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
46uint64_t metaslab_aliquot = 512ULL << 10;
47uint64_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 */
54int zfs_mg_alloc_failures = 0;
55
56SYSCTL_DECL(_vfs_zfs);
57SYSCTL_INT(_vfs_zfs, OID_AUTO, mg_alloc_failures, CTLFLAG_RDTUN,
58 &zfs_mg_alloc_failures, 0,
59 "Number of allowed allocation failures per vdev");
60TUNABLE_INT("vfs.zfs.mg_alloc_failures", &zfs_mg_alloc_failures);
61
62/*
63 * Metaslab debugging: when set, keeps all space maps in core to verify frees.
64 */
65static int metaslab_debug = 0;
66
67/*
68 * Minimum size which forces the dynamic allocator to change
69 * it's allocation strategy. Once the space map cannot satisfy
70 * an allocation of this size then it switches to using more
71 * aggressive strategy (i.e search by size rather than offset).
72 */
73uint64_t metaslab_df_alloc_threshold = SPA_MAXBLOCKSIZE;
74
75/*
76 * The minimum free space, in percent, which must be available
77 * in a space map to continue allocations in a first-fit fashion.
78 * Once the space_map's free space drops below this level we dynamically
79 * switch to using best-fit allocations.
80 */
81int metaslab_df_free_pct = 4;
82
83/*
84 * A metaslab is considered "free" if it contains a contiguous
85 * segment which is greater than metaslab_min_alloc_size.
86 */
87uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
88
89/*
90 * Max number of space_maps to prefetch.
91 */
92int metaslab_prefetch_limit = SPA_DVAS_PER_BP;
93
94/*
95 * Percentage bonus multiplier for metaslabs that are in the bonus area.
96 */
97int metaslab_smo_bonus_pct = 150;
98
99/*
100 * ==========================================================================
101 * Metaslab classes
102 * ==========================================================================
103 */
104metaslab_class_t *
105metaslab_class_create(spa_t *spa, space_map_ops_t *ops)
106{
107 metaslab_class_t *mc;
108
109 mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
110
111 mc->mc_spa = spa;
112 mc->mc_rotor = NULL;
113 mc->mc_ops = ops;
114
115 return (mc);
116}
117
118void
119metaslab_class_destroy(metaslab_class_t *mc)
120{
121 ASSERT(mc->mc_rotor == NULL);
122 ASSERT(mc->mc_alloc == 0);
123 ASSERT(mc->mc_deferred == 0);
124 ASSERT(mc->mc_space == 0);
125 ASSERT(mc->mc_dspace == 0);
126
127 kmem_free(mc, sizeof (metaslab_class_t));
128}
129
130int
131metaslab_class_validate(metaslab_class_t *mc)
132{
133 metaslab_group_t *mg;
134 vdev_t *vd;
135
136 /*
137 * Must hold one of the spa_config locks.
138 */
139 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
140 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
141
142 if ((mg = mc->mc_rotor) == NULL)
143 return (0);
144
145 do {
146 vd = mg->mg_vd;
147 ASSERT(vd->vdev_mg != NULL);
148 ASSERT3P(vd->vdev_top, ==, vd);
149 ASSERT3P(mg->mg_class, ==, mc);
150 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
151 } while ((mg = mg->mg_next) != mc->mc_rotor);
152
153 return (0);
154}
155
156void
157metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
158 int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
159{
160 atomic_add_64(&mc->mc_alloc, alloc_delta);
161 atomic_add_64(&mc->mc_deferred, defer_delta);
162 atomic_add_64(&mc->mc_space, space_delta);
163 atomic_add_64(&mc->mc_dspace, dspace_delta);
164}
165
166uint64_t
167metaslab_class_get_alloc(metaslab_class_t *mc)
168{
169 return (mc->mc_alloc);
170}
171
172uint64_t
173metaslab_class_get_deferred(metaslab_class_t *mc)
174{
175 return (mc->mc_deferred);
176}
177
178uint64_t
179metaslab_class_get_space(metaslab_class_t *mc)
180{
181 return (mc->mc_space);
182}
183
184uint64_t
185metaslab_class_get_dspace(metaslab_class_t *mc)
186{
187 return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
188}
189
190/*
191 * ==========================================================================
192 * Metaslab groups
193 * ==========================================================================
194 */
195static int
196metaslab_compare(const void *x1, const void *x2)
197{
198 const metaslab_t *m1 = x1;
199 const metaslab_t *m2 = x2;
200
201 if (m1->ms_weight < m2->ms_weight)
202 return (1);
203 if (m1->ms_weight > m2->ms_weight)
204 return (-1);
205
206 /*
207 * If the weights are identical, use the offset to force uniqueness.
208 */
209 if (m1->ms_map.sm_start < m2->ms_map.sm_start)
210 return (-1);
211 if (m1->ms_map.sm_start > m2->ms_map.sm_start)
212 return (1);
213
214 ASSERT3P(m1, ==, m2);
215
216 return (0);
217}
218
219metaslab_group_t *
220metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
221{
222 metaslab_group_t *mg;
223
224 mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
225 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
226 avl_create(&mg->mg_metaslab_tree, metaslab_compare,
227 sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
228 mg->mg_vd = vd;
229 mg->mg_class = mc;
230 mg->mg_activation_count = 0;
231
232 return (mg);
233}
234
235void
236metaslab_group_destroy(metaslab_group_t *mg)
237{
238 ASSERT(mg->mg_prev == NULL);
239 ASSERT(mg->mg_next == NULL);
240 /*
241 * We may have gone below zero with the activation count
242 * either because we never activated in the first place or
243 * because we're done, and possibly removing the vdev.
244 */
245 ASSERT(mg->mg_activation_count <= 0);
246
247 avl_destroy(&mg->mg_metaslab_tree);
248 mutex_destroy(&mg->mg_lock);
249 kmem_free(mg, sizeof (metaslab_group_t));
250}
251
252void
253metaslab_group_activate(metaslab_group_t *mg)
254{
255 metaslab_class_t *mc = mg->mg_class;
256 metaslab_group_t *mgprev, *mgnext;
257
258 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
259
260 ASSERT(mc->mc_rotor != mg);
261 ASSERT(mg->mg_prev == NULL);
262 ASSERT(mg->mg_next == NULL);
263 ASSERT(mg->mg_activation_count <= 0);
264
265 if (++mg->mg_activation_count <= 0)
266 return;
267
268 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
269
270 if ((mgprev = mc->mc_rotor) == NULL) {
271 mg->mg_prev = mg;
272 mg->mg_next = mg;
273 } else {
274 mgnext = mgprev->mg_next;
275 mg->mg_prev = mgprev;
276 mg->mg_next = mgnext;
277 mgprev->mg_next = mg;
278 mgnext->mg_prev = mg;
279 }
280 mc->mc_rotor = mg;
281}
282
283void
284metaslab_group_passivate(metaslab_group_t *mg)
285{
286 metaslab_class_t *mc = mg->mg_class;
287 metaslab_group_t *mgprev, *mgnext;
288
289 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
290
291 if (--mg->mg_activation_count != 0) {
292 ASSERT(mc->mc_rotor != mg);
293 ASSERT(mg->mg_prev == NULL);
294 ASSERT(mg->mg_next == NULL);
295 ASSERT(mg->mg_activation_count < 0);
296 return;
297 }
298
299 mgprev = mg->mg_prev;
300 mgnext = mg->mg_next;
301
302 if (mg == mgnext) {
303 mc->mc_rotor = NULL;
304 } else {
305 mc->mc_rotor = mgnext;
306 mgprev->mg_next = mgnext;
307 mgnext->mg_prev = mgprev;
308 }
309
310 mg->mg_prev = NULL;
311 mg->mg_next = NULL;
312}
313
314static void
315metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
316{
317 mutex_enter(&mg->mg_lock);
318 ASSERT(msp->ms_group == NULL);
319 msp->ms_group = mg;
320 msp->ms_weight = 0;
321 avl_add(&mg->mg_metaslab_tree, msp);
322 mutex_exit(&mg->mg_lock);
323}
324
325static void
326metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
327{
328 mutex_enter(&mg->mg_lock);
329 ASSERT(msp->ms_group == mg);
330 avl_remove(&mg->mg_metaslab_tree, msp);
331 msp->ms_group = NULL;
332 mutex_exit(&mg->mg_lock);
333}
334
335static void
336metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
337{
338 /*
339 * Although in principle the weight can be any value, in
340 * practice we do not use values in the range [1, 510].
341 */
342 ASSERT(weight >= SPA_MINBLOCKSIZE-1 || weight == 0);
343 ASSERT(MUTEX_HELD(&msp->ms_lock));
344
345 mutex_enter(&mg->mg_lock);
346 ASSERT(msp->ms_group == mg);
347 avl_remove(&mg->mg_metaslab_tree, msp);
348 msp->ms_weight = weight;
349 avl_add(&mg->mg_metaslab_tree, msp);
350 mutex_exit(&mg->mg_lock);
351}
352
353/*
354 * ==========================================================================
355 * Common allocator routines
356 * ==========================================================================
357 */
358static int
359metaslab_segsize_compare(const void *x1, const void *x2)
360{
361 const space_seg_t *s1 = x1;
362 const space_seg_t *s2 = x2;
363 uint64_t ss_size1 = s1->ss_end - s1->ss_start;
364 uint64_t ss_size2 = s2->ss_end - s2->ss_start;
365
366 if (ss_size1 < ss_size2)
367 return (-1);
368 if (ss_size1 > ss_size2)
369 return (1);
370
371 if (s1->ss_start < s2->ss_start)
372 return (-1);
373 if (s1->ss_start > s2->ss_start)
374 return (1);
375
376 return (0);
377}
378
379/*
380 * This is a helper function that can be used by the allocator to find
381 * a suitable block to allocate. This will search the specified AVL
382 * tree looking for a block that matches the specified criteria.
383 */
384static uint64_t
385metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
386 uint64_t align)
387{
388 space_seg_t *ss, ssearch;
389 avl_index_t where;
390
391 ssearch.ss_start = *cursor;
392 ssearch.ss_end = *cursor + size;
393
394 ss = avl_find(t, &ssearch, &where);
395 if (ss == NULL)
396 ss = avl_nearest(t, where, AVL_AFTER);
397
398 while (ss != NULL) {
399 uint64_t offset = P2ROUNDUP(ss->ss_start, align);
400
401 if (offset + size <= ss->ss_end) {
402 *cursor = offset + size;
403 return (offset);
404 }
405 ss = AVL_NEXT(t, ss);
406 }
407
408 /*
409 * If we know we've searched the whole map (*cursor == 0), give up.
410 * Otherwise, reset the cursor to the beginning and try again.
411 */
412 if (*cursor == 0)
413 return (-1ULL);
414
415 *cursor = 0;
416 return (metaslab_block_picker(t, cursor, size, align));
417}
418
419static void
420metaslab_pp_load(space_map_t *sm)
421{
422 space_seg_t *ss;
423
424 ASSERT(sm->sm_ppd == NULL);
425 sm->sm_ppd = kmem_zalloc(64 * sizeof (uint64_t), KM_SLEEP);
426
427 sm->sm_pp_root = kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
428 avl_create(sm->sm_pp_root, metaslab_segsize_compare,
429 sizeof (space_seg_t), offsetof(struct space_seg, ss_pp_node));
430
431 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
432 avl_add(sm->sm_pp_root, ss);
433}
434
435static void
436metaslab_pp_unload(space_map_t *sm)
437{
438 void *cookie = NULL;
439
440 kmem_free(sm->sm_ppd, 64 * sizeof (uint64_t));
441 sm->sm_ppd = NULL;
442
443 while (avl_destroy_nodes(sm->sm_pp_root, &cookie) != NULL) {
444 /* tear down the tree */
445 }
446
447 avl_destroy(sm->sm_pp_root);
448 kmem_free(sm->sm_pp_root, sizeof (avl_tree_t));
449 sm->sm_pp_root = NULL;
450}
451
452/* ARGSUSED */
453static void
454metaslab_pp_claim(space_map_t *sm, uint64_t start, uint64_t size)
455{
456 /* No need to update cursor */
457}
458
459/* ARGSUSED */
460static void
461metaslab_pp_free(space_map_t *sm, uint64_t start, uint64_t size)
462{
463 /* No need to update cursor */
464}
465
466/*
467 * Return the maximum contiguous segment within the metaslab.
