Print this page
7938 disable LBA weighting on files and SSDs
Reviewed by: Yuri Pankov <yuripv@gmx.com>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/fs/zfs/metaslab.c
+++ new/usr/src/uts/common/fs/zfs/metaslab.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21 /*
22 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 23 * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
24 24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 25 * Copyright (c) 2014 Integros [integros.com]
26 26 */
27 27
28 28 #include <sys/zfs_context.h>
29 29 #include <sys/dmu.h>
30 30 #include <sys/dmu_tx.h>
31 31 #include <sys/space_map.h>
32 32 #include <sys/metaslab_impl.h>
33 33 #include <sys/vdev_impl.h>
34 34 #include <sys/zio.h>
35 35 #include <sys/spa_impl.h>
36 36 #include <sys/zfeature.h>
37 37
38 38 #define GANG_ALLOCATION(flags) \
39 39 ((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER))
40 40
41 41 uint64_t metaslab_aliquot = 512ULL << 10;
42 42 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1; /* force gang blocks */
43 43
44 44 /*
45 45 * The in-core space map representation is more compact than its on-disk form.
46 46 * The zfs_condense_pct determines how much more compact the in-core
47 47 * space map representation must be before we compact it on-disk.
48 48 * Values should be greater than or equal to 100.
49 49 */
50 50 int zfs_condense_pct = 200;
51 51
52 52 /*
53 53 * Condensing a metaslab is not guaranteed to actually reduce the amount of
54 54 * space used on disk. In particular, a space map uses data in increments of
55 55 * MAX(1 << ashift, space_map_blksize), so a metaslab might use the
56 56 * same number of blocks after condensing. Since the goal of condensing is to
57 57 * reduce the number of IOPs required to read the space map, we only want to
58 58 * condense when we can be sure we will reduce the number of blocks used by the
59 59 * space map. Unfortunately, we cannot precisely compute whether or not this is
60 60 * the case in metaslab_should_condense since we are holding ms_lock. Instead,
61 61 * we apply the following heuristic: do not condense a spacemap unless the
62 62 * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
63 63 * blocks.
64 64 */
65 65 int zfs_metaslab_condense_block_threshold = 4;
66 66
67 67 /*
68 68 * The zfs_mg_noalloc_threshold defines which metaslab groups should
69 69 * be eligible for allocation. The value is defined as a percentage of
70 70 * free space. Metaslab groups that have more free space than
71 71 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
72 72 * a metaslab group's free space is less than or equal to the
73 73 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
74 74 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
75 75 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
76 76 * groups are allowed to accept allocations. Gang blocks are always
77 77 * eligible to allocate on any metaslab group. The default value of 0 means
78 78 * no metaslab group will be excluded based on this criterion.
79 79 */
80 80 int zfs_mg_noalloc_threshold = 0;
81 81
82 82 /*
83 83 * Metaslab groups are considered eligible for allocations if their
84 84 * fragmenation metric (measured as a percentage) is less than or equal to
85 85 * zfs_mg_fragmentation_threshold. If a metaslab group exceeds this threshold
86 86 * then it will be skipped unless all metaslab groups within the metaslab
87 87 * class have also crossed this threshold.
88 88 */
89 89 int zfs_mg_fragmentation_threshold = 85;
90 90
91 91 /*
92 92 * Allow metaslabs to keep their active state as long as their fragmentation
93 93 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
94 94 * active metaslab that exceeds this threshold will no longer keep its active
95 95 * status allowing better metaslabs to be selected.
96 96 */
97 97 int zfs_metaslab_fragmentation_threshold = 70;
98 98
99 99 /*
100 100 * When set will load all metaslabs when pool is first opened.
101 101 */
102 102 int metaslab_debug_load = 0;
103 103
104 104 /*
105 105 * When set will prevent metaslabs from being unloaded.
106 106 */
107 107 int metaslab_debug_unload = 0;
108 108
109 109 /*
110 110 * Minimum size which forces the dynamic allocator to change
111 111 * it's allocation strategy. Once the space map cannot satisfy
112 112 * an allocation of this size then it switches to using more
113 113 * aggressive strategy (i.e search by size rather than offset).
114 114 */
115 115 uint64_t metaslab_df_alloc_threshold = SPA_OLD_MAXBLOCKSIZE;
116 116
117 117 /*
118 118 * The minimum free space, in percent, which must be available
119 119 * in a space map to continue allocations in a first-fit fashion.
120 120 * Once the space map's free space drops below this level we dynamically
121 121 * switch to using best-fit allocations.
122 122 */
123 123 int metaslab_df_free_pct = 4;
124 124
125 125 /*
126 126 * A metaslab is considered "free" if it contains a contiguous
127 127 * segment which is greater than metaslab_min_alloc_size.
128 128 */
129 129 uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
130 130
131 131 /*
132 132 * Percentage of all cpus that can be used by the metaslab taskq.
133 133 */
134 134 int metaslab_load_pct = 50;
135 135
136 136 /*
137 137 * Determines how many txgs a metaslab may remain loaded without having any
138 138 * allocations from it. As long as a metaslab continues to be used we will
139 139 * keep it loaded.
140 140 */
141 141 int metaslab_unload_delay = TXG_SIZE * 2;
142 142
143 143 /*
144 144 * Max number of metaslabs per group to preload.
145 145 */
146 146 int metaslab_preload_limit = SPA_DVAS_PER_BP;
147 147
148 148 /*
149 149 * Enable/disable preloading of metaslab.
150 150 */
151 151 boolean_t metaslab_preload_enabled = B_TRUE;
152 152
153 153 /*
154 154 * Enable/disable fragmentation weighting on metaslabs.
155 155 */
156 156 boolean_t metaslab_fragmentation_factor_enabled = B_TRUE;
157 157
158 158 /*
159 159 * Enable/disable lba weighting (i.e. outer tracks are given preference).
160 160 */
161 161 boolean_t metaslab_lba_weighting_enabled = B_TRUE;
162 162
163 163 /*
164 164 * Enable/disable metaslab group biasing.
165 165 */
166 166 boolean_t metaslab_bias_enabled = B_TRUE;
167 167
168 168 /*
169 169 * Enable/disable segment-based metaslab selection.
170 170 */
171 171 boolean_t zfs_metaslab_segment_weight_enabled = B_TRUE;
172 172
173 173 /*
174 174 * When using segment-based metaslab selection, we will continue
175 175 * allocating from the active metaslab until we have exhausted
176 176 * zfs_metaslab_switch_threshold of its buckets.
177 177 */
178 178 int zfs_metaslab_switch_threshold = 2;
179 179
180 180 /*
181 181 * Internal switch to enable/disable the metaslab allocation tracing
182 182 * facility.
183 183 */
184 184 boolean_t metaslab_trace_enabled = B_TRUE;
185 185
186 186 /*
187 187 * Maximum entries that the metaslab allocation tracing facility will keep
188 188 * in a given list when running in non-debug mode. We limit the number
189 189 * of entries in non-debug mode to prevent us from using up too much memory.
190 190 * The limit should be sufficiently large that we don't expect any allocation
191 191 * to every exceed this value. In debug mode, the system will panic if this
192 192 * limit is ever reached allowing for further investigation.
193 193 */
194 194 uint64_t metaslab_trace_max_entries = 5000;
195 195
196 196 static uint64_t metaslab_weight(metaslab_t *);
197 197 static void metaslab_set_fragmentation(metaslab_t *);
198 198
199 199 kmem_cache_t *metaslab_alloc_trace_cache;
200 200
201 201 /*
202 202 * ==========================================================================
203 203 * Metaslab classes
204 204 * ==========================================================================
205 205 */
206 206 metaslab_class_t *
207 207 metaslab_class_create(spa_t *spa, metaslab_ops_t *ops)
208 208 {
209 209 metaslab_class_t *mc;
210 210
211 211 mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
212 212
213 213 mc->mc_spa = spa;
214 214 mc->mc_rotor = NULL;
215 215 mc->mc_ops = ops;
216 216 mutex_init(&mc->mc_lock, NULL, MUTEX_DEFAULT, NULL);
217 217 refcount_create_tracked(&mc->mc_alloc_slots);
218 218
219 219 return (mc);
220 220 }
221 221
222 222 void
223 223 metaslab_class_destroy(metaslab_class_t *mc)
224 224 {
225 225 ASSERT(mc->mc_rotor == NULL);
226 226 ASSERT(mc->mc_alloc == 0);
227 227 ASSERT(mc->mc_deferred == 0);
228 228 ASSERT(mc->mc_space == 0);
229 229 ASSERT(mc->mc_dspace == 0);
230 230
231 231 refcount_destroy(&mc->mc_alloc_slots);
232 232 mutex_destroy(&mc->mc_lock);
233 233 kmem_free(mc, sizeof (metaslab_class_t));
234 234 }
235 235
236 236 int
237 237 metaslab_class_validate(metaslab_class_t *mc)
238 238 {
239 239 metaslab_group_t *mg;
240 240 vdev_t *vd;
241 241
242 242 /*
243 243 * Must hold one of the spa_config locks.
244 244 */
245 245 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
246 246 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
247 247
248 248 if ((mg = mc->mc_rotor) == NULL)
249 249 return (0);
250 250
251 251 do {
252 252 vd = mg->mg_vd;
253 253 ASSERT(vd->vdev_mg != NULL);
254 254 ASSERT3P(vd->vdev_top, ==, vd);
255 255 ASSERT3P(mg->mg_class, ==, mc);
256 256 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
257 257 } while ((mg = mg->mg_next) != mc->mc_rotor);
258 258
259 259 return (0);
260 260 }
261 261
262 262 void
263 263 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
264 264 int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
265 265 {
266 266 atomic_add_64(&mc->mc_alloc, alloc_delta);
267 267 atomic_add_64(&mc->mc_deferred, defer_delta);
268 268 atomic_add_64(&mc->mc_space, space_delta);
269 269 atomic_add_64(&mc->mc_dspace, dspace_delta);
270 270 }
271 271
272 272 uint64_t
273 273 metaslab_class_get_alloc(metaslab_class_t *mc)
274 274 {
275 275 return (mc->mc_alloc);
276 276 }
277 277
278 278 uint64_t
279 279 metaslab_class_get_deferred(metaslab_class_t *mc)
280 280 {
281 281 return (mc->mc_deferred);
282 282 }
283 283
284 284 uint64_t
285 285 metaslab_class_get_space(metaslab_class_t *mc)
286 286 {
287 287 return (mc->mc_space);
288 288 }
289 289
290 290 uint64_t
291 291 metaslab_class_get_dspace(metaslab_class_t *mc)
292 292 {
293 293 return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
294 294 }
295 295
296 296 void
297 297 metaslab_class_histogram_verify(metaslab_class_t *mc)
298 298 {
299 299 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
300 300 uint64_t *mc_hist;
301 301 int i;
302 302
303 303 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
304 304 return;
305 305
306 306 mc_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
307 307 KM_SLEEP);
308 308
309 309 for (int c = 0; c < rvd->vdev_children; c++) {
310 310 vdev_t *tvd = rvd->vdev_child[c];
311 311 metaslab_group_t *mg = tvd->vdev_mg;
312 312
313 313 /*
314 314 * Skip any holes, uninitialized top-levels, or
315 315 * vdevs that are not in this metalab class.
316 316 */
317 317 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
318 318 mg->mg_class != mc) {
319 319 continue;
320 320 }
321 321
322 322 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
323 323 mc_hist[i] += mg->mg_histogram[i];
324 324 }
325 325
326 326 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
327 327 VERIFY3U(mc_hist[i], ==, mc->mc_histogram[i]);
328 328
329 329 kmem_free(mc_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
330 330 }
331 331
332 332 /*
333 333 * Calculate the metaslab class's fragmentation metric. The metric
334 334 * is weighted based on the space contribution of each metaslab group.
335 335 * The return value will be a number between 0 and 100 (inclusive), or
336 336 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
337 337 * zfs_frag_table for more information about the metric.
338 338 */
339 339 uint64_t
340 340 metaslab_class_fragmentation(metaslab_class_t *mc)
341 341 {
342 342 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
343 343 uint64_t fragmentation = 0;
344 344
345 345 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
346 346
347 347 for (int c = 0; c < rvd->vdev_children; c++) {
348 348 vdev_t *tvd = rvd->vdev_child[c];
349 349 metaslab_group_t *mg = tvd->vdev_mg;
350 350
351 351 /*
352 352 * Skip any holes, uninitialized top-levels, or
353 353 * vdevs that are not in this metalab class.
354 354 */
355 355 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
356 356 mg->mg_class != mc) {
357 357 continue;
358 358 }
359 359
360 360 /*
361 361 * If a metaslab group does not contain a fragmentation
362 362 * metric then just bail out.
363 363 */
364 364 if (mg->mg_fragmentation == ZFS_FRAG_INVALID) {
365 365 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
366 366 return (ZFS_FRAG_INVALID);
367 367 }
368 368
369 369 /*
370 370 * Determine how much this metaslab_group is contributing
371 371 * to the overall pool fragmentation metric.
372 372 */
373 373 fragmentation += mg->mg_fragmentation *
374 374 metaslab_group_get_space(mg);
375 375 }
376 376 fragmentation /= metaslab_class_get_space(mc);
377 377
378 378 ASSERT3U(fragmentation, <=, 100);
379 379 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
380 380 return (fragmentation);
381 381 }
382 382
383 383 /*
384 384 * Calculate the amount of expandable space that is available in
385 385 * this metaslab class. If a device is expanded then its expandable
386 386 * space will be the amount of allocatable space that is currently not
387 387 * part of this metaslab class.
388 388 */
389 389 uint64_t
390 390 metaslab_class_expandable_space(metaslab_class_t *mc)
391 391 {
392 392 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
393 393 uint64_t space = 0;
394 394
395 395 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
396 396 for (int c = 0; c < rvd->vdev_children; c++) {
397 397 uint64_t tspace;
398 398 vdev_t *tvd = rvd->vdev_child[c];
399 399 metaslab_group_t *mg = tvd->vdev_mg;
400 400
401 401 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
402 402 mg->mg_class != mc) {
403 403 continue;
404 404 }
405 405
406 406 /*
407 407 * Calculate if we have enough space to add additional
408 408 * metaslabs. We report the expandable space in terms
409 409 * of the metaslab size since that's the unit of expansion.
410 410 * Adjust by efi system partition size.
411 411 */
412 412 tspace = tvd->vdev_max_asize - tvd->vdev_asize;
413 413 if (tspace > mc->mc_spa->spa_bootsize) {
414 414 tspace -= mc->mc_spa->spa_bootsize;
415 415 }
416 416 space += P2ALIGN(tspace, 1ULL << tvd->vdev_ms_shift);
417 417 }
418 418 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
419 419 return (space);
420 420 }
421 421
422 422 static int
423 423 metaslab_compare(const void *x1, const void *x2)
424 424 {
425 425 const metaslab_t *m1 = x1;
426 426 const metaslab_t *m2 = x2;
427 427
428 428 if (m1->ms_weight < m2->ms_weight)
429 429 return (1);
430 430 if (m1->ms_weight > m2->ms_weight)
431 431 return (-1);
432 432
433 433 /*
434 434 * If the weights are identical, use the offset to force uniqueness.
435 435 */
436 436 if (m1->ms_start < m2->ms_start)
437 437 return (-1);
438 438 if (m1->ms_start > m2->ms_start)
439 439 return (1);
440 440
441 441 ASSERT3P(m1, ==, m2);
442 442
443 443 return (0);
444 444 }
445 445
446 446 /*
447 447 * Verify that the space accounting on disk matches the in-core range_trees.
448 448 */
449 449 void
450 450 metaslab_verify_space(metaslab_t *msp, uint64_t txg)
451 451 {
452 452 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
453 453 uint64_t allocated = 0;
454 454 uint64_t sm_free_space, msp_free_space;
455 455
456 456 ASSERT(MUTEX_HELD(&msp->ms_lock));
457 457
458 458 if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0)
459 459 return;
460 460
461 461 /*
462 462 * We can only verify the metaslab space when we're called
463 463 * from syncing context with a loaded metaslab that has an allocated
464 464 * space map. Calling this in non-syncing context does not
465 465 * provide a consistent view of the metaslab since we're performing
466 466 * allocations in the future.
467 467 */
468 468 if (txg != spa_syncing_txg(spa) || msp->ms_sm == NULL ||
469 469 !msp->ms_loaded)
470 470 return;
471 471
472 472 sm_free_space = msp->ms_size - space_map_allocated(msp->ms_sm) -
473 473 space_map_alloc_delta(msp->ms_sm);
474 474
475 475 /*
476 476 * Account for future allocations since we would have already
477 477 * deducted that space from the ms_freetree.
478 478 */
479 479 for (int t = 0; t < TXG_CONCURRENT_STATES; t++) {
480 480 allocated +=
481 481 range_tree_space(msp->ms_alloctree[(txg + t) & TXG_MASK]);
482 482 }
483 483
484 484 msp_free_space = range_tree_space(msp->ms_tree) + allocated +
485 485 msp->ms_deferspace + range_tree_space(msp->ms_freedtree);
486 486
487 487 VERIFY3U(sm_free_space, ==, msp_free_space);
488 488 }
489 489
490 490 /*
491 491 * ==========================================================================
492 492 * Metaslab groups
493 493 * ==========================================================================
494 494 */
495 495 /*
496 496 * Update the allocatable flag and the metaslab group's capacity.
497 497 * The allocatable flag is set to true if the capacity is below
498 498 * the zfs_mg_noalloc_threshold or has a fragmentation value that is
499 499 * greater than zfs_mg_fragmentation_threshold. If a metaslab group
500 500 * transitions from allocatable to non-allocatable or vice versa then the
501 501 * metaslab group's class is updated to reflect the transition.
