1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2013 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 */
26
27 #include <sys/zfs_context.h>
28 #include <sys/dmu.h>
29 #include <sys/dmu_tx.h>
30 #include <sys/space_map.h>
31 #include <sys/metaslab_impl.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/zio.h>
34
35 /*
36 * Allow allocations to switch to gang blocks quickly. We do this to
37 * avoid having to load lots of space_maps in a given txg. There are,
38 * however, some cases where we want to avoid "fast" ganging and instead
39 * we want to do an exhaustive search of all metaslabs on this device.
40 * Currently we don't allow any gang, zil, or dump device related allocations
41 * to "fast" gang.
42 */
43 #define CAN_FASTGANG(flags) \
44 (!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \
45 METASLAB_GANG_AVOID)))
46
47 uint64_t metaslab_aliquot = 512ULL << 10;
48 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1; /* force gang blocks */
49
50 /*
51 * The in-core space map representation is more compact than its on-disk form.
52 * The zfs_condense_pct determines how much more compact the in-core
53 * space_map representation must be before we compact it on-disk.
54 * Values should be greater than or equal to 100.
55 */
56 int zfs_condense_pct = 200;
57
58 /*
59 * This value defines the number of allowed allocation failures per vdev.
60 * If a device reaches this threshold in a given txg then we consider skipping
61 * allocations on that device.
62 */
63 int zfs_mg_alloc_failures;
64
65 /*
66 * Metaslab debugging: when set, keeps all space maps in core to verify frees.
67 */
68 static int metaslab_debug = 0;
69
70 /*
71 * Minimum size which forces the dynamic allocator to change
72 * it's allocation strategy. Once the space map cannot satisfy
73 * an allocation of this size then it switches to using more
74 * aggressive strategy (i.e search by size rather than offset).
75 */
76 uint64_t metaslab_df_alloc_threshold = SPA_MAXBLOCKSIZE;
77
78 /*
79 * The minimum free space, in percent, which must be available
80 * in a space map to continue allocations in a first-fit fashion.
81 * Once the space_map's free space drops below this level we dynamically
82 * switch to using best-fit allocations.
83 */
84 int metaslab_df_free_pct = 4;
85
86 /*
87 * A metaslab is considered "free" if it contains a contiguous
88 * segment which is greater than metaslab_min_alloc_size.
89 */
90 uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
91
92 /*
93 * Max number of space_maps to prefetch.
94 */
95 int metaslab_prefetch_limit = SPA_DVAS_PER_BP;
96
97 /*
98 * Percentage bonus multiplier for metaslabs that are in the bonus area.
99 */
100 int metaslab_smo_bonus_pct = 150;
101
102 /*
103 * Should we be willing to write data to degraded vdevs?
104 */
105 boolean_t zfs_write_to_degraded = B_FALSE;
106
107 /*
108 * ==========================================================================
109 * Metaslab classes
110 * ==========================================================================
111 */
112 metaslab_class_t *
113 metaslab_class_create(spa_t *spa, space_map_ops_t *ops)
114 {
115 metaslab_class_t *mc;
116
117 mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
118
119 mc->mc_spa = spa;
120 mc->mc_rotor = NULL;
121 mc->mc_ops = ops;
122
123 return (mc);
124 }
125
126 void
127 metaslab_class_destroy(metaslab_class_t *mc)
128 {
129 ASSERT(mc->mc_rotor == NULL);
130 ASSERT(mc->mc_alloc == 0);
131 ASSERT(mc->mc_deferred == 0);
132 ASSERT(mc->mc_space == 0);
133 ASSERT(mc->mc_dspace == 0);
134
135 kmem_free(mc, sizeof (metaslab_class_t));
136 }
137
138 int
139 metaslab_class_validate(metaslab_class_t *mc)
140 {
141 metaslab_group_t *mg;
142 vdev_t *vd;
143
144 /*
145 * Must hold one of the spa_config locks.
146 */
147 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
148 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
149
150 if ((mg = mc->mc_rotor) == NULL)
151 return (0);
152
153 do {
154 vd = mg->mg_vd;
155 ASSERT(vd->vdev_mg != NULL);
156 ASSERT3P(vd->vdev_top, ==, vd);
157 ASSERT3P(mg->mg_class, ==, mc);
158 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
159 } while ((mg = mg->mg_next) != mc->mc_rotor);
160
161 return (0);
162 }
163
164 void
165 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
166 int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
167 {
168 atomic_add_64(&mc->mc_alloc, alloc_delta);
169 atomic_add_64(&mc->mc_deferred, defer_delta);
170 atomic_add_64(&mc->mc_space, space_delta);
171 atomic_add_64(&mc->mc_dspace, dspace_delta);
172 }
173
174 uint64_t
175 metaslab_class_get_alloc(metaslab_class_t *mc)
176 {
177 return (mc->mc_alloc);
178 }
179
180 uint64_t
181 metaslab_class_get_deferred(metaslab_class_t *mc)
182 {
183 return (mc->mc_deferred);
184 }
185
186 uint64_t
187 metaslab_class_get_space(metaslab_class_t *mc)
188 {
189 return (mc->mc_space);
190 }
191
192 uint64_t
193 metaslab_class_get_dspace(metaslab_class_t *mc)
194 {
195 return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
196 }
197
198 /*
199 * ==========================================================================
200 * Metaslab groups
201 * ==========================================================================
202 */
203 static int
204 metaslab_compare(const void *x1, const void *x2)
205 {
206 const metaslab_t *m1 = x1;
207 const metaslab_t *m2 = x2;
208
209 if (m1->ms_weight < m2->ms_weight)
210 return (1);
211 if (m1->ms_weight > m2->ms_weight)
212 return (-1);
213
214 /*
215 * If the weights are identical, use the offset to force uniqueness.
216 */
217 if (m1->ms_map->sm_start < m2->ms_map->sm_start)
218 return (-1);
219 if (m1->ms_map->sm_start > m2->ms_map->sm_start)
220 return (1);
221
222 ASSERT3P(m1, ==, m2);
223
224 return (0);
225 }
226
227 metaslab_group_t *
228 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
229 {
230 metaslab_group_t *mg;
231
232 mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
233 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
234 avl_create(&mg->mg_metaslab_tree, metaslab_compare,
235 sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
236 mg->mg_vd = vd;
237 mg->mg_class = mc;
238 mg->mg_activation_count = 0;
239
240 return (mg);
241 }
242
243 void
244 metaslab_group_destroy(metaslab_group_t *mg)
245 {
246 ASSERT(mg->mg_prev == NULL);
247 ASSERT(mg->mg_next == NULL);
248 /*
249 * We may have gone below zero with the activation count
250 * either because we never activated in the first place or
251 * because we're done, and possibly removing the vdev.
