Print this page
4730 Destroy the metaslab group taskq in metaslab_group_destroy()
| 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) 2013 by Delphix. All rights reserved.
24 24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 25 */
26 26
27 27 #include <sys/zfs_context.h>
28 28 #include <sys/dmu.h>
29 29 #include <sys/dmu_tx.h>
30 30 #include <sys/space_map.h>
31 31 #include <sys/metaslab_impl.h>
32 32 #include <sys/vdev_impl.h>
33 33 #include <sys/zio.h>
34 34 #include <sys/spa_impl.h>
35 35
36 36 /*
37 37 * Allow allocations to switch to gang blocks quickly. We do this to
38 38 * avoid having to load lots of space_maps in a given txg. There are,
39 39 * however, some cases where we want to avoid "fast" ganging and instead
40 40 * we want to do an exhaustive search of all metaslabs on this device.
41 41 * Currently we don't allow any gang, zil, or dump device related allocations
42 42 * to "fast" gang.
43 43 */
44 44 #define CAN_FASTGANG(flags) \
45 45 (!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \
46 46 METASLAB_GANG_AVOID)))
47 47
48 48 #define METASLAB_WEIGHT_PRIMARY (1ULL << 63)
49 49 #define METASLAB_WEIGHT_SECONDARY (1ULL << 62)
50 50 #define METASLAB_ACTIVE_MASK \
51 51 (METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
52 52
53 53 uint64_t metaslab_aliquot = 512ULL << 10;
54 54 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1; /* force gang blocks */
55 55
56 56 /*
57 57 * The in-core space map representation is more compact than its on-disk form.
58 58 * The zfs_condense_pct determines how much more compact the in-core
59 59 * space_map representation must be before we compact it on-disk.
60 60 * Values should be greater than or equal to 100.
61 61 */
62 62 int zfs_condense_pct = 200;
63 63
64 64 /*
65 65 * This value defines the number of allowed allocation failures per vdev.
66 66 * If a device reaches this threshold in a given txg then we consider skipping
67 67 * allocations on that device. The value of zfs_mg_alloc_failures is computed
68 68 * in zio_init() unless it has been overridden in /etc/system.
69 69 */
70 70 int zfs_mg_alloc_failures = 0;
71 71
72 72 /*
73 73 * The zfs_mg_noalloc_threshold defines which metaslab groups should
74 74 * be eligible for allocation. The value is defined as a percentage of
75 75 * a free space. Metaslab groups that have more free space than
76 76 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
77 77 * a metaslab group's free space is less than or equal to the
78 78 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
79 79 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
80 80 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
81 81 * groups are allowed to accept allocations. Gang blocks are always
82 82 * eligible to allocate on any metaslab group. The default value of 0 means
83 83 * no metaslab group will be excluded based on this criterion.
84 84 */
85 85 int zfs_mg_noalloc_threshold = 0;
86 86
87 87 /*
88 88 * When set will load all metaslabs when pool is first opened.
89 89 */
90 90 int metaslab_debug_load = 0;
91 91
92 92 /*
93 93 * When set will prevent metaslabs from being unloaded.
94 94 */
95 95 int metaslab_debug_unload = 0;
96 96
97 97 /*
98 98 * Minimum size which forces the dynamic allocator to change
99 99 * it's allocation strategy. Once the space map cannot satisfy
100 100 * an allocation of this size then it switches to using more
101 101 * aggressive strategy (i.e search by size rather than offset).
102 102 */
103 103 uint64_t metaslab_df_alloc_threshold = SPA_MAXBLOCKSIZE;
104 104
105 105 /*
106 106 * The minimum free space, in percent, which must be available
107 107 * in a space map to continue allocations in a first-fit fashion.
108 108 * Once the space_map's free space drops below this level we dynamically
109 109 * switch to using best-fit allocations.
110 110 */
111 111 int metaslab_df_free_pct = 4;
112 112
113 113 /*
114 114 * A metaslab is considered "free" if it contains a contiguous
115 115 * segment which is greater than metaslab_min_alloc_size.
116 116 */
117 117 uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
118 118
119 119 /*
120 120 * Percentage of all cpus that can be used by the metaslab taskq.
121 121 */
122 122 int metaslab_load_pct = 50;
123 123
124 124 /*
125 125 * Determines how many txgs a metaslab may remain loaded without having any
126 126 * allocations from it. As long as a metaslab continues to be used we will
127 127 * keep it loaded.
128 128 */
129 129 int metaslab_unload_delay = TXG_SIZE * 2;
130 130
131 131 /*
132 132 * Should we be willing to write data to degraded vdevs?
133 133 */
134 134 boolean_t zfs_write_to_degraded = B_FALSE;
135 135
136 136 /*
137 137 * Max number of metaslabs per group to preload.
138 138 */
139 139 int metaslab_preload_limit = SPA_DVAS_PER_BP;
140 140
141 141 /*
142 142 * Enable/disable preloading of metaslab.
143 143 */
144 144 boolean_t metaslab_preload_enabled = B_TRUE;
145 145
146 146 /*
147 147 * Enable/disable additional weight factor for each metaslab.
148 148 */
149 149 boolean_t metaslab_weight_factor_enable = B_FALSE;
150 150
151 151
152 152 /*
153 153 * ==========================================================================
154 154 * Metaslab classes
155 155 * ==========================================================================
156 156 */
157 157 metaslab_class_t *
158 158 metaslab_class_create(spa_t *spa, metaslab_ops_t *ops)
159 159 {
160 160 metaslab_class_t *mc;
161 161
162 162 mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
163 163
164 164 mc->mc_spa = spa;
165 165 mc->mc_rotor = NULL;
166 166 mc->mc_ops = ops;
167 167
168 168 return (mc);
169 169 }
170 170
171 171 void
172 172 metaslab_class_destroy(metaslab_class_t *mc)
173 173 {
174 174 ASSERT(mc->mc_rotor == NULL);
175 175 ASSERT(mc->mc_alloc == 0);
176 176 ASSERT(mc->mc_deferred == 0);
177 177 ASSERT(mc->mc_space == 0);
178 178 ASSERT(mc->mc_dspace == 0);
179 179
180 180 kmem_free(mc, sizeof (metaslab_class_t));
181 181 }
182 182
183 183 int
184 184 metaslab_class_validate(metaslab_class_t *mc)
185 185 {
186 186 metaslab_group_t *mg;
187 187 vdev_t *vd;
188 188
189 189 /*
190 190 * Must hold one of the spa_config locks.
191 191 */
192 192 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
193 193 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
194 194
195 195 if ((mg = mc->mc_rotor) == NULL)
196 196 return (0);
197 197
198 198 do {
199 199 vd = mg->mg_vd;
200 200 ASSERT(vd->vdev_mg != NULL);
201 201 ASSERT3P(vd->vdev_top, ==, vd);
202 202 ASSERT3P(mg->mg_class, ==, mc);
203 203 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
204 204 } while ((mg = mg->mg_next) != mc->mc_rotor);
205 205
206 206 return (0);
207 207 }
208 208
209 209 void
210 210 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
211 211 int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
212 212 {
213 213 atomic_add_64(&mc->mc_alloc, alloc_delta);
214 214 atomic_add_64(&mc->mc_deferred, defer_delta);
215 215 atomic_add_64(&mc->mc_space, space_delta);
216 216 atomic_add_64(&mc->mc_dspace, dspace_delta);
217 217 }
218 218
219 219 uint64_t
220 220 metaslab_class_get_alloc(metaslab_class_t *mc)
221 221 {
222 222 return (mc->mc_alloc);
223 223 }
224 224
225 225 uint64_t
226 226 metaslab_class_get_deferred(metaslab_class_t *mc)
227 227 {
228 228 return (mc->mc_deferred);
229 229 }
230 230
231 231 uint64_t
232 232 metaslab_class_get_space(metaslab_class_t *mc)
233 233 {
234 234 return (mc->mc_space);
235 235 }
236 236
237 237 uint64_t
238 238 metaslab_class_get_dspace(metaslab_class_t *mc)
239 239 {
240 240 return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
241 241 }
242 242
243 243 /*
244 244 * ==========================================================================
245 245 * Metaslab groups
246 246 * ==========================================================================
247 247 */
248 248 static int
249 249 metaslab_compare(const void *x1, const void *x2)
250 250 {
251 251 const metaslab_t *m1 = x1;
252 252 const metaslab_t *m2 = x2;
253 253
254 254 if (m1->ms_weight < m2->ms_weight)
255 255 return (1);
256 256 if (m1->ms_weight > m2->ms_weight)
257 257 return (-1);
258 258
259 259 /*
260 260 * If the weights are identical, use the offset to force uniqueness.
261 261 */
262 262 if (m1->ms_start < m2->ms_start)
263 263 return (-1);
264 264 if (m1->ms_start > m2->ms_start)
265 265 return (1);
266 266
267 267 ASSERT3P(m1, ==, m2);
268 268
269 269 return (0);
270 270 }
271 271
272 272 /*
273 273 * Update the allocatable flag and the metaslab group's capacity.
274 274 * The allocatable flag is set to true if the capacity is below
275 275 * the zfs_mg_noalloc_threshold. If a metaslab group transitions
276 276 * from allocatable to non-allocatable or vice versa then the metaslab
277 277 * group's class is updated to reflect the transition.
278 278 */
279 279 static void
280 280 metaslab_group_alloc_update(metaslab_group_t *mg)
281 281 {
282 282 vdev_t *vd = mg->mg_vd;
283 283 metaslab_class_t *mc = mg->mg_class;
284 284 vdev_stat_t *vs = &vd->vdev_stat;
285 285 boolean_t was_allocatable;
286 286
287 287 ASSERT(vd == vd->vdev_top);
288 288
289 289 mutex_enter(&mg->mg_lock);
290 290 was_allocatable = mg->mg_allocatable;
291 291
292 292 mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
293 293 (vs->vs_space + 1);
294 294
295 295 mg->mg_allocatable = (mg->mg_free_capacity > zfs_mg_noalloc_threshold);
296 296
297 297 /*
298 298 * The mc_alloc_groups maintains a count of the number of
299 299 * groups in this metaslab class that are still above the
300 300 * zfs_mg_noalloc_threshold. This is used by the allocating
301 301 * threads to determine if they should avoid allocations to
302 302 * a given group. The allocator will avoid allocations to a group
303 303 * if that group has reached or is below the zfs_mg_noalloc_threshold
304 304 * and there are still other groups that are above the threshold.
305 305 * When a group transitions from allocatable to non-allocatable or
306 306 * vice versa we update the metaslab class to reflect that change.
307 307 * When the mc_alloc_groups value drops to 0 that means that all
308 308 * groups have reached the zfs_mg_noalloc_threshold making all groups
309 309 * eligible for allocations. This effectively means that all devices
310 310 * are balanced again.