502 502 */
503 503 static void
504 504 metaslab_group_alloc_update(metaslab_group_t *mg)
505 505 {
506 506 vdev_t *vd = mg->mg_vd;
507 507 metaslab_class_t *mc = mg->mg_class;
508 508 vdev_stat_t *vs = &vd->vdev_stat;
509 509 boolean_t was_allocatable;
510 510 boolean_t was_initialized;
511 511
512 512 ASSERT(vd == vd->vdev_top);
513 513
514 514 mutex_enter(&mg->mg_lock);
515 515 was_allocatable = mg->mg_allocatable;
516 516 was_initialized = mg->mg_initialized;
517 517
518 518 mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
519 519 (vs->vs_space + 1);
520 520
521 521 mutex_enter(&mc->mc_lock);
522 522
523 523 /*
524 524 * If the metaslab group was just added then it won't
525 525 * have any space until we finish syncing out this txg.
526 526 * At that point we will consider it initialized and available
527 527 * for allocations. We also don't consider non-activated
528 528 * metaslab groups (e.g. vdevs that are in the middle of being removed)
529 529 * to be initialized, because they can't be used for allocation.
530 530 */
531 531 mg->mg_initialized = metaslab_group_initialized(mg);
532 532 if (!was_initialized && mg->mg_initialized) {
533 533 mc->mc_groups++;
534 534 } else if (was_initialized && !mg->mg_initialized) {
535 535 ASSERT3U(mc->mc_groups, >, 0);
536 536 mc->mc_groups--;
537 537 }
538 538 if (mg->mg_initialized)
539 539 mg->mg_no_free_space = B_FALSE;
540 540
541 541 /*
542 542 * A metaslab group is considered allocatable if it has plenty
543 543 * of free space or is not heavily fragmented. We only take
544 544 * fragmentation into account if the metaslab group has a valid
545 545 * fragmentation metric (i.e. a value between 0 and 100).
546 546 */
547 547 mg->mg_allocatable = (mg->mg_activation_count > 0 &&
548 548 mg->mg_free_capacity > zfs_mg_noalloc_threshold &&
549 549 (mg->mg_fragmentation == ZFS_FRAG_INVALID ||
550 550 mg->mg_fragmentation <= zfs_mg_fragmentation_threshold));
551 551
552 552 /*
553 553 * The mc_alloc_groups maintains a count of the number of
554 554 * groups in this metaslab class that are still above the
555 555 * zfs_mg_noalloc_threshold. This is used by the allocating
556 556 * threads to determine if they should avoid allocations to
557 557 * a given group. The allocator will avoid allocations to a group
558 558 * if that group has reached or is below the zfs_mg_noalloc_threshold
559 559 * and there are still other groups that are above the threshold.
560 560 * When a group transitions from allocatable to non-allocatable or
561 561 * vice versa we update the metaslab class to reflect that change.
562 562 * When the mc_alloc_groups value drops to 0 that means that all
563 563 * groups have reached the zfs_mg_noalloc_threshold making all groups
564 564 * eligible for allocations. This effectively means that all devices
565 565 * are balanced again.
566 566 */
567 567 if (was_allocatable && !mg->mg_allocatable)
568 568 mc->mc_alloc_groups--;
569 569 else if (!was_allocatable && mg->mg_allocatable)
570 570 mc->mc_alloc_groups++;
571 571 mutex_exit(&mc->mc_lock);
572 572
573 573 mutex_exit(&mg->mg_lock);
574 574 }
575 575
576 576 metaslab_group_t *
577 577 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
578 578 {
579 579 metaslab_group_t *mg;
580 580
581 581 mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
582 582 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
583 583 avl_create(&mg->mg_metaslab_tree, metaslab_compare,
584 584 sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
585 585 mg->mg_vd = vd;
586 586 mg->mg_class = mc;
587 587 mg->mg_activation_count = 0;
588 588 mg->mg_initialized = B_FALSE;
589 589 mg->mg_no_free_space = B_TRUE;
590 590 refcount_create_tracked(&mg->mg_alloc_queue_depth);
591 591
592 592 mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct,
593 593 minclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT);
594 594
595 595 return (mg);
596 596 }
597 597
598 598 void
599 599 metaslab_group_destroy(metaslab_group_t *mg)
600 600 {
601 601 ASSERT(mg->mg_prev == NULL);
602 602 ASSERT(mg->mg_next == NULL);
603 603 /*
604 604 * We may have gone below zero with the activation count
605 605 * either because we never activated in the first place or
606 606 * because we're done, and possibly removing the vdev.
607 607 */
608 608 ASSERT(mg->mg_activation_count <= 0);
609 609
610 610 taskq_destroy(mg->mg_taskq);
611 611 avl_destroy(&mg->mg_metaslab_tree);
612 612 mutex_destroy(&mg->mg_lock);
613 613 refcount_destroy(&mg->mg_alloc_queue_depth);
614 614 kmem_free(mg, sizeof (metaslab_group_t));
615 615 }
616 616
617 617 void
618 618 metaslab_group_activate(metaslab_group_t *mg)
619 619 {
620 620 metaslab_class_t *mc = mg->mg_class;
621 621 metaslab_group_t *mgprev, *mgnext;
622 622
623 623 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
624 624
625 625 ASSERT(mc->mc_rotor != mg);
626 626 ASSERT(mg->mg_prev == NULL);
627 627 ASSERT(mg->mg_next == NULL);
628 628 ASSERT(mg->mg_activation_count <= 0);
629 629
630 630 if (++mg->mg_activation_count <= 0)
631 631 return;
632 632
633 633 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
634 634 metaslab_group_alloc_update(mg);
635 635
636 636 if ((mgprev = mc->mc_rotor) == NULL) {
637 637 mg->mg_prev = mg;
638 638 mg->mg_next = mg;
639 639 } else {
640 640 mgnext = mgprev->mg_next;
641 641 mg->mg_prev = mgprev;
642 642 mg->mg_next = mgnext;
643 643 mgprev->mg_next = mg;
644 644 mgnext->mg_prev = mg;
645 645 }
646 646 mc->mc_rotor = mg;
647 647 }
648 648
649 649 void
650 650 metaslab_group_passivate(metaslab_group_t *mg)
651 651 {
652 652 metaslab_class_t *mc = mg->mg_class;
653 653 metaslab_group_t *mgprev, *mgnext;
654 654
655 655 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
656 656
657 657 if (--mg->mg_activation_count != 0) {
658 658 ASSERT(mc->mc_rotor != mg);
659 659 ASSERT(mg->mg_prev == NULL);
660 660 ASSERT(mg->mg_next == NULL);
661 661 ASSERT(mg->mg_activation_count < 0);
662 662 return;
663 663 }
664 664
665 665 taskq_wait(mg->mg_taskq);
666 666 metaslab_group_alloc_update(mg);
667 667
668 668 mgprev = mg->mg_prev;
669 669 mgnext = mg->mg_next;
670 670
671 671 if (mg == mgnext) {
672 672 mc->mc_rotor = NULL;
673 673 } else {
674 674 mc->mc_rotor = mgnext;
675 675 mgprev->mg_next = mgnext;
676 676 mgnext->mg_prev = mgprev;
677 677 }
678 678
679 679 mg->mg_prev = NULL;
680 680 mg->mg_next = NULL;
681 681 }
682 682
683 683 boolean_t
684 684 metaslab_group_initialized(metaslab_group_t *mg)
685 685 {
686 686 vdev_t *vd = mg->mg_vd;
687 687 vdev_stat_t *vs = &vd->vdev_stat;
688 688
689 689 return (vs->vs_space != 0 && mg->mg_activation_count > 0);
690 690 }
691 691
692 692 uint64_t
693 693 metaslab_group_get_space(metaslab_group_t *mg)
694 694 {
695 695 return ((1ULL << mg->mg_vd->vdev_ms_shift) * mg->mg_vd->vdev_ms_count);
696 696 }
697 697
698 698 void
699 699 metaslab_group_histogram_verify(metaslab_group_t *mg)
700 700 {
701 701 uint64_t *mg_hist;
702 702 vdev_t *vd = mg->mg_vd;
703 703 uint64_t ashift = vd->vdev_ashift;
704 704 int i;
705 705
706 706 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
707 707 return;
708 708
709 709 mg_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
710 710 KM_SLEEP);
711 711
712 712 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE, >=,
713 713 SPACE_MAP_HISTOGRAM_SIZE + ashift);
714 714
715 715 for (int m = 0; m < vd->vdev_ms_count; m++) {
716 716 metaslab_t *msp = vd->vdev_ms[m];
717 717
718 718 if (msp->ms_sm == NULL)
719 719 continue;
720 720
721 721 for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++)
722 722 mg_hist[i + ashift] +=
723 723 msp->ms_sm->sm_phys->smp_histogram[i];
724 724 }
725 725
726 726 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i ++)
727 727 VERIFY3U(mg_hist[i], ==, mg->mg_histogram[i]);
728 728
729 729 kmem_free(mg_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
730 730 }
731 731
732 732 static void
733 733 metaslab_group_histogram_add(metaslab_group_t *mg, metaslab_t *msp)
734 734 {
735 735 metaslab_class_t *mc = mg->mg_class;
736 736 uint64_t ashift = mg->mg_vd->vdev_ashift;
737 737
738 738 ASSERT(MUTEX_HELD(&msp->ms_lock));
739 739 if (msp->ms_sm == NULL)
740 740 return;
741 741
742 742 mutex_enter(&mg->mg_lock);
743 743 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
744 744 mg->mg_histogram[i + ashift] +=
745 745 msp->ms_sm->sm_phys->smp_histogram[i];
746 746 mc->mc_histogram[i + ashift] +=
747 747 msp->ms_sm->sm_phys->smp_histogram[i];
748 748 }
749 749 mutex_exit(&mg->mg_lock);
750 750 }
751 751
752 752 void
753 753 metaslab_group_histogram_remove(metaslab_group_t *mg, metaslab_t *msp)
754 754 {
755 755 metaslab_class_t *mc = mg->mg_class;
756 756 uint64_t ashift = mg->mg_vd->vdev_ashift;
757 757
758 758 ASSERT(MUTEX_HELD(&msp->ms_lock));
759 759 if (msp->ms_sm == NULL)
760 760 return;
761 761
762 762 mutex_enter(&mg->mg_lock);
763 763 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
764 764 ASSERT3U(mg->mg_histogram[i + ashift], >=,
765 765 msp->ms_sm->sm_phys->smp_histogram[i]);
766 766 ASSERT3U(mc->mc_histogram[i + ashift], >=,
767 767 msp->ms_sm->sm_phys->smp_histogram[i]);
768 768
769 769 mg->mg_histogram[i + ashift] -=
770 770 msp->ms_sm->sm_phys->smp_histogram[i];
771 771 mc->mc_histogram[i + ashift] -=
772 772 msp->ms_sm->sm_phys->smp_histogram[i];
773 773 }
774 774 mutex_exit(&mg->mg_lock);
775 775 }
776 776
777 777 static void
778 778 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
779 779 {
780 780 ASSERT(msp->ms_group == NULL);
781 781 mutex_enter(&mg->mg_lock);
782 782 msp->ms_group = mg;
783 783 msp->ms_weight = 0;
784 784 avl_add(&mg->mg_metaslab_tree, msp);
785 785 mutex_exit(&mg->mg_lock);
786 786
787 787 mutex_enter(&msp->ms_lock);
788 788 metaslab_group_histogram_add(mg, msp);
789 789 mutex_exit(&msp->ms_lock);
790 790 }
791 791
792 792 static void
793 793 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
794 794 {
795 795 mutex_enter(&msp->ms_lock);
796 796 metaslab_group_histogram_remove(mg, msp);
797 797 mutex_exit(&msp->ms_lock);
798 798
799 799 mutex_enter(&mg->mg_lock);
800 800 ASSERT(msp->ms_group == mg);
801 801 avl_remove(&mg->mg_metaslab_tree, msp);
802 802 msp->ms_group = NULL;
803 803 mutex_exit(&mg->mg_lock);
804 804 }
805 805
806 806 static void
807 807 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
808 808 {
809 809 /*
810 810 * Although in principle the weight can be any value, in
811 811 * practice we do not use values in the range [1, 511].
812 812 */
813 813 ASSERT(weight >= SPA_MINBLOCKSIZE || weight == 0);
814 814 ASSERT(MUTEX_HELD(&msp->ms_lock));
815 815
816 816 mutex_enter(&mg->mg_lock);
817 817 ASSERT(msp->ms_group == mg);
818 818 avl_remove(&mg->mg_metaslab_tree, msp);
819 819 msp->ms_weight = weight;
820 820 avl_add(&mg->mg_metaslab_tree, msp);
821 821 mutex_exit(&mg->mg_lock);
822 822 }
823 823
824 824 /*
825 825 * Calculate the fragmentation for a given metaslab group. We can use
826 826 * a simple average here since all metaslabs within the group must have
827 827 * the same size. The return value will be a value between 0 and 100
828 828 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
829 829 * group have a fragmentation metric.
830 830 */
831 831 uint64_t
832 832 metaslab_group_fragmentation(metaslab_group_t *mg)
833 833 {
834 834 vdev_t *vd = mg->mg_vd;
835 835 uint64_t fragmentation = 0;
836 836 uint64_t valid_ms = 0;
837 837
838 838 for (int m = 0; m < vd->vdev_ms_count; m++) {
839 839 metaslab_t *msp = vd->vdev_ms[m];
840 840
841 841 if (msp->ms_fragmentation == ZFS_FRAG_INVALID)
842 842 continue;
843 843
844 844 valid_ms++;
845 845 fragmentation += msp->ms_fragmentation;
846 846 }
847 847
848 848 if (valid_ms <= vd->vdev_ms_count / 2)
849 849 return (ZFS_FRAG_INVALID);
850 850
851 851 fragmentation /= valid_ms;
852 852 ASSERT3U(fragmentation, <=, 100);
853 853 return (fragmentation);
854 854 }
855 855
856 856 /*
857 857 * Determine if a given metaslab group should skip allocations. A metaslab
858 858 * group should avoid allocations if its free capacity is less than the
859 859 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
860 860 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
861 861 * that can still handle allocations. If the allocation throttle is enabled
862 862 * then we skip allocations to devices that have reached their maximum
863 863 * allocation queue depth unless the selected metaslab group is the only
864 864 * eligible group remaining.
865 865 */
866 866 static boolean_t
867 867 metaslab_group_allocatable(metaslab_group_t *mg, metaslab_group_t *rotor,
868 868 uint64_t psize)
869 869 {
870 870 spa_t *spa = mg->mg_vd->vdev_spa;
871 871 metaslab_class_t *mc = mg->mg_class;
872 872
873 873 /*
874 874 * We can only consider skipping this metaslab group if it's
875 875 * in the normal metaslab class and there are other metaslab
876 876 * groups to select from. Otherwise, we always consider it eligible
877 877 * for allocations.
878 878 */
879 879 if (mc != spa_normal_class(spa) || mc->mc_groups <= 1)
880 880 return (B_TRUE);
881 881
882 882 /*
883 883 * If the metaslab group's mg_allocatable flag is set (see comments
884 884 * in metaslab_group_alloc_update() for more information) and
885 885 * the allocation throttle is disabled then allow allocations to this
886 886 * device. However, if the allocation throttle is enabled then
887 887 * check if we have reached our allocation limit (mg_alloc_queue_depth)
888 888 * to determine if we should allow allocations to this metaslab group.
889 889 * If all metaslab groups are no longer considered allocatable
890 890 * (mc_alloc_groups == 0) or we're trying to allocate the smallest
891 891 * gang block size then we allow allocations on this metaslab group
892 892 * regardless of the mg_allocatable or throttle settings.
893 893 */
894 894 if (mg->mg_allocatable) {
895 895 metaslab_group_t *mgp;
896 896 int64_t qdepth;
897 897 uint64_t qmax = mg->mg_max_alloc_queue_depth;
898 898
899 899 if (!mc->mc_alloc_throttle_enabled)
900 900 return (B_TRUE);
901 901
902 902 /*
903 903 * If this metaslab group does not have any free space, then
904 904 * there is no point in looking further.
905 905 */
906 906 if (mg->mg_no_free_space)
907 907 return (B_FALSE);
908 908
909 909 qdepth = refcount_count(&mg->mg_alloc_queue_depth);
910 910
911 911 /*
912 912 * If this metaslab group is below its qmax or it's
913 913 * the only allocatable metasable group, then attempt
914 914 * to allocate from it.
915 915 */
916 916 if (qdepth < qmax || mc->mc_alloc_groups == 1)
917 917 return (B_TRUE);
918 918 ASSERT3U(mc->mc_alloc_groups, >, 1);
919 919
920 920 /*
921 921 * Since this metaslab group is at or over its qmax, we
922 922 * need to determine if there are metaslab groups after this
923 923 * one that might be able to handle this allocation. This is
924 924 * racy since we can't hold the locks for all metaslab
925 925 * groups at the same time when we make this check.
926 926 */
927 927 for (mgp = mg->mg_next; mgp != rotor; mgp = mgp->mg_next) {
928 928 qmax = mgp->mg_max_alloc_queue_depth;
929 929
930 930 qdepth = refcount_count(&mgp->mg_alloc_queue_depth);
931 931
932 932 /*
933 933 * If there is another metaslab group that
934 934 * might be able to handle the allocation, then
935 935 * we return false so that we skip this group.
936 936 */
937 937 if (qdepth < qmax && !mgp->mg_no_free_space)
938 938 return (B_FALSE);
939 939 }
940 940
941 941 /*
942 942 * We didn't find another group to handle the allocation
943 943 * so we can't skip this metaslab group even though
944 944 * we are at or over our qmax.