252 */
253 ASSERT(mg->mg_activation_count <= 0);
254
255 avl_destroy(&mg->mg_metaslab_tree);
256 mutex_destroy(&mg->mg_lock);
257 kmem_free(mg, sizeof (metaslab_group_t));
258 }
259
260 void
261 metaslab_group_activate(metaslab_group_t *mg)
262 {
263 metaslab_class_t *mc = mg->mg_class;
264 metaslab_group_t *mgprev, *mgnext;
265
266 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
267
268 ASSERT(mc->mc_rotor != mg);
269 ASSERT(mg->mg_prev == NULL);
270 ASSERT(mg->mg_next == NULL);
271 ASSERT(mg->mg_activation_count <= 0);
272
273 if (++mg->mg_activation_count <= 0)
274 return;
275
276 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
277
278 if ((mgprev = mc->mc_rotor) == NULL) {
279 mg->mg_prev = mg;
280 mg->mg_next = mg;
281 } else {
282 mgnext = mgprev->mg_next;
283 mg->mg_prev = mgprev;
284 mg->mg_next = mgnext;
285 mgprev->mg_next = mg;
286 mgnext->mg_prev = mg;
287 }
288 mc->mc_rotor = mg;
289 }
290
291 void
292 metaslab_group_passivate(metaslab_group_t *mg)
293 {
294 metaslab_class_t *mc = mg->mg_class;
295 metaslab_group_t *mgprev, *mgnext;
296
297 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
298
299 if (--mg->mg_activation_count != 0) {
300 ASSERT(mc->mc_rotor != mg);
301 ASSERT(mg->mg_prev == NULL);
302 ASSERT(mg->mg_next == NULL);
303 ASSERT(mg->mg_activation_count < 0);
304 return;
305 }
306
307 mgprev = mg->mg_prev;
308 mgnext = mg->mg_next;
309
310 if (mg == mgnext) {
311 mc->mc_rotor = NULL;
312 } else {
313 mc->mc_rotor = mgnext;
314 mgprev->mg_next = mgnext;
315 mgnext->mg_prev = mgprev;
316 }
317
318 mg->mg_prev = NULL;
319 mg->mg_next = NULL;
320 }
321
322 static void
323 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
324 {
325 mutex_enter(&mg->mg_lock);
326 ASSERT(msp->ms_group == NULL);
327 msp->ms_group = mg;
328 msp->ms_weight = 0;
329 avl_add(&mg->mg_metaslab_tree, msp);
330 mutex_exit(&mg->mg_lock);
331 }
332
333 static void
334 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
335 {
336 mutex_enter(&mg->mg_lock);
337 ASSERT(msp->ms_group == mg);
338 avl_remove(&mg->mg_metaslab_tree, msp);
339 msp->ms_group = NULL;
340 mutex_exit(&mg->mg_lock);
341 }
342
343 static void
344 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
345 {
346 /*
347 * Although in principle the weight can be any value, in
348 * practice we do not use values in the range [1, 510].
349 */
350 ASSERT(weight >= SPA_MINBLOCKSIZE-1 || weight == 0);
351 ASSERT(MUTEX_HELD(&msp->ms_lock));
352
353 mutex_enter(&mg->mg_lock);
354 ASSERT(msp->ms_group == mg);
355 avl_remove(&mg->mg_metaslab_tree, msp);
356 msp->ms_weight = weight;
357 avl_add(&mg->mg_metaslab_tree, msp);
358 mutex_exit(&mg->mg_lock);
359 }
360
361 /*
362 * ==========================================================================
363 * Common allocator routines
364 * ==========================================================================
365 */
366 static int
367 metaslab_segsize_compare(const void *x1, const void *x2)
368 {
369 const space_seg_t *s1 = x1;
370 const space_seg_t *s2 = x2;
371 uint64_t ss_size1 = s1->ss_end - s1->ss_start;
372 uint64_t ss_size2 = s2->ss_end - s2->ss_start;
373
374 if (ss_size1 < ss_size2)
375 return (-1);
376 if (ss_size1 > ss_size2)
377 return (1);
378
379 if (s1->ss_start < s2->ss_start)
380 return (-1);
381 if (s1->ss_start > s2->ss_start)
382 return (1);
383
384 return (0);
385 }
386
387 /*
388 * This is a helper function that can be used by the allocator to find
389 * a suitable block to allocate. This will search the specified AVL
390 * tree looking for a block that matches the specified criteria.
391 */
392 static uint64_t
393 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
394 uint64_t align)
395 {
396 space_seg_t *ss, ssearch;
397 avl_index_t where;
398
399 ssearch.ss_start = *cursor;
400 ssearch.ss_end = *cursor + size;
401
402 ss = avl_find(t, &ssearch, &where);
403 if (ss == NULL)
404 ss = avl_nearest(t, where, AVL_AFTER);
405
406 while (ss != NULL) {
407 uint64_t offset = P2ROUNDUP(ss->ss_start, align);
408
409 if (offset + size <= ss->ss_end) {
410 *cursor = offset + size;
411 return (offset);
412 }
413 ss = AVL_NEXT(t, ss);
414 }
415
416 /*
417 * If we know we've searched the whole map (*cursor == 0), give up.
418 * Otherwise, reset the cursor to the beginning and try again.
419 */
420 if (*cursor == 0)
421 return (-1ULL);
422
423 *cursor = 0;
424 return (metaslab_block_picker(t, cursor, size, align));
425 }
426
427 static void
428 metaslab_pp_load(space_map_t *sm)
429 {
430 space_seg_t *ss;
431
432 ASSERT(sm->sm_ppd == NULL);
433 sm->sm_ppd = kmem_zalloc(64 * sizeof (uint64_t), KM_SLEEP);
434
435 sm->sm_pp_root = kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
436 avl_create(sm->sm_pp_root, metaslab_segsize_compare,
437 sizeof (space_seg_t), offsetof(struct space_seg, ss_pp_node));
438
439 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
440 avl_add(sm->sm_pp_root, ss);
441 }
442
443 static void
444 metaslab_pp_unload(space_map_t *sm)
445 {
446 void *cookie = NULL;
447
448 kmem_free(sm->sm_ppd, 64 * sizeof (uint64_t));
449 sm->sm_ppd = NULL;
450
451 while (avl_destroy_nodes(sm->sm_pp_root, &cookie) != NULL) {
452 /* tear down the tree */
453 }
454
455 avl_destroy(sm->sm_pp_root);
456 kmem_free(sm->sm_pp_root, sizeof (avl_tree_t));
457 sm->sm_pp_root = NULL;
458 }
459
460 /* ARGSUSED */
461 static void
462 metaslab_pp_claim(space_map_t *sm, uint64_t start, uint64_t size)
463 {
464 /* No need to update cursor */
465 }
466
467 /* ARGSUSED */
468 static void
469 metaslab_pp_free(space_map_t *sm, uint64_t start, uint64_t size)
470 {
471 /* No need to update cursor */
472 }
473
474 /*
475 * Return the maximum contiguous segment within the metaslab.
476 */
477 uint64_t
478 metaslab_pp_maxsize(space_map_t *sm)
479 {
480 avl_tree_t *t = sm->sm_pp_root;
481 space_seg_t *ss;
482
483 if (t == NULL || (ss = avl_last(t)) == NULL)
484 return (0ULL);
485
486 return (ss->ss_end - ss->ss_start);
487 }
488
489 /*
490 * ==========================================================================
491 * The first-fit block allocator
492 * ==========================================================================
493 */
494 static uint64_t
495 metaslab_ff_alloc(space_map_t *sm, uint64_t size)
496 {
497 avl_tree_t *t = &sm->sm_root;
498 uint64_t align = size & -size;
499 uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
500
501 return (metaslab_block_picker(t, cursor, size, align));
502 }
503
504 /* ARGSUSED */
505 boolean_t
506 metaslab_ff_fragmented(space_map_t *sm)
507 {
508 return (B_TRUE);
509 }
510
511 static space_map_ops_t metaslab_ff_ops = {
512 metaslab_pp_load,
513 metaslab_pp_unload,
514 metaslab_ff_alloc,
515 metaslab_pp_claim,
516 metaslab_pp_free,
517 metaslab_pp_maxsize,
518 metaslab_ff_fragmented
519 };
520
521 /*
522 * ==========================================================================
523 * Dynamic block allocator -
524 * Uses the first fit allocation scheme until space get low and then
525 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
526 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
527 * ==========================================================================
528 */
529 static uint64_t
530 metaslab_df_alloc(space_map_t *sm, uint64_t size)
531 {
532 avl_tree_t *t = &sm->sm_root;
533 uint64_t align = size & -size;
534 uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
535 uint64_t max_size = metaslab_pp_maxsize(sm);
536 int free_pct = sm->sm_space * 100 / sm->sm_size;
537
538 ASSERT(MUTEX_HELD(sm->sm_lock));
539 ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
540
541 if (max_size < size)
542 return (-1ULL);
543
544 /*
545 * If we're running low on space switch to using the size
546 * sorted AVL tree (best-fit).