468 */
469uint64_t
470metaslab_pp_maxsize(space_map_t *sm)
471{
472 avl_tree_t *t = sm->sm_pp_root;
473 space_seg_t *ss;
474
475 if (t == NULL || (ss = avl_last(t)) == NULL)
476 return (0ULL);
477
478 return (ss->ss_end - ss->ss_start);
479}
480
481/*
482 * ==========================================================================
483 * The first-fit block allocator
484 * ==========================================================================
485 */
486static uint64_t
487metaslab_ff_alloc(space_map_t *sm, uint64_t size)
488{
489 avl_tree_t *t = &sm->sm_root;
490 uint64_t align = size & -size;
491 uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
492
493 return (metaslab_block_picker(t, cursor, size, align));
494}
495
496/* ARGSUSED */
497boolean_t
498metaslab_ff_fragmented(space_map_t *sm)
499{
500 return (B_TRUE);
501}
502
503static space_map_ops_t metaslab_ff_ops = {
504 metaslab_pp_load,
505 metaslab_pp_unload,
506 metaslab_ff_alloc,
507 metaslab_pp_claim,
508 metaslab_pp_free,
509 metaslab_pp_maxsize,
510 metaslab_ff_fragmented
511};
512
513/*
514 * ==========================================================================
515 * Dynamic block allocator -
516 * Uses the first fit allocation scheme until space get low and then
517 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
518 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
519 * ==========================================================================
520 */
521static uint64_t
522metaslab_df_alloc(space_map_t *sm, uint64_t size)
523{
524 avl_tree_t *t = &sm->sm_root;
525 uint64_t align = size & -size;
526 uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
527 uint64_t max_size = metaslab_pp_maxsize(sm);
528 int free_pct = sm->sm_space * 100 / sm->sm_size;
529
530 ASSERT(MUTEX_HELD(sm->sm_lock));
531 ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
532
533 if (max_size < size)
534 return (-1ULL);
535
536 /*
537 * If we're running low on space switch to using the size
538 * sorted AVL tree (best-fit).
539 */
540 if (max_size < metaslab_df_alloc_threshold ||
541 free_pct < metaslab_df_free_pct) {
542 t = sm->sm_pp_root;
543 *cursor = 0;
544 }
545
546 return (metaslab_block_picker(t, cursor, size, 1ULL));
547}
548
549static boolean_t
550metaslab_df_fragmented(space_map_t *sm)
551{
552 uint64_t max_size = metaslab_pp_maxsize(sm);
553 int free_pct = sm->sm_space * 100 / sm->sm_size;
554
555 if (max_size >= metaslab_df_alloc_threshold &&
556 free_pct >= metaslab_df_free_pct)
557 return (B_FALSE);
558
559 return (B_TRUE);
560}
561
562static space_map_ops_t metaslab_df_ops = {
563 metaslab_pp_load,
564 metaslab_pp_unload,
565 metaslab_df_alloc,
566 metaslab_pp_claim,
567 metaslab_pp_free,
568 metaslab_pp_maxsize,
569 metaslab_df_fragmented
570};
571
572/*
573 * ==========================================================================
574 * Other experimental allocators
575 * ==========================================================================
576 */
577static uint64_t
578metaslab_cdf_alloc(space_map_t *sm, uint64_t size)
579{
580 avl_tree_t *t = &sm->sm_root;
581 uint64_t *cursor = (uint64_t *)sm->sm_ppd;
582 uint64_t *extent_end = (uint64_t *)sm->sm_ppd + 1;
583 uint64_t max_size = metaslab_pp_maxsize(sm);
584 uint64_t rsize = size;
585 uint64_t offset = 0;
586
587 ASSERT(MUTEX_HELD(sm->sm_lock));
588 ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
589
590 if (max_size < size)
591 return (-1ULL);
592
593 ASSERT3U(*extent_end, >=, *cursor);
594
595 /*
596 * If we're running low on space switch to using the size
597 * sorted AVL tree (best-fit).
598 */
599 if ((*cursor + size) > *extent_end) {
600
601 t = sm->sm_pp_root;
602 *cursor = *extent_end = 0;
603
604 if (max_size > 2 * SPA_MAXBLOCKSIZE)
605 rsize = MIN(metaslab_min_alloc_size, max_size);
606 offset = metaslab_block_picker(t, extent_end, rsize, 1ULL);
607 if (offset != -1)
608 *cursor = offset + size;
609 } else {
610 offset = metaslab_block_picker(t, cursor, rsize, 1ULL);
611 }
612 ASSERT3U(*cursor, <=, *extent_end);
613 return (offset);
614}
615
616static boolean_t
617metaslab_cdf_fragmented(space_map_t *sm)
618{
619 uint64_t max_size = metaslab_pp_maxsize(sm);
620
621 if (max_size > (metaslab_min_alloc_size * 10))
622 return (B_FALSE);
623 return (B_TRUE);
624}
625
626static space_map_ops_t metaslab_cdf_ops = {
627 metaslab_pp_load,
628 metaslab_pp_unload,
629 metaslab_cdf_alloc,
630 metaslab_pp_claim,
631 metaslab_pp_free,
632 metaslab_pp_maxsize,
633 metaslab_cdf_fragmented
634};
635
636uint64_t metaslab_ndf_clump_shift = 4;
637
638static uint64_t
639metaslab_ndf_alloc(space_map_t *sm, uint64_t size)
640{
641 avl_tree_t *t = &sm->sm_root;
642 avl_index_t where;
643 space_seg_t *ss, ssearch;
644 uint64_t hbit = highbit(size);
645 uint64_t *cursor = (uint64_t *)sm->sm_ppd + hbit - 1;
646 uint64_t max_size = metaslab_pp_maxsize(sm);
647
648 ASSERT(MUTEX_HELD(sm->sm_lock));
649 ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
650
651 if (max_size < size)
652 return (-1ULL);
653
654 ssearch.ss_start = *cursor;
655 ssearch.ss_end = *cursor + size;
656
657 ss = avl_find(t, &ssearch, &where);
658 if (ss == NULL || (ss->ss_start + size > ss->ss_end)) {
659 t = sm->sm_pp_root;
660
661 ssearch.ss_start = 0;
662 ssearch.ss_end = MIN(max_size,
663 1ULL << (hbit + metaslab_ndf_clump_shift));
664 ss = avl_find(t, &ssearch, &where);
665 if (ss == NULL)
666 ss = avl_nearest(t, where, AVL_AFTER);
667 ASSERT(ss != NULL);
668 }
669
670 if (ss != NULL) {
671 if (ss->ss_start + size <= ss->ss_end) {
672 *cursor = ss->ss_start + size;
673 return (ss->ss_start);
674 }
675 }
676 return (-1ULL);
677}
678
679static boolean_t
680metaslab_ndf_fragmented(space_map_t *sm)
681{
682 uint64_t max_size = metaslab_pp_maxsize(sm);
683
684 if (max_size > (metaslab_min_alloc_size << metaslab_ndf_clump_shift))
685 return (B_FALSE);
686 return (B_TRUE);
687}
688
689
690static space_map_ops_t metaslab_ndf_ops = {
691 metaslab_pp_load,
692 metaslab_pp_unload,
693 metaslab_ndf_alloc,
694 metaslab_pp_claim,
695 metaslab_pp_free,
696 metaslab_pp_maxsize,
697 metaslab_ndf_fragmented
698};
699
700space_map_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
701
702/*
703 * ==========================================================================
704 * Metaslabs
705 * ==========================================================================
706 */
707metaslab_t *
708metaslab_init(metaslab_group_t *mg, space_map_obj_t *smo,
709 uint64_t start, uint64_t size, uint64_t txg)
710{
711 vdev_t *vd = mg->mg_vd;
712 metaslab_t *msp;
713
714 msp = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
715 mutex_init(&msp->ms_lock, NULL, MUTEX_DEFAULT, NULL);
716
717 msp->ms_smo_syncing = *smo;
718
719 /*
720 * We create the main space map here, but we don't create the
721 * allocmaps and freemaps until metaslab_sync_done(). This serves
722 * two purposes: it allows metaslab_sync_done() to detect the
723 * addition of new space; and for debugging, it ensures that we'd
724 * data fault on any attempt to use this metaslab before it's ready.
725 */
726 space_map_create(&msp->ms_map, start, size,
727 vd->vdev_ashift, &msp->ms_lock);
728
729 metaslab_group_add(mg, msp);
730
731 if (metaslab_debug && smo->smo_object != 0) {
732 mutex_enter(&msp->ms_lock);
733 VERIFY(space_map_load(&msp->ms_map, mg->mg_class->mc_ops,
734 SM_FREE, smo, spa_meta_objset(vd->vdev_spa)) == 0);
735 mutex_exit(&msp->ms_lock);
736 }
737
738 /*
739 * If we're opening an existing pool (txg == 0) or creating
740 * a new one (txg == TXG_INITIAL), all space is available now.
741 * If we're adding space to an existing pool, the new space
742 * does not become available until after this txg has synced.
743 */
744 if (txg <= TXG_INITIAL)
745 metaslab_sync_done(msp, 0);
746
747 if (txg != 0) {
748 vdev_dirty(vd, 0, NULL, txg);
749 vdev_dirty(vd, VDD_METASLAB, msp, txg);
750 }
751
752 return (msp);
753}
754
755void
756metaslab_fini(metaslab_t *msp)
757{
758 metaslab_group_t *mg = msp->ms_group;
759
760 vdev_space_update(mg->mg_vd,
761 -msp->ms_smo.smo_alloc, 0, -msp->ms_map.sm_size);
762
763 metaslab_group_remove(mg, msp);
764
765 mutex_enter(&msp->ms_lock);
766
767 space_map_unload(&msp->ms_map);
768 space_map_destroy(&msp->ms_map);
769
770 for (int t = 0; t < TXG_SIZE; t++) {
771 space_map_destroy(&msp->ms_allocmap[t]);
772 space_map_destroy(&msp->ms_freemap[t]);
773 }
774
775 for (int t = 0; t < TXG_DEFER_SIZE; t++)
776 space_map_destroy(&msp->ms_defermap[t]);
777
778 ASSERT0(msp->ms_deferspace);
779
780 mutex_exit(&msp->ms_lock);
781 mutex_destroy(&msp->ms_lock);
782
783 kmem_free(msp, sizeof (metaslab_t));
784}
785
786#define METASLAB_WEIGHT_PRIMARY (1ULL << 63)
787#define METASLAB_WEIGHT_SECONDARY (1ULL << 62)
788#define METASLAB_ACTIVE_MASK \
789 (METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
790
791static uint64_t
792metaslab_weight(metaslab_t *msp)
793{
794 metaslab_group_t *mg = msp->ms_group;
795 space_map_t *sm = &msp->ms_map;
796 space_map_obj_t *smo = &msp->ms_smo;
797 vdev_t *vd = mg->mg_vd;
798 uint64_t weight, space;
799
800 ASSERT(MUTEX_HELD(&msp->ms_lock));
801
802 /*
803 * The baseline weight is the metaslab's free space.
804 */
805 space = sm->sm_size - smo->smo_alloc;
806 weight = space;
807
808 /*
809 * Modern disks have uniform bit density and constant angular velocity.
810 * Therefore, the outer recording zones are faster (higher bandwidth)
811 * than the inner zones by the ratio of outer to inner track diameter,
812 * which is typically around 2:1. We account for this by assigning
813 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
814 * In effect, this means that we'll select the metaslab with the most
815 * free bandwidth rather than simply the one with the most free space.
816 */
817 weight = 2 * weight -
818 ((sm->sm_start >> vd->vdev_ms_shift) * weight) / vd->vdev_ms_count;
819 ASSERT(weight >= space && weight <= 2 * space);
820
821 /*
822 * For locality, assign higher weight to metaslabs which have
823 * a lower offset than what we've already activated.
824 */
825 if (sm->sm_start <= mg->mg_bonus_area)
826 weight *= (metaslab_smo_bonus_pct / 100);
827 ASSERT(weight >= space &&
828 weight <= 2 * (metaslab_smo_bonus_pct / 100) * space);
829
830 if (sm->sm_loaded && !sm->sm_ops->smop_fragmented(sm)) {
831 /*
832 * If this metaslab is one we're actively using, adjust its
833 * weight to make it preferable to any inactive metaslab so
834 * we'll polish it off.
835 */
836 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
837 }
838 return (weight);
839}
840
841static void
842metaslab_prefetch(metaslab_group_t *mg)
843{
844 spa_t *spa = mg->mg_vd->vdev_spa;
845 metaslab_t *msp;
846 avl_tree_t *t = &mg->mg_metaslab_tree;
847 int m;
848
849 mutex_enter(&mg->mg_lock);
850
851 /*
852 * Prefetch the next potential metaslabs
853 */
854 for (msp = avl_first(t), m = 0; msp; msp = AVL_NEXT(t, msp), m++) {
855 space_map_t *sm = &msp->ms_map;
856 space_map_obj_t *smo = &msp->ms_smo;
857
858 /* If we have reached our prefetch limit then we're done */
859 if (m >= metaslab_prefetch_limit)
860 break;
861
862 if (!sm->sm_loaded && smo->smo_object != 0) {
863 mutex_exit(&mg->mg_lock);
864 dmu_prefetch(spa_meta_objset(spa), smo->smo_object,
865 0ULL, smo->smo_objsize);
866 mutex_enter(&mg->mg_lock);
867 }
868 }
869 mutex_exit(&mg->mg_lock);
870}
871
872static int
873metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
874{
875 metaslab_group_t *mg = msp->ms_group;
876 space_map_t *sm = &msp->ms_map;
877 space_map_ops_t *sm_ops = msp->ms_group->mg_class->mc_ops;
878
879 ASSERT(MUTEX_HELD(&msp->ms_lock));
880
881 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
882 space_map_load_wait(sm);
883 if (!sm->sm_loaded) {
884 space_map_obj_t *smo = &msp->ms_smo;
885
886 int error = space_map_load(sm, sm_ops, SM_FREE, smo,
887 spa_meta_objset(msp->ms_group->mg_vd->vdev_spa));
888 if (error) {
889 metaslab_group_sort(msp->ms_group, msp, 0);
890 return (error);
891 }
892 for (int t = 0; t < TXG_DEFER_SIZE; t++)
893 space_map_walk(&msp->ms_defermap[t],
894 space_map_claim, sm);
895
896 }
897
898 /*
899 * Track the bonus area as we activate new metaslabs.