945 945 */
946 946 return (B_TRUE);
947 947
948 948 } else if (mc->mc_alloc_groups == 0 || psize == SPA_MINBLOCKSIZE) {
949 949 return (B_TRUE);
950 950 }
951 951 return (B_FALSE);
952 952 }
953 953
954 954 /*
955 955 * ==========================================================================
956 956 * Range tree callbacks
957 957 * ==========================================================================
958 958 */
959 959
960 960 /*
961 961 * Comparison function for the private size-ordered tree. Tree is sorted
962 962 * by size, larger sizes at the end of the tree.
963 963 */
964 964 static int
965 965 metaslab_rangesize_compare(const void *x1, const void *x2)
966 966 {
967 967 const range_seg_t *r1 = x1;
968 968 const range_seg_t *r2 = x2;
969 969 uint64_t rs_size1 = r1->rs_end - r1->rs_start;
970 970 uint64_t rs_size2 = r2->rs_end - r2->rs_start;
971 971
972 972 if (rs_size1 < rs_size2)
973 973 return (-1);
974 974 if (rs_size1 > rs_size2)
975 975 return (1);
976 976
977 977 if (r1->rs_start < r2->rs_start)
978 978 return (-1);
979 979
980 980 if (r1->rs_start > r2->rs_start)
981 981 return (1);
982 982
983 983 return (0);
984 984 }
985 985
986 986 /*
987 987 * Create any block allocator specific components. The current allocators
988 988 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
989 989 */
990 990 static void
991 991 metaslab_rt_create(range_tree_t *rt, void *arg)
992 992 {
993 993 metaslab_t *msp = arg;
994 994
995 995 ASSERT3P(rt->rt_arg, ==, msp);
996 996 ASSERT(msp->ms_tree == NULL);
997 997
998 998 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
999 999 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
1000 1000 }
1001 1001
1002 1002 /*
1003 1003 * Destroy the block allocator specific components.
1004 1004 */
1005 1005 static void
1006 1006 metaslab_rt_destroy(range_tree_t *rt, void *arg)
1007 1007 {
1008 1008 metaslab_t *msp = arg;
1009 1009
1010 1010 ASSERT3P(rt->rt_arg, ==, msp);
1011 1011 ASSERT3P(msp->ms_tree, ==, rt);
1012 1012 ASSERT0(avl_numnodes(&msp->ms_size_tree));
1013 1013
1014 1014 avl_destroy(&msp->ms_size_tree);
1015 1015 }
1016 1016
1017 1017 static void
1018 1018 metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg)
1019 1019 {
1020 1020 metaslab_t *msp = arg;
1021 1021
1022 1022 ASSERT3P(rt->rt_arg, ==, msp);
1023 1023 ASSERT3P(msp->ms_tree, ==, rt);
1024 1024 VERIFY(!msp->ms_condensing);
1025 1025 avl_add(&msp->ms_size_tree, rs);
1026 1026 }
1027 1027
1028 1028 static void
1029 1029 metaslab_rt_remove(range_tree_t *rt, range_seg_t *rs, void *arg)
1030 1030 {
1031 1031 metaslab_t *msp = arg;
1032 1032
1033 1033 ASSERT3P(rt->rt_arg, ==, msp);
1034 1034 ASSERT3P(msp->ms_tree, ==, rt);
1035 1035 VERIFY(!msp->ms_condensing);
1036 1036 avl_remove(&msp->ms_size_tree, rs);
1037 1037 }
1038 1038
1039 1039 static void
1040 1040 metaslab_rt_vacate(range_tree_t *rt, void *arg)
1041 1041 {
1042 1042 metaslab_t *msp = arg;
1043 1043
1044 1044 ASSERT3P(rt->rt_arg, ==, msp);
1045 1045 ASSERT3P(msp->ms_tree, ==, rt);
1046 1046
1047 1047 /*
1048 1048 * Normally one would walk the tree freeing nodes along the way.
1049 1049 * Since the nodes are shared with the range trees we can avoid
1050 1050 * walking all nodes and just reinitialize the avl tree. The nodes
1051 1051 * will be freed by the range tree, so we don't want to free them here.
1052 1052 */
1053 1053 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
1054 1054 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
1055 1055 }
1056 1056
1057 1057 static range_tree_ops_t metaslab_rt_ops = {
1058 1058 metaslab_rt_create,
1059 1059 metaslab_rt_destroy,
1060 1060 metaslab_rt_add,
1061 1061 metaslab_rt_remove,
1062 1062 metaslab_rt_vacate
1063 1063 };
1064 1064
1065 1065 /*
1066 1066 * ==========================================================================
1067 1067 * Common allocator routines
1068 1068 * ==========================================================================
1069 1069 */
1070 1070
1071 1071 /*
1072 1072 * Return the maximum contiguous segment within the metaslab.
1073 1073 */
1074 1074 uint64_t
1075 1075 metaslab_block_maxsize(metaslab_t *msp)
1076 1076 {
1077 1077 avl_tree_t *t = &msp->ms_size_tree;
1078 1078 range_seg_t *rs;
1079 1079
1080 1080 if (t == NULL || (rs = avl_last(t)) == NULL)
1081 1081 return (0ULL);
1082 1082
1083 1083 return (rs->rs_end - rs->rs_start);
1084 1084 }
1085 1085
1086 1086 static range_seg_t *
1087 1087 metaslab_block_find(avl_tree_t *t, uint64_t start, uint64_t size)
1088 1088 {
1089 1089 range_seg_t *rs, rsearch;
1090 1090 avl_index_t where;
1091 1091
1092 1092 rsearch.rs_start = start;
1093 1093 rsearch.rs_end = start + size;
1094 1094
1095 1095 rs = avl_find(t, &rsearch, &where);
1096 1096 if (rs == NULL) {
1097 1097 rs = avl_nearest(t, where, AVL_AFTER);
1098 1098 }
1099 1099
1100 1100 return (rs);
1101 1101 }
1102 1102
1103 1103 /*
1104 1104 * This is a helper function that can be used by the allocator to find
1105 1105 * a suitable block to allocate. This will search the specified AVL
1106 1106 * tree looking for a block that matches the specified criteria.
1107 1107 */
1108 1108 static uint64_t
1109 1109 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
1110 1110 uint64_t align)
1111 1111 {
1112 1112 range_seg_t *rs = metaslab_block_find(t, *cursor, size);
1113 1113
1114 1114 while (rs != NULL) {
1115 1115 uint64_t offset = P2ROUNDUP(rs->rs_start, align);
1116 1116
1117 1117 if (offset + size <= rs->rs_end) {
1118 1118 *cursor = offset + size;
1119 1119 return (offset);
1120 1120 }
1121 1121 rs = AVL_NEXT(t, rs);
1122 1122 }
1123 1123
1124 1124 /*
1125 1125 * If we know we've searched the whole map (*cursor == 0), give up.
1126 1126 * Otherwise, reset the cursor to the beginning and try again.
1127 1127 */
1128 1128 if (*cursor == 0)
1129 1129 return (-1ULL);
1130 1130
1131 1131 *cursor = 0;
1132 1132 return (metaslab_block_picker(t, cursor, size, align));
1133 1133 }
1134 1134
1135 1135 /*
1136 1136 * ==========================================================================
1137 1137 * The first-fit block allocator
1138 1138 * ==========================================================================
1139 1139 */
1140 1140 static uint64_t
1141 1141 metaslab_ff_alloc(metaslab_t *msp, uint64_t size)
1142 1142 {
1143 1143 /*
1144 1144 * Find the largest power of 2 block size that evenly divides the
1145 1145 * requested size. This is used to try to allocate blocks with similar
1146 1146 * alignment from the same area of the metaslab (i.e. same cursor
1147 1147 * bucket) but it does not guarantee that other allocations sizes
1148 1148 * may exist in the same region.
1149 1149 */
1150 1150 uint64_t align = size & -size;
1151 1151 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1152 1152 avl_tree_t *t = &msp->ms_tree->rt_root;
1153 1153
1154 1154 return (metaslab_block_picker(t, cursor, size, align));
1155 1155 }
1156 1156
1157 1157 static metaslab_ops_t metaslab_ff_ops = {
1158 1158 metaslab_ff_alloc
1159 1159 };
1160 1160
1161 1161 /*
1162 1162 * ==========================================================================
1163 1163 * Dynamic block allocator -
1164 1164 * Uses the first fit allocation scheme until space get low and then
1165 1165 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1166 1166 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1167 1167 * ==========================================================================
1168 1168 */
1169 1169 static uint64_t
1170 1170 metaslab_df_alloc(metaslab_t *msp, uint64_t size)
1171 1171 {
1172 1172 /*
1173 1173 * Find the largest power of 2 block size that evenly divides the
1174 1174 * requested size. This is used to try to allocate blocks with similar
1175 1175 * alignment from the same area of the metaslab (i.e. same cursor
1176 1176 * bucket) but it does not guarantee that other allocations sizes
1177 1177 * may exist in the same region.
1178 1178 */
1179 1179 uint64_t align = size & -size;
1180 1180 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1181 1181 range_tree_t *rt = msp->ms_tree;
1182 1182 avl_tree_t *t = &rt->rt_root;
1183 1183 uint64_t max_size = metaslab_block_maxsize(msp);
1184 1184 int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
1185 1185
1186 1186 ASSERT(MUTEX_HELD(&msp->ms_lock));
1187 1187 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1188 1188
1189 1189 if (max_size < size)
1190 1190 return (-1ULL);
1191 1191
1192 1192 /*
1193 1193 * If we're running low on space switch to using the size
1194 1194 * sorted AVL tree (best-fit).
1195 1195 */
1196 1196 if (max_size < metaslab_df_alloc_threshold ||
1197 1197 free_pct < metaslab_df_free_pct) {
1198 1198 t = &msp->ms_size_tree;
1199 1199 *cursor = 0;
1200 1200 }
1201 1201
1202 1202 return (metaslab_block_picker(t, cursor, size, 1ULL));
1203 1203 }
1204 1204
1205 1205 static metaslab_ops_t metaslab_df_ops = {
1206 1206 metaslab_df_alloc
1207 1207 };
1208 1208
1209 1209 /*
1210 1210 * ==========================================================================
1211 1211 * Cursor fit block allocator -
1212 1212 * Select the largest region in the metaslab, set the cursor to the beginning
1213 1213 * of the range and the cursor_end to the end of the range. As allocations
1214 1214 * are made advance the cursor. Continue allocating from the cursor until
1215 1215 * the range is exhausted and then find a new range.
1216 1216 * ==========================================================================
1217 1217 */
1218 1218 static uint64_t
1219 1219 metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
1220 1220 {
1221 1221 range_tree_t *rt = msp->ms_tree;
1222 1222 avl_tree_t *t = &msp->ms_size_tree;
1223 1223 uint64_t *cursor = &msp->ms_lbas[0];
1224 1224 uint64_t *cursor_end = &msp->ms_lbas[1];
1225 1225 uint64_t offset = 0;
1226 1226
1227 1227 ASSERT(MUTEX_HELD(&msp->ms_lock));
1228 1228 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&rt->rt_root));
1229 1229
1230 1230 ASSERT3U(*cursor_end, >=, *cursor);
1231 1231
1232 1232 if ((*cursor + size) > *cursor_end) {
1233 1233 range_seg_t *rs;
1234 1234
1235 1235 rs = avl_last(&msp->ms_size_tree);
1236 1236 if (rs == NULL || (rs->rs_end - rs->rs_start) < size)
1237 1237 return (-1ULL);
1238 1238
1239 1239 *cursor = rs->rs_start;
1240 1240 *cursor_end = rs->rs_end;
1241 1241 }
1242 1242
1243 1243 offset = *cursor;
1244 1244 *cursor += size;
1245 1245
1246 1246 return (offset);
1247 1247 }
1248 1248
1249 1249 static metaslab_ops_t metaslab_cf_ops = {
1250 1250 metaslab_cf_alloc
1251 1251 };
1252 1252
1253 1253 /*
1254 1254 * ==========================================================================
1255 1255 * New dynamic fit allocator -
1256 1256 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1257 1257 * contiguous blocks. If no region is found then just use the largest segment
1258 1258 * that remains.
1259 1259 * ==========================================================================
1260 1260 */
1261 1261
1262 1262 /*
1263 1263 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1264 1264 * to request from the allocator.
1265 1265 */
1266 1266 uint64_t metaslab_ndf_clump_shift = 4;
1267 1267
1268 1268 static uint64_t
1269 1269 metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
1270 1270 {
1271 1271 avl_tree_t *t = &msp->ms_tree->rt_root;
1272 1272 avl_index_t where;
1273 1273 range_seg_t *rs, rsearch;
1274 1274 uint64_t hbit = highbit64(size);
1275 1275 uint64_t *cursor = &msp->ms_lbas[hbit - 1];
1276 1276 uint64_t max_size = metaslab_block_maxsize(msp);
1277 1277
1278 1278 ASSERT(MUTEX_HELD(&msp->ms_lock));
1279 1279 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1280 1280
1281 1281 if (max_size < size)
1282 1282 return (-1ULL);
1283 1283
1284 1284 rsearch.rs_start = *cursor;
1285 1285 rsearch.rs_end = *cursor + size;
1286 1286
1287 1287 rs = avl_find(t, &rsearch, &where);
1288 1288 if (rs == NULL || (rs->rs_end - rs->rs_start) < size) {
1289 1289 t = &msp->ms_size_tree;
1290 1290
1291 1291 rsearch.rs_start = 0;
1292 1292 rsearch.rs_end = MIN(max_size,
1293 1293 1ULL << (hbit + metaslab_ndf_clump_shift));
1294 1294 rs = avl_find(t, &rsearch, &where);
1295 1295 if (rs == NULL)
1296 1296 rs = avl_nearest(t, where, AVL_AFTER);
1297 1297 ASSERT(rs != NULL);
1298 1298 }
1299 1299
1300 1300 if ((rs->rs_end - rs->rs_start) >= size) {
1301 1301 *cursor = rs->rs_start + size;
1302 1302 return (rs->rs_start);
1303 1303 }
1304 1304 return (-1ULL);
1305 1305 }
1306 1306
1307 1307 static metaslab_ops_t metaslab_ndf_ops = {
1308 1308 metaslab_ndf_alloc
1309 1309 };
1310 1310
1311 1311 metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
1312 1312
1313 1313 /*
1314 1314 * ==========================================================================
1315 1315 * Metaslabs
1316 1316 * ==========================================================================
1317 1317 */
1318 1318
1319 1319 /*
1320 1320 * Wait for any in-progress metaslab loads to complete.
1321 1321 */
1322 1322 void
1323 1323 metaslab_load_wait(metaslab_t *msp)
1324 1324 {
1325 1325 ASSERT(MUTEX_HELD(&msp->ms_lock));
1326 1326
1327 1327 while (msp->ms_loading) {
1328 1328 ASSERT(!msp->ms_loaded);
1329 1329 cv_wait(&msp->ms_load_cv, &msp->ms_lock);
1330 1330 }
1331 1331 }
1332 1332
1333 1333 int
1334 1334 metaslab_load(metaslab_t *msp)
1335 1335 {
1336 1336 int error = 0;
1337 1337 boolean_t success = B_FALSE;
1338 1338
1339 1339 ASSERT(MUTEX_HELD(&msp->ms_lock));
1340 1340 ASSERT(!msp->ms_loaded);
1341 1341 ASSERT(!msp->ms_loading);
1342 1342
1343 1343 msp->ms_loading = B_TRUE;
1344 1344
1345 1345 /*
1346 1346 * If the space map has not been allocated yet, then treat
1347 1347 * all the space in the metaslab as free and add it to the
1348 1348 * ms_tree.
1349 1349 */
1350 1350 if (msp->ms_sm != NULL)
1351 1351 error = space_map_load(msp->ms_sm, msp->ms_tree, SM_FREE);
1352 1352 else
1353 1353 range_tree_add(msp->ms_tree, msp->ms_start, msp->ms_size);
1354 1354
1355 1355 success = (error == 0);
1356 1356 msp->ms_loading = B_FALSE;
1357 1357
1358 1358 if (success) {
1359 1359 ASSERT3P(msp->ms_group, !=, NULL);
1360 1360 msp->ms_loaded = B_TRUE;
1361 1361
1362 1362 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1363 1363 range_tree_walk(msp->ms_defertree[t],
1364 1364 range_tree_remove, msp->ms_tree);
1365 1365 }
1366 1366 msp->ms_max_size = metaslab_block_maxsize(msp);
1367 1367 }
1368 1368 cv_broadcast(&msp->ms_load_cv);
1369 1369 return (error);
1370 1370 }
1371 1371
1372 1372 void
1373 1373 metaslab_unload(metaslab_t *msp)
1374 1374 {
1375 1375 ASSERT(MUTEX_HELD(&msp->ms_lock));
1376 1376 range_tree_vacate(msp->ms_tree, NULL, NULL);
1377 1377 msp->ms_loaded = B_FALSE;
1378 1378 msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
1379 1379 msp->ms_max_size = 0;
1380 1380 }
1381 1381
1382 1382 int
1383 1383 metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, uint64_t txg,
1384 1384 metaslab_t **msp)
1385 1385 {
1386 1386 vdev_t *vd = mg->mg_vd;
1387 1387 objset_t *mos = vd->vdev_spa->spa_meta_objset;
1388 1388 metaslab_t *ms;
1389 1389 int error;
1390 1390
1391 1391 ms = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
1392 1392 mutex_init(&ms->ms_lock, NULL, MUTEX_DEFAULT, NULL);
1393 1393 cv_init(&ms->ms_load_cv, NULL, CV_DEFAULT, NULL);
1394 1394 ms->ms_id = id;
1395 1395 ms->ms_start = id << vd->vdev_ms_shift;
1396 1396 ms->ms_size = 1ULL << vd->vdev_ms_shift;
1397 1397
1398 1398 /*
1399 1399 * We only open space map objects that already exist. All others
1400 1400 * will be opened when we finally allocate an object for it.
