1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012 by Delphix. All rights reserved.
24 */
25
26 #include <sys/zfs_context.h>
27 #include <sys/dmu.h>
28 #include <sys/dmu_tx.h>
29 #include <sys/space_map.h>
30 #include <sys/metaslab_impl.h>
31 #include <sys/vdev_impl.h>
32 #include <sys/zio.h>
33
34 /*
35 * Allow allocations to switch to gang blocks quickly. We do this to
36 * avoid having to load lots of space_maps in a given txg. There are,
37 * however, some cases where we want to avoid "fast" ganging and instead
38 * we want to do an exhaustive search of all metaslabs on this device.
39 * Currently we don't allow any gang, zil, or dump device related allocations
40 * to "fast" gang.
41 */
42 #define CAN_FASTGANG(flags) \
43 (!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \
44 METASLAB_GANG_AVOID)))
45
46 uint64_t metaslab_aliquot = 512ULL << 10;
47 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1; /* force gang blocks */
48
49 /*
50 * This value defines the number of allowed allocation failures per vdev.
51 * If a device reaches this threshold in a given txg then we consider skipping
52 * allocations on that device.
53 */
54 int zfs_mg_alloc_failures;
55
56 /*
57 * Metaslab debugging: when set, keeps all space maps in core to verify frees.
58 */
59 static int metaslab_debug = 0;
60
61 /*
62 * Minimum size which forces the dynamic allocator to change
63 * it's allocation strategy. Once the space map cannot satisfy
64 * an allocation of this size then it switches to using more
65 * aggressive strategy (i.e search by size rather than offset).
66 */
67 uint64_t metaslab_df_alloc_threshold = SPA_MAXBLOCKSIZE;
68
69 /*
70 * The minimum free space, in percent, which must be available
71 * in a space map to continue allocations in a first-fit fashion.
72 * Once the space_map's free space drops below this level we dynamically
73 * switch to using best-fit allocations.
74 */
75 int metaslab_df_free_pct = 4;
76
77 /*
78 * A metaslab is considered "free" if it contains a contiguous
79 * segment which is greater than metaslab_min_alloc_size.
80 */
81 uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
82
83 /*
84 * Max number of space_maps to prefetch.
85 */
86 int metaslab_prefetch_limit = SPA_DVAS_PER_BP;
87
88 /*
89 * Percentage bonus multiplier for metaslabs that are in the bonus area.
90 */
91 int metaslab_smo_bonus_pct = 150;
92
93 /*
94 * ==========================================================================
95 * Metaslab classes
96 * ==========================================================================
97 */
98 metaslab_class_t *
99 metaslab_class_create(spa_t *spa, space_map_ops_t *ops)
100 {
101 metaslab_class_t *mc;
102
103 mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
104
105 mc->mc_spa = spa;
106 mc->mc_rotor = NULL;
107 mc->mc_ops = ops;
108
109 return (mc);
110 }
111
112 void
113 metaslab_class_destroy(metaslab_class_t *mc)
114 {
115 ASSERT(mc->mc_rotor == NULL);
116 ASSERT(mc->mc_alloc == 0);
117 ASSERT(mc->mc_deferred == 0);
118 ASSERT(mc->mc_space == 0);
119 ASSERT(mc->mc_dspace == 0);
120
121 kmem_free(mc, sizeof (metaslab_class_t));
122 }
123
124 int
125 metaslab_class_validate(metaslab_class_t *mc)
126 {
127 metaslab_group_t *mg;
128 vdev_t *vd;
129
130 /*
131 * Must hold one of the spa_config locks.
132 */
133 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
134 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
135
136 if ((mg = mc->mc_rotor) == NULL)
137 return (0);
138
139 do {
140 vd = mg->mg_vd;
141 ASSERT(vd->vdev_mg != NULL);
142 ASSERT3P(vd->vdev_top, ==, vd);
143 ASSERT3P(mg->mg_class, ==, mc);
144 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
145 } while ((mg = mg->mg_next) != mc->mc_rotor);
146
147 return (0);
148 }
149
150 void
151 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
152 int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
153 {
154 atomic_add_64(&mc->mc_alloc, alloc_delta);
155 atomic_add_64(&mc->mc_deferred, defer_delta);
156 atomic_add_64(&mc->mc_space, space_delta);
157 atomic_add_64(&mc->mc_dspace, dspace_delta);
158 }
159
160 uint64_t
161 metaslab_class_get_alloc(metaslab_class_t *mc)
162 {
163 return (mc->mc_alloc);
164 }
165
166 uint64_t
167 metaslab_class_get_deferred(metaslab_class_t *mc)
168 {
169 return (mc->mc_deferred);
170 }
171
172 uint64_t
173 metaslab_class_get_space(metaslab_class_t *mc)
174 {
175 return (mc->mc_space);
176 }
177
178 uint64_t
179 metaslab_class_get_dspace(metaslab_class_t *mc)
180 {
181 return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
182 }
183
184 /*
185 * ==========================================================================
186 * Metaslab groups
187 * ==========================================================================
188 */
189 static int
190 metaslab_compare(const void *x1, const void *x2)
191 {
192 const metaslab_t *m1 = x1;
193 const metaslab_t *m2 = x2;
194
195 if (m1->ms_weight < m2->ms_weight)
196 return (1);
197 if (m1->ms_weight > m2->ms_weight)
198 return (-1);
199
200 /*
201 * If the weights are identical, use the offset to force uniqueness.
202 */
203 if (m1->ms_map.sm_start < m2->ms_map.sm_start)
204 return (-1);
205 if (m1->ms_map.sm_start > m2->ms_map.sm_start)
206 return (1);
207
208 ASSERT3P(m1, ==, m2);
209
210 return (0);
211 }
212
213 metaslab_group_t *
214 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
215 {
216 metaslab_group_t *mg;
217
218 mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
219 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
220 avl_create(&mg->mg_metaslab_tree, metaslab_compare,
221 sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
222 mg->mg_vd = vd;
223 mg->mg_class = mc;
224 mg->mg_activation_count = 0;
225
226 return (mg);
227 }
228
229 void
230 metaslab_group_destroy(metaslab_group_t *mg)
231 {
232 ASSERT(mg->mg_prev == NULL);
233 ASSERT(mg->mg_next == NULL);
234 /*
235 * We may have gone below zero with the activation count
236 * either because we never activated in the first place or
237 * because we're done, and possibly removing the vdev.