900 */
901 if (sm->sm_start > mg->mg_bonus_area) {
902 mutex_enter(&mg->mg_lock);
903 mg->mg_bonus_area = sm->sm_start;
904 mutex_exit(&mg->mg_lock);
905 }
906
907 metaslab_group_sort(msp->ms_group, msp,
908 msp->ms_weight | activation_weight);
909 }
910 ASSERT(sm->sm_loaded);
911 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
912
913 return (0);
914}
915
916static void
917metaslab_passivate(metaslab_t *msp, uint64_t size)
918{
919 /*
920 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
921 * this metaslab again. In that case, it had better be empty,
922 * or we would be leaving space on the table.
923 */
924 ASSERT(size >= SPA_MINBLOCKSIZE || msp->ms_map.sm_space == 0);
925 metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
926 ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
927}
928
929/*
930 * Write a metaslab to disk in the context of the specified transaction group.
931 */
932void
933metaslab_sync(metaslab_t *msp, uint64_t txg)
934{
935 vdev_t *vd = msp->ms_group->mg_vd;
936 spa_t *spa = vd->vdev_spa;
937 objset_t *mos = spa_meta_objset(spa);
938 space_map_t *allocmap = &msp->ms_allocmap[txg & TXG_MASK];
939 space_map_t *freemap = &msp->ms_freemap[txg & TXG_MASK];
940 space_map_t *freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
941 space_map_t *sm = &msp->ms_map;
942 space_map_obj_t *smo = &msp->ms_smo_syncing;
943 dmu_buf_t *db;
944 dmu_tx_t *tx;
945
946 ASSERT(!vd->vdev_ishole);
947
948 if (allocmap->sm_space == 0 && freemap->sm_space == 0)
949 return;
950
951 /*
952 * The only state that can actually be changing concurrently with
953 * metaslab_sync() is the metaslab's ms_map. No other thread can
954 * be modifying this txg's allocmap, freemap, freed_map, or smo.
955 * Therefore, we only hold ms_lock to satify space_map ASSERTs.
956 * We drop it whenever we call into the DMU, because the DMU
957 * can call down to us (e.g. via zio_free()) at any time.
958 */
959
960 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
961
962 if (smo->smo_object == 0) {
963 ASSERT(smo->smo_objsize == 0);
964 ASSERT(smo->smo_alloc == 0);
965 smo->smo_object = dmu_object_alloc(mos,
966 DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
967 DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
968 ASSERT(smo->smo_object != 0);
969 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
970 (sm->sm_start >> vd->vdev_ms_shift),
971 sizeof (uint64_t), &smo->smo_object, tx);
972 }
973
974 mutex_enter(&msp->ms_lock);
975
976 space_map_walk(freemap, space_map_add, freed_map);
977
978 if (sm->sm_loaded && spa_sync_pass(spa) == 1 && smo->smo_objsize >=
979 2 * sizeof (uint64_t) * avl_numnodes(&sm->sm_root)) {
980 /*
981 * The in-core space map representation is twice as compact
982 * as the on-disk one, so it's time to condense the latter
983 * by generating a pure allocmap from first principles.
984 *
985 * This metaslab is 100% allocated,
986 * minus the content of the in-core map (sm),
987 * minus what's been freed this txg (freed_map),
988 * minus deferred frees (ms_defermap[]),
989 * minus allocations from txgs in the future
990 * (because they haven't been committed yet).
991 */
992 space_map_vacate(allocmap, NULL, NULL);
993 space_map_vacate(freemap, NULL, NULL);
994
995 space_map_add(allocmap, allocmap->sm_start, allocmap->sm_size);
996
997 space_map_walk(sm, space_map_remove, allocmap);
998 space_map_walk(freed_map, space_map_remove, allocmap);
999
1000 for (int t = 0; t < TXG_DEFER_SIZE; t++)
1001 space_map_walk(&msp->ms_defermap[t],
1002 space_map_remove, allocmap);
1003
1004 for (int t = 1; t < TXG_CONCURRENT_STATES; t++)
1005 space_map_walk(&msp->ms_allocmap[(txg + t) & TXG_MASK],
1006 space_map_remove, allocmap);
1007
1008 mutex_exit(&msp->ms_lock);
1009 space_map_truncate(smo, mos, tx);
1010 mutex_enter(&msp->ms_lock);
1011 }
1012
1013 space_map_sync(allocmap, SM_ALLOC, smo, mos, tx);
1014 space_map_sync(freemap, SM_FREE, smo, mos, tx);
1015
1016 mutex_exit(&msp->ms_lock);
1017
1018 VERIFY(0 == dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
1019 dmu_buf_will_dirty(db, tx);
1020 ASSERT3U(db->db_size, >=, sizeof (*smo));
1021 bcopy(smo, db->db_data, sizeof (*smo));
1022 dmu_buf_rele(db, FTAG);
1023
1024 dmu_tx_commit(tx);
1025}
1026
1027/*
1028 * Called after a transaction group has completely synced to mark
1029 * all of the metaslab's free space as usable.
1030 */
1031void
1032metaslab_sync_done(metaslab_t *msp, uint64_t txg)
1033{
1034 space_map_obj_t *smo = &msp->ms_smo;
1035 space_map_obj_t *smosync = &msp->ms_smo_syncing;
1036 space_map_t *sm = &msp->ms_map;
1037 space_map_t *freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
1038 space_map_t *defer_map = &msp->ms_defermap[txg % TXG_DEFER_SIZE];
1039 metaslab_group_t *mg = msp->ms_group;
1040 vdev_t *vd = mg->mg_vd;
1041 int64_t alloc_delta, defer_delta;
1042
1043 ASSERT(!vd->vdev_ishole);
1044
1045 mutex_enter(&msp->ms_lock);
1046
1047 /*
1048 * If this metaslab is just becoming available, initialize its
1049 * allocmaps and freemaps and add its capacity to the vdev.
1050 */
1051 if (freed_map->sm_size == 0) {
1052 for (int t = 0; t < TXG_SIZE; t++) {
1053 space_map_create(&msp->ms_allocmap[t], sm->sm_start,
1054 sm->sm_size, sm->sm_shift, sm->sm_lock);
1055 space_map_create(&msp->ms_freemap[t], sm->sm_start,
1056 sm->sm_size, sm->sm_shift, sm->sm_lock);
1057 }
1058
1059 for (int t = 0; t < TXG_DEFER_SIZE; t++)
1060 space_map_create(&msp->ms_defermap[t], sm->sm_start,
1061 sm->sm_size, sm->sm_shift, sm->sm_lock);
1062
1063 vdev_space_update(vd, 0, 0, sm->sm_size);
1064 }
1065
1066 alloc_delta = smosync->smo_alloc - smo->smo_alloc;
1067 defer_delta = freed_map->sm_space - defer_map->sm_space;
1068
1069 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
1070
1071 ASSERT(msp->ms_allocmap[txg & TXG_MASK].sm_space == 0);
1072 ASSERT(msp->ms_freemap[txg & TXG_MASK].sm_space == 0);
1073
1074 /*
1075 * If there's a space_map_load() in progress, wait for it to complete
1076 * so that we have a consistent view of the in-core space map.
1077 * Then, add defer_map (oldest deferred frees) to this map and
1078 * transfer freed_map (this txg's frees) to defer_map.
1079 */
1080 space_map_load_wait(sm);
1081 space_map_vacate(defer_map, sm->sm_loaded ? space_map_free : NULL, sm);
1082 space_map_vacate(freed_map, space_map_add, defer_map);
1083
1084 *smo = *smosync;
1085
1086 msp->ms_deferspace += defer_delta;
1087 ASSERT3S(msp->ms_deferspace, >=, 0);
1088 ASSERT3S(msp->ms_deferspace, <=, sm->sm_size);
1089 if (msp->ms_deferspace != 0) {
1090 /*
1091 * Keep syncing this metaslab until all deferred frees
1092 * are back in circulation.
1093 */
1094 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1095 }
1096
1097 /*
1098 * If the map is loaded but no longer active, evict it as soon as all
1099 * future allocations have synced. (If we unloaded it now and then
1100 * loaded a moment later, the map wouldn't reflect those allocations.)
1101 */
1102 if (sm->sm_loaded && (msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1103 int evictable = 1;
1104
1105 for (int t = 1; t < TXG_CONCURRENT_STATES; t++)
1106 if (msp->ms_allocmap[(txg + t) & TXG_MASK].sm_space)
1107 evictable = 0;
1108
1109 if (evictable && !metaslab_debug)
1110 space_map_unload(sm);
1111 }
1112
1113 metaslab_group_sort(mg, msp, metaslab_weight(msp));
1114
1115 mutex_exit(&msp->ms_lock);
1116}
1117
1118void
1119metaslab_sync_reassess(metaslab_group_t *mg)
1120{
1121 vdev_t *vd = mg->mg_vd;
1122 int64_t failures = mg->mg_alloc_failures;
1123
1124 /*
1125 * Re-evaluate all metaslabs which have lower offsets than the
1126 * bonus area.
1127 */
1128 for (int m = 0; m < vd->vdev_ms_count; m++) {
1129 metaslab_t *msp = vd->vdev_ms[m];
1130
1131 if (msp->ms_map.sm_start > mg->mg_bonus_area)
1132 break;
1133
1134 mutex_enter(&msp->ms_lock);
1135 metaslab_group_sort(mg, msp, metaslab_weight(msp));
1136 mutex_exit(&msp->ms_lock);
1137 }
1138
1139 atomic_add_64(&mg->mg_alloc_failures, -failures);
1140
1141 /*
1142 * Prefetch the next potential metaslabs
1143 */
1144 metaslab_prefetch(mg);
1145}
1146
1147static uint64_t
1148metaslab_distance(metaslab_t *msp, dva_t *dva)
1149{
1150 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
1151 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
1152 uint64_t start = msp->ms_map.sm_start >> ms_shift;
1153
1154 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
1155 return (1ULL << 63);
1156
1157 if (offset < start)
1158 return ((start - offset) << ms_shift);
1159 if (offset > start)
1160 return ((offset - start) << ms_shift);
1161 return (0);
1162}
1163
1164static uint64_t
1165metaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize,
1166 uint64_t txg, uint64_t min_distance, dva_t *dva, int d, int flags)
1167{
1168 spa_t *spa = mg->mg_vd->vdev_spa;
1169 metaslab_t *msp = NULL;
1170 uint64_t offset = -1ULL;
1171 avl_tree_t *t = &mg->mg_metaslab_tree;
1172 uint64_t activation_weight;
1173 uint64_t target_distance;
1174 int i;
1175
1176 activation_weight = METASLAB_WEIGHT_PRIMARY;
1177 for (i = 0; i < d; i++) {
1178 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
1179 activation_weight = METASLAB_WEIGHT_SECONDARY;
1180 break;
1181 }
1182 }
1183
1184 for (;;) {
1185 boolean_t was_active;
1186
1187 mutex_enter(&mg->mg_lock);
1188 for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
1189 if (msp->ms_weight < asize) {
1190 spa_dbgmsg(spa, "%s: failed to meet weight "
1191 "requirement: vdev %llu, txg %llu, mg %p, "
1192 "msp %p, psize %llu, asize %llu, "
1193 "failures %llu, weight %llu",
1194 spa_name(spa), mg->mg_vd->vdev_id, txg,
1195 mg, msp, psize, asize,
1196 mg->mg_alloc_failures, msp->ms_weight);
1197 mutex_exit(&mg->mg_lock);
1198 return (-1ULL);
1199 }
1200 was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
1201 if (activation_weight == METASLAB_WEIGHT_PRIMARY)
1202 break;
1203
1204 target_distance = min_distance +
1205 (msp->ms_smo.smo_alloc ? 0 : min_distance >> 1);
1206
1207 for (i = 0; i < d; i++)
1208 if (metaslab_distance(msp, &dva[i]) <
1209 target_distance)
1210 break;
1211 if (i == d)
1212 break;
1213 }
1214 mutex_exit(&mg->mg_lock);
1215 if (msp == NULL)
1216 return (-1ULL);
1217
1218 /*
1219 * If we've already reached the allowable number of failed
1220 * allocation attempts on this metaslab group then we
1221 * consider skipping it. We skip it only if we're allowed
1222 * to "fast" gang, the physical size is larger than
1223 * a gang block, and we're attempting to allocate from
1224 * the primary metaslab.
1225 */
1226 if (mg->mg_alloc_failures > zfs_mg_alloc_failures &&
1227 CAN_FASTGANG(flags) && psize > SPA_GANGBLOCKSIZE &&
1228 activation_weight == METASLAB_WEIGHT_PRIMARY) {
1229 spa_dbgmsg(spa, "%s: skipping metaslab group: "
1230 "vdev %llu, txg %llu, mg %p, psize %llu, "
1231 "asize %llu, failures %llu", spa_name(spa),
1232 mg->mg_vd->vdev_id, txg, mg, psize, asize,
1233 mg->mg_alloc_failures);
1234 return (-1ULL);
1235 }
1236
1237 mutex_enter(&msp->ms_lock);
1238
1239 /*
1240 * Ensure that the metaslab we have selected is still
1241 * capable of handling our request. It's possible that
1242 * another thread may have changed the weight while we
1243 * were blocked on the metaslab lock.