547 */
548 if (max_size < metaslab_df_alloc_threshold ||
549 free_pct < metaslab_df_free_pct) {
550 t = sm->sm_pp_root;
551 *cursor = 0;
552 }
553
554 return (metaslab_block_picker(t, cursor, size, 1ULL));
555 }
556
557 static boolean_t
558 metaslab_df_fragmented(space_map_t *sm)
559 {
560 uint64_t max_size = metaslab_pp_maxsize(sm);
561 int free_pct = sm->sm_space * 100 / sm->sm_size;
562
563 if (max_size >= metaslab_df_alloc_threshold &&
564 free_pct >= metaslab_df_free_pct)
565 return (B_FALSE);
566
567 return (B_TRUE);
568 }
569
570 static space_map_ops_t metaslab_df_ops = {
571 metaslab_pp_load,
572 metaslab_pp_unload,
573 metaslab_df_alloc,
574 metaslab_pp_claim,
575 metaslab_pp_free,
576 metaslab_pp_maxsize,
577 metaslab_df_fragmented
578 };
579
580 /*
581 * ==========================================================================
582 * Other experimental allocators
583 * ==========================================================================
584 */
585 static uint64_t
586 metaslab_cdf_alloc(space_map_t *sm, uint64_t size)
587 {
588 avl_tree_t *t = &sm->sm_root;
589 uint64_t *cursor = (uint64_t *)sm->sm_ppd;
590 uint64_t *extent_end = (uint64_t *)sm->sm_ppd + 1;
591 uint64_t max_size = metaslab_pp_maxsize(sm);
592 uint64_t rsize = size;
593 uint64_t offset = 0;
594
595 ASSERT(MUTEX_HELD(sm->sm_lock));
596 ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
597
598 if (max_size < size)
599 return (-1ULL);
600
601 ASSERT3U(*extent_end, >=, *cursor);
602
603 /*
604 * If we're running low on space switch to using the size
605 * sorted AVL tree (best-fit).
606 */
607 if ((*cursor + size) > *extent_end) {
608
609 t = sm->sm_pp_root;
610 *cursor = *extent_end = 0;
611
612 if (max_size > 2 * SPA_MAXBLOCKSIZE)
613 rsize = MIN(metaslab_min_alloc_size, max_size);
614 offset = metaslab_block_picker(t, extent_end, rsize, 1ULL);
615 if (offset != -1)
616 *cursor = offset + size;
617 } else {
618 offset = metaslab_block_picker(t, cursor, rsize, 1ULL);
619 }
620 ASSERT3U(*cursor, <=, *extent_end);
621 return (offset);
622 }
623
624 static boolean_t
625 metaslab_cdf_fragmented(space_map_t *sm)
626 {
627 uint64_t max_size = metaslab_pp_maxsize(sm);
628
629 if (max_size > (metaslab_min_alloc_size * 10))
630 return (B_FALSE);
631 return (B_TRUE);
632 }
633
634 static space_map_ops_t metaslab_cdf_ops = {
635 metaslab_pp_load,
636 metaslab_pp_unload,
637 metaslab_cdf_alloc,
638 metaslab_pp_claim,
639 metaslab_pp_free,
640 metaslab_pp_maxsize,
641 metaslab_cdf_fragmented
642 };
643
644 uint64_t metaslab_ndf_clump_shift = 4;
645
646 static uint64_t
647 metaslab_ndf_alloc(space_map_t *sm, uint64_t size)
648 {
649 avl_tree_t *t = &sm->sm_root;
650 avl_index_t where;
651 space_seg_t *ss, ssearch;
652 uint64_t hbit = highbit(size);
653 uint64_t *cursor = (uint64_t *)sm->sm_ppd + hbit - 1;
654 uint64_t max_size = metaslab_pp_maxsize(sm);
655
656 ASSERT(MUTEX_HELD(sm->sm_lock));
657 ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
658
659 if (max_size < size)
660 return (-1ULL);
661
662 ssearch.ss_start = *cursor;
663 ssearch.ss_end = *cursor + size;
664
665 ss = avl_find(t, &ssearch, &where);
666 if (ss == NULL || (ss->ss_start + size > ss->ss_end)) {
667 t = sm->sm_pp_root;
668
669 ssearch.ss_start = 0;
670 ssearch.ss_end = MIN(max_size,
671 1ULL << (hbit + metaslab_ndf_clump_shift));
672 ss = avl_find(t, &ssearch, &where);
673 if (ss == NULL)
674 ss = avl_nearest(t, where, AVL_AFTER);
675 ASSERT(ss != NULL);
676 }
677
678 if (ss != NULL) {
679 if (ss->ss_start + size <= ss->ss_end) {
680 *cursor = ss->ss_start + size;
681 return (ss->ss_start);
682 }
683 }
684 return (-1ULL);
685 }
686
687 static boolean_t
688 metaslab_ndf_fragmented(space_map_t *sm)
689 {
690 uint64_t max_size = metaslab_pp_maxsize(sm);
691
692 if (max_size > (metaslab_min_alloc_size << metaslab_ndf_clump_shift))
693 return (B_FALSE);
694 return (B_TRUE);
695 }
696
697
698 static space_map_ops_t metaslab_ndf_ops = {
699 metaslab_pp_load,
700 metaslab_pp_unload,
701 metaslab_ndf_alloc,
702 metaslab_pp_claim,
703 metaslab_pp_free,
704 metaslab_pp_maxsize,
705 metaslab_ndf_fragmented
706 };
707
708 space_map_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
709
710 /*
711 * ==========================================================================
712 * Metaslabs
713 * ==========================================================================
714 */
715 metaslab_t *
716 metaslab_init(metaslab_group_t *mg, space_map_obj_t *smo,
717 uint64_t start, uint64_t size, uint64_t txg)
718 {
719 vdev_t *vd = mg->mg_vd;
720 metaslab_t *msp;
721
722 msp = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
723 mutex_init(&msp->ms_lock, NULL, MUTEX_DEFAULT, NULL);
724
725 msp->ms_smo_syncing = *smo;
726
727 /*
728 * We create the main space map here, but we don't create the
729 * allocmaps and freemaps until metaslab_sync_done(). This serves
730 * two purposes: it allows metaslab_sync_done() to detect the
731 * addition of new space; and for debugging, it ensures that we'd
732 * data fault on any attempt to use this metaslab before it's ready.
733 */
734 msp->ms_map = kmem_zalloc(sizeof (space_map_t), KM_SLEEP);
735 space_map_create(msp->ms_map, start, size,
736 vd->vdev_ashift, &msp->ms_lock);
737
738 metaslab_group_add(mg, msp);
739
740 if (metaslab_debug && smo->smo_object != 0) {
741 mutex_enter(&msp->ms_lock);
742 VERIFY(space_map_load(msp->ms_map, mg->mg_class->mc_ops,
743 SM_FREE, smo, spa_meta_objset(vd->vdev_spa)) == 0);
744 mutex_exit(&msp->ms_lock);
745 }
746
747 /*
748 * If we're opening an existing pool (txg == 0) or creating
749 * a new one (txg == TXG_INITIAL), all space is available now.
750 * If we're adding space to an existing pool, the new space
751 * does not become available until after this txg has synced.
752 */
753 if (txg <= TXG_INITIAL)
754 metaslab_sync_done(msp, 0);
755
756 if (txg != 0) {
757 vdev_dirty(vd, 0, NULL, txg);
758 vdev_dirty(vd, VDD_METASLAB, msp, txg);
759 }
760
761 return (msp);
762 }
763
764 void
765 metaslab_fini(metaslab_t *msp)
766 {
767 metaslab_group_t *mg = msp->ms_group;
768
769 vdev_space_update(mg->mg_vd,
770 -msp->ms_smo.smo_alloc, 0, -msp->ms_map->sm_size);
771
772 metaslab_group_remove(mg, msp);
773
774 mutex_enter(&msp->ms_lock);
775
776 space_map_unload(msp->ms_map);
777 space_map_destroy(msp->ms_map);
778 kmem_free(msp->ms_map, sizeof (*msp->ms_map));
779
780 for (int t = 0; t < TXG_SIZE; t++) {
781 space_map_destroy(msp->ms_allocmap[t]);
782 space_map_destroy(msp->ms_freemap[t]);
783 kmem_free(msp->ms_allocmap[t], sizeof (*msp->ms_allocmap[t]));
784 kmem_free(msp->ms_freemap[t], sizeof (*msp->ms_freemap[t]));
785 }
786
787 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
788 space_map_destroy(msp->ms_defermap[t]);
789 kmem_free(msp->ms_defermap[t], sizeof (*msp->ms_defermap[t]));
790 }
791
792 ASSERT0(msp->ms_deferspace);
793
794 mutex_exit(&msp->ms_lock);
795 mutex_destroy(&msp->ms_lock);
796
797 kmem_free(msp, sizeof (metaslab_t));
798 }
799
800 #define METASLAB_WEIGHT_PRIMARY (1ULL << 63)
801 #define METASLAB_WEIGHT_SECONDARY (1ULL << 62)
802 #define METASLAB_ACTIVE_MASK \
803 (METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
804
805 static uint64_t
806 metaslab_weight(metaslab_t *msp)
807 {
808 metaslab_group_t *mg = msp->ms_group;
809 space_map_t *sm = msp->ms_map;
810 space_map_obj_t *smo = &msp->ms_smo;
811 vdev_t *vd = mg->mg_vd;
812 uint64_t weight, space;
813
814 ASSERT(MUTEX_HELD(&msp->ms_lock));
815
816 /*
817 * This vdev is in the process of being removed so there is nothing
818 * for us to do here.