238 */
239 ASSERT(mg->mg_activation_count <= 0);
240
241 avl_destroy(&mg->mg_metaslab_tree);
242 mutex_destroy(&mg->mg_lock);
243 kmem_free(mg, sizeof (metaslab_group_t));
244 }
245
246 void
247 metaslab_group_activate(metaslab_group_t *mg)
248 {
249 metaslab_class_t *mc = mg->mg_class;
250 metaslab_group_t *mgprev, *mgnext;
251
252 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
253
254 ASSERT(mc->mc_rotor != mg);
255 ASSERT(mg->mg_prev == NULL);
256 ASSERT(mg->mg_next == NULL);
257 ASSERT(mg->mg_activation_count <= 0);
258
259 if (++mg->mg_activation_count <= 0)
260 return;
261
262 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
263
264 if ((mgprev = mc->mc_rotor) == NULL) {
265 mg->mg_prev = mg;
266 mg->mg_next = mg;
267 } else {
268 mgnext = mgprev->mg_next;
269 mg->mg_prev = mgprev;
270 mg->mg_next = mgnext;
271 mgprev->mg_next = mg;
272 mgnext->mg_prev = mg;
273 }
274 mc->mc_rotor = mg;
275 }
276
277 void
278 metaslab_group_passivate(metaslab_group_t *mg)
279 {
280 metaslab_class_t *mc = mg->mg_class;
281 metaslab_group_t *mgprev, *mgnext;
282
283 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
284
285 if (--mg->mg_activation_count != 0) {
286 ASSERT(mc->mc_rotor != mg);
287 ASSERT(mg->mg_prev == NULL);
288 ASSERT(mg->mg_next == NULL);
289 ASSERT(mg->mg_activation_count < 0);
290 return;
291 }
292
293 mgprev = mg->mg_prev;
294 mgnext = mg->mg_next;
295
296 if (mg == mgnext) {
297 mc->mc_rotor = NULL;
298 } else {
299 mc->mc_rotor = mgnext;
300 mgprev->mg_next = mgnext;
301 mgnext->mg_prev = mgprev;
302 }
303
304 mg->mg_prev = NULL;
305 mg->mg_next = NULL;
306 }
307
308 static void
309 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
310 {
311 mutex_enter(&mg->mg_lock);
312 ASSERT(msp->ms_group == NULL);
313 msp->ms_group = mg;
314 msp->ms_weight = 0;
315 avl_add(&mg->mg_metaslab_tree, msp);
316 mutex_exit(&mg->mg_lock);
317 }
318
319 static void
320 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
321 {
322 mutex_enter(&mg->mg_lock);
323 ASSERT(msp->ms_group == mg);
324 avl_remove(&mg->mg_metaslab_tree, msp);
325 msp->ms_group = NULL;
326 mutex_exit(&mg->mg_lock);
327 }
328
329 static void
330 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
331 {
332 /*
333 * Although in principle the weight can be any value, in
334 * practice we do not use values in the range [1, 510].
335 */
336 ASSERT(weight >= SPA_MINBLOCKSIZE-1 || weight == 0);
337 ASSERT(MUTEX_HELD(&msp->ms_lock));
338
339 mutex_enter(&mg->mg_lock);
340 ASSERT(msp->ms_group == mg);
341 avl_remove(&mg->mg_metaslab_tree, msp);
342 msp->ms_weight = weight;
343 avl_add(&mg->mg_metaslab_tree, msp);
344 mutex_exit(&mg->mg_lock);
345 }
346
347 /*
348 * ==========================================================================
349 * Common allocator routines
350 * ==========================================================================
351 */
352 static int
353 metaslab_segsize_compare(const void *x1, const void *x2)
354 {
355 const space_seg_t *s1 = x1;
356 const space_seg_t *s2 = x2;
357 uint64_t ss_size1 = s1->ss_end - s1->ss_start;
358 uint64_t ss_size2 = s2->ss_end - s2->ss_start;
359
360 if (ss_size1 < ss_size2)
361 return (-1);
362 if (ss_size1 > ss_size2)
363 return (1);
364
365 if (s1->ss_start < s2->ss_start)
366 return (-1);
367 if (s1->ss_start > s2->ss_start)
368 return (1);
369
370 return (0);
371 }
372
373 /*
374 * This is a helper function that can be used by the allocator to find
375 * a suitable block to allocate. This will search the specified AVL
376 * tree looking for a block that matches the specified criteria.
377 */
378 static uint64_t
379 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
380 uint64_t align)
381 {
382 space_seg_t *ss, ssearch;
383 avl_index_t where;
384
385 ssearch.ss_start = *cursor;
386 ssearch.ss_end = *cursor + size;
387
388 ss = avl_find(t, &ssearch, &where);
389 if (ss == NULL)
390 ss = avl_nearest(t, where, AVL_AFTER);
391
392 while (ss != NULL) {
393 uint64_t offset = P2ROUNDUP(ss->ss_start, align);
394
395 if (offset + size <= ss->ss_end) {
396 *cursor = offset + size;
397 return (offset);
398 }
399 ss = AVL_NEXT(t, ss);
400 }
401
402 /*
403 * If we know we've searched the whole map (*cursor == 0), give up.
404 * Otherwise, reset the cursor to the beginning and try again.
405 */
406 if (*cursor == 0)
407 return (-1ULL);
408
409 *cursor = 0;
410 return (metaslab_block_picker(t, cursor, size, align));
411 }
412
413 static void
414 metaslab_pp_load(space_map_t *sm)
415 {
416 space_seg_t *ss;
417
418 ASSERT(sm->sm_ppd == NULL);
419 sm->sm_ppd = kmem_zalloc(64 * sizeof (uint64_t), KM_SLEEP);
420
421 sm->sm_pp_root = kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
422 avl_create(sm->sm_pp_root, metaslab_segsize_compare,
423 sizeof (space_seg_t), offsetof(struct space_seg, ss_pp_node));
424
425 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
426 avl_add(sm->sm_pp_root, ss);
427 }
428
429 static void
430 metaslab_pp_unload(space_map_t *sm)
431 {
432 void *cookie = NULL;
433
434 kmem_free(sm->sm_ppd, 64 * sizeof (uint64_t));
435 sm->sm_ppd = NULL;
436
437 while (avl_destroy_nodes(sm->sm_pp_root, &cookie) != NULL) {
438 /* tear down the tree */
439 }
440
441 avl_destroy(sm->sm_pp_root);
442 kmem_free(sm->sm_pp_root, sizeof (avl_tree_t));
443 sm->sm_pp_root = NULL;
444 }
445
446 /* ARGSUSED */
447 static void
448 metaslab_pp_claim(space_map_t *sm, uint64_t start, uint64_t size)
449 {
450 /* No need to update cursor */
451 }
452
453 /* ARGSUSED */
454 static void
455 metaslab_pp_free(space_map_t *sm, uint64_t start, uint64_t size)
456 {
457 /* No need to update cursor */
458 }
459
460 /*
461 * Return the maximum contiguous segment within the metaslab.
