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 (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 24 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 25 * Copyright 2014 Nexenta Systems, Inc. All rights reserved. 26 */ 27 28/* 29 * DVA-based Adjustable Replacement Cache 30 * 31 * While much of the theory of operation used here is 32 * based on the self-tuning, low overhead replacement cache 33 * presented by Megiddo and Modha at FAST 2003, there are some 34 * significant differences: 35 * 36 * 1. The Megiddo and Modha model assumes any page is evictable. 37 * Pages in its cache cannot be "locked" into memory. This makes 38 * the eviction algorithm simple: evict the last page in the list. 39 * This also make the performance characteristics easy to reason 40 * about. Our cache is not so simple. At any given moment, some 41 * subset of the blocks in the cache are un-evictable because we 42 * have handed out a reference to them. Blocks are only evictable 43 * when there are no external references active. This makes 44 * eviction far more problematic: we choose to evict the evictable 45 * blocks that are the "lowest" in the list. 46 * 47 * There are times when it is not possible to evict the requested 48 * space. In these circumstances we are unable to adjust the cache 49 * size. To prevent the cache growing unbounded at these times we 50 * implement a "cache throttle" that slows the flow of new data 51 * into the cache until we can make space available. 52 * 53 * 2. The Megiddo and Modha model assumes a fixed cache size. 54 * Pages are evicted when the cache is full and there is a cache 55 * miss. Our model has a variable sized cache. It grows with 56 * high use, but also tries to react to memory pressure from the 57 * operating system: decreasing its size when system memory is 58 * tight. 59 * 60 * 3. The Megiddo and Modha model assumes a fixed page size. All 61 * elements of the cache are therefore exactly the same size. So 62 * when adjusting the cache size following a cache miss, its simply 63 * a matter of choosing a single page to evict. In our model, we 64 * have variable sized cache blocks (rangeing from 512 bytes to 65 * 128K bytes). We therefore choose a set of blocks to evict to make 66 * space for a cache miss that approximates as closely as possible 67 * the space used by the new block. 68 * 69 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 70 * by N. Megiddo & D. Modha, FAST 2003 71 */ 72 73/* 74 * The locking model: 75 * 76 * A new reference to a cache buffer can be obtained in two 77 * ways: 1) via a hash table lookup using the DVA as a key, 78 * or 2) via one of the ARC lists. The arc_read() interface 79 * uses method 1, while the internal arc algorithms for 80 * adjusting the cache use method 2. We therefore provide two 81 * types of locks: 1) the hash table lock array, and 2) the 82 * arc list locks. 83 * 84 * Buffers do not have their own mutexs, rather they rely on the 85 * hash table mutexs for the bulk of their protection (i.e. most 86 * fields in the arc_buf_hdr_t are protected by these mutexs). 87 * 88 * buf_hash_find() returns the appropriate mutex (held) when it 89 * locates the requested buffer in the hash table. It returns 90 * NULL for the mutex if the buffer was not in the table. 91 * 92 * buf_hash_remove() expects the appropriate hash mutex to be 93 * already held before it is invoked. 94 * 95 * Each arc state also has a mutex which is used to protect the 96 * buffer list associated with the state. When attempting to 97 * obtain a hash table lock while holding an arc list lock you 98 * must use: mutex_tryenter() to avoid deadlock. Also note that 99 * the active state mutex must be held before the ghost state mutex. 100 * 101 * Arc buffers may have an associated eviction callback function. 102 * This function will be invoked prior to removing the buffer (e.g. 103 * in arc_do_user_evicts()). Note however that the data associated 104 * with the buffer may be evicted prior to the callback. The callback 105 * must be made with *no locks held* (to prevent deadlock). Additionally, 106 * the users of callbacks must ensure that their private data is 107 * protected from simultaneous callbacks from arc_clear_callback() 108 * and arc_do_user_evicts(). 109 * 110 * Note that the majority of the performance stats are manipulated 111 * with atomic operations. 112 * 113 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following: 114 * 115 * - L2ARC buflist creation 116 * - L2ARC buflist eviction 117 * - L2ARC write completion, which walks L2ARC buflists 118 * - ARC header destruction, as it removes from L2ARC buflists 119 * - ARC header release, as it removes from L2ARC buflists 120 */ 121 122#include <sys/spa.h> 123#include <sys/zio.h> 124#include <sys/zio_compress.h> 125#include <sys/zfs_context.h> 126#include <sys/arc.h> 127#include <sys/refcount.h> 128#include <sys/vdev.h> 129#include <sys/vdev_impl.h> 130#include <sys/dsl_pool.h> 131#ifdef _KERNEL 132#include <sys/dnlc.h> 133#endif 134#include <sys/callb.h> 135#include <sys/kstat.h> 136#include <sys/trim_map.h> 137#include <zfs_fletcher.h> 138#include <sys/sdt.h> 139 140#include <vm/vm_pageout.h> 141 142#ifdef illumos 143#ifndef _KERNEL 144/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 145boolean_t arc_watch = B_FALSE; 146int arc_procfd; 147#endif 148#endif /* illumos */ 149 150static kmutex_t arc_reclaim_thr_lock; 151static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 152static uint8_t arc_thread_exit; 153 154#define ARC_REDUCE_DNLC_PERCENT 3 155uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 156 157typedef enum arc_reclaim_strategy { 158 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 159 ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 160} arc_reclaim_strategy_t; 161 162/* 163 * The number of iterations through arc_evict_*() before we 164 * drop & reacquire the lock. 165 */ 166int arc_evict_iterations = 100; 167 168/* number of seconds before growing cache again */ 169static int arc_grow_retry = 60; 170 171/* shift of arc_c for calculating both min and max arc_p */ 172static int arc_p_min_shift = 4; 173 174/* log2(fraction of arc to reclaim) */ 175static int arc_shrink_shift = 5; 176 177/* 178 * minimum lifespan of a prefetch block in clock ticks 179 * (initialized in arc_init()) 180 */ 181static int arc_min_prefetch_lifespan; 182 183/* 184 * If this percent of memory is free, don't throttle. 185 */ 186int arc_lotsfree_percent = 10; 187 188static int arc_dead; 189extern int zfs_prefetch_disable; 190 191/* 192 * The arc has filled available memory and has now warmed up. 193 */ 194static boolean_t arc_warm; 195 196/* 197 * These tunables are for performance analysis. 198 */ 199uint64_t zfs_arc_max; 200uint64_t zfs_arc_min; 201uint64_t zfs_arc_meta_limit = 0; 202int zfs_arc_grow_retry = 0; 203int zfs_arc_shrink_shift = 0; 204int zfs_arc_p_min_shift = 0; 205int zfs_disable_dup_eviction = 0; 206uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 207 208TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max); 209TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 210TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 211TUNABLE_QUAD("vfs.zfs.arc_average_blocksize", &zfs_arc_average_blocksize); 212SYSCTL_DECL(_vfs_zfs); 213SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 214 "Maximum ARC size"); 215SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 216 "Minimum ARC size"); 217SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 218 &zfs_arc_average_blocksize, 0, 219 "ARC average blocksize"); 220 221/* 222 * Note that buffers can be in one of 6 states: 223 * ARC_anon - anonymous (discussed below) 224 * ARC_mru - recently used, currently cached 225 * ARC_mru_ghost - recentely used, no longer in cache 226 * ARC_mfu - frequently used, currently cached 227 * ARC_mfu_ghost - frequently used, no longer in cache 228 * ARC_l2c_only - exists in L2ARC but not other states 229 * When there are no active references to the buffer, they are 230 * are linked onto a list in one of these arc states. These are 231 * the only buffers that can be evicted or deleted. Within each 232 * state there are multiple lists, one for meta-data and one for 233 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 234 * etc.) is tracked separately so that it can be managed more 235 * explicitly: favored over data, limited explicitly. 236 * 237 * Anonymous buffers are buffers that are not associated with 238 * a DVA. These are buffers that hold dirty block copies 239 * before they are written to stable storage. By definition, 240 * they are "ref'd" and are considered part of arc_mru 241 * that cannot be freed. Generally, they will aquire a DVA 242 * as they are written and migrate onto the arc_mru list. 243 * 244 * The ARC_l2c_only state is for buffers that are in the second 245 * level ARC but no longer in any of the ARC_m* lists. The second 246 * level ARC itself may also contain buffers that are in any of 247 * the ARC_m* states - meaning that a buffer can exist in two 248 * places. The reason for the ARC_l2c_only state is to keep the 249 * buffer header in the hash table, so that reads that hit the 250 * second level ARC benefit from these fast lookups. 251 */ 252 253#define ARCS_LOCK_PAD CACHE_LINE_SIZE 254struct arcs_lock { 255 kmutex_t arcs_lock; 256#ifdef _KERNEL 257 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 258#endif 259}; 260 261/* 262 * must be power of two for mask use to work 263 * 264 */ 265#define ARC_BUFC_NUMDATALISTS 16 266#define ARC_BUFC_NUMMETADATALISTS 16 267#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 268 269typedef struct arc_state { 270 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 271 uint64_t arcs_size; /* total amount of data in this state */ 272 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 273 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 274} arc_state_t; 275 276#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 277 278/* The 6 states: */ 279static arc_state_t ARC_anon; 280static arc_state_t ARC_mru; 281static arc_state_t ARC_mru_ghost; 282static arc_state_t ARC_mfu; 283static arc_state_t ARC_mfu_ghost; 284static arc_state_t ARC_l2c_only; 285 286typedef struct arc_stats { 287 kstat_named_t arcstat_hits; 288 kstat_named_t arcstat_misses; 289 kstat_named_t arcstat_demand_data_hits; 290 kstat_named_t arcstat_demand_data_misses; 291 kstat_named_t arcstat_demand_metadata_hits; 292 kstat_named_t arcstat_demand_metadata_misses; 293 kstat_named_t arcstat_prefetch_data_hits; 294 kstat_named_t arcstat_prefetch_data_misses; 295 kstat_named_t arcstat_prefetch_metadata_hits; 296 kstat_named_t arcstat_prefetch_metadata_misses; 297 kstat_named_t arcstat_mru_hits; 298 kstat_named_t arcstat_mru_ghost_hits; 299 kstat_named_t arcstat_mfu_hits; 300 kstat_named_t arcstat_mfu_ghost_hits; 301 kstat_named_t arcstat_allocated; 302 kstat_named_t arcstat_deleted; 303 kstat_named_t arcstat_stolen; 304 kstat_named_t arcstat_recycle_miss; 305 /* 306 * Number of buffers that could not be evicted because the hash lock 307 * was held by another thread. The lock may not necessarily be held 308 * by something using the same buffer, since hash locks are shared 309 * by multiple buffers. 310 */ 311 kstat_named_t arcstat_mutex_miss; 312 /* 313 * Number of buffers skipped because they have I/O in progress, are 314 * indrect prefetch buffers that have not lived long enough, or are 315 * not from the spa we're trying to evict from. 316 */ 317 kstat_named_t arcstat_evict_skip; 318 kstat_named_t arcstat_evict_l2_cached; 319 kstat_named_t arcstat_evict_l2_eligible; 320 kstat_named_t arcstat_evict_l2_ineligible; 321 kstat_named_t arcstat_hash_elements; 322 kstat_named_t arcstat_hash_elements_max; 323 kstat_named_t arcstat_hash_collisions; 324 kstat_named_t arcstat_hash_chains; 325 kstat_named_t arcstat_hash_chain_max; 326 kstat_named_t arcstat_p; 327 kstat_named_t arcstat_c; 328 kstat_named_t arcstat_c_min; 329 kstat_named_t arcstat_c_max; 330 kstat_named_t arcstat_size; 331 kstat_named_t arcstat_hdr_size; 332 kstat_named_t arcstat_data_size; 333 kstat_named_t arcstat_other_size; 334 kstat_named_t arcstat_l2_hits; 335 kstat_named_t arcstat_l2_misses; 336 kstat_named_t arcstat_l2_feeds; 337 kstat_named_t arcstat_l2_rw_clash; 338 kstat_named_t arcstat_l2_read_bytes; 339 kstat_named_t arcstat_l2_write_bytes; 340 kstat_named_t arcstat_l2_writes_sent; 341 kstat_named_t arcstat_l2_writes_done; 342 kstat_named_t arcstat_l2_writes_error; 343 kstat_named_t arcstat_l2_writes_hdr_miss; 344 kstat_named_t arcstat_l2_evict_lock_retry; 345 kstat_named_t arcstat_l2_evict_reading; 346 kstat_named_t arcstat_l2_free_on_write; 347 kstat_named_t arcstat_l2_cdata_free_on_write; 348 kstat_named_t arcstat_l2_abort_lowmem; 349 kstat_named_t arcstat_l2_cksum_bad; 350 kstat_named_t arcstat_l2_io_error; 351 kstat_named_t arcstat_l2_size; 352 kstat_named_t arcstat_l2_asize; 353 kstat_named_t arcstat_l2_hdr_size; 354 kstat_named_t arcstat_l2_compress_successes; 355 kstat_named_t arcstat_l2_compress_zeros; 356 kstat_named_t arcstat_l2_compress_failures; 357 kstat_named_t arcstat_l2_write_trylock_fail; 358 kstat_named_t arcstat_l2_write_passed_headroom; 359 kstat_named_t arcstat_l2_write_spa_mismatch; 360 kstat_named_t arcstat_l2_write_in_l2; 361 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 362 kstat_named_t arcstat_l2_write_not_cacheable; 363 kstat_named_t arcstat_l2_write_full; 364 kstat_named_t arcstat_l2_write_buffer_iter; 365 kstat_named_t arcstat_l2_write_pios; 366 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 367 kstat_named_t arcstat_l2_write_buffer_list_iter; 368 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 369 kstat_named_t arcstat_memory_throttle_count; 370 kstat_named_t arcstat_duplicate_buffers; 371 kstat_named_t arcstat_duplicate_buffers_size; 372 kstat_named_t arcstat_duplicate_reads; 373} arc_stats_t; 374 375static arc_stats_t arc_stats = { 376 { "hits", KSTAT_DATA_UINT64 }, 377 { "misses", KSTAT_DATA_UINT64 }, 378 { "demand_data_hits", KSTAT_DATA_UINT64 }, 379 { "demand_data_misses", KSTAT_DATA_UINT64 }, 380 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 381 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 382 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 383 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 384 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 385 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 386 { "mru_hits", KSTAT_DATA_UINT64 }, 387 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 388 { "mfu_hits", KSTAT_DATA_UINT64 }, 389 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 390 { "allocated", KSTAT_DATA_UINT64 }, 391 { "deleted", KSTAT_DATA_UINT64 }, 392 { "stolen", KSTAT_DATA_UINT64 }, 393 { "recycle_miss", KSTAT_DATA_UINT64 }, 394 { "mutex_miss", KSTAT_DATA_UINT64 }, 395 { "evict_skip", KSTAT_DATA_UINT64 }, 396 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 397 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 398 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 399 { "hash_elements", KSTAT_DATA_UINT64 }, 400 { "hash_elements_max", KSTAT_DATA_UINT64 }, 401 { "hash_collisions", KSTAT_DATA_UINT64 }, 402 { "hash_chains", KSTAT_DATA_UINT64 }, 403 { "hash_chain_max", KSTAT_DATA_UINT64 }, 404 { "p", KSTAT_DATA_UINT64 }, 405 { "c", KSTAT_DATA_UINT64 }, 406 { "c_min", KSTAT_DATA_UINT64 }, 407 { "c_max", KSTAT_DATA_UINT64 }, 408 { "size", KSTAT_DATA_UINT64 }, 409 { "hdr_size", KSTAT_DATA_UINT64 }, 410 { "data_size", KSTAT_DATA_UINT64 }, 411 { "other_size", KSTAT_DATA_UINT64 }, 412 { "l2_hits", KSTAT_DATA_UINT64 }, 413 { "l2_misses", KSTAT_DATA_UINT64 }, 414 { "l2_feeds", KSTAT_DATA_UINT64 }, 415 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 416 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 417 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 418 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 419 { "l2_writes_done", KSTAT_DATA_UINT64 }, 420 { "l2_writes_error", KSTAT_DATA_UINT64 }, 421 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 422 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 423 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 424 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 425 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, 426 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 427 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 428 { "l2_io_error", KSTAT_DATA_UINT64 }, 429 { "l2_size", KSTAT_DATA_UINT64 }, 430 { "l2_asize", KSTAT_DATA_UINT64 }, 431 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 432 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 433 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 434 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 435 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 436 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 437 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 438 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 439 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 440 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 441 { "l2_write_full", KSTAT_DATA_UINT64 }, 442 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 443 { "l2_write_pios", KSTAT_DATA_UINT64 }, 444 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 445 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 446 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 447 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 448 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 449 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 450 { "duplicate_reads", KSTAT_DATA_UINT64 } 451}; 452 453#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 454 455#define ARCSTAT_INCR(stat, val) \ 456 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 457 458#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 459#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 460 461#define ARCSTAT_MAX(stat, val) { \ 462 uint64_t m; \ 463 while ((val) > (m = arc_stats.stat.value.ui64) && \ 464 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 465 continue; \ 466} 467 468#define ARCSTAT_MAXSTAT(stat) \ 469 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 470 471/* 472 * We define a macro to allow ARC hits/misses to be easily broken down by 473 * two separate conditions, giving a total of four different subtypes for 474 * each of hits and misses (so eight statistics total). 475 */ 476#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 477 if (cond1) { \ 478 if (cond2) { \ 479 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 480 } else { \ 481 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 482 } \ 483 } else { \ 484 if (cond2) { \ 485 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 486 } else { \ 487 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 488 } \ 489 } 490 491kstat_t *arc_ksp; 492static arc_state_t *arc_anon; 493static arc_state_t *arc_mru; 494static arc_state_t *arc_mru_ghost; 495static arc_state_t *arc_mfu; 496static arc_state_t *arc_mfu_ghost; 497static arc_state_t *arc_l2c_only; 498 499/* 500 * There are several ARC variables that are critical to export as kstats -- 501 * but we don't want to have to grovel around in the kstat whenever we wish to 502 * manipulate them. For these variables, we therefore define them to be in 503 * terms of the statistic variable. This assures that we are not introducing 504 * the possibility of inconsistency by having shadow copies of the variables, 505 * while still allowing the code to be readable. 