xref: /titanic_41/usr/src/uts/common/fs/zfs/metaslab.c (revision 4eaa471005973e11a6110b69fe990530b3b95a38)
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/spa_impl.h>
28 #include <sys/dmu.h>
29 #include <sys/dmu_tx.h>
30 #include <sys/space_map.h>
31 #include <sys/metaslab_impl.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/zio.h>
34 
35 uint64_t metaslab_aliquot = 512ULL << 10;
36 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1;	/* force gang blocks */
37 
38 /*
39  * Minimum size which forces the dynamic allocator to change
40  * it's allocation strategy. Once the space map cannot satisfy
41  * an allocation of this size then it switches to using more
42  * aggressive strategy (i.e search by size rather than offset).
43  */
44 uint64_t metaslab_df_alloc_threshold = SPA_MAXBLOCKSIZE;
45 
46 /*
47  * The minimum free space, in percent, which must be available
48  * in a space map to continue allocations in a first-fit fashion.
49  * Once the space_map's free space drops below this level we dynamically
50  * switch to using best-fit allocations.
51  */
52 int metaslab_df_free_pct = 30;
53 
54 /*
55  * ==========================================================================
56  * Metaslab classes
57  * ==========================================================================
58  */
59 metaslab_class_t *
60 metaslab_class_create(space_map_ops_t *ops)
61 {
62 	metaslab_class_t *mc;
63 
64 	mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
65 
66 	mc->mc_rotor = NULL;
67 	mc->mc_ops = ops;
68 
69 	return (mc);
70 }
71 
72 void
73 metaslab_class_destroy(metaslab_class_t *mc)
74 {
75 	metaslab_group_t *mg;
76 
77 	while ((mg = mc->mc_rotor) != NULL) {
78 		metaslab_class_remove(mc, mg);
79 		metaslab_group_destroy(mg);
80 	}
81 
82 	kmem_free(mc, sizeof (metaslab_class_t));
83 }
84 
85 void
86 metaslab_class_add(metaslab_class_t *mc, metaslab_group_t *mg)
87 {
88 	metaslab_group_t *mgprev, *mgnext;
89 
90 	ASSERT(mg->mg_class == NULL);
91 
92 	if ((mgprev = mc->mc_rotor) == NULL) {
93 		mg->mg_prev = mg;
94 		mg->mg_next = mg;
95 	} else {
96 		mgnext = mgprev->mg_next;
97 		mg->mg_prev = mgprev;
98 		mg->mg_next = mgnext;
99 		mgprev->mg_next = mg;
100 		mgnext->mg_prev = mg;
101 	}
102 	mc->mc_rotor = mg;
103 	mg->mg_class = mc;
104 }
105 
106 void
107 metaslab_class_remove(metaslab_class_t *mc, metaslab_group_t *mg)
108 {
109 	metaslab_group_t *mgprev, *mgnext;
110 
111 	ASSERT(mg->mg_class == mc);
112 
113 	mgprev = mg->mg_prev;
114 	mgnext = mg->mg_next;
115 
116 	if (mg == mgnext) {
117 		mc->mc_rotor = NULL;
118 	} else {
119 		mc->mc_rotor = mgnext;
120 		mgprev->mg_next = mgnext;
121 		mgnext->mg_prev = mgprev;
122 	}
123 
124 	mg->mg_prev = NULL;
125 	mg->mg_next = NULL;
126 	mg->mg_class = NULL;
127 }
128 
129 /*
130  * ==========================================================================
131  * Metaslab groups
132  * ==========================================================================
133  */
134 static int
135 metaslab_compare(const void *x1, const void *x2)
136 {
137 	const metaslab_t *m1 = x1;
138 	const metaslab_t *m2 = x2;
139 
140 	if (m1->ms_weight < m2->ms_weight)
141 		return (1);
142 	if (m1->ms_weight > m2->ms_weight)
143 		return (-1);
144 
145 	/*
146 	 * If the weights are identical, use the offset to force uniqueness.
147 	 */
148 	if (m1->ms_map.sm_start < m2->ms_map.sm_start)
149 		return (-1);
150 	if (m1->ms_map.sm_start > m2->ms_map.sm_start)
151 		return (1);
152 
153 	ASSERT3P(m1, ==, m2);
154 
155 	return (0);
156 }
157 
158 metaslab_group_t *
159 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
160 {
161 	metaslab_group_t *mg;
162 
163 	mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
164 	mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
165 	avl_create(&mg->mg_metaslab_tree, metaslab_compare,
166 	    sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
167 	mg->mg_aliquot = metaslab_aliquot * MAX(1, vd->vdev_children);
168 	mg->mg_vd = vd;
169 	metaslab_class_add(mc, mg);
170 
171 	return (mg);
172 }
173 
174 void
175 metaslab_group_destroy(metaslab_group_t *mg)
176 {
177 	avl_destroy(&mg->mg_metaslab_tree);
178 	mutex_destroy(&mg->mg_lock);
179 	kmem_free(mg, sizeof (metaslab_group_t));
180 }
181 
182 static void
183 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
184 {
185 	mutex_enter(&mg->mg_lock);
186 	ASSERT(msp->ms_group == NULL);
187 	msp->ms_group = mg;
188 	msp->ms_weight = 0;
189 	avl_add(&mg->mg_metaslab_tree, msp);
190 	mutex_exit(&mg->mg_lock);
191 }
192 
193 static void
194 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
195 {
196 	mutex_enter(&mg->mg_lock);
197 	ASSERT(msp->ms_group == mg);
198 	avl_remove(&mg->mg_metaslab_tree, msp);
199 	msp->ms_group = NULL;
200 	mutex_exit(&mg->mg_lock);
201 }
202 
203 static void
204 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
205 {
206 	/*
207 	 * Although in principle the weight can be any value, in
208 	 * practice we do not use values in the range [1, 510].