311 311 */
312 312 if (was_allocatable && !mg->mg_allocatable)
313 313 mc->mc_alloc_groups--;
314 314 else if (!was_allocatable && mg->mg_allocatable)
315 315 mc->mc_alloc_groups++;
316 316 mutex_exit(&mg->mg_lock);
317 317 }
318 318
319 319 metaslab_group_t *
320 320 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
321 321 {
322 322 metaslab_group_t *mg;
323 323
324 324 mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
325 325 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
326 326 avl_create(&mg->mg_metaslab_tree, metaslab_compare,
327 327 sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
328 328 mg->mg_vd = vd;
329 329 mg->mg_class = mc;
330 330 mg->mg_activation_count = 0;
331 331
332 332 mg->mg_taskq = taskq_create("metaslab_group_tasksq", metaslab_load_pct,
333 333 minclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT);
334 334
335 335 return (mg);
336 336 }
337 337
338 338 void
339 339 metaslab_group_destroy(metaslab_group_t *mg)
|
↓ open down ↓ |
339 lines elided |
↑ open up ↑ |
340 340 {
341 341 ASSERT(mg->mg_prev == NULL);
342 342 ASSERT(mg->mg_next == NULL);
343 343 /*
344 344 * We may have gone below zero with the activation count
345 345 * either because we never activated in the first place or
346 346 * because we're done, and possibly removing the vdev.
347 347 */
348 348 ASSERT(mg->mg_activation_count <= 0);
349 349
350 + if (mg->mg_taskq) taskq_destroy(mg->mg_taskq);
351 +
350 352 avl_destroy(&mg->mg_metaslab_tree);
351 353 mutex_destroy(&mg->mg_lock);
352 354 kmem_free(mg, sizeof (metaslab_group_t));
353 355 }
354 356
355 357 void
356 358 metaslab_group_activate(metaslab_group_t *mg)
357 359 {
358 360 metaslab_class_t *mc = mg->mg_class;
359 361 metaslab_group_t *mgprev, *mgnext;
360 362
361 363 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
362 364
363 365 ASSERT(mc->mc_rotor != mg);
364 366 ASSERT(mg->mg_prev == NULL);
365 367 ASSERT(mg->mg_next == NULL);
366 368 ASSERT(mg->mg_activation_count <= 0);
367 369
368 370 if (++mg->mg_activation_count <= 0)
369 371 return;
370 372
371 373 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
372 374 metaslab_group_alloc_update(mg);
373 375
374 376 if ((mgprev = mc->mc_rotor) == NULL) {
375 377 mg->mg_prev = mg;
376 378 mg->mg_next = mg;
377 379 } else {
378 380 mgnext = mgprev->mg_next;
379 381 mg->mg_prev = mgprev;
380 382 mg->mg_next = mgnext;
381 383 mgprev->mg_next = mg;
382 384 mgnext->mg_prev = mg;
383 385 }
384 386 mc->mc_rotor = mg;
385 387 }
386 388
387 389 void
388 390 metaslab_group_passivate(metaslab_group_t *mg)
389 391 {
390 392 metaslab_class_t *mc = mg->mg_class;
391 393 metaslab_group_t *mgprev, *mgnext;
392 394
393 395 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
394 396
395 397 if (--mg->mg_activation_count != 0) {
396 398 ASSERT(mc->mc_rotor != mg);
397 399 ASSERT(mg->mg_prev == NULL);
398 400 ASSERT(mg->mg_next == NULL);
399 401 ASSERT(mg->mg_activation_count < 0);
400 402 return;
401 403 }
402 404
403 405 taskq_wait(mg->mg_taskq);
404 406
405 407 mgprev = mg->mg_prev;
406 408 mgnext = mg->mg_next;
407 409
408 410 if (mg == mgnext) {
409 411 mc->mc_rotor = NULL;
410 412 } else {
411 413 mc->mc_rotor = mgnext;
412 414 mgprev->mg_next = mgnext;
413 415 mgnext->mg_prev = mgprev;
414 416 }
415 417
416 418 mg->mg_prev = NULL;
417 419 mg->mg_next = NULL;
418 420 }
419 421
420 422 static void
421 423 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
422 424 {
423 425 mutex_enter(&mg->mg_lock);
424 426 ASSERT(msp->ms_group == NULL);
425 427 msp->ms_group = mg;
426 428 msp->ms_weight = 0;
427 429 avl_add(&mg->mg_metaslab_tree, msp);
428 430 mutex_exit(&mg->mg_lock);
429 431 }
430 432
431 433 static void
432 434 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
433 435 {
434 436 mutex_enter(&mg->mg_lock);
435 437 ASSERT(msp->ms_group == mg);
436 438 avl_remove(&mg->mg_metaslab_tree, msp);
437 439 msp->ms_group = NULL;
438 440 mutex_exit(&mg->mg_lock);
439 441 }
440 442
441 443 static void
442 444 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
443 445 {
444 446 /*
445 447 * Although in principle the weight can be any value, in
446 448 * practice we do not use values in the range [1, 510].
447 449 */
448 450 ASSERT(weight >= SPA_MINBLOCKSIZE-1 || weight == 0);
449 451 ASSERT(MUTEX_HELD(&msp->ms_lock));
450 452
451 453 mutex_enter(&mg->mg_lock);
452 454 ASSERT(msp->ms_group == mg);
453 455 avl_remove(&mg->mg_metaslab_tree, msp);
454 456 msp->ms_weight = weight;
455 457 avl_add(&mg->mg_metaslab_tree, msp);
456 458 mutex_exit(&mg->mg_lock);
457 459 }
458 460
459 461 /*
460 462 * Determine if a given metaslab group should skip allocations. A metaslab
461 463 * group should avoid allocations if its used capacity has crossed the
462 464 * zfs_mg_noalloc_threshold and there is at least one metaslab group
463 465 * that can still handle allocations.
464 466 */
465 467 static boolean_t
466 468 metaslab_group_allocatable(metaslab_group_t *mg)
467 469 {
468 470 vdev_t *vd = mg->mg_vd;
469 471 spa_t *spa = vd->vdev_spa;
470 472 metaslab_class_t *mc = mg->mg_class;
471 473
472 474 /*
473 475 * A metaslab group is considered allocatable if its free capacity
474 476 * is greater than the set value of zfs_mg_noalloc_threshold, it's
475 477 * associated with a slog, or there are no other metaslab groups
476 478 * with free capacity greater than zfs_mg_noalloc_threshold.
477 479 */
478 480 return (mg->mg_free_capacity > zfs_mg_noalloc_threshold ||
479 481 mc != spa_normal_class(spa) || mc->mc_alloc_groups == 0);
480 482 }
481 483
482 484 /*
483 485 * ==========================================================================
484 486 * Range tree callbacks
485 487 * ==========================================================================
486 488 */
487 489
488 490 /*
489 491 * Comparison function for the private size-ordered tree. Tree is sorted
490 492 * by size, larger sizes at the end of the tree.
491 493 */
492 494 static int
493 495 metaslab_rangesize_compare(const void *x1, const void *x2)
494 496 {
495 497 const range_seg_t *r1 = x1;
496 498 const range_seg_t *r2 = x2;
497 499 uint64_t rs_size1 = r1->rs_end - r1->rs_start;
498 500 uint64_t rs_size2 = r2->rs_end - r2->rs_start;
499 501
500 502 if (rs_size1 < rs_size2)
501 503 return (-1);
502 504 if (rs_size1 > rs_size2)
503 505 return (1);
504 506
505 507 if (r1->rs_start < r2->rs_start)
506 508 return (-1);
507 509
508 510 if (r1->rs_start > r2->rs_start)
509 511 return (1);
510 512
511 513 return (0);
512 514 }
513 515
514 516 /*
515 517 * Create any block allocator specific components. The current allocators
516 518 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
517 519 */
518 520 static void
519 521 metaslab_rt_create(range_tree_t *rt, void *arg)
520 522 {
521 523 metaslab_t *msp = arg;
522 524
523 525 ASSERT3P(rt->rt_arg, ==, msp);
524 526 ASSERT(msp->ms_tree == NULL);
525 527
526 528 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
527 529 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
528 530 }
529 531
530 532 /*
531 533 * Destroy the block allocator specific components.
532 534 */
533 535 static void
534 536 metaslab_rt_destroy(range_tree_t *rt, void *arg)
535 537 {
536 538 metaslab_t *msp = arg;
537 539
538 540 ASSERT3P(rt->rt_arg, ==, msp);
539 541 ASSERT3P(msp->ms_tree, ==, rt);
540 542 ASSERT0(avl_numnodes(&msp->ms_size_tree));
541 543
542 544 avl_destroy(&msp->ms_size_tree);
543 545 }
544 546
545 547 static void
546 548 metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg)
547 549 {
548 550 metaslab_t *msp = arg;
549 551
550 552 ASSERT3P(rt->rt_arg, ==, msp);
551 553 ASSERT3P(msp->ms_tree, ==, rt);
552 554 VERIFY(!msp->ms_condensing);
553 555 avl_add(&msp->ms_size_tree, rs);
554 556 }
555 557
556 558 static void
557 559 metaslab_rt_remove(range_tree_t *rt, range_seg_t *rs, void *arg)
558 560 {
559 561 metaslab_t *msp = arg;
560 562
561 563 ASSERT3P(rt->rt_arg, ==, msp);
562 564 ASSERT3P(msp->ms_tree, ==, rt);
563 565 VERIFY(!msp->ms_condensing);
564 566 avl_remove(&msp->ms_size_tree, rs);
565 567 }
566 568
567 569 static void
568 570 metaslab_rt_vacate(range_tree_t *rt, void *arg)
569 571 {
570 572 metaslab_t *msp = arg;
571 573
572 574 ASSERT3P(rt->rt_arg, ==, msp);
573 575 ASSERT3P(msp->ms_tree, ==, rt);
574 576
575 577 /*
576 578 * Normally one would walk the tree freeing nodes along the way.
577 579 * Since the nodes are shared with the range trees we can avoid
578 580 * walking all nodes and just reinitialize the avl tree. The nodes
579 581 * will be freed by the range tree, so we don't want to free them here.
580 582 */
581 583 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
582 584 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
583 585 }
584 586
585 587 static range_tree_ops_t metaslab_rt_ops = {
586 588 metaslab_rt_create,
587 589 metaslab_rt_destroy,
588 590 metaslab_rt_add,
589 591 metaslab_rt_remove,
590 592 metaslab_rt_vacate
591 593 };
592 594
593 595 /*
594 596 * ==========================================================================
595 597 * Metaslab block operations
596 598 * ==========================================================================
597 599 */
598 600
599 601 /*
600 602 * Return the maximum contiguous segment within the metaslab.