1244 */
1245 if (msp->ms_weight < asize || (was_active &&
1246 !(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
1247 activation_weight == METASLAB_WEIGHT_PRIMARY)) {
1248 mutex_exit(&msp->ms_lock);
1249 continue;
1250 }
1251
1252 if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
1253 activation_weight == METASLAB_WEIGHT_PRIMARY) {
1254 metaslab_passivate(msp,
1255 msp->ms_weight & ~METASLAB_ACTIVE_MASK);
1256 mutex_exit(&msp->ms_lock);
1257 continue;
1258 }
1259
1260 if (metaslab_activate(msp, activation_weight) != 0) {
1261 mutex_exit(&msp->ms_lock);
1262 continue;
1263 }
1264
1265 if ((offset = space_map_alloc(&msp->ms_map, asize)) != -1ULL)
1266 break;
1267
1268 atomic_inc_64(&mg->mg_alloc_failures);
1269
1270 metaslab_passivate(msp, space_map_maxsize(&msp->ms_map));
1271
1272 mutex_exit(&msp->ms_lock);
1273 }
1274
1275 if (msp->ms_allocmap[txg & TXG_MASK].sm_space == 0)
1276 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
1277
1278 space_map_add(&msp->ms_allocmap[txg & TXG_MASK], offset, asize);
1279
1280 mutex_exit(&msp->ms_lock);
1281
1282 return (offset);
1283}
1284
1285/*
1286 * Allocate a block for the specified i/o.
1287 */
1288static int
1289metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
1290 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
1291{
1292 metaslab_group_t *mg, *rotor;
1293 vdev_t *vd;
1294 int dshift = 3;
1295 int all_zero;
1296 int zio_lock = B_FALSE;
1297 boolean_t allocatable;
1298 uint64_t offset = -1ULL;
1299 uint64_t asize;
1300 uint64_t distance;
1301
1302 ASSERT(!DVA_IS_VALID(&dva[d]));
1303
1304 /*
1305 * For testing, make some blocks above a certain size be gang blocks.
1306 */
1307 if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0)
1308 return (ENOSPC);
1309
1310 /*
1311 * Start at the rotor and loop through all mgs until we find something.
1312 * Note that there's no locking on mc_rotor or mc_aliquot because
1313 * nothing actually breaks if we miss a few updates -- we just won't
1314 * allocate quite as evenly. It all balances out over time.
1315 *
1316 * If we are doing ditto or log blocks, try to spread them across
1317 * consecutive vdevs. If we're forced to reuse a vdev before we've
1318 * allocated all of our ditto blocks, then try and spread them out on
1319 * that vdev as much as possible. If it turns out to not be possible,
1320 * gradually lower our standards until anything becomes acceptable.
1321 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
1322 * gives us hope of containing our fault domains to something we're
1323 * able to reason about. Otherwise, any two top-level vdev failures
1324 * will guarantee the loss of data. With consecutive allocation,
1325 * only two adjacent top-level vdev failures will result in data loss.
1326 *
1327 * If we are doing gang blocks (hintdva is non-NULL), try to keep
1328 * ourselves on the same vdev as our gang block header. That
1329 * way, we can hope for locality in vdev_cache, plus it makes our
1330 * fault domains something tractable.
1331 */
1332 if (hintdva) {
1333 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
1334
1335 /*
1336 * It's possible the vdev we're using as the hint no
1337 * longer exists (i.e. removed). Consult the rotor when
1338 * all else fails.
1339 */
1340 if (vd != NULL) {
1341 mg = vd->vdev_mg;
1342
1343 if (flags & METASLAB_HINTBP_AVOID &&
1344 mg->mg_next != NULL)
1345 mg = mg->mg_next;
1346 } else {
1347 mg = mc->mc_rotor;
1348 }
1349 } else if (d != 0) {
1350 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
1351 mg = vd->vdev_mg->mg_next;
1352 } else {
1353 mg = mc->mc_rotor;
1354 }
1355
1356 /*
1357 * If the hint put us into the wrong metaslab class, or into a
1358 * metaslab group that has been passivated, just follow the rotor.
1359 */
1360 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
1361 mg = mc->mc_rotor;
1362
1363 rotor = mg;
1364top:
1365 all_zero = B_TRUE;
1366 do {
1367 ASSERT(mg->mg_activation_count == 1);
1368
1369 vd = mg->mg_vd;
1370
1371 /*
1372 * Don't allocate from faulted devices.
1373 */
1374 if (zio_lock) {
1375 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
1376 allocatable = vdev_allocatable(vd);
1377 spa_config_exit(spa, SCL_ZIO, FTAG);
1378 } else {
1379 allocatable = vdev_allocatable(vd);
1380 }
1381 if (!allocatable)
1382 goto next;
1383
1384 /*
1385 * Avoid writing single-copy data to a failing vdev
1386 */
1387 if ((vd->vdev_stat.vs_write_errors > 0 ||
1388 vd->vdev_state < VDEV_STATE_HEALTHY) &&
1389 d == 0 && dshift == 3) {
1390 all_zero = B_FALSE;
1391 goto next;
1392 }
1393
1394 ASSERT(mg->mg_class == mc);
1395
1396 distance = vd->vdev_asize >> dshift;
1397 if (distance <= (1ULL << vd->vdev_ms_shift))
1398 distance = 0;
1399 else
1400 all_zero = B_FALSE;
1401
1402 asize = vdev_psize_to_asize(vd, psize);
1403 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
1404
1405 offset = metaslab_group_alloc(mg, psize, asize, txg, distance,
1406 dva, d, flags);
1407 if (offset != -1ULL) {
1408 /*
1409 * If we've just selected this metaslab group,
1410 * figure out whether the corresponding vdev is
1411 * over- or under-used relative to the pool,
1412 * and set an allocation bias to even it out.
1413 */
1414 if (mc->mc_aliquot == 0) {
1415 vdev_stat_t *vs = &vd->vdev_stat;
1416 int64_t vu, cu;
1417
1418 vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
1419 cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
1420
1421 /*
1422 * Calculate how much more or less we should
1423 * try to allocate from this device during
1424 * this iteration around the rotor.
1425 * For example, if a device is 80% full
1426 * and the pool is 20% full then we should
1427 * reduce allocations by 60% on this device.
1428 *
1429 * mg_bias = (20 - 80) * 512K / 100 = -307K
1430 *
1431 * This reduces allocations by 307K for this
1432 * iteration.
1433 */
1434 mg->mg_bias = ((cu - vu) *
1435 (int64_t)mg->mg_aliquot) / 100;
1436 }
1437
1438 if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
1439 mg->mg_aliquot + mg->mg_bias) {
1440 mc->mc_rotor = mg->mg_next;
1441 mc->mc_aliquot = 0;
1442 }
1443
1444 DVA_SET_VDEV(&dva[d], vd->vdev_id);
1445 DVA_SET_OFFSET(&dva[d], offset);
1446 DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
1447 DVA_SET_ASIZE(&dva[d], asize);
1448
1449 return (0);
1450 }
1451next:
1452 mc->mc_rotor = mg->mg_next;
1453 mc->mc_aliquot = 0;
1454 } while ((mg = mg->mg_next) != rotor);
1455
1456 if (!all_zero) {
1457 dshift++;
1458 ASSERT(dshift < 64);
1459 goto top;
1460 }
1461
1462 if (!allocatable && !zio_lock) {
1463 dshift = 3;
1464 zio_lock = B_TRUE;
1465 goto top;
1466 }
1467
1468 bzero(&dva[d], sizeof (dva_t));
1469
1470 return (ENOSPC);
1471}
1472
1473/*
1474 * Free the block represented by DVA in the context of the specified
1475 * transaction group.
1476 */
1477static void
1478metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
1479{
1480 uint64_t vdev = DVA_GET_VDEV(dva);
1481 uint64_t offset = DVA_GET_OFFSET(dva);
1482 uint64_t size = DVA_GET_ASIZE(dva);
1483 vdev_t *vd;
1484 metaslab_t *msp;
1485
1486 ASSERT(DVA_IS_VALID(dva));
1487
1488 if (txg > spa_freeze_txg(spa))
1489 return;
1490
1491 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
1492 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
1493 cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
1494 (u_longlong_t)vdev, (u_longlong_t)offset);
1495 ASSERT(0);
1496 return;
1497 }
1498
1499 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1500
1501 if (DVA_GET_GANG(dva))
1502 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1503
1504 mutex_enter(&msp->ms_lock);
1505
1506 if (now) {
1507 space_map_remove(&msp->ms_allocmap[txg & TXG_MASK],
1508 offset, size);
1509 space_map_free(&msp->ms_map, offset, size);
1510 } else {
1511 if (msp->ms_freemap[txg & TXG_MASK].sm_space == 0)
1512 vdev_dirty(vd, VDD_METASLAB, msp, txg);
1513 space_map_add(&msp->ms_freemap[txg & TXG_MASK], offset, size);
1514 }
1515
1516 mutex_exit(&msp->ms_lock);
1517}
1518
1519/*
1520 * Intent log support: upon opening the pool after a crash, notify the SPA
1521 * of blocks that the intent log has allocated for immediate write, but
1522 * which are still considered free by the SPA because the last transaction
1523 * group didn't commit yet.
1524 */
1525static int
1526metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
1527{
1528 uint64_t vdev = DVA_GET_VDEV(dva);
1529 uint64_t offset = DVA_GET_OFFSET(dva);
1530 uint64_t size = DVA_GET_ASIZE(dva);
1531 vdev_t *vd;
1532 metaslab_t *msp;
1533 int error = 0;
1534
1535 ASSERT(DVA_IS_VALID(dva));
1536
1537 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
1538 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
1539 return (ENXIO);
1540
1541 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1542
1543 if (DVA_GET_GANG(dva))
1544 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1545
1546 mutex_enter(&msp->ms_lock);
1547
1548 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_map.sm_loaded)
1549 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
1550
1551 if (error == 0 && !space_map_contains(&msp->ms_map, offset, size))
1552 error = ENOENT;
1553
1554 if (error || txg == 0) { /* txg == 0 indicates dry run */
1555 mutex_exit(&msp->ms_lock);
1556 return (error);
1557 }
1558
1559 space_map_claim(&msp->ms_map, offset, size);
1560
1561 if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
1562 if (msp->ms_allocmap[txg & TXG_MASK].sm_space == 0)
1563 vdev_dirty(vd, VDD_METASLAB, msp, txg);
1564 space_map_add(&msp->ms_allocmap[txg & TXG_MASK], offset, size);
1565 }
1566
1567 mutex_exit(&msp->ms_lock);
1568
1569 return (0);
1570}
1571
1572int
1573metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
1574 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
1575{
1576 dva_t *dva = bp->blk_dva;
1577 dva_t *hintdva = hintbp->blk_dva;
1578 int error = 0;
1579
1580 ASSERT(bp->blk_birth == 0);
1581 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
1582
1583 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
1584
1585 if (mc->mc_rotor == NULL) { /* no vdevs in this class */
1586 spa_config_exit(spa, SCL_ALLOC, FTAG);
1587 return (ENOSPC);
1588 }
1589
1590 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
1591 ASSERT(BP_GET_NDVAS(bp) == 0);
1592 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
1593
1594 for (int d = 0; d < ndvas; d++) {
1595 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
1596 txg, flags);
1597 if (error) {
1598 for (d--; d >= 0; d--) {
1599 metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
1600 bzero(&dva[d], sizeof (dva_t));
1601 }
1602 spa_config_exit(spa, SCL_ALLOC, FTAG);
1603 return (error);
1604 }
1605 }
1606 ASSERT(error == 0);
1607 ASSERT(BP_GET_NDVAS(bp) == ndvas);
1608
1609 spa_config_exit(spa, SCL_ALLOC, FTAG);
1610
1611 BP_SET_BIRTH(bp, txg, txg);
1612
1613 return (0);
1614}
1615
1616void
1617metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
1618{
1619 const dva_t *dva = bp->blk_dva;
1620 int ndvas = BP_GET_NDVAS(bp);
1621
1622 ASSERT(!BP_IS_HOLE(bp));
1623 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
1624
1625 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
1626
1627 for (int d = 0; d < ndvas; d++)
1628 metaslab_free_dva(spa, &dva[d], txg, now);
1629
1630 spa_config_exit(spa, SCL_FREE, FTAG);
1631}
1632
1633int
1634metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
1635{
1636 const dva_t *dva = bp->blk_dva;
1637 int ndvas = BP_GET_NDVAS(bp);
1638 int error = 0;
1639
1640 ASSERT(!BP_IS_HOLE(bp));
1641
1642 if (txg != 0) {
1643 /*
1644 * First do a dry run to make sure all DVAs are claimable,
1645 * so we don't have to unwind from partial failures below.