462 */
463 uint64_t
464 metaslab_pp_maxsize(space_map_t *sm)
465 {
466 avl_tree_t *t = sm->sm_pp_root;
467 space_seg_t *ss;
468
469 if (t == NULL || (ss = avl_last(t)) == NULL)
470 return (0ULL);
471
472 return (ss->ss_end - ss->ss_start);
473 }
474
475 /*
476 * ==========================================================================
477 * The first-fit block allocator
478 * ==========================================================================
479 */
480 static uint64_t
481 metaslab_ff_alloc(space_map_t *sm, uint64_t size)
482 {
483 avl_tree_t *t = &sm->sm_root;
484 uint64_t align = size & -size;
485 uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
486
487 return (metaslab_block_picker(t, cursor, size, align));
488 }
489
490 /* ARGSUSED */
491 boolean_t
492 metaslab_ff_fragmented(space_map_t *sm)
493 {
494 return (B_TRUE);
495 }
496
497 static space_map_ops_t metaslab_ff_ops = {
498 metaslab_pp_load,
499 metaslab_pp_unload,
500 metaslab_ff_alloc,
501 metaslab_pp_claim,
502 metaslab_pp_free,
503 metaslab_pp_maxsize,
504 metaslab_ff_fragmented
505 };
506
507 /*
508 * ==========================================================================
509 * Dynamic block allocator -
510 * Uses the first fit allocation scheme until space get low and then
511 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
512 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
513 * ==========================================================================
514 */
515 static uint64_t
516 metaslab_df_alloc(space_map_t *sm, uint64_t size)
517 {
518 avl_tree_t *t = &sm->sm_root;
519 uint64_t align = size & -size;
520 uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
521 uint64_t max_size = metaslab_pp_maxsize(sm);
522 int free_pct = sm->sm_space * 100 / sm->sm_size;
523
524 ASSERT(MUTEX_HELD(sm->sm_lock));
525 ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
526
527 if (max_size < size)
528 return (-1ULL);
529
530 /*
531 * If we're running low on space switch to using the size
532 * sorted AVL tree (best-fit).
533 */
534 if (max_size < metaslab_df_alloc_threshold ||
535 free_pct < metaslab_df_free_pct) {
536 t = sm->sm_pp_root;
537 *cursor = 0;
538 }
539
540 return (metaslab_block_picker(t, cursor, size, 1ULL));
541 }
542
543 static boolean_t
544 metaslab_df_fragmented(space_map_t *sm)
545 {
546 uint64_t max_size = metaslab_pp_maxsize(sm);
547 int free_pct = sm->sm_space * 100 / sm->sm_size;
548
549 if (max_size >= metaslab_df_alloc_threshold &&
550 free_pct >= metaslab_df_free_pct)
551 return (B_FALSE);
552
553 return (B_TRUE);
554 }
555
556 static space_map_ops_t metaslab_df_ops = {
557 metaslab_pp_load,
558 metaslab_pp_unload,
559 metaslab_df_alloc,
560 metaslab_pp_claim,
561 metaslab_pp_free,
562 metaslab_pp_maxsize,
563 metaslab_df_fragmented
564 };
565
566 /*
567 * ==========================================================================
568 * Other experimental allocators
569 * ==========================================================================
570 */
571 static uint64_t
572 metaslab_cdf_alloc(space_map_t *sm, uint64_t size)
573 {
574 avl_tree_t *t = &sm->sm_root;
575 uint64_t *cursor = (uint64_t *)sm->sm_ppd;
576 uint64_t *extent_end = (uint64_t *)sm->sm_ppd + 1;
577 uint64_t max_size = metaslab_pp_maxsize(sm);
578 uint64_t rsize = size;
579 uint64_t offset = 0;
580
581 ASSERT(MUTEX_HELD(sm->sm_lock));
582 ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
583
584 if (max_size < size)
585 return (-1ULL);
586
587 ASSERT3U(*extent_end, >=, *cursor);
588
589 /*
590 * If we're running low on space switch to using the size
591 * sorted AVL tree (best-fit).
592 */
593 if ((*cursor + size) > *extent_end) {
594
595 t = sm->sm_pp_root;
596 *cursor = *extent_end = 0;
597
598 if (max_size > 2 * SPA_MAXBLOCKSIZE)
599 rsize = MIN(metaslab_min_alloc_size, max_size);
600 offset = metaslab_block_picker(t, extent_end, rsize, 1ULL);
601 if (offset != -1)
602 *cursor = offset + size;
603 } else {
604 offset = metaslab_block_picker(t, cursor, rsize, 1ULL);
605 }
606 ASSERT3U(*cursor, <=, *extent_end);
607 return (offset);
608 }
609
610 static boolean_t
611 metaslab_cdf_fragmented(space_map_t *sm)
612 {
613 uint64_t max_size = metaslab_pp_maxsize(sm);
614
615 if (max_size > (metaslab_min_alloc_size * 10))
616 return (B_FALSE);
617 return (B_TRUE);
618 }
619
620 static space_map_ops_t metaslab_cdf_ops = {
621 metaslab_pp_load,
622 metaslab_pp_unload,
623 metaslab_cdf_alloc,
624 metaslab_pp_claim,
625 metaslab_pp_free,
626 metaslab_pp_maxsize,
627 metaslab_cdf_fragmented
628 };
629
630 uint64_t metaslab_ndf_clump_shift = 4;
631
632 static uint64_t
633 metaslab_ndf_alloc(space_map_t *sm, uint64_t size)
634 {
635 avl_tree_t *t = &sm->sm_root;
636 avl_index_t where;
637 space_seg_t *ss, ssearch;
638 uint64_t hbit = highbit(size);
639 uint64_t *cursor = (uint64_t *)sm->sm_ppd + hbit - 1;
640 uint64_t max_size = metaslab_pp_maxsize(sm);
641
642 ASSERT(MUTEX_HELD(sm->sm_lock));
643 ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
644
645 if (max_size < size)
646 return (-1ULL);
647
648 ssearch.ss_start = *cursor;
649 ssearch.ss_end = *cursor + size;
650
651 ss = avl_find(t, &ssearch, &where);
652 if (ss == NULL || (ss->ss_start + size > ss->ss_end)) {
653 t = sm->sm_pp_root;
654
655 ssearch.ss_start = 0;
656 ssearch.ss_end = MIN(max_size,
657 1ULL << (hbit + metaslab_ndf_clump_shift));
658 ss = avl_find(t, &ssearch, &where);
659 if (ss == NULL)
660 ss = avl_nearest(t, where, AVL_AFTER);
661 ASSERT(ss != NULL);
662 }
663
664 if (ss != NULL) {
665 if (ss->ss_start + size <= ss->ss_end) {
666 *cursor = ss->ss_start + size;
667 return (ss->ss_start);
668 }
669 }
670 return (-1ULL);
671 }
672
673 static boolean_t
674 metaslab_ndf_fragmented(space_map_t *sm)
675 {
676 uint64_t max_size = metaslab_pp_maxsize(sm);
677
678 if (max_size > (metaslab_min_alloc_size << metaslab_ndf_clump_shift))
679 return (B_FALSE);
680 return (B_TRUE);
681 }
682
683
684 static space_map_ops_t metaslab_ndf_ops = {
685 metaslab_pp_load,
686 metaslab_pp_unload,
687 metaslab_ndf_alloc,
688 metaslab_pp_claim,
689 metaslab_pp_free,
690 metaslab_pp_maxsize,
691 metaslab_ndf_fragmented
692 };
693
694 space_map_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
695
696 /*
697 * ==========================================================================
698 * Metaslabs
699 * ==========================================================================
700 */
701 metaslab_t *
702 metaslab_init(metaslab_group_t *mg, space_map_obj_t *smo,
703 uint64_t start, uint64_t size, uint64_t txg)
704 {
705 vdev_t *vd = mg->mg_vd;
706 metaslab_t *msp;
707
708 msp = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
709 mutex_init(&msp->ms_lock, NULL, MUTEX_DEFAULT, NULL);
710
711 msp->ms_smo_syncing = *smo;
712
713 /*
714 * We create the main space map here, but we don't create the
715 * allocmaps and freemaps until metaslab_sync_done(). This serves
716 * two purposes: it allows metaslab_sync_done() to detect the
717 * addition of new space; and for debugging, it ensures that we'd
718 * data fault on any attempt to use this metaslab before it's ready.
