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) 2011, 2015 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
26 */
27
28 #include <sys/zfs_context.h>
29 #include <sys/dmu.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/space_map.h>
32 #include <sys/metaslab_impl.h>
33 #include <sys/vdev_impl.h>
34 #include <sys/zio.h>
35 #include <sys/spa_impl.h>
36 #include <sys/zfeature.h>
37
38 SYSCTL_DECL(_vfs_zfs);
39 SYSCTL_NODE(_vfs_zfs, OID_AUTO, metaslab, CTLFLAG_RW, 0, "ZFS metaslab");
40
41 #define GANG_ALLOCATION(flags) \
42 ((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER))
43
44 uint64_t metaslab_aliquot = 512ULL << 10;
45 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1; /* force gang blocks */
46 SYSCTL_QUAD(_vfs_zfs_metaslab, OID_AUTO, gang_bang, CTLFLAG_RWTUN,
47 &metaslab_gang_bang, 0,
48 "Force gang block allocation for blocks larger than or equal to this value");
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 SYSCTL_INT(_vfs_zfs, OID_AUTO, condense_pct, CTLFLAG_RWTUN,
58 &zfs_condense_pct, 0,
59 "Condense on-disk spacemap when it is more than this many percents"
60 " of in-memory counterpart");
61
62 /*
63 * Condensing a metaslab is not guaranteed to actually reduce the amount of
64 * space used on disk. In particular, a space map uses data in increments of
65 * MAX(1 << ashift, space_map_blksize), so a metaslab might use the
66 * same number of blocks after condensing. Since the goal of condensing is to
67 * reduce the number of IOPs required to read the space map, we only want to
68 * condense when we can be sure we will reduce the number of blocks used by the
69 * space map. Unfortunately, we cannot precisely compute whether or not this is
70 * the case in metaslab_should_condense since we are holding ms_lock. Instead,
71 * we apply the following heuristic: do not condense a spacemap unless the
72 * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
73 * blocks.
74 */
75 int zfs_metaslab_condense_block_threshold = 4;
76
77 /*
78 * The zfs_mg_noalloc_threshold defines which metaslab groups should
79 * be eligible for allocation. The value is defined as a percentage of
80 * free space. Metaslab groups that have more free space than
81 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
82 * a metaslab group's free space is less than or equal to the
83 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
84 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
85 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
86 * groups are allowed to accept allocations. Gang blocks are always
87 * eligible to allocate on any metaslab group. The default value of 0 means
88 * no metaslab group will be excluded based on this criterion.
89 */
90 int zfs_mg_noalloc_threshold = 0;
91 SYSCTL_INT(_vfs_zfs, OID_AUTO, mg_noalloc_threshold, CTLFLAG_RWTUN,
92 &zfs_mg_noalloc_threshold, 0,
93 "Percentage of metaslab group size that should be free"
94 " to make it eligible for allocation");
95
96 /*
97 * Metaslab groups are considered eligible for allocations if their
98 * fragmenation metric (measured as a percentage) is less than or equal to
99 * zfs_mg_fragmentation_threshold. If a metaslab group exceeds this threshold
100 * then it will be skipped unless all metaslab groups within the metaslab
101 * class have also crossed this threshold.
102 */
103 int zfs_mg_fragmentation_threshold = 85;
104 SYSCTL_INT(_vfs_zfs, OID_AUTO, mg_fragmentation_threshold, CTLFLAG_RWTUN,
105 &zfs_mg_fragmentation_threshold, 0,
106 "Percentage of metaslab group size that should be considered "
107 "eligible for allocations unless all metaslab groups within the metaslab class "
108 "have also crossed this threshold");
109
110 /*
111 * Allow metaslabs to keep their active state as long as their fragmentation
112 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
113 * active metaslab that exceeds this threshold will no longer keep its active
114 * status allowing better metaslabs to be selected.
115 */
116 int zfs_metaslab_fragmentation_threshold = 70;
117 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, fragmentation_threshold, CTLFLAG_RWTUN,
118 &zfs_metaslab_fragmentation_threshold, 0,
119 "Maximum percentage of metaslab fragmentation level to keep their active state");
120
121 /*
122 * When set will load all metaslabs when pool is first opened.
123 */
124 int metaslab_debug_load = 0;
125 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, debug_load, CTLFLAG_RWTUN,
126 &metaslab_debug_load, 0,
127 "Load all metaslabs when pool is first opened");
128
129 /*
130 * When set will prevent metaslabs from being unloaded.
131 */
132 int metaslab_debug_unload = 0;
133 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, debug_unload, CTLFLAG_RWTUN,
134 &metaslab_debug_unload, 0,
135 "Prevent metaslabs from being unloaded");
136
137 /*
138 * Minimum size which forces the dynamic allocator to change
139 * it's allocation strategy. Once the space map cannot satisfy
140 * an allocation of this size then it switches to using more
141 * aggressive strategy (i.e search by size rather than offset).
142 */
143 uint64_t metaslab_df_alloc_threshold = SPA_OLD_MAXBLOCKSIZE;
144 SYSCTL_QUAD(_vfs_zfs_metaslab, OID_AUTO, df_alloc_threshold, CTLFLAG_RWTUN,
145 &metaslab_df_alloc_threshold, 0,
146 "Minimum size which forces the dynamic allocator to change it's allocation strategy");
147
148 /*
149 * The minimum free space, in percent, which must be available
150 * in a space map to continue allocations in a first-fit fashion.
151 * Once the space map's free space drops below this level we dynamically
152 * switch to using best-fit allocations.
153 */
154 int metaslab_df_free_pct = 4;
155 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, df_free_pct, CTLFLAG_RWTUN,
156 &metaslab_df_free_pct, 0,
157 "The minimum free space, in percent, which must be available in a "
158 "space map to continue allocations in a first-fit fashion");
159
160 /*
161 * A metaslab is considered "free" if it contains a contiguous
162 * segment which is greater than metaslab_min_alloc_size.
163 */
164 uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
165 SYSCTL_QUAD(_vfs_zfs_metaslab, OID_AUTO, min_alloc_size, CTLFLAG_RWTUN,
166 &metaslab_min_alloc_size, 0,
167 "A metaslab is considered \"free\" if it contains a contiguous "
168 "segment which is greater than vfs.zfs.metaslab.min_alloc_size");
169
170 /*
171 * Percentage of all cpus that can be used by the metaslab taskq.
172 */
173 int metaslab_load_pct = 50;
174 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, load_pct, CTLFLAG_RWTUN,
175 &metaslab_load_pct, 0,
176 "Percentage of cpus that can be used by the metaslab taskq");
177
178 /*
179 * Determines how many txgs a metaslab may remain loaded without having any
180 * allocations from it. As long as a metaslab continues to be used we will
181 * keep it loaded.
182 */
183 int metaslab_unload_delay = TXG_SIZE * 2;
184 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, unload_delay, CTLFLAG_RWTUN,
185 &metaslab_unload_delay, 0,
186 "Number of TXGs that an unused metaslab can be kept in memory");
187
188 /*
189 * Max number of metaslabs per group to preload.
190 */
191 int metaslab_preload_limit = SPA_DVAS_PER_BP;
192 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, preload_limit, CTLFLAG_RWTUN,
193 &metaslab_preload_limit, 0,
194 "Max number of metaslabs per group to preload");
195
196 /*
197 * Enable/disable preloading of metaslab.
198 */
199 boolean_t metaslab_preload_enabled = B_TRUE;
200 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, preload_enabled, CTLFLAG_RWTUN,
201 &metaslab_preload_enabled, 0,
202 "Max number of metaslabs per group to preload");
203
204 /*
205 * Enable/disable fragmentation weighting on metaslabs.
206 */
207 boolean_t metaslab_fragmentation_factor_enabled = B_TRUE;
208 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, fragmentation_factor_enabled, CTLFLAG_RWTUN,
209 &metaslab_fragmentation_factor_enabled, 0,
210 "Enable fragmentation weighting on metaslabs");
211
212 /*
213 * Enable/disable lba weighting (i.e. outer tracks are given preference).
214 */
215 boolean_t metaslab_lba_weighting_enabled = B_TRUE;
216 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, lba_weighting_enabled, CTLFLAG_RWTUN,
217 &metaslab_lba_weighting_enabled, 0,
218 "Enable LBA weighting (i.e. outer tracks are given preference)");
219
220 /*
221 * Enable/disable metaslab group biasing.
222 */
223 boolean_t metaslab_bias_enabled = B_TRUE;
224 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, bias_enabled, CTLFLAG_RWTUN,
225 &metaslab_bias_enabled, 0,
226 "Enable metaslab group biasing");
227
228 /*
229 * Enable/disable segment-based metaslab selection.
230 */
231 boolean_t zfs_metaslab_segment_weight_enabled = B_TRUE;
232
233 /*
234 * When using segment-based metaslab selection, we will continue
235 * allocating from the active metaslab until we have exhausted
236 * zfs_metaslab_switch_threshold of its buckets.
237 */
238 int zfs_metaslab_switch_threshold = 2;
239
240 /*
241 * Internal switch to enable/disable the metaslab allocation tracing
242 * facility.
243 */
244 boolean_t metaslab_trace_enabled = B_TRUE;
245
246 /*
247 * Maximum entries that the metaslab allocation tracing facility will keep
248 * in a given list when running in non-debug mode. We limit the number
249 * of entries in non-debug mode to prevent us from using up too much memory.
250 * The limit should be sufficiently large that we don't expect any allocation
251 * to every exceed this value. In debug mode, the system will panic if this
252 * limit is ever reached allowing for further investigation.
253 */
254 uint64_t metaslab_trace_max_entries = 5000;
255
256 static uint64_t metaslab_weight(metaslab_t *);
257 static void metaslab_set_fragmentation(metaslab_t *);
258
259 kmem_cache_t *metaslab_alloc_trace_cache;
260
261 /*
262 * ==========================================================================
263 * Metaslab classes
264 * ==========================================================================
265 */
266 metaslab_class_t *
metaslab_class_create(spa_t * spa,metaslab_ops_t * ops)267 metaslab_class_create(spa_t *spa, metaslab_ops_t *ops)
268 {
269 metaslab_class_t *mc;
270
271 mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
272
273 mc->mc_spa = spa;
274 mc->mc_rotor = NULL;
275 mc->mc_ops = ops;
276 mutex_init(&mc->mc_lock, NULL, MUTEX_DEFAULT, NULL);
277 refcount_create_tracked(&mc->mc_alloc_slots);
278
279 return (mc);
280 }
281
282 void
metaslab_class_destroy(metaslab_class_t * mc)283 metaslab_class_destroy(metaslab_class_t *mc)
284 {
285 ASSERT(mc->mc_rotor == NULL);
286 ASSERT(mc->mc_alloc == 0);
287 ASSERT(mc->mc_deferred == 0);
288 ASSERT(mc->mc_space == 0);
289 ASSERT(mc->mc_dspace == 0);
290
291 refcount_destroy(&mc->mc_alloc_slots);
292 mutex_destroy(&mc->mc_lock);
293 kmem_free(mc, sizeof (metaslab_class_t));
294 }
295
296 int
metaslab_class_validate(metaslab_class_t * mc)297 metaslab_class_validate(metaslab_class_t *mc)
298 {
299 metaslab_group_t *mg;
300 vdev_t *vd;
301
302 /*
303 * Must hold one of the spa_config locks.
304 */
305 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
306 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
307
308 if ((mg = mc->mc_rotor) == NULL)
309 return (0);
310
311 do {
312 vd = mg->mg_vd;
313 ASSERT(vd->vdev_mg != NULL);
314 ASSERT3P(vd->vdev_top, ==, vd);
315 ASSERT3P(mg->mg_class, ==, mc);
316 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
317 } while ((mg = mg->mg_next) != mc->mc_rotor);
318
319 return (0);
320 }
321
322 void
metaslab_class_space_update(metaslab_class_t * mc,int64_t alloc_delta,int64_t defer_delta,int64_t space_delta,int64_t dspace_delta)323 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
324 int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
325 {
326 atomic_add_64(&mc->mc_alloc, alloc_delta);
327 atomic_add_64(&mc->mc_deferred, defer_delta);
328 atomic_add_64(&mc->mc_space, space_delta);
329 atomic_add_64(&mc->mc_dspace, dspace_delta);
330 }
331
332 void
metaslab_class_minblocksize_update(metaslab_class_t * mc)333 metaslab_class_minblocksize_update(metaslab_class_t *mc)
334 {
335 metaslab_group_t *mg;
336 vdev_t *vd;
337 uint64_t minashift = UINT64_MAX;
338
339 if ((mg = mc->mc_rotor) == NULL) {
340 mc->mc_minblocksize = SPA_MINBLOCKSIZE;
341 return;
342 }
343
344 do {
345 vd = mg->mg_vd;
346 if (vd->vdev_ashift < minashift)
347 minashift = vd->vdev_ashift;
348 } while ((mg = mg->mg_next) != mc->mc_rotor);
349
350 mc->mc_minblocksize = 1ULL << minashift;
351 }
352
353 uint64_t
metaslab_class_get_alloc(metaslab_class_t * mc)354 metaslab_class_get_alloc(metaslab_class_t *mc)
355 {
356 return (mc->mc_alloc);
357 }
358
359 uint64_t
metaslab_class_get_deferred(metaslab_class_t * mc)360 metaslab_class_get_deferred(metaslab_class_t *mc)
361 {
362 return (mc->mc_deferred);
363 }
364
365 uint64_t
metaslab_class_get_space(metaslab_class_t * mc)366 metaslab_class_get_space(metaslab_class_t *mc)
367 {
368 return (mc->mc_space);
369 }
370
371 uint64_t
metaslab_class_get_dspace(metaslab_class_t * mc)372 metaslab_class_get_dspace(metaslab_class_t *mc)
373 {
374 return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
375 }
376
377 uint64_t
metaslab_class_get_minblocksize(metaslab_class_t * mc)378 metaslab_class_get_minblocksize(metaslab_class_t *mc)
379 {
380 return (mc->mc_minblocksize);
381 }
382
383 void
metaslab_class_histogram_verify(metaslab_class_t * mc)384 metaslab_class_histogram_verify(metaslab_class_t *mc)
385 {
386 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
387 uint64_t *mc_hist;
388 int i;
389
390 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
391 return;
392
393 mc_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
394 KM_SLEEP);
395
396 for (int c = 0; c < rvd->vdev_children; c++) {
397 vdev_t *tvd = rvd->vdev_child[c];
398 metaslab_group_t *mg = tvd->vdev_mg;
399
400 /*
401 * Skip any holes, uninitialized top-levels, or
402 * vdevs that are not in this metalab class.
403 */
404 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
405 mg->mg_class != mc) {
406 continue;
407 }
408
409 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
410 mc_hist[i] += mg->mg_histogram[i];
411 }
412
413 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
414 VERIFY3U(mc_hist[i], ==, mc->mc_histogram[i]);
415
416 kmem_free(mc_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
417 }
418
419 /*
420 * Calculate the metaslab class's fragmentation metric. The metric
421 * is weighted based on the space contribution of each metaslab group.
422 * The return value will be a number between 0 and 100 (inclusive), or
423 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
424 * zfs_frag_table for more information about the metric.
425 */
426 uint64_t
metaslab_class_fragmentation(metaslab_class_t * mc)427 metaslab_class_fragmentation(metaslab_class_t *mc)
428 {
429 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
430 uint64_t fragmentation = 0;
431
432 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
433
434 for (int c = 0; c < rvd->vdev_children; c++) {
435 vdev_t *tvd = rvd->vdev_child[c];
436 metaslab_group_t *mg = tvd->vdev_mg;
437
438 /*
439 * Skip any holes, uninitialized top-levels, or
440 * vdevs that are not in this metalab class.
441 */
442 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
443 mg->mg_class != mc) {
444 continue;
445 }
446
447 /*
448 * If a metaslab group does not contain a fragmentation
449 * metric then just bail out.
450 */
451 if (mg->mg_fragmentation == ZFS_FRAG_INVALID) {
452 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
453 return (ZFS_FRAG_INVALID);
454 }
455
456 /*
457 * Determine how much this metaslab_group is contributing
458 * to the overall pool fragmentation metric.