601 603 */
602 604 uint64_t
603 605 metaslab_block_maxsize(metaslab_t *msp)
604 606 {
605 607 avl_tree_t *t = &msp->ms_size_tree;
606 608 range_seg_t *rs;
607 609
608 610 if (t == NULL || (rs = avl_last(t)) == NULL)
609 611 return (0ULL);
610 612
611 613 return (rs->rs_end - rs->rs_start);
612 614 }
613 615
614 616 uint64_t
615 617 metaslab_block_alloc(metaslab_t *msp, uint64_t size)
616 618 {
617 619 uint64_t start;
618 620 range_tree_t *rt = msp->ms_tree;
619 621
620 622 VERIFY(!msp->ms_condensing);
621 623
622 624 start = msp->ms_ops->msop_alloc(msp, size);
623 625 if (start != -1ULL) {
624 626 vdev_t *vd = msp->ms_group->mg_vd;
625 627
626 628 VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
627 629 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
628 630 VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
629 631 range_tree_remove(rt, start, size);
630 632 }
631 633 return (start);
632 634 }
633 635
634 636 /*
635 637 * ==========================================================================
636 638 * Common allocator routines
637 639 * ==========================================================================
638 640 */
639 641
640 642 /*
641 643 * This is a helper function that can be used by the allocator to find
642 644 * a suitable block to allocate. This will search the specified AVL
643 645 * tree looking for a block that matches the specified criteria.
644 646 */
645 647 static uint64_t
646 648 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
647 649 uint64_t align)
648 650 {
649 651 range_seg_t *rs, rsearch;
650 652 avl_index_t where;
651 653
652 654 rsearch.rs_start = *cursor;
653 655 rsearch.rs_end = *cursor + size;
654 656
655 657 rs = avl_find(t, &rsearch, &where);
656 658 if (rs == NULL)
657 659 rs = avl_nearest(t, where, AVL_AFTER);
658 660
659 661 while (rs != NULL) {
660 662 uint64_t offset = P2ROUNDUP(rs->rs_start, align);
661 663
662 664 if (offset + size <= rs->rs_end) {
663 665 *cursor = offset + size;
664 666 return (offset);
665 667 }
666 668 rs = AVL_NEXT(t, rs);
667 669 }
668 670
669 671 /*
670 672 * If we know we've searched the whole map (*cursor == 0), give up.
671 673 * Otherwise, reset the cursor to the beginning and try again.
672 674 */
673 675 if (*cursor == 0)
674 676 return (-1ULL);
675 677
676 678 *cursor = 0;
677 679 return (metaslab_block_picker(t, cursor, size, align));
678 680 }
679 681
680 682 /*
681 683 * ==========================================================================
682 684 * The first-fit block allocator
683 685 * ==========================================================================
684 686 */
685 687 static uint64_t
686 688 metaslab_ff_alloc(metaslab_t *msp, uint64_t size)
687 689 {
688 690 /*
689 691 * Find the largest power of 2 block size that evenly divides the
690 692 * requested size. This is used to try to allocate blocks with similar
691 693 * alignment from the same area of the metaslab (i.e. same cursor
692 694 * bucket) but it does not guarantee that other allocations sizes
693 695 * may exist in the same region.
694 696 */
695 697 uint64_t align = size & -size;
696 698 uint64_t *cursor = &msp->ms_lbas[highbit(align) - 1];
697 699 avl_tree_t *t = &msp->ms_tree->rt_root;
698 700
699 701 return (metaslab_block_picker(t, cursor, size, align));
700 702 }
701 703
702 704 /* ARGSUSED */
703 705 static boolean_t
704 706 metaslab_ff_fragmented(metaslab_t *msp)
705 707 {
706 708 return (B_TRUE);
707 709 }
708 710
709 711 static metaslab_ops_t metaslab_ff_ops = {
710 712 metaslab_ff_alloc,
711 713 metaslab_ff_fragmented
712 714 };
713 715
714 716 /*
715 717 * ==========================================================================
716 718 * Dynamic block allocator -
717 719 * Uses the first fit allocation scheme until space get low and then
718 720 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
719 721 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
720 722 * ==========================================================================
721 723 */
722 724 static uint64_t
723 725 metaslab_df_alloc(metaslab_t *msp, uint64_t size)
724 726 {
725 727 /*
726 728 * Find the largest power of 2 block size that evenly divides the
727 729 * requested size. This is used to try to allocate blocks with similar
728 730 * alignment from the same area of the metaslab (i.e. same cursor
729 731 * bucket) but it does not guarantee that other allocations sizes
730 732 * may exist in the same region.
731 733 */
732 734 uint64_t align = size & -size;
733 735 uint64_t *cursor = &msp->ms_lbas[highbit(align) - 1];
734 736 range_tree_t *rt = msp->ms_tree;
735 737 avl_tree_t *t = &rt->rt_root;
736 738 uint64_t max_size = metaslab_block_maxsize(msp);
737 739 int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
738 740
739 741 ASSERT(MUTEX_HELD(&msp->ms_lock));
740 742 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
741 743
742 744 if (max_size < size)
743 745 return (-1ULL);
744 746
745 747 /*
746 748 * If we're running low on space switch to using the size
747 749 * sorted AVL tree (best-fit).
748 750 */
749 751 if (max_size < metaslab_df_alloc_threshold ||
750 752 free_pct < metaslab_df_free_pct) {
751 753 t = &msp->ms_size_tree;
752 754 *cursor = 0;
753 755 }
754 756
755 757 return (metaslab_block_picker(t, cursor, size, 1ULL));
756 758 }
757 759
758 760 static boolean_t
759 761 metaslab_df_fragmented(metaslab_t *msp)
760 762 {
761 763 range_tree_t *rt = msp->ms_tree;
762 764 uint64_t max_size = metaslab_block_maxsize(msp);
763 765 int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
764 766
765 767 if (max_size >= metaslab_df_alloc_threshold &&
766 768 free_pct >= metaslab_df_free_pct)
767 769 return (B_FALSE);
768 770
769 771 return (B_TRUE);
770 772 }
771 773
772 774 static metaslab_ops_t metaslab_df_ops = {
773 775 metaslab_df_alloc,
774 776 metaslab_df_fragmented
775 777 };
776 778
777 779 /*
778 780 * ==========================================================================
779 781 * Cursor fit block allocator -
780 782 * Select the largest region in the metaslab, set the cursor to the beginning
781 783 * of the range and the cursor_end to the end of the range. As allocations
782 784 * are made advance the cursor. Continue allocating from the cursor until
783 785 * the range is exhausted and then find a new range.
784 786 * ==========================================================================
785 787 */
786 788 static uint64_t
787 789 metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
788 790 {
789 791 range_tree_t *rt = msp->ms_tree;
790 792 avl_tree_t *t = &msp->ms_size_tree;
791 793 uint64_t *cursor = &msp->ms_lbas[0];
792 794 uint64_t *cursor_end = &msp->ms_lbas[1];
793 795 uint64_t offset = 0;
794 796
795 797 ASSERT(MUTEX_HELD(&msp->ms_lock));
796 798 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&rt->rt_root));
797 799
798 800 ASSERT3U(*cursor_end, >=, *cursor);
799 801
800 802 if ((*cursor + size) > *cursor_end) {
801 803 range_seg_t *rs;
802 804
803 805 rs = avl_last(&msp->ms_size_tree);
804 806 if (rs == NULL || (rs->rs_end - rs->rs_start) < size)
805 807 return (-1ULL);
806 808
807 809 *cursor = rs->rs_start;
808 810 *cursor_end = rs->rs_end;
809 811 }
810 812
811 813 offset = *cursor;
812 814 *cursor += size;
813 815
814 816 return (offset);
815 817 }
816 818
817 819 static boolean_t
818 820 metaslab_cf_fragmented(metaslab_t *msp)
819 821 {
820 822 return (metaslab_block_maxsize(msp) < metaslab_min_alloc_size);
821 823 }
822 824
823 825 static metaslab_ops_t metaslab_cf_ops = {
824 826 metaslab_cf_alloc,
825 827 metaslab_cf_fragmented
826 828 };
827 829
828 830 /*
829 831 * ==========================================================================
830 832 * New dynamic fit allocator -
831 833 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
832 834 * contiguous blocks. If no region is found then just use the largest segment
833 835 * that remains.
834 836 * ==========================================================================
835 837 */
836 838
837 839 /*
838 840 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
839 841 * to request from the allocator.
840 842 */
841 843 uint64_t metaslab_ndf_clump_shift = 4;
842 844
843 845 static uint64_t
844 846 metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
845 847 {
846 848 avl_tree_t *t = &msp->ms_tree->rt_root;
847 849 avl_index_t where;
848 850 range_seg_t *rs, rsearch;
849 851 uint64_t hbit = highbit(size);
850 852 uint64_t *cursor = &msp->ms_lbas[hbit - 1];
851 853 uint64_t max_size = metaslab_block_maxsize(msp);
852 854
853 855 ASSERT(MUTEX_HELD(&msp->ms_lock));
854 856 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
855 857
856 858 if (max_size < size)
857 859 return (-1ULL);
858 860
859 861 rsearch.rs_start = *cursor;
860 862 rsearch.rs_end = *cursor + size;
861 863
862 864 rs = avl_find(t, &rsearch, &where);
863 865 if (rs == NULL || (rs->rs_end - rs->rs_start) < size) {
864 866 t = &msp->ms_size_tree;
865 867
866 868 rsearch.rs_start = 0;
867 869 rsearch.rs_end = MIN(max_size,
868 870 1ULL << (hbit + metaslab_ndf_clump_shift));
869 871 rs = avl_find(t, &rsearch, &where);
870 872 if (rs == NULL)
871 873 rs = avl_nearest(t, where, AVL_AFTER);
872 874 ASSERT(rs != NULL);
873 875 }
874 876
875 877 if ((rs->rs_end - rs->rs_start) >= size) {
876 878 *cursor = rs->rs_start + size;
877 879 return (rs->rs_start);
878 880 }
879 881 return (-1ULL);
880 882 }
881 883
882 884 static boolean_t
883 885 metaslab_ndf_fragmented(metaslab_t *msp)
884 886 {
885 887 return (metaslab_block_maxsize(msp) <=
886 888 (metaslab_min_alloc_size << metaslab_ndf_clump_shift));
887 889 }
888 890
889 891 static metaslab_ops_t metaslab_ndf_ops = {
890 892 metaslab_ndf_alloc,
891 893 metaslab_ndf_fragmented
892 894 };
893 895
894 896 metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
895 897
896 898 /*
897 899 * ==========================================================================
898 900 * Metaslabs
899 901 * ==========================================================================
900 902 */
901 903
902 904 /*
903 905 * Wait for any in-progress metaslab loads to complete.
904 906 */
905 907 void
906 908 metaslab_load_wait(metaslab_t *msp)
907 909 {
908 910 ASSERT(MUTEX_HELD(&msp->ms_lock));
909 911
910 912 while (msp->ms_loading) {
911 913 ASSERT(!msp->ms_loaded);
912 914 cv_wait(&msp->ms_load_cv, &msp->ms_lock);
913 915 }
914 916 }
915 917
916 918 int
917 919 metaslab_load(metaslab_t *msp)
918 920 {
919 921 int error = 0;
920 922
921 923 ASSERT(MUTEX_HELD(&msp->ms_lock));
922 924 ASSERT(!msp->ms_loaded);
923 925 ASSERT(!msp->ms_loading);
924 926
925 927 msp->ms_loading = B_TRUE;
926 928
927 929 /*
928 930 * If the space map has not been allocated yet, then treat
929 931 * all the space in the metaslab as free and add it to the
930 932 * ms_tree.