719 */
720 space_map_create(&msp->ms_map, start, size,
721 vd->vdev_ashift, &msp->ms_lock);
722
723 metaslab_group_add(mg, msp);
724
725 if (metaslab_debug && smo->smo_object != 0) {
726 mutex_enter(&msp->ms_lock);
727 VERIFY(space_map_load(&msp->ms_map, mg->mg_class->mc_ops,
728 SM_FREE, smo, spa_meta_objset(vd->vdev_spa)) == 0);
729 mutex_exit(&msp->ms_lock);
730 }
731
732 /*
733 * If we're opening an existing pool (txg == 0) or creating
734 * a new one (txg == TXG_INITIAL), all space is available now.
735 * If we're adding space to an existing pool, the new space
736 * does not become available until after this txg has synced.
737 */
738 if (txg <= TXG_INITIAL)
739 metaslab_sync_done(msp, 0);
740
741 if (txg != 0) {
742 vdev_dirty(vd, 0, NULL, txg);
743 vdev_dirty(vd, VDD_METASLAB, msp, txg);
744 }
745
746 return (msp);
747 }
748
749 void
750 metaslab_fini(metaslab_t *msp)
751 {
752 metaslab_group_t *mg = msp->ms_group;
753
754 vdev_space_update(mg->mg_vd,
755 -msp->ms_smo.smo_alloc, 0, -msp->ms_map.sm_size);
756
757 metaslab_group_remove(mg, msp);
758
759 mutex_enter(&msp->ms_lock);
760
761 space_map_unload(&msp->ms_map);
762 space_map_destroy(&msp->ms_map);
763
764 for (int t = 0; t < TXG_SIZE; t++) {
765 space_map_destroy(&msp->ms_allocmap[t]);
766 space_map_destroy(&msp->ms_freemap[t]);
767 }
768
769 for (int t = 0; t < TXG_DEFER_SIZE; t++)
770 space_map_destroy(&msp->ms_defermap[t]);
771
772 ASSERT3S(msp->ms_deferspace, ==, 0);
773
774 mutex_exit(&msp->ms_lock);
775 mutex_destroy(&msp->ms_lock);
776
777 kmem_free(msp, sizeof (metaslab_t));
778 }
779
780 #define METASLAB_WEIGHT_PRIMARY (1ULL << 63)
781 #define METASLAB_WEIGHT_SECONDARY (1ULL << 62)
782 #define METASLAB_ACTIVE_MASK \
783 (METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
784
785 static uint64_t
786 metaslab_weight(metaslab_t *msp)
787 {
788 metaslab_group_t *mg = msp->ms_group;
789 space_map_t *sm = &msp->ms_map;
790 space_map_obj_t *smo = &msp->ms_smo;
791 vdev_t *vd = mg->mg_vd;
792 uint64_t weight, space;
793
794 ASSERT(MUTEX_HELD(&msp->ms_lock));
795
796 /*
797 * The baseline weight is the metaslab's free space.
798 */
799 space = sm->sm_size - smo->smo_alloc;
800 weight = space;
801
802 /*
803 * Modern disks have uniform bit density and constant angular velocity.
804 * Therefore, the outer recording zones are faster (higher bandwidth)
805 * than the inner zones by the ratio of outer to inner track diameter,
806 * which is typically around 2:1. We account for this by assigning
807 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
808 * In effect, this means that we'll select the metaslab with the most
809 * free bandwidth rather than simply the one with the most free space.
810 */
811 weight = 2 * weight -
812 ((sm->sm_start >> vd->vdev_ms_shift) * weight) / vd->vdev_ms_count;
813 ASSERT(weight >= space && weight <= 2 * space);
814
815 /*
816 * For locality, assign higher weight to metaslabs which have
817 * a lower offset than what we've already activated.
818 */
819 if (sm->sm_start <= mg->mg_bonus_area)
820 weight *= (metaslab_smo_bonus_pct / 100);
821 ASSERT(weight >= space &&
822 weight <= 2 * (metaslab_smo_bonus_pct / 100) * space);
823
824 if (sm->sm_loaded && !sm->sm_ops->smop_fragmented(sm)) {
825 /*
826 * If this metaslab is one we're actively using, adjust its
827 * weight to make it preferable to any inactive metaslab so
828 * we'll polish it off.
829 */
830 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
831 }
832 return (weight);
833 }
834
835 static void
836 metaslab_prefetch(metaslab_group_t *mg)
837 {
838 spa_t *spa = mg->mg_vd->vdev_spa;
839 metaslab_t *msp;
840 avl_tree_t *t = &mg->mg_metaslab_tree;
841 int m;
842
843 mutex_enter(&mg->mg_lock);
844
845 /*
846 * Prefetch the next potential metaslabs
847 */
848 for (msp = avl_first(t), m = 0; msp; msp = AVL_NEXT(t, msp), m++) {
849 space_map_t *sm = &msp->ms_map;
850 space_map_obj_t *smo = &msp->ms_smo;
851
852 /* If we have reached our prefetch limit then we're done */
853 if (m >= metaslab_prefetch_limit)
854 break;
855
856 if (!sm->sm_loaded && smo->smo_object != 0) {
857 mutex_exit(&mg->mg_lock);
858 dmu_prefetch(spa_meta_objset(spa), smo->smo_object,
859 0ULL, smo->smo_objsize);
860 mutex_enter(&mg->mg_lock);
861 }
862 }
863 mutex_exit(&mg->mg_lock);
864 }
865
866 static int
867 metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
868 {
869 metaslab_group_t *mg = msp->ms_group;
870 space_map_t *sm = &msp->ms_map;
871 space_map_ops_t *sm_ops = msp->ms_group->mg_class->mc_ops;
872
873 ASSERT(MUTEX_HELD(&msp->ms_lock));
874
875 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
876 space_map_load_wait(sm);
877 if (!sm->sm_loaded) {
878 int error = space_map_load(sm, sm_ops, SM_FREE,
879 &msp->ms_smo,
880 spa_meta_objset(msp->ms_group->mg_vd->vdev_spa));
881 if (error) {
882 metaslab_group_sort(msp->ms_group, msp, 0);
883 return (error);
884 }
885 for (int t = 0; t < TXG_DEFER_SIZE; t++)
886 space_map_walk(&msp->ms_defermap[t],
887 space_map_claim, sm);
888
889 }
890
891 /*
892 * Track the bonus area as we activate new metaslabs.