459 */
460 fragmentation += mg->mg_fragmentation *
461 metaslab_group_get_space(mg);
462 }
463 fragmentation /= metaslab_class_get_space(mc);
464
465 ASSERT3U(fragmentation, <=, 100);
466 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
467 return (fragmentation);
468 }
469
470 /*
471 * Calculate the amount of expandable space that is available in
472 * this metaslab class. If a device is expanded then its expandable
473 * space will be the amount of allocatable space that is currently not
474 * part of this metaslab class.
475 */
476 uint64_t
metaslab_class_expandable_space(metaslab_class_t * mc)477 metaslab_class_expandable_space(metaslab_class_t *mc)
478 {
479 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
480 uint64_t space = 0;
481
482 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
483 for (int c = 0; c < rvd->vdev_children; c++) {
484 vdev_t *tvd = rvd->vdev_child[c];
485 metaslab_group_t *mg = tvd->vdev_mg;
486
487 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
488 mg->mg_class != mc) {
489 continue;
490 }
491
492 /*
493 * Calculate if we have enough space to add additional
494 * metaslabs. We report the expandable space in terms
495 * of the metaslab size since that's the unit of expansion.
496 */
497 space += P2ALIGN(tvd->vdev_max_asize - tvd->vdev_asize,
498 1ULL << tvd->vdev_ms_shift);
499 }
500 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
501 return (space);
502 }
503
504 static int
metaslab_compare(const void * x1,const void * x2)505 metaslab_compare(const void *x1, const void *x2)
506 {
507 const metaslab_t *m1 = x1;
508 const metaslab_t *m2 = x2;
509
510 if (m1->ms_weight < m2->ms_weight)
511 return (1);
512 if (m1->ms_weight > m2->ms_weight)
513 return (-1);
514
515 /*
516 * If the weights are identical, use the offset to force uniqueness.
517 */
518 if (m1->ms_start < m2->ms_start)
519 return (-1);
520 if (m1->ms_start > m2->ms_start)
521 return (1);
522
523 ASSERT3P(m1, ==, m2);
524
525 return (0);
526 }
527
528 /*
529 * Verify that the space accounting on disk matches the in-core range_trees.
530 */
531 void
metaslab_verify_space(metaslab_t * msp,uint64_t txg)532 metaslab_verify_space(metaslab_t *msp, uint64_t txg)
533 {
534 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
535 uint64_t allocated = 0;
536 uint64_t freed = 0;
537 uint64_t sm_free_space, msp_free_space;
538
539 ASSERT(MUTEX_HELD(&msp->ms_lock));
540
541 if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0)
542 return;
543
544 /*
545 * We can only verify the metaslab space when we're called
546 * from syncing context with a loaded metaslab that has an allocated
547 * space map. Calling this in non-syncing context does not
548 * provide a consistent view of the metaslab since we're performing
549 * allocations in the future.
550 */
551 if (txg != spa_syncing_txg(spa) || msp->ms_sm == NULL ||
552 !msp->ms_loaded)
553 return;
554
555 sm_free_space = msp->ms_size - space_map_allocated(msp->ms_sm) -
556 space_map_alloc_delta(msp->ms_sm);
557
558 /*
559 * Account for future allocations since we would have already
560 * deducted that space from the ms_freetree.
561 */
562 for (int t = 0; t < TXG_CONCURRENT_STATES; t++) {
563 allocated +=
564 range_tree_space(msp->ms_alloctree[(txg + t) & TXG_MASK]);
565 }
566 freed = range_tree_space(msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK]);
567
568 msp_free_space = range_tree_space(msp->ms_tree) + allocated +
569 msp->ms_deferspace + freed;
570
571 VERIFY3U(sm_free_space, ==, msp_free_space);
572 }
573
574 /*
575 * ==========================================================================
576 * Metaslab groups
577 * ==========================================================================
578 */
579 /*
580 * Update the allocatable flag and the metaslab group's capacity.
581 * The allocatable flag is set to true if the capacity is below
582 * the zfs_mg_noalloc_threshold or has a fragmentation value that is
583 * greater than zfs_mg_fragmentation_threshold. If a metaslab group
584 * transitions from allocatable to non-allocatable or vice versa then the
585 * metaslab group's class is updated to reflect the transition.
586 */
587 static void
metaslab_group_alloc_update(metaslab_group_t * mg)588 metaslab_group_alloc_update(metaslab_group_t *mg)
589 {
590 vdev_t *vd = mg->mg_vd;
591 metaslab_class_t *mc = mg->mg_class;
592 vdev_stat_t *vs = &vd->vdev_stat;
593 boolean_t was_allocatable;
594 boolean_t was_initialized;
595
596 ASSERT(vd == vd->vdev_top);
597
598 mutex_enter(&mg->mg_lock);
599 was_allocatable = mg->mg_allocatable;
600 was_initialized = mg->mg_initialized;
601
602 mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
603 (vs->vs_space + 1);
604
605 mutex_enter(&mc->mc_lock);
606
607 /*
608 * If the metaslab group was just added then it won't
609 * have any space until we finish syncing out this txg.
610 * At that point we will consider it initialized and available
611 * for allocations. We also don't consider non-activated
612 * metaslab groups (e.g. vdevs that are in the middle of being removed)
613 * to be initialized, because they can't be used for allocation.
614 */
615 mg->mg_initialized = metaslab_group_initialized(mg);
616 if (!was_initialized && mg->mg_initialized) {
617 mc->mc_groups++;
618 } else if (was_initialized && !mg->mg_initialized) {
619 ASSERT3U(mc->mc_groups, >, 0);
620 mc->mc_groups--;
621 }
622 if (mg->mg_initialized)
623 mg->mg_no_free_space = B_FALSE;
624
625 /*
626 * A metaslab group is considered allocatable if it has plenty
627 * of free space or is not heavily fragmented. We only take
628 * fragmentation into account if the metaslab group has a valid
629 * fragmentation metric (i.e. a value between 0 and 100).
630 */
631 mg->mg_allocatable = (mg->mg_activation_count > 0 &&
632 mg->mg_free_capacity > zfs_mg_noalloc_threshold &&
633 (mg->mg_fragmentation == ZFS_FRAG_INVALID ||
634 mg->mg_fragmentation <= zfs_mg_fragmentation_threshold));
635
636 /*
637 * The mc_alloc_groups maintains a count of the number of
638 * groups in this metaslab class that are still above the
639 * zfs_mg_noalloc_threshold. This is used by the allocating
640 * threads to determine if they should avoid allocations to
641 * a given group. The allocator will avoid allocations to a group
642 * if that group has reached or is below the zfs_mg_noalloc_threshold
643 * and there are still other groups that are above the threshold.
644 * When a group transitions from allocatable to non-allocatable or
645 * vice versa we update the metaslab class to reflect that change.
646 * When the mc_alloc_groups value drops to 0 that means that all
647 * groups have reached the zfs_mg_noalloc_threshold making all groups
648 * eligible for allocations. This effectively means that all devices
649 * are balanced again.
650 */
651 if (was_allocatable && !mg->mg_allocatable)
652 mc->mc_alloc_groups--;
653 else if (!was_allocatable && mg->mg_allocatable)
654 mc->mc_alloc_groups++;
655 mutex_exit(&mc->mc_lock);
656
657 mutex_exit(&mg->mg_lock);
658 }
659
660 metaslab_group_t *
metaslab_group_create(metaslab_class_t * mc,vdev_t * vd)661 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
662 {
663 metaslab_group_t *mg;
664
665 mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
666 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
667 avl_create(&mg->mg_metaslab_tree, metaslab_compare,
668 sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
669 mg->mg_vd = vd;
670 mg->mg_class = mc;
671 mg->mg_activation_count = 0;
672 mg->mg_initialized = B_FALSE;
673 mg->mg_no_free_space = B_TRUE;
674 refcount_create_tracked(&mg->mg_alloc_queue_depth);
675
676 mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct,
677 minclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT);
678
679 return (mg);
680 }
681
682 void
metaslab_group_destroy(metaslab_group_t * mg)683 metaslab_group_destroy(metaslab_group_t *mg)
684 {
685 ASSERT(mg->mg_prev == NULL);
686 ASSERT(mg->mg_next == NULL);
687 /*
688 * We may have gone below zero with the activation count
689 * either because we never activated in the first place or
690 * because we're done, and possibly removing the vdev.
691 */
692 ASSERT(mg->mg_activation_count <= 0);
693
694 taskq_destroy(mg->mg_taskq);
695 avl_destroy(&mg->mg_metaslab_tree);
696 mutex_destroy(&mg->mg_lock);
697 refcount_destroy(&mg->mg_alloc_queue_depth);
698 kmem_free(mg, sizeof (metaslab_group_t));
699 }
700
701 void
metaslab_group_activate(metaslab_group_t * mg)702 metaslab_group_activate(metaslab_group_t *mg)
703 {
704 metaslab_class_t *mc = mg->mg_class;
705 metaslab_group_t *mgprev, *mgnext;
706
707 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
708
709 ASSERT(mc->mc_rotor != mg);
710 ASSERT(mg->mg_prev == NULL);
711 ASSERT(mg->mg_next == NULL);
712 ASSERT(mg->mg_activation_count <= 0);
713
714 if (++mg->mg_activation_count <= 0)
715 return;
716
717 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
718 metaslab_group_alloc_update(mg);
719
720 if ((mgprev = mc->mc_rotor) == NULL) {
721 mg->mg_prev = mg;
722 mg->mg_next = mg;
723 } else {
724 mgnext = mgprev->mg_next;
725 mg->mg_prev = mgprev;
726 mg->mg_next = mgnext;
727 mgprev->mg_next = mg;
728 mgnext->mg_prev = mg;
729 }
730 mc->mc_rotor = mg;
731 metaslab_class_minblocksize_update(mc);
732 }
733
734 void
metaslab_group_passivate(metaslab_group_t * mg)735 metaslab_group_passivate(metaslab_group_t *mg)
736 {
737 metaslab_class_t *mc = mg->mg_class;
738 metaslab_group_t *mgprev, *mgnext;
739
740 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
741
742 if (--mg->mg_activation_count != 0) {
743 ASSERT(mc->mc_rotor != mg);
744 ASSERT(mg->mg_prev == NULL);
745 ASSERT(mg->mg_next == NULL);
746 ASSERT(mg->mg_activation_count < 0);
747 return;
748 }
749
750 taskq_wait(mg->mg_taskq);
751 metaslab_group_alloc_update(mg);
752
753 mgprev = mg->mg_prev;
754 mgnext = mg->mg_next;
755
756 if (mg == mgnext) {
757 mc->mc_rotor = NULL;
758 } else {
759 mc->mc_rotor = mgnext;
760 mgprev->mg_next = mgnext;
761 mgnext->mg_prev = mgprev;
762 }
763
764 mg->mg_prev = NULL;
765 mg->mg_next = NULL;
766 metaslab_class_minblocksize_update(mc);
767 }
768
769 boolean_t
metaslab_group_initialized(metaslab_group_t * mg)770 metaslab_group_initialized(metaslab_group_t *mg)
771 {
772 vdev_t *vd = mg->mg_vd;
773 vdev_stat_t *vs = &vd->vdev_stat;
774
775 return (vs->vs_space != 0 && mg->mg_activation_count > 0);
776 }
777
778 uint64_t
metaslab_group_get_space(metaslab_group_t * mg)779 metaslab_group_get_space(metaslab_group_t *mg)
780 {
781 return ((1ULL << mg->mg_vd->vdev_ms_shift) * mg->mg_vd->vdev_ms_count);
782 }
783
784 void
metaslab_group_histogram_verify(metaslab_group_t * mg)785 metaslab_group_histogram_verify(metaslab_group_t *mg)
786 {
787 uint64_t *mg_hist;
788 vdev_t *vd = mg->mg_vd;
789 uint64_t ashift = vd->vdev_ashift;
790 int i;
791
792 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
793 return;
794
795 mg_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
796 KM_SLEEP);
797
798 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE, >=,
799 SPACE_MAP_HISTOGRAM_SIZE + ashift);
800
801 for (int m = 0; m < vd->vdev_ms_count; m++) {
802 metaslab_t *msp = vd->vdev_ms[m];
803
804 if (msp->ms_sm == NULL)
805 continue;
806
807 for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++)
808 mg_hist[i + ashift] +=
809 msp->ms_sm->sm_phys->smp_histogram[i];
810 }
811
812 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i ++)
813 VERIFY3U(mg_hist[i], ==, mg->mg_histogram[i]);
814
815 kmem_free(mg_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
816 }
817
818 static void
metaslab_group_histogram_add(metaslab_group_t * mg,metaslab_t * msp)819 metaslab_group_histogram_add(metaslab_group_t *mg, metaslab_t *msp)
820 {
821 metaslab_class_t *mc = mg->mg_class;
822 uint64_t ashift = mg->mg_vd->vdev_ashift;
823
824 ASSERT(MUTEX_HELD(&msp->ms_lock));
825 if (msp->ms_sm == NULL)
826 return;
827
828 mutex_enter(&mg->mg_lock);
829 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
830 mg->mg_histogram[i + ashift] +=
831 msp->ms_sm->sm_phys->smp_histogram[i];
832 mc->mc_histogram[i + ashift] +=
833 msp->ms_sm->sm_phys->smp_histogram[i];
834 }
835 mutex_exit(&mg->mg_lock);
836 }
837
838 void
metaslab_group_histogram_remove(metaslab_group_t * mg,metaslab_t * msp)839 metaslab_group_histogram_remove(metaslab_group_t *mg, metaslab_t *msp)
840 {
841 metaslab_class_t *mc = mg->mg_class;
842 uint64_t ashift = mg->mg_vd->vdev_ashift;
843
844 ASSERT(MUTEX_HELD(&msp->ms_lock));
845 if (msp->ms_sm == NULL)
846 return;
847
848 mutex_enter(&mg->mg_lock);
849 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
850 ASSERT3U(mg->mg_histogram[i + ashift], >=,
851 msp->ms_sm->sm_phys->smp_histogram[i]);
852 ASSERT3U(mc->mc_histogram[i + ashift], >=,
853 msp->ms_sm->sm_phys->smp_histogram[i]);
854
855 mg->mg_histogram[i + ashift] -=
856 msp->ms_sm->sm_phys->smp_histogram[i];
857 mc->mc_histogram[i + ashift] -=
858 msp->ms_sm->sm_phys->smp_histogram[i];
859 }
860 mutex_exit(&mg->mg_lock);
861 }
862
863 static void
metaslab_group_add(metaslab_group_t * mg,metaslab_t * msp)864 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
865 {
866 ASSERT(msp->ms_group == NULL);
867 mutex_enter(&mg->mg_lock);
868 msp->ms_group = mg;
869 msp->ms_weight = 0;
870 avl_add(&mg->mg_metaslab_tree, msp);
871 mutex_exit(&mg->mg_lock);
872
873 mutex_enter(&msp->ms_lock);
874 metaslab_group_histogram_add(mg, msp);
875 mutex_exit(&msp->ms_lock);
876 }
877
878 static void
metaslab_group_remove(metaslab_group_t * mg,metaslab_t * msp)879 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
880 {
881 mutex_enter(&msp->ms_lock);
882 metaslab_group_histogram_remove(mg, msp);
883 mutex_exit(&msp->ms_lock);
884
885 mutex_enter(&mg->mg_lock);
886 ASSERT(msp->ms_group == mg);
887 avl_remove(&mg->mg_metaslab_tree, msp);
888 msp->ms_group = NULL;
889 mutex_exit(&mg->mg_lock);
890 }
891
892 static void
metaslab_group_sort(metaslab_group_t * mg,metaslab_t * msp,uint64_t weight)893 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
894 {
895 /*
896 * Although in principle the weight can be any value, in
897 * practice we do not use values in the range [1, 511].
898 */
899 ASSERT(weight >= SPA_MINBLOCKSIZE || weight == 0);
900 ASSERT(MUTEX_HELD(&msp->ms_lock));
901
902 mutex_enter(&mg->mg_lock);
903 ASSERT(msp->ms_group == mg);
904 avl_remove(&mg->mg_metaslab_tree, msp);
905 msp->ms_weight = weight;
906 avl_add(&mg->mg_metaslab_tree, msp);
907 mutex_exit(&mg->mg_lock);
908 }
909
910 /*
911 * Calculate the fragmentation for a given metaslab group. We can use
912 * a simple average here since all metaslabs within the group must have
913 * the same size. The return value will be a value between 0 and 100
914 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
915 * group have a fragmentation metric.