506 */ 507#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 508#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 509#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 510#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 511#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 512 513#define L2ARC_IS_VALID_COMPRESS(_c_) \ 514 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 515 516static int arc_no_grow; /* Don't try to grow cache size */ 517static uint64_t arc_tempreserve; 518static uint64_t arc_loaned_bytes; 519static uint64_t arc_meta_used; 520static uint64_t arc_meta_limit; 521static uint64_t arc_meta_max = 0; 522SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RD, &arc_meta_used, 0, 523 "ARC metadata used"); 524SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RW, &arc_meta_limit, 0, 525 "ARC metadata limit"); 526 527typedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 528 529typedef struct arc_callback arc_callback_t; 530 531struct arc_callback { 532 void *acb_private; 533 arc_done_func_t *acb_done; 534 arc_buf_t *acb_buf; 535 zio_t *acb_zio_dummy; 536 arc_callback_t *acb_next; 537}; 538 539typedef struct arc_write_callback arc_write_callback_t; 540 541struct arc_write_callback { 542 void *awcb_private; 543 arc_done_func_t *awcb_ready; 544 arc_done_func_t *awcb_physdone; 545 arc_done_func_t *awcb_done; 546 arc_buf_t *awcb_buf; 547}; 548 549struct arc_buf_hdr { 550 /* protected by hash lock */ 551 dva_t b_dva; 552 uint64_t b_birth; 553 uint64_t b_cksum0; 554 555 kmutex_t b_freeze_lock; 556 zio_cksum_t *b_freeze_cksum; 557 void *b_thawed; 558 559 arc_buf_hdr_t *b_hash_next; 560 arc_buf_t *b_buf; 561 uint32_t b_flags; 562 uint32_t b_datacnt; 563 564 arc_callback_t *b_acb; 565 kcondvar_t b_cv; 566 567 /* immutable */ 568 arc_buf_contents_t b_type; 569 uint64_t b_size; 570 uint64_t b_spa; 571 572 /* protected by arc state mutex */ 573 arc_state_t *b_state; 574 list_node_t b_arc_node; 575 576 /* updated atomically */ 577 clock_t b_arc_access; 578 579 /* self protecting */ 580 refcount_t b_refcnt; 581 582 l2arc_buf_hdr_t *b_l2hdr; 583 list_node_t b_l2node; 584}; 585 586static arc_buf_t *arc_eviction_list; 587static kmutex_t arc_eviction_mtx; 588static arc_buf_hdr_t arc_eviction_hdr; 589static void arc_get_data_buf(arc_buf_t *buf); 590static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock); 591static int arc_evict_needed(arc_buf_contents_t type); 592static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes); 593#ifdef illumos 594static void arc_buf_watch(arc_buf_t *buf); 595#endif /* illumos */ 596 597static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab); 598 599#define GHOST_STATE(state) \ 600 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 601 (state) == arc_l2c_only) 602 603/* 604 * Private ARC flags. These flags are private ARC only flags that will show up 605 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can 606 * be passed in as arc_flags in things like arc_read. However, these flags 607 * should never be passed and should only be set by ARC code. When adding new 608 * public flags, make sure not to smash the private ones. 609 */ 610 611#define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */ 612#define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */ 613#define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */ 614#define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */ 615#define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */ 616#define ARC_INDIRECT (1 << 14) /* this is an indirect block */ 617#define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */ 618#define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */ 619#define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */ 620#define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */ 621 622#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE) 623#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS) 624#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR) 625#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH) 626#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ) 627#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE) 628#define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS) 629#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE) 630#define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \ 631 (hdr)->b_l2hdr != NULL) 632#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING) 633#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED) 634#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD) 635 636/* 637 * Other sizes 638 */ 639 640#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 641#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 642 643/* 644 * Hash table routines 645 */ 646 647#define HT_LOCK_PAD CACHE_LINE_SIZE 648 649struct ht_lock { 650 kmutex_t ht_lock; 651#ifdef _KERNEL 652 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 653#endif 654}; 655 656#define BUF_LOCKS 256 657typedef struct buf_hash_table { 658 uint64_t ht_mask; 659 arc_buf_hdr_t **ht_table; 660 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 661} buf_hash_table_t; 662 663static buf_hash_table_t buf_hash_table; 664 665#define BUF_HASH_INDEX(spa, dva, birth) \ 666 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 667#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 668#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 669#define HDR_LOCK(hdr) \ 670 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 671 672uint64_t zfs_crc64_table[256]; 673 674/* 675 * Level 2 ARC 676 */ 677 678#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 679#define L2ARC_HEADROOM 2 /* num of writes */ 680/* 681 * If we discover during ARC scan any buffers to be compressed, we boost 682 * our headroom for the next scanning cycle by this percentage multiple. 683 */ 684#define L2ARC_HEADROOM_BOOST 200 685#define L2ARC_FEED_SECS 1 /* caching interval secs */ 686#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 687 688#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 689#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 690 691/* L2ARC Performance Tunables */ 692uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 693uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 694uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 695uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 696uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 697uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 698boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 699boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 700boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 701 702SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 703 &l2arc_write_max, 0, "max write size"); 704SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 705 &l2arc_write_boost, 0, "extra write during warmup"); 706SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 707 &l2arc_headroom, 0, "number of dev writes"); 708SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 709 &l2arc_feed_secs, 0, "interval seconds"); 710SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 711 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 712 713SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 714 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 715SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 716 &l2arc_feed_again, 0, "turbo warmup"); 717SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 718 &l2arc_norw, 0, "no reads during writes"); 719 720SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 721 &ARC_anon.arcs_size, 0, "size of anonymous state"); 722SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 723 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 724SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 725 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 726 727SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 728 &ARC_mru.arcs_size, 0, "size of mru state"); 729SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 730 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 731SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 732 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 733 734SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 735 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 736SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 737 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 738 "size of metadata in mru ghost state"); 739SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 740 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 741 "size of data in mru ghost state"); 742 743SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 744 &ARC_mfu.arcs_size, 0, "size of mfu state"); 745SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 746 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 747SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 748 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 749 750SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 751 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 752SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 753 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 754 "size of metadata in mfu ghost state"); 755SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 756 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 757 "size of data in mfu ghost state"); 758 759SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 760 &ARC_l2c_only.arcs_size, 0, "size of mru state"); 761 762/* 763 * L2ARC Internals 764 */ 765typedef struct l2arc_dev { 766 vdev_t *l2ad_vdev; /* vdev */ 767 spa_t *l2ad_spa; /* spa */ 768 uint64_t l2ad_hand; /* next write location */ 769 uint64_t l2ad_start; /* first addr on device */ 770 uint64_t l2ad_end; /* last addr on device */ 771 uint64_t l2ad_evict; /* last addr eviction reached */ 772 boolean_t l2ad_first; /* first sweep through */ 773 boolean_t l2ad_writing; /* currently writing */ 774 list_t *l2ad_buflist; /* buffer list */ 775 list_node_t l2ad_node; /* device list node */ 776} l2arc_dev_t; 777 778static list_t L2ARC_dev_list; /* device list */ 779static list_t *l2arc_dev_list; /* device list pointer */ 780static kmutex_t l2arc_dev_mtx; /* device list mutex */ 781static l2arc_dev_t *l2arc_dev_last; /* last device used */ 782static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 783static list_t L2ARC_free_on_write; /* free after write buf list */ 784static list_t *l2arc_free_on_write; /* free after write list ptr */ 785static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 786static uint64_t l2arc_ndev; /* number of devices */ 787 788typedef struct l2arc_read_callback { 789 arc_buf_t *l2rcb_buf; /* read buffer */ 790 spa_t *l2rcb_spa; /* spa */ 791 blkptr_t l2rcb_bp; /* original blkptr */ 792 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 793 int l2rcb_flags; /* original flags */ 794 enum zio_compress l2rcb_compress; /* applied compress */ 795} l2arc_read_callback_t; 796 797typedef struct l2arc_write_callback { 798 l2arc_dev_t *l2wcb_dev; /* device info */ 799 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 800} l2arc_write_callback_t; 801 802struct l2arc_buf_hdr { 803 /* protected by arc_buf_hdr mutex */ 804 l2arc_dev_t *b_dev; /* L2ARC device */ 805 uint64_t b_daddr; /* disk address, offset byte */ 806 /* compression applied to buffer data */ 807 enum zio_compress b_compress; 808 /* real alloc'd buffer size depending on b_compress applied */ 809 int b_asize; 810 /* temporary buffer holder for in-flight compressed data */ 811 void *b_tmp_cdata; 812}; 813 814typedef struct l2arc_data_free { 815 /* protected by l2arc_free_on_write_mtx */ 816 void *l2df_data; 817 size_t l2df_size; 818 void (*l2df_func)(void *, size_t); 819 list_node_t l2df_list_node; 820} l2arc_data_free_t; 821 822static kmutex_t l2arc_feed_thr_lock; 823static kcondvar_t l2arc_feed_thr_cv; 824static uint8_t l2arc_thread_exit; 825 826static void l2arc_read_done(zio_t *zio); 827static void l2arc_hdr_stat_add(void); 828static void l2arc_hdr_stat_remove(void); 829 830static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr); 831static void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, 832 enum zio_compress c); 833static void l2arc_release_cdata_buf(arc_buf_hdr_t *ab); 834 835static uint64_t 836buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 837{ 838 uint8_t *vdva = (uint8_t *)dva; 839 uint64_t crc = -1ULL; 840 int i; 841 842 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 843 844 for (i = 0; i < sizeof (dva_t); i++) 845 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 846 847 crc ^= (spa>>8) ^ birth; 848 849 return (crc); 850} 851 852#define BUF_EMPTY(buf) \ 853 ((buf)->b_dva.dva_word[0] == 0 && \ 854 (buf)->b_dva.dva_word[1] == 0 && \ 855 (buf)->b_cksum0 == 0) 856 857#define BUF_EQUAL(spa, dva, birth, buf) \ 858 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 859 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 860 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 861 862static void 863buf_discard_identity(arc_buf_hdr_t *hdr) 864{ 865 hdr->b_dva.dva_word[0] = 0; 866 hdr->b_dva.dva_word[1] = 0; 867 hdr->b_birth = 0; 868 hdr->b_cksum0 = 0; 869} 870 871static arc_buf_hdr_t * 872buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 873{ 874 const dva_t *dva = BP_IDENTITY(bp); 875 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 876 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 877 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 878 arc_buf_hdr_t *buf; 879 880 mutex_enter(hash_lock); 881 for (buf = buf_hash_table.ht_table[idx]; buf != NULL; 882 buf = buf->b_hash_next) { 883 if (BUF_EQUAL(spa, dva, birth, buf)) { 884 *lockp = hash_lock; 885 return (buf); 886 } 887 } 888 mutex_exit(hash_lock); 889 *lockp = NULL; 890 return (NULL); 891} 892 893/* 894 * Insert an entry into the hash table. If there is already an element 895 * equal to elem in the hash table, then the already existing element 896 * will be returned and the new element will not be inserted. 897 * Otherwise returns NULL. 898 */ 899static arc_buf_hdr_t * 900buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp) 901{ 902 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 903 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 904 arc_buf_hdr_t *fbuf; 905 uint32_t i; 906 907 ASSERT(!DVA_IS_EMPTY(&buf->b_dva)); 908 ASSERT(buf->b_birth != 0); 909 ASSERT(!HDR_IN_HASH_TABLE(buf)); 910 *lockp = hash_lock; 911 mutex_enter(hash_lock); 912 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL; 913 fbuf = fbuf->b_hash_next, i++) { 914 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf)) 915 return (fbuf); 916 } 917 918 buf->b_hash_next = buf_hash_table.ht_table[idx]; 919 buf_hash_table.ht_table[idx] = buf; 920 buf->b_flags |= ARC_IN_HASH_TABLE; 921 922 /* collect some hash table performance data */ 923 if (i > 0) { 924 ARCSTAT_BUMP(arcstat_hash_collisions); 925 if (i == 1) 926 ARCSTAT_BUMP(arcstat_hash_chains); 927 928 ARCSTAT_MAX(arcstat_hash_chain_max, i); 929 } 930 931 ARCSTAT_BUMP(arcstat_hash_elements); 932 ARCSTAT_MAXSTAT(arcstat_hash_elements); 933 934 return (NULL); 935} 936 937static void 938buf_hash_remove(arc_buf_hdr_t *buf) 939{ 940 arc_buf_hdr_t *fbuf, **bufp; 941 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 942 943 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 944 ASSERT(HDR_IN_HASH_TABLE(buf)); 945 946 bufp = &buf_hash_table.ht_table[idx]; 947 while ((fbuf = *bufp) != buf) { 948 ASSERT(fbuf != NULL); 949 bufp = &fbuf->b_hash_next; 950 } 951 *bufp = buf->b_hash_next; 952 buf->b_hash_next = NULL; 953 buf->b_flags &= ~ARC_IN_HASH_TABLE; 954 955 /* collect some hash table performance data */ 956 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 957 958 if (buf_hash_table.ht_table[idx] && 959 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 960 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 961} 962 963/* 964 * Global data structures and functions for the buf kmem cache. 965 */ 966static kmem_cache_t *hdr_cache; 967static kmem_cache_t *buf_cache; 968 969static void 970buf_fini(void) 971{ 972 int i; 973 974 kmem_free(buf_hash_table.ht_table, 975 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 976 for (i = 0; i < BUF_LOCKS; i++) 977 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 978 kmem_cache_destroy(hdr_cache); 979 kmem_cache_destroy(buf_cache); 980} 981 982/* 983 * Constructor callback - called when the cache is empty 984 * and a new buf is requested. 985 */ 986/* ARGSUSED */ 987static int 988hdr_cons(void *vbuf, void *unused, int kmflag) 989{ 990 arc_buf_hdr_t *buf = vbuf; 991 992 bzero(buf, sizeof (arc_buf_hdr_t)); 993 refcount_create(&buf->b_refcnt); 994 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL); 995 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 996 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 997 998 return (0); 999} 1000 1001/* ARGSUSED */ 1002static int 1003buf_cons(void *vbuf, void *unused, int kmflag) 1004{ 1005 arc_buf_t *buf = vbuf; 1006 1007 bzero(buf, sizeof (arc_buf_t)); 1008 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1009 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1010 1011 return (0); 1012} 1013 1014/* 1015 * Destructor callback - called when a cached buf is 1016 * no longer required. 1017 */ 1018/* ARGSUSED */ 1019static void 1020hdr_dest(void *vbuf, void *unused) 1021{ 1022 arc_buf_hdr_t *buf = vbuf; 1023 1024 ASSERT(BUF_EMPTY(buf)); 1025 refcount_destroy(&buf->b_refcnt); 1026 cv_destroy(&buf->b_cv); 1027 mutex_destroy(&buf->b_freeze_lock); 1028 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1029} 1030 1031/* ARGSUSED */ 1032static void 1033buf_dest(void *vbuf, void *unused) 1034{ 1035 arc_buf_t *buf = vbuf; 1036 1037 mutex_destroy(&buf->b_evict_lock); 1038 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1039} 1040 1041/* 1042 * Reclaim callback -- invoked when memory is low. 1043 */ 1044/* ARGSUSED */ 1045static void 1046hdr_recl(void *unused) 1047{ 1048 dprintf("hdr_recl called\n"); 1049 /* 1050 * umem calls the reclaim func when we destroy the buf cache, 1051 * which is after we do arc_fini(). 1052 */ 1053 if (!arc_dead) 1054 cv_signal(&arc_reclaim_thr_cv); 1055} 1056 1057static void 1058buf_init(void) 1059{ 1060 uint64_t *ct; 1061 uint64_t hsize = 1ULL << 12; 1062 int i, j; 1063 1064 /* 1065 * The hash table is big enough to fill all of physical memory 1066 * with an average block size of zfs_arc_average_blocksize (default 8K). 1067 * By default, the table will take up 1068 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1069 */ 1070 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1071 hsize <<= 1; 1072retry: 1073 buf_hash_table.ht_mask = hsize - 1; 1074 buf_hash_table.ht_table = 1075 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1076 if (buf_hash_table.ht_table == NULL) { 1077 ASSERT(hsize > (1ULL << 8)); 1078 hsize >>= 1; 1079 goto retry; 1080 } 1081 1082 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1083 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1084 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1085 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1086 1087 for (i = 0; i < 256; i++) 1088 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1089 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1090 1091 for (i = 0; i < BUF_LOCKS; i++) { 1092 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1093 NULL, MUTEX_DEFAULT, NULL); 1094 } 1095} 1096 1097#define ARC_MINTIME (hz>>4) /* 62 ms */ 1098 1099static void 1100arc_cksum_verify(arc_buf_t *buf) 1101{ 1102 zio_cksum_t zc; 1103 1104 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1105 return; 1106 1107 mutex_enter(&buf->b_hdr->b_freeze_lock); 1108 if (buf->b_hdr->b_freeze_cksum == NULL || 1109 (buf->b_hdr->b_flags & ARC_IO_ERROR)) { 1110 mutex_exit(&buf->b_hdr->b_freeze_lock); 1111 return; 1112 } 1113 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1114 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1115 panic("buffer modified while frozen!"); 1116 mutex_exit(&buf->b_hdr->b_freeze_lock); 1117} 1118 1119static int 1120arc_cksum_equal(arc_buf_t *buf) 1121{ 1122 zio_cksum_t zc; 1123 int equal; 1124 1125 mutex_enter(&buf->b_hdr->b_freeze_lock); 1126 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1127 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1128 mutex_exit(&buf->b_hdr->b_freeze_lock); 1129 1130 return (equal); 1131} 1132 1133static void 1134arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1135{ 1136 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1137 return; 1138 1139 mutex_enter(&buf->b_hdr->b_freeze_lock); 1140 if (buf->b_hdr->b_freeze_cksum != NULL) { 1141 mutex_exit(&buf->b_hdr->b_freeze_lock); 1142 return; 1143 } 1144 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1145 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1146 buf->b_hdr->b_freeze_cksum); 1147 mutex_exit(&buf->b_hdr->b_freeze_lock); 1148#ifdef illumos 1149 arc_buf_watch(buf); 1150#endif /* illumos */ 1151} 1152 1153#ifdef illumos 1154#ifndef _KERNEL 1155typedef struct procctl { 1156 long cmd; 1157 prwatch_t prwatch; 1158} procctl_t; 1159#endif 1160 1161/* ARGSUSED */ 1162static void 1163arc_buf_unwatch(arc_buf_t *buf) 1164{ 1165#ifndef _KERNEL 1166 if (arc_watch) { 1167 int result; 1168 procctl_t ctl; 1169 ctl.cmd = PCWATCH; 1170 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1171 ctl.prwatch.pr_size = 0; 1172 ctl.prwatch.pr_wflags = 0; 1173 result = write(arc_procfd, &ctl, sizeof (ctl)); 1174 ASSERT3U(result, ==, sizeof (ctl)); 1175 } 1176#endif 1177} 1178 1179/* ARGSUSED */ 1180static void 1181arc_buf_watch(arc_buf_t *buf) 1182{ 1183#ifndef _KERNEL 1184 if (arc_watch) { 1185 int result; 1186 procctl_t ctl; 1187 ctl.cmd = PCWATCH; 1188 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1189 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1190 ctl.prwatch.pr_wflags = WA_WRITE; 1191 result = write(arc_procfd, &ctl, sizeof (ctl)); 1192 ASSERT3U(result, ==, sizeof (ctl)); 1193 } 1194#endif 1195} 1196#endif /* illumos */ 1197 1198void 1199arc_buf_thaw(arc_buf_t *buf) 1200{ 1201 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1202 if (buf->b_hdr->b_state != arc_anon) 1203 panic("modifying non-anon buffer!"); 1204 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS) 1205 panic("modifying buffer while i/o in progress!"); 1206 arc_cksum_verify(buf); 1207 } 1208 1209 mutex_enter(&buf->b_hdr->b_freeze_lock); 1210 if (buf->b_hdr->b_freeze_cksum != NULL) { 1211 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1212 buf->b_hdr->b_freeze_cksum = NULL; 1213 } 1214 1215 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1216 if (buf->b_hdr->b_thawed) 1217 kmem_free(buf->b_hdr->b_thawed, 1); 1218 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1219 } 1220 1221 mutex_exit(&buf->b_hdr->b_freeze_lock); 1222 1223#ifdef illumos 1224 arc_buf_unwatch(buf); 1225#endif /* illumos */ 1226} 1227 1228void 1229arc_buf_freeze(arc_buf_t *buf) 1230{ 1231 kmutex_t *hash_lock; 1232 1233 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1234 return; 1235 1236 hash_lock = HDR_LOCK(buf->b_hdr); 1237 mutex_enter(hash_lock); 1238 1239 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1240 buf->b_hdr->b_state == arc_anon); 1241 arc_cksum_compute(buf, B_FALSE); 1242 mutex_exit(hash_lock); 1243 1244} 1245 1246static void 1247get_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock) 1248{ 1249 uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth); 1250 1251 if (ab->b_type == ARC_BUFC_METADATA) 1252 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1253 else { 1254 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1255 buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1256 } 1257 1258 *list = &state->arcs_lists[buf_hashid]; 1259 *lock = ARCS_LOCK(state, buf_hashid); 1260} 1261 1262 1263static void 1264add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1265{ 1266 ASSERT(MUTEX_HELD(hash_lock)); 1267 1268 if ((refcount_add(&ab->b_refcnt, tag) == 1) && 1269 (ab->b_state != arc_anon)) { 1270 uint64_t delta = ab->b_size * ab->b_datacnt; 1271 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type]; 1272 list_t *list; 1273 kmutex_t *lock; 1274 1275 get_buf_info(ab, ab->b_state, &list, &lock); 1276 ASSERT(!MUTEX_HELD(lock)); 1277 mutex_enter(lock); 1278 ASSERT(list_link_active(&ab->b_arc_node)); 1279 list_remove(list, ab); 1280 if (GHOST_STATE(ab->b_state)) { 1281 ASSERT0(ab->b_datacnt); 1282 ASSERT3P(ab->b_buf, ==, NULL); 1283 delta = ab->b_size; 1284 } 1285 ASSERT(delta > 0); 1286 ASSERT3U(*size, >=, delta); 1287 atomic_add_64(size, -delta); 1288 mutex_exit(lock); 1289 /* remove the prefetch flag if we get a reference */ 1290 if (ab->b_flags & ARC_PREFETCH) 1291 ab->b_flags &= ~ARC_PREFETCH; 1292 } 1293} 1294 1295static int 1296remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1297{ 1298 int cnt; 1299 arc_state_t *state = ab->b_state; 1300 1301 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1302 ASSERT(!GHOST_STATE(state)); 1303 1304 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) && 1305 (state != arc_anon)) { 1306 uint64_t *size = &state->arcs_lsize[ab->b_type]; 1307 list_t *list; 1308 kmutex_t *lock; 1309 1310 get_buf_info(ab, state, &list, &lock); 1311 ASSERT(!MUTEX_HELD(lock)); 1312 mutex_enter(lock); 1313 ASSERT(!list_link_active(&ab->b_arc_node)); 1314 list_insert_head(list, ab); 1315 ASSERT(ab->b_datacnt > 0); 1316 atomic_add_64(size, ab->b_size * ab->b_datacnt); 1317 mutex_exit(lock); 1318 } 1319 return (cnt); 1320} 1321 1322/* 1323 * Move the supplied buffer to the indicated state. The mutex 1324 * for the buffer must be held by the caller. 