209 	 */
210 	ASSERT(weight >= SPA_MINBLOCKSIZE-1 || weight == 0);
211 	ASSERT(MUTEX_HELD(&msp->ms_lock));
212 
213 	mutex_enter(&mg->mg_lock);
214 	ASSERT(msp->ms_group == mg);
215 	avl_remove(&mg->mg_metaslab_tree, msp);
216 	msp->ms_weight = weight;
217 	avl_add(&mg->mg_metaslab_tree, msp);
218 	mutex_exit(&mg->mg_lock);
219 }
220 
221 /*
222  * This is a helper function that can be used by the allocator to find
223  * a suitable block to allocate. This will search the specified AVL
224  * tree looking for a block that matches the specified criteria.
225  */
226 static uint64_t
227 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
228     uint64_t align)
229 {
230 	space_seg_t *ss, ssearch;
231 	avl_index_t where;
232 
233 	ssearch.ss_start = *cursor;
234 	ssearch.ss_end = *cursor + size;
235 
236 	ss = avl_find(t, &ssearch, &where);
237 	if (ss == NULL)
238 		ss = avl_nearest(t, where, AVL_AFTER);
239 
240 	while (ss != NULL) {
241 		uint64_t offset = P2ROUNDUP(ss->ss_start, align);
242 
243 		if (offset + size <= ss->ss_end) {
244 			*cursor = offset + size;
245 			return (offset);
246 		}
247 		ss = AVL_NEXT(t, ss);
248 	}
249 
250 	/*
251 	 * If we know we've searched the whole map (*cursor == 0), give up.
252 	 * Otherwise, reset the cursor to the beginning and try again.
253 	 */
254 	if (*cursor == 0)
255 		return (-1ULL);
256 
257 	*cursor = 0;
258 	return (metaslab_block_picker(t, cursor, size, align));
259 }
260 
261 /*
262  * ==========================================================================
263  * The first-fit block allocator
264  * ==========================================================================
265  */
266 static void
267 metaslab_ff_load(space_map_t *sm)
268 {
269 	ASSERT(sm->sm_ppd == NULL);
270 	sm->sm_ppd = kmem_zalloc(64 * sizeof (uint64_t), KM_SLEEP);
271 	sm->sm_pp_root = NULL;
272 }
273 
274 static void
275 metaslab_ff_unload(space_map_t *sm)
276 {
277 	kmem_free(sm->sm_ppd, 64 * sizeof (uint64_t));
278 	sm->sm_ppd = NULL;
279 }
280 
281 static uint64_t
282 metaslab_ff_alloc(space_map_t *sm, uint64_t size)
283 {
284 	avl_tree_t *t = &sm->sm_root;
285 	uint64_t align = size & -size;
286 	uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
287 
288 	return (metaslab_block_picker(t, cursor, size, align));
289 }
290 
291 /* ARGSUSED */
292 static void
293 metaslab_ff_claim(space_map_t *sm, uint64_t start, uint64_t size)
294 {
295 	/* No need to update cursor */
296 }
297 
298 /* ARGSUSED */
299 static void
300 metaslab_ff_free(space_map_t *sm, uint64_t start, uint64_t size)
301 {
302 	/* No need to update cursor */
303 }
304 
305 static space_map_ops_t metaslab_ff_ops = {
306 	metaslab_ff_load,
307 	metaslab_ff_unload,
308 	metaslab_ff_alloc,
309 	metaslab_ff_claim,
310 	metaslab_ff_free,
311 	NULL	/* maxsize */
312 };
313 
314 /*
315  * Dynamic block allocator -
316  * Uses the first fit allocation scheme until space get low and then
317  * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
318  * and metaslab_df_free_pct to determine when to switch the allocation scheme.
