
847 			    sm->sm_size, sm->sm_shift, sm->sm_lock);
848 			space_map_create(&msp->ms_freemap[t], sm->sm_start,
849 			    sm->sm_size, sm->sm_shift, sm->sm_lock);
850 		}
851 
852 		for (int t = 0; t < TXG_DEFER_SIZE; t++)
853 			space_map_create(&msp->ms_defermap[t], sm->sm_start,
854 			    sm->sm_size, sm->sm_shift, sm->sm_lock);
855 
856 		vdev_space_update(vd, 0, 0, sm->sm_size);
857 	}
858 
859 	alloc_delta = smosync->smo_alloc - smo->smo_alloc;
860 	defer_delta = freed_map->sm_space - defer_map->sm_space;
861 
862 	vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
863 
864 	ASSERT(msp->ms_allocmap[txg & TXG_MASK].sm_space == 0);
865 	ASSERT(msp->ms_freemap[txg & TXG_MASK].sm_space == 0);
866 
867 	/*
868 	 * If there's a space_map_load() in progress, wait for it to complete
869 	 * so that we have a consistent view of the in-core space map.
870 	 * Then, add defer_map (oldest deferred frees) to this map and
871 	 * transfer freed_map (this txg's frees) to defer_map.
872 	 */
873 	space_map_load_wait(sm);
874 	space_map_vacate(defer_map, sm->sm_loaded ? space_map_free : NULL, sm);
875 	space_map_vacate(freed_map, space_map_add, defer_map);
876 
877 	*smo = *smosync;
878 
879 	msp->ms_deferspace += defer_delta;
880 	ASSERT3S(msp->ms_deferspace, >=, 0);
881 	ASSERT3S(msp->ms_deferspace, <=, sm->sm_size);
882 	if (msp->ms_deferspace != 0) {
883 		/*
884 		 * Keep syncing this metaslab until all deferred frees
885 		 * are back in circulation.
886 		 */
887 		vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
888 	}
889 
890 	/*
891 	 * If the map is loaded but no longer active, evict it as soon as all
892 	 * future allocations have synced.  (If we unloaded it now and then
893 	 * loaded a moment later, the map wouldn't reflect those allocations.)
894 	 */
895 	if (sm->sm_loaded && (msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
896 		int evictable = 1;
897 
898 		for (int t = 1; t < TXG_CONCURRENT_STATES; t++)
899 			if (msp->ms_allocmap[(txg + t) & TXG_MASK].sm_space)
900 				evictable = 0;
901 
902 		if (evictable && !metaslab_debug)
903 			space_map_unload(sm);
904 	}
905 
906 	metaslab_group_sort(mg, msp, metaslab_weight(msp));
907 
908 	mutex_exit(&msp->ms_lock);
909 }
910 
911 static uint64_t
912 metaslab_distance(metaslab_t *msp, dva_t *dva)
913 {
914 	uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
915 	uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
916 	uint64_t start = msp->ms_map.sm_start >> ms_shift;
917 
918 	if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
919 		return (1ULL << 63);
920 
921 	if (offset < start)
922 		return ((start - offset) << ms_shift);
923 	if (offset > start)
924 		return ((offset - start) << ms_shift);
925 	return (0);
926 }
927 
928 static uint64_t
929 metaslab_group_alloc(metaslab_group_t *mg, uint64_t size, uint64_t txg,
930     uint64_t min_distance, dva_t *dva, int d)
931 {
932 	metaslab_t *msp = NULL;
933 	uint64_t offset = -1ULL;
934 	avl_tree_t *t = &mg->mg_metaslab_tree;
935 	uint64_t activation_weight;
936 	uint64_t target_distance;
937 	int i;
938 
939 	activation_weight = METASLAB_WEIGHT_PRIMARY;
940 	for (i = 0; i < d; i++) {
941 		if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
942 			activation_weight = METASLAB_WEIGHT_SECONDARY;
943 			break;
944 		}
945 	}
946 
947 	for (;;) {
948 		boolean_t was_active;
949 
950 		mutex_enter(&mg->mg_lock);
951 		for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
952 			if (msp->ms_weight < size) {
953 				mutex_exit(&mg->mg_lock);
954 				return (-1ULL);
955 			}
956 
957 			was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
958 			if (activation_weight == METASLAB_WEIGHT_PRIMARY)
959 				break;
960 
961 			target_distance = min_distance +
962 			    (msp->ms_smo.smo_alloc ? 0 : min_distance >> 1);
963 
964 			for (i = 0; i < d; i++)
965 				if (metaslab_distance(msp, &dva[i]) <
966 				    target_distance)
967 					break;
968 			if (i == d)
969 				break;
970 		}
971 		mutex_exit(&mg->mg_lock);
972 		if (msp == NULL)
973 			return (-1ULL);
974 
975 		mutex_enter(&msp->ms_lock);
976 
977 		/*
978 		 * Ensure that the metaslab we have selected is still
979 		 * capable of handling our request. It's possible that
980 		 * another thread may have changed the weight while we
981 		 * were blocked on the metaslab lock.