1325 */ 1326static void 1327arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock) 1328{ 1329 arc_state_t *old_state = ab->b_state; 1330 int64_t refcnt = refcount_count(&ab->b_refcnt); 1331 uint64_t from_delta, to_delta; 1332 list_t *list; 1333 kmutex_t *lock; 1334 1335 ASSERT(MUTEX_HELD(hash_lock)); 1336 ASSERT3P(new_state, !=, old_state); 1337 ASSERT(refcnt == 0 || ab->b_datacnt > 0); 1338 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state)); 1339 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon); 1340 1341 from_delta = to_delta = ab->b_datacnt * ab->b_size; 1342 1343 /* 1344 * If this buffer is evictable, transfer it from the 1345 * old state list to the new state list. 1346 */ 1347 if (refcnt == 0) { 1348 if (old_state != arc_anon) { 1349 int use_mutex; 1350 uint64_t *size = &old_state->arcs_lsize[ab->b_type]; 1351 1352 get_buf_info(ab, old_state, &list, &lock); 1353 use_mutex = !MUTEX_HELD(lock); 1354 if (use_mutex) 1355 mutex_enter(lock); 1356 1357 ASSERT(list_link_active(&ab->b_arc_node)); 1358 list_remove(list, ab); 1359 1360 /* 1361 * If prefetching out of the ghost cache, 1362 * we will have a non-zero datacnt. 1363 */ 1364 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) { 1365 /* ghost elements have a ghost size */ 1366 ASSERT(ab->b_buf == NULL); 1367 from_delta = ab->b_size; 1368 } 1369 ASSERT3U(*size, >=, from_delta); 1370 atomic_add_64(size, -from_delta); 1371 1372 if (use_mutex) 1373 mutex_exit(lock); 1374 } 1375 if (new_state != arc_anon) { 1376 int use_mutex; 1377 uint64_t *size = &new_state->arcs_lsize[ab->b_type]; 1378 1379 get_buf_info(ab, new_state, &list, &lock); 1380 use_mutex = !MUTEX_HELD(lock); 1381 if (use_mutex) 1382 mutex_enter(lock); 1383 1384 list_insert_head(list, ab); 1385 1386 /* ghost elements have a ghost size */ 1387 if (GHOST_STATE(new_state)) { 1388 ASSERT(ab->b_datacnt == 0); 1389 ASSERT(ab->b_buf == NULL); 1390 to_delta = ab->b_size; 1391 } 1392 atomic_add_64(size, to_delta); 1393 1394 if (use_mutex) 1395 mutex_exit(lock); 1396 } 1397 } 1398 1399 ASSERT(!BUF_EMPTY(ab)); 1400 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab)) 1401 buf_hash_remove(ab); 1402 1403 /* adjust state sizes */ 1404 if (to_delta) 1405 atomic_add_64(&new_state->arcs_size, to_delta); 1406 if (from_delta) { 1407 ASSERT3U(old_state->arcs_size, >=, from_delta); 1408 atomic_add_64(&old_state->arcs_size, -from_delta); 1409 } 1410 ab->b_state = new_state; 1411 1412 /* adjust l2arc hdr stats */ 1413 if (new_state == arc_l2c_only) 1414 l2arc_hdr_stat_add(); 1415 else if (old_state == arc_l2c_only) 1416 l2arc_hdr_stat_remove(); 1417} 1418 1419void 1420arc_space_consume(uint64_t space, arc_space_type_t type) 1421{ 1422 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1423 1424 switch (type) { 1425 case ARC_SPACE_DATA: 1426 ARCSTAT_INCR(arcstat_data_size, space); 1427 break; 1428 case ARC_SPACE_OTHER: 1429 ARCSTAT_INCR(arcstat_other_size, space); 1430 break; 1431 case ARC_SPACE_HDRS: 1432 ARCSTAT_INCR(arcstat_hdr_size, space); 1433 break; 1434 case ARC_SPACE_L2HDRS: 1435 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1436 break; 1437 } 1438 1439 atomic_add_64(&arc_meta_used, space); 1440 atomic_add_64(&arc_size, space); 1441} 1442 1443void 1444arc_space_return(uint64_t space, arc_space_type_t type) 1445{ 1446 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1447 1448 switch (type) { 1449 case ARC_SPACE_DATA: 1450 ARCSTAT_INCR(arcstat_data_size, -space); 1451 break; 1452 case ARC_SPACE_OTHER: 1453 ARCSTAT_INCR(arcstat_other_size, -space); 1454 break; 1455 case ARC_SPACE_HDRS: 1456 ARCSTAT_INCR(arcstat_hdr_size, -space); 1457 break; 1458 case ARC_SPACE_L2HDRS: 1459 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1460 break; 1461 } 1462 1463 ASSERT(arc_meta_used >= space); 1464 if (arc_meta_max < arc_meta_used) 1465 arc_meta_max = arc_meta_used; 1466 atomic_add_64(&arc_meta_used, -space); 1467 ASSERT(arc_size >= space); 1468 atomic_add_64(&arc_size, -space); 1469} 1470 1471void * 1472arc_data_buf_alloc(uint64_t size) 1473{ 1474 if (arc_evict_needed(ARC_BUFC_DATA)) 1475 cv_signal(&arc_reclaim_thr_cv); 1476 atomic_add_64(&arc_size, size); 1477 return (zio_data_buf_alloc(size)); 1478} 1479 1480void 1481arc_data_buf_free(void *buf, uint64_t size) 1482{ 1483 zio_data_buf_free(buf, size); 1484 ASSERT(arc_size >= size); 1485 atomic_add_64(&arc_size, -size); 1486} 1487 1488arc_buf_t * 1489arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1490{ 1491 arc_buf_hdr_t *hdr; 1492 arc_buf_t *buf; 1493 1494 ASSERT3U(size, >, 0); 1495 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1496 ASSERT(BUF_EMPTY(hdr)); 1497 hdr->b_size = size; 1498 hdr->b_type = type; 1499 hdr->b_spa = spa_load_guid(spa); 1500 hdr->b_state = arc_anon; 1501 hdr->b_arc_access = 0; 1502 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1503 buf->b_hdr = hdr; 1504 buf->b_data = NULL; 1505 buf->b_efunc = NULL; 1506 buf->b_private = NULL; 1507 buf->b_next = NULL; 1508 hdr->b_buf = buf; 1509 arc_get_data_buf(buf); 1510 hdr->b_datacnt = 1; 1511 hdr->b_flags = 0; 1512 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1513 (void) refcount_add(&hdr->b_refcnt, tag); 1514 1515 return (buf); 1516} 1517 1518static char *arc_onloan_tag = "onloan"; 1519 1520/* 1521 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1522 * flight data by arc_tempreserve_space() until they are "returned". Loaned 1523 * buffers must be returned to the arc before they can be used by the DMU or 1524 * freed. 1525 */ 1526arc_buf_t * 1527arc_loan_buf(spa_t *spa, int size) 1528{ 1529 arc_buf_t *buf; 1530 1531 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1532 1533 atomic_add_64(&arc_loaned_bytes, size); 1534 return (buf); 1535} 1536 1537/* 1538 * Return a loaned arc buffer to the arc. 1539 */ 1540void 1541arc_return_buf(arc_buf_t *buf, void *tag) 1542{ 1543 arc_buf_hdr_t *hdr = buf->b_hdr; 1544 1545 ASSERT(buf->b_data != NULL); 1546 (void) refcount_add(&hdr->b_refcnt, tag); 1547 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1548 1549 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1550} 1551 1552/* Detach an arc_buf from a dbuf (tag) */ 1553void 1554arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1555{ 1556 arc_buf_hdr_t *hdr; 1557 1558 ASSERT(buf->b_data != NULL); 1559 hdr = buf->b_hdr; 1560 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1561 (void) refcount_remove(&hdr->b_refcnt, tag); 1562 buf->b_efunc = NULL; 1563 buf->b_private = NULL; 1564 1565 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1566} 1567 1568static arc_buf_t * 1569arc_buf_clone(arc_buf_t *from) 1570{ 1571 arc_buf_t *buf; 1572 arc_buf_hdr_t *hdr = from->b_hdr; 1573 uint64_t size = hdr->b_size; 1574 1575 ASSERT(hdr->b_state != arc_anon); 1576 1577 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1578 buf->b_hdr = hdr; 1579 buf->b_data = NULL; 1580 buf->b_efunc = NULL; 1581 buf->b_private = NULL; 1582 buf->b_next = hdr->b_buf; 1583 hdr->b_buf = buf; 1584 arc_get_data_buf(buf); 1585 bcopy(from->b_data, buf->b_data, size); 1586 1587 /* 1588 * This buffer already exists in the arc so create a duplicate 1589 * copy for the caller. If the buffer is associated with user data 1590 * then track the size and number of duplicates. These stats will be 1591 * updated as duplicate buffers are created and destroyed. 1592 */ 1593 if (hdr->b_type == ARC_BUFC_DATA) { 1594 ARCSTAT_BUMP(arcstat_duplicate_buffers); 1595 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1596 } 1597 hdr->b_datacnt += 1; 1598 return (buf); 1599} 1600 1601void 1602arc_buf_add_ref(arc_buf_t *buf, void* tag) 1603{ 1604 arc_buf_hdr_t *hdr; 1605 kmutex_t *hash_lock; 1606 1607 /* 1608 * Check to see if this buffer is evicted. Callers 1609 * must verify b_data != NULL to know if the add_ref 1610 * was successful. 1611 */ 1612 mutex_enter(&buf->b_evict_lock); 1613 if (buf->b_data == NULL) { 1614 mutex_exit(&buf->b_evict_lock); 1615 return; 1616 } 1617 hash_lock = HDR_LOCK(buf->b_hdr); 1618 mutex_enter(hash_lock); 1619 hdr = buf->b_hdr; 1620 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1621 mutex_exit(&buf->b_evict_lock); 1622 1623 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1624 add_reference(hdr, hash_lock, tag); 1625 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1626 arc_access(hdr, hash_lock); 1627 mutex_exit(hash_lock); 1628 ARCSTAT_BUMP(arcstat_hits); 1629 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 1630 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1631 data, metadata, hits); 1632} 1633 1634static void 1635arc_buf_free_on_write(void *data, size_t size, 1636 void (*free_func)(void *, size_t)) 1637{ 1638 l2arc_data_free_t *df; 1639 1640 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1641 df->l2df_data = data; 1642 df->l2df_size = size; 1643 df->l2df_func = free_func; 1644 mutex_enter(&l2arc_free_on_write_mtx); 1645 list_insert_head(l2arc_free_on_write, df); 1646 mutex_exit(&l2arc_free_on_write_mtx); 1647} 1648 1649/* 1650 * Free the arc data buffer. If it is an l2arc write in progress, 1651 * the buffer is placed on l2arc_free_on_write to be freed later. 1652 */ 1653static void 1654arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1655{ 1656 arc_buf_hdr_t *hdr = buf->b_hdr; 1657 1658 if (HDR_L2_WRITING(hdr)) { 1659 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 1660 ARCSTAT_BUMP(arcstat_l2_free_on_write); 1661 } else { 1662 free_func(buf->b_data, hdr->b_size); 1663 } 1664} 1665 1666/* 1667 * Free up buf->b_data and if 'remove' is set, then pull the 1668 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 1669 */ 1670static void 1671arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 1672{ 1673 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1674 1675 ASSERT(MUTEX_HELD(&l2arc_buflist_mtx)); 1676 1677 if (l2hdr->b_tmp_cdata == NULL) 1678 return; 1679 1680 ASSERT(HDR_L2_WRITING(hdr)); 1681 arc_buf_free_on_write(l2hdr->b_tmp_cdata, hdr->b_size, 1682 zio_data_buf_free); 1683 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 1684 l2hdr->b_tmp_cdata = NULL; 1685} 1686 1687static void 1688arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 1689{ 1690 arc_buf_t **bufp; 1691 1692 /* free up data associated with the buf */ 1693 if (buf->b_data) { 1694 arc_state_t *state = buf->b_hdr->b_state; 1695 uint64_t size = buf->b_hdr->b_size; 1696 arc_buf_contents_t type = buf->b_hdr->b_type; 1697 1698 arc_cksum_verify(buf); 1699#ifdef illumos 1700 arc_buf_unwatch(buf); 1701#endif /* illumos */ 1702 1703 if (!recycle) { 1704 if (type == ARC_BUFC_METADATA) { 1705 arc_buf_data_free(buf, zio_buf_free); 1706 arc_space_return(size, ARC_SPACE_DATA); 1707 } else { 1708 ASSERT(type == ARC_BUFC_DATA); 1709 arc_buf_data_free(buf, zio_data_buf_free); 1710 ARCSTAT_INCR(arcstat_data_size, -size); 1711 atomic_add_64(&arc_size, -size); 1712 } 1713 } 1714 if (list_link_active(&buf->b_hdr->b_arc_node)) { 1715 uint64_t *cnt = &state->arcs_lsize[type]; 1716 1717 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1718 ASSERT(state != arc_anon); 1719 1720 ASSERT3U(*cnt, >=, size); 1721 atomic_add_64(cnt, -size); 1722 } 1723 ASSERT3U(state->arcs_size, >=, size); 1724 atomic_add_64(&state->arcs_size, -size); 1725 buf->b_data = NULL; 1726 1727 /* 1728 * If we're destroying a duplicate buffer make sure 1729 * that the appropriate statistics are updated. 1730 */ 1731 if (buf->b_hdr->b_datacnt > 1 && 1732 buf->b_hdr->b_type == ARC_BUFC_DATA) { 1733 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1734 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1735 } 1736 ASSERT(buf->b_hdr->b_datacnt > 0); 1737 buf->b_hdr->b_datacnt -= 1; 1738 } 1739 1740 /* only remove the buf if requested */ 1741 if (!remove) 1742 return; 1743 1744 /* remove the buf from the hdr list */ 1745 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1746 continue; 1747 *bufp = buf->b_next; 1748 buf->b_next = NULL; 1749 1750 ASSERT(buf->b_efunc == NULL); 1751 1752 /* clean up the buf */ 1753 buf->b_hdr = NULL; 1754 kmem_cache_free(buf_cache, buf); 1755} 1756 1757static void 1758arc_hdr_destroy(arc_buf_hdr_t *hdr) 1759{ 1760 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1761 ASSERT3P(hdr->b_state, ==, arc_anon); 1762 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1763 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1764 1765 if (l2hdr != NULL) { 1766 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1767 /* 1768 * To prevent arc_free() and l2arc_evict() from 1769 * attempting to free the same buffer at the same time, 1770 * a FREE_IN_PROGRESS flag is given to arc_free() to 1771 * give it priority. l2arc_evict() can't destroy this 1772 * header while we are waiting on l2arc_buflist_mtx. 1773 * 1774 * The hdr may be removed from l2ad_buflist before we 1775 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1776 */ 1777 if (!buflist_held) { 1778 mutex_enter(&l2arc_buflist_mtx); 1779 l2hdr = hdr->b_l2hdr; 1780 } 1781 1782 if (l2hdr != NULL) { 1783 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1784 hdr->b_size, 0); 1785 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1786 arc_buf_l2_cdata_free(hdr); 1787 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1788 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1789 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 1790 -l2hdr->b_asize, 0, 0); 1791 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1792 if (hdr->b_state == arc_l2c_only) 1793 l2arc_hdr_stat_remove(); 1794 hdr->b_l2hdr = NULL; 1795 } 1796 1797 if (!buflist_held) 1798 mutex_exit(&l2arc_buflist_mtx); 1799 } 1800 1801 if (!BUF_EMPTY(hdr)) { 1802 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1803 buf_discard_identity(hdr); 1804 } 1805 while (hdr->b_buf) { 1806 arc_buf_t *buf = hdr->b_buf; 1807 1808 if (buf->b_efunc) { 1809 mutex_enter(&arc_eviction_mtx); 1810 mutex_enter(&buf->b_evict_lock); 1811 ASSERT(buf->b_hdr != NULL); 1812 arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1813 hdr->b_buf = buf->b_next; 1814 buf->b_hdr = &arc_eviction_hdr; 1815 buf->b_next = arc_eviction_list; 1816 arc_eviction_list = buf; 1817 mutex_exit(&buf->b_evict_lock); 1818 mutex_exit(&arc_eviction_mtx); 1819 } else { 1820 arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1821 } 1822 } 1823 if (hdr->b_freeze_cksum != NULL) { 1824 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1825 hdr->b_freeze_cksum = NULL; 1826 } 1827 if (hdr->b_thawed) { 1828 kmem_free(hdr->b_thawed, 1); 1829 hdr->b_thawed = NULL; 1830 } 1831 1832 ASSERT(!list_link_active(&hdr->b_arc_node)); 1833 ASSERT3P(hdr->b_hash_next, ==, NULL); 1834 ASSERT3P(hdr->b_acb, ==, NULL); 1835 kmem_cache_free(hdr_cache, hdr); 1836} 1837 1838void 1839arc_buf_free(arc_buf_t *buf, void *tag) 1840{ 1841 arc_buf_hdr_t *hdr = buf->b_hdr; 1842 int hashed = hdr->b_state != arc_anon; 1843 1844 ASSERT(buf->b_efunc == NULL); 1845 ASSERT(buf->b_data != NULL); 1846 1847 if (hashed) { 1848 kmutex_t *hash_lock = HDR_LOCK(hdr); 1849 1850 mutex_enter(hash_lock); 1851 hdr = buf->b_hdr; 1852 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1853 1854 (void) remove_reference(hdr, hash_lock, tag); 1855 if (hdr->b_datacnt > 1) { 1856 arc_buf_destroy(buf, FALSE, TRUE); 1857 } else { 1858 ASSERT(buf == hdr->b_buf); 1859 ASSERT(buf->b_efunc == NULL); 1860 hdr->b_flags |= ARC_BUF_AVAILABLE; 1861 } 1862 mutex_exit(hash_lock); 1863 } else if (HDR_IO_IN_PROGRESS(hdr)) { 1864 int destroy_hdr; 1865 /* 1866 * We are in the middle of an async write. Don't destroy 1867 * this buffer unless the write completes before we finish 1868 * decrementing the reference count. 1869 */ 1870 mutex_enter(&arc_eviction_mtx); 1871 (void) remove_reference(hdr, NULL, tag); 1872 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1873 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1874 mutex_exit(&arc_eviction_mtx); 1875 if (destroy_hdr) 1876 arc_hdr_destroy(hdr); 1877 } else { 1878 if (remove_reference(hdr, NULL, tag) > 0) 1879 arc_buf_destroy(buf, FALSE, TRUE); 1880 else 1881 arc_hdr_destroy(hdr); 1882 } 1883} 1884 1885boolean_t 1886arc_buf_remove_ref(arc_buf_t *buf, void* tag) 1887{ 1888 arc_buf_hdr_t *hdr = buf->b_hdr; 1889 kmutex_t *hash_lock = HDR_LOCK(hdr); 1890 boolean_t no_callback = (buf->b_efunc == NULL); 1891 1892 if (hdr->b_state == arc_anon) { 1893 ASSERT(hdr->b_datacnt == 1); 1894 arc_buf_free(buf, tag); 1895 return (no_callback); 1896 } 1897 1898 mutex_enter(hash_lock); 1899 hdr = buf->b_hdr; 1900 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1901 ASSERT(hdr->b_state != arc_anon); 1902 ASSERT(buf->b_data != NULL); 1903 1904 (void) remove_reference(hdr, hash_lock, tag); 1905 if (hdr->b_datacnt > 1) { 1906 if (no_callback) 1907 arc_buf_destroy(buf, FALSE, TRUE); 1908 } else if (no_callback) { 1909 ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1910 ASSERT(buf->b_efunc == NULL); 1911 hdr->b_flags |= ARC_BUF_AVAILABLE; 1912 } 1913 ASSERT(no_callback || hdr->b_datacnt > 1 || 1914 refcount_is_zero(&hdr->b_refcnt)); 1915 mutex_exit(hash_lock); 1916 return (no_callback); 1917} 1918 1919int 1920arc_buf_size(arc_buf_t *buf) 1921{ 1922 return (buf->b_hdr->b_size); 1923} 1924 1925/* 1926 * Called from the DMU to determine if the current buffer should be 1927 * evicted. In order to ensure proper locking, the eviction must be initiated 1928 * from the DMU. Return true if the buffer is associated with user data and 1929 * duplicate buffers still exist. 1930 */ 1931boolean_t 1932arc_buf_eviction_needed(arc_buf_t *buf) 1933{ 1934 arc_buf_hdr_t *hdr; 1935 boolean_t evict_needed = B_FALSE; 1936 1937 if (zfs_disable_dup_eviction) 1938 return (B_FALSE); 1939 1940 mutex_enter(&buf->b_evict_lock); 1941 hdr = buf->b_hdr; 1942 if (hdr == NULL) { 1943 /* 1944 * We are in arc_do_user_evicts(); let that function 1945 * perform the eviction. 1946 */ 1947 ASSERT(buf->b_data == NULL); 1948 mutex_exit(&buf->b_evict_lock); 1949 return (B_FALSE); 1950 } else if (buf->b_data == NULL) { 1951 /* 1952 * We have already been added to the arc eviction list; 1953 * recommend eviction. 1954 */ 1955 ASSERT3P(hdr, ==, &arc_eviction_hdr); 1956 mutex_exit(&buf->b_evict_lock); 1957 return (B_TRUE); 1958 } 1959 1960 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 1961 evict_needed = B_TRUE; 1962 1963 mutex_exit(&buf->b_evict_lock); 1964 return (evict_needed); 1965} 1966 1967/* 1968 * Evict buffers from list until we've removed the specified number of 1969 * bytes. Move the removed buffers to the appropriate evict state. 1970 * If the recycle flag is set, then attempt to "recycle" a buffer: 1971 * - look for a buffer to evict that is `bytes' long. 1972 * - return the data block from this buffer rather than freeing it. 1973 * This flag is used by callers that are trying to make space for a 1974 * new buffer in a full arc cache. 1975 * 1976 * This function makes a "best effort". It skips over any buffers 1977 * it can't get a hash_lock on, and so may not catch all candidates. 1978 * It may also return without evicting as much space as requested. 1979 */ 1980static void * 1981arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 1982 arc_buf_contents_t type) 1983{ 1984 arc_state_t *evicted_state; 1985 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 1986 int64_t bytes_remaining; 1987 arc_buf_hdr_t *ab, *ab_prev = NULL; 1988 list_t *evicted_list, *list, *evicted_list_start, *list_start; 1989 kmutex_t *lock, *evicted_lock; 1990 kmutex_t *hash_lock; 1991 boolean_t have_lock; 1992 void *stolen = NULL; 1993 arc_buf_hdr_t marker = { 0 }; 1994 int count = 0; 1995 static int evict_metadata_offset, evict_data_offset; 1996 int i, idx, offset, list_count, lists; 1997 1998 ASSERT(state == arc_mru || state == arc_mfu); 1999 2000 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2001 2002 if (type == ARC_BUFC_METADATA) { 2003 offset = 0; 2004 list_count = ARC_BUFC_NUMMETADATALISTS; 2005 list_start = &state->arcs_lists[0]; 2006 evicted_list_start = &evicted_state->arcs_lists[0]; 2007 idx = evict_metadata_offset; 2008 } else { 2009 offset = ARC_BUFC_NUMMETADATALISTS; 2010 list_start = &state->arcs_lists[offset]; 2011 evicted_list_start = &evicted_state->arcs_lists[offset]; 2012 list_count = ARC_BUFC_NUMDATALISTS; 2013 idx = evict_data_offset; 2014 } 2015 bytes_remaining = evicted_state->arcs_lsize[type]; 2016 lists = 0; 2017 2018evict_start: 2019 list = &list_start[idx]; 2020 evicted_list = &evicted_list_start[idx]; 2021 lock = ARCS_LOCK(state, (offset + idx)); 2022 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 2023 2024 mutex_enter(lock); 2025 mutex_enter(evicted_lock); 2026 2027 for (ab = list_tail(list); ab; ab = ab_prev) { 2028 ab_prev = list_prev(list, ab); 2029 bytes_remaining -= (ab->b_size * ab->b_datacnt); 2030 /* prefetch buffers have a minimum lifespan */ 2031 if (HDR_IO_IN_PROGRESS(ab) || 2032 (spa && ab->b_spa != spa) || 2033 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 2034 ddi_get_lbolt() - ab->b_arc_access < 2035 arc_min_prefetch_lifespan)) { 2036 skipped++; 2037 continue; 2038 } 2039 /* "lookahead" for better eviction candidate */ 2040 if (recycle && ab->b_size != bytes && 2041 ab_prev && ab_prev->b_size == bytes) 2042 continue; 2043 2044 /* ignore markers */ 2045 if (ab->b_spa == 0) 2046 continue; 2047 2048 /* 2049 * It may take a long time to evict all the bufs requested. 2050 * To avoid blocking all arc activity, periodically drop 2051 * the arcs_mtx and give other threads a chance to run 2052 * before reacquiring the lock. 2053 * 2054 * If we are looking for a buffer to recycle, we are in 2055 * the hot code path, so don't sleep. 2056 */ 2057 if (!recycle && count++ > arc_evict_iterations) { 2058 list_insert_after(list, ab, &marker); 2059 mutex_exit(evicted_lock); 2060 mutex_exit(lock); 2061 kpreempt(KPREEMPT_SYNC); 2062 mutex_enter(lock); 2063 mutex_enter(evicted_lock); 2064 ab_prev = list_prev(list, &marker); 2065 list_remove(list, &marker); 2066 count = 0; 2067 continue; 2068 } 2069 2070 hash_lock = HDR_LOCK(ab); 2071 have_lock = MUTEX_HELD(hash_lock); 2072 if (have_lock || mutex_tryenter(hash_lock)) { 2073 ASSERT0(refcount_count(&ab->b_refcnt)); 2074 ASSERT(ab->b_datacnt > 0); 2075 while (ab->b_buf) { 2076 arc_buf_t *buf = ab->b_buf; 2077 if (!mutex_tryenter(&buf->b_evict_lock)) { 2078 missed += 1; 2079 break; 2080 } 2081 if (buf->b_data) { 2082 bytes_evicted += ab->b_size; 2083 if (recycle && ab->b_type == type && 2084 ab->b_size == bytes && 2085 !HDR_L2_WRITING(ab)) { 2086 stolen = buf->b_data; 2087 recycle = FALSE; 2088 } 2089 } 2090 if (buf->b_efunc) { 2091 mutex_enter(&arc_eviction_mtx); 2092 arc_buf_destroy(buf, 2093 buf->b_data == stolen, FALSE); 2094 ab->b_buf = buf->b_next; 2095 buf->b_hdr = &arc_eviction_hdr; 2096 buf->b_next = arc_eviction_list; 2097 arc_eviction_list = buf; 2098 mutex_exit(&arc_eviction_mtx); 2099 mutex_exit(&buf->b_evict_lock); 2100 } else { 2101 mutex_exit(&buf->b_evict_lock); 2102 arc_buf_destroy(buf, 2103 buf->b_data == stolen, TRUE); 2104 } 2105 } 2106 2107 if (ab->b_l2hdr) { 2108 ARCSTAT_INCR(arcstat_evict_l2_cached, 2109 ab->b_size); 2110 } else { 2111 if (l2arc_write_eligible(ab->b_spa, ab)) { 2112 ARCSTAT_INCR(arcstat_evict_l2_eligible, 2113 ab->b_size); 2114 } else { 2115 ARCSTAT_INCR( 2116 arcstat_evict_l2_ineligible, 2117 ab->b_size); 2118 } 2119 } 2120 2121 if (ab->b_datacnt == 0) { 2122 arc_change_state(evicted_state, ab, hash_lock); 2123 ASSERT(HDR_IN_HASH_TABLE(ab)); 2124 ab->b_flags |= ARC_IN_HASH_TABLE; 2125 ab->b_flags &= ~ARC_BUF_AVAILABLE; 2126 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 2127 } 2128 if (!have_lock) 2129 mutex_exit(hash_lock); 2130 if (bytes >= 0 && bytes_evicted >= bytes) 2131 break; 2132 if (bytes_remaining > 0) { 2133 mutex_exit(evicted_lock); 2134 mutex_exit(lock); 2135 idx = ((idx + 1) & (list_count - 1)); 2136 lists++; 2137 goto evict_start; 2138 } 2139 } else { 2140 missed += 1; 2141 } 2142 } 2143 2144 mutex_exit(evicted_lock); 2145 mutex_exit(lock); 2146 2147 idx = ((idx + 1) & (list_count - 1)); 2148 lists++; 2149 2150 if (bytes_evicted < bytes) { 2151 if (lists < list_count) 2152 goto evict_start; 2153 else 2154 dprintf("only evicted %lld bytes from %x", 2155 (longlong_t)bytes_evicted, state); 2156 } 2157 if (type == ARC_BUFC_METADATA) 2158 evict_metadata_offset = idx; 2159 else 2160 evict_data_offset = idx; 2161 2162 if (skipped) 2163 ARCSTAT_INCR(arcstat_evict_skip, skipped); 2164 2165 if (missed) 2166 ARCSTAT_INCR(arcstat_mutex_miss, missed); 2167 2168 /* 2169 * Note: we have just evicted some data into the ghost state, 2170 * potentially putting the ghost size over the desired size. Rather 2171 * that evicting from the ghost list in this hot code path, leave 2172 * this chore to the arc_reclaim_thread(). 2173 */ 2174 2175 if (stolen) 2176 ARCSTAT_BUMP(arcstat_stolen); 2177 return (stolen); 2178} 2179 2180/* 2181 * Remove buffers from list until we've removed the specified number of 2182 * bytes. Destroy the buffers that are removed. 2183 */ 2184static void 2185arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2186{ 2187 arc_buf_hdr_t *ab, *ab_prev; 2188 arc_buf_hdr_t marker = { 0 }; 2189 list_t *list, *list_start; 2190 kmutex_t *hash_lock, *lock; 2191 uint64_t bytes_deleted = 0; 2192 uint64_t bufs_skipped = 0; 2193 int count = 0; 2194 static int evict_offset; 2195 int list_count, idx = evict_offset; 2196 int offset, lists = 0; 2197 2198 ASSERT(GHOST_STATE(state)); 2199 2200 /* 2201 * data lists come after metadata lists 2202 */ 2203 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2204 list_count = ARC_BUFC_NUMDATALISTS; 2205 offset = ARC_BUFC_NUMMETADATALISTS; 2206 2207evict_start: 2208 list = &list_start[idx]; 2209 lock = ARCS_LOCK(state, idx + offset); 2210 2211 mutex_enter(lock); 2212 for (ab = list_tail(list); ab; ab = ab_prev) { 2213 ab_prev = list_prev(list, ab); 2214 if (ab->b_type > ARC_BUFC_NUMTYPES) 2215 panic("invalid ab=%p", (void *)ab); 2216 if (spa && ab->b_spa != spa) 2217 continue; 2218 2219 /* ignore markers */ 2220 if (ab->b_spa == 0) 2221 continue; 2222 2223 hash_lock = HDR_LOCK(ab); 2224 /* caller may be trying to modify this buffer, skip it */ 2225 if (MUTEX_HELD(hash_lock)) 2226 continue; 2227 2228 /* 2229 * It may take a long time to evict all the bufs requested. 2230 * To avoid blocking all arc activity, periodically drop 2231 * the arcs_mtx and give other threads a chance to run 2232 * before reacquiring the lock. 2233 */ 2234 if (count++ > arc_evict_iterations) { 2235 list_insert_after(list, ab, &marker); 2236 mutex_exit(lock); 2237 kpreempt(KPREEMPT_SYNC); 2238 mutex_enter(lock); 2239 ab_prev = list_prev(list, &marker); 2240 list_remove(list, &marker); 2241 count = 0; 2242 continue; 2243 } 2244 if (mutex_tryenter(hash_lock)) { 2245 ASSERT(!HDR_IO_IN_PROGRESS(ab)); 2246 ASSERT(ab->b_buf == NULL); 2247 ARCSTAT_BUMP(arcstat_deleted); 2248 bytes_deleted += ab->b_size; 2249 2250 if (ab->b_l2hdr != NULL) { 2251 /* 2252 * This buffer is cached on the 2nd Level ARC; 2253 * don't destroy the header. 2254 */ 2255 arc_change_state(arc_l2c_only, ab, hash_lock); 2256 mutex_exit(hash_lock); 2257 } else { 2258 arc_change_state(arc_anon, ab, hash_lock); 2259 mutex_exit(hash_lock); 2260 arc_hdr_destroy(ab); 2261 } 2262 2263 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 2264 if (bytes >= 0 && bytes_deleted >= bytes) 2265 break; 2266 } else if (bytes < 0) { 2267 /* 2268 * Insert a list marker and then wait for the 2269 * hash lock to become available. Once its 2270 * available, restart from where we left off. 2271 */ 2272 list_insert_after(list, ab, &marker); 2273 mutex_exit(lock); 2274 mutex_enter(hash_lock); 2275 mutex_exit(hash_lock); 2276 mutex_enter(lock); 2277 ab_prev = list_prev(list, &marker); 2278 list_remove(list, &marker); 2279 } else { 2280 bufs_skipped += 1; 2281 } 2282 2283 } 2284 mutex_exit(lock); 2285 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2286 lists++; 2287 2288 if (lists < list_count) 2289 goto evict_start; 2290 2291 evict_offset = idx; 2292 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2293 (bytes < 0 || bytes_deleted < bytes)) { 2294 list_start = &state->arcs_lists[0]; 2295 list_count = ARC_BUFC_NUMMETADATALISTS; 2296 offset = lists = 0; 2297 goto evict_start; 2298 } 2299 2300 if (bufs_skipped) { 2301 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2302 ASSERT(bytes >= 0); 2303 } 2304 2305 if (bytes_deleted < bytes) 2306 dprintf("only deleted %lld bytes from %p", 2307 (longlong_t)bytes_deleted, state); 2308} 2309 2310static void 2311arc_adjust(void) 2312{ 2313 int64_t adjustment, delta; 2314 2315 /* 2316 * Adjust MRU size 2317 */ 2318 2319 adjustment = MIN((int64_t)(arc_size - arc_c), 2320 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2321 arc_p)); 2322 2323 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2324 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2325 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2326 adjustment -= delta; 2327 } 2328 2329 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2330 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2331 (void) arc_evict(arc_mru, 0, delta, FALSE, 2332 ARC_BUFC_METADATA); 2333 } 2334 2335 /* 2336 * Adjust MFU size 2337 */ 2338 2339 adjustment = arc_size - arc_c; 2340 2341 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2342 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2343 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2344 adjustment -= delta; 2345 } 2346 2347 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2348 int64_t delta = MIN(adjustment, 2349 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2350 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2351 ARC_BUFC_METADATA); 2352 } 2353 2354 /* 2355 * Adjust ghost lists 2356 */ 2357 2358 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2359 2360 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2361 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2362 arc_evict_ghost(arc_mru_ghost, 0, delta); 2363 } 2364 2365 adjustment = 2366 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2367 2368 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2369 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2370 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2371 } 2372} 2373 2374static void 2375arc_do_user_evicts(void) 2376{ 2377 static arc_buf_t *tmp_arc_eviction_list; 2378 2379 /* 2380 * Move list over to avoid LOR 2381 */ 2382restart: 2383 mutex_enter(&arc_eviction_mtx); 2384 tmp_arc_eviction_list = arc_eviction_list; 2385 arc_eviction_list = NULL; 2386 mutex_exit(&arc_eviction_mtx); 2387 2388 while (tmp_arc_eviction_list != NULL) { 2389 arc_buf_t *buf = tmp_arc_eviction_list; 2390 tmp_arc_eviction_list = buf->b_next; 2391 mutex_enter(&buf->b_evict_lock); 2392 buf->b_hdr = NULL; 2393 mutex_exit(&buf->b_evict_lock); 2394 2395 if (buf->b_efunc != NULL) 2396 VERIFY0(buf->b_efunc(buf->b_private)); 2397 2398 buf->b_efunc = NULL; 2399 buf->b_private = NULL; 2400 kmem_cache_free(buf_cache, buf); 2401 } 2402 2403 if (arc_eviction_list != NULL) 2404 goto restart; 2405} 2406 2407/* 2408 * Flush all *evictable* data from the cache for the given spa. 2409 * NOTE: this will not touch "active" (i.e. referenced) data. 2410 */ 2411void 2412arc_flush(spa_t *spa) 2413{ 2414 uint64_t guid = 0; 2415 2416 if (spa) 2417 guid = spa_load_guid(spa); 2418 2419 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2420 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2421 if (spa) 2422 break; 2423 } 2424 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2425 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2426 if (spa) 2427 break; 2428 } 2429 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2430 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2431 if (spa) 2432 break; 2433 } 2434 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2435 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2436 if (spa) 2437 break; 2438 } 2439 2440 arc_evict_ghost(arc_mru_ghost, guid, -1); 2441 arc_evict_ghost(arc_mfu_ghost, guid, -1); 2442 2443 mutex_enter(&arc_reclaim_thr_lock); 2444 arc_do_user_evicts(); 2445 mutex_exit(&arc_reclaim_thr_lock); 2446 ASSERT(spa || arc_eviction_list == NULL); 2447} 2448 2449void 2450arc_shrink(void) 2451{ 2452 if (arc_c > arc_c_min) { 2453 uint64_t to_free; 2454 2455#ifdef _KERNEL 2456 to_free = arc_c >> arc_shrink_shift; 2457#else 2458 to_free = arc_c >> arc_shrink_shift; 2459#endif 2460 if (arc_c > arc_c_min + to_free) 2461 atomic_add_64(&arc_c, -to_free); 2462 else 2463 arc_c = arc_c_min; 2464 2465 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2466 if (arc_c > arc_size) 2467 arc_c = MAX(arc_size, arc_c_min); 2468 if (arc_p > arc_c) 2469 arc_p = (arc_c >> 1); 2470 ASSERT(arc_c >= arc_c_min); 2471 ASSERT((int64_t)arc_p >= 0); 2472 } 2473 2474 if (arc_size > arc_c) 2475 arc_adjust(); 2476} 2477 2478static int needfree = 0; 2479 2480static int 2481arc_reclaim_needed(void) 2482{ 2483 2484#ifdef _KERNEL 2485 2486 if (needfree) 2487 return (1); 2488 2489 /* 2490 * Cooperate with pagedaemon when it's time for it to scan 2491 * and reclaim some pages. 2492 */ 2493 if (vm_paging_needed()) 2494 return (1); 2495 2496#ifdef sun 2497 /* 2498 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2499 */ 2500 extra = desfree; 2501 2502 /* 2503 * check that we're out of range of the pageout scanner. It starts to 2504 * schedule paging if freemem is less than lotsfree and needfree. 2505 * lotsfree is the high-water mark for pageout, and needfree is the 2506 * number of needed free pages. We add extra pages here to make sure 2507 * the scanner doesn't start up while we're freeing memory. 2508 */ 2509 if (freemem < lotsfree + needfree + extra) 2510 return (1); 2511 2512 /* 2513 * check to make sure that swapfs has enough space so that anon 2514 * reservations can still succeed. anon_resvmem() checks that the 2515 * availrmem is greater than swapfs_minfree, and the number of reserved 2516 * swap pages. We also add a bit of extra here just to prevent 2517 * circumstances from getting really dire. 2518 */ 2519 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2520 return (1); 2521 2522#if defined(__i386) 2523 /* 2524 * If we're on an i386 platform, it's possible that we'll exhaust the 2525 * kernel heap space before we ever run out of available physical 2526 * memory. Most checks of the size of the heap_area compare against 2527 * tune.t_minarmem, which is the minimum available real memory that we 2528 * can have in the system. However, this is generally fixed at 25 pages 2529 * which is so low that it's useless. In this comparison, we seek to 2530 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2531 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2532 * free) 2533 */ 2534 if (btop(vmem_size(heap_arena, VMEM_FREE)) < 2535 (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2)) 2536 return (1); 2537#endif 2538#else /* !sun */ 2539 if (kmem_used() > (kmem_size() * 3) / 4) 2540 return (1); 2541#endif /* sun */ 2542 2543#else 2544 if (spa_get_random(100) == 0) 2545 return (1); 2546#endif 2547 return (0); 2548} 2549 2550extern kmem_cache_t *zio_buf_cache[]; 2551extern kmem_cache_t *zio_data_buf_cache[]; 2552 2553static void 2554arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2555{ 2556 size_t i; 2557 kmem_cache_t *prev_cache = NULL; 2558 kmem_cache_t *prev_data_cache = NULL; 2559 2560#ifdef _KERNEL 2561 if (arc_meta_used >= arc_meta_limit) { 2562 /* 2563 * We are exceeding our meta-data cache limit. 2564 * Purge some DNLC entries to release holds on meta-data. 2565 */ 2566 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2567 } 2568#if defined(__i386) 2569 /* 2570 * Reclaim unused memory from all kmem caches. 2571 */ 2572 kmem_reap(); 2573#endif 2574#endif 2575 2576 /* 2577 * An aggressive reclamation will shrink the cache size as well as 2578 * reap free buffers from the arc kmem caches. 2579 */ 2580 if (strat == ARC_RECLAIM_AGGR) 2581 arc_shrink(); 2582 2583 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2584 if (zio_buf_cache[i] != prev_cache) { 2585 prev_cache = zio_buf_cache[i]; 2586 kmem_cache_reap_now(zio_buf_cache[i]); 2587 } 2588 if (zio_data_buf_cache[i] != prev_data_cache) { 2589 prev_data_cache = zio_data_buf_cache[i]; 2590 kmem_cache_reap_now(zio_data_buf_cache[i]); 2591 } 2592 } 2593 kmem_cache_reap_now(buf_cache); 2594 kmem_cache_reap_now(hdr_cache); 2595} 2596 2597static void 2598arc_reclaim_thread(void *dummy __unused) 2599{ 2600 clock_t growtime = 0; 2601 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2602 callb_cpr_t cpr; 2603 2604 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2605 2606 mutex_enter(&arc_reclaim_thr_lock); 2607 while (arc_thread_exit == 0) { 2608 if (arc_reclaim_needed()) { 2609 2610 if (arc_no_grow) { 2611 if (last_reclaim == ARC_RECLAIM_CONS) { 2612 last_reclaim = ARC_RECLAIM_AGGR; 2613 } else { 2614 last_reclaim = ARC_RECLAIM_CONS; 2615 } 2616 } else { 2617 arc_no_grow = TRUE; 2618 last_reclaim = ARC_RECLAIM_AGGR; 2619 membar_producer(); 2620 } 2621 2622 /* reset the growth delay for every reclaim */ 2623 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2624 2625 if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2626 /* 2627 * If needfree is TRUE our vm_lowmem hook 2628 * was called and in that case we must free some 2629 * memory, so switch to aggressive mode. 2630 */ 2631 arc_no_grow = TRUE; 2632 last_reclaim = ARC_RECLAIM_AGGR; 2633 } 2634 arc_kmem_reap_now(last_reclaim); 2635 arc_warm = B_TRUE; 2636 2637 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2638 arc_no_grow = FALSE; 2639 } 2640 2641 arc_adjust(); 2642 2643 if (arc_eviction_list != NULL) 2644 arc_do_user_evicts(); 2645 2646#ifdef _KERNEL 2647 if (needfree) { 2648 needfree = 0; 2649 wakeup(&needfree); 2650 } 2651#endif 2652 2653 /* block until needed, or one second, whichever is shorter */ 2654 CALLB_CPR_SAFE_BEGIN(&cpr); 2655 (void) cv_timedwait(&arc_reclaim_thr_cv, 2656 &arc_reclaim_thr_lock, hz); 2657 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2658 } 2659 2660 arc_thread_exit = 0; 2661 cv_broadcast(&arc_reclaim_thr_cv); 2662 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2663 thread_exit(); 2664} 2665 2666/* 2667 * Adapt arc info given the number of bytes we are trying to add and 2668 * the state that we are comming from. This function is only called 2669 * when we are adding new content to the cache. 2670 */ 2671static void 2672arc_adapt(int bytes, arc_state_t *state) 2673{ 2674 int mult; 2675 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2676 2677 if (state == arc_l2c_only) 2678 return; 2679 2680 ASSERT(bytes > 0); 2681 /* 2682 * Adapt the target size of the MRU list: 2683 * - if we just hit in the MRU ghost list, then increase 2684 * the target size of the MRU list. 2685 * - if we just hit in the MFU ghost list, then increase 2686 * the target size of the MFU list by decreasing the 2687 * target size of the MRU list. 2688 */ 2689 if (state == arc_mru_ghost) { 2690 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2691 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2692 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2693 2694 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2695 } else if (state == arc_mfu_ghost) { 2696 uint64_t delta; 2697 2698 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2699 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2700 mult = MIN(mult, 10); 2701 2702 delta = MIN(bytes * mult, arc_p); 2703 arc_p = MAX(arc_p_min, arc_p - delta); 2704 } 2705 ASSERT((int64_t)arc_p >= 0); 2706 2707 if (arc_reclaim_needed()) { 2708 cv_signal(&arc_reclaim_thr_cv); 2709 return; 2710 } 2711 2712 if (arc_no_grow) 2713 return; 2714 2715 if (arc_c >= arc_c_max) 2716 return; 2717 2718 /* 2719 * If we're within (2 * maxblocksize) bytes of the target 2720 * cache size, increment the target cache size 2721 */ 2722 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2723 atomic_add_64(&arc_c, (int64_t)bytes); 2724 if (arc_c > arc_c_max) 2725 arc_c = arc_c_max; 2726 else if (state == arc_anon) 2727 atomic_add_64(&arc_p, (int64_t)bytes); 2728 if (arc_p > arc_c) 2729 arc_p = arc_c; 2730 } 2731 ASSERT((int64_t)arc_p >= 0); 2732} 2733 2734/* 2735 * Check if the cache has reached its limits and eviction is required 2736 * prior to insert. 2737 */ 2738static int 2739arc_evict_needed(arc_buf_contents_t type) 2740{ 2741 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2742 return (1); 2743 2744#ifdef sun 2745#ifdef _KERNEL 2746 /* 2747 * If zio data pages are being allocated out of a separate heap segment, 2748 * then enforce that the size of available vmem for this area remains 2749 * above about 1/32nd free. 2750 */ 2751 if (type == ARC_BUFC_DATA && zio_arena != NULL && 2752 vmem_size(zio_arena, VMEM_FREE) < 2753 (vmem_size(zio_arena, VMEM_ALLOC) >> 5)) 2754 return (1); 2755#endif 2756#endif /* sun */ 2757 2758 if (arc_reclaim_needed()) 2759 return (1); 2760 2761 return (arc_size > arc_c); 2762} 2763 2764/* 2765 * The buffer, supplied as the first argument, needs a data block. 2766 * So, if we are at cache max, determine which cache should be victimized. 2767 * We have the following cases: 2768 * 2769 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2770 * In this situation if we're out of space, but the resident size of the MFU is 2771 * under the limit, victimize the MFU cache to satisfy this insertion request. 2772 * 2773 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2774 * Here, we've used up all of the available space for the MRU, so we need to 2775 * evict from our own cache instead. Evict from the set of resident MRU 2776 * entries. 2777 * 2778 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2779 * c minus p represents the MFU space in the cache, since p is the size of the 2780 * cache that is dedicated to the MRU. In this situation there's still space on 2781 * the MFU side, so the MRU side needs to be victimized. 2782 * 2783 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2784 * MFU's resident set is consuming more space than it has been allotted. In 2785 * this situation, we must victimize our own cache, the MFU, for this insertion. 2786 */ 2787static void 2788arc_get_data_buf(arc_buf_t *buf) 2789{ 2790 arc_state_t *state = buf->b_hdr->b_state; 2791 uint64_t size = buf->b_hdr->b_size; 2792 arc_buf_contents_t type = buf->b_hdr->b_type; 2793 2794 arc_adapt(size, state); 2795 2796 /* 2797 * We have not yet reached cache maximum size, 2798 * just allocate a new buffer. 2799 */ 2800 if (!arc_evict_needed(type)) { 2801 if (type == ARC_BUFC_METADATA) { 2802 buf->b_data = zio_buf_alloc(size); 2803 arc_space_consume(size, ARC_SPACE_DATA); 2804 } else { 2805 ASSERT(type == ARC_BUFC_DATA); 2806 buf->b_data = zio_data_buf_alloc(size); 2807 ARCSTAT_INCR(arcstat_data_size, size); 2808 atomic_add_64(&arc_size, size); 2809 } 2810 goto out; 2811 } 2812 2813 /* 2814 * If we are prefetching from the mfu ghost list, this buffer 2815 * will end up on the mru list; so steal space from there. 2816 */ 2817 if (state == arc_mfu_ghost) 2818 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2819 else if (state == arc_mru_ghost) 2820 state = arc_mru; 2821 2822 if (state == arc_mru || state == arc_anon) { 2823 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2824 state = (arc_mfu->arcs_lsize[type] >= size && 2825 arc_p > mru_used) ? arc_mfu : arc_mru; 2826 } else { 2827 /* MFU cases */ 2828 uint64_t mfu_space = arc_c - arc_p; 2829 state = (arc_mru->arcs_lsize[type] >= size && 2830 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2831 } 2832 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2833 if (type == ARC_BUFC_METADATA) { 2834 buf->b_data = zio_buf_alloc(size); 2835 arc_space_consume(size, ARC_SPACE_DATA); 2836 } else { 2837 ASSERT(type == ARC_BUFC_DATA); 2838 buf->b_data = zio_data_buf_alloc(size); 2839 ARCSTAT_INCR(arcstat_data_size, size); 2840 atomic_add_64(&arc_size, size); 2841 } 2842 ARCSTAT_BUMP(arcstat_recycle_miss); 2843 } 2844 ASSERT(buf->b_data != NULL); 2845out: 2846 /* 2847 * Update the state size. Note that ghost states have a 2848 * "ghost size" and so don't need to be updated. 2849 */ 2850 if (!GHOST_STATE(buf->b_hdr->b_state)) { 2851 arc_buf_hdr_t *hdr = buf->b_hdr; 2852 2853 atomic_add_64(&hdr->b_state->arcs_size, size); 2854 if (list_link_active(&hdr->b_arc_node)) { 2855 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2856 atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2857 } 2858 /* 2859 * If we are growing the cache, and we are adding anonymous 2860 * data, and we have outgrown arc_p, update arc_p 2861 */ 2862 if (arc_size < arc_c && hdr->b_state == arc_anon && 2863 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2864 arc_p = MIN(arc_c, arc_p + size); 2865 } 2866 ARCSTAT_BUMP(arcstat_allocated); 2867} 2868 2869/* 2870 * This routine is called whenever a buffer is accessed. 2871 * NOTE: the hash lock is dropped in this function. 2872 */ 2873static void 2874arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2875{ 2876 clock_t now; 2877 2878 ASSERT(MUTEX_HELD(hash_lock)); 2879 2880 if (buf->b_state == arc_anon) { 2881 /* 2882 * This buffer is not in the cache, and does not 2883 * appear in our "ghost" list. Add the new buffer 2884 * to the MRU state. 2885 */ 2886 2887 ASSERT(buf->b_arc_access == 0); 2888 buf->b_arc_access = ddi_get_lbolt(); 2889 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2890 arc_change_state(arc_mru, buf, hash_lock); 2891 2892 } else if (buf->b_state == arc_mru) { 2893 now = ddi_get_lbolt(); 2894 2895 /* 2896 * If this buffer is here because of a prefetch, then either: 2897 * - clear the flag if this is a "referencing" read 2898 * (any subsequent access will bump this into the MFU state). 2899 * or 2900 * - move the buffer to the head of the list if this is 2901 * another prefetch (to make it less likely to be evicted). 2902 */ 2903 if ((buf->b_flags & ARC_PREFETCH) != 0) { 2904 if (refcount_count(&buf->b_refcnt) == 0) { 2905 ASSERT(list_link_active(&buf->b_arc_node)); 2906 } else { 2907 buf->b_flags &= ~ARC_PREFETCH; 2908 ARCSTAT_BUMP(arcstat_mru_hits); 2909 } 2910 buf->b_arc_access = now; 2911 return; 2912 } 2913 2914 /* 2915 * This buffer has been "accessed" only once so far, 2916 * but it is still in the cache. Move it to the MFU 2917 * state. 2918 */ 2919 if (now > buf->b_arc_access + ARC_MINTIME) { 2920 /* 2921 * More than 125ms have passed since we 2922 * instantiated this buffer. Move it to the 2923 * most frequently used state. 2924 */ 2925 buf->b_arc_access = now; 2926 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2927 arc_change_state(arc_mfu, buf, hash_lock); 2928 } 2929 ARCSTAT_BUMP(arcstat_mru_hits); 2930 } else if (buf->b_state == arc_mru_ghost) { 2931 arc_state_t *new_state; 2932 /* 2933 * This buffer has been "accessed" recently, but 2934 * was evicted from the cache. Move it to the 2935 * MFU state. 2936 */ 2937 2938 if (buf->b_flags & ARC_PREFETCH) { 2939 new_state = arc_mru; 2940 if (refcount_count(&buf->b_refcnt) > 0) 2941 buf->b_flags &= ~ARC_PREFETCH; 2942 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2943 } else { 2944 new_state = arc_mfu; 2945 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2946 } 2947 2948 buf->b_arc_access = ddi_get_lbolt(); 2949 arc_change_state(new_state, buf, hash_lock); 2950 2951 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 2952 } else if (buf->b_state == arc_mfu) { 2953 /* 2954 * This buffer has been accessed more than once and is 2955 * still in the cache. Keep it in the MFU state. 2956 * 2957 * NOTE: an add_reference() that occurred when we did 2958 * the arc_read() will have kicked this off the list. 2959 * If it was a prefetch, we will explicitly move it to 2960 * the head of the list now. 2961 */ 2962 if ((buf->b_flags & ARC_PREFETCH) != 0) { 2963 ASSERT(refcount_count(&buf->b_refcnt) == 0); 2964 ASSERT(list_link_active(&buf->b_arc_node)); 2965 } 2966 ARCSTAT_BUMP(arcstat_mfu_hits); 2967 buf->b_arc_access = ddi_get_lbolt(); 2968 } else if (buf->b_state == arc_mfu_ghost) { 2969 arc_state_t *new_state = arc_mfu; 2970 /* 2971 * This buffer has been accessed more than once but has 2972 * been evicted from the cache. Move it back to the 2973 * MFU state. 2974 */ 2975 2976 if (buf->b_flags & ARC_PREFETCH) { 2977 /* 2978 * This is a prefetch access... 2979 * move this block back to the MRU state. 2980 */ 2981 ASSERT0(refcount_count(&buf->b_refcnt)); 2982 new_state = arc_mru; 2983 } 2984 2985 buf->b_arc_access = ddi_get_lbolt(); 2986 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2987 arc_change_state(new_state, buf, hash_lock); 2988 2989 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 2990 } else if (buf->b_state == arc_l2c_only) { 2991 /* 2992 * This buffer is on the 2nd Level ARC. 2993 */ 2994 2995 buf->b_arc_access = ddi_get_lbolt(); 2996 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2997 arc_change_state(arc_mfu, buf, hash_lock); 2998 } else { 2999 ASSERT(!"invalid arc state"); 3000 } 3001} 3002 3003/* a generic arc_done_func_t which you can use */ 3004/* ARGSUSED */ 3005void 3006arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 3007{ 3008 if (zio == NULL || zio->io_error == 0) 3009 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 3010 VERIFY(arc_buf_remove_ref(buf, arg)); 3011} 3012 3013/* a generic arc_done_func_t */ 3014void 3015arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 3016{ 3017 arc_buf_t **bufp = arg; 3018 if (zio && zio->io_error) { 3019 VERIFY(arc_buf_remove_ref(buf, arg)); 3020 *bufp = NULL; 3021 } else { 3022 *bufp = buf; 3023 ASSERT(buf->b_data); 3024 } 3025} 3026 3027static void 3028arc_read_done(zio_t *zio) 3029{ 3030 arc_buf_hdr_t *hdr; 3031 arc_buf_t *buf; 3032 arc_buf_t *abuf; /* buffer we're assigning to callback */ 3033 kmutex_t *hash_lock = NULL; 3034 arc_callback_t *callback_list, *acb; 3035 int freeable = FALSE; 3036 3037 buf = zio->io_private; 3038 hdr = buf->b_hdr; 3039 3040 /* 3041 * The hdr was inserted into hash-table and removed from lists 3042 * prior to starting I/O. We should find this header, since 3043 * it's in the hash table, and it should be legit since it's 3044 * not possible to evict it during the I/O. The only possible 3045 * reason for it not to be found is if we were freed during the 3046 * read. 3047 */ 3048 if (HDR_IN_HASH_TABLE(hdr)) { 3049 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3050 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3051 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3052 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3053 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3054 3055 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3056 &hash_lock); 3057 3058 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3059 hash_lock == NULL) || 3060 (found == hdr && 3061 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3062 (found == hdr && HDR_L2_READING(hdr))); 3063 } 3064 3065 hdr->b_flags &= ~ARC_L2_EVICTED; 3066 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 3067 hdr->b_flags &= ~ARC_L2CACHE; 3068 3069 /* byteswap if necessary */ 3070 callback_list = hdr->b_acb; 3071 ASSERT(callback_list != NULL); 3072 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3073 dmu_object_byteswap_t bswap = 3074 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3075 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3076 byteswap_uint64_array : 3077 dmu_ot_byteswap[bswap].ob_func; 3078 func(buf->b_data, hdr->b_size); 3079 } 3080 3081 arc_cksum_compute(buf, B_FALSE); 3082#ifdef illumos 3083 arc_buf_watch(buf); 3084#endif /* illumos */ 3085 3086 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3087 /* 3088 * Only call arc_access on anonymous buffers. This is because 3089 * if we've issued an I/O for an evicted buffer, we've already 3090 * called arc_access (to prevent any simultaneous readers from 3091 * getting confused). 3092 */ 3093 arc_access(hdr, hash_lock); 3094 } 3095 3096 /* create copies of the data buffer for the callers */ 3097 abuf = buf; 3098 for (acb = callback_list; acb; acb = acb->acb_next) { 3099 if (acb->acb_done) { 3100 if (abuf == NULL) { 3101 ARCSTAT_BUMP(arcstat_duplicate_reads); 3102 abuf = arc_buf_clone(buf); 3103 } 3104 acb->acb_buf = abuf; 3105 abuf = NULL; 3106 } 3107 } 3108 hdr->b_acb = NULL; 3109 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3110 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3111 if (abuf == buf) { 3112 ASSERT(buf->b_efunc == NULL); 3113 ASSERT(hdr->b_datacnt == 1); 3114 hdr->b_flags |= ARC_BUF_AVAILABLE; 3115 } 3116 3117 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3118 3119 if (zio->io_error != 0) { 3120 hdr->b_flags |= ARC_IO_ERROR; 3121 if (hdr->b_state != arc_anon) 3122 arc_change_state(arc_anon, hdr, hash_lock); 3123 if (HDR_IN_HASH_TABLE(hdr)) 3124 buf_hash_remove(hdr); 3125 freeable = refcount_is_zero(&hdr->b_refcnt); 3126 } 3127 3128 /* 3129 * Broadcast before we drop the hash_lock to avoid the possibility 3130 * that the hdr (and hence the cv) might be freed before we get to 3131 * the cv_broadcast(). 3132 */ 3133 cv_broadcast(&hdr->b_cv); 3134 3135 if (hash_lock) { 3136 mutex_exit(hash_lock); 3137 } else { 3138 /* 3139 * This block was freed while we waited for the read to 3140 * complete. It has been removed from the hash table and 3141 * moved to the anonymous state (so that it won't show up 3142 * in the cache). 3143 */ 3144 ASSERT3P(hdr->b_state, ==, arc_anon); 3145 freeable = refcount_is_zero(&hdr->b_refcnt); 3146 } 3147 3148 /* execute each callback and free its structure */ 3149 while ((acb = callback_list) != NULL) { 3150 if (acb->acb_done) 3151 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3152 3153 if (acb->acb_zio_dummy != NULL) { 3154 acb->acb_zio_dummy->io_error = zio->io_error; 3155 zio_nowait(acb->acb_zio_dummy); 3156 } 3157 3158 callback_list = acb->acb_next; 3159 kmem_free(acb, sizeof (arc_callback_t)); 3160 } 3161 3162 if (freeable) 3163 arc_hdr_destroy(hdr); 3164} 3165 3166/* 3167 * "Read" the block block at the specified DVA (in bp) via the 3168 * cache. If the block is found in the cache, invoke the provided 3169 * callback immediately and return. Note that the `zio' parameter 3170 * in the callback will be NULL in this case, since no IO was 3171 * required. If the block is not in the cache pass the read request 3172 * on to the spa with a substitute callback function, so that the 3173 * requested block will be added to the cache. 3174 * 3175 * If a read request arrives for a block that has a read in-progress, 3176 * either wait for the in-progress read to complete (and return the 3177 * results); or, if this is a read with a "done" func, add a record 3178 * to the read to invoke the "done" func when the read completes, 3179 * and return; or just return. 3180 * 3181 * arc_read_done() will invoke all the requested "done" functions 3182 * for readers of this block. 3183 */ 3184int 3185arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3186 void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags, 3187 const zbookmark_phys_t *zb) 3188{ 3189 arc_buf_hdr_t *hdr = NULL; 3190 arc_buf_t *buf = NULL; 3191 kmutex_t *hash_lock = NULL; 3192 zio_t *rzio; 3193 uint64_t guid = spa_load_guid(spa); 3194 3195 ASSERT(!BP_IS_EMBEDDED(bp) || 3196 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3197 3198top: 3199 if (!BP_IS_EMBEDDED(bp)) { 3200 /* 3201 * Embedded BP's have no DVA and require no I/O to "read". 3202 * Create an anonymous arc buf to back it. 3203 */ 3204 hdr = buf_hash_find(guid, bp, &hash_lock); 3205 } 3206 3207 if (hdr != NULL && hdr->b_datacnt > 0) { 3208 3209 *arc_flags |= ARC_CACHED; 3210 3211 if (HDR_IO_IN_PROGRESS(hdr)) { 3212 3213 if (*arc_flags & ARC_WAIT) { 3214 cv_wait(&hdr->b_cv, hash_lock); 3215 mutex_exit(hash_lock); 3216 goto top; 3217 } 3218 ASSERT(*arc_flags & ARC_NOWAIT); 3219 3220 if (done) { 3221 arc_callback_t *acb = NULL; 3222 3223 acb = kmem_zalloc(sizeof (arc_callback_t), 3224 KM_SLEEP); 3225 acb->acb_done = done; 3226 acb->acb_private = private; 3227 if (pio != NULL) 3228 acb->acb_zio_dummy = zio_null(pio, 3229 spa, NULL, NULL, NULL, zio_flags); 3230 3231 ASSERT(acb->acb_done != NULL); 3232 acb->acb_next = hdr->b_acb; 3233 hdr->b_acb = acb; 3234 add_reference(hdr, hash_lock, private); 3235 mutex_exit(hash_lock); 3236 return (0); 3237 } 3238 mutex_exit(hash_lock); 3239 return (0); 3240 } 3241 3242 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3243 3244 if (done) { 3245 add_reference(hdr, hash_lock, private); 3246 /* 3247 * If this block is already in use, create a new 3248 * copy of the data so that we will be guaranteed 3249 * that arc_release() will always succeed. 3250 */ 3251 buf = hdr->b_buf; 3252 ASSERT(buf); 3253 ASSERT(buf->b_data); 3254 if (HDR_BUF_AVAILABLE(hdr)) { 3255 ASSERT(buf->b_efunc == NULL); 3256 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3257 } else { 3258 buf = arc_buf_clone(buf); 3259 } 3260 3261 } else if (*arc_flags & ARC_PREFETCH && 3262 refcount_count(&hdr->b_refcnt) == 0) { 3263 hdr->b_flags |= ARC_PREFETCH; 3264 } 3265 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3266 arc_access(hdr, hash_lock); 3267 if (*arc_flags & ARC_L2CACHE) 3268 hdr->b_flags |= ARC_L2CACHE; 3269 if (*arc_flags & ARC_L2COMPRESS) 3270 hdr->b_flags |= ARC_L2COMPRESS; 3271 mutex_exit(hash_lock); 3272 ARCSTAT_BUMP(arcstat_hits); 3273 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3274 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3275 data, metadata, hits); 3276 3277 if (done) 3278 done(NULL, buf, private); 3279 } else { 3280 uint64_t size = BP_GET_LSIZE(bp); 3281 arc_callback_t *acb; 3282 vdev_t *vd = NULL; 3283 uint64_t addr = 0; 3284 boolean_t devw = B_FALSE; 3285 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3286 uint64_t b_asize = 0; 3287 3288 if (hdr == NULL) { 3289 /* this block is not in the cache */ 3290 arc_buf_hdr_t *exists = NULL; 3291 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3292 buf = arc_buf_alloc(spa, size, private, type); 3293 hdr = buf->b_hdr; 3294 if (!BP_IS_EMBEDDED(bp)) { 3295 hdr->b_dva = *BP_IDENTITY(bp); 3296 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3297 hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3298 exists = buf_hash_insert(hdr, &hash_lock); 3299 } 3300 if (exists != NULL) { 3301 /* somebody beat us to the hash insert */ 3302 mutex_exit(hash_lock); 3303 buf_discard_identity(hdr); 3304 (void) arc_buf_remove_ref(buf, private); 3305 goto top; /* restart the IO request */ 3306 } 3307 /* if this is a prefetch, we don't have a reference */ 3308 if (*arc_flags & ARC_PREFETCH) { 3309 (void) remove_reference(hdr, hash_lock, 3310 private); 3311 hdr->b_flags |= ARC_PREFETCH; 3312 } 3313 if (*arc_flags & ARC_L2CACHE) 3314 hdr->b_flags |= ARC_L2CACHE; 3315 if (*arc_flags & ARC_L2COMPRESS) 3316 hdr->b_flags |= ARC_L2COMPRESS; 3317 if (BP_GET_LEVEL(bp) > 0) 3318 hdr->b_flags |= ARC_INDIRECT; 3319 } else { 3320 /* this block is in the ghost cache */ 3321 ASSERT(GHOST_STATE(hdr->b_state)); 3322 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3323 ASSERT0(refcount_count(&hdr->b_refcnt)); 3324 ASSERT(hdr->b_buf == NULL); 3325 3326 /* if this is a prefetch, we don't have a reference */ 3327 if (*arc_flags & ARC_PREFETCH) 3328 hdr->b_flags |= ARC_PREFETCH; 3329 else 3330 add_reference(hdr, hash_lock, private); 3331 if (*arc_flags & ARC_L2CACHE) 3332 hdr->b_flags |= ARC_L2CACHE; 3333 if (*arc_flags & ARC_L2COMPRESS) 3334 hdr->b_flags |= ARC_L2COMPRESS; 3335 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3336 buf->b_hdr = hdr; 3337 buf->b_data = NULL; 3338 buf->b_efunc = NULL; 3339 buf->b_private = NULL; 3340 buf->b_next = NULL; 3341 hdr->b_buf = buf; 3342 ASSERT(hdr->b_datacnt == 0); 3343 hdr->b_datacnt = 1; 3344 arc_get_data_buf(buf); 3345 arc_access(hdr, hash_lock); 3346 } 3347 3348 ASSERT(!GHOST_STATE(hdr->b_state)); 3349 3350 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3351 acb->acb_done = done; 3352 acb->acb_private = private; 3353 3354 ASSERT(hdr->b_acb == NULL); 3355 hdr->b_acb = acb; 3356 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3357 3358 if (hdr->b_l2hdr != NULL && 3359 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3360 devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3361 addr = hdr->b_l2hdr->b_daddr; 3362 b_compress = hdr->b_l2hdr->b_compress; 3363 b_asize = hdr->b_l2hdr->b_asize; 3364 /* 3365 * Lock out device removal. 3366 */ 3367 if (vdev_is_dead(vd) || 3368 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3369 vd = NULL; 3370 } 3371 3372 if (hash_lock != NULL) 3373 mutex_exit(hash_lock); 3374 3375 /* 3376 * At this point, we have a level 1 cache miss. Try again in 3377 * L2ARC if possible. 3378 */ 3379 ASSERT3U(hdr->b_size, ==, size); 3380 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3381 uint64_t, size, zbookmark_phys_t *, zb); 3382 ARCSTAT_BUMP(arcstat_misses); 3383 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3384 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3385 data, metadata, misses); 3386#ifdef _KERNEL 3387 curthread->td_ru.ru_inblock++; 3388#endif 3389 3390 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3391 /* 3392 * Read from the L2ARC if the following are true: 3393 * 1. The L2ARC vdev was previously cached. 3394 * 2. This buffer still has L2ARC metadata. 3395 * 3. This buffer isn't currently writing to the L2ARC. 3396 * 4. The L2ARC entry wasn't evicted, which may 3397 * also have invalidated the vdev. 3398 * 5. This isn't prefetch and l2arc_noprefetch is set. 3399 */ 3400 if (hdr->b_l2hdr != NULL && 3401 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3402 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3403 l2arc_read_callback_t *cb; 3404 3405 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3406 ARCSTAT_BUMP(arcstat_l2_hits); 3407 3408 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3409 KM_SLEEP); 3410 cb->l2rcb_buf = buf; 3411 cb->l2rcb_spa = spa; 3412 cb->l2rcb_bp = *bp; 3413 cb->l2rcb_zb = *zb; 3414 cb->l2rcb_flags = zio_flags; 3415 cb->l2rcb_compress = b_compress; 3416 3417 ASSERT(addr >= VDEV_LABEL_START_SIZE && 3418 addr + size < vd->vdev_psize - 3419 VDEV_LABEL_END_SIZE); 3420 3421 /* 3422 * l2arc read. The SCL_L2ARC lock will be 3423 * released by l2arc_read_done(). 3424 * Issue a null zio if the underlying buffer 3425 * was squashed to zero size by compression. 3426 */ 3427 if (b_compress == ZIO_COMPRESS_EMPTY) { 3428 rzio = zio_null(pio, spa, vd, 3429 l2arc_read_done, cb, 3430 zio_flags | ZIO_FLAG_DONT_CACHE | 3431 ZIO_FLAG_CANFAIL | 3432 ZIO_FLAG_DONT_PROPAGATE | 3433 ZIO_FLAG_DONT_RETRY); 3434 } else { 3435 rzio = zio_read_phys(pio, vd, addr, 3436 b_asize, buf->b_data, 3437 ZIO_CHECKSUM_OFF, 3438 l2arc_read_done, cb, priority, 3439 zio_flags | ZIO_FLAG_DONT_CACHE | 3440 ZIO_FLAG_CANFAIL | 3441 ZIO_FLAG_DONT_PROPAGATE | 3442 ZIO_FLAG_DONT_RETRY, B_FALSE); 3443 } 3444 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3445 zio_t *, rzio); 3446 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3447 3448 if (*arc_flags & ARC_NOWAIT) { 3449 zio_nowait(rzio); 3450 return (0); 3451 } 3452 3453 ASSERT(*arc_flags & ARC_WAIT); 3454 if (zio_wait(rzio) == 0) 3455 return (0); 3456 3457 /* l2arc read error; goto zio_read() */ 3458 } else { 3459 DTRACE_PROBE1(l2arc__miss, 3460 arc_buf_hdr_t *, hdr); 3461 ARCSTAT_BUMP(arcstat_l2_misses); 3462 if (HDR_L2_WRITING(hdr)) 3463 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3464 spa_config_exit(spa, SCL_L2ARC, vd); 3465 } 3466 } else { 3467 if (vd != NULL) 3468 spa_config_exit(spa, SCL_L2ARC, vd); 3469 if (l2arc_ndev != 0) { 3470 DTRACE_PROBE1(l2arc__miss, 3471 arc_buf_hdr_t *, hdr); 3472 ARCSTAT_BUMP(arcstat_l2_misses); 3473 } 3474 } 3475 3476 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3477 arc_read_done, buf, priority, zio_flags, zb); 3478 3479 if (*arc_flags & ARC_WAIT) 3480 return (zio_wait(rzio)); 3481 3482 ASSERT(*arc_flags & ARC_NOWAIT); 3483 zio_nowait(rzio); 3484 } 3485 return (0); 3486} 3487 3488void 3489arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3490{ 3491 ASSERT(buf->b_hdr != NULL); 3492 ASSERT(buf->b_hdr->b_state != arc_anon); 3493 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3494 ASSERT(buf->b_efunc == NULL); 3495 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3496 3497 buf->b_efunc = func; 3498 buf->b_private = private; 3499} 3500 3501/* 3502 * Notify the arc that a block was freed, and thus will never be used again. 3503 */ 3504void 3505arc_freed(spa_t *spa, const blkptr_t *bp) 3506{ 3507 arc_buf_hdr_t *hdr; 3508 kmutex_t *hash_lock; 3509 uint64_t guid = spa_load_guid(spa); 3510 3511 ASSERT(!BP_IS_EMBEDDED(bp)); 3512 3513 hdr = buf_hash_find(guid, bp, &hash_lock); 3514 if (hdr == NULL) 3515 return; 3516 if (HDR_BUF_AVAILABLE(hdr)) { 3517 arc_buf_t *buf = hdr->b_buf; 3518 add_reference(hdr, hash_lock, FTAG); 3519 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3520 mutex_exit(hash_lock); 3521 3522 arc_release(buf, FTAG); 3523 (void) arc_buf_remove_ref(buf, FTAG); 3524 } else { 3525 mutex_exit(hash_lock); 3526 } 3527 3528} 3529 3530/* 3531 * Clear the user eviction callback set by arc_set_callback(), first calling 3532 * it if it exists. Because the presence of a callback keeps an arc_buf cached 3533 * clearing the callback may result in the arc_buf being destroyed. However, 3534 * it will not result in the *last* arc_buf being destroyed, hence the data 3535 * will remain cached in the ARC. We make a copy of the arc buffer here so 3536 * that we can process the callback without holding any locks. 3537 * 3538 * It's possible that the callback is already in the process of being cleared 3539 * by another thread. In this case we can not clear the callback. 3540 * 3541 * Returns B_TRUE if the callback was successfully called and cleared. 3542 */ 3543boolean_t 3544arc_clear_callback(arc_buf_t *buf) 3545{ 3546 arc_buf_hdr_t *hdr; 3547 kmutex_t *hash_lock; 3548 arc_evict_func_t *efunc = buf->b_efunc; 3549 void *private = buf->b_private; 3550 list_t *list, *evicted_list; 3551 kmutex_t *lock, *evicted_lock; 3552 3553 mutex_enter(&buf->b_evict_lock); 3554 hdr = buf->b_hdr; 3555 if (hdr == NULL) { 3556 /* 3557 * We are in arc_do_user_evicts(). 3558 */ 3559 ASSERT(buf->b_data == NULL); 3560 mutex_exit(&buf->b_evict_lock); 3561 return (B_FALSE); 3562 } else if (buf->b_data == NULL) { 3563 /* 3564 * We are on the eviction list; process this buffer now 3565 * but let arc_do_user_evicts() do the reaping. 3566 */ 3567 buf->b_efunc = NULL; 3568 mutex_exit(&buf->b_evict_lock); 3569 VERIFY0(efunc(private)); 3570 return (B_TRUE); 3571 } 3572 hash_lock = HDR_LOCK(hdr); 3573 mutex_enter(hash_lock); 3574 hdr = buf->b_hdr; 3575 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3576 3577 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3578 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3579 3580 buf->b_efunc = NULL; 3581 buf->b_private = NULL; 3582 3583 if (hdr->b_datacnt > 1) { 3584 mutex_exit(&buf->b_evict_lock); 3585 arc_buf_destroy(buf, FALSE, TRUE); 3586 } else { 3587 ASSERT(buf == hdr->b_buf); 3588 hdr->b_flags |= ARC_BUF_AVAILABLE; 3589 mutex_exit(&buf->b_evict_lock); 3590 } 3591 3592 mutex_exit(hash_lock); 3593 VERIFY0(efunc(private)); 3594 return (B_TRUE); 3595} 3596 3597/* 3598 * Release this buffer from the cache, making it an anonymous buffer. This 3599 * must be done after a read and prior to modifying the buffer contents. 3600 * If the buffer has more than one reference, we must make 3601 * a new hdr for the buffer. 3602 */ 3603void 3604arc_release(arc_buf_t *buf, void *tag) 3605{ 3606 arc_buf_hdr_t *hdr; 3607 kmutex_t *hash_lock = NULL; 3608 l2arc_buf_hdr_t *l2hdr; 3609 uint64_t buf_size; 3610 3611 /* 3612 * It would be nice to assert that if it's DMU metadata (level > 3613 * 0 || it's the dnode file), then it must be syncing context. 3614 * But we don't know that information at this level. 3615 */ 3616 3617 mutex_enter(&buf->b_evict_lock); 3618 hdr = buf->b_hdr; 3619 3620 /* this buffer is not on any list */ 3621 ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3622 3623 if (hdr->b_state == arc_anon) { 3624 /* this buffer is already released */ 3625 ASSERT(buf->b_efunc == NULL); 3626 } else { 3627 hash_lock = HDR_LOCK(hdr); 3628 mutex_enter(hash_lock); 3629 hdr = buf->b_hdr; 3630 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3631 } 3632 3633 l2hdr = hdr->b_l2hdr; 3634 if (l2hdr) { 3635 mutex_enter(&l2arc_buflist_mtx); 3636 arc_buf_l2_cdata_free(hdr); 3637 hdr->b_l2hdr = NULL; 3638 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3639 } 3640 buf_size = hdr->b_size; 3641 3642 /* 3643 * Do we have more than one buf? 3644 */ 3645 if (hdr->b_datacnt > 1) { 3646 arc_buf_hdr_t *nhdr; 3647 arc_buf_t **bufp; 3648 uint64_t blksz = hdr->b_size; 3649 uint64_t spa = hdr->b_spa; 3650 arc_buf_contents_t type = hdr->b_type; 3651 uint32_t flags = hdr->b_flags; 3652 3653 ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3654 /* 3655 * Pull the data off of this hdr and attach it to 3656 * a new anonymous hdr. 3657 */ 3658 (void) remove_reference(hdr, hash_lock, tag); 3659 bufp = &hdr->b_buf; 3660 while (*bufp != buf) 3661 bufp = &(*bufp)->b_next; 3662 *bufp = buf->b_next; 3663 buf->b_next = NULL; 3664 3665 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3666 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3667 if (refcount_is_zero(&hdr->b_refcnt)) { 3668 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3669 ASSERT3U(*size, >=, hdr->b_size); 3670 atomic_add_64(size, -hdr->b_size); 3671 } 3672 3673 /* 3674 * We're releasing a duplicate user data buffer, update 3675 * our statistics accordingly. 3676 */ 3677 if (hdr->b_type == ARC_BUFC_DATA) { 3678 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3679 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3680 -hdr->b_size); 3681 } 3682 hdr->b_datacnt -= 1; 3683 arc_cksum_verify(buf); 3684#ifdef illumos 3685 arc_buf_unwatch(buf); 3686#endif /* illumos */ 3687 3688 mutex_exit(hash_lock); 3689 3690 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3691 nhdr->b_size = blksz; 3692 nhdr->b_spa = spa; 3693 nhdr->b_type = type; 3694 nhdr->b_buf = buf; 3695 nhdr->b_state = arc_anon; 3696 nhdr->b_arc_access = 0; 3697 nhdr->b_flags = flags & ARC_L2_WRITING; 3698 nhdr->b_l2hdr = NULL; 3699 nhdr->b_datacnt = 1; 3700 nhdr->b_freeze_cksum = NULL; 3701 (void) refcount_add(&nhdr->b_refcnt, tag); 3702 buf->b_hdr = nhdr; 3703 mutex_exit(&buf->b_evict_lock); 3704 atomic_add_64(&arc_anon->arcs_size, blksz); 3705 } else { 3706 mutex_exit(&buf->b_evict_lock); 3707 ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3708 ASSERT(!list_link_active(&hdr->b_arc_node)); 3709 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3710 if (hdr->b_state != arc_anon) 3711 arc_change_state(arc_anon, hdr, hash_lock); 3712 hdr->b_arc_access = 0; 3713 if (hash_lock) 3714 mutex_exit(hash_lock); 3715 3716 buf_discard_identity(hdr); 3717 arc_buf_thaw(buf); 3718 } 3719 buf->b_efunc = NULL; 3720 buf->b_private = NULL; 3721 3722 if (l2hdr) { 3723 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3724 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 3725 -l2hdr->b_asize, 0, 0); 3726 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3727 hdr->b_size, 0); 3728 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3729 ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3730 mutex_exit(&l2arc_buflist_mtx); 3731 } 3732} 3733 3734int 3735arc_released(arc_buf_t *buf) 3736{ 3737 int released; 3738 3739 mutex_enter(&buf->b_evict_lock); 3740 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3741 mutex_exit(&buf->b_evict_lock); 3742 return (released); 3743} 3744 3745#ifdef ZFS_DEBUG 3746int 3747arc_referenced(arc_buf_t *buf) 3748{ 3749 int referenced; 3750 3751 mutex_enter(&buf->b_evict_lock); 3752 referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3753 mutex_exit(&buf->b_evict_lock); 3754 return (referenced); 3755} 3756#endif 3757 3758static void 3759arc_write_ready(zio_t *zio) 3760{ 3761 arc_write_callback_t *callback = zio->io_private; 3762 arc_buf_t *buf = callback->awcb_buf; 3763 arc_buf_hdr_t *hdr = buf->b_hdr; 3764 3765 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3766 callback->awcb_ready(zio, buf, callback->awcb_private); 3767 3768 /* 3769 * If the IO is already in progress, then this is a re-write 3770 * attempt, so we need to thaw and re-compute the cksum. 3771 * It is the responsibility of the callback to handle the 3772 * accounting for any re-write attempt. 3773 */ 3774 if (HDR_IO_IN_PROGRESS(hdr)) { 3775 mutex_enter(&hdr->b_freeze_lock); 3776 if (hdr->b_freeze_cksum != NULL) { 3777 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3778 hdr->b_freeze_cksum = NULL; 3779 } 3780 mutex_exit(&hdr->b_freeze_lock); 3781 } 3782 arc_cksum_compute(buf, B_FALSE); 3783 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3784} 3785 3786/* 3787 * The SPA calls this callback for each physical write that happens on behalf 3788 * of a logical write. See the comment in dbuf_write_physdone() for details. 3789 */ 3790static void 3791arc_write_physdone(zio_t *zio) 3792{ 3793 arc_write_callback_t *cb = zio->io_private; 3794 if (cb->awcb_physdone != NULL) 3795 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3796} 3797 3798static void 3799arc_write_done(zio_t *zio) 3800{ 3801 arc_write_callback_t *callback = zio->io_private; 3802 arc_buf_t *buf = callback->awcb_buf; 3803 arc_buf_hdr_t *hdr = buf->b_hdr; 3804 3805 ASSERT(hdr->b_acb == NULL); 3806 3807 if (zio->io_error == 0) { 3808 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 3809 buf_discard_identity(hdr); 3810 } else { 3811 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3812 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3813 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3814 } 3815 } else { 3816 ASSERT(BUF_EMPTY(hdr)); 3817 } 3818 3819 /* 3820 * If the block to be written was all-zero or compressed enough to be 3821 * embedded in the BP, no write was performed so there will be no 3822 * dva/birth/checksum. The buffer must therefore remain anonymous 3823 * (and uncached). 3824 */ 3825 if (!BUF_EMPTY(hdr)) { 3826 arc_buf_hdr_t *exists; 3827 kmutex_t *hash_lock; 3828 3829 ASSERT(zio->io_error == 0); 3830 3831 arc_cksum_verify(buf); 3832 3833 exists = buf_hash_insert(hdr, &hash_lock); 3834 if (exists) { 3835 /* 3836 * This can only happen if we overwrite for 3837 * sync-to-convergence, because we remove 3838 * buffers from the hash table when we arc_free(). 3839 */ 3840 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3841 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3842 panic("bad overwrite, hdr=%p exists=%p", 3843 (void *)hdr, (void *)exists); 3844 ASSERT(refcount_is_zero(&exists->b_refcnt)); 3845 arc_change_state(arc_anon, exists, hash_lock); 3846 mutex_exit(hash_lock); 3847 arc_hdr_destroy(exists); 3848 exists = buf_hash_insert(hdr, &hash_lock); 3849 ASSERT3P(exists, ==, NULL); 3850 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3851 /* nopwrite */ 3852 ASSERT(zio->io_prop.zp_nopwrite); 3853 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3854 panic("bad nopwrite, hdr=%p exists=%p", 3855 (void *)hdr, (void *)exists); 3856 } else { 3857 /* Dedup */ 3858 ASSERT(hdr->b_datacnt == 1); 3859 ASSERT(hdr->b_state == arc_anon); 3860 ASSERT(BP_GET_DEDUP(zio->io_bp)); 3861 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3862 } 3863 } 3864 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3865 /* if it's not anon, we are doing a scrub */ 3866 if (!exists && hdr->b_state == arc_anon) 3867 arc_access(hdr, hash_lock); 3868 mutex_exit(hash_lock); 3869 } else { 3870 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3871 } 3872 3873 ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3874 callback->awcb_done(zio, buf, callback->awcb_private); 3875 3876 kmem_free(callback, sizeof (arc_write_callback_t)); 3877} 3878 3879zio_t * 3880arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 3881 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 3882 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 3883 arc_done_func_t *done, void *private, zio_priority_t priority, 3884 int zio_flags, const zbookmark_phys_t *zb) 3885{ 3886 arc_buf_hdr_t *hdr = buf->b_hdr; 3887 arc_write_callback_t *callback; 3888 zio_t *zio; 3889 3890 ASSERT(ready != NULL); 3891 ASSERT(done != NULL); 3892 ASSERT(!HDR_IO_ERROR(hdr)); 3893 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3894 ASSERT(hdr->b_acb == NULL); 3895 if (l2arc) 3896 hdr->b_flags |= ARC_L2CACHE; 3897 if (l2arc_compress) 3898 hdr->b_flags |= ARC_L2COMPRESS; 3899 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 3900 callback->awcb_ready = ready; 3901 callback->awcb_physdone = physdone; 3902 callback->awcb_done = done; 3903 callback->awcb_private = private; 3904 callback->awcb_buf = buf; 3905 3906 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 3907 arc_write_ready, arc_write_physdone, arc_write_done, callback, 3908 priority, zio_flags, zb); 3909 3910 return (zio); 3911} 3912 3913static int 3914arc_memory_throttle(uint64_t reserve, uint64_t txg) 3915{ 3916#ifdef _KERNEL 3917 uint64_t available_memory = 3918 ptoa((uintmax_t)cnt.v_free_count + cnt.v_cache_count); 3919 static uint64_t page_load = 0; 3920 static uint64_t last_txg = 0; 3921 3922#ifdef sun 3923#if defined(__i386) 3924 available_memory = 3925 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 3926#endif 3927#endif /* sun */ 3928 3929 if (cnt.v_free_count + cnt.v_cache_count > 3930 (uint64_t)physmem * arc_lotsfree_percent / 100) 3931 return (0); 3932 3933 if (txg > last_txg) { 3934 last_txg = txg; 3935 page_load = 0; 3936 } 3937 /* 3938 * If we are in pageout, we know that memory is already tight, 3939 * the arc is already going to be evicting, so we just want to 3940 * continue to let page writes occur as quickly as possible. 3941 */ 3942 if (curproc == pageproc) { 3943 if (page_load > available_memory / 4) 3944 return (SET_ERROR(ERESTART)); 3945 /* Note: reserve is inflated, so we deflate */ 3946 page_load += reserve / 8; 3947 return (0); 3948 } else if (page_load > 0 && arc_reclaim_needed()) { 3949 /* memory is low, delay before restarting */ 3950 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3951 return (SET_ERROR(EAGAIN)); 3952 } 3953 page_load = 0; 3954#endif 3955 return (0); 3956} 3957 3958void 3959arc_tempreserve_clear(uint64_t reserve) 3960{ 3961 atomic_add_64(&arc_tempreserve, -reserve); 3962 ASSERT((int64_t)arc_tempreserve >= 0); 3963} 3964 3965int 3966arc_tempreserve_space(uint64_t reserve, uint64_t txg) 3967{ 3968 int error; 3969 uint64_t anon_size; 3970 3971 if (reserve > arc_c/4 && !arc_no_grow) 3972 arc_c = MIN(arc_c_max, reserve * 4); 3973 if (reserve > arc_c) 3974 return (SET_ERROR(ENOMEM)); 3975 3976 /* 3977 * Don't count loaned bufs as in flight dirty data to prevent long 3978 * network delays from blocking transactions that are ready to be 3979 * assigned to a txg. 3980 */ 3981 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 3982 3983 /* 3984 * Writes will, almost always, require additional memory allocations 3985 * in order to compress/encrypt/etc the data. We therefore need to 3986 * make sure that there is sufficient available memory for this. 3987 */ 3988 error = arc_memory_throttle(reserve, txg); 3989 if (error != 0) 3990 return (error); 3991 3992 /* 3993 * Throttle writes when the amount of dirty data in the cache 3994 * gets too large. We try to keep the cache less than half full 3995 * of dirty blocks so that our sync times don't grow too large. 3996 * Note: if two requests come in concurrently, we might let them 3997 * both succeed, when one of them should fail. Not a huge deal. 3998 */ 3999 4000 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 4001 anon_size > arc_c / 4) { 4002 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 4003 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 4004 arc_tempreserve>>10, 4005 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 4006 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 4007 reserve>>10, arc_c>>10); 4008 return (SET_ERROR(ERESTART)); 4009 } 4010 atomic_add_64(&arc_tempreserve, reserve); 4011 return (0); 4012} 4013 4014static kmutex_t arc_lowmem_lock; 4015#ifdef _KERNEL 4016static eventhandler_tag arc_event_lowmem = NULL; 4017 4018static void 4019arc_lowmem(void *arg __unused, int howto __unused) 4020{ 4021 4022 /* Serialize access via arc_lowmem_lock. */ 4023 mutex_enter(&arc_lowmem_lock); 4024 mutex_enter(&arc_reclaim_thr_lock); 4025 needfree = 1; 4026 cv_signal(&arc_reclaim_thr_cv); 4027 4028 /* 4029 * It is unsafe to block here in arbitrary threads, because we can come 4030 * here from ARC itself and may hold ARC locks and thus risk a deadlock 4031 * with ARC reclaim thread. 4032 */ 4033 if (curproc == pageproc) { 4034 while (needfree) 4035 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4036 } 4037 mutex_exit(&arc_reclaim_thr_lock); 4038 mutex_exit(&arc_lowmem_lock); 4039} 4040#endif 4041 4042void 4043arc_init(void) 4044{ 4045 int i, prefetch_tunable_set = 0; 4046 4047 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4048 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4049 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4050 4051 /* Convert seconds to clock ticks */ 4052 arc_min_prefetch_lifespan = 1 * hz; 4053 4054 /* Start out with 1/8 of all memory */ 4055 arc_c = kmem_size() / 8; 4056 4057#ifdef sun 4058#ifdef _KERNEL 4059 /* 4060 * On architectures where the physical memory can be larger 4061 * than the addressable space (intel in 32-bit mode), we may 4062 * need to limit the cache to 1/8 of VM size. 4063 */ 4064 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4065#endif 4066#endif /* sun */ 4067 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4068 arc_c_min = MAX(arc_c / 4, 64<<18); 4069 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4070 if (arc_c * 8 >= 1<<30) 4071 arc_c_max = (arc_c * 8) - (1<<30); 4072 else 4073 arc_c_max = arc_c_min; 4074 arc_c_max = MAX(arc_c * 5, arc_c_max); 4075 4076#ifdef _KERNEL 4077 /* 4078 * Allow the tunables to override our calculations if they are 4079 * reasonable (ie. over 16MB) 4080 */ 4081 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size()) 4082 arc_c_max = zfs_arc_max; 4083 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max) 4084 arc_c_min = zfs_arc_min; 4085#endif 4086 4087 arc_c = arc_c_max; 4088 arc_p = (arc_c >> 1); 4089 4090 /* limit meta-data to 1/4 of the arc capacity */ 4091 arc_meta_limit = arc_c_max / 4; 4092 4093 /* Allow the tunable to override if it is reasonable */ 4094 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4095 arc_meta_limit = zfs_arc_meta_limit; 4096 4097 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4098 arc_c_min = arc_meta_limit / 2; 4099 4100 if (zfs_arc_grow_retry > 0) 4101 arc_grow_retry = zfs_arc_grow_retry; 4102 4103 if (zfs_arc_shrink_shift > 0) 4104 arc_shrink_shift = zfs_arc_shrink_shift; 4105 4106 if (zfs_arc_p_min_shift > 0) 4107 arc_p_min_shift = zfs_arc_p_min_shift; 4108 4109 /* if kmem_flags are set, lets try to use less memory */ 4110 if (kmem_debugging()) 4111 arc_c = arc_c / 2; 4112 if (arc_c < arc_c_min) 4113 arc_c = arc_c_min; 4114 4115 zfs_arc_min = arc_c_min; 4116 zfs_arc_max = arc_c_max; 4117 4118 arc_anon = &ARC_anon; 4119 arc_mru = &ARC_mru; 4120 arc_mru_ghost = &ARC_mru_ghost; 4121 arc_mfu = &ARC_mfu; 4122 arc_mfu_ghost = &ARC_mfu_ghost; 4123 arc_l2c_only = &ARC_l2c_only; 4124 arc_size = 0; 4125 4126 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4127 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4128 NULL, MUTEX_DEFAULT, NULL); 4129 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4130 NULL, MUTEX_DEFAULT, NULL); 4131 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4132 NULL, MUTEX_DEFAULT, NULL); 4133 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4134 NULL, MUTEX_DEFAULT, NULL); 4135 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4136 NULL, MUTEX_DEFAULT, NULL); 4137 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4138 NULL, MUTEX_DEFAULT, NULL); 4139 4140 list_create(&arc_mru->arcs_lists[i], 4141 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4142 list_create(&arc_mru_ghost->arcs_lists[i], 4143 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4144 list_create(&arc_mfu->arcs_lists[i], 4145 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4146 list_create(&arc_mfu_ghost->arcs_lists[i], 4147 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4148 list_create(&arc_mfu_ghost->arcs_lists[i], 4149 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4150 list_create(&arc_l2c_only->arcs_lists[i], 4151 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4152 } 4153 4154 buf_init(); 4155 4156 arc_thread_exit = 0; 4157 arc_eviction_list = NULL; 4158 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4159 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4160 4161 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4162 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4163 4164 if (arc_ksp != NULL) { 4165 arc_ksp->ks_data = &arc_stats; 4166 kstat_install(arc_ksp); 4167 } 4168 4169 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4170 TS_RUN, minclsyspri); 4171 4172#ifdef _KERNEL 4173 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4174 EVENTHANDLER_PRI_FIRST); 4175#endif 4176 4177 arc_dead = FALSE; 4178 arc_warm = B_FALSE; 4179 4180 /* 4181 * Calculate maximum amount of dirty data per pool. 4182 * 4183 * If it has been set by /etc/system, take that. 4184 * Otherwise, use a percentage of physical memory defined by 4185 * zfs_dirty_data_max_percent (default 10%) with a cap at 4186 * zfs_dirty_data_max_max (default 4GB). 4187 */ 4188 if (zfs_dirty_data_max == 0) { 4189 zfs_dirty_data_max = ptob(physmem) * 4190 zfs_dirty_data_max_percent / 100; 4191 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4192 zfs_dirty_data_max_max); 4193 } 4194 4195#ifdef _KERNEL 4196 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4197 prefetch_tunable_set = 1; 4198 4199#ifdef __i386__ 4200 if (prefetch_tunable_set == 0) { 4201 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4202 "-- to enable,\n"); 4203 printf(" add \"vfs.zfs.prefetch_disable=0\" " 4204 "to /boot/loader.conf.\n"); 4205 zfs_prefetch_disable = 1; 4206 } 4207#else 4208 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4209 prefetch_tunable_set == 0) { 4210 printf("ZFS NOTICE: Prefetch is disabled by default if less " 4211 "than 4GB of RAM is present;\n" 4212 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4213 "to /boot/loader.conf.\n"); 4214 zfs_prefetch_disable = 1; 4215 } 4216#endif 4217 /* Warn about ZFS memory and address space requirements. */ 4218 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4219 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4220 "expect unstable behavior.\n"); 4221 } 4222 if (kmem_size() < 512 * (1 << 20)) { 4223 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4224 "expect unstable behavior.\n"); 4225 printf(" Consider tuning vm.kmem_size and " 4226 "vm.kmem_size_max\n"); 4227 printf(" in /boot/loader.conf.\n"); 4228 } 4229#endif 4230} 4231 4232void 4233arc_fini(void) 4234{ 4235 int i; 4236 4237 mutex_enter(&arc_reclaim_thr_lock); 4238 arc_thread_exit = 1; 4239 cv_signal(&arc_reclaim_thr_cv); 4240 while (arc_thread_exit != 0) 4241 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4242 mutex_exit(&arc_reclaim_thr_lock); 4243 4244 arc_flush(NULL); 4245 4246 arc_dead = TRUE; 4247 4248 if (arc_ksp != NULL) { 4249 kstat_delete(arc_ksp); 4250 arc_ksp = NULL; 4251 } 4252 4253 mutex_destroy(&arc_eviction_mtx); 4254 mutex_destroy(&arc_reclaim_thr_lock); 4255 cv_destroy(&arc_reclaim_thr_cv); 4256 4257 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4258 list_destroy(&arc_mru->arcs_lists[i]); 4259 list_destroy(&arc_mru_ghost->arcs_lists[i]); 4260 list_destroy(&arc_mfu->arcs_lists[i]); 4261 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4262 list_destroy(&arc_l2c_only->arcs_lists[i]); 4263 4264 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4265 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4266 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4267 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4268 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4269 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4270 } 4271 4272 buf_fini(); 4273 4274 ASSERT(arc_loaned_bytes == 0); 4275 4276 mutex_destroy(&arc_lowmem_lock); 4277#ifdef _KERNEL 4278 if (arc_event_lowmem != NULL) 4279 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4280#endif 4281} 4282 4283/* 4284 * Level 2 ARC 4285 * 4286 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4287 * It uses dedicated storage devices to hold cached data, which are populated 4288 * using large infrequent writes. The main role of this cache is to boost 4289 * the performance of random read workloads. The intended L2ARC devices 4290 * include short-stroked disks, solid state disks, and other media with 4291 * substantially faster read latency than disk. 4292 * 4293 * +-----------------------+ 4294 * | ARC | 4295 * +-----------------------+ 4296 * | ^ ^ 4297 * | | | 4298 * l2arc_feed_thread() arc_read() 4299 * | | | 4300 * | l2arc read | 4301 * V | | 4302 * +---------------+ | 4303 * | L2ARC | | 4304 * +---------------+ | 4305 * | ^ | 4306 * l2arc_write() | | 4307 * | | | 4308 * V | | 4309 * +-------+ +-------+ 4310 * | vdev | | vdev | 4311 * | cache | | cache | 4312 * +-------+ +-------+ 4313 * +=========+ .-----. 4314 * : L2ARC : |-_____-| 4315 * : devices : | Disks | 4316 * +=========+ `-_____-' 4317 * 4318 * Read requests are satisfied from the following sources, in order: 4319 * 4320 * 1) ARC 4321 * 2) vdev cache of L2ARC devices 4322 * 3) L2ARC devices 4323 * 4) vdev cache of disks 4324 * 5) disks 4325 * 4326 * Some L2ARC device types exhibit extremely slow write performance. 4327 * To accommodate for this there are some significant differences between 4328 * the L2ARC and traditional cache design: 4329 * 4330 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4331 * the ARC behave as usual, freeing buffers and placing headers on ghost 4332 * lists. The ARC does not send buffers to the L2ARC during eviction as 4333 * this would add inflated write latencies for all ARC memory pressure. 4334 * 4335 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4336 * It does this by periodically scanning buffers from the eviction-end of 4337 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4338 * not already there. It scans until a headroom of buffers is satisfied, 4339 * which itself is a buffer for ARC eviction. If a compressible buffer is 4340 * found during scanning and selected for writing to an L2ARC device, we 4341 * temporarily boost scanning headroom during the next scan cycle to make 4342 * sure we adapt to compression effects (which might significantly reduce 4343 * the data volume we write to L2ARC). The thread that does this is 4344 * l2arc_feed_thread(), illustrated below; example sizes are included to 4345 * provide a better sense of ratio than this diagram: 4346 * 4347 * head --> tail 4348 * +---------------------+----------+ 4349 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4350 * +---------------------+----------+ | o L2ARC eligible 4351 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4352 * +---------------------+----------+ | 4353 * 15.9 Gbytes ^ 32 Mbytes | 4354 * headroom | 4355 * l2arc_feed_thread() 4356 * | 4357 * l2arc write hand <--[oooo]--' 4358 * | 8 Mbyte 4359 * | write max 4360 * V 4361 * +==============================+ 4362 * L2ARC dev |####|#|###|###| |####| ... | 4363 * +==============================+ 4364 * 32 Gbytes 4365 * 4366 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4367 * evicted, then the L2ARC has cached a buffer much sooner than it probably 4368 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4369 * safe to say that this is an uncommon case, since buffers at the end of 4370 * the ARC lists have moved there due to inactivity. 4371 * 4372 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4373 * then the L2ARC simply misses copying some buffers. This serves as a 4374 * pressure valve to prevent heavy read workloads from both stalling the ARC 4375 * with waits and clogging the L2ARC with writes. This also helps prevent 4376 * the potential for the L2ARC to churn if it attempts to cache content too 4377 * quickly, such as during backups of the entire pool. 4378 * 4379 * 5. After system boot and before the ARC has filled main memory, there are 4380 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4381 * lists can remain mostly static. Instead of searching from tail of these 4382 * lists as pictured, the l2arc_feed_thread() will search from the list heads 4383 * for eligible buffers, greatly increasing its chance of finding them. 4384 * 4385 * The L2ARC device write speed is also boosted during this time so that 4386 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4387 * there are no L2ARC reads, and no fear of degrading read performance 4388 * through increased writes. 4389 * 4390 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4391 * the vdev queue can aggregate them into larger and fewer writes. Each 4392 * device is written to in a rotor fashion, sweeping writes through 4393 * available space then repeating. 4394 * 4395 * 7. The L2ARC does not store dirty content. It never needs to flush 4396 * write buffers back to disk based storage. 4397 * 4398 * 8. If an ARC buffer is written (and dirtied) which also exists in the 4399 * L2ARC, the now stale L2ARC buffer is immediately dropped. 4400 * 4401 * The performance of the L2ARC can be tweaked by a number of tunables, which 4402 * may be necessary for different workloads: 4403 * 4404 * l2arc_write_max max write bytes per interval 4405 * l2arc_write_boost extra write bytes during device warmup 4406 * l2arc_noprefetch skip caching prefetched buffers 4407 * l2arc_headroom number of max device writes to precache 4408 * l2arc_headroom_boost when we find compressed buffers during ARC 4409 * scanning, we multiply headroom by this 4410 * percentage factor for the next scan cycle, 4411 * since more compressed buffers are likely to 4412 * be present 4413 * l2arc_feed_secs seconds between L2ARC writing 4414 * 4415 * Tunables may be removed or added as future performance improvements are 4416 * integrated, and also may become zpool properties. 4417 * 4418 * There are three key functions that control how the L2ARC warms up: 4419 * 4420 * l2arc_write_eligible() check if a buffer is eligible to cache 4421 * l2arc_write_size() calculate how much to write 4422 * l2arc_write_interval() calculate sleep delay between writes 4423 * 4424 * These three functions determine what to write, how much, and how quickly 4425 * to send writes. 4426 */ 4427 4428static boolean_t 4429l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab) 4430{ 4431 /* 4432 * A buffer is *not* eligible for the L2ARC if it: 4433 * 1. belongs to a different spa. 4434 * 2. is already cached on the L2ARC. 4435 * 3. has an I/O in progress (it may be an incomplete read). 4436 * 4. is flagged not eligible (zfs property). 4437 */ 4438 if (ab->b_spa != spa_guid) { 4439 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4440 return (B_FALSE); 4441 } 4442 if (ab->b_l2hdr != NULL) { 4443 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4444 return (B_FALSE); 4445 } 4446 if (HDR_IO_IN_PROGRESS(ab)) { 4447 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4448 return (B_FALSE); 4449 } 4450 if (!HDR_L2CACHE(ab)) { 4451 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4452 return (B_FALSE); 4453 } 4454 4455 return (B_TRUE); 4456} 4457 4458static uint64_t 4459l2arc_write_size(void) 4460{ 4461 uint64_t size; 4462 4463 /* 4464 * Make sure our globals have meaningful values in case the user 4465 * altered them. 4466 */ 4467 size = l2arc_write_max; 4468 if (size == 0) { 4469 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4470 "be greater than zero, resetting it to the default (%d)", 4471 L2ARC_WRITE_SIZE); 4472 size = l2arc_write_max = L2ARC_WRITE_SIZE; 4473 } 4474 4475 if (arc_warm == B_FALSE) 4476 size += l2arc_write_boost; 4477 4478 return (size); 4479 4480} 4481 4482static clock_t 4483l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4484{ 4485 clock_t interval, next, now; 4486 4487 /* 4488 * If the ARC lists are busy, increase our write rate; if the 4489 * lists are stale, idle back. This is achieved by checking 4490 * how much we previously wrote - if it was more than half of 4491 * what we wanted, schedule the next write much sooner. 4492 */ 4493 if (l2arc_feed_again && wrote > (wanted / 2)) 4494 interval = (hz * l2arc_feed_min_ms) / 1000; 4495 else 4496 interval = hz * l2arc_feed_secs; 4497 4498 now = ddi_get_lbolt(); 4499 next = MAX(now, MIN(now + interval, began + interval)); 4500 4501 return (next); 4502} 4503 4504static void 4505l2arc_hdr_stat_add(void) 4506{ 4507 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4508 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4509} 4510 4511static void 4512l2arc_hdr_stat_remove(void) 4513{ 4514 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4515 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4516} 4517 4518/* 4519 * Cycle through L2ARC devices. This is how L2ARC load balances. 4520 * If a device is returned, this also returns holding the spa config lock. 4521 */ 4522static l2arc_dev_t * 4523l2arc_dev_get_next(void) 4524{ 4525 l2arc_dev_t *first, *next = NULL; 4526 4527 /* 4528 * Lock out the removal of spas (spa_namespace_lock), then removal 4529 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4530 * both locks will be dropped and a spa config lock held instead. 4531 */ 4532 mutex_enter(&spa_namespace_lock); 4533 mutex_enter(&l2arc_dev_mtx); 4534 4535 /* if there are no vdevs, there is nothing to do */ 4536 if (l2arc_ndev == 0) 4537 goto out; 4538 4539 first = NULL; 4540 next = l2arc_dev_last; 4541 do { 4542 /* loop around the list looking for a non-faulted vdev */ 4543 if (next == NULL) { 4544 next = list_head(l2arc_dev_list); 4545 } else { 4546 next = list_next(l2arc_dev_list, next); 4547 if (next == NULL) 4548 next = list_head(l2arc_dev_list); 4549 } 4550 4551 /* if we have come back to the start, bail out */ 4552 if (first == NULL) 4553 first = next; 4554 else if (next == first) 4555 break; 4556 4557 } while (vdev_is_dead(next->l2ad_vdev)); 4558 4559 /* if we were unable to find any usable vdevs, return NULL */ 4560 if (vdev_is_dead(next->l2ad_vdev)) 4561 next = NULL; 4562 4563 l2arc_dev_last = next; 4564 4565out: 4566 mutex_exit(&l2arc_dev_mtx); 4567 4568 /* 4569 * Grab the config lock to prevent the 'next' device from being 4570 * removed while we are writing to it. 4571 */ 4572 if (next != NULL) 4573 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4574 mutex_exit(&spa_namespace_lock); 4575 4576 return (next); 4577} 4578 4579/* 4580 * Free buffers that were tagged for destruction. 4581 */ 4582static void 4583l2arc_do_free_on_write() 4584{ 4585 list_t *buflist; 4586 l2arc_data_free_t *df, *df_prev; 4587 4588 mutex_enter(&l2arc_free_on_write_mtx); 4589 buflist = l2arc_free_on_write; 4590 4591 for (df = list_tail(buflist); df; df = df_prev) { 4592 df_prev = list_prev(buflist, df); 4593 ASSERT(df->l2df_data != NULL); 4594 ASSERT(df->l2df_func != NULL); 4595 df->l2df_func(df->l2df_data, df->l2df_size); 4596 list_remove(buflist, df); 4597 kmem_free(df, sizeof (l2arc_data_free_t)); 4598 } 4599 4600 mutex_exit(&l2arc_free_on_write_mtx); 4601} 4602 4603/* 4604 * A write to a cache device has completed. Update all headers to allow 4605 * reads from these buffers to begin. 4606 */ 4607static void 4608l2arc_write_done(zio_t *zio) 4609{ 4610 l2arc_write_callback_t *cb; 4611 l2arc_dev_t *dev; 4612 list_t *buflist; 4613 arc_buf_hdr_t *head, *ab, *ab_prev; 4614 l2arc_buf_hdr_t *abl2; 4615 kmutex_t *hash_lock; 4616 int64_t bytes_dropped = 0; 4617 4618 cb = zio->io_private; 4619 ASSERT(cb != NULL); 4620 dev = cb->l2wcb_dev; 4621 ASSERT(dev != NULL); 4622 head = cb->l2wcb_head; 4623 ASSERT(head != NULL); 4624 buflist = dev->l2ad_buflist; 4625 ASSERT(buflist != NULL); 4626 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4627 l2arc_write_callback_t *, cb); 4628 4629 if (zio->io_error != 0) 4630 ARCSTAT_BUMP(arcstat_l2_writes_error); 4631 4632 mutex_enter(&l2arc_buflist_mtx); 4633 4634 /* 4635 * All writes completed, or an error was hit. 4636 */ 4637 for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 4638 ab_prev = list_prev(buflist, ab); 4639 abl2 = ab->b_l2hdr; 4640 4641 /* 4642 * Release the temporary compressed buffer as soon as possible. 4643 */ 4644 if (abl2->b_compress != ZIO_COMPRESS_OFF) 4645 l2arc_release_cdata_buf(ab); 4646 4647 hash_lock = HDR_LOCK(ab); 4648 if (!mutex_tryenter(hash_lock)) { 4649 /* 4650 * This buffer misses out. It may be in a stage 4651 * of eviction. Its ARC_L2_WRITING flag will be 4652 * left set, denying reads to this buffer. 4653 */ 4654 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4655 continue; 4656 } 4657 4658 if (zio->io_error != 0) { 4659 /* 4660 * Error - drop L2ARC entry. 4661 */ 4662 list_remove(buflist, ab); 4663 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4664 bytes_dropped += abl2->b_asize; 4665 ab->b_l2hdr = NULL; 4666 trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4667 ab->b_size, 0); 4668 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4669 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4670 } 4671 4672 /* 4673 * Allow ARC to begin reads to this L2ARC entry. 4674 */ 4675 ab->b_flags &= ~ARC_L2_WRITING; 4676 4677 mutex_exit(hash_lock); 4678 } 4679 4680 atomic_inc_64(&l2arc_writes_done); 4681 list_remove(buflist, head); 4682 kmem_cache_free(hdr_cache, head); 4683 mutex_exit(&l2arc_buflist_mtx); 4684 4685 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 4686 4687 l2arc_do_free_on_write(); 4688 4689 kmem_free(cb, sizeof (l2arc_write_callback_t)); 4690} 4691 4692/* 4693 * A read to a cache device completed. Validate buffer contents before 4694 * handing over to the regular ARC routines. 4695 */ 4696static void 4697l2arc_read_done(zio_t *zio) 4698{ 4699 l2arc_read_callback_t *cb; 4700 arc_buf_hdr_t *hdr; 4701 arc_buf_t *buf; 4702 kmutex_t *hash_lock; 4703 int equal; 4704 4705 ASSERT(zio->io_vd != NULL); 4706 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4707 4708 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4709 4710 cb = zio->io_private; 4711 ASSERT(cb != NULL); 4712 buf = cb->l2rcb_buf; 4713 ASSERT(buf != NULL); 4714 4715 hash_lock = HDR_LOCK(buf->b_hdr); 4716 mutex_enter(hash_lock); 4717 hdr = buf->b_hdr; 4718 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4719 4720 /* 4721 * If the buffer was compressed, decompress it first. 4722 */ 4723 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4724 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4725 ASSERT(zio->io_data != NULL); 4726 4727 /* 4728 * Check this survived the L2ARC journey. 4729 */ 4730 equal = arc_cksum_equal(buf); 4731 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4732 mutex_exit(hash_lock); 4733 zio->io_private = buf; 4734 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4735 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4736 arc_read_done(zio); 4737 } else { 4738 mutex_exit(hash_lock); 4739 /* 4740 * Buffer didn't survive caching. Increment stats and 4741 * reissue to the original storage device. 4742 */ 4743 if (zio->io_error != 0) { 4744 ARCSTAT_BUMP(arcstat_l2_io_error); 4745 } else { 4746 zio->io_error = SET_ERROR(EIO); 4747 } 4748 if (!equal) 4749 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4750 4751 /* 4752 * If there's no waiter, issue an async i/o to the primary 4753 * storage now. If there *is* a waiter, the caller must 4754 * issue the i/o in a context where it's OK to block. 4755 */ 4756 if (zio->io_waiter == NULL) { 4757 zio_t *pio = zio_unique_parent(zio); 4758 4759 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4760 4761 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4762 buf->b_data, zio->io_size, arc_read_done, buf, 4763 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4764 } 4765 } 4766 4767 kmem_free(cb, sizeof (l2arc_read_callback_t)); 4768} 4769 4770/* 4771 * This is the list priority from which the L2ARC will search for pages to 4772 * cache. This is used within loops (0..3) to cycle through lists in the 4773 * desired order. This order can have a significant effect on cache 4774 * performance. 4775 * 4776 * Currently the metadata lists are hit first, MFU then MRU, followed by 4777 * the data lists. This function returns a locked list, and also returns 4778 * the lock pointer. 4779 */ 4780static list_t * 4781l2arc_list_locked(int list_num, kmutex_t **lock) 4782{ 4783 list_t *list = NULL; 4784 int idx; 4785 4786 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4787 4788 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4789 idx = list_num; 4790 list = &arc_mfu->arcs_lists[idx]; 4791 *lock = ARCS_LOCK(arc_mfu, idx); 4792 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4793 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4794 list = &arc_mru->arcs_lists[idx]; 4795 *lock = ARCS_LOCK(arc_mru, idx); 4796 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4797 ARC_BUFC_NUMDATALISTS)) { 4798 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4799 list = &arc_mfu->arcs_lists[idx]; 4800 *lock = ARCS_LOCK(arc_mfu, idx); 4801 } else { 4802 idx = list_num - ARC_BUFC_NUMLISTS; 4803 list = &arc_mru->arcs_lists[idx]; 4804 *lock = ARCS_LOCK(arc_mru, idx); 4805 } 4806 4807 ASSERT(!(MUTEX_HELD(*lock))); 4808 mutex_enter(*lock); 4809 return (list); 4810} 4811 4812/* 4813 * Evict buffers from the device write hand to the distance specified in 4814 * bytes. This distance may span populated buffers, it may span nothing. 4815 * This is clearing a region on the L2ARC device ready for writing. 4816 * If the 'all' boolean is set, every buffer is evicted. 4817 */ 4818static void 4819l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4820{ 4821 list_t *buflist; 4822 l2arc_buf_hdr_t *abl2; 4823 arc_buf_hdr_t *ab, *ab_prev; 4824 kmutex_t *hash_lock; 4825 uint64_t taddr; 4826 int64_t bytes_evicted = 0; 4827 4828 buflist = dev->l2ad_buflist; 4829 4830 if (buflist == NULL) 4831 return; 4832 4833 if (!all && dev->l2ad_first) { 4834 /* 4835 * This is the first sweep through the device. There is 4836 * nothing to evict. 4837 */ 4838 return; 4839 } 4840 4841 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4842 /* 4843 * When nearing the end of the device, evict to the end 4844 * before the device write hand jumps to the start. 4845 */ 4846 taddr = dev->l2ad_end; 4847 } else { 4848 taddr = dev->l2ad_hand + distance; 4849 } 4850 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4851 uint64_t, taddr, boolean_t, all); 4852 4853top: 4854 mutex_enter(&l2arc_buflist_mtx); 4855 for (ab = list_tail(buflist); ab; ab = ab_prev) { 4856 ab_prev = list_prev(buflist, ab); 4857 4858 hash_lock = HDR_LOCK(ab); 4859 if (!mutex_tryenter(hash_lock)) { 4860 /* 4861 * Missed the hash lock. Retry. 4862 */ 4863 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4864 mutex_exit(&l2arc_buflist_mtx); 4865 mutex_enter(hash_lock); 4866 mutex_exit(hash_lock); 4867 goto top; 4868 } 4869 4870 if (HDR_L2_WRITE_HEAD(ab)) { 4871 /* 4872 * We hit a write head node. Leave it for 4873 * l2arc_write_done(). 4874 */ 4875 list_remove(buflist, ab); 4876 mutex_exit(hash_lock); 4877 continue; 4878 } 4879 4880 if (!all && ab->b_l2hdr != NULL && 4881 (ab->b_l2hdr->b_daddr > taddr || 4882 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4883 /* 4884 * We've evicted to the target address, 4885 * or the end of the device. 4886 */ 4887 mutex_exit(hash_lock); 4888 break; 4889 } 4890 4891 if (HDR_FREE_IN_PROGRESS(ab)) { 4892 /* 4893 * Already on the path to destruction. 4894 */ 4895 mutex_exit(hash_lock); 4896 continue; 4897 } 4898 4899 if (ab->b_state == arc_l2c_only) { 4900 ASSERT(!HDR_L2_READING(ab)); 4901 /* 4902 * This doesn't exist in the ARC. Destroy. 4903 * arc_hdr_destroy() will call list_remove() 4904 * and decrement arcstat_l2_size. 4905 */ 4906 arc_change_state(arc_anon, ab, hash_lock); 4907 arc_hdr_destroy(ab); 4908 } else { 4909 /* 4910 * Invalidate issued or about to be issued 4911 * reads, since we may be about to write 4912 * over this location. 4913 */ 4914 if (HDR_L2_READING(ab)) { 4915 ARCSTAT_BUMP(arcstat_l2_evict_reading); 4916 ab->b_flags |= ARC_L2_EVICTED; 4917 } 4918 4919 /* 4920 * Tell ARC this no longer exists in L2ARC. 4921 */ 4922 if (ab->b_l2hdr != NULL) { 4923 abl2 = ab->b_l2hdr; 4924 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4925 bytes_evicted += abl2->b_asize; 4926 ab->b_l2hdr = NULL; 4927 /* 4928 * We are destroying l2hdr, so ensure that 4929 * its compressed buffer, if any, is not leaked. 4930 */ 4931 ASSERT(abl2->b_tmp_cdata == NULL); 4932 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4933 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4934 } 4935 list_remove(buflist, ab); 4936 4937 /* 4938 * This may have been leftover after a 4939 * failed write. 4940 */ 4941 ab->b_flags &= ~ARC_L2_WRITING; 4942 } 4943 mutex_exit(hash_lock); 4944 } 4945 mutex_exit(&l2arc_buflist_mtx); 4946 4947 vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); 4948 dev->l2ad_evict = taddr; 4949} 4950 4951/* 4952 * Find and write ARC buffers to the L2ARC device. 4953 * 4954 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 4955 * for reading until they have completed writing. 4956 * The headroom_boost is an in-out parameter used to maintain headroom boost 4957 * state between calls to this function. 4958 * 4959 * Returns the number of bytes actually written (which may be smaller than 4960 * the delta by which the device hand has changed due to alignment). 4961 */ 4962static uint64_t 4963l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 4964 boolean_t *headroom_boost) 4965{ 4966 arc_buf_hdr_t *ab, *ab_prev, *head; 4967 list_t *list; 4968 uint64_t write_asize, write_psize, write_sz, headroom, 4969 buf_compress_minsz; 4970 void *buf_data; 4971 kmutex_t *list_lock; 4972 boolean_t full; 4973 l2arc_write_callback_t *cb; 4974 zio_t *pio, *wzio; 4975 uint64_t guid = spa_load_guid(spa); 4976 const boolean_t do_headroom_boost = *headroom_boost; 4977 int try; 4978 4979 ASSERT(dev->l2ad_vdev != NULL); 4980 4981 /* Lower the flag now, we might want to raise it again later. */ 4982 *headroom_boost = B_FALSE; 4983 4984 pio = NULL; 4985 write_sz = write_asize = write_psize = 0; 4986 full = B_FALSE; 4987 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 4988 head->b_flags |= ARC_L2_WRITE_HEAD; 4989 4990 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 4991 /* 4992 * We will want to try to compress buffers that are at least 2x the 4993 * device sector size. 4994 */ 4995 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 4996 4997 /* 4998 * Copy buffers for L2ARC writing. 4999 */ 5000 mutex_enter(&l2arc_buflist_mtx); 5001 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 5002 uint64_t passed_sz = 0; 5003 5004 list = l2arc_list_locked(try, &list_lock); 5005 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 5006 5007 /* 5008 * L2ARC fast warmup. 5009 * 5010 * Until the ARC is warm and starts to evict, read from the 5011 * head of the ARC lists rather than the tail. 5012 */ 5013 if (arc_warm == B_FALSE) 5014 ab = list_head(list); 5015 else 5016 ab = list_tail(list); 5017 if (ab == NULL) 5018 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 5019 5020 headroom = target_sz * l2arc_headroom; 5021 if (do_headroom_boost) 5022 headroom = (headroom * l2arc_headroom_boost) / 100; 5023 5024 for (; ab; ab = ab_prev) { 5025 l2arc_buf_hdr_t *l2hdr; 5026 kmutex_t *hash_lock; 5027 uint64_t buf_sz; 5028 5029 if (arc_warm == B_FALSE) 5030 ab_prev = list_next(list, ab); 5031 else 5032 ab_prev = list_prev(list, ab); 5033 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size); 5034 5035 hash_lock = HDR_LOCK(ab); 5036 if (!mutex_tryenter(hash_lock)) { 5037 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5038 /* 5039 * Skip this buffer rather than waiting. 5040 */ 5041 continue; 5042 } 5043 5044 passed_sz += ab->b_size; 5045 if (passed_sz > headroom) { 5046 /* 5047 * Searched too far. 5048 */ 5049 mutex_exit(hash_lock); 5050 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5051 break; 5052 } 5053 5054 if (!l2arc_write_eligible(guid, ab)) { 5055 mutex_exit(hash_lock); 5056 continue; 5057 } 5058 5059 if ((write_sz + ab->b_size) > target_sz) { 5060 full = B_TRUE; 5061 mutex_exit(hash_lock); 5062 ARCSTAT_BUMP(arcstat_l2_write_full); 5063 break; 5064 } 5065 5066 if (pio == NULL) { 5067 /* 5068 * Insert a dummy header on the buflist so 5069 * l2arc_write_done() can find where the 5070 * write buffers begin without searching. 5071 */ 5072 list_insert_head(dev->l2ad_buflist, head); 5073 5074 cb = kmem_alloc( 5075 sizeof (l2arc_write_callback_t), KM_SLEEP); 5076 cb->l2wcb_dev = dev; 5077 cb->l2wcb_head = head; 5078 pio = zio_root(spa, l2arc_write_done, cb, 5079 ZIO_FLAG_CANFAIL); 5080 ARCSTAT_BUMP(arcstat_l2_write_pios); 5081 } 5082 5083 /* 5084 * Create and add a new L2ARC header. 5085 */ 5086 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5087 l2hdr->b_dev = dev; 5088 ab->b_flags |= ARC_L2_WRITING; 5089 5090 /* 5091 * Temporarily stash the data buffer in b_tmp_cdata. 5092 * The subsequent write step will pick it up from 5093 * there. This is because can't access ab->b_buf 5094 * without holding the hash_lock, which we in turn 5095 * can't access without holding the ARC list locks 5096 * (which we want to avoid during compression/writing). 5097 */ 5098 l2hdr->b_compress = ZIO_COMPRESS_OFF; 5099 l2hdr->b_asize = ab->b_size; 5100 l2hdr->b_tmp_cdata = ab->b_buf->b_data; 5101 5102 buf_sz = ab->b_size; 5103 ab->b_l2hdr = l2hdr; 5104 5105 list_insert_head(dev->l2ad_buflist, ab); 5106 5107 /* 5108 * Compute and store the buffer cksum before 5109 * writing. On debug the cksum is verified first. 5110 */ 5111 arc_cksum_verify(ab->b_buf); 5112 arc_cksum_compute(ab->b_buf, B_TRUE); 5113 5114 mutex_exit(hash_lock); 5115 5116 write_sz += buf_sz; 5117 } 5118 5119 mutex_exit(list_lock); 5120 5121 if (full == B_TRUE) 5122 break; 5123 } 5124 5125 /* No buffers selected for writing? */ 5126 if (pio == NULL) { 5127 ASSERT0(write_sz); 5128 mutex_exit(&l2arc_buflist_mtx); 5129 kmem_cache_free(hdr_cache, head); 5130 return (0); 5131 } 5132 5133 /* 5134 * Now start writing the buffers. We're starting at the write head 5135 * and work backwards, retracing the course of the buffer selector 5136 * loop above. 5137 */ 5138 for (ab = list_prev(dev->l2ad_buflist, head); ab; 5139 ab = list_prev(dev->l2ad_buflist, ab)) { 5140 l2arc_buf_hdr_t *l2hdr; 5141 uint64_t buf_sz; 5142 5143 /* 5144 * We shouldn't need to lock the buffer here, since we flagged 5145 * it as ARC_L2_WRITING in the previous step, but we must take 5146 * care to only access its L2 cache parameters. In particular, 5147 * ab->b_buf may be invalid by now due to ARC eviction. 5148 */ 5149 l2hdr = ab->b_l2hdr; 5150 l2hdr->b_daddr = dev->l2ad_hand; 5151 5152 if ((ab->b_flags & ARC_L2COMPRESS) && 5153 l2hdr->b_asize >= buf_compress_minsz) { 5154 if (l2arc_compress_buf(l2hdr)) { 5155 /* 5156 * If compression succeeded, enable headroom 5157 * boost on the next scan cycle. 5158 */ 5159 *headroom_boost = B_TRUE; 5160 } 5161 } 5162 5163 /* 5164 * Pick up the buffer data we had previously stashed away 5165 * (and now potentially also compressed). 5166 */ 5167 buf_data = l2hdr->b_tmp_cdata; 5168 buf_sz = l2hdr->b_asize; 5169 5170 /* 5171 * If the data has not been compressed, then clear b_tmp_cdata 5172 * to make sure that it points only to a temporary compression 5173 * buffer. 5174 */ 5175 if (!L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)) 5176 l2hdr->b_tmp_cdata = NULL; 5177 5178 /* Compression may have squashed the buffer to zero length. */ 5179 if (buf_sz != 0) { 5180 uint64_t buf_p_sz; 5181 5182 wzio = zio_write_phys(pio, dev->l2ad_vdev, 5183 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5184 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5185 ZIO_FLAG_CANFAIL, B_FALSE); 5186 5187 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5188 zio_t *, wzio); 5189 (void) zio_nowait(wzio); 5190 5191 write_asize += buf_sz; 5192 /* 5193 * Keep the clock hand suitably device-aligned. 5194 */ 5195 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5196 write_psize += buf_p_sz; 5197 dev->l2ad_hand += buf_p_sz; 5198 } 5199 } 5200 5201 mutex_exit(&l2arc_buflist_mtx); 5202 5203 ASSERT3U(write_asize, <=, target_sz); 5204 ARCSTAT_BUMP(arcstat_l2_writes_sent); 5205 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5206 ARCSTAT_INCR(arcstat_l2_size, write_sz); 5207 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 5208 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); 5209 5210 /* 5211 * Bump device hand to the device start if it is approaching the end. 5212 * l2arc_evict() will already have evicted ahead for this case. 5213 */ 5214 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5215 dev->l2ad_hand = dev->l2ad_start; 5216 dev->l2ad_evict = dev->l2ad_start; 5217 dev->l2ad_first = B_FALSE; 5218 } 5219 5220 dev->l2ad_writing = B_TRUE; 5221 (void) zio_wait(pio); 5222 dev->l2ad_writing = B_FALSE; 5223 5224 return (write_asize); 5225} 5226 5227/* 5228 * Compresses an L2ARC buffer. 5229 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5230 * size in l2hdr->b_asize. This routine tries to compress the data and 5231 * depending on the compression result there are three possible outcomes: 5232 * *) The buffer was incompressible. The original l2hdr contents were left 5233 * untouched and are ready for writing to an L2 device. 5234 * *) The buffer was all-zeros, so there is no need to write it to an L2 5235 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5236 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5237 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5238 * data buffer which holds the compressed data to be written, and b_asize 5239 * tells us how much data there is. b_compress is set to the appropriate 5240 * compression algorithm. Once writing is done, invoke 5241 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5242 * 5243 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5244 * buffer was incompressible). 5245 */ 5246static boolean_t 5247l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5248{ 5249 void *cdata; 5250 size_t csize, len, rounded; 5251 5252 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5253 ASSERT(l2hdr->b_tmp_cdata != NULL); 5254 5255 len = l2hdr->b_asize; 5256 cdata = zio_data_buf_alloc(len); 5257 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5258 cdata, l2hdr->b_asize); 5259 5260 rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE); 5261 if (rounded > csize) { 5262 bzero((char *)cdata + csize, rounded - csize); 5263 csize = rounded; 5264 } 5265 5266 if (csize == 0) { 5267 /* zero block, indicate that there's nothing to write */ 5268 zio_data_buf_free(cdata, len); 5269 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5270 l2hdr->b_asize = 0; 5271 l2hdr->b_tmp_cdata = NULL; 5272 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5273 return (B_TRUE); 5274 } else if (csize > 0 && csize < len) { 5275 /* 5276 * Compression succeeded, we'll keep the cdata around for 5277 * writing and release it afterwards. 5278 */ 5279 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5280 l2hdr->b_asize = csize; 5281 l2hdr->b_tmp_cdata = cdata; 5282 ARCSTAT_BUMP(arcstat_l2_compress_successes); 5283 return (B_TRUE); 5284 } else { 5285 /* 5286 * Compression failed, release the compressed buffer. 5287 * l2hdr will be left unmodified. 5288 */ 5289 zio_data_buf_free(cdata, len); 5290 ARCSTAT_BUMP(arcstat_l2_compress_failures); 5291 return (B_FALSE); 5292 } 5293} 5294 5295/* 5296 * Decompresses a zio read back from an l2arc device. On success, the 5297 * underlying zio's io_data buffer is overwritten by the uncompressed 5298 * version. On decompression error (corrupt compressed stream), the 5299 * zio->io_error value is set to signal an I/O error. 5300 * 5301 * Please note that the compressed data stream is not checksummed, so 5302 * if the underlying device is experiencing data corruption, we may feed 5303 * corrupt data to the decompressor, so the decompressor needs to be 5304 * able to handle this situation (LZ4 does). 5305 */ 5306static void 5307l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5308{ 5309 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5310 5311 if (zio->io_error != 0) { 5312 /* 5313 * An io error has occured, just restore the original io 5314 * size in preparation for a main pool read. 5315 */ 5316 zio->io_orig_size = zio->io_size = hdr->b_size; 5317 return; 5318 } 5319 5320 if (c == ZIO_COMPRESS_EMPTY) { 5321 /* 5322 * An empty buffer results in a null zio, which means we 5323 * need to fill its io_data after we're done restoring the 5324 * buffer's contents. 5325 */ 5326 ASSERT(hdr->b_buf != NULL); 5327 bzero(hdr->b_buf->b_data, hdr->b_size); 5328 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5329 } else { 5330 ASSERT(zio->io_data != NULL); 5331 /* 5332 * We copy the compressed data from the start of the arc buffer 5333 * (the zio_read will have pulled in only what we need, the 5334 * rest is garbage which we will overwrite at decompression) 5335 * and then decompress back to the ARC data buffer. This way we 5336 * can minimize copying by simply decompressing back over the 5337 * original compressed data (rather than decompressing to an 5338 * aux buffer and then copying back the uncompressed buffer, 5339 * which is likely to be much larger). 5340 */ 5341 uint64_t csize; 5342 void *cdata; 5343 5344 csize = zio->io_size; 5345 cdata = zio_data_buf_alloc(csize); 5346 bcopy(zio->io_data, cdata, csize); 5347 if (zio_decompress_data(c, cdata, zio->io_data, csize, 5348 hdr->b_size) != 0) 5349 zio->io_error = EIO; 5350 zio_data_buf_free(cdata, csize); 5351 } 5352 5353 /* Restore the expected uncompressed IO size. */ 5354 zio->io_orig_size = zio->io_size = hdr->b_size; 5355} 5356 5357/* 5358 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5359 * This buffer serves as a temporary holder of compressed data while 5360 * the buffer entry is being written to an l2arc device. Once that is 5361 * done, we can dispose of it. 5362 */ 5363static void 5364l2arc_release_cdata_buf(arc_buf_hdr_t *ab) 5365{ 5366 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr; 5367 5368 ASSERT(L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)); 5369 if (l2hdr->b_compress != ZIO_COMPRESS_EMPTY) { 5370 /* 5371 * If the data was compressed, then we've allocated a 5372 * temporary buffer for it, so now we need to release it. 5373 */ 5374 ASSERT(l2hdr->b_tmp_cdata != NULL); 5375 zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size); 5376 l2hdr->b_tmp_cdata = NULL; 5377 } else { 5378 ASSERT(l2hdr->b_tmp_cdata == NULL); 5379 } 5380} 5381 5382/* 5383 * This thread feeds the L2ARC at regular intervals. This is the beating 5384 * heart of the L2ARC. 5385 */ 5386static void 5387l2arc_feed_thread(void *dummy __unused) 5388{ 5389 callb_cpr_t cpr; 5390 l2arc_dev_t *dev; 5391 spa_t *spa; 5392 uint64_t size, wrote; 5393 clock_t begin, next = ddi_get_lbolt(); 5394 boolean_t headroom_boost = B_FALSE; 5395 5396 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5397 5398 mutex_enter(&l2arc_feed_thr_lock); 5399 5400 while (l2arc_thread_exit == 0) { 5401 CALLB_CPR_SAFE_BEGIN(&cpr); 5402 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5403 next - ddi_get_lbolt()); 5404 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5405 next = ddi_get_lbolt() + hz; 5406 5407 /* 5408 * Quick check for L2ARC devices. 5409 */ 5410 mutex_enter(&l2arc_dev_mtx); 5411 if (l2arc_ndev == 0) { 5412 mutex_exit(&l2arc_dev_mtx); 5413 continue; 5414 } 5415 mutex_exit(&l2arc_dev_mtx); 5416 begin = ddi_get_lbolt(); 5417 5418 /* 5419 * This selects the next l2arc device to write to, and in 5420 * doing so the next spa to feed from: dev->l2ad_spa. This 5421 * will return NULL if there are now no l2arc devices or if 5422 * they are all faulted. 5423 * 5424 * If a device is returned, its spa's config lock is also 5425 * held to prevent device removal. l2arc_dev_get_next() 5426 * will grab and release l2arc_dev_mtx. 5427 */ 5428 if ((dev = l2arc_dev_get_next()) == NULL) 5429 continue; 5430 5431 spa = dev->l2ad_spa; 5432 ASSERT(spa != NULL); 5433 5434 /* 5435 * If the pool is read-only then force the feed thread to 5436 * sleep a little longer. 5437 */ 5438 if (!spa_writeable(spa)) { 5439 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5440 spa_config_exit(spa, SCL_L2ARC, dev); 5441 continue; 5442 } 5443 5444 /* 5445 * Avoid contributing to memory pressure. 5446 */ 5447 if (arc_reclaim_needed()) { 5448 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5449 spa_config_exit(spa, SCL_L2ARC, dev); 5450 continue; 5451 } 5452 5453 ARCSTAT_BUMP(arcstat_l2_feeds); 5454 5455 size = l2arc_write_size(); 5456 5457 /* 5458 * Evict L2ARC buffers that will be overwritten. 5459 */ 5460 l2arc_evict(dev, size, B_FALSE); 5461 5462 /* 5463 * Write ARC buffers. 5464 */ 5465 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5466 5467 /* 5468 * Calculate interval between writes. 5469 */ 5470 next = l2arc_write_interval(begin, size, wrote); 5471 spa_config_exit(spa, SCL_L2ARC, dev); 5472 } 5473 5474 l2arc_thread_exit = 0; 5475 cv_broadcast(&l2arc_feed_thr_cv); 5476 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5477 thread_exit(); 5478} 5479 5480boolean_t 5481l2arc_vdev_present(vdev_t *vd) 5482{ 5483 l2arc_dev_t *dev; 5484 5485 mutex_enter(&l2arc_dev_mtx); 5486 for (dev = list_head(l2arc_dev_list); dev != NULL; 5487 dev = list_next(l2arc_dev_list, dev)) { 5488 if (dev->l2ad_vdev == vd) 5489 break; 5490 } 5491 mutex_exit(&l2arc_dev_mtx); 5492 5493 return (dev != NULL); 5494} 5495 5496/* 5497 * Add a vdev for use by the L2ARC. By this point the spa has already 5498 * validated the vdev and opened it. 5499 */ 5500void 5501l2arc_add_vdev(spa_t *spa, vdev_t *vd) 5502{ 5503 l2arc_dev_t *adddev; 5504 5505 ASSERT(!l2arc_vdev_present(vd)); 5506 5507 vdev_ashift_optimize(vd); 5508 5509 /* 5510 * Create a new l2arc device entry. 5511 */ 5512 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5513 adddev->l2ad_spa = spa; 5514 adddev->l2ad_vdev = vd; 5515 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5516 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5517 adddev->l2ad_hand = adddev->l2ad_start; 5518 adddev->l2ad_evict = adddev->l2ad_start; 5519 adddev->l2ad_first = B_TRUE; 5520 adddev->l2ad_writing = B_FALSE; 5521 5522 /* 5523 * This is a list of all ARC buffers that are still valid on the 5524 * device. 5525 */ 5526 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5527 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5528 offsetof(arc_buf_hdr_t, b_l2node)); 5529 5530 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5531 5532 /* 5533 * Add device to global list 5534 */ 5535 mutex_enter(&l2arc_dev_mtx); 5536 list_insert_head(l2arc_dev_list, adddev); 5537 atomic_inc_64(&l2arc_ndev); 5538 mutex_exit(&l2arc_dev_mtx); 5539} 5540 5541/* 5542 * Remove a vdev from the L2ARC. 5543 */ 5544void 5545l2arc_remove_vdev(vdev_t *vd) 5546{ 5547 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5548 5549 /* 5550 * Find the device by vdev 5551 */ 5552 mutex_enter(&l2arc_dev_mtx); 5553 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5554 nextdev = list_next(l2arc_dev_list, dev); 5555 if (vd == dev->l2ad_vdev) { 5556 remdev = dev; 5557 break; 5558 } 5559 } 5560 ASSERT(remdev != NULL); 5561 5562 /* 5563 * Remove device from global list 5564 */ 5565 list_remove(l2arc_dev_list, remdev); 5566 l2arc_dev_last = NULL; /* may have been invalidated */ 5567 atomic_dec_64(&l2arc_ndev); 5568 mutex_exit(&l2arc_dev_mtx); 5569 5570 /* 5571 * Clear all buflists and ARC references. L2ARC device flush. 5572 */ 5573 l2arc_evict(remdev, 0, B_TRUE); 5574 list_destroy(remdev->l2ad_buflist); 5575 kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5576 kmem_free(remdev, sizeof (l2arc_dev_t)); 5577} 5578 5579void 5580l2arc_init(void) 5581{ 5582 l2arc_thread_exit = 0; 5583 l2arc_ndev = 0; 5584 l2arc_writes_sent = 0; 5585 l2arc_writes_done = 0; 5586 5587 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5588 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5589 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5590 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5591 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5592 5593 l2arc_dev_list = &L2ARC_dev_list; 5594 l2arc_free_on_write = &L2ARC_free_on_write; 5595 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5596 offsetof(l2arc_dev_t, l2ad_node)); 5597 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5598 offsetof(l2arc_data_free_t, l2df_list_node)); 5599} 5600 5601void 5602l2arc_fini(void) 5603{ 5604 /* 5605 * This is called from dmu_fini(), which is called from spa_fini(); 5606 * Because of this, we can assume that all l2arc devices have 5607 * already been removed when the pools themselves were removed. 5608 */ 5609 5610 l2arc_do_free_on_write(); 5611 5612 mutex_destroy(&l2arc_feed_thr_lock); 5613 cv_destroy(&l2arc_feed_thr_cv); 5614 mutex_destroy(&l2arc_dev_mtx); 5615 mutex_destroy(&l2arc_buflist_mtx); 5616 mutex_destroy(&l2arc_free_on_write_mtx); 5617 5618 list_destroy(l2arc_dev_list); 5619 list_destroy(l2arc_free_on_write); 5620} 5621 5622void 5623l2arc_start(void) 5624{ 5625 if (!(spa_mode_global & FWRITE)) 5626 return; 5627 5628 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5629 TS_RUN, minclsyspri); 5630} 5631 5632void 5633l2arc_stop(void) 5634{ 5635 if (!(spa_mode_global & FWRITE)) 5636 return; 5637 5638 mutex_enter(&l2arc_feed_thr_lock); 5639 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5640 l2arc_thread_exit = 1; 5641 while (l2arc_thread_exit != 0) 5642 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5643 mutex_exit(&l2arc_feed_thr_lock); 5644} 5645