319  */
320 
321 uint64_t
322 metaslab_df_maxsize(space_map_t *sm)
323 {
324 	avl_tree_t *t = sm->sm_pp_root;
325 	space_seg_t *ss;
326 
327 	if (t == NULL || (ss = avl_last(t)) == NULL)
328 		return (0ULL);
329 
330 	return (ss->ss_end - ss->ss_start);
331 }
332 
333 static int
334 metaslab_df_seg_compare(const void *x1, const void *x2)
335 {
336 	const space_seg_t *s1 = x1;
337 	const space_seg_t *s2 = x2;
338 	uint64_t ss_size1 = s1->ss_end - s1->ss_start;
339 	uint64_t ss_size2 = s2->ss_end - s2->ss_start;
340 
341 	if (ss_size1 < ss_size2)
342 		return (-1);
343 	if (ss_size1 > ss_size2)
344 		return (1);
345 
346 	if (s1->ss_start < s2->ss_start)
347 		return (-1);
348 	if (s1->ss_start > s2->ss_start)
349 		return (1);
350 
351 	return (0);
352 }
353 
354 static void
355 metaslab_df_load(space_map_t *sm)
356 {
357 	space_seg_t *ss;
358 
359 	ASSERT(sm->sm_ppd == NULL);
360 	sm->sm_ppd = kmem_zalloc(64 * sizeof (uint64_t), KM_SLEEP);
361 
362 	sm->sm_pp_root = kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
363 	avl_create(sm->sm_pp_root, metaslab_df_seg_compare,
364 	    sizeof (space_seg_t), offsetof(struct space_seg, ss_pp_node));
365 
366 	for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
367 		avl_add(sm->sm_pp_root, ss);
368 }
369 
370 static void
371 metaslab_df_unload(space_map_t *sm)
372 {
373 	void *cookie = NULL;
374 
375 	kmem_free(sm->sm_ppd, 64 * sizeof (uint64_t));
376 	sm->sm_ppd = NULL;
377 
378 	while (avl_destroy_nodes(sm->sm_pp_root, &cookie) != NULL) {
379 		/* tear down the tree */
380 	}
381 
382 	avl_destroy(sm->sm_pp_root);
383 	kmem_free(sm->sm_pp_root, sizeof (avl_tree_t));
384 	sm->sm_pp_root = NULL;
385 }
386 
387 static uint64_t
388 metaslab_df_alloc(space_map_t *sm, uint64_t size)
389 {
390 	avl_tree_t *t = &sm->sm_root;
391 	uint64_t align = size & -size;
392 	uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
393 	uint64_t max_size = metaslab_df_maxsize(sm);
394 	int free_pct = sm->sm_space * 100 / sm->sm_size;
395 
396 	ASSERT(MUTEX_HELD(sm->sm_lock));
397 	ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
398 
399 	if (max_size < size)
400 		return (-1ULL);
401 
402 	/*
403 	 * If we're running low on space switch to using the size
404 	 * sorted AVL tree (best-fit).
405 	 */
406 	if (max_size < metaslab_df_alloc_threshold ||
407 	    free_pct < metaslab_df_free_pct) {
408 		t = sm->sm_pp_root;
409 		*cursor = 0;
410 	}
411 
412 	return (metaslab_block_picker(t, cursor, size, 1ULL));
413 }
414 
415 /* ARGSUSED */
416 static void
417 metaslab_df_claim(space_map_t *sm, uint64_t start, uint64_t size)
418 {
419 	/* No need to update cursor */
420 }
421 
422 /* ARGSUSED */
423 static void
424 metaslab_df_free(space_map_t *sm, uint64_t start, uint64_t size)
425 {
426 	/* No need to update cursor */
427 }
428 
429 static space_map_ops_t metaslab_df_ops = {
430 	metaslab_df_load,
431 	metaslab_df_unload,
432 	metaslab_df_alloc,
433 	metaslab_df_claim,
434 	metaslab_df_free,
435 	metaslab_df_maxsize
436 };
437 
438 space_map_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
439 
440 /*
441  * ==========================================================================
442  * Metaslabs
443  * ==========================================================================
444  */
445 metaslab_t *
446 metaslab_init(metaslab_group_t *mg, space_map_obj_t *smo,
447 	uint64_t start, uint64_t size, uint64_t txg)
448 {
449 	vdev_t *vd = mg->mg_vd;
450 	metaslab_t *msp;
451 
452 	msp = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
453 	mutex_init(&msp->ms_lock, NULL, MUTEX_DEFAULT, NULL);
454 
455 	msp->ms_smo_syncing = *smo;
456 
457 	/*
458 	 * We create the main space map here, but we don't create the
459 	 * allocmaps and freemaps until metaslab_sync_done().  This serves
460 	 * two purposes: it allows metaslab_sync_done() to detect the
461 	 * addition of new space; and for debugging, it ensures that we'd
462 	 * data fault on any attempt to use this metaslab before it's ready.
463 	 */
464 	space_map_create(&msp->ms_map, start, size,
465 	    vd->vdev_ashift, &msp->ms_lock);
466 
467 	metaslab_group_add(mg, msp);
468 
469 	/*
470 	 * If we're opening an existing pool (txg == 0) or creating
471 	 * a new one (txg == TXG_INITIAL), all space is available now.
472 	 * If we're adding space to an existing pool, the new space
473 	 * does not become available until after this txg has synced.
474 	 */
475 	if (txg <= TXG_INITIAL)
476 		metaslab_sync_done(msp, 0);
477 
478 	if (txg != 0) {
479 		/*
480 		 * The vdev is dirty, but the metaslab isn't -- it just needs
481 		 * to have metaslab_sync_done() invoked from vdev_sync_done().
482 		 * [We could just dirty the metaslab, but that would cause us
483 		 * to allocate a space map object for it, which is wasteful
484 		 * and would mess up the locality logic in metaslab_weight().]
