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 /* 22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26 /* 27 * Copyright (c) 2011, 2019 by Delphix. All rights reserved. 28 */ 29 30 #ifndef _SYS_METASLAB_IMPL_H 31 #define _SYS_METASLAB_IMPL_H 32 33 #include <sys/metaslab.h> 34 #include <sys/space_map.h> 35 #include <sys/range_tree.h> 36 #include <sys/vdev.h> 37 #include <sys/txg.h> 38 #include <sys/avl.h> 39 #include <sys/multilist.h> 40 41 #ifdef __cplusplus 42 extern "C" { 43 #endif 44 45 /* 46 * Metaslab allocation tracing record. 47 */ 48 typedef struct metaslab_alloc_trace { 49 list_node_t mat_list_node; 50 metaslab_group_t *mat_mg; 51 metaslab_t *mat_msp; 52 uint64_t mat_size; 53 uint64_t mat_weight; 54 uint32_t mat_dva_id; 55 uint64_t mat_offset; 56 int mat_allocator; 57 } metaslab_alloc_trace_t; 58 59 /* 60 * Used by the metaslab allocation tracing facility to indicate 61 * error conditions. These errors are stored to the offset member 62 * of the metaslab_alloc_trace_t record and displayed by mdb. 63 */ 64 typedef enum trace_alloc_type { 65 TRACE_ALLOC_FAILURE = -1ULL, 66 TRACE_TOO_SMALL = -2ULL, 67 TRACE_FORCE_GANG = -3ULL, 68 TRACE_NOT_ALLOCATABLE = -4ULL, 69 TRACE_GROUP_FAILURE = -5ULL, 70 TRACE_ENOSPC = -6ULL, 71 TRACE_CONDENSING = -7ULL, 72 TRACE_VDEV_ERROR = -8ULL, 73 TRACE_DISABLED = -9ULL, 74 } trace_alloc_type_t; 75 76 #define METASLAB_WEIGHT_PRIMARY (1ULL << 63) 77 #define METASLAB_WEIGHT_SECONDARY (1ULL << 62) 78 #define METASLAB_WEIGHT_CLAIM (1ULL << 61) 79 #define METASLAB_WEIGHT_TYPE (1ULL << 60) 80 #define METASLAB_ACTIVE_MASK \ 81 (METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY | \ 82 METASLAB_WEIGHT_CLAIM) 83 84 /* 85 * The metaslab weight is used to encode the amount of free space in a 86 * metaslab, such that the "best" metaslab appears first when sorting the 87 * metaslabs by weight. The weight (and therefore the "best" metaslab) can 88 * be determined in two different ways: by computing a weighted sum of all 89 * the free space in the metaslab (a space based weight) or by counting only 90 * the free segments of the largest size (a segment based weight). We prefer 91 * the segment based weight because it reflects how the free space is 92 * comprised, but we cannot always use it -- legacy pools do not have the 93 * space map histogram information necessary to determine the largest 94 * contiguous regions. Pools that have the space map histogram determine 95 * the segment weight by looking at each bucket in the histogram and 96 * determining the free space whose size in bytes is in the range: 97 * [2^i, 2^(i+1)) 98 * We then encode the largest index, i, that contains regions into the 99 * segment-weighted value. 100 * 101 * Space-based weight: 102 * 103 * 64 56 48 40 32 24 16 8 0 104 * +-------+-------+-------+-------+-------+-------+-------+-------+ 105 * |PSC1| weighted-free space | 106 * +-------+-------+-------+-------+-------+-------+-------+-------+ 107 * 108 * PS - indicates primary and secondary activation 109 * C - indicates activation for claimed block zio 110 * space - the fragmentation-weighted space 111 * 112 * Segment-based weight: 113 * 114 * 64 56 48 40 32 24 16 8 0 115 * +-------+-------+-------+-------+-------+-------+-------+-------+ 116 * |PSC0| idx| count of segments in region | 117 * +-------+-------+-------+-------+-------+-------+-------+-------+ 118 * 119 * PS - indicates primary and secondary activation 120 * C - indicates activation for claimed block zio 121 * idx - index for the highest bucket in the histogram 122 * count - number of segments in the specified bucket 123 */ 124 #define WEIGHT_GET_ACTIVE(weight) BF64_GET((weight), 61, 3) 125 #define WEIGHT_SET_ACTIVE(weight, x) BF64_SET((weight), 61, 3, x) 126 127 #define WEIGHT_IS_SPACEBASED(weight) \ 128 ((weight) == 0 || BF64_GET((weight), 60, 1)) 129 #define WEIGHT_SET_SPACEBASED(weight) BF64_SET((weight), 60, 1, 1) 130 131 /* 132 * These macros are only applicable to segment-based weighting. 133 */ 134 #define WEIGHT_GET_INDEX(weight) BF64_GET((weight), 54, 6) 135 #define WEIGHT_SET_INDEX(weight, x) BF64_SET((weight), 54, 6, x) 136 #define WEIGHT_GET_COUNT(weight) BF64_GET((weight), 0, 54) 137 #define WEIGHT_SET_COUNT(weight, x) BF64_SET((weight), 0, 54, x) 138 139 /* 140 * Per-allocator data structure. 141 */ 142 typedef struct metaslab_class_allocator { 143 metaslab_group_t *mca_rotor; 144 uint64_t mca_aliquot; 145 146 /* 147 * The allocation throttle works on a reservation system. Whenever 148 * an asynchronous zio wants to perform an allocation it must 149 * first reserve the number of blocks that it wants to allocate. 150 * If there aren't sufficient slots available for the pending zio 151 * then that I/O is throttled until more slots free up. The current 152 * number of reserved allocations is maintained by the mca_alloc_slots 153 * refcount. The mca_alloc_max_slots value determines the maximum 154 * number of allocations that the system allows. Gang blocks are 155 * allowed to reserve slots even if we've reached the maximum 156 * number of allocations allowed. 157 */ 158 uint64_t mca_alloc_max_slots; 159 zfs_refcount_t mca_alloc_slots; 160 } ____cacheline_aligned metaslab_class_allocator_t; 161 162 /* 163 * A metaslab class encompasses a category of allocatable top-level vdevs. 164 * Each top-level vdev is associated with a metaslab group which defines 165 * the allocatable region for that vdev. Examples of these categories include 166 * "normal" for data block allocations (i.e. main pool allocations) or "log" 167 * for allocations designated for intent log devices (i.e. slog devices). 168 * When a block allocation is requested from the SPA it is associated with a 169 * metaslab_class_t, and only top-level vdevs (i.e. metaslab groups) belonging 170 * to the class can be used to satisfy that request. Allocations are done 171 * by traversing the metaslab groups that are linked off of the mca_rotor field. 172 * This rotor points to the next metaslab group where allocations will be 173 * attempted. Allocating a block is a 3 step process -- select the metaslab 174 * group, select the metaslab, and then allocate the block. The metaslab 175 * class defines the low-level block allocator that will be used as the 176 * final step in allocation. These allocators are pluggable allowing each class 177 * to use a block allocator that best suits that class. 178 */ 179 struct metaslab_class { 180 kmutex_t mc_lock; 181 spa_t *mc_spa; 182 metaslab_ops_t *mc_ops; 183 184 /* 185 * Track the number of metaslab groups that have been initialized 186 * and can accept allocations. An initialized metaslab group is 187 * one has been completely added to the config (i.e. we have 188 * updated the MOS config and the space has been added to the pool). 189 */ 190 uint64_t mc_groups; 191 192 /* 193 * Toggle to enable/disable the allocation throttle. 194 */ 195 boolean_t mc_alloc_throttle_enabled; 196 197 uint64_t mc_alloc_groups; /* # of allocatable groups */ 198 199 uint64_t mc_alloc; /* total allocated space */ 200 uint64_t mc_deferred; /* total deferred frees */ 201 uint64_t mc_space; /* total space (alloc + free) */ 202 uint64_t mc_dspace; /* total deflated space */ 203 uint64_t mc_histogram[RANGE_TREE_HISTOGRAM_SIZE]; 204 205 /* 206 * List of all loaded metaslabs in the class, sorted in order of most 207 * recent use. 208 */ 209 multilist_t mc_metaslab_txg_list; 210 211 metaslab_class_allocator_t mc_allocator[]; 212 }; 213 214 /* 215 * Per-allocator data structure. 216 */ 217 typedef struct metaslab_group_allocator { 218 uint64_t mga_cur_max_alloc_queue_depth; 219 zfs_refcount_t mga_alloc_queue_depth; 220 metaslab_t *mga_primary; 221 metaslab_t *mga_secondary; 222 } metaslab_group_allocator_t; 223 224 /* 225 * Metaslab groups encapsulate all the allocatable regions (i.e. metaslabs) 226 * of a top-level vdev. They are linked together to form a circular linked 227 * list and can belong to only one metaslab class. Metaslab groups may become 228 * ineligible for allocations for a number of reasons such as limited free 229 * space, fragmentation, or going offline. When this happens the allocator will 230 * simply find the next metaslab group in the linked list and attempt 231 * to allocate from that group instead. 232 */ 233 struct metaslab_group { 234 kmutex_t mg_lock; 235 avl_tree_t mg_metaslab_tree; 236 uint64_t mg_aliquot; 237 boolean_t mg_allocatable; /* can we allocate? */ 238 uint64_t mg_ms_ready; 239 240 /* 241 * A metaslab group is considered to be initialized only after 242 * we have updated the MOS config and added the space to the pool. 243 * We only allow allocation attempts to a metaslab group if it 244 * has been initialized. 245 */ 246 boolean_t mg_initialized; 247 248 uint64_t mg_free_capacity; /* percentage free */ 249 int64_t mg_bias; 250 int64_t mg_activation_count; 251 metaslab_class_t *mg_class; 252 vdev_t *mg_vd; 253 taskq_t *mg_taskq; 254 metaslab_group_t *mg_prev; 255 metaslab_group_t *mg_next; 256 257 /* 258 * In order for the allocation throttle to function properly, we cannot 259 * have too many IOs going to each disk by default; the throttle 260 * operates by allocating more work to disks that finish quickly, so 261 * allocating larger chunks to each disk reduces its effectiveness. 262 * However, if the number of IOs going to each allocator is too small, 263 * we will not perform proper aggregation at the vdev_queue layer, 264 * also resulting in decreased performance. Therefore, we will use a 265 * ramp-up strategy. 266 * 267 * Each allocator in each metaslab group has a current queue depth 268 * (mg_alloc_queue_depth[allocator]) and a current max queue depth 269 * (mga_cur_max_alloc_queue_depth[allocator]), and each metaslab group 270 * has an absolute max queue depth (mg_max_alloc_queue_depth). We 271 * add IOs to an allocator until the mg_alloc_queue_depth for that 272 * allocator hits the cur_max. Every time an IO completes for a given 273 * allocator on a given metaslab group, we increment its cur_max until 274 * it reaches mg_max_alloc_queue_depth. The cur_max resets every txg to 275 * help protect against disks that decrease in performance over time. 276 * 277 * It's possible for an allocator to handle more allocations than 278 * its max. This can occur when gang blocks are required or when other 279 * groups are unable to handle their share of allocations. 280 */ 281 uint64_t mg_max_alloc_queue_depth; 282 283 /* 284 * A metalab group that can no longer allocate the minimum block 285 * size will set mg_no_free_space. Once a metaslab group is out 286 * of space then its share of work must be distributed to other 287 * groups. 288 */ 289 boolean_t mg_no_free_space; 290 291 uint64_t mg_allocations; 292 uint64_t mg_failed_allocations; 293 uint64_t mg_fragmentation; 294 uint64_t mg_histogram[RANGE_TREE_HISTOGRAM_SIZE]; 295 296 int mg_ms_disabled; 297 boolean_t mg_disabled_updating; 298 kmutex_t mg_ms_disabled_lock; 299 kcondvar_t mg_ms_disabled_cv; 300 301 int mg_allocators; 302 metaslab_group_allocator_t mg_allocator[]; 303 }; 304 305 /* 306 * This value defines the number of elements in the ms_lbas array. The value 307 * of 64 was chosen as it covers all power of 2 buckets up to UINT64_MAX. 308 * This is the equivalent of highbit(UINT64_MAX). 309 */ 310 #define MAX_LBAS 64 311 312 /* 313 * Each metaslab maintains a set of in-core trees to track metaslab 314 * operations. The in-core free tree (ms_allocatable) contains the list of 315 * free segments which are eligible for allocation. As blocks are 316 * allocated, the allocated segment are removed from the ms_allocatable and 317 * added to a per txg allocation tree (ms_allocating). As blocks are 318 * freed, they are added to the free tree (ms_freeing). These trees 319 * allow us to process all allocations and frees in syncing context 320 * where it is safe to update the on-disk space maps. An additional set 321 * of in-core trees is maintained to track deferred frees 322 * (ms_defer). Once a block is freed it will move from the 323 * ms_freed to the ms_defer tree. A deferred free means that a block 324 * has been freed but cannot be used by the pool until TXG_DEFER_SIZE 325 * transactions groups later. For example, a block that is freed in txg 326 * 50 will not be available for reallocation until txg 52 (50 + 327 * TXG_DEFER_SIZE). This provides a safety net for uberblock rollback. 328 * A pool could be safely rolled back TXG_DEFERS_SIZE transactions 329 * groups and ensure that no block has been reallocated. 330 * 331 * The simplified transition diagram looks like this: 332 * 333 * 334 * ALLOCATE 335 * | 336 * V 337 * free segment (ms_allocatable) -> ms_allocating[4] -> (write to space map) 338 * ^ 339 * | ms_freeing <--- FREE 340 * | | 341 * | v 342 * | ms_freed 343 * | | 344 * +-------- ms_defer[2] <-------+-------> (write to space map) 345 * 346 * 347 * Each metaslab's space is tracked in a single space map in the MOS, 348 * which is only updated in syncing context. Each time we sync a txg, 349 * we append the allocs and frees from that txg to the space map. The 350 * pool space is only updated once all metaslabs have finished syncing. 351 * 352 * To load the in-core free tree we read the space map from disk. This 353 * object contains a series of alloc and free records that are combined 354 * to make up the list of all free segments in this metaslab. These 355 * segments are represented in-core by the ms_allocatable and are stored 356 * in an AVL tree. 357 * 358 * As the space map grows (as a result of the appends) it will 359 * eventually become space-inefficient. When the metaslab's in-core 360 * free tree is zfs_condense_pct/100 times the size of the minimal 361 * on-disk representation, we rewrite it in its minimized form. If a 362 * metaslab needs to condense then we must set the ms_condensing flag to 363 * ensure that allocations are not performed on the metaslab that is 364 * being written. 365 */ 366 struct metaslab { 367 /* 368 * This is the main lock of the metaslab and its purpose is to 369 * coordinate our allocations and frees [e.g metaslab_block_alloc(), 370 * metaslab_free_concrete(), ..etc] with our various syncing 371 * procedures [e.g. metaslab_sync(), metaslab_sync_done(), ..etc]. 372 * 373 * The lock is also used during some miscellaneous operations like 374 * using the metaslab's histogram for the metaslab group's histogram 375 * aggregation, or marking the metaslab for initialization. 376 */ 377 kmutex_t ms_lock; 378 379 /* 380 * Acquired together with the ms_lock whenever we expect to 381 * write to metaslab data on-disk (i.e flushing entries to 382 * the metaslab's space map). It helps coordinate readers of 383 * the metaslab's space map [see spa_vdev_remove_thread()] 384 * with writers [see metaslab_sync() or metaslab_flush()]. 385 * 386 * Note that metaslab_load(), even though a reader, uses 387 * a completely different mechanism to deal with the reading 388 * of the metaslab's space map based on ms_synced_length. That 389 * said, the function still uses the ms_sync_lock after it 390 * has read the ms_sm [see relevant comment in metaslab_load() 391 * as to why]. 392 */ 393 kmutex_t ms_sync_lock; 394 395 kcondvar_t ms_load_cv; 396 space_map_t *ms_sm; 397 uint64_t ms_id; 398 uint64_t ms_start; 399 uint64_t ms_size; 400 uint64_t ms_fragmentation; 401 402 range_tree_t *ms_allocating[TXG_SIZE]; 403 range_tree_t *ms_allocatable; 404 uint64_t ms_allocated_this_txg; 405 uint64_t ms_allocating_total; 406 407 /* 408 * The following range trees are accessed only from syncing context. 409 * ms_free*tree only have entries while syncing, and are empty 410 * between syncs. 411 */ 412 range_tree_t *ms_freeing; /* to free this syncing txg */ 413 range_tree_t *ms_freed; /* already freed this syncing txg */ 414 range_tree_t *ms_defer[TXG_DEFER_SIZE]; 415 range_tree_t *ms_checkpointing; /* to add to the checkpoint */ 416 417 /* 418 * The ms_trim tree is the set of allocatable segments which are 419 * eligible for trimming. (When the metaslab is loaded, it's a 420 * subset of ms_allocatable.) It's kept in-core as long as the 421 * autotrim property is set and is not vacated when the metaslab 422 * is unloaded. Its purpose is to aggregate freed ranges to 423 * facilitate efficient trimming. 424 */ 425 range_tree_t *ms_trim; 426 427 boolean_t ms_condensing; /* condensing? */ 428 boolean_t ms_condense_wanted; 429 430 /* 431 * The number of consumers which have disabled the metaslab. 432 */ 433 uint64_t ms_disabled; 434 435 /* 436 * We must always hold the ms_lock when modifying ms_loaded 437 * and ms_loading. 438 */ 439 boolean_t ms_loaded; 440 boolean_t ms_loading; 441 kcondvar_t ms_flush_cv; 442 boolean_t ms_flushing; 443 444 /* 445 * The following histograms count entries that are in the 446 * metaslab's space map (and its histogram) but are not in 447 * ms_allocatable yet, because they are in ms_freed, ms_freeing, 448 * or ms_defer[]. 449 * 450 * When the metaslab is not loaded, its ms_weight needs to 451 * reflect what is allocatable (i.e. what will be part of 452 * ms_allocatable if it is loaded). The weight is computed from 453 * the spacemap histogram, but that includes ranges that are 454 * not yet allocatable (because they are in ms_freed, 455 * ms_freeing, or ms_defer[]). Therefore, when calculating the 456 * weight, we need to remove those ranges. 457 * 458 * The ranges in the ms_freed and ms_defer[] range trees are all 459 * present in the spacemap. However, the spacemap may have 460 * multiple entries to represent a contiguous range, because it 461 * is written across multiple sync passes, but the changes of 462 * all sync passes are consolidated into the range trees. 463 * Adjacent ranges that are freed in different sync passes of 464 * one txg will be represented separately (as 2 or more entries) 465 * in the space map (and its histogram), but these adjacent 466 * ranges will be consolidated (represented as one entry) in the 467 * ms_freed/ms_defer[] range trees (and their histograms). 468 * 469 * When calculating the weight, we can not simply subtract the 470 * range trees' histograms from the spacemap's histogram, 471 * because the range trees' histograms may have entries in 472 * higher buckets than the spacemap, due to consolidation. 473 * Instead we must subtract the exact entries that were added to 474 * the spacemap's histogram. ms_synchist and ms_deferhist[] 475 * represent these exact entries, so we can subtract them from 476 * the spacemap's histogram when calculating ms_weight. 477 * 478 * ms_synchist represents the same ranges as ms_freeing + 479 * ms_freed, but without consolidation across sync passes. 480 * 481 * ms_deferhist[i] represents the same ranges as ms_defer[i], 482 * but without consolidation across sync passes. 483 */ 484 uint64_t ms_synchist[SPACE_MAP_HISTOGRAM_SIZE]; 485 uint64_t ms_deferhist[TXG_DEFER_SIZE][SPACE_MAP_HISTOGRAM_SIZE]; 486 487 /* 488 * Tracks the exact amount of allocated space of this metaslab 489 * (and specifically the metaslab's space map) up to the most 490 * recently completed sync pass [see usage in metaslab_sync()]. 491 */ 492 uint64_t ms_allocated_space; 493 int64_t ms_deferspace; /* sum of ms_defermap[] space */ 494 uint64_t ms_weight; /* weight vs. others in group */ 495 uint64_t ms_activation_weight; /* activation weight */ 496 497 /* 498 * Track of whenever a metaslab is selected for loading or allocation. 499 * We use this value to determine how long the metaslab should 500 * stay cached. 501 */ 502 uint64_t ms_selected_txg; 503 /* 504 * ms_load/unload_time can be used for performance monitoring 505 * (e.g. by dtrace or mdb). 506 */ 507 hrtime_t ms_load_time; /* time last loaded */ 508 hrtime_t ms_unload_time; /* time last unloaded */ 509 hrtime_t ms_selected_time; /* time last allocated from */ 510 511 uint64_t ms_alloc_txg; /* last successful alloc (debug only) */ 512 uint64_t ms_max_size; /* maximum allocatable size */ 513 514 /* 515 * -1 if it's not active in an allocator, otherwise set to the allocator 516 * this metaslab is active for. 517 */ 518 int ms_allocator; 519 boolean_t ms_primary; /* Only valid if ms_allocator is not -1 */ 520 521 /* 522 * The metaslab block allocators can optionally use a size-ordered 523 * range tree and/or an array of LBAs. Not all allocators use 524 * this functionality. The ms_allocatable_by_size should always 525 * contain the same number of segments as the ms_allocatable. The 526 * only difference is that the ms_allocatable_by_size is ordered by 527 * segment sizes. 528 */ 529 zfs_btree_t ms_allocatable_by_size; 530 zfs_btree_t ms_unflushed_frees_by_size; 531 uint64_t ms_lbas[MAX_LBAS]; 532 533 metaslab_group_t *ms_group; /* metaslab group */ 534 avl_node_t ms_group_node; /* node in metaslab group tree */ 535 txg_node_t ms_txg_node; /* per-txg dirty metaslab links */ 536 avl_node_t ms_spa_txg_node; /* node in spa_metaslabs_by_txg */ 537 /* 538 * Node in metaslab class's selected txg list 539 */ 540 multilist_node_t ms_class_txg_node; 541 542 /* 543 * Allocs and frees that are committed to the vdev log spacemap but 544 * not yet to this metaslab's spacemap. 545 */ 546 range_tree_t *ms_unflushed_allocs; 547 range_tree_t *ms_unflushed_frees; 548 549 /* 550 * We have flushed entries up to but not including this TXG. In 551 * other words, all changes from this TXG and onward should not 552 * be in this metaslab's space map and must be read from the 553 * log space maps. 554 */ 555 uint64_t ms_unflushed_txg; 556 557 /* updated every time we are done syncing the metaslab's space map */ 558 uint64_t ms_synced_length; 559 560 boolean_t ms_new; 561 }; 562 563 typedef struct metaslab_unflushed_phys { 564 /* on-disk counterpart of ms_unflushed_txg */ 565 uint64_t msp_unflushed_txg; 566 } metaslab_unflushed_phys_t; 567 568 #ifdef __cplusplus 569 } 570 #endif 571 572 #endif /* _SYS_METASLAB_IMPL_H */ 573