1646 */
1647 if ((error = metaslab_claim(spa, bp, 0)) != 0)
1648 return (error);
1649 }
1650
1651 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
1652
1653 for (int d = 0; d < ndvas; d++)
1654 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
1655 break;
1656
1657 spa_config_exit(spa, SCL_ALLOC, FTAG);
1658
1659 ASSERT(error == 0 || txg == 0);
1660
1661 return (error);
1662}
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
46uint64_t metaslab_aliquot = 512ULL << 10;
47uint64_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 */
54int zfs_mg_alloc_failures = 0;
55
56SYSCTL_DECL(_vfs_zfs);
57SYSCTL_INT(_vfs_zfs, OID_AUTO, mg_alloc_failures, CTLFLAG_RDTUN,
58 &zfs_mg_alloc_failures, 0,
59 "Number of allowed allocation failures per vdev");
60TUNABLE_INT("vfs.zfs.mg_alloc_failures", &zfs_mg_alloc_failures);
61
62/*
63 * Metaslab debugging: when set, keeps all space maps in core to verify frees.
64 */
65static int metaslab_debug = 0;
66
67/*
68 * Minimum size which forces the dynamic allocator to change
69 * it's allocation strategy. Once the space map cannot satisfy
70 * an allocation of this size then it switches to using more
71 * aggressive strategy (i.e search by size rather than offset).
72 */
73uint64_t metaslab_df_alloc_threshold = SPA_MAXBLOCKSIZE;
74
75/*
76 * The minimum free space, in percent, which must be available
77 * in a space map to continue allocations in a first-fit fashion.
78 * Once the space_map's free space drops below this level we dynamically
79 * switch to using best-fit allocations.
80 */
81int metaslab_df_free_pct = 4;
82
83/*
84 * A metaslab is considered "free" if it contains a contiguous
85 * segment which is greater than metaslab_min_alloc_size.
86 */
87uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
88
89/*
90 * Max number of space_maps to prefetch.
91 */
92int metaslab_prefetch_limit = SPA_DVAS_PER_BP;
93
94/*
95 * Percentage bonus multiplier for metaslabs that are in the bonus area.
96 */
97int metaslab_smo_bonus_pct = 150;
98
99/*
100 * ==========================================================================
101 * Metaslab classes
102 * ==========================================================================
103 */
104metaslab_class_t *
105metaslab_class_create(spa_t *spa, space_map_ops_t *ops)
106{
107 metaslab_class_t *mc;
108
109 mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
110
111 mc->mc_spa = spa;
112 mc->mc_rotor = NULL;
113 mc->mc_ops = ops;
114
115 return (mc);
116}
117
118void
119metaslab_class_destroy(metaslab_class_t *mc)
120{
121 ASSERT(mc->mc_rotor == NULL);
122 ASSERT(mc->mc_alloc == 0);
123 ASSERT(mc->mc_deferred == 0);
124 ASSERT(mc->mc_space == 0);
125 ASSERT(mc->mc_dspace == 0);
126
127 kmem_free(mc, sizeof (metaslab_class_t));
128}
129
130int
131metaslab_class_validate(metaslab_class_t *mc)
132{
133 metaslab_group_t *mg;
134 vdev_t *vd;
135
136 /*
137 * Must hold one of the spa_config locks.
138 */
139 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
140 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
141
142 if ((mg = mc->mc_rotor) == NULL)
143 return (0);
144
145 do {
146 vd = mg->mg_vd;
147 ASSERT(vd->vdev_mg != NULL);
148 ASSERT3P(vd->vdev_top, ==, vd);
149 ASSERT3P(mg->mg_class, ==, mc);
150 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
151 } while ((mg = mg->mg_next) != mc->mc_rotor);
152
153 return (0);
154}
155
156void
157metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
158 int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
159{
160 atomic_add_64(&mc->mc_alloc, alloc_delta);
161 atomic_add_64(&mc->mc_deferred, defer_delta);
162 atomic_add_64(&mc->mc_space, space_delta);
163 atomic_add_64(&mc->mc_dspace, dspace_delta);
164}
165
166uint64_t
167metaslab_class_get_alloc(metaslab_class_t *mc)
168{
169 return (mc->mc_alloc);
170}
171
172uint64_t
173metaslab_class_get_deferred(metaslab_class_t *mc)
174{
175 return (mc->mc_deferred);
176}
177
178uint64_t
179metaslab_class_get_space(metaslab_class_t *mc)
180{
181 return (mc->mc_space);
182}
183
184uint64_t
185metaslab_class_get_dspace(metaslab_class_t *mc)
186{
187 return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
188}
189
190/*
191 * ==========================================================================
192 * Metaslab groups
193 * ==========================================================================
194 */
195static int
196metaslab_compare(const void *x1, const void *x2)
197{
198 const metaslab_t *m1 = x1;
199 const metaslab_t *m2 = x2;
200
201 if (m1->ms_weight < m2->ms_weight)
202 return (1);
203 if (m1->ms_weight > m2->ms_weight)
204 return (-1);
205
206 /*
207 * If the weights are identical, use the offset to force uniqueness.
208 */
209 if (m1->ms_map.sm_start < m2->ms_map.sm_start)
210 return (-1);
211 if (m1->ms_map.sm_start > m2->ms_map.sm_start)
212 return (1);
213
214 ASSERT3P(m1, ==, m2);
215
216 return (0);
217}
218
219metaslab_group_t *
220metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
221{
222 metaslab_group_t *mg;
223
224 mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
225 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
226 avl_create(&mg->mg_metaslab_tree, metaslab_compare,
227 sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
228 mg->mg_vd = vd;
229 mg->mg_class = mc;
230 mg->mg_activation_count = 0;
231
232 return (mg);
233}
234
235void
236metaslab_group_destroy(metaslab_group_t *mg)
237{
238 ASSERT(mg->mg_prev == NULL);
239 ASSERT(mg->mg_next == NULL);
240 /*
241 * We may have gone below zero with the activation count
242 * either because we never activated in the first place or
243 * because we're done, and possibly removing the vdev.
244 */
245 ASSERT(mg->mg_activation_count <= 0);
246
247 avl_destroy(&mg->mg_metaslab_tree);
248 mutex_destroy(&mg->mg_lock);
249 kmem_free(mg, sizeof (metaslab_group_t));
250}
251
252void
253metaslab_group_activate(metaslab_group_t *mg)
254{
255 metaslab_class_t *mc = mg->mg_class;
256 metaslab_group_t *mgprev, *mgnext;
257
258 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
259
260 ASSERT(mc->mc_rotor != mg);
261 ASSERT(mg->mg_prev == NULL);
262 ASSERT(mg->mg_next == NULL);
263 ASSERT(mg->mg_activation_count <= 0);
264
265 if (++mg->mg_activation_count <= 0)
266 return;
267
268 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
269
270 if ((mgprev = mc->mc_rotor) == NULL) {
271 mg->mg_prev = mg;
272 mg->mg_next = mg;
273 } else {
274 mgnext = mgprev->mg_next;
275 mg->mg_prev = mgprev;
276 mg->mg_next = mgnext;
277 mgprev->mg_next = mg;
278 mgnext->mg_prev = mg;
279 }
280 mc->mc_rotor = mg;
281}
282
283void
284metaslab_group_passivate(metaslab_group_t *mg)
285{
286 metaslab_class_t *mc = mg->mg_class;
287 metaslab_group_t *mgprev, *mgnext;
288
289 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
290
291 if (--mg->mg_activation_count != 0) {
292 ASSERT(mc->mc_rotor != mg);
293 ASSERT(mg->mg_prev == NULL);
294 ASSERT(mg->mg_next == NULL);
295 ASSERT(mg->mg_activation_count < 0);
296 return;
297 }
298
299 mgprev = mg->mg_prev;
300 mgnext = mg->mg_next;
301
302 if (mg == mgnext) {
303 mc->mc_rotor = NULL;
304 } else {
305 mc->mc_rotor = mgnext;
306 mgprev->mg_next = mgnext;
307 mgnext->mg_prev = mgprev;
308 }
309
310 mg->mg_prev = NULL;
311 mg->mg_next = NULL;
312}
313
314static void
315metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
316{
317 mutex_enter(&mg->mg_lock);
318 ASSERT(msp->ms_group == NULL);
319 msp->ms_group = mg;
320 msp->ms_weight = 0;
321 avl_add(&mg->mg_metaslab_tree, msp);
322 mutex_exit(&mg->mg_lock);
323}
324
325static void
326metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
327{
328 mutex_enter(&mg->mg_lock);
329 ASSERT(msp->ms_group == mg);
330 avl_remove(&mg->mg_metaslab_tree, msp);
331 msp->ms_group = NULL;
332 mutex_exit(&mg->mg_lock);
333}
334
335static void
336metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
337{
338 /*
339 * Although in principle the weight can be any value, in
340 * practice we do not use values in the range [1, 510].
341 */
342 ASSERT(weight >= SPA_MINBLOCKSIZE-1 || weight == 0);
343 ASSERT(MUTEX_HELD(&msp->ms_lock));
344
345 mutex_enter(&mg->mg_lock);
346 ASSERT(msp->ms_group == mg);
347 avl_remove(&mg->mg_metaslab_tree, msp);
348 msp->ms_weight = weight;
349 avl_add(&mg->mg_metaslab_tree, msp);
350 mutex_exit(&mg->mg_lock);
351}
352
353/*
354 * ==========================================================================
355 * Common allocator routines
356 * ==========================================================================
357 */
358static int
359metaslab_segsize_compare(const void *x1, const void *x2)
360{
361 const space_seg_t *s1 = x1;
362 const space_seg_t *s2 = x2;
363 uint64_t ss_size1 = s1->ss_end - s1->ss_start;
364 uint64_t ss_size2 = s2->ss_end - s2->ss_start;
365
366 if (ss_size1 < ss_size2)
367 return (-1);
368 if (ss_size1 > ss_size2)
369 return (1);
370
371 if (s1->ss_start < s2->ss_start)
372 return (-1);
373 if (s1->ss_start > s2->ss_start)
374 return (1);
375
376 return (0);
377}
378
379/*
380 * This is a helper function that can be used by the allocator to find
381 * a suitable block to allocate. This will search the specified AVL
382 * tree looking for a block that matches the specified criteria.
383 */
384static uint64_t
385metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
386 uint64_t align)
387{
388 space_seg_t *ss, ssearch;
389 avl_index_t where;
390
391 ssearch.ss_start = *cursor;
392 ssearch.ss_end = *cursor + size;
393
394 ss = avl_find(t, &ssearch, &where);
395 if (ss == NULL)
396 ss = avl_nearest(t, where, AVL_AFTER);
397
398 while (ss != NULL) {
399 uint64_t offset = P2ROUNDUP(ss->ss_start, align);
400
401 if (offset + size <= ss->ss_end) {
402 *cursor = offset + size;
403 return (offset);
404 }
405 ss = AVL_NEXT(t, ss);
406 }
407
408 /*
409 * If we know we've searched the whole map (*cursor == 0), give up.
410 * Otherwise, reset the cursor to the beginning and try again.
411 */
412 if (*cursor == 0)
413 return (-1ULL);
414
415 *cursor = 0;
416 return (metaslab_block_picker(t, cursor, size, align));
417}
418
419static void
420metaslab_pp_load(space_map_t *sm)
421{
422 space_seg_t *ss;
423
424 ASSERT(sm->sm_ppd == NULL);
425 sm->sm_ppd = kmem_zalloc(64 * sizeof (uint64_t), KM_SLEEP);
426
427 sm->sm_pp_root = kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
428 avl_create(sm->sm_pp_root, metaslab_segsize_compare,
429 sizeof (space_seg_t), offsetof(struct space_seg, ss_pp_node));
430
431 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
432 avl_add(sm->sm_pp_root, ss);
433}
434
435static void
436metaslab_pp_unload(space_map_t *sm)
437{
438 void *cookie = NULL;
439
440 kmem_free(sm->sm_ppd, 64 * sizeof (uint64_t));
441 sm->sm_ppd = NULL;
442
443 while (avl_destroy_nodes(sm->sm_pp_root, &cookie) != NULL) {
444 /* tear down the tree */
445 }
446
447 avl_destroy(sm->sm_pp_root);
448 kmem_free(sm->sm_pp_root, sizeof (avl_tree_t));
449 sm->sm_pp_root = NULL;
450}
451
452/* ARGSUSED */
453static void
454metaslab_pp_claim(space_map_t *sm, uint64_t start, uint64_t size)
455{
456 /* No need to update cursor */
457}
458
459/* ARGSUSED */
460static void
461metaslab_pp_free(space_map_t *sm, uint64_t start, uint64_t size)
462{
463 /* No need to update cursor */
464}
465
466/*
467 * Return the maximum contiguous segment within the metaslab.
468 */
469uint64_t
470metaslab_pp_maxsize(space_map_t *sm)
471{
472 avl_tree_t *t = sm->sm_pp_root;
473 space_seg_t *ss;
474
475 if (t == NULL || (ss = avl_last(t)) == NULL)
476 return (0ULL);
477
478 return (ss->ss_end - ss->ss_start);
479}
480
481/*
482 * ==========================================================================
483 * The first-fit block allocator
484 * ==========================================================================
485 */
486static uint64_t
487metaslab_ff_alloc(space_map_t *sm, uint64_t size)
488{
489 avl_tree_t *t = &sm->sm_root;
490 uint64_t align = size & -size;
491 uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
492
493 return (metaslab_block_picker(t, cursor, size, align));
494}
495
496/* ARGSUSED */
497boolean_t
498metaslab_ff_fragmented(space_map_t *sm)
499{
500 return (B_TRUE);
501}
502
503static space_map_ops_t metaslab_ff_ops = {
504 metaslab_pp_load,
505 metaslab_pp_unload,
506 metaslab_ff_alloc,
507 metaslab_pp_claim,
508 metaslab_pp_free,
509 metaslab_pp_maxsize,
510 metaslab_ff_fragmented
511};
512
513/*
514 * ==========================================================================
515 * Dynamic block allocator -
516 * Uses the first fit allocation scheme until space get low and then
517 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
518 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
519 * ==========================================================================
520 */
521static uint64_t
522metaslab_df_alloc(space_map_t *sm, uint64_t size)
523{
524 avl_tree_t *t = &sm->sm_root;
525 uint64_t align = size & -size;
526 uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
527 uint64_t max_size = metaslab_pp_maxsize(sm);
528 int free_pct = sm->sm_space * 100 / sm->sm_size;
529
530 ASSERT(MUTEX_HELD(sm->sm_lock));
531 ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
532
533 if (max_size < size)
534 return (-1ULL);
535
536 /*
537 * If we're running low on space switch to using the size
538 * sorted AVL tree (best-fit).
539 */
540 if (max_size < metaslab_df_alloc_threshold ||
541 free_pct < metaslab_df_free_pct) {
542 t = sm->sm_pp_root;
543 *cursor = 0;
544 }
545
546 return (metaslab_block_picker(t, cursor, size, 1ULL));
547}
548
549static boolean_t
550metaslab_df_fragmented(space_map_t *sm)
551{
552 uint64_t max_size = metaslab_pp_maxsize(sm);
553 int free_pct = sm->sm_space * 100 / sm->sm_size;
554
555 if (max_size >= metaslab_df_alloc_threshold &&
556 free_pct >= metaslab_df_free_pct)
557 return (B_FALSE);
558
559 return (B_TRUE);
560}
561
562static space_map_ops_t metaslab_df_ops = {
563 metaslab_pp_load,
564 metaslab_pp_unload,
565 metaslab_df_alloc,
566 metaslab_pp_claim,
567 metaslab_pp_free,
568 metaslab_pp_maxsize,
569 metaslab_df_fragmented
570};
571
572/*
573 * ==========================================================================
574 * Other experimental allocators
575 * ==========================================================================
576 */
577static uint64_t
578metaslab_cdf_alloc(space_map_t *sm, uint64_t size)
579{
580 avl_tree_t *t = &sm->sm_root;
581 uint64_t *cursor = (uint64_t *)sm->sm_ppd;
582 uint64_t *extent_end = (uint64_t *)sm->sm_ppd + 1;
583 uint64_t max_size = metaslab_pp_maxsize(sm);
584 uint64_t rsize = size;
585 uint64_t offset = 0;
586
587 ASSERT(MUTEX_HELD(sm->sm_lock));
588 ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
589
590 if (max_size < size)
591 return (-1ULL);
592
593 ASSERT3U(*extent_end, >=, *cursor);
594
595 /*
596 * If we're running low on space switch to using the size
597 * sorted AVL tree (best-fit).
598 */
599 if ((*cursor + size) > *extent_end) {
600
601 t = sm->sm_pp_root;
602 *cursor = *extent_end = 0;
603
604 if (max_size > 2 * SPA_MAXBLOCKSIZE)
605 rsize = MIN(metaslab_min_alloc_size, max_size);
606 offset = metaslab_block_picker(t, extent_end, rsize, 1ULL);
607 if (offset != -1)
608 *cursor = offset + size;
609 } else {
610 offset = metaslab_block_picker(t, cursor, rsize, 1ULL);
611 }
612 ASSERT3U(*cursor, <=, *extent_end);
613 return (offset);
614}
615
616static boolean_t
617metaslab_cdf_fragmented(space_map_t *sm)
618{
619 uint64_t max_size = metaslab_pp_maxsize(sm);
620
621 if (max_size > (metaslab_min_alloc_size * 10))
622 return (B_FALSE);
623 return (B_TRUE);
624}
625
626static space_map_ops_t metaslab_cdf_ops = {
627 metaslab_pp_load,
628 metaslab_pp_unload,
629 metaslab_cdf_alloc,
630 metaslab_pp_claim,
631 metaslab_pp_free,
632 metaslab_pp_maxsize,
633 metaslab_cdf_fragmented
634};
635
636uint64_t metaslab_ndf_clump_shift = 4;
637
638static uint64_t
639metaslab_ndf_alloc(space_map_t *sm, uint64_t size)
640{
641 avl_tree_t *t = &sm->sm_root;
642 avl_index_t where;
643 space_seg_t *ss, ssearch;
644 uint64_t hbit = highbit(size);
645 uint64_t *cursor = (uint64_t *)sm->sm_ppd + hbit - 1;
646 uint64_t max_size = metaslab_pp_maxsize(sm);
647
648 ASSERT(MUTEX_HELD(sm->sm_lock));
649 ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
650
651 if (max_size < size)
652 return (-1ULL);
653
654 ssearch.ss_start = *cursor;
655 ssearch.ss_end = *cursor + size;
656
657 ss = avl_find(t, &ssearch, &where);
658 if (ss == NULL || (ss->ss_start + size > ss->ss_end)) {
659 t = sm->sm_pp_root;
660
661 ssearch.ss_start = 0;
662 ssearch.ss_end = MIN(max_size,
663 1ULL << (hbit + metaslab_ndf_clump_shift));
664 ss = avl_find(t, &ssearch, &where);
665 if (ss == NULL)
666 ss = avl_nearest(t, where, AVL_AFTER);
667 ASSERT(ss != NULL);
668 }
669
670 if (ss != NULL) {
671 if (ss->ss_start + size <= ss->ss_end) {
672 *cursor = ss->ss_start + size;
673 return (ss->ss_start);
674 }
675 }
676 return (-1ULL);
677}
678
679static boolean_t
680metaslab_ndf_fragmented(space_map_t *sm)
681{
682 uint64_t max_size = metaslab_pp_maxsize(sm);
683
684 if (max_size > (metaslab_min_alloc_size << metaslab_ndf_clump_shift))
685 return (B_FALSE);
686 return (B_TRUE);
687}
688
689
690static space_map_ops_t metaslab_ndf_ops = {
691 metaslab_pp_load,
692 metaslab_pp_unload,
693 metaslab_ndf_alloc,
694 metaslab_pp_claim,
695 metaslab_pp_free,
696 metaslab_pp_maxsize,
697 metaslab_ndf_fragmented
698};
699
700space_map_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
701
702/*
703 * ==========================================================================
704 * Metaslabs
705 * ==========================================================================
706 */
707metaslab_t *
708metaslab_init(metaslab_group_t *mg, space_map_obj_t *smo,
709 uint64_t start, uint64_t size, uint64_t txg)
710{
711 vdev_t *vd = mg->mg_vd;
712 metaslab_t *msp;
713
714 msp = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
715 mutex_init(&msp->ms_lock, NULL, MUTEX_DEFAULT, NULL);
716
717 msp->ms_smo_syncing = *smo;
718
719 /*
720 * We create the main space map here, but we don't create the
721 * allocmaps and freemaps until metaslab_sync_done(). This serves
722 * two purposes: it allows metaslab_sync_done() to detect the
723 * addition of new space; and for debugging, it ensures that we'd
724 * data fault on any attempt to use this metaslab before it's ready.
725 */
726 space_map_create(&msp->ms_map, start, size,
727 vd->vdev_ashift, &msp->ms_lock);
728
729 metaslab_group_add(mg, msp);
730
731 if (metaslab_debug && smo->smo_object != 0) {
732 mutex_enter(&msp->ms_lock);
733 VERIFY(space_map_load(&msp->ms_map, mg->mg_class->mc_ops,
734 SM_FREE, smo, spa_meta_objset(vd->vdev_spa)) == 0);
735 mutex_exit(&msp->ms_lock);
736 }
737
738 /*
739 * If we're opening an existing pool (txg == 0) or creating
740 * a new one (txg == TXG_INITIAL), all space is available now.
741 * If we're adding space to an existing pool, the new space
742 * does not become available until after this txg has synced.
743 */
744 if (txg <= TXG_INITIAL)
745 metaslab_sync_done(msp, 0);
746
747 if (txg != 0) {
748 vdev_dirty(vd, 0, NULL, txg);
749 vdev_dirty(vd, VDD_METASLAB, msp, txg);
750 }
751
752 return (msp);
753}
754
755void
756metaslab_fini(metaslab_t *msp)
757{
758 metaslab_group_t *mg = msp->ms_group;
759
760 vdev_space_update(mg->mg_vd,
761 -msp->ms_smo.smo_alloc, 0, -msp->ms_map.sm_size);
762
763 metaslab_group_remove(mg, msp);
764
765 mutex_enter(&msp->ms_lock);
766
767 space_map_unload(&msp->ms_map);
768 space_map_destroy(&msp->ms_map);
769
770 for (int t = 0; t < TXG_SIZE; t++) {
771 space_map_destroy(&msp->ms_allocmap[t]);
772 space_map_destroy(&msp->ms_freemap[t]);
773 }
774
775 for (int t = 0; t < TXG_DEFER_SIZE; t++)
776 space_map_destroy(&msp->ms_defermap[t]);
777
778 ASSERT0(msp->ms_deferspace);
779
780 mutex_exit(&msp->ms_lock);
781 mutex_destroy(&msp->ms_lock);
782
783 kmem_free(msp, sizeof (metaslab_t));
784}
785
786#define METASLAB_WEIGHT_PRIMARY (1ULL << 63)
787#define METASLAB_WEIGHT_SECONDARY (1ULL << 62)
788#define METASLAB_ACTIVE_MASK \
789 (METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
790
791static uint64_t
792metaslab_weight(metaslab_t *msp)
793{
794 metaslab_group_t *mg = msp->ms_group;
795 space_map_t *sm = &msp->ms_map;
796 space_map_obj_t *smo = &msp->ms_smo;
797 vdev_t *vd = mg->mg_vd;
798 uint64_t weight, space;
799
800 ASSERT(MUTEX_HELD(&msp->ms_lock));
801
802 /*
803 * The baseline weight is the metaslab's free space.
804 */
805 space = sm->sm_size - smo->smo_alloc;
806 weight = space;
807
808 /*
809 * Modern disks have uniform bit density and constant angular velocity.
810 * Therefore, the outer recording zones are faster (higher bandwidth)
811 * than the inner zones by the ratio of outer to inner track diameter,
812 * which is typically around 2:1. We account for this by assigning
813 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
814 * In effect, this means that we'll select the metaslab with the most
815 * free bandwidth rather than simply the one with the most free space.
816 */
817 weight = 2 * weight -
818 ((sm->sm_start >> vd->vdev_ms_shift) * weight) / vd->vdev_ms_count;
819 ASSERT(weight >= space && weight <= 2 * space);
820
821 /*
822 * For locality, assign higher weight to metaslabs which have
823 * a lower offset than what we've already activated.
824 */
825 if (sm->sm_start <= mg->mg_bonus_area)
826 weight *= (metaslab_smo_bonus_pct / 100);
827 ASSERT(weight >= space &&
828 weight <= 2 * (metaslab_smo_bonus_pct / 100) * space);
829
830 if (sm->sm_loaded && !sm->sm_ops->smop_fragmented(sm)) {
831 /*
832 * If this metaslab is one we're actively using, adjust its
833 * weight to make it preferable to any inactive metaslab so
834 * we'll polish it off.
835 */
836 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
837 }
838 return (weight);
839}
840
841static void
842metaslab_prefetch(metaslab_group_t *mg)
843{
844 spa_t *spa = mg->mg_vd->vdev_spa;
845 metaslab_t *msp;
846 avl_tree_t *t = &mg->mg_metaslab_tree;
847 int m;
848
849 mutex_enter(&mg->mg_lock);
850
851 /*
852 * Prefetch the next potential metaslabs
853 */
854 for (msp = avl_first(t), m = 0; msp; msp = AVL_NEXT(t, msp), m++) {
855 space_map_t *sm = &msp->ms_map;
856 space_map_obj_t *smo = &msp->ms_smo;
857
858 /* If we have reached our prefetch limit then we're done */
859 if (m >= metaslab_prefetch_limit)
860 break;
861
862 if (!sm->sm_loaded && smo->smo_object != 0) {
863 mutex_exit(&mg->mg_lock);
864 dmu_prefetch(spa_meta_objset(spa), smo->smo_object,
865 0ULL, smo->smo_objsize);
866 mutex_enter(&mg->mg_lock);
867 }
868 }
869 mutex_exit(&mg->mg_lock);
870}
871
872static int
873metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
874{
875 metaslab_group_t *mg = msp->ms_group;
876 space_map_t *sm = &msp->ms_map;
877 space_map_ops_t *sm_ops = msp->ms_group->mg_class->mc_ops;
878
879 ASSERT(MUTEX_HELD(&msp->ms_lock));
880
881 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
882 space_map_load_wait(sm);
883 if (!sm->sm_loaded) {
884 space_map_obj_t *smo = &msp->ms_smo;
885
886 int error = space_map_load(sm, sm_ops, SM_FREE, smo,
887 spa_meta_objset(msp->ms_group->mg_vd->vdev_spa));
888 if (error) {
889 metaslab_group_sort(msp->ms_group, msp, 0);
890 return (error);
891 }
892 for (int t = 0; t < TXG_DEFER_SIZE; t++)
893 space_map_walk(&msp->ms_defermap[t],
894 space_map_claim, sm);
895
896 }
897
898 /*
899 * Track the bonus area as we activate new metaslabs.
900 */
901 if (sm->sm_start > mg->mg_bonus_area) {
902 mutex_enter(&mg->mg_lock);
903 mg->mg_bonus_area = sm->sm_start;
904 mutex_exit(&mg->mg_lock);
905 }
906
907 metaslab_group_sort(msp->ms_group, msp,
908 msp->ms_weight | activation_weight);
909 }
910 ASSERT(sm->sm_loaded);
911 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
912
913 return (0);
914}
915
916static void
917metaslab_passivate(metaslab_t *msp, uint64_t size)
918{
919 /*
920 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
921 * this metaslab again. In that case, it had better be empty,
922 * or we would be leaving space on the table.
923 */
924 ASSERT(size >= SPA_MINBLOCKSIZE || msp->ms_map.sm_space == 0);
925 metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
926 ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
927}
928
929/*
930 * Write a metaslab to disk in the context of the specified transaction group.
931 */
932void
933metaslab_sync(metaslab_t *msp, uint64_t txg)
934{
935 vdev_t *vd = msp->ms_group->mg_vd;
936 spa_t *spa = vd->vdev_spa;
937 objset_t *mos = spa_meta_objset(spa);
938 space_map_t *allocmap = &msp->ms_allocmap[txg & TXG_MASK];
939 space_map_t *freemap = &msp->ms_freemap[txg & TXG_MASK];
940 space_map_t *freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
941 space_map_t *sm = &msp->ms_map;
942 space_map_obj_t *smo = &msp->ms_smo_syncing;
943 dmu_buf_t *db;
944 dmu_tx_t *tx;
945
946 ASSERT(!vd->vdev_ishole);
947
948 if (allocmap->sm_space == 0 && freemap->sm_space == 0)
949 return;
950
951 /*
952 * The only state that can actually be changing concurrently with
953 * metaslab_sync() is the metaslab's ms_map. No other thread can
954 * be modifying this txg's allocmap, freemap, freed_map, or smo.
955 * Therefore, we only hold ms_lock to satify space_map ASSERTs.
956 * We drop it whenever we call into the DMU, because the DMU
957 * can call down to us (e.g. via zio_free()) at any time.
958 */
959
960 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
961
962 if (smo->smo_object == 0) {
963 ASSERT(smo->smo_objsize == 0);
964 ASSERT(smo->smo_alloc == 0);
965 smo->smo_object = dmu_object_alloc(mos,
966 DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
967 DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
968 ASSERT(smo->smo_object != 0);
969 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
970 (sm->sm_start >> vd->vdev_ms_shift),
971 sizeof (uint64_t), &smo->smo_object, tx);
972 }
973
974 mutex_enter(&msp->ms_lock);
975
976 space_map_walk(freemap, space_map_add, freed_map);
977
978 if (sm->sm_loaded && spa_sync_pass(spa) == 1 && smo->smo_objsize >=
979 2 * sizeof (uint64_t) * avl_numnodes(&sm->sm_root)) {
980 /*
981 * The in-core space map representation is twice as compact
982 * as the on-disk one, so it's time to condense the latter
983 * by generating a pure allocmap from first principles.
984 *
985 * This metaslab is 100% allocated,
986 * minus the content of the in-core map (sm),
987 * minus what's been freed this txg (freed_map),
988 * minus deferred frees (ms_defermap[]),
989 * minus allocations from txgs in the future
990 * (because they haven't been committed yet).
991 */
992 space_map_vacate(allocmap, NULL, NULL);
993 space_map_vacate(freemap, NULL, NULL);
994
995 space_map_add(allocmap, allocmap->sm_start, allocmap->sm_size);
996
997 space_map_walk(sm, space_map_remove, allocmap);
998 space_map_walk(freed_map, space_map_remove, allocmap);
999
1000 for (int t = 0; t < TXG_DEFER_SIZE; t++)
1001 space_map_walk(&msp->ms_defermap[t],
1002 space_map_remove, allocmap);
1003
1004 for (int t = 1; t < TXG_CONCURRENT_STATES; t++)
1005 space_map_walk(&msp->ms_allocmap[(txg + t) & TXG_MASK],
1006 space_map_remove, allocmap);
1007
1008 mutex_exit(&msp->ms_lock);
1009 space_map_truncate(smo, mos, tx);
1010 mutex_enter(&msp->ms_lock);
1011 }
1012
1013 space_map_sync(allocmap, SM_ALLOC, smo, mos, tx);
1014 space_map_sync(freemap, SM_FREE, smo, mos, tx);
1015
1016 mutex_exit(&msp->ms_lock);
1017
1018 VERIFY(0 == dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
1019 dmu_buf_will_dirty(db, tx);
1020 ASSERT3U(db->db_size, >=, sizeof (*smo));
1021 bcopy(smo, db->db_data, sizeof (*smo));
1022 dmu_buf_rele(db, FTAG);
1023
1024 dmu_tx_commit(tx);
1025}
1026
1027/*
1028 * Called after a transaction group has completely synced to mark
1029 * all of the metaslab's free space as usable.
1030 */
1031void
1032metaslab_sync_done(metaslab_t *msp, uint64_t txg)
1033{
1034 space_map_obj_t *smo = &msp->ms_smo;
1035 space_map_obj_t *smosync = &msp->ms_smo_syncing;
1036 space_map_t *sm = &msp->ms_map;
1037 space_map_t *freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
1038 space_map_t *defer_map = &msp->ms_defermap[txg % TXG_DEFER_SIZE];
1039 metaslab_group_t *mg = msp->ms_group;
1040 vdev_t *vd = mg->mg_vd;
1041 int64_t alloc_delta, defer_delta;
1042
1043 ASSERT(!vd->vdev_ishole);
1044
1045 mutex_enter(&msp->ms_lock);
1046
1047 /*
1048 * If this metaslab is just becoming available, initialize its
1049 * allocmaps and freemaps and add its capacity to the vdev.
1050 */
1051 if (freed_map->sm_size == 0) {
1052 for (int t = 0; t < TXG_SIZE; t++) {
1053 space_map_create(&msp->ms_allocmap[t], sm->sm_start,
1054 sm->sm_size, sm->sm_shift, sm->sm_lock);
1055 space_map_create(&msp->ms_freemap[t], sm->sm_start,
1056 sm->sm_size, sm->sm_shift, sm->sm_lock);
1057 }
1058
1059 for (int t = 0; t < TXG_DEFER_SIZE; t++)
1060 space_map_create(&msp->ms_defermap[t], sm->sm_start,
1061 sm->sm_size, sm->sm_shift, sm->sm_lock);
1062
1063 vdev_space_update(vd, 0, 0, sm->sm_size);
1064 }
1065
1066 alloc_delta = smosync->smo_alloc - smo->smo_alloc;
1067 defer_delta = freed_map->sm_space - defer_map->sm_space;
1068
1069 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
1070
1071 ASSERT(msp->ms_allocmap[txg & TXG_MASK].sm_space == 0);
1072 ASSERT(msp->ms_freemap[txg & TXG_MASK].sm_space == 0);
1073
1074 /*
1075 * If there's a space_map_load() in progress, wait for it to complete
1076 * so that we have a consistent view of the in-core space map.
1077 * Then, add defer_map (oldest deferred frees) to this map and
1078 * transfer freed_map (this txg's frees) to defer_map.
1079 */
1080 space_map_load_wait(sm);
1081 space_map_vacate(defer_map, sm->sm_loaded ? space_map_free : NULL, sm);
1082 space_map_vacate(freed_map, space_map_add, defer_map);
1083
1084 *smo = *smosync;
1085
1086 msp->ms_deferspace += defer_delta;
1087 ASSERT3S(msp->ms_deferspace, >=, 0);
1088 ASSERT3S(msp->ms_deferspace, <=, sm->sm_size);
1089 if (msp->ms_deferspace != 0) {
1090 /*
1091 * Keep syncing this metaslab until all deferred frees
1092 * are back in circulation.
1093 */
1094 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1095 }
1096
1097 /*
1098 * If the map is loaded but no longer active, evict it as soon as all
1099 * future allocations have synced. (If we unloaded it now and then
1100 * loaded a moment later, the map wouldn't reflect those allocations.)
1101 */
1102 if (sm->sm_loaded && (msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1103 int evictable = 1;
1104
1105 for (int t = 1; t < TXG_CONCURRENT_STATES; t++)
1106 if (msp->ms_allocmap[(txg + t) & TXG_MASK].sm_space)
1107 evictable = 0;
1108
1109 if (evictable && !metaslab_debug)
1110 space_map_unload(sm);
1111 }
1112
1113 metaslab_group_sort(mg, msp, metaslab_weight(msp));
1114
1115 mutex_exit(&msp->ms_lock);
1116}
1117
1118void
1119metaslab_sync_reassess(metaslab_group_t *mg)
1120{
1121 vdev_t *vd = mg->mg_vd;
1122 int64_t failures = mg->mg_alloc_failures;
1123
1124 /*
1125 * Re-evaluate all metaslabs which have lower offsets than the
1126 * bonus area.
1127 */
1128 for (int m = 0; m < vd->vdev_ms_count; m++) {
1129 metaslab_t *msp = vd->vdev_ms[m];
1130
1131 if (msp->ms_map.sm_start > mg->mg_bonus_area)
1132 break;
1133
1134 mutex_enter(&msp->ms_lock);
1135 metaslab_group_sort(mg, msp, metaslab_weight(msp));
1136 mutex_exit(&msp->ms_lock);
1137 }
1138
1139 atomic_add_64(&mg->mg_alloc_failures, -failures);
1140
1141 /*
1142 * Prefetch the next potential metaslabs
1143 */
1144 metaslab_prefetch(mg);
1145}
1146
1147static uint64_t
1148metaslab_distance(metaslab_t *msp, dva_t *dva)
1149{
1150 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
1151 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
1152 uint64_t start = msp->ms_map.sm_start >> ms_shift;
1153
1154 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
1155 return (1ULL << 63);
1156
1157 if (offset < start)
1158 return ((start - offset) << ms_shift);
1159 if (offset > start)
1160 return ((offset - start) << ms_shift);
1161 return (0);
1162}
1163
1164static uint64_t
1165metaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize,
1166 uint64_t txg, uint64_t min_distance, dva_t *dva, int d, int flags)
1167{
1168 spa_t *spa = mg->mg_vd->vdev_spa;
1169 metaslab_t *msp = NULL;
1170 uint64_t offset = -1ULL;
1171 avl_tree_t *t = &mg->mg_metaslab_tree;
1172 uint64_t activation_weight;
1173 uint64_t target_distance;
1174 int i;
1175
1176 activation_weight = METASLAB_WEIGHT_PRIMARY;
1177 for (i = 0; i < d; i++) {
1178 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
1179 activation_weight = METASLAB_WEIGHT_SECONDARY;
1180 break;
1181 }
1182 }
1183
1184 for (;;) {
1185 boolean_t was_active;
1186
1187 mutex_enter(&mg->mg_lock);
1188 for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
1189 if (msp->ms_weight < asize) {
1190 spa_dbgmsg(spa, "%s: failed to meet weight "
1191 "requirement: vdev %llu, txg %llu, mg %p, "
1192 "msp %p, psize %llu, asize %llu, "
1193 "failures %llu, weight %llu",
1194 spa_name(spa), mg->mg_vd->vdev_id, txg,
1195 mg, msp, psize, asize,
1196 mg->mg_alloc_failures, msp->ms_weight);
1197 mutex_exit(&mg->mg_lock);
1198 return (-1ULL);
1199 }
1200 was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
1201 if (activation_weight == METASLAB_WEIGHT_PRIMARY)
1202 break;
1203
1204 target_distance = min_distance +
1205 (msp->ms_smo.smo_alloc ? 0 : min_distance >> 1);
1206
1207 for (i = 0; i < d; i++)
1208 if (metaslab_distance(msp, &dva[i]) <
1209 target_distance)
1210 break;
1211 if (i == d)
1212 break;
1213 }
1214 mutex_exit(&mg->mg_lock);
1215 if (msp == NULL)
1216 return (-1ULL);
1217
1218 /*
1219 * If we've already reached the allowable number of failed
1220 * allocation attempts on this metaslab group then we
1221 * consider skipping it. We skip it only if we're allowed
1222 * to "fast" gang, the physical size is larger than
1223 * a gang block, and we're attempting to allocate from
1224 * the primary metaslab.
1225 */
1226 if (mg->mg_alloc_failures > zfs_mg_alloc_failures &&
1227 CAN_FASTGANG(flags) && psize > SPA_GANGBLOCKSIZE &&
1228 activation_weight == METASLAB_WEIGHT_PRIMARY) {
1229 spa_dbgmsg(spa, "%s: skipping metaslab group: "
1230 "vdev %llu, txg %llu, mg %p, psize %llu, "
1231 "asize %llu, failures %llu", spa_name(spa),
1232 mg->mg_vd->vdev_id, txg, mg, psize, asize,
1233 mg->mg_alloc_failures);
1234 return (-1ULL);
1235 }
1236
1237 mutex_enter(&msp->ms_lock);
1238
1239 /*
1240 * Ensure that the metaslab we have selected is still
1241 * capable of handling our request. It's possible that
1242 * another thread may have changed the weight while we
1243 * were blocked on the metaslab lock.