982 		 */
983 		if (msp->ms_weight < size || (was_active &&
984 		    !(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
985 		    activation_weight == METASLAB_WEIGHT_PRIMARY)) {
986 			mutex_exit(&msp->ms_lock);
987 			continue;
988 		}
989 
990 		if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
991 		    activation_weight == METASLAB_WEIGHT_PRIMARY) {
992 			metaslab_passivate(msp,
993 			    msp->ms_weight & ~METASLAB_ACTIVE_MASK);
994 			mutex_exit(&msp->ms_lock);
995 			continue;
996 		}
997 
998 		if (metaslab_activate(msp, activation_weight, size) != 0) {
999 			mutex_exit(&msp->ms_lock);
1000 			continue;
1001 		}
1002 
1003 		if ((offset = space_map_alloc(&msp->ms_map, size)) != -1ULL)
1004 			break;
1005 
1006 		metaslab_passivate(msp, size - 1);
1007 
1008 		mutex_exit(&msp->ms_lock);
1009 	}
1010 
1011 	if (msp->ms_allocmap[txg & TXG_MASK].sm_space == 0)
1012 		vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
1013 
1014 	space_map_add(&msp->ms_allocmap[txg & TXG_MASK], offset, size);
1015 
1016 	mutex_exit(&msp->ms_lock);
1017 
1018 	return (offset);
1019 }
1020 
1021 /*
1022  * Allocate a block for the specified i/o.
1023  */
1024 static int
1025 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
1026     dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
1027 {
1028 	metaslab_group_t *mg, *rotor;
1029 	vdev_t *vd;
1030 	int dshift = 3;
1031 	int all_zero;
1032 	int zio_lock = B_FALSE;
1033 	boolean_t allocatable;
1034 	uint64_t offset = -1ULL;
1035 	uint64_t asize;
1036 	uint64_t distance;
1037 
1038 	ASSERT(!DVA_IS_VALID(&dva[d]));
1039 
1040 	/*
1041 	 * For testing, make some blocks above a certain size be gang blocks.
1042 	 */
1043 	if (psize >= metaslab_gang_bang && (lbolt & 3) == 0)
1044 		return (ENOSPC);
1045 
1046 	/*
1047 	 * Start at the rotor and loop through all mgs until we find something.
1048 	 * Note that there's no locking on mc_rotor or mc_aliquot because
1049 	 * nothing actually breaks if we miss a few updates -- we just won't
1050 	 * allocate quite as evenly.  It all balances out over time.
1051 	 *
1052 	 * If we are doing ditto or log blocks, try to spread them across
1053 	 * consecutive vdevs.  If we're forced to reuse a vdev before we've
1054 	 * allocated all of our ditto blocks, then try and spread them out on
1055 	 * that vdev as much as possible.  If it turns out to not be possible,
1056 	 * gradually lower our standards until anything becomes acceptable.
