]>
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. |
4e21fd06 | 23 | * Copyright (c) 2011, 2016 by Delphix. All rights reserved. |
2e528b49 | 24 | * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. |
34dc7c2f BB |
25 | */ |
26 | ||
34dc7c2f | 27 | #include <sys/zfs_context.h> |
34dc7c2f BB |
28 | #include <sys/dmu.h> |
29 | #include <sys/dmu_tx.h> | |
30 | #include <sys/space_map.h> | |
31 | #include <sys/metaslab_impl.h> | |
32 | #include <sys/vdev_impl.h> | |
33 | #include <sys/zio.h> | |
93cf2076 | 34 | #include <sys/spa_impl.h> |
f3a7f661 | 35 | #include <sys/zfeature.h> |
34dc7c2f | 36 | |
d1d7e268 | 37 | #define WITH_DF_BLOCK_ALLOCATOR |
6d974228 | 38 | |
3dfb57a3 DB |
39 | #define GANG_ALLOCATION(flags) \ |
40 | ((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER)) | |
22c81dd8 | 41 | |
e8fe6684 ED |
42 | /* |
43 | * Metaslab granularity, in bytes. This is roughly similar to what would be | |
44 | * referred to as the "stripe size" in traditional RAID arrays. In normal | |
45 | * operation, we will try to write this amount of data to a top-level vdev | |
46 | * before moving on to the next one. | |
47 | */ | |
99b14de4 | 48 | unsigned long metaslab_aliquot = 512 << 10; |
e8fe6684 | 49 | |
34dc7c2f BB |
50 | uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1; /* force gang blocks */ |
51 | ||
e51be066 GW |
52 | /* |
53 | * The in-core space map representation is more compact than its on-disk form. | |
54 | * The zfs_condense_pct determines how much more compact the in-core | |
4e21fd06 | 55 | * space map representation must be before we compact it on-disk. |
e51be066 GW |
56 | * Values should be greater than or equal to 100. |
57 | */ | |
58 | int zfs_condense_pct = 200; | |
59 | ||
b02fe35d AR |
60 | /* |
61 | * Condensing a metaslab is not guaranteed to actually reduce the amount of | |
62 | * space used on disk. In particular, a space map uses data in increments of | |
96358617 | 63 | * MAX(1 << ashift, space_map_blksz), so a metaslab might use the |
b02fe35d AR |
64 | * same number of blocks after condensing. Since the goal of condensing is to |
65 | * reduce the number of IOPs required to read the space map, we only want to | |
66 | * condense when we can be sure we will reduce the number of blocks used by the | |
67 | * space map. Unfortunately, we cannot precisely compute whether or not this is | |
68 | * the case in metaslab_should_condense since we are holding ms_lock. Instead, | |
69 | * we apply the following heuristic: do not condense a spacemap unless the | |
70 | * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold | |
71 | * blocks. | |
72 | */ | |
73 | int zfs_metaslab_condense_block_threshold = 4; | |
74 | ||
ac72fac3 GW |
75 | /* |
76 | * The zfs_mg_noalloc_threshold defines which metaslab groups should | |
77 | * be eligible for allocation. The value is defined as a percentage of | |
f3a7f661 | 78 | * free space. Metaslab groups that have more free space than |
ac72fac3 GW |
79 | * zfs_mg_noalloc_threshold are always eligible for allocations. Once |
80 | * a metaslab group's free space is less than or equal to the | |
81 | * zfs_mg_noalloc_threshold the allocator will avoid allocating to that | |
82 | * group unless all groups in the pool have reached zfs_mg_noalloc_threshold. | |
83 | * Once all groups in the pool reach zfs_mg_noalloc_threshold then all | |
84 | * groups are allowed to accept allocations. Gang blocks are always | |
85 | * eligible to allocate on any metaslab group. The default value of 0 means | |
86 | * no metaslab group will be excluded based on this criterion. | |
87 | */ | |
88 | int zfs_mg_noalloc_threshold = 0; | |
6d974228 | 89 | |
f3a7f661 GW |
90 | /* |
91 | * Metaslab groups are considered eligible for allocations if their | |
92 | * fragmenation metric (measured as a percentage) is less than or equal to | |
93 | * zfs_mg_fragmentation_threshold. If a metaslab group exceeds this threshold | |
94 | * then it will be skipped unless all metaslab groups within the metaslab | |
95 | * class have also crossed this threshold. | |
96 | */ | |
97 | int zfs_mg_fragmentation_threshold = 85; | |
98 | ||
99 | /* | |
100 | * Allow metaslabs to keep their active state as long as their fragmentation | |
101 | * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An | |
102 | * active metaslab that exceeds this threshold will no longer keep its active | |
103 | * status allowing better metaslabs to be selected. | |
104 | */ | |
105 | int zfs_metaslab_fragmentation_threshold = 70; | |
106 | ||
428870ff | 107 | /* |
aa7d06a9 | 108 | * When set will load all metaslabs when pool is first opened. |
428870ff | 109 | */ |
aa7d06a9 GW |
110 | int metaslab_debug_load = 0; |
111 | ||
112 | /* | |
113 | * When set will prevent metaslabs from being unloaded. | |
114 | */ | |
115 | int metaslab_debug_unload = 0; | |
428870ff | 116 | |
9babb374 BB |
117 | /* |
118 | * Minimum size which forces the dynamic allocator to change | |
428870ff | 119 | * it's allocation strategy. Once the space map cannot satisfy |
9babb374 BB |
120 | * an allocation of this size then it switches to using more |
121 | * aggressive strategy (i.e search by size rather than offset). | |
122 | */ | |
4e21fd06 | 123 | uint64_t metaslab_df_alloc_threshold = SPA_OLD_MAXBLOCKSIZE; |
9babb374 BB |
124 | |
125 | /* | |
126 | * The minimum free space, in percent, which must be available | |
127 | * in a space map to continue allocations in a first-fit fashion. | |
4e21fd06 | 128 | * Once the space map's free space drops below this level we dynamically |
9babb374 BB |
129 | * switch to using best-fit allocations. |
130 | */ | |
428870ff BB |
131 | int metaslab_df_free_pct = 4; |
132 | ||
428870ff | 133 | /* |
93cf2076 | 134 | * Percentage of all cpus that can be used by the metaslab taskq. |
428870ff | 135 | */ |
93cf2076 | 136 | int metaslab_load_pct = 50; |
428870ff BB |
137 | |
138 | /* | |
93cf2076 GW |
139 | * Determines how many txgs a metaslab may remain loaded without having any |
140 | * allocations from it. As long as a metaslab continues to be used we will | |
141 | * keep it loaded. | |
428870ff | 142 | */ |
93cf2076 | 143 | int metaslab_unload_delay = TXG_SIZE * 2; |
9babb374 | 144 | |
93cf2076 GW |
145 | /* |
146 | * Max number of metaslabs per group to preload. | |
147 | */ | |
148 | int metaslab_preload_limit = SPA_DVAS_PER_BP; | |
149 | ||
150 | /* | |
151 | * Enable/disable preloading of metaslab. | |
152 | */ | |
f3a7f661 | 153 | int metaslab_preload_enabled = B_TRUE; |
93cf2076 GW |
154 | |
155 | /* | |
f3a7f661 | 156 | * Enable/disable fragmentation weighting on metaslabs. |
93cf2076 | 157 | */ |
f3a7f661 | 158 | int metaslab_fragmentation_factor_enabled = B_TRUE; |
93cf2076 | 159 | |
f3a7f661 GW |
160 | /* |
161 | * Enable/disable lba weighting (i.e. outer tracks are given preference). | |
162 | */ | |
163 | int metaslab_lba_weighting_enabled = B_TRUE; | |
164 | ||
165 | /* | |
166 | * Enable/disable metaslab group biasing. | |
167 | */ | |
168 | int metaslab_bias_enabled = B_TRUE; | |
169 | ||
4e21fd06 DB |
170 | |
171 | /* | |
172 | * Enable/disable segment-based metaslab selection. | |
173 | */ | |
174 | int zfs_metaslab_segment_weight_enabled = B_TRUE; | |
175 | ||
176 | /* | |
177 | * When using segment-based metaslab selection, we will continue | |
178 | * allocating from the active metaslab until we have exhausted | |
179 | * zfs_metaslab_switch_threshold of its buckets. | |
180 | */ | |
181 | int zfs_metaslab_switch_threshold = 2; | |
182 | ||
183 | /* | |
184 | * Internal switch to enable/disable the metaslab allocation tracing | |
185 | * facility. | |
186 | */ | |
187 | #ifdef _METASLAB_TRACING | |
188 | boolean_t metaslab_trace_enabled = B_TRUE; | |
189 | #endif | |
190 | ||
191 | /* | |
192 | * Maximum entries that the metaslab allocation tracing facility will keep | |
193 | * in a given list when running in non-debug mode. We limit the number | |
194 | * of entries in non-debug mode to prevent us from using up too much memory. | |
195 | * The limit should be sufficiently large that we don't expect any allocation | |
196 | * to every exceed this value. In debug mode, the system will panic if this | |
197 | * limit is ever reached allowing for further investigation. | |
198 | */ | |
199 | #ifdef _METASLAB_TRACING | |
200 | uint64_t metaslab_trace_max_entries = 5000; | |
201 | #endif | |
202 | ||
203 | static uint64_t metaslab_weight(metaslab_t *); | |
204 | static void metaslab_set_fragmentation(metaslab_t *); | |
205 | ||
206 | #ifdef _METASLAB_TRACING | |
207 | kmem_cache_t *metaslab_alloc_trace_cache; | |
208 | #endif | |
93cf2076 | 209 | |
34dc7c2f BB |
210 | /* |
211 | * ========================================================================== | |
212 | * Metaslab classes | |
213 | * ========================================================================== | |
214 | */ | |
215 | metaslab_class_t * | |
93cf2076 | 216 | metaslab_class_create(spa_t *spa, metaslab_ops_t *ops) |
34dc7c2f BB |
217 | { |
218 | metaslab_class_t *mc; | |
219 | ||
79c76d5b | 220 | mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP); |
34dc7c2f | 221 | |
428870ff | 222 | mc->mc_spa = spa; |
34dc7c2f | 223 | mc->mc_rotor = NULL; |
9babb374 | 224 | mc->mc_ops = ops; |
3dfb57a3 DB |
225 | mutex_init(&mc->mc_lock, NULL, MUTEX_DEFAULT, NULL); |
226 | refcount_create_tracked(&mc->mc_alloc_slots); | |
34dc7c2f BB |
227 | |
228 | return (mc); | |
229 | } | |
230 | ||
231 | void | |
232 | metaslab_class_destroy(metaslab_class_t *mc) | |
233 | { | |
428870ff BB |
234 | ASSERT(mc->mc_rotor == NULL); |
235 | ASSERT(mc->mc_alloc == 0); | |
236 | ASSERT(mc->mc_deferred == 0); | |
237 | ASSERT(mc->mc_space == 0); | |
238 | ASSERT(mc->mc_dspace == 0); | |
34dc7c2f | 239 | |
3dfb57a3 DB |
240 | refcount_destroy(&mc->mc_alloc_slots); |
241 | mutex_destroy(&mc->mc_lock); | |
34dc7c2f BB |
242 | kmem_free(mc, sizeof (metaslab_class_t)); |
243 | } | |
244 | ||
428870ff BB |
245 | int |
246 | metaslab_class_validate(metaslab_class_t *mc) | |
34dc7c2f | 247 | { |
428870ff BB |
248 | metaslab_group_t *mg; |
249 | vdev_t *vd; | |
34dc7c2f | 250 | |
428870ff BB |
251 | /* |
252 | * Must hold one of the spa_config locks. | |
253 | */ | |
254 | ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) || | |
255 | spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER)); | |
34dc7c2f | 256 | |
428870ff BB |
257 | if ((mg = mc->mc_rotor) == NULL) |
258 | return (0); | |
259 | ||
260 | do { | |
261 | vd = mg->mg_vd; | |
262 | ASSERT(vd->vdev_mg != NULL); | |
263 | ASSERT3P(vd->vdev_top, ==, vd); | |
264 | ASSERT3P(mg->mg_class, ==, mc); | |
265 | ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops); | |
266 | } while ((mg = mg->mg_next) != mc->mc_rotor); | |
267 | ||
268 | return (0); | |
34dc7c2f BB |
269 | } |
270 | ||
271 | void | |
428870ff BB |
272 | metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta, |
273 | int64_t defer_delta, int64_t space_delta, int64_t dspace_delta) | |
34dc7c2f | 274 | { |
428870ff BB |
275 | atomic_add_64(&mc->mc_alloc, alloc_delta); |
276 | atomic_add_64(&mc->mc_deferred, defer_delta); | |
277 | atomic_add_64(&mc->mc_space, space_delta); | |
278 | atomic_add_64(&mc->mc_dspace, dspace_delta); | |
279 | } | |
34dc7c2f | 280 | |
428870ff BB |
281 | uint64_t |
282 | metaslab_class_get_alloc(metaslab_class_t *mc) | |
283 | { | |
284 | return (mc->mc_alloc); | |
285 | } | |
34dc7c2f | 286 | |
428870ff BB |
287 | uint64_t |
288 | metaslab_class_get_deferred(metaslab_class_t *mc) | |
289 | { | |
290 | return (mc->mc_deferred); | |
291 | } | |
34dc7c2f | 292 | |
428870ff BB |
293 | uint64_t |
294 | metaslab_class_get_space(metaslab_class_t *mc) | |
295 | { | |
296 | return (mc->mc_space); | |
297 | } | |
34dc7c2f | 298 | |
428870ff BB |
299 | uint64_t |
300 | metaslab_class_get_dspace(metaslab_class_t *mc) | |
301 | { | |
302 | return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space); | |
34dc7c2f BB |
303 | } |
304 | ||
f3a7f661 GW |
305 | void |
306 | metaslab_class_histogram_verify(metaslab_class_t *mc) | |
307 | { | |
308 | vdev_t *rvd = mc->mc_spa->spa_root_vdev; | |
309 | uint64_t *mc_hist; | |
310 | int i, c; | |
311 | ||
312 | if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0) | |
313 | return; | |
314 | ||
315 | mc_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE, | |
79c76d5b | 316 | KM_SLEEP); |
f3a7f661 GW |
317 | |
318 | for (c = 0; c < rvd->vdev_children; c++) { | |
319 | vdev_t *tvd = rvd->vdev_child[c]; | |
320 | metaslab_group_t *mg = tvd->vdev_mg; | |
321 | ||
322 | /* | |
323 | * Skip any holes, uninitialized top-levels, or | |
324 | * vdevs that are not in this metalab class. | |
325 | */ | |
326 | if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 || | |
327 | mg->mg_class != mc) { | |
328 | continue; | |
329 | } | |
330 | ||
331 | for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++) | |
332 | mc_hist[i] += mg->mg_histogram[i]; | |
333 | } | |
334 | ||
335 | for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++) | |
336 | VERIFY3U(mc_hist[i], ==, mc->mc_histogram[i]); | |
337 | ||
338 | kmem_free(mc_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE); | |
339 | } | |
340 | ||
341 | /* | |
342 | * Calculate the metaslab class's fragmentation metric. The metric | |
343 | * is weighted based on the space contribution of each metaslab group. | |
344 | * The return value will be a number between 0 and 100 (inclusive), or | |
345 | * ZFS_FRAG_INVALID if the metric has not been set. See comment above the | |
346 | * zfs_frag_table for more information about the metric. | |
347 | */ | |
348 | uint64_t | |
349 | metaslab_class_fragmentation(metaslab_class_t *mc) | |
350 | { | |
351 | vdev_t *rvd = mc->mc_spa->spa_root_vdev; | |
352 | uint64_t fragmentation = 0; | |
353 | int c; | |
354 | ||
355 | spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER); | |
356 | ||
357 | for (c = 0; c < rvd->vdev_children; c++) { | |
358 | vdev_t *tvd = rvd->vdev_child[c]; | |
359 | metaslab_group_t *mg = tvd->vdev_mg; | |
360 | ||
361 | /* | |
362 | * Skip any holes, uninitialized top-levels, or | |
363 | * vdevs that are not in this metalab class. | |
364 | */ | |
365 | if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 || | |
366 | mg->mg_class != mc) { | |
367 | continue; | |
368 | } | |
369 | ||
370 | /* | |
371 | * If a metaslab group does not contain a fragmentation | |
372 | * metric then just bail out. | |
373 | */ | |
374 | if (mg->mg_fragmentation == ZFS_FRAG_INVALID) { | |
375 | spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG); | |
376 | return (ZFS_FRAG_INVALID); | |
377 | } | |
378 | ||
379 | /* | |
380 | * Determine how much this metaslab_group is contributing | |
381 | * to the overall pool fragmentation metric. | |
382 | */ | |
383 | fragmentation += mg->mg_fragmentation * | |
384 | metaslab_group_get_space(mg); | |
385 | } | |
386 | fragmentation /= metaslab_class_get_space(mc); | |
387 | ||
388 | ASSERT3U(fragmentation, <=, 100); | |
389 | spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG); | |
390 | return (fragmentation); | |
391 | } | |
392 | ||
393 | /* | |
394 | * Calculate the amount of expandable space that is available in | |
395 | * this metaslab class. If a device is expanded then its expandable | |
396 | * space will be the amount of allocatable space that is currently not | |
397 | * part of this metaslab class. | |
398 | */ | |
399 | uint64_t | |
400 | metaslab_class_expandable_space(metaslab_class_t *mc) | |
401 | { | |
402 | vdev_t *rvd = mc->mc_spa->spa_root_vdev; | |
403 | uint64_t space = 0; | |
404 | int c; | |
405 | ||
406 | spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER); | |
407 | for (c = 0; c < rvd->vdev_children; c++) { | |
408 | vdev_t *tvd = rvd->vdev_child[c]; | |
409 | metaslab_group_t *mg = tvd->vdev_mg; | |
410 | ||
411 | if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 || | |
412 | mg->mg_class != mc) { | |
413 | continue; | |
414 | } | |
415 | ||
416 | space += tvd->vdev_max_asize - tvd->vdev_asize; | |
417 | } | |
418 | spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG); | |
419 | return (space); | |
420 | } | |
421 | ||
34dc7c2f BB |
422 | static int |
423 | metaslab_compare(const void *x1, const void *x2) | |
424 | { | |
ee36c709 GN |
425 | const metaslab_t *m1 = (const metaslab_t *)x1; |
426 | const metaslab_t *m2 = (const metaslab_t *)x2; | |
34dc7c2f | 427 | |
ee36c709 GN |
428 | int cmp = AVL_CMP(m2->ms_weight, m1->ms_weight); |
429 | if (likely(cmp)) | |
430 | return (cmp); | |
34dc7c2f | 431 | |
ee36c709 | 432 | IMPLY(AVL_CMP(m1->ms_start, m2->ms_start) == 0, m1 == m2); |
34dc7c2f | 433 | |
ee36c709 | 434 | return (AVL_CMP(m1->ms_start, m2->ms_start)); |
34dc7c2f BB |
435 | } |
436 | ||
4e21fd06 DB |
437 | /* |
438 | * Verify that the space accounting on disk matches the in-core range_trees. | |
439 | */ | |
440 | void | |
441 | metaslab_verify_space(metaslab_t *msp, uint64_t txg) | |
442 | { | |
443 | spa_t *spa = msp->ms_group->mg_vd->vdev_spa; | |
444 | uint64_t allocated = 0; | |
445 | uint64_t freed = 0; | |
446 | uint64_t sm_free_space, msp_free_space; | |
447 | int t; | |
448 | ||
449 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
450 | ||
451 | if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0) | |
452 | return; | |
453 | ||
454 | /* | |
455 | * We can only verify the metaslab space when we're called | |
456 | * from syncing context with a loaded metaslab that has an allocated | |
457 | * space map. Calling this in non-syncing context does not | |
458 | * provide a consistent view of the metaslab since we're performing | |
459 | * allocations in the future. | |
460 | */ | |
461 | if (txg != spa_syncing_txg(spa) || msp->ms_sm == NULL || | |
462 | !msp->ms_loaded) | |
463 | return; | |
464 | ||
465 | sm_free_space = msp->ms_size - space_map_allocated(msp->ms_sm) - | |
466 | space_map_alloc_delta(msp->ms_sm); | |
467 | ||
468 | /* | |
469 | * Account for future allocations since we would have already | |
470 | * deducted that space from the ms_freetree. | |
471 | */ | |
472 | for (t = 0; t < TXG_CONCURRENT_STATES; t++) { | |
473 | allocated += | |
474 | range_tree_space(msp->ms_alloctree[(txg + t) & TXG_MASK]); | |
475 | } | |
476 | freed = range_tree_space(msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK]); | |
477 | ||
478 | msp_free_space = range_tree_space(msp->ms_tree) + allocated + | |
479 | msp->ms_deferspace + freed; | |
480 | ||
481 | VERIFY3U(sm_free_space, ==, msp_free_space); | |
482 | } | |
483 | ||
484 | /* | |
485 | * ========================================================================== | |
486 | * Metaslab groups | |
487 | * ========================================================================== | |
488 | */ | |
ac72fac3 GW |
489 | /* |
490 | * Update the allocatable flag and the metaslab group's capacity. | |
491 | * The allocatable flag is set to true if the capacity is below | |
3dfb57a3 DB |
492 | * the zfs_mg_noalloc_threshold or has a fragmentation value that is |
493 | * greater than zfs_mg_fragmentation_threshold. If a metaslab group | |
494 | * transitions from allocatable to non-allocatable or vice versa then the | |
495 | * metaslab group's class is updated to reflect the transition. | |
ac72fac3 GW |
496 | */ |
497 | static void | |
498 | metaslab_group_alloc_update(metaslab_group_t *mg) | |
499 | { | |
500 | vdev_t *vd = mg->mg_vd; | |
501 | metaslab_class_t *mc = mg->mg_class; | |
502 | vdev_stat_t *vs = &vd->vdev_stat; | |
503 | boolean_t was_allocatable; | |
3dfb57a3 | 504 | boolean_t was_initialized; |
ac72fac3 GW |
505 | |
506 | ASSERT(vd == vd->vdev_top); | |
507 | ||
508 | mutex_enter(&mg->mg_lock); | |
509 | was_allocatable = mg->mg_allocatable; | |
3dfb57a3 | 510 | was_initialized = mg->mg_initialized; |
ac72fac3 GW |
511 | |
512 | mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) / | |
513 | (vs->vs_space + 1); | |
514 | ||
3dfb57a3 DB |
515 | mutex_enter(&mc->mc_lock); |
516 | ||
517 | /* | |
518 | * If the metaslab group was just added then it won't | |
519 | * have any space until we finish syncing out this txg. | |
520 | * At that point we will consider it initialized and available | |
521 | * for allocations. We also don't consider non-activated | |
522 | * metaslab groups (e.g. vdevs that are in the middle of being removed) | |
523 | * to be initialized, because they can't be used for allocation. | |
524 | */ | |
525 | mg->mg_initialized = metaslab_group_initialized(mg); | |
526 | if (!was_initialized && mg->mg_initialized) { | |
527 | mc->mc_groups++; | |
528 | } else if (was_initialized && !mg->mg_initialized) { | |
529 | ASSERT3U(mc->mc_groups, >, 0); | |
530 | mc->mc_groups--; | |
531 | } | |
532 | if (mg->mg_initialized) | |
533 | mg->mg_no_free_space = B_FALSE; | |
534 | ||
f3a7f661 GW |
535 | /* |
536 | * A metaslab group is considered allocatable if it has plenty | |
537 | * of free space or is not heavily fragmented. We only take | |
538 | * fragmentation into account if the metaslab group has a valid | |
539 | * fragmentation metric (i.e. a value between 0 and 100). | |
540 | */ | |
3dfb57a3 DB |
541 | mg->mg_allocatable = (mg->mg_activation_count > 0 && |
542 | mg->mg_free_capacity > zfs_mg_noalloc_threshold && | |
f3a7f661 GW |
543 | (mg->mg_fragmentation == ZFS_FRAG_INVALID || |
544 | mg->mg_fragmentation <= zfs_mg_fragmentation_threshold)); | |
ac72fac3 GW |
545 | |
546 | /* | |
547 | * The mc_alloc_groups maintains a count of the number of | |
548 | * groups in this metaslab class that are still above the | |
549 | * zfs_mg_noalloc_threshold. This is used by the allocating | |
550 | * threads to determine if they should avoid allocations to | |
551 | * a given group. The allocator will avoid allocations to a group | |
552 | * if that group has reached or is below the zfs_mg_noalloc_threshold | |
553 | * and there are still other groups that are above the threshold. | |
554 | * When a group transitions from allocatable to non-allocatable or | |
555 | * vice versa we update the metaslab class to reflect that change. | |
556 | * When the mc_alloc_groups value drops to 0 that means that all | |
557 | * groups have reached the zfs_mg_noalloc_threshold making all groups | |
558 | * eligible for allocations. This effectively means that all devices | |
559 | * are balanced again. | |
560 | */ | |
561 | if (was_allocatable && !mg->mg_allocatable) | |
562 | mc->mc_alloc_groups--; | |
563 | else if (!was_allocatable && mg->mg_allocatable) | |
564 | mc->mc_alloc_groups++; | |
3dfb57a3 | 565 | mutex_exit(&mc->mc_lock); |
f3a7f661 | 566 | |
ac72fac3 GW |
567 | mutex_exit(&mg->mg_lock); |
568 | } | |
569 | ||
34dc7c2f BB |
570 | metaslab_group_t * |
571 | metaslab_group_create(metaslab_class_t *mc, vdev_t *vd) | |
572 | { | |
573 | metaslab_group_t *mg; | |
574 | ||
79c76d5b | 575 | mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP); |
34dc7c2f BB |
576 | mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL); |
577 | avl_create(&mg->mg_metaslab_tree, metaslab_compare, | |
578 | sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node)); | |
34dc7c2f | 579 | mg->mg_vd = vd; |
428870ff BB |
580 | mg->mg_class = mc; |
581 | mg->mg_activation_count = 0; | |
3dfb57a3 DB |
582 | mg->mg_initialized = B_FALSE; |
583 | mg->mg_no_free_space = B_TRUE; | |
584 | refcount_create_tracked(&mg->mg_alloc_queue_depth); | |
34dc7c2f | 585 | |
3c51c5cb | 586 | mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct, |
1229323d | 587 | maxclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT | TASKQ_DYNAMIC); |
93cf2076 | 588 | |
34dc7c2f BB |
589 | return (mg); |
590 | } | |
591 | ||
592 | void | |
593 | metaslab_group_destroy(metaslab_group_t *mg) | |
594 | { | |
428870ff BB |
595 | ASSERT(mg->mg_prev == NULL); |
596 | ASSERT(mg->mg_next == NULL); | |
597 | /* | |
598 | * We may have gone below zero with the activation count | |
599 | * either because we never activated in the first place or | |
600 | * because we're done, and possibly removing the vdev. | |
601 | */ | |
602 | ASSERT(mg->mg_activation_count <= 0); | |
603 | ||
3c51c5cb | 604 | taskq_destroy(mg->mg_taskq); |
34dc7c2f BB |
605 | avl_destroy(&mg->mg_metaslab_tree); |
606 | mutex_destroy(&mg->mg_lock); | |
3dfb57a3 | 607 | refcount_destroy(&mg->mg_alloc_queue_depth); |
34dc7c2f BB |
608 | kmem_free(mg, sizeof (metaslab_group_t)); |
609 | } | |
610 | ||
428870ff BB |
611 | void |
612 | metaslab_group_activate(metaslab_group_t *mg) | |
613 | { | |
614 | metaslab_class_t *mc = mg->mg_class; | |
615 | metaslab_group_t *mgprev, *mgnext; | |
616 | ||
617 | ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER)); | |
618 | ||
619 | ASSERT(mc->mc_rotor != mg); | |
620 | ASSERT(mg->mg_prev == NULL); | |
621 | ASSERT(mg->mg_next == NULL); | |
622 | ASSERT(mg->mg_activation_count <= 0); | |
623 | ||
624 | if (++mg->mg_activation_count <= 0) | |
625 | return; | |
626 | ||
627 | mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children); | |
ac72fac3 | 628 | metaslab_group_alloc_update(mg); |
428870ff BB |
629 | |
630 | if ((mgprev = mc->mc_rotor) == NULL) { | |
631 | mg->mg_prev = mg; | |
632 | mg->mg_next = mg; | |
633 | } else { | |
634 | mgnext = mgprev->mg_next; | |
635 | mg->mg_prev = mgprev; | |
636 | mg->mg_next = mgnext; | |
637 | mgprev->mg_next = mg; | |
638 | mgnext->mg_prev = mg; | |
639 | } | |
640 | mc->mc_rotor = mg; | |
641 | } | |
642 | ||
643 | void | |
644 | metaslab_group_passivate(metaslab_group_t *mg) | |
645 | { | |
646 | metaslab_class_t *mc = mg->mg_class; | |
647 | metaslab_group_t *mgprev, *mgnext; | |
648 | ||
649 | ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER)); | |
650 | ||
651 | if (--mg->mg_activation_count != 0) { | |
652 | ASSERT(mc->mc_rotor != mg); | |
653 | ASSERT(mg->mg_prev == NULL); | |
654 | ASSERT(mg->mg_next == NULL); | |
655 | ASSERT(mg->mg_activation_count < 0); | |
656 | return; | |
657 | } | |
658 | ||
c5528b9b | 659 | taskq_wait_outstanding(mg->mg_taskq, 0); |
f3a7f661 | 660 | metaslab_group_alloc_update(mg); |
93cf2076 | 661 | |
428870ff BB |
662 | mgprev = mg->mg_prev; |
663 | mgnext = mg->mg_next; | |
664 | ||
665 | if (mg == mgnext) { | |
666 | mc->mc_rotor = NULL; | |
667 | } else { | |
668 | mc->mc_rotor = mgnext; | |
669 | mgprev->mg_next = mgnext; | |
670 | mgnext->mg_prev = mgprev; | |
671 | } | |
672 | ||
673 | mg->mg_prev = NULL; | |
674 | mg->mg_next = NULL; | |
675 | } | |
676 | ||
3dfb57a3 DB |
677 | boolean_t |
678 | metaslab_group_initialized(metaslab_group_t *mg) | |
679 | { | |
680 | vdev_t *vd = mg->mg_vd; | |
681 | vdev_stat_t *vs = &vd->vdev_stat; | |
682 | ||
683 | return (vs->vs_space != 0 && mg->mg_activation_count > 0); | |
684 | } | |
685 | ||
f3a7f661 GW |
686 | uint64_t |
687 | metaslab_group_get_space(metaslab_group_t *mg) | |
688 | { | |
689 | return ((1ULL << mg->mg_vd->vdev_ms_shift) * mg->mg_vd->vdev_ms_count); | |
690 | } | |
691 | ||
692 | void | |
693 | metaslab_group_histogram_verify(metaslab_group_t *mg) | |
694 | { | |
695 | uint64_t *mg_hist; | |
696 | vdev_t *vd = mg->mg_vd; | |
697 | uint64_t ashift = vd->vdev_ashift; | |
698 | int i, m; | |
699 | ||
700 | if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0) | |
701 | return; | |
702 | ||
703 | mg_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE, | |
79c76d5b | 704 | KM_SLEEP); |
f3a7f661 GW |
705 | |
706 | ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE, >=, | |
707 | SPACE_MAP_HISTOGRAM_SIZE + ashift); | |
708 | ||
709 | for (m = 0; m < vd->vdev_ms_count; m++) { | |
710 | metaslab_t *msp = vd->vdev_ms[m]; | |
711 | ||
712 | if (msp->ms_sm == NULL) | |
713 | continue; | |
714 | ||
715 | for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) | |
716 | mg_hist[i + ashift] += | |
717 | msp->ms_sm->sm_phys->smp_histogram[i]; | |
718 | } | |
719 | ||
720 | for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i ++) | |
721 | VERIFY3U(mg_hist[i], ==, mg->mg_histogram[i]); | |
722 | ||
723 | kmem_free(mg_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE); | |
724 | } | |
725 | ||
34dc7c2f | 726 | static void |
f3a7f661 | 727 | metaslab_group_histogram_add(metaslab_group_t *mg, metaslab_t *msp) |
34dc7c2f | 728 | { |
f3a7f661 GW |
729 | metaslab_class_t *mc = mg->mg_class; |
730 | uint64_t ashift = mg->mg_vd->vdev_ashift; | |
731 | int i; | |
732 | ||
733 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
734 | if (msp->ms_sm == NULL) | |
735 | return; | |
736 | ||
34dc7c2f | 737 | mutex_enter(&mg->mg_lock); |
f3a7f661 GW |
738 | for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) { |
739 | mg->mg_histogram[i + ashift] += | |
740 | msp->ms_sm->sm_phys->smp_histogram[i]; | |
741 | mc->mc_histogram[i + ashift] += | |
742 | msp->ms_sm->sm_phys->smp_histogram[i]; | |
743 | } | |
744 | mutex_exit(&mg->mg_lock); | |
745 | } | |
746 | ||
747 | void | |
748 | metaslab_group_histogram_remove(metaslab_group_t *mg, metaslab_t *msp) | |
749 | { | |
750 | metaslab_class_t *mc = mg->mg_class; | |
751 | uint64_t ashift = mg->mg_vd->vdev_ashift; | |
752 | int i; | |
753 | ||
754 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
755 | if (msp->ms_sm == NULL) | |
756 | return; | |
757 | ||
758 | mutex_enter(&mg->mg_lock); | |
759 | for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) { | |
760 | ASSERT3U(mg->mg_histogram[i + ashift], >=, | |
761 | msp->ms_sm->sm_phys->smp_histogram[i]); | |
762 | ASSERT3U(mc->mc_histogram[i + ashift], >=, | |
763 | msp->ms_sm->sm_phys->smp_histogram[i]); | |
764 | ||
765 | mg->mg_histogram[i + ashift] -= | |
766 | msp->ms_sm->sm_phys->smp_histogram[i]; | |
767 | mc->mc_histogram[i + ashift] -= | |
768 | msp->ms_sm->sm_phys->smp_histogram[i]; | |
769 | } | |
770 | mutex_exit(&mg->mg_lock); | |
771 | } | |
772 | ||
773 | static void | |
774 | metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp) | |
775 | { | |
34dc7c2f | 776 | ASSERT(msp->ms_group == NULL); |
f3a7f661 | 777 | mutex_enter(&mg->mg_lock); |
34dc7c2f BB |
778 | msp->ms_group = mg; |
779 | msp->ms_weight = 0; | |
780 | avl_add(&mg->mg_metaslab_tree, msp); | |
781 | mutex_exit(&mg->mg_lock); | |
f3a7f661 GW |
782 | |
783 | mutex_enter(&msp->ms_lock); | |
784 | metaslab_group_histogram_add(mg, msp); | |
785 | mutex_exit(&msp->ms_lock); | |
34dc7c2f BB |
786 | } |
787 | ||
788 | static void | |
789 | metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp) | |
790 | { | |
f3a7f661 GW |
791 | mutex_enter(&msp->ms_lock); |
792 | metaslab_group_histogram_remove(mg, msp); | |
793 | mutex_exit(&msp->ms_lock); | |
794 | ||
34dc7c2f BB |
795 | mutex_enter(&mg->mg_lock); |
796 | ASSERT(msp->ms_group == mg); | |
797 | avl_remove(&mg->mg_metaslab_tree, msp); | |
798 | msp->ms_group = NULL; | |
799 | mutex_exit(&mg->mg_lock); | |
800 | } | |
801 | ||
802 | static void | |
803 | metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight) | |
804 | { | |
805 | /* | |
806 | * Although in principle the weight can be any value, in | |
f3a7f661 | 807 | * practice we do not use values in the range [1, 511]. |
34dc7c2f | 808 | */ |
f3a7f661 | 809 | ASSERT(weight >= SPA_MINBLOCKSIZE || weight == 0); |
34dc7c2f BB |
810 | ASSERT(MUTEX_HELD(&msp->ms_lock)); |
811 | ||
812 | mutex_enter(&mg->mg_lock); | |
813 | ASSERT(msp->ms_group == mg); | |
814 | avl_remove(&mg->mg_metaslab_tree, msp); | |
815 | msp->ms_weight = weight; | |
816 | avl_add(&mg->mg_metaslab_tree, msp); | |
817 | mutex_exit(&mg->mg_lock); | |
818 | } | |
819 | ||
f3a7f661 GW |
820 | /* |
821 | * Calculate the fragmentation for a given metaslab group. We can use | |
822 | * a simple average here since all metaslabs within the group must have | |
823 | * the same size. The return value will be a value between 0 and 100 | |
824 | * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this | |
825 | * group have a fragmentation metric. | |
826 | */ | |
827 | uint64_t | |
828 | metaslab_group_fragmentation(metaslab_group_t *mg) | |
829 | { | |
830 | vdev_t *vd = mg->mg_vd; | |
831 | uint64_t fragmentation = 0; | |
832 | uint64_t valid_ms = 0; | |
833 | int m; | |
834 | ||
835 | for (m = 0; m < vd->vdev_ms_count; m++) { | |
836 | metaslab_t *msp = vd->vdev_ms[m]; | |
837 | ||
838 | if (msp->ms_fragmentation == ZFS_FRAG_INVALID) | |
839 | continue; | |
840 | ||
841 | valid_ms++; | |
842 | fragmentation += msp->ms_fragmentation; | |
843 | } | |
844 | ||
845 | if (valid_ms <= vd->vdev_ms_count / 2) | |
846 | return (ZFS_FRAG_INVALID); | |
847 | ||
848 | fragmentation /= valid_ms; | |
849 | ASSERT3U(fragmentation, <=, 100); | |
850 | return (fragmentation); | |
851 | } | |
852 | ||
ac72fac3 GW |
853 | /* |
854 | * Determine if a given metaslab group should skip allocations. A metaslab | |
f3a7f661 GW |
855 | * group should avoid allocations if its free capacity is less than the |
856 | * zfs_mg_noalloc_threshold or its fragmentation metric is greater than | |
857 | * zfs_mg_fragmentation_threshold and there is at least one metaslab group | |
3dfb57a3 DB |
858 | * that can still handle allocations. If the allocation throttle is enabled |
859 | * then we skip allocations to devices that have reached their maximum | |
860 | * allocation queue depth unless the selected metaslab group is the only | |
861 | * eligible group remaining. | |
ac72fac3 GW |
862 | */ |
863 | static boolean_t | |
3dfb57a3 DB |
864 | metaslab_group_allocatable(metaslab_group_t *mg, metaslab_group_t *rotor, |
865 | uint64_t psize) | |
ac72fac3 | 866 | { |
3dfb57a3 | 867 | spa_t *spa = mg->mg_vd->vdev_spa; |
ac72fac3 GW |
868 | metaslab_class_t *mc = mg->mg_class; |
869 | ||
870 | /* | |
3dfb57a3 DB |
871 | * We can only consider skipping this metaslab group if it's |
872 | * in the normal metaslab class and there are other metaslab | |
873 | * groups to select from. Otherwise, we always consider it eligible | |
f3a7f661 | 874 | * for allocations. |
ac72fac3 | 875 | */ |
3dfb57a3 DB |
876 | if (mc != spa_normal_class(spa) || mc->mc_groups <= 1) |
877 | return (B_TRUE); | |
878 | ||
879 | /* | |
880 | * If the metaslab group's mg_allocatable flag is set (see comments | |
881 | * in metaslab_group_alloc_update() for more information) and | |
882 | * the allocation throttle is disabled then allow allocations to this | |
883 | * device. However, if the allocation throttle is enabled then | |
884 | * check if we have reached our allocation limit (mg_alloc_queue_depth) | |
885 | * to determine if we should allow allocations to this metaslab group. | |
886 | * If all metaslab groups are no longer considered allocatable | |
887 | * (mc_alloc_groups == 0) or we're trying to allocate the smallest | |
888 | * gang block size then we allow allocations on this metaslab group | |
889 | * regardless of the mg_allocatable or throttle settings. | |
890 | */ | |
891 | if (mg->mg_allocatable) { | |
892 | metaslab_group_t *mgp; | |
893 | int64_t qdepth; | |
894 | uint64_t qmax = mg->mg_max_alloc_queue_depth; | |
895 | ||
896 | if (!mc->mc_alloc_throttle_enabled) | |
897 | return (B_TRUE); | |
898 | ||
899 | /* | |
900 | * If this metaslab group does not have any free space, then | |
901 | * there is no point in looking further. | |
902 | */ | |
903 | if (mg->mg_no_free_space) | |
904 | return (B_FALSE); | |
905 | ||
906 | qdepth = refcount_count(&mg->mg_alloc_queue_depth); | |
907 | ||
908 | /* | |
909 | * If this metaslab group is below its qmax or it's | |
910 | * the only allocatable metasable group, then attempt | |
911 | * to allocate from it. | |
912 | */ | |
913 | if (qdepth < qmax || mc->mc_alloc_groups == 1) | |
914 | return (B_TRUE); | |
915 | ASSERT3U(mc->mc_alloc_groups, >, 1); | |
916 | ||
917 | /* | |
918 | * Since this metaslab group is at or over its qmax, we | |
919 | * need to determine if there are metaslab groups after this | |
920 | * one that might be able to handle this allocation. This is | |
921 | * racy since we can't hold the locks for all metaslab | |
922 | * groups at the same time when we make this check. | |
923 | */ | |
924 | for (mgp = mg->mg_next; mgp != rotor; mgp = mgp->mg_next) { | |
925 | qmax = mgp->mg_max_alloc_queue_depth; | |
926 | ||
927 | qdepth = refcount_count(&mgp->mg_alloc_queue_depth); | |
928 | ||
929 | /* | |
930 | * If there is another metaslab group that | |
931 | * might be able to handle the allocation, then | |
932 | * we return false so that we skip this group. | |
933 | */ | |
934 | if (qdepth < qmax && !mgp->mg_no_free_space) | |
935 | return (B_FALSE); | |
936 | } | |
937 | ||
938 | /* | |
939 | * We didn't find another group to handle the allocation | |
940 | * so we can't skip this metaslab group even though | |
941 | * we are at or over our qmax. | |
942 | */ | |
943 | return (B_TRUE); | |
944 | ||
945 | } else if (mc->mc_alloc_groups == 0 || psize == SPA_MINBLOCKSIZE) { | |
946 | return (B_TRUE); | |
947 | } | |
948 | return (B_FALSE); | |
ac72fac3 GW |
949 | } |
950 | ||
428870ff BB |
951 | /* |
952 | * ========================================================================== | |
93cf2076 | 953 | * Range tree callbacks |
428870ff BB |
954 | * ========================================================================== |
955 | */ | |
93cf2076 GW |
956 | |
957 | /* | |
958 | * Comparison function for the private size-ordered tree. Tree is sorted | |
959 | * by size, larger sizes at the end of the tree. | |
960 | */ | |
428870ff | 961 | static int |
93cf2076 | 962 | metaslab_rangesize_compare(const void *x1, const void *x2) |
428870ff | 963 | { |
93cf2076 GW |
964 | const range_seg_t *r1 = x1; |
965 | const range_seg_t *r2 = x2; | |
966 | uint64_t rs_size1 = r1->rs_end - r1->rs_start; | |
967 | uint64_t rs_size2 = r2->rs_end - r2->rs_start; | |
428870ff | 968 | |
ee36c709 GN |
969 | int cmp = AVL_CMP(rs_size1, rs_size2); |
970 | if (likely(cmp)) | |
971 | return (cmp); | |
428870ff | 972 | |
ee36c709 | 973 | return (AVL_CMP(r1->rs_start, r2->rs_start)); |
428870ff BB |
974 | } |
975 | ||
34dc7c2f | 976 | /* |
93cf2076 GW |
977 | * Create any block allocator specific components. The current allocators |
978 | * rely on using both a size-ordered range_tree_t and an array of uint64_t's. | |
34dc7c2f | 979 | */ |
93cf2076 GW |
980 | static void |
981 | metaslab_rt_create(range_tree_t *rt, void *arg) | |
34dc7c2f | 982 | { |
93cf2076 | 983 | metaslab_t *msp = arg; |
34dc7c2f | 984 | |
93cf2076 GW |
985 | ASSERT3P(rt->rt_arg, ==, msp); |
986 | ASSERT(msp->ms_tree == NULL); | |
34dc7c2f | 987 | |
93cf2076 GW |
988 | avl_create(&msp->ms_size_tree, metaslab_rangesize_compare, |
989 | sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node)); | |
9babb374 BB |
990 | } |
991 | ||
93cf2076 GW |
992 | /* |
993 | * Destroy the block allocator specific components. | |
994 | */ | |
9babb374 | 995 | static void |
93cf2076 | 996 | metaslab_rt_destroy(range_tree_t *rt, void *arg) |
9babb374 | 997 | { |
93cf2076 | 998 | metaslab_t *msp = arg; |
428870ff | 999 | |
93cf2076 GW |
1000 | ASSERT3P(rt->rt_arg, ==, msp); |
1001 | ASSERT3P(msp->ms_tree, ==, rt); | |
1002 | ASSERT0(avl_numnodes(&msp->ms_size_tree)); | |
428870ff | 1003 | |
93cf2076 | 1004 | avl_destroy(&msp->ms_size_tree); |
9babb374 BB |
1005 | } |
1006 | ||
1007 | static void | |
93cf2076 | 1008 | metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg) |
9babb374 | 1009 | { |
93cf2076 | 1010 | metaslab_t *msp = arg; |
9babb374 | 1011 | |
93cf2076 GW |
1012 | ASSERT3P(rt->rt_arg, ==, msp); |
1013 | ASSERT3P(msp->ms_tree, ==, rt); | |
1014 | VERIFY(!msp->ms_condensing); | |
1015 | avl_add(&msp->ms_size_tree, rs); | |
34dc7c2f BB |
1016 | } |
1017 | ||
34dc7c2f | 1018 | static void |
93cf2076 | 1019 | metaslab_rt_remove(range_tree_t *rt, range_seg_t *rs, void *arg) |
34dc7c2f | 1020 | { |
93cf2076 GW |
1021 | metaslab_t *msp = arg; |
1022 | ||
1023 | ASSERT3P(rt->rt_arg, ==, msp); | |
1024 | ASSERT3P(msp->ms_tree, ==, rt); | |
1025 | VERIFY(!msp->ms_condensing); | |
1026 | avl_remove(&msp->ms_size_tree, rs); | |
34dc7c2f BB |
1027 | } |
1028 | ||
34dc7c2f | 1029 | static void |
93cf2076 | 1030 | metaslab_rt_vacate(range_tree_t *rt, void *arg) |
34dc7c2f | 1031 | { |
93cf2076 GW |
1032 | metaslab_t *msp = arg; |
1033 | ||
1034 | ASSERT3P(rt->rt_arg, ==, msp); | |
1035 | ASSERT3P(msp->ms_tree, ==, rt); | |
1036 | ||
1037 | /* | |
1038 | * Normally one would walk the tree freeing nodes along the way. | |
1039 | * Since the nodes are shared with the range trees we can avoid | |
1040 | * walking all nodes and just reinitialize the avl tree. The nodes | |
1041 | * will be freed by the range tree, so we don't want to free them here. | |
1042 | */ | |
1043 | avl_create(&msp->ms_size_tree, metaslab_rangesize_compare, | |
1044 | sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node)); | |
34dc7c2f BB |
1045 | } |
1046 | ||
93cf2076 GW |
1047 | static range_tree_ops_t metaslab_rt_ops = { |
1048 | metaslab_rt_create, | |
1049 | metaslab_rt_destroy, | |
1050 | metaslab_rt_add, | |
1051 | metaslab_rt_remove, | |
1052 | metaslab_rt_vacate | |
1053 | }; | |
1054 | ||
1055 | /* | |
1056 | * ========================================================================== | |
4e21fd06 | 1057 | * Common allocator routines |
93cf2076 GW |
1058 | * ========================================================================== |
1059 | */ | |
1060 | ||
9babb374 | 1061 | /* |
428870ff | 1062 | * Return the maximum contiguous segment within the metaslab. |
9babb374 | 1063 | */ |
9babb374 | 1064 | uint64_t |
93cf2076 | 1065 | metaslab_block_maxsize(metaslab_t *msp) |
9babb374 | 1066 | { |
93cf2076 GW |
1067 | avl_tree_t *t = &msp->ms_size_tree; |
1068 | range_seg_t *rs; | |
9babb374 | 1069 | |
93cf2076 | 1070 | if (t == NULL || (rs = avl_last(t)) == NULL) |
9babb374 BB |
1071 | return (0ULL); |
1072 | ||
93cf2076 GW |
1073 | return (rs->rs_end - rs->rs_start); |
1074 | } | |
1075 | ||
4e21fd06 DB |
1076 | static range_seg_t * |
1077 | metaslab_block_find(avl_tree_t *t, uint64_t start, uint64_t size) | |
93cf2076 | 1078 | { |
4e21fd06 DB |
1079 | range_seg_t *rs, rsearch; |
1080 | avl_index_t where; | |
93cf2076 | 1081 | |
4e21fd06 DB |
1082 | rsearch.rs_start = start; |
1083 | rsearch.rs_end = start + size; | |
93cf2076 | 1084 | |
4e21fd06 DB |
1085 | rs = avl_find(t, &rsearch, &where); |
1086 | if (rs == NULL) { | |
1087 | rs = avl_nearest(t, where, AVL_AFTER); | |
93cf2076 | 1088 | } |
93cf2076 | 1089 | |
4e21fd06 DB |
1090 | return (rs); |
1091 | } | |
93cf2076 GW |
1092 | |
1093 | #if defined(WITH_FF_BLOCK_ALLOCATOR) || \ | |
1094 | defined(WITH_DF_BLOCK_ALLOCATOR) || \ | |
1095 | defined(WITH_CF_BLOCK_ALLOCATOR) | |
1096 | /* | |
1097 | * This is a helper function that can be used by the allocator to find | |
1098 | * a suitable block to allocate. This will search the specified AVL | |
1099 | * tree looking for a block that matches the specified criteria. | |
1100 | */ | |
1101 | static uint64_t | |
1102 | metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size, | |
1103 | uint64_t align) | |
1104 | { | |
4e21fd06 | 1105 | range_seg_t *rs = metaslab_block_find(t, *cursor, size); |
93cf2076 GW |
1106 | |
1107 | while (rs != NULL) { | |
1108 | uint64_t offset = P2ROUNDUP(rs->rs_start, align); | |
1109 | ||
1110 | if (offset + size <= rs->rs_end) { | |
1111 | *cursor = offset + size; | |
1112 | return (offset); | |
1113 | } | |
1114 | rs = AVL_NEXT(t, rs); | |
1115 | } | |
1116 | ||
1117 | /* | |
1118 | * If we know we've searched the whole map (*cursor == 0), give up. | |
1119 | * Otherwise, reset the cursor to the beginning and try again. | |
1120 | */ | |
1121 | if (*cursor == 0) | |
1122 | return (-1ULL); | |
1123 | ||
1124 | *cursor = 0; | |
1125 | return (metaslab_block_picker(t, cursor, size, align)); | |
9babb374 | 1126 | } |
93cf2076 | 1127 | #endif /* WITH_FF/DF/CF_BLOCK_ALLOCATOR */ |
9babb374 | 1128 | |
22c81dd8 | 1129 | #if defined(WITH_FF_BLOCK_ALLOCATOR) |
428870ff BB |
1130 | /* |
1131 | * ========================================================================== | |
1132 | * The first-fit block allocator | |
1133 | * ========================================================================== | |
1134 | */ | |
1135 | static uint64_t | |
93cf2076 | 1136 | metaslab_ff_alloc(metaslab_t *msp, uint64_t size) |
9babb374 | 1137 | { |
93cf2076 GW |
1138 | /* |
1139 | * Find the largest power of 2 block size that evenly divides the | |
1140 | * requested size. This is used to try to allocate blocks with similar | |
1141 | * alignment from the same area of the metaslab (i.e. same cursor | |
1142 | * bucket) but it does not guarantee that other allocations sizes | |
1143 | * may exist in the same region. | |
1144 | */ | |
428870ff | 1145 | uint64_t align = size & -size; |
9bd274dd | 1146 | uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1]; |
93cf2076 | 1147 | avl_tree_t *t = &msp->ms_tree->rt_root; |
9babb374 | 1148 | |
428870ff | 1149 | return (metaslab_block_picker(t, cursor, size, align)); |
9babb374 BB |
1150 | } |
1151 | ||
93cf2076 | 1152 | static metaslab_ops_t metaslab_ff_ops = { |
f3a7f661 | 1153 | metaslab_ff_alloc |
428870ff | 1154 | }; |
9babb374 | 1155 | |
93cf2076 | 1156 | metaslab_ops_t *zfs_metaslab_ops = &metaslab_ff_ops; |
22c81dd8 BB |
1157 | #endif /* WITH_FF_BLOCK_ALLOCATOR */ |
1158 | ||
1159 | #if defined(WITH_DF_BLOCK_ALLOCATOR) | |
428870ff BB |
1160 | /* |
1161 | * ========================================================================== | |
1162 | * Dynamic block allocator - | |
1163 | * Uses the first fit allocation scheme until space get low and then | |
1164 | * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold | |
1165 | * and metaslab_df_free_pct to determine when to switch the allocation scheme. | |
1166 | * ========================================================================== | |
1167 | */ | |
9babb374 | 1168 | static uint64_t |
93cf2076 | 1169 | metaslab_df_alloc(metaslab_t *msp, uint64_t size) |
9babb374 | 1170 | { |
93cf2076 GW |
1171 | /* |
1172 | * Find the largest power of 2 block size that evenly divides the | |
1173 | * requested size. This is used to try to allocate blocks with similar | |
1174 | * alignment from the same area of the metaslab (i.e. same cursor | |
1175 | * bucket) but it does not guarantee that other allocations sizes | |
1176 | * may exist in the same region. | |
1177 | */ | |
9babb374 | 1178 | uint64_t align = size & -size; |
9bd274dd | 1179 | uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1]; |
93cf2076 GW |
1180 | range_tree_t *rt = msp->ms_tree; |
1181 | avl_tree_t *t = &rt->rt_root; | |
1182 | uint64_t max_size = metaslab_block_maxsize(msp); | |
1183 | int free_pct = range_tree_space(rt) * 100 / msp->ms_size; | |
9babb374 | 1184 | |
93cf2076 GW |
1185 | ASSERT(MUTEX_HELD(&msp->ms_lock)); |
1186 | ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree)); | |
9babb374 BB |
1187 | |
1188 | if (max_size < size) | |
1189 | return (-1ULL); | |
1190 | ||
1191 | /* | |
1192 | * If we're running low on space switch to using the size | |
1193 | * sorted AVL tree (best-fit). | |
1194 | */ | |
1195 | if (max_size < metaslab_df_alloc_threshold || | |
1196 | free_pct < metaslab_df_free_pct) { | |
93cf2076 | 1197 | t = &msp->ms_size_tree; |
9babb374 BB |
1198 | *cursor = 0; |
1199 | } | |
1200 | ||
1201 | return (metaslab_block_picker(t, cursor, size, 1ULL)); | |
1202 | } | |
1203 | ||
93cf2076 | 1204 | static metaslab_ops_t metaslab_df_ops = { |
f3a7f661 | 1205 | metaslab_df_alloc |
34dc7c2f BB |
1206 | }; |
1207 | ||
93cf2076 | 1208 | metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops; |
22c81dd8 BB |
1209 | #endif /* WITH_DF_BLOCK_ALLOCATOR */ |
1210 | ||
93cf2076 | 1211 | #if defined(WITH_CF_BLOCK_ALLOCATOR) |
428870ff BB |
1212 | /* |
1213 | * ========================================================================== | |
93cf2076 GW |
1214 | * Cursor fit block allocator - |
1215 | * Select the largest region in the metaslab, set the cursor to the beginning | |
1216 | * of the range and the cursor_end to the end of the range. As allocations | |
1217 | * are made advance the cursor. Continue allocating from the cursor until | |
1218 | * the range is exhausted and then find a new range. | |
428870ff BB |
1219 | * ========================================================================== |
1220 | */ | |
1221 | static uint64_t | |
93cf2076 | 1222 | metaslab_cf_alloc(metaslab_t *msp, uint64_t size) |
428870ff | 1223 | { |
93cf2076 GW |
1224 | range_tree_t *rt = msp->ms_tree; |
1225 | avl_tree_t *t = &msp->ms_size_tree; | |
1226 | uint64_t *cursor = &msp->ms_lbas[0]; | |
1227 | uint64_t *cursor_end = &msp->ms_lbas[1]; | |
428870ff BB |
1228 | uint64_t offset = 0; |
1229 | ||
93cf2076 GW |
1230 | ASSERT(MUTEX_HELD(&msp->ms_lock)); |
1231 | ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&rt->rt_root)); | |
428870ff | 1232 | |
93cf2076 | 1233 | ASSERT3U(*cursor_end, >=, *cursor); |
428870ff | 1234 | |
93cf2076 GW |
1235 | if ((*cursor + size) > *cursor_end) { |
1236 | range_seg_t *rs; | |
428870ff | 1237 | |
93cf2076 GW |
1238 | rs = avl_last(&msp->ms_size_tree); |
1239 | if (rs == NULL || (rs->rs_end - rs->rs_start) < size) | |
1240 | return (-1ULL); | |
428870ff | 1241 | |
93cf2076 GW |
1242 | *cursor = rs->rs_start; |
1243 | *cursor_end = rs->rs_end; | |
428870ff | 1244 | } |
93cf2076 GW |
1245 | |
1246 | offset = *cursor; | |
1247 | *cursor += size; | |
1248 | ||
428870ff BB |
1249 | return (offset); |
1250 | } | |
1251 | ||
93cf2076 | 1252 | static metaslab_ops_t metaslab_cf_ops = { |
f3a7f661 | 1253 | metaslab_cf_alloc |
428870ff BB |
1254 | }; |
1255 | ||
93cf2076 GW |
1256 | metaslab_ops_t *zfs_metaslab_ops = &metaslab_cf_ops; |
1257 | #endif /* WITH_CF_BLOCK_ALLOCATOR */ | |
22c81dd8 BB |
1258 | |
1259 | #if defined(WITH_NDF_BLOCK_ALLOCATOR) | |
93cf2076 GW |
1260 | /* |
1261 | * ========================================================================== | |
1262 | * New dynamic fit allocator - | |
1263 | * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift | |
1264 | * contiguous blocks. If no region is found then just use the largest segment | |
1265 | * that remains. | |
1266 | * ========================================================================== | |
1267 | */ | |
1268 | ||
1269 | /* | |
1270 | * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift) | |
1271 | * to request from the allocator. | |
1272 | */ | |
428870ff BB |
1273 | uint64_t metaslab_ndf_clump_shift = 4; |
1274 | ||
1275 | static uint64_t | |
93cf2076 | 1276 | metaslab_ndf_alloc(metaslab_t *msp, uint64_t size) |
428870ff | 1277 | { |
93cf2076 | 1278 | avl_tree_t *t = &msp->ms_tree->rt_root; |
428870ff | 1279 | avl_index_t where; |
93cf2076 | 1280 | range_seg_t *rs, rsearch; |
9bd274dd | 1281 | uint64_t hbit = highbit64(size); |
93cf2076 GW |
1282 | uint64_t *cursor = &msp->ms_lbas[hbit - 1]; |
1283 | uint64_t max_size = metaslab_block_maxsize(msp); | |
428870ff | 1284 | |
93cf2076 GW |
1285 | ASSERT(MUTEX_HELD(&msp->ms_lock)); |
1286 | ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree)); | |
428870ff BB |
1287 | |
1288 | if (max_size < size) | |
1289 | return (-1ULL); | |
1290 | ||
93cf2076 GW |
1291 | rsearch.rs_start = *cursor; |
1292 | rsearch.rs_end = *cursor + size; | |
428870ff | 1293 | |
93cf2076 GW |
1294 | rs = avl_find(t, &rsearch, &where); |
1295 | if (rs == NULL || (rs->rs_end - rs->rs_start) < size) { | |
1296 | t = &msp->ms_size_tree; | |
428870ff | 1297 | |
93cf2076 GW |
1298 | rsearch.rs_start = 0; |
1299 | rsearch.rs_end = MIN(max_size, | |
428870ff | 1300 | 1ULL << (hbit + metaslab_ndf_clump_shift)); |
93cf2076 GW |
1301 | rs = avl_find(t, &rsearch, &where); |
1302 | if (rs == NULL) | |
1303 | rs = avl_nearest(t, where, AVL_AFTER); | |
1304 | ASSERT(rs != NULL); | |
428870ff BB |
1305 | } |
1306 | ||
93cf2076 GW |
1307 | if ((rs->rs_end - rs->rs_start) >= size) { |
1308 | *cursor = rs->rs_start + size; | |
1309 | return (rs->rs_start); | |
428870ff BB |
1310 | } |
1311 | return (-1ULL); | |
1312 | } | |
1313 | ||
93cf2076 | 1314 | static metaslab_ops_t metaslab_ndf_ops = { |
f3a7f661 | 1315 | metaslab_ndf_alloc |
428870ff BB |
1316 | }; |
1317 | ||
93cf2076 | 1318 | metaslab_ops_t *zfs_metaslab_ops = &metaslab_ndf_ops; |
22c81dd8 | 1319 | #endif /* WITH_NDF_BLOCK_ALLOCATOR */ |
9babb374 | 1320 | |
93cf2076 | 1321 | |
34dc7c2f BB |
1322 | /* |
1323 | * ========================================================================== | |
1324 | * Metaslabs | |
1325 | * ========================================================================== | |
1326 | */ | |
93cf2076 GW |
1327 | |
1328 | /* | |
1329 | * Wait for any in-progress metaslab loads to complete. | |
1330 | */ | |
1331 | void | |
1332 | metaslab_load_wait(metaslab_t *msp) | |
1333 | { | |
1334 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
1335 | ||
1336 | while (msp->ms_loading) { | |
1337 | ASSERT(!msp->ms_loaded); | |
1338 | cv_wait(&msp->ms_load_cv, &msp->ms_lock); | |
1339 | } | |
1340 | } | |
1341 | ||
1342 | int | |
1343 | metaslab_load(metaslab_t *msp) | |
1344 | { | |
1345 | int error = 0; | |
1346 | int t; | |
4e21fd06 | 1347 | boolean_t success = B_FALSE; |
93cf2076 GW |
1348 | |
1349 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
1350 | ASSERT(!msp->ms_loaded); | |
1351 | ASSERT(!msp->ms_loading); | |
1352 | ||
1353 | msp->ms_loading = B_TRUE; | |
1354 | ||
1355 | /* | |
1356 | * If the space map has not been allocated yet, then treat | |
1357 | * all the space in the metaslab as free and add it to the | |
1358 | * ms_tree. | |
1359 | */ | |
1360 | if (msp->ms_sm != NULL) | |
1361 | error = space_map_load(msp->ms_sm, msp->ms_tree, SM_FREE); | |
1362 | else | |
1363 | range_tree_add(msp->ms_tree, msp->ms_start, msp->ms_size); | |
1364 | ||
4e21fd06 | 1365 | success = (error == 0); |
93cf2076 GW |
1366 | msp->ms_loading = B_FALSE; |
1367 | ||
4e21fd06 DB |
1368 | if (success) { |
1369 | ASSERT3P(msp->ms_group, !=, NULL); | |
1370 | msp->ms_loaded = B_TRUE; | |
1371 | ||
93cf2076 GW |
1372 | for (t = 0; t < TXG_DEFER_SIZE; t++) { |
1373 | range_tree_walk(msp->ms_defertree[t], | |
1374 | range_tree_remove, msp->ms_tree); | |
1375 | } | |
4e21fd06 | 1376 | msp->ms_max_size = metaslab_block_maxsize(msp); |
93cf2076 GW |
1377 | } |
1378 | cv_broadcast(&msp->ms_load_cv); | |
1379 | return (error); | |
1380 | } | |
1381 | ||
1382 | void | |
1383 | metaslab_unload(metaslab_t *msp) | |
1384 | { | |
1385 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
1386 | range_tree_vacate(msp->ms_tree, NULL, NULL); | |
1387 | msp->ms_loaded = B_FALSE; | |
1388 | msp->ms_weight &= ~METASLAB_ACTIVE_MASK; | |
4e21fd06 | 1389 | msp->ms_max_size = 0; |
93cf2076 GW |
1390 | } |
1391 | ||
fb42a493 PS |
1392 | int |
1393 | metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, uint64_t txg, | |
1394 | metaslab_t **msp) | |
34dc7c2f BB |
1395 | { |
1396 | vdev_t *vd = mg->mg_vd; | |
93cf2076 | 1397 | objset_t *mos = vd->vdev_spa->spa_meta_objset; |
fb42a493 PS |
1398 | metaslab_t *ms; |
1399 | int error; | |
34dc7c2f | 1400 | |
79c76d5b | 1401 | ms = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP); |
fb42a493 PS |
1402 | mutex_init(&ms->ms_lock, NULL, MUTEX_DEFAULT, NULL); |
1403 | cv_init(&ms->ms_load_cv, NULL, CV_DEFAULT, NULL); | |
1404 | ms->ms_id = id; | |
1405 | ms->ms_start = id << vd->vdev_ms_shift; | |
1406 | ms->ms_size = 1ULL << vd->vdev_ms_shift; | |
34dc7c2f | 1407 | |
93cf2076 GW |
1408 | /* |
1409 | * We only open space map objects that already exist. All others | |
afe37326 | 1410 | * will be opened when we finally allocate an object for it. |
93cf2076 | 1411 | */ |
afe37326 | 1412 | if (object != 0) { |
fb42a493 PS |
1413 | error = space_map_open(&ms->ms_sm, mos, object, ms->ms_start, |
1414 | ms->ms_size, vd->vdev_ashift, &ms->ms_lock); | |
1415 | ||
1416 | if (error != 0) { | |
1417 | kmem_free(ms, sizeof (metaslab_t)); | |
1418 | return (error); | |
1419 | } | |
1420 | ||
1421 | ASSERT(ms->ms_sm != NULL); | |
93cf2076 | 1422 | } |
34dc7c2f BB |
1423 | |
1424 | /* | |
93cf2076 GW |
1425 | * We create the main range tree here, but we don't create the |
1426 | * alloctree and freetree until metaslab_sync_done(). This serves | |
34dc7c2f BB |
1427 | * two purposes: it allows metaslab_sync_done() to detect the |
1428 | * addition of new space; and for debugging, it ensures that we'd | |
1429 | * data fault on any attempt to use this metaslab before it's ready. | |
1430 | */ | |
fb42a493 PS |
1431 | ms->ms_tree = range_tree_create(&metaslab_rt_ops, ms, &ms->ms_lock); |
1432 | metaslab_group_add(mg, ms); | |
34dc7c2f | 1433 | |
4e21fd06 | 1434 | metaslab_set_fragmentation(ms); |
428870ff | 1435 | |
34dc7c2f BB |
1436 | /* |
1437 | * If we're opening an existing pool (txg == 0) or creating | |
1438 | * a new one (txg == TXG_INITIAL), all space is available now. | |
1439 | * If we're adding space to an existing pool, the new space | |
1440 | * does not become available until after this txg has synced. | |
4e21fd06 DB |
1441 | * The metaslab's weight will also be initialized when we sync |
1442 | * out this txg. This ensures that we don't attempt to allocate | |
1443 | * from it before we have initialized it completely. | |
34dc7c2f BB |
1444 | */ |
1445 | if (txg <= TXG_INITIAL) | |
fb42a493 | 1446 | metaslab_sync_done(ms, 0); |
34dc7c2f | 1447 | |
93cf2076 GW |
1448 | /* |
1449 | * If metaslab_debug_load is set and we're initializing a metaslab | |
4e21fd06 | 1450 | * that has an allocated space map object then load the its space |
93cf2076 GW |
1451 | * map so that can verify frees. |
1452 | */ | |
fb42a493 PS |
1453 | if (metaslab_debug_load && ms->ms_sm != NULL) { |
1454 | mutex_enter(&ms->ms_lock); | |
1455 | VERIFY0(metaslab_load(ms)); | |
1456 | mutex_exit(&ms->ms_lock); | |
93cf2076 GW |
1457 | } |
1458 | ||
34dc7c2f | 1459 | if (txg != 0) { |
34dc7c2f | 1460 | vdev_dirty(vd, 0, NULL, txg); |
fb42a493 | 1461 | vdev_dirty(vd, VDD_METASLAB, ms, txg); |
34dc7c2f BB |
1462 | } |
1463 | ||
fb42a493 PS |
1464 | *msp = ms; |
1465 | ||
1466 | return (0); | |
34dc7c2f BB |
1467 | } |
1468 | ||
1469 | void | |
1470 | metaslab_fini(metaslab_t *msp) | |
1471 | { | |
d6320ddb | 1472 | int t; |
34dc7c2f | 1473 | |
93cf2076 | 1474 | metaslab_group_t *mg = msp->ms_group; |
34dc7c2f BB |
1475 | |
1476 | metaslab_group_remove(mg, msp); | |
1477 | ||
1478 | mutex_enter(&msp->ms_lock); | |
93cf2076 GW |
1479 | VERIFY(msp->ms_group == NULL); |
1480 | vdev_space_update(mg->mg_vd, -space_map_allocated(msp->ms_sm), | |
1481 | 0, -msp->ms_size); | |
1482 | space_map_close(msp->ms_sm); | |
1483 | ||
1484 | metaslab_unload(msp); | |
1485 | range_tree_destroy(msp->ms_tree); | |
34dc7c2f | 1486 | |
d6320ddb | 1487 | for (t = 0; t < TXG_SIZE; t++) { |
93cf2076 GW |
1488 | range_tree_destroy(msp->ms_alloctree[t]); |
1489 | range_tree_destroy(msp->ms_freetree[t]); | |
34dc7c2f BB |
1490 | } |
1491 | ||
e51be066 | 1492 | for (t = 0; t < TXG_DEFER_SIZE; t++) { |
93cf2076 | 1493 | range_tree_destroy(msp->ms_defertree[t]); |
e51be066 | 1494 | } |
428870ff | 1495 | |
c99c9001 | 1496 | ASSERT0(msp->ms_deferspace); |
428870ff | 1497 | |
34dc7c2f | 1498 | mutex_exit(&msp->ms_lock); |
93cf2076 | 1499 | cv_destroy(&msp->ms_load_cv); |
34dc7c2f BB |
1500 | mutex_destroy(&msp->ms_lock); |
1501 | ||
1502 | kmem_free(msp, sizeof (metaslab_t)); | |
1503 | } | |
1504 | ||
f3a7f661 GW |
1505 | #define FRAGMENTATION_TABLE_SIZE 17 |
1506 | ||
93cf2076 | 1507 | /* |
f3a7f661 GW |
1508 | * This table defines a segment size based fragmentation metric that will |
1509 | * allow each metaslab to derive its own fragmentation value. This is done | |
1510 | * by calculating the space in each bucket of the spacemap histogram and | |
1511 | * multiplying that by the fragmetation metric in this table. Doing | |
1512 | * this for all buckets and dividing it by the total amount of free | |
1513 | * space in this metaslab (i.e. the total free space in all buckets) gives | |
1514 | * us the fragmentation metric. This means that a high fragmentation metric | |
1515 | * equates to most of the free space being comprised of small segments. | |
1516 | * Conversely, if the metric is low, then most of the free space is in | |
1517 | * large segments. A 10% change in fragmentation equates to approximately | |
1518 | * double the number of segments. | |
93cf2076 | 1519 | * |
f3a7f661 GW |
1520 | * This table defines 0% fragmented space using 16MB segments. Testing has |
1521 | * shown that segments that are greater than or equal to 16MB do not suffer | |
1522 | * from drastic performance problems. Using this value, we derive the rest | |
1523 | * of the table. Since the fragmentation value is never stored on disk, it | |
1524 | * is possible to change these calculations in the future. | |
1525 | */ | |
1526 | int zfs_frag_table[FRAGMENTATION_TABLE_SIZE] = { | |
1527 | 100, /* 512B */ | |
1528 | 100, /* 1K */ | |
1529 | 98, /* 2K */ | |
1530 | 95, /* 4K */ | |
1531 | 90, /* 8K */ | |
1532 | 80, /* 16K */ | |
1533 | 70, /* 32K */ | |
1534 | 60, /* 64K */ | |
1535 | 50, /* 128K */ | |
1536 | 40, /* 256K */ | |
1537 | 30, /* 512K */ | |
1538 | 20, /* 1M */ | |
1539 | 15, /* 2M */ | |
1540 | 10, /* 4M */ | |
1541 | 5, /* 8M */ | |
1542 | 0 /* 16M */ | |
1543 | }; | |
1544 | ||
1545 | /* | |
1546 | * Calclate the metaslab's fragmentation metric. A return value | |
1547 | * of ZFS_FRAG_INVALID means that the metaslab has not been upgraded and does | |
1548 | * not support this metric. Otherwise, the return value should be in the | |
1549 | * range [0, 100]. | |
93cf2076 | 1550 | */ |
4e21fd06 DB |
1551 | static void |
1552 | metaslab_set_fragmentation(metaslab_t *msp) | |
93cf2076 | 1553 | { |
f3a7f661 GW |
1554 | spa_t *spa = msp->ms_group->mg_vd->vdev_spa; |
1555 | uint64_t fragmentation = 0; | |
1556 | uint64_t total = 0; | |
1557 | boolean_t feature_enabled = spa_feature_is_enabled(spa, | |
1558 | SPA_FEATURE_SPACEMAP_HISTOGRAM); | |
93cf2076 GW |
1559 | int i; |
1560 | ||
4e21fd06 DB |
1561 | if (!feature_enabled) { |
1562 | msp->ms_fragmentation = ZFS_FRAG_INVALID; | |
1563 | return; | |
1564 | } | |
f3a7f661 | 1565 | |
93cf2076 | 1566 | /* |
f3a7f661 GW |
1567 | * A null space map means that the entire metaslab is free |
1568 | * and thus is not fragmented. | |
93cf2076 | 1569 | */ |
4e21fd06 DB |
1570 | if (msp->ms_sm == NULL) { |
1571 | msp->ms_fragmentation = 0; | |
1572 | return; | |
1573 | } | |
f3a7f661 GW |
1574 | |
1575 | /* | |
4e21fd06 | 1576 | * If this metaslab's space map has not been upgraded, flag it |
f3a7f661 GW |
1577 | * so that we upgrade next time we encounter it. |
1578 | */ | |
1579 | if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t)) { | |
93cf2076 GW |
1580 | vdev_t *vd = msp->ms_group->mg_vd; |
1581 | ||
8b0a0840 TC |
1582 | if (spa_writeable(vd->vdev_spa)) { |
1583 | uint64_t txg = spa_syncing_txg(spa); | |
1584 | ||
1585 | msp->ms_condense_wanted = B_TRUE; | |
1586 | vdev_dirty(vd, VDD_METASLAB, msp, txg + 1); | |
1587 | spa_dbgmsg(spa, "txg %llu, requesting force condense: " | |
1588 | "msp %p, vd %p", txg, msp, vd); | |
1589 | } | |
4e21fd06 DB |
1590 | msp->ms_fragmentation = ZFS_FRAG_INVALID; |
1591 | return; | |
93cf2076 GW |
1592 | } |
1593 | ||
f3a7f661 GW |
1594 | for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) { |
1595 | uint64_t space = 0; | |
1596 | uint8_t shift = msp->ms_sm->sm_shift; | |
4e21fd06 | 1597 | |
f3a7f661 GW |
1598 | int idx = MIN(shift - SPA_MINBLOCKSHIFT + i, |
1599 | FRAGMENTATION_TABLE_SIZE - 1); | |
93cf2076 | 1600 | |
93cf2076 GW |
1601 | if (msp->ms_sm->sm_phys->smp_histogram[i] == 0) |
1602 | continue; | |
1603 | ||
f3a7f661 GW |
1604 | space = msp->ms_sm->sm_phys->smp_histogram[i] << (i + shift); |
1605 | total += space; | |
1606 | ||
1607 | ASSERT3U(idx, <, FRAGMENTATION_TABLE_SIZE); | |
1608 | fragmentation += space * zfs_frag_table[idx]; | |
93cf2076 | 1609 | } |
f3a7f661 GW |
1610 | |
1611 | if (total > 0) | |
1612 | fragmentation /= total; | |
1613 | ASSERT3U(fragmentation, <=, 100); | |
4e21fd06 DB |
1614 | |
1615 | msp->ms_fragmentation = fragmentation; | |
93cf2076 | 1616 | } |
34dc7c2f | 1617 | |
f3a7f661 GW |
1618 | /* |
1619 | * Compute a weight -- a selection preference value -- for the given metaslab. | |
1620 | * This is based on the amount of free space, the level of fragmentation, | |
1621 | * the LBA range, and whether the metaslab is loaded. | |
1622 | */ | |
34dc7c2f | 1623 | static uint64_t |
4e21fd06 | 1624 | metaslab_space_weight(metaslab_t *msp) |
34dc7c2f BB |
1625 | { |
1626 | metaslab_group_t *mg = msp->ms_group; | |
34dc7c2f BB |
1627 | vdev_t *vd = mg->mg_vd; |
1628 | uint64_t weight, space; | |
1629 | ||
1630 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
4e21fd06 | 1631 | ASSERT(!vd->vdev_removing); |
c2e42f9d | 1632 | |
34dc7c2f BB |
1633 | /* |
1634 | * The baseline weight is the metaslab's free space. | |
1635 | */ | |
93cf2076 | 1636 | space = msp->ms_size - space_map_allocated(msp->ms_sm); |
f3a7f661 | 1637 | |
f3a7f661 GW |
1638 | if (metaslab_fragmentation_factor_enabled && |
1639 | msp->ms_fragmentation != ZFS_FRAG_INVALID) { | |
1640 | /* | |
1641 | * Use the fragmentation information to inversely scale | |
1642 | * down the baseline weight. We need to ensure that we | |
1643 | * don't exclude this metaslab completely when it's 100% | |
1644 | * fragmented. To avoid this we reduce the fragmented value | |
1645 | * by 1. | |
1646 | */ | |
1647 | space = (space * (100 - (msp->ms_fragmentation - 1))) / 100; | |
1648 | ||
1649 | /* | |
1650 | * If space < SPA_MINBLOCKSIZE, then we will not allocate from | |
1651 | * this metaslab again. The fragmentation metric may have | |
1652 | * decreased the space to something smaller than | |
1653 | * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE | |
1654 | * so that we can consume any remaining space. | |
1655 | */ | |
1656 | if (space > 0 && space < SPA_MINBLOCKSIZE) | |
1657 | space = SPA_MINBLOCKSIZE; | |
1658 | } | |
34dc7c2f BB |
1659 | weight = space; |
1660 | ||
1661 | /* | |
1662 | * Modern disks have uniform bit density and constant angular velocity. | |
1663 | * Therefore, the outer recording zones are faster (higher bandwidth) | |
1664 | * than the inner zones by the ratio of outer to inner track diameter, | |
1665 | * which is typically around 2:1. We account for this by assigning | |
1666 | * higher weight to lower metaslabs (multiplier ranging from 2x to 1x). | |
1667 | * In effect, this means that we'll select the metaslab with the most | |
1668 | * free bandwidth rather than simply the one with the most free space. | |
1669 | */ | |
fb40095f | 1670 | if (!vd->vdev_nonrot && metaslab_lba_weighting_enabled) { |
f3a7f661 GW |
1671 | weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count; |
1672 | ASSERT(weight >= space && weight <= 2 * space); | |
1673 | } | |
428870ff | 1674 | |
f3a7f661 GW |
1675 | /* |
1676 | * If this metaslab is one we're actively using, adjust its | |
1677 | * weight to make it preferable to any inactive metaslab so | |
1678 | * we'll polish it off. If the fragmentation on this metaslab | |
1679 | * has exceed our threshold, then don't mark it active. | |
1680 | */ | |
1681 | if (msp->ms_loaded && msp->ms_fragmentation != ZFS_FRAG_INVALID && | |
1682 | msp->ms_fragmentation <= zfs_metaslab_fragmentation_threshold) { | |
428870ff BB |
1683 | weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK); |
1684 | } | |
34dc7c2f | 1685 | |
4e21fd06 DB |
1686 | WEIGHT_SET_SPACEBASED(weight); |
1687 | return (weight); | |
1688 | } | |
1689 | ||
1690 | /* | |
1691 | * Return the weight of the specified metaslab, according to the segment-based | |
1692 | * weighting algorithm. The metaslab must be loaded. This function can | |
1693 | * be called within a sync pass since it relies only on the metaslab's | |
1694 | * range tree which is always accurate when the metaslab is loaded. | |
1695 | */ | |
1696 | static uint64_t | |
1697 | metaslab_weight_from_range_tree(metaslab_t *msp) | |
1698 | { | |
1699 | uint64_t weight = 0; | |
1700 | uint32_t segments = 0; | |
1701 | int i; | |
1702 | ||
1703 | ASSERT(msp->ms_loaded); | |
1704 | ||
1705 | for (i = RANGE_TREE_HISTOGRAM_SIZE - 1; i >= SPA_MINBLOCKSHIFT; i--) { | |
1706 | uint8_t shift = msp->ms_group->mg_vd->vdev_ashift; | |
1707 | int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1; | |
1708 | ||
1709 | segments <<= 1; | |
1710 | segments += msp->ms_tree->rt_histogram[i]; | |
1711 | ||
1712 | /* | |
1713 | * The range tree provides more precision than the space map | |
1714 | * and must be downgraded so that all values fit within the | |
1715 | * space map's histogram. This allows us to compare loaded | |
1716 | * vs. unloaded metaslabs to determine which metaslab is | |
1717 | * considered "best". | |
1718 | */ | |
1719 | if (i > max_idx) | |
1720 | continue; | |
1721 | ||
1722 | if (segments != 0) { | |
1723 | WEIGHT_SET_COUNT(weight, segments); | |
1724 | WEIGHT_SET_INDEX(weight, i); | |
1725 | WEIGHT_SET_ACTIVE(weight, 0); | |
1726 | break; | |
1727 | } | |
1728 | } | |
1729 | return (weight); | |
1730 | } | |
1731 | ||
1732 | /* | |
1733 | * Calculate the weight based on the on-disk histogram. This should only | |
1734 | * be called after a sync pass has completely finished since the on-disk | |
1735 | * information is updated in metaslab_sync(). | |
1736 | */ | |
1737 | static uint64_t | |
1738 | metaslab_weight_from_spacemap(metaslab_t *msp) | |
1739 | { | |
1740 | uint64_t weight = 0; | |
1741 | int i; | |
1742 | ||
1743 | for (i = SPACE_MAP_HISTOGRAM_SIZE - 1; i >= 0; i--) { | |
1744 | if (msp->ms_sm->sm_phys->smp_histogram[i] != 0) { | |
1745 | WEIGHT_SET_COUNT(weight, | |
1746 | msp->ms_sm->sm_phys->smp_histogram[i]); | |
1747 | WEIGHT_SET_INDEX(weight, i + | |
1748 | msp->ms_sm->sm_shift); | |
1749 | WEIGHT_SET_ACTIVE(weight, 0); | |
1750 | break; | |
1751 | } | |
1752 | } | |
1753 | return (weight); | |
1754 | } | |
1755 | ||
1756 | /* | |
1757 | * Compute a segment-based weight for the specified metaslab. The weight | |
1758 | * is determined by highest bucket in the histogram. The information | |
1759 | * for the highest bucket is encoded into the weight value. | |
1760 | */ | |
1761 | static uint64_t | |
1762 | metaslab_segment_weight(metaslab_t *msp) | |
1763 | { | |
1764 | metaslab_group_t *mg = msp->ms_group; | |
1765 | uint64_t weight = 0; | |
1766 | uint8_t shift = mg->mg_vd->vdev_ashift; | |
1767 | ||
1768 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
1769 | ||
1770 | /* | |
1771 | * The metaslab is completely free. | |
1772 | */ | |
1773 | if (space_map_allocated(msp->ms_sm) == 0) { | |
1774 | int idx = highbit64(msp->ms_size) - 1; | |
1775 | int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1; | |
1776 | ||
1777 | if (idx < max_idx) { | |
1778 | WEIGHT_SET_COUNT(weight, 1ULL); | |
1779 | WEIGHT_SET_INDEX(weight, idx); | |
1780 | } else { | |
1781 | WEIGHT_SET_COUNT(weight, 1ULL << (idx - max_idx)); | |
1782 | WEIGHT_SET_INDEX(weight, max_idx); | |
1783 | } | |
1784 | WEIGHT_SET_ACTIVE(weight, 0); | |
1785 | ASSERT(!WEIGHT_IS_SPACEBASED(weight)); | |
1786 | ||
1787 | return (weight); | |
1788 | } | |
1789 | ||
1790 | ASSERT3U(msp->ms_sm->sm_dbuf->db_size, ==, sizeof (space_map_phys_t)); | |
1791 | ||
1792 | /* | |
1793 | * If the metaslab is fully allocated then just make the weight 0. | |
1794 | */ | |
1795 | if (space_map_allocated(msp->ms_sm) == msp->ms_size) | |
1796 | return (0); | |
1797 | /* | |
1798 | * If the metaslab is already loaded, then use the range tree to | |
1799 | * determine the weight. Otherwise, we rely on the space map information | |
1800 | * to generate the weight. | |
1801 | */ | |
1802 | if (msp->ms_loaded) { | |
1803 | weight = metaslab_weight_from_range_tree(msp); | |
1804 | } else { | |
1805 | weight = metaslab_weight_from_spacemap(msp); | |
1806 | } | |
1807 | ||
1808 | /* | |
1809 | * If the metaslab was active the last time we calculated its weight | |
1810 | * then keep it active. We want to consume the entire region that | |
1811 | * is associated with this weight. | |
1812 | */ | |
1813 | if (msp->ms_activation_weight != 0 && weight != 0) | |
1814 | WEIGHT_SET_ACTIVE(weight, WEIGHT_GET_ACTIVE(msp->ms_weight)); | |
1815 | return (weight); | |
1816 | } | |
1817 | ||
1818 | /* | |
1819 | * Determine if we should attempt to allocate from this metaslab. If the | |
1820 | * metaslab has a maximum size then we can quickly determine if the desired | |
1821 | * allocation size can be satisfied. Otherwise, if we're using segment-based | |
1822 | * weighting then we can determine the maximum allocation that this metaslab | |
1823 | * can accommodate based on the index encoded in the weight. If we're using | |
1824 | * space-based weights then rely on the entire weight (excluding the weight | |
1825 | * type bit). | |
1826 | */ | |
1827 | boolean_t | |
1828 | metaslab_should_allocate(metaslab_t *msp, uint64_t asize) | |
1829 | { | |
1830 | boolean_t should_allocate; | |
1831 | ||
1832 | if (msp->ms_max_size != 0) | |
1833 | return (msp->ms_max_size >= asize); | |
1834 | ||
1835 | if (!WEIGHT_IS_SPACEBASED(msp->ms_weight)) { | |
1836 | /* | |
1837 | * The metaslab segment weight indicates segments in the | |
1838 | * range [2^i, 2^(i+1)), where i is the index in the weight. | |
1839 | * Since the asize might be in the middle of the range, we | |
1840 | * should attempt the allocation if asize < 2^(i+1). | |
1841 | */ | |
1842 | should_allocate = (asize < | |
1843 | 1ULL << (WEIGHT_GET_INDEX(msp->ms_weight) + 1)); | |
1844 | } else { | |
1845 | should_allocate = (asize <= | |
1846 | (msp->ms_weight & ~METASLAB_WEIGHT_TYPE)); | |
1847 | } | |
1848 | return (should_allocate); | |
1849 | } | |
1850 | static uint64_t | |
1851 | metaslab_weight(metaslab_t *msp) | |
1852 | { | |
1853 | vdev_t *vd = msp->ms_group->mg_vd; | |
1854 | spa_t *spa = vd->vdev_spa; | |
1855 | uint64_t weight; | |
1856 | ||
1857 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
1858 | ||
1859 | /* | |
1860 | * This vdev is in the process of being removed so there is nothing | |
1861 | * for us to do here. | |
1862 | */ | |
1863 | if (vd->vdev_removing) { | |
1864 | ASSERT0(space_map_allocated(msp->ms_sm)); | |
1865 | ASSERT0(vd->vdev_ms_shift); | |
1866 | return (0); | |
1867 | } | |
1868 | ||
1869 | metaslab_set_fragmentation(msp); | |
1870 | ||
1871 | /* | |
1872 | * Update the maximum size if the metaslab is loaded. This will | |
1873 | * ensure that we get an accurate maximum size if newly freed space | |
1874 | * has been added back into the free tree. | |
1875 | */ | |
1876 | if (msp->ms_loaded) | |
1877 | msp->ms_max_size = metaslab_block_maxsize(msp); | |
1878 | ||
1879 | /* | |
1880 | * Segment-based weighting requires space map histogram support. | |
1881 | */ | |
1882 | if (zfs_metaslab_segment_weight_enabled && | |
1883 | spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) && | |
1884 | (msp->ms_sm == NULL || msp->ms_sm->sm_dbuf->db_size == | |
1885 | sizeof (space_map_phys_t))) { | |
1886 | weight = metaslab_segment_weight(msp); | |
1887 | } else { | |
1888 | weight = metaslab_space_weight(msp); | |
1889 | } | |
93cf2076 | 1890 | return (weight); |
34dc7c2f BB |
1891 | } |
1892 | ||
1893 | static int | |
6d974228 | 1894 | metaslab_activate(metaslab_t *msp, uint64_t activation_weight) |
34dc7c2f | 1895 | { |
34dc7c2f BB |
1896 | ASSERT(MUTEX_HELD(&msp->ms_lock)); |
1897 | ||
1898 | if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) { | |
93cf2076 GW |
1899 | metaslab_load_wait(msp); |
1900 | if (!msp->ms_loaded) { | |
1901 | int error = metaslab_load(msp); | |
1902 | if (error) { | |
428870ff BB |
1903 | metaslab_group_sort(msp->ms_group, msp, 0); |
1904 | return (error); | |
1905 | } | |
34dc7c2f | 1906 | } |
9babb374 | 1907 | |
4e21fd06 | 1908 | msp->ms_activation_weight = msp->ms_weight; |
34dc7c2f BB |
1909 | metaslab_group_sort(msp->ms_group, msp, |
1910 | msp->ms_weight | activation_weight); | |
1911 | } | |
93cf2076 | 1912 | ASSERT(msp->ms_loaded); |
34dc7c2f BB |
1913 | ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK); |
1914 | ||
1915 | return (0); | |
1916 | } | |
1917 | ||
1918 | static void | |
4e21fd06 | 1919 | metaslab_passivate(metaslab_t *msp, uint64_t weight) |
34dc7c2f | 1920 | { |
4e21fd06 DB |
1921 | ASSERTV(uint64_t size = weight & ~METASLAB_WEIGHT_TYPE); |
1922 | ||
34dc7c2f BB |
1923 | /* |
1924 | * If size < SPA_MINBLOCKSIZE, then we will not allocate from | |
1925 | * this metaslab again. In that case, it had better be empty, | |
1926 | * or we would be leaving space on the table. | |
1927 | */ | |
4e21fd06 DB |
1928 | ASSERT(size >= SPA_MINBLOCKSIZE || |
1929 | range_tree_space(msp->ms_tree) == 0); | |
1930 | ASSERT0(weight & METASLAB_ACTIVE_MASK); | |
1931 | ||
1932 | msp->ms_activation_weight = 0; | |
1933 | metaslab_group_sort(msp->ms_group, msp, weight); | |
34dc7c2f BB |
1934 | ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0); |
1935 | } | |
1936 | ||
4e21fd06 DB |
1937 | /* |
1938 | * Segment-based metaslabs are activated once and remain active until | |
1939 | * we either fail an allocation attempt (similar to space-based metaslabs) | |
1940 | * or have exhausted the free space in zfs_metaslab_switch_threshold | |
1941 | * buckets since the metaslab was activated. This function checks to see | |
1942 | * if we've exhaused the zfs_metaslab_switch_threshold buckets in the | |
1943 | * metaslab and passivates it proactively. This will allow us to select a | |
1944 | * metaslab with a larger contiguous region, if any, remaining within this | |
1945 | * metaslab group. If we're in sync pass > 1, then we continue using this | |
1946 | * metaslab so that we don't dirty more block and cause more sync passes. | |
1947 | */ | |
1948 | void | |
1949 | metaslab_segment_may_passivate(metaslab_t *msp) | |
1950 | { | |
1951 | spa_t *spa = msp->ms_group->mg_vd->vdev_spa; | |
1952 | uint64_t weight; | |
1953 | int activation_idx, current_idx; | |
1954 | ||
1955 | if (WEIGHT_IS_SPACEBASED(msp->ms_weight) || spa_sync_pass(spa) > 1) | |
1956 | return; | |
1957 | ||
1958 | /* | |
1959 | * Since we are in the middle of a sync pass, the most accurate | |
1960 | * information that is accessible to us is the in-core range tree | |
1961 | * histogram; calculate the new weight based on that information. | |
1962 | */ | |
1963 | weight = metaslab_weight_from_range_tree(msp); | |
1964 | activation_idx = WEIGHT_GET_INDEX(msp->ms_activation_weight); | |
1965 | current_idx = WEIGHT_GET_INDEX(weight); | |
1966 | ||
1967 | if (current_idx <= activation_idx - zfs_metaslab_switch_threshold) | |
1968 | metaslab_passivate(msp, weight); | |
1969 | } | |
1970 | ||
93cf2076 GW |
1971 | static void |
1972 | metaslab_preload(void *arg) | |
1973 | { | |
1974 | metaslab_t *msp = arg; | |
1975 | spa_t *spa = msp->ms_group->mg_vd->vdev_spa; | |
1cd77734 | 1976 | fstrans_cookie_t cookie = spl_fstrans_mark(); |
93cf2076 | 1977 | |
080b3100 GW |
1978 | ASSERT(!MUTEX_HELD(&msp->ms_group->mg_lock)); |
1979 | ||
93cf2076 GW |
1980 | mutex_enter(&msp->ms_lock); |
1981 | metaslab_load_wait(msp); | |
1982 | if (!msp->ms_loaded) | |
1983 | (void) metaslab_load(msp); | |
4e21fd06 | 1984 | msp->ms_selected_txg = spa_syncing_txg(spa); |
93cf2076 | 1985 | mutex_exit(&msp->ms_lock); |
1cd77734 | 1986 | spl_fstrans_unmark(cookie); |
93cf2076 GW |
1987 | } |
1988 | ||
1989 | static void | |
1990 | metaslab_group_preload(metaslab_group_t *mg) | |
1991 | { | |
1992 | spa_t *spa = mg->mg_vd->vdev_spa; | |
1993 | metaslab_t *msp; | |
1994 | avl_tree_t *t = &mg->mg_metaslab_tree; | |
1995 | int m = 0; | |
1996 | ||
1997 | if (spa_shutting_down(spa) || !metaslab_preload_enabled) { | |
c5528b9b | 1998 | taskq_wait_outstanding(mg->mg_taskq, 0); |
93cf2076 GW |
1999 | return; |
2000 | } | |
93cf2076 | 2001 | |
080b3100 | 2002 | mutex_enter(&mg->mg_lock); |
93cf2076 | 2003 | /* |
080b3100 | 2004 | * Load the next potential metaslabs |
93cf2076 | 2005 | */ |
4e21fd06 | 2006 | for (msp = avl_first(t); msp != NULL; msp = AVL_NEXT(t, msp)) { |
f3a7f661 GW |
2007 | /* |
2008 | * We preload only the maximum number of metaslabs specified | |
2009 | * by metaslab_preload_limit. If a metaslab is being forced | |
2010 | * to condense then we preload it too. This will ensure | |
2011 | * that force condensing happens in the next txg. | |
2012 | */ | |
2013 | if (++m > metaslab_preload_limit && !msp->ms_condense_wanted) { | |
f3a7f661 GW |
2014 | continue; |
2015 | } | |
93cf2076 GW |
2016 | |
2017 | VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload, | |
48d3eb40 | 2018 | msp, TQ_SLEEP) != TASKQID_INVALID); |
93cf2076 GW |
2019 | } |
2020 | mutex_exit(&mg->mg_lock); | |
2021 | } | |
2022 | ||
e51be066 | 2023 | /* |
93cf2076 GW |
2024 | * Determine if the space map's on-disk footprint is past our tolerance |
2025 | * for inefficiency. We would like to use the following criteria to make | |
2026 | * our decision: | |
e51be066 GW |
2027 | * |
2028 | * 1. The size of the space map object should not dramatically increase as a | |
93cf2076 | 2029 | * result of writing out the free space range tree. |
e51be066 GW |
2030 | * |
2031 | * 2. The minimal on-disk space map representation is zfs_condense_pct/100 | |
93cf2076 GW |
2032 | * times the size than the free space range tree representation |
2033 | * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB). | |
e51be066 | 2034 | * |
b02fe35d AR |
2035 | * 3. The on-disk size of the space map should actually decrease. |
2036 | * | |
e51be066 GW |
2037 | * Checking the first condition is tricky since we don't want to walk |
2038 | * the entire AVL tree calculating the estimated on-disk size. Instead we | |
93cf2076 GW |
2039 | * use the size-ordered range tree in the metaslab and calculate the |
2040 | * size required to write out the largest segment in our free tree. If the | |
e51be066 GW |
2041 | * size required to represent that segment on disk is larger than the space |
2042 | * map object then we avoid condensing this map. | |
2043 | * | |
2044 | * To determine the second criterion we use a best-case estimate and assume | |
2045 | * each segment can be represented on-disk as a single 64-bit entry. We refer | |
2046 | * to this best-case estimate as the space map's minimal form. | |
b02fe35d AR |
2047 | * |
2048 | * Unfortunately, we cannot compute the on-disk size of the space map in this | |
2049 | * context because we cannot accurately compute the effects of compression, etc. | |
2050 | * Instead, we apply the heuristic described in the block comment for | |
2051 | * zfs_metaslab_condense_block_threshold - we only condense if the space used | |
2052 | * is greater than a threshold number of blocks. | |
e51be066 GW |
2053 | */ |
2054 | static boolean_t | |
2055 | metaslab_should_condense(metaslab_t *msp) | |
2056 | { | |
93cf2076 GW |
2057 | space_map_t *sm = msp->ms_sm; |
2058 | range_seg_t *rs; | |
b02fe35d AR |
2059 | uint64_t size, entries, segsz, object_size, optimal_size, record_size; |
2060 | dmu_object_info_t doi; | |
f4bae2ed | 2061 | uint64_t vdev_blocksize = 1ULL << msp->ms_group->mg_vd->vdev_ashift; |
e51be066 GW |
2062 | |
2063 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
93cf2076 | 2064 | ASSERT(msp->ms_loaded); |
e51be066 GW |
2065 | |
2066 | /* | |
93cf2076 | 2067 | * Use the ms_size_tree range tree, which is ordered by size, to |
f3a7f661 GW |
2068 | * obtain the largest segment in the free tree. We always condense |
2069 | * metaslabs that are empty and metaslabs for which a condense | |
2070 | * request has been made. | |
e51be066 | 2071 | */ |
93cf2076 | 2072 | rs = avl_last(&msp->ms_size_tree); |
f3a7f661 | 2073 | if (rs == NULL || msp->ms_condense_wanted) |
e51be066 GW |
2074 | return (B_TRUE); |
2075 | ||
2076 | /* | |
2077 | * Calculate the number of 64-bit entries this segment would | |
2078 | * require when written to disk. If this single segment would be | |
2079 | * larger on-disk than the entire current on-disk structure, then | |
2080 | * clearly condensing will increase the on-disk structure size. | |
2081 | */ | |
93cf2076 | 2082 | size = (rs->rs_end - rs->rs_start) >> sm->sm_shift; |
e51be066 GW |
2083 | entries = size / (MIN(size, SM_RUN_MAX)); |
2084 | segsz = entries * sizeof (uint64_t); | |
2085 | ||
b02fe35d AR |
2086 | optimal_size = sizeof (uint64_t) * avl_numnodes(&msp->ms_tree->rt_root); |
2087 | object_size = space_map_length(msp->ms_sm); | |
2088 | ||
2089 | dmu_object_info_from_db(sm->sm_dbuf, &doi); | |
2090 | record_size = MAX(doi.doi_data_block_size, vdev_blocksize); | |
2091 | ||
2092 | return (segsz <= object_size && | |
2093 | object_size >= (optimal_size * zfs_condense_pct / 100) && | |
2094 | object_size > zfs_metaslab_condense_block_threshold * record_size); | |
e51be066 GW |
2095 | } |
2096 | ||
2097 | /* | |
2098 | * Condense the on-disk space map representation to its minimized form. | |
2099 | * The minimized form consists of a small number of allocations followed by | |
93cf2076 | 2100 | * the entries of the free range tree. |
e51be066 GW |
2101 | */ |
2102 | static void | |
2103 | metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx) | |
2104 | { | |
2105 | spa_t *spa = msp->ms_group->mg_vd->vdev_spa; | |
93cf2076 GW |
2106 | range_tree_t *freetree = msp->ms_freetree[txg & TXG_MASK]; |
2107 | range_tree_t *condense_tree; | |
2108 | space_map_t *sm = msp->ms_sm; | |
e51be066 GW |
2109 | int t; |
2110 | ||
2111 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
2112 | ASSERT3U(spa_sync_pass(spa), ==, 1); | |
93cf2076 | 2113 | ASSERT(msp->ms_loaded); |
e51be066 | 2114 | |
f3a7f661 | 2115 | |
5f3d9c69 JS |
2116 | spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, vdev id %llu, " |
2117 | "spa %s, smp size %llu, segments %lu, forcing condense=%s", txg, | |
2118 | msp->ms_id, msp, msp->ms_group->mg_vd->vdev_id, | |
2119 | msp->ms_group->mg_vd->vdev_spa->spa_name, | |
2120 | space_map_length(msp->ms_sm), avl_numnodes(&msp->ms_tree->rt_root), | |
f3a7f661 GW |
2121 | msp->ms_condense_wanted ? "TRUE" : "FALSE"); |
2122 | ||
2123 | msp->ms_condense_wanted = B_FALSE; | |
e51be066 GW |
2124 | |
2125 | /* | |
93cf2076 | 2126 | * Create an range tree that is 100% allocated. We remove segments |
e51be066 GW |
2127 | * that have been freed in this txg, any deferred frees that exist, |
2128 | * and any allocation in the future. Removing segments should be | |
93cf2076 GW |
2129 | * a relatively inexpensive operation since we expect these trees to |
2130 | * have a small number of nodes. | |
e51be066 | 2131 | */ |
93cf2076 GW |
2132 | condense_tree = range_tree_create(NULL, NULL, &msp->ms_lock); |
2133 | range_tree_add(condense_tree, msp->ms_start, msp->ms_size); | |
e51be066 GW |
2134 | |
2135 | /* | |
93cf2076 | 2136 | * Remove what's been freed in this txg from the condense_tree. |
e51be066 | 2137 | * Since we're in sync_pass 1, we know that all the frees from |
93cf2076 | 2138 | * this txg are in the freetree. |
e51be066 | 2139 | */ |
93cf2076 | 2140 | range_tree_walk(freetree, range_tree_remove, condense_tree); |
e51be066 | 2141 | |
93cf2076 GW |
2142 | for (t = 0; t < TXG_DEFER_SIZE; t++) { |
2143 | range_tree_walk(msp->ms_defertree[t], | |
2144 | range_tree_remove, condense_tree); | |
2145 | } | |
e51be066 | 2146 | |
93cf2076 GW |
2147 | for (t = 1; t < TXG_CONCURRENT_STATES; t++) { |
2148 | range_tree_walk(msp->ms_alloctree[(txg + t) & TXG_MASK], | |
2149 | range_tree_remove, condense_tree); | |
2150 | } | |
e51be066 GW |
2151 | |
2152 | /* | |
2153 | * We're about to drop the metaslab's lock thus allowing | |
2154 | * other consumers to change it's content. Set the | |
93cf2076 | 2155 | * metaslab's ms_condensing flag to ensure that |
e51be066 GW |
2156 | * allocations on this metaslab do not occur while we're |
2157 | * in the middle of committing it to disk. This is only critical | |
93cf2076 | 2158 | * for the ms_tree as all other range trees use per txg |
e51be066 GW |
2159 | * views of their content. |
2160 | */ | |
93cf2076 | 2161 | msp->ms_condensing = B_TRUE; |
e51be066 GW |
2162 | |
2163 | mutex_exit(&msp->ms_lock); | |
93cf2076 | 2164 | space_map_truncate(sm, tx); |
e51be066 GW |
2165 | mutex_enter(&msp->ms_lock); |
2166 | ||
2167 | /* | |
4e21fd06 | 2168 | * While we would ideally like to create a space map representation |
e51be066 | 2169 | * that consists only of allocation records, doing so can be |
93cf2076 | 2170 | * prohibitively expensive because the in-core free tree can be |
e51be066 | 2171 | * large, and therefore computationally expensive to subtract |
93cf2076 GW |
2172 | * from the condense_tree. Instead we sync out two trees, a cheap |
2173 | * allocation only tree followed by the in-core free tree. While not | |
e51be066 GW |
2174 | * optimal, this is typically close to optimal, and much cheaper to |
2175 | * compute. | |
2176 | */ | |
93cf2076 GW |
2177 | space_map_write(sm, condense_tree, SM_ALLOC, tx); |
2178 | range_tree_vacate(condense_tree, NULL, NULL); | |
2179 | range_tree_destroy(condense_tree); | |
e51be066 | 2180 | |
93cf2076 GW |
2181 | space_map_write(sm, msp->ms_tree, SM_FREE, tx); |
2182 | msp->ms_condensing = B_FALSE; | |
e51be066 GW |
2183 | } |
2184 | ||
34dc7c2f BB |
2185 | /* |
2186 | * Write a metaslab to disk in the context of the specified transaction group. | |
2187 | */ | |
2188 | void | |
2189 | metaslab_sync(metaslab_t *msp, uint64_t txg) | |
2190 | { | |
93cf2076 GW |
2191 | metaslab_group_t *mg = msp->ms_group; |
2192 | vdev_t *vd = mg->mg_vd; | |
34dc7c2f | 2193 | spa_t *spa = vd->vdev_spa; |
428870ff | 2194 | objset_t *mos = spa_meta_objset(spa); |
93cf2076 GW |
2195 | range_tree_t *alloctree = msp->ms_alloctree[txg & TXG_MASK]; |
2196 | range_tree_t **freetree = &msp->ms_freetree[txg & TXG_MASK]; | |
2197 | range_tree_t **freed_tree = | |
2198 | &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK]; | |
34dc7c2f | 2199 | dmu_tx_t *tx; |
93cf2076 | 2200 | uint64_t object = space_map_object(msp->ms_sm); |
34dc7c2f | 2201 | |
428870ff BB |
2202 | ASSERT(!vd->vdev_ishole); |
2203 | ||
e51be066 GW |
2204 | /* |
2205 | * This metaslab has just been added so there's no work to do now. | |
2206 | */ | |
93cf2076 GW |
2207 | if (*freetree == NULL) { |
2208 | ASSERT3P(alloctree, ==, NULL); | |
e51be066 GW |
2209 | return; |
2210 | } | |
2211 | ||
93cf2076 GW |
2212 | ASSERT3P(alloctree, !=, NULL); |
2213 | ASSERT3P(*freetree, !=, NULL); | |
2214 | ASSERT3P(*freed_tree, !=, NULL); | |
e51be066 | 2215 | |
f3a7f661 GW |
2216 | /* |
2217 | * Normally, we don't want to process a metaslab if there | |
2218 | * are no allocations or frees to perform. However, if the metaslab | |
2219 | * is being forced to condense we need to let it through. | |
2220 | */ | |
93cf2076 | 2221 | if (range_tree_space(alloctree) == 0 && |
f3a7f661 GW |
2222 | range_tree_space(*freetree) == 0 && |
2223 | !msp->ms_condense_wanted) | |
428870ff | 2224 | return; |
34dc7c2f BB |
2225 | |
2226 | /* | |
2227 | * The only state that can actually be changing concurrently with | |
93cf2076 GW |
2228 | * metaslab_sync() is the metaslab's ms_tree. No other thread can |
2229 | * be modifying this txg's alloctree, freetree, freed_tree, or | |
2230 | * space_map_phys_t. Therefore, we only hold ms_lock to satify | |
4e21fd06 | 2231 | * space map ASSERTs. We drop it whenever we call into the DMU, |
93cf2076 GW |
2232 | * because the DMU can call down to us (e.g. via zio_free()) at |
2233 | * any time. | |
34dc7c2f | 2234 | */ |
428870ff BB |
2235 | |
2236 | tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg); | |
34dc7c2f | 2237 | |
93cf2076 GW |
2238 | if (msp->ms_sm == NULL) { |
2239 | uint64_t new_object; | |
2240 | ||
2241 | new_object = space_map_alloc(mos, tx); | |
2242 | VERIFY3U(new_object, !=, 0); | |
2243 | ||
2244 | VERIFY0(space_map_open(&msp->ms_sm, mos, new_object, | |
2245 | msp->ms_start, msp->ms_size, vd->vdev_ashift, | |
2246 | &msp->ms_lock)); | |
2247 | ASSERT(msp->ms_sm != NULL); | |
34dc7c2f BB |
2248 | } |
2249 | ||
428870ff BB |
2250 | mutex_enter(&msp->ms_lock); |
2251 | ||
96358617 | 2252 | /* |
4e21fd06 DB |
2253 | * Note: metaslab_condense() clears the space map's histogram. |
2254 | * Therefore we must verify and remove this histogram before | |
96358617 MA |
2255 | * condensing. |
2256 | */ | |
2257 | metaslab_group_histogram_verify(mg); | |
2258 | metaslab_class_histogram_verify(mg->mg_class); | |
2259 | metaslab_group_histogram_remove(mg, msp); | |
2260 | ||
93cf2076 | 2261 | if (msp->ms_loaded && spa_sync_pass(spa) == 1 && |
e51be066 GW |
2262 | metaslab_should_condense(msp)) { |
2263 | metaslab_condense(msp, txg, tx); | |
2264 | } else { | |
93cf2076 GW |
2265 | space_map_write(msp->ms_sm, alloctree, SM_ALLOC, tx); |
2266 | space_map_write(msp->ms_sm, *freetree, SM_FREE, tx); | |
e51be066 | 2267 | } |
428870ff | 2268 | |
93cf2076 | 2269 | if (msp->ms_loaded) { |
4e21fd06 DB |
2270 | int t; |
2271 | ||
93cf2076 GW |
2272 | /* |
2273 | * When the space map is loaded, we have an accruate | |
2274 | * histogram in the range tree. This gives us an opportunity | |
2275 | * to bring the space map's histogram up-to-date so we clear | |
2276 | * it first before updating it. | |
2277 | */ | |
2278 | space_map_histogram_clear(msp->ms_sm); | |
2279 | space_map_histogram_add(msp->ms_sm, msp->ms_tree, tx); | |
4e21fd06 DB |
2280 | |
2281 | /* | |
2282 | * Since we've cleared the histogram we need to add back | |
2283 | * any free space that has already been processed, plus | |
2284 | * any deferred space. This allows the on-disk histogram | |
2285 | * to accurately reflect all free space even if some space | |
2286 | * is not yet available for allocation (i.e. deferred). | |
2287 | */ | |
2288 | space_map_histogram_add(msp->ms_sm, *freed_tree, tx); | |
2289 | ||
93cf2076 | 2290 | /* |
4e21fd06 DB |
2291 | * Add back any deferred free space that has not been |
2292 | * added back into the in-core free tree yet. This will | |
2293 | * ensure that we don't end up with a space map histogram | |
2294 | * that is completely empty unless the metaslab is fully | |
2295 | * allocated. | |
93cf2076 | 2296 | */ |
4e21fd06 DB |
2297 | for (t = 0; t < TXG_DEFER_SIZE; t++) { |
2298 | space_map_histogram_add(msp->ms_sm, | |
2299 | msp->ms_defertree[t], tx); | |
2300 | } | |
93cf2076 | 2301 | } |
4e21fd06 DB |
2302 | |
2303 | /* | |
2304 | * Always add the free space from this sync pass to the space | |
2305 | * map histogram. We want to make sure that the on-disk histogram | |
2306 | * accounts for all free space. If the space map is not loaded, | |
2307 | * then we will lose some accuracy but will correct it the next | |
2308 | * time we load the space map. | |
2309 | */ | |
2310 | space_map_histogram_add(msp->ms_sm, *freetree, tx); | |
2311 | ||
f3a7f661 GW |
2312 | metaslab_group_histogram_add(mg, msp); |
2313 | metaslab_group_histogram_verify(mg); | |
2314 | metaslab_class_histogram_verify(mg->mg_class); | |
34dc7c2f | 2315 | |
e51be066 | 2316 | /* |
93cf2076 GW |
2317 | * For sync pass 1, we avoid traversing this txg's free range tree |
2318 | * and instead will just swap the pointers for freetree and | |
2319 | * freed_tree. We can safely do this since the freed_tree is | |
e51be066 GW |
2320 | * guaranteed to be empty on the initial pass. |
2321 | */ | |
2322 | if (spa_sync_pass(spa) == 1) { | |
93cf2076 | 2323 | range_tree_swap(freetree, freed_tree); |
e51be066 | 2324 | } else { |
93cf2076 | 2325 | range_tree_vacate(*freetree, range_tree_add, *freed_tree); |
34dc7c2f | 2326 | } |
f3a7f661 | 2327 | range_tree_vacate(alloctree, NULL, NULL); |
34dc7c2f | 2328 | |
93cf2076 | 2329 | ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK])); |
4e21fd06 | 2330 | ASSERT0(range_tree_space(msp->ms_alloctree[TXG_CLEAN(txg) & TXG_MASK])); |
93cf2076 | 2331 | ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK])); |
34dc7c2f BB |
2332 | |
2333 | mutex_exit(&msp->ms_lock); | |
2334 | ||
93cf2076 GW |
2335 | if (object != space_map_object(msp->ms_sm)) { |
2336 | object = space_map_object(msp->ms_sm); | |
2337 | dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) * | |
2338 | msp->ms_id, sizeof (uint64_t), &object, tx); | |
2339 | } | |
34dc7c2f BB |
2340 | dmu_tx_commit(tx); |
2341 | } | |
2342 | ||
2343 | /* | |
2344 | * Called after a transaction group has completely synced to mark | |
2345 | * all of the metaslab's free space as usable. | |
2346 | */ | |
2347 | void | |
2348 | metaslab_sync_done(metaslab_t *msp, uint64_t txg) | |
2349 | { | |
34dc7c2f BB |
2350 | metaslab_group_t *mg = msp->ms_group; |
2351 | vdev_t *vd = mg->mg_vd; | |
4e21fd06 | 2352 | spa_t *spa = vd->vdev_spa; |
93cf2076 GW |
2353 | range_tree_t **freed_tree; |
2354 | range_tree_t **defer_tree; | |
428870ff | 2355 | int64_t alloc_delta, defer_delta; |
4e21fd06 DB |
2356 | uint64_t free_space; |
2357 | boolean_t defer_allowed = B_TRUE; | |
d6320ddb | 2358 | int t; |
428870ff BB |
2359 | |
2360 | ASSERT(!vd->vdev_ishole); | |
34dc7c2f BB |
2361 | |
2362 | mutex_enter(&msp->ms_lock); | |
2363 | ||
2364 | /* | |
2365 | * If this metaslab is just becoming available, initialize its | |
93cf2076 GW |
2366 | * alloctrees, freetrees, and defertree and add its capacity to |
2367 | * the vdev. | |
34dc7c2f | 2368 | */ |
93cf2076 | 2369 | if (msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK] == NULL) { |
d6320ddb | 2370 | for (t = 0; t < TXG_SIZE; t++) { |
93cf2076 GW |
2371 | ASSERT(msp->ms_alloctree[t] == NULL); |
2372 | ASSERT(msp->ms_freetree[t] == NULL); | |
2373 | ||
2374 | msp->ms_alloctree[t] = range_tree_create(NULL, msp, | |
2375 | &msp->ms_lock); | |
2376 | msp->ms_freetree[t] = range_tree_create(NULL, msp, | |
2377 | &msp->ms_lock); | |
34dc7c2f | 2378 | } |
428870ff | 2379 | |
e51be066 | 2380 | for (t = 0; t < TXG_DEFER_SIZE; t++) { |
93cf2076 | 2381 | ASSERT(msp->ms_defertree[t] == NULL); |
e51be066 | 2382 | |
93cf2076 GW |
2383 | msp->ms_defertree[t] = range_tree_create(NULL, msp, |
2384 | &msp->ms_lock); | |
2385 | } | |
428870ff | 2386 | |
93cf2076 | 2387 | vdev_space_update(vd, 0, 0, msp->ms_size); |
34dc7c2f BB |
2388 | } |
2389 | ||
93cf2076 GW |
2390 | freed_tree = &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK]; |
2391 | defer_tree = &msp->ms_defertree[txg % TXG_DEFER_SIZE]; | |
2392 | ||
4e21fd06 DB |
2393 | free_space = metaslab_class_get_space(spa_normal_class(spa)) - |
2394 | metaslab_class_get_alloc(spa_normal_class(spa)); | |
2395 | if (free_space <= spa_get_slop_space(spa)) { | |
2396 | defer_allowed = B_FALSE; | |
2397 | } | |
2398 | ||
2399 | defer_delta = 0; | |
93cf2076 | 2400 | alloc_delta = space_map_alloc_delta(msp->ms_sm); |
4e21fd06 DB |
2401 | if (defer_allowed) { |
2402 | defer_delta = range_tree_space(*freed_tree) - | |
2403 | range_tree_space(*defer_tree); | |
2404 | } else { | |
2405 | defer_delta -= range_tree_space(*defer_tree); | |
2406 | } | |
428870ff BB |
2407 | |
2408 | vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0); | |
34dc7c2f | 2409 | |
93cf2076 GW |
2410 | ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK])); |
2411 | ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK])); | |
34dc7c2f BB |
2412 | |
2413 | /* | |
93cf2076 | 2414 | * If there's a metaslab_load() in progress, wait for it to complete |
34dc7c2f | 2415 | * so that we have a consistent view of the in-core space map. |
34dc7c2f | 2416 | */ |
93cf2076 | 2417 | metaslab_load_wait(msp); |
c2e42f9d GW |
2418 | |
2419 | /* | |
93cf2076 GW |
2420 | * Move the frees from the defer_tree back to the free |
2421 | * range tree (if it's loaded). Swap the freed_tree and the | |
2422 | * defer_tree -- this is safe to do because we've just emptied out | |
2423 | * the defer_tree. | |
c2e42f9d | 2424 | */ |
93cf2076 GW |
2425 | range_tree_vacate(*defer_tree, |
2426 | msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree); | |
4e21fd06 DB |
2427 | if (defer_allowed) { |
2428 | range_tree_swap(freed_tree, defer_tree); | |
2429 | } else { | |
2430 | range_tree_vacate(*freed_tree, | |
2431 | msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree); | |
2432 | } | |
34dc7c2f | 2433 | |
93cf2076 | 2434 | space_map_update(msp->ms_sm); |
34dc7c2f | 2435 | |
428870ff BB |
2436 | msp->ms_deferspace += defer_delta; |
2437 | ASSERT3S(msp->ms_deferspace, >=, 0); | |
93cf2076 | 2438 | ASSERT3S(msp->ms_deferspace, <=, msp->ms_size); |
428870ff BB |
2439 | if (msp->ms_deferspace != 0) { |
2440 | /* | |
2441 | * Keep syncing this metaslab until all deferred frees | |
2442 | * are back in circulation. | |
2443 | */ | |
2444 | vdev_dirty(vd, VDD_METASLAB, msp, txg + 1); | |
2445 | } | |
2446 | ||
4e21fd06 DB |
2447 | /* |
2448 | * Calculate the new weights before unloading any metaslabs. | |
2449 | * This will give us the most accurate weighting. | |
2450 | */ | |
2451 | metaslab_group_sort(mg, msp, metaslab_weight(msp)); | |
2452 | ||
2453 | /* | |
2454 | * If the metaslab is loaded and we've not tried to load or allocate | |
2455 | * from it in 'metaslab_unload_delay' txgs, then unload it. | |
2456 | */ | |
2457 | if (msp->ms_loaded && | |
2458 | msp->ms_selected_txg + metaslab_unload_delay < txg) { | |
2459 | ||
93cf2076 GW |
2460 | for (t = 1; t < TXG_CONCURRENT_STATES; t++) { |
2461 | VERIFY0(range_tree_space( | |
2462 | msp->ms_alloctree[(txg + t) & TXG_MASK])); | |
2463 | } | |
34dc7c2f | 2464 | |
93cf2076 GW |
2465 | if (!metaslab_debug_unload) |
2466 | metaslab_unload(msp); | |
34dc7c2f BB |
2467 | } |
2468 | ||
34dc7c2f BB |
2469 | mutex_exit(&msp->ms_lock); |
2470 | } | |
2471 | ||
428870ff BB |
2472 | void |
2473 | metaslab_sync_reassess(metaslab_group_t *mg) | |
2474 | { | |
1be627f5 | 2475 | metaslab_group_alloc_update(mg); |
f3a7f661 | 2476 | mg->mg_fragmentation = metaslab_group_fragmentation(mg); |
6d974228 | 2477 | |
428870ff | 2478 | /* |
93cf2076 | 2479 | * Preload the next potential metaslabs |
428870ff | 2480 | */ |
93cf2076 | 2481 | metaslab_group_preload(mg); |
428870ff BB |
2482 | } |
2483 | ||
34dc7c2f BB |
2484 | static uint64_t |
2485 | metaslab_distance(metaslab_t *msp, dva_t *dva) | |
2486 | { | |
2487 | uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift; | |
2488 | uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift; | |
93cf2076 | 2489 | uint64_t start = msp->ms_id; |
34dc7c2f BB |
2490 | |
2491 | if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva)) | |
2492 | return (1ULL << 63); | |
2493 | ||
2494 | if (offset < start) | |
2495 | return ((start - offset) << ms_shift); | |
2496 | if (offset > start) | |
2497 | return ((offset - start) << ms_shift); | |
2498 | return (0); | |
2499 | } | |
2500 | ||
4e21fd06 DB |
2501 | /* |
2502 | * ========================================================================== | |
2503 | * Metaslab allocation tracing facility | |
2504 | * ========================================================================== | |
2505 | */ | |
2506 | #ifdef _METASLAB_TRACING | |
2507 | kstat_t *metaslab_trace_ksp; | |
2508 | kstat_named_t metaslab_trace_over_limit; | |
2509 | ||
2510 | void | |
2511 | metaslab_alloc_trace_init(void) | |
2512 | { | |
2513 | ASSERT(metaslab_alloc_trace_cache == NULL); | |
2514 | metaslab_alloc_trace_cache = kmem_cache_create( | |
2515 | "metaslab_alloc_trace_cache", sizeof (metaslab_alloc_trace_t), | |
2516 | 0, NULL, NULL, NULL, NULL, NULL, 0); | |
2517 | metaslab_trace_ksp = kstat_create("zfs", 0, "metaslab_trace_stats", | |
2518 | "misc", KSTAT_TYPE_NAMED, 1, KSTAT_FLAG_VIRTUAL); | |
2519 | if (metaslab_trace_ksp != NULL) { | |
2520 | metaslab_trace_ksp->ks_data = &metaslab_trace_over_limit; | |
2521 | kstat_named_init(&metaslab_trace_over_limit, | |
2522 | "metaslab_trace_over_limit", KSTAT_DATA_UINT64); | |
2523 | kstat_install(metaslab_trace_ksp); | |
2524 | } | |
2525 | } | |
2526 | ||
2527 | void | |
2528 | metaslab_alloc_trace_fini(void) | |
2529 | { | |
2530 | if (metaslab_trace_ksp != NULL) { | |
2531 | kstat_delete(metaslab_trace_ksp); | |
2532 | metaslab_trace_ksp = NULL; | |
2533 | } | |
2534 | kmem_cache_destroy(metaslab_alloc_trace_cache); | |
2535 | metaslab_alloc_trace_cache = NULL; | |
2536 | } | |
2537 | ||
2538 | /* | |
2539 | * Add an allocation trace element to the allocation tracing list. | |
2540 | */ | |
2541 | static void | |
2542 | metaslab_trace_add(zio_alloc_list_t *zal, metaslab_group_t *mg, | |
2543 | metaslab_t *msp, uint64_t psize, uint32_t dva_id, uint64_t offset) | |
2544 | { | |
2545 | metaslab_alloc_trace_t *mat; | |
2546 | ||
2547 | if (!metaslab_trace_enabled) | |
2548 | return; | |
2549 | ||
2550 | /* | |
2551 | * When the tracing list reaches its maximum we remove | |
2552 | * the second element in the list before adding a new one. | |
2553 | * By removing the second element we preserve the original | |
2554 | * entry as a clue to what allocations steps have already been | |
2555 | * performed. | |
2556 | */ | |
2557 | if (zal->zal_size == metaslab_trace_max_entries) { | |
2558 | metaslab_alloc_trace_t *mat_next; | |
2559 | #ifdef DEBUG | |
2560 | panic("too many entries in allocation list"); | |
2561 | #endif | |
2562 | atomic_inc_64(&metaslab_trace_over_limit.value.ui64); | |
2563 | zal->zal_size--; | |
2564 | mat_next = list_next(&zal->zal_list, list_head(&zal->zal_list)); | |
2565 | list_remove(&zal->zal_list, mat_next); | |
2566 | kmem_cache_free(metaslab_alloc_trace_cache, mat_next); | |
2567 | } | |
2568 | ||
2569 | mat = kmem_cache_alloc(metaslab_alloc_trace_cache, KM_SLEEP); | |
2570 | list_link_init(&mat->mat_list_node); | |
2571 | mat->mat_mg = mg; | |
2572 | mat->mat_msp = msp; | |
2573 | mat->mat_size = psize; | |
2574 | mat->mat_dva_id = dva_id; | |
2575 | mat->mat_offset = offset; | |
2576 | mat->mat_weight = 0; | |
2577 | ||
2578 | if (msp != NULL) | |
2579 | mat->mat_weight = msp->ms_weight; | |
2580 | ||
2581 | /* | |
2582 | * The list is part of the zio so locking is not required. Only | |
2583 | * a single thread will perform allocations for a given zio. | |
2584 | */ | |
2585 | list_insert_tail(&zal->zal_list, mat); | |
2586 | zal->zal_size++; | |
2587 | ||
2588 | ASSERT3U(zal->zal_size, <=, metaslab_trace_max_entries); | |
2589 | } | |
2590 | ||
2591 | void | |
2592 | metaslab_trace_init(zio_alloc_list_t *zal) | |
2593 | { | |
2594 | list_create(&zal->zal_list, sizeof (metaslab_alloc_trace_t), | |
2595 | offsetof(metaslab_alloc_trace_t, mat_list_node)); | |
2596 | zal->zal_size = 0; | |
2597 | } | |
2598 | ||
2599 | void | |
2600 | metaslab_trace_fini(zio_alloc_list_t *zal) | |
2601 | { | |
2602 | metaslab_alloc_trace_t *mat; | |
2603 | ||
2604 | while ((mat = list_remove_head(&zal->zal_list)) != NULL) | |
2605 | kmem_cache_free(metaslab_alloc_trace_cache, mat); | |
2606 | list_destroy(&zal->zal_list); | |
2607 | zal->zal_size = 0; | |
2608 | } | |
2609 | #else | |
2610 | ||
2611 | #define metaslab_trace_add(zal, mg, msp, psize, id, off) | |
2612 | ||
2613 | void | |
2614 | metaslab_alloc_trace_init(void) | |
2615 | { | |
2616 | } | |
2617 | ||
2618 | void | |
2619 | metaslab_alloc_trace_fini(void) | |
2620 | { | |
2621 | } | |
2622 | ||
2623 | void | |
2624 | metaslab_trace_init(zio_alloc_list_t *zal) | |
2625 | { | |
2626 | } | |
2627 | ||
2628 | void | |
2629 | metaslab_trace_fini(zio_alloc_list_t *zal) | |
2630 | { | |
2631 | } | |
2632 | ||
2633 | #endif /* _METASLAB_TRACING */ | |
2634 | ||
3dfb57a3 DB |
2635 | /* |
2636 | * ========================================================================== | |
2637 | * Metaslab block operations | |
2638 | * ========================================================================== | |
2639 | */ | |
2640 | ||
2641 | static void | |
2642 | metaslab_group_alloc_increment(spa_t *spa, uint64_t vdev, void *tag, int flags) | |
2643 | { | |
2644 | metaslab_group_t *mg; | |
2645 | ||
2646 | if (!(flags & METASLAB_ASYNC_ALLOC) || | |
2647 | flags & METASLAB_DONT_THROTTLE) | |
2648 | return; | |
2649 | ||
2650 | mg = vdev_lookup_top(spa, vdev)->vdev_mg; | |
2651 | if (!mg->mg_class->mc_alloc_throttle_enabled) | |
2652 | return; | |
2653 | ||
2654 | (void) refcount_add(&mg->mg_alloc_queue_depth, tag); | |
2655 | } | |
2656 | ||
2657 | void | |
2658 | metaslab_group_alloc_decrement(spa_t *spa, uint64_t vdev, void *tag, int flags) | |
2659 | { | |
2660 | metaslab_group_t *mg; | |
2661 | ||
2662 | if (!(flags & METASLAB_ASYNC_ALLOC) || | |
2663 | flags & METASLAB_DONT_THROTTLE) | |
2664 | return; | |
2665 | ||
2666 | mg = vdev_lookup_top(spa, vdev)->vdev_mg; | |
2667 | if (!mg->mg_class->mc_alloc_throttle_enabled) | |
2668 | return; | |
2669 | ||
2670 | (void) refcount_remove(&mg->mg_alloc_queue_depth, tag); | |
2671 | } | |
2672 | ||
2673 | void | |
2674 | metaslab_group_alloc_verify(spa_t *spa, const blkptr_t *bp, void *tag) | |
2675 | { | |
2676 | #ifdef ZFS_DEBUG | |
2677 | const dva_t *dva = bp->blk_dva; | |
2678 | int ndvas = BP_GET_NDVAS(bp); | |
2679 | int d; | |
2680 | ||
2681 | for (d = 0; d < ndvas; d++) { | |
2682 | uint64_t vdev = DVA_GET_VDEV(&dva[d]); | |
2683 | metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg; | |
2684 | VERIFY(refcount_not_held(&mg->mg_alloc_queue_depth, tag)); | |
2685 | } | |
2686 | #endif | |
2687 | } | |
2688 | ||
34dc7c2f | 2689 | static uint64_t |
4e21fd06 DB |
2690 | metaslab_block_alloc(metaslab_t *msp, uint64_t size, uint64_t txg) |
2691 | { | |
2692 | uint64_t start; | |
2693 | range_tree_t *rt = msp->ms_tree; | |
2694 | metaslab_class_t *mc = msp->ms_group->mg_class; | |
2695 | ||
2696 | VERIFY(!msp->ms_condensing); | |
2697 | ||
2698 | start = mc->mc_ops->msop_alloc(msp, size); | |
2699 | if (start != -1ULL) { | |
2700 | metaslab_group_t *mg = msp->ms_group; | |
2701 | vdev_t *vd = mg->mg_vd; | |
2702 | ||
2703 | VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift)); | |
2704 | VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift)); | |
2705 | VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size); | |
2706 | range_tree_remove(rt, start, size); | |
2707 | ||
2708 | if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0) | |
2709 | vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg); | |
2710 | ||
2711 | range_tree_add(msp->ms_alloctree[txg & TXG_MASK], start, size); | |
2712 | ||
2713 | /* Track the last successful allocation */ | |
2714 | msp->ms_alloc_txg = txg; | |
2715 | metaslab_verify_space(msp, txg); | |
2716 | } | |
2717 | ||
2718 | /* | |
2719 | * Now that we've attempted the allocation we need to update the | |
2720 | * metaslab's maximum block size since it may have changed. | |
2721 | */ | |
2722 | msp->ms_max_size = metaslab_block_maxsize(msp); | |
2723 | return (start); | |
2724 | } | |
2725 | ||
2726 | static uint64_t | |
2727 | metaslab_group_alloc_normal(metaslab_group_t *mg, zio_alloc_list_t *zal, | |
2728 | uint64_t asize, uint64_t txg, uint64_t min_distance, dva_t *dva, int d) | |
34dc7c2f BB |
2729 | { |
2730 | metaslab_t *msp = NULL; | |
4e21fd06 | 2731 | metaslab_t *search; |
34dc7c2f | 2732 | uint64_t offset = -1ULL; |
34dc7c2f BB |
2733 | uint64_t activation_weight; |
2734 | uint64_t target_distance; | |
2735 | int i; | |
2736 | ||
2737 | activation_weight = METASLAB_WEIGHT_PRIMARY; | |
9babb374 BB |
2738 | for (i = 0; i < d; i++) { |
2739 | if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) { | |
34dc7c2f | 2740 | activation_weight = METASLAB_WEIGHT_SECONDARY; |
9babb374 BB |
2741 | break; |
2742 | } | |
2743 | } | |
34dc7c2f | 2744 | |
4e21fd06 DB |
2745 | search = kmem_alloc(sizeof (*search), KM_SLEEP); |
2746 | search->ms_weight = UINT64_MAX; | |
2747 | search->ms_start = 0; | |
34dc7c2f | 2748 | for (;;) { |
9babb374 | 2749 | boolean_t was_active; |
4e21fd06 DB |
2750 | avl_tree_t *t = &mg->mg_metaslab_tree; |
2751 | avl_index_t idx; | |
9babb374 | 2752 | |
34dc7c2f | 2753 | mutex_enter(&mg->mg_lock); |
4e21fd06 DB |
2754 | |
2755 | /* | |
2756 | * Find the metaslab with the highest weight that is less | |
2757 | * than what we've already tried. In the common case, this | |
2758 | * means that we will examine each metaslab at most once. | |
2759 | * Note that concurrent callers could reorder metaslabs | |
2760 | * by activation/passivation once we have dropped the mg_lock. | |
2761 | * If a metaslab is activated by another thread, and we fail | |
2762 | * to allocate from the metaslab we have selected, we may | |
2763 | * not try the newly-activated metaslab, and instead activate | |
2764 | * another metaslab. This is not optimal, but generally | |
2765 | * does not cause any problems (a possible exception being | |
2766 | * if every metaslab is completely full except for the | |
2767 | * the newly-activated metaslab which we fail to examine). | |
2768 | */ | |
2769 | msp = avl_find(t, search, &idx); | |
2770 | if (msp == NULL) | |
2771 | msp = avl_nearest(t, idx, AVL_AFTER); | |
2772 | for (; msp != NULL; msp = AVL_NEXT(t, msp)) { | |
2773 | ||
2774 | if (!metaslab_should_allocate(msp, asize)) { | |
2775 | metaslab_trace_add(zal, mg, msp, asize, d, | |
2776 | TRACE_TOO_SMALL); | |
2777 | continue; | |
34dc7c2f | 2778 | } |
7a614407 GW |
2779 | |
2780 | /* | |
2781 | * If the selected metaslab is condensing, skip it. | |
2782 | */ | |
93cf2076 | 2783 | if (msp->ms_condensing) |
7a614407 GW |
2784 | continue; |
2785 | ||
9babb374 | 2786 | was_active = msp->ms_weight & METASLAB_ACTIVE_MASK; |
34dc7c2f BB |
2787 | if (activation_weight == METASLAB_WEIGHT_PRIMARY) |
2788 | break; | |
2789 | ||
2790 | target_distance = min_distance + | |
93cf2076 GW |
2791 | (space_map_allocated(msp->ms_sm) != 0 ? 0 : |
2792 | min_distance >> 1); | |
34dc7c2f | 2793 | |
4e21fd06 | 2794 | for (i = 0; i < d; i++) { |
34dc7c2f BB |
2795 | if (metaslab_distance(msp, &dva[i]) < |
2796 | target_distance) | |
2797 | break; | |
4e21fd06 | 2798 | } |
34dc7c2f BB |
2799 | if (i == d) |
2800 | break; | |
2801 | } | |
2802 | mutex_exit(&mg->mg_lock); | |
4e21fd06 DB |
2803 | if (msp == NULL) { |
2804 | kmem_free(search, sizeof (*search)); | |
34dc7c2f | 2805 | return (-1ULL); |
4e21fd06 DB |
2806 | } |
2807 | search->ms_weight = msp->ms_weight; | |
2808 | search->ms_start = msp->ms_start + 1; | |
34dc7c2f | 2809 | |
ac72fac3 GW |
2810 | mutex_enter(&msp->ms_lock); |
2811 | ||
34dc7c2f BB |
2812 | /* |
2813 | * Ensure that the metaslab we have selected is still | |
2814 | * capable of handling our request. It's possible that | |
2815 | * another thread may have changed the weight while we | |
4e21fd06 DB |
2816 | * were blocked on the metaslab lock. We check the |
2817 | * active status first to see if we need to reselect | |
2818 | * a new metaslab. | |
34dc7c2f | 2819 | */ |
4e21fd06 | 2820 | if (was_active && !(msp->ms_weight & METASLAB_ACTIVE_MASK)) { |
34dc7c2f BB |
2821 | mutex_exit(&msp->ms_lock); |
2822 | continue; | |
2823 | } | |
2824 | ||
2825 | if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) && | |
2826 | activation_weight == METASLAB_WEIGHT_PRIMARY) { | |
2827 | metaslab_passivate(msp, | |
2828 | msp->ms_weight & ~METASLAB_ACTIVE_MASK); | |
2829 | mutex_exit(&msp->ms_lock); | |
2830 | continue; | |
2831 | } | |
2832 | ||
6d974228 | 2833 | if (metaslab_activate(msp, activation_weight) != 0) { |
34dc7c2f BB |
2834 | mutex_exit(&msp->ms_lock); |
2835 | continue; | |
2836 | } | |
4e21fd06 DB |
2837 | msp->ms_selected_txg = txg; |
2838 | ||
2839 | /* | |
2840 | * Now that we have the lock, recheck to see if we should | |
2841 | * continue to use this metaslab for this allocation. The | |
2842 | * the metaslab is now loaded so metaslab_should_allocate() can | |
2843 | * accurately determine if the allocation attempt should | |
2844 | * proceed. | |
2845 | */ | |
2846 | if (!metaslab_should_allocate(msp, asize)) { | |
2847 | /* Passivate this metaslab and select a new one. */ | |
2848 | metaslab_trace_add(zal, mg, msp, asize, d, | |
2849 | TRACE_TOO_SMALL); | |
2850 | goto next; | |
2851 | } | |
2852 | ||
34dc7c2f | 2853 | |
7a614407 GW |
2854 | /* |
2855 | * If this metaslab is currently condensing then pick again as | |
2856 | * we can't manipulate this metaslab until it's committed | |
2857 | * to disk. | |
2858 | */ | |
93cf2076 | 2859 | if (msp->ms_condensing) { |
4e21fd06 DB |
2860 | metaslab_trace_add(zal, mg, msp, asize, d, |
2861 | TRACE_CONDENSING); | |
7a614407 GW |
2862 | mutex_exit(&msp->ms_lock); |
2863 | continue; | |
2864 | } | |
2865 | ||
4e21fd06 DB |
2866 | offset = metaslab_block_alloc(msp, asize, txg); |
2867 | metaslab_trace_add(zal, mg, msp, asize, d, offset); | |
2868 | ||
2869 | if (offset != -1ULL) { | |
2870 | /* Proactively passivate the metaslab, if needed */ | |
2871 | metaslab_segment_may_passivate(msp); | |
34dc7c2f | 2872 | break; |
4e21fd06 DB |
2873 | } |
2874 | next: | |
2875 | ASSERT(msp->ms_loaded); | |
2876 | ||
2877 | /* | |
2878 | * We were unable to allocate from this metaslab so determine | |
2879 | * a new weight for this metaslab. Now that we have loaded | |
2880 | * the metaslab we can provide a better hint to the metaslab | |
2881 | * selector. | |
2882 | * | |
2883 | * For space-based metaslabs, we use the maximum block size. | |
2884 | * This information is only available when the metaslab | |
2885 | * is loaded and is more accurate than the generic free | |
2886 | * space weight that was calculated by metaslab_weight(). | |
2887 | * This information allows us to quickly compare the maximum | |
2888 | * available allocation in the metaslab to the allocation | |
2889 | * size being requested. | |
2890 | * | |
2891 | * For segment-based metaslabs, determine the new weight | |
2892 | * based on the highest bucket in the range tree. We | |
2893 | * explicitly use the loaded segment weight (i.e. the range | |
2894 | * tree histogram) since it contains the space that is | |
2895 | * currently available for allocation and is accurate | |
2896 | * even within a sync pass. | |
2897 | */ | |
2898 | if (WEIGHT_IS_SPACEBASED(msp->ms_weight)) { | |
2899 | uint64_t weight = metaslab_block_maxsize(msp); | |
2900 | WEIGHT_SET_SPACEBASED(weight); | |
2901 | metaslab_passivate(msp, weight); | |
2902 | } else { | |
2903 | metaslab_passivate(msp, | |
2904 | metaslab_weight_from_range_tree(msp)); | |
2905 | } | |
34dc7c2f | 2906 | |
4e21fd06 DB |
2907 | /* |
2908 | * We have just failed an allocation attempt, check | |
2909 | * that metaslab_should_allocate() agrees. Otherwise, | |
2910 | * we may end up in an infinite loop retrying the same | |
2911 | * metaslab. | |
2912 | */ | |
2913 | ASSERT(!metaslab_should_allocate(msp, asize)); | |
34dc7c2f BB |
2914 | mutex_exit(&msp->ms_lock); |
2915 | } | |
4e21fd06 DB |
2916 | mutex_exit(&msp->ms_lock); |
2917 | kmem_free(search, sizeof (*search)); | |
2918 | return (offset); | |
2919 | } | |
34dc7c2f | 2920 | |
4e21fd06 DB |
2921 | static uint64_t |
2922 | metaslab_group_alloc(metaslab_group_t *mg, zio_alloc_list_t *zal, | |
2923 | uint64_t asize, uint64_t txg, uint64_t min_distance, dva_t *dva, int d) | |
2924 | { | |
2925 | uint64_t offset; | |
2926 | ASSERT(mg->mg_initialized); | |
34dc7c2f | 2927 | |
4e21fd06 DB |
2928 | offset = metaslab_group_alloc_normal(mg, zal, asize, txg, |
2929 | min_distance, dva, d); | |
34dc7c2f | 2930 | |
4e21fd06 DB |
2931 | mutex_enter(&mg->mg_lock); |
2932 | if (offset == -1ULL) { | |
2933 | mg->mg_failed_allocations++; | |
2934 | metaslab_trace_add(zal, mg, NULL, asize, d, | |
2935 | TRACE_GROUP_FAILURE); | |
2936 | if (asize == SPA_GANGBLOCKSIZE) { | |
2937 | /* | |
2938 | * This metaslab group was unable to allocate | |
2939 | * the minimum gang block size so it must be out of | |
2940 | * space. We must notify the allocation throttle | |
2941 | * to start skipping allocation attempts to this | |
2942 | * metaslab group until more space becomes available. | |
2943 | * Note: this failure cannot be caused by the | |
2944 | * allocation throttle since the allocation throttle | |
2945 | * is only responsible for skipping devices and | |
2946 | * not failing block allocations. | |
2947 | */ | |
2948 | mg->mg_no_free_space = B_TRUE; | |
2949 | } | |
2950 | } | |
2951 | mg->mg_allocations++; | |
2952 | mutex_exit(&mg->mg_lock); | |
34dc7c2f BB |
2953 | return (offset); |
2954 | } | |
2955 | ||
4e21fd06 DB |
2956 | /* |
2957 | * If we have to write a ditto block (i.e. more than one DVA for a given BP) | |
2958 | * on the same vdev as an existing DVA of this BP, then try to allocate it | |
2959 | * at least (vdev_asize / (2 ^ ditto_same_vdev_distance_shift)) away from the | |
2960 | * existing DVAs. | |
2961 | */ | |
2962 | int ditto_same_vdev_distance_shift = 3; | |
2963 | ||
34dc7c2f BB |
2964 | /* |
2965 | * Allocate a block for the specified i/o. | |
2966 | */ | |
2967 | static int | |
2968 | metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize, | |
4e21fd06 DB |
2969 | dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags, |
2970 | zio_alloc_list_t *zal) | |
34dc7c2f | 2971 | { |
920dd524 | 2972 | metaslab_group_t *mg, *fast_mg, *rotor; |
34dc7c2f | 2973 | vdev_t *vd; |
4e21fd06 | 2974 | boolean_t try_hard = B_FALSE; |
34dc7c2f BB |
2975 | |
2976 | ASSERT(!DVA_IS_VALID(&dva[d])); | |
2977 | ||
2978 | /* | |
2979 | * For testing, make some blocks above a certain size be gang blocks. | |
2980 | */ | |
4e21fd06 DB |
2981 | if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0) { |
2982 | metaslab_trace_add(zal, NULL, NULL, psize, d, TRACE_FORCE_GANG); | |
2e528b49 | 2983 | return (SET_ERROR(ENOSPC)); |
4e21fd06 | 2984 | } |
34dc7c2f BB |
2985 | |
2986 | /* | |
2987 | * Start at the rotor and loop through all mgs until we find something. | |
428870ff | 2988 | * Note that there's no locking on mc_rotor or mc_aliquot because |
34dc7c2f BB |
2989 | * nothing actually breaks if we miss a few updates -- we just won't |
2990 | * allocate quite as evenly. It all balances out over time. | |
2991 | * | |
2992 | * If we are doing ditto or log blocks, try to spread them across | |
2993 | * consecutive vdevs. If we're forced to reuse a vdev before we've | |
2994 | * allocated all of our ditto blocks, then try and spread them out on | |
2995 | * that vdev as much as possible. If it turns out to not be possible, | |
2996 | * gradually lower our standards until anything becomes acceptable. | |
2997 | * Also, allocating on consecutive vdevs (as opposed to random vdevs) | |
2998 | * gives us hope of containing our fault domains to something we're | |
2999 | * able to reason about. Otherwise, any two top-level vdev failures | |
3000 | * will guarantee the loss of data. With consecutive allocation, | |
3001 | * only two adjacent top-level vdev failures will result in data loss. | |
3002 | * | |
3003 | * If we are doing gang blocks (hintdva is non-NULL), try to keep | |
3004 | * ourselves on the same vdev as our gang block header. That | |
3005 | * way, we can hope for locality in vdev_cache, plus it makes our | |
3006 | * fault domains something tractable. | |
3007 | */ | |
3008 | if (hintdva) { | |
3009 | vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d])); | |
428870ff BB |
3010 | |
3011 | /* | |
3012 | * It's possible the vdev we're using as the hint no | |
3013 | * longer exists (i.e. removed). Consult the rotor when | |
3014 | * all else fails. | |
3015 | */ | |
3016 | if (vd != NULL) { | |
34dc7c2f | 3017 | mg = vd->vdev_mg; |
428870ff BB |
3018 | |
3019 | if (flags & METASLAB_HINTBP_AVOID && | |
3020 | mg->mg_next != NULL) | |
3021 | mg = mg->mg_next; | |
3022 | } else { | |
3023 | mg = mc->mc_rotor; | |
3024 | } | |
34dc7c2f BB |
3025 | } else if (d != 0) { |
3026 | vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1])); | |
3027 | mg = vd->vdev_mg->mg_next; | |
920dd524 ED |
3028 | } else if (flags & METASLAB_FASTWRITE) { |
3029 | mg = fast_mg = mc->mc_rotor; | |
3030 | ||
3031 | do { | |
3032 | if (fast_mg->mg_vd->vdev_pending_fastwrite < | |
3033 | mg->mg_vd->vdev_pending_fastwrite) | |
3034 | mg = fast_mg; | |
3035 | } while ((fast_mg = fast_mg->mg_next) != mc->mc_rotor); | |
3036 | ||
34dc7c2f BB |
3037 | } else { |
3038 | mg = mc->mc_rotor; | |
3039 | } | |
3040 | ||
3041 | /* | |
428870ff BB |
3042 | * If the hint put us into the wrong metaslab class, or into a |
3043 | * metaslab group that has been passivated, just follow the rotor. | |
34dc7c2f | 3044 | */ |
428870ff | 3045 | if (mg->mg_class != mc || mg->mg_activation_count <= 0) |
34dc7c2f BB |
3046 | mg = mc->mc_rotor; |
3047 | ||
3048 | rotor = mg; | |
3049 | top: | |
34dc7c2f | 3050 | do { |
4e21fd06 | 3051 | boolean_t allocatable; |
3dfb57a3 | 3052 | uint64_t offset; |
4e21fd06 | 3053 | uint64_t distance, asize; |
428870ff | 3054 | |
3dfb57a3 | 3055 | ASSERT(mg->mg_activation_count == 1); |
34dc7c2f | 3056 | vd = mg->mg_vd; |
fb5f0bc8 | 3057 | |
34dc7c2f | 3058 | /* |
b128c09f | 3059 | * Don't allocate from faulted devices. |
34dc7c2f | 3060 | */ |
4e21fd06 | 3061 | if (try_hard) { |
fb5f0bc8 BB |
3062 | spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER); |
3063 | allocatable = vdev_allocatable(vd); | |
3064 | spa_config_exit(spa, SCL_ZIO, FTAG); | |
3065 | } else { | |
3066 | allocatable = vdev_allocatable(vd); | |
3067 | } | |
ac72fac3 GW |
3068 | |
3069 | /* | |
3070 | * Determine if the selected metaslab group is eligible | |
3dfb57a3 DB |
3071 | * for allocations. If we're ganging then don't allow |
3072 | * this metaslab group to skip allocations since that would | |
3073 | * inadvertently return ENOSPC and suspend the pool | |
ac72fac3 GW |
3074 | * even though space is still available. |
3075 | */ | |
4e21fd06 | 3076 | if (allocatable && !GANG_ALLOCATION(flags) && !try_hard) { |
3dfb57a3 DB |
3077 | allocatable = metaslab_group_allocatable(mg, rotor, |
3078 | psize); | |
3079 | } | |
ac72fac3 | 3080 | |
4e21fd06 DB |
3081 | if (!allocatable) { |
3082 | metaslab_trace_add(zal, mg, NULL, psize, d, | |
3083 | TRACE_NOT_ALLOCATABLE); | |
34dc7c2f | 3084 | goto next; |
4e21fd06 | 3085 | } |
fb5f0bc8 | 3086 | |
3dfb57a3 DB |
3087 | ASSERT(mg->mg_initialized); |
3088 | ||
34dc7c2f | 3089 | /* |
4e21fd06 DB |
3090 | * Avoid writing single-copy data to a failing, |
3091 | * non-redundant vdev, unless we've already tried all | |
3092 | * other vdevs. | |
34dc7c2f BB |
3093 | */ |
3094 | if ((vd->vdev_stat.vs_write_errors > 0 || | |
3095 | vd->vdev_state < VDEV_STATE_HEALTHY) && | |
4e21fd06 DB |
3096 | d == 0 && !try_hard && vd->vdev_children == 0) { |
3097 | metaslab_trace_add(zal, mg, NULL, psize, d, | |
3098 | TRACE_VDEV_ERROR); | |
34dc7c2f BB |
3099 | goto next; |
3100 | } | |
3101 | ||
3102 | ASSERT(mg->mg_class == mc); | |
3103 | ||
4e21fd06 DB |
3104 | /* |
3105 | * If we don't need to try hard, then require that the | |
3106 | * block be 1/8th of the device away from any other DVAs | |
3107 | * in this BP. If we are trying hard, allow any offset | |
3108 | * to be used (distance=0). | |
3109 | */ | |
3110 | distance = 0; | |
3111 | if (!try_hard) { | |
3112 | distance = vd->vdev_asize >> | |
3113 | ditto_same_vdev_distance_shift; | |
3114 | if (distance <= (1ULL << vd->vdev_ms_shift)) | |
3115 | distance = 0; | |
3116 | } | |
34dc7c2f BB |
3117 | |
3118 | asize = vdev_psize_to_asize(vd, psize); | |
3119 | ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0); | |
3120 | ||
4e21fd06 DB |
3121 | offset = metaslab_group_alloc(mg, zal, asize, txg, distance, |
3122 | dva, d); | |
3dfb57a3 | 3123 | |
34dc7c2f BB |
3124 | if (offset != -1ULL) { |
3125 | /* | |
3126 | * If we've just selected this metaslab group, | |
3127 | * figure out whether the corresponding vdev is | |
3128 | * over- or under-used relative to the pool, | |
3129 | * and set an allocation bias to even it out. | |
bb3250d0 ED |
3130 | * |
3131 | * Bias is also used to compensate for unequally | |
3132 | * sized vdevs so that space is allocated fairly. | |
34dc7c2f | 3133 | */ |
f3a7f661 | 3134 | if (mc->mc_aliquot == 0 && metaslab_bias_enabled) { |
34dc7c2f | 3135 | vdev_stat_t *vs = &vd->vdev_stat; |
bb3250d0 ED |
3136 | int64_t vs_free = vs->vs_space - vs->vs_alloc; |
3137 | int64_t mc_free = mc->mc_space - mc->mc_alloc; | |
3138 | int64_t ratio; | |
34dc7c2f BB |
3139 | |
3140 | /* | |
6d974228 GW |
3141 | * Calculate how much more or less we should |
3142 | * try to allocate from this device during | |
3143 | * this iteration around the rotor. | |
6d974228 | 3144 | * |
bb3250d0 ED |
3145 | * This basically introduces a zero-centered |
3146 | * bias towards the devices with the most | |
3147 | * free space, while compensating for vdev | |
3148 | * size differences. | |
3149 | * | |
3150 | * Examples: | |
3151 | * vdev V1 = 16M/128M | |
3152 | * vdev V2 = 16M/128M | |
3153 | * ratio(V1) = 100% ratio(V2) = 100% | |
3154 | * | |
3155 | * vdev V1 = 16M/128M | |
3156 | * vdev V2 = 64M/128M | |
3157 | * ratio(V1) = 127% ratio(V2) = 72% | |
6d974228 | 3158 | * |
bb3250d0 ED |
3159 | * vdev V1 = 16M/128M |
3160 | * vdev V2 = 64M/512M | |
3161 | * ratio(V1) = 40% ratio(V2) = 160% | |
34dc7c2f | 3162 | */ |
bb3250d0 ED |
3163 | ratio = (vs_free * mc->mc_alloc_groups * 100) / |
3164 | (mc_free + 1); | |
3165 | mg->mg_bias = ((ratio - 100) * | |
6d974228 | 3166 | (int64_t)mg->mg_aliquot) / 100; |
f3a7f661 GW |
3167 | } else if (!metaslab_bias_enabled) { |
3168 | mg->mg_bias = 0; | |
34dc7c2f BB |
3169 | } |
3170 | ||
920dd524 ED |
3171 | if ((flags & METASLAB_FASTWRITE) || |
3172 | atomic_add_64_nv(&mc->mc_aliquot, asize) >= | |
34dc7c2f BB |
3173 | mg->mg_aliquot + mg->mg_bias) { |
3174 | mc->mc_rotor = mg->mg_next; | |
428870ff | 3175 | mc->mc_aliquot = 0; |
34dc7c2f BB |
3176 | } |
3177 | ||
3178 | DVA_SET_VDEV(&dva[d], vd->vdev_id); | |
3179 | DVA_SET_OFFSET(&dva[d], offset); | |
e3e7cf60 D |
3180 | DVA_SET_GANG(&dva[d], |
3181 | ((flags & METASLAB_GANG_HEADER) ? 1 : 0)); | |
34dc7c2f BB |
3182 | DVA_SET_ASIZE(&dva[d], asize); |
3183 | ||
920dd524 ED |
3184 | if (flags & METASLAB_FASTWRITE) { |
3185 | atomic_add_64(&vd->vdev_pending_fastwrite, | |
3186 | psize); | |
920dd524 ED |
3187 | } |
3188 | ||
34dc7c2f BB |
3189 | return (0); |
3190 | } | |
3191 | next: | |
3192 | mc->mc_rotor = mg->mg_next; | |
428870ff | 3193 | mc->mc_aliquot = 0; |
34dc7c2f BB |
3194 | } while ((mg = mg->mg_next) != rotor); |
3195 | ||
4e21fd06 DB |
3196 | /* |
3197 | * If we haven't tried hard, do so now. | |
3198 | */ | |
3199 | if (!try_hard) { | |
3200 | try_hard = B_TRUE; | |
fb5f0bc8 BB |
3201 | goto top; |
3202 | } | |
3203 | ||
34dc7c2f BB |
3204 | bzero(&dva[d], sizeof (dva_t)); |
3205 | ||
4e21fd06 | 3206 | metaslab_trace_add(zal, rotor, NULL, psize, d, TRACE_ENOSPC); |
2e528b49 | 3207 | return (SET_ERROR(ENOSPC)); |
34dc7c2f BB |
3208 | } |
3209 | ||
3210 | /* | |
3211 | * Free the block represented by DVA in the context of the specified | |
3212 | * transaction group. | |
3213 | */ | |
3214 | static void | |
3215 | metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now) | |
3216 | { | |
3217 | uint64_t vdev = DVA_GET_VDEV(dva); | |
3218 | uint64_t offset = DVA_GET_OFFSET(dva); | |
3219 | uint64_t size = DVA_GET_ASIZE(dva); | |
3220 | vdev_t *vd; | |
3221 | metaslab_t *msp; | |
3222 | ||
34dc7c2f BB |
3223 | if (txg > spa_freeze_txg(spa)) |
3224 | return; | |
3225 | ||
7d2868d5 | 3226 | if ((vd = vdev_lookup_top(spa, vdev)) == NULL || !DVA_IS_VALID(dva) || |
34dc7c2f | 3227 | (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) { |
7d2868d5 BB |
3228 | zfs_panic_recover("metaslab_free_dva(): bad DVA %llu:%llu:%llu", |
3229 | (u_longlong_t)vdev, (u_longlong_t)offset, | |
3230 | (u_longlong_t)size); | |
34dc7c2f BB |
3231 | return; |
3232 | } | |
3233 | ||
3234 | msp = vd->vdev_ms[offset >> vd->vdev_ms_shift]; | |
3235 | ||
3236 | if (DVA_GET_GANG(dva)) | |
3237 | size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); | |
3238 | ||
3239 | mutex_enter(&msp->ms_lock); | |
3240 | ||
3241 | if (now) { | |
93cf2076 | 3242 | range_tree_remove(msp->ms_alloctree[txg & TXG_MASK], |
34dc7c2f | 3243 | offset, size); |
93cf2076 GW |
3244 | |
3245 | VERIFY(!msp->ms_condensing); | |
3246 | VERIFY3U(offset, >=, msp->ms_start); | |
3247 | VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size); | |
3248 | VERIFY3U(range_tree_space(msp->ms_tree) + size, <=, | |
3249 | msp->ms_size); | |
3250 | VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift)); | |
3251 | VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift)); | |
3252 | range_tree_add(msp->ms_tree, offset, size); | |
4e21fd06 | 3253 | msp->ms_max_size = metaslab_block_maxsize(msp); |
34dc7c2f | 3254 | } else { |
93cf2076 | 3255 | if (range_tree_space(msp->ms_freetree[txg & TXG_MASK]) == 0) |
34dc7c2f | 3256 | vdev_dirty(vd, VDD_METASLAB, msp, txg); |
93cf2076 GW |
3257 | range_tree_add(msp->ms_freetree[txg & TXG_MASK], |
3258 | offset, size); | |
34dc7c2f BB |
3259 | } |
3260 | ||
3261 | mutex_exit(&msp->ms_lock); | |
3262 | } | |
3263 | ||
3264 | /* | |
3265 | * Intent log support: upon opening the pool after a crash, notify the SPA | |
3266 | * of blocks that the intent log has allocated for immediate write, but | |
3267 | * which are still considered free by the SPA because the last transaction | |
3268 | * group didn't commit yet. | |
3269 | */ | |
3270 | static int | |
3271 | metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg) | |
3272 | { | |
3273 | uint64_t vdev = DVA_GET_VDEV(dva); | |
3274 | uint64_t offset = DVA_GET_OFFSET(dva); | |
3275 | uint64_t size = DVA_GET_ASIZE(dva); | |
3276 | vdev_t *vd; | |
3277 | metaslab_t *msp; | |
428870ff | 3278 | int error = 0; |
34dc7c2f BB |
3279 | |
3280 | ASSERT(DVA_IS_VALID(dva)); | |
3281 | ||
3282 | if ((vd = vdev_lookup_top(spa, vdev)) == NULL || | |
3283 | (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) | |
2e528b49 | 3284 | return (SET_ERROR(ENXIO)); |
34dc7c2f BB |
3285 | |
3286 | msp = vd->vdev_ms[offset >> vd->vdev_ms_shift]; | |
3287 | ||
3288 | if (DVA_GET_GANG(dva)) | |
3289 | size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); | |
3290 | ||
3291 | mutex_enter(&msp->ms_lock); | |
3292 | ||
93cf2076 | 3293 | if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded) |
6d974228 | 3294 | error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY); |
428870ff | 3295 | |
93cf2076 | 3296 | if (error == 0 && !range_tree_contains(msp->ms_tree, offset, size)) |
2e528b49 | 3297 | error = SET_ERROR(ENOENT); |
428870ff | 3298 | |
b128c09f | 3299 | if (error || txg == 0) { /* txg == 0 indicates dry run */ |
34dc7c2f BB |
3300 | mutex_exit(&msp->ms_lock); |
3301 | return (error); | |
3302 | } | |
3303 | ||
93cf2076 GW |
3304 | VERIFY(!msp->ms_condensing); |
3305 | VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift)); | |
3306 | VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift)); | |
3307 | VERIFY3U(range_tree_space(msp->ms_tree) - size, <=, msp->ms_size); | |
3308 | range_tree_remove(msp->ms_tree, offset, size); | |
b128c09f | 3309 | |
fb5f0bc8 | 3310 | if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */ |
93cf2076 | 3311 | if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0) |
b128c09f | 3312 | vdev_dirty(vd, VDD_METASLAB, msp, txg); |
93cf2076 | 3313 | range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, size); |
b128c09f | 3314 | } |
34dc7c2f BB |
3315 | |
3316 | mutex_exit(&msp->ms_lock); | |
3317 | ||
3318 | return (0); | |
3319 | } | |
3320 | ||
3dfb57a3 DB |
3321 | /* |
3322 | * Reserve some allocation slots. The reservation system must be called | |
3323 | * before we call into the allocator. If there aren't any available slots | |
3324 | * then the I/O will be throttled until an I/O completes and its slots are | |
3325 | * freed up. The function returns true if it was successful in placing | |
3326 | * the reservation. | |
3327 | */ | |
3328 | boolean_t | |
3329 | metaslab_class_throttle_reserve(metaslab_class_t *mc, int slots, zio_t *zio, | |
3330 | int flags) | |
3331 | { | |
3332 | uint64_t available_slots = 0; | |
3333 | uint64_t reserved_slots; | |
3334 | boolean_t slot_reserved = B_FALSE; | |
3335 | ||
3336 | ASSERT(mc->mc_alloc_throttle_enabled); | |
3337 | mutex_enter(&mc->mc_lock); | |
3338 | ||
3339 | reserved_slots = refcount_count(&mc->mc_alloc_slots); | |
3340 | if (reserved_slots < mc->mc_alloc_max_slots) | |
3341 | available_slots = mc->mc_alloc_max_slots - reserved_slots; | |
3342 | ||
3343 | if (slots <= available_slots || GANG_ALLOCATION(flags)) { | |
3344 | int d; | |
3345 | ||
3346 | /* | |
3347 | * We reserve the slots individually so that we can unreserve | |
3348 | * them individually when an I/O completes. | |
3349 | */ | |
3350 | for (d = 0; d < slots; d++) { | |
3351 | reserved_slots = refcount_add(&mc->mc_alloc_slots, zio); | |
3352 | } | |
3353 | zio->io_flags |= ZIO_FLAG_IO_ALLOCATING; | |
3354 | slot_reserved = B_TRUE; | |
3355 | } | |
3356 | ||
3357 | mutex_exit(&mc->mc_lock); | |
3358 | return (slot_reserved); | |
3359 | } | |
3360 | ||
3361 | void | |
3362 | metaslab_class_throttle_unreserve(metaslab_class_t *mc, int slots, zio_t *zio) | |
3363 | { | |
3364 | int d; | |
3365 | ||
3366 | ASSERT(mc->mc_alloc_throttle_enabled); | |
3367 | mutex_enter(&mc->mc_lock); | |
3368 | for (d = 0; d < slots; d++) { | |
3369 | (void) refcount_remove(&mc->mc_alloc_slots, zio); | |
3370 | } | |
3371 | mutex_exit(&mc->mc_lock); | |
3372 | } | |
3373 | ||
34dc7c2f BB |
3374 | int |
3375 | metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp, | |
4e21fd06 DB |
3376 | int ndvas, uint64_t txg, blkptr_t *hintbp, int flags, |
3377 | zio_alloc_list_t *zal, zio_t *zio) | |
34dc7c2f BB |
3378 | { |
3379 | dva_t *dva = bp->blk_dva; | |
3380 | dva_t *hintdva = hintbp->blk_dva; | |
d6320ddb | 3381 | int d, error = 0; |
34dc7c2f | 3382 | |
b128c09f | 3383 | ASSERT(bp->blk_birth == 0); |
428870ff | 3384 | ASSERT(BP_PHYSICAL_BIRTH(bp) == 0); |
b128c09f BB |
3385 | |
3386 | spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER); | |
3387 | ||
3388 | if (mc->mc_rotor == NULL) { /* no vdevs in this class */ | |
3389 | spa_config_exit(spa, SCL_ALLOC, FTAG); | |
2e528b49 | 3390 | return (SET_ERROR(ENOSPC)); |
b128c09f | 3391 | } |
34dc7c2f BB |
3392 | |
3393 | ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa)); | |
3394 | ASSERT(BP_GET_NDVAS(bp) == 0); | |
3395 | ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp)); | |
4e21fd06 | 3396 | ASSERT3P(zal, !=, NULL); |
34dc7c2f | 3397 | |
d6320ddb | 3398 | for (d = 0; d < ndvas; d++) { |
34dc7c2f | 3399 | error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva, |
4e21fd06 | 3400 | txg, flags, zal); |
93cf2076 | 3401 | if (error != 0) { |
34dc7c2f BB |
3402 | for (d--; d >= 0; d--) { |
3403 | metaslab_free_dva(spa, &dva[d], txg, B_TRUE); | |
3dfb57a3 DB |
3404 | metaslab_group_alloc_decrement(spa, |
3405 | DVA_GET_VDEV(&dva[d]), zio, flags); | |
34dc7c2f BB |
3406 | bzero(&dva[d], sizeof (dva_t)); |
3407 | } | |
b128c09f | 3408 | spa_config_exit(spa, SCL_ALLOC, FTAG); |
34dc7c2f | 3409 | return (error); |
3dfb57a3 DB |
3410 | } else { |
3411 | /* | |
3412 | * Update the metaslab group's queue depth | |
3413 | * based on the newly allocated dva. | |
3414 | */ | |
3415 | metaslab_group_alloc_increment(spa, | |
3416 | DVA_GET_VDEV(&dva[d]), zio, flags); | |
34dc7c2f | 3417 | } |
3dfb57a3 | 3418 | |
34dc7c2f BB |
3419 | } |
3420 | ASSERT(error == 0); | |
3421 | ASSERT(BP_GET_NDVAS(bp) == ndvas); | |
3422 | ||
b128c09f BB |
3423 | spa_config_exit(spa, SCL_ALLOC, FTAG); |
3424 | ||
efe7978d | 3425 | BP_SET_BIRTH(bp, txg, 0); |
b128c09f | 3426 | |
34dc7c2f BB |
3427 | return (0); |
3428 | } | |
3429 | ||
3430 | void | |
3431 | metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now) | |
3432 | { | |
3433 | const dva_t *dva = bp->blk_dva; | |
d6320ddb | 3434 | int d, ndvas = BP_GET_NDVAS(bp); |
34dc7c2f BB |
3435 | |
3436 | ASSERT(!BP_IS_HOLE(bp)); | |
428870ff | 3437 | ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa)); |
b128c09f BB |
3438 | |
3439 | spa_config_enter(spa, SCL_FREE, FTAG, RW_READER); | |
34dc7c2f | 3440 | |
d6320ddb | 3441 | for (d = 0; d < ndvas; d++) |
34dc7c2f | 3442 | metaslab_free_dva(spa, &dva[d], txg, now); |
b128c09f BB |
3443 | |
3444 | spa_config_exit(spa, SCL_FREE, FTAG); | |
34dc7c2f BB |
3445 | } |
3446 | ||
3447 | int | |
3448 | metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg) | |
3449 | { | |
3450 | const dva_t *dva = bp->blk_dva; | |
3451 | int ndvas = BP_GET_NDVAS(bp); | |
d6320ddb | 3452 | int d, error = 0; |
34dc7c2f BB |
3453 | |
3454 | ASSERT(!BP_IS_HOLE(bp)); | |
3455 | ||
b128c09f BB |
3456 | if (txg != 0) { |
3457 | /* | |
3458 | * First do a dry run to make sure all DVAs are claimable, | |
3459 | * so we don't have to unwind from partial failures below. | |
3460 | */ | |
3461 | if ((error = metaslab_claim(spa, bp, 0)) != 0) | |
3462 | return (error); | |
3463 | } | |
3464 | ||
3465 | spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER); | |
3466 | ||
d6320ddb | 3467 | for (d = 0; d < ndvas; d++) |
34dc7c2f | 3468 | if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0) |
b128c09f BB |
3469 | break; |
3470 | ||
3471 | spa_config_exit(spa, SCL_ALLOC, FTAG); | |
3472 | ||
3473 | ASSERT(error == 0 || txg == 0); | |
34dc7c2f | 3474 | |
b128c09f | 3475 | return (error); |
34dc7c2f | 3476 | } |
920dd524 | 3477 | |
d1d7e268 MK |
3478 | void |
3479 | metaslab_fastwrite_mark(spa_t *spa, const blkptr_t *bp) | |
920dd524 ED |
3480 | { |
3481 | const dva_t *dva = bp->blk_dva; | |
3482 | int ndvas = BP_GET_NDVAS(bp); | |
3483 | uint64_t psize = BP_GET_PSIZE(bp); | |
3484 | int d; | |
3485 | vdev_t *vd; | |
3486 | ||
3487 | ASSERT(!BP_IS_HOLE(bp)); | |
9b67f605 | 3488 | ASSERT(!BP_IS_EMBEDDED(bp)); |
920dd524 ED |
3489 | ASSERT(psize > 0); |
3490 | ||
3491 | spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER); | |
3492 | ||
3493 | for (d = 0; d < ndvas; d++) { | |
3494 | if ((vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d]))) == NULL) | |
3495 | continue; | |
3496 | atomic_add_64(&vd->vdev_pending_fastwrite, psize); | |
3497 | } | |
3498 | ||
3499 | spa_config_exit(spa, SCL_VDEV, FTAG); | |
3500 | } | |
3501 | ||
d1d7e268 MK |
3502 | void |
3503 | metaslab_fastwrite_unmark(spa_t *spa, const blkptr_t *bp) | |
920dd524 ED |
3504 | { |
3505 | const dva_t *dva = bp->blk_dva; | |
3506 | int ndvas = BP_GET_NDVAS(bp); | |
3507 | uint64_t psize = BP_GET_PSIZE(bp); | |
3508 | int d; | |
3509 | vdev_t *vd; | |
3510 | ||
3511 | ASSERT(!BP_IS_HOLE(bp)); | |
9b67f605 | 3512 | ASSERT(!BP_IS_EMBEDDED(bp)); |
920dd524 ED |
3513 | ASSERT(psize > 0); |
3514 | ||
3515 | spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER); | |
3516 | ||
3517 | for (d = 0; d < ndvas; d++) { | |
3518 | if ((vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d]))) == NULL) | |
3519 | continue; | |
3520 | ASSERT3U(vd->vdev_pending_fastwrite, >=, psize); | |
3521 | atomic_sub_64(&vd->vdev_pending_fastwrite, psize); | |
3522 | } | |
3523 | ||
3524 | spa_config_exit(spa, SCL_VDEV, FTAG); | |
3525 | } | |
30b92c1d | 3526 | |
13fe0198 MA |
3527 | void |
3528 | metaslab_check_free(spa_t *spa, const blkptr_t *bp) | |
3529 | { | |
3530 | int i, j; | |
3531 | ||
3532 | if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0) | |
3533 | return; | |
3534 | ||
3535 | spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER); | |
3536 | for (i = 0; i < BP_GET_NDVAS(bp); i++) { | |
93cf2076 GW |
3537 | uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]); |
3538 | vdev_t *vd = vdev_lookup_top(spa, vdev); | |
3539 | uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]); | |
13fe0198 | 3540 | uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]); |
93cf2076 | 3541 | metaslab_t *msp = vd->vdev_ms[offset >> vd->vdev_ms_shift]; |
13fe0198 | 3542 | |
93cf2076 GW |
3543 | if (msp->ms_loaded) |
3544 | range_tree_verify(msp->ms_tree, offset, size); | |
13fe0198 MA |
3545 | |
3546 | for (j = 0; j < TXG_SIZE; j++) | |
93cf2076 | 3547 | range_tree_verify(msp->ms_freetree[j], offset, size); |
13fe0198 | 3548 | for (j = 0; j < TXG_DEFER_SIZE; j++) |
93cf2076 | 3549 | range_tree_verify(msp->ms_defertree[j], offset, size); |
13fe0198 MA |
3550 | } |
3551 | spa_config_exit(spa, SCL_VDEV, FTAG); | |
3552 | } | |
3553 | ||
30b92c1d | 3554 | #if defined(_KERNEL) && defined(HAVE_SPL) |
02730c33 | 3555 | /* CSTYLED */ |
99b14de4 | 3556 | module_param(metaslab_aliquot, ulong, 0644); |
99b14de4 ED |
3557 | MODULE_PARM_DESC(metaslab_aliquot, |
3558 | "allocation granularity (a.k.a. stripe size)"); | |
02730c33 BB |
3559 | |
3560 | module_param(metaslab_debug_load, int, 0644); | |
93cf2076 GW |
3561 | MODULE_PARM_DESC(metaslab_debug_load, |
3562 | "load all metaslabs when pool is first opened"); | |
02730c33 BB |
3563 | |
3564 | module_param(metaslab_debug_unload, int, 0644); | |
1ce04573 BB |
3565 | MODULE_PARM_DESC(metaslab_debug_unload, |
3566 | "prevent metaslabs from being unloaded"); | |
02730c33 BB |
3567 | |
3568 | module_param(metaslab_preload_enabled, int, 0644); | |
f3a7f661 GW |
3569 | MODULE_PARM_DESC(metaslab_preload_enabled, |
3570 | "preload potential metaslabs during reassessment"); | |
f4a4046b | 3571 | |
02730c33 | 3572 | module_param(zfs_mg_noalloc_threshold, int, 0644); |
f4a4046b TC |
3573 | MODULE_PARM_DESC(zfs_mg_noalloc_threshold, |
3574 | "percentage of free space for metaslab group to allow allocation"); | |
02730c33 BB |
3575 | |
3576 | module_param(zfs_mg_fragmentation_threshold, int, 0644); | |
f3a7f661 GW |
3577 | MODULE_PARM_DESC(zfs_mg_fragmentation_threshold, |
3578 | "fragmentation for metaslab group to allow allocation"); | |
3579 | ||
02730c33 | 3580 | module_param(zfs_metaslab_fragmentation_threshold, int, 0644); |
f3a7f661 GW |
3581 | MODULE_PARM_DESC(zfs_metaslab_fragmentation_threshold, |
3582 | "fragmentation for metaslab to allow allocation"); | |
02730c33 BB |
3583 | |
3584 | module_param(metaslab_fragmentation_factor_enabled, int, 0644); | |
f3a7f661 GW |
3585 | MODULE_PARM_DESC(metaslab_fragmentation_factor_enabled, |
3586 | "use the fragmentation metric to prefer less fragmented metaslabs"); | |
02730c33 BB |
3587 | |
3588 | module_param(metaslab_lba_weighting_enabled, int, 0644); | |
f3a7f661 GW |
3589 | MODULE_PARM_DESC(metaslab_lba_weighting_enabled, |
3590 | "prefer metaslabs with lower LBAs"); | |
02730c33 BB |
3591 | |
3592 | module_param(metaslab_bias_enabled, int, 0644); | |
f3a7f661 GW |
3593 | MODULE_PARM_DESC(metaslab_bias_enabled, |
3594 | "enable metaslab group biasing"); | |
4e21fd06 DB |
3595 | |
3596 | module_param(zfs_metaslab_segment_weight_enabled, int, 0644); | |
3597 | MODULE_PARM_DESC(zfs_metaslab_segment_weight_enabled, | |
3598 | "enable segment-based metaslab selection"); | |
3599 | ||
3600 | module_param(zfs_metaslab_switch_threshold, int, 0644); | |
3601 | MODULE_PARM_DESC(zfs_metaslab_switch_threshold, | |
3602 | "segment-based metaslab selection maximum buckets before switching"); | |
30b92c1d | 3603 | #endif /* _KERNEL && HAVE_SPL */ |