1244 */
1245 if (msp->ms_weight < asize || (was_active &&
1246 !(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
1247 activation_weight == METASLAB_WEIGHT_PRIMARY)) {
1248 mutex_exit(&msp->ms_lock);
1249 continue;
1250 }
1251
1252 if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
1253 activation_weight == METASLAB_WEIGHT_PRIMARY) {
1254 metaslab_passivate(msp,
1255 msp->ms_weight & ~METASLAB_ACTIVE_MASK);
1256 mutex_exit(&msp->ms_lock);
1257 continue;
1258 }
1259
1260 if (metaslab_activate(msp, activation_weight) != 0) {
1261 mutex_exit(&msp->ms_lock);
1262 continue;
1263 }
1264
1265 if ((offset = space_map_alloc(&msp->ms_map, asize)) != -1ULL)
1266 break;
1267
1268 atomic_inc_64(&mg->mg_alloc_failures);
1269
1270 metaslab_passivate(msp, space_map_maxsize(&msp->ms_map));
1271
1272 mutex_exit(&msp->ms_lock);
1273 }
1274
1275 if (msp->ms_allocmap[txg & TXG_MASK].sm_space == 0)
1276 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
1277
1278 space_map_add(&msp->ms_allocmap[txg & TXG_MASK], offset, asize);
1279
1280 mutex_exit(&msp->ms_lock);
1281
1282 return (offset);
1283}
1284
1285/*
1286 * Allocate a block for the specified i/o.
1287 */
1288static int
1289metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
1290 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
1291{
1292 metaslab_group_t *mg, *rotor;
1293 vdev_t *vd;
1294 int dshift = 3;
1295 int all_zero;
1296 int zio_lock = B_FALSE;
1297 boolean_t allocatable;
1298 uint64_t offset = -1ULL;
1299 uint64_t asize;
1300 uint64_t distance;
1301
1302 ASSERT(!DVA_IS_VALID(&dva[d]));
1303
1304 /*
1305 * For testing, make some blocks above a certain size be gang blocks.
1306 */
1307 if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0)
1308 return (ENOSPC);
1309
1310 /*
1311 * Start at the rotor and loop through all mgs until we find something.
1312 * Note that there's no locking on mc_rotor or mc_aliquot because
1313 * nothing actually breaks if we miss a few updates -- we just won't
1314 * allocate quite as evenly. It all balances out over time.
1315 *
1316 * If we are doing ditto or log blocks, try to spread them across
1317 * consecutive vdevs. If we're forced to reuse a vdev before we've
1318 * allocated all of our ditto blocks, then try and spread them out on
1319 * that vdev as much as possible. If it turns out to not be possible,
1320 * gradually lower our standards until anything becomes acceptable.
1321 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
1322 * gives us hope of containing our fault domains to something we're
1323 * able to reason about. Otherwise, any two top-level vdev failures
1324 * will guarantee the loss of data. With consecutive allocation,
1325 * only two adjacent top-level vdev failures will result in data loss.
1326 *
1327 * If we are doing gang blocks (hintdva is non-NULL), try to keep
1328 * ourselves on the same vdev as our gang block header. That
1329 * way, we can hope for locality in vdev_cache, plus it makes our
1330 * fault domains something tractable.
1331 */
1332 if (hintdva) {
1333 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
1334
1335 /*
1336 * It's possible the vdev we're using as the hint no
1337 * longer exists (i.e. removed). Consult the rotor when
1338 * all else fails.
1339 */
1340 if (vd != NULL) {
1341 mg = vd->vdev_mg;
1342
1343 if (flags & METASLAB_HINTBP_AVOID &&
1344 mg->mg_next != NULL)
1345 mg = mg->mg_next;
1346 } else {
1347 mg = mc->mc_rotor;
1348 }
1349 } else if (d != 0) {
1350 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
1351 mg = vd->vdev_mg->mg_next;
1352 } else {
1353 mg = mc->mc_rotor;
1354 }
1355
1356 /*
1357 * If the hint put us into the wrong metaslab class, or into a
1358 * metaslab group that has been passivated, just follow the rotor.
1359 */
1360 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
1361 mg = mc->mc_rotor;
1362
1363 rotor = mg;
1364top:
1365 all_zero = B_TRUE;
1366 do {
1367 ASSERT(mg->mg_activation_count == 1);
1368
1369 vd = mg->mg_vd;
1370
1371 /*
1372 * Don't allocate from faulted devices.
1373 */
1374 if (zio_lock) {
1375 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
1376 allocatable = vdev_allocatable(vd);
1377 spa_config_exit(spa, SCL_ZIO, FTAG);
1378 } else {
1379 allocatable = vdev_allocatable(vd);
1380 }
1381 if (!allocatable)
1382 goto next;
1383
1384 /*
1385 * Avoid writing single-copy data to a failing vdev
1386 */
1387 if ((vd->vdev_stat.vs_write_errors > 0 ||
1388 vd->vdev_state < VDEV_STATE_HEALTHY) &&
1389 d == 0 && dshift == 3) {
1390 all_zero = B_FALSE;
1391 goto next;
1392 }
1393
1394 ASSERT(mg->mg_class == mc);
1395
1396 distance = vd->vdev_asize >> dshift;
1397 if (distance <= (1ULL << vd->vdev_ms_shift))
1398 distance = 0;
1399 else
1400 all_zero = B_FALSE;
1401
1402 asize = vdev_psize_to_asize(vd, psize);
1403 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
1404
1405 offset = metaslab_group_alloc(mg, psize, asize, txg, distance,
1406 dva, d, flags);
1407 if (offset != -1ULL) {
1408 /*
1409 * If we've just selected this metaslab group,
1410 * figure out whether the corresponding vdev is
1411 * over- or under-used relative to the pool,
1412 * and set an allocation bias to even it out.
1413 */
1414 if (mc->mc_aliquot == 0) {
1415 vdev_stat_t *vs = &vd->vdev_stat;
1416 int64_t vu, cu;
1417
1418 vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
1419 cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
1420
1421 /*
1422 * Calculate how much more or less we should
1423 * try to allocate from this device during
1424 * this iteration around the rotor.
1425 * For example, if a device is 80% full
1426 * and the pool is 20% full then we should
1427 * reduce allocations by 60% on this device.
1428 *
1429 * mg_bias = (20 - 80) * 512K / 100 = -307K
1430 *
1431 * This reduces allocations by 307K for this
1432 * iteration.
1433 */
1434 mg->mg_bias = ((cu - vu) *
1435 (int64_t)mg->mg_aliquot) / 100;
1436 }
1437
1438 if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
1439 mg->mg_aliquot + mg->mg_bias) {
1440 mc->mc_rotor = mg->mg_next;
1441 mc->mc_aliquot = 0;
1442 }
1443
1444 DVA_SET_VDEV(&dva[d], vd->vdev_id);
1445 DVA_SET_OFFSET(&dva[d], offset);
1446 DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
1447 DVA_SET_ASIZE(&dva[d], asize);
1448
1449 return (0);
1450 }
1451next:
1452 mc->mc_rotor = mg->mg_next;
1453 mc->mc_aliquot = 0;
1454 } while ((mg = mg->mg_next) != rotor);
1455
1456 if (!all_zero) {
1457 dshift++;
1458 ASSERT(dshift < 64);
1459 goto top;
1460 }
1461
1462 if (!allocatable && !zio_lock) {
1463 dshift = 3;
1464 zio_lock = B_TRUE;
1465 goto top;
1466 }
1467
1468 bzero(&dva[d], sizeof (dva_t));
1469
1470 return (ENOSPC);
1471}
1472
1473/*
1474 * Free the block represented by DVA in the context of the specified
1475 * transaction group.
1476 */
1477static void
1478metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
1479{
1480 uint64_t vdev = DVA_GET_VDEV(dva);
1481 uint64_t offset = DVA_GET_OFFSET(dva);
1482 uint64_t size = DVA_GET_ASIZE(dva);
1483 vdev_t *vd;
1484 metaslab_t *msp;
1485
1486 ASSERT(DVA_IS_VALID(dva));
1487
1488 if (txg > spa_freeze_txg(spa))
1489 return;
1490
1491 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
1492 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
1493 cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
1494 (u_longlong_t)vdev, (u_longlong_t)offset);
1495 ASSERT(0);
1496 return;
1497 }
1498
1499 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1500
1501 if (DVA_GET_GANG(dva))
1502 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1503
1504 mutex_enter(&msp->ms_lock);
1505
1506 if (now) {
1507 space_map_remove(&msp->ms_allocmap[txg & TXG_MASK],
1508 offset, size);
1509 space_map_free(&msp->ms_map, offset, size);
1510 } else {
1511 if (msp->ms_freemap[txg & TXG_MASK].sm_space == 0)
1512 vdev_dirty(vd, VDD_METASLAB, msp, txg);
1513 space_map_add(&msp->ms_freemap[txg & TXG_MASK], offset, size);
1514 }
1515
1516 mutex_exit(&msp->ms_lock);
1517}
1518
1519/*
1520 * Intent log support: upon opening the pool after a crash, notify the SPA
1521 * of blocks that the intent log has allocated for immediate write, but
1522 * which are still considered free by the SPA because the last transaction
1523 * group didn't commit yet.
1524 */
1525static int
1526metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
1527{
1528 uint64_t vdev = DVA_GET_VDEV(dva);
1529 uint64_t offset = DVA_GET_OFFSET(dva);
1530 uint64_t size = DVA_GET_ASIZE(dva);
1531 vdev_t *vd;
1532 metaslab_t *msp;
1533 int error = 0;
1534
1535 ASSERT(DVA_IS_VALID(dva));
1536
1537 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
1538 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
1539 return (ENXIO);
1540
1541 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1542
1543 if (DVA_GET_GANG(dva))
1544 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1545
1546 mutex_enter(&msp->ms_lock);
1547
1548 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_map.sm_loaded)
1549 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
1550
1551 if (error == 0 && !space_map_contains(&msp->ms_map, offset, size))
1552 error = ENOENT;
1553
1554 if (error || txg == 0) { /* txg == 0 indicates dry run */
1555 mutex_exit(&msp->ms_lock);
1556 return (error);
1557 }
1558
1559 space_map_claim(&msp->ms_map, offset, size);
1560
1561 if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
1562 if (msp->ms_allocmap[txg & TXG_MASK].sm_space == 0)
1563 vdev_dirty(vd, VDD_METASLAB, msp, txg);
1564 space_map_add(&msp->ms_allocmap[txg & TXG_MASK], offset, size);
1565 }
1566
1567 mutex_exit(&msp->ms_lock);
1568
1569 return (0);
1570}
1571
1572int
1573metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
1574 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
1575{
1576 dva_t *dva = bp->blk_dva;
1577 dva_t *hintdva = hintbp->blk_dva;
1578 int error = 0;
1579
1580 ASSERT(bp->blk_birth == 0);
1581 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
1582
1583 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
1584
1585 if (mc->mc_rotor == NULL) { /* no vdevs in this class */
1586 spa_config_exit(spa, SCL_ALLOC, FTAG);
1587 return (ENOSPC);
1588 }
1589
1590 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
1591 ASSERT(BP_GET_NDVAS(bp) == 0);
1592 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
1593
1594 for (int d = 0; d < ndvas; d++) {
1595 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
1596 txg, flags);
1597 if (error) {
1598 for (d--; d >= 0; d--) {
1599 metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
1600 bzero(&dva[d], sizeof (dva_t));
1601 }
1602 spa_config_exit(spa, SCL_ALLOC, FTAG);
1603 return (error);
1604 }
1605 }
1606 ASSERT(error == 0);
1607 ASSERT(BP_GET_NDVAS(bp) == ndvas);
1608
1609 spa_config_exit(spa, SCL_ALLOC, FTAG);
1610
1611 BP_SET_BIRTH(bp, txg, txg);
1612
1613 return (0);
1614}
1615
1616void
1617metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
1618{
1619 const dva_t *dva = bp->blk_dva;
1620 int ndvas = BP_GET_NDVAS(bp);
1621
1622 ASSERT(!BP_IS_HOLE(bp));
1623 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
1624
1625 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
1626
1627 for (int d = 0; d < ndvas; d++)
1628 metaslab_free_dva(spa, &dva[d], txg, now);
1629
1630 spa_config_exit(spa, SCL_FREE, FTAG);
1631}
1632
1633int
1634metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
1635{
1636 const dva_t *dva = bp->blk_dva;
1637 int ndvas = BP_GET_NDVAS(bp);
1638 int error = 0;
1639
1640 ASSERT(!BP_IS_HOLE(bp));
1641
1642 if (txg != 0) {
1643 /*
1644 * First do a dry run to make sure all DVAs are claimable,
1645 * so we don't have to unwind from partial failures below.
1646 */
1647 if ((error = metaslab_claim(spa, bp, 0)) != 0)
1648 return (error);
1649 }
1650
1651 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
1652
1653 for (int d = 0; d < ndvas; d++)
1654 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
1655 break;
1656
1657 spa_config_exit(spa, SCL_ALLOC, FTAG);
1658
1659 ASSERT(error == 0 || txg == 0);
1660
1661 return (error);
1662}