1401 1401 */
1402 1402 if (object != 0) {
1403 1403 error = space_map_open(&ms->ms_sm, mos, object, ms->ms_start,
1404 1404 ms->ms_size, vd->vdev_ashift, &ms->ms_lock);
1405 1405
1406 1406 if (error != 0) {
1407 1407 kmem_free(ms, sizeof (metaslab_t));
1408 1408 return (error);
1409 1409 }
1410 1410
1411 1411 ASSERT(ms->ms_sm != NULL);
1412 1412 }
1413 1413
1414 1414 /*
1415 1415 * We create the main range tree here, but we don't create the
1416 1416 * other range trees until metaslab_sync_done(). This serves
1417 1417 * two purposes: it allows metaslab_sync_done() to detect the
1418 1418 * addition of new space; and for debugging, it ensures that we'd
1419 1419 * data fault on any attempt to use this metaslab before it's ready.
1420 1420 */
1421 1421 ms->ms_tree = range_tree_create(&metaslab_rt_ops, ms, &ms->ms_lock);
1422 1422 metaslab_group_add(mg, ms);
1423 1423
1424 1424 metaslab_set_fragmentation(ms);
1425 1425
1426 1426 /*
1427 1427 * If we're opening an existing pool (txg == 0) or creating
1428 1428 * a new one (txg == TXG_INITIAL), all space is available now.
1429 1429 * If we're adding space to an existing pool, the new space
1430 1430 * does not become available until after this txg has synced.
1431 1431 * The metaslab's weight will also be initialized when we sync
1432 1432 * out this txg. This ensures that we don't attempt to allocate
1433 1433 * from it before we have initialized it completely.
1434 1434 */
1435 1435 if (txg <= TXG_INITIAL)
1436 1436 metaslab_sync_done(ms, 0);
1437 1437
1438 1438 /*
1439 1439 * If metaslab_debug_load is set and we're initializing a metaslab
1440 1440 * that has an allocated space map object then load the its space
1441 1441 * map so that can verify frees.
1442 1442 */
1443 1443 if (metaslab_debug_load && ms->ms_sm != NULL) {
1444 1444 mutex_enter(&ms->ms_lock);
1445 1445 VERIFY0(metaslab_load(ms));
1446 1446 mutex_exit(&ms->ms_lock);
1447 1447 }
1448 1448
1449 1449 if (txg != 0) {
1450 1450 vdev_dirty(vd, 0, NULL, txg);
1451 1451 vdev_dirty(vd, VDD_METASLAB, ms, txg);
1452 1452 }
1453 1453
1454 1454 *msp = ms;
1455 1455
1456 1456 return (0);
1457 1457 }
1458 1458
1459 1459 void
1460 1460 metaslab_fini(metaslab_t *msp)
1461 1461 {
1462 1462 metaslab_group_t *mg = msp->ms_group;
1463 1463
1464 1464 metaslab_group_remove(mg, msp);
1465 1465
1466 1466 mutex_enter(&msp->ms_lock);
1467 1467 VERIFY(msp->ms_group == NULL);
1468 1468 vdev_space_update(mg->mg_vd, -space_map_allocated(msp->ms_sm),
1469 1469 0, -msp->ms_size);
1470 1470 space_map_close(msp->ms_sm);
1471 1471
1472 1472 metaslab_unload(msp);
1473 1473 range_tree_destroy(msp->ms_tree);
1474 1474 range_tree_destroy(msp->ms_freeingtree);
1475 1475 range_tree_destroy(msp->ms_freedtree);
1476 1476
1477 1477 for (int t = 0; t < TXG_SIZE; t++) {
1478 1478 range_tree_destroy(msp->ms_alloctree[t]);
1479 1479 }
1480 1480
1481 1481 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1482 1482 range_tree_destroy(msp->ms_defertree[t]);
1483 1483 }
1484 1484
1485 1485 ASSERT0(msp->ms_deferspace);
1486 1486
1487 1487 mutex_exit(&msp->ms_lock);
1488 1488 cv_destroy(&msp->ms_load_cv);
1489 1489 mutex_destroy(&msp->ms_lock);
1490 1490
1491 1491 kmem_free(msp, sizeof (metaslab_t));
1492 1492 }
1493 1493
1494 1494 #define FRAGMENTATION_TABLE_SIZE 17
1495 1495
1496 1496 /*
1497 1497 * This table defines a segment size based fragmentation metric that will
1498 1498 * allow each metaslab to derive its own fragmentation value. This is done
1499 1499 * by calculating the space in each bucket of the spacemap histogram and
1500 1500 * multiplying that by the fragmetation metric in this table. Doing
1501 1501 * this for all buckets and dividing it by the total amount of free
1502 1502 * space in this metaslab (i.e. the total free space in all buckets) gives
1503 1503 * us the fragmentation metric. This means that a high fragmentation metric
1504 1504 * equates to most of the free space being comprised of small segments.
1505 1505 * Conversely, if the metric is low, then most of the free space is in
1506 1506 * large segments. A 10% change in fragmentation equates to approximately
1507 1507 * double the number of segments.
1508 1508 *
1509 1509 * This table defines 0% fragmented space using 16MB segments. Testing has
1510 1510 * shown that segments that are greater than or equal to 16MB do not suffer
1511 1511 * from drastic performance problems. Using this value, we derive the rest
1512 1512 * of the table. Since the fragmentation value is never stored on disk, it
1513 1513 * is possible to change these calculations in the future.
1514 1514 */
1515 1515 int zfs_frag_table[FRAGMENTATION_TABLE_SIZE] = {
1516 1516 100, /* 512B */
1517 1517 100, /* 1K */
1518 1518 98, /* 2K */
1519 1519 95, /* 4K */
1520 1520 90, /* 8K */
1521 1521 80, /* 16K */
1522 1522 70, /* 32K */
1523 1523 60, /* 64K */
1524 1524 50, /* 128K */
1525 1525 40, /* 256K */
1526 1526 30, /* 512K */
1527 1527 20, /* 1M */
1528 1528 15, /* 2M */
1529 1529 10, /* 4M */
1530 1530 5, /* 8M */
1531 1531 0 /* 16M */
1532 1532 };
1533 1533
1534 1534 /*
1535 1535 * Calclate the metaslab's fragmentation metric. A return value
1536 1536 * of ZFS_FRAG_INVALID means that the metaslab has not been upgraded and does
1537 1537 * not support this metric. Otherwise, the return value should be in the
1538 1538 * range [0, 100].
1539 1539 */
1540 1540 static void
1541 1541 metaslab_set_fragmentation(metaslab_t *msp)
1542 1542 {
1543 1543 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1544 1544 uint64_t fragmentation = 0;
1545 1545 uint64_t total = 0;
1546 1546 boolean_t feature_enabled = spa_feature_is_enabled(spa,
1547 1547 SPA_FEATURE_SPACEMAP_HISTOGRAM);
1548 1548
1549 1549 if (!feature_enabled) {
1550 1550 msp->ms_fragmentation = ZFS_FRAG_INVALID;
1551 1551 return;
1552 1552 }
1553 1553
1554 1554 /*
1555 1555 * A null space map means that the entire metaslab is free
1556 1556 * and thus is not fragmented.
1557 1557 */
1558 1558 if (msp->ms_sm == NULL) {
1559 1559 msp->ms_fragmentation = 0;
1560 1560 return;
1561 1561 }
1562 1562
1563 1563 /*
1564 1564 * If this metaslab's space map has not been upgraded, flag it
1565 1565 * so that we upgrade next time we encounter it.
1566 1566 */
1567 1567 if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t)) {
1568 1568 uint64_t txg = spa_syncing_txg(spa);
1569 1569 vdev_t *vd = msp->ms_group->mg_vd;
1570 1570
1571 1571 /*
1572 1572 * If we've reached the final dirty txg, then we must
1573 1573 * be shutting down the pool. We don't want to dirty
1574 1574 * any data past this point so skip setting the condense
1575 1575 * flag. We can retry this action the next time the pool
1576 1576 * is imported.
1577 1577 */
1578 1578 if (spa_writeable(spa) && txg < spa_final_dirty_txg(spa)) {
1579 1579 msp->ms_condense_wanted = B_TRUE;
1580 1580 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1581 1581 spa_dbgmsg(spa, "txg %llu, requesting force condense: "
1582 1582 "ms_id %llu, vdev_id %llu", txg, msp->ms_id,
1583 1583 vd->vdev_id);
1584 1584 }
1585 1585 msp->ms_fragmentation = ZFS_FRAG_INVALID;
1586 1586 return;
1587 1587 }
1588 1588
1589 1589 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
1590 1590 uint64_t space = 0;
1591 1591 uint8_t shift = msp->ms_sm->sm_shift;
1592 1592
1593 1593 int idx = MIN(shift - SPA_MINBLOCKSHIFT + i,
1594 1594 FRAGMENTATION_TABLE_SIZE - 1);
1595 1595
1596 1596 if (msp->ms_sm->sm_phys->smp_histogram[i] == 0)
1597 1597 continue;
1598 1598
1599 1599 space = msp->ms_sm->sm_phys->smp_histogram[i] << (i + shift);
1600 1600 total += space;
1601 1601
1602 1602 ASSERT3U(idx, <, FRAGMENTATION_TABLE_SIZE);
1603 1603 fragmentation += space * zfs_frag_table[idx];
1604 1604 }
1605 1605
1606 1606 if (total > 0)
1607 1607 fragmentation /= total;
1608 1608 ASSERT3U(fragmentation, <=, 100);
1609 1609
1610 1610 msp->ms_fragmentation = fragmentation;
1611 1611 }
1612 1612
1613 1613 /*
1614 1614 * Compute a weight -- a selection preference value -- for the given metaslab.
1615 1615 * This is based on the amount of free space, the level of fragmentation,
1616 1616 * the LBA range, and whether the metaslab is loaded.
1617 1617 */
1618 1618 static uint64_t
1619 1619 metaslab_space_weight(metaslab_t *msp)
1620 1620 {
1621 1621 metaslab_group_t *mg = msp->ms_group;
1622 1622 vdev_t *vd = mg->mg_vd;
1623 1623 uint64_t weight, space;
1624 1624
1625 1625 ASSERT(MUTEX_HELD(&msp->ms_lock));
1626 1626 ASSERT(!vd->vdev_removing);
1627 1627
1628 1628 /*
1629 1629 * The baseline weight is the metaslab's free space.
1630 1630 */
1631 1631 space = msp->ms_size - space_map_allocated(msp->ms_sm);
1632 1632
1633 1633 if (metaslab_fragmentation_factor_enabled &&
1634 1634 msp->ms_fragmentation != ZFS_FRAG_INVALID) {
1635 1635 /*
1636 1636 * Use the fragmentation information to inversely scale
1637 1637 * down the baseline weight. We need to ensure that we
1638 1638 * don't exclude this metaslab completely when it's 100%
1639 1639 * fragmented. To avoid this we reduce the fragmented value
1640 1640 * by 1.
1641 1641 */
1642 1642 space = (space * (100 - (msp->ms_fragmentation - 1))) / 100;
1643 1643
1644 1644 /*
1645 1645 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
1646 1646 * this metaslab again. The fragmentation metric may have
1647 1647 * decreased the space to something smaller than
1648 1648 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
1649 1649 * so that we can consume any remaining space.
1650 1650 */
1651 1651 if (space > 0 && space < SPA_MINBLOCKSIZE)
1652 1652 space = SPA_MINBLOCKSIZE;
1653 1653 }
1654 1654 weight = space;
|
↓ open down ↓ |
1654 lines elided |
↑ open up ↑ |
1655 1655
1656 1656 /*
1657 1657 * Modern disks have uniform bit density and constant angular velocity.
1658 1658 * Therefore, the outer recording zones are faster (higher bandwidth)
1659 1659 * than the inner zones by the ratio of outer to inner track diameter,
1660 1660 * which is typically around 2:1. We account for this by assigning
1661 1661 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1662 1662 * In effect, this means that we'll select the metaslab with the most
1663 1663 * free bandwidth rather than simply the one with the most free space.
1664 1664 */
1665 - if (metaslab_lba_weighting_enabled) {
1665 + if (!vd->vdev_nonrot && metaslab_lba_weighting_enabled) {
1666 1666 weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count;
1667 1667 ASSERT(weight >= space && weight <= 2 * space);
1668 1668 }
1669 1669
1670 1670 /*
1671 1671 * If this metaslab is one we're actively using, adjust its
1672 1672 * weight to make it preferable to any inactive metaslab so
1673 1673 * we'll polish it off. If the fragmentation on this metaslab
1674 1674 * has exceed our threshold, then don't mark it active.
1675 1675 */
1676 1676 if (msp->ms_loaded && msp->ms_fragmentation != ZFS_FRAG_INVALID &&
1677 1677 msp->ms_fragmentation <= zfs_metaslab_fragmentation_threshold) {
1678 1678 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
1679 1679 }
1680 1680
1681 1681 WEIGHT_SET_SPACEBASED(weight);
1682 1682 return (weight);
1683 1683 }
1684 1684
1685 1685 /*
1686 1686 * Return the weight of the specified metaslab, according to the segment-based
1687 1687 * weighting algorithm. The metaslab must be loaded. This function can
1688 1688 * be called within a sync pass since it relies only on the metaslab's
1689 1689 * range tree which is always accurate when the metaslab is loaded.
1690 1690 */
1691 1691 static uint64_t
1692 1692 metaslab_weight_from_range_tree(metaslab_t *msp)
1693 1693 {
1694 1694 uint64_t weight = 0;
1695 1695 uint32_t segments = 0;
1696 1696
1697 1697 ASSERT(msp->ms_loaded);
1698 1698
1699 1699 for (int i = RANGE_TREE_HISTOGRAM_SIZE - 1; i >= SPA_MINBLOCKSHIFT;
1700 1700 i--) {
1701 1701 uint8_t shift = msp->ms_group->mg_vd->vdev_ashift;
1702 1702 int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1;
1703 1703
1704 1704 segments <<= 1;
1705 1705 segments += msp->ms_tree->rt_histogram[i];
1706 1706
1707 1707 /*
1708 1708 * The range tree provides more precision than the space map
1709 1709 * and must be downgraded so that all values fit within the
1710 1710 * space map's histogram. This allows us to compare loaded
1711 1711 * vs. unloaded metaslabs to determine which metaslab is
1712 1712 * considered "best".
1713 1713 */
1714 1714 if (i > max_idx)
1715 1715 continue;
1716 1716
1717 1717 if (segments != 0) {
1718 1718 WEIGHT_SET_COUNT(weight, segments);
1719 1719 WEIGHT_SET_INDEX(weight, i);
1720 1720 WEIGHT_SET_ACTIVE(weight, 0);
1721 1721 break;
1722 1722 }
1723 1723 }
1724 1724 return (weight);
1725 1725 }
1726 1726
1727 1727 /*
1728 1728 * Calculate the weight based on the on-disk histogram. This should only
1729 1729 * be called after a sync pass has completely finished since the on-disk
1730 1730 * information is updated in metaslab_sync().
1731 1731 */
1732 1732 static uint64_t
1733 1733 metaslab_weight_from_spacemap(metaslab_t *msp)
1734 1734 {
1735 1735 uint64_t weight = 0;
1736 1736
1737 1737 for (int i = SPACE_MAP_HISTOGRAM_SIZE - 1; i >= 0; i--) {
1738 1738 if (msp->ms_sm->sm_phys->smp_histogram[i] != 0) {
1739 1739 WEIGHT_SET_COUNT(weight,
1740 1740 msp->ms_sm->sm_phys->smp_histogram[i]);
1741 1741 WEIGHT_SET_INDEX(weight, i +
1742 1742 msp->ms_sm->sm_shift);
1743 1743 WEIGHT_SET_ACTIVE(weight, 0);
1744 1744 break;
1745 1745 }
1746 1746 }
1747 1747 return (weight);
1748 1748 }
1749 1749
1750 1750 /*
1751 1751 * Compute a segment-based weight for the specified metaslab. The weight
1752 1752 * is determined by highest bucket in the histogram. The information
1753 1753 * for the highest bucket is encoded into the weight value.
1754 1754 */
1755 1755 static uint64_t
1756 1756 metaslab_segment_weight(metaslab_t *msp)
1757 1757 {
1758 1758 metaslab_group_t *mg = msp->ms_group;
1759 1759 uint64_t weight = 0;
1760 1760 uint8_t shift = mg->mg_vd->vdev_ashift;
1761 1761
1762 1762 ASSERT(MUTEX_HELD(&msp->ms_lock));
1763 1763
1764 1764 /*
1765 1765 * The metaslab is completely free.
1766 1766 */
1767 1767 if (space_map_allocated(msp->ms_sm) == 0) {
1768 1768 int idx = highbit64(msp->ms_size) - 1;
1769 1769 int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1;
1770 1770
1771 1771 if (idx < max_idx) {
1772 1772 WEIGHT_SET_COUNT(weight, 1ULL);
1773 1773 WEIGHT_SET_INDEX(weight, idx);
1774 1774 } else {
1775 1775 WEIGHT_SET_COUNT(weight, 1ULL << (idx - max_idx));
1776 1776 WEIGHT_SET_INDEX(weight, max_idx);
1777 1777 }
1778 1778 WEIGHT_SET_ACTIVE(weight, 0);
1779 1779 ASSERT(!WEIGHT_IS_SPACEBASED(weight));
1780 1780
1781 1781 return (weight);
1782 1782 }
1783 1783
1784 1784 ASSERT3U(msp->ms_sm->sm_dbuf->db_size, ==, sizeof (space_map_phys_t));
1785 1785
1786 1786 /*
1787 1787 * If the metaslab is fully allocated then just make the weight 0.
1788 1788 */
1789 1789 if (space_map_allocated(msp->ms_sm) == msp->ms_size)
1790 1790 return (0);
1791 1791 /*
1792 1792 * If the metaslab is already loaded, then use the range tree to
1793 1793 * determine the weight. Otherwise, we rely on the space map information
1794 1794 * to generate the weight.
