]>
Commit | Line | Data |
---|---|---|
34dc7c2f BB |
1 | /* |
2 | * CDDL HEADER START | |
3 | * | |
4 | * The contents of this file are subject to the terms of the | |
5 | * Common Development and Distribution License (the "License"). | |
6 | * You may not use this file except in compliance with the License. | |
7 | * | |
8 | * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE | |
9 | * or http://www.opensolaris.org/os/licensing. | |
10 | * See the License for the specific language governing permissions | |
11 | * and limitations under the License. | |
12 | * | |
13 | * When distributing Covered Code, include this CDDL HEADER in each | |
14 | * file and include the License file at usr/src/OPENSOLARIS.LICENSE. | |
15 | * If applicable, add the following below this CDDL HEADER, with the | |
16 | * fields enclosed by brackets "[]" replaced with your own identifying | |
17 | * information: Portions Copyright [yyyy] [name of copyright owner] | |
18 | * | |
19 | * CDDL HEADER END | |
20 | */ | |
21 | /* | |
428870ff | 22 | * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. |
ebf8e3a2 | 23 | * Copyright (c) 2012 by Delphix. All rights reserved. |
34dc7c2f BB |
24 | */ |
25 | ||
34dc7c2f | 26 | #include <sys/zfs_context.h> |
34dc7c2f BB |
27 | #include <sys/dmu.h> |
28 | #include <sys/dmu_tx.h> | |
29 | #include <sys/space_map.h> | |
30 | #include <sys/metaslab_impl.h> | |
31 | #include <sys/vdev_impl.h> | |
32 | #include <sys/zio.h> | |
33 | ||
6d974228 GW |
34 | #define WITH_DF_BLOCK_ALLOCATOR |
35 | ||
36 | /* | |
37 | * Allow allocations to switch to gang blocks quickly. We do this to | |
38 | * avoid having to load lots of space_maps in a given txg. There are, | |
39 | * however, some cases where we want to avoid "fast" ganging and instead | |
40 | * we want to do an exhaustive search of all metaslabs on this device. | |
ebf8e3a2 | 41 | * Currently we don't allow any gang, zil, or dump device related allocations |
6d974228 GW |
42 | * to "fast" gang. |
43 | */ | |
44 | #define CAN_FASTGANG(flags) \ | |
45 | (!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \ | |
46 | METASLAB_GANG_AVOID))) | |
22c81dd8 | 47 | |
34dc7c2f BB |
48 | uint64_t metaslab_aliquot = 512ULL << 10; |
49 | uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1; /* force gang blocks */ | |
50 | ||
e51be066 GW |
51 | /* |
52 | * The in-core space map representation is more compact than its on-disk form. | |
53 | * The zfs_condense_pct determines how much more compact the in-core | |
54 | * space_map representation must be before we compact it on-disk. | |
55 | * Values should be greater than or equal to 100. | |
56 | */ | |
57 | int zfs_condense_pct = 200; | |
58 | ||
6d974228 GW |
59 | /* |
60 | * This value defines the number of allowed allocation failures per vdev. | |
61 | * If a device reaches this threshold in a given txg then we consider skipping | |
62 | * allocations on that device. | |
63 | */ | |
64 | int zfs_mg_alloc_failures; | |
65 | ||
428870ff BB |
66 | /* |
67 | * Metaslab debugging: when set, keeps all space maps in core to verify frees. | |
68 | */ | |
30b92c1d | 69 | int metaslab_debug = 0; |
428870ff | 70 | |
9babb374 BB |
71 | /* |
72 | * Minimum size which forces the dynamic allocator to change | |
428870ff | 73 | * it's allocation strategy. Once the space map cannot satisfy |
9babb374 BB |
74 | * an allocation of this size then it switches to using more |
75 | * aggressive strategy (i.e search by size rather than offset). | |
76 | */ | |
77 | uint64_t metaslab_df_alloc_threshold = SPA_MAXBLOCKSIZE; | |
78 | ||
79 | /* | |
80 | * The minimum free space, in percent, which must be available | |
81 | * in a space map to continue allocations in a first-fit fashion. | |
82 | * Once the space_map's free space drops below this level we dynamically | |
83 | * switch to using best-fit allocations. | |
84 | */ | |
428870ff BB |
85 | int metaslab_df_free_pct = 4; |
86 | ||
87 | /* | |
88 | * A metaslab is considered "free" if it contains a contiguous | |
89 | * segment which is greater than metaslab_min_alloc_size. | |
90 | */ | |
91 | uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS; | |
92 | ||
93 | /* | |
94 | * Max number of space_maps to prefetch. | |
95 | */ | |
96 | int metaslab_prefetch_limit = SPA_DVAS_PER_BP; | |
97 | ||
98 | /* | |
99 | * Percentage bonus multiplier for metaslabs that are in the bonus area. | |
100 | */ | |
101 | int metaslab_smo_bonus_pct = 150; | |
9babb374 | 102 | |
34dc7c2f BB |
103 | /* |
104 | * ========================================================================== | |
105 | * Metaslab classes | |
106 | * ========================================================================== | |
107 | */ | |
108 | metaslab_class_t * | |
428870ff | 109 | metaslab_class_create(spa_t *spa, space_map_ops_t *ops) |
34dc7c2f BB |
110 | { |
111 | metaslab_class_t *mc; | |
112 | ||
b8d06fca | 113 | mc = kmem_zalloc(sizeof (metaslab_class_t), KM_PUSHPAGE); |
34dc7c2f | 114 | |
428870ff | 115 | mc->mc_spa = spa; |
34dc7c2f | 116 | mc->mc_rotor = NULL; |
9babb374 | 117 | mc->mc_ops = ops; |
920dd524 | 118 | mutex_init(&mc->mc_fastwrite_lock, NULL, MUTEX_DEFAULT, NULL); |
34dc7c2f BB |
119 | |
120 | return (mc); | |
121 | } | |
122 | ||
123 | void | |
124 | metaslab_class_destroy(metaslab_class_t *mc) | |
125 | { | |
428870ff BB |
126 | ASSERT(mc->mc_rotor == NULL); |
127 | ASSERT(mc->mc_alloc == 0); | |
128 | ASSERT(mc->mc_deferred == 0); | |
129 | ASSERT(mc->mc_space == 0); | |
130 | ASSERT(mc->mc_dspace == 0); | |
34dc7c2f | 131 | |
920dd524 | 132 | mutex_destroy(&mc->mc_fastwrite_lock); |
34dc7c2f BB |
133 | kmem_free(mc, sizeof (metaslab_class_t)); |
134 | } | |
135 | ||
428870ff BB |
136 | int |
137 | metaslab_class_validate(metaslab_class_t *mc) | |
34dc7c2f | 138 | { |
428870ff BB |
139 | metaslab_group_t *mg; |
140 | vdev_t *vd; | |
34dc7c2f | 141 | |
428870ff BB |
142 | /* |
143 | * Must hold one of the spa_config locks. | |
144 | */ | |
145 | ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) || | |
146 | spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER)); | |
34dc7c2f | 147 | |
428870ff BB |
148 | if ((mg = mc->mc_rotor) == NULL) |
149 | return (0); | |
150 | ||
151 | do { | |
152 | vd = mg->mg_vd; | |
153 | ASSERT(vd->vdev_mg != NULL); | |
154 | ASSERT3P(vd->vdev_top, ==, vd); | |
155 | ASSERT3P(mg->mg_class, ==, mc); | |
156 | ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops); | |
157 | } while ((mg = mg->mg_next) != mc->mc_rotor); | |
158 | ||
159 | return (0); | |
34dc7c2f BB |
160 | } |
161 | ||
162 | void | |
428870ff BB |
163 | metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta, |
164 | int64_t defer_delta, int64_t space_delta, int64_t dspace_delta) | |
34dc7c2f | 165 | { |
428870ff BB |
166 | atomic_add_64(&mc->mc_alloc, alloc_delta); |
167 | atomic_add_64(&mc->mc_deferred, defer_delta); | |
168 | atomic_add_64(&mc->mc_space, space_delta); | |
169 | atomic_add_64(&mc->mc_dspace, dspace_delta); | |
170 | } | |
34dc7c2f | 171 | |
428870ff BB |
172 | uint64_t |
173 | metaslab_class_get_alloc(metaslab_class_t *mc) | |
174 | { | |
175 | return (mc->mc_alloc); | |
176 | } | |
34dc7c2f | 177 | |
428870ff BB |
178 | uint64_t |
179 | metaslab_class_get_deferred(metaslab_class_t *mc) | |
180 | { | |
181 | return (mc->mc_deferred); | |
182 | } | |
34dc7c2f | 183 | |
428870ff BB |
184 | uint64_t |
185 | metaslab_class_get_space(metaslab_class_t *mc) | |
186 | { | |
187 | return (mc->mc_space); | |
188 | } | |
34dc7c2f | 189 | |
428870ff BB |
190 | uint64_t |
191 | metaslab_class_get_dspace(metaslab_class_t *mc) | |
192 | { | |
193 | return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space); | |
34dc7c2f BB |
194 | } |
195 | ||
196 | /* | |
197 | * ========================================================================== | |
198 | * Metaslab groups | |
199 | * ========================================================================== | |
200 | */ | |
201 | static int | |
202 | metaslab_compare(const void *x1, const void *x2) | |
203 | { | |
204 | const metaslab_t *m1 = x1; | |
205 | const metaslab_t *m2 = x2; | |
206 | ||
207 | if (m1->ms_weight < m2->ms_weight) | |
208 | return (1); | |
209 | if (m1->ms_weight > m2->ms_weight) | |
210 | return (-1); | |
211 | ||
212 | /* | |
213 | * If the weights are identical, use the offset to force uniqueness. | |
214 | */ | |
e51be066 | 215 | if (m1->ms_map->sm_start < m2->ms_map->sm_start) |
34dc7c2f | 216 | return (-1); |
e51be066 | 217 | if (m1->ms_map->sm_start > m2->ms_map->sm_start) |
34dc7c2f BB |
218 | return (1); |
219 | ||
220 | ASSERT3P(m1, ==, m2); | |
221 | ||
222 | return (0); | |
223 | } | |
224 | ||
225 | metaslab_group_t * | |
226 | metaslab_group_create(metaslab_class_t *mc, vdev_t *vd) | |
227 | { | |
228 | metaslab_group_t *mg; | |
229 | ||
b8d06fca | 230 | mg = kmem_zalloc(sizeof (metaslab_group_t), KM_PUSHPAGE); |
34dc7c2f BB |
231 | mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL); |
232 | avl_create(&mg->mg_metaslab_tree, metaslab_compare, | |
233 | sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node)); | |
34dc7c2f | 234 | mg->mg_vd = vd; |
428870ff BB |
235 | mg->mg_class = mc; |
236 | mg->mg_activation_count = 0; | |
34dc7c2f BB |
237 | |
238 | return (mg); | |
239 | } | |
240 | ||
241 | void | |
242 | metaslab_group_destroy(metaslab_group_t *mg) | |
243 | { | |
428870ff BB |
244 | ASSERT(mg->mg_prev == NULL); |
245 | ASSERT(mg->mg_next == NULL); | |
246 | /* | |
247 | * We may have gone below zero with the activation count | |
248 | * either because we never activated in the first place or | |
249 | * because we're done, and possibly removing the vdev. | |
250 | */ | |
251 | ASSERT(mg->mg_activation_count <= 0); | |
252 | ||
34dc7c2f BB |
253 | avl_destroy(&mg->mg_metaslab_tree); |
254 | mutex_destroy(&mg->mg_lock); | |
255 | kmem_free(mg, sizeof (metaslab_group_t)); | |
256 | } | |
257 | ||
428870ff BB |
258 | void |
259 | metaslab_group_activate(metaslab_group_t *mg) | |
260 | { | |
261 | metaslab_class_t *mc = mg->mg_class; | |
262 | metaslab_group_t *mgprev, *mgnext; | |
263 | ||
264 | ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER)); | |
265 | ||
266 | ASSERT(mc->mc_rotor != mg); | |
267 | ASSERT(mg->mg_prev == NULL); | |
268 | ASSERT(mg->mg_next == NULL); | |
269 | ASSERT(mg->mg_activation_count <= 0); | |
270 | ||
271 | if (++mg->mg_activation_count <= 0) | |
272 | return; | |
273 | ||
274 | mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children); | |
275 | ||
276 | if ((mgprev = mc->mc_rotor) == NULL) { | |
277 | mg->mg_prev = mg; | |
278 | mg->mg_next = mg; | |
279 | } else { | |
280 | mgnext = mgprev->mg_next; | |
281 | mg->mg_prev = mgprev; | |
282 | mg->mg_next = mgnext; | |
283 | mgprev->mg_next = mg; | |
284 | mgnext->mg_prev = mg; | |
285 | } | |
286 | mc->mc_rotor = mg; | |
287 | } | |
288 | ||
289 | void | |
290 | metaslab_group_passivate(metaslab_group_t *mg) | |
291 | { | |
292 | metaslab_class_t *mc = mg->mg_class; | |
293 | metaslab_group_t *mgprev, *mgnext; | |
294 | ||
295 | ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER)); | |
296 | ||
297 | if (--mg->mg_activation_count != 0) { | |
298 | ASSERT(mc->mc_rotor != mg); | |
299 | ASSERT(mg->mg_prev == NULL); | |
300 | ASSERT(mg->mg_next == NULL); | |
301 | ASSERT(mg->mg_activation_count < 0); | |
302 | return; | |
303 | } | |
304 | ||
305 | mgprev = mg->mg_prev; | |
306 | mgnext = mg->mg_next; | |
307 | ||
308 | if (mg == mgnext) { | |
309 | mc->mc_rotor = NULL; | |
310 | } else { | |
311 | mc->mc_rotor = mgnext; | |
312 | mgprev->mg_next = mgnext; | |
313 | mgnext->mg_prev = mgprev; | |
314 | } | |
315 | ||
316 | mg->mg_prev = NULL; | |
317 | mg->mg_next = NULL; | |
318 | } | |
319 | ||
34dc7c2f BB |
320 | static void |
321 | metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp) | |
322 | { | |
323 | mutex_enter(&mg->mg_lock); | |
324 | ASSERT(msp->ms_group == NULL); | |
325 | msp->ms_group = mg; | |
326 | msp->ms_weight = 0; | |
327 | avl_add(&mg->mg_metaslab_tree, msp); | |
328 | mutex_exit(&mg->mg_lock); | |
329 | } | |
330 | ||
331 | static void | |
332 | metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp) | |
333 | { | |
334 | mutex_enter(&mg->mg_lock); | |
335 | ASSERT(msp->ms_group == mg); | |
336 | avl_remove(&mg->mg_metaslab_tree, msp); | |
337 | msp->ms_group = NULL; | |
338 | mutex_exit(&mg->mg_lock); | |
339 | } | |
340 | ||
341 | static void | |
342 | metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight) | |
343 | { | |
344 | /* | |
345 | * Although in principle the weight can be any value, in | |
346 | * practice we do not use values in the range [1, 510]. | |
347 | */ | |
348 | ASSERT(weight >= SPA_MINBLOCKSIZE-1 || weight == 0); | |
349 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
350 | ||
351 | mutex_enter(&mg->mg_lock); | |
352 | ASSERT(msp->ms_group == mg); | |
353 | avl_remove(&mg->mg_metaslab_tree, msp); | |
354 | msp->ms_weight = weight; | |
355 | avl_add(&mg->mg_metaslab_tree, msp); | |
356 | mutex_exit(&mg->mg_lock); | |
357 | } | |
358 | ||
428870ff BB |
359 | /* |
360 | * ========================================================================== | |
361 | * Common allocator routines | |
362 | * ========================================================================== | |
363 | */ | |
364 | static int | |
365 | metaslab_segsize_compare(const void *x1, const void *x2) | |
366 | { | |
367 | const space_seg_t *s1 = x1; | |
368 | const space_seg_t *s2 = x2; | |
369 | uint64_t ss_size1 = s1->ss_end - s1->ss_start; | |
370 | uint64_t ss_size2 = s2->ss_end - s2->ss_start; | |
371 | ||
372 | if (ss_size1 < ss_size2) | |
373 | return (-1); | |
374 | if (ss_size1 > ss_size2) | |
375 | return (1); | |
376 | ||
377 | if (s1->ss_start < s2->ss_start) | |
378 | return (-1); | |
379 | if (s1->ss_start > s2->ss_start) | |
380 | return (1); | |
381 | ||
382 | return (0); | |
383 | } | |
384 | ||
22c81dd8 BB |
385 | #if defined(WITH_FF_BLOCK_ALLOCATOR) || \ |
386 | defined(WITH_DF_BLOCK_ALLOCATOR) || \ | |
387 | defined(WITH_CDF_BLOCK_ALLOCATOR) | |
34dc7c2f | 388 | /* |
9babb374 BB |
389 | * This is a helper function that can be used by the allocator to find |
390 | * a suitable block to allocate. This will search the specified AVL | |
391 | * tree looking for a block that matches the specified criteria. | |
34dc7c2f | 392 | */ |
34dc7c2f | 393 | static uint64_t |
9babb374 BB |
394 | metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size, |
395 | uint64_t align) | |
34dc7c2f | 396 | { |
34dc7c2f BB |
397 | space_seg_t *ss, ssearch; |
398 | avl_index_t where; | |
399 | ||
400 | ssearch.ss_start = *cursor; | |
401 | ssearch.ss_end = *cursor + size; | |
402 | ||
403 | ss = avl_find(t, &ssearch, &where); | |
404 | if (ss == NULL) | |
405 | ss = avl_nearest(t, where, AVL_AFTER); | |
406 | ||
407 | while (ss != NULL) { | |
408 | uint64_t offset = P2ROUNDUP(ss->ss_start, align); | |
409 | ||
410 | if (offset + size <= ss->ss_end) { | |
411 | *cursor = offset + size; | |
412 | return (offset); | |
413 | } | |
414 | ss = AVL_NEXT(t, ss); | |
415 | } | |
416 | ||
417 | /* | |
418 | * If we know we've searched the whole map (*cursor == 0), give up. | |
419 | * Otherwise, reset the cursor to the beginning and try again. | |
420 | */ | |
421 | if (*cursor == 0) | |
422 | return (-1ULL); | |
423 | ||
424 | *cursor = 0; | |
9babb374 BB |
425 | return (metaslab_block_picker(t, cursor, size, align)); |
426 | } | |
22c81dd8 | 427 | #endif /* WITH_FF/DF/CDF_BLOCK_ALLOCATOR */ |
9babb374 | 428 | |
9babb374 | 429 | static void |
428870ff | 430 | metaslab_pp_load(space_map_t *sm) |
9babb374 | 431 | { |
428870ff BB |
432 | space_seg_t *ss; |
433 | ||
9babb374 | 434 | ASSERT(sm->sm_ppd == NULL); |
b8d06fca | 435 | sm->sm_ppd = kmem_zalloc(64 * sizeof (uint64_t), KM_PUSHPAGE); |
428870ff | 436 | |
b8d06fca | 437 | sm->sm_pp_root = kmem_alloc(sizeof (avl_tree_t), KM_PUSHPAGE); |
428870ff BB |
438 | avl_create(sm->sm_pp_root, metaslab_segsize_compare, |
439 | sizeof (space_seg_t), offsetof(struct space_seg, ss_pp_node)); | |
440 | ||
441 | for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss)) | |
442 | avl_add(sm->sm_pp_root, ss); | |
9babb374 BB |
443 | } |
444 | ||
445 | static void | |
428870ff | 446 | metaslab_pp_unload(space_map_t *sm) |
9babb374 | 447 | { |
428870ff BB |
448 | void *cookie = NULL; |
449 | ||
9babb374 BB |
450 | kmem_free(sm->sm_ppd, 64 * sizeof (uint64_t)); |
451 | sm->sm_ppd = NULL; | |
9babb374 | 452 | |
428870ff BB |
453 | while (avl_destroy_nodes(sm->sm_pp_root, &cookie) != NULL) { |
454 | /* tear down the tree */ | |
455 | } | |
9babb374 | 456 | |
428870ff BB |
457 | avl_destroy(sm->sm_pp_root); |
458 | kmem_free(sm->sm_pp_root, sizeof (avl_tree_t)); | |
459 | sm->sm_pp_root = NULL; | |
34dc7c2f BB |
460 | } |
461 | ||
462 | /* ARGSUSED */ | |
463 | static void | |
428870ff | 464 | metaslab_pp_claim(space_map_t *sm, uint64_t start, uint64_t size) |
34dc7c2f BB |
465 | { |
466 | /* No need to update cursor */ | |
467 | } | |
468 | ||
469 | /* ARGSUSED */ | |
470 | static void | |
428870ff | 471 | metaslab_pp_free(space_map_t *sm, uint64_t start, uint64_t size) |
34dc7c2f BB |
472 | { |
473 | /* No need to update cursor */ | |
474 | } | |
475 | ||
9babb374 | 476 | /* |
428870ff | 477 | * Return the maximum contiguous segment within the metaslab. |
9babb374 | 478 | */ |
9babb374 | 479 | uint64_t |
428870ff | 480 | metaslab_pp_maxsize(space_map_t *sm) |
9babb374 BB |
481 | { |
482 | avl_tree_t *t = sm->sm_pp_root; | |
483 | space_seg_t *ss; | |
484 | ||
485 | if (t == NULL || (ss = avl_last(t)) == NULL) | |
486 | return (0ULL); | |
487 | ||
488 | return (ss->ss_end - ss->ss_start); | |
489 | } | |
490 | ||
22c81dd8 | 491 | #if defined(WITH_FF_BLOCK_ALLOCATOR) |
428870ff BB |
492 | /* |
493 | * ========================================================================== | |
494 | * The first-fit block allocator | |
495 | * ========================================================================== | |
496 | */ | |
497 | static uint64_t | |
498 | metaslab_ff_alloc(space_map_t *sm, uint64_t size) | |
9babb374 | 499 | { |
428870ff BB |
500 | avl_tree_t *t = &sm->sm_root; |
501 | uint64_t align = size & -size; | |
502 | uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1; | |
9babb374 | 503 | |
428870ff | 504 | return (metaslab_block_picker(t, cursor, size, align)); |
9babb374 BB |
505 | } |
506 | ||
428870ff BB |
507 | /* ARGSUSED */ |
508 | boolean_t | |
509 | metaslab_ff_fragmented(space_map_t *sm) | |
9babb374 | 510 | { |
428870ff | 511 | return (B_TRUE); |
9babb374 BB |
512 | } |
513 | ||
428870ff BB |
514 | static space_map_ops_t metaslab_ff_ops = { |
515 | metaslab_pp_load, | |
516 | metaslab_pp_unload, | |
517 | metaslab_ff_alloc, | |
518 | metaslab_pp_claim, | |
519 | metaslab_pp_free, | |
520 | metaslab_pp_maxsize, | |
521 | metaslab_ff_fragmented | |
522 | }; | |
9babb374 | 523 | |
22c81dd8 BB |
524 | space_map_ops_t *zfs_metaslab_ops = &metaslab_ff_ops; |
525 | #endif /* WITH_FF_BLOCK_ALLOCATOR */ | |
526 | ||
527 | #if defined(WITH_DF_BLOCK_ALLOCATOR) | |
428870ff BB |
528 | /* |
529 | * ========================================================================== | |
530 | * Dynamic block allocator - | |
531 | * Uses the first fit allocation scheme until space get low and then | |
532 | * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold | |
533 | * and metaslab_df_free_pct to determine when to switch the allocation scheme. | |
534 | * ========================================================================== | |
535 | */ | |
9babb374 BB |
536 | static uint64_t |
537 | metaslab_df_alloc(space_map_t *sm, uint64_t size) | |
538 | { | |
539 | avl_tree_t *t = &sm->sm_root; | |
540 | uint64_t align = size & -size; | |
541 | uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1; | |
428870ff | 542 | uint64_t max_size = metaslab_pp_maxsize(sm); |
9babb374 BB |
543 | int free_pct = sm->sm_space * 100 / sm->sm_size; |
544 | ||
545 | ASSERT(MUTEX_HELD(sm->sm_lock)); | |
546 | ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root)); | |
547 | ||
548 | if (max_size < size) | |
549 | return (-1ULL); | |
550 | ||
551 | /* | |
552 | * If we're running low on space switch to using the size | |
553 | * sorted AVL tree (best-fit). | |
554 | */ | |
555 | if (max_size < metaslab_df_alloc_threshold || | |
556 | free_pct < metaslab_df_free_pct) { | |
557 | t = sm->sm_pp_root; | |
558 | *cursor = 0; | |
559 | } | |
560 | ||
561 | return (metaslab_block_picker(t, cursor, size, 1ULL)); | |
562 | } | |
563 | ||
428870ff BB |
564 | static boolean_t |
565 | metaslab_df_fragmented(space_map_t *sm) | |
9babb374 | 566 | { |
428870ff BB |
567 | uint64_t max_size = metaslab_pp_maxsize(sm); |
568 | int free_pct = sm->sm_space * 100 / sm->sm_size; | |
9babb374 | 569 | |
428870ff BB |
570 | if (max_size >= metaslab_df_alloc_threshold && |
571 | free_pct >= metaslab_df_free_pct) | |
572 | return (B_FALSE); | |
573 | ||
574 | return (B_TRUE); | |
9babb374 BB |
575 | } |
576 | ||
577 | static space_map_ops_t metaslab_df_ops = { | |
428870ff BB |
578 | metaslab_pp_load, |
579 | metaslab_pp_unload, | |
9babb374 | 580 | metaslab_df_alloc, |
428870ff BB |
581 | metaslab_pp_claim, |
582 | metaslab_pp_free, | |
583 | metaslab_pp_maxsize, | |
584 | metaslab_df_fragmented | |
34dc7c2f BB |
585 | }; |
586 | ||
22c81dd8 BB |
587 | space_map_ops_t *zfs_metaslab_ops = &metaslab_df_ops; |
588 | #endif /* WITH_DF_BLOCK_ALLOCATOR */ | |
589 | ||
428870ff BB |
590 | /* |
591 | * ========================================================================== | |
592 | * Other experimental allocators | |
593 | * ========================================================================== | |
594 | */ | |
22c81dd8 | 595 | #if defined(WITH_CDF_BLOCK_ALLOCATOR) |
428870ff BB |
596 | static uint64_t |
597 | metaslab_cdf_alloc(space_map_t *sm, uint64_t size) | |
598 | { | |
599 | avl_tree_t *t = &sm->sm_root; | |
600 | uint64_t *cursor = (uint64_t *)sm->sm_ppd; | |
601 | uint64_t *extent_end = (uint64_t *)sm->sm_ppd + 1; | |
602 | uint64_t max_size = metaslab_pp_maxsize(sm); | |
603 | uint64_t rsize = size; | |
604 | uint64_t offset = 0; | |
605 | ||
606 | ASSERT(MUTEX_HELD(sm->sm_lock)); | |
607 | ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root)); | |
608 | ||
609 | if (max_size < size) | |
610 | return (-1ULL); | |
611 | ||
612 | ASSERT3U(*extent_end, >=, *cursor); | |
613 | ||
614 | /* | |
615 | * If we're running low on space switch to using the size | |
616 | * sorted AVL tree (best-fit). | |
617 | */ | |
618 | if ((*cursor + size) > *extent_end) { | |
619 | ||
620 | t = sm->sm_pp_root; | |
621 | *cursor = *extent_end = 0; | |
622 | ||
623 | if (max_size > 2 * SPA_MAXBLOCKSIZE) | |
624 | rsize = MIN(metaslab_min_alloc_size, max_size); | |
625 | offset = metaslab_block_picker(t, extent_end, rsize, 1ULL); | |
626 | if (offset != -1) | |
627 | *cursor = offset + size; | |
628 | } else { | |
629 | offset = metaslab_block_picker(t, cursor, rsize, 1ULL); | |
630 | } | |
631 | ASSERT3U(*cursor, <=, *extent_end); | |
632 | return (offset); | |
633 | } | |
634 | ||
635 | static boolean_t | |
636 | metaslab_cdf_fragmented(space_map_t *sm) | |
637 | { | |
638 | uint64_t max_size = metaslab_pp_maxsize(sm); | |
639 | ||
640 | if (max_size > (metaslab_min_alloc_size * 10)) | |
641 | return (B_FALSE); | |
642 | return (B_TRUE); | |
643 | } | |
644 | ||
645 | static space_map_ops_t metaslab_cdf_ops = { | |
646 | metaslab_pp_load, | |
647 | metaslab_pp_unload, | |
648 | metaslab_cdf_alloc, | |
649 | metaslab_pp_claim, | |
650 | metaslab_pp_free, | |
651 | metaslab_pp_maxsize, | |
652 | metaslab_cdf_fragmented | |
653 | }; | |
654 | ||
22c81dd8 BB |
655 | space_map_ops_t *zfs_metaslab_ops = &metaslab_cdf_ops; |
656 | #endif /* WITH_CDF_BLOCK_ALLOCATOR */ | |
657 | ||
658 | #if defined(WITH_NDF_BLOCK_ALLOCATOR) | |
428870ff BB |
659 | uint64_t metaslab_ndf_clump_shift = 4; |
660 | ||
661 | static uint64_t | |
662 | metaslab_ndf_alloc(space_map_t *sm, uint64_t size) | |
663 | { | |
664 | avl_tree_t *t = &sm->sm_root; | |
665 | avl_index_t where; | |
666 | space_seg_t *ss, ssearch; | |
667 | uint64_t hbit = highbit(size); | |
668 | uint64_t *cursor = (uint64_t *)sm->sm_ppd + hbit - 1; | |
669 | uint64_t max_size = metaslab_pp_maxsize(sm); | |
670 | ||
671 | ASSERT(MUTEX_HELD(sm->sm_lock)); | |
672 | ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root)); | |
673 | ||
674 | if (max_size < size) | |
675 | return (-1ULL); | |
676 | ||
677 | ssearch.ss_start = *cursor; | |
678 | ssearch.ss_end = *cursor + size; | |
679 | ||
680 | ss = avl_find(t, &ssearch, &where); | |
681 | if (ss == NULL || (ss->ss_start + size > ss->ss_end)) { | |
682 | t = sm->sm_pp_root; | |
683 | ||
684 | ssearch.ss_start = 0; | |
685 | ssearch.ss_end = MIN(max_size, | |
686 | 1ULL << (hbit + metaslab_ndf_clump_shift)); | |
687 | ss = avl_find(t, &ssearch, &where); | |
688 | if (ss == NULL) | |
689 | ss = avl_nearest(t, where, AVL_AFTER); | |
690 | ASSERT(ss != NULL); | |
691 | } | |
692 | ||
693 | if (ss != NULL) { | |
694 | if (ss->ss_start + size <= ss->ss_end) { | |
695 | *cursor = ss->ss_start + size; | |
696 | return (ss->ss_start); | |
697 | } | |
698 | } | |
699 | return (-1ULL); | |
700 | } | |
701 | ||
702 | static boolean_t | |
703 | metaslab_ndf_fragmented(space_map_t *sm) | |
704 | { | |
705 | uint64_t max_size = metaslab_pp_maxsize(sm); | |
706 | ||
707 | if (max_size > (metaslab_min_alloc_size << metaslab_ndf_clump_shift)) | |
708 | return (B_FALSE); | |
709 | return (B_TRUE); | |
710 | } | |
711 | ||
712 | ||
713 | static space_map_ops_t metaslab_ndf_ops = { | |
714 | metaslab_pp_load, | |
715 | metaslab_pp_unload, | |
716 | metaslab_ndf_alloc, | |
717 | metaslab_pp_claim, | |
718 | metaslab_pp_free, | |
719 | metaslab_pp_maxsize, | |
720 | metaslab_ndf_fragmented | |
721 | }; | |
722 | ||
723 | space_map_ops_t *zfs_metaslab_ops = &metaslab_ndf_ops; | |
22c81dd8 | 724 | #endif /* WITH_NDF_BLOCK_ALLOCATOR */ |
9babb374 | 725 | |
34dc7c2f BB |
726 | /* |
727 | * ========================================================================== | |
728 | * Metaslabs | |
729 | * ========================================================================== | |
730 | */ | |
731 | metaslab_t * | |
732 | metaslab_init(metaslab_group_t *mg, space_map_obj_t *smo, | |
733 | uint64_t start, uint64_t size, uint64_t txg) | |
734 | { | |
735 | vdev_t *vd = mg->mg_vd; | |
736 | metaslab_t *msp; | |
737 | ||
b8d06fca | 738 | msp = kmem_zalloc(sizeof (metaslab_t), KM_PUSHPAGE); |
34dc7c2f BB |
739 | mutex_init(&msp->ms_lock, NULL, MUTEX_DEFAULT, NULL); |
740 | ||
741 | msp->ms_smo_syncing = *smo; | |
742 | ||
743 | /* | |
744 | * We create the main space map here, but we don't create the | |
745 | * allocmaps and freemaps until metaslab_sync_done(). This serves | |
746 | * two purposes: it allows metaslab_sync_done() to detect the | |
747 | * addition of new space; and for debugging, it ensures that we'd | |
748 | * data fault on any attempt to use this metaslab before it's ready. | |
749 | */ | |
e51be066 GW |
750 | msp->ms_map = kmem_zalloc(sizeof (space_map_t), KM_PUSHPAGE); |
751 | space_map_create(msp->ms_map, start, size, | |
34dc7c2f BB |
752 | vd->vdev_ashift, &msp->ms_lock); |
753 | ||
754 | metaslab_group_add(mg, msp); | |
755 | ||
428870ff BB |
756 | if (metaslab_debug && smo->smo_object != 0) { |
757 | mutex_enter(&msp->ms_lock); | |
e51be066 | 758 | VERIFY(space_map_load(msp->ms_map, mg->mg_class->mc_ops, |
428870ff BB |
759 | SM_FREE, smo, spa_meta_objset(vd->vdev_spa)) == 0); |
760 | mutex_exit(&msp->ms_lock); | |
761 | } | |
762 | ||
34dc7c2f BB |
763 | /* |
764 | * If we're opening an existing pool (txg == 0) or creating | |
765 | * a new one (txg == TXG_INITIAL), all space is available now. | |
766 | * If we're adding space to an existing pool, the new space | |
767 | * does not become available until after this txg has synced. | |
768 | */ | |
769 | if (txg <= TXG_INITIAL) | |
770 | metaslab_sync_done(msp, 0); | |
771 | ||
772 | if (txg != 0) { | |
34dc7c2f | 773 | vdev_dirty(vd, 0, NULL, txg); |
428870ff | 774 | vdev_dirty(vd, VDD_METASLAB, msp, txg); |
34dc7c2f BB |
775 | } |
776 | ||
777 | return (msp); | |
778 | } | |
779 | ||
780 | void | |
781 | metaslab_fini(metaslab_t *msp) | |
782 | { | |
783 | metaslab_group_t *mg = msp->ms_group; | |
d6320ddb | 784 | int t; |
34dc7c2f | 785 | |
428870ff | 786 | vdev_space_update(mg->mg_vd, |
e51be066 | 787 | -msp->ms_smo.smo_alloc, 0, -msp->ms_map->sm_size); |
34dc7c2f BB |
788 | |
789 | metaslab_group_remove(mg, msp); | |
790 | ||
791 | mutex_enter(&msp->ms_lock); | |
792 | ||
e51be066 GW |
793 | space_map_unload(msp->ms_map); |
794 | space_map_destroy(msp->ms_map); | |
795 | kmem_free(msp->ms_map, sizeof (*msp->ms_map)); | |
34dc7c2f | 796 | |
d6320ddb | 797 | for (t = 0; t < TXG_SIZE; t++) { |
e51be066 GW |
798 | space_map_destroy(msp->ms_allocmap[t]); |
799 | space_map_destroy(msp->ms_freemap[t]); | |
800 | kmem_free(msp->ms_allocmap[t], sizeof (*msp->ms_allocmap[t])); | |
801 | kmem_free(msp->ms_freemap[t], sizeof (*msp->ms_freemap[t])); | |
34dc7c2f BB |
802 | } |
803 | ||
e51be066 GW |
804 | for (t = 0; t < TXG_DEFER_SIZE; t++) { |
805 | space_map_destroy(msp->ms_defermap[t]); | |
806 | kmem_free(msp->ms_defermap[t], sizeof (*msp->ms_defermap[t])); | |
807 | } | |
428870ff | 808 | |
c99c9001 | 809 | ASSERT0(msp->ms_deferspace); |
428870ff | 810 | |
34dc7c2f BB |
811 | mutex_exit(&msp->ms_lock); |
812 | mutex_destroy(&msp->ms_lock); | |
813 | ||
814 | kmem_free(msp, sizeof (metaslab_t)); | |
815 | } | |
816 | ||
817 | #define METASLAB_WEIGHT_PRIMARY (1ULL << 63) | |
818 | #define METASLAB_WEIGHT_SECONDARY (1ULL << 62) | |
819 | #define METASLAB_ACTIVE_MASK \ | |
820 | (METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY) | |
34dc7c2f BB |
821 | |
822 | static uint64_t | |
823 | metaslab_weight(metaslab_t *msp) | |
824 | { | |
825 | metaslab_group_t *mg = msp->ms_group; | |
e51be066 | 826 | space_map_t *sm = msp->ms_map; |
34dc7c2f BB |
827 | space_map_obj_t *smo = &msp->ms_smo; |
828 | vdev_t *vd = mg->mg_vd; | |
829 | uint64_t weight, space; | |
830 | ||
831 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
832 | ||
833 | /* | |
834 | * The baseline weight is the metaslab's free space. | |
835 | */ | |
836 | space = sm->sm_size - smo->smo_alloc; | |
837 | weight = space; | |
838 | ||
839 | /* | |
840 | * Modern disks have uniform bit density and constant angular velocity. | |
841 | * Therefore, the outer recording zones are faster (higher bandwidth) | |
842 | * than the inner zones by the ratio of outer to inner track diameter, | |
843 | * which is typically around 2:1. We account for this by assigning | |
844 | * higher weight to lower metaslabs (multiplier ranging from 2x to 1x). | |
845 | * In effect, this means that we'll select the metaslab with the most | |
846 | * free bandwidth rather than simply the one with the most free space. | |
847 | */ | |
848 | weight = 2 * weight - | |
849 | ((sm->sm_start >> vd->vdev_ms_shift) * weight) / vd->vdev_ms_count; | |
850 | ASSERT(weight >= space && weight <= 2 * space); | |
851 | ||
852 | /* | |
428870ff BB |
853 | * For locality, assign higher weight to metaslabs which have |
854 | * a lower offset than what we've already activated. | |
34dc7c2f | 855 | */ |
428870ff BB |
856 | if (sm->sm_start <= mg->mg_bonus_area) |
857 | weight *= (metaslab_smo_bonus_pct / 100); | |
34dc7c2f | 858 | ASSERT(weight >= space && |
428870ff BB |
859 | weight <= 2 * (metaslab_smo_bonus_pct / 100) * space); |
860 | ||
861 | if (sm->sm_loaded && !sm->sm_ops->smop_fragmented(sm)) { | |
862 | /* | |
863 | * If this metaslab is one we're actively using, adjust its | |
864 | * weight to make it preferable to any inactive metaslab so | |
865 | * we'll polish it off. | |
866 | */ | |
867 | weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK); | |
868 | } | |
869 | return (weight); | |
870 | } | |
871 | ||
872 | static void | |
873 | metaslab_prefetch(metaslab_group_t *mg) | |
874 | { | |
875 | spa_t *spa = mg->mg_vd->vdev_spa; | |
876 | metaslab_t *msp; | |
877 | avl_tree_t *t = &mg->mg_metaslab_tree; | |
878 | int m; | |
879 | ||
880 | mutex_enter(&mg->mg_lock); | |
34dc7c2f BB |
881 | |
882 | /* | |
428870ff | 883 | * Prefetch the next potential metaslabs |
34dc7c2f | 884 | */ |
428870ff | 885 | for (msp = avl_first(t), m = 0; msp; msp = AVL_NEXT(t, msp), m++) { |
e51be066 | 886 | space_map_t *sm = msp->ms_map; |
428870ff | 887 | space_map_obj_t *smo = &msp->ms_smo; |
34dc7c2f | 888 | |
428870ff BB |
889 | /* If we have reached our prefetch limit then we're done */ |
890 | if (m >= metaslab_prefetch_limit) | |
891 | break; | |
892 | ||
893 | if (!sm->sm_loaded && smo->smo_object != 0) { | |
894 | mutex_exit(&mg->mg_lock); | |
895 | dmu_prefetch(spa_meta_objset(spa), smo->smo_object, | |
896 | 0ULL, smo->smo_objsize); | |
897 | mutex_enter(&mg->mg_lock); | |
898 | } | |
899 | } | |
900 | mutex_exit(&mg->mg_lock); | |
34dc7c2f BB |
901 | } |
902 | ||
903 | static int | |
6d974228 | 904 | metaslab_activate(metaslab_t *msp, uint64_t activation_weight) |
34dc7c2f | 905 | { |
428870ff | 906 | metaslab_group_t *mg = msp->ms_group; |
e51be066 | 907 | space_map_t *sm = msp->ms_map; |
9babb374 | 908 | space_map_ops_t *sm_ops = msp->ms_group->mg_class->mc_ops; |
d6320ddb | 909 | int t; |
34dc7c2f BB |
910 | |
911 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
912 | ||
913 | if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) { | |
428870ff BB |
914 | space_map_load_wait(sm); |
915 | if (!sm->sm_loaded) { | |
55d85d5a GW |
916 | space_map_obj_t *smo = &msp->ms_smo; |
917 | ||
918 | int error = space_map_load(sm, sm_ops, SM_FREE, smo, | |
428870ff BB |
919 | spa_meta_objset(msp->ms_group->mg_vd->vdev_spa)); |
920 | if (error) { | |
921 | metaslab_group_sort(msp->ms_group, msp, 0); | |
922 | return (error); | |
923 | } | |
d6320ddb | 924 | for (t = 0; t < TXG_DEFER_SIZE; t++) |
e51be066 | 925 | space_map_walk(msp->ms_defermap[t], |
428870ff BB |
926 | space_map_claim, sm); |
927 | ||
928 | } | |
929 | ||
930 | /* | |
931 | * Track the bonus area as we activate new metaslabs. | |
932 | */ | |
933 | if (sm->sm_start > mg->mg_bonus_area) { | |
934 | mutex_enter(&mg->mg_lock); | |
935 | mg->mg_bonus_area = sm->sm_start; | |
936 | mutex_exit(&mg->mg_lock); | |
34dc7c2f | 937 | } |
9babb374 | 938 | |
34dc7c2f BB |
939 | metaslab_group_sort(msp->ms_group, msp, |
940 | msp->ms_weight | activation_weight); | |
941 | } | |
942 | ASSERT(sm->sm_loaded); | |
943 | ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK); | |
944 | ||
945 | return (0); | |
946 | } | |
947 | ||
948 | static void | |
949 | metaslab_passivate(metaslab_t *msp, uint64_t size) | |
950 | { | |
951 | /* | |
952 | * If size < SPA_MINBLOCKSIZE, then we will not allocate from | |
953 | * this metaslab again. In that case, it had better be empty, | |
954 | * or we would be leaving space on the table. | |
955 | */ | |
e51be066 | 956 | ASSERT(size >= SPA_MINBLOCKSIZE || msp->ms_map->sm_space == 0); |
34dc7c2f BB |
957 | metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size)); |
958 | ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0); | |
959 | } | |
960 | ||
e51be066 GW |
961 | /* |
962 | * Determine if the in-core space map representation can be condensed on-disk. | |
963 | * We would like to use the following criteria to make our decision: | |
964 | * | |
965 | * 1. The size of the space map object should not dramatically increase as a | |
966 | * result of writing out our in-core free map. | |
967 | * | |
968 | * 2. The minimal on-disk space map representation is zfs_condense_pct/100 | |
969 | * times the size than the in-core representation (i.e. zfs_condense_pct = 110 | |
970 | * and in-core = 1MB, minimal = 1.1.MB). | |
971 | * | |
972 | * Checking the first condition is tricky since we don't want to walk | |
973 | * the entire AVL tree calculating the estimated on-disk size. Instead we | |
974 | * use the size-ordered AVL tree in the space map and calculate the | |
975 | * size required for the largest segment in our in-core free map. If the | |
976 | * size required to represent that segment on disk is larger than the space | |
977 | * map object then we avoid condensing this map. | |
978 | * | |
979 | * To determine the second criterion we use a best-case estimate and assume | |
980 | * each segment can be represented on-disk as a single 64-bit entry. We refer | |
981 | * to this best-case estimate as the space map's minimal form. | |
982 | */ | |
983 | static boolean_t | |
984 | metaslab_should_condense(metaslab_t *msp) | |
985 | { | |
986 | space_map_t *sm = msp->ms_map; | |
987 | space_map_obj_t *smo = &msp->ms_smo_syncing; | |
988 | space_seg_t *ss; | |
989 | uint64_t size, entries, segsz; | |
990 | ||
991 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
992 | ASSERT(sm->sm_loaded); | |
993 | ||
994 | /* | |
995 | * Use the sm_pp_root AVL tree, which is ordered by size, to obtain | |
996 | * the largest segment in the in-core free map. If the tree is | |
997 | * empty then we should condense the map. | |
998 | */ | |
999 | ss = avl_last(sm->sm_pp_root); | |
1000 | if (ss == NULL) | |
1001 | return (B_TRUE); | |
1002 | ||
1003 | /* | |
1004 | * Calculate the number of 64-bit entries this segment would | |
1005 | * require when written to disk. If this single segment would be | |
1006 | * larger on-disk than the entire current on-disk structure, then | |
1007 | * clearly condensing will increase the on-disk structure size. | |
1008 | */ | |
1009 | size = (ss->ss_end - ss->ss_start) >> sm->sm_shift; | |
1010 | entries = size / (MIN(size, SM_RUN_MAX)); | |
1011 | segsz = entries * sizeof (uint64_t); | |
1012 | ||
1013 | return (segsz <= smo->smo_objsize && | |
1014 | smo->smo_objsize >= (zfs_condense_pct * | |
1015 | sizeof (uint64_t) * avl_numnodes(&sm->sm_root)) / 100); | |
1016 | } | |
1017 | ||
1018 | /* | |
1019 | * Condense the on-disk space map representation to its minimized form. | |
1020 | * The minimized form consists of a small number of allocations followed by | |
1021 | * the in-core free map. | |
1022 | */ | |
1023 | static void | |
1024 | metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx) | |
1025 | { | |
1026 | spa_t *spa = msp->ms_group->mg_vd->vdev_spa; | |
1027 | space_map_t *freemap = msp->ms_freemap[txg & TXG_MASK]; | |
1028 | space_map_t condense_map; | |
1029 | space_map_t *sm = msp->ms_map; | |
1030 | objset_t *mos = spa_meta_objset(spa); | |
1031 | space_map_obj_t *smo = &msp->ms_smo_syncing; | |
1032 | int t; | |
1033 | ||
1034 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
1035 | ASSERT3U(spa_sync_pass(spa), ==, 1); | |
1036 | ASSERT(sm->sm_loaded); | |
1037 | ||
1038 | spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, " | |
1039 | "smo size %llu, segments %lu", txg, | |
1040 | (msp->ms_map->sm_start / msp->ms_map->sm_size), msp, | |
1041 | smo->smo_objsize, avl_numnodes(&sm->sm_root)); | |
1042 | ||
1043 | /* | |
1044 | * Create an map that is a 100% allocated map. We remove segments | |
1045 | * that have been freed in this txg, any deferred frees that exist, | |
1046 | * and any allocation in the future. Removing segments should be | |
1047 | * a relatively inexpensive operation since we expect these maps to | |
1048 | * a small number of nodes. | |
1049 | */ | |
1050 | space_map_create(&condense_map, sm->sm_start, sm->sm_size, | |
1051 | sm->sm_shift, sm->sm_lock); | |
1052 | space_map_add(&condense_map, condense_map.sm_start, | |
1053 | condense_map.sm_size); | |
1054 | ||
1055 | /* | |
1056 | * Remove what's been freed in this txg from the condense_map. | |
1057 | * Since we're in sync_pass 1, we know that all the frees from | |
1058 | * this txg are in the freemap. | |
1059 | */ | |
1060 | space_map_walk(freemap, space_map_remove, &condense_map); | |
1061 | ||
1062 | for (t = 0; t < TXG_DEFER_SIZE; t++) | |
1063 | space_map_walk(msp->ms_defermap[t], | |
1064 | space_map_remove, &condense_map); | |
1065 | ||
1066 | for (t = 1; t < TXG_CONCURRENT_STATES; t++) | |
1067 | space_map_walk(msp->ms_allocmap[(txg + t) & TXG_MASK], | |
1068 | space_map_remove, &condense_map); | |
1069 | ||
1070 | /* | |
1071 | * We're about to drop the metaslab's lock thus allowing | |
1072 | * other consumers to change it's content. Set the | |
1073 | * space_map's sm_condensing flag to ensure that | |
1074 | * allocations on this metaslab do not occur while we're | |
1075 | * in the middle of committing it to disk. This is only critical | |
1076 | * for the ms_map as all other space_maps use per txg | |
1077 | * views of their content. | |
1078 | */ | |
1079 | sm->sm_condensing = B_TRUE; | |
1080 | ||
1081 | mutex_exit(&msp->ms_lock); | |
1082 | space_map_truncate(smo, mos, tx); | |
1083 | mutex_enter(&msp->ms_lock); | |
1084 | ||
1085 | /* | |
1086 | * While we would ideally like to create a space_map representation | |
1087 | * that consists only of allocation records, doing so can be | |
1088 | * prohibitively expensive because the in-core free map can be | |
1089 | * large, and therefore computationally expensive to subtract | |
1090 | * from the condense_map. Instead we sync out two maps, a cheap | |
1091 | * allocation only map followed by the in-core free map. While not | |
1092 | * optimal, this is typically close to optimal, and much cheaper to | |
1093 | * compute. | |
1094 | */ | |
1095 | space_map_sync(&condense_map, SM_ALLOC, smo, mos, tx); | |
1096 | space_map_vacate(&condense_map, NULL, NULL); | |
1097 | space_map_destroy(&condense_map); | |
1098 | ||
1099 | space_map_sync(sm, SM_FREE, smo, mos, tx); | |
1100 | sm->sm_condensing = B_FALSE; | |
1101 | ||
1102 | spa_dbgmsg(spa, "condensed: txg %llu, msp[%llu] %p, " | |
1103 | "smo size %llu", txg, | |
1104 | (msp->ms_map->sm_start / msp->ms_map->sm_size), msp, | |
1105 | smo->smo_objsize); | |
1106 | } | |
1107 | ||
34dc7c2f BB |
1108 | /* |
1109 | * Write a metaslab to disk in the context of the specified transaction group. | |
1110 | */ | |
1111 | void | |
1112 | metaslab_sync(metaslab_t *msp, uint64_t txg) | |
1113 | { | |
1114 | vdev_t *vd = msp->ms_group->mg_vd; | |
1115 | spa_t *spa = vd->vdev_spa; | |
428870ff | 1116 | objset_t *mos = spa_meta_objset(spa); |
e51be066 GW |
1117 | space_map_t *allocmap = msp->ms_allocmap[txg & TXG_MASK]; |
1118 | space_map_t **freemap = &msp->ms_freemap[txg & TXG_MASK]; | |
1119 | space_map_t **freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK]; | |
1120 | space_map_t *sm = msp->ms_map; | |
34dc7c2f BB |
1121 | space_map_obj_t *smo = &msp->ms_smo_syncing; |
1122 | dmu_buf_t *db; | |
1123 | dmu_tx_t *tx; | |
34dc7c2f | 1124 | |
428870ff BB |
1125 | ASSERT(!vd->vdev_ishole); |
1126 | ||
e51be066 GW |
1127 | /* |
1128 | * This metaslab has just been added so there's no work to do now. | |
1129 | */ | |
1130 | if (*freemap == NULL) { | |
1131 | ASSERT3P(allocmap, ==, NULL); | |
1132 | return; | |
1133 | } | |
1134 | ||
1135 | ASSERT3P(allocmap, !=, NULL); | |
1136 | ASSERT3P(*freemap, !=, NULL); | |
1137 | ASSERT3P(*freed_map, !=, NULL); | |
1138 | ||
1139 | if (allocmap->sm_space == 0 && (*freemap)->sm_space == 0) | |
428870ff | 1140 | return; |
34dc7c2f BB |
1141 | |
1142 | /* | |
1143 | * The only state that can actually be changing concurrently with | |
1144 | * metaslab_sync() is the metaslab's ms_map. No other thread can | |
1145 | * be modifying this txg's allocmap, freemap, freed_map, or smo. | |
1146 | * Therefore, we only hold ms_lock to satify space_map ASSERTs. | |
1147 | * We drop it whenever we call into the DMU, because the DMU | |
1148 | * can call down to us (e.g. via zio_free()) at any time. | |
1149 | */ | |
428870ff BB |
1150 | |
1151 | tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg); | |
34dc7c2f BB |
1152 | |
1153 | if (smo->smo_object == 0) { | |
1154 | ASSERT(smo->smo_objsize == 0); | |
1155 | ASSERT(smo->smo_alloc == 0); | |
34dc7c2f BB |
1156 | smo->smo_object = dmu_object_alloc(mos, |
1157 | DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT, | |
1158 | DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx); | |
1159 | ASSERT(smo->smo_object != 0); | |
1160 | dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) * | |
1161 | (sm->sm_start >> vd->vdev_ms_shift), | |
1162 | sizeof (uint64_t), &smo->smo_object, tx); | |
34dc7c2f BB |
1163 | } |
1164 | ||
428870ff BB |
1165 | mutex_enter(&msp->ms_lock); |
1166 | ||
e51be066 GW |
1167 | if (sm->sm_loaded && spa_sync_pass(spa) == 1 && |
1168 | metaslab_should_condense(msp)) { | |
1169 | metaslab_condense(msp, txg, tx); | |
1170 | } else { | |
1171 | space_map_sync(allocmap, SM_ALLOC, smo, mos, tx); | |
1172 | space_map_sync(*freemap, SM_FREE, smo, mos, tx); | |
1173 | } | |
428870ff | 1174 | |
e51be066 | 1175 | space_map_vacate(allocmap, NULL, NULL); |
34dc7c2f | 1176 | |
e51be066 GW |
1177 | /* |
1178 | * For sync pass 1, we avoid walking the entire space map and | |
1179 | * instead will just swap the pointers for freemap and | |
1180 | * freed_map. We can safely do this since the freed_map is | |
1181 | * guaranteed to be empty on the initial pass. | |
1182 | */ | |
1183 | if (spa_sync_pass(spa) == 1) { | |
1184 | ASSERT0((*freed_map)->sm_space); | |
1185 | ASSERT0(avl_numnodes(&(*freed_map)->sm_root)); | |
1186 | space_map_swap(freemap, freed_map); | |
1187 | } else { | |
1188 | space_map_vacate(*freemap, space_map_add, *freed_map); | |
34dc7c2f BB |
1189 | } |
1190 | ||
e51be066 GW |
1191 | ASSERT0(msp->ms_allocmap[txg & TXG_MASK]->sm_space); |
1192 | ASSERT0(msp->ms_freemap[txg & TXG_MASK]->sm_space); | |
34dc7c2f BB |
1193 | |
1194 | mutex_exit(&msp->ms_lock); | |
1195 | ||
e51be066 | 1196 | VERIFY0(dmu_bonus_hold(mos, smo->smo_object, FTAG, &db)); |
34dc7c2f BB |
1197 | dmu_buf_will_dirty(db, tx); |
1198 | ASSERT3U(db->db_size, >=, sizeof (*smo)); | |
1199 | bcopy(smo, db->db_data, sizeof (*smo)); | |
1200 | dmu_buf_rele(db, FTAG); | |
1201 | ||
1202 | dmu_tx_commit(tx); | |
1203 | } | |
1204 | ||
1205 | /* | |
1206 | * Called after a transaction group has completely synced to mark | |
1207 | * all of the metaslab's free space as usable. | |
1208 | */ | |
1209 | void | |
1210 | metaslab_sync_done(metaslab_t *msp, uint64_t txg) | |
1211 | { | |
1212 | space_map_obj_t *smo = &msp->ms_smo; | |
1213 | space_map_obj_t *smosync = &msp->ms_smo_syncing; | |
e51be066 GW |
1214 | space_map_t *sm = msp->ms_map; |
1215 | space_map_t *freed_map = msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK]; | |
1216 | space_map_t *defer_map = msp->ms_defermap[txg % TXG_DEFER_SIZE]; | |
34dc7c2f BB |
1217 | metaslab_group_t *mg = msp->ms_group; |
1218 | vdev_t *vd = mg->mg_vd; | |
428870ff | 1219 | int64_t alloc_delta, defer_delta; |
d6320ddb | 1220 | int t; |
428870ff BB |
1221 | |
1222 | ASSERT(!vd->vdev_ishole); | |
34dc7c2f BB |
1223 | |
1224 | mutex_enter(&msp->ms_lock); | |
1225 | ||
1226 | /* | |
1227 | * If this metaslab is just becoming available, initialize its | |
e51be066 | 1228 | * allocmaps, freemaps, and defermap and add its capacity to the vdev. |
34dc7c2f | 1229 | */ |
e51be066 GW |
1230 | if (freed_map == NULL) { |
1231 | ASSERT(defer_map == NULL); | |
d6320ddb | 1232 | for (t = 0; t < TXG_SIZE; t++) { |
e51be066 GW |
1233 | msp->ms_allocmap[t] = kmem_zalloc(sizeof (space_map_t), |
1234 | KM_PUSHPAGE); | |
1235 | space_map_create(msp->ms_allocmap[t], sm->sm_start, | |
34dc7c2f | 1236 | sm->sm_size, sm->sm_shift, sm->sm_lock); |
e51be066 GW |
1237 | msp->ms_freemap[t] = kmem_zalloc(sizeof (space_map_t), |
1238 | KM_PUSHPAGE); | |
1239 | space_map_create(msp->ms_freemap[t], sm->sm_start, | |
34dc7c2f BB |
1240 | sm->sm_size, sm->sm_shift, sm->sm_lock); |
1241 | } | |
428870ff | 1242 | |
e51be066 GW |
1243 | for (t = 0; t < TXG_DEFER_SIZE; t++) { |
1244 | msp->ms_defermap[t] = kmem_zalloc(sizeof (space_map_t), | |
1245 | KM_PUSHPAGE); | |
1246 | space_map_create(msp->ms_defermap[t], sm->sm_start, | |
428870ff | 1247 | sm->sm_size, sm->sm_shift, sm->sm_lock); |
e51be066 GW |
1248 | } |
1249 | ||
1250 | freed_map = msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK]; | |
1251 | defer_map = msp->ms_defermap[txg % TXG_DEFER_SIZE]; | |
428870ff BB |
1252 | |
1253 | vdev_space_update(vd, 0, 0, sm->sm_size); | |
34dc7c2f BB |
1254 | } |
1255 | ||
428870ff BB |
1256 | alloc_delta = smosync->smo_alloc - smo->smo_alloc; |
1257 | defer_delta = freed_map->sm_space - defer_map->sm_space; | |
1258 | ||
1259 | vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0); | |
34dc7c2f | 1260 | |
e51be066 GW |
1261 | ASSERT(msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0); |
1262 | ASSERT(msp->ms_freemap[txg & TXG_MASK]->sm_space == 0); | |
34dc7c2f BB |
1263 | |
1264 | /* | |
1265 | * If there's a space_map_load() in progress, wait for it to complete | |
1266 | * so that we have a consistent view of the in-core space map. | |
428870ff BB |
1267 | * Then, add defer_map (oldest deferred frees) to this map and |
1268 | * transfer freed_map (this txg's frees) to defer_map. | |
34dc7c2f BB |
1269 | */ |
1270 | space_map_load_wait(sm); | |
428870ff BB |
1271 | space_map_vacate(defer_map, sm->sm_loaded ? space_map_free : NULL, sm); |
1272 | space_map_vacate(freed_map, space_map_add, defer_map); | |
34dc7c2f BB |
1273 | |
1274 | *smo = *smosync; | |
1275 | ||
428870ff BB |
1276 | msp->ms_deferspace += defer_delta; |
1277 | ASSERT3S(msp->ms_deferspace, >=, 0); | |
1278 | ASSERT3S(msp->ms_deferspace, <=, sm->sm_size); | |
1279 | if (msp->ms_deferspace != 0) { | |
1280 | /* | |
1281 | * Keep syncing this metaslab until all deferred frees | |
1282 | * are back in circulation. | |
1283 | */ | |
1284 | vdev_dirty(vd, VDD_METASLAB, msp, txg + 1); | |
1285 | } | |
1286 | ||
34dc7c2f BB |
1287 | /* |
1288 | * If the map is loaded but no longer active, evict it as soon as all | |
1289 | * future allocations have synced. (If we unloaded it now and then | |
1290 | * loaded a moment later, the map wouldn't reflect those allocations.) | |
1291 | */ | |
1292 | if (sm->sm_loaded && (msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) { | |
1293 | int evictable = 1; | |
1294 | ||
d6320ddb | 1295 | for (t = 1; t < TXG_CONCURRENT_STATES; t++) |
e51be066 | 1296 | if (msp->ms_allocmap[(txg + t) & TXG_MASK]->sm_space) |
34dc7c2f BB |
1297 | evictable = 0; |
1298 | ||
428870ff | 1299 | if (evictable && !metaslab_debug) |
34dc7c2f BB |
1300 | space_map_unload(sm); |
1301 | } | |
1302 | ||
1303 | metaslab_group_sort(mg, msp, metaslab_weight(msp)); | |
1304 | ||
1305 | mutex_exit(&msp->ms_lock); | |
1306 | } | |
1307 | ||
428870ff BB |
1308 | void |
1309 | metaslab_sync_reassess(metaslab_group_t *mg) | |
1310 | { | |
1311 | vdev_t *vd = mg->mg_vd; | |
6d974228 | 1312 | int64_t failures = mg->mg_alloc_failures; |
d6320ddb | 1313 | int m; |
428870ff BB |
1314 | |
1315 | /* | |
1316 | * Re-evaluate all metaslabs which have lower offsets than the | |
1317 | * bonus area. | |
1318 | */ | |
d6320ddb | 1319 | for (m = 0; m < vd->vdev_ms_count; m++) { |
428870ff BB |
1320 | metaslab_t *msp = vd->vdev_ms[m]; |
1321 | ||
e51be066 | 1322 | if (msp->ms_map->sm_start > mg->mg_bonus_area) |
428870ff BB |
1323 | break; |
1324 | ||
1325 | mutex_enter(&msp->ms_lock); | |
1326 | metaslab_group_sort(mg, msp, metaslab_weight(msp)); | |
1327 | mutex_exit(&msp->ms_lock); | |
1328 | } | |
1329 | ||
6d974228 GW |
1330 | atomic_add_64(&mg->mg_alloc_failures, -failures); |
1331 | ||
428870ff BB |
1332 | /* |
1333 | * Prefetch the next potential metaslabs | |
1334 | */ | |
1335 | metaslab_prefetch(mg); | |
1336 | } | |
1337 | ||
34dc7c2f BB |
1338 | static uint64_t |
1339 | metaslab_distance(metaslab_t *msp, dva_t *dva) | |
1340 | { | |
1341 | uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift; | |
1342 | uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift; | |
e51be066 | 1343 | uint64_t start = msp->ms_map->sm_start >> ms_shift; |
34dc7c2f BB |
1344 | |
1345 | if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva)) | |
1346 | return (1ULL << 63); | |
1347 | ||
1348 | if (offset < start) | |
1349 | return ((start - offset) << ms_shift); | |
1350 | if (offset > start) | |
1351 | return ((offset - start) << ms_shift); | |
1352 | return (0); | |
1353 | } | |
1354 | ||
1355 | static uint64_t | |
6d974228 GW |
1356 | metaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize, |
1357 | uint64_t txg, uint64_t min_distance, dva_t *dva, int d, int flags) | |
34dc7c2f | 1358 | { |
6d974228 | 1359 | spa_t *spa = mg->mg_vd->vdev_spa; |
34dc7c2f BB |
1360 | metaslab_t *msp = NULL; |
1361 | uint64_t offset = -1ULL; | |
1362 | avl_tree_t *t = &mg->mg_metaslab_tree; | |
1363 | uint64_t activation_weight; | |
1364 | uint64_t target_distance; | |
1365 | int i; | |
1366 | ||
1367 | activation_weight = METASLAB_WEIGHT_PRIMARY; | |
9babb374 BB |
1368 | for (i = 0; i < d; i++) { |
1369 | if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) { | |
34dc7c2f | 1370 | activation_weight = METASLAB_WEIGHT_SECONDARY; |
9babb374 BB |
1371 | break; |
1372 | } | |
1373 | } | |
34dc7c2f BB |
1374 | |
1375 | for (;;) { | |
9babb374 BB |
1376 | boolean_t was_active; |
1377 | ||
34dc7c2f BB |
1378 | mutex_enter(&mg->mg_lock); |
1379 | for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) { | |
6d974228 GW |
1380 | if (msp->ms_weight < asize) { |
1381 | spa_dbgmsg(spa, "%s: failed to meet weight " | |
1382 | "requirement: vdev %llu, txg %llu, mg %p, " | |
1383 | "msp %p, psize %llu, asize %llu, " | |
1384 | "failures %llu, weight %llu", | |
1385 | spa_name(spa), mg->mg_vd->vdev_id, txg, | |
1386 | mg, msp, psize, asize, | |
1387 | mg->mg_alloc_failures, msp->ms_weight); | |
34dc7c2f BB |
1388 | mutex_exit(&mg->mg_lock); |
1389 | return (-1ULL); | |
1390 | } | |
7a614407 GW |
1391 | |
1392 | /* | |
1393 | * If the selected metaslab is condensing, skip it. | |
1394 | */ | |
1395 | if (msp->ms_map->sm_condensing) | |
1396 | continue; | |
1397 | ||
9babb374 | 1398 | was_active = msp->ms_weight & METASLAB_ACTIVE_MASK; |
34dc7c2f BB |
1399 | if (activation_weight == METASLAB_WEIGHT_PRIMARY) |
1400 | break; | |
1401 | ||
1402 | target_distance = min_distance + | |
1403 | (msp->ms_smo.smo_alloc ? 0 : min_distance >> 1); | |
1404 | ||
1405 | for (i = 0; i < d; i++) | |
1406 | if (metaslab_distance(msp, &dva[i]) < | |
1407 | target_distance) | |
1408 | break; | |
1409 | if (i == d) | |
1410 | break; | |
1411 | } | |
1412 | mutex_exit(&mg->mg_lock); | |
1413 | if (msp == NULL) | |
1414 | return (-1ULL); | |
1415 | ||
6d974228 GW |
1416 | /* |
1417 | * If we've already reached the allowable number of failed | |
1418 | * allocation attempts on this metaslab group then we | |
1419 | * consider skipping it. We skip it only if we're allowed | |
1420 | * to "fast" gang, the physical size is larger than | |
1421 | * a gang block, and we're attempting to allocate from | |
1422 | * the primary metaslab. | |
1423 | */ | |
1424 | if (mg->mg_alloc_failures > zfs_mg_alloc_failures && | |
1425 | CAN_FASTGANG(flags) && psize > SPA_GANGBLOCKSIZE && | |
1426 | activation_weight == METASLAB_WEIGHT_PRIMARY) { | |
1427 | spa_dbgmsg(spa, "%s: skipping metaslab group: " | |
1428 | "vdev %llu, txg %llu, mg %p, psize %llu, " | |
1429 | "asize %llu, failures %llu", spa_name(spa), | |
1430 | mg->mg_vd->vdev_id, txg, mg, psize, asize, | |
1431 | mg->mg_alloc_failures); | |
1432 | return (-1ULL); | |
1433 | } | |
1434 | ||
34dc7c2f BB |
1435 | mutex_enter(&msp->ms_lock); |
1436 | ||
1437 | /* | |
1438 | * Ensure that the metaslab we have selected is still | |
1439 | * capable of handling our request. It's possible that | |
1440 | * another thread may have changed the weight while we | |
1441 | * were blocked on the metaslab lock. | |
1442 | */ | |
6d974228 | 1443 | if (msp->ms_weight < asize || (was_active && |
9babb374 BB |
1444 | !(msp->ms_weight & METASLAB_ACTIVE_MASK) && |
1445 | activation_weight == METASLAB_WEIGHT_PRIMARY)) { | |
34dc7c2f BB |
1446 | mutex_exit(&msp->ms_lock); |
1447 | continue; | |
1448 | } | |
1449 | ||
1450 | if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) && | |
1451 | activation_weight == METASLAB_WEIGHT_PRIMARY) { | |
1452 | metaslab_passivate(msp, | |
1453 | msp->ms_weight & ~METASLAB_ACTIVE_MASK); | |
1454 | mutex_exit(&msp->ms_lock); | |
1455 | continue; | |
1456 | } | |
1457 | ||
6d974228 | 1458 | if (metaslab_activate(msp, activation_weight) != 0) { |
34dc7c2f BB |
1459 | mutex_exit(&msp->ms_lock); |
1460 | continue; | |
1461 | } | |
1462 | ||
7a614407 GW |
1463 | /* |
1464 | * If this metaslab is currently condensing then pick again as | |
1465 | * we can't manipulate this metaslab until it's committed | |
1466 | * to disk. | |
1467 | */ | |
1468 | if (msp->ms_map->sm_condensing) { | |
1469 | mutex_exit(&msp->ms_lock); | |
1470 | continue; | |
1471 | } | |
1472 | ||
e51be066 | 1473 | if ((offset = space_map_alloc(msp->ms_map, asize)) != -1ULL) |
34dc7c2f BB |
1474 | break; |
1475 | ||
6d974228 GW |
1476 | atomic_inc_64(&mg->mg_alloc_failures); |
1477 | ||
e51be066 | 1478 | metaslab_passivate(msp, space_map_maxsize(msp->ms_map)); |
34dc7c2f BB |
1479 | |
1480 | mutex_exit(&msp->ms_lock); | |
1481 | } | |
1482 | ||
e51be066 | 1483 | if (msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0) |
34dc7c2f BB |
1484 | vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg); |
1485 | ||
e51be066 | 1486 | space_map_add(msp->ms_allocmap[txg & TXG_MASK], offset, asize); |
34dc7c2f BB |
1487 | |
1488 | mutex_exit(&msp->ms_lock); | |
1489 | ||
1490 | return (offset); | |
1491 | } | |
1492 | ||
1493 | /* | |
1494 | * Allocate a block for the specified i/o. | |
1495 | */ | |
1496 | static int | |
1497 | metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize, | |
b128c09f | 1498 | dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags) |
34dc7c2f | 1499 | { |
920dd524 | 1500 | metaslab_group_t *mg, *fast_mg, *rotor; |
34dc7c2f BB |
1501 | vdev_t *vd; |
1502 | int dshift = 3; | |
1503 | int all_zero; | |
fb5f0bc8 BB |
1504 | int zio_lock = B_FALSE; |
1505 | boolean_t allocatable; | |
34dc7c2f BB |
1506 | uint64_t offset = -1ULL; |
1507 | uint64_t asize; | |
1508 | uint64_t distance; | |
1509 | ||
1510 | ASSERT(!DVA_IS_VALID(&dva[d])); | |
1511 | ||
1512 | /* | |
1513 | * For testing, make some blocks above a certain size be gang blocks. | |
1514 | */ | |
428870ff | 1515 | if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0) |
34dc7c2f BB |
1516 | return (ENOSPC); |
1517 | ||
920dd524 ED |
1518 | if (flags & METASLAB_FASTWRITE) |
1519 | mutex_enter(&mc->mc_fastwrite_lock); | |
1520 | ||
34dc7c2f BB |
1521 | /* |
1522 | * Start at the rotor and loop through all mgs until we find something. | |
428870ff | 1523 | * Note that there's no locking on mc_rotor or mc_aliquot because |
34dc7c2f BB |
1524 | * nothing actually breaks if we miss a few updates -- we just won't |
1525 | * allocate quite as evenly. It all balances out over time. | |
1526 | * | |
1527 | * If we are doing ditto or log blocks, try to spread them across | |
1528 | * consecutive vdevs. If we're forced to reuse a vdev before we've | |
1529 | * allocated all of our ditto blocks, then try and spread them out on | |
1530 | * that vdev as much as possible. If it turns out to not be possible, | |
1531 | * gradually lower our standards until anything becomes acceptable. | |
1532 | * Also, allocating on consecutive vdevs (as opposed to random vdevs) | |
1533 | * gives us hope of containing our fault domains to something we're | |
1534 | * able to reason about. Otherwise, any two top-level vdev failures | |
1535 | * will guarantee the loss of data. With consecutive allocation, | |
1536 | * only two adjacent top-level vdev failures will result in data loss. | |
1537 | * | |
1538 | * If we are doing gang blocks (hintdva is non-NULL), try to keep | |
1539 | * ourselves on the same vdev as our gang block header. That | |
1540 | * way, we can hope for locality in vdev_cache, plus it makes our | |
1541 | * fault domains something tractable. | |
1542 | */ | |
1543 | if (hintdva) { | |
1544 | vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d])); | |
428870ff BB |
1545 | |
1546 | /* | |
1547 | * It's possible the vdev we're using as the hint no | |
1548 | * longer exists (i.e. removed). Consult the rotor when | |
1549 | * all else fails. | |
1550 | */ | |
1551 | if (vd != NULL) { | |
34dc7c2f | 1552 | mg = vd->vdev_mg; |
428870ff BB |
1553 | |
1554 | if (flags & METASLAB_HINTBP_AVOID && | |
1555 | mg->mg_next != NULL) | |
1556 | mg = mg->mg_next; | |
1557 | } else { | |
1558 | mg = mc->mc_rotor; | |
1559 | } | |
34dc7c2f BB |
1560 | } else if (d != 0) { |
1561 | vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1])); | |
1562 | mg = vd->vdev_mg->mg_next; | |
920dd524 ED |
1563 | } else if (flags & METASLAB_FASTWRITE) { |
1564 | mg = fast_mg = mc->mc_rotor; | |
1565 | ||
1566 | do { | |
1567 | if (fast_mg->mg_vd->vdev_pending_fastwrite < | |
1568 | mg->mg_vd->vdev_pending_fastwrite) | |
1569 | mg = fast_mg; | |
1570 | } while ((fast_mg = fast_mg->mg_next) != mc->mc_rotor); | |
1571 | ||
34dc7c2f BB |
1572 | } else { |
1573 | mg = mc->mc_rotor; | |
1574 | } | |
1575 | ||
1576 | /* | |
428870ff BB |
1577 | * If the hint put us into the wrong metaslab class, or into a |
1578 | * metaslab group that has been passivated, just follow the rotor. | |
34dc7c2f | 1579 | */ |
428870ff | 1580 | if (mg->mg_class != mc || mg->mg_activation_count <= 0) |
34dc7c2f BB |
1581 | mg = mc->mc_rotor; |
1582 | ||
1583 | rotor = mg; | |
1584 | top: | |
1585 | all_zero = B_TRUE; | |
1586 | do { | |
428870ff BB |
1587 | ASSERT(mg->mg_activation_count == 1); |
1588 | ||
34dc7c2f | 1589 | vd = mg->mg_vd; |
fb5f0bc8 | 1590 | |
34dc7c2f | 1591 | /* |
b128c09f | 1592 | * Don't allocate from faulted devices. |
34dc7c2f | 1593 | */ |
fb5f0bc8 BB |
1594 | if (zio_lock) { |
1595 | spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER); | |
1596 | allocatable = vdev_allocatable(vd); | |
1597 | spa_config_exit(spa, SCL_ZIO, FTAG); | |
1598 | } else { | |
1599 | allocatable = vdev_allocatable(vd); | |
1600 | } | |
1601 | if (!allocatable) | |
34dc7c2f | 1602 | goto next; |
fb5f0bc8 | 1603 | |
34dc7c2f BB |
1604 | /* |
1605 | * Avoid writing single-copy data to a failing vdev | |
1606 | */ | |
1607 | if ((vd->vdev_stat.vs_write_errors > 0 || | |
1608 | vd->vdev_state < VDEV_STATE_HEALTHY) && | |
1609 | d == 0 && dshift == 3) { | |
1610 | all_zero = B_FALSE; | |
1611 | goto next; | |
1612 | } | |
1613 | ||
1614 | ASSERT(mg->mg_class == mc); | |
1615 | ||
1616 | distance = vd->vdev_asize >> dshift; | |
1617 | if (distance <= (1ULL << vd->vdev_ms_shift)) | |
1618 | distance = 0; | |
1619 | else | |
1620 | all_zero = B_FALSE; | |
1621 | ||
1622 | asize = vdev_psize_to_asize(vd, psize); | |
1623 | ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0); | |
1624 | ||
6d974228 GW |
1625 | offset = metaslab_group_alloc(mg, psize, asize, txg, distance, |
1626 | dva, d, flags); | |
34dc7c2f BB |
1627 | if (offset != -1ULL) { |
1628 | /* | |
1629 | * If we've just selected this metaslab group, | |
1630 | * figure out whether the corresponding vdev is | |
1631 | * over- or under-used relative to the pool, | |
1632 | * and set an allocation bias to even it out. | |
1633 | */ | |
428870ff | 1634 | if (mc->mc_aliquot == 0) { |
34dc7c2f | 1635 | vdev_stat_t *vs = &vd->vdev_stat; |
428870ff | 1636 | int64_t vu, cu; |
34dc7c2f | 1637 | |
6d974228 GW |
1638 | vu = (vs->vs_alloc * 100) / (vs->vs_space + 1); |
1639 | cu = (mc->mc_alloc * 100) / (mc->mc_space + 1); | |
34dc7c2f BB |
1640 | |
1641 | /* | |
6d974228 GW |
1642 | * Calculate how much more or less we should |
1643 | * try to allocate from this device during | |
1644 | * this iteration around the rotor. | |
1645 | * For example, if a device is 80% full | |
1646 | * and the pool is 20% full then we should | |
1647 | * reduce allocations by 60% on this device. | |
1648 | * | |
1649 | * mg_bias = (20 - 80) * 512K / 100 = -307K | |
1650 | * | |
1651 | * This reduces allocations by 307K for this | |
1652 | * iteration. | |
34dc7c2f | 1653 | */ |
428870ff | 1654 | mg->mg_bias = ((cu - vu) * |
6d974228 | 1655 | (int64_t)mg->mg_aliquot) / 100; |
34dc7c2f BB |
1656 | } |
1657 | ||
920dd524 ED |
1658 | if ((flags & METASLAB_FASTWRITE) || |
1659 | atomic_add_64_nv(&mc->mc_aliquot, asize) >= | |
34dc7c2f BB |
1660 | mg->mg_aliquot + mg->mg_bias) { |
1661 | mc->mc_rotor = mg->mg_next; | |
428870ff | 1662 | mc->mc_aliquot = 0; |
34dc7c2f BB |
1663 | } |
1664 | ||
1665 | DVA_SET_VDEV(&dva[d], vd->vdev_id); | |
1666 | DVA_SET_OFFSET(&dva[d], offset); | |
b128c09f | 1667 | DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER)); |
34dc7c2f BB |
1668 | DVA_SET_ASIZE(&dva[d], asize); |
1669 | ||
920dd524 ED |
1670 | if (flags & METASLAB_FASTWRITE) { |
1671 | atomic_add_64(&vd->vdev_pending_fastwrite, | |
1672 | psize); | |
1673 | mutex_exit(&mc->mc_fastwrite_lock); | |
1674 | } | |
1675 | ||
34dc7c2f BB |
1676 | return (0); |
1677 | } | |
1678 | next: | |
1679 | mc->mc_rotor = mg->mg_next; | |
428870ff | 1680 | mc->mc_aliquot = 0; |
34dc7c2f BB |
1681 | } while ((mg = mg->mg_next) != rotor); |
1682 | ||
1683 | if (!all_zero) { | |
1684 | dshift++; | |
1685 | ASSERT(dshift < 64); | |
1686 | goto top; | |
1687 | } | |
1688 | ||
9babb374 | 1689 | if (!allocatable && !zio_lock) { |
fb5f0bc8 BB |
1690 | dshift = 3; |
1691 | zio_lock = B_TRUE; | |
1692 | goto top; | |
1693 | } | |
1694 | ||
34dc7c2f BB |
1695 | bzero(&dva[d], sizeof (dva_t)); |
1696 | ||
920dd524 ED |
1697 | if (flags & METASLAB_FASTWRITE) |
1698 | mutex_exit(&mc->mc_fastwrite_lock); | |
34dc7c2f BB |
1699 | return (ENOSPC); |
1700 | } | |
1701 | ||
1702 | /* | |
1703 | * Free the block represented by DVA in the context of the specified | |
1704 | * transaction group. | |
1705 | */ | |
1706 | static void | |
1707 | metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now) | |
1708 | { | |
1709 | uint64_t vdev = DVA_GET_VDEV(dva); | |
1710 | uint64_t offset = DVA_GET_OFFSET(dva); | |
1711 | uint64_t size = DVA_GET_ASIZE(dva); | |
1712 | vdev_t *vd; | |
1713 | metaslab_t *msp; | |
1714 | ||
1715 | ASSERT(DVA_IS_VALID(dva)); | |
1716 | ||
1717 | if (txg > spa_freeze_txg(spa)) | |
1718 | return; | |
1719 | ||
1720 | if ((vd = vdev_lookup_top(spa, vdev)) == NULL || | |
1721 | (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) { | |
1722 | cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu", | |
1723 | (u_longlong_t)vdev, (u_longlong_t)offset); | |
1724 | ASSERT(0); | |
1725 | return; | |
1726 | } | |
1727 | ||
1728 | msp = vd->vdev_ms[offset >> vd->vdev_ms_shift]; | |
1729 | ||
1730 | if (DVA_GET_GANG(dva)) | |
1731 | size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); | |
1732 | ||
1733 | mutex_enter(&msp->ms_lock); | |
1734 | ||
1735 | if (now) { | |
e51be066 | 1736 | space_map_remove(msp->ms_allocmap[txg & TXG_MASK], |
34dc7c2f | 1737 | offset, size); |
e51be066 | 1738 | space_map_free(msp->ms_map, offset, size); |
34dc7c2f | 1739 | } else { |
e51be066 | 1740 | if (msp->ms_freemap[txg & TXG_MASK]->sm_space == 0) |
34dc7c2f | 1741 | vdev_dirty(vd, VDD_METASLAB, msp, txg); |
e51be066 | 1742 | space_map_add(msp->ms_freemap[txg & TXG_MASK], offset, size); |
34dc7c2f BB |
1743 | } |
1744 | ||
1745 | mutex_exit(&msp->ms_lock); | |
1746 | } | |
1747 | ||
1748 | /* | |
1749 | * Intent log support: upon opening the pool after a crash, notify the SPA | |
1750 | * of blocks that the intent log has allocated for immediate write, but | |
1751 | * which are still considered free by the SPA because the last transaction | |
1752 | * group didn't commit yet. | |
1753 | */ | |
1754 | static int | |
1755 | metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg) | |
1756 | { | |
1757 | uint64_t vdev = DVA_GET_VDEV(dva); | |
1758 | uint64_t offset = DVA_GET_OFFSET(dva); | |
1759 | uint64_t size = DVA_GET_ASIZE(dva); | |
1760 | vdev_t *vd; | |
1761 | metaslab_t *msp; | |
428870ff | 1762 | int error = 0; |
34dc7c2f BB |
1763 | |
1764 | ASSERT(DVA_IS_VALID(dva)); | |
1765 | ||
1766 | if ((vd = vdev_lookup_top(spa, vdev)) == NULL || | |
1767 | (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) | |
1768 | return (ENXIO); | |
1769 | ||
1770 | msp = vd->vdev_ms[offset >> vd->vdev_ms_shift]; | |
1771 | ||
1772 | if (DVA_GET_GANG(dva)) | |
1773 | size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); | |
1774 | ||
1775 | mutex_enter(&msp->ms_lock); | |
1776 | ||
e51be066 | 1777 | if ((txg != 0 && spa_writeable(spa)) || !msp->ms_map->sm_loaded) |
6d974228 | 1778 | error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY); |
428870ff | 1779 | |
e51be066 | 1780 | if (error == 0 && !space_map_contains(msp->ms_map, offset, size)) |
428870ff BB |
1781 | error = ENOENT; |
1782 | ||
b128c09f | 1783 | if (error || txg == 0) { /* txg == 0 indicates dry run */ |
34dc7c2f BB |
1784 | mutex_exit(&msp->ms_lock); |
1785 | return (error); | |
1786 | } | |
1787 | ||
e51be066 | 1788 | space_map_claim(msp->ms_map, offset, size); |
b128c09f | 1789 | |
fb5f0bc8 | 1790 | if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */ |
e51be066 | 1791 | if (msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0) |
b128c09f | 1792 | vdev_dirty(vd, VDD_METASLAB, msp, txg); |
e51be066 | 1793 | space_map_add(msp->ms_allocmap[txg & TXG_MASK], offset, size); |
b128c09f | 1794 | } |
34dc7c2f BB |
1795 | |
1796 | mutex_exit(&msp->ms_lock); | |
1797 | ||
1798 | return (0); | |
1799 | } | |
1800 | ||
1801 | int | |
1802 | metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp, | |
b128c09f | 1803 | int ndvas, uint64_t txg, blkptr_t *hintbp, int flags) |
34dc7c2f BB |
1804 | { |
1805 | dva_t *dva = bp->blk_dva; | |
1806 | dva_t *hintdva = hintbp->blk_dva; | |
d6320ddb | 1807 | int d, error = 0; |
34dc7c2f | 1808 | |
b128c09f | 1809 | ASSERT(bp->blk_birth == 0); |
428870ff | 1810 | ASSERT(BP_PHYSICAL_BIRTH(bp) == 0); |
b128c09f BB |
1811 | |
1812 | spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER); | |
1813 | ||
1814 | if (mc->mc_rotor == NULL) { /* no vdevs in this class */ | |
1815 | spa_config_exit(spa, SCL_ALLOC, FTAG); | |
34dc7c2f | 1816 | return (ENOSPC); |
b128c09f | 1817 | } |
34dc7c2f BB |
1818 | |
1819 | ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa)); | |
1820 | ASSERT(BP_GET_NDVAS(bp) == 0); | |
1821 | ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp)); | |
1822 | ||
d6320ddb | 1823 | for (d = 0; d < ndvas; d++) { |
34dc7c2f | 1824 | error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva, |
b128c09f | 1825 | txg, flags); |
34dc7c2f BB |
1826 | if (error) { |
1827 | for (d--; d >= 0; d--) { | |
1828 | metaslab_free_dva(spa, &dva[d], txg, B_TRUE); | |
1829 | bzero(&dva[d], sizeof (dva_t)); | |
1830 | } | |
b128c09f | 1831 | spa_config_exit(spa, SCL_ALLOC, FTAG); |
34dc7c2f BB |
1832 | return (error); |
1833 | } | |
1834 | } | |
1835 | ASSERT(error == 0); | |
1836 | ASSERT(BP_GET_NDVAS(bp) == ndvas); | |
1837 | ||
b128c09f BB |
1838 | spa_config_exit(spa, SCL_ALLOC, FTAG); |
1839 | ||
428870ff | 1840 | BP_SET_BIRTH(bp, txg, txg); |
b128c09f | 1841 | |
34dc7c2f BB |
1842 | return (0); |
1843 | } | |
1844 | ||
1845 | void | |
1846 | metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now) | |
1847 | { | |
1848 | const dva_t *dva = bp->blk_dva; | |
d6320ddb | 1849 | int d, ndvas = BP_GET_NDVAS(bp); |
34dc7c2f BB |
1850 | |
1851 | ASSERT(!BP_IS_HOLE(bp)); | |
428870ff | 1852 | ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa)); |
b128c09f BB |
1853 | |
1854 | spa_config_enter(spa, SCL_FREE, FTAG, RW_READER); | |
34dc7c2f | 1855 | |
d6320ddb | 1856 | for (d = 0; d < ndvas; d++) |
34dc7c2f | 1857 | metaslab_free_dva(spa, &dva[d], txg, now); |
b128c09f BB |
1858 | |
1859 | spa_config_exit(spa, SCL_FREE, FTAG); | |
34dc7c2f BB |
1860 | } |
1861 | ||
1862 | int | |
1863 | metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg) | |
1864 | { | |
1865 | const dva_t *dva = bp->blk_dva; | |
1866 | int ndvas = BP_GET_NDVAS(bp); | |
d6320ddb | 1867 | int d, error = 0; |
34dc7c2f BB |
1868 | |
1869 | ASSERT(!BP_IS_HOLE(bp)); | |
1870 | ||
b128c09f BB |
1871 | if (txg != 0) { |
1872 | /* | |
1873 | * First do a dry run to make sure all DVAs are claimable, | |
1874 | * so we don't have to unwind from partial failures below. | |
1875 | */ | |
1876 | if ((error = metaslab_claim(spa, bp, 0)) != 0) | |
1877 | return (error); | |
1878 | } | |
1879 | ||
1880 | spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER); | |
1881 | ||
d6320ddb | 1882 | for (d = 0; d < ndvas; d++) |
34dc7c2f | 1883 | if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0) |
b128c09f BB |
1884 | break; |
1885 | ||
1886 | spa_config_exit(spa, SCL_ALLOC, FTAG); | |
1887 | ||
1888 | ASSERT(error == 0 || txg == 0); | |
34dc7c2f | 1889 | |
b128c09f | 1890 | return (error); |
34dc7c2f | 1891 | } |
920dd524 ED |
1892 | |
1893 | void metaslab_fastwrite_mark(spa_t *spa, const blkptr_t *bp) | |
1894 | { | |
1895 | const dva_t *dva = bp->blk_dva; | |
1896 | int ndvas = BP_GET_NDVAS(bp); | |
1897 | uint64_t psize = BP_GET_PSIZE(bp); | |
1898 | int d; | |
1899 | vdev_t *vd; | |
1900 | ||
1901 | ASSERT(!BP_IS_HOLE(bp)); | |
1902 | ASSERT(psize > 0); | |
1903 | ||
1904 | spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER); | |
1905 | ||
1906 | for (d = 0; d < ndvas; d++) { | |
1907 | if ((vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d]))) == NULL) | |
1908 | continue; | |
1909 | atomic_add_64(&vd->vdev_pending_fastwrite, psize); | |
1910 | } | |
1911 | ||
1912 | spa_config_exit(spa, SCL_VDEV, FTAG); | |
1913 | } | |
1914 | ||
1915 | void metaslab_fastwrite_unmark(spa_t *spa, const blkptr_t *bp) | |
1916 | { | |
1917 | const dva_t *dva = bp->blk_dva; | |
1918 | int ndvas = BP_GET_NDVAS(bp); | |
1919 | uint64_t psize = BP_GET_PSIZE(bp); | |
1920 | int d; | |
1921 | vdev_t *vd; | |
1922 | ||
1923 | ASSERT(!BP_IS_HOLE(bp)); | |
1924 | ASSERT(psize > 0); | |
1925 | ||
1926 | spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER); | |
1927 | ||
1928 | for (d = 0; d < ndvas; d++) { | |
1929 | if ((vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d]))) == NULL) | |
1930 | continue; | |
1931 | ASSERT3U(vd->vdev_pending_fastwrite, >=, psize); | |
1932 | atomic_sub_64(&vd->vdev_pending_fastwrite, psize); | |
1933 | } | |
1934 | ||
1935 | spa_config_exit(spa, SCL_VDEV, FTAG); | |
1936 | } | |
30b92c1d | 1937 | |
13fe0198 MA |
1938 | static void |
1939 | checkmap(space_map_t *sm, uint64_t off, uint64_t size) | |
1940 | { | |
1941 | space_seg_t *ss; | |
1942 | avl_index_t where; | |
1943 | ||
1944 | mutex_enter(sm->sm_lock); | |
1945 | ss = space_map_find(sm, off, size, &where); | |
1946 | if (ss != NULL) | |
1947 | panic("freeing free block; ss=%p", (void *)ss); | |
1948 | mutex_exit(sm->sm_lock); | |
1949 | } | |
1950 | ||
1951 | void | |
1952 | metaslab_check_free(spa_t *spa, const blkptr_t *bp) | |
1953 | { | |
1954 | int i, j; | |
1955 | ||
1956 | if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0) | |
1957 | return; | |
1958 | ||
1959 | spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER); | |
1960 | for (i = 0; i < BP_GET_NDVAS(bp); i++) { | |
1961 | uint64_t vdid = DVA_GET_VDEV(&bp->blk_dva[i]); | |
1962 | vdev_t *vd = vdev_lookup_top(spa, vdid); | |
1963 | uint64_t off = DVA_GET_OFFSET(&bp->blk_dva[i]); | |
1964 | uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]); | |
1965 | metaslab_t *ms = vd->vdev_ms[off >> vd->vdev_ms_shift]; | |
1966 | ||
1967 | if (ms->ms_map->sm_loaded) | |
1968 | checkmap(ms->ms_map, off, size); | |
1969 | ||
1970 | for (j = 0; j < TXG_SIZE; j++) | |
1971 | checkmap(ms->ms_freemap[j], off, size); | |
1972 | for (j = 0; j < TXG_DEFER_SIZE; j++) | |
1973 | checkmap(ms->ms_defermap[j], off, size); | |
1974 | } | |
1975 | spa_config_exit(spa, SCL_VDEV, FTAG); | |
1976 | } | |
1977 | ||
30b92c1d BB |
1978 | #if defined(_KERNEL) && defined(HAVE_SPL) |
1979 | module_param(metaslab_debug, int, 0644); | |
1980 | MODULE_PARM_DESC(metaslab_debug, "keep space maps in core to verify frees"); | |
1981 | #endif /* _KERNEL && HAVE_SPL */ |