819 */
820 if (vd->vdev_removing) {
821 ASSERT0(smo->smo_alloc);
822 ASSERT0(vd->vdev_ms_shift);
823 return (0);
824 }
825
826 /*
827 * The baseline weight is the metaslab's free space.
828 */
829 space = sm->sm_size - smo->smo_alloc;
830 weight = space;
831
832 /*
833 * Modern disks have uniform bit density and constant angular velocity.
834 * Therefore, the outer recording zones are faster (higher bandwidth)
835 * than the inner zones by the ratio of outer to inner track diameter,
836 * which is typically around 2:1. We account for this by assigning
837 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
838 * In effect, this means that we'll select the metaslab with the most
839 * free bandwidth rather than simply the one with the most free space.
840 */
841 weight = 2 * weight -
842 ((sm->sm_start >> vd->vdev_ms_shift) * weight) / vd->vdev_ms_count;
843 ASSERT(weight >= space && weight <= 2 * space);
844
845 /*
846 * For locality, assign higher weight to metaslabs which have
847 * a lower offset than what we've already activated.
848 */
849 if (sm->sm_start <= mg->mg_bonus_area)
850 weight *= (metaslab_smo_bonus_pct / 100);
851 ASSERT(weight >= space &&
852 weight <= 2 * (metaslab_smo_bonus_pct / 100) * space);
853
854 if (sm->sm_loaded && !sm->sm_ops->smop_fragmented(sm)) {
855 /*
856 * If this metaslab is one we're actively using, adjust its
857 * weight to make it preferable to any inactive metaslab so
858 * we'll polish it off.
859 */
860 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
861 }
862 return (weight);
863 }
864
865 static void
866 metaslab_prefetch(metaslab_group_t *mg)
867 {
868 spa_t *spa = mg->mg_vd->vdev_spa;
869 metaslab_t *msp;
870 avl_tree_t *t = &mg->mg_metaslab_tree;
871 int m;
872
873 mutex_enter(&mg->mg_lock);
874
875 /*
876 * Prefetch the next potential metaslabs
877 */
878 for (msp = avl_first(t), m = 0; msp; msp = AVL_NEXT(t, msp), m++) {
879 space_map_t *sm = msp->ms_map;
880 space_map_obj_t *smo = &msp->ms_smo;
881
882 /* If we have reached our prefetch limit then we're done */
883 if (m >= metaslab_prefetch_limit)
884 break;
885
886 if (!sm->sm_loaded && smo->smo_object != 0) {
887 mutex_exit(&mg->mg_lock);
888 dmu_prefetch(spa_meta_objset(spa), smo->smo_object,
889 0ULL, smo->smo_objsize);
890 mutex_enter(&mg->mg_lock);
891 }
892 }
893 mutex_exit(&mg->mg_lock);
894 }
895
896 static int
897 metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
898 {
899 metaslab_group_t *mg = msp->ms_group;
900 space_map_t *sm = msp->ms_map;
901 space_map_ops_t *sm_ops = msp->ms_group->mg_class->mc_ops;
902
903 ASSERT(MUTEX_HELD(&msp->ms_lock));
904
905 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
906 space_map_load_wait(sm);
907 if (!sm->sm_loaded) {
908 space_map_obj_t *smo = &msp->ms_smo;
909
910 int error = space_map_load(sm, sm_ops, SM_FREE, smo,
911 spa_meta_objset(msp->ms_group->mg_vd->vdev_spa));
912 if (error) {
913 metaslab_group_sort(msp->ms_group, msp, 0);
914 return (error);
915 }
916 for (int t = 0; t < TXG_DEFER_SIZE; t++)
917 space_map_walk(msp->ms_defermap[t],
918 space_map_claim, sm);
919
920 }
921
922 /*
923 * Track the bonus area as we activate new metaslabs.
924 */
925 if (sm->sm_start > mg->mg_bonus_area) {
926 mutex_enter(&mg->mg_lock);
927 mg->mg_bonus_area = sm->sm_start;
928 mutex_exit(&mg->mg_lock);
929 }
930
931 metaslab_group_sort(msp->ms_group, msp,
932 msp->ms_weight | activation_weight);
933 }
934 ASSERT(sm->sm_loaded);
935 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
936
937 return (0);
938 }
939
940 static void
941 metaslab_passivate(metaslab_t *msp, uint64_t size)
942 {
943 /*
944 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
945 * this metaslab again. In that case, it had better be empty,
946 * or we would be leaving space on the table.
947 */
948 ASSERT(size >= SPA_MINBLOCKSIZE || msp->ms_map->sm_space == 0);
949 metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
950 ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
951 }
952
953 /*
954 * Determine if the in-core space map representation can be condensed on-disk.
955 * We would like to use the following criteria to make our decision:
956 *
957 * 1. The size of the space map object should not dramatically increase as a
958 * result of writing out our in-core free map.
959 *
960 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
961 * times the size than the in-core representation (i.e. zfs_condense_pct = 110
962 * and in-core = 1MB, minimal = 1.1.MB).
963 *
964 * Checking the first condition is tricky since we don't want to walk
965 * the entire AVL tree calculating the estimated on-disk size. Instead we
966 * use the size-ordered AVL tree in the space map and calculate the
967 * size required for the largest segment in our in-core free map. If the
968 * size required to represent that segment on disk is larger than the space
969 * map object then we avoid condensing this map.
970 *
971 * To determine the second criterion we use a best-case estimate and assume
972 * each segment can be represented on-disk as a single 64-bit entry. We refer
973 * to this best-case estimate as the space map's minimal form.
974 */
975 static boolean_t
976 metaslab_should_condense(metaslab_t *msp)
977 {
978 space_map_t *sm = msp->ms_map;
979 space_map_obj_t *smo = &msp->ms_smo_syncing;
980 space_seg_t *ss;
981 uint64_t size, entries, segsz;
982
983 ASSERT(MUTEX_HELD(&msp->ms_lock));
984 ASSERT(sm->sm_loaded);
985
986 /*
987 * Use the sm_pp_root AVL tree, which is ordered by size, to obtain
988 * the largest segment in the in-core free map. If the tree is
989 * empty then we should condense the map.
990 */
991 ss = avl_last(sm->sm_pp_root);
992 if (ss == NULL)
993 return (B_TRUE);
994
995 /*
996 * Calculate the number of 64-bit entries this segment would
997 * require when written to disk. If this single segment would be
998 * larger on-disk than the entire current on-disk structure, then
999 * clearly condensing will increase the on-disk structure size.
1000 */
1001 size = (ss->ss_end - ss->ss_start) >> sm->sm_shift;
1002 entries = size / (MIN(size, SM_RUN_MAX));
1003 segsz = entries * sizeof (uint64_t);
1004
1005 return (segsz <= smo->smo_objsize &&
1006 smo->smo_objsize >= (zfs_condense_pct *
1007 sizeof (uint64_t) * avl_numnodes(&sm->sm_root)) / 100);
1008 }
1009
1010 /*
1011 * Condense the on-disk space map representation to its minimized form.