893 */
894 if (sm->sm_start > mg->mg_bonus_area) {
895 mutex_enter(&mg->mg_lock);
896 mg->mg_bonus_area = sm->sm_start;
897 mutex_exit(&mg->mg_lock);
898 }
899
900 metaslab_group_sort(msp->ms_group, msp,
901 msp->ms_weight | activation_weight);
902 }
903 ASSERT(sm->sm_loaded);
904 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
905
906 return (0);
907 }
908
909 static void
910 metaslab_passivate(metaslab_t *msp, uint64_t size)
911 {
912 /*
913 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
914 * this metaslab again. In that case, it had better be empty,
915 * or we would be leaving space on the table.
916 */
917 ASSERT(size >= SPA_MINBLOCKSIZE || msp->ms_map.sm_space == 0);
918 metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
919 ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
920 }
921
922 /*
923 * Write a metaslab to disk in the context of the specified transaction group.
924 */
925 void
926 metaslab_sync(metaslab_t *msp, uint64_t txg)
927 {
928 vdev_t *vd = msp->ms_group->mg_vd;
929 spa_t *spa = vd->vdev_spa;
930 objset_t *mos = spa_meta_objset(spa);
931 space_map_t *allocmap = &msp->ms_allocmap[txg & TXG_MASK];
932 space_map_t *freemap = &msp->ms_freemap[txg & TXG_MASK];
933 space_map_t *freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
934 space_map_t *sm = &msp->ms_map;
935 space_map_obj_t *smo = &msp->ms_smo_syncing;
936 dmu_buf_t *db;
937 dmu_tx_t *tx;
938
939 ASSERT(!vd->vdev_ishole);
940
941 if (allocmap->sm_space == 0 && freemap->sm_space == 0)
942 return;
943
944 /*
945 * The only state that can actually be changing concurrently with
946 * metaslab_sync() is the metaslab's ms_map. No other thread can
947 * be modifying this txg's allocmap, freemap, freed_map, or smo.
948 * Therefore, we only hold ms_lock to satify space_map ASSERTs.
949 * We drop it whenever we call into the DMU, because the DMU
950 * can call down to us (e.g. via zio_free()) at any time.
951 */
952
953 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
954
955 if (smo->smo_object == 0) {
956 ASSERT(smo->smo_objsize == 0);
957 ASSERT(smo->smo_alloc == 0);
958 smo->smo_object = dmu_object_alloc(mos,
959 DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
960 DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
961 ASSERT(smo->smo_object != 0);
962 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
963 (sm->sm_start >> vd->vdev_ms_shift),
964 sizeof (uint64_t), &smo->smo_object, tx);
965 }
966
967 mutex_enter(&msp->ms_lock);
968
969 space_map_walk(freemap, space_map_add, freed_map);
970
971 if (sm->sm_loaded && spa_sync_pass(spa) == 1 && smo->smo_objsize >=
972 2 * sizeof (uint64_t) * avl_numnodes(&sm->sm_root)) {
973 /*
974 * The in-core space map representation is twice as compact
975 * as the on-disk one, so it's time to condense the latter
976 * by generating a pure allocmap from first principles.
977 *
978 * This metaslab is 100% allocated,
979 * minus the content of the in-core map (sm),
980 * minus what's been freed this txg (freed_map),
981 * minus deferred frees (ms_defermap[]),
982 * minus allocations from txgs in the future
983 * (because they haven't been committed yet).
984 */
985 space_map_vacate(allocmap, NULL, NULL);
986 space_map_vacate(freemap, NULL, NULL);
987
988 space_map_add(allocmap, allocmap->sm_start, allocmap->sm_size);
989
990 space_map_walk(sm, space_map_remove, allocmap);
991 space_map_walk(freed_map, space_map_remove, allocmap);
992
993 for (int t = 0; t < TXG_DEFER_SIZE; t++)
994 space_map_walk(&msp->ms_defermap[t],
995 space_map_remove, allocmap);
996
997 for (int t = 1; t < TXG_CONCURRENT_STATES; t++)
998 space_map_walk(&msp->ms_allocmap[(txg + t) & TXG_MASK],
999 space_map_remove, allocmap);
1000
1001 mutex_exit(&msp->ms_lock);
1002 space_map_truncate(smo, mos, tx);
1003 mutex_enter(&msp->ms_lock);
1004 }
1005
1006 space_map_sync(allocmap, SM_ALLOC, smo, mos, tx);
1007 space_map_sync(freemap, SM_FREE, smo, mos, tx);
1008
1009 mutex_exit(&msp->ms_lock);
1010
1011 VERIFY(0 == dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
1012 dmu_buf_will_dirty(db, tx);
1013 ASSERT3U(db->db_size, >=, sizeof (*smo));
1014 bcopy(smo, db->db_data, sizeof (*smo));
1015 dmu_buf_rele(db, FTAG);
1016
1017 dmu_tx_commit(tx);
1018 }
1019
1020 /*
1021 * Called after a transaction group has completely synced to mark
1022 * all of the metaslab's free space as usable.
1023 */
1024 void
1025 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
1026 {
1027 space_map_obj_t *smo = &msp->ms_smo;
1028 space_map_obj_t *smosync = &msp->ms_smo_syncing;
1029 space_map_t *sm = &msp->ms_map;
1030 space_map_t *freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
1031 space_map_t *defer_map = &msp->ms_defermap[txg % TXG_DEFER_SIZE];
1032 metaslab_group_t *mg = msp->ms_group;
1033 vdev_t *vd = mg->mg_vd;
1034 int64_t alloc_delta, defer_delta;
1035
1036 ASSERT(!vd->vdev_ishole);
1037
1038 mutex_enter(&msp->ms_lock);
1039
1040 /*
1041 * If this metaslab is just becoming available, initialize its
1042 * allocmaps and freemaps and add its capacity to the vdev.
1043 */
1044 if (freed_map->sm_size == 0) {
1045 for (int t = 0; t < TXG_SIZE; t++) {
1046 space_map_create(&msp->ms_allocmap[t], sm->sm_start,
1047 sm->sm_size, sm->sm_shift, sm->sm_lock);
1048 space_map_create(&msp->ms_freemap[t], sm->sm_start,
1049 sm->sm_size, sm->sm_shift, sm->sm_lock);
1050 }
1051
1052 for (int t = 0; t < TXG_DEFER_SIZE; t++)
1053 space_map_create(&msp->ms_defermap[t], sm->sm_start,
1054 sm->sm_size, sm->sm_shift, sm->sm_lock);
1055
1056 vdev_space_update(vd, 0, 0, sm->sm_size);
1057 }
1058
1059 alloc_delta = smosync->smo_alloc - smo->smo_alloc;
1060 defer_delta = freed_map->sm_space - defer_map->sm_space;
1061
1062 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
1063
1064 ASSERT(msp->ms_allocmap[txg & TXG_MASK].sm_space == 0);
1065 ASSERT(msp->ms_freemap[txg & TXG_MASK].sm_space == 0);
1066
1067 /*
1068 * If there's a space_map_load() in progress, wait for it to complete
1069 * so that we have a consistent view of the in-core space map.
1070 * Then, add defer_map (oldest deferred frees) to this map and
1071 * transfer freed_map (this txg's frees) to defer_map.
1072 */
1073 space_map_load_wait(sm);
1074 space_map_vacate(defer_map, sm->sm_loaded ? space_map_free : NULL, sm);
1075 space_map_vacate(freed_map, space_map_add, defer_map);
1076
1077 *smo = *smosync;
1078
1079 msp->ms_deferspace += defer_delta;
1080 ASSERT3S(msp->ms_deferspace, >=, 0);
1081 ASSERT3S(msp->ms_deferspace, <=, sm->sm_size);
1082 if (msp->ms_deferspace != 0) {
1083 /*
1084 * Keep syncing this metaslab until all deferred frees
1085 * are back in circulation.
1086 */
1087 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1088 }
1089
1090 /*
1091 * If the map is loaded but no longer active, evict it as soon as all
1092 * future allocations have synced. (If we unloaded it now and then
1093 * loaded a moment later, the map wouldn't reflect those allocations.)
1094 */
1095 if (sm->sm_loaded && (msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1096 int evictable = 1;
1097
1098 for (int t = 1; t < TXG_CONCURRENT_STATES; t++)
1099 if (msp->ms_allocmap[(txg + t) & TXG_MASK].sm_space)
1100 evictable = 0;
1101
1102 if (evictable && !metaslab_debug)
1103 space_map_unload(sm);
1104 }
1105
1106 metaslab_group_sort(mg, msp, metaslab_weight(msp));
1107
1108 mutex_exit(&msp->ms_lock);
1109 }
1110
1111 void
1112 metaslab_sync_reassess(metaslab_group_t *mg)
1113 {
1114 vdev_t *vd = mg->mg_vd;
1115 int64_t failures = mg->mg_alloc_failures;
1116
1117 /*
1118 * Re-evaluate all metaslabs which have lower offsets than the
1119 * bonus area.
1120 */
1121 for (int m = 0; m < vd->vdev_ms_count; m++) {
1122 metaslab_t *msp = vd->vdev_ms[m];
1123
1124 if (msp->ms_map.sm_start > mg->mg_bonus_area)
1125 break;
1126
1127 mutex_enter(&msp->ms_lock);
1128 metaslab_group_sort(mg, msp, metaslab_weight(msp));
1129 mutex_exit(&msp->ms_lock);
1130 }
1131
1132 atomic_add_64(&mg->mg_alloc_failures, -failures);
1133
1134 /*
1135 * Prefetch the next potential metaslabs
1136 */
1137 metaslab_prefetch(mg);
1138 }
1139
1140 static uint64_t
1141 metaslab_distance(metaslab_t *msp, dva_t *dva)
1142 {
1143 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
1144 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
1145 uint64_t start = msp->ms_map.sm_start >> ms_shift;
1146
1147 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
1148 return (1ULL << 63);
1149
1150 if (offset < start)
1151 return ((start - offset) << ms_shift);
1152 if (offset > start)
1153 return ((offset - start) << ms_shift);
1154 return (0);
1155 }
1156
1157 static uint64_t
1158 metaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize,
1159 uint64_t txg, uint64_t min_distance, dva_t *dva, int d, int flags)
1160 {
1161 spa_t *spa = mg->mg_vd->vdev_spa;
1162 metaslab_t *msp = NULL;
1163 uint64_t offset = -1ULL;
1164 avl_tree_t *t = &mg->mg_metaslab_tree;
1165 uint64_t activation_weight;
1166 uint64_t target_distance;
1167 int i;
1168
1169 activation_weight = METASLAB_WEIGHT_PRIMARY;
1170 for (i = 0; i < d; i++) {
1171 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
1172 activation_weight = METASLAB_WEIGHT_SECONDARY;
1173 break;
1174 }
1175 }
1176
1177 for (;;) {
1178 boolean_t was_active;
1179
1180 mutex_enter(&mg->mg_lock);
1181 for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
1182 if (msp->ms_weight < asize) {
1183 spa_dbgmsg(spa, "%s: failed to meet weight "
1184 "requirement: vdev %llu, txg %llu, mg %p, "
1185 "msp %p, psize %llu, asize %llu, "
1186 "failures %llu, weight %llu",
1187 spa_name(spa), mg->mg_vd->vdev_id, txg,
1188 mg, msp, psize, asize,
1189 mg->mg_alloc_failures, msp->ms_weight);
1190 mutex_exit(&mg->mg_lock);
1191 return (-1ULL);
1192 }
1193 was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
1194 if (activation_weight == METASLAB_WEIGHT_PRIMARY)
1195 break;
1196
1197 target_distance = min_distance +
1198 (msp->ms_smo.smo_alloc ? 0 : min_distance >> 1);
1199
1200 for (i = 0; i < d; i++)
1201 if (metaslab_distance(msp, &dva[i]) <
1202 target_distance)
1203 break;
1204 if (i == d)
1205 break;
1206 }
1207 mutex_exit(&mg->mg_lock);
1208 if (msp == NULL)
1209 return (-1ULL);
1210
1211 /*
1212 * If we've already reached the allowable number of failed
1213 * allocation attempts on this metaslab group then we
1214 * consider skipping it. We skip it only if we're allowed
1215 * to "fast" gang, the physical size is larger than
1216 * a gang block, and we're attempting to allocate from
1217 * the primary metaslab.
1218 */
1219 if (mg->mg_alloc_failures > zfs_mg_alloc_failures &&
1220 CAN_FASTGANG(flags) && psize > SPA_GANGBLOCKSIZE &&
1221 activation_weight == METASLAB_WEIGHT_PRIMARY) {
1222 spa_dbgmsg(spa, "%s: skipping metaslab group: "
1223 "vdev %llu, txg %llu, mg %p, psize %llu, "
1224 "asize %llu, failures %llu", spa_name(spa),
1225 mg->mg_vd->vdev_id, txg, mg, psize, asize,
1226 mg->mg_alloc_failures);
1227 return (-1ULL);
1228 }
1229
1230 mutex_enter(&msp->ms_lock);
1231
1232 /*
1233 * Ensure that the metaslab we have selected is still
1234 * capable of handling our request. It's possible that
1235 * another thread may have changed the weight while we
1236 * were blocked on the metaslab lock.
1237 */
1238 if (msp->ms_weight < asize || (was_active &&
1239 !(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
1240 activation_weight == METASLAB_WEIGHT_PRIMARY)) {
1241 mutex_exit(&msp->ms_lock);
1242 continue;
1243 }
1244
1245 if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
1246 activation_weight == METASLAB_WEIGHT_PRIMARY) {
1247 metaslab_passivate(msp,
1248 msp->ms_weight & ~METASLAB_ACTIVE_MASK);
1249 mutex_exit(&msp->ms_lock);
1250 continue;
1251 }
1252
1253 if (metaslab_activate(msp, activation_weight) != 0) {
1254 mutex_exit(&msp->ms_lock);
1255 continue;
1256 }
1257
1258 if ((offset = space_map_alloc(&msp->ms_map, asize)) != -1ULL)
1259 break;
1260
1261 atomic_inc_64(&mg->mg_alloc_failures);
1262
1263 metaslab_passivate(msp, space_map_maxsize(&msp->ms_map));
1264
1265 mutex_exit(&msp->ms_lock);
1266 }
1267
1268 if (msp->ms_allocmap[txg & TXG_MASK].sm_space == 0)
1269 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
1270
1271 space_map_add(&msp->ms_allocmap[txg & TXG_MASK], offset, asize);
1272
1273 mutex_exit(&msp->ms_lock);
1274
1275 return (offset);
1276 }
1277
1278 /*
1279 * Allocate a block for the specified i/o.
1280 */
1281 static int
1282 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
1283 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
1284 {
1285 metaslab_group_t *mg, *rotor;
1286 vdev_t *vd;
1287 int dshift = 3;
1288 int all_zero;
1289 int zio_lock = B_FALSE;
1290 boolean_t allocatable;
1291 uint64_t offset = -1ULL;
1292 uint64_t asize;
1293 uint64_t distance;
1294
1295 ASSERT(!DVA_IS_VALID(&dva[d]));
1296
1297 /*
1298 * For testing, make some blocks above a certain size be gang blocks.
1299 */
1300 if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0)
1301 return (ENOSPC);
1302
1303 /*
1304 * Start at the rotor and loop through all mgs until we find something.
1305 * Note that there's no locking on mc_rotor or mc_aliquot because
1306 * nothing actually breaks if we miss a few updates -- we just won't
1307 * allocate quite as evenly. It all balances out over time.
1308 *
1309 * If we are doing ditto or log blocks, try to spread them across
1310 * consecutive vdevs. If we're forced to reuse a vdev before we've
1311 * allocated all of our ditto blocks, then try and spread them out on
1312 * that vdev as much as possible. If it turns out to not be possible,
1313 * gradually lower our standards until anything becomes acceptable.
1314 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
1315 * gives us hope of containing our fault domains to something we're
1316 * able to reason about. Otherwise, any two top-level vdev failures
1317 * will guarantee the loss of data. With consecutive allocation,
1318 * only two adjacent top-level vdev failures will result in data loss.
1319 *
1320 * If we are doing gang blocks (hintdva is non-NULL), try to keep
1321 * ourselves on the same vdev as our gang block header. That
1322 * way, we can hope for locality in vdev_cache, plus it makes our
1323 * fault domains something tractable.
1324 */
1325 if (hintdva) {
1326 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
1327
1328 /*
1329 * It's possible the vdev we're using as the hint no
1330 * longer exists (i.e. removed). Consult the rotor when
1331 * all else fails.
1332 */
1333 if (vd != NULL) {
1334 mg = vd->vdev_mg;
1335
1336 if (flags & METASLAB_HINTBP_AVOID &&
1337 mg->mg_next != NULL)
1338 mg = mg->mg_next;
1339 } else {
1340 mg = mc->mc_rotor;
1341 }
1342 } else if (d != 0) {
1343 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
1344 mg = vd->vdev_mg->mg_next;
1345 } else {
1346 mg = mc->mc_rotor;
1347 }
1348
1349 /*
1350 * If the hint put us into the wrong metaslab class, or into a
1351 * metaslab group that has been passivated, just follow the rotor.
1352 */
1353 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
1354 mg = mc->mc_rotor;
1355
1356 rotor = mg;
1357 top:
1358 all_zero = B_TRUE;
1359 do {
1360 ASSERT(mg->mg_activation_count == 1);
1361
1362 vd = mg->mg_vd;
1363
1364 /*
1365 * Don't allocate from faulted devices.
1366 */
1367 if (zio_lock) {
1368 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
1369 allocatable = vdev_allocatable(vd);
1370 spa_config_exit(spa, SCL_ZIO, FTAG);
1371 } else {
1372 allocatable = vdev_allocatable(vd);
1373 }
1374 if (!allocatable)
1375 goto next;
1376
1377 /*
1378 * Avoid writing single-copy data to a failing vdev
1379 */
1380 if ((vd->vdev_stat.vs_write_errors > 0 ||
1381 vd->vdev_state < VDEV_STATE_HEALTHY) &&
1382 d == 0 && dshift == 3) {
1383 all_zero = B_FALSE;
1384 goto next;
1385 }
1386
1387 ASSERT(mg->mg_class == mc);
1388
1389 distance = vd->vdev_asize >> dshift;
1390 if (distance <= (1ULL << vd->vdev_ms_shift))
1391 distance = 0;
1392 else
1393 all_zero = B_FALSE;
1394
1395 asize = vdev_psize_to_asize(vd, psize);
1396 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
1397
1398 offset = metaslab_group_alloc(mg, psize, asize, txg, distance,
1399 dva, d, flags);
1400 if (offset != -1ULL) {
1401 /*
1402 * If we've just selected this metaslab group,
1403 * figure out whether the corresponding vdev is
1404 * over- or under-used relative to the pool,
1405 * and set an allocation bias to even it out.
1406 */
1407 if (mc->mc_aliquot == 0) {
1408 vdev_stat_t *vs = &vd->vdev_stat;
1409 int64_t vu, cu;
1410
1411 vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
1412 cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
1413
1414 /*
1415 * Calculate how much more or less we should
1416 * try to allocate from this device during
1417 * this iteration around the rotor.
1418 * For example, if a device is 80% full
1419 * and the pool is 20% full then we should
1420 * reduce allocations by 60% on this device.
1421 *
1422 * mg_bias = (20 - 80) * 512K / 100 = -307K
1423 *
1424 * This reduces allocations by 307K for this
1425 * iteration.
1426 */
1427 mg->mg_bias = ((cu - vu) *
1428 (int64_t)mg->mg_aliquot) / 100;
1429 }
1430
1431 if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
1432 mg->mg_aliquot + mg->mg_bias) {
1433 mc->mc_rotor = mg->mg_next;
1434 mc->mc_aliquot = 0;
1435 }
1436
1437 DVA_SET_VDEV(&dva[d], vd->vdev_id);
1438 DVA_SET_OFFSET(&dva[d], offset);
1439 DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
1440 DVA_SET_ASIZE(&dva[d], asize);
1441
1442 return (0);
1443 }
1444 next:
1445 mc->mc_rotor = mg->mg_next;
1446 mc->mc_aliquot = 0;
1447 } while ((mg = mg->mg_next) != rotor);
1448
1449 if (!all_zero) {
1450 dshift++;
1451 ASSERT(dshift < 64);
1452 goto top;
1453 }
1454
1455 if (!allocatable && !zio_lock) {
1456 dshift = 3;
1457 zio_lock = B_TRUE;
1458 goto top;
1459 }
1460
1461 bzero(&dva[d], sizeof (dva_t));
1462
1463 return (ENOSPC);
1464 }
1465
1466 /*
1467 * Free the block represented by DVA in the context of the specified
1468 * transaction group.
1469 */
1470 static void
1471 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
1472 {
1473 uint64_t vdev = DVA_GET_VDEV(dva);
1474 uint64_t offset = DVA_GET_OFFSET(dva);
1475 uint64_t size = DVA_GET_ASIZE(dva);
1476 vdev_t *vd;
1477 metaslab_t *msp;
1478
1479 ASSERT(DVA_IS_VALID(dva));
1480
1481 if (txg > spa_freeze_txg(spa))
1482 return;
1483
1484 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
1485 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
1486 cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
1487 (u_longlong_t)vdev, (u_longlong_t)offset);
1488 ASSERT(0);
1489 return;
1490 }
1491
1492 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1493
1494 if (DVA_GET_GANG(dva))
1495 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1496
1497 mutex_enter(&msp->ms_lock);
1498
1499 if (now) {
1500 space_map_remove(&msp->ms_allocmap[txg & TXG_MASK],
1501 offset, size);
1502 space_map_free(&msp->ms_map, offset, size);
1503 } else {
1504 if (msp->ms_freemap[txg & TXG_MASK].sm_space == 0)
1505 vdev_dirty(vd, VDD_METASLAB, msp, txg);
1506 space_map_add(&msp->ms_freemap[txg & TXG_MASK], offset, size);
1507 }
1508
1509 mutex_exit(&msp->ms_lock);
1510 }
1511
1512 /*
1513 * Intent log support: upon opening the pool after a crash, notify the SPA
1514 * of blocks that the intent log has allocated for immediate write, but
1515 * which are still considered free by the SPA because the last transaction
1516 * group didn't commit yet.
1517 */
1518 static int
1519 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
1520 {
1521 uint64_t vdev = DVA_GET_VDEV(dva);
1522 uint64_t offset = DVA_GET_OFFSET(dva);
1523 uint64_t size = DVA_GET_ASIZE(dva);
1524 vdev_t *vd;
1525 metaslab_t *msp;
1526 int error = 0;
1527
1528 ASSERT(DVA_IS_VALID(dva));
1529
1530 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
1531 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
1532 return (ENXIO);
1533
1534 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1535
1536 if (DVA_GET_GANG(dva))
1537 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1538
1539 mutex_enter(&msp->ms_lock);
1540
1541 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_map.sm_loaded)
1542 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
1543
1544 if (error == 0 && !space_map_contains(&msp->ms_map, offset, size))
1545 error = ENOENT;
1546
1547 if (error || txg == 0) { /* txg == 0 indicates dry run */
1548 mutex_exit(&msp->ms_lock);
1549 return (error);
1550 }
1551
1552 space_map_claim(&msp->ms_map, offset, size);
1553
1554 if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
1555 if (msp->ms_allocmap[txg & TXG_MASK].sm_space == 0)
1556 vdev_dirty(vd, VDD_METASLAB, msp, txg);
1557 space_map_add(&msp->ms_allocmap[txg & TXG_MASK], offset, size);
1558 }
1559
1560 mutex_exit(&msp->ms_lock);
1561
1562 return (0);
1563 }
1564
1565 int
1566 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
1567 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
1568 {
1569 dva_t *dva = bp->blk_dva;
1570 dva_t *hintdva = hintbp->blk_dva;
1571 int error = 0;
1572
1573 ASSERT(bp->blk_birth == 0);
1574 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
1575
1576 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
1577
1578 if (mc->mc_rotor == NULL) { /* no vdevs in this class */
1579 spa_config_exit(spa, SCL_ALLOC, FTAG);
1580 return (ENOSPC);
1581 }
1582
1583 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
1584 ASSERT(BP_GET_NDVAS(bp) == 0);
1585 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
1586
1587 for (int d = 0; d < ndvas; d++) {
1588 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
1589 txg, flags);
1590 if (error) {
1591 for (d--; d >= 0; d--) {
1592 metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
1593 bzero(&dva[d], sizeof (dva_t));
1594 }
1595 spa_config_exit(spa, SCL_ALLOC, FTAG);
1596 return (error);
1597 }
1598 }
1599 ASSERT(error == 0);
1600 ASSERT(BP_GET_NDVAS(bp) == ndvas);
1601
1602 spa_config_exit(spa, SCL_ALLOC, FTAG);
1603
1604 BP_SET_BIRTH(bp, txg, txg);
1605
1606 return (0);
1607 }
1608
1609 void
1610 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
1611 {
1612 const dva_t *dva = bp->blk_dva;
1613 int ndvas = BP_GET_NDVAS(bp);
1614
1615 ASSERT(!BP_IS_HOLE(bp));
1616 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
1617
1618 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
1619
1620 for (int d = 0; d < ndvas; d++)
1621 metaslab_free_dva(spa, &dva[d], txg, now);
1622
1623 spa_config_exit(spa, SCL_FREE, FTAG);
1624 }
1625
1626 int
1627 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
1628 {
1629 const dva_t *dva = bp->blk_dva;
1630 int ndvas = BP_GET_NDVAS(bp);
1631 int error = 0;
1632
1633 ASSERT(!BP_IS_HOLE(bp));
1634
1635 if (txg != 0) {
1636 /*
1637 * First do a dry run to make sure all DVAs are claimable,
1638 * so we don't have to unwind from partial failures below.
1639 */
1640 if ((error = metaslab_claim(spa, bp, 0)) != 0)
1641 return (error);
1642 }
1643
1644 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
1645
1646 for (int d = 0; d < ndvas; d++)
1647 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
1648 break;
1649
1650 spa_config_exit(spa, SCL_ALLOC, FTAG);
1651
1652 ASSERT(error == 0 || txg == 0);
1653
1654 return (error);
1655 }