916 */
917 uint64_t
metaslab_group_fragmentation(metaslab_group_t * mg)918 metaslab_group_fragmentation(metaslab_group_t *mg)
919 {
920 vdev_t *vd = mg->mg_vd;
921 uint64_t fragmentation = 0;
922 uint64_t valid_ms = 0;
923
924 for (int m = 0; m < vd->vdev_ms_count; m++) {
925 metaslab_t *msp = vd->vdev_ms[m];
926
927 if (msp->ms_fragmentation == ZFS_FRAG_INVALID)
928 continue;
929
930 valid_ms++;
931 fragmentation += msp->ms_fragmentation;
932 }
933
934 if (valid_ms <= vd->vdev_ms_count / 2)
935 return (ZFS_FRAG_INVALID);
936
937 fragmentation /= valid_ms;
938 ASSERT3U(fragmentation, <=, 100);
939 return (fragmentation);
940 }
941
942 /*
943 * Determine if a given metaslab group should skip allocations. A metaslab
944 * group should avoid allocations if its free capacity is less than the
945 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
946 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
947 * that can still handle allocations. If the allocation throttle is enabled
948 * then we skip allocations to devices that have reached their maximum
949 * allocation queue depth unless the selected metaslab group is the only
950 * eligible group remaining.
951 */
952 static boolean_t
metaslab_group_allocatable(metaslab_group_t * mg,metaslab_group_t * rotor,uint64_t psize)953 metaslab_group_allocatable(metaslab_group_t *mg, metaslab_group_t *rotor,
954 uint64_t psize)
955 {
956 spa_t *spa = mg->mg_vd->vdev_spa;
957 metaslab_class_t *mc = mg->mg_class;
958
959 /*
960 * We can only consider skipping this metaslab group if it's
961 * in the normal metaslab class and there are other metaslab
962 * groups to select from. Otherwise, we always consider it eligible
963 * for allocations.
964 */
965 if (mc != spa_normal_class(spa) || mc->mc_groups <= 1)
966 return (B_TRUE);
967
968 /*
969 * If the metaslab group's mg_allocatable flag is set (see comments
970 * in metaslab_group_alloc_update() for more information) and
971 * the allocation throttle is disabled then allow allocations to this
972 * device. However, if the allocation throttle is enabled then
973 * check if we have reached our allocation limit (mg_alloc_queue_depth)
974 * to determine if we should allow allocations to this metaslab group.
975 * If all metaslab groups are no longer considered allocatable
976 * (mc_alloc_groups == 0) or we're trying to allocate the smallest
977 * gang block size then we allow allocations on this metaslab group
978 * regardless of the mg_allocatable or throttle settings.
979 */
980 if (mg->mg_allocatable) {
981 metaslab_group_t *mgp;
982 int64_t qdepth;
983 uint64_t qmax = mg->mg_max_alloc_queue_depth;
984
985 if (!mc->mc_alloc_throttle_enabled)
986 return (B_TRUE);
987
988 /*
989 * If this metaslab group does not have any free space, then
990 * there is no point in looking further.
991 */
992 if (mg->mg_no_free_space)
993 return (B_FALSE);
994
995 qdepth = refcount_count(&mg->mg_alloc_queue_depth);
996
997 /*
998 * If this metaslab group is below its qmax or it's
999 * the only allocatable metasable group, then attempt
1000 * to allocate from it.
1001 */
1002 if (qdepth < qmax || mc->mc_alloc_groups == 1)
1003 return (B_TRUE);
1004 ASSERT3U(mc->mc_alloc_groups, >, 1);
1005
1006 /*
1007 * Since this metaslab group is at or over its qmax, we
1008 * need to determine if there are metaslab groups after this
1009 * one that might be able to handle this allocation. This is
1010 * racy since we can't hold the locks for all metaslab
1011 * groups at the same time when we make this check.
1012 */
1013 for (mgp = mg->mg_next; mgp != rotor; mgp = mgp->mg_next) {
1014 qmax = mgp->mg_max_alloc_queue_depth;
1015
1016 qdepth = refcount_count(&mgp->mg_alloc_queue_depth);
1017
1018 /*
1019 * If there is another metaslab group that
1020 * might be able to handle the allocation, then
1021 * we return false so that we skip this group.
1022 */
1023 if (qdepth < qmax && !mgp->mg_no_free_space)
1024 return (B_FALSE);
1025 }
1026
1027 /*
1028 * We didn't find another group to handle the allocation
1029 * so we can't skip this metaslab group even though
1030 * we are at or over our qmax.
1031 */
1032 return (B_TRUE);
1033
1034 } else if (mc->mc_alloc_groups == 0 || psize == SPA_MINBLOCKSIZE) {
1035 return (B_TRUE);
1036 }
1037 return (B_FALSE);
1038 }
1039
1040 /*
1041 * ==========================================================================
1042 * Range tree callbacks
1043 * ==========================================================================
1044 */
1045
1046 /*
1047 * Comparison function for the private size-ordered tree. Tree is sorted
1048 * by size, larger sizes at the end of the tree.
1049 */
1050 static int
metaslab_rangesize_compare(const void * x1,const void * x2)1051 metaslab_rangesize_compare(const void *x1, const void *x2)
1052 {
1053 const range_seg_t *r1 = x1;
1054 const range_seg_t *r2 = x2;
1055 uint64_t rs_size1 = r1->rs_end - r1->rs_start;
1056 uint64_t rs_size2 = r2->rs_end - r2->rs_start;
1057
1058 if (rs_size1 < rs_size2)
1059 return (-1);
1060 if (rs_size1 > rs_size2)
1061 return (1);
1062
1063 if (r1->rs_start < r2->rs_start)
1064 return (-1);
1065
1066 if (r1->rs_start > r2->rs_start)
1067 return (1);
1068
1069 return (0);
1070 }
1071
1072 /*
1073 * Create any block allocator specific components. The current allocators
1074 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
1075 */
1076 static void
metaslab_rt_create(range_tree_t * rt,void * arg)1077 metaslab_rt_create(range_tree_t *rt, void *arg)
1078 {
1079 metaslab_t *msp = arg;
1080
1081 ASSERT3P(rt->rt_arg, ==, msp);
1082 ASSERT(msp->ms_tree == NULL);
1083
1084 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
1085 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
1086 }
1087
1088 /*
1089 * Destroy the block allocator specific components.
1090 */
1091 static void
metaslab_rt_destroy(range_tree_t * rt,void * arg)1092 metaslab_rt_destroy(range_tree_t *rt, void *arg)
1093 {
1094 metaslab_t *msp = arg;
1095
1096 ASSERT3P(rt->rt_arg, ==, msp);
1097 ASSERT3P(msp->ms_tree, ==, rt);
1098 ASSERT0(avl_numnodes(&msp->ms_size_tree));
1099
1100 avl_destroy(&msp->ms_size_tree);
1101 }
1102
1103 static void
metaslab_rt_add(range_tree_t * rt,range_seg_t * rs,void * arg)1104 metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg)
1105 {
1106 metaslab_t *msp = arg;
1107
1108 ASSERT3P(rt->rt_arg, ==, msp);
1109 ASSERT3P(msp->ms_tree, ==, rt);
1110 VERIFY(!msp->ms_condensing);
1111 avl_add(&msp->ms_size_tree, rs);
1112 }
1113
1114 static void
metaslab_rt_remove(range_tree_t * rt,range_seg_t * rs,void * arg)1115 metaslab_rt_remove(range_tree_t *rt, range_seg_t *rs, void *arg)
1116 {
1117 metaslab_t *msp = arg;
1118
1119 ASSERT3P(rt->rt_arg, ==, msp);
1120 ASSERT3P(msp->ms_tree, ==, rt);
1121 VERIFY(!msp->ms_condensing);
1122 avl_remove(&msp->ms_size_tree, rs);
1123 }
1124
1125 static void
metaslab_rt_vacate(range_tree_t * rt,void * arg)1126 metaslab_rt_vacate(range_tree_t *rt, void *arg)
1127 {
1128 metaslab_t *msp = arg;
1129
1130 ASSERT3P(rt->rt_arg, ==, msp);
1131 ASSERT3P(msp->ms_tree, ==, rt);
1132
1133 /*
1134 * Normally one would walk the tree freeing nodes along the way.
1135 * Since the nodes are shared with the range trees we can avoid
1136 * walking all nodes and just reinitialize the avl tree. The nodes
1137 * will be freed by the range tree, so we don't want to free them here.
1138 */
1139 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
1140 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
1141 }
1142
1143 static range_tree_ops_t metaslab_rt_ops = {
1144 metaslab_rt_create,
1145 metaslab_rt_destroy,
1146 metaslab_rt_add,
1147 metaslab_rt_remove,
1148 metaslab_rt_vacate
1149 };
1150
1151 /*
1152 * ==========================================================================
1153 * Common allocator routines
1154 * ==========================================================================
1155 */
1156
1157 /*
1158 * Return the maximum contiguous segment within the metaslab.
1159 */
1160 uint64_t
metaslab_block_maxsize(metaslab_t * msp)1161 metaslab_block_maxsize(metaslab_t *msp)
1162 {
1163 avl_tree_t *t = &msp->ms_size_tree;
1164 range_seg_t *rs;
1165
1166 if (t == NULL || (rs = avl_last(t)) == NULL)
1167 return (0ULL);
1168
1169 return (rs->rs_end - rs->rs_start);
1170 }
1171
1172 static range_seg_t *
metaslab_block_find(avl_tree_t * t,uint64_t start,uint64_t size)1173 metaslab_block_find(avl_tree_t *t, uint64_t start, uint64_t size)
1174 {
1175 range_seg_t *rs, rsearch;
1176 avl_index_t where;
1177
1178 rsearch.rs_start = start;
1179 rsearch.rs_end = start + size;
1180
1181 rs = avl_find(t, &rsearch, &where);
1182 if (rs == NULL) {
1183 rs = avl_nearest(t, where, AVL_AFTER);
1184 }
1185
1186 return (rs);
1187 }
1188
1189 /*
1190 * This is a helper function that can be used by the allocator to find
1191 * a suitable block to allocate. This will search the specified AVL
1192 * tree looking for a block that matches the specified criteria.
1193 */
1194 static uint64_t
metaslab_block_picker(avl_tree_t * t,uint64_t * cursor,uint64_t size,uint64_t align)1195 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
1196 uint64_t align)
1197 {
1198 range_seg_t *rs = metaslab_block_find(t, *cursor, size);
1199
1200 while (rs != NULL) {
1201 uint64_t offset = P2ROUNDUP(rs->rs_start, align);
1202
1203 if (offset + size <= rs->rs_end) {
1204 *cursor = offset + size;
1205 return (offset);
1206 }
1207 rs = AVL_NEXT(t, rs);
1208 }
1209
1210 /*
1211 * If we know we've searched the whole map (*cursor == 0), give up.
1212 * Otherwise, reset the cursor to the beginning and try again.
1213 */
1214 if (*cursor == 0)
1215 return (-1ULL);
1216
1217 *cursor = 0;
1218 return (metaslab_block_picker(t, cursor, size, align));
1219 }
1220
1221 /*
1222 * ==========================================================================
1223 * The first-fit block allocator
1224 * ==========================================================================
1225 */
1226 static uint64_t
metaslab_ff_alloc(metaslab_t * msp,uint64_t size)1227 metaslab_ff_alloc(metaslab_t *msp, uint64_t size)
1228 {
1229 /*
1230 * Find the largest power of 2 block size that evenly divides the
1231 * requested size. This is used to try to allocate blocks with similar
1232 * alignment from the same area of the metaslab (i.e. same cursor
1233 * bucket) but it does not guarantee that other allocations sizes
1234 * may exist in the same region.
1235 */
1236 uint64_t align = size & -size;
1237 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1238 avl_tree_t *t = &msp->ms_tree->rt_root;
1239
1240 return (metaslab_block_picker(t, cursor, size, align));
1241 }
1242
1243 static metaslab_ops_t metaslab_ff_ops = {
1244 metaslab_ff_alloc
1245 };
1246
1247 /*
1248 * ==========================================================================
1249 * Dynamic block allocator -
1250 * Uses the first fit allocation scheme until space get low and then
1251 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1252 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1253 * ==========================================================================
1254 */
1255 static uint64_t
metaslab_df_alloc(metaslab_t * msp,uint64_t size)1256 metaslab_df_alloc(metaslab_t *msp, uint64_t size)
1257 {
1258 /*
1259 * Find the largest power of 2 block size that evenly divides the
1260 * requested size. This is used to try to allocate blocks with similar
1261 * alignment from the same area of the metaslab (i.e. same cursor
1262 * bucket) but it does not guarantee that other allocations sizes
1263 * may exist in the same region.
1264 */
1265 uint64_t align = size & -size;
1266 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1267 range_tree_t *rt = msp->ms_tree;
1268 avl_tree_t *t = &rt->rt_root;
1269 uint64_t max_size = metaslab_block_maxsize(msp);
1270 int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
1271
1272 ASSERT(MUTEX_HELD(&msp->ms_lock));
1273 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1274
1275 if (max_size < size)
1276 return (-1ULL);
1277
1278 /*
1279 * If we're running low on space switch to using the size
1280 * sorted AVL tree (best-fit).
1281 */
1282 if (max_size < metaslab_df_alloc_threshold ||
1283 free_pct < metaslab_df_free_pct) {
1284 t = &msp->ms_size_tree;
1285 *cursor = 0;
1286 }
1287
1288 return (metaslab_block_picker(t, cursor, size, 1ULL));
1289 }
1290
1291 static metaslab_ops_t metaslab_df_ops = {
1292 metaslab_df_alloc
1293 };
1294
1295 /*
1296 * ==========================================================================
1297 * Cursor fit block allocator -
1298 * Select the largest region in the metaslab, set the cursor to the beginning
1299 * of the range and the cursor_end to the end of the range. As allocations
1300 * are made advance the cursor. Continue allocating from the cursor until
1301 * the range is exhausted and then find a new range.
1302 * ==========================================================================
1303 */
1304 static uint64_t
metaslab_cf_alloc(metaslab_t * msp,uint64_t size)1305 metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
1306 {
1307 range_tree_t *rt = msp->ms_tree;
1308 avl_tree_t *t = &msp->ms_size_tree;
1309 uint64_t *cursor = &msp->ms_lbas[0];
1310 uint64_t *cursor_end = &msp->ms_lbas[1];
1311 uint64_t offset = 0;
1312
1313 ASSERT(MUTEX_HELD(&msp->ms_lock));
1314 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&rt->rt_root));
1315
1316 ASSERT3U(*cursor_end, >=, *cursor);
1317
1318 if ((*cursor + size) > *cursor_end) {
1319 range_seg_t *rs;
1320
1321 rs = avl_last(&msp->ms_size_tree);
1322 if (rs == NULL || (rs->rs_end - rs->rs_start) < size)
1323 return (-1ULL);
1324
1325 *cursor = rs->rs_start;
1326 *cursor_end = rs->rs_end;
1327 }
1328
1329 offset = *cursor;
1330 *cursor += size;
1331
1332 return (offset);
1333 }
1334
1335 static metaslab_ops_t metaslab_cf_ops = {
1336 metaslab_cf_alloc
1337 };
1338
1339 /*
1340 * ==========================================================================
1341 * New dynamic fit allocator -
1342 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1343 * contiguous blocks. If no region is found then just use the largest segment
1344 * that remains.
1345 * ==========================================================================
1346 */
1347
1348 /*
1349 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1350 * to request from the allocator.
1351 */
1352 uint64_t metaslab_ndf_clump_shift = 4;
1353
1354 static uint64_t
metaslab_ndf_alloc(metaslab_t * msp,uint64_t size)1355 metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
1356 {
1357 avl_tree_t *t = &msp->ms_tree->rt_root;
1358 avl_index_t where;
1359 range_seg_t *rs, rsearch;
1360 uint64_t hbit = highbit64(size);
1361 uint64_t *cursor = &msp->ms_lbas[hbit - 1];
1362 uint64_t max_size = metaslab_block_maxsize(msp);
1363
1364 ASSERT(MUTEX_HELD(&msp->ms_lock));
1365 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1366
1367 if (max_size < size)
1368 return (-1ULL);
1369
1370 rsearch.rs_start = *cursor;
1371 rsearch.rs_end = *cursor + size;
1372
1373 rs = avl_find(t, &rsearch, &where);
1374 if (rs == NULL || (rs->rs_end - rs->rs_start) < size) {
1375 t = &msp->ms_size_tree;
1376
1377 rsearch.rs_start = 0;
1378 rsearch.rs_end = MIN(max_size,
1379 1ULL << (hbit + metaslab_ndf_clump_shift));
1380 rs = avl_find(t, &rsearch, &where);
1381 if (rs == NULL)
1382 rs = avl_nearest(t, where, AVL_AFTER);
1383 ASSERT(rs != NULL);
1384 }
1385
1386 if ((rs->rs_end - rs->rs_start) >= size) {
1387 *cursor = rs->rs_start + size;
1388 return (rs->rs_start);
1389 }
1390 return (-1ULL);
1391 }
1392
1393 static metaslab_ops_t metaslab_ndf_ops = {
1394 metaslab_ndf_alloc
1395 };
1396
1397 metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
1398
1399 /*
1400 * ==========================================================================
1401 * Metaslabs
1402 * ==========================================================================
1403 */
1404
1405 /*
1406 * Wait for any in-progress metaslab loads to complete.
1407 */
1408 void
metaslab_load_wait(metaslab_t * msp)1409 metaslab_load_wait(metaslab_t *msp)
1410 {
1411 ASSERT(MUTEX_HELD(&msp->ms_lock));
1412
1413 while (msp->ms_loading) {
1414 ASSERT(!msp->ms_loaded);
1415 cv_wait(&msp->ms_load_cv, &msp->ms_lock);
1416 }
1417 }
1418
1419 int
metaslab_load(metaslab_t * msp)1420 metaslab_load(metaslab_t *msp)
1421 {
1422 int error = 0;
1423 boolean_t success = B_FALSE;
1424
1425 ASSERT(MUTEX_HELD(&msp->ms_lock));
1426 ASSERT(!msp->ms_loaded);
1427 ASSERT(!msp->ms_loading);
1428
1429 msp->ms_loading = B_TRUE;
1430
1431 /*
1432 * If the space map has not been allocated yet, then treat
1433 * all the space in the metaslab as free and add it to the
1434 * ms_tree.
1435 */
1436 if (msp->ms_sm != NULL)
1437 error = space_map_load(msp->ms_sm, msp->ms_tree, SM_FREE);
1438 else
1439 range_tree_add(msp->ms_tree, msp->ms_start, msp->ms_size);
1440
1441 success = (error == 0);
1442 msp->ms_loading = B_FALSE;
1443
1444 if (success) {
1445 ASSERT3P(msp->ms_group, !=, NULL);
1446 msp->ms_loaded = B_TRUE;
1447
1448 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1449 range_tree_walk(msp->ms_defertree[t],
1450 range_tree_remove, msp->ms_tree);
1451 }
1452 msp->ms_max_size = metaslab_block_maxsize(msp);
1453 }
1454 cv_broadcast(&msp->ms_load_cv);
1455 return (error);
1456 }
1457
1458 void
metaslab_unload(metaslab_t * msp)1459 metaslab_unload(metaslab_t *msp)
1460 {
1461 ASSERT(MUTEX_HELD(&msp->ms_lock));
1462 range_tree_vacate(msp->ms_tree, NULL, NULL);
1463 msp->ms_loaded = B_FALSE;
1464 msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
1465 msp->ms_max_size = 0;
1466 }
1467
1468 int
metaslab_init(metaslab_group_t * mg,uint64_t id,uint64_t object,uint64_t txg,metaslab_t ** msp)1469 metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, uint64_t txg,
1470 metaslab_t **msp)
1471 {
1472 vdev_t *vd = mg->mg_vd;
1473 objset_t *mos = vd->vdev_spa->spa_meta_objset;
1474 metaslab_t *ms;
1475 int error;
1476
1477 ms = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
1478 mutex_init(&ms->ms_lock, NULL, MUTEX_DEFAULT, NULL);
1479 cv_init(&ms->ms_load_cv, NULL, CV_DEFAULT, NULL);
1480 ms->ms_id = id;
1481 ms->ms_start = id << vd->vdev_ms_shift;
1482 ms->ms_size = 1ULL << vd->vdev_ms_shift;
1483
1484 /*
1485 * We only open space map objects that already exist. All others
1486 * will be opened when we finally allocate an object for it.
1487 */
1488 if (object != 0) {
1489 error = space_map_open(&ms->ms_sm, mos, object, ms->ms_start,
1490 ms->ms_size, vd->vdev_ashift, &ms->ms_lock);
1491
1492 if (error != 0) {
1493 kmem_free(ms, sizeof (metaslab_t));
1494 return (error);
1495 }
1496
1497 ASSERT(ms->ms_sm != NULL);
1498 }
1499
1500 /*
1501 * We create the main range tree here, but we don't create the
1502 * alloctree and freetree until metaslab_sync_done(). This serves
1503 * two purposes: it allows metaslab_sync_done() to detect the
1504 * addition of new space; and for debugging, it ensures that we'd
1505 * data fault on any attempt to use this metaslab before it's ready.
1506 */
1507 ms->ms_tree = range_tree_create(&metaslab_rt_ops, ms, &ms->ms_lock);
1508 metaslab_group_add(mg, ms);
1509
1510 metaslab_set_fragmentation(ms);
1511
1512 /*
1513 * If we're opening an existing pool (txg == 0) or creating
1514 * a new one (txg == TXG_INITIAL), all space is available now.
1515 * If we're adding space to an existing pool, the new space
1516 * does not become available until after this txg has synced.
1517 * The metaslab's weight will also be initialized when we sync
1518 * out this txg. This ensures that we don't attempt to allocate
1519 * from it before we have initialized it completely.
1520 */
1521 if (txg <= TXG_INITIAL)
1522 metaslab_sync_done(ms, 0);
1523
1524 /*
1525 * If metaslab_debug_load is set and we're initializing a metaslab
1526 * that has an allocated space map object then load the its space
1527 * map so that can verify frees.
1528 */
1529 if (metaslab_debug_load && ms->ms_sm != NULL) {
1530 mutex_enter(&ms->ms_lock);
1531 VERIFY0(metaslab_load(ms));
1532 mutex_exit(&ms->ms_lock);
1533 }
1534
1535 if (txg != 0) {
1536 vdev_dirty(vd, 0, NULL, txg);
1537 vdev_dirty(vd, VDD_METASLAB, ms, txg);
1538 }
1539
1540 *msp = ms;
1541
1542 return (0);
1543 }
1544
1545 void
metaslab_fini(metaslab_t * msp)1546 metaslab_fini(metaslab_t *msp)
1547 {
1548 metaslab_group_t *mg = msp->ms_group;
1549
1550 metaslab_group_remove(mg, msp);
1551
1552 mutex_enter(&msp->ms_lock);
1553 VERIFY(msp->ms_group == NULL);
1554 vdev_space_update(mg->mg_vd, -space_map_allocated(msp->ms_sm),
1555 0, -msp->ms_size);
1556 space_map_close(msp->ms_sm);
1557
1558 metaslab_unload(msp);
1559 range_tree_destroy(msp->ms_tree);
1560
1561 for (int t = 0; t < TXG_SIZE; t++) {
1562 range_tree_destroy(msp->ms_alloctree[t]);
1563 range_tree_destroy(msp->ms_freetree[t]);
1564 }
1565
1566 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1567 range_tree_destroy(msp->ms_defertree[t]);
1568 }
1569
1570 ASSERT0(msp->ms_deferspace);
1571
1572 mutex_exit(&msp->ms_lock);
1573 cv_destroy(&msp->ms_load_cv);
1574 mutex_destroy(&msp->ms_lock);
1575
1576 kmem_free(msp, sizeof (metaslab_t));
1577 }
1578
1579 #define FRAGMENTATION_TABLE_SIZE 17
1580
1581 /*
1582 * This table defines a segment size based fragmentation metric that will
1583 * allow each metaslab to derive its own fragmentation value. This is done
1584 * by calculating the space in each bucket of the spacemap histogram and
1585 * multiplying that by the fragmetation metric in this table. Doing
1586 * this for all buckets and dividing it by the total amount of free
1587 * space in this metaslab (i.e. the total free space in all buckets) gives
1588 * us the fragmentation metric. This means that a high fragmentation metric
1589 * equates to most of the free space being comprised of small segments.
1590 * Conversely, if the metric is low, then most of the free space is in
1591 * large segments. A 10% change in fragmentation equates to approximately
1592 * double the number of segments.
1593 *
1594 * This table defines 0% fragmented space using 16MB segments. Testing has
1595 * shown that segments that are greater than or equal to 16MB do not suffer
1596 * from drastic performance problems. Using this value, we derive the rest
1597 * of the table. Since the fragmentation value is never stored on disk, it
1598 * is possible to change these calculations in the future.
1599 */
1600 int zfs_frag_table[FRAGMENTATION_TABLE_SIZE] = {
1601 100, /* 512B */
1602 100, /* 1K */
1603 98, /* 2K */
1604 95, /* 4K */
1605 90, /* 8K */
1606 80, /* 16K */
1607 70, /* 32K */
1608 60, /* 64K */
1609 50, /* 128K */
1610 40, /* 256K */
1611 30, /* 512K */
1612 20, /* 1M */
1613 15, /* 2M */
1614 10, /* 4M */
1615 5, /* 8M */
1616 0 /* 16M */
1617 };
1618
1619 /*
1620 * Calclate the metaslab's fragmentation metric. A return value
1621 * of ZFS_FRAG_INVALID means that the metaslab has not been upgraded and does
1622 * not support this metric. Otherwise, the return value should be in the
1623 * range [0, 100].
1624 */
1625 static void
metaslab_set_fragmentation(metaslab_t * msp)1626 metaslab_set_fragmentation(metaslab_t *msp)
1627 {
1628 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1629 uint64_t fragmentation = 0;
1630 uint64_t total = 0;
1631 boolean_t feature_enabled = spa_feature_is_enabled(spa,
1632 SPA_FEATURE_SPACEMAP_HISTOGRAM);
1633
1634 if (!feature_enabled) {
1635 msp->ms_fragmentation = ZFS_FRAG_INVALID;
1636 return;
1637 }
1638
1639 /*
1640 * A null space map means that the entire metaslab is free
1641 * and thus is not fragmented.
1642 */
1643 if (msp->ms_sm == NULL) {
1644 msp->ms_fragmentation = 0;
1645 return;
1646 }
1647
1648 /*
1649 * If this metaslab's space map has not been upgraded, flag it
1650 * so that we upgrade next time we encounter it.
1651 */
1652 if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t)) {
1653 uint64_t txg = spa_syncing_txg(spa);
1654 vdev_t *vd = msp->ms_group->mg_vd;
1655
1656 if (spa_writeable(spa)) {
1657 msp->ms_condense_wanted = B_TRUE;
1658 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1659 spa_dbgmsg(spa, "txg %llu, requesting force condense: "
1660 "msp %p, vd %p", txg, msp, vd);
1661 }
1662 msp->ms_fragmentation = ZFS_FRAG_INVALID;
1663 return;
1664 }
1665
1666 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
1667 uint64_t space = 0;
1668 uint8_t shift = msp->ms_sm->sm_shift;
1669
1670 int idx = MIN(shift - SPA_MINBLOCKSHIFT + i,
1671 FRAGMENTATION_TABLE_SIZE - 1);
1672
1673 if (msp->ms_sm->sm_phys->smp_histogram[i] == 0)
1674 continue;
1675
1676 space = msp->ms_sm->sm_phys->smp_histogram[i] << (i + shift);
1677 total += space;
1678
1679 ASSERT3U(idx, <, FRAGMENTATION_TABLE_SIZE);
1680 fragmentation += space * zfs_frag_table[idx];
1681 }
1682
1683 if (total > 0)
1684 fragmentation /= total;
1685 ASSERT3U(fragmentation, <=, 100);
1686
1687 msp->ms_fragmentation = fragmentation;
1688 }
1689
1690 /*
1691 * Compute a weight -- a selection preference value -- for the given metaslab.
1692 * This is based on the amount of free space, the level of fragmentation,
1693 * the LBA range, and whether the metaslab is loaded.
1694 */
1695 static uint64_t
metaslab_space_weight(metaslab_t * msp)1696 metaslab_space_weight(metaslab_t *msp)
1697 {
1698 metaslab_group_t *mg = msp->ms_group;
1699 vdev_t *vd = mg->mg_vd;
1700 uint64_t weight, space;
1701
1702 ASSERT(MUTEX_HELD(&msp->ms_lock));
1703 ASSERT(!vd->vdev_removing);
1704
1705 /*
1706 * The baseline weight is the metaslab's free space.
1707 */
1708 space = msp->ms_size - space_map_allocated(msp->ms_sm);
1709
1710 if (metaslab_fragmentation_factor_enabled &&
1711 msp->ms_fragmentation != ZFS_FRAG_INVALID) {
1712 /*
1713 * Use the fragmentation information to inversely scale
1714 * down the baseline weight. We need to ensure that we
1715 * don't exclude this metaslab completely when it's 100%
1716 * fragmented. To avoid this we reduce the fragmented value
1717 * by 1.
1718 */
1719 space = (space * (100 - (msp->ms_fragmentation - 1))) / 100;
1720
1721 /*
1722 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
1723 * this metaslab again. The fragmentation metric may have
1724 * decreased the space to something smaller than
1725 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
1726 * so that we can consume any remaining space.
1727 */
1728 if (space > 0 && space < SPA_MINBLOCKSIZE)
1729 space = SPA_MINBLOCKSIZE;
1730 }
1731 weight = space;
1732
1733 /*
1734 * Modern disks have uniform bit density and constant angular velocity.
1735 * Therefore, the outer recording zones are faster (higher bandwidth)
1736 * than the inner zones by the ratio of outer to inner track diameter,
1737 * which is typically around 2:1. We account for this by assigning
1738 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1739 * In effect, this means that we'll select the metaslab with the most
1740 * free bandwidth rather than simply the one with the most free space.
1741 */
1742 if (metaslab_lba_weighting_enabled) {
1743 weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count;
1744 ASSERT(weight >= space && weight <= 2 * space);
1745 }
1746
1747 /*
1748 * If this metaslab is one we're actively using, adjust its
1749 * weight to make it preferable to any inactive metaslab so
1750 * we'll polish it off. If the fragmentation on this metaslab
1751 * has exceed our threshold, then don't mark it active.
1752 */
1753 if (msp->ms_loaded && msp->ms_fragmentation != ZFS_FRAG_INVALID &&
1754 msp->ms_fragmentation <= zfs_metaslab_fragmentation_threshold) {
1755 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
1756 }
1757
1758 WEIGHT_SET_SPACEBASED(weight);
1759 return (weight);
1760 }
1761
1762 /*
1763 * Return the weight of the specified metaslab, according to the segment-based
1764 * weighting algorithm. The metaslab must be loaded. This function can
1765 * be called within a sync pass since it relies only on the metaslab's
1766 * range tree which is always accurate when the metaslab is loaded.
1767 */
1768 static uint64_t
metaslab_weight_from_range_tree(metaslab_t * msp)1769 metaslab_weight_from_range_tree(metaslab_t *msp)
1770 {
1771 uint64_t weight = 0;
1772 uint32_t segments = 0;
1773
1774 ASSERT(msp->ms_loaded);
1775
1776 for (int i = RANGE_TREE_HISTOGRAM_SIZE - 1; i >= SPA_MINBLOCKSHIFT;
1777 i--) {
1778 uint8_t shift = msp->ms_group->mg_vd->vdev_ashift;
1779 int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1;
1780
1781 segments <<= 1;
1782 segments += msp->ms_tree->rt_histogram[i];
1783
1784 /*
1785 * The range tree provides more precision than the space map
1786 * and must be downgraded so that all values fit within the
1787 * space map's histogram. This allows us to compare loaded
1788 * vs. unloaded metaslabs to determine which metaslab is
1789 * considered "best".
1790 */
1791 if (i > max_idx)
1792 continue;
1793
1794 if (segments != 0) {
1795 WEIGHT_SET_COUNT(weight, segments);
1796 WEIGHT_SET_INDEX(weight, i);
1797 WEIGHT_SET_ACTIVE(weight, 0);
1798 break;
1799 }
1800 }
1801 return (weight);
1802 }
1803
1804 /*
1805 * Calculate the weight based on the on-disk histogram. This should only
1806 * be called after a sync pass has completely finished since the on-disk
1807 * information is updated in metaslab_sync().
1808 */
1809 static uint64_t
metaslab_weight_from_spacemap(metaslab_t * msp)1810 metaslab_weight_from_spacemap(metaslab_t *msp)
1811 {
1812 uint64_t weight = 0;
1813
1814 for (int i = SPACE_MAP_HISTOGRAM_SIZE - 1; i >= 0; i--) {
1815 if (msp->ms_sm->sm_phys->smp_histogram[i] != 0) {
1816 WEIGHT_SET_COUNT(weight,
1817 msp->ms_sm->sm_phys->smp_histogram[i]);
1818 WEIGHT_SET_INDEX(weight, i +
1819 msp->ms_sm->sm_shift);
1820 WEIGHT_SET_ACTIVE(weight, 0);
1821 break;
1822 }
1823 }
1824 return (weight);
1825 }
1826
1827 /*
1828 * Compute a segment-based weight for the specified metaslab. The weight
1829 * is determined by highest bucket in the histogram. The information
1830 * for the highest bucket is encoded into the weight value.
1831 */
1832 static uint64_t
metaslab_segment_weight(metaslab_t * msp)1833 metaslab_segment_weight(metaslab_t *msp)
1834 {
1835 metaslab_group_t *mg = msp->ms_group;
1836 uint64_t weight = 0;
1837 uint8_t shift = mg->mg_vd->vdev_ashift;
1838
1839 ASSERT(MUTEX_HELD(&msp->ms_lock));
1840
1841 /*
1842 * The metaslab is completely free.
1843 */
1844 if (space_map_allocated(msp->ms_sm) == 0) {
1845 int idx = highbit64(msp->ms_size) - 1;
1846 int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1;
1847
1848 if (idx < max_idx) {
1849 WEIGHT_SET_COUNT(weight, 1ULL);
1850 WEIGHT_SET_INDEX(weight, idx);
1851 } else {
1852 WEIGHT_SET_COUNT(weight, 1ULL << (idx - max_idx));
1853 WEIGHT_SET_INDEX(weight, max_idx);
1854 }
1855 WEIGHT_SET_ACTIVE(weight, 0);
1856 ASSERT(!WEIGHT_IS_SPACEBASED(weight));
1857
1858 return (weight);
1859 }
1860
1861 ASSERT3U(msp->ms_sm->sm_dbuf->db_size, ==, sizeof (space_map_phys_t));
1862
1863 /*
1864 * If the metaslab is fully allocated then just make the weight 0.
1865 */
1866 if (space_map_allocated(msp->ms_sm) == msp->ms_size)
1867 return (0);
1868 /*
1869 * If the metaslab is already loaded, then use the range tree to
1870 * determine the weight. Otherwise, we rely on the space map information
1871 * to generate the weight.
1872 */
1873 if (msp->ms_loaded) {
1874 weight = metaslab_weight_from_range_tree(msp);
1875 } else {
1876 weight = metaslab_weight_from_spacemap(msp);
1877 }
1878
1879 /*
1880 * If the metaslab was active the last time we calculated its weight
1881 * then keep it active. We want to consume the entire region that
1882 * is associated with this weight.
1883 */
1884 if (msp->ms_activation_weight != 0 && weight != 0)
1885 WEIGHT_SET_ACTIVE(weight, WEIGHT_GET_ACTIVE(msp->ms_weight));
1886 return (weight);
1887 }
1888
1889 /*
1890 * Determine if we should attempt to allocate from this metaslab. If the
1891 * metaslab has a maximum size then we can quickly determine if the desired
1892 * allocation size can be satisfied. Otherwise, if we're using segment-based
1893 * weighting then we can determine the maximum allocation that this metaslab
1894 * can accommodate based on the index encoded in the weight. If we're using
1895 * space-based weights then rely on the entire weight (excluding the weight
1896 * type bit).
1897 */
1898 boolean_t
metaslab_should_allocate(metaslab_t * msp,uint64_t asize)1899 metaslab_should_allocate(metaslab_t *msp, uint64_t asize)
1900 {
1901 boolean_t should_allocate;
1902
1903 if (msp->ms_max_size != 0)
1904 return (msp->ms_max_size >= asize);
1905
1906 if (!WEIGHT_IS_SPACEBASED(msp->ms_weight)) {
1907 /*
1908 * The metaslab segment weight indicates segments in the
1909 * range [2^i, 2^(i+1)), where i is the index in the weight.
1910 * Since the asize might be in the middle of the range, we
1911 * should attempt the allocation if asize < 2^(i+1).
1912 */
1913 should_allocate = (asize <
1914 1ULL << (WEIGHT_GET_INDEX(msp->ms_weight) + 1));
1915 } else {
1916 should_allocate = (asize <=
1917 (msp->ms_weight & ~METASLAB_WEIGHT_TYPE));
1918 }
1919 return (should_allocate);
1920 }
1921
1922 static uint64_t
metaslab_weight(metaslab_t * msp)1923 metaslab_weight(metaslab_t *msp)
1924 {
1925 vdev_t *vd = msp->ms_group->mg_vd;
1926 spa_t *spa = vd->vdev_spa;
1927 uint64_t weight;
1928
1929 ASSERT(MUTEX_HELD(&msp->ms_lock));
1930
1931 /*
1932 * This vdev is in the process of being removed so there is nothing
1933 * for us to do here.
1934 */
1935 if (vd->vdev_removing) {
1936 ASSERT0(space_map_allocated(msp->ms_sm));
1937 ASSERT0(vd->vdev_ms_shift);
1938 return (0);
1939 }
1940
1941 metaslab_set_fragmentation(msp);
1942
1943 /*
1944 * Update the maximum size if the metaslab is loaded. This will
1945 * ensure that we get an accurate maximum size if newly freed space
1946 * has been added back into the free tree.
1947 */
1948 if (msp->ms_loaded)
1949 msp->ms_max_size = metaslab_block_maxsize(msp);
1950
1951 /*
1952 * Segment-based weighting requires space map histogram support.
1953 */
1954 if (zfs_metaslab_segment_weight_enabled &&
1955 spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
1956 (msp->ms_sm == NULL || msp->ms_sm->sm_dbuf->db_size ==
1957 sizeof (space_map_phys_t))) {
1958 weight = metaslab_segment_weight(msp);
1959 } else {
1960 weight = metaslab_space_weight(msp);
1961 }
1962 return (weight);
1963 }
1964
1965 static int
metaslab_activate(metaslab_t * msp,uint64_t activation_weight)1966 metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
1967 {
1968 ASSERT(MUTEX_HELD(&msp->ms_lock));
1969
1970 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1971 metaslab_load_wait(msp);
1972 if (!msp->ms_loaded) {
1973 int error = metaslab_load(msp);
1974 if (error) {
1975 metaslab_group_sort(msp->ms_group, msp, 0);
1976 return (error);
1977 }
1978 }
1979
1980 msp->ms_activation_weight = msp->ms_weight;
1981 metaslab_group_sort(msp->ms_group, msp,
1982 msp->ms_weight | activation_weight);
1983 }
1984 ASSERT(msp->ms_loaded);
1985 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
1986
1987 return (0);
1988 }
1989
1990 static void
metaslab_passivate(metaslab_t * msp,uint64_t weight)1991 metaslab_passivate(metaslab_t *msp, uint64_t weight)
1992 {
1993 uint64_t size = weight & ~METASLAB_WEIGHT_TYPE;
1994
1995 /*
1996 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1997 * this metaslab again. In that case, it had better be empty,
1998 * or we would be leaving space on the table.
1999 */
2000 ASSERT(size >= SPA_MINBLOCKSIZE ||
2001 range_tree_space(msp->ms_tree) == 0);
2002 ASSERT0(weight & METASLAB_ACTIVE_MASK);
2003
2004 msp->ms_activation_weight = 0;
2005 metaslab_group_sort(msp->ms_group, msp, weight);
2006 ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
2007 }
2008
2009 /*
2010 * Segment-based metaslabs are activated once and remain active until
2011 * we either fail an allocation attempt (similar to space-based metaslabs)
2012 * or have exhausted the free space in zfs_metaslab_switch_threshold
2013 * buckets since the metaslab was activated. This function checks to see
2014 * if we've exhaused the zfs_metaslab_switch_threshold buckets in the
2015 * metaslab and passivates it proactively. This will allow us to select a
2016 * metaslabs with larger contiguous region if any remaining within this
2017 * metaslab group. If we're in sync pass > 1, then we continue using this
2018 * metaslab so that we don't dirty more block and cause more sync passes.
2019 */
2020 void
metaslab_segment_may_passivate(metaslab_t * msp)2021 metaslab_segment_may_passivate(metaslab_t *msp)
2022 {
2023 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2024
2025 if (WEIGHT_IS_SPACEBASED(msp->ms_weight) || spa_sync_pass(spa) > 1)
2026 return;
2027
2028 /*
2029 * Since we are in the middle of a sync pass, the most accurate
2030 * information that is accessible to us is the in-core range tree
2031 * histogram; calculate the new weight based on that information.
2032 */
2033 uint64_t weight = metaslab_weight_from_range_tree(msp);
2034 int activation_idx = WEIGHT_GET_INDEX(msp->ms_activation_weight);
2035 int current_idx = WEIGHT_GET_INDEX(weight);
2036
2037 if (current_idx <= activation_idx - zfs_metaslab_switch_threshold)
2038 metaslab_passivate(msp, weight);
2039 }
2040
2041 static void
metaslab_preload(void * arg)2042 metaslab_preload(void *arg)
2043 {
2044 metaslab_t *msp = arg;
2045 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2046
2047 ASSERT(!MUTEX_HELD(&msp->ms_group->mg_lock));
2048
2049 mutex_enter(&msp->ms_lock);
2050 metaslab_load_wait(msp);
2051 if (!msp->ms_loaded)
2052 (void) metaslab_load(msp);
2053 msp->ms_selected_txg = spa_syncing_txg(spa);
2054 mutex_exit(&msp->ms_lock);
2055 }
2056
2057 static void
metaslab_group_preload(metaslab_group_t * mg)2058 metaslab_group_preload(metaslab_group_t *mg)
2059 {
2060 spa_t *spa = mg->mg_vd->vdev_spa;
2061 metaslab_t *msp;
2062 avl_tree_t *t = &mg->mg_metaslab_tree;
2063 int m = 0;
2064
2065 if (spa_shutting_down(spa) || !metaslab_preload_enabled) {
2066 taskq_wait(mg->mg_taskq);
2067 return;
2068 }
2069
2070 mutex_enter(&mg->mg_lock);
2071 /*
2072 * Load the next potential metaslabs
2073 */
2074 for (msp = avl_first(t); msp != NULL; msp = AVL_NEXT(t, msp)) {
2075 /*
2076 * We preload only the maximum number of metaslabs specified
2077 * by metaslab_preload_limit. If a metaslab is being forced
2078 * to condense then we preload it too. This will ensure
2079 * that force condensing happens in the next txg.
2080 */
2081 if (++m > metaslab_preload_limit && !msp->ms_condense_wanted) {
2082 continue;
2083 }
2084
2085 VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
2086 msp, TQ_SLEEP) != 0);
2087 }
2088 mutex_exit(&mg->mg_lock);
2089 }
2090
2091 /*
2092 * Determine if the space map's on-disk footprint is past our tolerance
2093 * for inefficiency. We would like to use the following criteria to make
2094 * our decision:
2095 *
2096 * 1. The size of the space map object should not dramatically increase as a
2097 * result of writing out the free space range tree.
2098 *
2099 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
2100 * times the size than the free space range tree representation
2101 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
2102 *
2103 * 3. The on-disk size of the space map should actually decrease.
2104 *
2105 * Checking the first condition is tricky since we don't want to walk
2106 * the entire AVL tree calculating the estimated on-disk size. Instead we
2107 * use the size-ordered range tree in the metaslab and calculate the
2108 * size required to write out the largest segment in our free tree. If the
2109 * size required to represent that segment on disk is larger than the space
2110 * map object then we avoid condensing this map.
2111 *
2112 * To determine the second criterion we use a best-case estimate and assume
2113 * each segment can be represented on-disk as a single 64-bit entry. We refer
2114 * to this best-case estimate as the space map's minimal form.
2115 *
2116 * Unfortunately, we cannot compute the on-disk size of the space map in this
2117 * context because we cannot accurately compute the effects of compression, etc.
2118 * Instead, we apply the heuristic described in the block comment for
2119 * zfs_metaslab_condense_block_threshold - we only condense if the space used
2120 * is greater than a threshold number of blocks.
2121 */
2122 static boolean_t
metaslab_should_condense(metaslab_t * msp)2123 metaslab_should_condense(metaslab_t *msp)
2124 {
2125 space_map_t *sm = msp->ms_sm;
2126 range_seg_t *rs;
2127 uint64_t size, entries, segsz, object_size, optimal_size, record_size;
2128 dmu_object_info_t doi;
2129 uint64_t vdev_blocksize = 1 << msp->ms_group->mg_vd->vdev_ashift;
2130
2131 ASSERT(MUTEX_HELD(&msp->ms_lock));
2132 ASSERT(msp->ms_loaded);
2133
2134 /*
2135 * Use the ms_size_tree range tree, which is ordered by size, to
2136 * obtain the largest segment in the free tree. We always condense
2137 * metaslabs that are empty and metaslabs for which a condense
2138 * request has been made.
2139 */
2140 rs = avl_last(&msp->ms_size_tree);
2141 if (rs == NULL || msp->ms_condense_wanted)
2142 return (B_TRUE);
2143
2144 /*
2145 * Calculate the number of 64-bit entries this segment would
2146 * require when written to disk. If this single segment would be
2147 * larger on-disk than the entire current on-disk structure, then
2148 * clearly condensing will increase the on-disk structure size.
2149 */
2150 size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
2151 entries = size / (MIN(size, SM_RUN_MAX));
2152 segsz = entries * sizeof (uint64_t);
2153
2154 optimal_size = sizeof (uint64_t) * avl_numnodes(&msp->ms_tree->rt_root);
2155 object_size = space_map_length(msp->ms_sm);
2156
2157 dmu_object_info_from_db(sm->sm_dbuf, &doi);
2158 record_size = MAX(doi.doi_data_block_size, vdev_blocksize);
2159
2160 return (segsz <= object_size &&
2161 object_size >= (optimal_size * zfs_condense_pct / 100) &&
2162 object_size > zfs_metaslab_condense_block_threshold * record_size);
2163 }
2164
2165 /*
2166 * Condense the on-disk space map representation to its minimized form.
2167 * The minimized form consists of a small number of allocations followed by
2168 * the entries of the free range tree.
2169 */
2170 static void
metaslab_condense(metaslab_t * msp,uint64_t txg,dmu_tx_t * tx)2171 metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
2172 {
2173 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2174 range_tree_t *freetree = msp->ms_freetree[txg & TXG_MASK];
2175 range_tree_t *condense_tree;
2176 space_map_t *sm = msp->ms_sm;
2177
2178 ASSERT(MUTEX_HELD(&msp->ms_lock));
2179 ASSERT3U(spa_sync_pass(spa), ==, 1);
2180 ASSERT(msp->ms_loaded);
2181
2182
2183 spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, vdev id %llu, "
2184 "spa %s, smp size %llu, segments %lu, forcing condense=%s", txg,
2185 msp->ms_id, msp, msp->ms_group->mg_vd->vdev_id,
2186 msp->ms_group->mg_vd->vdev_spa->spa_name,
2187 space_map_length(msp->ms_sm), avl_numnodes(&msp->ms_tree->rt_root),
2188 msp->ms_condense_wanted ? "TRUE" : "FALSE");
2189
2190 msp->ms_condense_wanted = B_FALSE;
2191
2192 /*
2193 * Create an range tree that is 100% allocated. We remove segments
2194 * that have been freed in this txg, any deferred frees that exist,
2195 * and any allocation in the future. Removing segments should be
2196 * a relatively inexpensive operation since we expect these trees to
2197 * have a small number of nodes.
2198 */
2199 condense_tree = range_tree_create(NULL, NULL, &msp->ms_lock);
2200 range_tree_add(condense_tree, msp->ms_start, msp->ms_size);
2201
2202 /*
2203 * Remove what's been freed in this txg from the condense_tree.
2204 * Since we're in sync_pass 1, we know that all the frees from
2205 * this txg are in the freetree.
2206 */
2207 range_tree_walk(freetree, range_tree_remove, condense_tree);
2208
2209 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2210 range_tree_walk(msp->ms_defertree[t],
2211 range_tree_remove, condense_tree);
2212 }
2213
2214 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
2215 range_tree_walk(msp->ms_alloctree[(txg + t) & TXG_MASK],
2216 range_tree_remove, condense_tree);
2217 }
2218
2219 /*
2220 * We're about to drop the metaslab's lock thus allowing
2221 * other consumers to change it's content. Set the
2222 * metaslab's ms_condensing flag to ensure that
2223 * allocations on this metaslab do not occur while we're
2224 * in the middle of committing it to disk. This is only critical
2225 * for the ms_tree as all other range trees use per txg
2226 * views of their content.
2227 */
2228 msp->ms_condensing = B_TRUE;
2229
2230 mutex_exit(&msp->ms_lock);
2231 space_map_truncate(sm, tx);
2232 mutex_enter(&msp->ms_lock);
2233
2234 /*
2235 * While we would ideally like to create a space map representation
2236 * that consists only of allocation records, doing so can be
2237 * prohibitively expensive because the in-core free tree can be
2238 * large, and therefore computationally expensive to subtract
2239 * from the condense_tree. Instead we sync out two trees, a cheap
2240 * allocation only tree followed by the in-core free tree. While not
2241 * optimal, this is typically close to optimal, and much cheaper to
2242 * compute.
2243 */
2244 space_map_write(sm, condense_tree, SM_ALLOC, tx);
2245 range_tree_vacate(condense_tree, NULL, NULL);
2246 range_tree_destroy(condense_tree);
2247
2248 space_map_write(sm, msp->ms_tree, SM_FREE, tx);
2249 msp->ms_condensing = B_FALSE;
2250 }
2251
2252 /*
2253 * Write a metaslab to disk in the context of the specified transaction group.
2254 */
2255 void
metaslab_sync(metaslab_t * msp,uint64_t txg)2256 metaslab_sync(metaslab_t *msp, uint64_t txg)
2257 {
2258 metaslab_group_t *mg = msp->ms_group;
2259 vdev_t *vd = mg->mg_vd;
2260 spa_t *spa = vd->vdev_spa;
2261 objset_t *mos = spa_meta_objset(spa);
2262 range_tree_t *alloctree = msp->ms_alloctree[txg & TXG_MASK];
2263 range_tree_t **freetree = &msp->ms_freetree[txg & TXG_MASK];
2264 range_tree_t **freed_tree =
2265 &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
2266 dmu_tx_t *tx;
2267 uint64_t object = space_map_object(msp->ms_sm);
2268
2269 ASSERT(!vd->vdev_ishole);
2270
2271 /*
2272 * This metaslab has just been added so there's no work to do now.
2273 */
2274 if (*freetree == NULL) {
2275 ASSERT3P(alloctree, ==, NULL);
2276 return;
2277 }
2278
2279 ASSERT3P(alloctree, !=, NULL);
2280 ASSERT3P(*freetree, !=, NULL);
2281 ASSERT3P(*freed_tree, !=, NULL);
2282
2283 /*
2284 * Normally, we don't want to process a metaslab if there
2285 * are no allocations or frees to perform. However, if the metaslab
2286 * is being forced to condense we need to let it through.
2287 */
2288 if (range_tree_space(alloctree) == 0 &&
2289 range_tree_space(*freetree) == 0 &&
2290 !msp->ms_condense_wanted)
2291 return;
2292
2293 /*
2294 * The only state that can actually be changing concurrently with
2295 * metaslab_sync() is the metaslab's ms_tree. No other thread can
2296 * be modifying this txg's alloctree, freetree, freed_tree, or
2297 * space_map_phys_t. Therefore, we only hold ms_lock to satify
2298 * space map ASSERTs. We drop it whenever we call into the DMU,
2299 * because the DMU can call down to us (e.g. via zio_free()) at
2300 * any time.
2301 */
2302
2303 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2304
2305 if (msp->ms_sm == NULL) {
2306 uint64_t new_object;
2307
2308 new_object = space_map_alloc(mos, tx);
2309 VERIFY3U(new_object, !=, 0);
2310
2311 VERIFY0(space_map_open(&msp->ms_sm, mos, new_object,
2312 msp->ms_start, msp->ms_size, vd->vdev_ashift,
2313 &msp->ms_lock));
2314 ASSERT(msp->ms_sm != NULL);
2315 }
2316
2317 mutex_enter(&msp->ms_lock);
2318
2319 /*
2320 * Note: metaslab_condense() clears the space map's histogram.
2321 * Therefore we must verify and remove this histogram before
2322 * condensing.
2323 */
2324 metaslab_group_histogram_verify(mg);
2325 metaslab_class_histogram_verify(mg->mg_class);
2326 metaslab_group_histogram_remove(mg, msp);
2327
2328 if (msp->ms_loaded && spa_sync_pass(spa) == 1 &&
2329 metaslab_should_condense(msp)) {
2330 metaslab_condense(msp, txg, tx);
2331 } else {
2332 space_map_write(msp->ms_sm, alloctree, SM_ALLOC, tx);
2333 space_map_write(msp->ms_sm, *freetree, SM_FREE, tx);
2334 }
2335
2336 if (msp->ms_loaded) {
2337 /*
2338 * When the space map is loaded, we have an accruate
2339 * histogram in the range tree. This gives us an opportunity
2340 * to bring the space map's histogram up-to-date so we clear
2341 * it first before updating it.
2342 */
2343 space_map_histogram_clear(msp->ms_sm);
2344 space_map_histogram_add(msp->ms_sm, msp->ms_tree, tx);
2345
2346 /*
2347 * Since we've cleared the histogram we need to add back
2348 * any free space that has already been processed, plus
2349 * any deferred space. This allows the on-disk histogram
2350 * to accurately reflect all free space even if some space
2351 * is not yet available for allocation (i.e. deferred).
2352 */
2353 space_map_histogram_add(msp->ms_sm, *freed_tree, tx);
2354
2355 /*
2356 * Add back any deferred free space that has not been
2357 * added back into the in-core free tree yet. This will
2358 * ensure that we don't end up with a space map histogram
2359 * that is completely empty unless the metaslab is fully
2360 * allocated.
2361 */
2362 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2363 space_map_histogram_add(msp->ms_sm,
2364 msp->ms_defertree[t], tx);
2365 }
2366 }
2367
2368 /*
2369 * Always add the free space from this sync pass to the space
2370 * map histogram. We want to make sure that the on-disk histogram
2371 * accounts for all free space. If the space map is not loaded,
2372 * then we will lose some accuracy but will correct it the next
2373 * time we load the space map.
2374 */
2375 space_map_histogram_add(msp->ms_sm, *freetree, tx);
2376
2377 metaslab_group_histogram_add(mg, msp);
2378 metaslab_group_histogram_verify(mg);
2379 metaslab_class_histogram_verify(mg->mg_class);
2380
2381 /*
2382 * For sync pass 1, we avoid traversing this txg's free range tree
2383 * and instead will just swap the pointers for freetree and
2384 * freed_tree. We can safely do this since the freed_tree is
2385 * guaranteed to be empty on the initial pass.
2386 */
2387 if (spa_sync_pass(spa) == 1) {
2388 range_tree_swap(freetree, freed_tree);
2389 } else {
2390 range_tree_vacate(*freetree, range_tree_add, *freed_tree);
2391 }
2392 range_tree_vacate(alloctree, NULL, NULL);
2393
2394 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
2395 ASSERT0(range_tree_space(msp->ms_alloctree[TXG_CLEAN(txg) & TXG_MASK]));
2396 ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
2397
2398 mutex_exit(&msp->ms_lock);
2399
2400 if (object != space_map_object(msp->ms_sm)) {
2401 object = space_map_object(msp->ms_sm);
2402 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
2403 msp->ms_id, sizeof (uint64_t), &object, tx);
2404 }
2405 dmu_tx_commit(tx);
2406 }
2407
2408 /*
2409 * Called after a transaction group has completely synced to mark
2410 * all of the metaslab's free space as usable.
2411 */
2412 void
metaslab_sync_done(metaslab_t * msp,uint64_t txg)2413 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
2414 {
2415 metaslab_group_t *mg = msp->ms_group;
2416 vdev_t *vd = mg->mg_vd;
2417 spa_t *spa = vd->vdev_spa;
2418 range_tree_t **freed_tree;
2419 range_tree_t **defer_tree;
2420 int64_t alloc_delta, defer_delta;
2421 boolean_t defer_allowed = B_TRUE;
2422
2423 ASSERT(!vd->vdev_ishole);
2424
2425 mutex_enter(&msp->ms_lock);
2426
2427 /*
2428 * If this metaslab is just becoming available, initialize its
2429 * alloctrees, freetrees, and defertree and add its capacity to
2430 * the vdev.
2431 */
2432 if (msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK] == NULL) {
2433 for (int t = 0; t < TXG_SIZE; t++) {
2434 ASSERT(msp->ms_alloctree[t] == NULL);
2435 ASSERT(msp->ms_freetree[t] == NULL);
2436
2437 msp->ms_alloctree[t] = range_tree_create(NULL, msp,
2438 &msp->ms_lock);
2439 msp->ms_freetree[t] = range_tree_create(NULL, msp,
2440 &msp->ms_lock);
2441 }
2442
2443 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2444 ASSERT(msp->ms_defertree[t] == NULL);
2445
2446 msp->ms_defertree[t] = range_tree_create(NULL, msp,
2447 &msp->ms_lock);
2448 }
2449
2450 vdev_space_update(vd, 0, 0, msp->ms_size);
2451 }
2452
2453 freed_tree = &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
2454 defer_tree = &msp->ms_defertree[txg % TXG_DEFER_SIZE];
2455
2456 uint64_t free_space = metaslab_class_get_space(spa_normal_class(spa)) -
2457 metaslab_class_get_alloc(spa_normal_class(spa));
2458 if (free_space <= spa_get_slop_space(spa)) {
2459 defer_allowed = B_FALSE;
2460 }
2461
2462 defer_delta = 0;
2463 alloc_delta = space_map_alloc_delta(msp->ms_sm);
2464 if (defer_allowed) {
2465 defer_delta = range_tree_space(*freed_tree) -
2466 range_tree_space(*defer_tree);
2467 } else {
2468 defer_delta -= range_tree_space(*defer_tree);
2469 }
2470
2471 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
2472
2473 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
2474 ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
2475
2476 /*
2477 * If there's a metaslab_load() in progress, wait for it to complete
2478 * so that we have a consistent view of the in-core space map.
2479 */
2480 metaslab_load_wait(msp);
2481
2482 /*
2483 * Move the frees from the defer_tree back to the free
2484 * range tree (if it's loaded). Swap the freed_tree and the
2485 * defer_tree -- this is safe to do because we've just emptied out
2486 * the defer_tree.
2487 */
2488 range_tree_vacate(*defer_tree,
2489 msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
2490 if (defer_allowed) {
2491 range_tree_swap(freed_tree, defer_tree);
2492 } else {
2493 range_tree_vacate(*freed_tree,
2494 msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
2495 }
2496
2497 space_map_update(msp->ms_sm);
2498
2499 msp->ms_deferspace += defer_delta;
2500 ASSERT3S(msp->ms_deferspace, >=, 0);
2501 ASSERT3S(msp->ms_deferspace, <=, msp->ms_size);
2502 if (msp->ms_deferspace != 0) {
2503 /*
2504 * Keep syncing this metaslab until all deferred frees
2505 * are back in circulation.
2506 */
2507 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
2508 }
2509
2510 /*
2511 * Calculate the new weights before unloading any metaslabs.
2512 * This will give us the most accurate weighting.
2513 */
2514 metaslab_group_sort(mg, msp, metaslab_weight(msp));
2515
2516 /*
2517 * If the metaslab is loaded and we've not tried to load or allocate
2518 * from it in 'metaslab_unload_delay' txgs, then unload it.
2519 */
2520 if (msp->ms_loaded &&
2521 msp->ms_selected_txg + metaslab_unload_delay < txg) {
2522 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
2523 VERIFY0(range_tree_space(
2524 msp->ms_alloctree[(txg + t) & TXG_MASK]));
2525 }
2526
2527 if (!metaslab_debug_unload)
2528 metaslab_unload(msp);
2529 }
2530
2531 mutex_exit(&msp->ms_lock);
2532 }
2533
2534 void
metaslab_sync_reassess(metaslab_group_t * mg)2535 metaslab_sync_reassess(metaslab_group_t *mg)
2536 {
2537 metaslab_group_alloc_update(mg);
2538 mg->mg_fragmentation = metaslab_group_fragmentation(mg);
2539
2540 /*
2541 * Preload the next potential metaslabs
2542 */
2543 metaslab_group_preload(mg);
2544 }
2545
2546 static uint64_t
metaslab_distance(metaslab_t * msp,dva_t * dva)2547 metaslab_distance(metaslab_t *msp, dva_t *dva)
2548 {
2549 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
2550 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
2551 uint64_t start = msp->ms_id;
2552
2553 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
2554 return (1ULL << 63);
2555
2556 if (offset < start)
2557 return ((start - offset) << ms_shift);
2558 if (offset > start)
2559 return ((offset - start) << ms_shift);
2560 return (0);
2561 }
2562
2563 /*
2564 * ==========================================================================
2565 * Metaslab allocation tracing facility
2566 * ==========================================================================
2567 */
2568 kstat_t *metaslab_trace_ksp;
2569 kstat_named_t metaslab_trace_over_limit;
2570
2571 void
metaslab_alloc_trace_init(void)2572 metaslab_alloc_trace_init(void)
2573 {
2574 ASSERT(metaslab_alloc_trace_cache == NULL);
2575 metaslab_alloc_trace_cache = kmem_cache_create(
2576 "metaslab_alloc_trace_cache", sizeof (metaslab_alloc_trace_t),
2577 0, NULL, NULL, NULL, NULL, NULL, 0);
2578 metaslab_trace_ksp = kstat_create("zfs", 0, "metaslab_trace_stats",
2579 "misc", KSTAT_TYPE_NAMED, 1, KSTAT_FLAG_VIRTUAL);
2580 if (metaslab_trace_ksp != NULL) {
2581 metaslab_trace_ksp->ks_data = &metaslab_trace_over_limit;
2582 kstat_named_init(&metaslab_trace_over_limit,
2583 "metaslab_trace_over_limit", KSTAT_DATA_UINT64);
2584 kstat_install(metaslab_trace_ksp);
2585 }
2586 }
2587
2588 void
metaslab_alloc_trace_fini(void)2589 metaslab_alloc_trace_fini(void)
2590 {
2591 if (metaslab_trace_ksp != NULL) {
2592 kstat_delete(metaslab_trace_ksp);
2593 metaslab_trace_ksp = NULL;
2594 }
2595 kmem_cache_destroy(metaslab_alloc_trace_cache);
2596 metaslab_alloc_trace_cache = NULL;
2597 }
2598
2599 /*
2600 * Add an allocation trace element to the allocation tracing list.
2601 */
2602 static void
metaslab_trace_add(zio_alloc_list_t * zal,metaslab_group_t * mg,metaslab_t * msp,uint64_t psize,uint32_t dva_id,uint64_t offset)2603 metaslab_trace_add(zio_alloc_list_t *zal, metaslab_group_t *mg,
2604 metaslab_t *msp, uint64_t psize, uint32_t dva_id, uint64_t offset)
2605 {
2606 if (!metaslab_trace_enabled)
2607 return;
2608
2609 /*
2610 * When the tracing list reaches its maximum we remove
2611 * the second element in the list before adding a new one.
2612 * By removing the second element we preserve the original
2613 * entry as a clue to what allocations steps have already been
2614 * performed.
2615 */
2616 if (zal->zal_size == metaslab_trace_max_entries) {
2617 metaslab_alloc_trace_t *mat_next;
2618 #ifdef DEBUG
2619 panic("too many entries in allocation list");
2620 #endif
2621 atomic_inc_64(&metaslab_trace_over_limit.value.ui64);
2622 zal->zal_size--;
2623 mat_next = list_next(&zal->zal_list, list_head(&zal->zal_list));
2624 list_remove(&zal->zal_list, mat_next);
2625 kmem_cache_free(metaslab_alloc_trace_cache, mat_next);
2626 }
2627
2628 metaslab_alloc_trace_t *mat =
2629 kmem_cache_alloc(metaslab_alloc_trace_cache, KM_SLEEP);
2630 list_link_init(&mat->mat_list_node);
2631 mat->mat_mg = mg;
2632 mat->mat_msp = msp;
2633 mat->mat_size = psize;
2634 mat->mat_dva_id = dva_id;
2635 mat->mat_offset = offset;
2636 mat->mat_weight = 0;
2637
2638 if (msp != NULL)
2639 mat->mat_weight = msp->ms_weight;
2640
2641 /*
2642 * The list is part of the zio so locking is not required. Only
2643 * a single thread will perform allocations for a given zio.
2644 */
2645 list_insert_tail(&zal->zal_list, mat);
2646 zal->zal_size++;
2647
2648 ASSERT3U(zal->zal_size, <=, metaslab_trace_max_entries);
2649 }
2650
2651 void
metaslab_trace_init(zio_alloc_list_t * zal)2652 metaslab_trace_init(zio_alloc_list_t *zal)
2653 {
2654 list_create(&zal->zal_list, sizeof (metaslab_alloc_trace_t),
2655 offsetof(metaslab_alloc_trace_t, mat_list_node));
2656 zal->zal_size = 0;
2657 }
2658
2659 void
metaslab_trace_fini(zio_alloc_list_t * zal)2660 metaslab_trace_fini(zio_alloc_list_t *zal)
2661 {
2662 metaslab_alloc_trace_t *mat;
2663
2664 while ((mat = list_remove_head(&zal->zal_list)) != NULL)
2665 kmem_cache_free(metaslab_alloc_trace_cache, mat);
2666 list_destroy(&zal->zal_list);
2667 zal->zal_size = 0;
2668 }
2669
2670 /*
2671 * ==========================================================================
2672 * Metaslab block operations
2673 * ==========================================================================
2674 */
2675
2676 static void
metaslab_group_alloc_increment(spa_t * spa,uint64_t vdev,void * tag,int flags)2677 metaslab_group_alloc_increment(spa_t *spa, uint64_t vdev, void *tag, int flags)
2678 {
2679 if (!(flags & METASLAB_ASYNC_ALLOC) ||
2680 flags & METASLAB_DONT_THROTTLE)
2681 return;
2682
2683 metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
2684 if (!mg->mg_class->mc_alloc_throttle_enabled)
2685 return;
2686
2687 (void) refcount_add(&mg->mg_alloc_queue_depth, tag);
2688 }
2689
2690 void
metaslab_group_alloc_decrement(spa_t * spa,uint64_t vdev,void * tag,int flags)2691 metaslab_group_alloc_decrement(spa_t *spa, uint64_t vdev, void *tag, int flags)
2692 {
2693 if (!(flags & METASLAB_ASYNC_ALLOC) ||
2694 flags & METASLAB_DONT_THROTTLE)
2695 return;
2696
2697 metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
2698 if (!mg->mg_class->mc_alloc_throttle_enabled)
2699 return;
2700
2701 (void) refcount_remove(&mg->mg_alloc_queue_depth, tag);
2702 }
2703
2704 void
metaslab_group_alloc_verify(spa_t * spa,const blkptr_t * bp,void * tag)2705 metaslab_group_alloc_verify(spa_t *spa, const blkptr_t *bp, void *tag)
2706 {
2707 #ifdef ZFS_DEBUG
2708 const dva_t *dva = bp->blk_dva;
2709 int ndvas = BP_GET_NDVAS(bp);
2710
2711 for (int d = 0; d < ndvas; d++) {
2712 uint64_t vdev = DVA_GET_VDEV(&dva[d]);
2713 metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
2714 VERIFY(refcount_not_held(&mg->mg_alloc_queue_depth, tag));
2715 }
2716 #endif
2717 }
2718
2719 static uint64_t
metaslab_block_alloc(metaslab_t * msp,uint64_t size,uint64_t txg)2720 metaslab_block_alloc(metaslab_t *msp, uint64_t size, uint64_t txg)
2721 {
2722 uint64_t start;
2723 range_tree_t *rt = msp->ms_tree;
2724 metaslab_class_t *mc = msp->ms_group->mg_class;
2725
2726 VERIFY(!msp->ms_condensing);
2727
2728 start = mc->mc_ops->msop_alloc(msp, size);
2729 if (start != -1ULL) {
2730 metaslab_group_t *mg = msp->ms_group;
2731 vdev_t *vd = mg->mg_vd;
2732
2733 VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
2734 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2735 VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
2736 range_tree_remove(rt, start, size);
2737
2738 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
2739 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
2740
2741 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], start, size);
2742
2743 /* Track the last successful allocation */
2744 msp->ms_alloc_txg = txg;
2745 metaslab_verify_space(msp, txg);
2746 }
2747
2748 /*
2749 * Now that we've attempted the allocation we need to update the
2750 * metaslab's maximum block size since it may have changed.
2751 */
2752 msp->ms_max_size = metaslab_block_maxsize(msp);
2753 return (start);
2754 }
2755
2756 static uint64_t
metaslab_group_alloc_normal(metaslab_group_t * mg,zio_alloc_list_t * zal,uint64_t asize,uint64_t txg,uint64_t min_distance,dva_t * dva,int d)2757 metaslab_group_alloc_normal(metaslab_group_t *mg, zio_alloc_list_t *zal,
2758 uint64_t asize, uint64_t txg, uint64_t min_distance, dva_t *dva, int d)
2759 {
2760 metaslab_t *msp = NULL;
2761 uint64_t offset = -1ULL;
2762 uint64_t activation_weight;
2763 uint64_t target_distance;
2764 int i;
2765
2766 activation_weight = METASLAB_WEIGHT_PRIMARY;
2767 for (i = 0; i < d; i++) {
2768 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
2769 activation_weight = METASLAB_WEIGHT_SECONDARY;
2770 break;
2771 }
2772 }
2773
2774 metaslab_t *search = kmem_alloc(sizeof (*search), KM_SLEEP);
2775 search->ms_weight = UINT64_MAX;
2776 search->ms_start = 0;
2777 for (;;) {
2778 boolean_t was_active;
2779 avl_tree_t *t = &mg->mg_metaslab_tree;
2780 avl_index_t idx;
2781
2782 mutex_enter(&mg->mg_lock);
2783
2784 /*
2785 * Find the metaslab with the highest weight that is less
2786 * than what we've already tried. In the common case, this
2787 * means that we will examine each metaslab at most once.
2788 * Note that concurrent callers could reorder metaslabs
2789 * by activation/passivation once we have dropped the mg_lock.
2790 * If a metaslab is activated by another thread, and we fail
2791 * to allocate from the metaslab we have selected, we may
2792 * not try the newly-activated metaslab, and instead activate
2793 * another metaslab. This is not optimal, but generally
2794 * does not cause any problems (a possible exception being
2795 * if every metaslab is completely full except for the
2796 * the newly-activated metaslab which we fail to examine).
2797 */
2798 msp = avl_find(t, search, &idx);
2799 if (msp == NULL)
2800 msp = avl_nearest(t, idx, AVL_AFTER);
2801 for (; msp != NULL; msp = AVL_NEXT(t, msp)) {
2802
2803 if (!metaslab_should_allocate(msp, asize)) {
2804 metaslab_trace_add(zal, mg, msp, asize, d,
2805 TRACE_TOO_SMALL);
2806 continue;
2807 }
2808
2809 /*
2810 * If the selected metaslab is condensing, skip it.
2811 */
2812 if (msp->ms_condensing)
2813 continue;
2814
2815 was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
2816 if (activation_weight == METASLAB_WEIGHT_PRIMARY)
2817 break;
2818
2819 target_distance = min_distance +
2820 (space_map_allocated(msp->ms_sm) != 0 ? 0 :
2821 min_distance >> 1);
2822
2823 for (i = 0; i < d; i++) {
2824 if (metaslab_distance(msp, &dva[i]) <
2825 target_distance)
2826 break;
2827 }
2828 if (i == d)
2829 break;
2830 }
2831 mutex_exit(&mg->mg_lock);
2832 if (msp == NULL) {
2833 kmem_free(search, sizeof (*search));
2834 return (-1ULL);
2835 }
2836 search->ms_weight = msp->ms_weight;
2837 search->ms_start = msp->ms_start + 1;
2838
2839 mutex_enter(&msp->ms_lock);
2840
2841 /*
2842 * Ensure that the metaslab we have selected is still
2843 * capable of handling our request. It's possible that
2844 * another thread may have changed the weight while we
2845 * were blocked on the metaslab lock. We check the
2846 * active status first to see if we need to reselect
2847 * a new metaslab.
2848 */
2849 if (was_active && !(msp->ms_weight & METASLAB_ACTIVE_MASK)) {
2850 mutex_exit(&msp->ms_lock);
2851 continue;
2852 }
2853
2854 if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
2855 activation_weight == METASLAB_WEIGHT_PRIMARY) {
2856 metaslab_passivate(msp,
2857 msp->ms_weight & ~METASLAB_ACTIVE_MASK);
2858 mutex_exit(&msp->ms_lock);
2859 continue;
2860 }
2861
2862 if (metaslab_activate(msp, activation_weight) != 0) {
2863 mutex_exit(&msp->ms_lock);
2864 continue;
2865 }
2866 msp->ms_selected_txg = txg;
2867
2868 /*
2869 * Now that we have the lock, recheck to see if we should
2870 * continue to use this metaslab for this allocation. The
2871 * the metaslab is now loaded so metaslab_should_allocate() can
2872 * accurately determine if the allocation attempt should
2873 * proceed.
2874 */
2875 if (!metaslab_should_allocate(msp, asize)) {
2876 /* Passivate this metaslab and select a new one. */
2877 metaslab_trace_add(zal, mg, msp, asize, d,
2878 TRACE_TOO_SMALL);
2879 goto next;
2880 }
2881
2882 /*
2883 * If this metaslab is currently condensing then pick again as
2884 * we can't manipulate this metaslab until it's committed
2885 * to disk.
2886 */
2887 if (msp->ms_condensing) {
2888 metaslab_trace_add(zal, mg, msp, asize, d,
2889 TRACE_CONDENSING);
2890 mutex_exit(&msp->ms_lock);
2891 continue;
2892 }
2893
2894 offset = metaslab_block_alloc(msp, asize, txg);
2895 metaslab_trace_add(zal, mg, msp, asize, d, offset);
2896
2897 if (offset != -1ULL) {
2898 /* Proactively passivate the metaslab, if needed */
2899 metaslab_segment_may_passivate(msp);
2900 break;
2901 }
2902 next:
2903 ASSERT(msp->ms_loaded);
2904
2905 /*
2906 * We were unable to allocate from this metaslab so determine
2907 * a new weight for this metaslab. Now that we have loaded
2908 * the metaslab we can provide a better hint to the metaslab
2909 * selector.
2910 *
2911 * For space-based metaslabs, we use the maximum block size.
2912 * This information is only available when the metaslab
2913 * is loaded and is more accurate than the generic free
2914 * space weight that was calculated by metaslab_weight().
2915 * This information allows us to quickly compare the maximum
2916 * available allocation in the metaslab to the allocation
2917 * size being requested.
2918 *
2919 * For segment-based metaslabs, determine the new weight
2920 * based on the highest bucket in the range tree. We
2921 * explicitly use the loaded segment weight (i.e. the range
2922 * tree histogram) since it contains the space that is
2923 * currently available for allocation and is accurate
2924 * even within a sync pass.
2925 */
2926 if (WEIGHT_IS_SPACEBASED(msp->ms_weight)) {
2927 uint64_t weight = metaslab_block_maxsize(msp);
2928 WEIGHT_SET_SPACEBASED(weight);
2929 metaslab_passivate(msp, weight);
2930 } else {
2931 metaslab_passivate(msp,
2932 metaslab_weight_from_range_tree(msp));
2933 }
2934
2935 /*
2936 * We have just failed an allocation attempt, check
2937 * that metaslab_should_allocate() agrees. Otherwise,
2938 * we may end up in an infinite loop retrying the same
2939 * metaslab.
2940 */
2941 ASSERT(!metaslab_should_allocate(msp, asize));
2942 mutex_exit(&msp->ms_lock);
2943 }
2944 mutex_exit(&msp->ms_lock);
2945 kmem_free(search, sizeof (*search));
2946 return (offset);
2947 }
2948
2949 static uint64_t
metaslab_group_alloc(metaslab_group_t * mg,zio_alloc_list_t * zal,uint64_t asize,uint64_t txg,uint64_t min_distance,dva_t * dva,int d)2950 metaslab_group_alloc(metaslab_group_t *mg, zio_alloc_list_t *zal,
2951 uint64_t asize, uint64_t txg, uint64_t min_distance, dva_t *dva, int d)
2952 {
2953 uint64_t offset;
2954 ASSERT(mg->mg_initialized);
2955
2956 offset = metaslab_group_alloc_normal(mg, zal, asize, txg,
2957 min_distance, dva, d);
2958
2959 mutex_enter(&mg->mg_lock);
2960 if (offset == -1ULL) {
2961 mg->mg_failed_allocations++;
2962 metaslab_trace_add(zal, mg, NULL, asize, d,
2963 TRACE_GROUP_FAILURE);
2964 if (asize == SPA_GANGBLOCKSIZE) {
2965 /*
2966 * This metaslab group was unable to allocate
2967 * the minimum gang block size so it must be out of
2968 * space. We must notify the allocation throttle
2969 * to start skipping allocation attempts to this
2970 * metaslab group until more space becomes available.
2971 * Note: this failure cannot be caused by the
2972 * allocation throttle since the allocation throttle
2973 * is only responsible for skipping devices and
2974 * not failing block allocations.
2975 */
2976 mg->mg_no_free_space = B_TRUE;
2977 }
2978 }
2979 mg->mg_allocations++;
2980 mutex_exit(&mg->mg_lock);
2981 return (offset);
2982 }
2983
2984 /*
2985 * If we have to write a ditto block (i.e. more than one DVA for a given BP)
2986 * on the same vdev as an existing DVA of this BP, then try to allocate it
2987 * at least (vdev_asize / (2 ^ ditto_same_vdev_distance_shift)) away from the
2988 * existing DVAs.
2989 */
2990 int ditto_same_vdev_distance_shift = 3;
2991
2992 /*
2993 * Allocate a block for the specified i/o.
2994 */
2995 static int
metaslab_alloc_dva(spa_t * spa,metaslab_class_t * mc,uint64_t psize,dva_t * dva,int d,dva_t * hintdva,uint64_t txg,int flags,zio_alloc_list_t * zal)2996 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
2997 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags,
2998 zio_alloc_list_t *zal)
2999 {
3000 metaslab_group_t *mg, *rotor;
3001 vdev_t *vd;
3002 boolean_t try_hard = B_FALSE;
3003
3004 ASSERT(!DVA_IS_VALID(&dva[d]));
3005
3006 /*
3007 * For testing, make some blocks above a certain size be gang blocks.
3008 */
3009 if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0) {
3010 metaslab_trace_add(zal, NULL, NULL, psize, d, TRACE_FORCE_GANG);
3011 return (SET_ERROR(ENOSPC));
3012 }
3013
3014 /*
3015 * Start at the rotor and loop through all mgs until we find something.
3016 * Note that there's no locking on mc_rotor or mc_aliquot because
3017 * nothing actually breaks if we miss a few updates -- we just won't
3018 * allocate quite as evenly. It all balances out over time.
3019 *
3020 * If we are doing ditto or log blocks, try to spread them across
3021 * consecutive vdevs. If we're forced to reuse a vdev before we've
3022 * allocated all of our ditto blocks, then try and spread them out on
3023 * that vdev as much as possible. If it turns out to not be possible,
3024 * gradually lower our standards until anything becomes acceptable.
3025 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
3026 * gives us hope of containing our fault domains to something we're
3027 * able to reason about. Otherwise, any two top-level vdev failures
3028 * will guarantee the loss of data. With consecutive allocation,
3029 * only two adjacent top-level vdev failures will result in data loss.
3030 *
3031 * If we are doing gang blocks (hintdva is non-NULL), try to keep
3032 * ourselves on the same vdev as our gang block header. That
3033 * way, we can hope for locality in vdev_cache, plus it makes our
3034 * fault domains something tractable.
3035 */
3036 if (hintdva) {
3037 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
3038
3039 /*
3040 * It's possible the vdev we're using as the hint no
3041 * longer exists (i.e. removed). Consult the rotor when
3042 * all else fails.
3043 */
3044 if (vd != NULL) {
3045 mg = vd->vdev_mg;
3046
3047 if (flags & METASLAB_HINTBP_AVOID &&
3048 mg->mg_next != NULL)
3049 mg = mg->mg_next;
3050 } else {
3051 mg = mc->mc_rotor;
3052 }
3053 } else if (d != 0) {
3054 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
3055 mg = vd->vdev_mg->mg_next;
3056 } else {
3057 mg = mc->mc_rotor;
3058 }
3059
3060 /*
3061 * If the hint put us into the wrong metaslab class, or into a
3062 * metaslab group that has been passivated, just follow the rotor.
3063 */
3064 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
3065 mg = mc->mc_rotor;
3066
3067 rotor = mg;
3068 top:
3069 do {
3070 boolean_t allocatable;
3071
3072 ASSERT(mg->mg_activation_count == 1);
3073 vd = mg->mg_vd;
3074
3075 /*
3076 * Don't allocate from faulted devices.
3077 */
3078 if (try_hard) {
3079 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
3080 allocatable = vdev_allocatable(vd);
3081 spa_config_exit(spa, SCL_ZIO, FTAG);
3082 } else {
3083 allocatable = vdev_allocatable(vd);
3084 }
3085
3086 /*
3087 * Determine if the selected metaslab group is eligible
3088 * for allocations. If we're ganging then don't allow
3089 * this metaslab group to skip allocations since that would
3090 * inadvertently return ENOSPC and suspend the pool
3091 * even though space is still available.
3092 */
3093 if (allocatable && !GANG_ALLOCATION(flags) && !try_hard) {
3094 allocatable = metaslab_group_allocatable(mg, rotor,
3095 psize);
3096 }
3097
3098 if (!allocatable) {
3099 metaslab_trace_add(zal, mg, NULL, psize, d,
3100 TRACE_NOT_ALLOCATABLE);
3101 goto next;
3102 }
3103
3104 ASSERT(mg->mg_initialized);
3105
3106 /*
3107 * Avoid writing single-copy data to a failing,
3108 * non-redundant vdev, unless we've already tried all
3109 * other vdevs.
3110 */
3111 if ((vd->vdev_stat.vs_write_errors > 0 ||
3112 vd->vdev_state < VDEV_STATE_HEALTHY) &&
3113 d == 0 && !try_hard && vd->vdev_children == 0) {
3114 metaslab_trace_add(zal, mg, NULL, psize, d,
3115 TRACE_VDEV_ERROR);
3116 goto next;
3117 }
3118
3119 ASSERT(mg->mg_class == mc);
3120
3121 /*
3122 * If we don't need to try hard, then require that the
3123 * block be 1/8th of the device away from any other DVAs
3124 * in this BP. If we are trying hard, allow any offset
3125 * to be used (distance=0).
3126 */
3127 uint64_t distance = 0;
3128 if (!try_hard) {
3129 distance = vd->vdev_asize >>
3130 ditto_same_vdev_distance_shift;
3131 if (distance <= (1ULL << vd->vdev_ms_shift))
3132 distance = 0;
3133 }
3134
3135 uint64_t asize = vdev_psize_to_asize(vd, psize);
3136 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
3137
3138 uint64_t offset = metaslab_group_alloc(mg, zal, asize, txg,
3139 distance, dva, d);
3140
3141 if (offset != -1ULL) {
3142 /*
3143 * If we've just selected this metaslab group,
3144 * figure out whether the corresponding vdev is
3145 * over- or under-used relative to the pool,
3146 * and set an allocation bias to even it out.
3147 */
3148 if (mc->mc_aliquot == 0 && metaslab_bias_enabled) {
3149 vdev_stat_t *vs = &vd->vdev_stat;
3150 int64_t vu, cu;
3151
3152 vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
3153 cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
3154
3155 /*
3156 * Calculate how much more or less we should
3157 * try to allocate from this device during
3158 * this iteration around the rotor.
3159 * For example, if a device is 80% full
3160 * and the pool is 20% full then we should
3161 * reduce allocations by 60% on this device.
3162 *
3163 * mg_bias = (20 - 80) * 512K / 100 = -307K
3164 *
3165 * This reduces allocations by 307K for this
3166 * iteration.
3167 */
3168 mg->mg_bias = ((cu - vu) *
3169 (int64_t)mg->mg_aliquot) / 100;
3170 } else if (!metaslab_bias_enabled) {
3171 mg->mg_bias = 0;
3172 }
3173
3174 if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
3175 mg->mg_aliquot + mg->mg_bias) {
3176 mc->mc_rotor = mg->mg_next;
3177 mc->mc_aliquot = 0;
3178 }
3179
3180 DVA_SET_VDEV(&dva[d], vd->vdev_id);
3181 DVA_SET_OFFSET(&dva[d], offset);
3182 DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
3183 DVA_SET_ASIZE(&dva[d], asize);
3184
3185 return (0);
3186 }
3187 next:
3188 mc->mc_rotor = mg->mg_next;
3189 mc->mc_aliquot = 0;
3190 } while ((mg = mg->mg_next) != rotor);
3191
3192 /*
3193 * If we haven't tried hard, do so now.
3194 */
3195 if (!try_hard) {
3196 try_hard = B_TRUE;
3197 goto top;
3198 }
3199
3200 bzero(&dva[d], sizeof (dva_t));
3201
3202 metaslab_trace_add(zal, rotor, NULL, psize, d, TRACE_ENOSPC);
3203 return (SET_ERROR(ENOSPC));
3204 }
3205
3206 /*
3207 * Free the block represented by DVA in the context of the specified
3208 * transaction group.
3209 */
3210 static void
metaslab_free_dva(spa_t * spa,const dva_t * dva,uint64_t txg,boolean_t now)3211 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
3212 {
3213 uint64_t vdev = DVA_GET_VDEV(dva);
3214 uint64_t offset = DVA_GET_OFFSET(dva);
3215 uint64_t size = DVA_GET_ASIZE(dva);
3216 vdev_t *vd;
3217 metaslab_t *msp;
3218
3219 ASSERT(DVA_IS_VALID(dva));
3220
3221 if (txg > spa_freeze_txg(spa))
3222 return;
3223
3224 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
3225 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
3226 cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
3227 (u_longlong_t)vdev, (u_longlong_t)offset);
3228 ASSERT(0);
3229 return;
3230 }
3231
3232 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3233
3234 if (DVA_GET_GANG(dva))
3235 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
3236
3237 mutex_enter(&msp->ms_lock);
3238
3239 if (now) {
3240 range_tree_remove(msp->ms_alloctree[txg & TXG_MASK],
3241 offset, size);
3242
3243 VERIFY(!msp->ms_condensing);
3244 VERIFY3U(offset, >=, msp->ms_start);
3245 VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size);
3246 VERIFY3U(range_tree_space(msp->ms_tree) + size, <=,
3247 msp->ms_size);
3248 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
3249 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
3250 range_tree_add(msp->ms_tree, offset, size);
3251 msp->ms_max_size = metaslab_block_maxsize(msp);
3252 } else {
3253 if (range_tree_space(msp->ms_freetree[txg & TXG_MASK]) == 0)
3254 vdev_dirty(vd, VDD_METASLAB, msp, txg);
3255 range_tree_add(msp->ms_freetree[txg & TXG_MASK],
3256 offset, size);
3257 }
3258
3259 mutex_exit(&msp->ms_lock);
3260 }
3261
3262 /*
3263 * Intent log support: upon opening the pool after a crash, notify the SPA
3264 * of blocks that the intent log has allocated for immediate write, but
3265 * which are still considered free by the SPA because the last transaction
3266 * group didn't commit yet.
3267 */
3268 static int
metaslab_claim_dva(spa_t * spa,const dva_t * dva,uint64_t txg)3269 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
3270 {
3271 uint64_t vdev = DVA_GET_VDEV(dva);
3272 uint64_t offset = DVA_GET_OFFSET(dva);
3273 uint64_t size = DVA_GET_ASIZE(dva);
3274 vdev_t *vd;
3275 metaslab_t *msp;
3276 int error = 0;
3277
3278 ASSERT(DVA_IS_VALID(dva));
3279
3280 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
3281 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
3282 return (SET_ERROR(ENXIO));
3283
3284 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3285
3286 if (DVA_GET_GANG(dva))
3287 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
3288
3289 mutex_enter(&msp->ms_lock);
3290
3291 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded)
3292 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
3293
3294 if (error == 0 && !range_tree_contains(msp->ms_tree, offset, size))
3295 error = SET_ERROR(ENOENT);
3296
3297 if (error || txg == 0) { /* txg == 0 indicates dry run */
3298 mutex_exit(&msp->ms_lock);
3299 return (error);
3300 }
3301
3302 VERIFY(!msp->ms_condensing);
3303 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
3304 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
3305 VERIFY3U(range_tree_space(msp->ms_tree) - size, <=, msp->ms_size);
3306 range_tree_remove(msp->ms_tree, offset, size);
3307
3308 if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
3309 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
3310 vdev_dirty(vd, VDD_METASLAB, msp, txg);
3311 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, size);
3312 }
3313
3314 mutex_exit(&msp->ms_lock);
3315
3316 return (0);
3317 }
3318
3319 /*
3320 * Reserve some allocation slots. The reservation system must be called
3321 * before we call into the allocator. If there aren't any available slots
3322 * then the I/O will be throttled until an I/O completes and its slots are
3323 * freed up. The function returns true if it was successful in placing
3324 * the reservation.
3325 */
3326 boolean_t
metaslab_class_throttle_reserve(metaslab_class_t * mc,int slots,zio_t * zio,int flags)3327 metaslab_class_throttle_reserve(metaslab_class_t *mc, int slots, zio_t *zio,
3328 int flags)
3329 {
3330 uint64_t available_slots = 0;
3331 boolean_t slot_reserved = B_FALSE;
3332
3333 ASSERT(mc->mc_alloc_throttle_enabled);
3334 mutex_enter(&mc->mc_lock);
3335
3336 uint64_t reserved_slots = refcount_count(&mc->mc_alloc_slots);
3337 if (reserved_slots < mc->mc_alloc_max_slots)
3338 available_slots = mc->mc_alloc_max_slots - reserved_slots;
3339
3340 if (slots <= available_slots || GANG_ALLOCATION(flags)) {
3341 /*
3342 * We reserve the slots individually so that we can unreserve
3343 * them individually when an I/O completes.
3344 */
3345 for (int d = 0; d < slots; d++) {
3346 reserved_slots = refcount_add(&mc->mc_alloc_slots, zio);
3347 }
3348 zio->io_flags |= ZIO_FLAG_IO_ALLOCATING;
3349 slot_reserved = B_TRUE;
3350 }
3351
3352 mutex_exit(&mc->mc_lock);
3353 return (slot_reserved);
3354 }
3355
3356 void
metaslab_class_throttle_unreserve(metaslab_class_t * mc,int slots,zio_t * zio)3357 metaslab_class_throttle_unreserve(metaslab_class_t *mc, int slots, zio_t *zio)
3358 {
3359 ASSERT(mc->mc_alloc_throttle_enabled);
3360 mutex_enter(&mc->mc_lock);
3361 for (int d = 0; d < slots; d++) {
3362 (void) refcount_remove(&mc->mc_alloc_slots, zio);
3363 }
3364 mutex_exit(&mc->mc_lock);
3365 }
3366
3367 int
metaslab_alloc(spa_t * spa,metaslab_class_t * mc,uint64_t psize,blkptr_t * bp,int ndvas,uint64_t txg,blkptr_t * hintbp,int flags,zio_alloc_list_t * zal,zio_t * zio)3368 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
3369 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags,
3370 zio_alloc_list_t *zal, zio_t *zio)
3371 {
3372 dva_t *dva = bp->blk_dva;
3373 dva_t *hintdva = (hintbp != NULL) ? hintbp->blk_dva : NULL;
3374 int error = 0;
3375
3376 ASSERT(bp->blk_birth == 0);
3377 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
3378
3379 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
3380
3381 if (mc->mc_rotor == NULL) { /* no vdevs in this class */
3382 spa_config_exit(spa, SCL_ALLOC, FTAG);
3383 return (SET_ERROR(ENOSPC));
3384 }
3385
3386 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
3387 ASSERT(BP_GET_NDVAS(bp) == 0);
3388 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
3389 ASSERT3P(zal, !=, NULL);
3390
3391 for (int d = 0; d < ndvas; d++) {
3392 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
3393 txg, flags, zal);
3394 if (error != 0) {
3395 for (d--; d >= 0; d--) {
3396 metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
3397 metaslab_group_alloc_decrement(spa,
3398 DVA_GET_VDEV(&dva[d]), zio, flags);
3399 bzero(&dva[d], sizeof (dva_t));
3400 }
3401 spa_config_exit(spa, SCL_ALLOC, FTAG);
3402 return (error);
3403 } else {
3404 /*
3405 * Update the metaslab group's queue depth
3406 * based on the newly allocated dva.
3407 */
3408 metaslab_group_alloc_increment(spa,
3409 DVA_GET_VDEV(&dva[d]), zio, flags);
3410 }
3411
3412 }
3413 ASSERT(error == 0);
3414 ASSERT(BP_GET_NDVAS(bp) == ndvas);
3415
3416 spa_config_exit(spa, SCL_ALLOC, FTAG);
3417
3418 BP_SET_BIRTH(bp, txg, txg);
3419
3420 return (0);
3421 }
3422
3423 void
metaslab_free(spa_t * spa,const blkptr_t * bp,uint64_t txg,boolean_t now)3424 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
3425 {
3426 const dva_t *dva = bp->blk_dva;
3427 int ndvas = BP_GET_NDVAS(bp);
3428
3429 ASSERT(!BP_IS_HOLE(bp));
3430 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
3431
3432 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
3433
3434 for (int d = 0; d < ndvas; d++)
3435 metaslab_free_dva(spa, &dva[d], txg, now);
3436
3437 spa_config_exit(spa, SCL_FREE, FTAG);
3438 }
3439
3440 int
metaslab_claim(spa_t * spa,const blkptr_t * bp,uint64_t txg)3441 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
3442 {
3443 const dva_t *dva = bp->blk_dva;
3444 int ndvas = BP_GET_NDVAS(bp);
3445 int error = 0;
3446
3447 ASSERT(!BP_IS_HOLE(bp));
3448
3449 if (txg != 0) {
3450 /*
3451 * First do a dry run to make sure all DVAs are claimable,
3452 * so we don't have to unwind from partial failures below.
3453 */
3454 if ((error = metaslab_claim(spa, bp, 0)) != 0)
3455 return (error);
3456 }
3457
3458 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
3459
3460 for (int d = 0; d < ndvas; d++)
3461 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
3462 break;
3463
3464 spa_config_exit(spa, SCL_ALLOC, FTAG);
3465
3466 ASSERT(error == 0 || txg == 0);
3467
3468 return (error);
3469 }
3470
3471 void
metaslab_check_free(spa_t * spa,const blkptr_t * bp)3472 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
3473 {
3474 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
3475 return;
3476
3477 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
3478 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
3479 uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
3480 vdev_t *vd = vdev_lookup_top(spa, vdev);
3481 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
3482 uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
3483 metaslab_t *msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3484
3485 if (msp->ms_loaded)
3486 range_tree_verify(msp->ms_tree, offset, size);
3487
3488 for (int j = 0; j < TXG_SIZE; j++)
3489 range_tree_verify(msp->ms_freetree[j], offset, size);
3490 for (int j = 0; j < TXG_DEFER_SIZE; j++)
3491 range_tree_verify(msp->ms_defertree[j], offset, size);
3492 }
3493 spa_config_exit(spa, SCL_VDEV, FTAG);
3494 }
3495