1012 * The minimized form consists of a small number of allocations followed by
1013 * the in-core free map.
1014 */
1015 static void
1016 metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
1017 {
1018 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1019 space_map_t *freemap = msp->ms_freemap[txg & TXG_MASK];
1020 space_map_t condense_map;
1021 space_map_t *sm = msp->ms_map;
1022 objset_t *mos = spa_meta_objset(spa);
1023 space_map_obj_t *smo = &msp->ms_smo_syncing;
1024
1025 ASSERT(MUTEX_HELD(&msp->ms_lock));
1026 ASSERT3U(spa_sync_pass(spa), ==, 1);
1027 ASSERT(sm->sm_loaded);
1028
1029 spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, "
1030 "smo size %llu, segments %lu", txg,
1031 (msp->ms_map->sm_start / msp->ms_map->sm_size), msp,
1032 smo->smo_objsize, avl_numnodes(&sm->sm_root));
1033
1034 /*
1035 * Create an map that is a 100% allocated map. We remove segments
1036 * that have been freed in this txg, any deferred frees that exist,
1037 * and any allocation in the future. Removing segments should be
1038 * a relatively inexpensive operation since we expect these maps to
1039 * a small number of nodes.
1040 */
1041 space_map_create(&condense_map, sm->sm_start, sm->sm_size,
1042 sm->sm_shift, sm->sm_lock);
1043 space_map_add(&condense_map, condense_map.sm_start,
1044 condense_map.sm_size);
1045
1046 /*
1047 * Remove what's been freed in this txg from the condense_map.
1048 * Since we're in sync_pass 1, we know that all the frees from
1049 * this txg are in the freemap.
1050 */
1051 space_map_walk(freemap, space_map_remove, &condense_map);
1052
1053 for (int t = 0; t < TXG_DEFER_SIZE; t++)
1054 space_map_walk(msp->ms_defermap[t],
1055 space_map_remove, &condense_map);
1056
1057 for (int t = 1; t < TXG_CONCURRENT_STATES; t++)
1058 space_map_walk(msp->ms_allocmap[(txg + t) & TXG_MASK],
1059 space_map_remove, &condense_map);
1060
1061 /*
1062 * We're about to drop the metaslab's lock thus allowing
1063 * other consumers to change it's content. Set the
1064 * space_map's sm_condensing flag to ensure that
1065 * allocations on this metaslab do not occur while we're
1066 * in the middle of committing it to disk. This is only critical
1067 * for the ms_map as all other space_maps use per txg
1068 * views of their content.
1069 */
1070 sm->sm_condensing = B_TRUE;
1071
1072 mutex_exit(&msp->ms_lock);
1073 space_map_truncate(smo, mos, tx);
1074 mutex_enter(&msp->ms_lock);
1075
1076 /*
1077 * While we would ideally like to create a space_map representation
1078 * that consists only of allocation records, doing so can be
1079 * prohibitively expensive because the in-core free map can be
1080 * large, and therefore computationally expensive to subtract
1081 * from the condense_map. Instead we sync out two maps, a cheap
1082 * allocation only map followed by the in-core free map. While not
1083 * optimal, this is typically close to optimal, and much cheaper to
1084 * compute.
1085 */
1086 space_map_sync(&condense_map, SM_ALLOC, smo, mos, tx);
1087 space_map_vacate(&condense_map, NULL, NULL);
1088 space_map_destroy(&condense_map);
1089
1090 space_map_sync(sm, SM_FREE, smo, mos, tx);
1091 sm->sm_condensing = B_FALSE;
1092
1093 spa_dbgmsg(spa, "condensed: txg %llu, msp[%llu] %p, "
1094 "smo size %llu", txg,
1095 (msp->ms_map->sm_start / msp->ms_map->sm_size), msp,
1096 smo->smo_objsize);
1097 }
1098
1099 /*
1100 * Write a metaslab to disk in the context of the specified transaction group.
1101 */
1102 void
1103 metaslab_sync(metaslab_t *msp, uint64_t txg)
1104 {
1105 vdev_t *vd = msp->ms_group->mg_vd;
1106 spa_t *spa = vd->vdev_spa;
1107 objset_t *mos = spa_meta_objset(spa);
1108 space_map_t *allocmap = msp->ms_allocmap[txg & TXG_MASK];
1109 space_map_t **freemap = &msp->ms_freemap[txg & TXG_MASK];
1110 space_map_t **freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
1111 space_map_t *sm = msp->ms_map;
1112 space_map_obj_t *smo = &msp->ms_smo_syncing;
1113 dmu_buf_t *db;
1114 dmu_tx_t *tx;
1115
1116 ASSERT(!vd->vdev_ishole);
1117
1118 /*
1119 * This metaslab has just been added so there's no work to do now.
1120 */
1121 if (*freemap == NULL) {
1122 ASSERT3P(allocmap, ==, NULL);
1123 return;
1124 }
1125
1126 ASSERT3P(allocmap, !=, NULL);
1127 ASSERT3P(*freemap, !=, NULL);
1128 ASSERT3P(*freed_map, !=, NULL);
1129
1130 if (allocmap->sm_space == 0 && (*freemap)->sm_space == 0)
1131 return;
1132
1133 /*
1134 * The only state that can actually be changing concurrently with
1135 * metaslab_sync() is the metaslab's ms_map. No other thread can
1136 * be modifying this txg's allocmap, freemap, freed_map, or smo.
1137 * Therefore, we only hold ms_lock to satify space_map ASSERTs.
1138 * We drop it whenever we call into the DMU, because the DMU
1139 * can call down to us (e.g. via zio_free()) at any time.
1140 */
1141
1142 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
1143
1144 if (smo->smo_object == 0) {
1145 ASSERT(smo->smo_objsize == 0);
1146 ASSERT(smo->smo_alloc == 0);
1147 smo->smo_object = dmu_object_alloc(mos,
1148 DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
1149 DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
1150 ASSERT(smo->smo_object != 0);
1151 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
1152 (sm->sm_start >> vd->vdev_ms_shift),
1153 sizeof (uint64_t), &smo->smo_object, tx);
1154 }
1155
1156 mutex_enter(&msp->ms_lock);
1157
1158 if (sm->sm_loaded && spa_sync_pass(spa) == 1 &&
1159 metaslab_should_condense(msp)) {
1160 metaslab_condense(msp, txg, tx);
1161 } else {
1162 space_map_sync(allocmap, SM_ALLOC, smo, mos, tx);
1163 space_map_sync(*freemap, SM_FREE, smo, mos, tx);
1164 }
1165
1166 space_map_vacate(allocmap, NULL, NULL);
1167
1168 /*
1169 * For sync pass 1, we avoid walking the entire space map and
1170 * instead will just swap the pointers for freemap and
1171 * freed_map. We can safely do this since the freed_map is
1172 * guaranteed to be empty on the initial pass.
1173 */
1174 if (spa_sync_pass(spa) == 1) {
1175 ASSERT0((*freed_map)->sm_space);
1176 ASSERT0(avl_numnodes(&(*freed_map)->sm_root));
1177 space_map_swap(freemap, freed_map);
1178 } else {
1179 space_map_vacate(*freemap, space_map_add, *freed_map);
1180 }
1181
1182 ASSERT0(msp->ms_allocmap[txg & TXG_MASK]->sm_space);
1183 ASSERT0(msp->ms_freemap[txg & TXG_MASK]->sm_space);
1184
1185 mutex_exit(&msp->ms_lock);
1186
1187 VERIFY0(dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
1188 dmu_buf_will_dirty(db, tx);
1189 ASSERT3U(db->db_size, >=, sizeof (*smo));
1190 bcopy(smo, db->db_data, sizeof (*smo));
1191 dmu_buf_rele(db, FTAG);
1192
1193 dmu_tx_commit(tx);
1194 }
1195
1196 /*
1197 * Called after a transaction group has completely synced to mark
1198 * all of the metaslab's free space as usable.
1199 */
1200 void
1201 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
1202 {
1203 space_map_obj_t *smo = &msp->ms_smo;
1204 space_map_obj_t *smosync = &msp->ms_smo_syncing;
1205 space_map_t *sm = msp->ms_map;
1206 space_map_t **freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
1207 space_map_t **defer_map = &msp->ms_defermap[txg % TXG_DEFER_SIZE];
1208 metaslab_group_t *mg = msp->ms_group;
1209 vdev_t *vd = mg->mg_vd;
1210 int64_t alloc_delta, defer_delta;
1211
1212 ASSERT(!vd->vdev_ishole);
1213
1214 mutex_enter(&msp->ms_lock);
1215
1216 /*
1217 * If this metaslab is just becoming available, initialize its
1218 * allocmaps, freemaps, and defermap and add its capacity to the vdev.
1219 */
1220 if (*freed_map == NULL) {
1221 ASSERT(*defer_map == NULL);
1222 for (int t = 0; t < TXG_SIZE; t++) {
1223 msp->ms_allocmap[t] = kmem_zalloc(sizeof (space_map_t),
1224 KM_SLEEP);
1225 space_map_create(msp->ms_allocmap[t], sm->sm_start,
1226 sm->sm_size, sm->sm_shift, sm->sm_lock);
1227 msp->ms_freemap[t] = kmem_zalloc(sizeof (space_map_t),
1228 KM_SLEEP);
1229 space_map_create(msp->ms_freemap[t], sm->sm_start,
1230 sm->sm_size, sm->sm_shift, sm->sm_lock);
1231 }
1232
1233 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1234 msp->ms_defermap[t] = kmem_zalloc(sizeof (space_map_t),
1235 KM_SLEEP);
1236 space_map_create(msp->ms_defermap[t], sm->sm_start,
1237 sm->sm_size, sm->sm_shift, sm->sm_lock);
1238 }
1239
1240 freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
1241 defer_map = &msp->ms_defermap[txg % TXG_DEFER_SIZE];
1242
1243 vdev_space_update(vd, 0, 0, sm->sm_size);
1244 }
1245
1246 alloc_delta = smosync->smo_alloc - smo->smo_alloc;
1247 defer_delta = (*freed_map)->sm_space - (*defer_map)->sm_space;
1248
1249 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
1250
1251 ASSERT(msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0);
1252 ASSERT(msp->ms_freemap[txg & TXG_MASK]->sm_space == 0);
1253
1254 /*
1255 * If there's a space_map_load() in progress, wait for it to complete
1256 * so that we have a consistent view of the in-core space map.
1257 */
1258 space_map_load_wait(sm);
1259
1260 /*
1261 * Move the frees from the defer_map to this map (if it's loaded).
1262 * Swap the freed_map and the defer_map -- this is safe to do
1263 * because we've just emptied out the defer_map.
1264 */
1265 space_map_vacate(*defer_map, sm->sm_loaded ? space_map_free : NULL, sm);
1266 ASSERT0((*defer_map)->sm_space);
1267 ASSERT0(avl_numnodes(&(*defer_map)->sm_root));
1268 space_map_swap(freed_map, defer_map);
1269
1270 *smo = *smosync;
1271
1272 msp->ms_deferspace += defer_delta;
1273 ASSERT3S(msp->ms_deferspace, >=, 0);
1274 ASSERT3S(msp->ms_deferspace, <=, sm->sm_size);
1275 if (msp->ms_deferspace != 0) {
1276 /*
1277 * Keep syncing this metaslab until all deferred frees
1278 * are back in circulation.
1279 */
1280 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1281 }
1282
1283 /*
1284 * If the map is loaded but no longer active, evict it as soon as all
1285 * future allocations have synced. (If we unloaded it now and then
1286 * loaded a moment later, the map wouldn't reflect those allocations.)
1287 */
1288 if (sm->sm_loaded && (msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1289 int evictable = 1;
1290
1291 for (int t = 1; t < TXG_CONCURRENT_STATES; t++)
1292 if (msp->ms_allocmap[(txg + t) & TXG_MASK]->sm_space)
1293 evictable = 0;
1294
1295 if (evictable && !metaslab_debug)
1296 space_map_unload(sm);
1297 }
1298
1299 metaslab_group_sort(mg, msp, metaslab_weight(msp));
1300
1301 mutex_exit(&msp->ms_lock);
1302 }
1303
1304 void
1305 metaslab_sync_reassess(metaslab_group_t *mg)
1306 {
1307 vdev_t *vd = mg->mg_vd;
1308 int64_t failures = mg->mg_alloc_failures;
1309
1310 /*
1311 * Re-evaluate all metaslabs which have lower offsets than the
1312 * bonus area.
1313 */
1314 for (int m = 0; m < vd->vdev_ms_count; m++) {
1315 metaslab_t *msp = vd->vdev_ms[m];
1316
1317 if (msp->ms_map->sm_start > mg->mg_bonus_area)
1318 break;
1319
1320 mutex_enter(&msp->ms_lock);
1321 metaslab_group_sort(mg, msp, metaslab_weight(msp));
1322 mutex_exit(&msp->ms_lock);
1323 }
1324
1325 atomic_add_64(&mg->mg_alloc_failures, -failures);
1326
1327 /*
1328 * Prefetch the next potential metaslabs
1329 */
1330 metaslab_prefetch(mg);
1331 }
1332
1333 static uint64_t
1334 metaslab_distance(metaslab_t *msp, dva_t *dva)
1335 {
1336 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
1337 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
1338 uint64_t start = msp->ms_map->sm_start >> ms_shift;
1339
1340 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
1341 return (1ULL << 63);
1342
1343 if (offset < start)
1344 return ((start - offset) << ms_shift);
1345 if (offset > start)
1346 return ((offset - start) << ms_shift);
1347 return (0);
1348 }
1349
1350 static uint64_t
1351 metaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize,
1352 uint64_t txg, uint64_t min_distance, dva_t *dva, int d, int flags)
1353 {
1354 spa_t *spa = mg->mg_vd->vdev_spa;
1355 metaslab_t *msp = NULL;
1356 uint64_t offset = -1ULL;
1357 avl_tree_t *t = &mg->mg_metaslab_tree;
1358 uint64_t activation_weight;
1359 uint64_t target_distance;
1360 int i;
1361
1362 activation_weight = METASLAB_WEIGHT_PRIMARY;
1363 for (i = 0; i < d; i++) {
1364 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
1365 activation_weight = METASLAB_WEIGHT_SECONDARY;
1366 break;
1367 }
1368 }
1369
1370 for (;;) {
1371 boolean_t was_active;
1372
1373 mutex_enter(&mg->mg_lock);
1374 for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
1375 if (msp->ms_weight < asize) {
1376 spa_dbgmsg(spa, "%s: failed to meet weight "
1377 "requirement: vdev %llu, txg %llu, mg %p, "
1378 "msp %p, psize %llu, asize %llu, "
1379 "failures %llu, weight %llu",
1380 spa_name(spa), mg->mg_vd->vdev_id, txg,
1381 mg, msp, psize, asize,
1382 mg->mg_alloc_failures, msp->ms_weight);
1383 mutex_exit(&mg->mg_lock);
1384 return (-1ULL);
1385 }
1386
1387 /*
1388 * If the selected metaslab is condensing, skip it.
1389 */
1390 if (msp->ms_map->sm_condensing)
1391 continue;
1392
1393 was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
1394 if (activation_weight == METASLAB_WEIGHT_PRIMARY)
1395 break;
1396
1397 target_distance = min_distance +
1398 (msp->ms_smo.smo_alloc ? 0 : min_distance >> 1);
1399
1400 for (i = 0; i < d; i++)
1401 if (metaslab_distance(msp, &dva[i]) <
1402 target_distance)
1403 break;
1404 if (i == d)
1405 break;
1406 }
1407 mutex_exit(&mg->mg_lock);
1408 if (msp == NULL)
1409 return (-1ULL);
1410
1411 /*
1412 * If we've already reached the allowable number of failed
1413 * allocation attempts on this metaslab group then we
1414 * consider skipping it. We skip it only if we're allowed
1415 * to "fast" gang, the physical size is larger than
1416 * a gang block, and we're attempting to allocate from
1417 * the primary metaslab.
1418 */
1419 if (mg->mg_alloc_failures > zfs_mg_alloc_failures &&
1420 CAN_FASTGANG(flags) && psize > SPA_GANGBLOCKSIZE &&
1421 activation_weight == METASLAB_WEIGHT_PRIMARY) {
1422 spa_dbgmsg(spa, "%s: skipping metaslab group: "
1423 "vdev %llu, txg %llu, mg %p, psize %llu, "
1424 "asize %llu, failures %llu", spa_name(spa),
1425 mg->mg_vd->vdev_id, txg, mg, psize, asize,
1426 mg->mg_alloc_failures);
1427 return (-1ULL);
1428 }
1429
1430 mutex_enter(&msp->ms_lock);
1431
1432 /*
1433 * Ensure that the metaslab we have selected is still
1434 * capable of handling our request. It's possible that
1435 * another thread may have changed the weight while we
1436 * were blocked on the metaslab lock.
1437 */
1438 if (msp->ms_weight < asize || (was_active &&
1439 !(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
1440 activation_weight == METASLAB_WEIGHT_PRIMARY)) {
1441 mutex_exit(&msp->ms_lock);
1442 continue;
1443 }
1444
1445 if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
1446 activation_weight == METASLAB_WEIGHT_PRIMARY) {
1447 metaslab_passivate(msp,
1448 msp->ms_weight & ~METASLAB_ACTIVE_MASK);
1449 mutex_exit(&msp->ms_lock);
1450 continue;
1451 }
1452
1453 if (metaslab_activate(msp, activation_weight) != 0) {
1454 mutex_exit(&msp->ms_lock);
1455 continue;
1456 }
1457
1458 /*
1459 * If this metaslab is currently condensing then pick again as
1460 * we can't manipulate this metaslab until it's committed
1461 * to disk.
1462 */
1463 if (msp->ms_map->sm_condensing) {
1464 mutex_exit(&msp->ms_lock);
1465 continue;
1466 }
1467
1468 if ((offset = space_map_alloc(msp->ms_map, asize)) != -1ULL)
1469 break;
1470
1471 atomic_inc_64(&mg->mg_alloc_failures);
1472
1473 metaslab_passivate(msp, space_map_maxsize(msp->ms_map));
1474
1475 mutex_exit(&msp->ms_lock);
1476 }
1477
1478 if (msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0)
1479 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
1480
1481 space_map_add(msp->ms_allocmap[txg & TXG_MASK], offset, asize);
1482
1483 mutex_exit(&msp->ms_lock);
1484
1485 return (offset);
1486 }
1487
1488 /*
1489 * Allocate a block for the specified i/o.
1490 */
1491 static int
1492 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
1493 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
1494 {
1495 metaslab_group_t *mg, *rotor;
1496 vdev_t *vd;
1497 int dshift = 3;
1498 int all_zero;
1499 int zio_lock = B_FALSE;
1500 boolean_t allocatable;
1501 uint64_t offset = -1ULL;
1502 uint64_t asize;
1503 uint64_t distance;
1504
1505 ASSERT(!DVA_IS_VALID(&dva[d]));
1506
1507 /*
1508 * For testing, make some blocks above a certain size be gang blocks.
1509 */
1510 if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0)
1511 return (SET_ERROR(ENOSPC));
1512
1513 /*
1514 * Start at the rotor and loop through all mgs until we find something.
1515 * Note that there's no locking on mc_rotor or mc_aliquot because
1516 * nothing actually breaks if we miss a few updates -- we just won't
1517 * allocate quite as evenly. It all balances out over time.
1518 *
1519 * If we are doing ditto or log blocks, try to spread them across
1520 * consecutive vdevs. If we're forced to reuse a vdev before we've
1521 * allocated all of our ditto blocks, then try and spread them out on
1522 * that vdev as much as possible. If it turns out to not be possible,
1523 * gradually lower our standards until anything becomes acceptable.
1524 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
1525 * gives us hope of containing our fault domains to something we're
1526 * able to reason about. Otherwise, any two top-level vdev failures
1527 * will guarantee the loss of data. With consecutive allocation,
1528 * only two adjacent top-level vdev failures will result in data loss.
1529 *
1530 * If we are doing gang blocks (hintdva is non-NULL), try to keep
1531 * ourselves on the same vdev as our gang block header. That
1532 * way, we can hope for locality in vdev_cache, plus it makes our
1533 * fault domains something tractable.
1534 */
1535 if (hintdva) {
1536 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
1537
1538 /*
1539 * It's possible the vdev we're using as the hint no
1540 * longer exists (i.e. removed). Consult the rotor when
1541 * all else fails.
1542 */
1543 if (vd != NULL) {
1544 mg = vd->vdev_mg;
1545
1546 if (flags & METASLAB_HINTBP_AVOID &&
1547 mg->mg_next != NULL)
1548 mg = mg->mg_next;
1549 } else {
1550 mg = mc->mc_rotor;
1551 }
1552 } else if (d != 0) {
1553 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
1554 mg = vd->vdev_mg->mg_next;
1555 } else {
1556 mg = mc->mc_rotor;
1557 }
1558
1559 /*
1560 * If the hint put us into the wrong metaslab class, or into a
1561 * metaslab group that has been passivated, just follow the rotor.
1562 */
1563 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
1564 mg = mc->mc_rotor;
1565
1566 rotor = mg;
1567 top:
1568 all_zero = B_TRUE;
1569 do {
1570 ASSERT(mg->mg_activation_count == 1);
1571
1572 vd = mg->mg_vd;
1573
1574 /*
1575 * Don't allocate from faulted devices.
1576 */
1577 if (zio_lock) {
1578 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
1579 allocatable = vdev_allocatable(vd);
1580 spa_config_exit(spa, SCL_ZIO, FTAG);
1581 } else {
1582 allocatable = vdev_allocatable(vd);
1583 }
1584 if (!allocatable)
1585 goto next;
1586
1587 /*
1588 * Avoid writing single-copy data to a failing vdev
1589 * unless the user instructs us that it is okay.
1590 */
1591 if ((vd->vdev_stat.vs_write_errors > 0 ||
1592 vd->vdev_state < VDEV_STATE_HEALTHY) &&
1593 d == 0 && dshift == 3 &&
1594 !(zfs_write_to_degraded && vd->vdev_state ==
1595 VDEV_STATE_DEGRADED)) {
1596 all_zero = B_FALSE;
1597 goto next;
1598 }
1599
1600 ASSERT(mg->mg_class == mc);
1601
1602 distance = vd->vdev_asize >> dshift;
1603 if (distance <= (1ULL << vd->vdev_ms_shift))
1604 distance = 0;
1605 else
1606 all_zero = B_FALSE;
1607
1608 asize = vdev_psize_to_asize(vd, psize);
1609 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
1610
1611 offset = metaslab_group_alloc(mg, psize, asize, txg, distance,
1612 dva, d, flags);
1613 if (offset != -1ULL) {
1614 /*
1615 * If we've just selected this metaslab group,
1616 * figure out whether the corresponding vdev is
1617 * over- or under-used relative to the pool,
1618 * and set an allocation bias to even it out.
1619 */
1620 if (mc->mc_aliquot == 0) {
1621 vdev_stat_t *vs = &vd->vdev_stat;
1622 int64_t vu, cu;
1623
1624 vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
1625 cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
1626
1627 /*
1628 * Calculate how much more or less we should
1629 * try to allocate from this device during
1630 * this iteration around the rotor.
1631 * For example, if a device is 80% full
1632 * and the pool is 20% full then we should
1633 * reduce allocations by 60% on this device.
1634 *
1635 * mg_bias = (20 - 80) * 512K / 100 = -307K
1636 *
1637 * This reduces allocations by 307K for this
1638 * iteration.
1639 */
1640 mg->mg_bias = ((cu - vu) *
1641 (int64_t)mg->mg_aliquot) / 100;
1642 }
1643
1644 if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
1645 mg->mg_aliquot + mg->mg_bias) {
1646 mc->mc_rotor = mg->mg_next;
1647 mc->mc_aliquot = 0;
1648 }
1649
1650 DVA_SET_VDEV(&dva[d], vd->vdev_id);
1651 DVA_SET_OFFSET(&dva[d], offset);
1652 DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
1653 DVA_SET_ASIZE(&dva[d], asize);
1654
1655 return (0);
1656 }
1657 next:
1658 mc->mc_rotor = mg->mg_next;
1659 mc->mc_aliquot = 0;
1660 } while ((mg = mg->mg_next) != rotor);
1661
1662 if (!all_zero) {
1663 dshift++;
1664 ASSERT(dshift < 64);
1665 goto top;
1666 }
1667
1668 if (!allocatable && !zio_lock) {
1669 dshift = 3;
1670 zio_lock = B_TRUE;
1671 goto top;
1672 }
1673
1674 bzero(&dva[d], sizeof (dva_t));
1675
1676 return (SET_ERROR(ENOSPC));
1677 }
1678
1679 /*
1680 * Free the block represented by DVA in the context of the specified
1681 * transaction group.
1682 */
1683 static void
1684 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
1685 {
1686 uint64_t vdev = DVA_GET_VDEV(dva);
1687 uint64_t offset = DVA_GET_OFFSET(dva);
1688 uint64_t size = DVA_GET_ASIZE(dva);
1689 vdev_t *vd;
1690 metaslab_t *msp;
1691
1692 ASSERT(DVA_IS_VALID(dva));
1693
1694 if (txg > spa_freeze_txg(spa))
1695 return;
1696
1697 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
1698 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
1699 cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
1700 (u_longlong_t)vdev, (u_longlong_t)offset);
1701 ASSERT(0);
1702 return;
1703 }
1704
1705 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1706
1707 if (DVA_GET_GANG(dva))
1708 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1709
1710 mutex_enter(&msp->ms_lock);
1711
1712 if (now) {
1713 space_map_remove(msp->ms_allocmap[txg & TXG_MASK],
1714 offset, size);
1715 space_map_free(msp->ms_map, offset, size);
1716 } else {
1717 if (msp->ms_freemap[txg & TXG_MASK]->sm_space == 0)
1718 vdev_dirty(vd, VDD_METASLAB, msp, txg);
1719 space_map_add(msp->ms_freemap[txg & TXG_MASK], offset, size);
1720 }
1721
1722 mutex_exit(&msp->ms_lock);
1723 }
1724
1725 /*
1726 * Intent log support: upon opening the pool after a crash, notify the SPA
1727 * of blocks that the intent log has allocated for immediate write, but
1728 * which are still considered free by the SPA because the last transaction
1729 * group didn't commit yet.
1730 */
1731 static int
1732 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
1733 {
1734 uint64_t vdev = DVA_GET_VDEV(dva);
1735 uint64_t offset = DVA_GET_OFFSET(dva);
1736 uint64_t size = DVA_GET_ASIZE(dva);
1737 vdev_t *vd;
1738 metaslab_t *msp;
1739 int error = 0;
1740
1741 ASSERT(DVA_IS_VALID(dva));
1742
1743 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
1744 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
1745 return (SET_ERROR(ENXIO));
1746
1747 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1748
1749 if (DVA_GET_GANG(dva))
1750 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1751
1752 mutex_enter(&msp->ms_lock);
1753
1754 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_map->sm_loaded)
1755 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
1756
1757 if (error == 0 && !space_map_contains(msp->ms_map, offset, size))
1758 error = SET_ERROR(ENOENT);
1759
1760 if (error || txg == 0) { /* txg == 0 indicates dry run */
1761 mutex_exit(&msp->ms_lock);
1762 return (error);
1763 }
1764
1765 space_map_claim(msp->ms_map, offset, size);
1766
1767 if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
1768 if (msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0)
1769 vdev_dirty(vd, VDD_METASLAB, msp, txg);
1770 space_map_add(msp->ms_allocmap[txg & TXG_MASK], offset, size);
1771 }
1772
1773 mutex_exit(&msp->ms_lock);
1774
1775 return (0);
1776 }
1777
1778 int
1779 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
1780 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
1781 {
1782 dva_t *dva = bp->blk_dva;
1783 dva_t *hintdva = hintbp->blk_dva;
1784 int error = 0;
1785
1786 ASSERT(bp->blk_birth == 0);
1787 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
1788
1789 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
1790
1791 if (mc->mc_rotor == NULL) { /* no vdevs in this class */
1792 spa_config_exit(spa, SCL_ALLOC, FTAG);
1793 return (SET_ERROR(ENOSPC));
1794 }
1795
1796 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
1797 ASSERT(BP_GET_NDVAS(bp) == 0);
1798 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
1799
1800 for (int d = 0; d < ndvas; d++) {
1801 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
1802 txg, flags);
1803 if (error) {
1804 for (d--; d >= 0; d--) {
1805 metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
1806 bzero(&dva[d], sizeof (dva_t));
1807 }
1808 spa_config_exit(spa, SCL_ALLOC, FTAG);
1809 return (error);
1810 }
1811 }
1812 ASSERT(error == 0);
1813 ASSERT(BP_GET_NDVAS(bp) == ndvas);
1814
1815 spa_config_exit(spa, SCL_ALLOC, FTAG);
1816
1817 BP_SET_BIRTH(bp, txg, txg);
1818
1819 return (0);
1820 }
1821
1822 void
1823 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
1824 {
1825 const dva_t *dva = bp->blk_dva;
1826 int ndvas = BP_GET_NDVAS(bp);
1827
1828 ASSERT(!BP_IS_HOLE(bp));
1829 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
1830
1831 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
1832
1833 for (int d = 0; d < ndvas; d++)
1834 metaslab_free_dva(spa, &dva[d], txg, now);
1835
1836 spa_config_exit(spa, SCL_FREE, FTAG);
1837 }
1838
1839 int
1840 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
1841 {
1842 const dva_t *dva = bp->blk_dva;
1843 int ndvas = BP_GET_NDVAS(bp);
1844 int error = 0;
1845
1846 ASSERT(!BP_IS_HOLE(bp));
1847
1848 if (txg != 0) {
1849 /*
1850 * First do a dry run to make sure all DVAs are claimable,
1851 * so we don't have to unwind from partial failures below.
1852 */
1853 if ((error = metaslab_claim(spa, bp, 0)) != 0)
1854 return (error);
1855 }
1856
1857 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
1858
1859 for (int d = 0; d < ndvas; d++)
1860 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
1861 break;
1862
1863 spa_config_exit(spa, SCL_ALLOC, FTAG);
1864
1865 ASSERT(error == 0 || txg == 0);
1866
1867 return (error);
1868 }
1869
1870 static void
1871 checkmap(space_map_t *sm, uint64_t off, uint64_t size)
1872 {
1873 space_seg_t *ss;
1874 avl_index_t where;
1875
1876 mutex_enter(sm->sm_lock);
1877 ss = space_map_find(sm, off, size, &where);
1878 if (ss != NULL)
1879 panic("freeing free block; ss=%p", (void *)ss);
1880 mutex_exit(sm->sm_lock);
1881 }
1882
1883 void
1884 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
1885 {
1886 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
1887 return;
1888
1889 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
1890 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
1891 uint64_t vdid = DVA_GET_VDEV(&bp->blk_dva[i]);
1892 vdev_t *vd = vdev_lookup_top(spa, vdid);
1893 uint64_t off = DVA_GET_OFFSET(&bp->blk_dva[i]);
1894 uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
1895 metaslab_t *ms = vd->vdev_ms[off >> vd->vdev_ms_shift];
1896
1897 if (ms->ms_map->sm_loaded)
1898 checkmap(ms->ms_map, off, size);
1899
1900 for (int j = 0; j < TXG_SIZE; j++)
1901 checkmap(ms->ms_freemap[j], off, size);
1902 for (int j = 0; j < TXG_DEFER_SIZE; j++)
1903 checkmap(ms->ms_defermap[j], off, size);
1904 }
1905 spa_config_exit(spa, SCL_VDEV, FTAG);
1906 }