1795 1795 */
1796 1796 if (msp->ms_loaded) {
1797 1797 weight = metaslab_weight_from_range_tree(msp);
1798 1798 } else {
1799 1799 weight = metaslab_weight_from_spacemap(msp);
1800 1800 }
1801 1801
1802 1802 /*
1803 1803 * If the metaslab was active the last time we calculated its weight
1804 1804 * then keep it active. We want to consume the entire region that
1805 1805 * is associated with this weight.
1806 1806 */
1807 1807 if (msp->ms_activation_weight != 0 && weight != 0)
1808 1808 WEIGHT_SET_ACTIVE(weight, WEIGHT_GET_ACTIVE(msp->ms_weight));
1809 1809 return (weight);
1810 1810 }
1811 1811
1812 1812 /*
1813 1813 * Determine if we should attempt to allocate from this metaslab. If the
1814 1814 * metaslab has a maximum size then we can quickly determine if the desired
1815 1815 * allocation size can be satisfied. Otherwise, if we're using segment-based
1816 1816 * weighting then we can determine the maximum allocation that this metaslab
1817 1817 * can accommodate based on the index encoded in the weight. If we're using
1818 1818 * space-based weights then rely on the entire weight (excluding the weight
1819 1819 * type bit).
1820 1820 */
1821 1821 boolean_t
1822 1822 metaslab_should_allocate(metaslab_t *msp, uint64_t asize)
1823 1823 {
1824 1824 boolean_t should_allocate;
1825 1825
1826 1826 if (msp->ms_max_size != 0)
1827 1827 return (msp->ms_max_size >= asize);
1828 1828
1829 1829 if (!WEIGHT_IS_SPACEBASED(msp->ms_weight)) {
1830 1830 /*
1831 1831 * The metaslab segment weight indicates segments in the
1832 1832 * range [2^i, 2^(i+1)), where i is the index in the weight.
1833 1833 * Since the asize might be in the middle of the range, we
1834 1834 * should attempt the allocation if asize < 2^(i+1).
1835 1835 */
1836 1836 should_allocate = (asize <
1837 1837 1ULL << (WEIGHT_GET_INDEX(msp->ms_weight) + 1));
1838 1838 } else {
1839 1839 should_allocate = (asize <=
1840 1840 (msp->ms_weight & ~METASLAB_WEIGHT_TYPE));
1841 1841 }
1842 1842 return (should_allocate);
1843 1843 }
1844 1844
1845 1845 static uint64_t
1846 1846 metaslab_weight(metaslab_t *msp)
1847 1847 {
1848 1848 vdev_t *vd = msp->ms_group->mg_vd;
1849 1849 spa_t *spa = vd->vdev_spa;
1850 1850 uint64_t weight;
1851 1851
1852 1852 ASSERT(MUTEX_HELD(&msp->ms_lock));
1853 1853
1854 1854 /*
1855 1855 * This vdev is in the process of being removed so there is nothing
1856 1856 * for us to do here.
1857 1857 */
1858 1858 if (vd->vdev_removing) {
1859 1859 ASSERT0(space_map_allocated(msp->ms_sm));
1860 1860 ASSERT0(vd->vdev_ms_shift);
1861 1861 return (0);
1862 1862 }
1863 1863
1864 1864 metaslab_set_fragmentation(msp);
1865 1865
1866 1866 /*
1867 1867 * Update the maximum size if the metaslab is loaded. This will
1868 1868 * ensure that we get an accurate maximum size if newly freed space
1869 1869 * has been added back into the free tree.
1870 1870 */
1871 1871 if (msp->ms_loaded)
1872 1872 msp->ms_max_size = metaslab_block_maxsize(msp);
1873 1873
1874 1874 /*
1875 1875 * Segment-based weighting requires space map histogram support.
1876 1876 */
1877 1877 if (zfs_metaslab_segment_weight_enabled &&
1878 1878 spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
1879 1879 (msp->ms_sm == NULL || msp->ms_sm->sm_dbuf->db_size ==
1880 1880 sizeof (space_map_phys_t))) {
1881 1881 weight = metaslab_segment_weight(msp);
1882 1882 } else {
1883 1883 weight = metaslab_space_weight(msp);
1884 1884 }
1885 1885 return (weight);
1886 1886 }
1887 1887
1888 1888 static int
1889 1889 metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
1890 1890 {
1891 1891 ASSERT(MUTEX_HELD(&msp->ms_lock));
1892 1892
1893 1893 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1894 1894 metaslab_load_wait(msp);
1895 1895 if (!msp->ms_loaded) {
1896 1896 int error = metaslab_load(msp);
1897 1897 if (error) {
1898 1898 metaslab_group_sort(msp->ms_group, msp, 0);
1899 1899 return (error);
1900 1900 }
1901 1901 }
1902 1902
1903 1903 msp->ms_activation_weight = msp->ms_weight;
1904 1904 metaslab_group_sort(msp->ms_group, msp,
1905 1905 msp->ms_weight | activation_weight);
1906 1906 }
1907 1907 ASSERT(msp->ms_loaded);
1908 1908 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
1909 1909
1910 1910 return (0);
1911 1911 }
1912 1912
1913 1913 static void
1914 1914 metaslab_passivate(metaslab_t *msp, uint64_t weight)
1915 1915 {
1916 1916 uint64_t size = weight & ~METASLAB_WEIGHT_TYPE;
1917 1917
1918 1918 /*
1919 1919 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1920 1920 * this metaslab again. In that case, it had better be empty,
1921 1921 * or we would be leaving space on the table.
1922 1922 */
1923 1923 ASSERT(size >= SPA_MINBLOCKSIZE ||
1924 1924 range_tree_space(msp->ms_tree) == 0);
1925 1925 ASSERT0(weight & METASLAB_ACTIVE_MASK);
1926 1926
1927 1927 msp->ms_activation_weight = 0;
1928 1928 metaslab_group_sort(msp->ms_group, msp, weight);
1929 1929 ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
1930 1930 }
1931 1931
1932 1932 /*
1933 1933 * Segment-based metaslabs are activated once and remain active until
1934 1934 * we either fail an allocation attempt (similar to space-based metaslabs)
1935 1935 * or have exhausted the free space in zfs_metaslab_switch_threshold
1936 1936 * buckets since the metaslab was activated. This function checks to see
1937 1937 * if we've exhaused the zfs_metaslab_switch_threshold buckets in the
1938 1938 * metaslab and passivates it proactively. This will allow us to select a
1939 1939 * metaslabs with larger contiguous region if any remaining within this
1940 1940 * metaslab group. If we're in sync pass > 1, then we continue using this
1941 1941 * metaslab so that we don't dirty more block and cause more sync passes.
1942 1942 */
1943 1943 void
1944 1944 metaslab_segment_may_passivate(metaslab_t *msp)
1945 1945 {
1946 1946 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1947 1947
1948 1948 if (WEIGHT_IS_SPACEBASED(msp->ms_weight) || spa_sync_pass(spa) > 1)
1949 1949 return;
1950 1950
1951 1951 /*
1952 1952 * Since we are in the middle of a sync pass, the most accurate
1953 1953 * information that is accessible to us is the in-core range tree
1954 1954 * histogram; calculate the new weight based on that information.
1955 1955 */
1956 1956 uint64_t weight = metaslab_weight_from_range_tree(msp);
1957 1957 int activation_idx = WEIGHT_GET_INDEX(msp->ms_activation_weight);
1958 1958 int current_idx = WEIGHT_GET_INDEX(weight);
1959 1959
1960 1960 if (current_idx <= activation_idx - zfs_metaslab_switch_threshold)
1961 1961 metaslab_passivate(msp, weight);
1962 1962 }
1963 1963
1964 1964 static void
1965 1965 metaslab_preload(void *arg)
1966 1966 {
1967 1967 metaslab_t *msp = arg;
1968 1968 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1969 1969
1970 1970 ASSERT(!MUTEX_HELD(&msp->ms_group->mg_lock));
1971 1971
1972 1972 mutex_enter(&msp->ms_lock);
1973 1973 metaslab_load_wait(msp);
1974 1974 if (!msp->ms_loaded)
1975 1975 (void) metaslab_load(msp);
1976 1976 msp->ms_selected_txg = spa_syncing_txg(spa);
1977 1977 mutex_exit(&msp->ms_lock);
1978 1978 }
1979 1979
1980 1980 static void
1981 1981 metaslab_group_preload(metaslab_group_t *mg)
1982 1982 {
1983 1983 spa_t *spa = mg->mg_vd->vdev_spa;
1984 1984 metaslab_t *msp;
1985 1985 avl_tree_t *t = &mg->mg_metaslab_tree;
1986 1986 int m = 0;
1987 1987
1988 1988 if (spa_shutting_down(spa) || !metaslab_preload_enabled) {
1989 1989 taskq_wait(mg->mg_taskq);
1990 1990 return;
1991 1991 }
1992 1992
1993 1993 mutex_enter(&mg->mg_lock);
1994 1994 /*
1995 1995 * Load the next potential metaslabs
1996 1996 */
1997 1997 for (msp = avl_first(t); msp != NULL; msp = AVL_NEXT(t, msp)) {
1998 1998 /*
1999 1999 * We preload only the maximum number of metaslabs specified
2000 2000 * by metaslab_preload_limit. If a metaslab is being forced
2001 2001 * to condense then we preload it too. This will ensure
2002 2002 * that force condensing happens in the next txg.
2003 2003 */
2004 2004 if (++m > metaslab_preload_limit && !msp->ms_condense_wanted) {
2005 2005 continue;
2006 2006 }
2007 2007
2008 2008 VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
2009 2009 msp, TQ_SLEEP) != NULL);
2010 2010 }
2011 2011 mutex_exit(&mg->mg_lock);
2012 2012 }
2013 2013
2014 2014 /*
2015 2015 * Determine if the space map's on-disk footprint is past our tolerance
2016 2016 * for inefficiency. We would like to use the following criteria to make
2017 2017 * our decision:
2018 2018 *
2019 2019 * 1. The size of the space map object should not dramatically increase as a
2020 2020 * result of writing out the free space range tree.
2021 2021 *
2022 2022 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
2023 2023 * times the size than the free space range tree representation
2024 2024 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
2025 2025 *
2026 2026 * 3. The on-disk size of the space map should actually decrease.
2027 2027 *
2028 2028 * Checking the first condition is tricky since we don't want to walk
2029 2029 * the entire AVL tree calculating the estimated on-disk size. Instead we
2030 2030 * use the size-ordered range tree in the metaslab and calculate the
2031 2031 * size required to write out the largest segment in our free tree. If the
2032 2032 * size required to represent that segment on disk is larger than the space
2033 2033 * map object then we avoid condensing this map.
2034 2034 *
2035 2035 * To determine the second criterion we use a best-case estimate and assume
2036 2036 * each segment can be represented on-disk as a single 64-bit entry. We refer
2037 2037 * to this best-case estimate as the space map's minimal form.
2038 2038 *
2039 2039 * Unfortunately, we cannot compute the on-disk size of the space map in this
2040 2040 * context because we cannot accurately compute the effects of compression, etc.
2041 2041 * Instead, we apply the heuristic described in the block comment for
2042 2042 * zfs_metaslab_condense_block_threshold - we only condense if the space used
2043 2043 * is greater than a threshold number of blocks.
2044 2044 */
2045 2045 static boolean_t
2046 2046 metaslab_should_condense(metaslab_t *msp)
2047 2047 {
2048 2048 space_map_t *sm = msp->ms_sm;
2049 2049 range_seg_t *rs;
2050 2050 uint64_t size, entries, segsz, object_size, optimal_size, record_size;
2051 2051 dmu_object_info_t doi;
2052 2052 uint64_t vdev_blocksize = 1 << msp->ms_group->mg_vd->vdev_ashift;
2053 2053
2054 2054 ASSERT(MUTEX_HELD(&msp->ms_lock));
2055 2055 ASSERT(msp->ms_loaded);
2056 2056
2057 2057 /*
2058 2058 * Use the ms_size_tree range tree, which is ordered by size, to
2059 2059 * obtain the largest segment in the free tree. We always condense
2060 2060 * metaslabs that are empty and metaslabs for which a condense
2061 2061 * request has been made.
2062 2062 */
2063 2063 rs = avl_last(&msp->ms_size_tree);
2064 2064 if (rs == NULL || msp->ms_condense_wanted)
2065 2065 return (B_TRUE);
2066 2066
2067 2067 /*
2068 2068 * Calculate the number of 64-bit entries this segment would
2069 2069 * require when written to disk. If this single segment would be
2070 2070 * larger on-disk than the entire current on-disk structure, then
2071 2071 * clearly condensing will increase the on-disk structure size.
2072 2072 */
2073 2073 size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
2074 2074 entries = size / (MIN(size, SM_RUN_MAX));
2075 2075 segsz = entries * sizeof (uint64_t);
2076 2076
2077 2077 optimal_size = sizeof (uint64_t) * avl_numnodes(&msp->ms_tree->rt_root);
2078 2078 object_size = space_map_length(msp->ms_sm);
2079 2079
2080 2080 dmu_object_info_from_db(sm->sm_dbuf, &doi);
2081 2081 record_size = MAX(doi.doi_data_block_size, vdev_blocksize);
2082 2082
2083 2083 return (segsz <= object_size &&
2084 2084 object_size >= (optimal_size * zfs_condense_pct / 100) &&
2085 2085 object_size > zfs_metaslab_condense_block_threshold * record_size);
2086 2086 }
2087 2087
2088 2088 /*
2089 2089 * Condense the on-disk space map representation to its minimized form.
2090 2090 * The minimized form consists of a small number of allocations followed by
2091 2091 * the entries of the free range tree.
2092 2092 */
2093 2093 static void
2094 2094 metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
2095 2095 {
2096 2096 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2097 2097 range_tree_t *condense_tree;
2098 2098 space_map_t *sm = msp->ms_sm;
2099 2099
2100 2100 ASSERT(MUTEX_HELD(&msp->ms_lock));
2101 2101 ASSERT3U(spa_sync_pass(spa), ==, 1);
2102 2102 ASSERT(msp->ms_loaded);
2103 2103
2104 2104
2105 2105 spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, vdev id %llu, "
2106 2106 "spa %s, smp size %llu, segments %lu, forcing condense=%s", txg,
2107 2107 msp->ms_id, msp, msp->ms_group->mg_vd->vdev_id,
2108 2108 msp->ms_group->mg_vd->vdev_spa->spa_name,
2109 2109 space_map_length(msp->ms_sm), avl_numnodes(&msp->ms_tree->rt_root),
2110 2110 msp->ms_condense_wanted ? "TRUE" : "FALSE");
2111 2111
2112 2112 msp->ms_condense_wanted = B_FALSE;
2113 2113
2114 2114 /*
2115 2115 * Create an range tree that is 100% allocated. We remove segments
2116 2116 * that have been freed in this txg, any deferred frees that exist,
2117 2117 * and any allocation in the future. Removing segments should be
2118 2118 * a relatively inexpensive operation since we expect these trees to
2119 2119 * have a small number of nodes.
2120 2120 */
2121 2121 condense_tree = range_tree_create(NULL, NULL, &msp->ms_lock);
2122 2122 range_tree_add(condense_tree, msp->ms_start, msp->ms_size);
2123 2123
2124 2124 /*
2125 2125 * Remove what's been freed in this txg from the condense_tree.
2126 2126 * Since we're in sync_pass 1, we know that all the frees from
2127 2127 * this txg are in the freeingtree.
2128 2128 */
2129 2129 range_tree_walk(msp->ms_freeingtree, range_tree_remove, condense_tree);
2130 2130
2131 2131 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2132 2132 range_tree_walk(msp->ms_defertree[t],
2133 2133 range_tree_remove, condense_tree);
2134 2134 }
2135 2135
2136 2136 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
2137 2137 range_tree_walk(msp->ms_alloctree[(txg + t) & TXG_MASK],
2138 2138 range_tree_remove, condense_tree);
2139 2139 }
2140 2140
2141 2141 /*
2142 2142 * We're about to drop the metaslab's lock thus allowing
2143 2143 * other consumers to change it's content. Set the
2144 2144 * metaslab's ms_condensing flag to ensure that
2145 2145 * allocations on this metaslab do not occur while we're
2146 2146 * in the middle of committing it to disk. This is only critical
2147 2147 * for the ms_tree as all other range trees use per txg
2148 2148 * views of their content.
2149 2149 */
2150 2150 msp->ms_condensing = B_TRUE;
2151 2151
2152 2152 mutex_exit(&msp->ms_lock);
2153 2153 space_map_truncate(sm, tx);
2154 2154 mutex_enter(&msp->ms_lock);
2155 2155
2156 2156 /*
2157 2157 * While we would ideally like to create a space map representation
2158 2158 * that consists only of allocation records, doing so can be
2159 2159 * prohibitively expensive because the in-core free tree can be
2160 2160 * large, and therefore computationally expensive to subtract
2161 2161 * from the condense_tree. Instead we sync out two trees, a cheap
2162 2162 * allocation only tree followed by the in-core free tree. While not
2163 2163 * optimal, this is typically close to optimal, and much cheaper to
2164 2164 * compute.
2165 2165 */
2166 2166 space_map_write(sm, condense_tree, SM_ALLOC, tx);
2167 2167 range_tree_vacate(condense_tree, NULL, NULL);
2168 2168 range_tree_destroy(condense_tree);
2169 2169
2170 2170 space_map_write(sm, msp->ms_tree, SM_FREE, tx);
2171 2171 msp->ms_condensing = B_FALSE;
2172 2172 }
2173 2173
2174 2174 /*
2175 2175 * Write a metaslab to disk in the context of the specified transaction group.