485 		 */
486 		ASSERT(TXG_CLEAN(txg) == spa_last_synced_txg(vd->vdev_spa));
487 		vdev_dirty(vd, 0, NULL, txg);
488 		vdev_dirty(vd, VDD_METASLAB, msp, TXG_CLEAN(txg));
489 	}
490 
491 	return (msp);
492 }
493 
494 void
495 metaslab_fini(metaslab_t *msp)
496 {
497 	metaslab_group_t *mg = msp->ms_group;
498 	int t;
499 
500 	vdev_space_update(mg->mg_vd, -msp->ms_map.sm_size,
501 	    -msp->ms_smo.smo_alloc, B_TRUE);
502 
503 	metaslab_group_remove(mg, msp);
504 
505 	mutex_enter(&msp->ms_lock);
506 
507 	space_map_unload(&msp->ms_map);
508 	space_map_destroy(&msp->ms_map);
509 
510 	for (t = 0; t < TXG_SIZE; t++) {
511 		space_map_destroy(&msp->ms_allocmap[t]);
512 		space_map_destroy(&msp->ms_freemap[t]);
513 	}
514 
515 	mutex_exit(&msp->ms_lock);
516 	mutex_destroy(&msp->ms_lock);
517 
518 	kmem_free(msp, sizeof (metaslab_t));
519 }
520 
521 #define	METASLAB_WEIGHT_PRIMARY		(1ULL << 63)
522 #define	METASLAB_WEIGHT_SECONDARY	(1ULL << 62)
523 #define	METASLAB_ACTIVE_MASK		\
524 	(METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
525 #define	METASLAB_SMO_BONUS_MULTIPLIER	2
526 
527 static uint64_t
528 metaslab_weight(metaslab_t *msp)
529 {
530 	metaslab_group_t *mg = msp->ms_group;
531 	space_map_t *sm = &msp->ms_map;
532 	space_map_obj_t *smo = &msp->ms_smo;
533 	vdev_t *vd = mg->mg_vd;
534 	uint64_t weight, space;
535 
536 	ASSERT(MUTEX_HELD(&msp->ms_lock));
537 
538 	/*
539 	 * The baseline weight is the metaslab's free space.
540 	 */
541 	space = sm->sm_size - smo->smo_alloc;
542 	weight = space;
543 
544 	/*
545 	 * Modern disks have uniform bit density and constant angular velocity.
546 	 * Therefore, the outer recording zones are faster (higher bandwidth)
547 	 * than the inner zones by the ratio of outer to inner track diameter,
548 	 * which is typically around 2:1.  We account for this by assigning
549 	 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
550 	 * In effect, this means that we'll select the metaslab with the most
551 	 * free bandwidth rather than simply the one with the most free space.
552 	 */
553 	weight = 2 * weight -
554 	    ((sm->sm_start >> vd->vdev_ms_shift) * weight) / vd->vdev_ms_count;
555 	ASSERT(weight >= space && weight <= 2 * space);
556 
557 	/*
558 	 * For locality, assign higher weight to metaslabs we've used before.
559 	 */
560 	if (smo->smo_object != 0)
561 		weight *= METASLAB_SMO_BONUS_MULTIPLIER;
562 	ASSERT(weight >= space &&
563 	    weight <= 2 * METASLAB_SMO_BONUS_MULTIPLIER * space);
564 
565 	/*
566 	 * If this metaslab is one we're actively using, adjust its weight to
567 	 * make it preferable to any inactive metaslab so we'll polish it off.
568 	 */
569 	weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
570 
571 	return (weight);
572 }
573 
574 static int
575 metaslab_activate(metaslab_t *msp, uint64_t activation_weight, uint64_t size)
576 {
577 	space_map_t *sm = &msp->ms_map;
578 	space_map_ops_t *sm_ops = msp->ms_group->mg_class->mc_ops;
579 
580 	ASSERT(MUTEX_HELD(&msp->ms_lock));
581 
582 	if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
583 		int error = space_map_load(sm, sm_ops, SM_FREE, &msp->ms_smo,
584 		    msp->ms_group->mg_vd->vdev_spa->spa_meta_objset);
585 		if (error) {
586 			metaslab_group_sort(msp->ms_group, msp, 0);
587 			return (error);
588 		}
589 
590 		/*
591 		 * If we were able to load the map then make sure
592 		 * that this map is still able to satisfy our request.
593 		 */
594 		if (msp->ms_weight < size)
595 			return (ENOSPC);
596 
597 		metaslab_group_sort(msp->ms_group, msp,
598 		    msp->ms_weight | activation_weight);
599 	}
600 	ASSERT(sm->sm_loaded);
601 	ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
602 
603 	return (0);
604 }
605 
606 static void
607 metaslab_passivate(metaslab_t *msp, uint64_t size)
608 {
609 	/*
610 	 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
611 	 * this metaslab again.  In that case, it had better be empty,
612 	 * or we would be leaving space on the table.
613 	 */
614 	ASSERT(size >= SPA_MINBLOCKSIZE || msp->ms_map.sm_space == 0);
615 	metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
616 	ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
617 }
618 
619 /*
620  * Write a metaslab to disk in the context of the specified transaction group.
621  */
622 void
623 metaslab_sync(metaslab_t *msp, uint64_t txg)
624 {
625 	vdev_t *vd = msp->ms_group->mg_vd;
626 	spa_t *spa = vd->vdev_spa;
627 	objset_t *mos = spa->spa_meta_objset;
628 	space_map_t *allocmap = &msp->ms_allocmap[txg & TXG_MASK];
629 	space_map_t *freemap = &msp->ms_freemap[txg & TXG_MASK];
630 	space_map_t *freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
631 	space_map_t *sm = &msp->ms_map;
632 	space_map_obj_t *smo = &msp->ms_smo_syncing;
633 	dmu_buf_t *db;
634 	dmu_tx_t *tx;
635 	int t;
636 
637 	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
638 
639 	/*
640 	 * The only state that can actually be changing concurrently with
641 	 * metaslab_sync() is the metaslab's ms_map.  No other thread can
642 	 * be modifying this txg's allocmap, freemap, freed_map, or smo.
643 	 * Therefore, we only hold ms_lock to satify space_map ASSERTs.
644 	 * We drop it whenever we call into the DMU, because the DMU
645 	 * can call down to us (e.g. via zio_free()) at any time.
646 	 */
647 	mutex_enter(&msp->ms_lock);
648 
649 	if (smo->smo_object == 0) {
650 		ASSERT(smo->smo_objsize == 0);
651 		ASSERT(smo->smo_alloc == 0);
652 		mutex_exit(&msp->ms_lock);
653 		smo->smo_object = dmu_object_alloc(mos,
654 		    DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
655 		    DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
656 		ASSERT(smo->smo_object != 0);
657 		dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
658 		    (sm->sm_start >> vd->vdev_ms_shift),
659 		    sizeof (uint64_t), &smo->smo_object, tx);
660 		mutex_enter(&msp->ms_lock);
661 	}
662 
663 	space_map_walk(freemap, space_map_add, freed_map);
664 
665 	if (sm->sm_loaded && spa_sync_pass(spa) == 1 && smo->smo_objsize >=
666 	    2 * sizeof (uint64_t) * avl_numnodes(&sm->sm_root)) {
667 		/*
668 		 * The in-core space map representation is twice as compact
669 		 * as the on-disk one, so it's time to condense the latter
670 		 * by generating a pure allocmap from first principles.
671 		 *
672 		 * This metaslab is 100% allocated,
673 		 * minus the content of the in-core map (sm),
674 		 * minus what's been freed this txg (freed_map),
675 		 * minus allocations from txgs in the future
676 		 * (because they haven't been committed yet).
677 		 */
678 		space_map_vacate(allocmap, NULL, NULL);
679 		space_map_vacate(freemap, NULL, NULL);
680 
681 		space_map_add(allocmap, allocmap->sm_start, allocmap->sm_size);
682 
683 		space_map_walk(sm, space_map_remove, allocmap);
684 		space_map_walk(freed_map, space_map_remove, allocmap);
685 
686 		for (t = 1; t < TXG_CONCURRENT_STATES; t++)
687 			space_map_walk(&msp->ms_allocmap[(txg + t) & TXG_MASK],
688 			    space_map_remove, allocmap);
689 
690 		mutex_exit(&msp->ms_lock);
691 		space_map_truncate(smo, mos, tx);
692 		mutex_enter(&msp->ms_lock);
693 	}
694 
695 	space_map_sync(allocmap, SM_ALLOC, smo, mos, tx);
696 	space_map_sync(freemap, SM_FREE, smo, mos, tx);
697 
698 	mutex_exit(&msp->ms_lock);
699 
700 	VERIFY(0 == dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
701 	dmu_buf_will_dirty(db, tx);
702 	ASSERT3U(db->db_size, >=, sizeof (*smo));
703 	bcopy(smo, db->db_data, sizeof (*smo));
704 	dmu_buf_rele(db, FTAG);
705 
706 	dmu_tx_commit(tx);
707 }
708 
709 /*
710  * Called after a transaction group has completely synced to mark
711  * all of the metaslab's free space as usable.
712  */
713 void
714 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
715 {
716 	space_map_obj_t *smo = &msp->ms_smo;
717 	space_map_obj_t *smosync = &msp->ms_smo_syncing;
718 	space_map_t *sm = &msp->ms_map;
719 	space_map_t *freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
720 	metaslab_group_t *mg = msp->ms_group;
721 	vdev_t *vd = mg->mg_vd;
722 	int t;
723 
724 	mutex_enter(&msp->ms_lock);
725 
726 	/*
727 	 * If this metaslab is just becoming available, initialize its
728 	 * allocmaps and freemaps and add its capacity to the vdev.
729 	 */
730 	if (freed_map->sm_size == 0) {
731 		for (t = 0; t < TXG_SIZE; t++) {
732 			space_map_create(&msp->ms_allocmap[t], sm->sm_start,
733 			    sm->sm_size, sm->sm_shift, sm->sm_lock);
734 			space_map_create(&msp->ms_freemap[t], sm->sm_start,
735 			    sm->sm_size, sm->sm_shift, sm->sm_lock);
736 		}
737 		vdev_space_update(vd, sm->sm_size, 0, B_TRUE);
738 	}
739 
740 	vdev_space_update(vd, 0, smosync->smo_alloc - smo->smo_alloc, B_TRUE);
741 
742 	ASSERT(msp->ms_allocmap[txg & TXG_MASK].sm_space == 0);
743 	ASSERT(msp->ms_freemap[txg & TXG_MASK].sm_space == 0);
744 
745 	/*
746 	 * If there's a space_map_load() in progress, wait for it to complete
747 	 * so that we have a consistent view of the in-core space map.
748 	 * Then, add everything we freed in this txg to the map.
749 	 */
750 	space_map_load_wait(sm);
751 	space_map_vacate(freed_map, sm->sm_loaded ? space_map_free : NULL, sm);
752 
753 	*smo = *smosync;
754 
755 	/*
756 	 * If the map is loaded but no longer active, evict it as soon as all
757 	 * future allocations have synced.  (If we unloaded it now and then
758 	 * loaded a moment later, the map wouldn't reflect those allocations.)
759 	 */
760 	if (sm->sm_loaded && (msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
761 		int evictable = 1;
762 
763 		for (t = 1; t < TXG_CONCURRENT_STATES; t++)
764 			if (msp->ms_allocmap[(txg + t) & TXG_MASK].sm_space)
765 				evictable = 0;
766 
767 		if (evictable)
768 			space_map_unload(sm);
769 	}
770 
771 	metaslab_group_sort(mg, msp, metaslab_weight(msp));
772 
773 	mutex_exit(&msp->ms_lock);
774 }
775 
776 static uint64_t
777 metaslab_distance(metaslab_t *msp, dva_t *dva)
778 {
779 	uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
780 	uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
781 	uint64_t start = msp->ms_map.sm_start >> ms_shift;
782 
783 	if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
784 		return (1ULL << 63);
785 
786 	if (offset < start)
787 		return ((start - offset) << ms_shift);
788 	if (offset > start)
789 		return ((offset - start) << ms_shift);
790 	return (0);
791 }
792 
793 static uint64_t
794 metaslab_group_alloc(metaslab_group_t *mg, uint64_t size, uint64_t txg,
795     uint64_t min_distance, dva_t *dva, int d)
796 {
797 	metaslab_t *msp = NULL;
798 	uint64_t offset = -1ULL;
799 	avl_tree_t *t = &mg->mg_metaslab_tree;
800 	uint64_t activation_weight;
801 	uint64_t target_distance;
802 	int i;
803 
804 	activation_weight = METASLAB_WEIGHT_PRIMARY;
805 	for (i = 0; i < d; i++) {
806 		if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
807 			activation_weight = METASLAB_WEIGHT_SECONDARY;
808 			break;
809 		}
810 	}
811 
812 	for (;;) {
813 		boolean_t was_active;
814 
815 		mutex_enter(&mg->mg_lock);
816 		for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
817 			if (msp->ms_weight < size) {
818 				mutex_exit(&mg->mg_lock);
819 				return (-1ULL);
820 			}
821 
822 			was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
823 			if (activation_weight == METASLAB_WEIGHT_PRIMARY)
824 				break;
825 
826 			target_distance = min_distance +
827 			    (msp->ms_smo.smo_alloc ? 0 : min_distance >> 1);
828 
829 			for (i = 0; i < d; i++)
830 				if (metaslab_distance(msp, &dva[i]) <
831 				    target_distance)
832 					break;
833 			if (i == d)
834 				break;
835 		}
836 		mutex_exit(&mg->mg_lock);
837 		if (msp == NULL)
838 			return (-1ULL);
839 
840 		mutex_enter(&msp->ms_lock);
841 
842 		/*
843 		 * Ensure that the metaslab we have selected is still
844 		 * capable of handling our request. It's possible that
845 		 * another thread may have changed the weight while we
846 		 * were blocked on the metaslab lock.
847 		 */
848 		if (msp->ms_weight < size || (was_active &&
849 		    !(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
850 		    activation_weight == METASLAB_WEIGHT_PRIMARY)) {
851 			mutex_exit(&msp->ms_lock);
852 			continue;
853 		}
854 
855 		if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
856 		    activation_weight == METASLAB_WEIGHT_PRIMARY) {
857 			metaslab_passivate(msp,
858 			    msp->ms_weight & ~METASLAB_ACTIVE_MASK);
859 			mutex_exit(&msp->ms_lock);
860 			continue;
861 		}
862 
863 		if (metaslab_activate(msp, activation_weight, size) != 0) {
864 			mutex_exit(&msp->ms_lock);
865 			continue;
866 		}
867 
868 		if ((offset = space_map_alloc(&msp->ms_map, size)) != -1ULL)
869 			break;
870 
871 		metaslab_passivate(msp, size - 1);
872 
873 		mutex_exit(&msp->ms_lock);
874 	}
875 
876 	if (msp->ms_allocmap[txg & TXG_MASK].sm_space == 0)
877 		vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
878 
879 	space_map_add(&msp->ms_allocmap[txg & TXG_MASK], offset, size);
880 
881 	mutex_exit(&msp->ms_lock);
882 
883 	return (offset);
884 }
885 
886 /*
887  * Allocate a block for the specified i/o.
888  */
889 static int
890 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
891     dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
892 {
893 	metaslab_group_t *mg, *rotor;
894 	vdev_t *vd;
895 	int dshift = 3;
896 	int all_zero;
897 	int zio_lock = B_FALSE;
898 	boolean_t allocatable;
899 	uint64_t offset = -1ULL;
900 	uint64_t asize;
901 	uint64_t distance;
902 
903 	ASSERT(!DVA_IS_VALID(&dva[d]));
904 
905 	/*
906 	 * For testing, make some blocks above a certain size be gang blocks.
907 	 */
908 	if (psize >= metaslab_gang_bang && (lbolt & 3) == 0)
909 		return (ENOSPC);
910 
911 	/*
912 	 * Start at the rotor and loop through all mgs until we find something.
913 	 * Note that there's no locking on mc_rotor or mc_allocated because
914 	 * nothing actually breaks if we miss a few updates -- we just won't
915 	 * allocate quite as evenly.  It all balances out over time.
916 	 *
917 	 * If we are doing ditto or log blocks, try to spread them across
918 	 * consecutive vdevs.  If we're forced to reuse a vdev before we've
919 	 * allocated all of our ditto blocks, then try and spread them out on
920 	 * that vdev as much as possible.  If it turns out to not be possible,
921 	 * gradually lower our standards until anything becomes acceptable.
922 	 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
923 	 * gives us hope of containing our fault domains to something we're
924 	 * able to reason about.  Otherwise, any two top-level vdev failures
925 	 * will guarantee the loss of data.  With consecutive allocation,
926 	 * only two adjacent top-level vdev failures will result in data loss.
927 	 *
928 	 * If we are doing gang blocks (hintdva is non-NULL), try to keep
929 	 * ourselves on the same vdev as our gang block header.  That
930 	 * way, we can hope for locality in vdev_cache, plus it makes our
931 	 * fault domains something tractable.
932 	 */
933 	if (hintdva) {
934 		vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
935 		if (flags & METASLAB_HINTBP_AVOID)
936 			mg = vd->vdev_mg->mg_next;
937 		else
938 			mg = vd->vdev_mg;
939 	} else if (d != 0) {
940 		vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
941 		mg = vd->vdev_mg->mg_next;
942 	} else {
943 		mg = mc->mc_rotor;
944 	}
945 
946 	/*
947 	 * If the hint put us into the wrong class, just follow the rotor.
948 	 */
949 	if (mg->mg_class != mc)
950 		mg = mc->mc_rotor;
951 
952 	rotor = mg;
953 top:
954 	all_zero = B_TRUE;
955 	do {
956 		vd = mg->mg_vd;
957 
958 		/*
959 		 * Don't allocate from faulted devices.
960 		 */
961 		if (zio_lock) {
962 			spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
963 			allocatable = vdev_allocatable(vd);
964 			spa_config_exit(spa, SCL_ZIO, FTAG);
965 		} else {
966 			allocatable = vdev_allocatable(vd);
967 		}
968 		if (!allocatable)
969 			goto next;
970 
971 		/*
972 		 * Avoid writing single-copy data to a failing vdev
973 		 */
974 		if ((vd->vdev_stat.vs_write_errors > 0 ||
975 		    vd->vdev_state < VDEV_STATE_HEALTHY) &&
976 		    d == 0 && dshift == 3) {
977 			all_zero = B_FALSE;
978 			goto next;
979 		}
980 
981 		ASSERT(mg->mg_class == mc);
982 
983 		distance = vd->vdev_asize >> dshift;
984 		if (distance <= (1ULL << vd->vdev_ms_shift))
985 			distance = 0;
986 		else
987 			all_zero = B_FALSE;
988 
989 		asize = vdev_psize_to_asize(vd, psize);
990 		ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
991 
992 		offset = metaslab_group_alloc(mg, asize, txg, distance, dva, d);
993 		if (offset != -1ULL) {
994 			/*
995 			 * If we've just selected this metaslab group,
996 			 * figure out whether the corresponding vdev is
997 			 * over- or under-used relative to the pool,
998 			 * and set an allocation bias to even it out.
999 			 */
1000 			if (mc->mc_allocated == 0) {
1001 				vdev_stat_t *vs = &vd->vdev_stat;
1002 				uint64_t alloc, space;
1003 				int64_t vu, su;
1004 
1005 				alloc = spa_get_alloc(spa);
1006 				space = spa_get_space(spa);
1007 
1008 				/*
1009 				 * Determine percent used in units of 0..1024.
1010 				 * (This is just to avoid floating point.)
1011 				 */
1012 				vu = (vs->vs_alloc << 10) / (vs->vs_space + 1);
1013 				su = (alloc << 10) / (space + 1);
1014 
1015 				/*
1016 				 * Bias by at most +/- 25% of the aliquot.
1017 				 */
1018 				mg->mg_bias = ((su - vu) *
1019 				    (int64_t)mg->mg_aliquot) / (1024 * 4);
1020 			}
1021 
1022 			if (atomic_add_64_nv(&mc->mc_allocated, asize) >=
1023 			    mg->mg_aliquot + mg->mg_bias) {
1024 				mc->mc_rotor = mg->mg_next;
1025 				mc->mc_allocated = 0;
1026 			}
1027 
1028 			DVA_SET_VDEV(&dva[d], vd->vdev_id);
1029 			DVA_SET_OFFSET(&dva[d], offset);
1030 			DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
1031 			DVA_SET_ASIZE(&dva[d], asize);
1032 
1033 			return (0);
1034 		}
1035 next:
1036 		mc->mc_rotor = mg->mg_next;
1037 		mc->mc_allocated = 0;
1038 	} while ((mg = mg->mg_next) != rotor);
1039 
1040 	if (!all_zero) {
1041 		dshift++;
1042 		ASSERT(dshift < 64);
1043 		goto top;
1044 	}
1045 
1046 	if (!allocatable && !zio_lock) {
1047 		dshift = 3;
1048 		zio_lock = B_TRUE;
1049 		goto top;
1050 	}
1051 
1052 	bzero(&dva[d], sizeof (dva_t));
1053 
1054 	return (ENOSPC);
1055 }
1056 
1057 /*
1058  * Free the block represented by DVA in the context of the specified
1059  * transaction group.
1060  */
1061 static void
1062 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
1063 {
1064 	uint64_t vdev = DVA_GET_VDEV(dva);
1065 	uint64_t offset = DVA_GET_OFFSET(dva);
1066 	uint64_t size = DVA_GET_ASIZE(dva);
1067 	vdev_t *vd;
1068 	metaslab_t *msp;
1069 
1070 	ASSERT(DVA_IS_VALID(dva));
1071 
1072 	if (txg > spa_freeze_txg(spa))
1073 		return;
1074 
1075 	if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
1076 	    (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
1077 		cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
1078 		    (u_longlong_t)vdev, (u_longlong_t)offset);
1079 		ASSERT(0);
1080 		return;
1081 	}
1082 
1083 	msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1084 
1085 	if (DVA_GET_GANG(dva))
1086 		size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1087 
1088 	mutex_enter(&msp->ms_lock);
1089 
1090 	if (now) {
1091 		space_map_remove(&msp->ms_allocmap[txg & TXG_MASK],
1092 		    offset, size);
1093 		space_map_free(&msp->ms_map, offset, size);
1094 	} else {
1095 		if (msp->ms_freemap[txg & TXG_MASK].sm_space == 0)
1096 			vdev_dirty(vd, VDD_METASLAB, msp, txg);
1097 		space_map_add(&msp->ms_freemap[txg & TXG_MASK], offset, size);
1098 	}
1099 
1100 	mutex_exit(&msp->ms_lock);
1101 }
1102 
1103 /*
1104  * Intent log support: upon opening the pool after a crash, notify the SPA
1105  * of blocks that the intent log has allocated for immediate write, but
1106  * which are still considered free by the SPA because the last transaction
1107  * group didn't commit yet.
1108  */
1109 static int
1110 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
1111 {
1112 	uint64_t vdev = DVA_GET_VDEV(dva);
1113 	uint64_t offset = DVA_GET_OFFSET(dva);
1114 	uint64_t size = DVA_GET_ASIZE(dva);
1115 	vdev_t *vd;
1116 	metaslab_t *msp;
1117 	int error;
1118 
1119 	ASSERT(DVA_IS_VALID(dva));
1120 
1121 	if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
1122 	    (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
1123 		return (ENXIO);
1124 
1125 	msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1126 
1127 	if (DVA_GET_GANG(dva))
1128 		size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1129 
1130 	mutex_enter(&msp->ms_lock);
1131 
1132 	error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY, 0);
1133 	if (error || txg == 0) {	/* txg == 0 indicates dry run */
1134 		mutex_exit(&msp->ms_lock);
1135 		return (error);
1136 	}
1137 
1138 	space_map_claim(&msp->ms_map, offset, size);
1139 
1140 	if (spa_writeable(spa)) {	/* don't dirty if we're zdb(1M) */
1141 		if (msp->ms_allocmap[txg & TXG_MASK].sm_space == 0)
1142 			vdev_dirty(vd, VDD_METASLAB, msp, txg);
1143 		space_map_add(&msp->ms_allocmap[txg & TXG_MASK], offset, size);
1144 	}
1145 
1146 	mutex_exit(&msp->ms_lock);
1147 
1148 	return (0);
1149 }
1150 
1151 int
1152 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
1153     int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
1154 {
1155 	dva_t *dva = bp->blk_dva;
1156 	dva_t *hintdva = hintbp->blk_dva;
1157 	int error = 0;
1158 
1159 	ASSERT(bp->blk_birth == 0);
1160 
1161 	spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
1162 
1163 	if (mc->mc_rotor == NULL) {	/* no vdevs in this class */
1164 		spa_config_exit(spa, SCL_ALLOC, FTAG);
1165 		return (ENOSPC);
1166 	}
1167 
1168 	ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
1169 	ASSERT(BP_GET_NDVAS(bp) == 0);
1170 	ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
1171 
1172 	for (int d = 0; d < ndvas; d++) {
1173 		error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
1174 		    txg, flags);
1175 		if (error) {
1176 			for (d--; d >= 0; d--) {
1177 				metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
1178 				bzero(&dva[d], sizeof (dva_t));
1179 			}
1180 			spa_config_exit(spa, SCL_ALLOC, FTAG);
1181 			return (error);
1182 		}
1183 	}
1184 	ASSERT(error == 0);
1185 	ASSERT(BP_GET_NDVAS(bp) == ndvas);
1186 
1187 	spa_config_exit(spa, SCL_ALLOC, FTAG);
1188 
1189 	bp->blk_birth = txg;
1190 
1191 	return (0);
1192 }
1193 
1194 void
1195 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
1196 {
1197 	const dva_t *dva = bp->blk_dva;
1198 	int ndvas = BP_GET_NDVAS(bp);
1199 
1200 	ASSERT(!BP_IS_HOLE(bp));
1201 	ASSERT(!now || bp->blk_birth >= spa->spa_syncing_txg);
1202 
1203 	spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
1204 
1205 	for (int d = 0; d < ndvas; d++)
1206 		metaslab_free_dva(spa, &dva[d], txg, now);
1207 
1208 	spa_config_exit(spa, SCL_FREE, FTAG);
1209 }
1210 
1211 int
1212 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
1213 {
1214 	const dva_t *dva = bp->blk_dva;
1215 	int ndvas = BP_GET_NDVAS(bp);
1216 	int error = 0;
1217 
1218 	ASSERT(!BP_IS_HOLE(bp));
1219 
1220 	if (txg != 0) {
1221 		/*
1222 		 * First do a dry run to make sure all DVAs are claimable,
1223 		 * so we don't have to unwind from partial failures below.
1224 		 */
1225 		if ((error = metaslab_claim(spa, bp, 0)) != 0)
1226 			return (error);
1227 	}
1228 
1229 	spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
1230 
1231 	for (int d = 0; d < ndvas; d++)
1232 		if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
1233 			break;
1234 
1235 	spa_config_exit(spa, SCL_ALLOC, FTAG);
1236 
1237 	ASSERT(error == 0 || txg == 0);
1238 
1239 	return (error);
1240 }
1241