1057 	 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
1058 	 * gives us hope of containing our fault domains to something we're
1059 	 * able to reason about.  Otherwise, any two top-level vdev failures
1060 	 * will guarantee the loss of data.  With consecutive allocation,
1061 	 * only two adjacent top-level vdev failures will result in data loss.
1062 	 *
1063 	 * If we are doing gang blocks (hintdva is non-NULL), try to keep
1064 	 * ourselves on the same vdev as our gang block header.  That
1065 	 * way, we can hope for locality in vdev_cache, plus it makes our
1066 	 * fault domains something tractable.
1067 	 */
1068 	if (hintdva) {
1069 		vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
1070 
1071 		/*
1072 		 * It's possible the vdev we're using as the hint no
1073 		 * longer exists (i.e. removed). Consult the rotor when
1074 		 * all else fails.
1075 		 */
1076 		if (vd != NULL) {
1077 			mg = vd->vdev_mg;
1078 
1079 			if (flags & METASLAB_HINTBP_AVOID &&
1080 			    mg->mg_next != NULL)
1081 				mg = mg->mg_next;
1082 		} else {
1083 			mg = mc->mc_rotor;
1084 		}
1085 	} else if (d != 0) {
1086 		vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
1087 		mg = vd->vdev_mg->mg_next;
1088 	} else {
1089 		mg = mc->mc_rotor;
1090 	}
1091 
1092 	/*
1093 	 * If the hint put us into the wrong metaslab class, or into a
1094 	 * metaslab group that has been passivated, just follow the rotor.
1095 	 */
1096 	if (mg->mg_class != mc || mg->mg_activation_count <= 0)
1097 		mg = mc->mc_rotor;
1098 
1099 	rotor = mg;
1100 top:
1101 	all_zero = B_TRUE;
1102 	do {
1103 		ASSERT(mg->mg_activation_count == 1);
1104 
1105 		vd = mg->mg_vd;
1106 
1107 		/*
1108 		 * Don't allocate from faulted devices.
1109 		 */
1110 		if (zio_lock) {
1111 			spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
1112 			allocatable = vdev_allocatable(vd);
1113 			spa_config_exit(spa, SCL_ZIO, FTAG);
1114 		} else {
1115 			allocatable = vdev_allocatable(vd);
1116 		}
1117 		if (!allocatable)
1118 			goto next;
1119 
1120 		/*
1121 		 * Avoid writing single-copy data to a failing vdev
1122 		 */
1123 		if ((vd->vdev_stat.vs_write_errors > 0 ||
1124 		    vd->vdev_state < VDEV_STATE_HEALTHY) &&
1125 		    d == 0 && dshift == 3) {
1126 			all_zero = B_FALSE;
1127 			goto next;
1128 		}
1129 
1130 		ASSERT(mg->mg_class == mc);
1131 
1132 		distance = vd->vdev_asize >> dshift;
1133 		if (distance <= (1ULL << vd->vdev_ms_shift))
1134 			distance = 0;
1135 		else
1136 			all_zero = B_FALSE;
1137 
1138 		asize = vdev_psize_to_asize(vd, psize);
1139 		ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
1140 
1141 		offset = metaslab_group_alloc(mg, asize, txg, distance, dva, d);
1142 		if (offset != -1ULL) {
1143 			/*
1144 			 * If we've just selected this metaslab group,
1145 			 * figure out whether the corresponding vdev is
1146 			 * over- or under-used relative to the pool,
1147 			 * and set an allocation bias to even it out.
1148 			 */
1149 			if (mc->mc_aliquot == 0) {
1150 				vdev_stat_t *vs = &vd->vdev_stat;
1151 				int64_t vu, cu;
1152 
1153 				/*
1154 				 * Determine percent used in units of 0..1024.
1155 				 * (This is just to avoid floating point.)
1156 				 */
1157 				vu = (vs->vs_alloc << 10) / (vs->vs_space + 1);
1158 				cu = (mc->mc_alloc << 10) / (mc->mc_space + 1);
1159 
1160 				/*
1161 				 * Bias by at most +/- 25% of the aliquot.
1162 				 */
1163 				mg->mg_bias = ((cu - vu) *
1164 				    (int64_t)mg->mg_aliquot) / (1024 * 4);
1165 			}
1166 
1167 			if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
1168 			    mg->mg_aliquot + mg->mg_bias) {
1169 				mc->mc_rotor = mg->mg_next;
1170 				mc->mc_aliquot = 0;
1171 			}
1172 
1173 			DVA_SET_VDEV(&dva[d], vd->vdev_id);
1174 			DVA_SET_OFFSET(&dva[d], offset);
1175 			DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
1176 			DVA_SET_ASIZE(&dva[d], asize);
1177 
1178 			return (0);
1179 		}
1180 next:
1181 		mc->mc_rotor = mg->mg_next;
1182 		mc->mc_aliquot = 0;
1183 	} while ((mg = mg->mg_next) != rotor);
1184 
1185 	if (!all_zero) {
1186 		dshift++;
1187 		ASSERT(dshift < 64);
1188 		goto top;
1189 	}
1190 
1191 	if (!allocatable && !zio_lock) {
1192 		dshift = 3;
1193 		zio_lock = B_TRUE;
1194 		goto top;
1195 	}
1196 
1197 	bzero(&dva[d], sizeof (dva_t));
1198 
1199 	return (ENOSPC);
1200 }
1201 
1202 /*
1203  * Free the block represented by DVA in the context of the specified
1204  * transaction group.
1205  */
1206 static void
1207 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
1208 {
1209 	uint64_t vdev = DVA_GET_VDEV(dva);
1210 	uint64_t offset = DVA_GET_OFFSET(dva);
1211 	uint64_t size = DVA_GET_ASIZE(dva);
1212 	vdev_t *vd;
1213 	metaslab_t *msp;
1214 
1215 	ASSERT(DVA_IS_VALID(dva));
1216 
1217 	if (txg > spa_freeze_txg(spa))
1218 		return;
1219 
1220 	if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
1221 	    (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
1222 		cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
1223 		    (u_longlong_t)vdev, (u_longlong_t)offset);
1224 		ASSERT(0);
1225 		return;
1226 	}
1227 
1228 	msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1229 
1230 	if (DVA_GET_GANG(dva))
1231 		size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1232 
1233 	mutex_enter(&msp->ms_lock);
1234 
1235 	if (now) {
1236 		space_map_remove(&msp->ms_allocmap[txg & TXG_MASK],
1237 		    offset, size);
1238 		space_map_free(&msp->ms_map, offset, size);
1239 	} else {
1240 		if (msp->ms_freemap[txg & TXG_MASK].sm_space == 0)
1241 			vdev_dirty(vd, VDD_METASLAB, msp, txg);
1242 		space_map_add(&msp->ms_freemap[txg & TXG_MASK], offset, size);
1243 	}
1244 
1245 	mutex_exit(&msp->ms_lock);
1246 }
1247 
1248 /*
1249  * Intent log support: upon opening the pool after a crash, notify the SPA
1250  * of blocks that the intent log has allocated for immediate write, but
1251  * which are still considered free by the SPA because the last transaction
1252  * group didn't commit yet.
1253  */
1254 static int
1255 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
1256 {
1257 	uint64_t vdev = DVA_GET_VDEV(dva);
1258 	uint64_t offset = DVA_GET_OFFSET(dva);
1259 	uint64_t size = DVA_GET_ASIZE(dva);
1260 	vdev_t *vd;
1261 	metaslab_t *msp;
1262 	int error = 0;
1263 
1264 	ASSERT(DVA_IS_VALID(dva));
1265 
1266 	if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
1267 	    (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
1268 		return (ENXIO);
1269 
1270 	msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1271 
1272 	if (DVA_GET_GANG(dva))
1273 		size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1274 
1275 	mutex_enter(&msp->ms_lock);
1276 
1277 	if ((txg != 0 && spa_writeable(spa)) || !msp->ms_map.sm_loaded)
1278 		error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY, 0);
1279 
1280 	if (error == 0 && !space_map_contains(&msp->ms_map, offset, size))
1281 		error = ENOENT;
1282 
1283 	if (error || txg == 0) {	/* txg == 0 indicates dry run */
1284 		mutex_exit(&msp->ms_lock);
1285 		return (error);
1286 	}
1287 
1288 	space_map_claim(&msp->ms_map, offset, size);
1289 
1290 	if (spa_writeable(spa)) {	/* don't dirty if we're zdb(1M) */
1291 		if (msp->ms_allocmap[txg & TXG_MASK].sm_space == 0)
1292 			vdev_dirty(vd, VDD_METASLAB, msp, txg);
1293 		space_map_add(&msp->ms_allocmap[txg & TXG_MASK], offset, size);
1294 	}
1295 
1296 	mutex_exit(&msp->ms_lock);
1297 
1298 	return (0);
1299 }
1300 
1301 int
1302 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
1303     int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
1304 {
1305 	dva_t *dva = bp->blk_dva;
1306 	dva_t *hintdva = hintbp->blk_dva;
1307 	int error = 0;
1308 
1309 	ASSERT(bp->blk_birth == 0);
1310 	ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
1311 
1312 	spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
1313 
1314 	if (mc->mc_rotor == NULL) {	/* no vdevs in this class */
1315 		spa_config_exit(spa, SCL_ALLOC, FTAG);
1316 		return (ENOSPC);
1317 	}
1318 
1319 	ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
1320 	ASSERT(BP_GET_NDVAS(bp) == 0);
1321 	ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
1322 
1323 	for (int d = 0; d < ndvas; d++) {
1324 		error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
1325 		    txg, flags);
1326 		if (error) {
1327 			for (d--; d >= 0; d--) {
1328 				metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
1329 				bzero(&dva[d], sizeof (dva_t));
1330 			}
1331 			spa_config_exit(spa, SCL_ALLOC, FTAG);
1332 			return (error);
1333 		}
1334 	}
1335 	ASSERT(error == 0);
1336 	ASSERT(BP_GET_NDVAS(bp) == ndvas);
1337 
1338 	spa_config_exit(spa, SCL_ALLOC, FTAG);
1339 
1340 	BP_SET_BIRTH(bp, txg, txg);
1341 
1342 	return (0);
1343 }
1344 
1345 void
1346 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
1347 {
1348 	const dva_t *dva = bp->blk_dva;
1349 	int ndvas = BP_GET_NDVAS(bp);
1350 
1351 	ASSERT(!BP_IS_HOLE(bp));
1352 	ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
1353 
1354 	spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
1355 
1356 	for (int d = 0; d < ndvas; d++)
1357 		metaslab_free_dva(spa, &dva[d], txg, now);
1358 
1359 	spa_config_exit(spa, SCL_FREE, FTAG);
1360 }
1361 
1362 int
1363 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
1364 {
1365 	const dva_t *dva = bp->blk_dva;
1366 	int ndvas = BP_GET_NDVAS(bp);
1367 	int error = 0;
1368 
1369 	ASSERT(!BP_IS_HOLE(bp));
1370 
1371 	if (txg != 0) {
1372 		/*
1373 		 * First do a dry run to make sure all DVAs are claimable,
1374 		 * so we don't have to unwind from partial failures below.
1375 		 */
1376 		if ((error = metaslab_claim(spa, bp, 0)) != 0)
1377 			return (error);
1378 	}
1379 
1380 	spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
1381 
1382 	for (int d = 0; d < ndvas; d++)
1383 		if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
1384 			break;
1385 
1386 	spa_config_exit(spa, SCL_ALLOC, FTAG);
1387 
1388 	ASSERT(error == 0 || txg == 0);
1389 
1390 	return (error);
1391 }
1392