2176 2176 */
2177 2177 void
2178 2178 metaslab_sync(metaslab_t *msp, uint64_t txg)
2179 2179 {
2180 2180 metaslab_group_t *mg = msp->ms_group;
2181 2181 vdev_t *vd = mg->mg_vd;
2182 2182 spa_t *spa = vd->vdev_spa;
2183 2183 objset_t *mos = spa_meta_objset(spa);
2184 2184 range_tree_t *alloctree = msp->ms_alloctree[txg & TXG_MASK];
2185 2185 dmu_tx_t *tx;
2186 2186 uint64_t object = space_map_object(msp->ms_sm);
2187 2187
2188 2188 ASSERT(!vd->vdev_ishole);
2189 2189
2190 2190 /*
2191 2191 * This metaslab has just been added so there's no work to do now.
2192 2192 */
2193 2193 if (msp->ms_freeingtree == NULL) {
2194 2194 ASSERT3P(alloctree, ==, NULL);
2195 2195 return;
2196 2196 }
2197 2197
2198 2198 ASSERT3P(alloctree, !=, NULL);
2199 2199 ASSERT3P(msp->ms_freeingtree, !=, NULL);
2200 2200 ASSERT3P(msp->ms_freedtree, !=, NULL);
2201 2201
2202 2202 /*
2203 2203 * Normally, we don't want to process a metaslab if there
2204 2204 * are no allocations or frees to perform. However, if the metaslab
2205 2205 * is being forced to condense and it's loaded, we need to let it
2206 2206 * through.
2207 2207 */
2208 2208 if (range_tree_space(alloctree) == 0 &&
2209 2209 range_tree_space(msp->ms_freeingtree) == 0 &&
2210 2210 !(msp->ms_loaded && msp->ms_condense_wanted))
2211 2211 return;
2212 2212
2213 2213
2214 2214 VERIFY(txg <= spa_final_dirty_txg(spa));
2215 2215
2216 2216 /*
2217 2217 * The only state that can actually be changing concurrently with
2218 2218 * metaslab_sync() is the metaslab's ms_tree. No other thread can
2219 2219 * be modifying this txg's alloctree, freeingtree, freedtree, or
2220 2220 * space_map_phys_t. Therefore, we only hold ms_lock to satify
2221 2221 * space map ASSERTs. We drop it whenever we call into the DMU,
2222 2222 * because the DMU can call down to us (e.g. via zio_free()) at
2223 2223 * any time.
2224 2224 */
2225 2225
2226 2226 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2227 2227
2228 2228 if (msp->ms_sm == NULL) {
2229 2229 uint64_t new_object;
2230 2230
2231 2231 new_object = space_map_alloc(mos, tx);
2232 2232 VERIFY3U(new_object, !=, 0);
2233 2233
2234 2234 VERIFY0(space_map_open(&msp->ms_sm, mos, new_object,
2235 2235 msp->ms_start, msp->ms_size, vd->vdev_ashift,
2236 2236 &msp->ms_lock));
2237 2237 ASSERT(msp->ms_sm != NULL);
2238 2238 }
2239 2239
2240 2240 mutex_enter(&msp->ms_lock);
2241 2241
2242 2242 /*
2243 2243 * Note: metaslab_condense() clears the space map's histogram.
2244 2244 * Therefore we must verify and remove this histogram before
2245 2245 * condensing.
2246 2246 */
2247 2247 metaslab_group_histogram_verify(mg);
2248 2248 metaslab_class_histogram_verify(mg->mg_class);
2249 2249 metaslab_group_histogram_remove(mg, msp);
2250 2250
2251 2251 if (msp->ms_loaded && spa_sync_pass(spa) == 1 &&
2252 2252 metaslab_should_condense(msp)) {
2253 2253 metaslab_condense(msp, txg, tx);
2254 2254 } else {
2255 2255 space_map_write(msp->ms_sm, alloctree, SM_ALLOC, tx);
2256 2256 space_map_write(msp->ms_sm, msp->ms_freeingtree, SM_FREE, tx);
2257 2257 }
2258 2258
2259 2259 if (msp->ms_loaded) {
2260 2260 /*
2261 2261 * When the space map is loaded, we have an accruate
2262 2262 * histogram in the range tree. This gives us an opportunity
2263 2263 * to bring the space map's histogram up-to-date so we clear
2264 2264 * it first before updating it.
2265 2265 */
2266 2266 space_map_histogram_clear(msp->ms_sm);
2267 2267 space_map_histogram_add(msp->ms_sm, msp->ms_tree, tx);
2268 2268
2269 2269 /*
2270 2270 * Since we've cleared the histogram we need to add back
2271 2271 * any free space that has already been processed, plus
2272 2272 * any deferred space. This allows the on-disk histogram
2273 2273 * to accurately reflect all free space even if some space
2274 2274 * is not yet available for allocation (i.e. deferred).
2275 2275 */
2276 2276 space_map_histogram_add(msp->ms_sm, msp->ms_freedtree, tx);
2277 2277
2278 2278 /*
2279 2279 * Add back any deferred free space that has not been
2280 2280 * added back into the in-core free tree yet. This will
2281 2281 * ensure that we don't end up with a space map histogram
2282 2282 * that is completely empty unless the metaslab is fully
2283 2283 * allocated.
2284 2284 */
2285 2285 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2286 2286 space_map_histogram_add(msp->ms_sm,
2287 2287 msp->ms_defertree[t], tx);
2288 2288 }
2289 2289 }
2290 2290
2291 2291 /*
2292 2292 * Always add the free space from this sync pass to the space
2293 2293 * map histogram. We want to make sure that the on-disk histogram
2294 2294 * accounts for all free space. If the space map is not loaded,
2295 2295 * then we will lose some accuracy but will correct it the next
2296 2296 * time we load the space map.
2297 2297 */
2298 2298 space_map_histogram_add(msp->ms_sm, msp->ms_freeingtree, tx);
2299 2299
2300 2300 metaslab_group_histogram_add(mg, msp);
2301 2301 metaslab_group_histogram_verify(mg);
2302 2302 metaslab_class_histogram_verify(mg->mg_class);
2303 2303
2304 2304 /*
2305 2305 * For sync pass 1, we avoid traversing this txg's free range tree
2306 2306 * and instead will just swap the pointers for freeingtree and
2307 2307 * freedtree. We can safely do this since the freed_tree is
2308 2308 * guaranteed to be empty on the initial pass.
2309 2309 */
2310 2310 if (spa_sync_pass(spa) == 1) {
2311 2311 range_tree_swap(&msp->ms_freeingtree, &msp->ms_freedtree);
2312 2312 } else {
2313 2313 range_tree_vacate(msp->ms_freeingtree,
2314 2314 range_tree_add, msp->ms_freedtree);
2315 2315 }
2316 2316 range_tree_vacate(alloctree, NULL, NULL);
2317 2317
2318 2318 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
2319 2319 ASSERT0(range_tree_space(msp->ms_alloctree[TXG_CLEAN(txg) & TXG_MASK]));
2320 2320 ASSERT0(range_tree_space(msp->ms_freeingtree));
2321 2321
2322 2322 mutex_exit(&msp->ms_lock);
2323 2323
2324 2324 if (object != space_map_object(msp->ms_sm)) {
2325 2325 object = space_map_object(msp->ms_sm);
2326 2326 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
2327 2327 msp->ms_id, sizeof (uint64_t), &object, tx);
2328 2328 }
2329 2329 dmu_tx_commit(tx);
2330 2330 }
2331 2331
2332 2332 /*
2333 2333 * Called after a transaction group has completely synced to mark
2334 2334 * all of the metaslab's free space as usable.
2335 2335 */
2336 2336 void
2337 2337 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
2338 2338 {
2339 2339 metaslab_group_t *mg = msp->ms_group;
2340 2340 vdev_t *vd = mg->mg_vd;
2341 2341 spa_t *spa = vd->vdev_spa;
2342 2342 range_tree_t **defer_tree;
2343 2343 int64_t alloc_delta, defer_delta;
2344 2344 boolean_t defer_allowed = B_TRUE;
2345 2345
2346 2346 ASSERT(!vd->vdev_ishole);
2347 2347
2348 2348 mutex_enter(&msp->ms_lock);
2349 2349
2350 2350 /*
2351 2351 * If this metaslab is just becoming available, initialize its
2352 2352 * range trees and add its capacity to the vdev.
2353 2353 */
2354 2354 if (msp->ms_freedtree == NULL) {
2355 2355 for (int t = 0; t < TXG_SIZE; t++) {
2356 2356 ASSERT(msp->ms_alloctree[t] == NULL);
2357 2357
2358 2358 msp->ms_alloctree[t] = range_tree_create(NULL, msp,
2359 2359 &msp->ms_lock);
2360 2360 }
2361 2361
2362 2362 ASSERT3P(msp->ms_freeingtree, ==, NULL);
2363 2363 msp->ms_freeingtree = range_tree_create(NULL, msp,
2364 2364 &msp->ms_lock);
2365 2365
2366 2366 ASSERT3P(msp->ms_freedtree, ==, NULL);
2367 2367 msp->ms_freedtree = range_tree_create(NULL, msp,
2368 2368 &msp->ms_lock);
2369 2369
2370 2370 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2371 2371 ASSERT(msp->ms_defertree[t] == NULL);
2372 2372
2373 2373 msp->ms_defertree[t] = range_tree_create(NULL, msp,
2374 2374 &msp->ms_lock);
2375 2375 }
2376 2376
2377 2377 vdev_space_update(vd, 0, 0, msp->ms_size);
2378 2378 }
2379 2379
2380 2380 defer_tree = &msp->ms_defertree[txg % TXG_DEFER_SIZE];
2381 2381
2382 2382 uint64_t free_space = metaslab_class_get_space(spa_normal_class(spa)) -
2383 2383 metaslab_class_get_alloc(spa_normal_class(spa));
2384 2384 if (free_space <= spa_get_slop_space(spa)) {
2385 2385 defer_allowed = B_FALSE;
2386 2386 }
2387 2387
2388 2388 defer_delta = 0;
2389 2389 alloc_delta = space_map_alloc_delta(msp->ms_sm);
2390 2390 if (defer_allowed) {
2391 2391 defer_delta = range_tree_space(msp->ms_freedtree) -
2392 2392 range_tree_space(*defer_tree);
2393 2393 } else {
2394 2394 defer_delta -= range_tree_space(*defer_tree);
2395 2395 }
2396 2396
2397 2397 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
2398 2398
2399 2399 /*
2400 2400 * If there's a metaslab_load() in progress, wait for it to complete
2401 2401 * so that we have a consistent view of the in-core space map.
2402 2402 */
2403 2403 metaslab_load_wait(msp);
2404 2404
2405 2405 /*
2406 2406 * Move the frees from the defer_tree back to the free
2407 2407 * range tree (if it's loaded). Swap the freed_tree and the
2408 2408 * defer_tree -- this is safe to do because we've just emptied out
2409 2409 * the defer_tree.
2410 2410 */
2411 2411 range_tree_vacate(*defer_tree,
2412 2412 msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
2413 2413 if (defer_allowed) {
2414 2414 range_tree_swap(&msp->ms_freedtree, defer_tree);
2415 2415 } else {
2416 2416 range_tree_vacate(msp->ms_freedtree,
2417 2417 msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
2418 2418 }
2419 2419
2420 2420 space_map_update(msp->ms_sm);
2421 2421
2422 2422 msp->ms_deferspace += defer_delta;
2423 2423 ASSERT3S(msp->ms_deferspace, >=, 0);
2424 2424 ASSERT3S(msp->ms_deferspace, <=, msp->ms_size);
2425 2425 if (msp->ms_deferspace != 0) {
2426 2426 /*
2427 2427 * Keep syncing this metaslab until all deferred frees
2428 2428 * are back in circulation.
2429 2429 */
2430 2430 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
2431 2431 }
2432 2432
2433 2433 /*
2434 2434 * Calculate the new weights before unloading any metaslabs.
2435 2435 * This will give us the most accurate weighting.
2436 2436 */
2437 2437 metaslab_group_sort(mg, msp, metaslab_weight(msp));
2438 2438
2439 2439 /*
2440 2440 * If the metaslab is loaded and we've not tried to load or allocate
2441 2441 * from it in 'metaslab_unload_delay' txgs, then unload it.
2442 2442 */
2443 2443 if (msp->ms_loaded &&
2444 2444 msp->ms_selected_txg + metaslab_unload_delay < txg) {
2445 2445 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
2446 2446 VERIFY0(range_tree_space(
2447 2447 msp->ms_alloctree[(txg + t) & TXG_MASK]));
2448 2448 }
2449 2449
2450 2450 if (!metaslab_debug_unload)
2451 2451 metaslab_unload(msp);
2452 2452 }
2453 2453
2454 2454 mutex_exit(&msp->ms_lock);
2455 2455 }
2456 2456
2457 2457 void
2458 2458 metaslab_sync_reassess(metaslab_group_t *mg)
2459 2459 {
2460 2460 metaslab_group_alloc_update(mg);
2461 2461 mg->mg_fragmentation = metaslab_group_fragmentation(mg);
2462 2462
2463 2463 /*
2464 2464 * Preload the next potential metaslabs
2465 2465 */
2466 2466 metaslab_group_preload(mg);
2467 2467 }
2468 2468
2469 2469 static uint64_t
2470 2470 metaslab_distance(metaslab_t *msp, dva_t *dva)
2471 2471 {
2472 2472 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
2473 2473 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
2474 2474 uint64_t start = msp->ms_id;
2475 2475
2476 2476 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
2477 2477 return (1ULL << 63);
2478 2478
2479 2479 if (offset < start)
2480 2480 return ((start - offset) << ms_shift);
2481 2481 if (offset > start)
2482 2482 return ((offset - start) << ms_shift);
2483 2483 return (0);
2484 2484 }
2485 2485
2486 2486 /*
2487 2487 * ==========================================================================
2488 2488 * Metaslab allocation tracing facility
2489 2489 * ==========================================================================
2490 2490 */
2491 2491 kstat_t *metaslab_trace_ksp;
2492 2492 kstat_named_t metaslab_trace_over_limit;
2493 2493
2494 2494 void
2495 2495 metaslab_alloc_trace_init(void)
2496 2496 {
2497 2497 ASSERT(metaslab_alloc_trace_cache == NULL);
2498 2498 metaslab_alloc_trace_cache = kmem_cache_create(
2499 2499 "metaslab_alloc_trace_cache", sizeof (metaslab_alloc_trace_t),
2500 2500 0, NULL, NULL, NULL, NULL, NULL, 0);
2501 2501 metaslab_trace_ksp = kstat_create("zfs", 0, "metaslab_trace_stats",
2502 2502 "misc", KSTAT_TYPE_NAMED, 1, KSTAT_FLAG_VIRTUAL);
2503 2503 if (metaslab_trace_ksp != NULL) {
2504 2504 metaslab_trace_ksp->ks_data = &metaslab_trace_over_limit;
2505 2505 kstat_named_init(&metaslab_trace_over_limit,
2506 2506 "metaslab_trace_over_limit", KSTAT_DATA_UINT64);
2507 2507 kstat_install(metaslab_trace_ksp);
2508 2508 }
2509 2509 }
2510 2510
2511 2511 void
2512 2512 metaslab_alloc_trace_fini(void)
2513 2513 {
2514 2514 if (metaslab_trace_ksp != NULL) {
2515 2515 kstat_delete(metaslab_trace_ksp);
2516 2516 metaslab_trace_ksp = NULL;
2517 2517 }
2518 2518 kmem_cache_destroy(metaslab_alloc_trace_cache);
2519 2519 metaslab_alloc_trace_cache = NULL;
2520 2520 }
2521 2521
2522 2522 /*
2523 2523 * Add an allocation trace element to the allocation tracing list.
2524 2524 */
2525 2525 static void
2526 2526 metaslab_trace_add(zio_alloc_list_t *zal, metaslab_group_t *mg,
2527 2527 metaslab_t *msp, uint64_t psize, uint32_t dva_id, uint64_t offset)
2528 2528 {
2529 2529 if (!metaslab_trace_enabled)
2530 2530 return;
2531 2531
2532 2532 /*
2533 2533 * When the tracing list reaches its maximum we remove
2534 2534 * the second element in the list before adding a new one.
2535 2535 * By removing the second element we preserve the original
2536 2536 * entry as a clue to what allocations steps have already been
2537 2537 * performed.
2538 2538 */
2539 2539 if (zal->zal_size == metaslab_trace_max_entries) {
2540 2540 metaslab_alloc_trace_t *mat_next;
2541 2541 #ifdef DEBUG
2542 2542 panic("too many entries in allocation list");
2543 2543 #endif
2544 2544 atomic_inc_64(&metaslab_trace_over_limit.value.ui64);
2545 2545 zal->zal_size--;
2546 2546 mat_next = list_next(&zal->zal_list, list_head(&zal->zal_list));
2547 2547 list_remove(&zal->zal_list, mat_next);
2548 2548 kmem_cache_free(metaslab_alloc_trace_cache, mat_next);
2549 2549 }
2550 2550
2551 2551 metaslab_alloc_trace_t *mat =
2552 2552 kmem_cache_alloc(metaslab_alloc_trace_cache, KM_SLEEP);
2553 2553 list_link_init(&mat->mat_list_node);
2554 2554 mat->mat_mg = mg;
2555 2555 mat->mat_msp = msp;
2556 2556 mat->mat_size = psize;
2557 2557 mat->mat_dva_id = dva_id;
2558 2558 mat->mat_offset = offset;
2559 2559 mat->mat_weight = 0;
2560 2560
2561 2561 if (msp != NULL)
2562 2562 mat->mat_weight = msp->ms_weight;
2563 2563
2564 2564 /*
2565 2565 * The list is part of the zio so locking is not required. Only
2566 2566 * a single thread will perform allocations for a given zio.