931 933 */
932 934 if (msp->ms_sm != NULL)
933 935 error = space_map_load(msp->ms_sm, msp->ms_tree, SM_FREE);
934 936 else
935 937 range_tree_add(msp->ms_tree, msp->ms_start, msp->ms_size);
936 938
937 939 msp->ms_loaded = (error == 0);
938 940 msp->ms_loading = B_FALSE;
939 941
940 942 if (msp->ms_loaded) {
941 943 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
942 944 range_tree_walk(msp->ms_defertree[t],
943 945 range_tree_remove, msp->ms_tree);
944 946 }
945 947 }
946 948 cv_broadcast(&msp->ms_load_cv);
947 949 return (error);
948 950 }
949 951
950 952 void
951 953 metaslab_unload(metaslab_t *msp)
952 954 {
953 955 ASSERT(MUTEX_HELD(&msp->ms_lock));
954 956 range_tree_vacate(msp->ms_tree, NULL, NULL);
955 957 msp->ms_loaded = B_FALSE;
956 958 msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
957 959 }
958 960
959 961 metaslab_t *
960 962 metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, uint64_t txg)
961 963 {
962 964 vdev_t *vd = mg->mg_vd;
963 965 objset_t *mos = vd->vdev_spa->spa_meta_objset;
964 966 metaslab_t *msp;
965 967
966 968 msp = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
967 969 mutex_init(&msp->ms_lock, NULL, MUTEX_DEFAULT, NULL);
968 970 cv_init(&msp->ms_load_cv, NULL, CV_DEFAULT, NULL);
969 971 msp->ms_id = id;
970 972 msp->ms_start = id << vd->vdev_ms_shift;
971 973 msp->ms_size = 1ULL << vd->vdev_ms_shift;
972 974
973 975 /*
974 976 * We only open space map objects that already exist. All others
975 977 * will be opened when we finally allocate an object for it.
976 978 */
977 979 if (object != 0) {
978 980 VERIFY0(space_map_open(&msp->ms_sm, mos, object, msp->ms_start,
979 981 msp->ms_size, vd->vdev_ashift, &msp->ms_lock));
980 982 ASSERT(msp->ms_sm != NULL);
981 983 }
982 984
983 985 /*
984 986 * We create the main range tree here, but we don't create the
985 987 * alloctree and freetree until metaslab_sync_done(). This serves
986 988 * two purposes: it allows metaslab_sync_done() to detect the
987 989 * addition of new space; and for debugging, it ensures that we'd
988 990 * data fault on any attempt to use this metaslab before it's ready.
989 991 */
990 992 msp->ms_tree = range_tree_create(&metaslab_rt_ops, msp, &msp->ms_lock);
991 993 metaslab_group_add(mg, msp);
992 994
993 995 msp->ms_ops = mg->mg_class->mc_ops;
994 996
995 997 /*
996 998 * If we're opening an existing pool (txg == 0) or creating
997 999 * a new one (txg == TXG_INITIAL), all space is available now.
998 1000 * If we're adding space to an existing pool, the new space
999 1001 * does not become available until after this txg has synced.
1000 1002 */
1001 1003 if (txg <= TXG_INITIAL)
1002 1004 metaslab_sync_done(msp, 0);
1003 1005
1004 1006 /*
1005 1007 * If metaslab_debug_load is set and we're initializing a metaslab
1006 1008 * that has an allocated space_map object then load the its space
1007 1009 * map so that can verify frees.
1008 1010 */
1009 1011 if (metaslab_debug_load && msp->ms_sm != NULL) {
1010 1012 mutex_enter(&msp->ms_lock);
1011 1013 VERIFY0(metaslab_load(msp));
1012 1014 mutex_exit(&msp->ms_lock);
1013 1015 }
1014 1016
1015 1017 if (txg != 0) {
1016 1018 vdev_dirty(vd, 0, NULL, txg);
1017 1019 vdev_dirty(vd, VDD_METASLAB, msp, txg);
1018 1020 }
1019 1021
1020 1022 return (msp);
1021 1023 }
1022 1024
1023 1025 void
1024 1026 metaslab_fini(metaslab_t *msp)
1025 1027 {
1026 1028 metaslab_group_t *mg = msp->ms_group;
1027 1029
1028 1030 metaslab_group_remove(mg, msp);
1029 1031
1030 1032 mutex_enter(&msp->ms_lock);
1031 1033
1032 1034 VERIFY(msp->ms_group == NULL);
1033 1035 vdev_space_update(mg->mg_vd, -space_map_allocated(msp->ms_sm),
1034 1036 0, -msp->ms_size);
1035 1037 space_map_close(msp->ms_sm);
1036 1038
1037 1039 metaslab_unload(msp);
1038 1040 range_tree_destroy(msp->ms_tree);
1039 1041
1040 1042 for (int t = 0; t < TXG_SIZE; t++) {
1041 1043 range_tree_destroy(msp->ms_alloctree[t]);
1042 1044 range_tree_destroy(msp->ms_freetree[t]);
1043 1045 }
1044 1046
1045 1047 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1046 1048 range_tree_destroy(msp->ms_defertree[t]);
1047 1049 }
1048 1050
1049 1051 ASSERT0(msp->ms_deferspace);
1050 1052
1051 1053 mutex_exit(&msp->ms_lock);
1052 1054 cv_destroy(&msp->ms_load_cv);
1053 1055 mutex_destroy(&msp->ms_lock);
1054 1056
1055 1057 kmem_free(msp, sizeof (metaslab_t));
1056 1058 }
1057 1059
1058 1060 /*
1059 1061 * Apply a weighting factor based on the histogram information for this
1060 1062 * metaslab. The current weighting factor is somewhat arbitrary and requires
1061 1063 * additional investigation. The implementation provides a measure of
1062 1064 * "weighted" free space and gives a higher weighting for larger contiguous
1063 1065 * regions. The weighting factor is determined by counting the number of
1064 1066 * sm_shift sectors that exist in each region represented by the histogram.
1065 1067 * That value is then multiplied by the power of 2 exponent and the sm_shift
1066 1068 * value.
1067 1069 *
1068 1070 * For example, assume the 2^21 histogram bucket has 4 2MB regions and the
1069 1071 * metaslab has an sm_shift value of 9 (512B):
1070 1072 *
1071 1073 * 1) calculate the number of sm_shift sectors in the region:
1072 1074 * 2^21 / 2^9 = 2^12 = 4096 * 4 (number of regions) = 16384
1073 1075 * 2) multiply by the power of 2 exponent and the sm_shift value:
1074 1076 * 16384 * 21 * 9 = 3096576
1075 1077 * This value will be added to the weighting of the metaslab.
1076 1078 */
1077 1079 static uint64_t
1078 1080 metaslab_weight_factor(metaslab_t *msp)
1079 1081 {
1080 1082 uint64_t factor = 0;
1081 1083 uint64_t sectors;
1082 1084 int i;
1083 1085
1084 1086 /*
1085 1087 * A null space map means that the entire metaslab is free,
1086 1088 * calculate a weight factor that spans the entire size of the
1087 1089 * metaslab.
1088 1090 */
1089 1091 if (msp->ms_sm == NULL) {
1090 1092 vdev_t *vd = msp->ms_group->mg_vd;
1091 1093
1092 1094 i = highbit(msp->ms_size) - 1;
1093 1095 sectors = msp->ms_size >> vd->vdev_ashift;
1094 1096 return (sectors * i * vd->vdev_ashift);
1095 1097 }
1096 1098
1097 1099 if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
1098 1100 return (0);
1099 1101
1100 1102 for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE(msp->ms_sm); i++) {
1101 1103 if (msp->ms_sm->sm_phys->smp_histogram[i] == 0)
1102 1104 continue;
1103 1105
1104 1106 /*
1105 1107 * Determine the number of sm_shift sectors in the region
1106 1108 * indicated by the histogram. For example, given an
1107 1109 * sm_shift value of 9 (512 bytes) and i = 4 then we know
1108 1110 * that we're looking at an 8K region in the histogram
1109 1111 * (i.e. 9 + 4 = 13, 2^13 = 8192). To figure out the
1110 1112 * number of sm_shift sectors (512 bytes in this example),
1111 1113 * we would take 8192 / 512 = 16. Since the histogram
1112 1114 * is offset by sm_shift we can simply use the value of
1113 1115 * of i to calculate this (i.e. 2^i = 16 where i = 4).
1114 1116 */
1115 1117 sectors = msp->ms_sm->sm_phys->smp_histogram[i] << i;
1116 1118 factor += (i + msp->ms_sm->sm_shift) * sectors;
1117 1119 }
1118 1120 return (factor * msp->ms_sm->sm_shift);
1119 1121 }
1120 1122
1121 1123 static uint64_t
1122 1124 metaslab_weight(metaslab_t *msp)
1123 1125 {
1124 1126 metaslab_group_t *mg = msp->ms_group;
1125 1127 vdev_t *vd = mg->mg_vd;
1126 1128 uint64_t weight, space;
1127 1129
1128 1130 ASSERT(MUTEX_HELD(&msp->ms_lock));
1129 1131
1130 1132 /*
1131 1133 * This vdev is in the process of being removed so there is nothing
1132 1134 * for us to do here.
1133 1135 */
1134 1136 if (vd->vdev_removing) {
1135 1137 ASSERT0(space_map_allocated(msp->ms_sm));
1136 1138 ASSERT0(vd->vdev_ms_shift);
1137 1139 return (0);
1138 1140 }
1139 1141
1140 1142 /*
1141 1143 * The baseline weight is the metaslab's free space.
1142 1144 */
1143 1145 space = msp->ms_size - space_map_allocated(msp->ms_sm);
1144 1146 weight = space;
1145 1147
1146 1148 /*
1147 1149 * Modern disks have uniform bit density and constant angular velocity.
1148 1150 * Therefore, the outer recording zones are faster (higher bandwidth)
1149 1151 * than the inner zones by the ratio of outer to inner track diameter,
1150 1152 * which is typically around 2:1. We account for this by assigning
1151 1153 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1152 1154 * In effect, this means that we'll select the metaslab with the most
1153 1155 * free bandwidth rather than simply the one with the most free space.
1154 1156 */
1155 1157 weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count;
1156 1158 ASSERT(weight >= space && weight <= 2 * space);
1157 1159
1158 1160 msp->ms_factor = metaslab_weight_factor(msp);
1159 1161 if (metaslab_weight_factor_enable)
1160 1162 weight += msp->ms_factor;
1161 1163
1162 1164 if (msp->ms_loaded && !msp->ms_ops->msop_fragmented(msp)) {
1163 1165 /*
1164 1166 * If this metaslab is one we're actively using, adjust its
1165 1167 * weight to make it preferable to any inactive metaslab so
1166 1168 * we'll polish it off.
1167 1169 */
1168 1170 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
1169 1171 }
1170 1172
1171 1173 return (weight);
1172 1174 }
1173 1175
1174 1176 static int
1175 1177 metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
1176 1178 {
1177 1179 ASSERT(MUTEX_HELD(&msp->ms_lock));
1178 1180
1179 1181 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1180 1182 metaslab_load_wait(msp);
1181 1183 if (!msp->ms_loaded) {
1182 1184 int error = metaslab_load(msp);
1183 1185 if (error) {
1184 1186 metaslab_group_sort(msp->ms_group, msp, 0);
1185 1187 return (error);
1186 1188 }
1187 1189 }
1188 1190
1189 1191 metaslab_group_sort(msp->ms_group, msp,
1190 1192 msp->ms_weight | activation_weight);
1191 1193 }
1192 1194 ASSERT(msp->ms_loaded);
1193 1195 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
1194 1196
1195 1197 return (0);
1196 1198 }
1197 1199
1198 1200 static void
1199 1201 metaslab_passivate(metaslab_t *msp, uint64_t size)
1200 1202 {
1201 1203 /*
1202 1204 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1203 1205 * this metaslab again. In that case, it had better be empty,
1204 1206 * or we would be leaving space on the table.
1205 1207 */
1206 1208 ASSERT(size >= SPA_MINBLOCKSIZE || range_tree_space(msp->ms_tree) == 0);
1207 1209 metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
1208 1210 ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
1209 1211 }
1210 1212
1211 1213 static void
1212 1214 metaslab_preload(void *arg)
1213 1215 {
1214 1216 metaslab_t *msp = arg;
1215 1217 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1216 1218
1217 1219 mutex_enter(&msp->ms_lock);
1218 1220 metaslab_load_wait(msp);
1219 1221 if (!msp->ms_loaded)
1220 1222 (void) metaslab_load(msp);
1221 1223
1222 1224 /*
1223 1225 * Set the ms_access_txg value so that we don't unload it right away.
1224 1226 */
1225 1227 msp->ms_access_txg = spa_syncing_txg(spa) + metaslab_unload_delay + 1;
1226 1228 mutex_exit(&msp->ms_lock);
1227 1229 }
1228 1230
1229 1231 static void
1230 1232 metaslab_group_preload(metaslab_group_t *mg)
1231 1233 {
1232 1234 spa_t *spa = mg->mg_vd->vdev_spa;
1233 1235 metaslab_t *msp;
1234 1236 avl_tree_t *t = &mg->mg_metaslab_tree;
1235 1237 int m = 0;
1236 1238
1237 1239 if (spa_shutting_down(spa) || !metaslab_preload_enabled) {
1238 1240 taskq_wait(mg->mg_taskq);
1239 1241 return;
1240 1242 }
1241 1243 mutex_enter(&mg->mg_lock);
1242 1244
1243 1245 /*
1244 1246 * Prefetch the next potential metaslabs
1245 1247 */
1246 1248 for (msp = avl_first(t); msp != NULL; msp = AVL_NEXT(t, msp)) {
1247 1249
1248 1250 /* If we have reached our preload limit then we're done */
1249 1251 if (++m > metaslab_preload_limit)
1250 1252 break;
1251 1253
1252 1254 VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
1253 1255 msp, TQ_SLEEP) != NULL);
1254 1256 }
1255 1257 mutex_exit(&mg->mg_lock);
1256 1258 }
1257 1259
1258 1260 /*
1259 1261 * Determine if the space map's on-disk footprint is past our tolerance
1260 1262 * for inefficiency. We would like to use the following criteria to make
1261 1263 * our decision:
1262 1264 *
1263 1265 * 1. The size of the space map object should not dramatically increase as a
1264 1266 * result of writing out the free space range tree.
1265 1267 *
1266 1268 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
1267 1269 * times the size than the free space range tree representation
1268 1270 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
1269 1271 *
1270 1272 * Checking the first condition is tricky since we don't want to walk
1271 1273 * the entire AVL tree calculating the estimated on-disk size. Instead we
1272 1274 * use the size-ordered range tree in the metaslab and calculate the
1273 1275 * size required to write out the largest segment in our free tree. If the
1274 1276 * size required to represent that segment on disk is larger than the space
1275 1277 * map object then we avoid condensing this map.
1276 1278 *
1277 1279 * To determine the second criterion we use a best-case estimate and assume
1278 1280 * each segment can be represented on-disk as a single 64-bit entry. We refer
1279 1281 * to this best-case estimate as the space map's minimal form.
1280 1282 */
1281 1283 static boolean_t
1282 1284 metaslab_should_condense(metaslab_t *msp)
1283 1285 {
1284 1286 space_map_t *sm = msp->ms_sm;
1285 1287 range_seg_t *rs;
1286 1288 uint64_t size, entries, segsz;
1287 1289
1288 1290 ASSERT(MUTEX_HELD(&msp->ms_lock));
1289 1291 ASSERT(msp->ms_loaded);
1290 1292
1291 1293 /*
1292 1294 * Use the ms_size_tree range tree, which is ordered by size, to
1293 1295 * obtain the largest segment in the free tree. If the tree is empty
1294 1296 * then we should condense the map.
1295 1297 */
1296 1298 rs = avl_last(&msp->ms_size_tree);
1297 1299 if (rs == NULL)
1298 1300 return (B_TRUE);
1299 1301
1300 1302 /*
1301 1303 * Calculate the number of 64-bit entries this segment would
1302 1304 * require when written to disk. If this single segment would be
1303 1305 * larger on-disk than the entire current on-disk structure, then
1304 1306 * clearly condensing will increase the on-disk structure size.
1305 1307 */
1306 1308 size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
1307 1309 entries = size / (MIN(size, SM_RUN_MAX));
1308 1310 segsz = entries * sizeof (uint64_t);
1309 1311
1310 1312 return (segsz <= space_map_length(msp->ms_sm) &&
1311 1313 space_map_length(msp->ms_sm) >= (zfs_condense_pct *
1312 1314 sizeof (uint64_t) * avl_numnodes(&msp->ms_tree->rt_root)) / 100);
1313 1315 }
1314 1316
1315 1317 /*
1316 1318 * Condense the on-disk space map representation to its minimized form.
1317 1319 * The minimized form consists of a small number of allocations followed by
1318 1320 * the entries of the free range tree.
1319 1321 */
1320 1322 static void
1321 1323 metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
1322 1324 {
1323 1325 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1324 1326 range_tree_t *freetree = msp->ms_freetree[txg & TXG_MASK];
1325 1327 range_tree_t *condense_tree;
1326 1328 space_map_t *sm = msp->ms_sm;
1327 1329
1328 1330 ASSERT(MUTEX_HELD(&msp->ms_lock));
1329 1331 ASSERT3U(spa_sync_pass(spa), ==, 1);
1330 1332 ASSERT(msp->ms_loaded);
1331 1333
1332 1334 spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, "
1333 1335 "smp size %llu, segments %lu", txg, msp->ms_id, msp,
1334 1336 space_map_length(msp->ms_sm), avl_numnodes(&msp->ms_tree->rt_root));
1335 1337
1336 1338 /*
1337 1339 * Create an range tree that is 100% allocated. We remove segments
1338 1340 * that have been freed in this txg, any deferred frees that exist,
1339 1341 * and any allocation in the future. Removing segments should be
1340 1342 * a relatively inexpensive operation since we expect these trees to
1341 1343 * have a small number of nodes.
1342 1344 */
1343 1345 condense_tree = range_tree_create(NULL, NULL, &msp->ms_lock);
1344 1346 range_tree_add(condense_tree, msp->ms_start, msp->ms_size);
1345 1347
1346 1348 /*
1347 1349 * Remove what's been freed in this txg from the condense_tree.
1348 1350 * Since we're in sync_pass 1, we know that all the frees from
1349 1351 * this txg are in the freetree.
1350 1352 */
1351 1353 range_tree_walk(freetree, range_tree_remove, condense_tree);
1352 1354
1353 1355 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1354 1356 range_tree_walk(msp->ms_defertree[t],
1355 1357 range_tree_remove, condense_tree);
1356 1358 }
1357 1359
1358 1360 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
1359 1361 range_tree_walk(msp->ms_alloctree[(txg + t) & TXG_MASK],
1360 1362 range_tree_remove, condense_tree);
1361 1363 }
1362 1364
1363 1365 /*
1364 1366 * We're about to drop the metaslab's lock thus allowing
1365 1367 * other consumers to change it's content. Set the
1366 1368 * metaslab's ms_condensing flag to ensure that
1367 1369 * allocations on this metaslab do not occur while we're
1368 1370 * in the middle of committing it to disk. This is only critical
1369 1371 * for the ms_tree as all other range trees use per txg
1370 1372 * views of their content.
1371 1373 */
1372 1374 msp->ms_condensing = B_TRUE;
1373 1375
1374 1376 mutex_exit(&msp->ms_lock);
1375 1377 space_map_truncate(sm, tx);
1376 1378 mutex_enter(&msp->ms_lock);
1377 1379
1378 1380 /*
1379 1381 * While we would ideally like to create a space_map representation
1380 1382 * that consists only of allocation records, doing so can be
1381 1383 * prohibitively expensive because the in-core free tree can be
1382 1384 * large, and therefore computationally expensive to subtract
1383 1385 * from the condense_tree. Instead we sync out two trees, a cheap
1384 1386 * allocation only tree followed by the in-core free tree. While not
1385 1387 * optimal, this is typically close to optimal, and much cheaper to
1386 1388 * compute.
1387 1389 */
1388 1390 space_map_write(sm, condense_tree, SM_ALLOC, tx);
1389 1391 range_tree_vacate(condense_tree, NULL, NULL);
1390 1392 range_tree_destroy(condense_tree);
1391 1393
1392 1394 space_map_write(sm, msp->ms_tree, SM_FREE, tx);
1393 1395 msp->ms_condensing = B_FALSE;
1394 1396 }
1395 1397
1396 1398 /*
1397 1399 * Write a metaslab to disk in the context of the specified transaction group.
1398 1400 */
1399 1401 void
1400 1402 metaslab_sync(metaslab_t *msp, uint64_t txg)
1401 1403 {
1402 1404 metaslab_group_t *mg = msp->ms_group;
1403 1405 vdev_t *vd = mg->mg_vd;
1404 1406 spa_t *spa = vd->vdev_spa;
1405 1407 objset_t *mos = spa_meta_objset(spa);
1406 1408 range_tree_t *alloctree = msp->ms_alloctree[txg & TXG_MASK];
1407 1409 range_tree_t **freetree = &msp->ms_freetree[txg & TXG_MASK];
1408 1410 range_tree_t **freed_tree =
1409 1411 &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
1410 1412 dmu_tx_t *tx;
1411 1413 uint64_t object = space_map_object(msp->ms_sm);
1412 1414
1413 1415 ASSERT(!vd->vdev_ishole);
1414 1416
1415 1417 /*
1416 1418 * This metaslab has just been added so there's no work to do now.
1417 1419 */
1418 1420 if (*freetree == NULL) {
1419 1421 ASSERT3P(alloctree, ==, NULL);
1420 1422 return;
1421 1423 }
1422 1424
1423 1425 ASSERT3P(alloctree, !=, NULL);
1424 1426 ASSERT3P(*freetree, !=, NULL);
1425 1427 ASSERT3P(*freed_tree, !=, NULL);
1426 1428
1427 1429 if (range_tree_space(alloctree) == 0 &&
1428 1430 range_tree_space(*freetree) == 0)
1429 1431 return;
1430 1432
1431 1433 /*
1432 1434 * The only state that can actually be changing concurrently with
1433 1435 * metaslab_sync() is the metaslab's ms_tree. No other thread can
1434 1436 * be modifying this txg's alloctree, freetree, freed_tree, or
1435 1437 * space_map_phys_t. Therefore, we only hold ms_lock to satify
1436 1438 * space_map ASSERTs. We drop it whenever we call into the DMU,
1437 1439 * because the DMU can call down to us (e.g. via zio_free()) at
1438 1440 * any time.
1439 1441 */
1440 1442
1441 1443 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
1442 1444
1443 1445 if (msp->ms_sm == NULL) {
1444 1446 uint64_t new_object;
1445 1447
1446 1448 new_object = space_map_alloc(mos, tx);
1447 1449 VERIFY3U(new_object, !=, 0);
1448 1450
1449 1451 VERIFY0(space_map_open(&msp->ms_sm, mos, new_object,
1450 1452 msp->ms_start, msp->ms_size, vd->vdev_ashift,
1451 1453 &msp->ms_lock));
1452 1454 ASSERT(msp->ms_sm != NULL);
1453 1455 }
1454 1456
1455 1457 mutex_enter(&msp->ms_lock);
1456 1458
1457 1459 if (msp->ms_loaded && spa_sync_pass(spa) == 1 &&
1458 1460 metaslab_should_condense(msp)) {
1459 1461 metaslab_condense(msp, txg, tx);
1460 1462 } else {
1461 1463 space_map_write(msp->ms_sm, alloctree, SM_ALLOC, tx);
1462 1464 space_map_write(msp->ms_sm, *freetree, SM_FREE, tx);
1463 1465 }
1464 1466
1465 1467 range_tree_vacate(alloctree, NULL, NULL);
1466 1468
1467 1469 if (msp->ms_loaded) {
1468 1470 /*
1469 1471 * When the space map is loaded, we have an accruate
1470 1472 * histogram in the range tree. This gives us an opportunity
1471 1473 * to bring the space map's histogram up-to-date so we clear
1472 1474 * it first before updating it.
1473 1475 */
1474 1476 space_map_histogram_clear(msp->ms_sm);
1475 1477 space_map_histogram_add(msp->ms_sm, msp->ms_tree, tx);
1476 1478 } else {
1477 1479 /*
1478 1480 * Since the space map is not loaded we simply update the
1479 1481 * exisiting histogram with what was freed in this txg. This
1480 1482 * means that the on-disk histogram may not have an accurate
1481 1483 * view of the free space but it's close enough to allow
1482 1484 * us to make allocation decisions.
1483 1485 */
1484 1486 space_map_histogram_add(msp->ms_sm, *freetree, tx);
1485 1487 }
1486 1488
1487 1489 /*
1488 1490 * For sync pass 1, we avoid traversing this txg's free range tree
1489 1491 * and instead will just swap the pointers for freetree and
1490 1492 * freed_tree. We can safely do this since the freed_tree is
1491 1493 * guaranteed to be empty on the initial pass.
1492 1494 */
1493 1495 if (spa_sync_pass(spa) == 1) {
1494 1496 range_tree_swap(freetree, freed_tree);
1495 1497 } else {
1496 1498 range_tree_vacate(*freetree, range_tree_add, *freed_tree);
1497 1499 }
1498 1500
1499 1501 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
1500 1502 ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
1501 1503
1502 1504 mutex_exit(&msp->ms_lock);
1503 1505
1504 1506 if (object != space_map_object(msp->ms_sm)) {
1505 1507 object = space_map_object(msp->ms_sm);
1506 1508 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
1507 1509 msp->ms_id, sizeof (uint64_t), &object, tx);
1508 1510 }
1509 1511 dmu_tx_commit(tx);
1510 1512 }
1511 1513
1512 1514 /*
1513 1515 * Called after a transaction group has completely synced to mark
1514 1516 * all of the metaslab's free space as usable.
1515 1517 */
1516 1518 void
1517 1519 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
1518 1520 {
1519 1521 metaslab_group_t *mg = msp->ms_group;
1520 1522 vdev_t *vd = mg->mg_vd;
1521 1523 range_tree_t **freed_tree;
1522 1524 range_tree_t **defer_tree;
1523 1525 int64_t alloc_delta, defer_delta;
1524 1526
1525 1527 ASSERT(!vd->vdev_ishole);
1526 1528
1527 1529 mutex_enter(&msp->ms_lock);
1528 1530
1529 1531 /*
1530 1532 * If this metaslab is just becoming available, initialize its
1531 1533 * alloctrees, freetrees, and defertree and add its capacity to
1532 1534 * the vdev.
1533 1535 */
1534 1536 if (msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK] == NULL) {
1535 1537 for (int t = 0; t < TXG_SIZE; t++) {
1536 1538 ASSERT(msp->ms_alloctree[t] == NULL);
1537 1539 ASSERT(msp->ms_freetree[t] == NULL);
1538 1540
1539 1541 msp->ms_alloctree[t] = range_tree_create(NULL, msp,
1540 1542 &msp->ms_lock);
1541 1543 msp->ms_freetree[t] = range_tree_create(NULL, msp,
1542 1544 &msp->ms_lock);
1543 1545 }
1544 1546
1545 1547 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1546 1548 ASSERT(msp->ms_defertree[t] == NULL);
1547 1549
1548 1550 msp->ms_defertree[t] = range_tree_create(NULL, msp,
1549 1551 &msp->ms_lock);
1550 1552 }
1551 1553
1552 1554 vdev_space_update(vd, 0, 0, msp->ms_size);
1553 1555 }
1554 1556
1555 1557 freed_tree = &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
1556 1558 defer_tree = &msp->ms_defertree[txg % TXG_DEFER_SIZE];
1557 1559
1558 1560 alloc_delta = space_map_alloc_delta(msp->ms_sm);
1559 1561 defer_delta = range_tree_space(*freed_tree) -
1560 1562 range_tree_space(*defer_tree);
1561 1563
1562 1564 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
1563 1565
1564 1566 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
1565 1567 ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
1566 1568
1567 1569 /*
1568 1570 * If there's a metaslab_load() in progress, wait for it to complete
1569 1571 * so that we have a consistent view of the in-core space map.
1570 1572 */
1571 1573 metaslab_load_wait(msp);
1572 1574
1573 1575 /*
1574 1576 * Move the frees from the defer_tree back to the free
1575 1577 * range tree (if it's loaded). Swap the freed_tree and the
1576 1578 * defer_tree -- this is safe to do because we've just emptied out
1577 1579 * the defer_tree.
1578 1580 */
1579 1581 range_tree_vacate(*defer_tree,
1580 1582 msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
1581 1583 range_tree_swap(freed_tree, defer_tree);
1582 1584
1583 1585 space_map_update(msp->ms_sm);
1584 1586
1585 1587 msp->ms_deferspace += defer_delta;
1586 1588 ASSERT3S(msp->ms_deferspace, >=, 0);
1587 1589 ASSERT3S(msp->ms_deferspace, <=, msp->ms_size);
1588 1590 if (msp->ms_deferspace != 0) {
1589 1591 /*
1590 1592 * Keep syncing this metaslab until all deferred frees
1591 1593 * are back in circulation.
1592 1594 */
1593 1595 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1594 1596 }
1595 1597
1596 1598 if (msp->ms_loaded && msp->ms_access_txg < txg) {
1597 1599 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
1598 1600 VERIFY0(range_tree_space(
1599 1601 msp->ms_alloctree[(txg + t) & TXG_MASK]));
1600 1602 }
1601 1603
1602 1604 if (!metaslab_debug_unload)
1603 1605 metaslab_unload(msp);
1604 1606 }
1605 1607
1606 1608 metaslab_group_sort(mg, msp, metaslab_weight(msp));
1607 1609 mutex_exit(&msp->ms_lock);
1608 1610
1609 1611 }
1610 1612
1611 1613 void
1612 1614 metaslab_sync_reassess(metaslab_group_t *mg)
1613 1615 {
1614 1616 int64_t failures = mg->mg_alloc_failures;
1615 1617
1616 1618 metaslab_group_alloc_update(mg);
1617 1619 atomic_add_64(&mg->mg_alloc_failures, -failures);
1618 1620
1619 1621 /*
1620 1622 * Preload the next potential metaslabs
1621 1623 */
1622 1624 metaslab_group_preload(mg);
1623 1625 }
1624 1626
1625 1627 static uint64_t
1626 1628 metaslab_distance(metaslab_t *msp, dva_t *dva)
1627 1629 {
1628 1630 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
1629 1631 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
1630 1632 uint64_t start = msp->ms_id;
1631 1633
1632 1634 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
1633 1635 return (1ULL << 63);
1634 1636
1635 1637 if (offset < start)
1636 1638 return ((start - offset) << ms_shift);
1637 1639 if (offset > start)
1638 1640 return ((offset - start) << ms_shift);
1639 1641 return (0);
1640 1642 }
1641 1643
1642 1644 static uint64_t
1643 1645 metaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize,
1644 1646 uint64_t txg, uint64_t min_distance, dva_t *dva, int d, int flags)
1645 1647 {
1646 1648 spa_t *spa = mg->mg_vd->vdev_spa;
1647 1649 metaslab_t *msp = NULL;
1648 1650 uint64_t offset = -1ULL;
1649 1651 avl_tree_t *t = &mg->mg_metaslab_tree;
1650 1652 uint64_t activation_weight;
1651 1653 uint64_t target_distance;
1652 1654 int i;
1653 1655
1654 1656 activation_weight = METASLAB_WEIGHT_PRIMARY;
1655 1657 for (i = 0; i < d; i++) {
1656 1658 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
1657 1659 activation_weight = METASLAB_WEIGHT_SECONDARY;
1658 1660 break;
1659 1661 }
1660 1662 }
1661 1663
1662 1664 for (;;) {
1663 1665 boolean_t was_active;
1664 1666
1665 1667 mutex_enter(&mg->mg_lock);
1666 1668 for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
1667 1669 if (msp->ms_weight < asize) {
1668 1670 spa_dbgmsg(spa, "%s: failed to meet weight "
1669 1671 "requirement: vdev %llu, txg %llu, mg %p, "
1670 1672 "msp %p, psize %llu, asize %llu, "
1671 1673 "failures %llu, weight %llu",
1672 1674 spa_name(spa), mg->mg_vd->vdev_id, txg,
1673 1675 mg, msp, psize, asize,
1674 1676 mg->mg_alloc_failures, msp->ms_weight);
1675 1677 mutex_exit(&mg->mg_lock);
1676 1678 return (-1ULL);
1677 1679 }
1678 1680
1679 1681 /*
1680 1682 * If the selected metaslab is condensing, skip it.
1681 1683 */
1682 1684 if (msp->ms_condensing)
1683 1685 continue;
1684 1686
1685 1687 was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
1686 1688 if (activation_weight == METASLAB_WEIGHT_PRIMARY)
1687 1689 break;
1688 1690
1689 1691 target_distance = min_distance +
1690 1692 (space_map_allocated(msp->ms_sm) != 0 ? 0 :
1691 1693 min_distance >> 1);
1692 1694
1693 1695 for (i = 0; i < d; i++)
1694 1696 if (metaslab_distance(msp, &dva[i]) <
1695 1697 target_distance)
1696 1698 break;
1697 1699 if (i == d)
1698 1700 break;
1699 1701 }
1700 1702 mutex_exit(&mg->mg_lock);
1701 1703 if (msp == NULL)
1702 1704 return (-1ULL);
1703 1705
1704 1706 mutex_enter(&msp->ms_lock);
1705 1707
1706 1708 /*
1707 1709 * If we've already reached the allowable number of failed
1708 1710 * allocation attempts on this metaslab group then we
1709 1711 * consider skipping it. We skip it only if we're allowed
1710 1712 * to "fast" gang, the physical size is larger than
1711 1713 * a gang block, and we're attempting to allocate from
1712 1714 * the primary metaslab.
1713 1715 */
1714 1716 if (mg->mg_alloc_failures > zfs_mg_alloc_failures &&
1715 1717 CAN_FASTGANG(flags) && psize > SPA_GANGBLOCKSIZE &&
1716 1718 activation_weight == METASLAB_WEIGHT_PRIMARY) {
1717 1719 spa_dbgmsg(spa, "%s: skipping metaslab group: "
1718 1720 "vdev %llu, txg %llu, mg %p, msp[%llu] %p, "
1719 1721 "psize %llu, asize %llu, failures %llu",
1720 1722 spa_name(spa), mg->mg_vd->vdev_id, txg, mg,
1721 1723 msp->ms_id, msp, psize, asize,
1722 1724 mg->mg_alloc_failures);
1723 1725 mutex_exit(&msp->ms_lock);
1724 1726 return (-1ULL);
1725 1727 }
1726 1728
1727 1729 /*
1728 1730 * Ensure that the metaslab we have selected is still
1729 1731 * capable of handling our request. It's possible that
1730 1732 * another thread may have changed the weight while we
1731 1733 * were blocked on the metaslab lock.
1732 1734 */
1733 1735 if (msp->ms_weight < asize || (was_active &&
1734 1736 !(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
1735 1737 activation_weight == METASLAB_WEIGHT_PRIMARY)) {
1736 1738 mutex_exit(&msp->ms_lock);
1737 1739 continue;
1738 1740 }
1739 1741
1740 1742 if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
1741 1743 activation_weight == METASLAB_WEIGHT_PRIMARY) {
1742 1744 metaslab_passivate(msp,
1743 1745 msp->ms_weight & ~METASLAB_ACTIVE_MASK);
1744 1746 mutex_exit(&msp->ms_lock);
1745 1747 continue;
1746 1748 }
1747 1749
1748 1750 if (metaslab_activate(msp, activation_weight) != 0) {
1749 1751 mutex_exit(&msp->ms_lock);
1750 1752 continue;
1751 1753 }
1752 1754
1753 1755 /*
1754 1756 * If this metaslab is currently condensing then pick again as
1755 1757 * we can't manipulate this metaslab until it's committed
1756 1758 * to disk.
1757 1759 */
1758 1760 if (msp->ms_condensing) {
1759 1761 mutex_exit(&msp->ms_lock);
1760 1762 continue;
1761 1763 }
1762 1764
1763 1765 if ((offset = metaslab_block_alloc(msp, asize)) != -1ULL)
1764 1766 break;
1765 1767
1766 1768 atomic_inc_64(&mg->mg_alloc_failures);
1767 1769
1768 1770 metaslab_passivate(msp, metaslab_block_maxsize(msp));
1769 1771 mutex_exit(&msp->ms_lock);
1770 1772 }
1771 1773
1772 1774 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
1773 1775 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
1774 1776
1775 1777 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, asize);
1776 1778 msp->ms_access_txg = txg + metaslab_unload_delay;
1777 1779
1778 1780 mutex_exit(&msp->ms_lock);
1779 1781
1780 1782 return (offset);
1781 1783 }
1782 1784
1783 1785 /*
1784 1786 * Allocate a block for the specified i/o.
1785 1787 */
1786 1788 static int
1787 1789 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
1788 1790 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
1789 1791 {
1790 1792 metaslab_group_t *mg, *rotor;
1791 1793 vdev_t *vd;
1792 1794 int dshift = 3;
1793 1795 int all_zero;
1794 1796 int zio_lock = B_FALSE;
1795 1797 boolean_t allocatable;
1796 1798 uint64_t offset = -1ULL;
1797 1799 uint64_t asize;
1798 1800 uint64_t distance;
1799 1801
1800 1802 ASSERT(!DVA_IS_VALID(&dva[d]));
1801 1803
1802 1804 /*
1803 1805 * For testing, make some blocks above a certain size be gang blocks.
1804 1806 */
1805 1807 if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0)
1806 1808 return (SET_ERROR(ENOSPC));
1807 1809
1808 1810 /*
1809 1811 * Start at the rotor and loop through all mgs until we find something.
1810 1812 * Note that there's no locking on mc_rotor or mc_aliquot because
1811 1813 * nothing actually breaks if we miss a few updates -- we just won't
1812 1814 * allocate quite as evenly. It all balances out over time.
1813 1815 *
1814 1816 * If we are doing ditto or log blocks, try to spread them across
1815 1817 * consecutive vdevs. If we're forced to reuse a vdev before we've
1816 1818 * allocated all of our ditto blocks, then try and spread them out on
1817 1819 * that vdev as much as possible. If it turns out to not be possible,
1818 1820 * gradually lower our standards until anything becomes acceptable.
1819 1821 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
1820 1822 * gives us hope of containing our fault domains to something we're
1821 1823 * able to reason about. Otherwise, any two top-level vdev failures
1822 1824 * will guarantee the loss of data. With consecutive allocation,
1823 1825 * only two adjacent top-level vdev failures will result in data loss.
1824 1826 *
1825 1827 * If we are doing gang blocks (hintdva is non-NULL), try to keep
1826 1828 * ourselves on the same vdev as our gang block header. That
1827 1829 * way, we can hope for locality in vdev_cache, plus it makes our
1828 1830 * fault domains something tractable.
1829 1831 */
1830 1832 if (hintdva) {
1831 1833 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
1832 1834
1833 1835 /*
1834 1836 * It's possible the vdev we're using as the hint no
1835 1837 * longer exists (i.e. removed). Consult the rotor when
1836 1838 * all else fails.
1837 1839 */
1838 1840 if (vd != NULL) {
1839 1841 mg = vd->vdev_mg;
1840 1842
1841 1843 if (flags & METASLAB_HINTBP_AVOID &&
1842 1844 mg->mg_next != NULL)
1843 1845 mg = mg->mg_next;
1844 1846 } else {
1845 1847 mg = mc->mc_rotor;
1846 1848 }
1847 1849 } else if (d != 0) {
1848 1850 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
1849 1851 mg = vd->vdev_mg->mg_next;
1850 1852 } else {
1851 1853 mg = mc->mc_rotor;
1852 1854 }
1853 1855
1854 1856 /*
1855 1857 * If the hint put us into the wrong metaslab class, or into a
1856 1858 * metaslab group that has been passivated, just follow the rotor.
1857 1859 */
1858 1860 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
1859 1861 mg = mc->mc_rotor;
1860 1862
1861 1863 rotor = mg;
1862 1864 top:
1863 1865 all_zero = B_TRUE;
1864 1866 do {
1865 1867 ASSERT(mg->mg_activation_count == 1);
1866 1868
1867 1869 vd = mg->mg_vd;
1868 1870
1869 1871 /*
1870 1872 * Don't allocate from faulted devices.
1871 1873 */
1872 1874 if (zio_lock) {
1873 1875 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
1874 1876 allocatable = vdev_allocatable(vd);
1875 1877 spa_config_exit(spa, SCL_ZIO, FTAG);
1876 1878 } else {
1877 1879 allocatable = vdev_allocatable(vd);
1878 1880 }
1879 1881
1880 1882 /*
1881 1883 * Determine if the selected metaslab group is eligible
1882 1884 * for allocations. If we're ganging or have requested
1883 1885 * an allocation for the smallest gang block size
1884 1886 * then we don't want to avoid allocating to the this
1885 1887 * metaslab group. If we're in this condition we should
1886 1888 * try to allocate from any device possible so that we
1887 1889 * don't inadvertently return ENOSPC and suspend the pool
1888 1890 * even though space is still available.
1889 1891 */
1890 1892 if (allocatable && CAN_FASTGANG(flags) &&
1891 1893 psize > SPA_GANGBLOCKSIZE)
1892 1894 allocatable = metaslab_group_allocatable(mg);
1893 1895
1894 1896 if (!allocatable)
1895 1897 goto next;
1896 1898
1897 1899 /*
1898 1900 * Avoid writing single-copy data to a failing vdev
1899 1901 * unless the user instructs us that it is okay.
1900 1902 */
1901 1903 if ((vd->vdev_stat.vs_write_errors > 0 ||
1902 1904 vd->vdev_state < VDEV_STATE_HEALTHY) &&
1903 1905 d == 0 && dshift == 3 &&
1904 1906 !(zfs_write_to_degraded && vd->vdev_state ==
1905 1907 VDEV_STATE_DEGRADED)) {
1906 1908 all_zero = B_FALSE;
1907 1909 goto next;
1908 1910 }
1909 1911
1910 1912 ASSERT(mg->mg_class == mc);
1911 1913
1912 1914 distance = vd->vdev_asize >> dshift;
1913 1915 if (distance <= (1ULL << vd->vdev_ms_shift))
1914 1916 distance = 0;
1915 1917 else
1916 1918 all_zero = B_FALSE;
1917 1919
1918 1920 asize = vdev_psize_to_asize(vd, psize);
1919 1921 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
1920 1922
1921 1923 offset = metaslab_group_alloc(mg, psize, asize, txg, distance,
1922 1924 dva, d, flags);
1923 1925 if (offset != -1ULL) {
1924 1926 /*
1925 1927 * If we've just selected this metaslab group,
1926 1928 * figure out whether the corresponding vdev is
1927 1929 * over- or under-used relative to the pool,
1928 1930 * and set an allocation bias to even it out.
1929 1931 */
1930 1932 if (mc->mc_aliquot == 0) {
1931 1933 vdev_stat_t *vs = &vd->vdev_stat;
1932 1934 int64_t vu, cu;
1933 1935
1934 1936 vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
1935 1937 cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
1936 1938
1937 1939 /*
1938 1940 * Calculate how much more or less we should
1939 1941 * try to allocate from this device during
1940 1942 * this iteration around the rotor.
1941 1943 * For example, if a device is 80% full
1942 1944 * and the pool is 20% full then we should
1943 1945 * reduce allocations by 60% on this device.
1944 1946 *
1945 1947 * mg_bias = (20 - 80) * 512K / 100 = -307K
1946 1948 *
1947 1949 * This reduces allocations by 307K for this
1948 1950 * iteration.
1949 1951 */
1950 1952 mg->mg_bias = ((cu - vu) *
1951 1953 (int64_t)mg->mg_aliquot) / 100;
1952 1954 }
1953 1955
1954 1956 if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
1955 1957 mg->mg_aliquot + mg->mg_bias) {
1956 1958 mc->mc_rotor = mg->mg_next;
1957 1959 mc->mc_aliquot = 0;
1958 1960 }
1959 1961
1960 1962 DVA_SET_VDEV(&dva[d], vd->vdev_id);
1961 1963 DVA_SET_OFFSET(&dva[d], offset);
1962 1964 DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
1963 1965 DVA_SET_ASIZE(&dva[d], asize);
1964 1966
1965 1967 return (0);
1966 1968 }
1967 1969 next:
1968 1970 mc->mc_rotor = mg->mg_next;
1969 1971 mc->mc_aliquot = 0;
1970 1972 } while ((mg = mg->mg_next) != rotor);
1971 1973
1972 1974 if (!all_zero) {
1973 1975 dshift++;
1974 1976 ASSERT(dshift < 64);
1975 1977 goto top;
1976 1978 }
1977 1979
1978 1980 if (!allocatable && !zio_lock) {
1979 1981 dshift = 3;
1980 1982 zio_lock = B_TRUE;
1981 1983 goto top;
1982 1984 }
1983 1985
1984 1986 bzero(&dva[d], sizeof (dva_t));
1985 1987
1986 1988 return (SET_ERROR(ENOSPC));
1987 1989 }
1988 1990
1989 1991 /*
1990 1992 * Free the block represented by DVA in the context of the specified
1991 1993 * transaction group.
1992 1994 */
1993 1995 static void
1994 1996 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
1995 1997 {
1996 1998 uint64_t vdev = DVA_GET_VDEV(dva);
1997 1999 uint64_t offset = DVA_GET_OFFSET(dva);
1998 2000 uint64_t size = DVA_GET_ASIZE(dva);
1999 2001 vdev_t *vd;
2000 2002 metaslab_t *msp;
2001 2003
2002 2004 ASSERT(DVA_IS_VALID(dva));
2003 2005
2004 2006 if (txg > spa_freeze_txg(spa))
2005 2007 return;
2006 2008
2007 2009 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
2008 2010 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
2009 2011 cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
2010 2012 (u_longlong_t)vdev, (u_longlong_t)offset);
2011 2013 ASSERT(0);
2012 2014 return;
2013 2015 }
2014 2016
2015 2017 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2016 2018
2017 2019 if (DVA_GET_GANG(dva))
2018 2020 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
2019 2021
2020 2022 mutex_enter(&msp->ms_lock);
2021 2023
2022 2024 if (now) {
2023 2025 range_tree_remove(msp->ms_alloctree[txg & TXG_MASK],
2024 2026 offset, size);
2025 2027
2026 2028 VERIFY(!msp->ms_condensing);
2027 2029 VERIFY3U(offset, >=, msp->ms_start);
2028 2030 VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size);
2029 2031 VERIFY3U(range_tree_space(msp->ms_tree) + size, <=,
2030 2032 msp->ms_size);
2031 2033 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
2032 2034 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2033 2035 range_tree_add(msp->ms_tree, offset, size);
2034 2036 } else {
2035 2037 if (range_tree_space(msp->ms_freetree[txg & TXG_MASK]) == 0)
2036 2038 vdev_dirty(vd, VDD_METASLAB, msp, txg);
2037 2039 range_tree_add(msp->ms_freetree[txg & TXG_MASK],
2038 2040 offset, size);
2039 2041 }
2040 2042
2041 2043 mutex_exit(&msp->ms_lock);
2042 2044 }
2043 2045
2044 2046 /*
2045 2047 * Intent log support: upon opening the pool after a crash, notify the SPA
2046 2048 * of blocks that the intent log has allocated for immediate write, but
2047 2049 * which are still considered free by the SPA because the last transaction
2048 2050 * group didn't commit yet.
2049 2051 */
2050 2052 static int
2051 2053 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
2052 2054 {
2053 2055 uint64_t vdev = DVA_GET_VDEV(dva);
2054 2056 uint64_t offset = DVA_GET_OFFSET(dva);
2055 2057 uint64_t size = DVA_GET_ASIZE(dva);
2056 2058 vdev_t *vd;
2057 2059 metaslab_t *msp;
2058 2060 int error = 0;
2059 2061
2060 2062 ASSERT(DVA_IS_VALID(dva));
2061 2063
2062 2064 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
2063 2065 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
2064 2066 return (SET_ERROR(ENXIO));
2065 2067
2066 2068 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2067 2069
2068 2070 if (DVA_GET_GANG(dva))
2069 2071 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
2070 2072
2071 2073 mutex_enter(&msp->ms_lock);
2072 2074
2073 2075 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded)
2074 2076 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
2075 2077
2076 2078 if (error == 0 && !range_tree_contains(msp->ms_tree, offset, size))
2077 2079 error = SET_ERROR(ENOENT);
2078 2080
2079 2081 if (error || txg == 0) { /* txg == 0 indicates dry run */
2080 2082 mutex_exit(&msp->ms_lock);
2081 2083 return (error);
2082 2084 }
2083 2085
2084 2086 VERIFY(!msp->ms_condensing);
2085 2087 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
2086 2088 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2087 2089 VERIFY3U(range_tree_space(msp->ms_tree) - size, <=, msp->ms_size);
2088 2090 range_tree_remove(msp->ms_tree, offset, size);
2089 2091
2090 2092 if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
2091 2093 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
2092 2094 vdev_dirty(vd, VDD_METASLAB, msp, txg);
2093 2095 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, size);
2094 2096 }
2095 2097
2096 2098 mutex_exit(&msp->ms_lock);
2097 2099
2098 2100 return (0);
2099 2101 }
2100 2102
2101 2103 int
2102 2104 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
2103 2105 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
2104 2106 {
2105 2107 dva_t *dva = bp->blk_dva;
2106 2108 dva_t *hintdva = hintbp->blk_dva;
2107 2109 int error = 0;
2108 2110
2109 2111 ASSERT(bp->blk_birth == 0);
2110 2112 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
2111 2113
2112 2114 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
2113 2115
2114 2116 if (mc->mc_rotor == NULL) { /* no vdevs in this class */
2115 2117 spa_config_exit(spa, SCL_ALLOC, FTAG);
2116 2118 return (SET_ERROR(ENOSPC));
2117 2119 }
2118 2120
2119 2121 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
2120 2122 ASSERT(BP_GET_NDVAS(bp) == 0);
2121 2123 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
2122 2124
2123 2125 for (int d = 0; d < ndvas; d++) {
2124 2126 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
2125 2127 txg, flags);
2126 2128 if (error != 0) {
2127 2129 for (d--; d >= 0; d--) {
2128 2130 metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
2129 2131 bzero(&dva[d], sizeof (dva_t));
2130 2132 }
2131 2133 spa_config_exit(spa, SCL_ALLOC, FTAG);
2132 2134 return (error);
2133 2135 }
2134 2136 }
2135 2137 ASSERT(error == 0);
2136 2138 ASSERT(BP_GET_NDVAS(bp) == ndvas);
2137 2139
2138 2140 spa_config_exit(spa, SCL_ALLOC, FTAG);
2139 2141
2140 2142 BP_SET_BIRTH(bp, txg, txg);
2141 2143
2142 2144 return (0);
2143 2145 }
2144 2146
2145 2147 void
2146 2148 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
2147 2149 {
2148 2150 const dva_t *dva = bp->blk_dva;
2149 2151 int ndvas = BP_GET_NDVAS(bp);
2150 2152
2151 2153 ASSERT(!BP_IS_HOLE(bp));
2152 2154 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
2153 2155
2154 2156 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
2155 2157
2156 2158 for (int d = 0; d < ndvas; d++)
2157 2159 metaslab_free_dva(spa, &dva[d], txg, now);
2158 2160
2159 2161 spa_config_exit(spa, SCL_FREE, FTAG);
2160 2162 }
2161 2163
2162 2164 int
2163 2165 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
2164 2166 {
2165 2167 const dva_t *dva = bp->blk_dva;
2166 2168 int ndvas = BP_GET_NDVAS(bp);
2167 2169 int error = 0;
2168 2170
2169 2171 ASSERT(!BP_IS_HOLE(bp));
2170 2172
2171 2173 if (txg != 0) {
2172 2174 /*
2173 2175 * First do a dry run to make sure all DVAs are claimable,
2174 2176 * so we don't have to unwind from partial failures below.
2175 2177 */
2176 2178 if ((error = metaslab_claim(spa, bp, 0)) != 0)
2177 2179 return (error);
2178 2180 }
2179 2181
2180 2182 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
2181 2183
2182 2184 for (int d = 0; d < ndvas; d++)
2183 2185 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
2184 2186 break;
2185 2187
2186 2188 spa_config_exit(spa, SCL_ALLOC, FTAG);
2187 2189
2188 2190 ASSERT(error == 0 || txg == 0);
2189 2191
2190 2192 return (error);
2191 2193 }
2192 2194
2193 2195 void
2194 2196 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
2195 2197 {
2196 2198 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
2197 2199 return;
2198 2200
2199 2201 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
2200 2202 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
2201 2203 uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
2202 2204 vdev_t *vd = vdev_lookup_top(spa, vdev);
2203 2205 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
2204 2206 uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
2205 2207 metaslab_t *msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2206 2208
2207 2209 if (msp->ms_loaded)
2208 2210 range_tree_verify(msp->ms_tree, offset, size);
2209 2211
2210 2212 for (int j = 0; j < TXG_SIZE; j++)
2211 2213 range_tree_verify(msp->ms_freetree[j], offset, size);
2212 2214 for (int j = 0; j < TXG_DEFER_SIZE; j++)
2213 2215 range_tree_verify(msp->ms_defertree[j], offset, size);
2214 2216 }
2215 2217 spa_config_exit(spa, SCL_VDEV, FTAG);
2216 2218 }
|
↓ open down ↓ |
1857 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX