]>
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. |
893a6d62 | 23 | * Copyright (c) 2011, 2019 by Delphix. All rights reserved. |
2e528b49 | 24 | * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. |
cc99f275 | 25 | * Copyright (c) 2017, Intel Corporation. |
34dc7c2f BB |
26 | */ |
27 | ||
34dc7c2f | 28 | #include <sys/zfs_context.h> |
34dc7c2f BB |
29 | #include <sys/dmu.h> |
30 | #include <sys/dmu_tx.h> | |
31 | #include <sys/space_map.h> | |
32 | #include <sys/metaslab_impl.h> | |
33 | #include <sys/vdev_impl.h> | |
34 | #include <sys/zio.h> | |
93cf2076 | 35 | #include <sys/spa_impl.h> |
f3a7f661 | 36 | #include <sys/zfeature.h> |
a1d477c2 | 37 | #include <sys/vdev_indirect_mapping.h> |
d2734cce | 38 | #include <sys/zap.h> |
34dc7c2f | 39 | |
d1d7e268 | 40 | #define WITH_DF_BLOCK_ALLOCATOR |
6d974228 | 41 | |
3dfb57a3 DB |
42 | #define GANG_ALLOCATION(flags) \ |
43 | ((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER)) | |
22c81dd8 | 44 | |
e8fe6684 ED |
45 | /* |
46 | * Metaslab granularity, in bytes. This is roughly similar to what would be | |
47 | * referred to as the "stripe size" in traditional RAID arrays. In normal | |
48 | * operation, we will try to write this amount of data to a top-level vdev | |
49 | * before moving on to the next one. | |
50 | */ | |
99b14de4 | 51 | unsigned long metaslab_aliquot = 512 << 10; |
e8fe6684 | 52 | |
d830d479 MA |
53 | /* |
54 | * For testing, make some blocks above a certain size be gang blocks. | |
55 | */ | |
56 | unsigned long metaslab_force_ganging = SPA_MAXBLOCKSIZE + 1; | |
34dc7c2f | 57 | |
d2734cce | 58 | /* |
93e28d66 SD |
59 | * In pools where the log space map feature is not enabled we touch |
60 | * multiple metaslabs (and their respective space maps) with each | |
61 | * transaction group. Thus, we benefit from having a small space map | |
d2734cce | 62 | * block size since it allows us to issue more I/O operations scattered |
93e28d66 SD |
63 | * around the disk. So a sane default for the space map block size |
64 | * is 8~16K. | |
d2734cce | 65 | */ |
93e28d66 SD |
66 | int zfs_metaslab_sm_blksz_no_log = (1 << 14); |
67 | ||
68 | /* | |
69 | * When the log space map feature is enabled, we accumulate a lot of | |
70 | * changes per metaslab that are flushed once in a while so we benefit | |
71 | * from a bigger block size like 128K for the metaslab space maps. | |
72 | */ | |
73 | int zfs_metaslab_sm_blksz_with_log = (1 << 17); | |
d2734cce | 74 | |
e51be066 GW |
75 | /* |
76 | * The in-core space map representation is more compact than its on-disk form. | |
77 | * The zfs_condense_pct determines how much more compact the in-core | |
4e21fd06 | 78 | * space map representation must be before we compact it on-disk. |
e51be066 GW |
79 | * Values should be greater than or equal to 100. |
80 | */ | |
81 | int zfs_condense_pct = 200; | |
82 | ||
b02fe35d AR |
83 | /* |
84 | * Condensing a metaslab is not guaranteed to actually reduce the amount of | |
85 | * space used on disk. In particular, a space map uses data in increments of | |
96358617 | 86 | * MAX(1 << ashift, space_map_blksz), so a metaslab might use the |
b02fe35d AR |
87 | * same number of blocks after condensing. Since the goal of condensing is to |
88 | * reduce the number of IOPs required to read the space map, we only want to | |
89 | * condense when we can be sure we will reduce the number of blocks used by the | |
90 | * space map. Unfortunately, we cannot precisely compute whether or not this is | |
91 | * the case in metaslab_should_condense since we are holding ms_lock. Instead, | |
92 | * we apply the following heuristic: do not condense a spacemap unless the | |
93 | * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold | |
94 | * blocks. | |
95 | */ | |
96 | int zfs_metaslab_condense_block_threshold = 4; | |
97 | ||
ac72fac3 GW |
98 | /* |
99 | * The zfs_mg_noalloc_threshold defines which metaslab groups should | |
100 | * be eligible for allocation. The value is defined as a percentage of | |
f3a7f661 | 101 | * free space. Metaslab groups that have more free space than |
ac72fac3 GW |
102 | * zfs_mg_noalloc_threshold are always eligible for allocations. Once |
103 | * a metaslab group's free space is less than or equal to the | |
104 | * zfs_mg_noalloc_threshold the allocator will avoid allocating to that | |
105 | * group unless all groups in the pool have reached zfs_mg_noalloc_threshold. | |
106 | * Once all groups in the pool reach zfs_mg_noalloc_threshold then all | |
107 | * groups are allowed to accept allocations. Gang blocks are always | |
108 | * eligible to allocate on any metaslab group. The default value of 0 means | |
109 | * no metaslab group will be excluded based on this criterion. | |
110 | */ | |
111 | int zfs_mg_noalloc_threshold = 0; | |
6d974228 | 112 | |
f3a7f661 GW |
113 | /* |
114 | * Metaslab groups are considered eligible for allocations if their | |
e1cfd73f | 115 | * fragmentation metric (measured as a percentage) is less than or |
cb020f0d SD |
116 | * equal to zfs_mg_fragmentation_threshold. If a metaslab group |
117 | * exceeds this threshold then it will be skipped unless all metaslab | |
118 | * groups within the metaslab class have also crossed this threshold. | |
119 | * | |
120 | * This tunable was introduced to avoid edge cases where we continue | |
121 | * allocating from very fragmented disks in our pool while other, less | |
122 | * fragmented disks, exists. On the other hand, if all disks in the | |
123 | * pool are uniformly approaching the threshold, the threshold can | |
124 | * be a speed bump in performance, where we keep switching the disks | |
125 | * that we allocate from (e.g. we allocate some segments from disk A | |
126 | * making it bypassing the threshold while freeing segments from disk | |
127 | * B getting its fragmentation below the threshold). | |
128 | * | |
129 | * Empirically, we've seen that our vdev selection for allocations is | |
130 | * good enough that fragmentation increases uniformly across all vdevs | |
131 | * the majority of the time. Thus we set the threshold percentage high | |
132 | * enough to avoid hitting the speed bump on pools that are being pushed | |
133 | * to the edge. | |
f3a7f661 | 134 | */ |
cb020f0d | 135 | int zfs_mg_fragmentation_threshold = 95; |
f3a7f661 GW |
136 | |
137 | /* | |
138 | * Allow metaslabs to keep their active state as long as their fragmentation | |
139 | * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An | |
140 | * active metaslab that exceeds this threshold will no longer keep its active | |
141 | * status allowing better metaslabs to be selected. | |
142 | */ | |
143 | int zfs_metaslab_fragmentation_threshold = 70; | |
144 | ||
428870ff | 145 | /* |
aa7d06a9 | 146 | * When set will load all metaslabs when pool is first opened. |
428870ff | 147 | */ |
aa7d06a9 GW |
148 | int metaslab_debug_load = 0; |
149 | ||
150 | /* | |
151 | * When set will prevent metaslabs from being unloaded. | |
152 | */ | |
153 | int metaslab_debug_unload = 0; | |
428870ff | 154 | |
9babb374 BB |
155 | /* |
156 | * Minimum size which forces the dynamic allocator to change | |
428870ff | 157 | * it's allocation strategy. Once the space map cannot satisfy |
9babb374 BB |
158 | * an allocation of this size then it switches to using more |
159 | * aggressive strategy (i.e search by size rather than offset). | |
160 | */ | |
4e21fd06 | 161 | uint64_t metaslab_df_alloc_threshold = SPA_OLD_MAXBLOCKSIZE; |
9babb374 BB |
162 | |
163 | /* | |
164 | * The minimum free space, in percent, which must be available | |
165 | * in a space map to continue allocations in a first-fit fashion. | |
4e21fd06 | 166 | * Once the space map's free space drops below this level we dynamically |
9babb374 BB |
167 | * switch to using best-fit allocations. |
168 | */ | |
428870ff BB |
169 | int metaslab_df_free_pct = 4; |
170 | ||
d3230d76 MA |
171 | /* |
172 | * Maximum distance to search forward from the last offset. Without this | |
173 | * limit, fragmented pools can see >100,000 iterations and | |
174 | * metaslab_block_picker() becomes the performance limiting factor on | |
175 | * high-performance storage. | |
176 | * | |
177 | * With the default setting of 16MB, we typically see less than 500 | |
178 | * iterations, even with very fragmented, ashift=9 pools. The maximum number | |
179 | * of iterations possible is: | |
180 | * metaslab_df_max_search / (2 * (1<<ashift)) | |
181 | * With the default setting of 16MB this is 16*1024 (with ashift=9) or | |
182 | * 2048 (with ashift=12). | |
183 | */ | |
184 | int metaslab_df_max_search = 16 * 1024 * 1024; | |
185 | ||
186 | /* | |
187 | * If we are not searching forward (due to metaslab_df_max_search, | |
188 | * metaslab_df_free_pct, or metaslab_df_alloc_threshold), this tunable | |
189 | * controls what segment is used. If it is set, we will use the largest free | |
190 | * segment. If it is not set, we will use a segment of exactly the requested | |
191 | * size (or larger). | |
192 | */ | |
193 | int metaslab_df_use_largest_segment = B_FALSE; | |
194 | ||
428870ff | 195 | /* |
93cf2076 | 196 | * Percentage of all cpus that can be used by the metaslab taskq. |
428870ff | 197 | */ |
93cf2076 | 198 | int metaslab_load_pct = 50; |
428870ff BB |
199 | |
200 | /* | |
eef0f4d8 PD |
201 | * These tunables control how long a metaslab will remain loaded after the |
202 | * last allocation from it. A metaslab can't be unloaded until at least | |
203 | * metaslab_unload_delay TXG's and metaslab_unload_delay_ms milliseconds | |
204 | * have elapsed. However, zfs_metaslab_mem_limit may cause it to be | |
205 | * unloaded sooner. These settings are intended to be generous -- to keep | |
206 | * metaslabs loaded for a long time, reducing the rate of metaslab loading. | |
428870ff | 207 | */ |
eef0f4d8 PD |
208 | int metaslab_unload_delay = 32; |
209 | int metaslab_unload_delay_ms = 10 * 60 * 1000; /* ten minutes */ | |
9babb374 | 210 | |
93cf2076 GW |
211 | /* |
212 | * Max number of metaslabs per group to preload. | |
213 | */ | |
eef0f4d8 | 214 | int metaslab_preload_limit = 10; |
93cf2076 GW |
215 | |
216 | /* | |
217 | * Enable/disable preloading of metaslab. | |
218 | */ | |
f3a7f661 | 219 | int metaslab_preload_enabled = B_TRUE; |
93cf2076 GW |
220 | |
221 | /* | |
f3a7f661 | 222 | * Enable/disable fragmentation weighting on metaslabs. |
93cf2076 | 223 | */ |
f3a7f661 | 224 | int metaslab_fragmentation_factor_enabled = B_TRUE; |
93cf2076 | 225 | |
f3a7f661 GW |
226 | /* |
227 | * Enable/disable lba weighting (i.e. outer tracks are given preference). | |
228 | */ | |
229 | int metaslab_lba_weighting_enabled = B_TRUE; | |
230 | ||
231 | /* | |
232 | * Enable/disable metaslab group biasing. | |
233 | */ | |
234 | int metaslab_bias_enabled = B_TRUE; | |
235 | ||
a1d477c2 MA |
236 | /* |
237 | * Enable/disable remapping of indirect DVAs to their concrete vdevs. | |
238 | */ | |
239 | boolean_t zfs_remap_blkptr_enable = B_TRUE; | |
240 | ||
4e21fd06 DB |
241 | /* |
242 | * Enable/disable segment-based metaslab selection. | |
243 | */ | |
244 | int zfs_metaslab_segment_weight_enabled = B_TRUE; | |
245 | ||
246 | /* | |
247 | * When using segment-based metaslab selection, we will continue | |
248 | * allocating from the active metaslab until we have exhausted | |
249 | * zfs_metaslab_switch_threshold of its buckets. | |
250 | */ | |
251 | int zfs_metaslab_switch_threshold = 2; | |
252 | ||
253 | /* | |
254 | * Internal switch to enable/disable the metaslab allocation tracing | |
255 | * facility. | |
256 | */ | |
257 | #ifdef _METASLAB_TRACING | |
258 | boolean_t metaslab_trace_enabled = B_TRUE; | |
259 | #endif | |
260 | ||
261 | /* | |
262 | * Maximum entries that the metaslab allocation tracing facility will keep | |
263 | * in a given list when running in non-debug mode. We limit the number | |
264 | * of entries in non-debug mode to prevent us from using up too much memory. | |
265 | * The limit should be sufficiently large that we don't expect any allocation | |
266 | * to every exceed this value. In debug mode, the system will panic if this | |
267 | * limit is ever reached allowing for further investigation. | |
268 | */ | |
269 | #ifdef _METASLAB_TRACING | |
270 | uint64_t metaslab_trace_max_entries = 5000; | |
271 | #endif | |
272 | ||
1b939560 BB |
273 | /* |
274 | * Maximum number of metaslabs per group that can be disabled | |
275 | * simultaneously. | |
276 | */ | |
277 | int max_disabled_ms = 3; | |
278 | ||
f09fda50 PD |
279 | /* |
280 | * Maximum percentage of memory to use on storing loaded metaslabs. If loading | |
281 | * a metaslab would take it over this percentage, the oldest selected metaslab | |
282 | * is automatically unloaded. | |
283 | */ | |
eef0f4d8 PD |
284 | int zfs_metaslab_mem_limit = 25; |
285 | ||
286 | /* | |
287 | * Time (in seconds) to respect ms_max_size when the metaslab is not loaded. | |
288 | * To avoid 64-bit overflow, don't set above UINT32_MAX. | |
289 | */ | |
290 | unsigned long zfs_metaslab_max_size_cache_sec = 3600; /* 1 hour */ | |
f09fda50 | 291 | |
65a91b16 SD |
292 | static uint64_t metaslab_weight(metaslab_t *, boolean_t); |
293 | static void metaslab_set_fragmentation(metaslab_t *, boolean_t); | |
d2734cce | 294 | static void metaslab_free_impl(vdev_t *, uint64_t, uint64_t, boolean_t); |
a1d477c2 | 295 | static void metaslab_check_free_impl(vdev_t *, uint64_t, uint64_t); |
4e21fd06 | 296 | |
492f64e9 PD |
297 | static void metaslab_passivate(metaslab_t *msp, uint64_t weight); |
298 | static uint64_t metaslab_weight_from_range_tree(metaslab_t *msp); | |
93e28d66 | 299 | static void metaslab_flush_update(metaslab_t *, dmu_tx_t *); |
f09fda50 PD |
300 | static unsigned int metaslab_idx_func(multilist_t *, void *); |
301 | static void metaslab_evict(metaslab_t *, uint64_t); | |
4e21fd06 DB |
302 | #ifdef _METASLAB_TRACING |
303 | kmem_cache_t *metaslab_alloc_trace_cache; | |
304 | #endif | |
93cf2076 | 305 | |
34dc7c2f BB |
306 | /* |
307 | * ========================================================================== | |
308 | * Metaslab classes | |
309 | * ========================================================================== | |
310 | */ | |
311 | metaslab_class_t * | |
93cf2076 | 312 | metaslab_class_create(spa_t *spa, metaslab_ops_t *ops) |
34dc7c2f BB |
313 | { |
314 | metaslab_class_t *mc; | |
315 | ||
79c76d5b | 316 | mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP); |
34dc7c2f | 317 | |
428870ff | 318 | mc->mc_spa = spa; |
34dc7c2f | 319 | mc->mc_rotor = NULL; |
9babb374 | 320 | mc->mc_ops = ops; |
3dfb57a3 | 321 | mutex_init(&mc->mc_lock, NULL, MUTEX_DEFAULT, NULL); |
f09fda50 PD |
322 | mc->mc_metaslab_txg_list = multilist_create(sizeof (metaslab_t), |
323 | offsetof(metaslab_t, ms_class_txg_node), metaslab_idx_func); | |
492f64e9 | 324 | mc->mc_alloc_slots = kmem_zalloc(spa->spa_alloc_count * |
c13060e4 | 325 | sizeof (zfs_refcount_t), KM_SLEEP); |
492f64e9 PD |
326 | mc->mc_alloc_max_slots = kmem_zalloc(spa->spa_alloc_count * |
327 | sizeof (uint64_t), KM_SLEEP); | |
328 | for (int i = 0; i < spa->spa_alloc_count; i++) | |
424fd7c3 | 329 | zfs_refcount_create_tracked(&mc->mc_alloc_slots[i]); |
34dc7c2f BB |
330 | |
331 | return (mc); | |
332 | } | |
333 | ||
334 | void | |
335 | metaslab_class_destroy(metaslab_class_t *mc) | |
336 | { | |
428870ff BB |
337 | ASSERT(mc->mc_rotor == NULL); |
338 | ASSERT(mc->mc_alloc == 0); | |
339 | ASSERT(mc->mc_deferred == 0); | |
340 | ASSERT(mc->mc_space == 0); | |
341 | ASSERT(mc->mc_dspace == 0); | |
34dc7c2f | 342 | |
492f64e9 | 343 | for (int i = 0; i < mc->mc_spa->spa_alloc_count; i++) |
424fd7c3 | 344 | zfs_refcount_destroy(&mc->mc_alloc_slots[i]); |
492f64e9 | 345 | kmem_free(mc->mc_alloc_slots, mc->mc_spa->spa_alloc_count * |
c13060e4 | 346 | sizeof (zfs_refcount_t)); |
492f64e9 PD |
347 | kmem_free(mc->mc_alloc_max_slots, mc->mc_spa->spa_alloc_count * |
348 | sizeof (uint64_t)); | |
3dfb57a3 | 349 | mutex_destroy(&mc->mc_lock); |
f09fda50 | 350 | multilist_destroy(mc->mc_metaslab_txg_list); |
34dc7c2f BB |
351 | kmem_free(mc, sizeof (metaslab_class_t)); |
352 | } | |
353 | ||
428870ff BB |
354 | int |
355 | metaslab_class_validate(metaslab_class_t *mc) | |
34dc7c2f | 356 | { |
428870ff BB |
357 | metaslab_group_t *mg; |
358 | vdev_t *vd; | |
34dc7c2f | 359 | |
428870ff BB |
360 | /* |
361 | * Must hold one of the spa_config locks. | |
362 | */ | |
363 | ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) || | |
364 | spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER)); | |
34dc7c2f | 365 | |
428870ff BB |
366 | if ((mg = mc->mc_rotor) == NULL) |
367 | return (0); | |
368 | ||
369 | do { | |
370 | vd = mg->mg_vd; | |
371 | ASSERT(vd->vdev_mg != NULL); | |
372 | ASSERT3P(vd->vdev_top, ==, vd); | |
373 | ASSERT3P(mg->mg_class, ==, mc); | |
374 | ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops); | |
375 | } while ((mg = mg->mg_next) != mc->mc_rotor); | |
376 | ||
377 | return (0); | |
34dc7c2f BB |
378 | } |
379 | ||
cc99f275 | 380 | static void |
428870ff BB |
381 | metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta, |
382 | int64_t defer_delta, int64_t space_delta, int64_t dspace_delta) | |
34dc7c2f | 383 | { |
428870ff BB |
384 | atomic_add_64(&mc->mc_alloc, alloc_delta); |
385 | atomic_add_64(&mc->mc_deferred, defer_delta); | |
386 | atomic_add_64(&mc->mc_space, space_delta); | |
387 | atomic_add_64(&mc->mc_dspace, dspace_delta); | |
388 | } | |
34dc7c2f | 389 | |
428870ff BB |
390 | uint64_t |
391 | metaslab_class_get_alloc(metaslab_class_t *mc) | |
392 | { | |
393 | return (mc->mc_alloc); | |
394 | } | |
34dc7c2f | 395 | |
428870ff BB |
396 | uint64_t |
397 | metaslab_class_get_deferred(metaslab_class_t *mc) | |
398 | { | |
399 | return (mc->mc_deferred); | |
400 | } | |
34dc7c2f | 401 | |
428870ff BB |
402 | uint64_t |
403 | metaslab_class_get_space(metaslab_class_t *mc) | |
404 | { | |
405 | return (mc->mc_space); | |
406 | } | |
34dc7c2f | 407 | |
428870ff BB |
408 | uint64_t |
409 | metaslab_class_get_dspace(metaslab_class_t *mc) | |
410 | { | |
411 | return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space); | |
34dc7c2f BB |
412 | } |
413 | ||
f3a7f661 GW |
414 | void |
415 | metaslab_class_histogram_verify(metaslab_class_t *mc) | |
416 | { | |
cc99f275 DB |
417 | spa_t *spa = mc->mc_spa; |
418 | vdev_t *rvd = spa->spa_root_vdev; | |
f3a7f661 | 419 | uint64_t *mc_hist; |
1c27024e | 420 | int i; |
f3a7f661 GW |
421 | |
422 | if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0) | |
423 | return; | |
424 | ||
425 | mc_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE, | |
79c76d5b | 426 | KM_SLEEP); |
f3a7f661 | 427 | |
1c27024e | 428 | for (int c = 0; c < rvd->vdev_children; c++) { |
f3a7f661 GW |
429 | vdev_t *tvd = rvd->vdev_child[c]; |
430 | metaslab_group_t *mg = tvd->vdev_mg; | |
431 | ||
432 | /* | |
433 | * Skip any holes, uninitialized top-levels, or | |
434 | * vdevs that are not in this metalab class. | |
435 | */ | |
a1d477c2 | 436 | if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 || |
f3a7f661 GW |
437 | mg->mg_class != mc) { |
438 | continue; | |
439 | } | |
440 | ||
441 | for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++) | |
442 | mc_hist[i] += mg->mg_histogram[i]; | |
443 | } | |
444 | ||
445 | for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++) | |
446 | VERIFY3U(mc_hist[i], ==, mc->mc_histogram[i]); | |
447 | ||
448 | kmem_free(mc_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE); | |
449 | } | |
450 | ||
451 | /* | |
452 | * Calculate the metaslab class's fragmentation metric. The metric | |
453 | * is weighted based on the space contribution of each metaslab group. | |
454 | * The return value will be a number between 0 and 100 (inclusive), or | |
455 | * ZFS_FRAG_INVALID if the metric has not been set. See comment above the | |
456 | * zfs_frag_table for more information about the metric. | |
457 | */ | |
458 | uint64_t | |
459 | metaslab_class_fragmentation(metaslab_class_t *mc) | |
460 | { | |
461 | vdev_t *rvd = mc->mc_spa->spa_root_vdev; | |
462 | uint64_t fragmentation = 0; | |
f3a7f661 GW |
463 | |
464 | spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER); | |
465 | ||
1c27024e | 466 | for (int c = 0; c < rvd->vdev_children; c++) { |
f3a7f661 GW |
467 | vdev_t *tvd = rvd->vdev_child[c]; |
468 | metaslab_group_t *mg = tvd->vdev_mg; | |
469 | ||
470 | /* | |
a1d477c2 MA |
471 | * Skip any holes, uninitialized top-levels, |
472 | * or vdevs that are not in this metalab class. | |
f3a7f661 | 473 | */ |
a1d477c2 | 474 | if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 || |
f3a7f661 GW |
475 | mg->mg_class != mc) { |
476 | continue; | |
477 | } | |
478 | ||
479 | /* | |
480 | * If a metaslab group does not contain a fragmentation | |
481 | * metric then just bail out. | |
482 | */ | |
483 | if (mg->mg_fragmentation == ZFS_FRAG_INVALID) { | |
484 | spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG); | |
485 | return (ZFS_FRAG_INVALID); | |
486 | } | |
487 | ||
488 | /* | |
489 | * Determine how much this metaslab_group is contributing | |
490 | * to the overall pool fragmentation metric. | |
491 | */ | |
492 | fragmentation += mg->mg_fragmentation * | |
493 | metaslab_group_get_space(mg); | |
494 | } | |
495 | fragmentation /= metaslab_class_get_space(mc); | |
496 | ||
497 | ASSERT3U(fragmentation, <=, 100); | |
498 | spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG); | |
499 | return (fragmentation); | |
500 | } | |
501 | ||
502 | /* | |
503 | * Calculate the amount of expandable space that is available in | |
504 | * this metaslab class. If a device is expanded then its expandable | |
505 | * space will be the amount of allocatable space that is currently not | |
506 | * part of this metaslab class. | |
507 | */ | |
508 | uint64_t | |
509 | metaslab_class_expandable_space(metaslab_class_t *mc) | |
510 | { | |
511 | vdev_t *rvd = mc->mc_spa->spa_root_vdev; | |
512 | uint64_t space = 0; | |
f3a7f661 GW |
513 | |
514 | spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER); | |
1c27024e | 515 | for (int c = 0; c < rvd->vdev_children; c++) { |
f3a7f661 GW |
516 | vdev_t *tvd = rvd->vdev_child[c]; |
517 | metaslab_group_t *mg = tvd->vdev_mg; | |
518 | ||
a1d477c2 | 519 | if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 || |
f3a7f661 GW |
520 | mg->mg_class != mc) { |
521 | continue; | |
522 | } | |
523 | ||
0f676dc2 GM |
524 | /* |
525 | * Calculate if we have enough space to add additional | |
526 | * metaslabs. We report the expandable space in terms | |
527 | * of the metaslab size since that's the unit of expansion. | |
528 | */ | |
529 | space += P2ALIGN(tvd->vdev_max_asize - tvd->vdev_asize, | |
530 | 1ULL << tvd->vdev_ms_shift); | |
f3a7f661 GW |
531 | } |
532 | spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG); | |
533 | return (space); | |
534 | } | |
535 | ||
f09fda50 PD |
536 | void |
537 | metaslab_class_evict_old(metaslab_class_t *mc, uint64_t txg) | |
538 | { | |
539 | multilist_t *ml = mc->mc_metaslab_txg_list; | |
540 | for (int i = 0; i < multilist_get_num_sublists(ml); i++) { | |
541 | multilist_sublist_t *mls = multilist_sublist_lock(ml, i); | |
542 | metaslab_t *msp = multilist_sublist_head(mls); | |
543 | multilist_sublist_unlock(mls); | |
544 | while (msp != NULL) { | |
545 | mutex_enter(&msp->ms_lock); | |
f09fda50 PD |
546 | |
547 | /* | |
548 | * If the metaslab has been removed from the list | |
549 | * (which could happen if we were at the memory limit | |
550 | * and it was evicted during this loop), then we can't | |
551 | * proceed and we should restart the sublist. | |
552 | */ | |
553 | if (!multilist_link_active(&msp->ms_class_txg_node)) { | |
554 | mutex_exit(&msp->ms_lock); | |
555 | i--; | |
556 | break; | |
557 | } | |
558 | mls = multilist_sublist_lock(ml, i); | |
559 | metaslab_t *next_msp = multilist_sublist_next(mls, msp); | |
560 | multilist_sublist_unlock(mls); | |
eef0f4d8 PD |
561 | if (txg > |
562 | msp->ms_selected_txg + metaslab_unload_delay && | |
563 | gethrtime() > msp->ms_selected_time + | |
564 | (uint64_t)MSEC2NSEC(metaslab_unload_delay_ms)) { | |
565 | metaslab_evict(msp, txg); | |
566 | } else { | |
567 | /* | |
568 | * Once we've hit a metaslab selected too | |
569 | * recently to evict, we're done evicting for | |
570 | * now. | |
571 | */ | |
572 | mutex_exit(&msp->ms_lock); | |
573 | break; | |
574 | } | |
f09fda50 PD |
575 | mutex_exit(&msp->ms_lock); |
576 | msp = next_msp; | |
577 | } | |
578 | } | |
579 | } | |
580 | ||
34dc7c2f BB |
581 | static int |
582 | metaslab_compare(const void *x1, const void *x2) | |
583 | { | |
ee36c709 GN |
584 | const metaslab_t *m1 = (const metaslab_t *)x1; |
585 | const metaslab_t *m2 = (const metaslab_t *)x2; | |
34dc7c2f | 586 | |
492f64e9 PD |
587 | int sort1 = 0; |
588 | int sort2 = 0; | |
589 | if (m1->ms_allocator != -1 && m1->ms_primary) | |
590 | sort1 = 1; | |
591 | else if (m1->ms_allocator != -1 && !m1->ms_primary) | |
592 | sort1 = 2; | |
593 | if (m2->ms_allocator != -1 && m2->ms_primary) | |
594 | sort2 = 1; | |
595 | else if (m2->ms_allocator != -1 && !m2->ms_primary) | |
596 | sort2 = 2; | |
597 | ||
598 | /* | |
599 | * Sort inactive metaslabs first, then primaries, then secondaries. When | |
600 | * selecting a metaslab to allocate from, an allocator first tries its | |
601 | * primary, then secondary active metaslab. If it doesn't have active | |
602 | * metaslabs, or can't allocate from them, it searches for an inactive | |
603 | * metaslab to activate. If it can't find a suitable one, it will steal | |
604 | * a primary or secondary metaslab from another allocator. | |
605 | */ | |
606 | if (sort1 < sort2) | |
607 | return (-1); | |
608 | if (sort1 > sort2) | |
609 | return (1); | |
610 | ||
ee36c709 GN |
611 | int cmp = AVL_CMP(m2->ms_weight, m1->ms_weight); |
612 | if (likely(cmp)) | |
613 | return (cmp); | |
34dc7c2f | 614 | |
ee36c709 | 615 | IMPLY(AVL_CMP(m1->ms_start, m2->ms_start) == 0, m1 == m2); |
34dc7c2f | 616 | |
ee36c709 | 617 | return (AVL_CMP(m1->ms_start, m2->ms_start)); |
34dc7c2f BB |
618 | } |
619 | ||
4e21fd06 DB |
620 | /* |
621 | * ========================================================================== | |
622 | * Metaslab groups | |
623 | * ========================================================================== | |
624 | */ | |
ac72fac3 GW |
625 | /* |
626 | * Update the allocatable flag and the metaslab group's capacity. | |
627 | * The allocatable flag is set to true if the capacity is below | |
3dfb57a3 DB |
628 | * the zfs_mg_noalloc_threshold or has a fragmentation value that is |
629 | * greater than zfs_mg_fragmentation_threshold. If a metaslab group | |
630 | * transitions from allocatable to non-allocatable or vice versa then the | |
631 | * metaslab group's class is updated to reflect the transition. | |
ac72fac3 GW |
632 | */ |
633 | static void | |
634 | metaslab_group_alloc_update(metaslab_group_t *mg) | |
635 | { | |
636 | vdev_t *vd = mg->mg_vd; | |
637 | metaslab_class_t *mc = mg->mg_class; | |
638 | vdev_stat_t *vs = &vd->vdev_stat; | |
639 | boolean_t was_allocatable; | |
3dfb57a3 | 640 | boolean_t was_initialized; |
ac72fac3 GW |
641 | |
642 | ASSERT(vd == vd->vdev_top); | |
a1d477c2 MA |
643 | ASSERT3U(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_READER), ==, |
644 | SCL_ALLOC); | |
ac72fac3 GW |
645 | |
646 | mutex_enter(&mg->mg_lock); | |
647 | was_allocatable = mg->mg_allocatable; | |
3dfb57a3 | 648 | was_initialized = mg->mg_initialized; |
ac72fac3 GW |
649 | |
650 | mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) / | |
651 | (vs->vs_space + 1); | |
652 | ||
3dfb57a3 DB |
653 | mutex_enter(&mc->mc_lock); |
654 | ||
655 | /* | |
656 | * If the metaslab group was just added then it won't | |
657 | * have any space until we finish syncing out this txg. | |
658 | * At that point we will consider it initialized and available | |
659 | * for allocations. We also don't consider non-activated | |
660 | * metaslab groups (e.g. vdevs that are in the middle of being removed) | |
661 | * to be initialized, because they can't be used for allocation. | |
662 | */ | |
663 | mg->mg_initialized = metaslab_group_initialized(mg); | |
664 | if (!was_initialized && mg->mg_initialized) { | |
665 | mc->mc_groups++; | |
666 | } else if (was_initialized && !mg->mg_initialized) { | |
667 | ASSERT3U(mc->mc_groups, >, 0); | |
668 | mc->mc_groups--; | |
669 | } | |
670 | if (mg->mg_initialized) | |
671 | mg->mg_no_free_space = B_FALSE; | |
672 | ||
f3a7f661 GW |
673 | /* |
674 | * A metaslab group is considered allocatable if it has plenty | |
675 | * of free space or is not heavily fragmented. We only take | |
676 | * fragmentation into account if the metaslab group has a valid | |
677 | * fragmentation metric (i.e. a value between 0 and 100). | |
678 | */ | |
3dfb57a3 DB |
679 | mg->mg_allocatable = (mg->mg_activation_count > 0 && |
680 | mg->mg_free_capacity > zfs_mg_noalloc_threshold && | |
f3a7f661 GW |
681 | (mg->mg_fragmentation == ZFS_FRAG_INVALID || |
682 | mg->mg_fragmentation <= zfs_mg_fragmentation_threshold)); | |
ac72fac3 GW |
683 | |
684 | /* | |
685 | * The mc_alloc_groups maintains a count of the number of | |
686 | * groups in this metaslab class that are still above the | |
687 | * zfs_mg_noalloc_threshold. This is used by the allocating | |
688 | * threads to determine if they should avoid allocations to | |
689 | * a given group. The allocator will avoid allocations to a group | |
690 | * if that group has reached or is below the zfs_mg_noalloc_threshold | |
691 | * and there are still other groups that are above the threshold. | |
692 | * When a group transitions from allocatable to non-allocatable or | |
693 | * vice versa we update the metaslab class to reflect that change. | |
694 | * When the mc_alloc_groups value drops to 0 that means that all | |
695 | * groups have reached the zfs_mg_noalloc_threshold making all groups | |
696 | * eligible for allocations. This effectively means that all devices | |
697 | * are balanced again. | |
698 | */ | |
699 | if (was_allocatable && !mg->mg_allocatable) | |
700 | mc->mc_alloc_groups--; | |
701 | else if (!was_allocatable && mg->mg_allocatable) | |
702 | mc->mc_alloc_groups++; | |
3dfb57a3 | 703 | mutex_exit(&mc->mc_lock); |
f3a7f661 | 704 | |
ac72fac3 GW |
705 | mutex_exit(&mg->mg_lock); |
706 | } | |
707 | ||
93e28d66 SD |
708 | int |
709 | metaslab_sort_by_flushed(const void *va, const void *vb) | |
710 | { | |
711 | const metaslab_t *a = va; | |
712 | const metaslab_t *b = vb; | |
713 | ||
714 | int cmp = AVL_CMP(a->ms_unflushed_txg, b->ms_unflushed_txg); | |
715 | if (likely(cmp)) | |
716 | return (cmp); | |
717 | ||
718 | uint64_t a_vdev_id = a->ms_group->mg_vd->vdev_id; | |
719 | uint64_t b_vdev_id = b->ms_group->mg_vd->vdev_id; | |
720 | cmp = AVL_CMP(a_vdev_id, b_vdev_id); | |
721 | if (cmp) | |
722 | return (cmp); | |
723 | ||
724 | return (AVL_CMP(a->ms_id, b->ms_id)); | |
725 | } | |
726 | ||
34dc7c2f | 727 | metaslab_group_t * |
492f64e9 | 728 | metaslab_group_create(metaslab_class_t *mc, vdev_t *vd, int allocators) |
34dc7c2f BB |
729 | { |
730 | metaslab_group_t *mg; | |
731 | ||
79c76d5b | 732 | mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP); |
34dc7c2f | 733 | mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL); |
1b939560 BB |
734 | mutex_init(&mg->mg_ms_disabled_lock, NULL, MUTEX_DEFAULT, NULL); |
735 | cv_init(&mg->mg_ms_disabled_cv, NULL, CV_DEFAULT, NULL); | |
492f64e9 PD |
736 | mg->mg_primaries = kmem_zalloc(allocators * sizeof (metaslab_t *), |
737 | KM_SLEEP); | |
738 | mg->mg_secondaries = kmem_zalloc(allocators * sizeof (metaslab_t *), | |
739 | KM_SLEEP); | |
34dc7c2f | 740 | avl_create(&mg->mg_metaslab_tree, metaslab_compare, |
93e28d66 | 741 | sizeof (metaslab_t), offsetof(metaslab_t, ms_group_node)); |
34dc7c2f | 742 | mg->mg_vd = vd; |
428870ff BB |
743 | mg->mg_class = mc; |
744 | mg->mg_activation_count = 0; | |
3dfb57a3 DB |
745 | mg->mg_initialized = B_FALSE; |
746 | mg->mg_no_free_space = B_TRUE; | |
492f64e9 PD |
747 | mg->mg_allocators = allocators; |
748 | ||
c13060e4 TS |
749 | mg->mg_alloc_queue_depth = kmem_zalloc(allocators * |
750 | sizeof (zfs_refcount_t), KM_SLEEP); | |
492f64e9 PD |
751 | mg->mg_cur_max_alloc_queue_depth = kmem_zalloc(allocators * |
752 | sizeof (uint64_t), KM_SLEEP); | |
753 | for (int i = 0; i < allocators; i++) { | |
424fd7c3 | 754 | zfs_refcount_create_tracked(&mg->mg_alloc_queue_depth[i]); |
492f64e9 PD |
755 | mg->mg_cur_max_alloc_queue_depth[i] = 0; |
756 | } | |
34dc7c2f | 757 | |
3c51c5cb | 758 | mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct, |
1229323d | 759 | maxclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT | TASKQ_DYNAMIC); |
93cf2076 | 760 | |
34dc7c2f BB |
761 | return (mg); |
762 | } | |
763 | ||
764 | void | |
765 | metaslab_group_destroy(metaslab_group_t *mg) | |
766 | { | |
428870ff BB |
767 | ASSERT(mg->mg_prev == NULL); |
768 | ASSERT(mg->mg_next == NULL); | |
769 | /* | |
770 | * We may have gone below zero with the activation count | |
771 | * either because we never activated in the first place or | |
772 | * because we're done, and possibly removing the vdev. | |
773 | */ | |
774 | ASSERT(mg->mg_activation_count <= 0); | |
775 | ||
3c51c5cb | 776 | taskq_destroy(mg->mg_taskq); |
34dc7c2f | 777 | avl_destroy(&mg->mg_metaslab_tree); |
492f64e9 PD |
778 | kmem_free(mg->mg_primaries, mg->mg_allocators * sizeof (metaslab_t *)); |
779 | kmem_free(mg->mg_secondaries, mg->mg_allocators * | |
780 | sizeof (metaslab_t *)); | |
34dc7c2f | 781 | mutex_destroy(&mg->mg_lock); |
1b939560 BB |
782 | mutex_destroy(&mg->mg_ms_disabled_lock); |
783 | cv_destroy(&mg->mg_ms_disabled_cv); | |
492f64e9 PD |
784 | |
785 | for (int i = 0; i < mg->mg_allocators; i++) { | |
424fd7c3 | 786 | zfs_refcount_destroy(&mg->mg_alloc_queue_depth[i]); |
492f64e9 PD |
787 | mg->mg_cur_max_alloc_queue_depth[i] = 0; |
788 | } | |
789 | kmem_free(mg->mg_alloc_queue_depth, mg->mg_allocators * | |
c13060e4 | 790 | sizeof (zfs_refcount_t)); |
492f64e9 PD |
791 | kmem_free(mg->mg_cur_max_alloc_queue_depth, mg->mg_allocators * |
792 | sizeof (uint64_t)); | |
793 | ||
34dc7c2f BB |
794 | kmem_free(mg, sizeof (metaslab_group_t)); |
795 | } | |
796 | ||
428870ff BB |
797 | void |
798 | metaslab_group_activate(metaslab_group_t *mg) | |
799 | { | |
800 | metaslab_class_t *mc = mg->mg_class; | |
801 | metaslab_group_t *mgprev, *mgnext; | |
802 | ||
a1d477c2 | 803 | ASSERT3U(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER), !=, 0); |
428870ff BB |
804 | |
805 | ASSERT(mc->mc_rotor != mg); | |
806 | ASSERT(mg->mg_prev == NULL); | |
807 | ASSERT(mg->mg_next == NULL); | |
808 | ASSERT(mg->mg_activation_count <= 0); | |
809 | ||
810 | if (++mg->mg_activation_count <= 0) | |
811 | return; | |
812 | ||
813 | mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children); | |
ac72fac3 | 814 | metaslab_group_alloc_update(mg); |
428870ff BB |
815 | |
816 | if ((mgprev = mc->mc_rotor) == NULL) { | |
817 | mg->mg_prev = mg; | |
818 | mg->mg_next = mg; | |
819 | } else { | |
820 | mgnext = mgprev->mg_next; | |
821 | mg->mg_prev = mgprev; | |
822 | mg->mg_next = mgnext; | |
823 | mgprev->mg_next = mg; | |
824 | mgnext->mg_prev = mg; | |
825 | } | |
826 | mc->mc_rotor = mg; | |
827 | } | |
828 | ||
a1d477c2 MA |
829 | /* |
830 | * Passivate a metaslab group and remove it from the allocation rotor. | |
831 | * Callers must hold both the SCL_ALLOC and SCL_ZIO lock prior to passivating | |
832 | * a metaslab group. This function will momentarily drop spa_config_locks | |
833 | * that are lower than the SCL_ALLOC lock (see comment below). | |
834 | */ | |
428870ff BB |
835 | void |
836 | metaslab_group_passivate(metaslab_group_t *mg) | |
837 | { | |
838 | metaslab_class_t *mc = mg->mg_class; | |
a1d477c2 | 839 | spa_t *spa = mc->mc_spa; |
428870ff | 840 | metaslab_group_t *mgprev, *mgnext; |
a1d477c2 | 841 | int locks = spa_config_held(spa, SCL_ALL, RW_WRITER); |
428870ff | 842 | |
a1d477c2 MA |
843 | ASSERT3U(spa_config_held(spa, SCL_ALLOC | SCL_ZIO, RW_WRITER), ==, |
844 | (SCL_ALLOC | SCL_ZIO)); | |
428870ff BB |
845 | |
846 | if (--mg->mg_activation_count != 0) { | |
847 | ASSERT(mc->mc_rotor != mg); | |
848 | ASSERT(mg->mg_prev == NULL); | |
849 | ASSERT(mg->mg_next == NULL); | |
850 | ASSERT(mg->mg_activation_count < 0); | |
851 | return; | |
852 | } | |
853 | ||
a1d477c2 MA |
854 | /* |
855 | * The spa_config_lock is an array of rwlocks, ordered as | |
856 | * follows (from highest to lowest): | |
857 | * SCL_CONFIG > SCL_STATE > SCL_L2ARC > SCL_ALLOC > | |
858 | * SCL_ZIO > SCL_FREE > SCL_VDEV | |
859 | * (For more information about the spa_config_lock see spa_misc.c) | |
860 | * The higher the lock, the broader its coverage. When we passivate | |
861 | * a metaslab group, we must hold both the SCL_ALLOC and the SCL_ZIO | |
862 | * config locks. However, the metaslab group's taskq might be trying | |
863 | * to preload metaslabs so we must drop the SCL_ZIO lock and any | |
864 | * lower locks to allow the I/O to complete. At a minimum, | |
865 | * we continue to hold the SCL_ALLOC lock, which prevents any future | |
866 | * allocations from taking place and any changes to the vdev tree. | |
867 | */ | |
868 | spa_config_exit(spa, locks & ~(SCL_ZIO - 1), spa); | |
c5528b9b | 869 | taskq_wait_outstanding(mg->mg_taskq, 0); |
a1d477c2 | 870 | spa_config_enter(spa, locks & ~(SCL_ZIO - 1), spa, RW_WRITER); |
f3a7f661 | 871 | metaslab_group_alloc_update(mg); |
492f64e9 PD |
872 | for (int i = 0; i < mg->mg_allocators; i++) { |
873 | metaslab_t *msp = mg->mg_primaries[i]; | |
874 | if (msp != NULL) { | |
875 | mutex_enter(&msp->ms_lock); | |
876 | metaslab_passivate(msp, | |
877 | metaslab_weight_from_range_tree(msp)); | |
878 | mutex_exit(&msp->ms_lock); | |
879 | } | |
880 | msp = mg->mg_secondaries[i]; | |
881 | if (msp != NULL) { | |
882 | mutex_enter(&msp->ms_lock); | |
883 | metaslab_passivate(msp, | |
884 | metaslab_weight_from_range_tree(msp)); | |
885 | mutex_exit(&msp->ms_lock); | |
886 | } | |
887 | } | |
93cf2076 | 888 | |
428870ff BB |
889 | mgprev = mg->mg_prev; |
890 | mgnext = mg->mg_next; | |
891 | ||
892 | if (mg == mgnext) { | |
893 | mc->mc_rotor = NULL; | |
894 | } else { | |
895 | mc->mc_rotor = mgnext; | |
896 | mgprev->mg_next = mgnext; | |
897 | mgnext->mg_prev = mgprev; | |
898 | } | |
899 | ||
900 | mg->mg_prev = NULL; | |
901 | mg->mg_next = NULL; | |
902 | } | |
903 | ||
3dfb57a3 DB |
904 | boolean_t |
905 | metaslab_group_initialized(metaslab_group_t *mg) | |
906 | { | |
907 | vdev_t *vd = mg->mg_vd; | |
908 | vdev_stat_t *vs = &vd->vdev_stat; | |
909 | ||
910 | return (vs->vs_space != 0 && mg->mg_activation_count > 0); | |
911 | } | |
912 | ||
f3a7f661 GW |
913 | uint64_t |
914 | metaslab_group_get_space(metaslab_group_t *mg) | |
915 | { | |
916 | return ((1ULL << mg->mg_vd->vdev_ms_shift) * mg->mg_vd->vdev_ms_count); | |
917 | } | |
918 | ||
919 | void | |
920 | metaslab_group_histogram_verify(metaslab_group_t *mg) | |
921 | { | |
922 | uint64_t *mg_hist; | |
923 | vdev_t *vd = mg->mg_vd; | |
924 | uint64_t ashift = vd->vdev_ashift; | |
1c27024e | 925 | int i; |
f3a7f661 GW |
926 | |
927 | if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0) | |
928 | return; | |
929 | ||
930 | mg_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE, | |
79c76d5b | 931 | KM_SLEEP); |
f3a7f661 GW |
932 | |
933 | ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE, >=, | |
934 | SPACE_MAP_HISTOGRAM_SIZE + ashift); | |
935 | ||
1c27024e | 936 | for (int m = 0; m < vd->vdev_ms_count; m++) { |
f3a7f661 GW |
937 | metaslab_t *msp = vd->vdev_ms[m]; |
938 | ||
cc99f275 DB |
939 | /* skip if not active or not a member */ |
940 | if (msp->ms_sm == NULL || msp->ms_group != mg) | |
f3a7f661 GW |
941 | continue; |
942 | ||
943 | for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) | |
944 | mg_hist[i + ashift] += | |
945 | msp->ms_sm->sm_phys->smp_histogram[i]; | |
946 | } | |
947 | ||
948 | for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i ++) | |
949 | VERIFY3U(mg_hist[i], ==, mg->mg_histogram[i]); | |
950 | ||
951 | kmem_free(mg_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE); | |
952 | } | |
953 | ||
34dc7c2f | 954 | static void |
f3a7f661 | 955 | metaslab_group_histogram_add(metaslab_group_t *mg, metaslab_t *msp) |
34dc7c2f | 956 | { |
f3a7f661 GW |
957 | metaslab_class_t *mc = mg->mg_class; |
958 | uint64_t ashift = mg->mg_vd->vdev_ashift; | |
f3a7f661 GW |
959 | |
960 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
961 | if (msp->ms_sm == NULL) | |
962 | return; | |
963 | ||
34dc7c2f | 964 | mutex_enter(&mg->mg_lock); |
1c27024e | 965 | for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) { |
f3a7f661 GW |
966 | mg->mg_histogram[i + ashift] += |
967 | msp->ms_sm->sm_phys->smp_histogram[i]; | |
968 | mc->mc_histogram[i + ashift] += | |
969 | msp->ms_sm->sm_phys->smp_histogram[i]; | |
970 | } | |
971 | mutex_exit(&mg->mg_lock); | |
972 | } | |
973 | ||
974 | void | |
975 | metaslab_group_histogram_remove(metaslab_group_t *mg, metaslab_t *msp) | |
976 | { | |
977 | metaslab_class_t *mc = mg->mg_class; | |
978 | uint64_t ashift = mg->mg_vd->vdev_ashift; | |
f3a7f661 GW |
979 | |
980 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
981 | if (msp->ms_sm == NULL) | |
982 | return; | |
983 | ||
984 | mutex_enter(&mg->mg_lock); | |
1c27024e | 985 | for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) { |
f3a7f661 GW |
986 | ASSERT3U(mg->mg_histogram[i + ashift], >=, |
987 | msp->ms_sm->sm_phys->smp_histogram[i]); | |
988 | ASSERT3U(mc->mc_histogram[i + ashift], >=, | |
989 | msp->ms_sm->sm_phys->smp_histogram[i]); | |
990 | ||
991 | mg->mg_histogram[i + ashift] -= | |
992 | msp->ms_sm->sm_phys->smp_histogram[i]; | |
993 | mc->mc_histogram[i + ashift] -= | |
994 | msp->ms_sm->sm_phys->smp_histogram[i]; | |
995 | } | |
996 | mutex_exit(&mg->mg_lock); | |
997 | } | |
998 | ||
999 | static void | |
1000 | metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp) | |
1001 | { | |
34dc7c2f | 1002 | ASSERT(msp->ms_group == NULL); |
f3a7f661 | 1003 | mutex_enter(&mg->mg_lock); |
34dc7c2f BB |
1004 | msp->ms_group = mg; |
1005 | msp->ms_weight = 0; | |
1006 | avl_add(&mg->mg_metaslab_tree, msp); | |
1007 | mutex_exit(&mg->mg_lock); | |
f3a7f661 GW |
1008 | |
1009 | mutex_enter(&msp->ms_lock); | |
1010 | metaslab_group_histogram_add(mg, msp); | |
1011 | mutex_exit(&msp->ms_lock); | |
34dc7c2f BB |
1012 | } |
1013 | ||
1014 | static void | |
1015 | metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp) | |
1016 | { | |
f3a7f661 GW |
1017 | mutex_enter(&msp->ms_lock); |
1018 | metaslab_group_histogram_remove(mg, msp); | |
1019 | mutex_exit(&msp->ms_lock); | |
1020 | ||
34dc7c2f BB |
1021 | mutex_enter(&mg->mg_lock); |
1022 | ASSERT(msp->ms_group == mg); | |
1023 | avl_remove(&mg->mg_metaslab_tree, msp); | |
f09fda50 PD |
1024 | |
1025 | metaslab_class_t *mc = msp->ms_group->mg_class; | |
1026 | multilist_sublist_t *mls = | |
1027 | multilist_sublist_lock_obj(mc->mc_metaslab_txg_list, msp); | |
1028 | if (multilist_link_active(&msp->ms_class_txg_node)) | |
1029 | multilist_sublist_remove(mls, msp); | |
1030 | multilist_sublist_unlock(mls); | |
1031 | ||
34dc7c2f BB |
1032 | msp->ms_group = NULL; |
1033 | mutex_exit(&mg->mg_lock); | |
1034 | } | |
1035 | ||
492f64e9 PD |
1036 | static void |
1037 | metaslab_group_sort_impl(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight) | |
1038 | { | |
679b0f2a | 1039 | ASSERT(MUTEX_HELD(&msp->ms_lock)); |
492f64e9 PD |
1040 | ASSERT(MUTEX_HELD(&mg->mg_lock)); |
1041 | ASSERT(msp->ms_group == mg); | |
679b0f2a | 1042 | |
492f64e9 PD |
1043 | avl_remove(&mg->mg_metaslab_tree, msp); |
1044 | msp->ms_weight = weight; | |
1045 | avl_add(&mg->mg_metaslab_tree, msp); | |
1046 | ||
1047 | } | |
1048 | ||
34dc7c2f BB |
1049 | static void |
1050 | metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight) | |
1051 | { | |
1052 | /* | |
1053 | * Although in principle the weight can be any value, in | |
f3a7f661 | 1054 | * practice we do not use values in the range [1, 511]. |
34dc7c2f | 1055 | */ |
f3a7f661 | 1056 | ASSERT(weight >= SPA_MINBLOCKSIZE || weight == 0); |
34dc7c2f BB |
1057 | ASSERT(MUTEX_HELD(&msp->ms_lock)); |
1058 | ||
1059 | mutex_enter(&mg->mg_lock); | |
492f64e9 | 1060 | metaslab_group_sort_impl(mg, msp, weight); |
34dc7c2f BB |
1061 | mutex_exit(&mg->mg_lock); |
1062 | } | |
1063 | ||
f3a7f661 GW |
1064 | /* |
1065 | * Calculate the fragmentation for a given metaslab group. We can use | |
1066 | * a simple average here since all metaslabs within the group must have | |
1067 | * the same size. The return value will be a value between 0 and 100 | |
1068 | * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this | |
1069 | * group have a fragmentation metric. | |
1070 | */ | |
1071 | uint64_t | |
1072 | metaslab_group_fragmentation(metaslab_group_t *mg) | |
1073 | { | |
1074 | vdev_t *vd = mg->mg_vd; | |
1075 | uint64_t fragmentation = 0; | |
1076 | uint64_t valid_ms = 0; | |
f3a7f661 | 1077 | |
1c27024e | 1078 | for (int m = 0; m < vd->vdev_ms_count; m++) { |
f3a7f661 GW |
1079 | metaslab_t *msp = vd->vdev_ms[m]; |
1080 | ||
1081 | if (msp->ms_fragmentation == ZFS_FRAG_INVALID) | |
1082 | continue; | |
cc99f275 DB |
1083 | if (msp->ms_group != mg) |
1084 | continue; | |
f3a7f661 GW |
1085 | |
1086 | valid_ms++; | |
1087 | fragmentation += msp->ms_fragmentation; | |
1088 | } | |
1089 | ||
cc99f275 | 1090 | if (valid_ms <= mg->mg_vd->vdev_ms_count / 2) |
f3a7f661 GW |
1091 | return (ZFS_FRAG_INVALID); |
1092 | ||
1093 | fragmentation /= valid_ms; | |
1094 | ASSERT3U(fragmentation, <=, 100); | |
1095 | return (fragmentation); | |
1096 | } | |
1097 | ||
ac72fac3 GW |
1098 | /* |
1099 | * Determine if a given metaslab group should skip allocations. A metaslab | |
f3a7f661 GW |
1100 | * group should avoid allocations if its free capacity is less than the |
1101 | * zfs_mg_noalloc_threshold or its fragmentation metric is greater than | |
1102 | * zfs_mg_fragmentation_threshold and there is at least one metaslab group | |
3dfb57a3 DB |
1103 | * that can still handle allocations. If the allocation throttle is enabled |
1104 | * then we skip allocations to devices that have reached their maximum | |
1105 | * allocation queue depth unless the selected metaslab group is the only | |
1106 | * eligible group remaining. | |
ac72fac3 GW |
1107 | */ |
1108 | static boolean_t | |
3dfb57a3 | 1109 | metaslab_group_allocatable(metaslab_group_t *mg, metaslab_group_t *rotor, |
c197a77c | 1110 | uint64_t psize, int allocator, int d) |
ac72fac3 | 1111 | { |
3dfb57a3 | 1112 | spa_t *spa = mg->mg_vd->vdev_spa; |
ac72fac3 GW |
1113 | metaslab_class_t *mc = mg->mg_class; |
1114 | ||
1115 | /* | |
3dfb57a3 DB |
1116 | * We can only consider skipping this metaslab group if it's |
1117 | * in the normal metaslab class and there are other metaslab | |
1118 | * groups to select from. Otherwise, we always consider it eligible | |
f3a7f661 | 1119 | * for allocations. |
ac72fac3 | 1120 | */ |
cc99f275 DB |
1121 | if ((mc != spa_normal_class(spa) && |
1122 | mc != spa_special_class(spa) && | |
1123 | mc != spa_dedup_class(spa)) || | |
1124 | mc->mc_groups <= 1) | |
3dfb57a3 DB |
1125 | return (B_TRUE); |
1126 | ||
1127 | /* | |
1128 | * If the metaslab group's mg_allocatable flag is set (see comments | |
1129 | * in metaslab_group_alloc_update() for more information) and | |
1130 | * the allocation throttle is disabled then allow allocations to this | |
1131 | * device. However, if the allocation throttle is enabled then | |
1132 | * check if we have reached our allocation limit (mg_alloc_queue_depth) | |
1133 | * to determine if we should allow allocations to this metaslab group. | |
1134 | * If all metaslab groups are no longer considered allocatable | |
1135 | * (mc_alloc_groups == 0) or we're trying to allocate the smallest | |
1136 | * gang block size then we allow allocations on this metaslab group | |
1137 | * regardless of the mg_allocatable or throttle settings. | |
1138 | */ | |
1139 | if (mg->mg_allocatable) { | |
1140 | metaslab_group_t *mgp; | |
1141 | int64_t qdepth; | |
492f64e9 | 1142 | uint64_t qmax = mg->mg_cur_max_alloc_queue_depth[allocator]; |
3dfb57a3 DB |
1143 | |
1144 | if (!mc->mc_alloc_throttle_enabled) | |
1145 | return (B_TRUE); | |
1146 | ||
1147 | /* | |
1148 | * If this metaslab group does not have any free space, then | |
1149 | * there is no point in looking further. | |
1150 | */ | |
1151 | if (mg->mg_no_free_space) | |
1152 | return (B_FALSE); | |
1153 | ||
c197a77c | 1154 | /* |
1155 | * Relax allocation throttling for ditto blocks. Due to | |
1156 | * random imbalances in allocation it tends to push copies | |
1157 | * to one vdev, that looks a bit better at the moment. | |
1158 | */ | |
1159 | qmax = qmax * (4 + d) / 4; | |
1160 | ||
424fd7c3 TS |
1161 | qdepth = zfs_refcount_count( |
1162 | &mg->mg_alloc_queue_depth[allocator]); | |
3dfb57a3 DB |
1163 | |
1164 | /* | |
1165 | * If this metaslab group is below its qmax or it's | |
1166 | * the only allocatable metasable group, then attempt | |
1167 | * to allocate from it. | |
1168 | */ | |
1169 | if (qdepth < qmax || mc->mc_alloc_groups == 1) | |
1170 | return (B_TRUE); | |
1171 | ASSERT3U(mc->mc_alloc_groups, >, 1); | |
1172 | ||
1173 | /* | |
1174 | * Since this metaslab group is at or over its qmax, we | |
1175 | * need to determine if there are metaslab groups after this | |
1176 | * one that might be able to handle this allocation. This is | |
1177 | * racy since we can't hold the locks for all metaslab | |
1178 | * groups at the same time when we make this check. | |
1179 | */ | |
1180 | for (mgp = mg->mg_next; mgp != rotor; mgp = mgp->mg_next) { | |
492f64e9 | 1181 | qmax = mgp->mg_cur_max_alloc_queue_depth[allocator]; |
c197a77c | 1182 | qmax = qmax * (4 + d) / 4; |
424fd7c3 | 1183 | qdepth = zfs_refcount_count( |
492f64e9 | 1184 | &mgp->mg_alloc_queue_depth[allocator]); |
3dfb57a3 DB |
1185 | |
1186 | /* | |
1187 | * If there is another metaslab group that | |
1188 | * might be able to handle the allocation, then | |
1189 | * we return false so that we skip this group. | |
1190 | */ | |
1191 | if (qdepth < qmax && !mgp->mg_no_free_space) | |
1192 | return (B_FALSE); | |
1193 | } | |
1194 | ||
1195 | /* | |
1196 | * We didn't find another group to handle the allocation | |
1197 | * so we can't skip this metaslab group even though | |
1198 | * we are at or over our qmax. | |
1199 | */ | |
1200 | return (B_TRUE); | |
1201 | ||
1202 | } else if (mc->mc_alloc_groups == 0 || psize == SPA_MINBLOCKSIZE) { | |
1203 | return (B_TRUE); | |
1204 | } | |
1205 | return (B_FALSE); | |
ac72fac3 GW |
1206 | } |
1207 | ||
428870ff BB |
1208 | /* |
1209 | * ========================================================================== | |
93cf2076 | 1210 | * Range tree callbacks |
428870ff BB |
1211 | * ========================================================================== |
1212 | */ | |
93cf2076 GW |
1213 | |
1214 | /* | |
1215 | * Comparison function for the private size-ordered tree. Tree is sorted | |
1216 | * by size, larger sizes at the end of the tree. | |
1217 | */ | |
428870ff | 1218 | static int |
93cf2076 | 1219 | metaslab_rangesize_compare(const void *x1, const void *x2) |
428870ff | 1220 | { |
93cf2076 GW |
1221 | const range_seg_t *r1 = x1; |
1222 | const range_seg_t *r2 = x2; | |
1223 | uint64_t rs_size1 = r1->rs_end - r1->rs_start; | |
1224 | uint64_t rs_size2 = r2->rs_end - r2->rs_start; | |
428870ff | 1225 | |
ee36c709 GN |
1226 | int cmp = AVL_CMP(rs_size1, rs_size2); |
1227 | if (likely(cmp)) | |
1228 | return (cmp); | |
428870ff | 1229 | |
ee36c709 | 1230 | return (AVL_CMP(r1->rs_start, r2->rs_start)); |
428870ff BB |
1231 | } |
1232 | ||
93cf2076 GW |
1233 | /* |
1234 | * ========================================================================== | |
4e21fd06 | 1235 | * Common allocator routines |
93cf2076 GW |
1236 | * ========================================================================== |
1237 | */ | |
1238 | ||
9babb374 | 1239 | /* |
428870ff | 1240 | * Return the maximum contiguous segment within the metaslab. |
9babb374 | 1241 | */ |
9babb374 | 1242 | uint64_t |
c81f1790 | 1243 | metaslab_largest_allocatable(metaslab_t *msp) |
9babb374 | 1244 | { |
d2734cce | 1245 | avl_tree_t *t = &msp->ms_allocatable_by_size; |
93cf2076 | 1246 | range_seg_t *rs; |
9babb374 | 1247 | |
c81f1790 PD |
1248 | if (t == NULL) |
1249 | return (0); | |
1250 | rs = avl_last(t); | |
1251 | if (rs == NULL) | |
1252 | return (0); | |
9babb374 | 1253 | |
93cf2076 GW |
1254 | return (rs->rs_end - rs->rs_start); |
1255 | } | |
1256 | ||
c81f1790 PD |
1257 | /* |
1258 | * Return the maximum contiguous segment within the unflushed frees of this | |
1259 | * metaslab. | |
1260 | */ | |
1261 | uint64_t | |
1262 | metaslab_largest_unflushed_free(metaslab_t *msp) | |
1263 | { | |
1264 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
1265 | ||
1266 | if (msp->ms_unflushed_frees == NULL) | |
1267 | return (0); | |
1268 | ||
1269 | range_seg_t *rs = avl_last(&msp->ms_unflushed_frees_by_size); | |
1270 | if (rs == NULL) | |
1271 | return (0); | |
1272 | ||
1273 | /* | |
1274 | * When a range is freed from the metaslab, that range is added to | |
1275 | * both the unflushed frees and the deferred frees. While the block | |
1276 | * will eventually be usable, if the metaslab were loaded the range | |
1277 | * would not be added to the ms_allocatable tree until TXG_DEFER_SIZE | |
1278 | * txgs had passed. As a result, when attempting to estimate an upper | |
1279 | * bound for the largest currently-usable free segment in the | |
1280 | * metaslab, we need to not consider any ranges currently in the defer | |
1281 | * trees. This algorithm approximates the largest available chunk in | |
1282 | * the largest range in the unflushed_frees tree by taking the first | |
1283 | * chunk. While this may be a poor estimate, it should only remain so | |
1284 | * briefly and should eventually self-correct as frees are no longer | |
1285 | * deferred. Similar logic applies to the ms_freed tree. See | |
1286 | * metaslab_load() for more details. | |
1287 | * | |
e1cfd73f | 1288 | * There are two primary sources of inaccuracy in this estimate. Both |
c81f1790 PD |
1289 | * are tolerated for performance reasons. The first source is that we |
1290 | * only check the largest segment for overlaps. Smaller segments may | |
1291 | * have more favorable overlaps with the other trees, resulting in | |
1292 | * larger usable chunks. Second, we only look at the first chunk in | |
1293 | * the largest segment; there may be other usable chunks in the | |
1294 | * largest segment, but we ignore them. | |
1295 | */ | |
1296 | uint64_t rstart = rs->rs_start; | |
1297 | uint64_t rsize = rs->rs_end - rstart; | |
1298 | for (int t = 0; t < TXG_DEFER_SIZE; t++) { | |
1299 | uint64_t start = 0; | |
1300 | uint64_t size = 0; | |
1301 | boolean_t found = range_tree_find_in(msp->ms_defer[t], rstart, | |
1302 | rsize, &start, &size); | |
1303 | if (found) { | |
1304 | if (rstart == start) | |
1305 | return (0); | |
1306 | rsize = start - rstart; | |
1307 | } | |
1308 | } | |
1309 | ||
1310 | uint64_t start = 0; | |
1311 | uint64_t size = 0; | |
1312 | boolean_t found = range_tree_find_in(msp->ms_freed, rstart, | |
1313 | rsize, &start, &size); | |
1314 | if (found) | |
1315 | rsize = start - rstart; | |
1316 | ||
1317 | return (rsize); | |
1318 | } | |
1319 | ||
4e21fd06 DB |
1320 | static range_seg_t * |
1321 | metaslab_block_find(avl_tree_t *t, uint64_t start, uint64_t size) | |
93cf2076 | 1322 | { |
4e21fd06 DB |
1323 | range_seg_t *rs, rsearch; |
1324 | avl_index_t where; | |
93cf2076 | 1325 | |
4e21fd06 DB |
1326 | rsearch.rs_start = start; |
1327 | rsearch.rs_end = start + size; | |
93cf2076 | 1328 | |
4e21fd06 DB |
1329 | rs = avl_find(t, &rsearch, &where); |
1330 | if (rs == NULL) { | |
1331 | rs = avl_nearest(t, where, AVL_AFTER); | |
93cf2076 | 1332 | } |
93cf2076 | 1333 | |
4e21fd06 DB |
1334 | return (rs); |
1335 | } | |
93cf2076 | 1336 | |
d3230d76 | 1337 | #if defined(WITH_DF_BLOCK_ALLOCATOR) || \ |
93cf2076 GW |
1338 | defined(WITH_CF_BLOCK_ALLOCATOR) |
1339 | /* | |
1340 | * This is a helper function that can be used by the allocator to find | |
1341 | * a suitable block to allocate. This will search the specified AVL | |
1342 | * tree looking for a block that matches the specified criteria. | |
1343 | */ | |
1344 | static uint64_t | |
1345 | metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size, | |
d3230d76 | 1346 | uint64_t max_search) |
93cf2076 | 1347 | { |
4e21fd06 | 1348 | range_seg_t *rs = metaslab_block_find(t, *cursor, size); |
d3230d76 | 1349 | uint64_t first_found; |
93cf2076 | 1350 | |
d3230d76 MA |
1351 | if (rs != NULL) |
1352 | first_found = rs->rs_start; | |
93cf2076 | 1353 | |
d3230d76 MA |
1354 | while (rs != NULL && rs->rs_start - first_found <= max_search) { |
1355 | uint64_t offset = rs->rs_start; | |
93cf2076 GW |
1356 | if (offset + size <= rs->rs_end) { |
1357 | *cursor = offset + size; | |
1358 | return (offset); | |
1359 | } | |
1360 | rs = AVL_NEXT(t, rs); | |
1361 | } | |
1362 | ||
93cf2076 | 1363 | *cursor = 0; |
d3230d76 | 1364 | return (-1ULL); |
9babb374 | 1365 | } |
d3230d76 | 1366 | #endif /* WITH_DF/CF_BLOCK_ALLOCATOR */ |
22c81dd8 BB |
1367 | |
1368 | #if defined(WITH_DF_BLOCK_ALLOCATOR) | |
428870ff BB |
1369 | /* |
1370 | * ========================================================================== | |
d3230d76 MA |
1371 | * Dynamic Fit (df) block allocator |
1372 | * | |
1373 | * Search for a free chunk of at least this size, starting from the last | |
1374 | * offset (for this alignment of block) looking for up to | |
1375 | * metaslab_df_max_search bytes (16MB). If a large enough free chunk is not | |
1376 | * found within 16MB, then return a free chunk of exactly the requested size (or | |
1377 | * larger). | |
1378 | * | |
1379 | * If it seems like searching from the last offset will be unproductive, skip | |
1380 | * that and just return a free chunk of exactly the requested size (or larger). | |
1381 | * This is based on metaslab_df_alloc_threshold and metaslab_df_free_pct. This | |
1382 | * mechanism is probably not very useful and may be removed in the future. | |
1383 | * | |
1384 | * The behavior when not searching can be changed to return the largest free | |
1385 | * chunk, instead of a free chunk of exactly the requested size, by setting | |
1386 | * metaslab_df_use_largest_segment. | |
428870ff BB |
1387 | * ========================================================================== |
1388 | */ | |
9babb374 | 1389 | static uint64_t |
93cf2076 | 1390 | metaslab_df_alloc(metaslab_t *msp, uint64_t size) |
9babb374 | 1391 | { |
93cf2076 GW |
1392 | /* |
1393 | * Find the largest power of 2 block size that evenly divides the | |
1394 | * requested size. This is used to try to allocate blocks with similar | |
1395 | * alignment from the same area of the metaslab (i.e. same cursor | |
1396 | * bucket) but it does not guarantee that other allocations sizes | |
1397 | * may exist in the same region. | |
1398 | */ | |
9babb374 | 1399 | uint64_t align = size & -size; |
9bd274dd | 1400 | uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1]; |
d2734cce | 1401 | range_tree_t *rt = msp->ms_allocatable; |
93cf2076 | 1402 | int free_pct = range_tree_space(rt) * 100 / msp->ms_size; |
d3230d76 | 1403 | uint64_t offset; |
9babb374 | 1404 | |
93cf2076 | 1405 | ASSERT(MUTEX_HELD(&msp->ms_lock)); |
d3230d76 | 1406 | ASSERT3U(avl_numnodes(&rt->rt_root), ==, |
d2734cce | 1407 | avl_numnodes(&msp->ms_allocatable_by_size)); |
9babb374 | 1408 | |
9babb374 | 1409 | /* |
d3230d76 MA |
1410 | * If we're running low on space, find a segment based on size, |
1411 | * rather than iterating based on offset. | |
9babb374 | 1412 | */ |
c81f1790 | 1413 | if (metaslab_largest_allocatable(msp) < metaslab_df_alloc_threshold || |
9babb374 | 1414 | free_pct < metaslab_df_free_pct) { |
d3230d76 MA |
1415 | offset = -1; |
1416 | } else { | |
1417 | offset = metaslab_block_picker(&rt->rt_root, | |
1418 | cursor, size, metaslab_df_max_search); | |
9babb374 BB |
1419 | } |
1420 | ||
d3230d76 MA |
1421 | if (offset == -1) { |
1422 | range_seg_t *rs; | |
1423 | if (metaslab_df_use_largest_segment) { | |
1424 | /* use largest free segment */ | |
1425 | rs = avl_last(&msp->ms_allocatable_by_size); | |
1426 | } else { | |
1427 | /* use segment of this size, or next largest */ | |
1428 | rs = metaslab_block_find(&msp->ms_allocatable_by_size, | |
1429 | 0, size); | |
1430 | } | |
1431 | if (rs != NULL && rs->rs_start + size <= rs->rs_end) { | |
1432 | offset = rs->rs_start; | |
1433 | *cursor = offset + size; | |
1434 | } | |
1435 | } | |
1436 | ||
1437 | return (offset); | |
9babb374 BB |
1438 | } |
1439 | ||
93cf2076 | 1440 | static metaslab_ops_t metaslab_df_ops = { |
f3a7f661 | 1441 | metaslab_df_alloc |
34dc7c2f BB |
1442 | }; |
1443 | ||
93cf2076 | 1444 | metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops; |
22c81dd8 BB |
1445 | #endif /* WITH_DF_BLOCK_ALLOCATOR */ |
1446 | ||
93cf2076 | 1447 | #if defined(WITH_CF_BLOCK_ALLOCATOR) |
428870ff BB |
1448 | /* |
1449 | * ========================================================================== | |
93cf2076 GW |
1450 | * Cursor fit block allocator - |
1451 | * Select the largest region in the metaslab, set the cursor to the beginning | |
1452 | * of the range and the cursor_end to the end of the range. As allocations | |
1453 | * are made advance the cursor. Continue allocating from the cursor until | |
1454 | * the range is exhausted and then find a new range. | |
428870ff BB |
1455 | * ========================================================================== |
1456 | */ | |
1457 | static uint64_t | |
93cf2076 | 1458 | metaslab_cf_alloc(metaslab_t *msp, uint64_t size) |
428870ff | 1459 | { |
d2734cce SD |
1460 | range_tree_t *rt = msp->ms_allocatable; |
1461 | avl_tree_t *t = &msp->ms_allocatable_by_size; | |
93cf2076 GW |
1462 | uint64_t *cursor = &msp->ms_lbas[0]; |
1463 | uint64_t *cursor_end = &msp->ms_lbas[1]; | |
428870ff BB |
1464 | uint64_t offset = 0; |
1465 | ||
93cf2076 GW |
1466 | ASSERT(MUTEX_HELD(&msp->ms_lock)); |
1467 | ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&rt->rt_root)); | |
428870ff | 1468 | |
93cf2076 | 1469 | ASSERT3U(*cursor_end, >=, *cursor); |
428870ff | 1470 | |
93cf2076 GW |
1471 | if ((*cursor + size) > *cursor_end) { |
1472 | range_seg_t *rs; | |
428870ff | 1473 | |
d2734cce | 1474 | rs = avl_last(&msp->ms_allocatable_by_size); |
93cf2076 GW |
1475 | if (rs == NULL || (rs->rs_end - rs->rs_start) < size) |
1476 | return (-1ULL); | |
428870ff | 1477 | |
93cf2076 GW |
1478 | *cursor = rs->rs_start; |
1479 | *cursor_end = rs->rs_end; | |
428870ff | 1480 | } |
93cf2076 GW |
1481 | |
1482 | offset = *cursor; | |
1483 | *cursor += size; | |
1484 | ||
428870ff BB |
1485 | return (offset); |
1486 | } | |
1487 | ||
93cf2076 | 1488 | static metaslab_ops_t metaslab_cf_ops = { |
f3a7f661 | 1489 | metaslab_cf_alloc |
428870ff BB |
1490 | }; |
1491 | ||
93cf2076 GW |
1492 | metaslab_ops_t *zfs_metaslab_ops = &metaslab_cf_ops; |
1493 | #endif /* WITH_CF_BLOCK_ALLOCATOR */ | |
22c81dd8 BB |
1494 | |
1495 | #if defined(WITH_NDF_BLOCK_ALLOCATOR) | |
93cf2076 GW |
1496 | /* |
1497 | * ========================================================================== | |
1498 | * New dynamic fit allocator - | |
1499 | * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift | |
1500 | * contiguous blocks. If no region is found then just use the largest segment | |
1501 | * that remains. | |
1502 | * ========================================================================== | |
1503 | */ | |
1504 | ||
1505 | /* | |
1506 | * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift) | |
1507 | * to request from the allocator. | |
1508 | */ | |
428870ff BB |
1509 | uint64_t metaslab_ndf_clump_shift = 4; |
1510 | ||
1511 | static uint64_t | |
93cf2076 | 1512 | metaslab_ndf_alloc(metaslab_t *msp, uint64_t size) |
428870ff | 1513 | { |
d2734cce | 1514 | avl_tree_t *t = &msp->ms_allocatable->rt_root; |
428870ff | 1515 | avl_index_t where; |
93cf2076 | 1516 | range_seg_t *rs, rsearch; |
9bd274dd | 1517 | uint64_t hbit = highbit64(size); |
93cf2076 | 1518 | uint64_t *cursor = &msp->ms_lbas[hbit - 1]; |
c81f1790 | 1519 | uint64_t max_size = metaslab_largest_allocatable(msp); |
428870ff | 1520 | |
93cf2076 | 1521 | ASSERT(MUTEX_HELD(&msp->ms_lock)); |
d2734cce SD |
1522 | ASSERT3U(avl_numnodes(t), ==, |
1523 | avl_numnodes(&msp->ms_allocatable_by_size)); | |
428870ff BB |
1524 | |
1525 | if (max_size < size) | |
1526 | return (-1ULL); | |
1527 | ||
93cf2076 GW |
1528 | rsearch.rs_start = *cursor; |
1529 | rsearch.rs_end = *cursor + size; | |
428870ff | 1530 | |
93cf2076 GW |
1531 | rs = avl_find(t, &rsearch, &where); |
1532 | if (rs == NULL || (rs->rs_end - rs->rs_start) < size) { | |
d2734cce | 1533 | t = &msp->ms_allocatable_by_size; |
428870ff | 1534 | |
93cf2076 GW |
1535 | rsearch.rs_start = 0; |
1536 | rsearch.rs_end = MIN(max_size, | |
428870ff | 1537 | 1ULL << (hbit + metaslab_ndf_clump_shift)); |
93cf2076 GW |
1538 | rs = avl_find(t, &rsearch, &where); |
1539 | if (rs == NULL) | |
1540 | rs = avl_nearest(t, where, AVL_AFTER); | |
1541 | ASSERT(rs != NULL); | |
428870ff BB |
1542 | } |
1543 | ||
93cf2076 GW |
1544 | if ((rs->rs_end - rs->rs_start) >= size) { |
1545 | *cursor = rs->rs_start + size; | |
1546 | return (rs->rs_start); | |
428870ff BB |
1547 | } |
1548 | return (-1ULL); | |
1549 | } | |
1550 | ||
93cf2076 | 1551 | static metaslab_ops_t metaslab_ndf_ops = { |
f3a7f661 | 1552 | metaslab_ndf_alloc |
428870ff BB |
1553 | }; |
1554 | ||
93cf2076 | 1555 | metaslab_ops_t *zfs_metaslab_ops = &metaslab_ndf_ops; |
22c81dd8 | 1556 | #endif /* WITH_NDF_BLOCK_ALLOCATOR */ |
9babb374 | 1557 | |
93cf2076 | 1558 | |
34dc7c2f BB |
1559 | /* |
1560 | * ========================================================================== | |
1561 | * Metaslabs | |
1562 | * ========================================================================== | |
1563 | */ | |
93cf2076 | 1564 | |
93e28d66 SD |
1565 | /* |
1566 | * Wait for any in-progress metaslab loads to complete. | |
1567 | */ | |
1568 | void | |
1569 | metaslab_load_wait(metaslab_t *msp) | |
1570 | { | |
1571 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
1572 | ||
1573 | while (msp->ms_loading) { | |
1574 | ASSERT(!msp->ms_loaded); | |
1575 | cv_wait(&msp->ms_load_cv, &msp->ms_lock); | |
1576 | } | |
1577 | } | |
1578 | ||
1579 | /* | |
1580 | * Wait for any in-progress flushing to complete. | |
1581 | */ | |
1582 | void | |
1583 | metaslab_flush_wait(metaslab_t *msp) | |
1584 | { | |
1585 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
1586 | ||
1587 | while (msp->ms_flushing) | |
1588 | cv_wait(&msp->ms_flush_cv, &msp->ms_lock); | |
1589 | } | |
1590 | ||
f09fda50 PD |
1591 | static unsigned int |
1592 | metaslab_idx_func(multilist_t *ml, void *arg) | |
1593 | { | |
1594 | metaslab_t *msp = arg; | |
1595 | return (msp->ms_id % multilist_get_num_sublists(ml)); | |
1596 | } | |
1597 | ||
93e28d66 SD |
1598 | uint64_t |
1599 | metaslab_allocated_space(metaslab_t *msp) | |
1600 | { | |
1601 | return (msp->ms_allocated_space); | |
1602 | } | |
1603 | ||
1604 | /* | |
1605 | * Verify that the space accounting on disk matches the in-core range_trees. | |
1606 | */ | |
1607 | static void | |
1608 | metaslab_verify_space(metaslab_t *msp, uint64_t txg) | |
1609 | { | |
1610 | spa_t *spa = msp->ms_group->mg_vd->vdev_spa; | |
1611 | uint64_t allocating = 0; | |
1612 | uint64_t sm_free_space, msp_free_space; | |
1613 | ||
1614 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
1615 | ASSERT(!msp->ms_condensing); | |
1616 | ||
1617 | if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0) | |
1618 | return; | |
1619 | ||
1620 | /* | |
1621 | * We can only verify the metaslab space when we're called | |
1622 | * from syncing context with a loaded metaslab that has an | |
1623 | * allocated space map. Calling this in non-syncing context | |
1624 | * does not provide a consistent view of the metaslab since | |
1625 | * we're performing allocations in the future. | |
1626 | */ | |
1627 | if (txg != spa_syncing_txg(spa) || msp->ms_sm == NULL || | |
1628 | !msp->ms_loaded) | |
1629 | return; | |
1630 | ||
1631 | /* | |
1632 | * Even though the smp_alloc field can get negative, | |
1633 | * when it comes to a metaslab's space map, that should | |
1634 | * never be the case. | |
1635 | */ | |
1636 | ASSERT3S(space_map_allocated(msp->ms_sm), >=, 0); | |
1637 | ||
1638 | ASSERT3U(space_map_allocated(msp->ms_sm), >=, | |
1639 | range_tree_space(msp->ms_unflushed_frees)); | |
1640 | ||
1641 | ASSERT3U(metaslab_allocated_space(msp), ==, | |
1642 | space_map_allocated(msp->ms_sm) + | |
1643 | range_tree_space(msp->ms_unflushed_allocs) - | |
1644 | range_tree_space(msp->ms_unflushed_frees)); | |
1645 | ||
1646 | sm_free_space = msp->ms_size - metaslab_allocated_space(msp); | |
1647 | ||
1648 | /* | |
1649 | * Account for future allocations since we would have | |
1650 | * already deducted that space from the ms_allocatable. | |
1651 | */ | |
1652 | for (int t = 0; t < TXG_CONCURRENT_STATES; t++) { | |
1653 | allocating += | |
1654 | range_tree_space(msp->ms_allocating[(txg + t) & TXG_MASK]); | |
1655 | } | |
f09fda50 PD |
1656 | ASSERT3U(allocating + msp->ms_allocated_this_txg, ==, |
1657 | msp->ms_allocating_total); | |
93e28d66 SD |
1658 | |
1659 | ASSERT3U(msp->ms_deferspace, ==, | |
1660 | range_tree_space(msp->ms_defer[0]) + | |
1661 | range_tree_space(msp->ms_defer[1])); | |
1662 | ||
1663 | msp_free_space = range_tree_space(msp->ms_allocatable) + allocating + | |
1664 | msp->ms_deferspace + range_tree_space(msp->ms_freed); | |
1665 | ||
1666 | VERIFY3U(sm_free_space, ==, msp_free_space); | |
1667 | } | |
1668 | ||
928e8ad4 SD |
1669 | static void |
1670 | metaslab_aux_histograms_clear(metaslab_t *msp) | |
1671 | { | |
1672 | /* | |
1673 | * Auxiliary histograms are only cleared when resetting them, | |
1674 | * which can only happen while the metaslab is loaded. | |
1675 | */ | |
1676 | ASSERT(msp->ms_loaded); | |
1677 | ||
1678 | bzero(msp->ms_synchist, sizeof (msp->ms_synchist)); | |
1679 | for (int t = 0; t < TXG_DEFER_SIZE; t++) | |
1680 | bzero(msp->ms_deferhist[t], sizeof (msp->ms_deferhist[t])); | |
1681 | } | |
1682 | ||
1683 | static void | |
1684 | metaslab_aux_histogram_add(uint64_t *histogram, uint64_t shift, | |
1685 | range_tree_t *rt) | |
1686 | { | |
1687 | /* | |
1688 | * This is modeled after space_map_histogram_add(), so refer to that | |
1689 | * function for implementation details. We want this to work like | |
1690 | * the space map histogram, and not the range tree histogram, as we | |
1691 | * are essentially constructing a delta that will be later subtracted | |
1692 | * from the space map histogram. | |
1693 | */ | |
1694 | int idx = 0; | |
1695 | for (int i = shift; i < RANGE_TREE_HISTOGRAM_SIZE; i++) { | |
1696 | ASSERT3U(i, >=, idx + shift); | |
1697 | histogram[idx] += rt->rt_histogram[i] << (i - idx - shift); | |
1698 | ||
1699 | if (idx < SPACE_MAP_HISTOGRAM_SIZE - 1) { | |
1700 | ASSERT3U(idx + shift, ==, i); | |
1701 | idx++; | |
1702 | ASSERT3U(idx, <, SPACE_MAP_HISTOGRAM_SIZE); | |
1703 | } | |
1704 | } | |
1705 | } | |
1706 | ||
1707 | /* | |
1708 | * Called at every sync pass that the metaslab gets synced. | |
1709 | * | |
1710 | * The reason is that we want our auxiliary histograms to be updated | |
1711 | * wherever the metaslab's space map histogram is updated. This way | |
1712 | * we stay consistent on which parts of the metaslab space map's | |
1713 | * histogram are currently not available for allocations (e.g because | |
1714 | * they are in the defer, freed, and freeing trees). | |
1715 | */ | |
1716 | static void | |
1717 | metaslab_aux_histograms_update(metaslab_t *msp) | |
1718 | { | |
1719 | space_map_t *sm = msp->ms_sm; | |
1720 | ASSERT(sm != NULL); | |
1721 | ||
1722 | /* | |
1723 | * This is similar to the metaslab's space map histogram updates | |
1724 | * that take place in metaslab_sync(). The only difference is that | |
1725 | * we only care about segments that haven't made it into the | |
1726 | * ms_allocatable tree yet. | |
1727 | */ | |
1728 | if (msp->ms_loaded) { | |
1729 | metaslab_aux_histograms_clear(msp); | |
1730 | ||
1731 | metaslab_aux_histogram_add(msp->ms_synchist, | |
1732 | sm->sm_shift, msp->ms_freed); | |
1733 | ||
1734 | for (int t = 0; t < TXG_DEFER_SIZE; t++) { | |
1735 | metaslab_aux_histogram_add(msp->ms_deferhist[t], | |
1736 | sm->sm_shift, msp->ms_defer[t]); | |
1737 | } | |
1738 | } | |
1739 | ||
1740 | metaslab_aux_histogram_add(msp->ms_synchist, | |
1741 | sm->sm_shift, msp->ms_freeing); | |
1742 | } | |
1743 | ||
1744 | /* | |
1745 | * Called every time we are done syncing (writing to) the metaslab, | |
1746 | * i.e. at the end of each sync pass. | |
1747 | * [see the comment in metaslab_impl.h for ms_synchist, ms_deferhist] | |
1748 | */ | |
1749 | static void | |
1750 | metaslab_aux_histograms_update_done(metaslab_t *msp, boolean_t defer_allowed) | |
1751 | { | |
1752 | spa_t *spa = msp->ms_group->mg_vd->vdev_spa; | |
1753 | space_map_t *sm = msp->ms_sm; | |
1754 | ||
1755 | if (sm == NULL) { | |
1756 | /* | |
1757 | * We came here from metaslab_init() when creating/opening a | |
1758 | * pool, looking at a metaslab that hasn't had any allocations | |
1759 | * yet. | |
1760 | */ | |
1761 | return; | |
1762 | } | |
1763 | ||
1764 | /* | |
1765 | * This is similar to the actions that we take for the ms_freed | |
1766 | * and ms_defer trees in metaslab_sync_done(). | |
1767 | */ | |
1768 | uint64_t hist_index = spa_syncing_txg(spa) % TXG_DEFER_SIZE; | |
1769 | if (defer_allowed) { | |
1770 | bcopy(msp->ms_synchist, msp->ms_deferhist[hist_index], | |
1771 | sizeof (msp->ms_synchist)); | |
1772 | } else { | |
1773 | bzero(msp->ms_deferhist[hist_index], | |
1774 | sizeof (msp->ms_deferhist[hist_index])); | |
1775 | } | |
1776 | bzero(msp->ms_synchist, sizeof (msp->ms_synchist)); | |
1777 | } | |
1778 | ||
1779 | /* | |
1780 | * Ensure that the metaslab's weight and fragmentation are consistent | |
1781 | * with the contents of the histogram (either the range tree's histogram | |
1782 | * or the space map's depending whether the metaslab is loaded). | |
1783 | */ | |
1784 | static void | |
1785 | metaslab_verify_weight_and_frag(metaslab_t *msp) | |
1786 | { | |
1787 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
1788 | ||
1789 | if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0) | |
1790 | return; | |
1791 | ||
2fcf4481 SD |
1792 | /* |
1793 | * We can end up here from vdev_remove_complete(), in which case we | |
1794 | * cannot do these assertions because we hold spa config locks and | |
1795 | * thus we are not allowed to read from the DMU. | |
1796 | * | |
1797 | * We check if the metaslab group has been removed and if that's | |
1798 | * the case we return immediately as that would mean that we are | |
1799 | * here from the aforementioned code path. | |
1800 | */ | |
928e8ad4 SD |
1801 | if (msp->ms_group == NULL) |
1802 | return; | |
1803 | ||
1804 | /* | |
1805 | * Devices being removed always return a weight of 0 and leave | |
1806 | * fragmentation and ms_max_size as is - there is nothing for | |
1807 | * us to verify here. | |
1808 | */ | |
1809 | vdev_t *vd = msp->ms_group->mg_vd; | |
1810 | if (vd->vdev_removing) | |
1811 | return; | |
1812 | ||
1813 | /* | |
1814 | * If the metaslab is dirty it probably means that we've done | |
1815 | * some allocations or frees that have changed our histograms | |
1816 | * and thus the weight. | |
1817 | */ | |
1818 | for (int t = 0; t < TXG_SIZE; t++) { | |
1819 | if (txg_list_member(&vd->vdev_ms_list, msp, t)) | |
1820 | return; | |
1821 | } | |
1822 | ||
1823 | /* | |
1824 | * This verification checks that our in-memory state is consistent | |
1825 | * with what's on disk. If the pool is read-only then there aren't | |
1826 | * any changes and we just have the initially-loaded state. | |
1827 | */ | |
1828 | if (!spa_writeable(msp->ms_group->mg_vd->vdev_spa)) | |
1829 | return; | |
1830 | ||
1831 | /* some extra verification for in-core tree if you can */ | |
1832 | if (msp->ms_loaded) { | |
1833 | range_tree_stat_verify(msp->ms_allocatable); | |
1834 | VERIFY(space_map_histogram_verify(msp->ms_sm, | |
1835 | msp->ms_allocatable)); | |
1836 | } | |
1837 | ||
1838 | uint64_t weight = msp->ms_weight; | |
1839 | uint64_t was_active = msp->ms_weight & METASLAB_ACTIVE_MASK; | |
1840 | boolean_t space_based = WEIGHT_IS_SPACEBASED(msp->ms_weight); | |
1841 | uint64_t frag = msp->ms_fragmentation; | |
1842 | uint64_t max_segsize = msp->ms_max_size; | |
1843 | ||
1844 | msp->ms_weight = 0; | |
1845 | msp->ms_fragmentation = 0; | |
928e8ad4 SD |
1846 | |
1847 | /* | |
65a91b16 SD |
1848 | * This function is used for verification purposes and thus should |
1849 | * not introduce any side-effects/mutations on the system's state. | |
1850 | * | |
1851 | * Regardless of whether metaslab_weight() thinks this metaslab | |
1852 | * should be active or not, we want to ensure that the actual weight | |
1853 | * (and therefore the value of ms_weight) would be the same if it | |
1854 | * was to be recalculated at this point. | |
1855 | * | |
1856 | * In addition we set the nodirty flag so metaslab_weight() does | |
1857 | * not dirty the metaslab for future TXGs (e.g. when trying to | |
1858 | * force condensing to upgrade the metaslab spacemaps). | |
928e8ad4 | 1859 | */ |
65a91b16 | 1860 | msp->ms_weight = metaslab_weight(msp, B_TRUE) | was_active; |
928e8ad4 SD |
1861 | |
1862 | VERIFY3U(max_segsize, ==, msp->ms_max_size); | |
1863 | ||
1864 | /* | |
1865 | * If the weight type changed then there is no point in doing | |
1866 | * verification. Revert fields to their original values. | |
1867 | */ | |
1868 | if ((space_based && !WEIGHT_IS_SPACEBASED(msp->ms_weight)) || | |
1869 | (!space_based && WEIGHT_IS_SPACEBASED(msp->ms_weight))) { | |
1870 | msp->ms_fragmentation = frag; | |
1871 | msp->ms_weight = weight; | |
1872 | return; | |
1873 | } | |
1874 | ||
1875 | VERIFY3U(msp->ms_fragmentation, ==, frag); | |
1876 | VERIFY3U(msp->ms_weight, ==, weight); | |
1877 | } | |
1878 | ||
f09fda50 PD |
1879 | /* |
1880 | * If we're over the zfs_metaslab_mem_limit, select the loaded metaslab from | |
1881 | * this class that was used longest ago, and attempt to unload it. We don't | |
1882 | * want to spend too much time in this loop to prevent performance | |
e1cfd73f | 1883 | * degradation, and we expect that most of the time this operation will |
f09fda50 PD |
1884 | * succeed. Between that and the normal unloading processing during txg sync, |
1885 | * we expect this to keep the metaslab memory usage under control. | |
1886 | */ | |
1887 | static void | |
1888 | metaslab_potentially_evict(metaslab_class_t *mc) | |
1889 | { | |
1890 | #ifdef _KERNEL | |
1891 | uint64_t allmem = arc_all_memory(); | |
1892 | extern kmem_cache_t *range_seg_cache; | |
1893 | uint64_t inuse = range_seg_cache->skc_obj_total; | |
1894 | uint64_t size = range_seg_cache->skc_obj_size; | |
1895 | int tries = 0; | |
1896 | for (; allmem * zfs_metaslab_mem_limit / 100 < inuse * size && | |
1897 | tries < multilist_get_num_sublists(mc->mc_metaslab_txg_list) * 2; | |
1898 | tries++) { | |
1899 | unsigned int idx = multilist_get_random_index( | |
1900 | mc->mc_metaslab_txg_list); | |
1901 | multilist_sublist_t *mls = | |
1902 | multilist_sublist_lock(mc->mc_metaslab_txg_list, idx); | |
1903 | metaslab_t *msp = multilist_sublist_head(mls); | |
1904 | multilist_sublist_unlock(mls); | |
1905 | while (msp != NULL && allmem * zfs_metaslab_mem_limit / 100 < | |
1906 | inuse * size) { | |
1907 | VERIFY3P(mls, ==, multilist_sublist_lock( | |
1908 | mc->mc_metaslab_txg_list, idx)); | |
1909 | ASSERT3U(idx, ==, | |
1910 | metaslab_idx_func(mc->mc_metaslab_txg_list, msp)); | |
1911 | ||
1912 | if (!multilist_link_active(&msp->ms_class_txg_node)) { | |
1913 | multilist_sublist_unlock(mls); | |
1914 | break; | |
1915 | } | |
1916 | metaslab_t *next_msp = multilist_sublist_next(mls, msp); | |
1917 | multilist_sublist_unlock(mls); | |
1918 | /* | |
1919 | * If the metaslab is currently loading there are two | |
1920 | * cases. If it's the metaslab we're evicting, we | |
1921 | * can't continue on or we'll panic when we attempt to | |
1922 | * recursively lock the mutex. If it's another | |
1923 | * metaslab that's loading, it can be safely skipped, | |
1924 | * since we know it's very new and therefore not a | |
1925 | * good eviction candidate. We check later once the | |
1926 | * lock is held that the metaslab is fully loaded | |
1927 | * before actually unloading it. | |
1928 | */ | |
1929 | if (msp->ms_loading) { | |
1930 | msp = next_msp; | |
1931 | inuse = range_seg_cache->skc_obj_total; | |
1932 | continue; | |
1933 | } | |
1934 | /* | |
1935 | * We can't unload metaslabs with no spacemap because | |
1936 | * they're not ready to be unloaded yet. We can't | |
1937 | * unload metaslabs with outstanding allocations | |
1938 | * because doing so could cause the metaslab's weight | |
1939 | * to decrease while it's unloaded, which violates an | |
1940 | * invariant that we use to prevent unnecessary | |
1941 | * loading. We also don't unload metaslabs that are | |
1942 | * currently active because they are high-weight | |
1943 | * metaslabs that are likely to be used in the near | |
1944 | * future. | |
1945 | */ | |
1946 | mutex_enter(&msp->ms_lock); | |
1947 | if (msp->ms_allocator == -1 && msp->ms_sm != NULL && | |
1948 | msp->ms_allocating_total == 0) { | |
1949 | metaslab_unload(msp); | |
1950 | } | |
1951 | mutex_exit(&msp->ms_lock); | |
1952 | msp = next_msp; | |
1953 | inuse = range_seg_cache->skc_obj_total; | |
1954 | } | |
1955 | } | |
1956 | #endif | |
1957 | } | |
1958 | ||
b194fab0 SD |
1959 | static int |
1960 | metaslab_load_impl(metaslab_t *msp) | |
93cf2076 GW |
1961 | { |
1962 | int error = 0; | |
93cf2076 GW |
1963 | |
1964 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
b194fab0 | 1965 | ASSERT(msp->ms_loading); |
425d3237 | 1966 | ASSERT(!msp->ms_condensing); |
93cf2076 | 1967 | |
a1d477c2 | 1968 | /* |
425d3237 SD |
1969 | * We temporarily drop the lock to unblock other operations while we |
1970 | * are reading the space map. Therefore, metaslab_sync() and | |
1971 | * metaslab_sync_done() can run at the same time as we do. | |
1972 | * | |
93e28d66 SD |
1973 | * If we are using the log space maps, metaslab_sync() can't write to |
1974 | * the metaslab's space map while we are loading as we only write to | |
1975 | * it when we are flushing the metaslab, and that can't happen while | |
1976 | * we are loading it. | |
1977 | * | |
1978 | * If we are not using log space maps though, metaslab_sync() can | |
1979 | * append to the space map while we are loading. Therefore we load | |
1980 | * only entries that existed when we started the load. Additionally, | |
1981 | * metaslab_sync_done() has to wait for the load to complete because | |
1982 | * there are potential races like metaslab_load() loading parts of the | |
1983 | * space map that are currently being appended by metaslab_sync(). If | |
1984 | * we didn't, the ms_allocatable would have entries that | |
1985 | * metaslab_sync_done() would try to re-add later. | |
425d3237 SD |
1986 | * |
1987 | * That's why before dropping the lock we remember the synced length | |
1988 | * of the metaslab and read up to that point of the space map, | |
1989 | * ignoring entries appended by metaslab_sync() that happen after we | |
1990 | * drop the lock. | |
a1d477c2 | 1991 | */ |
425d3237 | 1992 | uint64_t length = msp->ms_synced_length; |
a1d477c2 | 1993 | mutex_exit(&msp->ms_lock); |
93cf2076 | 1994 | |
93e28d66 | 1995 | hrtime_t load_start = gethrtime(); |
d2734cce | 1996 | if (msp->ms_sm != NULL) { |
425d3237 SD |
1997 | error = space_map_load_length(msp->ms_sm, msp->ms_allocatable, |
1998 | SM_FREE, length); | |
d2734cce | 1999 | } else { |
425d3237 SD |
2000 | /* |
2001 | * The space map has not been allocated yet, so treat | |
2002 | * all the space in the metaslab as free and add it to the | |
2003 | * ms_allocatable tree. | |
2004 | */ | |
d2734cce SD |
2005 | range_tree_add(msp->ms_allocatable, |
2006 | msp->ms_start, msp->ms_size); | |
93e28d66 SD |
2007 | |
2008 | if (msp->ms_freed != NULL) { | |
2009 | /* | |
2010 | * If the ms_sm doesn't exist, this means that this | |
2011 | * metaslab hasn't gone through metaslab_sync() and | |
2012 | * thus has never been dirtied. So we shouldn't | |
2013 | * expect any unflushed allocs or frees from previous | |
2014 | * TXGs. | |
2015 | * | |
2016 | * Note: ms_freed and all the other trees except for | |
2017 | * the ms_allocatable, can be NULL at this point only | |
2018 | * if this is a new metaslab of a vdev that just got | |
2019 | * expanded. | |
2020 | */ | |
2021 | ASSERT(range_tree_is_empty(msp->ms_unflushed_allocs)); | |
2022 | ASSERT(range_tree_is_empty(msp->ms_unflushed_frees)); | |
2023 | } | |
d2734cce | 2024 | } |
93cf2076 | 2025 | |
425d3237 SD |
2026 | /* |
2027 | * We need to grab the ms_sync_lock to prevent metaslab_sync() from | |
93e28d66 SD |
2028 | * changing the ms_sm (or log_sm) and the metaslab's range trees |
2029 | * while we are about to use them and populate the ms_allocatable. | |
2030 | * The ms_lock is insufficient for this because metaslab_sync() doesn't | |
2031 | * hold the ms_lock while writing the ms_checkpointing tree to disk. | |
425d3237 SD |
2032 | */ |
2033 | mutex_enter(&msp->ms_sync_lock); | |
a1d477c2 | 2034 | mutex_enter(&msp->ms_lock); |
93e28d66 | 2035 | |
425d3237 | 2036 | ASSERT(!msp->ms_condensing); |
93e28d66 | 2037 | ASSERT(!msp->ms_flushing); |
93cf2076 | 2038 | |
8eef9976 SD |
2039 | if (error != 0) { |
2040 | mutex_exit(&msp->ms_sync_lock); | |
b194fab0 | 2041 | return (error); |
8eef9976 | 2042 | } |
4e21fd06 | 2043 | |
b194fab0 SD |
2044 | ASSERT3P(msp->ms_group, !=, NULL); |
2045 | msp->ms_loaded = B_TRUE; | |
2046 | ||
2047 | /* | |
93e28d66 SD |
2048 | * Apply all the unflushed changes to ms_allocatable right |
2049 | * away so any manipulations we do below have a clear view | |
2050 | * of what is allocated and what is free. | |
2051 | */ | |
2052 | range_tree_walk(msp->ms_unflushed_allocs, | |
2053 | range_tree_remove, msp->ms_allocatable); | |
2054 | range_tree_walk(msp->ms_unflushed_frees, | |
2055 | range_tree_add, msp->ms_allocatable); | |
2056 | ||
2057 | msp->ms_loaded = B_TRUE; | |
2058 | ||
2059 | ASSERT3P(msp->ms_group, !=, NULL); | |
2060 | spa_t *spa = msp->ms_group->mg_vd->vdev_spa; | |
2061 | if (spa_syncing_log_sm(spa) != NULL) { | |
2062 | ASSERT(spa_feature_is_enabled(spa, | |
2063 | SPA_FEATURE_LOG_SPACEMAP)); | |
2064 | ||
2065 | /* | |
2066 | * If we use a log space map we add all the segments | |
2067 | * that are in ms_unflushed_frees so they are available | |
2068 | * for allocation. | |
2069 | * | |
2070 | * ms_allocatable needs to contain all free segments | |
2071 | * that are ready for allocations (thus not segments | |
2072 | * from ms_freeing, ms_freed, and the ms_defer trees). | |
2073 | * But if we grab the lock in this code path at a sync | |
2074 | * pass later that 1, then it also contains the | |
2075 | * segments of ms_freed (they were added to it earlier | |
2076 | * in this path through ms_unflushed_frees). So we | |
2077 | * need to remove all the segments that exist in | |
2078 | * ms_freed from ms_allocatable as they will be added | |
2079 | * later in metaslab_sync_done(). | |
2080 | * | |
2081 | * When there's no log space map, the ms_allocatable | |
2082 | * correctly doesn't contain any segments that exist | |
2083 | * in ms_freed [see ms_synced_length]. | |
2084 | */ | |
2085 | range_tree_walk(msp->ms_freed, | |
2086 | range_tree_remove, msp->ms_allocatable); | |
2087 | } | |
2088 | ||
2089 | /* | |
2090 | * If we are not using the log space map, ms_allocatable | |
2091 | * contains the segments that exist in the ms_defer trees | |
2092 | * [see ms_synced_length]. Thus we need to remove them | |
2093 | * from ms_allocatable as they will be added again in | |
425d3237 | 2094 | * metaslab_sync_done(). |
93e28d66 SD |
2095 | * |
2096 | * If we are using the log space map, ms_allocatable still | |
2097 | * contains the segments that exist in the ms_defer trees. | |
2098 | * Not because it read them through the ms_sm though. But | |
2099 | * because these segments are part of ms_unflushed_frees | |
2100 | * whose segments we add to ms_allocatable earlier in this | |
2101 | * code path. | |
b194fab0 | 2102 | */ |
425d3237 SD |
2103 | for (int t = 0; t < TXG_DEFER_SIZE; t++) { |
2104 | range_tree_walk(msp->ms_defer[t], | |
2105 | range_tree_remove, msp->ms_allocatable); | |
93cf2076 | 2106 | } |
425d3237 | 2107 | |
928e8ad4 SD |
2108 | /* |
2109 | * Call metaslab_recalculate_weight_and_sort() now that the | |
2110 | * metaslab is loaded so we get the metaslab's real weight. | |
2111 | * | |
2112 | * Unless this metaslab was created with older software and | |
2113 | * has not yet been converted to use segment-based weight, we | |
2114 | * expect the new weight to be better or equal to the weight | |
2115 | * that the metaslab had while it was not loaded. This is | |
2116 | * because the old weight does not take into account the | |
2117 | * consolidation of adjacent segments between TXGs. [see | |
2118 | * comment for ms_synchist and ms_deferhist[] for more info] | |
2119 | */ | |
2120 | uint64_t weight = msp->ms_weight; | |
c81f1790 | 2121 | uint64_t max_size = msp->ms_max_size; |
928e8ad4 SD |
2122 | metaslab_recalculate_weight_and_sort(msp); |
2123 | if (!WEIGHT_IS_SPACEBASED(weight)) | |
2124 | ASSERT3U(weight, <=, msp->ms_weight); | |
c81f1790 PD |
2125 | msp->ms_max_size = metaslab_largest_allocatable(msp); |
2126 | ASSERT3U(max_size, <=, msp->ms_max_size); | |
93e28d66 | 2127 | hrtime_t load_end = gethrtime(); |
c81f1790 | 2128 | msp->ms_load_time = load_end; |
93e28d66 SD |
2129 | if (zfs_flags & ZFS_DEBUG_LOG_SPACEMAP) { |
2130 | zfs_dbgmsg("loading: txg %llu, spa %s, vdev_id %llu, " | |
2131 | "ms_id %llu, smp_length %llu, " | |
2132 | "unflushed_allocs %llu, unflushed_frees %llu, " | |
2133 | "freed %llu, defer %llu + %llu, " | |
c81f1790 PD |
2134 | "loading_time %lld ms, ms_max_size %llu, " |
2135 | "max size error %llu", | |
93e28d66 SD |
2136 | spa_syncing_txg(spa), spa_name(spa), |
2137 | msp->ms_group->mg_vd->vdev_id, msp->ms_id, | |
2138 | space_map_length(msp->ms_sm), | |
2139 | range_tree_space(msp->ms_unflushed_allocs), | |
2140 | range_tree_space(msp->ms_unflushed_frees), | |
2141 | range_tree_space(msp->ms_freed), | |
2142 | range_tree_space(msp->ms_defer[0]), | |
2143 | range_tree_space(msp->ms_defer[1]), | |
c81f1790 PD |
2144 | (longlong_t)((load_end - load_start) / 1000000), |
2145 | msp->ms_max_size, msp->ms_max_size - max_size); | |
93e28d66 SD |
2146 | } |
2147 | ||
425d3237 SD |
2148 | metaslab_verify_space(msp, spa_syncing_txg(spa)); |
2149 | mutex_exit(&msp->ms_sync_lock); | |
b194fab0 SD |
2150 | return (0); |
2151 | } | |
2152 | ||
2153 | int | |
2154 | metaslab_load(metaslab_t *msp) | |
2155 | { | |
2156 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
2157 | ||
2158 | /* | |
2159 | * There may be another thread loading the same metaslab, if that's | |
2160 | * the case just wait until the other thread is done and return. | |
2161 | */ | |
2162 | metaslab_load_wait(msp); | |
2163 | if (msp->ms_loaded) | |
2164 | return (0); | |
2165 | VERIFY(!msp->ms_loading); | |
425d3237 | 2166 | ASSERT(!msp->ms_condensing); |
b194fab0 | 2167 | |
93e28d66 SD |
2168 | /* |
2169 | * We set the loading flag BEFORE potentially dropping the lock to | |
2170 | * wait for an ongoing flush (see ms_flushing below). This way other | |
2171 | * threads know that there is already a thread that is loading this | |
2172 | * metaslab. | |
2173 | */ | |
b194fab0 | 2174 | msp->ms_loading = B_TRUE; |
93e28d66 SD |
2175 | |
2176 | /* | |
2177 | * Wait for any in-progress flushing to finish as we drop the ms_lock | |
2178 | * both here (during space_map_load()) and in metaslab_flush() (when | |
2179 | * we flush our changes to the ms_sm). | |
2180 | */ | |
2181 | if (msp->ms_flushing) | |
2182 | metaslab_flush_wait(msp); | |
2183 | ||
2184 | /* | |
2185 | * In the possibility that we were waiting for the metaslab to be | |
2186 | * flushed (where we temporarily dropped the ms_lock), ensure that | |
2187 | * no one else loaded the metaslab somehow. | |
2188 | */ | |
2189 | ASSERT(!msp->ms_loaded); | |
2190 | ||
f09fda50 PD |
2191 | /* |
2192 | * If we're loading a metaslab in the normal class, consider evicting | |
2193 | * another one to keep our memory usage under the limit defined by the | |
2194 | * zfs_metaslab_mem_limit tunable. | |
2195 | */ | |
2196 | if (spa_normal_class(msp->ms_group->mg_class->mc_spa) == | |
2197 | msp->ms_group->mg_class) { | |
2198 | metaslab_potentially_evict(msp->ms_group->mg_class); | |
2199 | } | |
2200 | ||
b194fab0 | 2201 | int error = metaslab_load_impl(msp); |
93e28d66 SD |
2202 | |
2203 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
b194fab0 | 2204 | msp->ms_loading = B_FALSE; |
93cf2076 | 2205 | cv_broadcast(&msp->ms_load_cv); |
b194fab0 | 2206 | |
93cf2076 GW |
2207 | return (error); |
2208 | } | |
2209 | ||
2210 | void | |
2211 | metaslab_unload(metaslab_t *msp) | |
2212 | { | |
2213 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
928e8ad4 | 2214 | |
f09fda50 PD |
2215 | /* |
2216 | * This can happen if a metaslab is selected for eviction (in | |
2217 | * metaslab_potentially_evict) and then unloaded during spa_sync (via | |
2218 | * metaslab_class_evict_old). | |
2219 | */ | |
2220 | if (!msp->ms_loaded) | |
2221 | return; | |
928e8ad4 | 2222 | |
d2734cce | 2223 | range_tree_vacate(msp->ms_allocatable, NULL, NULL); |
93cf2076 | 2224 | msp->ms_loaded = B_FALSE; |
c81f1790 | 2225 | msp->ms_unload_time = gethrtime(); |
928e8ad4 | 2226 | |
679b0f2a | 2227 | msp->ms_activation_weight = 0; |
93cf2076 | 2228 | msp->ms_weight &= ~METASLAB_ACTIVE_MASK; |
928e8ad4 | 2229 | |
f09fda50 PD |
2230 | if (msp->ms_group != NULL) { |
2231 | metaslab_class_t *mc = msp->ms_group->mg_class; | |
2232 | multilist_sublist_t *mls = | |
2233 | multilist_sublist_lock_obj(mc->mc_metaslab_txg_list, msp); | |
2234 | if (multilist_link_active(&msp->ms_class_txg_node)) | |
2235 | multilist_sublist_remove(mls, msp); | |
2236 | multilist_sublist_unlock(mls); | |
2237 | } | |
2238 | ||
928e8ad4 SD |
2239 | /* |
2240 | * We explicitly recalculate the metaslab's weight based on its space | |
2241 | * map (as it is now not loaded). We want unload metaslabs to always | |
2242 | * have their weights calculated from the space map histograms, while | |
2243 | * loaded ones have it calculated from their in-core range tree | |
2244 | * [see metaslab_load()]. This way, the weight reflects the information | |
93e28d66 | 2245 | * available in-core, whether it is loaded or not. |
928e8ad4 SD |
2246 | * |
2247 | * If ms_group == NULL means that we came here from metaslab_fini(), | |
2248 | * at which point it doesn't make sense for us to do the recalculation | |
2249 | * and the sorting. | |
2250 | */ | |
2251 | if (msp->ms_group != NULL) | |
2252 | metaslab_recalculate_weight_and_sort(msp); | |
93cf2076 GW |
2253 | } |
2254 | ||
f09fda50 PD |
2255 | void |
2256 | metaslab_set_selected_txg(metaslab_t *msp, uint64_t txg) | |
2257 | { | |
2258 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
2259 | metaslab_class_t *mc = msp->ms_group->mg_class; | |
2260 | multilist_sublist_t *mls = | |
2261 | multilist_sublist_lock_obj(mc->mc_metaslab_txg_list, msp); | |
2262 | if (multilist_link_active(&msp->ms_class_txg_node)) | |
2263 | multilist_sublist_remove(mls, msp); | |
2264 | msp->ms_selected_txg = txg; | |
eef0f4d8 | 2265 | msp->ms_selected_time = gethrtime(); |
f09fda50 PD |
2266 | multilist_sublist_insert_tail(mls, msp); |
2267 | multilist_sublist_unlock(mls); | |
2268 | } | |
2269 | ||
93e28d66 | 2270 | void |
cc99f275 DB |
2271 | metaslab_space_update(vdev_t *vd, metaslab_class_t *mc, int64_t alloc_delta, |
2272 | int64_t defer_delta, int64_t space_delta) | |
2273 | { | |
2274 | vdev_space_update(vd, alloc_delta, defer_delta, space_delta); | |
2275 | ||
2276 | ASSERT3P(vd->vdev_spa->spa_root_vdev, ==, vd->vdev_parent); | |
2277 | ASSERT(vd->vdev_ms_count != 0); | |
2278 | ||
2279 | metaslab_class_space_update(mc, alloc_delta, defer_delta, space_delta, | |
2280 | vdev_deflated_space(vd, space_delta)); | |
2281 | } | |
2282 | ||
fb42a493 | 2283 | int |
93e28d66 SD |
2284 | metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, |
2285 | uint64_t txg, metaslab_t **msp) | |
34dc7c2f BB |
2286 | { |
2287 | vdev_t *vd = mg->mg_vd; | |
cc99f275 DB |
2288 | spa_t *spa = vd->vdev_spa; |
2289 | objset_t *mos = spa->spa_meta_objset; | |
fb42a493 PS |
2290 | metaslab_t *ms; |
2291 | int error; | |
34dc7c2f | 2292 | |
79c76d5b | 2293 | ms = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP); |
fb42a493 | 2294 | mutex_init(&ms->ms_lock, NULL, MUTEX_DEFAULT, NULL); |
a1d477c2 | 2295 | mutex_init(&ms->ms_sync_lock, NULL, MUTEX_DEFAULT, NULL); |
fb42a493 | 2296 | cv_init(&ms->ms_load_cv, NULL, CV_DEFAULT, NULL); |
93e28d66 | 2297 | cv_init(&ms->ms_flush_cv, NULL, CV_DEFAULT, NULL); |
f09fda50 | 2298 | multilist_link_init(&ms->ms_class_txg_node); |
619f0976 | 2299 | |
fb42a493 PS |
2300 | ms->ms_id = id; |
2301 | ms->ms_start = id << vd->vdev_ms_shift; | |
2302 | ms->ms_size = 1ULL << vd->vdev_ms_shift; | |
492f64e9 PD |
2303 | ms->ms_allocator = -1; |
2304 | ms->ms_new = B_TRUE; | |
34dc7c2f | 2305 | |
93cf2076 GW |
2306 | /* |
2307 | * We only open space map objects that already exist. All others | |
afe37326 | 2308 | * will be opened when we finally allocate an object for it. |
425d3237 SD |
2309 | * |
2310 | * Note: | |
2311 | * When called from vdev_expand(), we can't call into the DMU as | |
2312 | * we are holding the spa_config_lock as a writer and we would | |
2313 | * deadlock [see relevant comment in vdev_metaslab_init()]. in | |
2314 | * that case, the object parameter is zero though, so we won't | |
2315 | * call into the DMU. | |
93cf2076 | 2316 | */ |
afe37326 | 2317 | if (object != 0) { |
fb42a493 | 2318 | error = space_map_open(&ms->ms_sm, mos, object, ms->ms_start, |
a1d477c2 | 2319 | ms->ms_size, vd->vdev_ashift); |
fb42a493 PS |
2320 | |
2321 | if (error != 0) { | |
2322 | kmem_free(ms, sizeof (metaslab_t)); | |
2323 | return (error); | |
2324 | } | |
2325 | ||
2326 | ASSERT(ms->ms_sm != NULL); | |
425d3237 | 2327 | ms->ms_allocated_space = space_map_allocated(ms->ms_sm); |
93cf2076 | 2328 | } |
34dc7c2f BB |
2329 | |
2330 | /* | |
425d3237 | 2331 | * We create the ms_allocatable here, but we don't create the |
258553d3 | 2332 | * other range trees until metaslab_sync_done(). This serves |
34dc7c2f | 2333 | * two purposes: it allows metaslab_sync_done() to detect the |
425d3237 SD |
2334 | * addition of new space; and for debugging, it ensures that |
2335 | * we'd data fault on any attempt to use this metaslab before | |
2336 | * it's ready. | |
34dc7c2f | 2337 | */ |
d2734cce SD |
2338 | ms->ms_allocatable = range_tree_create_impl(&rt_avl_ops, |
2339 | &ms->ms_allocatable_by_size, metaslab_rangesize_compare, 0); | |
34dc7c2f | 2340 | |
1b939560 BB |
2341 | ms->ms_trim = range_tree_create(NULL, NULL); |
2342 | ||
2343 | metaslab_group_add(mg, ms); | |
65a91b16 | 2344 | metaslab_set_fragmentation(ms, B_FALSE); |
428870ff | 2345 | |
34dc7c2f BB |
2346 | /* |
2347 | * If we're opening an existing pool (txg == 0) or creating | |
2348 | * a new one (txg == TXG_INITIAL), all space is available now. | |
2349 | * If we're adding space to an existing pool, the new space | |
2350 | * does not become available until after this txg has synced. | |
4e21fd06 DB |
2351 | * The metaslab's weight will also be initialized when we sync |
2352 | * out this txg. This ensures that we don't attempt to allocate | |
2353 | * from it before we have initialized it completely. | |
34dc7c2f | 2354 | */ |
425d3237 | 2355 | if (txg <= TXG_INITIAL) { |
fb42a493 | 2356 | metaslab_sync_done(ms, 0); |
425d3237 SD |
2357 | metaslab_space_update(vd, mg->mg_class, |
2358 | metaslab_allocated_space(ms), 0, 0); | |
2359 | } | |
34dc7c2f BB |
2360 | |
2361 | if (txg != 0) { | |
34dc7c2f | 2362 | vdev_dirty(vd, 0, NULL, txg); |
fb42a493 | 2363 | vdev_dirty(vd, VDD_METASLAB, ms, txg); |
34dc7c2f BB |
2364 | } |
2365 | ||
fb42a493 PS |
2366 | *msp = ms; |
2367 | ||
2368 | return (0); | |
34dc7c2f BB |
2369 | } |
2370 | ||
93e28d66 SD |
2371 | static void |
2372 | metaslab_fini_flush_data(metaslab_t *msp) | |
2373 | { | |
2374 | spa_t *spa = msp->ms_group->mg_vd->vdev_spa; | |
2375 | ||
2376 | if (metaslab_unflushed_txg(msp) == 0) { | |
2377 | ASSERT3P(avl_find(&spa->spa_metaslabs_by_flushed, msp, NULL), | |
2378 | ==, NULL); | |
2379 | return; | |
2380 | } | |
2381 | ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)); | |
2382 | ||
2383 | mutex_enter(&spa->spa_flushed_ms_lock); | |
2384 | avl_remove(&spa->spa_metaslabs_by_flushed, msp); | |
2385 | mutex_exit(&spa->spa_flushed_ms_lock); | |
2386 | ||
2387 | spa_log_sm_decrement_mscount(spa, metaslab_unflushed_txg(msp)); | |
2388 | spa_log_summary_decrement_mscount(spa, metaslab_unflushed_txg(msp)); | |
2389 | } | |
2390 | ||
2391 | uint64_t | |
2392 | metaslab_unflushed_changes_memused(metaslab_t *ms) | |
2393 | { | |
2394 | return ((range_tree_numsegs(ms->ms_unflushed_allocs) + | |
2395 | range_tree_numsegs(ms->ms_unflushed_frees)) * | |
2396 | sizeof (range_seg_t)); | |
2397 | } | |
2398 | ||
34dc7c2f BB |
2399 | void |
2400 | metaslab_fini(metaslab_t *msp) | |
2401 | { | |
93cf2076 | 2402 | metaslab_group_t *mg = msp->ms_group; |
cc99f275 | 2403 | vdev_t *vd = mg->mg_vd; |
93e28d66 SD |
2404 | spa_t *spa = vd->vdev_spa; |
2405 | ||
2406 | metaslab_fini_flush_data(msp); | |
34dc7c2f BB |
2407 | |
2408 | metaslab_group_remove(mg, msp); | |
2409 | ||
2410 | mutex_enter(&msp->ms_lock); | |
93cf2076 | 2411 | VERIFY(msp->ms_group == NULL); |
cc99f275 | 2412 | metaslab_space_update(vd, mg->mg_class, |
425d3237 | 2413 | -metaslab_allocated_space(msp), 0, -msp->ms_size); |
cc99f275 | 2414 | |
93cf2076 | 2415 | space_map_close(msp->ms_sm); |
93e28d66 | 2416 | msp->ms_sm = NULL; |
93cf2076 GW |
2417 | |
2418 | metaslab_unload(msp); | |
d2734cce SD |
2419 | range_tree_destroy(msp->ms_allocatable); |
2420 | range_tree_destroy(msp->ms_freeing); | |
2421 | range_tree_destroy(msp->ms_freed); | |
34dc7c2f | 2422 | |
93e28d66 SD |
2423 | ASSERT3U(spa->spa_unflushed_stats.sus_memused, >=, |
2424 | metaslab_unflushed_changes_memused(msp)); | |
2425 | spa->spa_unflushed_stats.sus_memused -= | |
2426 | metaslab_unflushed_changes_memused(msp); | |
2427 | range_tree_vacate(msp->ms_unflushed_allocs, NULL, NULL); | |
2428 | range_tree_destroy(msp->ms_unflushed_allocs); | |
2429 | range_tree_vacate(msp->ms_unflushed_frees, NULL, NULL); | |
2430 | range_tree_destroy(msp->ms_unflushed_frees); | |
2431 | ||
1c27024e | 2432 | for (int t = 0; t < TXG_SIZE; t++) { |
d2734cce | 2433 | range_tree_destroy(msp->ms_allocating[t]); |
34dc7c2f BB |
2434 | } |
2435 | ||
1c27024e | 2436 | for (int t = 0; t < TXG_DEFER_SIZE; t++) { |
d2734cce | 2437 | range_tree_destroy(msp->ms_defer[t]); |
e51be066 | 2438 | } |
c99c9001 | 2439 | ASSERT0(msp->ms_deferspace); |
428870ff | 2440 | |
d2734cce SD |
2441 | range_tree_destroy(msp->ms_checkpointing); |
2442 | ||
928e8ad4 SD |
2443 | for (int t = 0; t < TXG_SIZE; t++) |
2444 | ASSERT(!txg_list_member(&vd->vdev_ms_list, msp, t)); | |
2445 | ||
1b939560 BB |
2446 | range_tree_vacate(msp->ms_trim, NULL, NULL); |
2447 | range_tree_destroy(msp->ms_trim); | |
2448 | ||
34dc7c2f | 2449 | mutex_exit(&msp->ms_lock); |
93cf2076 | 2450 | cv_destroy(&msp->ms_load_cv); |
93e28d66 | 2451 | cv_destroy(&msp->ms_flush_cv); |
34dc7c2f | 2452 | mutex_destroy(&msp->ms_lock); |
a1d477c2 | 2453 | mutex_destroy(&msp->ms_sync_lock); |
492f64e9 | 2454 | ASSERT3U(msp->ms_allocator, ==, -1); |
34dc7c2f BB |
2455 | |
2456 | kmem_free(msp, sizeof (metaslab_t)); | |
2457 | } | |
2458 | ||
f3a7f661 GW |
2459 | #define FRAGMENTATION_TABLE_SIZE 17 |
2460 | ||
93cf2076 | 2461 | /* |
f3a7f661 GW |
2462 | * This table defines a segment size based fragmentation metric that will |
2463 | * allow each metaslab to derive its own fragmentation value. This is done | |
2464 | * by calculating the space in each bucket of the spacemap histogram and | |
928e8ad4 | 2465 | * multiplying that by the fragmentation metric in this table. Doing |
f3a7f661 GW |
2466 | * this for all buckets and dividing it by the total amount of free |
2467 | * space in this metaslab (i.e. the total free space in all buckets) gives | |
2468 | * us the fragmentation metric. This means that a high fragmentation metric | |
2469 | * equates to most of the free space being comprised of small segments. | |
2470 | * Conversely, if the metric is low, then most of the free space is in | |
2471 | * large segments. A 10% change in fragmentation equates to approximately | |
2472 | * double the number of segments. | |
93cf2076 | 2473 | * |
f3a7f661 GW |
2474 | * This table defines 0% fragmented space using 16MB segments. Testing has |
2475 | * shown that segments that are greater than or equal to 16MB do not suffer | |
2476 | * from drastic performance problems. Using this value, we derive the rest | |
2477 | * of the table. Since the fragmentation value is never stored on disk, it | |
2478 | * is possible to change these calculations in the future. | |
2479 | */ | |
2480 | int zfs_frag_table[FRAGMENTATION_TABLE_SIZE] = { | |
2481 | 100, /* 512B */ | |
2482 | 100, /* 1K */ | |
2483 | 98, /* 2K */ | |
2484 | 95, /* 4K */ | |
2485 | 90, /* 8K */ | |
2486 | 80, /* 16K */ | |
2487 | 70, /* 32K */ | |
2488 | 60, /* 64K */ | |
2489 | 50, /* 128K */ | |
2490 | 40, /* 256K */ | |
2491 | 30, /* 512K */ | |
2492 | 20, /* 1M */ | |
2493 | 15, /* 2M */ | |
2494 | 10, /* 4M */ | |
2495 | 5, /* 8M */ | |
2496 | 0 /* 16M */ | |
2497 | }; | |
2498 | ||
2499 | /* | |
425d3237 SD |
2500 | * Calculate the metaslab's fragmentation metric and set ms_fragmentation. |
2501 | * Setting this value to ZFS_FRAG_INVALID means that the metaslab has not | |
2502 | * been upgraded and does not support this metric. Otherwise, the return | |
2503 | * value should be in the range [0, 100]. | |
93cf2076 | 2504 | */ |
4e21fd06 | 2505 | static void |
65a91b16 | 2506 | metaslab_set_fragmentation(metaslab_t *msp, boolean_t nodirty) |
93cf2076 | 2507 | { |
f3a7f661 GW |
2508 | spa_t *spa = msp->ms_group->mg_vd->vdev_spa; |
2509 | uint64_t fragmentation = 0; | |
2510 | uint64_t total = 0; | |
2511 | boolean_t feature_enabled = spa_feature_is_enabled(spa, | |
2512 | SPA_FEATURE_SPACEMAP_HISTOGRAM); | |
93cf2076 | 2513 | |
4e21fd06 DB |
2514 | if (!feature_enabled) { |
2515 | msp->ms_fragmentation = ZFS_FRAG_INVALID; | |
2516 | return; | |
2517 | } | |
f3a7f661 | 2518 | |
93cf2076 | 2519 | /* |
f3a7f661 GW |
2520 | * A null space map means that the entire metaslab is free |
2521 | * and thus is not fragmented. | |
93cf2076 | 2522 | */ |
4e21fd06 DB |
2523 | if (msp->ms_sm == NULL) { |
2524 | msp->ms_fragmentation = 0; | |
2525 | return; | |
2526 | } | |
f3a7f661 GW |
2527 | |
2528 | /* | |
4e21fd06 | 2529 | * If this metaslab's space map has not been upgraded, flag it |
f3a7f661 GW |
2530 | * so that we upgrade next time we encounter it. |
2531 | */ | |
2532 | if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t)) { | |
3b7f360c | 2533 | uint64_t txg = spa_syncing_txg(spa); |
93cf2076 GW |
2534 | vdev_t *vd = msp->ms_group->mg_vd; |
2535 | ||
3b7f360c GW |
2536 | /* |
2537 | * If we've reached the final dirty txg, then we must | |
2538 | * be shutting down the pool. We don't want to dirty | |
2539 | * any data past this point so skip setting the condense | |
2540 | * flag. We can retry this action the next time the pool | |
65a91b16 SD |
2541 | * is imported. We also skip marking this metaslab for |
2542 | * condensing if the caller has explicitly set nodirty. | |
3b7f360c | 2543 | */ |
65a91b16 SD |
2544 | if (!nodirty && |
2545 | spa_writeable(spa) && txg < spa_final_dirty_txg(spa)) { | |
8b0a0840 TC |
2546 | msp->ms_condense_wanted = B_TRUE; |
2547 | vdev_dirty(vd, VDD_METASLAB, msp, txg + 1); | |
964c2d69 | 2548 | zfs_dbgmsg("txg %llu, requesting force condense: " |
3b7f360c GW |
2549 | "ms_id %llu, vdev_id %llu", txg, msp->ms_id, |
2550 | vd->vdev_id); | |
8b0a0840 | 2551 | } |
4e21fd06 DB |
2552 | msp->ms_fragmentation = ZFS_FRAG_INVALID; |
2553 | return; | |
93cf2076 GW |
2554 | } |
2555 | ||
1c27024e | 2556 | for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) { |
f3a7f661 GW |
2557 | uint64_t space = 0; |
2558 | uint8_t shift = msp->ms_sm->sm_shift; | |
4e21fd06 | 2559 | |
f3a7f661 GW |
2560 | int idx = MIN(shift - SPA_MINBLOCKSHIFT + i, |
2561 | FRAGMENTATION_TABLE_SIZE - 1); | |
93cf2076 | 2562 | |
93cf2076 GW |
2563 | if (msp->ms_sm->sm_phys->smp_histogram[i] == 0) |
2564 | continue; | |
2565 | ||
f3a7f661 GW |
2566 | space = msp->ms_sm->sm_phys->smp_histogram[i] << (i + shift); |
2567 | total += space; | |
2568 | ||
2569 | ASSERT3U(idx, <, FRAGMENTATION_TABLE_SIZE); | |
2570 | fragmentation += space * zfs_frag_table[idx]; | |
93cf2076 | 2571 | } |
f3a7f661 GW |
2572 | |
2573 | if (total > 0) | |
2574 | fragmentation /= total; | |
2575 | ASSERT3U(fragmentation, <=, 100); | |
4e21fd06 DB |
2576 | |
2577 | msp->ms_fragmentation = fragmentation; | |
93cf2076 | 2578 | } |
34dc7c2f | 2579 | |
f3a7f661 GW |
2580 | /* |
2581 | * Compute a weight -- a selection preference value -- for the given metaslab. | |
2582 | * This is based on the amount of free space, the level of fragmentation, | |
2583 | * the LBA range, and whether the metaslab is loaded. | |
2584 | */ | |
34dc7c2f | 2585 | static uint64_t |
4e21fd06 | 2586 | metaslab_space_weight(metaslab_t *msp) |
34dc7c2f BB |
2587 | { |
2588 | metaslab_group_t *mg = msp->ms_group; | |
34dc7c2f BB |
2589 | vdev_t *vd = mg->mg_vd; |
2590 | uint64_t weight, space; | |
2591 | ||
2592 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
c2e42f9d | 2593 | |
34dc7c2f BB |
2594 | /* |
2595 | * The baseline weight is the metaslab's free space. | |
2596 | */ | |
425d3237 | 2597 | space = msp->ms_size - metaslab_allocated_space(msp); |
f3a7f661 | 2598 | |
f3a7f661 GW |
2599 | if (metaslab_fragmentation_factor_enabled && |
2600 | msp->ms_fragmentation != ZFS_FRAG_INVALID) { | |
2601 | /* | |
2602 | * Use the fragmentation information to inversely scale | |
2603 | * down the baseline weight. We need to ensure that we | |
2604 | * don't exclude this metaslab completely when it's 100% | |
2605 | * fragmented. To avoid this we reduce the fragmented value | |
2606 | * by 1. | |
2607 | */ | |
2608 | space = (space * (100 - (msp->ms_fragmentation - 1))) / 100; | |
2609 | ||
2610 | /* | |
2611 | * If space < SPA_MINBLOCKSIZE, then we will not allocate from | |
2612 | * this metaslab again. The fragmentation metric may have | |
2613 | * decreased the space to something smaller than | |
2614 | * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE | |
2615 | * so that we can consume any remaining space. | |
2616 | */ | |
2617 | if (space > 0 && space < SPA_MINBLOCKSIZE) | |
2618 | space = SPA_MINBLOCKSIZE; | |
2619 | } | |
34dc7c2f BB |
2620 | weight = space; |
2621 | ||
2622 | /* | |
2623 | * Modern disks have uniform bit density and constant angular velocity. | |
2624 | * Therefore, the outer recording zones are faster (higher bandwidth) | |
2625 | * than the inner zones by the ratio of outer to inner track diameter, | |
2626 | * which is typically around 2:1. We account for this by assigning | |
2627 | * higher weight to lower metaslabs (multiplier ranging from 2x to 1x). | |
2628 | * In effect, this means that we'll select the metaslab with the most | |
2629 | * free bandwidth rather than simply the one with the most free space. | |
2630 | */ | |
fb40095f | 2631 | if (!vd->vdev_nonrot && metaslab_lba_weighting_enabled) { |
f3a7f661 GW |
2632 | weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count; |
2633 | ASSERT(weight >= space && weight <= 2 * space); | |
2634 | } | |
428870ff | 2635 | |
f3a7f661 GW |
2636 | /* |
2637 | * If this metaslab is one we're actively using, adjust its | |
2638 | * weight to make it preferable to any inactive metaslab so | |
2639 | * we'll polish it off. If the fragmentation on this metaslab | |
2640 | * has exceed our threshold, then don't mark it active. | |
2641 | */ | |
2642 | if (msp->ms_loaded && msp->ms_fragmentation != ZFS_FRAG_INVALID && | |
2643 | msp->ms_fragmentation <= zfs_metaslab_fragmentation_threshold) { | |
428870ff BB |
2644 | weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK); |
2645 | } | |
34dc7c2f | 2646 | |
4e21fd06 DB |
2647 | WEIGHT_SET_SPACEBASED(weight); |
2648 | return (weight); | |
2649 | } | |
2650 | ||
2651 | /* | |
2652 | * Return the weight of the specified metaslab, according to the segment-based | |
2653 | * weighting algorithm. The metaslab must be loaded. This function can | |
2654 | * be called within a sync pass since it relies only on the metaslab's | |
2655 | * range tree which is always accurate when the metaslab is loaded. | |
2656 | */ | |
2657 | static uint64_t | |
2658 | metaslab_weight_from_range_tree(metaslab_t *msp) | |
2659 | { | |
2660 | uint64_t weight = 0; | |
2661 | uint32_t segments = 0; | |
4e21fd06 DB |
2662 | |
2663 | ASSERT(msp->ms_loaded); | |
2664 | ||
1c27024e DB |
2665 | for (int i = RANGE_TREE_HISTOGRAM_SIZE - 1; i >= SPA_MINBLOCKSHIFT; |
2666 | i--) { | |
4e21fd06 DB |
2667 | uint8_t shift = msp->ms_group->mg_vd->vdev_ashift; |
2668 | int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1; | |
2669 | ||
2670 | segments <<= 1; | |
d2734cce | 2671 | segments += msp->ms_allocatable->rt_histogram[i]; |
4e21fd06 DB |
2672 | |
2673 | /* | |
2674 | * The range tree provides more precision than the space map | |
2675 | * and must be downgraded so that all values fit within the | |
2676 | * space map's histogram. This allows us to compare loaded | |
2677 | * vs. unloaded metaslabs to determine which metaslab is | |
2678 | * considered "best". | |
2679 | */ | |
2680 | if (i > max_idx) | |
2681 | continue; | |
2682 | ||
2683 | if (segments != 0) { | |
2684 | WEIGHT_SET_COUNT(weight, segments); | |
2685 | WEIGHT_SET_INDEX(weight, i); | |
2686 | WEIGHT_SET_ACTIVE(weight, 0); | |
2687 | break; | |
2688 | } | |
2689 | } | |
2690 | return (weight); | |
2691 | } | |
2692 | ||
2693 | /* | |
93e28d66 SD |
2694 | * Calculate the weight based on the on-disk histogram. Should be applied |
2695 | * only to unloaded metaslabs (i.e no incoming allocations) in-order to | |
2696 | * give results consistent with the on-disk state | |
4e21fd06 DB |
2697 | */ |
2698 | static uint64_t | |
2699 | metaslab_weight_from_spacemap(metaslab_t *msp) | |
2700 | { | |
928e8ad4 SD |
2701 | space_map_t *sm = msp->ms_sm; |
2702 | ASSERT(!msp->ms_loaded); | |
2703 | ASSERT(sm != NULL); | |
2704 | ASSERT3U(space_map_object(sm), !=, 0); | |
2705 | ASSERT3U(sm->sm_dbuf->db_size, ==, sizeof (space_map_phys_t)); | |
4e21fd06 | 2706 | |
928e8ad4 SD |
2707 | /* |
2708 | * Create a joint histogram from all the segments that have made | |
2709 | * it to the metaslab's space map histogram, that are not yet | |
2710 | * available for allocation because they are still in the freeing | |
2711 | * pipeline (e.g. freeing, freed, and defer trees). Then subtract | |
2712 | * these segments from the space map's histogram to get a more | |
2713 | * accurate weight. | |
2714 | */ | |
2715 | uint64_t deferspace_histogram[SPACE_MAP_HISTOGRAM_SIZE] = {0}; | |
2716 | for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) | |
2717 | deferspace_histogram[i] += msp->ms_synchist[i]; | |
2718 | for (int t = 0; t < TXG_DEFER_SIZE; t++) { | |
2719 | for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) { | |
2720 | deferspace_histogram[i] += msp->ms_deferhist[t][i]; | |
2721 | } | |
2722 | } | |
2723 | ||
2724 | uint64_t weight = 0; | |
1c27024e | 2725 | for (int i = SPACE_MAP_HISTOGRAM_SIZE - 1; i >= 0; i--) { |
928e8ad4 SD |
2726 | ASSERT3U(sm->sm_phys->smp_histogram[i], >=, |
2727 | deferspace_histogram[i]); | |
2728 | uint64_t count = | |
2729 | sm->sm_phys->smp_histogram[i] - deferspace_histogram[i]; | |
2730 | if (count != 0) { | |
2731 | WEIGHT_SET_COUNT(weight, count); | |
2732 | WEIGHT_SET_INDEX(weight, i + sm->sm_shift); | |
4e21fd06 DB |
2733 | WEIGHT_SET_ACTIVE(weight, 0); |
2734 | break; | |
2735 | } | |
2736 | } | |
2737 | return (weight); | |
2738 | } | |
2739 | ||
2740 | /* | |
2741 | * Compute a segment-based weight for the specified metaslab. The weight | |
2742 | * is determined by highest bucket in the histogram. The information | |
2743 | * for the highest bucket is encoded into the weight value. | |
2744 | */ | |
2745 | static uint64_t | |
2746 | metaslab_segment_weight(metaslab_t *msp) | |
2747 | { | |
2748 | metaslab_group_t *mg = msp->ms_group; | |
2749 | uint64_t weight = 0; | |
2750 | uint8_t shift = mg->mg_vd->vdev_ashift; | |
2751 | ||
2752 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
2753 | ||
2754 | /* | |
2755 | * The metaslab is completely free. | |
2756 | */ | |
425d3237 | 2757 | if (metaslab_allocated_space(msp) == 0) { |
4e21fd06 DB |
2758 | int idx = highbit64(msp->ms_size) - 1; |
2759 | int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1; | |
2760 | ||
2761 | if (idx < max_idx) { | |
2762 | WEIGHT_SET_COUNT(weight, 1ULL); | |
2763 | WEIGHT_SET_INDEX(weight, idx); | |
2764 | } else { | |
2765 | WEIGHT_SET_COUNT(weight, 1ULL << (idx - max_idx)); | |
2766 | WEIGHT_SET_INDEX(weight, max_idx); | |
2767 | } | |
2768 | WEIGHT_SET_ACTIVE(weight, 0); | |
2769 | ASSERT(!WEIGHT_IS_SPACEBASED(weight)); | |
4e21fd06 DB |
2770 | return (weight); |
2771 | } | |
2772 | ||
2773 | ASSERT3U(msp->ms_sm->sm_dbuf->db_size, ==, sizeof (space_map_phys_t)); | |
2774 | ||
2775 | /* | |
2776 | * If the metaslab is fully allocated then just make the weight 0. | |
2777 | */ | |
425d3237 | 2778 | if (metaslab_allocated_space(msp) == msp->ms_size) |
4e21fd06 DB |
2779 | return (0); |
2780 | /* | |
2781 | * If the metaslab is already loaded, then use the range tree to | |
2782 | * determine the weight. Otherwise, we rely on the space map information | |
2783 | * to generate the weight. | |
2784 | */ | |
2785 | if (msp->ms_loaded) { | |
2786 | weight = metaslab_weight_from_range_tree(msp); | |
2787 | } else { | |
2788 | weight = metaslab_weight_from_spacemap(msp); | |
2789 | } | |
2790 | ||
2791 | /* | |
2792 | * If the metaslab was active the last time we calculated its weight | |
2793 | * then keep it active. We want to consume the entire region that | |
2794 | * is associated with this weight. | |
2795 | */ | |
2796 | if (msp->ms_activation_weight != 0 && weight != 0) | |
2797 | WEIGHT_SET_ACTIVE(weight, WEIGHT_GET_ACTIVE(msp->ms_weight)); | |
2798 | return (weight); | |
2799 | } | |
2800 | ||
2801 | /* | |
2802 | * Determine if we should attempt to allocate from this metaslab. If the | |
7f319089 SD |
2803 | * metaslab is loaded, then we can determine if the desired allocation |
2804 | * can be satisfied by looking at the size of the maximum free segment | |
2805 | * on that metaslab. Otherwise, we make our decision based on the metaslab's | |
2806 | * weight. For segment-based weighting we can determine the maximum | |
2807 | * allocation based on the index encoded in its value. For space-based | |
2808 | * weights we rely on the entire weight (excluding the weight-type bit). | |
4e21fd06 DB |
2809 | */ |
2810 | boolean_t | |
c81f1790 | 2811 | metaslab_should_allocate(metaslab_t *msp, uint64_t asize, boolean_t try_hard) |
4e21fd06 | 2812 | { |
c81f1790 PD |
2813 | /* |
2814 | * If the metaslab is loaded, ms_max_size is definitive and we can use | |
2815 | * the fast check. If it's not, the ms_max_size is a lower bound (once | |
2816 | * set), and we should use the fast check as long as we're not in | |
2817 | * try_hard and it's been less than zfs_metaslab_max_size_cache_sec | |
2818 | * seconds since the metaslab was unloaded. | |
2819 | */ | |
2820 | if (msp->ms_loaded || | |
2821 | (msp->ms_max_size != 0 && !try_hard && gethrtime() < | |
2822 | msp->ms_unload_time + SEC2NSEC(zfs_metaslab_max_size_cache_sec))) | |
4e21fd06 DB |
2823 | return (msp->ms_max_size >= asize); |
2824 | ||
679b0f2a | 2825 | boolean_t should_allocate; |
4e21fd06 DB |
2826 | if (!WEIGHT_IS_SPACEBASED(msp->ms_weight)) { |
2827 | /* | |
2828 | * The metaslab segment weight indicates segments in the | |
2829 | * range [2^i, 2^(i+1)), where i is the index in the weight. | |
2830 | * Since the asize might be in the middle of the range, we | |
2831 | * should attempt the allocation if asize < 2^(i+1). | |
2832 | */ | |
2833 | should_allocate = (asize < | |
2834 | 1ULL << (WEIGHT_GET_INDEX(msp->ms_weight) + 1)); | |
2835 | } else { | |
2836 | should_allocate = (asize <= | |
2837 | (msp->ms_weight & ~METASLAB_WEIGHT_TYPE)); | |
2838 | } | |
679b0f2a | 2839 | |
4e21fd06 DB |
2840 | return (should_allocate); |
2841 | } | |
65a91b16 | 2842 | |
4e21fd06 | 2843 | static uint64_t |
65a91b16 | 2844 | metaslab_weight(metaslab_t *msp, boolean_t nodirty) |
4e21fd06 DB |
2845 | { |
2846 | vdev_t *vd = msp->ms_group->mg_vd; | |
2847 | spa_t *spa = vd->vdev_spa; | |
2848 | uint64_t weight; | |
2849 | ||
2850 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
2851 | ||
65a91b16 | 2852 | metaslab_set_fragmentation(msp, nodirty); |
4e21fd06 DB |
2853 | |
2854 | /* | |
c81f1790 | 2855 | * Update the maximum size. If the metaslab is loaded, this will |
4e21fd06 | 2856 | * ensure that we get an accurate maximum size if newly freed space |
c81f1790 PD |
2857 | * has been added back into the free tree. If the metaslab is |
2858 | * unloaded, we check if there's a larger free segment in the | |
2859 | * unflushed frees. This is a lower bound on the largest allocatable | |
2860 | * segment size. Coalescing of adjacent entries may reveal larger | |
2861 | * allocatable segments, but we aren't aware of those until loading | |
2862 | * the space map into a range tree. | |
4e21fd06 | 2863 | */ |
c81f1790 PD |
2864 | if (msp->ms_loaded) { |
2865 | msp->ms_max_size = metaslab_largest_allocatable(msp); | |
2866 | } else { | |
2867 | msp->ms_max_size = MAX(msp->ms_max_size, | |
2868 | metaslab_largest_unflushed_free(msp)); | |
2869 | } | |
4e21fd06 DB |
2870 | |
2871 | /* | |
2872 | * Segment-based weighting requires space map histogram support. | |
2873 | */ | |
2874 | if (zfs_metaslab_segment_weight_enabled && | |
2875 | spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) && | |
2876 | (msp->ms_sm == NULL || msp->ms_sm->sm_dbuf->db_size == | |
2877 | sizeof (space_map_phys_t))) { | |
2878 | weight = metaslab_segment_weight(msp); | |
2879 | } else { | |
2880 | weight = metaslab_space_weight(msp); | |
2881 | } | |
93cf2076 | 2882 | return (weight); |
34dc7c2f BB |
2883 | } |
2884 | ||
928e8ad4 SD |
2885 | void |
2886 | metaslab_recalculate_weight_and_sort(metaslab_t *msp) | |
2887 | { | |
679b0f2a PD |
2888 | ASSERT(MUTEX_HELD(&msp->ms_lock)); |
2889 | ||
928e8ad4 SD |
2890 | /* note: we preserve the mask (e.g. indication of primary, etc..) */ |
2891 | uint64_t was_active = msp->ms_weight & METASLAB_ACTIVE_MASK; | |
2892 | metaslab_group_sort(msp->ms_group, msp, | |
65a91b16 | 2893 | metaslab_weight(msp, B_FALSE) | was_active); |
928e8ad4 SD |
2894 | } |
2895 | ||
34dc7c2f | 2896 | static int |
492f64e9 PD |
2897 | metaslab_activate_allocator(metaslab_group_t *mg, metaslab_t *msp, |
2898 | int allocator, uint64_t activation_weight) | |
2899 | { | |
679b0f2a PD |
2900 | ASSERT(MUTEX_HELD(&msp->ms_lock)); |
2901 | ||
492f64e9 PD |
2902 | /* |
2903 | * If we're activating for the claim code, we don't want to actually | |
2904 | * set the metaslab up for a specific allocator. | |
2905 | */ | |
f09fda50 PD |
2906 | if (activation_weight == METASLAB_WEIGHT_CLAIM) { |
2907 | ASSERT0(msp->ms_activation_weight); | |
2908 | msp->ms_activation_weight = msp->ms_weight; | |
2909 | metaslab_group_sort(mg, msp, msp->ms_weight | | |
2910 | activation_weight); | |
492f64e9 | 2911 | return (0); |
f09fda50 | 2912 | } |
679b0f2a | 2913 | |
492f64e9 PD |
2914 | metaslab_t **arr = (activation_weight == METASLAB_WEIGHT_PRIMARY ? |
2915 | mg->mg_primaries : mg->mg_secondaries); | |
2916 | ||
492f64e9 PD |
2917 | mutex_enter(&mg->mg_lock); |
2918 | if (arr[allocator] != NULL) { | |
2919 | mutex_exit(&mg->mg_lock); | |
2920 | return (EEXIST); | |
2921 | } | |
2922 | ||
2923 | arr[allocator] = msp; | |
2924 | ASSERT3S(msp->ms_allocator, ==, -1); | |
2925 | msp->ms_allocator = allocator; | |
2926 | msp->ms_primary = (activation_weight == METASLAB_WEIGHT_PRIMARY); | |
f09fda50 PD |
2927 | |
2928 | ASSERT0(msp->ms_activation_weight); | |
2929 | msp->ms_activation_weight = msp->ms_weight; | |
2930 | metaslab_group_sort_impl(mg, msp, | |
2931 | msp->ms_weight | activation_weight); | |
2932 | ||
492f64e9 PD |
2933 | mutex_exit(&mg->mg_lock); |
2934 | ||
2935 | return (0); | |
2936 | } | |
2937 | ||
2938 | static int | |
2939 | metaslab_activate(metaslab_t *msp, int allocator, uint64_t activation_weight) | |
34dc7c2f | 2940 | { |
34dc7c2f BB |
2941 | ASSERT(MUTEX_HELD(&msp->ms_lock)); |
2942 | ||
679b0f2a PD |
2943 | /* |
2944 | * The current metaslab is already activated for us so there | |
2945 | * is nothing to do. Already activated though, doesn't mean | |
2946 | * that this metaslab is activated for our allocator nor our | |
2947 | * requested activation weight. The metaslab could have started | |
2948 | * as an active one for our allocator but changed allocators | |
2949 | * while we were waiting to grab its ms_lock or we stole it | |
2950 | * [see find_valid_metaslab()]. This means that there is a | |
2951 | * possibility of passivating a metaslab of another allocator | |
2952 | * or from a different activation mask, from this thread. | |
2953 | */ | |
2954 | if ((msp->ms_weight & METASLAB_ACTIVE_MASK) != 0) { | |
2955 | ASSERT(msp->ms_loaded); | |
2956 | return (0); | |
2957 | } | |
2958 | ||
2959 | int error = metaslab_load(msp); | |
2960 | if (error != 0) { | |
2961 | metaslab_group_sort(msp->ms_group, msp, 0); | |
2962 | return (error); | |
2963 | } | |
2964 | ||
2965 | /* | |
2966 | * When entering metaslab_load() we may have dropped the | |
2967 | * ms_lock because we were loading this metaslab, or we | |
2968 | * were waiting for another thread to load it for us. In | |
2969 | * that scenario, we recheck the weight of the metaslab | |
2970 | * to see if it was activated by another thread. | |
2971 | * | |
2972 | * If the metaslab was activated for another allocator or | |
2973 | * it was activated with a different activation weight (e.g. | |
2974 | * we wanted to make it a primary but it was activated as | |
2975 | * secondary) we return error (EBUSY). | |
2976 | * | |
2977 | * If the metaslab was activated for the same allocator | |
2978 | * and requested activation mask, skip activating it. | |
2979 | */ | |
2980 | if ((msp->ms_weight & METASLAB_ACTIVE_MASK) != 0) { | |
2981 | if (msp->ms_allocator != allocator) | |
2982 | return (EBUSY); | |
2983 | ||
2984 | if ((msp->ms_weight & activation_weight) == 0) | |
7ab96299 | 2985 | return (SET_ERROR(EBUSY)); |
9babb374 | 2986 | |
679b0f2a PD |
2987 | EQUIV((activation_weight == METASLAB_WEIGHT_PRIMARY), |
2988 | msp->ms_primary); | |
2989 | return (0); | |
34dc7c2f | 2990 | } |
679b0f2a | 2991 | |
fe0ea848 PD |
2992 | /* |
2993 | * If the metaslab has literally 0 space, it will have weight 0. In | |
2994 | * that case, don't bother activating it. This can happen if the | |
2995 | * metaslab had space during find_valid_metaslab, but another thread | |
2996 | * loaded it and used all that space while we were waiting to grab the | |
2997 | * lock. | |
2998 | */ | |
2999 | if (msp->ms_weight == 0) { | |
3000 | ASSERT0(range_tree_space(msp->ms_allocatable)); | |
3001 | return (SET_ERROR(ENOSPC)); | |
3002 | } | |
3003 | ||
679b0f2a PD |
3004 | if ((error = metaslab_activate_allocator(msp->ms_group, msp, |
3005 | allocator, activation_weight)) != 0) { | |
3006 | return (error); | |
3007 | } | |
3008 | ||
93cf2076 | 3009 | ASSERT(msp->ms_loaded); |
34dc7c2f BB |
3010 | ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK); |
3011 | ||
3012 | return (0); | |
3013 | } | |
3014 | ||
492f64e9 PD |
3015 | static void |
3016 | metaslab_passivate_allocator(metaslab_group_t *mg, metaslab_t *msp, | |
3017 | uint64_t weight) | |
3018 | { | |
3019 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
679b0f2a PD |
3020 | ASSERT(msp->ms_loaded); |
3021 | ||
492f64e9 PD |
3022 | if (msp->ms_weight & METASLAB_WEIGHT_CLAIM) { |
3023 | metaslab_group_sort(mg, msp, weight); | |
3024 | return; | |
3025 | } | |
3026 | ||
3027 | mutex_enter(&mg->mg_lock); | |
3028 | ASSERT3P(msp->ms_group, ==, mg); | |
679b0f2a PD |
3029 | ASSERT3S(0, <=, msp->ms_allocator); |
3030 | ASSERT3U(msp->ms_allocator, <, mg->mg_allocators); | |
3031 | ||
492f64e9 | 3032 | if (msp->ms_primary) { |
492f64e9 PD |
3033 | ASSERT3P(mg->mg_primaries[msp->ms_allocator], ==, msp); |
3034 | ASSERT(msp->ms_weight & METASLAB_WEIGHT_PRIMARY); | |
3035 | mg->mg_primaries[msp->ms_allocator] = NULL; | |
3036 | } else { | |
492f64e9 | 3037 | ASSERT3P(mg->mg_secondaries[msp->ms_allocator], ==, msp); |
679b0f2a | 3038 | ASSERT(msp->ms_weight & METASLAB_WEIGHT_SECONDARY); |
492f64e9 PD |
3039 | mg->mg_secondaries[msp->ms_allocator] = NULL; |
3040 | } | |
3041 | msp->ms_allocator = -1; | |
3042 | metaslab_group_sort_impl(mg, msp, weight); | |
3043 | mutex_exit(&mg->mg_lock); | |
3044 | } | |
3045 | ||
34dc7c2f | 3046 | static void |
4e21fd06 | 3047 | metaslab_passivate(metaslab_t *msp, uint64_t weight) |
34dc7c2f | 3048 | { |
4e21fd06 DB |
3049 | ASSERTV(uint64_t size = weight & ~METASLAB_WEIGHT_TYPE); |
3050 | ||
34dc7c2f BB |
3051 | /* |
3052 | * If size < SPA_MINBLOCKSIZE, then we will not allocate from | |
3053 | * this metaslab again. In that case, it had better be empty, | |
3054 | * or we would be leaving space on the table. | |
3055 | */ | |
94d49e8f TC |
3056 | ASSERT(!WEIGHT_IS_SPACEBASED(msp->ms_weight) || |
3057 | size >= SPA_MINBLOCKSIZE || | |
d2734cce | 3058 | range_tree_space(msp->ms_allocatable) == 0); |
4e21fd06 DB |
3059 | ASSERT0(weight & METASLAB_ACTIVE_MASK); |
3060 | ||
679b0f2a | 3061 | ASSERT(msp->ms_activation_weight != 0); |
4e21fd06 | 3062 | msp->ms_activation_weight = 0; |
492f64e9 | 3063 | metaslab_passivate_allocator(msp->ms_group, msp, weight); |
679b0f2a | 3064 | ASSERT0(msp->ms_weight & METASLAB_ACTIVE_MASK); |
34dc7c2f BB |
3065 | } |
3066 | ||
4e21fd06 DB |
3067 | /* |
3068 | * Segment-based metaslabs are activated once and remain active until | |
3069 | * we either fail an allocation attempt (similar to space-based metaslabs) | |
3070 | * or have exhausted the free space in zfs_metaslab_switch_threshold | |
3071 | * buckets since the metaslab was activated. This function checks to see | |
e1cfd73f | 3072 | * if we've exhausted the zfs_metaslab_switch_threshold buckets in the |
4e21fd06 DB |
3073 | * metaslab and passivates it proactively. This will allow us to select a |
3074 | * metaslab with a larger contiguous region, if any, remaining within this | |
3075 | * metaslab group. If we're in sync pass > 1, then we continue using this | |
3076 | * metaslab so that we don't dirty more block and cause more sync passes. | |
3077 | */ | |
3078 | void | |
3079 | metaslab_segment_may_passivate(metaslab_t *msp) | |
3080 | { | |
3081 | spa_t *spa = msp->ms_group->mg_vd->vdev_spa; | |
4e21fd06 DB |
3082 | |
3083 | if (WEIGHT_IS_SPACEBASED(msp->ms_weight) || spa_sync_pass(spa) > 1) | |
3084 | return; | |
3085 | ||
3086 | /* | |
3087 | * Since we are in the middle of a sync pass, the most accurate | |
3088 | * information that is accessible to us is the in-core range tree | |
3089 | * histogram; calculate the new weight based on that information. | |
3090 | */ | |
1c27024e DB |
3091 | uint64_t weight = metaslab_weight_from_range_tree(msp); |
3092 | int activation_idx = WEIGHT_GET_INDEX(msp->ms_activation_weight); | |
3093 | int current_idx = WEIGHT_GET_INDEX(weight); | |
4e21fd06 DB |
3094 | |
3095 | if (current_idx <= activation_idx - zfs_metaslab_switch_threshold) | |
3096 | metaslab_passivate(msp, weight); | |
3097 | } | |
3098 | ||
93cf2076 GW |
3099 | static void |
3100 | metaslab_preload(void *arg) | |
3101 | { | |
3102 | metaslab_t *msp = arg; | |
f09fda50 PD |
3103 | metaslab_class_t *mc = msp->ms_group->mg_class; |
3104 | spa_t *spa = mc->mc_spa; | |
1cd77734 | 3105 | fstrans_cookie_t cookie = spl_fstrans_mark(); |
93cf2076 | 3106 | |
080b3100 GW |
3107 | ASSERT(!MUTEX_HELD(&msp->ms_group->mg_lock)); |
3108 | ||
93cf2076 | 3109 | mutex_enter(&msp->ms_lock); |
b194fab0 | 3110 | (void) metaslab_load(msp); |
f09fda50 | 3111 | metaslab_set_selected_txg(msp, spa_syncing_txg(spa)); |
93cf2076 | 3112 | mutex_exit(&msp->ms_lock); |
1cd77734 | 3113 | spl_fstrans_unmark(cookie); |
93cf2076 GW |
3114 | } |
3115 | ||
3116 | static void | |
3117 | metaslab_group_preload(metaslab_group_t *mg) | |
3118 | { | |
3119 | spa_t *spa = mg->mg_vd->vdev_spa; | |
3120 | metaslab_t *msp; | |
3121 | avl_tree_t *t = &mg->mg_metaslab_tree; | |
3122 | int m = 0; | |
3123 | ||
3124 | if (spa_shutting_down(spa) || !metaslab_preload_enabled) { | |
c5528b9b | 3125 | taskq_wait_outstanding(mg->mg_taskq, 0); |
93cf2076 GW |
3126 | return; |
3127 | } | |
93cf2076 | 3128 | |
080b3100 | 3129 | mutex_enter(&mg->mg_lock); |
a1d477c2 | 3130 | |
93cf2076 | 3131 | /* |
080b3100 | 3132 | * Load the next potential metaslabs |
93cf2076 | 3133 | */ |
4e21fd06 | 3134 | for (msp = avl_first(t); msp != NULL; msp = AVL_NEXT(t, msp)) { |
a1d477c2 MA |
3135 | ASSERT3P(msp->ms_group, ==, mg); |
3136 | ||
f3a7f661 GW |
3137 | /* |
3138 | * We preload only the maximum number of metaslabs specified | |
3139 | * by metaslab_preload_limit. If a metaslab is being forced | |
3140 | * to condense then we preload it too. This will ensure | |
3141 | * that force condensing happens in the next txg. | |
3142 | */ | |
3143 | if (++m > metaslab_preload_limit && !msp->ms_condense_wanted) { | |
f3a7f661 GW |
3144 | continue; |
3145 | } | |
93cf2076 GW |
3146 | |
3147 | VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload, | |
48d3eb40 | 3148 | msp, TQ_SLEEP) != TASKQID_INVALID); |
93cf2076 GW |
3149 | } |
3150 | mutex_exit(&mg->mg_lock); | |
3151 | } | |
3152 | ||
e51be066 | 3153 | /* |
93e28d66 SD |
3154 | * Determine if the space map's on-disk footprint is past our tolerance for |
3155 | * inefficiency. We would like to use the following criteria to make our | |
3156 | * decision: | |
e51be066 | 3157 | * |
93e28d66 SD |
3158 | * 1. Do not condense if the size of the space map object would dramatically |
3159 | * increase as a result of writing out the free space range tree. | |
e51be066 | 3160 | * |
93e28d66 SD |
3161 | * 2. Condense if the on on-disk space map representation is at least |
3162 | * zfs_condense_pct/100 times the size of the optimal representation | |
3163 | * (i.e. zfs_condense_pct = 110 and in-core = 1MB, optimal = 1.1MB). | |
e51be066 | 3164 | * |
93e28d66 SD |
3165 | * 3. Do not condense if the on-disk size of the space map does not actually |
3166 | * decrease. | |
b02fe35d | 3167 | * |
b02fe35d AR |
3168 | * Unfortunately, we cannot compute the on-disk size of the space map in this |
3169 | * context because we cannot accurately compute the effects of compression, etc. | |
3170 | * Instead, we apply the heuristic described in the block comment for | |
3171 | * zfs_metaslab_condense_block_threshold - we only condense if the space used | |
3172 | * is greater than a threshold number of blocks. | |
e51be066 GW |
3173 | */ |
3174 | static boolean_t | |
3175 | metaslab_should_condense(metaslab_t *msp) | |
3176 | { | |
93cf2076 | 3177 | space_map_t *sm = msp->ms_sm; |
d2734cce SD |
3178 | vdev_t *vd = msp->ms_group->mg_vd; |
3179 | uint64_t vdev_blocksize = 1 << vd->vdev_ashift; | |
e51be066 GW |
3180 | |
3181 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
93cf2076 | 3182 | ASSERT(msp->ms_loaded); |
93e28d66 SD |
3183 | ASSERT(sm != NULL); |
3184 | ASSERT3U(spa_sync_pass(vd->vdev_spa), ==, 1); | |
d2734cce SD |
3185 | |
3186 | /* | |
4d044c4c SD |
3187 | * We always condense metaslabs that are empty and metaslabs for |
3188 | * which a condense request has been made. | |
e51be066 | 3189 | */ |
4d044c4c SD |
3190 | if (avl_is_empty(&msp->ms_allocatable_by_size) || |
3191 | msp->ms_condense_wanted) | |
e51be066 GW |
3192 | return (B_TRUE); |
3193 | ||
93e28d66 SD |
3194 | uint64_t record_size = MAX(sm->sm_blksz, vdev_blocksize); |
3195 | uint64_t object_size = space_map_length(sm); | |
4d044c4c SD |
3196 | uint64_t optimal_size = space_map_estimate_optimal_size(sm, |
3197 | msp->ms_allocatable, SM_NO_VDEVID); | |
b02fe35d | 3198 | |
4d044c4c | 3199 | return (object_size >= (optimal_size * zfs_condense_pct / 100) && |
b02fe35d | 3200 | object_size > zfs_metaslab_condense_block_threshold * record_size); |
e51be066 GW |
3201 | } |
3202 | ||
3203 | /* | |
3204 | * Condense the on-disk space map representation to its minimized form. | |
93e28d66 SD |
3205 | * The minimized form consists of a small number of allocations followed |
3206 | * by the entries of the free range tree (ms_allocatable). The condensed | |
3207 | * spacemap contains all the entries of previous TXGs (including those in | |
3208 | * the pool-wide log spacemaps; thus this is effectively a superset of | |
3209 | * metaslab_flush()), but this TXG's entries still need to be written. | |
e51be066 GW |
3210 | */ |
3211 | static void | |
93e28d66 | 3212 | metaslab_condense(metaslab_t *msp, dmu_tx_t *tx) |
e51be066 | 3213 | { |
93cf2076 GW |
3214 | range_tree_t *condense_tree; |
3215 | space_map_t *sm = msp->ms_sm; | |
93e28d66 SD |
3216 | uint64_t txg = dmu_tx_get_txg(tx); |
3217 | spa_t *spa = msp->ms_group->mg_vd->vdev_spa; | |
e51be066 GW |
3218 | |
3219 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
93cf2076 | 3220 | ASSERT(msp->ms_loaded); |
93e28d66 | 3221 | ASSERT(msp->ms_sm != NULL); |
e51be066 | 3222 | |
93e28d66 SD |
3223 | /* |
3224 | * In order to condense the space map, we need to change it so it | |
3225 | * only describes which segments are currently allocated and free. | |
3226 | * | |
3227 | * All the current free space resides in the ms_allocatable, all | |
3228 | * the ms_defer trees, and all the ms_allocating trees. We ignore | |
3229 | * ms_freed because it is empty because we're in sync pass 1. We | |
3230 | * ignore ms_freeing because these changes are not yet reflected | |
3231 | * in the spacemap (they will be written later this txg). | |
3232 | * | |
3233 | * So to truncate the space map to represent all the entries of | |
3234 | * previous TXGs we do the following: | |
3235 | * | |
3236 | * 1] We create a range tree (condense tree) that is 100% allocated. | |
3237 | * 2] We remove from it all segments found in the ms_defer trees | |
3238 | * as those segments are marked as free in the original space | |
3239 | * map. We do the same with the ms_allocating trees for the same | |
3240 | * reason. Removing these segments should be a relatively | |
3241 | * inexpensive operation since we expect these trees to have a | |
3242 | * small number of nodes. | |
3243 | * 3] We vacate any unflushed allocs as they should already exist | |
3244 | * in the condense tree. Then we vacate any unflushed frees as | |
3245 | * they should already be part of ms_allocatable. | |
3246 | * 4] At this point, we would ideally like to remove all segments | |
3247 | * in the ms_allocatable tree from the condense tree. This way | |
3248 | * we would write all the entries of the condense tree as the | |
3249 | * condensed space map, which would only contain allocated | |
3250 | * segments with everything else assumed to be freed. | |
3251 | * | |
3252 | * Doing so can be prohibitively expensive as ms_allocatable can | |
3253 | * be large, and therefore computationally expensive to subtract | |
3254 | * from the condense_tree. Instead we first sync out the | |
3255 | * condense_tree and then the ms_allocatable, in the condensed | |
3256 | * space map. While this is not optimal, it is typically close to | |
3257 | * optimal and more importantly much cheaper to compute. | |
3258 | * | |
3259 | * 5] Finally, as both of the unflushed trees were written to our | |
3260 | * new and condensed metaslab space map, we basically flushed | |
3261 | * all the unflushed changes to disk, thus we call | |
3262 | * metaslab_flush_update(). | |
3263 | */ | |
3264 | ASSERT3U(spa_sync_pass(spa), ==, 1); | |
3265 | ASSERT(range_tree_is_empty(msp->ms_freed)); /* since it is pass 1 */ | |
f3a7f661 | 3266 | |
a887d653 | 3267 | zfs_dbgmsg("condensing: txg %llu, msp[%llu] %px, vdev id %llu, " |
5f3d9c69 JS |
3268 | "spa %s, smp size %llu, segments %lu, forcing condense=%s", txg, |
3269 | msp->ms_id, msp, msp->ms_group->mg_vd->vdev_id, | |
93e28d66 | 3270 | spa->spa_name, space_map_length(msp->ms_sm), |
d2734cce | 3271 | avl_numnodes(&msp->ms_allocatable->rt_root), |
f3a7f661 GW |
3272 | msp->ms_condense_wanted ? "TRUE" : "FALSE"); |
3273 | ||
3274 | msp->ms_condense_wanted = B_FALSE; | |
e51be066 | 3275 | |
a1d477c2 | 3276 | condense_tree = range_tree_create(NULL, NULL); |
93cf2076 | 3277 | range_tree_add(condense_tree, msp->ms_start, msp->ms_size); |
e51be066 | 3278 | |
1c27024e | 3279 | for (int t = 0; t < TXG_DEFER_SIZE; t++) { |
d2734cce | 3280 | range_tree_walk(msp->ms_defer[t], |
93cf2076 GW |
3281 | range_tree_remove, condense_tree); |
3282 | } | |
e51be066 | 3283 | |
93e28d66 | 3284 | for (int t = 0; t < TXG_CONCURRENT_STATES; t++) { |
d2734cce | 3285 | range_tree_walk(msp->ms_allocating[(txg + t) & TXG_MASK], |
93cf2076 GW |
3286 | range_tree_remove, condense_tree); |
3287 | } | |
e51be066 | 3288 | |
93e28d66 SD |
3289 | ASSERT3U(spa->spa_unflushed_stats.sus_memused, >=, |
3290 | metaslab_unflushed_changes_memused(msp)); | |
3291 | spa->spa_unflushed_stats.sus_memused -= | |
3292 | metaslab_unflushed_changes_memused(msp); | |
3293 | range_tree_vacate(msp->ms_unflushed_allocs, NULL, NULL); | |
3294 | range_tree_vacate(msp->ms_unflushed_frees, NULL, NULL); | |
3295 | ||
e51be066 | 3296 | /* |
93e28d66 SD |
3297 | * We're about to drop the metaslab's lock thus allowing other |
3298 | * consumers to change it's content. Set the metaslab's ms_condensing | |
3299 | * flag to ensure that allocations on this metaslab do not occur | |
3300 | * while we're in the middle of committing it to disk. This is only | |
3301 | * critical for ms_allocatable as all other range trees use per TXG | |
e51be066 GW |
3302 | * views of their content. |
3303 | */ | |
93cf2076 | 3304 | msp->ms_condensing = B_TRUE; |
e51be066 GW |
3305 | |
3306 | mutex_exit(&msp->ms_lock); | |
93e28d66 SD |
3307 | uint64_t object = space_map_object(msp->ms_sm); |
3308 | space_map_truncate(sm, | |
3309 | spa_feature_is_enabled(spa, SPA_FEATURE_LOG_SPACEMAP) ? | |
3310 | zfs_metaslab_sm_blksz_with_log : zfs_metaslab_sm_blksz_no_log, tx); | |
3311 | ||
3312 | /* | |
3313 | * space_map_truncate() may have reallocated the spacemap object. | |
3314 | * If so, update the vdev_ms_array. | |
3315 | */ | |
3316 | if (space_map_object(msp->ms_sm) != object) { | |
3317 | object = space_map_object(msp->ms_sm); | |
3318 | dmu_write(spa->spa_meta_objset, | |
3319 | msp->ms_group->mg_vd->vdev_ms_array, sizeof (uint64_t) * | |
3320 | msp->ms_id, sizeof (uint64_t), &object, tx); | |
3321 | } | |
e51be066 GW |
3322 | |
3323 | /* | |
93e28d66 SD |
3324 | * Note: |
3325 | * When the log space map feature is enabled, each space map will | |
3326 | * always have ALLOCS followed by FREES for each sync pass. This is | |
3327 | * typically true even when the log space map feature is disabled, | |
3328 | * except from the case where a metaslab goes through metaslab_sync() | |
3329 | * and gets condensed. In that case the metaslab's space map will have | |
3330 | * ALLOCS followed by FREES (due to condensing) followed by ALLOCS | |
3331 | * followed by FREES (due to space_map_write() in metaslab_sync()) for | |
3332 | * sync pass 1. | |
e51be066 | 3333 | */ |
4d044c4c | 3334 | space_map_write(sm, condense_tree, SM_ALLOC, SM_NO_VDEVID, tx); |
93e28d66 SD |
3335 | space_map_write(sm, msp->ms_allocatable, SM_FREE, SM_NO_VDEVID, tx); |
3336 | ||
93cf2076 GW |
3337 | range_tree_vacate(condense_tree, NULL, NULL); |
3338 | range_tree_destroy(condense_tree); | |
a1d477c2 | 3339 | mutex_enter(&msp->ms_lock); |
93e28d66 | 3340 | |
93cf2076 | 3341 | msp->ms_condensing = B_FALSE; |
93e28d66 SD |
3342 | metaslab_flush_update(msp, tx); |
3343 | } | |
3344 | ||
3345 | /* | |
3346 | * Called when the metaslab has been flushed (its own spacemap now reflects | |
3347 | * all the contents of the pool-wide spacemap log). Updates the metaslab's | |
3348 | * metadata and any pool-wide related log space map data (e.g. summary, | |
3349 | * obsolete logs, etc..) to reflect that. | |
3350 | */ | |
3351 | static void | |
3352 | metaslab_flush_update(metaslab_t *msp, dmu_tx_t *tx) | |
3353 | { | |
3354 | metaslab_group_t *mg = msp->ms_group; | |
3355 | spa_t *spa = mg->mg_vd->vdev_spa; | |
3356 | ||
3357 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
3358 | ||
3359 | ASSERT3U(spa_sync_pass(spa), ==, 1); | |
3360 | ASSERT(range_tree_is_empty(msp->ms_unflushed_allocs)); | |
3361 | ASSERT(range_tree_is_empty(msp->ms_unflushed_frees)); | |
3362 | ||
3363 | /* | |
3364 | * Just because a metaslab got flushed, that doesn't mean that | |
3365 | * it will pass through metaslab_sync_done(). Thus, make sure to | |
3366 | * update ms_synced_length here in case it doesn't. | |
3367 | */ | |
3368 | msp->ms_synced_length = space_map_length(msp->ms_sm); | |
3369 | ||
3370 | /* | |
3371 | * We may end up here from metaslab_condense() without the | |
3372 | * feature being active. In that case this is a no-op. | |
3373 | */ | |
3374 | if (!spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) | |
3375 | return; | |
3376 | ||
3377 | ASSERT(spa_syncing_log_sm(spa) != NULL); | |
3378 | ASSERT(msp->ms_sm != NULL); | |
3379 | ASSERT(metaslab_unflushed_txg(msp) != 0); | |
3380 | ASSERT3P(avl_find(&spa->spa_metaslabs_by_flushed, msp, NULL), ==, msp); | |
3381 | ||
3382 | VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(spa)); | |
3383 | ||
3384 | /* update metaslab's position in our flushing tree */ | |
3385 | uint64_t ms_prev_flushed_txg = metaslab_unflushed_txg(msp); | |
3386 | mutex_enter(&spa->spa_flushed_ms_lock); | |
3387 | avl_remove(&spa->spa_metaslabs_by_flushed, msp); | |
3388 | metaslab_set_unflushed_txg(msp, spa_syncing_txg(spa), tx); | |
3389 | avl_add(&spa->spa_metaslabs_by_flushed, msp); | |
3390 | mutex_exit(&spa->spa_flushed_ms_lock); | |
3391 | ||
3392 | /* update metaslab counts of spa_log_sm_t nodes */ | |
3393 | spa_log_sm_decrement_mscount(spa, ms_prev_flushed_txg); | |
3394 | spa_log_sm_increment_current_mscount(spa); | |
3395 | ||
3396 | /* cleanup obsolete logs if any */ | |
3397 | uint64_t log_blocks_before = spa_log_sm_nblocks(spa); | |
3398 | spa_cleanup_old_sm_logs(spa, tx); | |
3399 | uint64_t log_blocks_after = spa_log_sm_nblocks(spa); | |
3400 | VERIFY3U(log_blocks_after, <=, log_blocks_before); | |
3401 | ||
3402 | /* update log space map summary */ | |
3403 | uint64_t blocks_gone = log_blocks_before - log_blocks_after; | |
3404 | spa_log_summary_add_flushed_metaslab(spa); | |
3405 | spa_log_summary_decrement_mscount(spa, ms_prev_flushed_txg); | |
3406 | spa_log_summary_decrement_blkcount(spa, blocks_gone); | |
3407 | } | |
3408 | ||
3409 | boolean_t | |
3410 | metaslab_flush(metaslab_t *msp, dmu_tx_t *tx) | |
3411 | { | |
3412 | spa_t *spa = msp->ms_group->mg_vd->vdev_spa; | |
3413 | ||
3414 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
3415 | ASSERT3U(spa_sync_pass(spa), ==, 1); | |
3416 | ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)); | |
3417 | ||
3418 | ASSERT(msp->ms_sm != NULL); | |
3419 | ASSERT(metaslab_unflushed_txg(msp) != 0); | |
3420 | ASSERT(avl_find(&spa->spa_metaslabs_by_flushed, msp, NULL) != NULL); | |
3421 | ||
3422 | /* | |
3423 | * There is nothing wrong with flushing the same metaslab twice, as | |
3424 | * this codepath should work on that case. However, the current | |
3425 | * flushing scheme makes sure to avoid this situation as we would be | |
3426 | * making all these calls without having anything meaningful to write | |
3427 | * to disk. We assert this behavior here. | |
3428 | */ | |
3429 | ASSERT3U(metaslab_unflushed_txg(msp), <, dmu_tx_get_txg(tx)); | |
3430 | ||
3431 | /* | |
3432 | * We can not flush while loading, because then we would | |
3433 | * not load the ms_unflushed_{allocs,frees}. | |
3434 | */ | |
3435 | if (msp->ms_loading) | |
3436 | return (B_FALSE); | |
3437 | ||
3438 | metaslab_verify_space(msp, dmu_tx_get_txg(tx)); | |
3439 | metaslab_verify_weight_and_frag(msp); | |
3440 | ||
3441 | /* | |
3442 | * Metaslab condensing is effectively flushing. Therefore if the | |
3443 | * metaslab can be condensed we can just condense it instead of | |
3444 | * flushing it. | |
3445 | * | |
3446 | * Note that metaslab_condense() does call metaslab_flush_update() | |
3447 | * so we can just return immediately after condensing. We also | |
3448 | * don't need to care about setting ms_flushing or broadcasting | |
3449 | * ms_flush_cv, even if we temporarily drop the ms_lock in | |
3450 | * metaslab_condense(), as the metaslab is already loaded. | |
3451 | */ | |
3452 | if (msp->ms_loaded && metaslab_should_condense(msp)) { | |
3453 | metaslab_group_t *mg = msp->ms_group; | |
3454 | ||
3455 | /* | |
3456 | * For all histogram operations below refer to the | |
3457 | * comments of metaslab_sync() where we follow a | |
3458 | * similar procedure. | |
3459 | */ | |
3460 | metaslab_group_histogram_verify(mg); | |
3461 | metaslab_class_histogram_verify(mg->mg_class); | |
3462 | metaslab_group_histogram_remove(mg, msp); | |
3463 | ||
3464 | metaslab_condense(msp, tx); | |
3465 | ||
3466 | space_map_histogram_clear(msp->ms_sm); | |
3467 | space_map_histogram_add(msp->ms_sm, msp->ms_allocatable, tx); | |
3468 | ASSERT(range_tree_is_empty(msp->ms_freed)); | |
3469 | for (int t = 0; t < TXG_DEFER_SIZE; t++) { | |
3470 | space_map_histogram_add(msp->ms_sm, | |
3471 | msp->ms_defer[t], tx); | |
3472 | } | |
3473 | metaslab_aux_histograms_update(msp); | |
3474 | ||
3475 | metaslab_group_histogram_add(mg, msp); | |
3476 | metaslab_group_histogram_verify(mg); | |
3477 | metaslab_class_histogram_verify(mg->mg_class); | |
3478 | ||
3479 | metaslab_verify_space(msp, dmu_tx_get_txg(tx)); | |
3480 | ||
3481 | /* | |
3482 | * Since we recreated the histogram (and potentially | |
3483 | * the ms_sm too while condensing) ensure that the | |
3484 | * weight is updated too because we are not guaranteed | |
3485 | * that this metaslab is dirty and will go through | |
3486 | * metaslab_sync_done(). | |
3487 | */ | |
3488 | metaslab_recalculate_weight_and_sort(msp); | |
3489 | return (B_TRUE); | |
3490 | } | |
3491 | ||
3492 | msp->ms_flushing = B_TRUE; | |
3493 | uint64_t sm_len_before = space_map_length(msp->ms_sm); | |
3494 | ||
3495 | mutex_exit(&msp->ms_lock); | |
3496 | space_map_write(msp->ms_sm, msp->ms_unflushed_allocs, SM_ALLOC, | |
3497 | SM_NO_VDEVID, tx); | |
3498 | space_map_write(msp->ms_sm, msp->ms_unflushed_frees, SM_FREE, | |
3499 | SM_NO_VDEVID, tx); | |
3500 | mutex_enter(&msp->ms_lock); | |
3501 | ||
3502 | uint64_t sm_len_after = space_map_length(msp->ms_sm); | |
3503 | if (zfs_flags & ZFS_DEBUG_LOG_SPACEMAP) { | |
3504 | zfs_dbgmsg("flushing: txg %llu, spa %s, vdev_id %llu, " | |
3505 | "ms_id %llu, unflushed_allocs %llu, unflushed_frees %llu, " | |
3506 | "appended %llu bytes", dmu_tx_get_txg(tx), spa_name(spa), | |
3507 | msp->ms_group->mg_vd->vdev_id, msp->ms_id, | |
3508 | range_tree_space(msp->ms_unflushed_allocs), | |
3509 | range_tree_space(msp->ms_unflushed_frees), | |
3510 | (sm_len_after - sm_len_before)); | |
3511 | } | |
3512 | ||
3513 | ASSERT3U(spa->spa_unflushed_stats.sus_memused, >=, | |
3514 | metaslab_unflushed_changes_memused(msp)); | |
3515 | spa->spa_unflushed_stats.sus_memused -= | |
3516 | metaslab_unflushed_changes_memused(msp); | |
3517 | range_tree_vacate(msp->ms_unflushed_allocs, NULL, NULL); | |
3518 | range_tree_vacate(msp->ms_unflushed_frees, NULL, NULL); | |
3519 | ||
3520 | metaslab_verify_space(msp, dmu_tx_get_txg(tx)); | |
3521 | metaslab_verify_weight_and_frag(msp); | |
3522 | ||
3523 | metaslab_flush_update(msp, tx); | |
3524 | ||
3525 | metaslab_verify_space(msp, dmu_tx_get_txg(tx)); | |
3526 | metaslab_verify_weight_and_frag(msp); | |
3527 | ||
3528 | msp->ms_flushing = B_FALSE; | |
3529 | cv_broadcast(&msp->ms_flush_cv); | |
3530 | return (B_TRUE); | |
e51be066 GW |
3531 | } |
3532 | ||
34dc7c2f BB |
3533 | /* |
3534 | * Write a metaslab to disk in the context of the specified transaction group. | |
3535 | */ | |
3536 | void | |
3537 | metaslab_sync(metaslab_t *msp, uint64_t txg) | |
3538 | { | |
93cf2076 GW |
3539 | metaslab_group_t *mg = msp->ms_group; |
3540 | vdev_t *vd = mg->mg_vd; | |
34dc7c2f | 3541 | spa_t *spa = vd->vdev_spa; |
428870ff | 3542 | objset_t *mos = spa_meta_objset(spa); |
d2734cce | 3543 | range_tree_t *alloctree = msp->ms_allocating[txg & TXG_MASK]; |
34dc7c2f | 3544 | dmu_tx_t *tx; |
34dc7c2f | 3545 | |
428870ff BB |
3546 | ASSERT(!vd->vdev_ishole); |
3547 | ||
e51be066 GW |
3548 | /* |
3549 | * This metaslab has just been added so there's no work to do now. | |
3550 | */ | |
d2734cce | 3551 | if (msp->ms_freeing == NULL) { |
93cf2076 | 3552 | ASSERT3P(alloctree, ==, NULL); |
e51be066 GW |
3553 | return; |
3554 | } | |
3555 | ||
93cf2076 | 3556 | ASSERT3P(alloctree, !=, NULL); |
d2734cce SD |
3557 | ASSERT3P(msp->ms_freeing, !=, NULL); |
3558 | ASSERT3P(msp->ms_freed, !=, NULL); | |
3559 | ASSERT3P(msp->ms_checkpointing, !=, NULL); | |
1b939560 | 3560 | ASSERT3P(msp->ms_trim, !=, NULL); |
e51be066 | 3561 | |
f3a7f661 | 3562 | /* |
d2734cce SD |
3563 | * Normally, we don't want to process a metaslab if there are no |
3564 | * allocations or frees to perform. However, if the metaslab is being | |
475aa97c PD |
3565 | * forced to condense, it's loaded and we're not beyond the final |
3566 | * dirty txg, we need to let it through. Not condensing beyond the | |
3567 | * final dirty txg prevents an issue where metaslabs that need to be | |
3568 | * condensed but were loaded for other reasons could cause a panic | |
3569 | * here. By only checking the txg in that branch of the conditional, | |
3570 | * we preserve the utility of the VERIFY statements in all other | |
3571 | * cases. | |
f3a7f661 | 3572 | */ |
d2734cce SD |
3573 | if (range_tree_is_empty(alloctree) && |
3574 | range_tree_is_empty(msp->ms_freeing) && | |
3575 | range_tree_is_empty(msp->ms_checkpointing) && | |
475aa97c PD |
3576 | !(msp->ms_loaded && msp->ms_condense_wanted && |
3577 | txg <= spa_final_dirty_txg(spa))) | |
428870ff | 3578 | return; |
34dc7c2f | 3579 | |
3b7f360c GW |
3580 | |
3581 | VERIFY(txg <= spa_final_dirty_txg(spa)); | |
3582 | ||
34dc7c2f | 3583 | /* |
425d3237 SD |
3584 | * The only state that can actually be changing concurrently |
3585 | * with metaslab_sync() is the metaslab's ms_allocatable. No | |
3586 | * other thread can be modifying this txg's alloc, freeing, | |
d2734cce | 3587 | * freed, or space_map_phys_t. We drop ms_lock whenever we |
425d3237 SD |
3588 | * could call into the DMU, because the DMU can call down to |
3589 | * us (e.g. via zio_free()) at any time. | |
a1d477c2 MA |
3590 | * |
3591 | * The spa_vdev_remove_thread() can be reading metaslab state | |
425d3237 SD |
3592 | * concurrently, and it is locked out by the ms_sync_lock. |
3593 | * Note that the ms_lock is insufficient for this, because it | |
3594 | * is dropped by space_map_write(). | |
34dc7c2f | 3595 | */ |
428870ff | 3596 | tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg); |
34dc7c2f | 3597 | |
93e28d66 SD |
3598 | /* |
3599 | * Generate a log space map if one doesn't exist already. | |
3600 | */ | |
3601 | spa_generate_syncing_log_sm(spa, tx); | |
93cf2076 | 3602 | |
93e28d66 SD |
3603 | if (msp->ms_sm == NULL) { |
3604 | uint64_t new_object = space_map_alloc(mos, | |
3605 | spa_feature_is_enabled(spa, SPA_FEATURE_LOG_SPACEMAP) ? | |
3606 | zfs_metaslab_sm_blksz_with_log : | |
3607 | zfs_metaslab_sm_blksz_no_log, tx); | |
93cf2076 GW |
3608 | VERIFY3U(new_object, !=, 0); |
3609 | ||
93e28d66 SD |
3610 | dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) * |
3611 | msp->ms_id, sizeof (uint64_t), &new_object, tx); | |
3612 | ||
93cf2076 | 3613 | VERIFY0(space_map_open(&msp->ms_sm, mos, new_object, |
a1d477c2 | 3614 | msp->ms_start, msp->ms_size, vd->vdev_ashift)); |
93cf2076 | 3615 | ASSERT(msp->ms_sm != NULL); |
93e28d66 SD |
3616 | |
3617 | ASSERT(range_tree_is_empty(msp->ms_unflushed_allocs)); | |
3618 | ASSERT(range_tree_is_empty(msp->ms_unflushed_frees)); | |
425d3237 | 3619 | ASSERT0(metaslab_allocated_space(msp)); |
34dc7c2f BB |
3620 | } |
3621 | ||
93e28d66 SD |
3622 | if (metaslab_unflushed_txg(msp) == 0 && |
3623 | spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) { | |
3624 | ASSERT(spa_syncing_log_sm(spa) != NULL); | |
3625 | ||
3626 | metaslab_set_unflushed_txg(msp, spa_syncing_txg(spa), tx); | |
3627 | spa_log_sm_increment_current_mscount(spa); | |
3628 | spa_log_summary_add_flushed_metaslab(spa); | |
3629 | ||
3630 | ASSERT(msp->ms_sm != NULL); | |
3631 | mutex_enter(&spa->spa_flushed_ms_lock); | |
3632 | avl_add(&spa->spa_metaslabs_by_flushed, msp); | |
3633 | mutex_exit(&spa->spa_flushed_ms_lock); | |
3634 | ||
3635 | ASSERT(range_tree_is_empty(msp->ms_unflushed_allocs)); | |
3636 | ASSERT(range_tree_is_empty(msp->ms_unflushed_frees)); | |
3637 | } | |
3638 | ||
d2734cce SD |
3639 | if (!range_tree_is_empty(msp->ms_checkpointing) && |
3640 | vd->vdev_checkpoint_sm == NULL) { | |
3641 | ASSERT(spa_has_checkpoint(spa)); | |
3642 | ||
3643 | uint64_t new_object = space_map_alloc(mos, | |
93e28d66 | 3644 | zfs_vdev_standard_sm_blksz, tx); |
d2734cce SD |
3645 | VERIFY3U(new_object, !=, 0); |
3646 | ||
3647 | VERIFY0(space_map_open(&vd->vdev_checkpoint_sm, | |
3648 | mos, new_object, 0, vd->vdev_asize, vd->vdev_ashift)); | |
3649 | ASSERT3P(vd->vdev_checkpoint_sm, !=, NULL); | |
3650 | ||
3651 | /* | |
3652 | * We save the space map object as an entry in vdev_top_zap | |
3653 | * so it can be retrieved when the pool is reopened after an | |
3654 | * export or through zdb. | |
3655 | */ | |
3656 | VERIFY0(zap_add(vd->vdev_spa->spa_meta_objset, | |
3657 | vd->vdev_top_zap, VDEV_TOP_ZAP_POOL_CHECKPOINT_SM, | |
3658 | sizeof (new_object), 1, &new_object, tx)); | |
3659 | } | |
3660 | ||
a1d477c2 | 3661 | mutex_enter(&msp->ms_sync_lock); |
428870ff BB |
3662 | mutex_enter(&msp->ms_lock); |
3663 | ||
96358617 | 3664 | /* |
4e21fd06 DB |
3665 | * Note: metaslab_condense() clears the space map's histogram. |
3666 | * Therefore we must verify and remove this histogram before | |
96358617 MA |
3667 | * condensing. |
3668 | */ | |
3669 | metaslab_group_histogram_verify(mg); | |
3670 | metaslab_class_histogram_verify(mg->mg_class); | |
3671 | metaslab_group_histogram_remove(mg, msp); | |
3672 | ||
93e28d66 SD |
3673 | if (spa->spa_sync_pass == 1 && msp->ms_loaded && |
3674 | metaslab_should_condense(msp)) | |
3675 | metaslab_condense(msp, tx); | |
3676 | ||
3677 | /* | |
3678 | * We'll be going to disk to sync our space accounting, thus we | |
3679 | * drop the ms_lock during that time so allocations coming from | |
3680 | * open-context (ZIL) for future TXGs do not block. | |
3681 | */ | |
3682 | mutex_exit(&msp->ms_lock); | |
3683 | space_map_t *log_sm = spa_syncing_log_sm(spa); | |
3684 | if (log_sm != NULL) { | |
3685 | ASSERT(spa_feature_is_enabled(spa, SPA_FEATURE_LOG_SPACEMAP)); | |
3686 | ||
3687 | space_map_write(log_sm, alloctree, SM_ALLOC, | |
3688 | vd->vdev_id, tx); | |
3689 | space_map_write(log_sm, msp->ms_freeing, SM_FREE, | |
3690 | vd->vdev_id, tx); | |
3691 | mutex_enter(&msp->ms_lock); | |
3692 | ||
3693 | ASSERT3U(spa->spa_unflushed_stats.sus_memused, >=, | |
3694 | metaslab_unflushed_changes_memused(msp)); | |
3695 | spa->spa_unflushed_stats.sus_memused -= | |
3696 | metaslab_unflushed_changes_memused(msp); | |
3697 | range_tree_remove_xor_add(alloctree, | |
3698 | msp->ms_unflushed_frees, msp->ms_unflushed_allocs); | |
3699 | range_tree_remove_xor_add(msp->ms_freeing, | |
3700 | msp->ms_unflushed_allocs, msp->ms_unflushed_frees); | |
3701 | spa->spa_unflushed_stats.sus_memused += | |
3702 | metaslab_unflushed_changes_memused(msp); | |
e51be066 | 3703 | } else { |
93e28d66 SD |
3704 | ASSERT(!spa_feature_is_enabled(spa, SPA_FEATURE_LOG_SPACEMAP)); |
3705 | ||
4d044c4c SD |
3706 | space_map_write(msp->ms_sm, alloctree, SM_ALLOC, |
3707 | SM_NO_VDEVID, tx); | |
3708 | space_map_write(msp->ms_sm, msp->ms_freeing, SM_FREE, | |
3709 | SM_NO_VDEVID, tx); | |
a1d477c2 | 3710 | mutex_enter(&msp->ms_lock); |
e51be066 | 3711 | } |
428870ff | 3712 | |
425d3237 SD |
3713 | msp->ms_allocated_space += range_tree_space(alloctree); |
3714 | ASSERT3U(msp->ms_allocated_space, >=, | |
3715 | range_tree_space(msp->ms_freeing)); | |
3716 | msp->ms_allocated_space -= range_tree_space(msp->ms_freeing); | |
3717 | ||
d2734cce SD |
3718 | if (!range_tree_is_empty(msp->ms_checkpointing)) { |
3719 | ASSERT(spa_has_checkpoint(spa)); | |
3720 | ASSERT3P(vd->vdev_checkpoint_sm, !=, NULL); | |
3721 | ||
3722 | /* | |
3723 | * Since we are doing writes to disk and the ms_checkpointing | |
3724 | * tree won't be changing during that time, we drop the | |
93e28d66 SD |
3725 | * ms_lock while writing to the checkpoint space map, for the |
3726 | * same reason mentioned above. | |
d2734cce SD |
3727 | */ |
3728 | mutex_exit(&msp->ms_lock); | |
3729 | space_map_write(vd->vdev_checkpoint_sm, | |
4d044c4c | 3730 | msp->ms_checkpointing, SM_FREE, SM_NO_VDEVID, tx); |
d2734cce | 3731 | mutex_enter(&msp->ms_lock); |
d2734cce SD |
3732 | |
3733 | spa->spa_checkpoint_info.sci_dspace += | |
3734 | range_tree_space(msp->ms_checkpointing); | |
3735 | vd->vdev_stat.vs_checkpoint_space += | |
3736 | range_tree_space(msp->ms_checkpointing); | |
3737 | ASSERT3U(vd->vdev_stat.vs_checkpoint_space, ==, | |
425d3237 | 3738 | -space_map_allocated(vd->vdev_checkpoint_sm)); |
d2734cce SD |
3739 | |
3740 | range_tree_vacate(msp->ms_checkpointing, NULL, NULL); | |
3741 | } | |
3742 | ||
93cf2076 GW |
3743 | if (msp->ms_loaded) { |
3744 | /* | |
a1d477c2 | 3745 | * When the space map is loaded, we have an accurate |
93cf2076 GW |
3746 | * histogram in the range tree. This gives us an opportunity |
3747 | * to bring the space map's histogram up-to-date so we clear | |
3748 | * it first before updating it. | |
3749 | */ | |
3750 | space_map_histogram_clear(msp->ms_sm); | |
d2734cce | 3751 | space_map_histogram_add(msp->ms_sm, msp->ms_allocatable, tx); |
4e21fd06 DB |
3752 | |
3753 | /* | |
3754 | * Since we've cleared the histogram we need to add back | |
3755 | * any free space that has already been processed, plus | |
3756 | * any deferred space. This allows the on-disk histogram | |
3757 | * to accurately reflect all free space even if some space | |
3758 | * is not yet available for allocation (i.e. deferred). | |
3759 | */ | |
d2734cce | 3760 | space_map_histogram_add(msp->ms_sm, msp->ms_freed, tx); |
4e21fd06 | 3761 | |
93cf2076 | 3762 | /* |
4e21fd06 DB |
3763 | * Add back any deferred free space that has not been |
3764 | * added back into the in-core free tree yet. This will | |
3765 | * ensure that we don't end up with a space map histogram | |
3766 | * that is completely empty unless the metaslab is fully | |
3767 | * allocated. | |
93cf2076 | 3768 | */ |
1c27024e | 3769 | for (int t = 0; t < TXG_DEFER_SIZE; t++) { |
4e21fd06 | 3770 | space_map_histogram_add(msp->ms_sm, |
d2734cce | 3771 | msp->ms_defer[t], tx); |
4e21fd06 | 3772 | } |
93cf2076 | 3773 | } |
4e21fd06 DB |
3774 | |
3775 | /* | |
3776 | * Always add the free space from this sync pass to the space | |
3777 | * map histogram. We want to make sure that the on-disk histogram | |
3778 | * accounts for all free space. If the space map is not loaded, | |
3779 | * then we will lose some accuracy but will correct it the next | |
3780 | * time we load the space map. | |
3781 | */ | |
d2734cce | 3782 | space_map_histogram_add(msp->ms_sm, msp->ms_freeing, tx); |
928e8ad4 | 3783 | metaslab_aux_histograms_update(msp); |
4e21fd06 | 3784 | |
f3a7f661 GW |
3785 | metaslab_group_histogram_add(mg, msp); |
3786 | metaslab_group_histogram_verify(mg); | |
3787 | metaslab_class_histogram_verify(mg->mg_class); | |
34dc7c2f | 3788 | |
e51be066 | 3789 | /* |
93cf2076 | 3790 | * For sync pass 1, we avoid traversing this txg's free range tree |
425d3237 SD |
3791 | * and instead will just swap the pointers for freeing and freed. |
3792 | * We can safely do this since the freed_tree is guaranteed to be | |
3793 | * empty on the initial pass. | |
93e28d66 SD |
3794 | * |
3795 | * Keep in mind that even if we are currently using a log spacemap | |
3796 | * we want current frees to end up in the ms_allocatable (but not | |
3797 | * get appended to the ms_sm) so their ranges can be reused as usual. | |
e51be066 GW |
3798 | */ |
3799 | if (spa_sync_pass(spa) == 1) { | |
d2734cce | 3800 | range_tree_swap(&msp->ms_freeing, &msp->ms_freed); |
425d3237 | 3801 | ASSERT0(msp->ms_allocated_this_txg); |
e51be066 | 3802 | } else { |
d2734cce SD |
3803 | range_tree_vacate(msp->ms_freeing, |
3804 | range_tree_add, msp->ms_freed); | |
34dc7c2f | 3805 | } |
425d3237 | 3806 | msp->ms_allocated_this_txg += range_tree_space(alloctree); |
f3a7f661 | 3807 | range_tree_vacate(alloctree, NULL, NULL); |
34dc7c2f | 3808 | |
d2734cce SD |
3809 | ASSERT0(range_tree_space(msp->ms_allocating[txg & TXG_MASK])); |
3810 | ASSERT0(range_tree_space(msp->ms_allocating[TXG_CLEAN(txg) | |
3811 | & TXG_MASK])); | |
3812 | ASSERT0(range_tree_space(msp->ms_freeing)); | |
3813 | ASSERT0(range_tree_space(msp->ms_checkpointing)); | |
34dc7c2f BB |
3814 | |
3815 | mutex_exit(&msp->ms_lock); | |
3816 | ||
93e28d66 SD |
3817 | /* |
3818 | * Verify that the space map object ID has been recorded in the | |
3819 | * vdev_ms_array. | |
3820 | */ | |
3821 | uint64_t object; | |
3822 | VERIFY0(dmu_read(mos, vd->vdev_ms_array, | |
3823 | msp->ms_id * sizeof (uint64_t), sizeof (uint64_t), &object, 0)); | |
3824 | VERIFY3U(object, ==, space_map_object(msp->ms_sm)); | |
3825 | ||
a1d477c2 | 3826 | mutex_exit(&msp->ms_sync_lock); |
34dc7c2f BB |
3827 | dmu_tx_commit(tx); |
3828 | } | |
3829 | ||
f09fda50 PD |
3830 | static void |
3831 | metaslab_evict(metaslab_t *msp, uint64_t txg) | |
893a6d62 | 3832 | { |
f09fda50 PD |
3833 | if (!msp->ms_loaded || msp->ms_disabled != 0) |
3834 | return; | |
893a6d62 | 3835 | |
f09fda50 PD |
3836 | for (int t = 1; t < TXG_CONCURRENT_STATES; t++) { |
3837 | VERIFY0(range_tree_space( | |
3838 | msp->ms_allocating[(txg + t) & TXG_MASK])); | |
893a6d62 | 3839 | } |
f09fda50 PD |
3840 | if (msp->ms_allocator != -1) |
3841 | metaslab_passivate(msp, msp->ms_weight & ~METASLAB_ACTIVE_MASK); | |
3842 | ||
3843 | if (!metaslab_debug_unload) | |
3844 | metaslab_unload(msp); | |
893a6d62 PD |
3845 | } |
3846 | ||
34dc7c2f BB |
3847 | /* |
3848 | * Called after a transaction group has completely synced to mark | |
3849 | * all of the metaslab's free space as usable. | |
3850 | */ | |
3851 | void | |
3852 | metaslab_sync_done(metaslab_t *msp, uint64_t txg) | |
3853 | { | |
34dc7c2f BB |
3854 | metaslab_group_t *mg = msp->ms_group; |
3855 | vdev_t *vd = mg->mg_vd; | |
4e21fd06 | 3856 | spa_t *spa = vd->vdev_spa; |
93cf2076 | 3857 | range_tree_t **defer_tree; |
428870ff | 3858 | int64_t alloc_delta, defer_delta; |
4e21fd06 | 3859 | boolean_t defer_allowed = B_TRUE; |
428870ff BB |
3860 | |
3861 | ASSERT(!vd->vdev_ishole); | |
34dc7c2f BB |
3862 | |
3863 | mutex_enter(&msp->ms_lock); | |
3864 | ||
3865 | /* | |
3866 | * If this metaslab is just becoming available, initialize its | |
258553d3 | 3867 | * range trees and add its capacity to the vdev. |
34dc7c2f | 3868 | */ |
d2734cce | 3869 | if (msp->ms_freed == NULL) { |
1c27024e | 3870 | for (int t = 0; t < TXG_SIZE; t++) { |
d2734cce | 3871 | ASSERT(msp->ms_allocating[t] == NULL); |
93cf2076 | 3872 | |
d2734cce | 3873 | msp->ms_allocating[t] = range_tree_create(NULL, NULL); |
34dc7c2f | 3874 | } |
428870ff | 3875 | |
d2734cce SD |
3876 | ASSERT3P(msp->ms_freeing, ==, NULL); |
3877 | msp->ms_freeing = range_tree_create(NULL, NULL); | |
258553d3 | 3878 | |
d2734cce SD |
3879 | ASSERT3P(msp->ms_freed, ==, NULL); |
3880 | msp->ms_freed = range_tree_create(NULL, NULL); | |
258553d3 | 3881 | |
1c27024e | 3882 | for (int t = 0; t < TXG_DEFER_SIZE; t++) { |
93e28d66 | 3883 | ASSERT3P(msp->ms_defer[t], ==, NULL); |
d2734cce | 3884 | msp->ms_defer[t] = range_tree_create(NULL, NULL); |
93cf2076 | 3885 | } |
428870ff | 3886 | |
d2734cce SD |
3887 | ASSERT3P(msp->ms_checkpointing, ==, NULL); |
3888 | msp->ms_checkpointing = range_tree_create(NULL, NULL); | |
3889 | ||
93e28d66 SD |
3890 | ASSERT3P(msp->ms_unflushed_allocs, ==, NULL); |
3891 | msp->ms_unflushed_allocs = range_tree_create(NULL, NULL); | |
3892 | ASSERT3P(msp->ms_unflushed_frees, ==, NULL); | |
c81f1790 PD |
3893 | msp->ms_unflushed_frees = range_tree_create_impl(&rt_avl_ops, |
3894 | &msp->ms_unflushed_frees_by_size, | |
3895 | metaslab_rangesize_compare, 0); | |
93e28d66 | 3896 | |
cc99f275 | 3897 | metaslab_space_update(vd, mg->mg_class, 0, 0, msp->ms_size); |
34dc7c2f | 3898 | } |
d2734cce SD |
3899 | ASSERT0(range_tree_space(msp->ms_freeing)); |
3900 | ASSERT0(range_tree_space(msp->ms_checkpointing)); | |
34dc7c2f | 3901 | |
d2734cce | 3902 | defer_tree = &msp->ms_defer[txg % TXG_DEFER_SIZE]; |
93cf2076 | 3903 | |
1c27024e | 3904 | uint64_t free_space = metaslab_class_get_space(spa_normal_class(spa)) - |
4e21fd06 | 3905 | metaslab_class_get_alloc(spa_normal_class(spa)); |
a1d477c2 | 3906 | if (free_space <= spa_get_slop_space(spa) || vd->vdev_removing) { |
4e21fd06 DB |
3907 | defer_allowed = B_FALSE; |
3908 | } | |
3909 | ||
3910 | defer_delta = 0; | |
425d3237 SD |
3911 | alloc_delta = msp->ms_allocated_this_txg - |
3912 | range_tree_space(msp->ms_freed); | |
93e28d66 | 3913 | |
4e21fd06 | 3914 | if (defer_allowed) { |
d2734cce | 3915 | defer_delta = range_tree_space(msp->ms_freed) - |
4e21fd06 DB |
3916 | range_tree_space(*defer_tree); |
3917 | } else { | |
3918 | defer_delta -= range_tree_space(*defer_tree); | |
3919 | } | |
cc99f275 DB |
3920 | metaslab_space_update(vd, mg->mg_class, alloc_delta + defer_delta, |
3921 | defer_delta, 0); | |
34dc7c2f | 3922 | |
93e28d66 SD |
3923 | if (spa_syncing_log_sm(spa) == NULL) { |
3924 | /* | |
3925 | * If there's a metaslab_load() in progress and we don't have | |
3926 | * a log space map, it means that we probably wrote to the | |
3927 | * metaslab's space map. If this is the case, we need to | |
3928 | * make sure that we wait for the load to complete so that we | |
3929 | * have a consistent view at the in-core side of the metaslab. | |
3930 | */ | |
3931 | metaslab_load_wait(msp); | |
3932 | } else { | |
3933 | ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)); | |
3934 | } | |
c2e42f9d | 3935 | |
1b939560 BB |
3936 | /* |
3937 | * When auto-trimming is enabled, free ranges which are added to | |
3938 | * ms_allocatable are also be added to ms_trim. The ms_trim tree is | |
3939 | * periodically consumed by the vdev_autotrim_thread() which issues | |
3940 | * trims for all ranges and then vacates the tree. The ms_trim tree | |
3941 | * can be discarded at any time with the sole consequence of recent | |
3942 | * frees not being trimmed. | |
3943 | */ | |
3944 | if (spa_get_autotrim(spa) == SPA_AUTOTRIM_ON) { | |
3945 | range_tree_walk(*defer_tree, range_tree_add, msp->ms_trim); | |
3946 | if (!defer_allowed) { | |
3947 | range_tree_walk(msp->ms_freed, range_tree_add, | |
3948 | msp->ms_trim); | |
3949 | } | |
3950 | } else { | |
3951 | range_tree_vacate(msp->ms_trim, NULL, NULL); | |
3952 | } | |
3953 | ||
c2e42f9d | 3954 | /* |
93cf2076 | 3955 | * Move the frees from the defer_tree back to the free |
d2734cce SD |
3956 | * range tree (if it's loaded). Swap the freed_tree and |
3957 | * the defer_tree -- this is safe to do because we've | |
3958 | * just emptied out the defer_tree. | |
c2e42f9d | 3959 | */ |
93cf2076 | 3960 | range_tree_vacate(*defer_tree, |
d2734cce | 3961 | msp->ms_loaded ? range_tree_add : NULL, msp->ms_allocatable); |
4e21fd06 | 3962 | if (defer_allowed) { |
d2734cce | 3963 | range_tree_swap(&msp->ms_freed, defer_tree); |
4e21fd06 | 3964 | } else { |
d2734cce SD |
3965 | range_tree_vacate(msp->ms_freed, |
3966 | msp->ms_loaded ? range_tree_add : NULL, | |
3967 | msp->ms_allocatable); | |
4e21fd06 | 3968 | } |
425d3237 SD |
3969 | |
3970 | msp->ms_synced_length = space_map_length(msp->ms_sm); | |
34dc7c2f | 3971 | |
428870ff BB |
3972 | msp->ms_deferspace += defer_delta; |
3973 | ASSERT3S(msp->ms_deferspace, >=, 0); | |
93cf2076 | 3974 | ASSERT3S(msp->ms_deferspace, <=, msp->ms_size); |
428870ff BB |
3975 | if (msp->ms_deferspace != 0) { |
3976 | /* | |
3977 | * Keep syncing this metaslab until all deferred frees | |
3978 | * are back in circulation. | |
3979 | */ | |
3980 | vdev_dirty(vd, VDD_METASLAB, msp, txg + 1); | |
3981 | } | |
928e8ad4 | 3982 | metaslab_aux_histograms_update_done(msp, defer_allowed); |
428870ff | 3983 | |
492f64e9 PD |
3984 | if (msp->ms_new) { |
3985 | msp->ms_new = B_FALSE; | |
3986 | mutex_enter(&mg->mg_lock); | |
3987 | mg->mg_ms_ready++; | |
3988 | mutex_exit(&mg->mg_lock); | |
3989 | } | |
928e8ad4 | 3990 | |
4e21fd06 | 3991 | /* |
928e8ad4 SD |
3992 | * Re-sort metaslab within its group now that we've adjusted |
3993 | * its allocatable space. | |
4e21fd06 | 3994 | */ |
928e8ad4 | 3995 | metaslab_recalculate_weight_and_sort(msp); |
4e21fd06 | 3996 | |
d2734cce SD |
3997 | ASSERT0(range_tree_space(msp->ms_allocating[txg & TXG_MASK])); |
3998 | ASSERT0(range_tree_space(msp->ms_freeing)); | |
3999 | ASSERT0(range_tree_space(msp->ms_freed)); | |
4000 | ASSERT0(range_tree_space(msp->ms_checkpointing)); | |
f09fda50 | 4001 | msp->ms_allocating_total -= msp->ms_allocated_this_txg; |
425d3237 | 4002 | msp->ms_allocated_this_txg = 0; |
34dc7c2f BB |
4003 | mutex_exit(&msp->ms_lock); |
4004 | } | |
4005 | ||
428870ff BB |
4006 | void |
4007 | metaslab_sync_reassess(metaslab_group_t *mg) | |
4008 | { | |
a1d477c2 MA |
4009 | spa_t *spa = mg->mg_class->mc_spa; |
4010 | ||
4011 | spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER); | |
1be627f5 | 4012 | metaslab_group_alloc_update(mg); |
f3a7f661 | 4013 | mg->mg_fragmentation = metaslab_group_fragmentation(mg); |
6d974228 | 4014 | |
428870ff | 4015 | /* |
a1d477c2 MA |
4016 | * Preload the next potential metaslabs but only on active |
4017 | * metaslab groups. We can get into a state where the metaslab | |
4018 | * is no longer active since we dirty metaslabs as we remove a | |
4019 | * a device, thus potentially making the metaslab group eligible | |
4020 | * for preloading. | |
428870ff | 4021 | */ |
a1d477c2 MA |
4022 | if (mg->mg_activation_count > 0) { |
4023 | metaslab_group_preload(mg); | |
4024 | } | |
4025 | spa_config_exit(spa, SCL_ALLOC, FTAG); | |
428870ff BB |
4026 | } |
4027 | ||
cc99f275 DB |
4028 | /* |
4029 | * When writing a ditto block (i.e. more than one DVA for a given BP) on | |
4030 | * the same vdev as an existing DVA of this BP, then try to allocate it | |
4031 | * on a different metaslab than existing DVAs (i.e. a unique metaslab). | |
4032 | */ | |
4033 | static boolean_t | |
4034 | metaslab_is_unique(metaslab_t *msp, dva_t *dva) | |
34dc7c2f | 4035 | { |
cc99f275 DB |
4036 | uint64_t dva_ms_id; |
4037 | ||
4038 | if (DVA_GET_ASIZE(dva) == 0) | |
4039 | return (B_TRUE); | |
34dc7c2f BB |
4040 | |
4041 | if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva)) | |
cc99f275 | 4042 | return (B_TRUE); |
34dc7c2f | 4043 | |
cc99f275 DB |
4044 | dva_ms_id = DVA_GET_OFFSET(dva) >> msp->ms_group->mg_vd->vdev_ms_shift; |
4045 | ||
4046 | return (msp->ms_id != dva_ms_id); | |
34dc7c2f BB |
4047 | } |
4048 | ||
4e21fd06 DB |
4049 | /* |
4050 | * ========================================================================== | |
4051 | * Metaslab allocation tracing facility | |
4052 | * ========================================================================== | |
4053 | */ | |
4054 | #ifdef _METASLAB_TRACING | |
4055 | kstat_t *metaslab_trace_ksp; | |
4056 | kstat_named_t metaslab_trace_over_limit; | |
4057 | ||
4058 | void | |
4059 | metaslab_alloc_trace_init(void) | |
4060 | { | |
4061 | ASSERT(metaslab_alloc_trace_cache == NULL); | |
4062 | metaslab_alloc_trace_cache = kmem_cache_create( | |
4063 | "metaslab_alloc_trace_cache", sizeof (metaslab_alloc_trace_t), | |
4064 | 0, NULL, NULL, NULL, NULL, NULL, 0); | |
4065 | metaslab_trace_ksp = kstat_create("zfs", 0, "metaslab_trace_stats", | |
4066 | "misc", KSTAT_TYPE_NAMED, 1, KSTAT_FLAG_VIRTUAL); | |
4067 | if (metaslab_trace_ksp != NULL) { | |
4068 | metaslab_trace_ksp->ks_data = &metaslab_trace_over_limit; | |
4069 | kstat_named_init(&metaslab_trace_over_limit, | |
4070 | "metaslab_trace_over_limit", KSTAT_DATA_UINT64); | |
4071 | kstat_install(metaslab_trace_ksp); | |
4072 | } | |
4073 | } | |
4074 | ||
4075 | void | |
4076 | metaslab_alloc_trace_fini(void) | |
4077 | { | |
4078 | if (metaslab_trace_ksp != NULL) { | |
4079 | kstat_delete(metaslab_trace_ksp); | |
4080 | metaslab_trace_ksp = NULL; | |
4081 | } | |
4082 | kmem_cache_destroy(metaslab_alloc_trace_cache); | |
4083 | metaslab_alloc_trace_cache = NULL; | |
4084 | } | |
4085 | ||
4086 | /* | |
4087 | * Add an allocation trace element to the allocation tracing list. | |
4088 | */ | |
4089 | static void | |
4090 | metaslab_trace_add(zio_alloc_list_t *zal, metaslab_group_t *mg, | |
492f64e9 PD |
4091 | metaslab_t *msp, uint64_t psize, uint32_t dva_id, uint64_t offset, |
4092 | int allocator) | |
4e21fd06 DB |
4093 | { |
4094 | metaslab_alloc_trace_t *mat; | |
4095 | ||
4096 | if (!metaslab_trace_enabled) | |
4097 | return; | |
4098 | ||
4099 | /* | |
4100 | * When the tracing list reaches its maximum we remove | |
4101 | * the second element in the list before adding a new one. | |
4102 | * By removing the second element we preserve the original | |
4103 | * entry as a clue to what allocations steps have already been | |
4104 | * performed. | |
4105 | */ | |
4106 | if (zal->zal_size == metaslab_trace_max_entries) { | |
4107 | metaslab_alloc_trace_t *mat_next; | |
4108 | #ifdef DEBUG | |
4109 | panic("too many entries in allocation list"); | |
4110 | #endif | |
4111 | atomic_inc_64(&metaslab_trace_over_limit.value.ui64); | |
4112 | zal->zal_size--; | |
4113 | mat_next = list_next(&zal->zal_list, list_head(&zal->zal_list)); | |
4114 | list_remove(&zal->zal_list, mat_next); | |
4115 | kmem_cache_free(metaslab_alloc_trace_cache, mat_next); | |
4116 | } | |
4117 | ||
4118 | mat = kmem_cache_alloc(metaslab_alloc_trace_cache, KM_SLEEP); | |
4119 | list_link_init(&mat->mat_list_node); | |
4120 | mat->mat_mg = mg; | |
4121 | mat->mat_msp = msp; | |
4122 | mat->mat_size = psize; | |
4123 | mat->mat_dva_id = dva_id; | |
4124 | mat->mat_offset = offset; | |
4125 | mat->mat_weight = 0; | |
492f64e9 | 4126 | mat->mat_allocator = allocator; |
4e21fd06 DB |
4127 | |
4128 | if (msp != NULL) | |
4129 | mat->mat_weight = msp->ms_weight; | |
4130 | ||
4131 | /* | |
4132 | * The list is part of the zio so locking is not required. Only | |
4133 | * a single thread will perform allocations for a given zio. | |
4134 | */ | |
4135 | list_insert_tail(&zal->zal_list, mat); | |
4136 | zal->zal_size++; | |
4137 | ||
4138 | ASSERT3U(zal->zal_size, <=, metaslab_trace_max_entries); | |
4139 | } | |
4140 | ||
4141 | void | |
4142 | metaslab_trace_init(zio_alloc_list_t *zal) | |
4143 | { | |
4144 | list_create(&zal->zal_list, sizeof (metaslab_alloc_trace_t), | |
4145 | offsetof(metaslab_alloc_trace_t, mat_list_node)); | |
4146 | zal->zal_size = 0; | |
4147 | } | |
4148 | ||
4149 | void | |
4150 | metaslab_trace_fini(zio_alloc_list_t *zal) | |
4151 | { | |
4152 | metaslab_alloc_trace_t *mat; | |
4153 | ||
4154 | while ((mat = list_remove_head(&zal->zal_list)) != NULL) | |
4155 | kmem_cache_free(metaslab_alloc_trace_cache, mat); | |
4156 | list_destroy(&zal->zal_list); | |
4157 | zal->zal_size = 0; | |
4158 | } | |
4159 | #else | |
4160 | ||
492f64e9 | 4161 | #define metaslab_trace_add(zal, mg, msp, psize, id, off, alloc) |
4e21fd06 DB |
4162 | |
4163 | void | |
4164 | metaslab_alloc_trace_init(void) | |
4165 | { | |
4166 | } | |
4167 | ||
4168 | void | |
4169 | metaslab_alloc_trace_fini(void) | |
4170 | { | |
4171 | } | |
4172 | ||
4173 | void | |
4174 | metaslab_trace_init(zio_alloc_list_t *zal) | |
4175 | { | |
4176 | } | |
4177 | ||
4178 | void | |
4179 | metaslab_trace_fini(zio_alloc_list_t *zal) | |
4180 | { | |
4181 | } | |
4182 | ||
4183 | #endif /* _METASLAB_TRACING */ | |
4184 | ||
3dfb57a3 DB |
4185 | /* |
4186 | * ========================================================================== | |
4187 | * Metaslab block operations | |
4188 | * ========================================================================== | |
4189 | */ | |
4190 | ||
4191 | static void | |
492f64e9 PD |
4192 | metaslab_group_alloc_increment(spa_t *spa, uint64_t vdev, void *tag, int flags, |
4193 | int allocator) | |
3dfb57a3 | 4194 | { |
3dfb57a3 | 4195 | if (!(flags & METASLAB_ASYNC_ALLOC) || |
492f64e9 | 4196 | (flags & METASLAB_DONT_THROTTLE)) |
3dfb57a3 DB |
4197 | return; |
4198 | ||
1c27024e | 4199 | metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg; |
3dfb57a3 DB |
4200 | if (!mg->mg_class->mc_alloc_throttle_enabled) |
4201 | return; | |
4202 | ||
c13060e4 | 4203 | (void) zfs_refcount_add(&mg->mg_alloc_queue_depth[allocator], tag); |
492f64e9 PD |
4204 | } |
4205 | ||
4206 | static void | |
4207 | metaslab_group_increment_qdepth(metaslab_group_t *mg, int allocator) | |
4208 | { | |
4209 | uint64_t max = mg->mg_max_alloc_queue_depth; | |
4210 | uint64_t cur = mg->mg_cur_max_alloc_queue_depth[allocator]; | |
4211 | while (cur < max) { | |
4212 | if (atomic_cas_64(&mg->mg_cur_max_alloc_queue_depth[allocator], | |
4213 | cur, cur + 1) == cur) { | |
4214 | atomic_inc_64( | |
4215 | &mg->mg_class->mc_alloc_max_slots[allocator]); | |
4216 | return; | |
4217 | } | |
4218 | cur = mg->mg_cur_max_alloc_queue_depth[allocator]; | |
4219 | } | |
3dfb57a3 DB |
4220 | } |
4221 | ||
4222 | void | |
492f64e9 PD |
4223 | metaslab_group_alloc_decrement(spa_t *spa, uint64_t vdev, void *tag, int flags, |
4224 | int allocator, boolean_t io_complete) | |
3dfb57a3 | 4225 | { |
3dfb57a3 | 4226 | if (!(flags & METASLAB_ASYNC_ALLOC) || |
492f64e9 | 4227 | (flags & METASLAB_DONT_THROTTLE)) |
3dfb57a3 DB |
4228 | return; |
4229 | ||
1c27024e | 4230 | metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg; |
3dfb57a3 DB |
4231 | if (!mg->mg_class->mc_alloc_throttle_enabled) |
4232 | return; | |
4233 | ||
424fd7c3 | 4234 | (void) zfs_refcount_remove(&mg->mg_alloc_queue_depth[allocator], tag); |
492f64e9 PD |
4235 | if (io_complete) |
4236 | metaslab_group_increment_qdepth(mg, allocator); | |
3dfb57a3 DB |
4237 | } |
4238 | ||
4239 | void | |
492f64e9 PD |
4240 | metaslab_group_alloc_verify(spa_t *spa, const blkptr_t *bp, void *tag, |
4241 | int allocator) | |
3dfb57a3 DB |
4242 | { |
4243 | #ifdef ZFS_DEBUG | |
4244 | const dva_t *dva = bp->blk_dva; | |
4245 | int ndvas = BP_GET_NDVAS(bp); | |
3dfb57a3 | 4246 | |
1c27024e | 4247 | for (int d = 0; d < ndvas; d++) { |
3dfb57a3 DB |
4248 | uint64_t vdev = DVA_GET_VDEV(&dva[d]); |
4249 | metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg; | |
424fd7c3 TS |
4250 | VERIFY(zfs_refcount_not_held( |
4251 | &mg->mg_alloc_queue_depth[allocator], tag)); | |
3dfb57a3 DB |
4252 | } |
4253 | #endif | |
4254 | } | |
4255 | ||
34dc7c2f | 4256 | static uint64_t |
4e21fd06 DB |
4257 | metaslab_block_alloc(metaslab_t *msp, uint64_t size, uint64_t txg) |
4258 | { | |
4259 | uint64_t start; | |
d2734cce | 4260 | range_tree_t *rt = msp->ms_allocatable; |
4e21fd06 DB |
4261 | metaslab_class_t *mc = msp->ms_group->mg_class; |
4262 | ||
93e28d66 | 4263 | ASSERT(MUTEX_HELD(&msp->ms_lock)); |
4e21fd06 | 4264 | VERIFY(!msp->ms_condensing); |
1b939560 | 4265 | VERIFY0(msp->ms_disabled); |
4e21fd06 DB |
4266 | |
4267 | start = mc->mc_ops->msop_alloc(msp, size); | |
4268 | if (start != -1ULL) { | |
4269 | metaslab_group_t *mg = msp->ms_group; | |
4270 | vdev_t *vd = mg->mg_vd; | |
4271 | ||
4272 | VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift)); | |
4273 | VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift)); | |
4274 | VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size); | |
4275 | range_tree_remove(rt, start, size); | |
1b939560 | 4276 | range_tree_clear(msp->ms_trim, start, size); |
4e21fd06 | 4277 | |
d2734cce | 4278 | if (range_tree_is_empty(msp->ms_allocating[txg & TXG_MASK])) |
4e21fd06 DB |
4279 | vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg); |
4280 | ||
d2734cce | 4281 | range_tree_add(msp->ms_allocating[txg & TXG_MASK], start, size); |
f09fda50 | 4282 | msp->ms_allocating_total += size; |
4e21fd06 DB |
4283 | |
4284 | /* Track the last successful allocation */ | |
4285 | msp->ms_alloc_txg = txg; | |
4286 | metaslab_verify_space(msp, txg); | |
4287 | } | |
4288 | ||
4289 | /* | |
4290 | * Now that we've attempted the allocation we need to update the | |
4291 | * metaslab's maximum block size since it may have changed. | |
4292 | */ | |
c81f1790 | 4293 | msp->ms_max_size = metaslab_largest_allocatable(msp); |
4e21fd06 DB |
4294 | return (start); |
4295 | } | |
4296 | ||
492f64e9 PD |
4297 | /* |
4298 | * Find the metaslab with the highest weight that is less than what we've | |
4299 | * already tried. In the common case, this means that we will examine each | |
4300 | * metaslab at most once. Note that concurrent callers could reorder metaslabs | |
4301 | * by activation/passivation once we have dropped the mg_lock. If a metaslab is | |
4302 | * activated by another thread, and we fail to allocate from the metaslab we | |
4303 | * have selected, we may not try the newly-activated metaslab, and instead | |
4304 | * activate another metaslab. This is not optimal, but generally does not cause | |
4305 | * any problems (a possible exception being if every metaslab is completely full | |
e1cfd73f | 4306 | * except for the newly-activated metaslab which we fail to examine). |
492f64e9 PD |
4307 | */ |
4308 | static metaslab_t * | |
4309 | find_valid_metaslab(metaslab_group_t *mg, uint64_t activation_weight, | |
cc99f275 | 4310 | dva_t *dva, int d, boolean_t want_unique, uint64_t asize, int allocator, |
c81f1790 PD |
4311 | boolean_t try_hard, zio_alloc_list_t *zal, metaslab_t *search, |
4312 | boolean_t *was_active) | |
492f64e9 PD |
4313 | { |
4314 | avl_index_t idx; | |
4315 | avl_tree_t *t = &mg->mg_metaslab_tree; | |
4316 | metaslab_t *msp = avl_find(t, search, &idx); | |
4317 | if (msp == NULL) | |
4318 | msp = avl_nearest(t, idx, AVL_AFTER); | |
4319 | ||
4320 | for (; msp != NULL; msp = AVL_NEXT(t, msp)) { | |
4321 | int i; | |
c81f1790 | 4322 | if (!metaslab_should_allocate(msp, asize, try_hard)) { |
492f64e9 PD |
4323 | metaslab_trace_add(zal, mg, msp, asize, d, |
4324 | TRACE_TOO_SMALL, allocator); | |
4325 | continue; | |
4326 | } | |
4327 | ||
4328 | /* | |
1b939560 BB |
4329 | * If the selected metaslab is condensing or disabled, |
4330 | * skip it. | |
492f64e9 | 4331 | */ |
1b939560 | 4332 | if (msp->ms_condensing || msp->ms_disabled > 0) |
492f64e9 PD |
4333 | continue; |
4334 | ||
4335 | *was_active = msp->ms_allocator != -1; | |
4336 | /* | |
4337 | * If we're activating as primary, this is our first allocation | |
4338 | * from this disk, so we don't need to check how close we are. | |
4339 | * If the metaslab under consideration was already active, | |
4340 | * we're getting desperate enough to steal another allocator's | |
4341 | * metaslab, so we still don't care about distances. | |
4342 | */ | |
4343 | if (activation_weight == METASLAB_WEIGHT_PRIMARY || *was_active) | |
4344 | break; | |
4345 | ||
492f64e9 | 4346 | for (i = 0; i < d; i++) { |
cc99f275 DB |
4347 | if (want_unique && |
4348 | !metaslab_is_unique(msp, &dva[i])) | |
4349 | break; /* try another metaslab */ | |
492f64e9 PD |
4350 | } |
4351 | if (i == d) | |
4352 | break; | |
4353 | } | |
4354 | ||
4355 | if (msp != NULL) { | |
4356 | search->ms_weight = msp->ms_weight; | |
4357 | search->ms_start = msp->ms_start + 1; | |
4358 | search->ms_allocator = msp->ms_allocator; | |
4359 | search->ms_primary = msp->ms_primary; | |
4360 | } | |
4361 | return (msp); | |
4362 | } | |
4363 | ||
679b0f2a PD |
4364 | void |
4365 | metaslab_active_mask_verify(metaslab_t *msp) | |
4366 | { | |
4367 | ASSERT(MUTEX_HELD(&msp->ms_lock)); | |
4368 | ||
4369 | if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0) | |
4370 | return; | |
4371 | ||
4372 | if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) | |
4373 | return; | |
4374 | ||
4375 | if (msp->ms_weight & METASLAB_WEIGHT_PRIMARY) { | |
4376 | VERIFY0(msp->ms_weight & METASLAB_WEIGHT_SECONDARY); | |
4377 | VERIFY0(msp->ms_weight & METASLAB_WEIGHT_CLAIM); | |
4378 | VERIFY3S(msp->ms_allocator, !=, -1); | |
4379 | VERIFY(msp->ms_primary); | |
4380 | return; | |
4381 | } | |
4382 | ||
4383 | if (msp->ms_weight & METASLAB_WEIGHT_SECONDARY) { | |
4384 | VERIFY0(msp->ms_weight & METASLAB_WEIGHT_PRIMARY); | |
4385 | VERIFY0(msp->ms_weight & METASLAB_WEIGHT_CLAIM); | |
4386 | VERIFY3S(msp->ms_allocator, !=, -1); | |
4387 | VERIFY(!msp->ms_primary); | |
4388 | return; | |
4389 | } | |
4390 | ||
4391 | if (msp->ms_weight & METASLAB_WEIGHT_CLAIM) { | |
4392 | VERIFY0(msp->ms_weight & METASLAB_WEIGHT_PRIMARY); | |
4393 | VERIFY0(msp->ms_weight & METASLAB_WEIGHT_SECONDARY); | |
4394 | VERIFY3S(msp->ms_allocator, ==, -1); | |
4395 | return; | |
4396 | } | |
4397 | } | |
4398 | ||
492f64e9 | 4399 | /* ARGSUSED */ |
4e21fd06 DB |
4400 | static uint64_t |
4401 | metaslab_group_alloc_normal(metaslab_group_t *mg, zio_alloc_list_t *zal, | |
c81f1790 PD |
4402 | uint64_t asize, uint64_t txg, boolean_t want_unique, dva_t *dva, int d, |
4403 | int allocator, boolean_t try_hard) | |
34dc7c2f BB |
4404 | { |
4405 | metaslab_t *msp = NULL; | |
4406 | uint64_t offset = -1ULL; | |
34dc7c2f | 4407 | |
679b0f2a | 4408 | uint64_t activation_weight = METASLAB_WEIGHT_PRIMARY; |
492f64e9 PD |
4409 | for (int i = 0; i < d; i++) { |
4410 | if (activation_weight == METASLAB_WEIGHT_PRIMARY && | |
4411 | DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) { | |
34dc7c2f | 4412 | activation_weight = METASLAB_WEIGHT_SECONDARY; |
492f64e9 PD |
4413 | } else if (activation_weight == METASLAB_WEIGHT_SECONDARY && |
4414 | DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) { | |
e38afd34 | 4415 | activation_weight = METASLAB_WEIGHT_CLAIM; |
9babb374 BB |
4416 | break; |
4417 | } | |
4418 | } | |
34dc7c2f | 4419 | |
492f64e9 PD |
4420 | /* |
4421 | * If we don't have enough metaslabs active to fill the entire array, we | |
4422 | * just use the 0th slot. | |
4423 | */ | |
e38afd34 | 4424 | if (mg->mg_ms_ready < mg->mg_allocators * 3) |
492f64e9 | 4425 | allocator = 0; |
492f64e9 PD |
4426 | |
4427 | ASSERT3U(mg->mg_vd->vdev_ms_count, >=, 2); | |
4428 | ||
1c27024e | 4429 | metaslab_t *search = kmem_alloc(sizeof (*search), KM_SLEEP); |
4e21fd06 DB |
4430 | search->ms_weight = UINT64_MAX; |
4431 | search->ms_start = 0; | |
492f64e9 PD |
4432 | /* |
4433 | * At the end of the metaslab tree are the already-active metaslabs, | |
4434 | * first the primaries, then the secondaries. When we resume searching | |
4435 | * through the tree, we need to consider ms_allocator and ms_primary so | |
4436 | * we start in the location right after where we left off, and don't | |
4437 | * accidentally loop forever considering the same metaslabs. | |
4438 | */ | |
4439 | search->ms_allocator = -1; | |
4440 | search->ms_primary = B_TRUE; | |
34dc7c2f | 4441 | for (;;) { |
492f64e9 | 4442 | boolean_t was_active = B_FALSE; |
9babb374 | 4443 | |
34dc7c2f | 4444 | mutex_enter(&mg->mg_lock); |
4e21fd06 | 4445 | |
492f64e9 PD |
4446 | if (activation_weight == METASLAB_WEIGHT_PRIMARY && |
4447 | mg->mg_primaries[allocator] != NULL) { | |
4448 | msp = mg->mg_primaries[allocator]; | |
679b0f2a PD |
4449 | |
4450 | /* | |
4451 | * Even though we don't hold the ms_lock for the | |
4452 | * primary metaslab, those fields should not | |
e1cfd73f | 4453 | * change while we hold the mg_lock. Thus it is |
679b0f2a PD |
4454 | * safe to make assertions on them. |
4455 | */ | |
4456 | ASSERT(msp->ms_primary); | |
4457 | ASSERT3S(msp->ms_allocator, ==, allocator); | |
4458 | ASSERT(msp->ms_loaded); | |
4459 | ||
492f64e9 | 4460 | was_active = B_TRUE; |
f09fda50 | 4461 | ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK); |
492f64e9 | 4462 | } else if (activation_weight == METASLAB_WEIGHT_SECONDARY && |
e38afd34 | 4463 | mg->mg_secondaries[allocator] != NULL) { |
492f64e9 | 4464 | msp = mg->mg_secondaries[allocator]; |
679b0f2a PD |
4465 | |
4466 | /* | |
4467 | * See comment above about the similar assertions | |
4468 | * for the primary metaslab. | |
4469 | */ | |
4470 | ASSERT(!msp->ms_primary); | |
4471 | ASSERT3S(msp->ms_allocator, ==, allocator); | |
4472 | ASSERT(msp->ms_loaded); | |
4473 | ||
492f64e9 | 4474 | was_active = B_TRUE; |
f09fda50 | 4475 | ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK); |
492f64e9 PD |
4476 | } else { |
4477 | msp = find_valid_metaslab(mg, activation_weight, dva, d, | |
c81f1790 PD |
4478 | want_unique, asize, allocator, try_hard, zal, |
4479 | search, &was_active); | |
34dc7c2f | 4480 | } |
492f64e9 | 4481 | |
34dc7c2f | 4482 | mutex_exit(&mg->mg_lock); |
4e21fd06 DB |
4483 | if (msp == NULL) { |
4484 | kmem_free(search, sizeof (*search)); | |
34dc7c2f | 4485 | return (-1ULL); |
4e21fd06 | 4486 | } |
ac72fac3 | 4487 | mutex_enter(&msp->ms_lock); |
679b0f2a PD |
4488 | |
4489 | metaslab_active_mask_verify(msp); | |
4490 | ||
4491 | /* | |
4492 | * This code is disabled out because of issues with | |
4493 | * tracepoints in non-gpl kernel modules. | |
4494 | */ | |
4495 | #if 0 | |
4496 | DTRACE_PROBE3(ms__activation__attempt, | |
4497 | metaslab_t *, msp, uint64_t, activation_weight, | |
4498 | boolean_t, was_active); | |
4499 | #endif | |
4500 | ||
34dc7c2f BB |
4501 | /* |
4502 | * Ensure that the metaslab we have selected is still | |
4503 | * capable of handling our request. It's possible that | |
4504 | * another thread may have changed the weight while we | |
4e21fd06 | 4505 | * were blocked on the metaslab lock. We check the |
f09fda50 | 4506 | * active status first to see if we need to set_selected_txg |
4e21fd06 | 4507 | * a new metaslab. |
34dc7c2f | 4508 | */ |
4e21fd06 | 4509 | if (was_active && !(msp->ms_weight & METASLAB_ACTIVE_MASK)) { |
679b0f2a | 4510 | ASSERT3S(msp->ms_allocator, ==, -1); |
34dc7c2f BB |
4511 | mutex_exit(&msp->ms_lock); |
4512 | continue; | |
4513 | } | |
4514 | ||
492f64e9 | 4515 | /* |
679b0f2a PD |
4516 | * If the metaslab was activated for another allocator |
4517 | * while we were waiting in the ms_lock above, or it's | |
4518 | * a primary and we're seeking a secondary (or vice versa), | |
4519 | * we go back and select a new metaslab. | |
492f64e9 PD |
4520 | */ |
4521 | if (!was_active && (msp->ms_weight & METASLAB_ACTIVE_MASK) && | |
4522 | (msp->ms_allocator != -1) && | |
4523 | (msp->ms_allocator != allocator || ((activation_weight == | |
4524 | METASLAB_WEIGHT_PRIMARY) != msp->ms_primary))) { | |
679b0f2a PD |
4525 | ASSERT(msp->ms_loaded); |
4526 | ASSERT((msp->ms_weight & METASLAB_WEIGHT_CLAIM) || | |
4527 | msp->ms_allocator != -1); | |
492f64e9 PD |
4528 | mutex_exit(&msp->ms_lock); |
4529 | continue; | |
4530 | } | |
4531 | ||
679b0f2a PD |
4532 | /* |
4533 | * This metaslab was used for claiming regions allocated | |
4534 | * by the ZIL during pool import. Once these regions are | |
4535 | * claimed we don't need to keep the CLAIM bit set | |
4536 | * anymore. Passivate this metaslab to zero its activation | |
4537 | * mask. | |
4538 | */ | |
e38afd34 | 4539 | if (msp->ms_weight & METASLAB_WEIGHT_CLAIM && |
4540 | activation_weight != METASLAB_WEIGHT_CLAIM) { | |
679b0f2a PD |
4541 | ASSERT(msp->ms_loaded); |
4542 | ASSERT3S(msp->ms_allocator, ==, -1); | |
492f64e9 PD |
4543 | metaslab_passivate(msp, msp->ms_weight & |
4544 | ~METASLAB_WEIGHT_CLAIM); | |
34dc7c2f BB |
4545 | mutex_exit(&msp->ms_lock); |
4546 | continue; | |
4547 | } | |
4548 | ||
f09fda50 | 4549 | metaslab_set_selected_txg(msp, txg); |
679b0f2a PD |
4550 | |
4551 | int activation_error = | |
4552 | metaslab_activate(msp, allocator, activation_weight); | |
4553 | metaslab_active_mask_verify(msp); | |
4554 | ||
4555 | /* | |
4556 | * If the metaslab was activated by another thread for | |
4557 | * another allocator or activation_weight (EBUSY), or it | |
4558 | * failed because another metaslab was assigned as primary | |
4559 | * for this allocator (EEXIST) we continue using this | |
4560 | * metaslab for our allocation, rather than going on to a | |
4561 | * worse metaslab (we waited for that metaslab to be loaded | |
4562 | * after all). | |
4563 | * | |
fe0ea848 PD |
4564 | * If the activation failed due to an I/O error or ENOSPC we |
4565 | * skip to the next metaslab. | |
679b0f2a PD |
4566 | */ |
4567 | boolean_t activated; | |
4568 | if (activation_error == 0) { | |
4569 | activated = B_TRUE; | |
4570 | } else if (activation_error == EBUSY || | |
4571 | activation_error == EEXIST) { | |
4572 | activated = B_FALSE; | |
4573 | } else { | |
34dc7c2f BB |
4574 | mutex_exit(&msp->ms_lock); |
4575 | continue; | |
4576 | } | |
679b0f2a | 4577 | ASSERT(msp->ms_loaded); |
4e21fd06 DB |
4578 | |
4579 | /* | |
4580 | * Now that we have the lock, recheck to see if we should | |
4581 | * continue to use this metaslab for this allocation. The | |
679b0f2a PD |
4582 | * the metaslab is now loaded so metaslab_should_allocate() |
4583 | * can accurately determine if the allocation attempt should | |
4e21fd06 DB |
4584 | * proceed. |
4585 | */ | |
c81f1790 | 4586 | if (!metaslab_should_allocate(msp, asize, try_hard)) { |
4e21fd06 DB |
4587 | /* Passivate this metaslab and select a new one. */ |
4588 | metaslab_trace_add(zal, mg, msp, asize, d, | |
492f64e9 | 4589 | TRACE_TOO_SMALL, allocator); |
4e21fd06 DB |
4590 | goto next; |
4591 | } | |
4592 | ||
7a614407 | 4593 | /* |
679b0f2a PD |
4594 | * If this metaslab is currently condensing then pick again |
4595 | * as we can't manipulate this metaslab until it's committed | |
619f0976 GW |
4596 | * to disk. If this metaslab is being initialized, we shouldn't |
4597 | * allocate from it since the allocated region might be | |
4598 | * overwritten after allocation. | |
7a614407 | 4599 | */ |
93cf2076 | 4600 | if (msp->ms_condensing) { |
4e21fd06 | 4601 | metaslab_trace_add(zal, mg, msp, asize, d, |
492f64e9 | 4602 | TRACE_CONDENSING, allocator); |
679b0f2a PD |
4603 | if (activated) { |
4604 | metaslab_passivate(msp, msp->ms_weight & | |
4605 | ~METASLAB_ACTIVE_MASK); | |
4606 | } | |
7a614407 GW |
4607 | mutex_exit(&msp->ms_lock); |
4608 | continue; | |
1b939560 | 4609 | } else if (msp->ms_disabled > 0) { |
619f0976 | 4610 | metaslab_trace_add(zal, mg, msp, asize, d, |
1b939560 | 4611 | TRACE_DISABLED, allocator); |
679b0f2a PD |
4612 | if (activated) { |
4613 | metaslab_passivate(msp, msp->ms_weight & | |
4614 | ~METASLAB_ACTIVE_MASK); | |
4615 | } | |
619f0976 GW |
4616 | mutex_exit(&msp->ms_lock); |
4617 | continue; | |
7a614407 GW |
4618 | } |
4619 | ||
4e21fd06 | 4620 | offset = metaslab_block_alloc(msp, asize, txg); |
492f64e9 | 4621 | metaslab_trace_add(zal, mg, msp, asize, d, offset, allocator); |
4e21fd06 DB |
4622 | |
4623 | if (offset != -1ULL) { | |
4624 | /* Proactively passivate the metaslab, if needed */ | |
679b0f2a PD |
4625 | if (activated) |
4626 | metaslab_segment_may_passivate(msp); | |
34dc7c2f | 4627 | break; |
4e21fd06 DB |
4628 | } |
4629 | next: | |
4630 | ASSERT(msp->ms_loaded); | |
4631 | ||
679b0f2a PD |
4632 | /* |
4633 | * This code is disabled out because of issues with | |
4634 | * tracepoints in non-gpl kernel modules. | |
4635 | */ | |
4636 | #if 0 | |
4637 | DTRACE_PROBE2(ms__alloc__failure, metaslab_t *, msp, | |
4638 | uint64_t, asize); | |
4639 | #endif | |
4640 | ||
4e21fd06 DB |
4641 | /* |
4642 | * We were unable to allocate from this metaslab so determine | |
4643 | * a new weight for this metaslab. Now that we have loaded | |
4644 | * the metaslab we can provide a better hint to the metaslab | |
4645 | * selector. | |
4646 | * | |
4647 | * For space-based metaslabs, we use the maximum block size. | |
4648 | * This information is only available when the metaslab | |
4649 | * is loaded and is more accurate than the generic free | |
4650 | * space weight that was calculated by metaslab_weight(). | |
4651 | * This information allows us to quickly compare the maximum | |
4652 | * available allocation in the metaslab to the allocation | |
4653 | * size being requested. | |
4654 | * | |
4655 | * For segment-based metaslabs, determine the new weight | |
4656 | * based on the highest bucket in the range tree. We | |
4657 | * explicitly use the loaded segment weight (i.e. the range | |
4658 | * tree histogram) since it contains the space that is | |
4659 | * currently available for allocation and is accurate | |
4660 | * even within a sync pass. | |
4661 | */ | |
679b0f2a | 4662 | uint64_t weight; |
4e21fd06 | 4663 | if (WEIGHT_IS_SPACEBASED(msp->ms_weight)) { |
c81f1790 | 4664 | weight = metaslab_largest_allocatable(msp); |
4e21fd06 | 4665 | WEIGHT_SET_SPACEBASED(weight); |
679b0f2a PD |
4666 | } else { |
4667 | weight = metaslab_weight_from_range_tree(msp); | |
4668 | } | |
4669 | ||
4670 | if (activated) { | |
4e21fd06 DB |
4671 | metaslab_passivate(msp, weight); |
4672 | } else { | |
679b0f2a PD |
4673 | /* |
4674 | * For the case where we use the metaslab that is | |
4675 | * active for another allocator we want to make | |
4676 | * sure that we retain the activation mask. | |
4677 | * | |
4678 | * Note that we could attempt to use something like | |
4679 | * metaslab_recalculate_weight_and_sort() that | |
4680 | * retains the activation mask here. That function | |
4681 | * uses metaslab_weight() to set the weight though | |
4682 | * which is not as accurate as the calculations | |
4683 | * above. | |
4684 | */ | |
4685 | weight |= msp->ms_weight & METASLAB_ACTIVE_MASK; | |
4686 | metaslab_group_sort(mg, msp, weight); | |
4e21fd06 | 4687 | } |
679b0f2a | 4688 | metaslab_active_mask_verify(msp); |
34dc7c2f | 4689 | |
4e21fd06 DB |
4690 | /* |
4691 | * We have just failed an allocation attempt, check | |
4692 | * that metaslab_should_allocate() agrees. Otherwise, | |
4693 | * we may end up in an infinite loop retrying the same | |
4694 | * metaslab. | |
4695 | */ | |
c81f1790 | 4696 | ASSERT(!metaslab_should_allocate(msp, asize, try_hard)); |
cc99f275 | 4697 | |
34dc7c2f BB |
4698 | mutex_exit(&msp->ms_lock); |
4699 | } | |
4e21fd06 DB |
4700 | mutex_exit(&msp->ms_lock); |
4701 | kmem_free(search, sizeof (*search)); | |
4702 | return (offset); | |
4703 | } | |
34dc7c2f | 4704 | |
4e21fd06 DB |
4705 | static uint64_t |
4706 | metaslab_group_alloc(metaslab_group_t *mg, zio_alloc_list_t *zal, | |
c81f1790 PD |
4707 | uint64_t asize, uint64_t txg, boolean_t want_unique, dva_t *dva, int d, |
4708 | int allocator, boolean_t try_hard) | |
4e21fd06 DB |
4709 | { |
4710 | uint64_t offset; | |
4711 | ASSERT(mg->mg_initialized); | |
34dc7c2f | 4712 | |
cc99f275 | 4713 | offset = metaslab_group_alloc_normal(mg, zal, asize, txg, want_unique, |
c81f1790 | 4714 | dva, d, allocator, try_hard); |
34dc7c2f | 4715 | |
4e21fd06 DB |
4716 | mutex_enter(&mg->mg_lock); |
4717 | if (offset == -1ULL) { | |
4718 | mg->mg_failed_allocations++; | |
4719 | metaslab_trace_add(zal, mg, NULL, asize, d, | |
492f64e9 | 4720 | TRACE_GROUP_FAILURE, allocator); |
4e21fd06 DB |
4721 | if (asize == SPA_GANGBLOCKSIZE) { |
4722 | /* | |
4723 | * This metaslab group was unable to allocate | |
4724 | * the minimum gang block size so it must be out of | |
4725 | * space. We must notify the allocation throttle | |
4726 | * to start skipping allocation attempts to this | |
4727 | * metaslab group until more space becomes available. | |
4728 | * Note: this failure cannot be caused by the | |
4729 | * allocation throttle since the allocation throttle | |
4730 | * is only responsible for skipping devices and | |
4731 | * not failing block allocations. | |
4732 | */ | |
4733 | mg->mg_no_free_space = B_TRUE; | |
4734 | } | |
4735 | } | |
4736 | mg->mg_allocations++; | |
4737 | mutex_exit(&mg->mg_lock); | |
34dc7c2f BB |
4738 | return (offset); |
4739 | } | |
4740 | ||
4741 | /* | |
4742 | * Allocate a block for the specified i/o. | |
4743 | */ | |
a1d477c2 | 4744 | int |
34dc7c2f | 4745 | metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize, |
4e21fd06 | 4746 | dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags, |
492f64e9 | 4747 | zio_alloc_list_t *zal, int allocator) |
34dc7c2f | 4748 | { |
920dd524 | 4749 | metaslab_group_t *mg, *fast_mg, *rotor; |
34dc7c2f | 4750 | vdev_t *vd; |
4e21fd06 | 4751 | boolean_t try_hard = B_FALSE; |
34dc7c2f BB |
4752 | |
4753 | ASSERT(!DVA_IS_VALID(&dva[d])); | |
4754 | ||
4755 | /* | |
4756 | * For testing, make some blocks above a certain size be gang blocks. | |
09b85f2d BB |
4757 | * This will result in more split blocks when using device removal, |
4758 | * and a large number of split blocks coupled with ztest-induced | |
4759 | * damage can result in extremely long reconstruction times. This | |
4760 | * will also test spilling from special to normal. | |
34dc7c2f | 4761 | */ |
09b85f2d | 4762 | if (psize >= metaslab_force_ganging && (spa_get_random(100) < 3)) { |
492f64e9 PD |
4763 | metaslab_trace_add(zal, NULL, NULL, psize, d, TRACE_FORCE_GANG, |
4764 | allocator); | |
2e528b49 | 4765 | return (SET_ERROR(ENOSPC)); |
4e21fd06 | 4766 | } |
34dc7c2f BB |
4767 | |
4768 | /* | |
4769 | * Start at the rotor and loop through all mgs until we find something. | |
428870ff | 4770 | * Note that there's no locking on mc_rotor or mc_aliquot because |
34dc7c2f BB |
4771 | * nothing actually breaks if we miss a few updates -- we just won't |
4772 | * allocate quite as evenly. It all balances out over time. | |
4773 | * | |
4774 | * If we are doing ditto or log blocks, try to spread them across | |
4775 | * consecutive vdevs. If we're forced to reuse a vdev before we've | |
4776 | * allocated all of our ditto blocks, then try and spread them out on | |
4777 | * that vdev as much as possible. If it turns out to not be possible, | |
4778 | * gradually lower our standards until anything becomes acceptable. | |
4779 | * Also, allocating on consecutive vdevs (as opposed to random vdevs) | |
4780 | * gives us hope of containing our fault domains to something we're | |
4781 | * able to reason about. Otherwise, any two top-level vdev failures | |
4782 | * will guarantee the loss of data. With consecutive allocation, | |
4783 | * only two adjacent top-level vdev failures will result in data loss. | |
4784 | * | |
4785 | * If we are doing gang blocks (hintdva is non-NULL), try to keep | |
4786 | * ourselves on the same vdev as our gang block header. That | |
4787 | * way, we can hope for locality in vdev_cache, plus it makes our | |
4788 | * fault domains something tractable. | |
4789 | */ | |
4790 | if (hintdva) { | |
4791 | vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d])); | |
428870ff BB |
4792 | |
4793 | /* | |
4794 | * It's possible the vdev we're using as the hint no | |
a1d477c2 MA |
4795 | * longer exists or its mg has been closed (e.g. by |
4796 | * device removal). Consult the rotor when | |
428870ff BB |
4797 | * all else fails. |
4798 | */ | |
a1d477c2 | 4799 | if (vd != NULL && vd->vdev_mg != NULL) { |
34dc7c2f | 4800 | mg = vd->vdev_mg; |
428870ff BB |
4801 | |
4802 | if (flags & METASLAB_HINTBP_AVOID && | |
4803 | mg->mg_next != NULL) | |
4804 | mg = mg->mg_next; | |
4805 | } else { | |
4806 | mg = mc->mc_rotor; | |
4807 | } | |
34dc7c2f BB |
4808 | } else if (d != 0) { |
4809 | vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1])); | |
4810 | mg = vd->vdev_mg->mg_next; | |
920dd524 ED |
4811 | } else if (flags & METASLAB_FASTWRITE) { |
4812 | mg = fast_mg = mc->mc_rotor; | |
4813 | ||
4814 | do { | |
4815 | if (fast_mg->mg_vd->vdev_pending_fastwrite < | |
4816 | mg->mg_vd->vdev_pending_fastwrite) | |
4817 | mg = fast_mg; | |
4818 | } while ((fast_mg = fast_mg->mg_next) != mc->mc_rotor); | |
4819 | ||
34dc7c2f | 4820 | } else { |
cc99f275 | 4821 | ASSERT(mc->mc_rotor != NULL); |
34dc7c2f BB |
4822 | mg = mc->mc_rotor; |
4823 | } | |
4824 | ||
4825 | /* | |
428870ff BB |
4826 | * If the hint put us into the wrong metaslab class, or into a |
4827 | * metaslab group that has been passivated, just follow the rotor. | |
34dc7c2f | 4828 | */ |
428870ff | 4829 | if (mg->mg_class != mc || mg->mg_activation_count <= 0) |
34dc7c2f BB |
4830 | mg = mc->mc_rotor; |
4831 | ||
4832 | rotor = mg; | |
4833 | top: | |
34dc7c2f | 4834 | do { |
4e21fd06 | 4835 | boolean_t allocatable; |
428870ff | 4836 | |
3dfb57a3 | 4837 | ASSERT(mg->mg_activation_count == 1); |
34dc7c2f | 4838 | vd = mg->mg_vd; |
fb5f0bc8 | 4839 | |
34dc7c2f | 4840 | /* |
b128c09f | 4841 | * Don't allocate from faulted devices. |
34dc7c2f | 4842 | */ |
4e21fd06 | 4843 | if (try_hard) { |
fb5f0bc8 BB |
4844 | spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER); |
4845 | allocatable = vdev_allocatable(vd); | |
4846 | spa_config_exit(spa, SCL_ZIO, FTAG); | |
4847 | } else { | |
4848 | allocatable = vdev_allocatable(vd); | |
4849 | } | |
ac72fac3 GW |
4850 | |
4851 | /* | |
4852 | * Determine if the selected metaslab group is eligible | |
3dfb57a3 DB |
4853 | * for allocations. If we're ganging then don't allow |
4854 | * this metaslab group to skip allocations since that would | |
4855 | * inadvertently return ENOSPC and suspend the pool | |
ac72fac3 GW |
4856 | * even though space is still available. |
4857 | */ | |
4e21fd06 | 4858 | if (allocatable && !GANG_ALLOCATION(flags) && !try_hard) { |
3dfb57a3 | 4859 | allocatable = metaslab_group_allocatable(mg, rotor, |
c197a77c | 4860 | psize, allocator, d); |
3dfb57a3 | 4861 | } |
ac72fac3 | 4862 | |
4e21fd06 DB |
4863 | if (!allocatable) { |
4864 | metaslab_trace_add(zal, mg, NULL, psize, d, | |
492f64e9 | 4865 | TRACE_NOT_ALLOCATABLE, allocator); |
34dc7c2f | 4866 | goto next; |
4e21fd06 | 4867 | } |
fb5f0bc8 | 4868 | |
3dfb57a3 DB |
4869 | ASSERT(mg->mg_initialized); |
4870 | ||
34dc7c2f | 4871 | /* |
4e21fd06 DB |
4872 | * Avoid writing single-copy data to a failing, |
4873 | * non-redundant vdev, unless we've already tried all | |
4874 | * other vdevs. | |
34dc7c2f BB |
4875 | */ |
4876 | if ((vd->vdev_stat.vs_write_errors > 0 || | |
4877 | vd->vdev_state < VDEV_STATE_HEALTHY) && | |
4e21fd06 DB |
4878 | d == 0 && !try_hard && vd->vdev_children == 0) { |
4879 | metaslab_trace_add(zal, mg, NULL, psize, d, | |
492f64e9 | 4880 | TRACE_VDEV_ERROR, allocator); |
34dc7c2f BB |
4881 | goto next; |
4882 | } | |
4883 | ||
4884 | ASSERT(mg->mg_class == mc); | |
4885 | ||
1c27024e | 4886 | uint64_t asize = vdev_psize_to_asize(vd, psize); |
34dc7c2f BB |
4887 | ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0); |
4888 | ||
cc99f275 DB |
4889 | /* |
4890 | * If we don't need to try hard, then require that the | |
e1cfd73f | 4891 | * block be on a different metaslab from any other DVAs |
cc99f275 DB |
4892 | * in this BP (unique=true). If we are trying hard, then |
4893 | * allow any metaslab to be used (unique=false). | |
4894 | */ | |
1c27024e | 4895 | uint64_t offset = metaslab_group_alloc(mg, zal, asize, txg, |
c81f1790 | 4896 | !try_hard, dva, d, allocator, try_hard); |
3dfb57a3 | 4897 | |
34dc7c2f BB |
4898 | if (offset != -1ULL) { |
4899 | /* | |
4900 | * If we've just selected this metaslab group, | |
4901 | * figure out whether the corresponding vdev is | |
4902 | * over- or under-used relative to the pool, | |
4903 | * and set an allocation bias to even it out. | |
bb3250d0 ED |
4904 | * |
4905 | * Bias is also used to compensate for unequally | |
4906 | * sized vdevs so that space is allocated fairly. | |
34dc7c2f | 4907 | */ |
f3a7f661 | 4908 | if (mc->mc_aliquot == 0 && metaslab_bias_enabled) { |
34dc7c2f | 4909 | vdev_stat_t *vs = &vd->vdev_stat; |
bb3250d0 ED |
4910 | int64_t vs_free = vs->vs_space - vs->vs_alloc; |
4911 | int64_t mc_free = mc->mc_space - mc->mc_alloc; | |
4912 | int64_t ratio; | |
34dc7c2f BB |
4913 | |
4914 | /* | |
6d974228 GW |
4915 | * Calculate how much more or less we should |
4916 | * try to allocate from this device during | |
4917 | * this iteration around the rotor. | |
6d974228 | 4918 | * |
bb3250d0 ED |
4919 | * This basically introduces a zero-centered |
4920 | * bias towards the devices with the most | |
4921 | * free space, while compensating for vdev | |
4922 | * size differences. | |
4923 | * | |
4924 | * Examples: | |
4925 | * vdev V1 = 16M/128M | |
4926 | * vdev V2 = 16M/128M | |
4927 | * ratio(V1) = 100% ratio(V2) = 100% | |
4928 | * | |
4929 | * vdev V1 = 16M/128M | |
4930 | * vdev V2 = 64M/128M | |
4931 | * ratio(V1) = 127% ratio(V2) = 72% | |
6d974228 | 4932 | * |
bb3250d0 ED |
4933 | * vdev V1 = 16M/128M |
4934 | * vdev V2 = 64M/512M | |
4935 | * ratio(V1) = 40% ratio(V2) = 160% | |
34dc7c2f | 4936 | */ |
bb3250d0 ED |
4937 | ratio = (vs_free * mc->mc_alloc_groups * 100) / |
4938 | (mc_free + 1); | |
4939 | mg->mg_bias = ((ratio - 100) * | |
6d974228 | 4940 | (int64_t)mg->mg_aliquot) / 100; |
f3a7f661 GW |
4941 | } else if (!metaslab_bias_enabled) { |
4942 | mg->mg_bias = 0; | |
34dc7c2f BB |
4943 | } |
4944 | ||
920dd524 ED |
4945 | if ((flags & METASLAB_FASTWRITE) || |
4946 | atomic_add_64_nv(&mc->mc_aliquot, asize) >= | |
34dc7c2f BB |
4947 | mg->mg_aliquot + mg->mg_bias) { |
4948 | mc->mc_rotor = mg->mg_next; | |
428870ff | 4949 | mc->mc_aliquot = 0; |
34dc7c2f BB |
4950 | } |
4951 | ||
4952 | DVA_SET_VDEV(&dva[d], vd->vdev_id); | |
4953 | DVA_SET_OFFSET(&dva[d], offset); | |
e3e7cf60 D |
4954 | DVA_SET_GANG(&dva[d], |
4955 | ((flags & METASLAB_GANG_HEADER) ? 1 : 0)); | |
34dc7c2f BB |
4956 | DVA_SET_ASIZE(&dva[d], asize); |
4957 | ||
920dd524 ED |
4958 | if (flags & METASLAB_FASTWRITE) { |
4959 | atomic_add_64(&vd->vdev_pending_fastwrite, | |
4960 | psize); | |
920dd524 ED |
4961 | } |
4962 | ||
34dc7c2f BB |
4963 | return (0); |
4964 | } | |
4965 | next: | |
4966 | mc->mc_rotor = mg->mg_next; | |
428870ff | 4967 | mc->mc_aliquot = 0; |
34dc7c2f BB |
4968 | } while ((mg = mg->mg_next) != rotor); |
4969 | ||
4e21fd06 DB |
4970 | /* |
4971 | * If we haven't tried hard, do so now. | |
4972 | */ | |
4973 | if (!try_hard) { | |
4974 | try_hard = B_TRUE; | |
fb5f0bc8 BB |
4975 | goto top; |
4976 | } | |
4977 | ||
34dc7c2f BB |
4978 | bzero(&dva[d], sizeof (dva_t)); |
4979 | ||
492f64e9 | 4980 | metaslab_trace_add(zal, rotor, NULL, psize, d, TRACE_ENOSPC, allocator); |
2e528b49 | 4981 | return (SET_ERROR(ENOSPC)); |
34dc7c2f BB |
4982 | } |
4983 | ||
a1d477c2 MA |
4984 | void |
4985 | metaslab_free_concrete(vdev_t *vd, uint64_t offset, uint64_t asize, | |
d2734cce | 4986 | boolean_t checkpoint) |
a1d477c2 MA |
4987 | { |
4988 | metaslab_t *msp; | |
d2734cce | 4989 | spa_t *spa = vd->vdev_spa; |
a1d477c2 | 4990 | |
a1d477c2 MA |
4991 | ASSERT(vdev_is_concrete(vd)); |
4992 | ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0); | |
4993 | ASSERT3U(offset >> vd->vdev_ms_shift, <, vd->vdev_ms_count); | |
4994 | ||
4995 | msp = vd->vdev_ms[offset >> vd->vdev_ms_shift]; | |
4996 | ||
4997 | VERIFY(!msp->ms_condensing); | |
4998 | VERIFY3U(offset, >=, msp->ms_start); | |
4999 | VERIFY3U(offset + asize, <=, msp->ms_start + msp->ms_size); | |
5000 | VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift)); | |
5001 | VERIFY0(P2PHASE(asize, 1ULL << vd->vdev_ashift)); | |
5002 | ||
5003 | metaslab_check_free_impl(vd, offset, asize); | |
d2734cce | 5004 | |
a1d477c2 | 5005 | mutex_enter(&msp->ms_lock); |
d2734cce SD |
5006 | if (range_tree_is_empty(msp->ms_freeing) && |
5007 | range_tree_is_empty(msp->ms_checkpointing)) { | |
5008 | vdev_dirty(vd, VDD_METASLAB, msp, spa_syncing_txg(spa)); | |
5009 | } | |
5010 | ||
5011 | if (checkpoint) { | |
5012 | ASSERT(spa_has_checkpoint(spa)); | |
5013 | range_tree_add(msp->ms_checkpointing, offset, asize); | |
5014 | } else { | |
5015 | range_tree_add(msp->ms_freeing, offset, asize); | |
a1d477c2 | 5016 | } |
a1d477c2 MA |
5017 | mutex_exit(&msp->ms_lock); |
5018 | } | |
5019 | ||
5020 | /* ARGSUSED */ | |
5021 | void | |
5022 | metaslab_free_impl_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset, | |
5023 | uint64_t size, void *arg) | |
5024 | { | |
d2734cce SD |
5025 | boolean_t *checkpoint = arg; |
5026 | ||
5027 | ASSERT3P(checkpoint, !=, NULL); | |
a1d477c2 MA |
5028 | |
5029 | if (vd->vdev_ops->vdev_op_remap != NULL) | |
d2734cce | 5030 | vdev_indirect_mark_obsolete(vd, offset, size); |
a1d477c2 | 5031 | else |
d2734cce | 5032 | metaslab_free_impl(vd, offset, size, *checkpoint); |
a1d477c2 MA |
5033 | } |
5034 | ||
5035 | static void | |
5036 | metaslab_free_impl(vdev_t *vd, uint64_t offset, uint64_t size, | |
d2734cce | 5037 | boolean_t checkpoint) |
a1d477c2 MA |
5038 | { |
5039 | spa_t *spa = vd->vdev_spa; | |
5040 | ||
5041 | ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0); | |
5042 | ||
d2734cce | 5043 | if (spa_syncing_txg(spa) > spa_freeze_txg(spa)) |
a1d477c2 MA |
5044 | return; |
5045 | ||
5046 | if (spa->spa_vdev_removal != NULL && | |
9e052db4 | 5047 | spa->spa_vdev_removal->svr_vdev_id == vd->vdev_id && |
a1d477c2 MA |
5048 | vdev_is_concrete(vd)) { |
5049 | /* | |
5050 | * Note: we check if the vdev is concrete because when | |
5051 | * we complete the removal, we first change the vdev to be | |
5052 | * an indirect vdev (in open context), and then (in syncing | |
5053 | * context) clear spa_vdev_removal. | |
5054 | */ | |
d2734cce | 5055 | free_from_removing_vdev(vd, offset, size); |
a1d477c2 | 5056 | } else if (vd->vdev_ops->vdev_op_remap != NULL) { |
d2734cce | 5057 | vdev_indirect_mark_obsolete(vd, offset, size); |
a1d477c2 | 5058 | vd->vdev_ops->vdev_op_remap(vd, offset, size, |
d2734cce | 5059 | metaslab_free_impl_cb, &checkpoint); |
a1d477c2 | 5060 | } else { |
d2734cce | 5061 | metaslab_free_concrete(vd, offset, size, checkpoint); |
a1d477c2 MA |
5062 | } |
5063 | } | |
5064 | ||
5065 | typedef struct remap_blkptr_cb_arg { | |
5066 | blkptr_t *rbca_bp; | |
5067 | spa_remap_cb_t rbca_cb; | |
5068 | vdev_t *rbca_remap_vd; | |
5069 | uint64_t rbca_remap_offset; | |
5070 | void *rbca_cb_arg; | |
5071 | } remap_blkptr_cb_arg_t; | |
5072 | ||
5073 | void | |
5074 | remap_blkptr_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset, | |
5075 | uint64_t size, void *arg) | |
5076 | { | |
5077 | remap_blkptr_cb_arg_t *rbca = arg; | |
5078 | blkptr_t *bp = rbca->rbca_bp; | |
5079 | ||
5080 | /* We can not remap split blocks. */ | |
5081 | if (size != DVA_GET_ASIZE(&bp->blk_dva[0])) | |
5082 | return; | |
5083 | ASSERT0(inner_offset); | |
5084 | ||
5085 | if (rbca->rbca_cb != NULL) { | |
5086 | /* | |
5087 | * At this point we know that we are not handling split | |
5088 | * blocks and we invoke the callback on the previous | |
5089 | * vdev which must be indirect. | |
5090 | */ | |
5091 | ASSERT3P(rbca->rbca_remap_vd->vdev_ops, ==, &vdev_indirect_ops); | |
5092 | ||
5093 | rbca->rbca_cb(rbca->rbca_remap_vd->vdev_id, | |
5094 | rbca->rbca_remap_offset, size, rbca->rbca_cb_arg); | |
5095 | ||
5096 | /* set up remap_blkptr_cb_arg for the next call */ | |
5097 | rbca->rbca_remap_vd = vd; | |
5098 | rbca->rbca_remap_offset = offset; | |
5099 | } | |
5100 | ||
5101 | /* | |
5102 | * The phys birth time is that of dva[0]. This ensures that we know | |
5103 | * when each dva was written, so that resilver can determine which | |
5104 | * blocks need to be scrubbed (i.e. those written during the time | |
5105 | * the vdev was offline). It also ensures that the key used in | |
5106 | * the ARC hash table is unique (i.e. dva[0] + phys_birth). If | |
5107 | * we didn't change the phys_birth, a lookup in the ARC for a | |
5108 | * remapped BP could find the data that was previously stored at | |
5109 | * this vdev + offset. | |
5110 | */ | |
5111 | vdev_t *oldvd = vdev_lookup_top(vd->vdev_spa, | |
5112 | DVA_GET_VDEV(&bp->blk_dva[0])); | |
5113 | vdev_indirect_births_t *vib = oldvd->vdev_indirect_births; | |
5114 | bp->blk_phys_birth = vdev_indirect_births_physbirth(vib, | |
5115 | DVA_GET_OFFSET(&bp->blk_dva[0]), DVA_GET_ASIZE(&bp->blk_dva[0])); | |
5116 | ||
5117 | DVA_SET_VDEV(&bp->blk_dva[0], vd->vdev_id); | |
5118 | DVA_SET_OFFSET(&bp->blk_dva[0], offset); | |
5119 | } | |
5120 | ||
34dc7c2f | 5121 | /* |
a1d477c2 MA |
5122 | * If the block pointer contains any indirect DVAs, modify them to refer to |
5123 | * concrete DVAs. Note that this will sometimes not be possible, leaving | |
5124 | * the indirect DVA in place. This happens if the indirect DVA spans multiple | |
5125 | * segments in the mapping (i.e. it is a "split block"). | |
5126 | * | |
5127 | * If the BP was remapped, calls the callback on the original dva (note the | |
5128 | * callback can be called multiple times if the original indirect DVA refers | |
5129 | * to another indirect DVA, etc). | |
5130 | * | |
5131 | * Returns TRUE if the BP was remapped. | |
34dc7c2f | 5132 | */ |
a1d477c2 MA |
5133 | boolean_t |
5134 | spa_remap_blkptr(spa_t *spa, blkptr_t *bp, spa_remap_cb_t callback, void *arg) | |
34dc7c2f | 5135 | { |
a1d477c2 MA |
5136 | remap_blkptr_cb_arg_t rbca; |
5137 | ||
5138 | if (!zfs_remap_blkptr_enable) | |
5139 | return (B_FALSE); | |
5140 | ||
5141 | if (!spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS)) | |
5142 | return (B_FALSE); | |
5143 | ||
5144 | /* | |
5145 | * Dedup BP's can not be remapped, because ddt_phys_select() depends | |
5146 | * on DVA[0] being the same in the BP as in the DDT (dedup table). | |
5147 | */ | |
5148 | if (BP_GET_DEDUP(bp)) | |
5149 | return (B_FALSE); | |
5150 | ||
5151 | /* | |
5152 | * Gang blocks can not be remapped, because | |
5153 | * zio_checksum_gang_verifier() depends on the DVA[0] that's in | |
5154 | * the BP used to read the gang block header (GBH) being the same | |
5155 | * as the DVA[0] that we allocated for the GBH. | |
5156 | */ | |
5157 | if (BP_IS_GANG(bp)) | |
5158 | return (B_FALSE); | |
5159 | ||
5160 | /* | |
5161 | * Embedded BP's have no DVA to remap. | |
5162 | */ | |
5163 | if (BP_GET_NDVAS(bp) < 1) | |
5164 | return (B_FALSE); | |
5165 | ||
5166 | /* | |
5167 | * Note: we only remap dva[0]. If we remapped other dvas, we | |
5168 | * would no longer know what their phys birth txg is. | |
5169 | */ | |
5170 | dva_t *dva = &bp->blk_dva[0]; | |
5171 | ||
34dc7c2f BB |
5172 | uint64_t offset = DVA_GET_OFFSET(dva); |
5173 | uint64_t size = DVA_GET_ASIZE(dva); | |
a1d477c2 MA |
5174 | vdev_t *vd = vdev_lookup_top(spa, DVA_GET_VDEV(dva)); |
5175 | ||
5176 | if (vd->vdev_ops->vdev_op_remap == NULL) | |
5177 | return (B_FALSE); | |
5178 | ||
5179 | rbca.rbca_bp = bp; | |
5180 | rbca.rbca_cb = callback; | |
5181 | rbca.rbca_remap_vd = vd; | |
5182 | rbca.rbca_remap_offset = offset; | |
5183 | rbca.rbca_cb_arg = arg; | |
5184 | ||
5185 | /* | |
5186 | * remap_blkptr_cb() will be called in order for each level of | |
5187 | * indirection, until a concrete vdev is reached or a split block is | |
5188 | * encountered. old_vd and old_offset are updated within the callback | |
5189 | * as we go from the one indirect vdev to the next one (either concrete | |
5190 | * or indirect again) in that order. | |
5191 | */ | |
5192 | vd->vdev_ops->vdev_op_remap(vd, offset, size, remap_blkptr_cb, &rbca); | |
5193 | ||
5194 | /* Check if the DVA wasn't remapped because it is a split block */ | |
5195 | if (DVA_GET_VDEV(&rbca.rbca_bp->blk_dva[0]) == vd->vdev_id) | |
5196 | return (B_FALSE); | |
5197 | ||
5198 | return (B_TRUE); | |
5199 | } | |
5200 | ||
5201 | /* | |
5202 | * Undo the allocation of a DVA which happened in the given transaction group. | |
5203 | */ | |
5204 | void | |
5205 | metaslab_unalloc_dva(spa_t *spa, const dva_t *dva, uint64_t txg) | |
5206 | { | |
34dc7c2f | 5207 | metaslab_t *msp; |
a1d477c2 MA |
5208 | vdev_t *vd; |
5209 | uint64_t vdev = DVA_GET_VDEV(dva); | |
5210 | uint64_t offset = DVA_GET_OFFSET(dva); | |
5211 | uint64_t size = DVA_GET_ASIZE(dva); | |
5212 | ||
5213 | ASSERT(DVA_IS_VALID(dva)); | |
5214 | ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0); | |
34dc7c2f | 5215 | |
34dc7c2f BB |
5216 | if (txg > spa_freeze_txg(spa)) |
5217 | return; | |
5218 | ||
7d2868d5 | 5219 | if ((vd = vdev_lookup_top(spa, vdev)) == NULL || !DVA_IS_VALID(dva) || |
34dc7c2f | 5220 | (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) { |
7d2868d5 BB |
5221 | zfs_panic_recover("metaslab_free_dva(): bad DVA %llu:%llu:%llu", |
5222 | (u_longlong_t)vdev, (u_longlong_t)offset, | |
5223 | (u_longlong_t)size); | |
34dc7c2f BB |
5224 | return; |
5225 | } | |
5226 | ||
a1d477c2 MA |
5227 | ASSERT(!vd->vdev_removing); |
5228 | ASSERT(vdev_is_concrete(vd)); | |
5229 | ASSERT0(vd->vdev_indirect_config.vic_mapping_object); | |
5230 | ASSERT3P(vd->vdev_indirect_mapping, ==, NULL); | |
34dc7c2f BB |
5231 | |
5232 | if (DVA_GET_GANG(dva)) | |
5233 | size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); | |
5234 | ||
a1d477c2 | 5235 | msp = vd->vdev_ms[offset >> vd->vdev_ms_shift]; |
93cf2076 | 5236 | |
a1d477c2 | 5237 | mutex_enter(&msp->ms_lock); |
d2734cce | 5238 | range_tree_remove(msp->ms_allocating[txg & TXG_MASK], |
a1d477c2 | 5239 | offset, size); |
f09fda50 | 5240 | msp->ms_allocating_total -= size; |
34dc7c2f | 5241 | |
a1d477c2 MA |
5242 | VERIFY(!msp->ms_condensing); |
5243 | VERIFY3U(offset, >=, msp->ms_start); | |
5244 | VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size); | |
d2734cce | 5245 | VERIFY3U(range_tree_space(msp->ms_allocatable) + size, <=, |
a1d477c2 MA |
5246 | msp->ms_size); |
5247 | VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift)); | |
5248 | VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift)); | |
d2734cce | 5249 | range_tree_add(msp->ms_allocatable, offset, size); |
34dc7c2f BB |
5250 | mutex_exit(&msp->ms_lock); |
5251 | } | |
5252 | ||
5253 | /* | |
d2734cce | 5254 | * Free the block represented by the given DVA. |
34dc7c2f | 5255 | */ |
a1d477c2 | 5256 | void |
d2734cce | 5257 | metaslab_free_dva(spa_t *spa, const dva_t *dva, boolean_t checkpoint) |
34dc7c2f BB |
5258 | { |
5259 | uint64_t vdev = DVA_GET_VDEV(dva); | |
5260 | uint64_t offset = DVA_GET_OFFSET(dva); | |
5261 | uint64_t size = DVA_GET_ASIZE(dva); | |
a1d477c2 | 5262 | vdev_t *vd = vdev_lookup_top(spa, vdev); |
34dc7c2f BB |
5263 | |
5264 | ASSERT(DVA_IS_VALID(dva)); | |
a1d477c2 | 5265 | ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0); |
34dc7c2f | 5266 | |
a1d477c2 | 5267 | if (DVA_GET_GANG(dva)) { |
34dc7c2f | 5268 | size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); |
34dc7c2f BB |
5269 | } |
5270 | ||
d2734cce | 5271 | metaslab_free_impl(vd, offset, size, checkpoint); |
34dc7c2f BB |
5272 | } |
5273 | ||
3dfb57a3 DB |
5274 | /* |
5275 | * Reserve some allocation slots. The reservation system must be called | |
5276 | * before we call into the allocator. If there aren't any available slots | |
5277 | * then the I/O will be throttled until an I/O completes and its slots are | |
5278 | * freed up. The function returns true if it was successful in placing | |
5279 | * the reservation. | |
5280 | */ | |
5281 | boolean_t | |
492f64e9 PD |
5282 | metaslab_class_throttle_reserve(metaslab_class_t *mc, int slots, int allocator, |
5283 | zio_t *zio, int flags) | |
3dfb57a3 DB |
5284 | { |
5285 | uint64_t available_slots = 0; | |
3dfb57a3 | 5286 | boolean_t slot_reserved = B_FALSE; |
492f64e9 | 5287 | uint64_t max = mc->mc_alloc_max_slots[allocator]; |
3dfb57a3 DB |
5288 | |
5289 | ASSERT(mc->mc_alloc_throttle_enabled); | |
5290 | mutex_enter(&mc->mc_lock); | |
5291 | ||
492f64e9 | 5292 | uint64_t reserved_slots = |
424fd7c3 | 5293 | zfs_refcount_count(&mc->mc_alloc_slots[allocator]); |
492f64e9 PD |
5294 | if (reserved_slots < max) |
5295 | available_slots = max - reserved_slots; | |
3dfb57a3 | 5296 | |
cc99f275 DB |
5297 | if (slots <= available_slots || GANG_ALLOCATION(flags) || |
5298 | flags & METASLAB_MUST_RESERVE) { | |
3dfb57a3 DB |
5299 | /* |
5300 | * We reserve the slots individually so that we can unreserve | |
5301 | * them individually when an I/O completes. | |
5302 | */ | |
1c27024e | 5303 | for (int d = 0; d < slots; d++) { |
492f64e9 | 5304 | reserved_slots = |
c13060e4 | 5305 | zfs_refcount_add(&mc->mc_alloc_slots[allocator], |
492f64e9 | 5306 | zio); |
3dfb57a3 DB |
5307 | } |
5308 | zio->io_flags |= ZIO_FLAG_IO_ALLOCATING; | |
5309 | slot_reserved = B_TRUE; | |
5310 | } | |
5311 | ||
5312 | mutex_exit(&mc->mc_lock); | |
5313 | return (slot_reserved); | |
5314 | } | |
5315 | ||
5316 | void | |
492f64e9 PD |
5317 | metaslab_class_throttle_unreserve(metaslab_class_t *mc, int slots, |
5318 | int allocator, zio_t *zio) | |
3dfb57a3 | 5319 | { |
3dfb57a3 DB |
5320 | ASSERT(mc->mc_alloc_throttle_enabled); |
5321 | mutex_enter(&mc->mc_lock); | |
1c27024e | 5322 | for (int d = 0; d < slots; d++) { |
424fd7c3 | 5323 | (void) zfs_refcount_remove(&mc->mc_alloc_slots[allocator], |
492f64e9 | 5324 | zio); |
3dfb57a3 DB |
5325 | } |
5326 | mutex_exit(&mc->mc_lock); | |
5327 | } | |
5328 | ||
a1d477c2 MA |
5329 | static int |
5330 | metaslab_claim_concrete(vdev_t *vd, uint64_t offset, uint64_t size, | |
5331 | uint64_t txg) | |
5332 | { | |
5333 | metaslab_t *msp; | |
5334 | spa_t *spa = vd->vdev_spa; | |
5335 | int error = 0; | |
5336 | ||
5337 | if (offset >> vd->vdev_ms_shift >= vd->vdev_ms_count) | |
7ab96299 | 5338 | return (SET_ERROR(ENXIO)); |
a1d477c2 MA |
5339 | |
5340 | ASSERT3P(vd->vdev_ms, !=, NULL); | |
5341 | msp = vd->vdev_ms[offset >> vd->vdev_ms_shift]; | |
5342 | ||
5343 | mutex_enter(&msp->ms_lock); | |
5344 | ||
7ab96299 | 5345 | if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded) { |
492f64e9 | 5346 | error = metaslab_activate(msp, 0, METASLAB_WEIGHT_CLAIM); |
7ab96299 TC |
5347 | if (error == EBUSY) { |
5348 | ASSERT(msp->ms_loaded); | |
5349 | ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK); | |
5350 | error = 0; | |
5351 | } | |
5352 | } | |
a1d477c2 | 5353 | |
d2734cce SD |
5354 | if (error == 0 && |
5355 | !range_tree_contains(msp->ms_allocatable, offset, size)) | |
a1d477c2 MA |
5356 | error = SET_ERROR(ENOENT); |
5357 | ||
5358 | if (error || txg == 0) { /* txg == 0 indicates dry run */ | |
5359 | mutex_exit(&msp->ms_lock); | |
5360 | return (error); | |
5361 | } | |
5362 | ||
5363 | VERIFY(!msp->ms_condensing); | |
5364 | VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift)); | |
5365 | VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift)); | |
d2734cce SD |
5366 | VERIFY3U(range_tree_space(msp->ms_allocatable) - size, <=, |
5367 | msp->ms_size); | |
5368 | range_tree_remove(msp->ms_allocatable, offset, size); | |
1b939560 | 5369 | range_tree_clear(msp->ms_trim, offset, size); |
a1d477c2 MA |
5370 | |
5371 | if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */ | |
f09fda50 PD |
5372 | metaslab_class_t *mc = msp->ms_group->mg_class; |
5373 | multilist_sublist_t *mls = | |
5374 | multilist_sublist_lock_obj(mc->mc_metaslab_txg_list, msp); | |
5375 | if (!multilist_link_active(&msp->ms_class_txg_node)) { | |
5376 | msp->ms_selected_txg = txg; | |
5377 | multilist_sublist_insert_head(mls, msp); | |
5378 | } | |
5379 | multilist_sublist_unlock(mls); | |
5380 | ||
d2734cce | 5381 | if (range_tree_is_empty(msp->ms_allocating[txg & TXG_MASK])) |
a1d477c2 | 5382 | vdev_dirty(vd, VDD_METASLAB, msp, txg); |
d2734cce SD |
5383 | range_tree_add(msp->ms_allocating[txg & TXG_MASK], |
5384 | offset, size); | |
f09fda50 | 5385 | msp->ms_allocating_total += size; |
a1d477c2 MA |
5386 | } |
5387 | ||
5388 | mutex_exit(&msp->ms_lock); | |
5389 | ||
5390 | return (0); | |
5391 | } | |
5392 | ||
5393 | typedef struct metaslab_claim_cb_arg_t { | |
5394 | uint64_t mcca_txg; | |
5395 | int mcca_error; | |
5396 | } metaslab_claim_cb_arg_t; | |
5397 | ||
5398 | /* ARGSUSED */ | |
5399 | static void | |
5400 | metaslab_claim_impl_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset, | |
5401 | uint64_t size, void *arg) | |
5402 | { | |
5403 | metaslab_claim_cb_arg_t *mcca_arg = arg; | |
5404 | ||
5405 | if (mcca_arg->mcca_error == 0) { | |
5406 | mcca_arg->mcca_error = metaslab_claim_concrete(vd, offset, | |
5407 | size, mcca_arg->mcca_txg); | |
5408 | } | |
5409 | } | |
5410 | ||
5411 | int | |
5412 | metaslab_claim_impl(vdev_t *vd, uint64_t offset, uint64_t size, uint64_t txg) | |
5413 | { | |
5414 | if (vd->vdev_ops->vdev_op_remap != NULL) { | |
5415 | metaslab_claim_cb_arg_t arg; | |
5416 | ||
5417 | /* | |
5418 | * Only zdb(1M) can claim on indirect vdevs. This is used | |
5419 | * to detect leaks of mapped space (that are not accounted | |
5420 | * for in the obsolete counts, spacemap, or bpobj). | |
5421 | */ | |
5422 | ASSERT(!spa_writeable(vd->vdev_spa)); | |
5423 | arg.mcca_error = 0; | |
5424 | arg.mcca_txg = txg; | |
5425 | ||
5426 | vd->vdev_ops->vdev_op_remap(vd, offset, size, | |
5427 | metaslab_claim_impl_cb, &arg); | |
5428 | ||
5429 | if (arg.mcca_error == 0) { | |
5430 | arg.mcca_error = metaslab_claim_concrete(vd, | |
5431 | offset, size, txg); | |
5432 | } | |
5433 | return (arg.mcca_error); | |
5434 | } else { | |
5435 | return (metaslab_claim_concrete(vd, offset, size, txg)); | |
5436 | } | |
5437 | } | |
5438 | ||
5439 | /* | |
5440 | * Intent log support: upon opening the pool after a crash, notify the SPA | |
5441 | * of blocks that the intent log has allocated for immediate write, but | |
5442 | * which are still considered free by the SPA because the last transaction | |
5443 | * group didn't commit yet. | |
5444 | */ | |
5445 | static int | |
5446 | metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg) | |
5447 | { | |
5448 | uint64_t vdev = DVA_GET_VDEV(dva); | |
5449 | uint64_t offset = DVA_GET_OFFSET(dva); | |
5450 | uint64_t size = DVA_GET_ASIZE(dva); | |
5451 | vdev_t *vd; | |
5452 | ||
5453 | if ((vd = vdev_lookup_top(spa, vdev)) == NULL) { | |
5454 | return (SET_ERROR(ENXIO)); | |
5455 | } | |
5456 | ||
5457 | ASSERT(DVA_IS_VALID(dva)); | |
5458 | ||
5459 | if (DVA_GET_GANG(dva)) | |
5460 | size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); | |
5461 | ||
5462 | return (metaslab_claim_impl(vd, offset, size, txg)); | |
5463 | } | |
5464 | ||
34dc7c2f BB |
5465 | int |
5466 | metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp, | |
4e21fd06 | 5467 | int ndvas, uint64_t txg, blkptr_t *hintbp, int flags, |
492f64e9 | 5468 | zio_alloc_list_t *zal, zio_t *zio, int allocator) |
34dc7c2f BB |
5469 | { |
5470 | dva_t *dva = bp->blk_dva; | |
928e8ad4 | 5471 | dva_t *hintdva = (hintbp != NULL) ? hintbp->blk_dva : NULL; |
1c27024e | 5472 | int error = 0; |
34dc7c2f | 5473 | |
b128c09f | 5474 | ASSERT(bp->blk_birth == 0); |
428870ff | 5475 | ASSERT(BP_PHYSICAL_BIRTH(bp) == 0); |
b128c09f BB |
5476 | |
5477 | spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER); | |
5478 | ||
5479 | if (mc->mc_rotor == NULL) { /* no vdevs in this class */ | |
5480 | spa_config_exit(spa, SCL_ALLOC, FTAG); | |
2e528b49 | 5481 | return (SET_ERROR(ENOSPC)); |
b128c09f | 5482 | } |
34dc7c2f BB |
5483 | |
5484 | ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa)); | |
5485 | ASSERT(BP_GET_NDVAS(bp) == 0); | |
5486 | ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp)); | |
4e21fd06 | 5487 | ASSERT3P(zal, !=, NULL); |
34dc7c2f | 5488 | |
1c27024e | 5489 | for (int d = 0; d < ndvas; d++) { |
34dc7c2f | 5490 | error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva, |
492f64e9 | 5491 | txg, flags, zal, allocator); |
93cf2076 | 5492 | if (error != 0) { |
34dc7c2f | 5493 | for (d--; d >= 0; d--) { |
a1d477c2 | 5494 | metaslab_unalloc_dva(spa, &dva[d], txg); |
3dfb57a3 | 5495 | metaslab_group_alloc_decrement(spa, |
492f64e9 PD |
5496 | DVA_GET_VDEV(&dva[d]), zio, flags, |
5497 | allocator, B_FALSE); | |
34dc7c2f BB |
5498 | bzero(&dva[d], sizeof (dva_t)); |
5499 | } | |
b128c09f | 5500 | spa_config_exit(spa, SCL_ALLOC, FTAG); |
34dc7c2f | 5501 | return (error); |
3dfb57a3 DB |
5502 | } else { |
5503 | /* | |
5504 | * Update the metaslab group's queue depth | |
5505 | * based on the newly allocated dva. | |
5506 | */ | |
5507 | metaslab_group_alloc_increment(spa, | |
492f64e9 | 5508 | DVA_GET_VDEV(&dva[d]), zio, flags, allocator); |
34dc7c2f | 5509 | } |
3dfb57a3 | 5510 | |
34dc7c2f BB |
5511 | } |
5512 | ASSERT(error == 0); | |
5513 | ASSERT(BP_GET_NDVAS(bp) == ndvas); | |
5514 | ||
b128c09f BB |
5515 | spa_config_exit(spa, SCL_ALLOC, FTAG); |
5516 | ||
efe7978d | 5517 | BP_SET_BIRTH(bp, txg, 0); |
b128c09f | 5518 | |
34dc7c2f BB |
5519 | return (0); |
5520 | } | |
5521 | ||
5522 | void | |
5523 | metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now) | |
5524 | { | |
5525 | const dva_t *dva = bp->blk_dva; | |
1c27024e | 5526 | int ndvas = BP_GET_NDVAS(bp); |
34dc7c2f BB |
5527 | |
5528 | ASSERT(!BP_IS_HOLE(bp)); | |
428870ff | 5529 | ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa)); |
b128c09f | 5530 | |
d2734cce SD |
5531 | /* |
5532 | * If we have a checkpoint for the pool we need to make sure that | |
5533 | * the blocks that we free that are part of the checkpoint won't be | |
5534 | * reused until the checkpoint is discarded or we revert to it. | |
5535 | * | |
5536 | * The checkpoint flag is passed down the metaslab_free code path | |
5537 | * and is set whenever we want to add a block to the checkpoint's | |
5538 | * accounting. That is, we "checkpoint" blocks that existed at the | |
5539 | * time the checkpoint was created and are therefore referenced by | |
5540 | * the checkpointed uberblock. | |
5541 | * | |
5542 | * Note that, we don't checkpoint any blocks if the current | |
5543 | * syncing txg <= spa_checkpoint_txg. We want these frees to sync | |
5544 | * normally as they will be referenced by the checkpointed uberblock. | |
5545 | */ | |
5546 | boolean_t checkpoint = B_FALSE; | |
5547 | if (bp->blk_birth <= spa->spa_checkpoint_txg && | |
5548 | spa_syncing_txg(spa) > spa->spa_checkpoint_txg) { | |
5549 | /* | |
5550 | * At this point, if the block is part of the checkpoint | |
5551 | * there is no way it was created in the current txg. | |
5552 | */ | |
5553 | ASSERT(!now); | |
5554 | ASSERT3U(spa_syncing_txg(spa), ==, txg); | |
5555 | checkpoint = B_TRUE; | |
5556 | } | |
5557 | ||
b128c09f | 5558 | spa_config_enter(spa, SCL_FREE, FTAG, RW_READER); |
34dc7c2f | 5559 | |
a1d477c2 MA |
5560 | for (int d = 0; d < ndvas; d++) { |
5561 | if (now) { | |
5562 | metaslab_unalloc_dva(spa, &dva[d], txg); | |
5563 | } else { | |
d2734cce SD |
5564 | ASSERT3U(txg, ==, spa_syncing_txg(spa)); |
5565 | metaslab_free_dva(spa, &dva[d], checkpoint); | |
a1d477c2 MA |
5566 | } |
5567 | } | |
b128c09f BB |
5568 | |
5569 | spa_config_exit(spa, SCL_FREE, FTAG); | |
34dc7c2f BB |
5570 | } |
5571 | ||
5572 | int | |
5573 | metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg) | |
5574 | { | |
5575 | const dva_t *dva = bp->blk_dva; | |
5576 | int ndvas = BP_GET_NDVAS(bp); | |
1c27024e | 5577 | int error = 0; |
34dc7c2f BB |
5578 | |
5579 | ASSERT(!BP_IS_HOLE(bp)); | |
5580 | ||
b128c09f BB |
5581 | if (txg != 0) { |
5582 | /* | |
5583 | * First do a dry run to make sure all DVAs are claimable, | |
5584 | * so we don't have to unwind from partial failures below. | |
5585 | */ | |
5586 | if ((error = metaslab_claim(spa, bp, 0)) != 0) | |
5587 | return (error); | |
5588 | } | |
5589 | ||
5590 | spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER); | |
5591 | ||
cc99f275 DB |
5592 | for (int d = 0; d < ndvas; d++) { |
5593 | error = metaslab_claim_dva(spa, &dva[d], txg); | |
5594 | if (error != 0) | |
b128c09f | 5595 | break; |
cc99f275 | 5596 | } |
b128c09f BB |
5597 | |
5598 | spa_config_exit(spa, SCL_ALLOC, FTAG); | |
5599 | ||
5600 | ASSERT(error == 0 || txg == 0); | |
34dc7c2f | 5601 | |
b128c09f | 5602 | return (error); |
34dc7c2f | 5603 | } |
920dd524 | 5604 | |
d1d7e268 MK |
5605 | void |
5606 | metaslab_fastwrite_mark(spa_t *spa, const blkptr_t *bp) | |
920dd524 ED |
5607 | { |
5608 | const dva_t *dva = bp->blk_dva; | |
5609 | int ndvas = BP_GET_NDVAS(bp); | |
5610 | uint64_t psize = BP_GET_PSIZE(bp); | |
5611 | int d; | |
5612 | vdev_t *vd; | |
5613 | ||
5614 | ASSERT(!BP_IS_HOLE(bp)); | |
9b67f605 | 5615 | ASSERT(!BP_IS_EMBEDDED(bp)); |
920dd524 ED |
5616 | ASSERT(psize > 0); |
5617 | ||
5618 | spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER); | |
5619 | ||
5620 | for (d = 0; d < ndvas; d++) { | |
5621 | if ((vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d]))) == NULL) | |
5622 | continue; | |
5623 | atomic_add_64(&vd->vdev_pending_fastwrite, psize); | |
5624 | } | |
5625 | ||
5626 | spa_config_exit(spa, SCL_VDEV, FTAG); | |
5627 | } | |
5628 | ||
d1d7e268 MK |
5629 | void |
5630 | metaslab_fastwrite_unmark(spa_t *spa, const blkptr_t *bp) | |
920dd524 ED |
5631 | { |
5632 | const dva_t *dva = bp->blk_dva; | |
5633 | int ndvas = BP_GET_NDVAS(bp); | |
5634 | uint64_t psize = BP_GET_PSIZE(bp); | |
5635 | int d; | |
5636 | vdev_t *vd; | |
5637 | ||
5638 | ASSERT(!BP_IS_HOLE(bp)); | |
9b67f605 | 5639 | ASSERT(!BP_IS_EMBEDDED(bp)); |
920dd524 ED |
5640 | ASSERT(psize > 0); |
5641 | ||
5642 | spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER); | |
5643 | ||
5644 | for (d = 0; d < ndvas; d++) { | |
5645 | if ((vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d]))) == NULL) | |
5646 | continue; | |
5647 | ASSERT3U(vd->vdev_pending_fastwrite, >=, psize); | |
5648 | atomic_sub_64(&vd->vdev_pending_fastwrite, psize); | |
5649 | } | |
5650 | ||
5651 | spa_config_exit(spa, SCL_VDEV, FTAG); | |
5652 | } | |
30b92c1d | 5653 | |
a1d477c2 MA |
5654 | /* ARGSUSED */ |
5655 | static void | |
5656 | metaslab_check_free_impl_cb(uint64_t inner, vdev_t *vd, uint64_t offset, | |
5657 | uint64_t size, void *arg) | |
5658 | { | |
5659 | if (vd->vdev_ops == &vdev_indirect_ops) | |
5660 | return; | |
5661 | ||
5662 | metaslab_check_free_impl(vd, offset, size); | |
5663 | } | |
5664 | ||
5665 | static void | |
5666 | metaslab_check_free_impl(vdev_t *vd, uint64_t offset, uint64_t size) | |
5667 | { | |
5668 | metaslab_t *msp; | |
5669 | ASSERTV(spa_t *spa = vd->vdev_spa); | |
5670 | ||
5671 | if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0) | |
5672 | return; | |
5673 | ||
5674 | if (vd->vdev_ops->vdev_op_remap != NULL) { | |
5675 | vd->vdev_ops->vdev_op_remap(vd, offset, size, | |
5676 | metaslab_check_free_impl_cb, NULL); | |
5677 | return; | |
5678 | } | |
5679 | ||
5680 | ASSERT(vdev_is_concrete(vd)); | |
5681 | ASSERT3U(offset >> vd->vdev_ms_shift, <, vd->vdev_ms_count); | |
5682 | ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0); | |
5683 | ||
5684 | msp = vd->vdev_ms[offset >> vd->vdev_ms_shift]; | |
5685 | ||
5686 | mutex_enter(&msp->ms_lock); | |
df72b8be SD |
5687 | if (msp->ms_loaded) { |
5688 | range_tree_verify_not_present(msp->ms_allocatable, | |
5689 | offset, size); | |
5690 | } | |
a1d477c2 | 5691 | |
93e28d66 SD |
5692 | /* |
5693 | * Check all segments that currently exist in the freeing pipeline. | |
5694 | * | |
5695 | * It would intuitively make sense to also check the current allocating | |
5696 | * tree since metaslab_unalloc_dva() exists for extents that are | |
e1cfd73f | 5697 | * allocated and freed in the same sync pass within the same txg. |
93e28d66 SD |
5698 | * Unfortunately there are places (e.g. the ZIL) where we allocate a |
5699 | * segment but then we free part of it within the same txg | |
5700 | * [see zil_sync()]. Thus, we don't call range_tree_verify() in the | |
5701 | * current allocating tree. | |
5702 | */ | |
df72b8be SD |
5703 | range_tree_verify_not_present(msp->ms_freeing, offset, size); |
5704 | range_tree_verify_not_present(msp->ms_checkpointing, offset, size); | |
5705 | range_tree_verify_not_present(msp->ms_freed, offset, size); | |
a1d477c2 | 5706 | for (int j = 0; j < TXG_DEFER_SIZE; j++) |
df72b8be | 5707 | range_tree_verify_not_present(msp->ms_defer[j], offset, size); |
93e28d66 | 5708 | range_tree_verify_not_present(msp->ms_trim, offset, size); |
a1d477c2 MA |
5709 | mutex_exit(&msp->ms_lock); |
5710 | } | |
5711 | ||
13fe0198 MA |
5712 | void |
5713 | metaslab_check_free(spa_t *spa, const blkptr_t *bp) | |
5714 | { | |
13fe0198 MA |
5715 | if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0) |
5716 | return; | |
5717 | ||
5718 | spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER); | |
1c27024e | 5719 | for (int i = 0; i < BP_GET_NDVAS(bp); i++) { |
93cf2076 GW |
5720 | uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]); |
5721 | vdev_t *vd = vdev_lookup_top(spa, vdev); | |
5722 | uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]); | |
13fe0198 | 5723 | uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]); |
13fe0198 | 5724 | |
a1d477c2 MA |
5725 | if (DVA_GET_GANG(&bp->blk_dva[i])) |
5726 | size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); | |
5727 | ||
5728 | ASSERT3P(vd, !=, NULL); | |
13fe0198 | 5729 | |
a1d477c2 | 5730 | metaslab_check_free_impl(vd, offset, size); |
13fe0198 MA |
5731 | } |
5732 | spa_config_exit(spa, SCL_VDEV, FTAG); | |
5733 | } | |
5734 | ||
1b939560 BB |
5735 | static void |
5736 | metaslab_group_disable_wait(metaslab_group_t *mg) | |
5737 | { | |
5738 | ASSERT(MUTEX_HELD(&mg->mg_ms_disabled_lock)); | |
5739 | while (mg->mg_disabled_updating) { | |
5740 | cv_wait(&mg->mg_ms_disabled_cv, &mg->mg_ms_disabled_lock); | |
5741 | } | |
5742 | } | |
5743 | ||
5744 | static void | |
5745 | metaslab_group_disabled_increment(metaslab_group_t *mg) | |
5746 | { | |
5747 | ASSERT(MUTEX_HELD(&mg->mg_ms_disabled_lock)); | |
5748 | ASSERT(mg->mg_disabled_updating); | |
5749 | ||
5750 | while (mg->mg_ms_disabled >= max_disabled_ms) { | |
5751 | cv_wait(&mg->mg_ms_disabled_cv, &mg->mg_ms_disabled_lock); | |
5752 | } | |
5753 | mg->mg_ms_disabled++; | |
5754 | ASSERT3U(mg->mg_ms_disabled, <=, max_disabled_ms); | |
5755 | } | |
5756 | ||
5757 | /* | |
5758 | * Mark the metaslab as disabled to prevent any allocations on this metaslab. | |
5759 | * We must also track how many metaslabs are currently disabled within a | |
5760 | * metaslab group and limit them to prevent allocation failures from | |
5761 | * occurring because all metaslabs are disabled. | |
5762 | */ | |
5763 | void | |
5764 | metaslab_disable(metaslab_t *msp) | |
5765 | { | |
5766 | ASSERT(!MUTEX_HELD(&msp->ms_lock)); | |
5767 | metaslab_group_t *mg = msp->ms_group; | |
5768 | ||
5769 | mutex_enter(&mg->mg_ms_disabled_lock); | |
5770 | ||
5771 | /* | |
5772 | * To keep an accurate count of how many threads have disabled | |
5773 | * a specific metaslab group, we only allow one thread to mark | |
5774 | * the metaslab group at a time. This ensures that the value of | |
5775 | * ms_disabled will be accurate when we decide to mark a metaslab | |
5776 | * group as disabled. To do this we force all other threads | |
5777 | * to wait till the metaslab's mg_disabled_updating flag is no | |
5778 | * longer set. | |
5779 | */ | |
5780 | metaslab_group_disable_wait(mg); | |
5781 | mg->mg_disabled_updating = B_TRUE; | |
5782 | if (msp->ms_disabled == 0) { | |
5783 | metaslab_group_disabled_increment(mg); | |
5784 | } | |
5785 | mutex_enter(&msp->ms_lock); | |
5786 | msp->ms_disabled++; | |
5787 | mutex_exit(&msp->ms_lock); | |
5788 | ||
5789 | mg->mg_disabled_updating = B_FALSE; | |
5790 | cv_broadcast(&mg->mg_ms_disabled_cv); | |
5791 | mutex_exit(&mg->mg_ms_disabled_lock); | |
5792 | } | |
5793 | ||
5794 | void | |
f09fda50 | 5795 | metaslab_enable(metaslab_t *msp, boolean_t sync, boolean_t unload) |
1b939560 BB |
5796 | { |
5797 | metaslab_group_t *mg = msp->ms_group; | |
5798 | spa_t *spa = mg->mg_vd->vdev_spa; | |
5799 | ||
5800 | /* | |
5801 | * Wait for the outstanding IO to be synced to prevent newly | |
5802 | * allocated blocks from being overwritten. This used by | |
5803 | * initialize and TRIM which are modifying unallocated space. | |
5804 | */ | |
5805 | if (sync) | |
5806 | txg_wait_synced(spa_get_dsl(spa), 0); | |
5807 | ||
5808 | mutex_enter(&mg->mg_ms_disabled_lock); | |
5809 | mutex_enter(&msp->ms_lock); | |
5810 | if (--msp->ms_disabled == 0) { | |
5811 | mg->mg_ms_disabled--; | |
5812 | cv_broadcast(&mg->mg_ms_disabled_cv); | |
f09fda50 PD |
5813 | if (unload) |
5814 | metaslab_unload(msp); | |
1b939560 BB |
5815 | } |
5816 | mutex_exit(&msp->ms_lock); | |
5817 | mutex_exit(&mg->mg_ms_disabled_lock); | |
5818 | } | |
5819 | ||
93e28d66 SD |
5820 | static void |
5821 | metaslab_update_ondisk_flush_data(metaslab_t *ms, dmu_tx_t *tx) | |
5822 | { | |
5823 | vdev_t *vd = ms->ms_group->mg_vd; | |
5824 | spa_t *spa = vd->vdev_spa; | |
5825 | objset_t *mos = spa_meta_objset(spa); | |
5826 | ||
5827 | ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)); | |
5828 | ||
5829 | metaslab_unflushed_phys_t entry = { | |
5830 | .msp_unflushed_txg = metaslab_unflushed_txg(ms), | |
5831 | }; | |
5832 | uint64_t entry_size = sizeof (entry); | |
5833 | uint64_t entry_offset = ms->ms_id * entry_size; | |
5834 | ||
5835 | uint64_t object = 0; | |
5836 | int err = zap_lookup(mos, vd->vdev_top_zap, | |
5837 | VDEV_TOP_ZAP_MS_UNFLUSHED_PHYS_TXGS, sizeof (uint64_t), 1, | |
5838 | &object); | |
5839 | if (err == ENOENT) { | |
5840 | object = dmu_object_alloc(mos, DMU_OTN_UINT64_METADATA, | |
5841 | SPA_OLD_MAXBLOCKSIZE, DMU_OT_NONE, 0, tx); | |
5842 | VERIFY0(zap_add(mos, vd->vdev_top_zap, | |
5843 | VDEV_TOP_ZAP_MS_UNFLUSHED_PHYS_TXGS, sizeof (uint64_t), 1, | |
5844 | &object, tx)); | |
5845 | } else { | |
5846 | VERIFY0(err); | |
5847 | } | |
5848 | ||
5849 | dmu_write(spa_meta_objset(spa), object, entry_offset, entry_size, | |
5850 | &entry, tx); | |
5851 | } | |
5852 | ||
5853 | void | |
5854 | metaslab_set_unflushed_txg(metaslab_t *ms, uint64_t txg, dmu_tx_t *tx) | |
5855 | { | |
5856 | spa_t *spa = ms->ms_group->mg_vd->vdev_spa; | |
5857 | ||
5858 | if (!spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) | |
5859 | return; | |
5860 | ||
5861 | ms->ms_unflushed_txg = txg; | |
5862 | metaslab_update_ondisk_flush_data(ms, tx); | |
5863 | } | |
5864 | ||
5865 | uint64_t | |
5866 | metaslab_unflushed_txg(metaslab_t *ms) | |
5867 | { | |
5868 | return (ms->ms_unflushed_txg); | |
5869 | } | |
5870 | ||
03fdcb9a MM |
5871 | ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, aliquot, ULONG, ZMOD_RW, |
5872 | "Allocation granularity (a.k.a. stripe size)"); | |
02730c33 | 5873 | |
03fdcb9a MM |
5874 | ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, debug_load, INT, ZMOD_RW, |
5875 | "Load all metaslabs when pool is first opened"); | |
02730c33 | 5876 | |
03fdcb9a MM |
5877 | ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, debug_unload, INT, ZMOD_RW, |
5878 | "Prevent metaslabs from being unloaded"); | |
f4a4046b | 5879 | |
03fdcb9a MM |
5880 | ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, preload_enabled, INT, ZMOD_RW, |
5881 | "Preload potential metaslabs during reassessment"); | |
eef0f4d8 | 5882 | |
03fdcb9a MM |
5883 | ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, unload_delay, INT, ZMOD_RW, |
5884 | "Delay in txgs after metaslab was last used before unloading"); | |
eef0f4d8 | 5885 | |
03fdcb9a MM |
5886 | ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, unload_delay_ms, INT, ZMOD_RW, |
5887 | "Delay in milliseconds after metaslab was last used before unloading"); | |
02730c33 | 5888 | |
03fdcb9a MM |
5889 | /* BEGIN CSTYLED */ |
5890 | ZFS_MODULE_PARAM(zfs_mg, zfs_mg_, noalloc_threshold, INT, ZMOD_RW, | |
5891 | "Percentage of metaslab group size that should be free to make it " | |
5892 | "eligible for allocation"); | |
f3a7f661 | 5893 | |
03fdcb9a MM |
5894 | ZFS_MODULE_PARAM(zfs_mg, zfs_mg_, fragmentation_threshold, INT, ZMOD_RW, |
5895 | "Percentage of metaslab group size that should be considered eligible " | |
5896 | "for allocations unless all metaslab groups within the metaslab class " | |
5897 | "have also crossed this threshold"); | |
02730c33 | 5898 | |
03fdcb9a MM |
5899 | ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, fragmentation_threshold, INT, |
5900 | ZMOD_RW, "Fragmentation for metaslab to allow allocation"); | |
02730c33 | 5901 | |
03fdcb9a MM |
5902 | ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, fragmentation_factor_enabled, INT, ZMOD_RW, |
5903 | "Use the fragmentation metric to prefer less fragmented metaslabs"); | |
5904 | /* END CSTYLED */ | |
02730c33 | 5905 | |
03fdcb9a MM |
5906 | ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, lba_weighting_enabled, INT, ZMOD_RW, |
5907 | "Prefer metaslabs with lower LBAs"); | |
4e21fd06 | 5908 | |
03fdcb9a MM |
5909 | ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, bias_enabled, INT, ZMOD_RW, |
5910 | "Enable metaslab group biasing"); | |
4e21fd06 | 5911 | |
03fdcb9a MM |
5912 | ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, segment_weight_enabled, INT, |
5913 | ZMOD_RW, "Enable segment-based metaslab selection"); | |
a1d477c2 | 5914 | |
03fdcb9a MM |
5915 | ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, switch_threshold, INT, ZMOD_RW, |
5916 | "Segment-based metaslab selection maximum buckets before switching"); | |
d3230d76 | 5917 | |
03fdcb9a MM |
5918 | ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, force_ganging, ULONG, ZMOD_RW, |
5919 | "Blocks larger than this size are forced to be gang blocks"); | |
d3230d76 | 5920 | |
03fdcb9a MM |
5921 | ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, df_max_search, INT, ZMOD_RW, |
5922 | "Max distance (bytes) to search forward before using size tree"); | |
c81f1790 | 5923 | |
03fdcb9a MM |
5924 | ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, df_use_largest_segment, INT, ZMOD_RW, |
5925 | "When looking in size tree, use largest segment instead of exact fit"); | |
f09fda50 | 5926 | |
03fdcb9a MM |
5927 | ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, max_size_cache_sec, ULONG, |
5928 | ZMOD_RW, "How long to trust the cached max chunk size of a metaslab"); | |
cc99f275 | 5929 | |
03fdcb9a MM |
5930 | ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, mem_limit, INT, ZMOD_RW, |
5931 | "Percentage of memory that can be used to store metaslab range trees"); |