2567 2567 */
2568 2568 list_insert_tail(&zal->zal_list, mat);
2569 2569 zal->zal_size++;
2570 2570
2571 2571 ASSERT3U(zal->zal_size, <=, metaslab_trace_max_entries);
2572 2572 }
2573 2573
2574 2574 void
2575 2575 metaslab_trace_init(zio_alloc_list_t *zal)
2576 2576 {
2577 2577 list_create(&zal->zal_list, sizeof (metaslab_alloc_trace_t),
2578 2578 offsetof(metaslab_alloc_trace_t, mat_list_node));
2579 2579 zal->zal_size = 0;
2580 2580 }
2581 2581
2582 2582 void
2583 2583 metaslab_trace_fini(zio_alloc_list_t *zal)
2584 2584 {
2585 2585 metaslab_alloc_trace_t *mat;
2586 2586
2587 2587 while ((mat = list_remove_head(&zal->zal_list)) != NULL)
2588 2588 kmem_cache_free(metaslab_alloc_trace_cache, mat);
2589 2589 list_destroy(&zal->zal_list);
2590 2590 zal->zal_size = 0;
2591 2591 }
2592 2592
2593 2593 /*
2594 2594 * ==========================================================================
2595 2595 * Metaslab block operations
2596 2596 * ==========================================================================
2597 2597 */
2598 2598
2599 2599 static void
2600 2600 metaslab_group_alloc_increment(spa_t *spa, uint64_t vdev, void *tag, int flags)
2601 2601 {
2602 2602 if (!(flags & METASLAB_ASYNC_ALLOC) ||
2603 2603 flags & METASLAB_DONT_THROTTLE)
2604 2604 return;
2605 2605
2606 2606 metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
2607 2607 if (!mg->mg_class->mc_alloc_throttle_enabled)
2608 2608 return;
2609 2609
2610 2610 (void) refcount_add(&mg->mg_alloc_queue_depth, tag);
2611 2611 }
2612 2612
2613 2613 void
2614 2614 metaslab_group_alloc_decrement(spa_t *spa, uint64_t vdev, void *tag, int flags)
2615 2615 {
2616 2616 if (!(flags & METASLAB_ASYNC_ALLOC) ||
2617 2617 flags & METASLAB_DONT_THROTTLE)
2618 2618 return;
2619 2619
2620 2620 metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
2621 2621 if (!mg->mg_class->mc_alloc_throttle_enabled)
2622 2622 return;
2623 2623
2624 2624 (void) refcount_remove(&mg->mg_alloc_queue_depth, tag);
2625 2625 }
2626 2626
2627 2627 void
2628 2628 metaslab_group_alloc_verify(spa_t *spa, const blkptr_t *bp, void *tag)
2629 2629 {
2630 2630 #ifdef ZFS_DEBUG
2631 2631 const dva_t *dva = bp->blk_dva;
2632 2632 int ndvas = BP_GET_NDVAS(bp);
2633 2633
2634 2634 for (int d = 0; d < ndvas; d++) {
2635 2635 uint64_t vdev = DVA_GET_VDEV(&dva[d]);
2636 2636 metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
2637 2637 VERIFY(refcount_not_held(&mg->mg_alloc_queue_depth, tag));
2638 2638 }
2639 2639 #endif
2640 2640 }
2641 2641
2642 2642 static uint64_t
2643 2643 metaslab_block_alloc(metaslab_t *msp, uint64_t size, uint64_t txg)
2644 2644 {
2645 2645 uint64_t start;
2646 2646 range_tree_t *rt = msp->ms_tree;
2647 2647 metaslab_class_t *mc = msp->ms_group->mg_class;
2648 2648
2649 2649 VERIFY(!msp->ms_condensing);
2650 2650
2651 2651 start = mc->mc_ops->msop_alloc(msp, size);
2652 2652 if (start != -1ULL) {
2653 2653 metaslab_group_t *mg = msp->ms_group;
2654 2654 vdev_t *vd = mg->mg_vd;
2655 2655
2656 2656 VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
2657 2657 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2658 2658 VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
2659 2659 range_tree_remove(rt, start, size);
2660 2660
2661 2661 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
2662 2662 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
2663 2663
2664 2664 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], start, size);
2665 2665
2666 2666 /* Track the last successful allocation */
2667 2667 msp->ms_alloc_txg = txg;
2668 2668 metaslab_verify_space(msp, txg);
2669 2669 }
2670 2670
2671 2671 /*
2672 2672 * Now that we've attempted the allocation we need to update the
2673 2673 * metaslab's maximum block size since it may have changed.
2674 2674 */
2675 2675 msp->ms_max_size = metaslab_block_maxsize(msp);
2676 2676 return (start);
2677 2677 }
2678 2678
2679 2679 static uint64_t
2680 2680 metaslab_group_alloc_normal(metaslab_group_t *mg, zio_alloc_list_t *zal,
2681 2681 uint64_t asize, uint64_t txg, uint64_t min_distance, dva_t *dva, int d)
2682 2682 {
2683 2683 metaslab_t *msp = NULL;
2684 2684 uint64_t offset = -1ULL;
2685 2685 uint64_t activation_weight;
2686 2686 uint64_t target_distance;
2687 2687 int i;
2688 2688
2689 2689 activation_weight = METASLAB_WEIGHT_PRIMARY;
2690 2690 for (i = 0; i < d; i++) {
2691 2691 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
2692 2692 activation_weight = METASLAB_WEIGHT_SECONDARY;
2693 2693 break;
2694 2694 }
2695 2695 }
2696 2696
2697 2697 metaslab_t *search = kmem_alloc(sizeof (*search), KM_SLEEP);
2698 2698 search->ms_weight = UINT64_MAX;
2699 2699 search->ms_start = 0;
2700 2700 for (;;) {
2701 2701 boolean_t was_active;
2702 2702 avl_tree_t *t = &mg->mg_metaslab_tree;
2703 2703 avl_index_t idx;
2704 2704
2705 2705 mutex_enter(&mg->mg_lock);
2706 2706
2707 2707 /*
2708 2708 * Find the metaslab with the highest weight that is less
2709 2709 * than what we've already tried. In the common case, this
2710 2710 * means that we will examine each metaslab at most once.
2711 2711 * Note that concurrent callers could reorder metaslabs
2712 2712 * by activation/passivation once we have dropped the mg_lock.
2713 2713 * If a metaslab is activated by another thread, and we fail
2714 2714 * to allocate from the metaslab we have selected, we may
2715 2715 * not try the newly-activated metaslab, and instead activate
2716 2716 * another metaslab. This is not optimal, but generally
2717 2717 * does not cause any problems (a possible exception being
2718 2718 * if every metaslab is completely full except for the
2719 2719 * the newly-activated metaslab which we fail to examine).
2720 2720 */
2721 2721 msp = avl_find(t, search, &idx);
2722 2722 if (msp == NULL)
2723 2723 msp = avl_nearest(t, idx, AVL_AFTER);
2724 2724 for (; msp != NULL; msp = AVL_NEXT(t, msp)) {
2725 2725
2726 2726 if (!metaslab_should_allocate(msp, asize)) {
2727 2727 metaslab_trace_add(zal, mg, msp, asize, d,
2728 2728 TRACE_TOO_SMALL);
2729 2729 continue;
2730 2730 }
2731 2731
2732 2732 /*
2733 2733 * If the selected metaslab is condensing, skip it.
2734 2734 */
2735 2735 if (msp->ms_condensing)
2736 2736 continue;
2737 2737
2738 2738 was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
2739 2739 if (activation_weight == METASLAB_WEIGHT_PRIMARY)
2740 2740 break;
2741 2741
2742 2742 target_distance = min_distance +
2743 2743 (space_map_allocated(msp->ms_sm) != 0 ? 0 :
2744 2744 min_distance >> 1);
2745 2745
2746 2746 for (i = 0; i < d; i++) {
2747 2747 if (metaslab_distance(msp, &dva[i]) <
2748 2748 target_distance)
2749 2749 break;
2750 2750 }
2751 2751 if (i == d)
2752 2752 break;
2753 2753 }
2754 2754 mutex_exit(&mg->mg_lock);
2755 2755 if (msp == NULL) {
2756 2756 kmem_free(search, sizeof (*search));
2757 2757 return (-1ULL);
2758 2758 }
2759 2759 search->ms_weight = msp->ms_weight;
2760 2760 search->ms_start = msp->ms_start + 1;
2761 2761
2762 2762 mutex_enter(&msp->ms_lock);
2763 2763
2764 2764 /*
2765 2765 * Ensure that the metaslab we have selected is still
2766 2766 * capable of handling our request. It's possible that
2767 2767 * another thread may have changed the weight while we
2768 2768 * were blocked on the metaslab lock. We check the
2769 2769 * active status first to see if we need to reselect
2770 2770 * a new metaslab.
2771 2771 */
2772 2772 if (was_active && !(msp->ms_weight & METASLAB_ACTIVE_MASK)) {
2773 2773 mutex_exit(&msp->ms_lock);
2774 2774 continue;
2775 2775 }
2776 2776
2777 2777 if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
2778 2778 activation_weight == METASLAB_WEIGHT_PRIMARY) {
2779 2779 metaslab_passivate(msp,
2780 2780 msp->ms_weight & ~METASLAB_ACTIVE_MASK);
2781 2781 mutex_exit(&msp->ms_lock);
2782 2782 continue;
2783 2783 }
2784 2784
2785 2785 if (metaslab_activate(msp, activation_weight) != 0) {
2786 2786 mutex_exit(&msp->ms_lock);
2787 2787 continue;
2788 2788 }
2789 2789 msp->ms_selected_txg = txg;
2790 2790
2791 2791 /*
2792 2792 * Now that we have the lock, recheck to see if we should
2793 2793 * continue to use this metaslab for this allocation. The
2794 2794 * the metaslab is now loaded so metaslab_should_allocate() can
2795 2795 * accurately determine if the allocation attempt should
2796 2796 * proceed.
2797 2797 */
2798 2798 if (!metaslab_should_allocate(msp, asize)) {
2799 2799 /* Passivate this metaslab and select a new one. */
2800 2800 metaslab_trace_add(zal, mg, msp, asize, d,
2801 2801 TRACE_TOO_SMALL);
2802 2802 goto next;
2803 2803 }
2804 2804
2805 2805 /*
2806 2806 * If this metaslab is currently condensing then pick again as
2807 2807 * we can't manipulate this metaslab until it's committed
2808 2808 * to disk.
2809 2809 */
2810 2810 if (msp->ms_condensing) {
2811 2811 metaslab_trace_add(zal, mg, msp, asize, d,
2812 2812 TRACE_CONDENSING);
2813 2813 mutex_exit(&msp->ms_lock);
2814 2814 continue;
2815 2815 }
2816 2816
2817 2817 offset = metaslab_block_alloc(msp, asize, txg);
2818 2818 metaslab_trace_add(zal, mg, msp, asize, d, offset);
2819 2819
2820 2820 if (offset != -1ULL) {
2821 2821 /* Proactively passivate the metaslab, if needed */
2822 2822 metaslab_segment_may_passivate(msp);
2823 2823 break;
2824 2824 }
2825 2825 next:
2826 2826 ASSERT(msp->ms_loaded);
2827 2827
2828 2828 /*
2829 2829 * We were unable to allocate from this metaslab so determine
2830 2830 * a new weight for this metaslab. Now that we have loaded
2831 2831 * the metaslab we can provide a better hint to the metaslab
2832 2832 * selector.
2833 2833 *
2834 2834 * For space-based metaslabs, we use the maximum block size.
2835 2835 * This information is only available when the metaslab
2836 2836 * is loaded and is more accurate than the generic free
2837 2837 * space weight that was calculated by metaslab_weight().
2838 2838 * This information allows us to quickly compare the maximum
2839 2839 * available allocation in the metaslab to the allocation
2840 2840 * size being requested.
2841 2841 *
2842 2842 * For segment-based metaslabs, determine the new weight
2843 2843 * based on the highest bucket in the range tree. We
2844 2844 * explicitly use the loaded segment weight (i.e. the range
2845 2845 * tree histogram) since it contains the space that is
2846 2846 * currently available for allocation and is accurate
2847 2847 * even within a sync pass.
2848 2848 */
2849 2849 if (WEIGHT_IS_SPACEBASED(msp->ms_weight)) {
2850 2850 uint64_t weight = metaslab_block_maxsize(msp);
2851 2851 WEIGHT_SET_SPACEBASED(weight);
2852 2852 metaslab_passivate(msp, weight);
2853 2853 } else {
2854 2854 metaslab_passivate(msp,
2855 2855 metaslab_weight_from_range_tree(msp));
2856 2856 }
2857 2857
2858 2858 /*
2859 2859 * We have just failed an allocation attempt, check
2860 2860 * that metaslab_should_allocate() agrees. Otherwise,
2861 2861 * we may end up in an infinite loop retrying the same
2862 2862 * metaslab.
2863 2863 */
2864 2864 ASSERT(!metaslab_should_allocate(msp, asize));
2865 2865 mutex_exit(&msp->ms_lock);
2866 2866 }
2867 2867 mutex_exit(&msp->ms_lock);
2868 2868 kmem_free(search, sizeof (*search));
2869 2869 return (offset);
2870 2870 }
2871 2871
2872 2872 static uint64_t
2873 2873 metaslab_group_alloc(metaslab_group_t *mg, zio_alloc_list_t *zal,
2874 2874 uint64_t asize, uint64_t txg, uint64_t min_distance, dva_t *dva, int d)
2875 2875 {
2876 2876 uint64_t offset;
2877 2877 ASSERT(mg->mg_initialized);
2878 2878
2879 2879 offset = metaslab_group_alloc_normal(mg, zal, asize, txg,
2880 2880 min_distance, dva, d);
2881 2881
2882 2882 mutex_enter(&mg->mg_lock);
2883 2883 if (offset == -1ULL) {
2884 2884 mg->mg_failed_allocations++;
2885 2885 metaslab_trace_add(zal, mg, NULL, asize, d,
2886 2886 TRACE_GROUP_FAILURE);
2887 2887 if (asize == SPA_GANGBLOCKSIZE) {
2888 2888 /*
2889 2889 * This metaslab group was unable to allocate
2890 2890 * the minimum gang block size so it must be out of
2891 2891 * space. We must notify the allocation throttle
2892 2892 * to start skipping allocation attempts to this
2893 2893 * metaslab group until more space becomes available.
2894 2894 * Note: this failure cannot be caused by the
2895 2895 * allocation throttle since the allocation throttle
2896 2896 * is only responsible for skipping devices and
2897 2897 * not failing block allocations.
2898 2898 */
2899 2899 mg->mg_no_free_space = B_TRUE;
2900 2900 }
2901 2901 }
2902 2902 mg->mg_allocations++;
2903 2903 mutex_exit(&mg->mg_lock);
2904 2904 return (offset);
2905 2905 }
2906 2906
2907 2907 /*
2908 2908 * If we have to write a ditto block (i.e. more than one DVA for a given BP)
2909 2909 * on the same vdev as an existing DVA of this BP, then try to allocate it
2910 2910 * at least (vdev_asize / (2 ^ ditto_same_vdev_distance_shift)) away from the
2911 2911 * existing DVAs.
2912 2912 */
2913 2913 int ditto_same_vdev_distance_shift = 3;
2914 2914
2915 2915 /*
2916 2916 * Allocate a block for the specified i/o.
2917 2917 */
2918 2918 static int
2919 2919 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
2920 2920 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags,
2921 2921 zio_alloc_list_t *zal)
2922 2922 {
2923 2923 metaslab_group_t *mg, *rotor;
2924 2924 vdev_t *vd;
2925 2925 boolean_t try_hard = B_FALSE;
2926 2926
2927 2927 ASSERT(!DVA_IS_VALID(&dva[d]));
2928 2928
2929 2929 /*
2930 2930 * For testing, make some blocks above a certain size be gang blocks.
2931 2931 */
2932 2932 if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0) {
2933 2933 metaslab_trace_add(zal, NULL, NULL, psize, d, TRACE_FORCE_GANG);
2934 2934 return (SET_ERROR(ENOSPC));
2935 2935 }
2936 2936
2937 2937 /*
2938 2938 * Start at the rotor and loop through all mgs until we find something.
2939 2939 * Note that there's no locking on mc_rotor or mc_aliquot because
2940 2940 * nothing actually breaks if we miss a few updates -- we just won't
2941 2941 * allocate quite as evenly. It all balances out over time.
2942 2942 *
2943 2943 * If we are doing ditto or log blocks, try to spread them across
2944 2944 * consecutive vdevs. If we're forced to reuse a vdev before we've
2945 2945 * allocated all of our ditto blocks, then try and spread them out on
2946 2946 * that vdev as much as possible. If it turns out to not be possible,
2947 2947 * gradually lower our standards until anything becomes acceptable.
2948 2948 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
2949 2949 * gives us hope of containing our fault domains to something we're
2950 2950 * able to reason about. Otherwise, any two top-level vdev failures
2951 2951 * will guarantee the loss of data. With consecutive allocation,
2952 2952 * only two adjacent top-level vdev failures will result in data loss.
2953 2953 *
2954 2954 * If we are doing gang blocks (hintdva is non-NULL), try to keep
2955 2955 * ourselves on the same vdev as our gang block header. That
2956 2956 * way, we can hope for locality in vdev_cache, plus it makes our
2957 2957 * fault domains something tractable.
2958 2958 */
2959 2959 if (hintdva) {
2960 2960 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
2961 2961
2962 2962 /*
2963 2963 * It's possible the vdev we're using as the hint no
2964 2964 * longer exists (i.e. removed). Consult the rotor when
2965 2965 * all else fails.
2966 2966 */
2967 2967 if (vd != NULL) {
2968 2968 mg = vd->vdev_mg;
2969 2969
2970 2970 if (flags & METASLAB_HINTBP_AVOID &&
2971 2971 mg->mg_next != NULL)
2972 2972 mg = mg->mg_next;
2973 2973 } else {
2974 2974 mg = mc->mc_rotor;
2975 2975 }
2976 2976 } else if (d != 0) {
2977 2977 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
2978 2978 mg = vd->vdev_mg->mg_next;
2979 2979 } else {
2980 2980 mg = mc->mc_rotor;
2981 2981 }
2982 2982
2983 2983 /*
2984 2984 * If the hint put us into the wrong metaslab class, or into a
2985 2985 * metaslab group that has been passivated, just follow the rotor.
2986 2986 */
2987 2987 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
2988 2988 mg = mc->mc_rotor;
2989 2989
2990 2990 rotor = mg;
2991 2991 top:
2992 2992 do {
2993 2993 boolean_t allocatable;
2994 2994
2995 2995 ASSERT(mg->mg_activation_count == 1);
2996 2996 vd = mg->mg_vd;
2997 2997
2998 2998 /*
2999 2999 * Don't allocate from faulted devices.
3000 3000 */
3001 3001 if (try_hard) {
3002 3002 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
3003 3003 allocatable = vdev_allocatable(vd);
3004 3004 spa_config_exit(spa, SCL_ZIO, FTAG);
3005 3005 } else {
3006 3006 allocatable = vdev_allocatable(vd);
3007 3007 }
3008 3008
3009 3009 /*
3010 3010 * Determine if the selected metaslab group is eligible
3011 3011 * for allocations. If we're ganging then don't allow
3012 3012 * this metaslab group to skip allocations since that would
3013 3013 * inadvertently return ENOSPC and suspend the pool
3014 3014 * even though space is still available.
3015 3015 */
3016 3016 if (allocatable && !GANG_ALLOCATION(flags) && !try_hard) {
3017 3017 allocatable = metaslab_group_allocatable(mg, rotor,
3018 3018 psize);
3019 3019 }
3020 3020
3021 3021 if (!allocatable) {
3022 3022 metaslab_trace_add(zal, mg, NULL, psize, d,
3023 3023 TRACE_NOT_ALLOCATABLE);
3024 3024 goto next;
3025 3025 }
3026 3026
3027 3027 ASSERT(mg->mg_initialized);
3028 3028
3029 3029 /*
3030 3030 * Avoid writing single-copy data to a failing,
3031 3031 * non-redundant vdev, unless we've already tried all
3032 3032 * other vdevs.
3033 3033 */
3034 3034 if ((vd->vdev_stat.vs_write_errors > 0 ||
3035 3035 vd->vdev_state < VDEV_STATE_HEALTHY) &&
3036 3036 d == 0 && !try_hard && vd->vdev_children == 0) {
3037 3037 metaslab_trace_add(zal, mg, NULL, psize, d,
3038 3038 TRACE_VDEV_ERROR);
3039 3039 goto next;
3040 3040 }
3041 3041
3042 3042 ASSERT(mg->mg_class == mc);
3043 3043
3044 3044 /*
3045 3045 * If we don't need to try hard, then require that the
3046 3046 * block be 1/8th of the device away from any other DVAs
3047 3047 * in this BP. If we are trying hard, allow any offset
3048 3048 * to be used (distance=0).
3049 3049 */
3050 3050 uint64_t distance = 0;
3051 3051 if (!try_hard) {
3052 3052 distance = vd->vdev_asize >>
3053 3053 ditto_same_vdev_distance_shift;
3054 3054 if (distance <= (1ULL << vd->vdev_ms_shift))
3055 3055 distance = 0;
3056 3056 }
3057 3057
3058 3058 uint64_t asize = vdev_psize_to_asize(vd, psize);
3059 3059 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
3060 3060
3061 3061 uint64_t offset = metaslab_group_alloc(mg, zal, asize, txg,
3062 3062 distance, dva, d);
3063 3063
3064 3064 if (offset != -1ULL) {
3065 3065 /*
3066 3066 * If we've just selected this metaslab group,
3067 3067 * figure out whether the corresponding vdev is
3068 3068 * over- or under-used relative to the pool,
3069 3069 * and set an allocation bias to even it out.
3070 3070 */
3071 3071 if (mc->mc_aliquot == 0 && metaslab_bias_enabled) {
3072 3072 vdev_stat_t *vs = &vd->vdev_stat;
3073 3073 int64_t vu, cu;
3074 3074
3075 3075 vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
3076 3076 cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
3077 3077
3078 3078 /*
3079 3079 * Calculate how much more or less we should
3080 3080 * try to allocate from this device during
3081 3081 * this iteration around the rotor.
3082 3082 * For example, if a device is 80% full
3083 3083 * and the pool is 20% full then we should
3084 3084 * reduce allocations by 60% on this device.
3085 3085 *
3086 3086 * mg_bias = (20 - 80) * 512K / 100 = -307K
3087 3087 *
3088 3088 * This reduces allocations by 307K for this
3089 3089 * iteration.
3090 3090 */
3091 3091 mg->mg_bias = ((cu - vu) *
3092 3092 (int64_t)mg->mg_aliquot) / 100;
3093 3093 } else if (!metaslab_bias_enabled) {
3094 3094 mg->mg_bias = 0;
3095 3095 }
3096 3096
3097 3097 if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
3098 3098 mg->mg_aliquot + mg->mg_bias) {
3099 3099 mc->mc_rotor = mg->mg_next;
3100 3100 mc->mc_aliquot = 0;
3101 3101 }
3102 3102
3103 3103 DVA_SET_VDEV(&dva[d], vd->vdev_id);
3104 3104 DVA_SET_OFFSET(&dva[d], offset);
3105 3105 DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
3106 3106 DVA_SET_ASIZE(&dva[d], asize);
3107 3107
3108 3108 return (0);
3109 3109 }
3110 3110 next:
3111 3111 mc->mc_rotor = mg->mg_next;
3112 3112 mc->mc_aliquot = 0;
3113 3113 } while ((mg = mg->mg_next) != rotor);
3114 3114
3115 3115 /*
3116 3116 * If we haven't tried hard, do so now.
3117 3117 */
3118 3118 if (!try_hard) {
3119 3119 try_hard = B_TRUE;
3120 3120 goto top;
3121 3121 }
3122 3122
3123 3123 bzero(&dva[d], sizeof (dva_t));
3124 3124
3125 3125 metaslab_trace_add(zal, rotor, NULL, psize, d, TRACE_ENOSPC);
3126 3126 return (SET_ERROR(ENOSPC));
3127 3127 }
3128 3128
3129 3129 /*
3130 3130 * Free the block represented by DVA in the context of the specified
3131 3131 * transaction group.
3132 3132 */
3133 3133 static void
3134 3134 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
3135 3135 {
3136 3136 uint64_t vdev = DVA_GET_VDEV(dva);
3137 3137 uint64_t offset = DVA_GET_OFFSET(dva);
3138 3138 uint64_t size = DVA_GET_ASIZE(dva);
3139 3139 vdev_t *vd;
3140 3140 metaslab_t *msp;
3141 3141
3142 3142 ASSERT(DVA_IS_VALID(dva));
3143 3143
3144 3144 if (txg > spa_freeze_txg(spa))
3145 3145 return;
3146 3146
3147 3147 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
3148 3148 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
3149 3149 cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
3150 3150 (u_longlong_t)vdev, (u_longlong_t)offset);
3151 3151 ASSERT(0);
3152 3152 return;
3153 3153 }
3154 3154
3155 3155 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3156 3156
3157 3157 if (DVA_GET_GANG(dva))
3158 3158 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
3159 3159
3160 3160 mutex_enter(&msp->ms_lock);
3161 3161
3162 3162 if (now) {
3163 3163 range_tree_remove(msp->ms_alloctree[txg & TXG_MASK],
3164 3164 offset, size);
3165 3165
3166 3166 VERIFY(!msp->ms_condensing);
3167 3167 VERIFY3U(offset, >=, msp->ms_start);
3168 3168 VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size);
3169 3169 VERIFY3U(range_tree_space(msp->ms_tree) + size, <=,
3170 3170 msp->ms_size);
3171 3171 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
3172 3172 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
3173 3173 range_tree_add(msp->ms_tree, offset, size);
3174 3174 msp->ms_max_size = metaslab_block_maxsize(msp);
3175 3175 } else {
3176 3176 VERIFY3U(txg, ==, spa->spa_syncing_txg);
3177 3177 if (range_tree_space(msp->ms_freeingtree) == 0)
3178 3178 vdev_dirty(vd, VDD_METASLAB, msp, txg);
3179 3179 range_tree_add(msp->ms_freeingtree, offset, size);
3180 3180 }
3181 3181
3182 3182 mutex_exit(&msp->ms_lock);
3183 3183 }
3184 3184
3185 3185 /*
3186 3186 * Intent log support: upon opening the pool after a crash, notify the SPA
3187 3187 * of blocks that the intent log has allocated for immediate write, but
3188 3188 * which are still considered free by the SPA because the last transaction
3189 3189 * group didn't commit yet.
3190 3190 */
3191 3191 static int
3192 3192 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
3193 3193 {
3194 3194 uint64_t vdev = DVA_GET_VDEV(dva);
3195 3195 uint64_t offset = DVA_GET_OFFSET(dva);
3196 3196 uint64_t size = DVA_GET_ASIZE(dva);
3197 3197 vdev_t *vd;
3198 3198 metaslab_t *msp;
3199 3199 int error = 0;
3200 3200
3201 3201 ASSERT(DVA_IS_VALID(dva));
3202 3202
3203 3203 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
3204 3204 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
3205 3205 return (SET_ERROR(ENXIO));
3206 3206
3207 3207 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3208 3208
3209 3209 if (DVA_GET_GANG(dva))
3210 3210 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
3211 3211
3212 3212 mutex_enter(&msp->ms_lock);
3213 3213
3214 3214 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded)
3215 3215 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
3216 3216
3217 3217 if (error == 0 && !range_tree_contains(msp->ms_tree, offset, size))
3218 3218 error = SET_ERROR(ENOENT);
3219 3219
3220 3220 if (error || txg == 0) { /* txg == 0 indicates dry run */
3221 3221 mutex_exit(&msp->ms_lock);
3222 3222 return (error);
3223 3223 }
3224 3224
3225 3225 VERIFY(!msp->ms_condensing);
3226 3226 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
3227 3227 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
3228 3228 VERIFY3U(range_tree_space(msp->ms_tree) - size, <=, msp->ms_size);
3229 3229 range_tree_remove(msp->ms_tree, offset, size);
3230 3230
3231 3231 if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
3232 3232 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
3233 3233 vdev_dirty(vd, VDD_METASLAB, msp, txg);
3234 3234 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, size);
3235 3235 }
3236 3236
3237 3237 mutex_exit(&msp->ms_lock);
3238 3238
3239 3239 return (0);
3240 3240 }
3241 3241
3242 3242 /*
3243 3243 * Reserve some allocation slots. The reservation system must be called
3244 3244 * before we call into the allocator. If there aren't any available slots
3245 3245 * then the I/O will be throttled until an I/O completes and its slots are
3246 3246 * freed up. The function returns true if it was successful in placing
3247 3247 * the reservation.
3248 3248 */
3249 3249 boolean_t
3250 3250 metaslab_class_throttle_reserve(metaslab_class_t *mc, int slots, zio_t *zio,
3251 3251 int flags)
3252 3252 {
3253 3253 uint64_t available_slots = 0;
3254 3254 boolean_t slot_reserved = B_FALSE;
3255 3255
3256 3256 ASSERT(mc->mc_alloc_throttle_enabled);
3257 3257 mutex_enter(&mc->mc_lock);
3258 3258
3259 3259 uint64_t reserved_slots = refcount_count(&mc->mc_alloc_slots);
3260 3260 if (reserved_slots < mc->mc_alloc_max_slots)
3261 3261 available_slots = mc->mc_alloc_max_slots - reserved_slots;
3262 3262
3263 3263 if (slots <= available_slots || GANG_ALLOCATION(flags)) {
3264 3264 /*
3265 3265 * We reserve the slots individually so that we can unreserve
3266 3266 * them individually when an I/O completes.
3267 3267 */
3268 3268 for (int d = 0; d < slots; d++) {
3269 3269 reserved_slots = refcount_add(&mc->mc_alloc_slots, zio);
3270 3270 }
3271 3271 zio->io_flags |= ZIO_FLAG_IO_ALLOCATING;
3272 3272 slot_reserved = B_TRUE;
3273 3273 }
3274 3274
3275 3275 mutex_exit(&mc->mc_lock);
3276 3276 return (slot_reserved);
3277 3277 }
3278 3278
3279 3279 void
3280 3280 metaslab_class_throttle_unreserve(metaslab_class_t *mc, int slots, zio_t *zio)
3281 3281 {
3282 3282 ASSERT(mc->mc_alloc_throttle_enabled);
3283 3283 mutex_enter(&mc->mc_lock);
3284 3284 for (int d = 0; d < slots; d++) {
3285 3285 (void) refcount_remove(&mc->mc_alloc_slots, zio);
3286 3286 }
3287 3287 mutex_exit(&mc->mc_lock);
3288 3288 }
3289 3289
3290 3290 int
3291 3291 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
3292 3292 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags,
3293 3293 zio_alloc_list_t *zal, zio_t *zio)
3294 3294 {
3295 3295 dva_t *dva = bp->blk_dva;
3296 3296 dva_t *hintdva = hintbp->blk_dva;
3297 3297 int error = 0;
3298 3298
3299 3299 ASSERT(bp->blk_birth == 0);
3300 3300 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
3301 3301
3302 3302 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
3303 3303
3304 3304 if (mc->mc_rotor == NULL) { /* no vdevs in this class */
3305 3305 spa_config_exit(spa, SCL_ALLOC, FTAG);
3306 3306 return (SET_ERROR(ENOSPC));
3307 3307 }
3308 3308
3309 3309 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
3310 3310 ASSERT(BP_GET_NDVAS(bp) == 0);
3311 3311 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
3312 3312 ASSERT3P(zal, !=, NULL);
3313 3313
3314 3314 for (int d = 0; d < ndvas; d++) {
3315 3315 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
3316 3316 txg, flags, zal);
3317 3317 if (error != 0) {
3318 3318 for (d--; d >= 0; d--) {
3319 3319 metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
3320 3320 metaslab_group_alloc_decrement(spa,
3321 3321 DVA_GET_VDEV(&dva[d]), zio, flags);
3322 3322 bzero(&dva[d], sizeof (dva_t));
3323 3323 }
3324 3324 spa_config_exit(spa, SCL_ALLOC, FTAG);
3325 3325 return (error);
3326 3326 } else {
3327 3327 /*
3328 3328 * Update the metaslab group's queue depth
3329 3329 * based on the newly allocated dva.
3330 3330 */
3331 3331 metaslab_group_alloc_increment(spa,
3332 3332 DVA_GET_VDEV(&dva[d]), zio, flags);
3333 3333 }
3334 3334
3335 3335 }
3336 3336 ASSERT(error == 0);
3337 3337 ASSERT(BP_GET_NDVAS(bp) == ndvas);
3338 3338
3339 3339 spa_config_exit(spa, SCL_ALLOC, FTAG);
3340 3340
3341 3341 BP_SET_BIRTH(bp, txg, txg);
3342 3342
3343 3343 return (0);
3344 3344 }
3345 3345
3346 3346 void
3347 3347 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
3348 3348 {
3349 3349 const dva_t *dva = bp->blk_dva;
3350 3350 int ndvas = BP_GET_NDVAS(bp);
3351 3351
3352 3352 ASSERT(!BP_IS_HOLE(bp));
3353 3353 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
3354 3354
3355 3355 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
3356 3356
3357 3357 for (int d = 0; d < ndvas; d++)
3358 3358 metaslab_free_dva(spa, &dva[d], txg, now);
3359 3359
3360 3360 spa_config_exit(spa, SCL_FREE, FTAG);
3361 3361 }
3362 3362
3363 3363 int
3364 3364 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
3365 3365 {
3366 3366 const dva_t *dva = bp->blk_dva;
3367 3367 int ndvas = BP_GET_NDVAS(bp);
3368 3368 int error = 0;
3369 3369
3370 3370 ASSERT(!BP_IS_HOLE(bp));
3371 3371
3372 3372 if (txg != 0) {
3373 3373 /*
3374 3374 * First do a dry run to make sure all DVAs are claimable,
3375 3375 * so we don't have to unwind from partial failures below.
3376 3376 */
3377 3377 if ((error = metaslab_claim(spa, bp, 0)) != 0)
3378 3378 return (error);
3379 3379 }
3380 3380
3381 3381 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
3382 3382
3383 3383 for (int d = 0; d < ndvas; d++)
3384 3384 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
3385 3385 break;
3386 3386
3387 3387 spa_config_exit(spa, SCL_ALLOC, FTAG);
3388 3388
3389 3389 ASSERT(error == 0 || txg == 0);
3390 3390
3391 3391 return (error);
3392 3392 }
3393 3393
3394 3394 void
3395 3395 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
3396 3396 {
3397 3397 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
3398 3398 return;
3399 3399
3400 3400 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
3401 3401 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
3402 3402 uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
3403 3403 vdev_t *vd = vdev_lookup_top(spa, vdev);
3404 3404 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
3405 3405 uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
3406 3406 metaslab_t *msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3407 3407
3408 3408 if (msp->ms_loaded)
3409 3409 range_tree_verify(msp->ms_tree, offset, size);
3410 3410
3411 3411 range_tree_verify(msp->ms_freeingtree, offset, size);
3412 3412 range_tree_verify(msp->ms_freedtree, offset, size);
3413 3413 for (int j = 0; j < TXG_DEFER_SIZE; j++)
3414 3414 range_tree_verify(msp->ms_defertree[j], offset, size);
3415 3415 }
3416 3416 spa_config_exit(spa, SCL_VDEV, FTAG);
3417 3417 }
|
↓ open down ↓ |
1742 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX