arc.c revision 269846
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_abort_lowmem; 348 kstat_named_t arcstat_l2_cksum_bad; 349 kstat_named_t arcstat_l2_io_error; 350 kstat_named_t arcstat_l2_size; 351 kstat_named_t arcstat_l2_asize; 352 kstat_named_t arcstat_l2_hdr_size; 353 kstat_named_t arcstat_l2_compress_successes; 354 kstat_named_t arcstat_l2_compress_zeros; 355 kstat_named_t arcstat_l2_compress_failures; 356 kstat_named_t arcstat_l2_write_trylock_fail; 357 kstat_named_t arcstat_l2_write_passed_headroom; 358 kstat_named_t arcstat_l2_write_spa_mismatch; 359 kstat_named_t arcstat_l2_write_in_l2; 360 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 361 kstat_named_t arcstat_l2_write_not_cacheable; 362 kstat_named_t arcstat_l2_write_full; 363 kstat_named_t arcstat_l2_write_buffer_iter; 364 kstat_named_t arcstat_l2_write_pios; 365 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 366 kstat_named_t arcstat_l2_write_buffer_list_iter; 367 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 368 kstat_named_t arcstat_memory_throttle_count; 369 kstat_named_t arcstat_duplicate_buffers; 370 kstat_named_t arcstat_duplicate_buffers_size; 371 kstat_named_t arcstat_duplicate_reads; 372} arc_stats_t; 373 374static arc_stats_t arc_stats = { 375 { "hits", KSTAT_DATA_UINT64 }, 376 { "misses", KSTAT_DATA_UINT64 }, 377 { "demand_data_hits", KSTAT_DATA_UINT64 }, 378 { "demand_data_misses", KSTAT_DATA_UINT64 }, 379 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 380 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 381 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 382 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 383 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 384 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 385 { "mru_hits", KSTAT_DATA_UINT64 }, 386 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 387 { "mfu_hits", KSTAT_DATA_UINT64 }, 388 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 389 { "allocated", KSTAT_DATA_UINT64 }, 390 { "deleted", KSTAT_DATA_UINT64 }, 391 { "stolen", KSTAT_DATA_UINT64 }, 392 { "recycle_miss", KSTAT_DATA_UINT64 }, 393 { "mutex_miss", KSTAT_DATA_UINT64 }, 394 { "evict_skip", KSTAT_DATA_UINT64 }, 395 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 396 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 397 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 398 { "hash_elements", KSTAT_DATA_UINT64 }, 399 { "hash_elements_max", KSTAT_DATA_UINT64 }, 400 { "hash_collisions", KSTAT_DATA_UINT64 }, 401 { "hash_chains", KSTAT_DATA_UINT64 }, 402 { "hash_chain_max", KSTAT_DATA_UINT64 }, 403 { "p", KSTAT_DATA_UINT64 }, 404 { "c", KSTAT_DATA_UINT64 }, 405 { "c_min", KSTAT_DATA_UINT64 }, 406 { "c_max", KSTAT_DATA_UINT64 }, 407 { "size", KSTAT_DATA_UINT64 }, 408 { "hdr_size", KSTAT_DATA_UINT64 }, 409 { "data_size", KSTAT_DATA_UINT64 }, 410 { "other_size", KSTAT_DATA_UINT64 }, 411 { "l2_hits", KSTAT_DATA_UINT64 }, 412 { "l2_misses", KSTAT_DATA_UINT64 }, 413 { "l2_feeds", KSTAT_DATA_UINT64 }, 414 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 415 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 416 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 417 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 418 { "l2_writes_done", KSTAT_DATA_UINT64 }, 419 { "l2_writes_error", KSTAT_DATA_UINT64 }, 420 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 421 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 422 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 423 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 424 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 425 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 426 { "l2_io_error", KSTAT_DATA_UINT64 }, 427 { "l2_size", KSTAT_DATA_UINT64 }, 428 { "l2_asize", KSTAT_DATA_UINT64 }, 429 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 430 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 431 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 432 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 433 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 434 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 435 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 436 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 437 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 438 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 439 { "l2_write_full", KSTAT_DATA_UINT64 }, 440 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 441 { "l2_write_pios", KSTAT_DATA_UINT64 }, 442 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 443 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 444 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 445 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 446 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 447 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 448 { "duplicate_reads", KSTAT_DATA_UINT64 } 449}; 450 451#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 452 453#define ARCSTAT_INCR(stat, val) \ 454 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 455 456#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 457#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 458 459#define ARCSTAT_MAX(stat, val) { \ 460 uint64_t m; \ 461 while ((val) > (m = arc_stats.stat.value.ui64) && \ 462 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 463 continue; \ 464} 465 466#define ARCSTAT_MAXSTAT(stat) \ 467 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 468 469/* 470 * We define a macro to allow ARC hits/misses to be easily broken down by 471 * two separate conditions, giving a total of four different subtypes for 472 * each of hits and misses (so eight statistics total). 473 */ 474#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 475 if (cond1) { \ 476 if (cond2) { \ 477 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 478 } else { \ 479 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 480 } \ 481 } else { \ 482 if (cond2) { \ 483 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 484 } else { \ 485 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 486 } \ 487 } 488 489kstat_t *arc_ksp; 490static arc_state_t *arc_anon; 491static arc_state_t *arc_mru; 492static arc_state_t *arc_mru_ghost; 493static arc_state_t *arc_mfu; 494static arc_state_t *arc_mfu_ghost; 495static arc_state_t *arc_l2c_only; 496 497/* 498 * There are several ARC variables that are critical to export as kstats -- 499 * but we don't want to have to grovel around in the kstat whenever we wish to 500 * manipulate them. For these variables, we therefore define them to be in 501 * terms of the statistic variable. This assures that we are not introducing 502 * the possibility of inconsistency by having shadow copies of the variables, 503 * while still allowing the code to be readable. 504 */ 505#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 506#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 507#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 508#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 509#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 510 511#define L2ARC_IS_VALID_COMPRESS(_c_) \ 512 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 513 514static int arc_no_grow; /* Don't try to grow cache size */ 515static uint64_t arc_tempreserve; 516static uint64_t arc_loaned_bytes; 517static uint64_t arc_meta_used; 518static uint64_t arc_meta_limit; 519static uint64_t arc_meta_max = 0; 520SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RD, &arc_meta_used, 0, 521 "ARC metadata used"); 522SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RW, &arc_meta_limit, 0, 523 "ARC metadata limit"); 524 525typedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 526 527typedef struct arc_callback arc_callback_t; 528 529struct arc_callback { 530 void *acb_private; 531 arc_done_func_t *acb_done; 532 arc_buf_t *acb_buf; 533 zio_t *acb_zio_dummy; 534 arc_callback_t *acb_next; 535}; 536 537typedef struct arc_write_callback arc_write_callback_t; 538 539struct arc_write_callback { 540 void *awcb_private; 541 arc_done_func_t *awcb_ready; 542 arc_done_func_t *awcb_physdone; 543 arc_done_func_t *awcb_done; 544 arc_buf_t *awcb_buf; 545}; 546 547struct arc_buf_hdr { 548 /* protected by hash lock */ 549 dva_t b_dva; 550 uint64_t b_birth; 551 uint64_t b_cksum0; 552 553 kmutex_t b_freeze_lock; 554 zio_cksum_t *b_freeze_cksum; 555 void *b_thawed; 556 557 arc_buf_hdr_t *b_hash_next; 558 arc_buf_t *b_buf; 559 uint32_t b_flags; 560 uint32_t b_datacnt; 561 562 arc_callback_t *b_acb; 563 kcondvar_t b_cv; 564 565 /* immutable */ 566 arc_buf_contents_t b_type; 567 uint64_t b_size; 568 uint64_t b_spa; 569 570 /* protected by arc state mutex */ 571 arc_state_t *b_state; 572 list_node_t b_arc_node; 573 574 /* updated atomically */ 575 clock_t b_arc_access; 576 577 /* self protecting */ 578 refcount_t b_refcnt; 579 580 l2arc_buf_hdr_t *b_l2hdr; 581 list_node_t b_l2node; 582}; 583 584static arc_buf_t *arc_eviction_list; 585static kmutex_t arc_eviction_mtx; 586static arc_buf_hdr_t arc_eviction_hdr; 587static void arc_get_data_buf(arc_buf_t *buf); 588static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock); 589static int arc_evict_needed(arc_buf_contents_t type); 590static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes); 591#ifdef illumos 592static void arc_buf_watch(arc_buf_t *buf); 593#endif /* illumos */ 594 595static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab); 596 597#define GHOST_STATE(state) \ 598 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 599 (state) == arc_l2c_only) 600 601/* 602 * Private ARC flags. These flags are private ARC only flags that will show up 603 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can 604 * be passed in as arc_flags in things like arc_read. However, these flags 605 * should never be passed and should only be set by ARC code. When adding new 606 * public flags, make sure not to smash the private ones. 607 */ 608 609#define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */ 610#define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */ 611#define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */ 612#define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */ 613#define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */ 614#define ARC_INDIRECT (1 << 14) /* this is an indirect block */ 615#define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */ 616#define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */ 617#define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */ 618#define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */ 619 620#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE) 621#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS) 622#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR) 623#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH) 624#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ) 625#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE) 626#define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS) 627#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE) 628#define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \ 629 (hdr)->b_l2hdr != NULL) 630#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING) 631#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED) 632#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD) 633 634/* 635 * Other sizes 636 */ 637 638#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 639#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 640 641/* 642 * Hash table routines 643 */ 644 645#define HT_LOCK_PAD CACHE_LINE_SIZE 646 647struct ht_lock { 648 kmutex_t ht_lock; 649#ifdef _KERNEL 650 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 651#endif 652}; 653 654#define BUF_LOCKS 256 655typedef struct buf_hash_table { 656 uint64_t ht_mask; 657 arc_buf_hdr_t **ht_table; 658 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 659} buf_hash_table_t; 660 661static buf_hash_table_t buf_hash_table; 662 663#define BUF_HASH_INDEX(spa, dva, birth) \ 664 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 665#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 666#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 667#define HDR_LOCK(hdr) \ 668 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 669 670uint64_t zfs_crc64_table[256]; 671 672/* 673 * Level 2 ARC 674 */ 675 676#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 677#define L2ARC_HEADROOM 2 /* num of writes */ 678/* 679 * If we discover during ARC scan any buffers to be compressed, we boost 680 * our headroom for the next scanning cycle by this percentage multiple. 681 */ 682#define L2ARC_HEADROOM_BOOST 200 683#define L2ARC_FEED_SECS 1 /* caching interval secs */ 684#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 685 686#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 687#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 688 689/* L2ARC Performance Tunables */ 690uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 691uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 692uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 693uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 694uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 695uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 696boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 697boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 698boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 699 700SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 701 &l2arc_write_max, 0, "max write size"); 702SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 703 &l2arc_write_boost, 0, "extra write during warmup"); 704SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 705 &l2arc_headroom, 0, "number of dev writes"); 706SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 707 &l2arc_feed_secs, 0, "interval seconds"); 708SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 709 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 710 711SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 712 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 713SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 714 &l2arc_feed_again, 0, "turbo warmup"); 715SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 716 &l2arc_norw, 0, "no reads during writes"); 717 718SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 719 &ARC_anon.arcs_size, 0, "size of anonymous state"); 720SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 721 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 722SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 723 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 724 725SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 726 &ARC_mru.arcs_size, 0, "size of mru state"); 727SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 728 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 729SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 730 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 731 732SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 733 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 734SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 735 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 736 "size of metadata in mru ghost state"); 737SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 738 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 739 "size of data in mru ghost state"); 740 741SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 742 &ARC_mfu.arcs_size, 0, "size of mfu state"); 743SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 744 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 745SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 746 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 747 748SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 749 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 750SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 751 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 752 "size of metadata in mfu ghost state"); 753SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 754 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 755 "size of data in mfu ghost state"); 756 757SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 758 &ARC_l2c_only.arcs_size, 0, "size of mru state"); 759 760/* 761 * L2ARC Internals 762 */ 763typedef struct l2arc_dev { 764 vdev_t *l2ad_vdev; /* vdev */ 765 spa_t *l2ad_spa; /* spa */ 766 uint64_t l2ad_hand; /* next write location */ 767 uint64_t l2ad_start; /* first addr on device */ 768 uint64_t l2ad_end; /* last addr on device */ 769 uint64_t l2ad_evict; /* last addr eviction reached */ 770 boolean_t l2ad_first; /* first sweep through */ 771 boolean_t l2ad_writing; /* currently writing */ 772 list_t *l2ad_buflist; /* buffer list */ 773 list_node_t l2ad_node; /* device list node */ 774} l2arc_dev_t; 775 776static list_t L2ARC_dev_list; /* device list */ 777static list_t *l2arc_dev_list; /* device list pointer */ 778static kmutex_t l2arc_dev_mtx; /* device list mutex */ 779static l2arc_dev_t *l2arc_dev_last; /* last device used */ 780static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 781static list_t L2ARC_free_on_write; /* free after write buf list */ 782static list_t *l2arc_free_on_write; /* free after write list ptr */ 783static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 784static uint64_t l2arc_ndev; /* number of devices */ 785 786typedef struct l2arc_read_callback { 787 arc_buf_t *l2rcb_buf; /* read buffer */ 788 spa_t *l2rcb_spa; /* spa */ 789 blkptr_t l2rcb_bp; /* original blkptr */ 790 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 791 int l2rcb_flags; /* original flags */ 792 enum zio_compress l2rcb_compress; /* applied compress */ 793} l2arc_read_callback_t; 794 795typedef struct l2arc_write_callback { 796 l2arc_dev_t *l2wcb_dev; /* device info */ 797 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 798} l2arc_write_callback_t; 799 800struct l2arc_buf_hdr { 801 /* protected by arc_buf_hdr mutex */ 802 l2arc_dev_t *b_dev; /* L2ARC device */ 803 uint64_t b_daddr; /* disk address, offset byte */ 804 /* compression applied to buffer data */ 805 enum zio_compress b_compress; 806 /* real alloc'd buffer size depending on b_compress applied */ 807 int b_asize; 808 /* temporary buffer holder for in-flight compressed data */ 809 void *b_tmp_cdata; 810}; 811 812typedef struct l2arc_data_free { 813 /* protected by l2arc_free_on_write_mtx */ 814 void *l2df_data; 815 size_t l2df_size; 816 void (*l2df_func)(void *, size_t); 817 list_node_t l2df_list_node; 818} l2arc_data_free_t; 819 820static kmutex_t l2arc_feed_thr_lock; 821static kcondvar_t l2arc_feed_thr_cv; 822static uint8_t l2arc_thread_exit; 823 824static void l2arc_read_done(zio_t *zio); 825static void l2arc_hdr_stat_add(void); 826static void l2arc_hdr_stat_remove(void); 827 828static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr); 829static void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, 830 enum zio_compress c); 831static void l2arc_release_cdata_buf(arc_buf_hdr_t *ab); 832 833static uint64_t 834buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 835{ 836 uint8_t *vdva = (uint8_t *)dva; 837 uint64_t crc = -1ULL; 838 int i; 839 840 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 841 842 for (i = 0; i < sizeof (dva_t); i++) 843 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 844 845 crc ^= (spa>>8) ^ birth; 846 847 return (crc); 848} 849 850#define BUF_EMPTY(buf) \ 851 ((buf)->b_dva.dva_word[0] == 0 && \ 852 (buf)->b_dva.dva_word[1] == 0 && \ 853 (buf)->b_cksum0 == 0) 854 855#define BUF_EQUAL(spa, dva, birth, buf) \ 856 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 857 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 858 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 859 860static void 861buf_discard_identity(arc_buf_hdr_t *hdr) 862{ 863 hdr->b_dva.dva_word[0] = 0; 864 hdr->b_dva.dva_word[1] = 0; 865 hdr->b_birth = 0; 866 hdr->b_cksum0 = 0; 867} 868 869static arc_buf_hdr_t * 870buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 871{ 872 const dva_t *dva = BP_IDENTITY(bp); 873 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 874 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 875 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 876 arc_buf_hdr_t *buf; 877 878 mutex_enter(hash_lock); 879 for (buf = buf_hash_table.ht_table[idx]; buf != NULL; 880 buf = buf->b_hash_next) { 881 if (BUF_EQUAL(spa, dva, birth, buf)) { 882 *lockp = hash_lock; 883 return (buf); 884 } 885 } 886 mutex_exit(hash_lock); 887 *lockp = NULL; 888 return (NULL); 889} 890 891/* 892 * Insert an entry into the hash table. If there is already an element 893 * equal to elem in the hash table, then the already existing element 894 * will be returned and the new element will not be inserted. 895 * Otherwise returns NULL. 896 */ 897static arc_buf_hdr_t * 898buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp) 899{ 900 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 901 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 902 arc_buf_hdr_t *fbuf; 903 uint32_t i; 904 905 ASSERT(!DVA_IS_EMPTY(&buf->b_dva)); 906 ASSERT(buf->b_birth != 0); 907 ASSERT(!HDR_IN_HASH_TABLE(buf)); 908 *lockp = hash_lock; 909 mutex_enter(hash_lock); 910 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL; 911 fbuf = fbuf->b_hash_next, i++) { 912 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf)) 913 return (fbuf); 914 } 915 916 buf->b_hash_next = buf_hash_table.ht_table[idx]; 917 buf_hash_table.ht_table[idx] = buf; 918 buf->b_flags |= ARC_IN_HASH_TABLE; 919 920 /* collect some hash table performance data */ 921 if (i > 0) { 922 ARCSTAT_BUMP(arcstat_hash_collisions); 923 if (i == 1) 924 ARCSTAT_BUMP(arcstat_hash_chains); 925 926 ARCSTAT_MAX(arcstat_hash_chain_max, i); 927 } 928 929 ARCSTAT_BUMP(arcstat_hash_elements); 930 ARCSTAT_MAXSTAT(arcstat_hash_elements); 931 932 return (NULL); 933} 934 935static void 936buf_hash_remove(arc_buf_hdr_t *buf) 937{ 938 arc_buf_hdr_t *fbuf, **bufp; 939 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 940 941 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 942 ASSERT(HDR_IN_HASH_TABLE(buf)); 943 944 bufp = &buf_hash_table.ht_table[idx]; 945 while ((fbuf = *bufp) != buf) { 946 ASSERT(fbuf != NULL); 947 bufp = &fbuf->b_hash_next; 948 } 949 *bufp = buf->b_hash_next; 950 buf->b_hash_next = NULL; 951 buf->b_flags &= ~ARC_IN_HASH_TABLE; 952 953 /* collect some hash table performance data */ 954 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 955 956 if (buf_hash_table.ht_table[idx] && 957 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 958 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 959} 960 961/* 962 * Global data structures and functions for the buf kmem cache. 963 */ 964static kmem_cache_t *hdr_cache; 965static kmem_cache_t *buf_cache; 966 967static void 968buf_fini(void) 969{ 970 int i; 971 972 kmem_free(buf_hash_table.ht_table, 973 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 974 for (i = 0; i < BUF_LOCKS; i++) 975 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 976 kmem_cache_destroy(hdr_cache); 977 kmem_cache_destroy(buf_cache); 978} 979 980/* 981 * Constructor callback - called when the cache is empty 982 * and a new buf is requested. 983 */ 984/* ARGSUSED */ 985static int 986hdr_cons(void *vbuf, void *unused, int kmflag) 987{ 988 arc_buf_hdr_t *buf = vbuf; 989 990 bzero(buf, sizeof (arc_buf_hdr_t)); 991 refcount_create(&buf->b_refcnt); 992 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL); 993 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 994 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 995 996 return (0); 997} 998 999/* ARGSUSED */ 1000static int 1001buf_cons(void *vbuf, void *unused, int kmflag) 1002{ 1003 arc_buf_t *buf = vbuf; 1004 1005 bzero(buf, sizeof (arc_buf_t)); 1006 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1007 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1008 1009 return (0); 1010} 1011 1012/* 1013 * Destructor callback - called when a cached buf is 1014 * no longer required. 1015 */ 1016/* ARGSUSED */ 1017static void 1018hdr_dest(void *vbuf, void *unused) 1019{ 1020 arc_buf_hdr_t *buf = vbuf; 1021 1022 ASSERT(BUF_EMPTY(buf)); 1023 refcount_destroy(&buf->b_refcnt); 1024 cv_destroy(&buf->b_cv); 1025 mutex_destroy(&buf->b_freeze_lock); 1026 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1027} 1028 1029/* ARGSUSED */ 1030static void 1031buf_dest(void *vbuf, void *unused) 1032{ 1033 arc_buf_t *buf = vbuf; 1034 1035 mutex_destroy(&buf->b_evict_lock); 1036 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1037} 1038 1039/* 1040 * Reclaim callback -- invoked when memory is low. 1041 */ 1042/* ARGSUSED */ 1043static void 1044hdr_recl(void *unused) 1045{ 1046 dprintf("hdr_recl called\n"); 1047 /* 1048 * umem calls the reclaim func when we destroy the buf cache, 1049 * which is after we do arc_fini(). 1050 */ 1051 if (!arc_dead) 1052 cv_signal(&arc_reclaim_thr_cv); 1053} 1054 1055static void 1056buf_init(void) 1057{ 1058 uint64_t *ct; 1059 uint64_t hsize = 1ULL << 12; 1060 int i, j; 1061 1062 /* 1063 * The hash table is big enough to fill all of physical memory 1064 * with an average block size of zfs_arc_average_blocksize (default 8K). 1065 * By default, the table will take up 1066 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1067 */ 1068 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1069 hsize <<= 1; 1070retry: 1071 buf_hash_table.ht_mask = hsize - 1; 1072 buf_hash_table.ht_table = 1073 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1074 if (buf_hash_table.ht_table == NULL) { 1075 ASSERT(hsize > (1ULL << 8)); 1076 hsize >>= 1; 1077 goto retry; 1078 } 1079 1080 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1081 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1082 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1083 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1084 1085 for (i = 0; i < 256; i++) 1086 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1087 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1088 1089 for (i = 0; i < BUF_LOCKS; i++) { 1090 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1091 NULL, MUTEX_DEFAULT, NULL); 1092 } 1093} 1094 1095#define ARC_MINTIME (hz>>4) /* 62 ms */ 1096 1097static void 1098arc_cksum_verify(arc_buf_t *buf) 1099{ 1100 zio_cksum_t zc; 1101 1102 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1103 return; 1104 1105 mutex_enter(&buf->b_hdr->b_freeze_lock); 1106 if (buf->b_hdr->b_freeze_cksum == NULL || 1107 (buf->b_hdr->b_flags & ARC_IO_ERROR)) { 1108 mutex_exit(&buf->b_hdr->b_freeze_lock); 1109 return; 1110 } 1111 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1112 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1113 panic("buffer modified while frozen!"); 1114 mutex_exit(&buf->b_hdr->b_freeze_lock); 1115} 1116 1117static int 1118arc_cksum_equal(arc_buf_t *buf) 1119{ 1120 zio_cksum_t zc; 1121 int equal; 1122 1123 mutex_enter(&buf->b_hdr->b_freeze_lock); 1124 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1125 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1126 mutex_exit(&buf->b_hdr->b_freeze_lock); 1127 1128 return (equal); 1129} 1130 1131static void 1132arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1133{ 1134 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1135 return; 1136 1137 mutex_enter(&buf->b_hdr->b_freeze_lock); 1138 if (buf->b_hdr->b_freeze_cksum != NULL) { 1139 mutex_exit(&buf->b_hdr->b_freeze_lock); 1140 return; 1141 } 1142 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1143 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1144 buf->b_hdr->b_freeze_cksum); 1145 mutex_exit(&buf->b_hdr->b_freeze_lock); 1146#ifdef illumos 1147 arc_buf_watch(buf); 1148#endif /* illumos */ 1149} 1150 1151#ifdef illumos 1152#ifndef _KERNEL 1153typedef struct procctl { 1154 long cmd; 1155 prwatch_t prwatch; 1156} procctl_t; 1157#endif 1158 1159/* ARGSUSED */ 1160static void 1161arc_buf_unwatch(arc_buf_t *buf) 1162{ 1163#ifndef _KERNEL 1164 if (arc_watch) { 1165 int result; 1166 procctl_t ctl; 1167 ctl.cmd = PCWATCH; 1168 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1169 ctl.prwatch.pr_size = 0; 1170 ctl.prwatch.pr_wflags = 0; 1171 result = write(arc_procfd, &ctl, sizeof (ctl)); 1172 ASSERT3U(result, ==, sizeof (ctl)); 1173 } 1174#endif 1175} 1176 1177/* ARGSUSED */ 1178static void 1179arc_buf_watch(arc_buf_t *buf) 1180{ 1181#ifndef _KERNEL 1182 if (arc_watch) { 1183 int result; 1184 procctl_t ctl; 1185 ctl.cmd = PCWATCH; 1186 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1187 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1188 ctl.prwatch.pr_wflags = WA_WRITE; 1189 result = write(arc_procfd, &ctl, sizeof (ctl)); 1190 ASSERT3U(result, ==, sizeof (ctl)); 1191 } 1192#endif 1193} 1194#endif /* illumos */ 1195 1196void 1197arc_buf_thaw(arc_buf_t *buf) 1198{ 1199 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1200 if (buf->b_hdr->b_state != arc_anon) 1201 panic("modifying non-anon buffer!"); 1202 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS) 1203 panic("modifying buffer while i/o in progress!"); 1204 arc_cksum_verify(buf); 1205 } 1206 1207 mutex_enter(&buf->b_hdr->b_freeze_lock); 1208 if (buf->b_hdr->b_freeze_cksum != NULL) { 1209 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1210 buf->b_hdr->b_freeze_cksum = NULL; 1211 } 1212 1213 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1214 if (buf->b_hdr->b_thawed) 1215 kmem_free(buf->b_hdr->b_thawed, 1); 1216 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1217 } 1218 1219 mutex_exit(&buf->b_hdr->b_freeze_lock); 1220 1221#ifdef illumos 1222 arc_buf_unwatch(buf); 1223#endif /* illumos */ 1224} 1225 1226void 1227arc_buf_freeze(arc_buf_t *buf) 1228{ 1229 kmutex_t *hash_lock; 1230 1231 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1232 return; 1233 1234 hash_lock = HDR_LOCK(buf->b_hdr); 1235 mutex_enter(hash_lock); 1236 1237 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1238 buf->b_hdr->b_state == arc_anon); 1239 arc_cksum_compute(buf, B_FALSE); 1240 mutex_exit(hash_lock); 1241 1242} 1243 1244static void 1245get_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock) 1246{ 1247 uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth); 1248 1249 if (ab->b_type == ARC_BUFC_METADATA) 1250 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1251 else { 1252 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1253 buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1254 } 1255 1256 *list = &state->arcs_lists[buf_hashid]; 1257 *lock = ARCS_LOCK(state, buf_hashid); 1258} 1259 1260 1261static void 1262add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1263{ 1264 ASSERT(MUTEX_HELD(hash_lock)); 1265 1266 if ((refcount_add(&ab->b_refcnt, tag) == 1) && 1267 (ab->b_state != arc_anon)) { 1268 uint64_t delta = ab->b_size * ab->b_datacnt; 1269 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type]; 1270 list_t *list; 1271 kmutex_t *lock; 1272 1273 get_buf_info(ab, ab->b_state, &list, &lock); 1274 ASSERT(!MUTEX_HELD(lock)); 1275 mutex_enter(lock); 1276 ASSERT(list_link_active(&ab->b_arc_node)); 1277 list_remove(list, ab); 1278 if (GHOST_STATE(ab->b_state)) { 1279 ASSERT0(ab->b_datacnt); 1280 ASSERT3P(ab->b_buf, ==, NULL); 1281 delta = ab->b_size; 1282 } 1283 ASSERT(delta > 0); 1284 ASSERT3U(*size, >=, delta); 1285 atomic_add_64(size, -delta); 1286 mutex_exit(lock); 1287 /* remove the prefetch flag if we get a reference */ 1288 if (ab->b_flags & ARC_PREFETCH) 1289 ab->b_flags &= ~ARC_PREFETCH; 1290 } 1291} 1292 1293static int 1294remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1295{ 1296 int cnt; 1297 arc_state_t *state = ab->b_state; 1298 1299 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1300 ASSERT(!GHOST_STATE(state)); 1301 1302 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) && 1303 (state != arc_anon)) { 1304 uint64_t *size = &state->arcs_lsize[ab->b_type]; 1305 list_t *list; 1306 kmutex_t *lock; 1307 1308 get_buf_info(ab, state, &list, &lock); 1309 ASSERT(!MUTEX_HELD(lock)); 1310 mutex_enter(lock); 1311 ASSERT(!list_link_active(&ab->b_arc_node)); 1312 list_insert_head(list, ab); 1313 ASSERT(ab->b_datacnt > 0); 1314 atomic_add_64(size, ab->b_size * ab->b_datacnt); 1315 mutex_exit(lock); 1316 } 1317 return (cnt); 1318} 1319 1320/* 1321 * Move the supplied buffer to the indicated state. The mutex 1322 * for the buffer must be held by the caller. 1323 */ 1324static void 1325arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock) 1326{ 1327 arc_state_t *old_state = ab->b_state; 1328 int64_t refcnt = refcount_count(&ab->b_refcnt); 1329 uint64_t from_delta, to_delta; 1330 list_t *list; 1331 kmutex_t *lock; 1332 1333 ASSERT(MUTEX_HELD(hash_lock)); 1334 ASSERT3P(new_state, !=, old_state); 1335 ASSERT(refcnt == 0 || ab->b_datacnt > 0); 1336 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state)); 1337 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon); 1338 1339 from_delta = to_delta = ab->b_datacnt * ab->b_size; 1340 1341 /* 1342 * If this buffer is evictable, transfer it from the 1343 * old state list to the new state list. 1344 */ 1345 if (refcnt == 0) { 1346 if (old_state != arc_anon) { 1347 int use_mutex; 1348 uint64_t *size = &old_state->arcs_lsize[ab->b_type]; 1349 1350 get_buf_info(ab, old_state, &list, &lock); 1351 use_mutex = !MUTEX_HELD(lock); 1352 if (use_mutex) 1353 mutex_enter(lock); 1354 1355 ASSERT(list_link_active(&ab->b_arc_node)); 1356 list_remove(list, ab); 1357 1358 /* 1359 * If prefetching out of the ghost cache, 1360 * we will have a non-zero datacnt. 1361 */ 1362 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) { 1363 /* ghost elements have a ghost size */ 1364 ASSERT(ab->b_buf == NULL); 1365 from_delta = ab->b_size; 1366 } 1367 ASSERT3U(*size, >=, from_delta); 1368 atomic_add_64(size, -from_delta); 1369 1370 if (use_mutex) 1371 mutex_exit(lock); 1372 } 1373 if (new_state != arc_anon) { 1374 int use_mutex; 1375 uint64_t *size = &new_state->arcs_lsize[ab->b_type]; 1376 1377 get_buf_info(ab, new_state, &list, &lock); 1378 use_mutex = !MUTEX_HELD(lock); 1379 if (use_mutex) 1380 mutex_enter(lock); 1381 1382 list_insert_head(list, ab); 1383 1384 /* ghost elements have a ghost size */ 1385 if (GHOST_STATE(new_state)) { 1386 ASSERT(ab->b_datacnt == 0); 1387 ASSERT(ab->b_buf == NULL); 1388 to_delta = ab->b_size; 1389 } 1390 atomic_add_64(size, to_delta); 1391 1392 if (use_mutex) 1393 mutex_exit(lock); 1394 } 1395 } 1396 1397 ASSERT(!BUF_EMPTY(ab)); 1398 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab)) 1399 buf_hash_remove(ab); 1400 1401 /* adjust state sizes */ 1402 if (to_delta) 1403 atomic_add_64(&new_state->arcs_size, to_delta); 1404 if (from_delta) { 1405 ASSERT3U(old_state->arcs_size, >=, from_delta); 1406 atomic_add_64(&old_state->arcs_size, -from_delta); 1407 } 1408 ab->b_state = new_state; 1409 1410 /* adjust l2arc hdr stats */ 1411 if (new_state == arc_l2c_only) 1412 l2arc_hdr_stat_add(); 1413 else if (old_state == arc_l2c_only) 1414 l2arc_hdr_stat_remove(); 1415} 1416 1417void 1418arc_space_consume(uint64_t space, arc_space_type_t type) 1419{ 1420 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1421 1422 switch (type) { 1423 case ARC_SPACE_DATA: 1424 ARCSTAT_INCR(arcstat_data_size, space); 1425 break; 1426 case ARC_SPACE_OTHER: 1427 ARCSTAT_INCR(arcstat_other_size, space); 1428 break; 1429 case ARC_SPACE_HDRS: 1430 ARCSTAT_INCR(arcstat_hdr_size, space); 1431 break; 1432 case ARC_SPACE_L2HDRS: 1433 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1434 break; 1435 } 1436 1437 atomic_add_64(&arc_meta_used, space); 1438 atomic_add_64(&arc_size, space); 1439} 1440 1441void 1442arc_space_return(uint64_t space, arc_space_type_t type) 1443{ 1444 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1445 1446 switch (type) { 1447 case ARC_SPACE_DATA: 1448 ARCSTAT_INCR(arcstat_data_size, -space); 1449 break; 1450 case ARC_SPACE_OTHER: 1451 ARCSTAT_INCR(arcstat_other_size, -space); 1452 break; 1453 case ARC_SPACE_HDRS: 1454 ARCSTAT_INCR(arcstat_hdr_size, -space); 1455 break; 1456 case ARC_SPACE_L2HDRS: 1457 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1458 break; 1459 } 1460 1461 ASSERT(arc_meta_used >= space); 1462 if (arc_meta_max < arc_meta_used) 1463 arc_meta_max = arc_meta_used; 1464 atomic_add_64(&arc_meta_used, -space); 1465 ASSERT(arc_size >= space); 1466 atomic_add_64(&arc_size, -space); 1467} 1468 1469void * 1470arc_data_buf_alloc(uint64_t size) 1471{ 1472 if (arc_evict_needed(ARC_BUFC_DATA)) 1473 cv_signal(&arc_reclaim_thr_cv); 1474 atomic_add_64(&arc_size, size); 1475 return (zio_data_buf_alloc(size)); 1476} 1477 1478void 1479arc_data_buf_free(void *buf, uint64_t size) 1480{ 1481 zio_data_buf_free(buf, size); 1482 ASSERT(arc_size >= size); 1483 atomic_add_64(&arc_size, -size); 1484} 1485 1486arc_buf_t * 1487arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1488{ 1489 arc_buf_hdr_t *hdr; 1490 arc_buf_t *buf; 1491 1492 ASSERT3U(size, >, 0); 1493 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1494 ASSERT(BUF_EMPTY(hdr)); 1495 hdr->b_size = size; 1496 hdr->b_type = type; 1497 hdr->b_spa = spa_load_guid(spa); 1498 hdr->b_state = arc_anon; 1499 hdr->b_arc_access = 0; 1500 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1501 buf->b_hdr = hdr; 1502 buf->b_data = NULL; 1503 buf->b_efunc = NULL; 1504 buf->b_private = NULL; 1505 buf->b_next = NULL; 1506 hdr->b_buf = buf; 1507 arc_get_data_buf(buf); 1508 hdr->b_datacnt = 1; 1509 hdr->b_flags = 0; 1510 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1511 (void) refcount_add(&hdr->b_refcnt, tag); 1512 1513 return (buf); 1514} 1515 1516static char *arc_onloan_tag = "onloan"; 1517 1518/* 1519 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1520 * flight data by arc_tempreserve_space() until they are "returned". Loaned 1521 * buffers must be returned to the arc before they can be used by the DMU or 1522 * freed. 1523 */ 1524arc_buf_t * 1525arc_loan_buf(spa_t *spa, int size) 1526{ 1527 arc_buf_t *buf; 1528 1529 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1530 1531 atomic_add_64(&arc_loaned_bytes, size); 1532 return (buf); 1533} 1534 1535/* 1536 * Return a loaned arc buffer to the arc. 1537 */ 1538void 1539arc_return_buf(arc_buf_t *buf, void *tag) 1540{ 1541 arc_buf_hdr_t *hdr = buf->b_hdr; 1542 1543 ASSERT(buf->b_data != NULL); 1544 (void) refcount_add(&hdr->b_refcnt, tag); 1545 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1546 1547 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1548} 1549 1550/* Detach an arc_buf from a dbuf (tag) */ 1551void 1552arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1553{ 1554 arc_buf_hdr_t *hdr; 1555 1556 ASSERT(buf->b_data != NULL); 1557 hdr = buf->b_hdr; 1558 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1559 (void) refcount_remove(&hdr->b_refcnt, tag); 1560 buf->b_efunc = NULL; 1561 buf->b_private = NULL; 1562 1563 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1564} 1565 1566static arc_buf_t * 1567arc_buf_clone(arc_buf_t *from) 1568{ 1569 arc_buf_t *buf; 1570 arc_buf_hdr_t *hdr = from->b_hdr; 1571 uint64_t size = hdr->b_size; 1572 1573 ASSERT(hdr->b_state != arc_anon); 1574 1575 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1576 buf->b_hdr = hdr; 1577 buf->b_data = NULL; 1578 buf->b_efunc = NULL; 1579 buf->b_private = NULL; 1580 buf->b_next = hdr->b_buf; 1581 hdr->b_buf = buf; 1582 arc_get_data_buf(buf); 1583 bcopy(from->b_data, buf->b_data, size); 1584 1585 /* 1586 * This buffer already exists in the arc so create a duplicate 1587 * copy for the caller. If the buffer is associated with user data 1588 * then track the size and number of duplicates. These stats will be 1589 * updated as duplicate buffers are created and destroyed. 1590 */ 1591 if (hdr->b_type == ARC_BUFC_DATA) { 1592 ARCSTAT_BUMP(arcstat_duplicate_buffers); 1593 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1594 } 1595 hdr->b_datacnt += 1; 1596 return (buf); 1597} 1598 1599void 1600arc_buf_add_ref(arc_buf_t *buf, void* tag) 1601{ 1602 arc_buf_hdr_t *hdr; 1603 kmutex_t *hash_lock; 1604 1605 /* 1606 * Check to see if this buffer is evicted. Callers 1607 * must verify b_data != NULL to know if the add_ref 1608 * was successful. 1609 */ 1610 mutex_enter(&buf->b_evict_lock); 1611 if (buf->b_data == NULL) { 1612 mutex_exit(&buf->b_evict_lock); 1613 return; 1614 } 1615 hash_lock = HDR_LOCK(buf->b_hdr); 1616 mutex_enter(hash_lock); 1617 hdr = buf->b_hdr; 1618 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1619 mutex_exit(&buf->b_evict_lock); 1620 1621 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1622 add_reference(hdr, hash_lock, tag); 1623 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1624 arc_access(hdr, hash_lock); 1625 mutex_exit(hash_lock); 1626 ARCSTAT_BUMP(arcstat_hits); 1627 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 1628 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1629 data, metadata, hits); 1630} 1631 1632/* 1633 * Free the arc data buffer. If it is an l2arc write in progress, 1634 * the buffer is placed on l2arc_free_on_write to be freed later. 1635 */ 1636static void 1637arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1638{ 1639 arc_buf_hdr_t *hdr = buf->b_hdr; 1640 1641 if (HDR_L2_WRITING(hdr)) { 1642 l2arc_data_free_t *df; 1643 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1644 df->l2df_data = buf->b_data; 1645 df->l2df_size = hdr->b_size; 1646 df->l2df_func = free_func; 1647 mutex_enter(&l2arc_free_on_write_mtx); 1648 list_insert_head(l2arc_free_on_write, df); 1649 mutex_exit(&l2arc_free_on_write_mtx); 1650 ARCSTAT_BUMP(arcstat_l2_free_on_write); 1651 } else { 1652 free_func(buf->b_data, hdr->b_size); 1653 } 1654} 1655 1656/* 1657 * Free up buf->b_data and if 'remove' is set, then pull the 1658 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 1659 */ 1660static void 1661arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 1662{ 1663 arc_buf_t **bufp; 1664 1665 /* free up data associated with the buf */ 1666 if (buf->b_data) { 1667 arc_state_t *state = buf->b_hdr->b_state; 1668 uint64_t size = buf->b_hdr->b_size; 1669 arc_buf_contents_t type = buf->b_hdr->b_type; 1670 1671 arc_cksum_verify(buf); 1672#ifdef illumos 1673 arc_buf_unwatch(buf); 1674#endif /* illumos */ 1675 1676 if (!recycle) { 1677 if (type == ARC_BUFC_METADATA) { 1678 arc_buf_data_free(buf, zio_buf_free); 1679 arc_space_return(size, ARC_SPACE_DATA); 1680 } else { 1681 ASSERT(type == ARC_BUFC_DATA); 1682 arc_buf_data_free(buf, zio_data_buf_free); 1683 ARCSTAT_INCR(arcstat_data_size, -size); 1684 atomic_add_64(&arc_size, -size); 1685 } 1686 } 1687 if (list_link_active(&buf->b_hdr->b_arc_node)) { 1688 uint64_t *cnt = &state->arcs_lsize[type]; 1689 1690 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1691 ASSERT(state != arc_anon); 1692 1693 ASSERT3U(*cnt, >=, size); 1694 atomic_add_64(cnt, -size); 1695 } 1696 ASSERT3U(state->arcs_size, >=, size); 1697 atomic_add_64(&state->arcs_size, -size); 1698 buf->b_data = NULL; 1699 1700 /* 1701 * If we're destroying a duplicate buffer make sure 1702 * that the appropriate statistics are updated. 1703 */ 1704 if (buf->b_hdr->b_datacnt > 1 && 1705 buf->b_hdr->b_type == ARC_BUFC_DATA) { 1706 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1707 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1708 } 1709 ASSERT(buf->b_hdr->b_datacnt > 0); 1710 buf->b_hdr->b_datacnt -= 1; 1711 } 1712 1713 /* only remove the buf if requested */ 1714 if (!remove) 1715 return; 1716 1717 /* remove the buf from the hdr list */ 1718 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1719 continue; 1720 *bufp = buf->b_next; 1721 buf->b_next = NULL; 1722 1723 ASSERT(buf->b_efunc == NULL); 1724 1725 /* clean up the buf */ 1726 buf->b_hdr = NULL; 1727 kmem_cache_free(buf_cache, buf); 1728} 1729 1730static void 1731arc_hdr_destroy(arc_buf_hdr_t *hdr) 1732{ 1733 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1734 ASSERT3P(hdr->b_state, ==, arc_anon); 1735 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1736 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1737 1738 if (l2hdr != NULL) { 1739 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1740 /* 1741 * To prevent arc_free() and l2arc_evict() from 1742 * attempting to free the same buffer at the same time, 1743 * a FREE_IN_PROGRESS flag is given to arc_free() to 1744 * give it priority. l2arc_evict() can't destroy this 1745 * header while we are waiting on l2arc_buflist_mtx. 1746 * 1747 * The hdr may be removed from l2ad_buflist before we 1748 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1749 */ 1750 if (!buflist_held) { 1751 mutex_enter(&l2arc_buflist_mtx); 1752 l2hdr = hdr->b_l2hdr; 1753 } 1754 1755 if (l2hdr != NULL) { 1756 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1757 hdr->b_size, 0); 1758 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1759 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1760 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1761 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 1762 -l2hdr->b_asize, 0, 0); 1763 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1764 if (hdr->b_state == arc_l2c_only) 1765 l2arc_hdr_stat_remove(); 1766 hdr->b_l2hdr = NULL; 1767 } 1768 1769 if (!buflist_held) 1770 mutex_exit(&l2arc_buflist_mtx); 1771 } 1772 1773 if (!BUF_EMPTY(hdr)) { 1774 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1775 buf_discard_identity(hdr); 1776 } 1777 while (hdr->b_buf) { 1778 arc_buf_t *buf = hdr->b_buf; 1779 1780 if (buf->b_efunc) { 1781 mutex_enter(&arc_eviction_mtx); 1782 mutex_enter(&buf->b_evict_lock); 1783 ASSERT(buf->b_hdr != NULL); 1784 arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1785 hdr->b_buf = buf->b_next; 1786 buf->b_hdr = &arc_eviction_hdr; 1787 buf->b_next = arc_eviction_list; 1788 arc_eviction_list = buf; 1789 mutex_exit(&buf->b_evict_lock); 1790 mutex_exit(&arc_eviction_mtx); 1791 } else { 1792 arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1793 } 1794 } 1795 if (hdr->b_freeze_cksum != NULL) { 1796 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1797 hdr->b_freeze_cksum = NULL; 1798 } 1799 if (hdr->b_thawed) { 1800 kmem_free(hdr->b_thawed, 1); 1801 hdr->b_thawed = NULL; 1802 } 1803 1804 ASSERT(!list_link_active(&hdr->b_arc_node)); 1805 ASSERT3P(hdr->b_hash_next, ==, NULL); 1806 ASSERT3P(hdr->b_acb, ==, NULL); 1807 kmem_cache_free(hdr_cache, hdr); 1808} 1809 1810void 1811arc_buf_free(arc_buf_t *buf, void *tag) 1812{ 1813 arc_buf_hdr_t *hdr = buf->b_hdr; 1814 int hashed = hdr->b_state != arc_anon; 1815 1816 ASSERT(buf->b_efunc == NULL); 1817 ASSERT(buf->b_data != NULL); 1818 1819 if (hashed) { 1820 kmutex_t *hash_lock = HDR_LOCK(hdr); 1821 1822 mutex_enter(hash_lock); 1823 hdr = buf->b_hdr; 1824 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1825 1826 (void) remove_reference(hdr, hash_lock, tag); 1827 if (hdr->b_datacnt > 1) { 1828 arc_buf_destroy(buf, FALSE, TRUE); 1829 } else { 1830 ASSERT(buf == hdr->b_buf); 1831 ASSERT(buf->b_efunc == NULL); 1832 hdr->b_flags |= ARC_BUF_AVAILABLE; 1833 } 1834 mutex_exit(hash_lock); 1835 } else if (HDR_IO_IN_PROGRESS(hdr)) { 1836 int destroy_hdr; 1837 /* 1838 * We are in the middle of an async write. Don't destroy 1839 * this buffer unless the write completes before we finish 1840 * decrementing the reference count. 1841 */ 1842 mutex_enter(&arc_eviction_mtx); 1843 (void) remove_reference(hdr, NULL, tag); 1844 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1845 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1846 mutex_exit(&arc_eviction_mtx); 1847 if (destroy_hdr) 1848 arc_hdr_destroy(hdr); 1849 } else { 1850 if (remove_reference(hdr, NULL, tag) > 0) 1851 arc_buf_destroy(buf, FALSE, TRUE); 1852 else 1853 arc_hdr_destroy(hdr); 1854 } 1855} 1856 1857boolean_t 1858arc_buf_remove_ref(arc_buf_t *buf, void* tag) 1859{ 1860 arc_buf_hdr_t *hdr = buf->b_hdr; 1861 kmutex_t *hash_lock = HDR_LOCK(hdr); 1862 boolean_t no_callback = (buf->b_efunc == NULL); 1863 1864 if (hdr->b_state == arc_anon) { 1865 ASSERT(hdr->b_datacnt == 1); 1866 arc_buf_free(buf, tag); 1867 return (no_callback); 1868 } 1869 1870 mutex_enter(hash_lock); 1871 hdr = buf->b_hdr; 1872 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1873 ASSERT(hdr->b_state != arc_anon); 1874 ASSERT(buf->b_data != NULL); 1875 1876 (void) remove_reference(hdr, hash_lock, tag); 1877 if (hdr->b_datacnt > 1) { 1878 if (no_callback) 1879 arc_buf_destroy(buf, FALSE, TRUE); 1880 } else if (no_callback) { 1881 ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1882 ASSERT(buf->b_efunc == NULL); 1883 hdr->b_flags |= ARC_BUF_AVAILABLE; 1884 } 1885 ASSERT(no_callback || hdr->b_datacnt > 1 || 1886 refcount_is_zero(&hdr->b_refcnt)); 1887 mutex_exit(hash_lock); 1888 return (no_callback); 1889} 1890 1891int 1892arc_buf_size(arc_buf_t *buf) 1893{ 1894 return (buf->b_hdr->b_size); 1895} 1896 1897/* 1898 * Called from the DMU to determine if the current buffer should be 1899 * evicted. In order to ensure proper locking, the eviction must be initiated 1900 * from the DMU. Return true if the buffer is associated with user data and 1901 * duplicate buffers still exist. 1902 */ 1903boolean_t 1904arc_buf_eviction_needed(arc_buf_t *buf) 1905{ 1906 arc_buf_hdr_t *hdr; 1907 boolean_t evict_needed = B_FALSE; 1908 1909 if (zfs_disable_dup_eviction) 1910 return (B_FALSE); 1911 1912 mutex_enter(&buf->b_evict_lock); 1913 hdr = buf->b_hdr; 1914 if (hdr == NULL) { 1915 /* 1916 * We are in arc_do_user_evicts(); let that function 1917 * perform the eviction. 1918 */ 1919 ASSERT(buf->b_data == NULL); 1920 mutex_exit(&buf->b_evict_lock); 1921 return (B_FALSE); 1922 } else if (buf->b_data == NULL) { 1923 /* 1924 * We have already been added to the arc eviction list; 1925 * recommend eviction. 1926 */ 1927 ASSERT3P(hdr, ==, &arc_eviction_hdr); 1928 mutex_exit(&buf->b_evict_lock); 1929 return (B_TRUE); 1930 } 1931 1932 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 1933 evict_needed = B_TRUE; 1934 1935 mutex_exit(&buf->b_evict_lock); 1936 return (evict_needed); 1937} 1938 1939/* 1940 * Evict buffers from list until we've removed the specified number of 1941 * bytes. Move the removed buffers to the appropriate evict state. 1942 * If the recycle flag is set, then attempt to "recycle" a buffer: 1943 * - look for a buffer to evict that is `bytes' long. 1944 * - return the data block from this buffer rather than freeing it. 1945 * This flag is used by callers that are trying to make space for a 1946 * new buffer in a full arc cache. 1947 * 1948 * This function makes a "best effort". It skips over any buffers 1949 * it can't get a hash_lock on, and so may not catch all candidates. 1950 * It may also return without evicting as much space as requested. 1951 */ 1952static void * 1953arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 1954 arc_buf_contents_t type) 1955{ 1956 arc_state_t *evicted_state; 1957 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 1958 int64_t bytes_remaining; 1959 arc_buf_hdr_t *ab, *ab_prev = NULL; 1960 list_t *evicted_list, *list, *evicted_list_start, *list_start; 1961 kmutex_t *lock, *evicted_lock; 1962 kmutex_t *hash_lock; 1963 boolean_t have_lock; 1964 void *stolen = NULL; 1965 arc_buf_hdr_t marker = { 0 }; 1966 int count = 0; 1967 static int evict_metadata_offset, evict_data_offset; 1968 int i, idx, offset, list_count, lists; 1969 1970 ASSERT(state == arc_mru || state == arc_mfu); 1971 1972 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 1973 1974 if (type == ARC_BUFC_METADATA) { 1975 offset = 0; 1976 list_count = ARC_BUFC_NUMMETADATALISTS; 1977 list_start = &state->arcs_lists[0]; 1978 evicted_list_start = &evicted_state->arcs_lists[0]; 1979 idx = evict_metadata_offset; 1980 } else { 1981 offset = ARC_BUFC_NUMMETADATALISTS; 1982 list_start = &state->arcs_lists[offset]; 1983 evicted_list_start = &evicted_state->arcs_lists[offset]; 1984 list_count = ARC_BUFC_NUMDATALISTS; 1985 idx = evict_data_offset; 1986 } 1987 bytes_remaining = evicted_state->arcs_lsize[type]; 1988 lists = 0; 1989 1990evict_start: 1991 list = &list_start[idx]; 1992 evicted_list = &evicted_list_start[idx]; 1993 lock = ARCS_LOCK(state, (offset + idx)); 1994 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 1995 1996 mutex_enter(lock); 1997 mutex_enter(evicted_lock); 1998 1999 for (ab = list_tail(list); ab; ab = ab_prev) { 2000 ab_prev = list_prev(list, ab); 2001 bytes_remaining -= (ab->b_size * ab->b_datacnt); 2002 /* prefetch buffers have a minimum lifespan */ 2003 if (HDR_IO_IN_PROGRESS(ab) || 2004 (spa && ab->b_spa != spa) || 2005 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 2006 ddi_get_lbolt() - ab->b_arc_access < 2007 arc_min_prefetch_lifespan)) { 2008 skipped++; 2009 continue; 2010 } 2011 /* "lookahead" for better eviction candidate */ 2012 if (recycle && ab->b_size != bytes && 2013 ab_prev && ab_prev->b_size == bytes) 2014 continue; 2015 2016 /* ignore markers */ 2017 if (ab->b_spa == 0) 2018 continue; 2019 2020 /* 2021 * It may take a long time to evict all the bufs requested. 2022 * To avoid blocking all arc activity, periodically drop 2023 * the arcs_mtx and give other threads a chance to run 2024 * before reacquiring the lock. 2025 * 2026 * If we are looking for a buffer to recycle, we are in 2027 * the hot code path, so don't sleep. 2028 */ 2029 if (!recycle && count++ > arc_evict_iterations) { 2030 list_insert_after(list, ab, &marker); 2031 mutex_exit(evicted_lock); 2032 mutex_exit(lock); 2033 kpreempt(KPREEMPT_SYNC); 2034 mutex_enter(lock); 2035 mutex_enter(evicted_lock); 2036 ab_prev = list_prev(list, &marker); 2037 list_remove(list, &marker); 2038 count = 0; 2039 continue; 2040 } 2041 2042 hash_lock = HDR_LOCK(ab); 2043 have_lock = MUTEX_HELD(hash_lock); 2044 if (have_lock || mutex_tryenter(hash_lock)) { 2045 ASSERT0(refcount_count(&ab->b_refcnt)); 2046 ASSERT(ab->b_datacnt > 0); 2047 while (ab->b_buf) { 2048 arc_buf_t *buf = ab->b_buf; 2049 if (!mutex_tryenter(&buf->b_evict_lock)) { 2050 missed += 1; 2051 break; 2052 } 2053 if (buf->b_data) { 2054 bytes_evicted += ab->b_size; 2055 if (recycle && ab->b_type == type && 2056 ab->b_size == bytes && 2057 !HDR_L2_WRITING(ab)) { 2058 stolen = buf->b_data; 2059 recycle = FALSE; 2060 } 2061 } 2062 if (buf->b_efunc) { 2063 mutex_enter(&arc_eviction_mtx); 2064 arc_buf_destroy(buf, 2065 buf->b_data == stolen, FALSE); 2066 ab->b_buf = buf->b_next; 2067 buf->b_hdr = &arc_eviction_hdr; 2068 buf->b_next = arc_eviction_list; 2069 arc_eviction_list = buf; 2070 mutex_exit(&arc_eviction_mtx); 2071 mutex_exit(&buf->b_evict_lock); 2072 } else { 2073 mutex_exit(&buf->b_evict_lock); 2074 arc_buf_destroy(buf, 2075 buf->b_data == stolen, TRUE); 2076 } 2077 } 2078 2079 if (ab->b_l2hdr) { 2080 ARCSTAT_INCR(arcstat_evict_l2_cached, 2081 ab->b_size); 2082 } else { 2083 if (l2arc_write_eligible(ab->b_spa, ab)) { 2084 ARCSTAT_INCR(arcstat_evict_l2_eligible, 2085 ab->b_size); 2086 } else { 2087 ARCSTAT_INCR( 2088 arcstat_evict_l2_ineligible, 2089 ab->b_size); 2090 } 2091 } 2092 2093 if (ab->b_datacnt == 0) { 2094 arc_change_state(evicted_state, ab, hash_lock); 2095 ASSERT(HDR_IN_HASH_TABLE(ab)); 2096 ab->b_flags |= ARC_IN_HASH_TABLE; 2097 ab->b_flags &= ~ARC_BUF_AVAILABLE; 2098 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 2099 } 2100 if (!have_lock) 2101 mutex_exit(hash_lock); 2102 if (bytes >= 0 && bytes_evicted >= bytes) 2103 break; 2104 if (bytes_remaining > 0) { 2105 mutex_exit(evicted_lock); 2106 mutex_exit(lock); 2107 idx = ((idx + 1) & (list_count - 1)); 2108 lists++; 2109 goto evict_start; 2110 } 2111 } else { 2112 missed += 1; 2113 } 2114 } 2115 2116 mutex_exit(evicted_lock); 2117 mutex_exit(lock); 2118 2119 idx = ((idx + 1) & (list_count - 1)); 2120 lists++; 2121 2122 if (bytes_evicted < bytes) { 2123 if (lists < list_count) 2124 goto evict_start; 2125 else 2126 dprintf("only evicted %lld bytes from %x", 2127 (longlong_t)bytes_evicted, state); 2128 } 2129 if (type == ARC_BUFC_METADATA) 2130 evict_metadata_offset = idx; 2131 else 2132 evict_data_offset = idx; 2133 2134 if (skipped) 2135 ARCSTAT_INCR(arcstat_evict_skip, skipped); 2136 2137 if (missed) 2138 ARCSTAT_INCR(arcstat_mutex_miss, missed); 2139 2140 /* 2141 * Note: we have just evicted some data into the ghost state, 2142 * potentially putting the ghost size over the desired size. Rather 2143 * that evicting from the ghost list in this hot code path, leave 2144 * this chore to the arc_reclaim_thread(). 2145 */ 2146 2147 if (stolen) 2148 ARCSTAT_BUMP(arcstat_stolen); 2149 return (stolen); 2150} 2151 2152/* 2153 * Remove buffers from list until we've removed the specified number of 2154 * bytes. Destroy the buffers that are removed. 2155 */ 2156static void 2157arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2158{ 2159 arc_buf_hdr_t *ab, *ab_prev; 2160 arc_buf_hdr_t marker = { 0 }; 2161 list_t *list, *list_start; 2162 kmutex_t *hash_lock, *lock; 2163 uint64_t bytes_deleted = 0; 2164 uint64_t bufs_skipped = 0; 2165 int count = 0; 2166 static int evict_offset; 2167 int list_count, idx = evict_offset; 2168 int offset, lists = 0; 2169 2170 ASSERT(GHOST_STATE(state)); 2171 2172 /* 2173 * data lists come after metadata lists 2174 */ 2175 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2176 list_count = ARC_BUFC_NUMDATALISTS; 2177 offset = ARC_BUFC_NUMMETADATALISTS; 2178 2179evict_start: 2180 list = &list_start[idx]; 2181 lock = ARCS_LOCK(state, idx + offset); 2182 2183 mutex_enter(lock); 2184 for (ab = list_tail(list); ab; ab = ab_prev) { 2185 ab_prev = list_prev(list, ab); 2186 if (ab->b_type > ARC_BUFC_NUMTYPES) 2187 panic("invalid ab=%p", (void *)ab); 2188 if (spa && ab->b_spa != spa) 2189 continue; 2190 2191 /* ignore markers */ 2192 if (ab->b_spa == 0) 2193 continue; 2194 2195 hash_lock = HDR_LOCK(ab); 2196 /* caller may be trying to modify this buffer, skip it */ 2197 if (MUTEX_HELD(hash_lock)) 2198 continue; 2199 2200 /* 2201 * It may take a long time to evict all the bufs requested. 2202 * To avoid blocking all arc activity, periodically drop 2203 * the arcs_mtx and give other threads a chance to run 2204 * before reacquiring the lock. 2205 */ 2206 if (count++ > arc_evict_iterations) { 2207 list_insert_after(list, ab, &marker); 2208 mutex_exit(lock); 2209 kpreempt(KPREEMPT_SYNC); 2210 mutex_enter(lock); 2211 ab_prev = list_prev(list, &marker); 2212 list_remove(list, &marker); 2213 count = 0; 2214 continue; 2215 } 2216 if (mutex_tryenter(hash_lock)) { 2217 ASSERT(!HDR_IO_IN_PROGRESS(ab)); 2218 ASSERT(ab->b_buf == NULL); 2219 ARCSTAT_BUMP(arcstat_deleted); 2220 bytes_deleted += ab->b_size; 2221 2222 if (ab->b_l2hdr != NULL) { 2223 /* 2224 * This buffer is cached on the 2nd Level ARC; 2225 * don't destroy the header. 2226 */ 2227 arc_change_state(arc_l2c_only, ab, hash_lock); 2228 mutex_exit(hash_lock); 2229 } else { 2230 arc_change_state(arc_anon, ab, hash_lock); 2231 mutex_exit(hash_lock); 2232 arc_hdr_destroy(ab); 2233 } 2234 2235 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 2236 if (bytes >= 0 && bytes_deleted >= bytes) 2237 break; 2238 } else if (bytes < 0) { 2239 /* 2240 * Insert a list marker and then wait for the 2241 * hash lock to become available. Once its 2242 * available, restart from where we left off. 2243 */ 2244 list_insert_after(list, ab, &marker); 2245 mutex_exit(lock); 2246 mutex_enter(hash_lock); 2247 mutex_exit(hash_lock); 2248 mutex_enter(lock); 2249 ab_prev = list_prev(list, &marker); 2250 list_remove(list, &marker); 2251 } else { 2252 bufs_skipped += 1; 2253 } 2254 2255 } 2256 mutex_exit(lock); 2257 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2258 lists++; 2259 2260 if (lists < list_count) 2261 goto evict_start; 2262 2263 evict_offset = idx; 2264 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2265 (bytes < 0 || bytes_deleted < bytes)) { 2266 list_start = &state->arcs_lists[0]; 2267 list_count = ARC_BUFC_NUMMETADATALISTS; 2268 offset = lists = 0; 2269 goto evict_start; 2270 } 2271 2272 if (bufs_skipped) { 2273 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2274 ASSERT(bytes >= 0); 2275 } 2276 2277 if (bytes_deleted < bytes) 2278 dprintf("only deleted %lld bytes from %p", 2279 (longlong_t)bytes_deleted, state); 2280} 2281 2282static void 2283arc_adjust(void) 2284{ 2285 int64_t adjustment, delta; 2286 2287 /* 2288 * Adjust MRU size 2289 */ 2290 2291 adjustment = MIN((int64_t)(arc_size - arc_c), 2292 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2293 arc_p)); 2294 2295 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2296 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2297 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2298 adjustment -= delta; 2299 } 2300 2301 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2302 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2303 (void) arc_evict(arc_mru, 0, delta, FALSE, 2304 ARC_BUFC_METADATA); 2305 } 2306 2307 /* 2308 * Adjust MFU size 2309 */ 2310 2311 adjustment = arc_size - arc_c; 2312 2313 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2314 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2315 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2316 adjustment -= delta; 2317 } 2318 2319 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2320 int64_t delta = MIN(adjustment, 2321 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2322 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2323 ARC_BUFC_METADATA); 2324 } 2325 2326 /* 2327 * Adjust ghost lists 2328 */ 2329 2330 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2331 2332 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2333 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2334 arc_evict_ghost(arc_mru_ghost, 0, delta); 2335 } 2336 2337 adjustment = 2338 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2339 2340 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2341 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2342 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2343 } 2344} 2345 2346static void 2347arc_do_user_evicts(void) 2348{ 2349 static arc_buf_t *tmp_arc_eviction_list; 2350 2351 /* 2352 * Move list over to avoid LOR 2353 */ 2354restart: 2355 mutex_enter(&arc_eviction_mtx); 2356 tmp_arc_eviction_list = arc_eviction_list; 2357 arc_eviction_list = NULL; 2358 mutex_exit(&arc_eviction_mtx); 2359 2360 while (tmp_arc_eviction_list != NULL) { 2361 arc_buf_t *buf = tmp_arc_eviction_list; 2362 tmp_arc_eviction_list = buf->b_next; 2363 mutex_enter(&buf->b_evict_lock); 2364 buf->b_hdr = NULL; 2365 mutex_exit(&buf->b_evict_lock); 2366 2367 if (buf->b_efunc != NULL) 2368 VERIFY0(buf->b_efunc(buf->b_private)); 2369 2370 buf->b_efunc = NULL; 2371 buf->b_private = NULL; 2372 kmem_cache_free(buf_cache, buf); 2373 } 2374 2375 if (arc_eviction_list != NULL) 2376 goto restart; 2377} 2378 2379/* 2380 * Flush all *evictable* data from the cache for the given spa. 2381 * NOTE: this will not touch "active" (i.e. referenced) data. 2382 */ 2383void 2384arc_flush(spa_t *spa) 2385{ 2386 uint64_t guid = 0; 2387 2388 if (spa) 2389 guid = spa_load_guid(spa); 2390 2391 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2392 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2393 if (spa) 2394 break; 2395 } 2396 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2397 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2398 if (spa) 2399 break; 2400 } 2401 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2402 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2403 if (spa) 2404 break; 2405 } 2406 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2407 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2408 if (spa) 2409 break; 2410 } 2411 2412 arc_evict_ghost(arc_mru_ghost, guid, -1); 2413 arc_evict_ghost(arc_mfu_ghost, guid, -1); 2414 2415 mutex_enter(&arc_reclaim_thr_lock); 2416 arc_do_user_evicts(); 2417 mutex_exit(&arc_reclaim_thr_lock); 2418 ASSERT(spa || arc_eviction_list == NULL); 2419} 2420 2421void 2422arc_shrink(void) 2423{ 2424 if (arc_c > arc_c_min) { 2425 uint64_t to_free; 2426 2427#ifdef _KERNEL 2428 to_free = arc_c >> arc_shrink_shift; 2429#else 2430 to_free = arc_c >> arc_shrink_shift; 2431#endif 2432 if (arc_c > arc_c_min + to_free) 2433 atomic_add_64(&arc_c, -to_free); 2434 else 2435 arc_c = arc_c_min; 2436 2437 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2438 if (arc_c > arc_size) 2439 arc_c = MAX(arc_size, arc_c_min); 2440 if (arc_p > arc_c) 2441 arc_p = (arc_c >> 1); 2442 ASSERT(arc_c >= arc_c_min); 2443 ASSERT((int64_t)arc_p >= 0); 2444 } 2445 2446 if (arc_size > arc_c) 2447 arc_adjust(); 2448} 2449 2450static int needfree = 0; 2451 2452static int 2453arc_reclaim_needed(void) 2454{ 2455 2456#ifdef _KERNEL 2457 2458 if (needfree) 2459 return (1); 2460 2461 /* 2462 * Cooperate with pagedaemon when it's time for it to scan 2463 * and reclaim some pages. 2464 */ 2465 if (vm_paging_needed()) 2466 return (1); 2467 2468#ifdef sun 2469 /* 2470 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2471 */ 2472 extra = desfree; 2473 2474 /* 2475 * check that we're out of range of the pageout scanner. It starts to 2476 * schedule paging if freemem is less than lotsfree and needfree. 2477 * lotsfree is the high-water mark for pageout, and needfree is the 2478 * number of needed free pages. We add extra pages here to make sure 2479 * the scanner doesn't start up while we're freeing memory. 2480 */ 2481 if (freemem < lotsfree + needfree + extra) 2482 return (1); 2483 2484 /* 2485 * check to make sure that swapfs has enough space so that anon 2486 * reservations can still succeed. anon_resvmem() checks that the 2487 * availrmem is greater than swapfs_minfree, and the number of reserved 2488 * swap pages. We also add a bit of extra here just to prevent 2489 * circumstances from getting really dire. 2490 */ 2491 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2492 return (1); 2493 2494#if defined(__i386) 2495 /* 2496 * If we're on an i386 platform, it's possible that we'll exhaust the 2497 * kernel heap space before we ever run out of available physical 2498 * memory. Most checks of the size of the heap_area compare against 2499 * tune.t_minarmem, which is the minimum available real memory that we 2500 * can have in the system. However, this is generally fixed at 25 pages 2501 * which is so low that it's useless. In this comparison, we seek to 2502 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2503 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2504 * free) 2505 */ 2506 if (btop(vmem_size(heap_arena, VMEM_FREE)) < 2507 (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2)) 2508 return (1); 2509#endif 2510#else /* !sun */ 2511 if (kmem_used() > (kmem_size() * 3) / 4) 2512 return (1); 2513#endif /* sun */ 2514 2515#else 2516 if (spa_get_random(100) == 0) 2517 return (1); 2518#endif 2519 return (0); 2520} 2521 2522extern kmem_cache_t *zio_buf_cache[]; 2523extern kmem_cache_t *zio_data_buf_cache[]; 2524 2525static void 2526arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2527{ 2528 size_t i; 2529 kmem_cache_t *prev_cache = NULL; 2530 kmem_cache_t *prev_data_cache = NULL; 2531 2532#ifdef _KERNEL 2533 if (arc_meta_used >= arc_meta_limit) { 2534 /* 2535 * We are exceeding our meta-data cache limit. 2536 * Purge some DNLC entries to release holds on meta-data. 2537 */ 2538 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2539 } 2540#if defined(__i386) 2541 /* 2542 * Reclaim unused memory from all kmem caches. 2543 */ 2544 kmem_reap(); 2545#endif 2546#endif 2547 2548 /* 2549 * An aggressive reclamation will shrink the cache size as well as 2550 * reap free buffers from the arc kmem caches. 2551 */ 2552 if (strat == ARC_RECLAIM_AGGR) 2553 arc_shrink(); 2554 2555 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2556 if (zio_buf_cache[i] != prev_cache) { 2557 prev_cache = zio_buf_cache[i]; 2558 kmem_cache_reap_now(zio_buf_cache[i]); 2559 } 2560 if (zio_data_buf_cache[i] != prev_data_cache) { 2561 prev_data_cache = zio_data_buf_cache[i]; 2562 kmem_cache_reap_now(zio_data_buf_cache[i]); 2563 } 2564 } 2565 kmem_cache_reap_now(buf_cache); 2566 kmem_cache_reap_now(hdr_cache); 2567} 2568 2569static void 2570arc_reclaim_thread(void *dummy __unused) 2571{ 2572 clock_t growtime = 0; 2573 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2574 callb_cpr_t cpr; 2575 2576 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2577 2578 mutex_enter(&arc_reclaim_thr_lock); 2579 while (arc_thread_exit == 0) { 2580 if (arc_reclaim_needed()) { 2581 2582 if (arc_no_grow) { 2583 if (last_reclaim == ARC_RECLAIM_CONS) { 2584 last_reclaim = ARC_RECLAIM_AGGR; 2585 } else { 2586 last_reclaim = ARC_RECLAIM_CONS; 2587 } 2588 } else { 2589 arc_no_grow = TRUE; 2590 last_reclaim = ARC_RECLAIM_AGGR; 2591 membar_producer(); 2592 } 2593 2594 /* reset the growth delay for every reclaim */ 2595 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2596 2597 if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2598 /* 2599 * If needfree is TRUE our vm_lowmem hook 2600 * was called and in that case we must free some 2601 * memory, so switch to aggressive mode. 2602 */ 2603 arc_no_grow = TRUE; 2604 last_reclaim = ARC_RECLAIM_AGGR; 2605 } 2606 arc_kmem_reap_now(last_reclaim); 2607 arc_warm = B_TRUE; 2608 2609 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2610 arc_no_grow = FALSE; 2611 } 2612 2613 arc_adjust(); 2614 2615 if (arc_eviction_list != NULL) 2616 arc_do_user_evicts(); 2617 2618#ifdef _KERNEL 2619 if (needfree) { 2620 needfree = 0; 2621 wakeup(&needfree); 2622 } 2623#endif 2624 2625 /* block until needed, or one second, whichever is shorter */ 2626 CALLB_CPR_SAFE_BEGIN(&cpr); 2627 (void) cv_timedwait(&arc_reclaim_thr_cv, 2628 &arc_reclaim_thr_lock, hz); 2629 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2630 } 2631 2632 arc_thread_exit = 0; 2633 cv_broadcast(&arc_reclaim_thr_cv); 2634 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2635 thread_exit(); 2636} 2637 2638/* 2639 * Adapt arc info given the number of bytes we are trying to add and 2640 * the state that we are comming from. This function is only called 2641 * when we are adding new content to the cache. 2642 */ 2643static void 2644arc_adapt(int bytes, arc_state_t *state) 2645{ 2646 int mult; 2647 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2648 2649 if (state == arc_l2c_only) 2650 return; 2651 2652 ASSERT(bytes > 0); 2653 /* 2654 * Adapt the target size of the MRU list: 2655 * - if we just hit in the MRU ghost list, then increase 2656 * the target size of the MRU list. 2657 * - if we just hit in the MFU ghost list, then increase 2658 * the target size of the MFU list by decreasing the 2659 * target size of the MRU list. 2660 */ 2661 if (state == arc_mru_ghost) { 2662 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2663 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2664 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2665 2666 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2667 } else if (state == arc_mfu_ghost) { 2668 uint64_t delta; 2669 2670 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2671 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2672 mult = MIN(mult, 10); 2673 2674 delta = MIN(bytes * mult, arc_p); 2675 arc_p = MAX(arc_p_min, arc_p - delta); 2676 } 2677 ASSERT((int64_t)arc_p >= 0); 2678 2679 if (arc_reclaim_needed()) { 2680 cv_signal(&arc_reclaim_thr_cv); 2681 return; 2682 } 2683 2684 if (arc_no_grow) 2685 return; 2686 2687 if (arc_c >= arc_c_max) 2688 return; 2689 2690 /* 2691 * If we're within (2 * maxblocksize) bytes of the target 2692 * cache size, increment the target cache size 2693 */ 2694 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2695 atomic_add_64(&arc_c, (int64_t)bytes); 2696 if (arc_c > arc_c_max) 2697 arc_c = arc_c_max; 2698 else if (state == arc_anon) 2699 atomic_add_64(&arc_p, (int64_t)bytes); 2700 if (arc_p > arc_c) 2701 arc_p = arc_c; 2702 } 2703 ASSERT((int64_t)arc_p >= 0); 2704} 2705 2706/* 2707 * Check if the cache has reached its limits and eviction is required 2708 * prior to insert. 2709 */ 2710static int 2711arc_evict_needed(arc_buf_contents_t type) 2712{ 2713 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2714 return (1); 2715 2716#ifdef sun 2717#ifdef _KERNEL 2718 /* 2719 * If zio data pages are being allocated out of a separate heap segment, 2720 * then enforce that the size of available vmem for this area remains 2721 * above about 1/32nd free. 2722 */ 2723 if (type == ARC_BUFC_DATA && zio_arena != NULL && 2724 vmem_size(zio_arena, VMEM_FREE) < 2725 (vmem_size(zio_arena, VMEM_ALLOC) >> 5)) 2726 return (1); 2727#endif 2728#endif /* sun */ 2729 2730 if (arc_reclaim_needed()) 2731 return (1); 2732 2733 return (arc_size > arc_c); 2734} 2735 2736/* 2737 * The buffer, supplied as the first argument, needs a data block. 2738 * So, if we are at cache max, determine which cache should be victimized. 2739 * We have the following cases: 2740 * 2741 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2742 * In this situation if we're out of space, but the resident size of the MFU is 2743 * under the limit, victimize the MFU cache to satisfy this insertion request. 2744 * 2745 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2746 * Here, we've used up all of the available space for the MRU, so we need to 2747 * evict from our own cache instead. Evict from the set of resident MRU 2748 * entries. 2749 * 2750 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2751 * c minus p represents the MFU space in the cache, since p is the size of the 2752 * cache that is dedicated to the MRU. In this situation there's still space on 2753 * the MFU side, so the MRU side needs to be victimized. 2754 * 2755 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2756 * MFU's resident set is consuming more space than it has been allotted. In 2757 * this situation, we must victimize our own cache, the MFU, for this insertion. 2758 */ 2759static void 2760arc_get_data_buf(arc_buf_t *buf) 2761{ 2762 arc_state_t *state = buf->b_hdr->b_state; 2763 uint64_t size = buf->b_hdr->b_size; 2764 arc_buf_contents_t type = buf->b_hdr->b_type; 2765 2766 arc_adapt(size, state); 2767 2768 /* 2769 * We have not yet reached cache maximum size, 2770 * just allocate a new buffer. 2771 */ 2772 if (!arc_evict_needed(type)) { 2773 if (type == ARC_BUFC_METADATA) { 2774 buf->b_data = zio_buf_alloc(size); 2775 arc_space_consume(size, ARC_SPACE_DATA); 2776 } else { 2777 ASSERT(type == ARC_BUFC_DATA); 2778 buf->b_data = zio_data_buf_alloc(size); 2779 ARCSTAT_INCR(arcstat_data_size, size); 2780 atomic_add_64(&arc_size, size); 2781 } 2782 goto out; 2783 } 2784 2785 /* 2786 * If we are prefetching from the mfu ghost list, this buffer 2787 * will end up on the mru list; so steal space from there. 2788 */ 2789 if (state == arc_mfu_ghost) 2790 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2791 else if (state == arc_mru_ghost) 2792 state = arc_mru; 2793 2794 if (state == arc_mru || state == arc_anon) { 2795 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2796 state = (arc_mfu->arcs_lsize[type] >= size && 2797 arc_p > mru_used) ? arc_mfu : arc_mru; 2798 } else { 2799 /* MFU cases */ 2800 uint64_t mfu_space = arc_c - arc_p; 2801 state = (arc_mru->arcs_lsize[type] >= size && 2802 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2803 } 2804 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2805 if (type == ARC_BUFC_METADATA) { 2806 buf->b_data = zio_buf_alloc(size); 2807 arc_space_consume(size, ARC_SPACE_DATA); 2808 } else { 2809 ASSERT(type == ARC_BUFC_DATA); 2810 buf->b_data = zio_data_buf_alloc(size); 2811 ARCSTAT_INCR(arcstat_data_size, size); 2812 atomic_add_64(&arc_size, size); 2813 } 2814 ARCSTAT_BUMP(arcstat_recycle_miss); 2815 } 2816 ASSERT(buf->b_data != NULL); 2817out: 2818 /* 2819 * Update the state size. Note that ghost states have a 2820 * "ghost size" and so don't need to be updated. 2821 */ 2822 if (!GHOST_STATE(buf->b_hdr->b_state)) { 2823 arc_buf_hdr_t *hdr = buf->b_hdr; 2824 2825 atomic_add_64(&hdr->b_state->arcs_size, size); 2826 if (list_link_active(&hdr->b_arc_node)) { 2827 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2828 atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2829 } 2830 /* 2831 * If we are growing the cache, and we are adding anonymous 2832 * data, and we have outgrown arc_p, update arc_p 2833 */ 2834 if (arc_size < arc_c && hdr->b_state == arc_anon && 2835 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2836 arc_p = MIN(arc_c, arc_p + size); 2837 } 2838 ARCSTAT_BUMP(arcstat_allocated); 2839} 2840 2841/* 2842 * This routine is called whenever a buffer is accessed. 2843 * NOTE: the hash lock is dropped in this function. 2844 */ 2845static void 2846arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2847{ 2848 clock_t now; 2849 2850 ASSERT(MUTEX_HELD(hash_lock)); 2851 2852 if (buf->b_state == arc_anon) { 2853 /* 2854 * This buffer is not in the cache, and does not 2855 * appear in our "ghost" list. Add the new buffer 2856 * to the MRU state. 2857 */ 2858 2859 ASSERT(buf->b_arc_access == 0); 2860 buf->b_arc_access = ddi_get_lbolt(); 2861 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2862 arc_change_state(arc_mru, buf, hash_lock); 2863 2864 } else if (buf->b_state == arc_mru) { 2865 now = ddi_get_lbolt(); 2866 2867 /* 2868 * If this buffer is here because of a prefetch, then either: 2869 * - clear the flag if this is a "referencing" read 2870 * (any subsequent access will bump this into the MFU state). 2871 * or 2872 * - move the buffer to the head of the list if this is 2873 * another prefetch (to make it less likely to be evicted). 2874 */ 2875 if ((buf->b_flags & ARC_PREFETCH) != 0) { 2876 if (refcount_count(&buf->b_refcnt) == 0) { 2877 ASSERT(list_link_active(&buf->b_arc_node)); 2878 } else { 2879 buf->b_flags &= ~ARC_PREFETCH; 2880 ARCSTAT_BUMP(arcstat_mru_hits); 2881 } 2882 buf->b_arc_access = now; 2883 return; 2884 } 2885 2886 /* 2887 * This buffer has been "accessed" only once so far, 2888 * but it is still in the cache. Move it to the MFU 2889 * state. 2890 */ 2891 if (now > buf->b_arc_access + ARC_MINTIME) { 2892 /* 2893 * More than 125ms have passed since we 2894 * instantiated this buffer. Move it to the 2895 * most frequently used state. 2896 */ 2897 buf->b_arc_access = now; 2898 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2899 arc_change_state(arc_mfu, buf, hash_lock); 2900 } 2901 ARCSTAT_BUMP(arcstat_mru_hits); 2902 } else if (buf->b_state == arc_mru_ghost) { 2903 arc_state_t *new_state; 2904 /* 2905 * This buffer has been "accessed" recently, but 2906 * was evicted from the cache. Move it to the 2907 * MFU state. 2908 */ 2909 2910 if (buf->b_flags & ARC_PREFETCH) { 2911 new_state = arc_mru; 2912 if (refcount_count(&buf->b_refcnt) > 0) 2913 buf->b_flags &= ~ARC_PREFETCH; 2914 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2915 } else { 2916 new_state = arc_mfu; 2917 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2918 } 2919 2920 buf->b_arc_access = ddi_get_lbolt(); 2921 arc_change_state(new_state, buf, hash_lock); 2922 2923 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 2924 } else if (buf->b_state == arc_mfu) { 2925 /* 2926 * This buffer has been accessed more than once and is 2927 * still in the cache. Keep it in the MFU state. 2928 * 2929 * NOTE: an add_reference() that occurred when we did 2930 * the arc_read() will have kicked this off the list. 2931 * If it was a prefetch, we will explicitly move it to 2932 * the head of the list now. 2933 */ 2934 if ((buf->b_flags & ARC_PREFETCH) != 0) { 2935 ASSERT(refcount_count(&buf->b_refcnt) == 0); 2936 ASSERT(list_link_active(&buf->b_arc_node)); 2937 } 2938 ARCSTAT_BUMP(arcstat_mfu_hits); 2939 buf->b_arc_access = ddi_get_lbolt(); 2940 } else if (buf->b_state == arc_mfu_ghost) { 2941 arc_state_t *new_state = arc_mfu; 2942 /* 2943 * This buffer has been accessed more than once but has 2944 * been evicted from the cache. Move it back to the 2945 * MFU state. 2946 */ 2947 2948 if (buf->b_flags & ARC_PREFETCH) { 2949 /* 2950 * This is a prefetch access... 2951 * move this block back to the MRU state. 2952 */ 2953 ASSERT0(refcount_count(&buf->b_refcnt)); 2954 new_state = arc_mru; 2955 } 2956 2957 buf->b_arc_access = ddi_get_lbolt(); 2958 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2959 arc_change_state(new_state, buf, hash_lock); 2960 2961 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 2962 } else if (buf->b_state == arc_l2c_only) { 2963 /* 2964 * This buffer is on the 2nd Level ARC. 2965 */ 2966 2967 buf->b_arc_access = ddi_get_lbolt(); 2968 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2969 arc_change_state(arc_mfu, buf, hash_lock); 2970 } else { 2971 ASSERT(!"invalid arc state"); 2972 } 2973} 2974 2975/* a generic arc_done_func_t which you can use */ 2976/* ARGSUSED */ 2977void 2978arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 2979{ 2980 if (zio == NULL || zio->io_error == 0) 2981 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 2982 VERIFY(arc_buf_remove_ref(buf, arg)); 2983} 2984 2985/* a generic arc_done_func_t */ 2986void 2987arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 2988{ 2989 arc_buf_t **bufp = arg; 2990 if (zio && zio->io_error) { 2991 VERIFY(arc_buf_remove_ref(buf, arg)); 2992 *bufp = NULL; 2993 } else { 2994 *bufp = buf; 2995 ASSERT(buf->b_data); 2996 } 2997} 2998 2999static void 3000arc_read_done(zio_t *zio) 3001{ 3002 arc_buf_hdr_t *hdr; 3003 arc_buf_t *buf; 3004 arc_buf_t *abuf; /* buffer we're assigning to callback */ 3005 kmutex_t *hash_lock = NULL; 3006 arc_callback_t *callback_list, *acb; 3007 int freeable = FALSE; 3008 3009 buf = zio->io_private; 3010 hdr = buf->b_hdr; 3011 3012 /* 3013 * The hdr was inserted into hash-table and removed from lists 3014 * prior to starting I/O. We should find this header, since 3015 * it's in the hash table, and it should be legit since it's 3016 * not possible to evict it during the I/O. The only possible 3017 * reason for it not to be found is if we were freed during the 3018 * read. 3019 */ 3020 if (HDR_IN_HASH_TABLE(hdr)) { 3021 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3022 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3023 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3024 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3025 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3026 3027 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3028 &hash_lock); 3029 3030 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3031 hash_lock == NULL) || 3032 (found == hdr && 3033 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3034 (found == hdr && HDR_L2_READING(hdr))); 3035 } 3036 3037 hdr->b_flags &= ~ARC_L2_EVICTED; 3038 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 3039 hdr->b_flags &= ~ARC_L2CACHE; 3040 3041 /* byteswap if necessary */ 3042 callback_list = hdr->b_acb; 3043 ASSERT(callback_list != NULL); 3044 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3045 dmu_object_byteswap_t bswap = 3046 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3047 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3048 byteswap_uint64_array : 3049 dmu_ot_byteswap[bswap].ob_func; 3050 func(buf->b_data, hdr->b_size); 3051 } 3052 3053 arc_cksum_compute(buf, B_FALSE); 3054#ifdef illumos 3055 arc_buf_watch(buf); 3056#endif /* illumos */ 3057 3058 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3059 /* 3060 * Only call arc_access on anonymous buffers. This is because 3061 * if we've issued an I/O for an evicted buffer, we've already 3062 * called arc_access (to prevent any simultaneous readers from 3063 * getting confused). 3064 */ 3065 arc_access(hdr, hash_lock); 3066 } 3067 3068 /* create copies of the data buffer for the callers */ 3069 abuf = buf; 3070 for (acb = callback_list; acb; acb = acb->acb_next) { 3071 if (acb->acb_done) { 3072 if (abuf == NULL) { 3073 ARCSTAT_BUMP(arcstat_duplicate_reads); 3074 abuf = arc_buf_clone(buf); 3075 } 3076 acb->acb_buf = abuf; 3077 abuf = NULL; 3078 } 3079 } 3080 hdr->b_acb = NULL; 3081 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3082 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3083 if (abuf == buf) { 3084 ASSERT(buf->b_efunc == NULL); 3085 ASSERT(hdr->b_datacnt == 1); 3086 hdr->b_flags |= ARC_BUF_AVAILABLE; 3087 } 3088 3089 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3090 3091 if (zio->io_error != 0) { 3092 hdr->b_flags |= ARC_IO_ERROR; 3093 if (hdr->b_state != arc_anon) 3094 arc_change_state(arc_anon, hdr, hash_lock); 3095 if (HDR_IN_HASH_TABLE(hdr)) 3096 buf_hash_remove(hdr); 3097 freeable = refcount_is_zero(&hdr->b_refcnt); 3098 } 3099 3100 /* 3101 * Broadcast before we drop the hash_lock to avoid the possibility 3102 * that the hdr (and hence the cv) might be freed before we get to 3103 * the cv_broadcast(). 3104 */ 3105 cv_broadcast(&hdr->b_cv); 3106 3107 if (hash_lock) { 3108 mutex_exit(hash_lock); 3109 } else { 3110 /* 3111 * This block was freed while we waited for the read to 3112 * complete. It has been removed from the hash table and 3113 * moved to the anonymous state (so that it won't show up 3114 * in the cache). 3115 */ 3116 ASSERT3P(hdr->b_state, ==, arc_anon); 3117 freeable = refcount_is_zero(&hdr->b_refcnt); 3118 } 3119 3120 /* execute each callback and free its structure */ 3121 while ((acb = callback_list) != NULL) { 3122 if (acb->acb_done) 3123 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3124 3125 if (acb->acb_zio_dummy != NULL) { 3126 acb->acb_zio_dummy->io_error = zio->io_error; 3127 zio_nowait(acb->acb_zio_dummy); 3128 } 3129 3130 callback_list = acb->acb_next; 3131 kmem_free(acb, sizeof (arc_callback_t)); 3132 } 3133 3134 if (freeable) 3135 arc_hdr_destroy(hdr); 3136} 3137 3138/* 3139 * "Read" the block block at the specified DVA (in bp) via the 3140 * cache. If the block is found in the cache, invoke the provided 3141 * callback immediately and return. Note that the `zio' parameter 3142 * in the callback will be NULL in this case, since no IO was 3143 * required. If the block is not in the cache pass the read request 3144 * on to the spa with a substitute callback function, so that the 3145 * requested block will be added to the cache. 3146 * 3147 * If a read request arrives for a block that has a read in-progress, 3148 * either wait for the in-progress read to complete (and return the 3149 * results); or, if this is a read with a "done" func, add a record 3150 * to the read to invoke the "done" func when the read completes, 3151 * and return; or just return. 3152 * 3153 * arc_read_done() will invoke all the requested "done" functions 3154 * for readers of this block. 3155 */ 3156int 3157arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3158 void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags, 3159 const zbookmark_phys_t *zb) 3160{ 3161 arc_buf_hdr_t *hdr = NULL; 3162 arc_buf_t *buf = NULL; 3163 kmutex_t *hash_lock = NULL; 3164 zio_t *rzio; 3165 uint64_t guid = spa_load_guid(spa); 3166 3167 ASSERT(!BP_IS_EMBEDDED(bp) || 3168 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3169 3170top: 3171 if (!BP_IS_EMBEDDED(bp)) { 3172 /* 3173 * Embedded BP's have no DVA and require no I/O to "read". 3174 * Create an anonymous arc buf to back it. 3175 */ 3176 hdr = buf_hash_find(guid, bp, &hash_lock); 3177 } 3178 3179 if (hdr != NULL && hdr->b_datacnt > 0) { 3180 3181 *arc_flags |= ARC_CACHED; 3182 3183 if (HDR_IO_IN_PROGRESS(hdr)) { 3184 3185 if (*arc_flags & ARC_WAIT) { 3186 cv_wait(&hdr->b_cv, hash_lock); 3187 mutex_exit(hash_lock); 3188 goto top; 3189 } 3190 ASSERT(*arc_flags & ARC_NOWAIT); 3191 3192 if (done) { 3193 arc_callback_t *acb = NULL; 3194 3195 acb = kmem_zalloc(sizeof (arc_callback_t), 3196 KM_SLEEP); 3197 acb->acb_done = done; 3198 acb->acb_private = private; 3199 if (pio != NULL) 3200 acb->acb_zio_dummy = zio_null(pio, 3201 spa, NULL, NULL, NULL, zio_flags); 3202 3203 ASSERT(acb->acb_done != NULL); 3204 acb->acb_next = hdr->b_acb; 3205 hdr->b_acb = acb; 3206 add_reference(hdr, hash_lock, private); 3207 mutex_exit(hash_lock); 3208 return (0); 3209 } 3210 mutex_exit(hash_lock); 3211 return (0); 3212 } 3213 3214 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3215 3216 if (done) { 3217 add_reference(hdr, hash_lock, private); 3218 /* 3219 * If this block is already in use, create a new 3220 * copy of the data so that we will be guaranteed 3221 * that arc_release() will always succeed. 3222 */ 3223 buf = hdr->b_buf; 3224 ASSERT(buf); 3225 ASSERT(buf->b_data); 3226 if (HDR_BUF_AVAILABLE(hdr)) { 3227 ASSERT(buf->b_efunc == NULL); 3228 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3229 } else { 3230 buf = arc_buf_clone(buf); 3231 } 3232 3233 } else if (*arc_flags & ARC_PREFETCH && 3234 refcount_count(&hdr->b_refcnt) == 0) { 3235 hdr->b_flags |= ARC_PREFETCH; 3236 } 3237 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3238 arc_access(hdr, hash_lock); 3239 if (*arc_flags & ARC_L2CACHE) 3240 hdr->b_flags |= ARC_L2CACHE; 3241 if (*arc_flags & ARC_L2COMPRESS) 3242 hdr->b_flags |= ARC_L2COMPRESS; 3243 mutex_exit(hash_lock); 3244 ARCSTAT_BUMP(arcstat_hits); 3245 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3246 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3247 data, metadata, hits); 3248 3249 if (done) 3250 done(NULL, buf, private); 3251 } else { 3252 uint64_t size = BP_GET_LSIZE(bp); 3253 arc_callback_t *acb; 3254 vdev_t *vd = NULL; 3255 uint64_t addr = 0; 3256 boolean_t devw = B_FALSE; 3257 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3258 uint64_t b_asize = 0; 3259 3260 if (hdr == NULL) { 3261 /* this block is not in the cache */ 3262 arc_buf_hdr_t *exists = NULL; 3263 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3264 buf = arc_buf_alloc(spa, size, private, type); 3265 hdr = buf->b_hdr; 3266 if (!BP_IS_EMBEDDED(bp)) { 3267 hdr->b_dva = *BP_IDENTITY(bp); 3268 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3269 hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3270 exists = buf_hash_insert(hdr, &hash_lock); 3271 } 3272 if (exists != NULL) { 3273 /* somebody beat us to the hash insert */ 3274 mutex_exit(hash_lock); 3275 buf_discard_identity(hdr); 3276 (void) arc_buf_remove_ref(buf, private); 3277 goto top; /* restart the IO request */ 3278 } 3279 /* if this is a prefetch, we don't have a reference */ 3280 if (*arc_flags & ARC_PREFETCH) { 3281 (void) remove_reference(hdr, hash_lock, 3282 private); 3283 hdr->b_flags |= ARC_PREFETCH; 3284 } 3285 if (*arc_flags & ARC_L2CACHE) 3286 hdr->b_flags |= ARC_L2CACHE; 3287 if (*arc_flags & ARC_L2COMPRESS) 3288 hdr->b_flags |= ARC_L2COMPRESS; 3289 if (BP_GET_LEVEL(bp) > 0) 3290 hdr->b_flags |= ARC_INDIRECT; 3291 } else { 3292 /* this block is in the ghost cache */ 3293 ASSERT(GHOST_STATE(hdr->b_state)); 3294 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3295 ASSERT0(refcount_count(&hdr->b_refcnt)); 3296 ASSERT(hdr->b_buf == NULL); 3297 3298 /* if this is a prefetch, we don't have a reference */ 3299 if (*arc_flags & ARC_PREFETCH) 3300 hdr->b_flags |= ARC_PREFETCH; 3301 else 3302 add_reference(hdr, hash_lock, private); 3303 if (*arc_flags & ARC_L2CACHE) 3304 hdr->b_flags |= ARC_L2CACHE; 3305 if (*arc_flags & ARC_L2COMPRESS) 3306 hdr->b_flags |= ARC_L2COMPRESS; 3307 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3308 buf->b_hdr = hdr; 3309 buf->b_data = NULL; 3310 buf->b_efunc = NULL; 3311 buf->b_private = NULL; 3312 buf->b_next = NULL; 3313 hdr->b_buf = buf; 3314 ASSERT(hdr->b_datacnt == 0); 3315 hdr->b_datacnt = 1; 3316 arc_get_data_buf(buf); 3317 arc_access(hdr, hash_lock); 3318 } 3319 3320 ASSERT(!GHOST_STATE(hdr->b_state)); 3321 3322 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3323 acb->acb_done = done; 3324 acb->acb_private = private; 3325 3326 ASSERT(hdr->b_acb == NULL); 3327 hdr->b_acb = acb; 3328 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3329 3330 if (hdr->b_l2hdr != NULL && 3331 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3332 devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3333 addr = hdr->b_l2hdr->b_daddr; 3334 b_compress = hdr->b_l2hdr->b_compress; 3335 b_asize = hdr->b_l2hdr->b_asize; 3336 /* 3337 * Lock out device removal. 3338 */ 3339 if (vdev_is_dead(vd) || 3340 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3341 vd = NULL; 3342 } 3343 3344 if (hash_lock != NULL) 3345 mutex_exit(hash_lock); 3346 3347 /* 3348 * At this point, we have a level 1 cache miss. Try again in 3349 * L2ARC if possible. 3350 */ 3351 ASSERT3U(hdr->b_size, ==, size); 3352 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3353 uint64_t, size, zbookmark_phys_t *, zb); 3354 ARCSTAT_BUMP(arcstat_misses); 3355 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3356 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3357 data, metadata, misses); 3358#ifdef _KERNEL 3359 curthread->td_ru.ru_inblock++; 3360#endif 3361 3362 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3363 /* 3364 * Read from the L2ARC if the following are true: 3365 * 1. The L2ARC vdev was previously cached. 3366 * 2. This buffer still has L2ARC metadata. 3367 * 3. This buffer isn't currently writing to the L2ARC. 3368 * 4. The L2ARC entry wasn't evicted, which may 3369 * also have invalidated the vdev. 3370 * 5. This isn't prefetch and l2arc_noprefetch is set. 3371 */ 3372 if (hdr->b_l2hdr != NULL && 3373 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3374 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3375 l2arc_read_callback_t *cb; 3376 3377 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3378 ARCSTAT_BUMP(arcstat_l2_hits); 3379 3380 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3381 KM_SLEEP); 3382 cb->l2rcb_buf = buf; 3383 cb->l2rcb_spa = spa; 3384 cb->l2rcb_bp = *bp; 3385 cb->l2rcb_zb = *zb; 3386 cb->l2rcb_flags = zio_flags; 3387 cb->l2rcb_compress = b_compress; 3388 3389 ASSERT(addr >= VDEV_LABEL_START_SIZE && 3390 addr + size < vd->vdev_psize - 3391 VDEV_LABEL_END_SIZE); 3392 3393 /* 3394 * l2arc read. The SCL_L2ARC lock will be 3395 * released by l2arc_read_done(). 3396 * Issue a null zio if the underlying buffer 3397 * was squashed to zero size by compression. 3398 */ 3399 if (b_compress == ZIO_COMPRESS_EMPTY) { 3400 rzio = zio_null(pio, spa, vd, 3401 l2arc_read_done, cb, 3402 zio_flags | ZIO_FLAG_DONT_CACHE | 3403 ZIO_FLAG_CANFAIL | 3404 ZIO_FLAG_DONT_PROPAGATE | 3405 ZIO_FLAG_DONT_RETRY); 3406 } else { 3407 rzio = zio_read_phys(pio, vd, addr, 3408 b_asize, buf->b_data, 3409 ZIO_CHECKSUM_OFF, 3410 l2arc_read_done, cb, priority, 3411 zio_flags | ZIO_FLAG_DONT_CACHE | 3412 ZIO_FLAG_CANFAIL | 3413 ZIO_FLAG_DONT_PROPAGATE | 3414 ZIO_FLAG_DONT_RETRY, B_FALSE); 3415 } 3416 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3417 zio_t *, rzio); 3418 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3419 3420 if (*arc_flags & ARC_NOWAIT) { 3421 zio_nowait(rzio); 3422 return (0); 3423 } 3424 3425 ASSERT(*arc_flags & ARC_WAIT); 3426 if (zio_wait(rzio) == 0) 3427 return (0); 3428 3429 /* l2arc read error; goto zio_read() */ 3430 } else { 3431 DTRACE_PROBE1(l2arc__miss, 3432 arc_buf_hdr_t *, hdr); 3433 ARCSTAT_BUMP(arcstat_l2_misses); 3434 if (HDR_L2_WRITING(hdr)) 3435 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3436 spa_config_exit(spa, SCL_L2ARC, vd); 3437 } 3438 } else { 3439 if (vd != NULL) 3440 spa_config_exit(spa, SCL_L2ARC, vd); 3441 if (l2arc_ndev != 0) { 3442 DTRACE_PROBE1(l2arc__miss, 3443 arc_buf_hdr_t *, hdr); 3444 ARCSTAT_BUMP(arcstat_l2_misses); 3445 } 3446 } 3447 3448 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3449 arc_read_done, buf, priority, zio_flags, zb); 3450 3451 if (*arc_flags & ARC_WAIT) 3452 return (zio_wait(rzio)); 3453 3454 ASSERT(*arc_flags & ARC_NOWAIT); 3455 zio_nowait(rzio); 3456 } 3457 return (0); 3458} 3459 3460void 3461arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3462{ 3463 ASSERT(buf->b_hdr != NULL); 3464 ASSERT(buf->b_hdr->b_state != arc_anon); 3465 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3466 ASSERT(buf->b_efunc == NULL); 3467 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3468 3469 buf->b_efunc = func; 3470 buf->b_private = private; 3471} 3472 3473/* 3474 * Notify the arc that a block was freed, and thus will never be used again. 3475 */ 3476void 3477arc_freed(spa_t *spa, const blkptr_t *bp) 3478{ 3479 arc_buf_hdr_t *hdr; 3480 kmutex_t *hash_lock; 3481 uint64_t guid = spa_load_guid(spa); 3482 3483 ASSERT(!BP_IS_EMBEDDED(bp)); 3484 3485 hdr = buf_hash_find(guid, bp, &hash_lock); 3486 if (hdr == NULL) 3487 return; 3488 if (HDR_BUF_AVAILABLE(hdr)) { 3489 arc_buf_t *buf = hdr->b_buf; 3490 add_reference(hdr, hash_lock, FTAG); 3491 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3492 mutex_exit(hash_lock); 3493 3494 arc_release(buf, FTAG); 3495 (void) arc_buf_remove_ref(buf, FTAG); 3496 } else { 3497 mutex_exit(hash_lock); 3498 } 3499 3500} 3501 3502/* 3503 * Clear the user eviction callback set by arc_set_callback(), first calling 3504 * it if it exists. Because the presence of a callback keeps an arc_buf cached 3505 * clearing the callback may result in the arc_buf being destroyed. However, 3506 * it will not result in the *last* arc_buf being destroyed, hence the data 3507 * will remain cached in the ARC. We make a copy of the arc buffer here so 3508 * that we can process the callback without holding any locks. 3509 * 3510 * It's possible that the callback is already in the process of being cleared 3511 * by another thread. In this case we can not clear the callback. 3512 * 3513 * Returns B_TRUE if the callback was successfully called and cleared. 3514 */ 3515boolean_t 3516arc_clear_callback(arc_buf_t *buf) 3517{ 3518 arc_buf_hdr_t *hdr; 3519 kmutex_t *hash_lock; 3520 arc_evict_func_t *efunc = buf->b_efunc; 3521 void *private = buf->b_private; 3522 list_t *list, *evicted_list; 3523 kmutex_t *lock, *evicted_lock; 3524 3525 mutex_enter(&buf->b_evict_lock); 3526 hdr = buf->b_hdr; 3527 if (hdr == NULL) { 3528 /* 3529 * We are in arc_do_user_evicts(). 3530 */ 3531 ASSERT(buf->b_data == NULL); 3532 mutex_exit(&buf->b_evict_lock); 3533 return (B_FALSE); 3534 } else if (buf->b_data == NULL) { 3535 /* 3536 * We are on the eviction list; process this buffer now 3537 * but let arc_do_user_evicts() do the reaping. 3538 */ 3539 buf->b_efunc = NULL; 3540 mutex_exit(&buf->b_evict_lock); 3541 VERIFY0(efunc(private)); 3542 return (B_TRUE); 3543 } 3544 hash_lock = HDR_LOCK(hdr); 3545 mutex_enter(hash_lock); 3546 hdr = buf->b_hdr; 3547 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3548 3549 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3550 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3551 3552 buf->b_efunc = NULL; 3553 buf->b_private = NULL; 3554 3555 if (hdr->b_datacnt > 1) { 3556 mutex_exit(&buf->b_evict_lock); 3557 arc_buf_destroy(buf, FALSE, TRUE); 3558 } else { 3559 ASSERT(buf == hdr->b_buf); 3560 hdr->b_flags |= ARC_BUF_AVAILABLE; 3561 mutex_exit(&buf->b_evict_lock); 3562 } 3563 3564 mutex_exit(hash_lock); 3565 VERIFY0(efunc(private)); 3566 return (B_TRUE); 3567} 3568 3569/* 3570 * Release this buffer from the cache, making it an anonymous buffer. This 3571 * must be done after a read and prior to modifying the buffer contents. 3572 * If the buffer has more than one reference, we must make 3573 * a new hdr for the buffer. 3574 */ 3575void 3576arc_release(arc_buf_t *buf, void *tag) 3577{ 3578 arc_buf_hdr_t *hdr; 3579 kmutex_t *hash_lock = NULL; 3580 l2arc_buf_hdr_t *l2hdr; 3581 uint64_t buf_size; 3582 3583 /* 3584 * It would be nice to assert that if it's DMU metadata (level > 3585 * 0 || it's the dnode file), then it must be syncing context. 3586 * But we don't know that information at this level. 3587 */ 3588 3589 mutex_enter(&buf->b_evict_lock); 3590 hdr = buf->b_hdr; 3591 3592 /* this buffer is not on any list */ 3593 ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3594 3595 if (hdr->b_state == arc_anon) { 3596 /* this buffer is already released */ 3597 ASSERT(buf->b_efunc == NULL); 3598 } else { 3599 hash_lock = HDR_LOCK(hdr); 3600 mutex_enter(hash_lock); 3601 hdr = buf->b_hdr; 3602 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3603 } 3604 3605 l2hdr = hdr->b_l2hdr; 3606 if (l2hdr) { 3607 mutex_enter(&l2arc_buflist_mtx); 3608 hdr->b_l2hdr = NULL; 3609 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3610 } 3611 buf_size = hdr->b_size; 3612 3613 /* 3614 * Do we have more than one buf? 3615 */ 3616 if (hdr->b_datacnt > 1) { 3617 arc_buf_hdr_t *nhdr; 3618 arc_buf_t **bufp; 3619 uint64_t blksz = hdr->b_size; 3620 uint64_t spa = hdr->b_spa; 3621 arc_buf_contents_t type = hdr->b_type; 3622 uint32_t flags = hdr->b_flags; 3623 3624 ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3625 /* 3626 * Pull the data off of this hdr and attach it to 3627 * a new anonymous hdr. 3628 */ 3629 (void) remove_reference(hdr, hash_lock, tag); 3630 bufp = &hdr->b_buf; 3631 while (*bufp != buf) 3632 bufp = &(*bufp)->b_next; 3633 *bufp = buf->b_next; 3634 buf->b_next = NULL; 3635 3636 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3637 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3638 if (refcount_is_zero(&hdr->b_refcnt)) { 3639 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3640 ASSERT3U(*size, >=, hdr->b_size); 3641 atomic_add_64(size, -hdr->b_size); 3642 } 3643 3644 /* 3645 * We're releasing a duplicate user data buffer, update 3646 * our statistics accordingly. 3647 */ 3648 if (hdr->b_type == ARC_BUFC_DATA) { 3649 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3650 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3651 -hdr->b_size); 3652 } 3653 hdr->b_datacnt -= 1; 3654 arc_cksum_verify(buf); 3655#ifdef illumos 3656 arc_buf_unwatch(buf); 3657#endif /* illumos */ 3658 3659 mutex_exit(hash_lock); 3660 3661 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3662 nhdr->b_size = blksz; 3663 nhdr->b_spa = spa; 3664 nhdr->b_type = type; 3665 nhdr->b_buf = buf; 3666 nhdr->b_state = arc_anon; 3667 nhdr->b_arc_access = 0; 3668 nhdr->b_flags = flags & ARC_L2_WRITING; 3669 nhdr->b_l2hdr = NULL; 3670 nhdr->b_datacnt = 1; 3671 nhdr->b_freeze_cksum = NULL; 3672 (void) refcount_add(&nhdr->b_refcnt, tag); 3673 buf->b_hdr = nhdr; 3674 mutex_exit(&buf->b_evict_lock); 3675 atomic_add_64(&arc_anon->arcs_size, blksz); 3676 } else { 3677 mutex_exit(&buf->b_evict_lock); 3678 ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3679 ASSERT(!list_link_active(&hdr->b_arc_node)); 3680 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3681 if (hdr->b_state != arc_anon) 3682 arc_change_state(arc_anon, hdr, hash_lock); 3683 hdr->b_arc_access = 0; 3684 if (hash_lock) 3685 mutex_exit(hash_lock); 3686 3687 buf_discard_identity(hdr); 3688 arc_buf_thaw(buf); 3689 } 3690 buf->b_efunc = NULL; 3691 buf->b_private = NULL; 3692 3693 if (l2hdr) { 3694 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3695 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 3696 -l2hdr->b_asize, 0, 0); 3697 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3698 hdr->b_size, 0); 3699 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3700 ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3701 mutex_exit(&l2arc_buflist_mtx); 3702 } 3703} 3704 3705int 3706arc_released(arc_buf_t *buf) 3707{ 3708 int released; 3709 3710 mutex_enter(&buf->b_evict_lock); 3711 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3712 mutex_exit(&buf->b_evict_lock); 3713 return (released); 3714} 3715 3716#ifdef ZFS_DEBUG 3717int 3718arc_referenced(arc_buf_t *buf) 3719{ 3720 int referenced; 3721 3722 mutex_enter(&buf->b_evict_lock); 3723 referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3724 mutex_exit(&buf->b_evict_lock); 3725 return (referenced); 3726} 3727#endif 3728 3729static void 3730arc_write_ready(zio_t *zio) 3731{ 3732 arc_write_callback_t *callback = zio->io_private; 3733 arc_buf_t *buf = callback->awcb_buf; 3734 arc_buf_hdr_t *hdr = buf->b_hdr; 3735 3736 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3737 callback->awcb_ready(zio, buf, callback->awcb_private); 3738 3739 /* 3740 * If the IO is already in progress, then this is a re-write 3741 * attempt, so we need to thaw and re-compute the cksum. 3742 * It is the responsibility of the callback to handle the 3743 * accounting for any re-write attempt. 3744 */ 3745 if (HDR_IO_IN_PROGRESS(hdr)) { 3746 mutex_enter(&hdr->b_freeze_lock); 3747 if (hdr->b_freeze_cksum != NULL) { 3748 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3749 hdr->b_freeze_cksum = NULL; 3750 } 3751 mutex_exit(&hdr->b_freeze_lock); 3752 } 3753 arc_cksum_compute(buf, B_FALSE); 3754 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3755} 3756 3757/* 3758 * The SPA calls this callback for each physical write that happens on behalf 3759 * of a logical write. See the comment in dbuf_write_physdone() for details. 3760 */ 3761static void 3762arc_write_physdone(zio_t *zio) 3763{ 3764 arc_write_callback_t *cb = zio->io_private; 3765 if (cb->awcb_physdone != NULL) 3766 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3767} 3768 3769static void 3770arc_write_done(zio_t *zio) 3771{ 3772 arc_write_callback_t *callback = zio->io_private; 3773 arc_buf_t *buf = callback->awcb_buf; 3774 arc_buf_hdr_t *hdr = buf->b_hdr; 3775 3776 ASSERT(hdr->b_acb == NULL); 3777 3778 if (zio->io_error == 0) { 3779 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 3780 buf_discard_identity(hdr); 3781 } else { 3782 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3783 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3784 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3785 } 3786 } else { 3787 ASSERT(BUF_EMPTY(hdr)); 3788 } 3789 3790 /* 3791 * If the block to be written was all-zero or compressed enough to be 3792 * embedded in the BP, no write was performed so there will be no 3793 * dva/birth/checksum. The buffer must therefore remain anonymous 3794 * (and uncached). 3795 */ 3796 if (!BUF_EMPTY(hdr)) { 3797 arc_buf_hdr_t *exists; 3798 kmutex_t *hash_lock; 3799 3800 ASSERT(zio->io_error == 0); 3801 3802 arc_cksum_verify(buf); 3803 3804 exists = buf_hash_insert(hdr, &hash_lock); 3805 if (exists) { 3806 /* 3807 * This can only happen if we overwrite for 3808 * sync-to-convergence, because we remove 3809 * buffers from the hash table when we arc_free(). 3810 */ 3811 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3812 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3813 panic("bad overwrite, hdr=%p exists=%p", 3814 (void *)hdr, (void *)exists); 3815 ASSERT(refcount_is_zero(&exists->b_refcnt)); 3816 arc_change_state(arc_anon, exists, hash_lock); 3817 mutex_exit(hash_lock); 3818 arc_hdr_destroy(exists); 3819 exists = buf_hash_insert(hdr, &hash_lock); 3820 ASSERT3P(exists, ==, NULL); 3821 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3822 /* nopwrite */ 3823 ASSERT(zio->io_prop.zp_nopwrite); 3824 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3825 panic("bad nopwrite, hdr=%p exists=%p", 3826 (void *)hdr, (void *)exists); 3827 } else { 3828 /* Dedup */ 3829 ASSERT(hdr->b_datacnt == 1); 3830 ASSERT(hdr->b_state == arc_anon); 3831 ASSERT(BP_GET_DEDUP(zio->io_bp)); 3832 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3833 } 3834 } 3835 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3836 /* if it's not anon, we are doing a scrub */ 3837 if (!exists && hdr->b_state == arc_anon) 3838 arc_access(hdr, hash_lock); 3839 mutex_exit(hash_lock); 3840 } else { 3841 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3842 } 3843 3844 ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3845 callback->awcb_done(zio, buf, callback->awcb_private); 3846 3847 kmem_free(callback, sizeof (arc_write_callback_t)); 3848} 3849 3850zio_t * 3851arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 3852 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 3853 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 3854 arc_done_func_t *done, void *private, zio_priority_t priority, 3855 int zio_flags, const zbookmark_phys_t *zb) 3856{ 3857 arc_buf_hdr_t *hdr = buf->b_hdr; 3858 arc_write_callback_t *callback; 3859 zio_t *zio; 3860 3861 ASSERT(ready != NULL); 3862 ASSERT(done != NULL); 3863 ASSERT(!HDR_IO_ERROR(hdr)); 3864 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3865 ASSERT(hdr->b_acb == NULL); 3866 if (l2arc) 3867 hdr->b_flags |= ARC_L2CACHE; 3868 if (l2arc_compress) 3869 hdr->b_flags |= ARC_L2COMPRESS; 3870 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 3871 callback->awcb_ready = ready; 3872 callback->awcb_physdone = physdone; 3873 callback->awcb_done = done; 3874 callback->awcb_private = private; 3875 callback->awcb_buf = buf; 3876 3877 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 3878 arc_write_ready, arc_write_physdone, arc_write_done, callback, 3879 priority, zio_flags, zb); 3880 3881 return (zio); 3882} 3883 3884static int 3885arc_memory_throttle(uint64_t reserve, uint64_t txg) 3886{ 3887#ifdef _KERNEL 3888 uint64_t available_memory = 3889 ptoa((uintmax_t)cnt.v_free_count + cnt.v_cache_count); 3890 static uint64_t page_load = 0; 3891 static uint64_t last_txg = 0; 3892 3893#ifdef sun 3894#if defined(__i386) 3895 available_memory = 3896 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 3897#endif 3898#endif /* sun */ 3899 3900 if (cnt.v_free_count + cnt.v_cache_count > 3901 (uint64_t)physmem * arc_lotsfree_percent / 100) 3902 return (0); 3903 3904 if (txg > last_txg) { 3905 last_txg = txg; 3906 page_load = 0; 3907 } 3908 /* 3909 * If we are in pageout, we know that memory is already tight, 3910 * the arc is already going to be evicting, so we just want to 3911 * continue to let page writes occur as quickly as possible. 3912 */ 3913 if (curproc == pageproc) { 3914 if (page_load > available_memory / 4) 3915 return (SET_ERROR(ERESTART)); 3916 /* Note: reserve is inflated, so we deflate */ 3917 page_load += reserve / 8; 3918 return (0); 3919 } else if (page_load > 0 && arc_reclaim_needed()) { 3920 /* memory is low, delay before restarting */ 3921 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3922 return (SET_ERROR(EAGAIN)); 3923 } 3924 page_load = 0; 3925#endif 3926 return (0); 3927} 3928 3929void 3930arc_tempreserve_clear(uint64_t reserve) 3931{ 3932 atomic_add_64(&arc_tempreserve, -reserve); 3933 ASSERT((int64_t)arc_tempreserve >= 0); 3934} 3935 3936int 3937arc_tempreserve_space(uint64_t reserve, uint64_t txg) 3938{ 3939 int error; 3940 uint64_t anon_size; 3941 3942 if (reserve > arc_c/4 && !arc_no_grow) 3943 arc_c = MIN(arc_c_max, reserve * 4); 3944 if (reserve > arc_c) 3945 return (SET_ERROR(ENOMEM)); 3946 3947 /* 3948 * Don't count loaned bufs as in flight dirty data to prevent long 3949 * network delays from blocking transactions that are ready to be 3950 * assigned to a txg. 3951 */ 3952 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 3953 3954 /* 3955 * Writes will, almost always, require additional memory allocations 3956 * in order to compress/encrypt/etc the data. We therefore need to 3957 * make sure that there is sufficient available memory for this. 3958 */ 3959 error = arc_memory_throttle(reserve, txg); 3960 if (error != 0) 3961 return (error); 3962 3963 /* 3964 * Throttle writes when the amount of dirty data in the cache 3965 * gets too large. We try to keep the cache less than half full 3966 * of dirty blocks so that our sync times don't grow too large. 3967 * Note: if two requests come in concurrently, we might let them 3968 * both succeed, when one of them should fail. Not a huge deal. 3969 */ 3970 3971 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 3972 anon_size > arc_c / 4) { 3973 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 3974 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 3975 arc_tempreserve>>10, 3976 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 3977 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 3978 reserve>>10, arc_c>>10); 3979 return (SET_ERROR(ERESTART)); 3980 } 3981 atomic_add_64(&arc_tempreserve, reserve); 3982 return (0); 3983} 3984 3985static kmutex_t arc_lowmem_lock; 3986#ifdef _KERNEL 3987static eventhandler_tag arc_event_lowmem = NULL; 3988 3989static void 3990arc_lowmem(void *arg __unused, int howto __unused) 3991{ 3992 3993 /* Serialize access via arc_lowmem_lock. */ 3994 mutex_enter(&arc_lowmem_lock); 3995 mutex_enter(&arc_reclaim_thr_lock); 3996 needfree = 1; 3997 cv_signal(&arc_reclaim_thr_cv); 3998 3999 /* 4000 * It is unsafe to block here in arbitrary threads, because we can come 4001 * here from ARC itself and may hold ARC locks and thus risk a deadlock 4002 * with ARC reclaim thread. 4003 */ 4004 if (curproc == pageproc) { 4005 while (needfree) 4006 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4007 } 4008 mutex_exit(&arc_reclaim_thr_lock); 4009 mutex_exit(&arc_lowmem_lock); 4010} 4011#endif 4012 4013void 4014arc_init(void) 4015{ 4016 int i, prefetch_tunable_set = 0; 4017 4018 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4019 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4020 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4021 4022 /* Convert seconds to clock ticks */ 4023 arc_min_prefetch_lifespan = 1 * hz; 4024 4025 /* Start out with 1/8 of all memory */ 4026 arc_c = kmem_size() / 8; 4027 4028#ifdef sun 4029#ifdef _KERNEL 4030 /* 4031 * On architectures where the physical memory can be larger 4032 * than the addressable space (intel in 32-bit mode), we may 4033 * need to limit the cache to 1/8 of VM size. 4034 */ 4035 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4036#endif 4037#endif /* sun */ 4038 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4039 arc_c_min = MAX(arc_c / 4, 64<<18); 4040 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4041 if (arc_c * 8 >= 1<<30) 4042 arc_c_max = (arc_c * 8) - (1<<30); 4043 else 4044 arc_c_max = arc_c_min; 4045 arc_c_max = MAX(arc_c * 5, arc_c_max); 4046 4047#ifdef _KERNEL 4048 /* 4049 * Allow the tunables to override our calculations if they are 4050 * reasonable (ie. over 16MB) 4051 */ 4052 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size()) 4053 arc_c_max = zfs_arc_max; 4054 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max) 4055 arc_c_min = zfs_arc_min; 4056#endif 4057 4058 arc_c = arc_c_max; 4059 arc_p = (arc_c >> 1); 4060 4061 /* limit meta-data to 1/4 of the arc capacity */ 4062 arc_meta_limit = arc_c_max / 4; 4063 4064 /* Allow the tunable to override if it is reasonable */ 4065 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4066 arc_meta_limit = zfs_arc_meta_limit; 4067 4068 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4069 arc_c_min = arc_meta_limit / 2; 4070 4071 if (zfs_arc_grow_retry > 0) 4072 arc_grow_retry = zfs_arc_grow_retry; 4073 4074 if (zfs_arc_shrink_shift > 0) 4075 arc_shrink_shift = zfs_arc_shrink_shift; 4076 4077 if (zfs_arc_p_min_shift > 0) 4078 arc_p_min_shift = zfs_arc_p_min_shift; 4079 4080 /* if kmem_flags are set, lets try to use less memory */ 4081 if (kmem_debugging()) 4082 arc_c = arc_c / 2; 4083 if (arc_c < arc_c_min) 4084 arc_c = arc_c_min; 4085 4086 zfs_arc_min = arc_c_min; 4087 zfs_arc_max = arc_c_max; 4088 4089 arc_anon = &ARC_anon; 4090 arc_mru = &ARC_mru; 4091 arc_mru_ghost = &ARC_mru_ghost; 4092 arc_mfu = &ARC_mfu; 4093 arc_mfu_ghost = &ARC_mfu_ghost; 4094 arc_l2c_only = &ARC_l2c_only; 4095 arc_size = 0; 4096 4097 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4098 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4099 NULL, MUTEX_DEFAULT, NULL); 4100 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4101 NULL, MUTEX_DEFAULT, NULL); 4102 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4103 NULL, MUTEX_DEFAULT, NULL); 4104 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4105 NULL, MUTEX_DEFAULT, NULL); 4106 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4107 NULL, MUTEX_DEFAULT, NULL); 4108 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4109 NULL, MUTEX_DEFAULT, NULL); 4110 4111 list_create(&arc_mru->arcs_lists[i], 4112 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4113 list_create(&arc_mru_ghost->arcs_lists[i], 4114 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4115 list_create(&arc_mfu->arcs_lists[i], 4116 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4117 list_create(&arc_mfu_ghost->arcs_lists[i], 4118 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4119 list_create(&arc_mfu_ghost->arcs_lists[i], 4120 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4121 list_create(&arc_l2c_only->arcs_lists[i], 4122 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4123 } 4124 4125 buf_init(); 4126 4127 arc_thread_exit = 0; 4128 arc_eviction_list = NULL; 4129 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4130 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4131 4132 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4133 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4134 4135 if (arc_ksp != NULL) { 4136 arc_ksp->ks_data = &arc_stats; 4137 kstat_install(arc_ksp); 4138 } 4139 4140 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4141 TS_RUN, minclsyspri); 4142 4143#ifdef _KERNEL 4144 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4145 EVENTHANDLER_PRI_FIRST); 4146#endif 4147 4148 arc_dead = FALSE; 4149 arc_warm = B_FALSE; 4150 4151 /* 4152 * Calculate maximum amount of dirty data per pool. 4153 * 4154 * If it has been set by /etc/system, take that. 4155 * Otherwise, use a percentage of physical memory defined by 4156 * zfs_dirty_data_max_percent (default 10%) with a cap at 4157 * zfs_dirty_data_max_max (default 4GB). 4158 */ 4159 if (zfs_dirty_data_max == 0) { 4160 zfs_dirty_data_max = ptob(physmem) * 4161 zfs_dirty_data_max_percent / 100; 4162 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4163 zfs_dirty_data_max_max); 4164 } 4165 4166#ifdef _KERNEL 4167 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4168 prefetch_tunable_set = 1; 4169 4170#ifdef __i386__ 4171 if (prefetch_tunable_set == 0) { 4172 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4173 "-- to enable,\n"); 4174 printf(" add \"vfs.zfs.prefetch_disable=0\" " 4175 "to /boot/loader.conf.\n"); 4176 zfs_prefetch_disable = 1; 4177 } 4178#else 4179 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4180 prefetch_tunable_set == 0) { 4181 printf("ZFS NOTICE: Prefetch is disabled by default if less " 4182 "than 4GB of RAM is present;\n" 4183 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4184 "to /boot/loader.conf.\n"); 4185 zfs_prefetch_disable = 1; 4186 } 4187#endif 4188 /* Warn about ZFS memory and address space requirements. */ 4189 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4190 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4191 "expect unstable behavior.\n"); 4192 } 4193 if (kmem_size() < 512 * (1 << 20)) { 4194 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4195 "expect unstable behavior.\n"); 4196 printf(" Consider tuning vm.kmem_size and " 4197 "vm.kmem_size_max\n"); 4198 printf(" in /boot/loader.conf.\n"); 4199 } 4200#endif 4201} 4202 4203void 4204arc_fini(void) 4205{ 4206 int i; 4207 4208 mutex_enter(&arc_reclaim_thr_lock); 4209 arc_thread_exit = 1; 4210 cv_signal(&arc_reclaim_thr_cv); 4211 while (arc_thread_exit != 0) 4212 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4213 mutex_exit(&arc_reclaim_thr_lock); 4214 4215 arc_flush(NULL); 4216 4217 arc_dead = TRUE; 4218 4219 if (arc_ksp != NULL) { 4220 kstat_delete(arc_ksp); 4221 arc_ksp = NULL; 4222 } 4223 4224 mutex_destroy(&arc_eviction_mtx); 4225 mutex_destroy(&arc_reclaim_thr_lock); 4226 cv_destroy(&arc_reclaim_thr_cv); 4227 4228 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4229 list_destroy(&arc_mru->arcs_lists[i]); 4230 list_destroy(&arc_mru_ghost->arcs_lists[i]); 4231 list_destroy(&arc_mfu->arcs_lists[i]); 4232 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4233 list_destroy(&arc_l2c_only->arcs_lists[i]); 4234 4235 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4236 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4237 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4238 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4239 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4240 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4241 } 4242 4243 buf_fini(); 4244 4245 ASSERT(arc_loaned_bytes == 0); 4246 4247 mutex_destroy(&arc_lowmem_lock); 4248#ifdef _KERNEL 4249 if (arc_event_lowmem != NULL) 4250 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4251#endif 4252} 4253 4254/* 4255 * Level 2 ARC 4256 * 4257 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4258 * It uses dedicated storage devices to hold cached data, which are populated 4259 * using large infrequent writes. The main role of this cache is to boost 4260 * the performance of random read workloads. The intended L2ARC devices 4261 * include short-stroked disks, solid state disks, and other media with 4262 * substantially faster read latency than disk. 4263 * 4264 * +-----------------------+ 4265 * | ARC | 4266 * +-----------------------+ 4267 * | ^ ^ 4268 * | | | 4269 * l2arc_feed_thread() arc_read() 4270 * | | | 4271 * | l2arc read | 4272 * V | | 4273 * +---------------+ | 4274 * | L2ARC | | 4275 * +---------------+ | 4276 * | ^ | 4277 * l2arc_write() | | 4278 * | | | 4279 * V | | 4280 * +-------+ +-------+ 4281 * | vdev | | vdev | 4282 * | cache | | cache | 4283 * +-------+ +-------+ 4284 * +=========+ .-----. 4285 * : L2ARC : |-_____-| 4286 * : devices : | Disks | 4287 * +=========+ `-_____-' 4288 * 4289 * Read requests are satisfied from the following sources, in order: 4290 * 4291 * 1) ARC 4292 * 2) vdev cache of L2ARC devices 4293 * 3) L2ARC devices 4294 * 4) vdev cache of disks 4295 * 5) disks 4296 * 4297 * Some L2ARC device types exhibit extremely slow write performance. 4298 * To accommodate for this there are some significant differences between 4299 * the L2ARC and traditional cache design: 4300 * 4301 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4302 * the ARC behave as usual, freeing buffers and placing headers on ghost 4303 * lists. The ARC does not send buffers to the L2ARC during eviction as 4304 * this would add inflated write latencies for all ARC memory pressure. 4305 * 4306 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4307 * It does this by periodically scanning buffers from the eviction-end of 4308 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4309 * not already there. It scans until a headroom of buffers is satisfied, 4310 * which itself is a buffer for ARC eviction. If a compressible buffer is 4311 * found during scanning and selected for writing to an L2ARC device, we 4312 * temporarily boost scanning headroom during the next scan cycle to make 4313 * sure we adapt to compression effects (which might significantly reduce 4314 * the data volume we write to L2ARC). The thread that does this is 4315 * l2arc_feed_thread(), illustrated below; example sizes are included to 4316 * provide a better sense of ratio than this diagram: 4317 * 4318 * head --> tail 4319 * +---------------------+----------+ 4320 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4321 * +---------------------+----------+ | o L2ARC eligible 4322 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4323 * +---------------------+----------+ | 4324 * 15.9 Gbytes ^ 32 Mbytes | 4325 * headroom | 4326 * l2arc_feed_thread() 4327 * | 4328 * l2arc write hand <--[oooo]--' 4329 * | 8 Mbyte 4330 * | write max 4331 * V 4332 * +==============================+ 4333 * L2ARC dev |####|#|###|###| |####| ... | 4334 * +==============================+ 4335 * 32 Gbytes 4336 * 4337 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4338 * evicted, then the L2ARC has cached a buffer much sooner than it probably 4339 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4340 * safe to say that this is an uncommon case, since buffers at the end of 4341 * the ARC lists have moved there due to inactivity. 4342 * 4343 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4344 * then the L2ARC simply misses copying some buffers. This serves as a 4345 * pressure valve to prevent heavy read workloads from both stalling the ARC 4346 * with waits and clogging the L2ARC with writes. This also helps prevent 4347 * the potential for the L2ARC to churn if it attempts to cache content too 4348 * quickly, such as during backups of the entire pool. 4349 * 4350 * 5. After system boot and before the ARC has filled main memory, there are 4351 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4352 * lists can remain mostly static. Instead of searching from tail of these 4353 * lists as pictured, the l2arc_feed_thread() will search from the list heads 4354 * for eligible buffers, greatly increasing its chance of finding them. 4355 * 4356 * The L2ARC device write speed is also boosted during this time so that 4357 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4358 * there are no L2ARC reads, and no fear of degrading read performance 4359 * through increased writes. 4360 * 4361 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4362 * the vdev queue can aggregate them into larger and fewer writes. Each 4363 * device is written to in a rotor fashion, sweeping writes through 4364 * available space then repeating. 4365 * 4366 * 7. The L2ARC does not store dirty content. It never needs to flush 4367 * write buffers back to disk based storage. 4368 * 4369 * 8. If an ARC buffer is written (and dirtied) which also exists in the 4370 * L2ARC, the now stale L2ARC buffer is immediately dropped. 4371 * 4372 * The performance of the L2ARC can be tweaked by a number of tunables, which 4373 * may be necessary for different workloads: 4374 * 4375 * l2arc_write_max max write bytes per interval 4376 * l2arc_write_boost extra write bytes during device warmup 4377 * l2arc_noprefetch skip caching prefetched buffers 4378 * l2arc_headroom number of max device writes to precache 4379 * l2arc_headroom_boost when we find compressed buffers during ARC 4380 * scanning, we multiply headroom by this 4381 * percentage factor for the next scan cycle, 4382 * since more compressed buffers are likely to 4383 * be present 4384 * l2arc_feed_secs seconds between L2ARC writing 4385 * 4386 * Tunables may be removed or added as future performance improvements are 4387 * integrated, and also may become zpool properties. 4388 * 4389 * There are three key functions that control how the L2ARC warms up: 4390 * 4391 * l2arc_write_eligible() check if a buffer is eligible to cache 4392 * l2arc_write_size() calculate how much to write 4393 * l2arc_write_interval() calculate sleep delay between writes 4394 * 4395 * These three functions determine what to write, how much, and how quickly 4396 * to send writes. 4397 */ 4398 4399static boolean_t 4400l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab) 4401{ 4402 /* 4403 * A buffer is *not* eligible for the L2ARC if it: 4404 * 1. belongs to a different spa. 4405 * 2. is already cached on the L2ARC. 4406 * 3. has an I/O in progress (it may be an incomplete read). 4407 * 4. is flagged not eligible (zfs property). 4408 */ 4409 if (ab->b_spa != spa_guid) { 4410 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4411 return (B_FALSE); 4412 } 4413 if (ab->b_l2hdr != NULL) { 4414 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4415 return (B_FALSE); 4416 } 4417 if (HDR_IO_IN_PROGRESS(ab)) { 4418 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4419 return (B_FALSE); 4420 } 4421 if (!HDR_L2CACHE(ab)) { 4422 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4423 return (B_FALSE); 4424 } 4425 4426 return (B_TRUE); 4427} 4428 4429static uint64_t 4430l2arc_write_size(void) 4431{ 4432 uint64_t size; 4433 4434 /* 4435 * Make sure our globals have meaningful values in case the user 4436 * altered them. 4437 */ 4438 size = l2arc_write_max; 4439 if (size == 0) { 4440 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4441 "be greater than zero, resetting it to the default (%d)", 4442 L2ARC_WRITE_SIZE); 4443 size = l2arc_write_max = L2ARC_WRITE_SIZE; 4444 } 4445 4446 if (arc_warm == B_FALSE) 4447 size += l2arc_write_boost; 4448 4449 return (size); 4450 4451} 4452 4453static clock_t 4454l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4455{ 4456 clock_t interval, next, now; 4457 4458 /* 4459 * If the ARC lists are busy, increase our write rate; if the 4460 * lists are stale, idle back. This is achieved by checking 4461 * how much we previously wrote - if it was more than half of 4462 * what we wanted, schedule the next write much sooner. 4463 */ 4464 if (l2arc_feed_again && wrote > (wanted / 2)) 4465 interval = (hz * l2arc_feed_min_ms) / 1000; 4466 else 4467 interval = hz * l2arc_feed_secs; 4468 4469 now = ddi_get_lbolt(); 4470 next = MAX(now, MIN(now + interval, began + interval)); 4471 4472 return (next); 4473} 4474 4475static void 4476l2arc_hdr_stat_add(void) 4477{ 4478 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4479 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4480} 4481 4482static void 4483l2arc_hdr_stat_remove(void) 4484{ 4485 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4486 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4487} 4488 4489/* 4490 * Cycle through L2ARC devices. This is how L2ARC load balances. 4491 * If a device is returned, this also returns holding the spa config lock. 4492 */ 4493static l2arc_dev_t * 4494l2arc_dev_get_next(void) 4495{ 4496 l2arc_dev_t *first, *next = NULL; 4497 4498 /* 4499 * Lock out the removal of spas (spa_namespace_lock), then removal 4500 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4501 * both locks will be dropped and a spa config lock held instead. 4502 */ 4503 mutex_enter(&spa_namespace_lock); 4504 mutex_enter(&l2arc_dev_mtx); 4505 4506 /* if there are no vdevs, there is nothing to do */ 4507 if (l2arc_ndev == 0) 4508 goto out; 4509 4510 first = NULL; 4511 next = l2arc_dev_last; 4512 do { 4513 /* loop around the list looking for a non-faulted vdev */ 4514 if (next == NULL) { 4515 next = list_head(l2arc_dev_list); 4516 } else { 4517 next = list_next(l2arc_dev_list, next); 4518 if (next == NULL) 4519 next = list_head(l2arc_dev_list); 4520 } 4521 4522 /* if we have come back to the start, bail out */ 4523 if (first == NULL) 4524 first = next; 4525 else if (next == first) 4526 break; 4527 4528 } while (vdev_is_dead(next->l2ad_vdev)); 4529 4530 /* if we were unable to find any usable vdevs, return NULL */ 4531 if (vdev_is_dead(next->l2ad_vdev)) 4532 next = NULL; 4533 4534 l2arc_dev_last = next; 4535 4536out: 4537 mutex_exit(&l2arc_dev_mtx); 4538 4539 /* 4540 * Grab the config lock to prevent the 'next' device from being 4541 * removed while we are writing to it. 4542 */ 4543 if (next != NULL) 4544 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4545 mutex_exit(&spa_namespace_lock); 4546 4547 return (next); 4548} 4549 4550/* 4551 * Free buffers that were tagged for destruction. 4552 */ 4553static void 4554l2arc_do_free_on_write() 4555{ 4556 list_t *buflist; 4557 l2arc_data_free_t *df, *df_prev; 4558 4559 mutex_enter(&l2arc_free_on_write_mtx); 4560 buflist = l2arc_free_on_write; 4561 4562 for (df = list_tail(buflist); df; df = df_prev) { 4563 df_prev = list_prev(buflist, df); 4564 ASSERT(df->l2df_data != NULL); 4565 ASSERT(df->l2df_func != NULL); 4566 df->l2df_func(df->l2df_data, df->l2df_size); 4567 list_remove(buflist, df); 4568 kmem_free(df, sizeof (l2arc_data_free_t)); 4569 } 4570 4571 mutex_exit(&l2arc_free_on_write_mtx); 4572} 4573 4574/* 4575 * A write to a cache device has completed. Update all headers to allow 4576 * reads from these buffers to begin. 4577 */ 4578static void 4579l2arc_write_done(zio_t *zio) 4580{ 4581 l2arc_write_callback_t *cb; 4582 l2arc_dev_t *dev; 4583 list_t *buflist; 4584 arc_buf_hdr_t *head, *ab, *ab_prev; 4585 l2arc_buf_hdr_t *abl2; 4586 kmutex_t *hash_lock; 4587 int64_t bytes_dropped = 0; 4588 4589 cb = zio->io_private; 4590 ASSERT(cb != NULL); 4591 dev = cb->l2wcb_dev; 4592 ASSERT(dev != NULL); 4593 head = cb->l2wcb_head; 4594 ASSERT(head != NULL); 4595 buflist = dev->l2ad_buflist; 4596 ASSERT(buflist != NULL); 4597 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4598 l2arc_write_callback_t *, cb); 4599 4600 if (zio->io_error != 0) 4601 ARCSTAT_BUMP(arcstat_l2_writes_error); 4602 4603 mutex_enter(&l2arc_buflist_mtx); 4604 4605 /* 4606 * All writes completed, or an error was hit. 4607 */ 4608 for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 4609 ab_prev = list_prev(buflist, ab); 4610 abl2 = ab->b_l2hdr; 4611 4612 /* 4613 * Release the temporary compressed buffer as soon as possible. 4614 */ 4615 if (abl2->b_compress != ZIO_COMPRESS_OFF) 4616 l2arc_release_cdata_buf(ab); 4617 4618 hash_lock = HDR_LOCK(ab); 4619 if (!mutex_tryenter(hash_lock)) { 4620 /* 4621 * This buffer misses out. It may be in a stage 4622 * of eviction. Its ARC_L2_WRITING flag will be 4623 * left set, denying reads to this buffer. 4624 */ 4625 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4626 continue; 4627 } 4628 4629 if (zio->io_error != 0) { 4630 /* 4631 * Error - drop L2ARC entry. 4632 */ 4633 list_remove(buflist, ab); 4634 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4635 bytes_dropped += abl2->b_asize; 4636 ab->b_l2hdr = NULL; 4637 trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4638 ab->b_size, 0); 4639 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4640 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4641 } 4642 4643 /* 4644 * Allow ARC to begin reads to this L2ARC entry. 4645 */ 4646 ab->b_flags &= ~ARC_L2_WRITING; 4647 4648 mutex_exit(hash_lock); 4649 } 4650 4651 atomic_inc_64(&l2arc_writes_done); 4652 list_remove(buflist, head); 4653 kmem_cache_free(hdr_cache, head); 4654 mutex_exit(&l2arc_buflist_mtx); 4655 4656 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 4657 4658 l2arc_do_free_on_write(); 4659 4660 kmem_free(cb, sizeof (l2arc_write_callback_t)); 4661} 4662 4663/* 4664 * A read to a cache device completed. Validate buffer contents before 4665 * handing over to the regular ARC routines. 4666 */ 4667static void 4668l2arc_read_done(zio_t *zio) 4669{ 4670 l2arc_read_callback_t *cb; 4671 arc_buf_hdr_t *hdr; 4672 arc_buf_t *buf; 4673 kmutex_t *hash_lock; 4674 int equal; 4675 4676 ASSERT(zio->io_vd != NULL); 4677 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4678 4679 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4680 4681 cb = zio->io_private; 4682 ASSERT(cb != NULL); 4683 buf = cb->l2rcb_buf; 4684 ASSERT(buf != NULL); 4685 4686 hash_lock = HDR_LOCK(buf->b_hdr); 4687 mutex_enter(hash_lock); 4688 hdr = buf->b_hdr; 4689 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4690 4691 /* 4692 * If the buffer was compressed, decompress it first. 4693 */ 4694 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4695 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4696 ASSERT(zio->io_data != NULL); 4697 4698 /* 4699 * Check this survived the L2ARC journey. 4700 */ 4701 equal = arc_cksum_equal(buf); 4702 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4703 mutex_exit(hash_lock); 4704 zio->io_private = buf; 4705 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4706 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4707 arc_read_done(zio); 4708 } else { 4709 mutex_exit(hash_lock); 4710 /* 4711 * Buffer didn't survive caching. Increment stats and 4712 * reissue to the original storage device. 4713 */ 4714 if (zio->io_error != 0) { 4715 ARCSTAT_BUMP(arcstat_l2_io_error); 4716 } else { 4717 zio->io_error = SET_ERROR(EIO); 4718 } 4719 if (!equal) 4720 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4721 4722 /* 4723 * If there's no waiter, issue an async i/o to the primary 4724 * storage now. If there *is* a waiter, the caller must 4725 * issue the i/o in a context where it's OK to block. 4726 */ 4727 if (zio->io_waiter == NULL) { 4728 zio_t *pio = zio_unique_parent(zio); 4729 4730 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4731 4732 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4733 buf->b_data, zio->io_size, arc_read_done, buf, 4734 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4735 } 4736 } 4737 4738 kmem_free(cb, sizeof (l2arc_read_callback_t)); 4739} 4740 4741/* 4742 * This is the list priority from which the L2ARC will search for pages to 4743 * cache. This is used within loops (0..3) to cycle through lists in the 4744 * desired order. This order can have a significant effect on cache 4745 * performance. 4746 * 4747 * Currently the metadata lists are hit first, MFU then MRU, followed by 4748 * the data lists. This function returns a locked list, and also returns 4749 * the lock pointer. 4750 */ 4751static list_t * 4752l2arc_list_locked(int list_num, kmutex_t **lock) 4753{ 4754 list_t *list = NULL; 4755 int idx; 4756 4757 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4758 4759 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4760 idx = list_num; 4761 list = &arc_mfu->arcs_lists[idx]; 4762 *lock = ARCS_LOCK(arc_mfu, idx); 4763 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4764 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4765 list = &arc_mru->arcs_lists[idx]; 4766 *lock = ARCS_LOCK(arc_mru, idx); 4767 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4768 ARC_BUFC_NUMDATALISTS)) { 4769 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4770 list = &arc_mfu->arcs_lists[idx]; 4771 *lock = ARCS_LOCK(arc_mfu, idx); 4772 } else { 4773 idx = list_num - ARC_BUFC_NUMLISTS; 4774 list = &arc_mru->arcs_lists[idx]; 4775 *lock = ARCS_LOCK(arc_mru, idx); 4776 } 4777 4778 ASSERT(!(MUTEX_HELD(*lock))); 4779 mutex_enter(*lock); 4780 return (list); 4781} 4782 4783/* 4784 * Evict buffers from the device write hand to the distance specified in 4785 * bytes. This distance may span populated buffers, it may span nothing. 4786 * This is clearing a region on the L2ARC device ready for writing. 4787 * If the 'all' boolean is set, every buffer is evicted. 4788 */ 4789static void 4790l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4791{ 4792 list_t *buflist; 4793 l2arc_buf_hdr_t *abl2; 4794 arc_buf_hdr_t *ab, *ab_prev; 4795 kmutex_t *hash_lock; 4796 uint64_t taddr; 4797 int64_t bytes_evicted = 0; 4798 4799 buflist = dev->l2ad_buflist; 4800 4801 if (buflist == NULL) 4802 return; 4803 4804 if (!all && dev->l2ad_first) { 4805 /* 4806 * This is the first sweep through the device. There is 4807 * nothing to evict. 4808 */ 4809 return; 4810 } 4811 4812 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4813 /* 4814 * When nearing the end of the device, evict to the end 4815 * before the device write hand jumps to the start. 4816 */ 4817 taddr = dev->l2ad_end; 4818 } else { 4819 taddr = dev->l2ad_hand + distance; 4820 } 4821 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4822 uint64_t, taddr, boolean_t, all); 4823 4824top: 4825 mutex_enter(&l2arc_buflist_mtx); 4826 for (ab = list_tail(buflist); ab; ab = ab_prev) { 4827 ab_prev = list_prev(buflist, ab); 4828 4829 hash_lock = HDR_LOCK(ab); 4830 if (!mutex_tryenter(hash_lock)) { 4831 /* 4832 * Missed the hash lock. Retry. 4833 */ 4834 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4835 mutex_exit(&l2arc_buflist_mtx); 4836 mutex_enter(hash_lock); 4837 mutex_exit(hash_lock); 4838 goto top; 4839 } 4840 4841 if (HDR_L2_WRITE_HEAD(ab)) { 4842 /* 4843 * We hit a write head node. Leave it for 4844 * l2arc_write_done(). 4845 */ 4846 list_remove(buflist, ab); 4847 mutex_exit(hash_lock); 4848 continue; 4849 } 4850 4851 if (!all && ab->b_l2hdr != NULL && 4852 (ab->b_l2hdr->b_daddr > taddr || 4853 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4854 /* 4855 * We've evicted to the target address, 4856 * or the end of the device. 4857 */ 4858 mutex_exit(hash_lock); 4859 break; 4860 } 4861 4862 if (HDR_FREE_IN_PROGRESS(ab)) { 4863 /* 4864 * Already on the path to destruction. 4865 */ 4866 mutex_exit(hash_lock); 4867 continue; 4868 } 4869 4870 if (ab->b_state == arc_l2c_only) { 4871 ASSERT(!HDR_L2_READING(ab)); 4872 /* 4873 * This doesn't exist in the ARC. Destroy. 4874 * arc_hdr_destroy() will call list_remove() 4875 * and decrement arcstat_l2_size. 4876 */ 4877 arc_change_state(arc_anon, ab, hash_lock); 4878 arc_hdr_destroy(ab); 4879 } else { 4880 /* 4881 * Invalidate issued or about to be issued 4882 * reads, since we may be about to write 4883 * over this location. 4884 */ 4885 if (HDR_L2_READING(ab)) { 4886 ARCSTAT_BUMP(arcstat_l2_evict_reading); 4887 ab->b_flags |= ARC_L2_EVICTED; 4888 } 4889 4890 /* 4891 * Tell ARC this no longer exists in L2ARC. 4892 */ 4893 if (ab->b_l2hdr != NULL) { 4894 abl2 = ab->b_l2hdr; 4895 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4896 bytes_evicted += abl2->b_asize; 4897 ab->b_l2hdr = NULL; 4898 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4899 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4900 } 4901 list_remove(buflist, ab); 4902 4903 /* 4904 * This may have been leftover after a 4905 * failed write. 4906 */ 4907 ab->b_flags &= ~ARC_L2_WRITING; 4908 } 4909 mutex_exit(hash_lock); 4910 } 4911 mutex_exit(&l2arc_buflist_mtx); 4912 4913 vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); 4914 dev->l2ad_evict = taddr; 4915} 4916 4917/* 4918 * Find and write ARC buffers to the L2ARC device. 4919 * 4920 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 4921 * for reading until they have completed writing. 4922 * The headroom_boost is an in-out parameter used to maintain headroom boost 4923 * state between calls to this function. 4924 * 4925 * Returns the number of bytes actually written (which may be smaller than 4926 * the delta by which the device hand has changed due to alignment). 4927 */ 4928static uint64_t 4929l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 4930 boolean_t *headroom_boost) 4931{ 4932 arc_buf_hdr_t *ab, *ab_prev, *head; 4933 list_t *list; 4934 uint64_t write_asize, write_psize, write_sz, headroom, 4935 buf_compress_minsz; 4936 void *buf_data; 4937 kmutex_t *list_lock; 4938 boolean_t full; 4939 l2arc_write_callback_t *cb; 4940 zio_t *pio, *wzio; 4941 uint64_t guid = spa_load_guid(spa); 4942 const boolean_t do_headroom_boost = *headroom_boost; 4943 int try; 4944 4945 ASSERT(dev->l2ad_vdev != NULL); 4946 4947 /* Lower the flag now, we might want to raise it again later. */ 4948 *headroom_boost = B_FALSE; 4949 4950 pio = NULL; 4951 write_sz = write_asize = write_psize = 0; 4952 full = B_FALSE; 4953 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 4954 head->b_flags |= ARC_L2_WRITE_HEAD; 4955 4956 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 4957 /* 4958 * We will want to try to compress buffers that are at least 2x the 4959 * device sector size. 4960 */ 4961 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 4962 4963 /* 4964 * Copy buffers for L2ARC writing. 4965 */ 4966 mutex_enter(&l2arc_buflist_mtx); 4967 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 4968 uint64_t passed_sz = 0; 4969 4970 list = l2arc_list_locked(try, &list_lock); 4971 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 4972 4973 /* 4974 * L2ARC fast warmup. 4975 * 4976 * Until the ARC is warm and starts to evict, read from the 4977 * head of the ARC lists rather than the tail. 4978 */ 4979 if (arc_warm == B_FALSE) 4980 ab = list_head(list); 4981 else 4982 ab = list_tail(list); 4983 if (ab == NULL) 4984 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 4985 4986 headroom = target_sz * l2arc_headroom; 4987 if (do_headroom_boost) 4988 headroom = (headroom * l2arc_headroom_boost) / 100; 4989 4990 for (; ab; ab = ab_prev) { 4991 l2arc_buf_hdr_t *l2hdr; 4992 kmutex_t *hash_lock; 4993 uint64_t buf_sz; 4994 4995 if (arc_warm == B_FALSE) 4996 ab_prev = list_next(list, ab); 4997 else 4998 ab_prev = list_prev(list, ab); 4999 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size); 5000 5001 hash_lock = HDR_LOCK(ab); 5002 if (!mutex_tryenter(hash_lock)) { 5003 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5004 /* 5005 * Skip this buffer rather than waiting. 5006 */ 5007 continue; 5008 } 5009 5010 passed_sz += ab->b_size; 5011 if (passed_sz > headroom) { 5012 /* 5013 * Searched too far. 5014 */ 5015 mutex_exit(hash_lock); 5016 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5017 break; 5018 } 5019 5020 if (!l2arc_write_eligible(guid, ab)) { 5021 mutex_exit(hash_lock); 5022 continue; 5023 } 5024 5025 if ((write_sz + ab->b_size) > target_sz) { 5026 full = B_TRUE; 5027 mutex_exit(hash_lock); 5028 ARCSTAT_BUMP(arcstat_l2_write_full); 5029 break; 5030 } 5031 5032 if (pio == NULL) { 5033 /* 5034 * Insert a dummy header on the buflist so 5035 * l2arc_write_done() can find where the 5036 * write buffers begin without searching. 5037 */ 5038 list_insert_head(dev->l2ad_buflist, head); 5039 5040 cb = kmem_alloc( 5041 sizeof (l2arc_write_callback_t), KM_SLEEP); 5042 cb->l2wcb_dev = dev; 5043 cb->l2wcb_head = head; 5044 pio = zio_root(spa, l2arc_write_done, cb, 5045 ZIO_FLAG_CANFAIL); 5046 ARCSTAT_BUMP(arcstat_l2_write_pios); 5047 } 5048 5049 /* 5050 * Create and add a new L2ARC header. 5051 */ 5052 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5053 l2hdr->b_dev = dev; 5054 ab->b_flags |= ARC_L2_WRITING; 5055 5056 /* 5057 * Temporarily stash the data buffer in b_tmp_cdata. 5058 * The subsequent write step will pick it up from 5059 * there. This is because can't access ab->b_buf 5060 * without holding the hash_lock, which we in turn 5061 * can't access without holding the ARC list locks 5062 * (which we want to avoid during compression/writing). 5063 */ 5064 l2hdr->b_compress = ZIO_COMPRESS_OFF; 5065 l2hdr->b_asize = ab->b_size; 5066 l2hdr->b_tmp_cdata = ab->b_buf->b_data; 5067 5068 buf_sz = ab->b_size; 5069 ab->b_l2hdr = l2hdr; 5070 5071 list_insert_head(dev->l2ad_buflist, ab); 5072 5073 /* 5074 * Compute and store the buffer cksum before 5075 * writing. On debug the cksum is verified first. 5076 */ 5077 arc_cksum_verify(ab->b_buf); 5078 arc_cksum_compute(ab->b_buf, B_TRUE); 5079 5080 mutex_exit(hash_lock); 5081 5082 write_sz += buf_sz; 5083 } 5084 5085 mutex_exit(list_lock); 5086 5087 if (full == B_TRUE) 5088 break; 5089 } 5090 5091 /* No buffers selected for writing? */ 5092 if (pio == NULL) { 5093 ASSERT0(write_sz); 5094 mutex_exit(&l2arc_buflist_mtx); 5095 kmem_cache_free(hdr_cache, head); 5096 return (0); 5097 } 5098 5099 /* 5100 * Now start writing the buffers. We're starting at the write head 5101 * and work backwards, retracing the course of the buffer selector 5102 * loop above. 5103 */ 5104 for (ab = list_prev(dev->l2ad_buflist, head); ab; 5105 ab = list_prev(dev->l2ad_buflist, ab)) { 5106 l2arc_buf_hdr_t *l2hdr; 5107 uint64_t buf_sz; 5108 5109 /* 5110 * We shouldn't need to lock the buffer here, since we flagged 5111 * it as ARC_L2_WRITING in the previous step, but we must take 5112 * care to only access its L2 cache parameters. In particular, 5113 * ab->b_buf may be invalid by now due to ARC eviction. 5114 */ 5115 l2hdr = ab->b_l2hdr; 5116 l2hdr->b_daddr = dev->l2ad_hand; 5117 5118 if ((ab->b_flags & ARC_L2COMPRESS) && 5119 l2hdr->b_asize >= buf_compress_minsz) { 5120 if (l2arc_compress_buf(l2hdr)) { 5121 /* 5122 * If compression succeeded, enable headroom 5123 * boost on the next scan cycle. 5124 */ 5125 *headroom_boost = B_TRUE; 5126 } 5127 } 5128 5129 /* 5130 * Pick up the buffer data we had previously stashed away 5131 * (and now potentially also compressed). 5132 */ 5133 buf_data = l2hdr->b_tmp_cdata; 5134 buf_sz = l2hdr->b_asize; 5135 5136 /* Compression may have squashed the buffer to zero length. */ 5137 if (buf_sz != 0) { 5138 uint64_t buf_p_sz; 5139 5140 wzio = zio_write_phys(pio, dev->l2ad_vdev, 5141 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5142 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5143 ZIO_FLAG_CANFAIL, B_FALSE); 5144 5145 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5146 zio_t *, wzio); 5147 (void) zio_nowait(wzio); 5148 5149 write_asize += buf_sz; 5150 /* 5151 * Keep the clock hand suitably device-aligned. 5152 */ 5153 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5154 write_psize += buf_p_sz; 5155 dev->l2ad_hand += buf_p_sz; 5156 } 5157 } 5158 5159 mutex_exit(&l2arc_buflist_mtx); 5160 5161 ASSERT3U(write_asize, <=, target_sz); 5162 ARCSTAT_BUMP(arcstat_l2_writes_sent); 5163 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5164 ARCSTAT_INCR(arcstat_l2_size, write_sz); 5165 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 5166 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 5167 5168 /* 5169 * Bump device hand to the device start if it is approaching the end. 5170 * l2arc_evict() will already have evicted ahead for this case. 5171 */ 5172 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5173 dev->l2ad_hand = dev->l2ad_start; 5174 dev->l2ad_evict = dev->l2ad_start; 5175 dev->l2ad_first = B_FALSE; 5176 } 5177 5178 dev->l2ad_writing = B_TRUE; 5179 (void) zio_wait(pio); 5180 dev->l2ad_writing = B_FALSE; 5181 5182 return (write_asize); 5183} 5184 5185/* 5186 * Compresses an L2ARC buffer. 5187 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5188 * size in l2hdr->b_asize. This routine tries to compress the data and 5189 * depending on the compression result there are three possible outcomes: 5190 * *) The buffer was incompressible. The original l2hdr contents were left 5191 * untouched and are ready for writing to an L2 device. 5192 * *) The buffer was all-zeros, so there is no need to write it to an L2 5193 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5194 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5195 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5196 * data buffer which holds the compressed data to be written, and b_asize 5197 * tells us how much data there is. b_compress is set to the appropriate 5198 * compression algorithm. Once writing is done, invoke 5199 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5200 * 5201 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5202 * buffer was incompressible). 5203 */ 5204static boolean_t 5205l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5206{ 5207 void *cdata; 5208 size_t csize, len, rounded; 5209 5210 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5211 ASSERT(l2hdr->b_tmp_cdata != NULL); 5212 5213 len = l2hdr->b_asize; 5214 cdata = zio_data_buf_alloc(len); 5215 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5216 cdata, l2hdr->b_asize); 5217 5218 rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE); 5219 if (rounded > csize) { 5220 bzero((char *)cdata + csize, rounded - csize); 5221 csize = rounded; 5222 } 5223 5224 if (csize == 0) { 5225 /* zero block, indicate that there's nothing to write */ 5226 zio_data_buf_free(cdata, len); 5227 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5228 l2hdr->b_asize = 0; 5229 l2hdr->b_tmp_cdata = NULL; 5230 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5231 return (B_TRUE); 5232 } else if (csize > 0 && csize < len) { 5233 /* 5234 * Compression succeeded, we'll keep the cdata around for 5235 * writing and release it afterwards. 5236 */ 5237 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5238 l2hdr->b_asize = csize; 5239 l2hdr->b_tmp_cdata = cdata; 5240 ARCSTAT_BUMP(arcstat_l2_compress_successes); 5241 return (B_TRUE); 5242 } else { 5243 /* 5244 * Compression failed, release the compressed buffer. 5245 * l2hdr will be left unmodified. 5246 */ 5247 zio_data_buf_free(cdata, len); 5248 ARCSTAT_BUMP(arcstat_l2_compress_failures); 5249 return (B_FALSE); 5250 } 5251} 5252 5253/* 5254 * Decompresses a zio read back from an l2arc device. On success, the 5255 * underlying zio's io_data buffer is overwritten by the uncompressed 5256 * version. On decompression error (corrupt compressed stream), the 5257 * zio->io_error value is set to signal an I/O error. 5258 * 5259 * Please note that the compressed data stream is not checksummed, so 5260 * if the underlying device is experiencing data corruption, we may feed 5261 * corrupt data to the decompressor, so the decompressor needs to be 5262 * able to handle this situation (LZ4 does). 5263 */ 5264static void 5265l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5266{ 5267 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5268 5269 if (zio->io_error != 0) { 5270 /* 5271 * An io error has occured, just restore the original io 5272 * size in preparation for a main pool read. 5273 */ 5274 zio->io_orig_size = zio->io_size = hdr->b_size; 5275 return; 5276 } 5277 5278 if (c == ZIO_COMPRESS_EMPTY) { 5279 /* 5280 * An empty buffer results in a null zio, which means we 5281 * need to fill its io_data after we're done restoring the 5282 * buffer's contents. 5283 */ 5284 ASSERT(hdr->b_buf != NULL); 5285 bzero(hdr->b_buf->b_data, hdr->b_size); 5286 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5287 } else { 5288 ASSERT(zio->io_data != NULL); 5289 /* 5290 * We copy the compressed data from the start of the arc buffer 5291 * (the zio_read will have pulled in only what we need, the 5292 * rest is garbage which we will overwrite at decompression) 5293 * and then decompress back to the ARC data buffer. This way we 5294 * can minimize copying by simply decompressing back over the 5295 * original compressed data (rather than decompressing to an 5296 * aux buffer and then copying back the uncompressed buffer, 5297 * which is likely to be much larger). 5298 */ 5299 uint64_t csize; 5300 void *cdata; 5301 5302 csize = zio->io_size; 5303 cdata = zio_data_buf_alloc(csize); 5304 bcopy(zio->io_data, cdata, csize); 5305 if (zio_decompress_data(c, cdata, zio->io_data, csize, 5306 hdr->b_size) != 0) 5307 zio->io_error = EIO; 5308 zio_data_buf_free(cdata, csize); 5309 } 5310 5311 /* Restore the expected uncompressed IO size. */ 5312 zio->io_orig_size = zio->io_size = hdr->b_size; 5313} 5314 5315/* 5316 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5317 * This buffer serves as a temporary holder of compressed data while 5318 * the buffer entry is being written to an l2arc device. Once that is 5319 * done, we can dispose of it. 5320 */ 5321static void 5322l2arc_release_cdata_buf(arc_buf_hdr_t *ab) 5323{ 5324 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr; 5325 5326 if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) { 5327 /* 5328 * If the data was compressed, then we've allocated a 5329 * temporary buffer for it, so now we need to release it. 5330 */ 5331 ASSERT(l2hdr->b_tmp_cdata != NULL); 5332 zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size); 5333 } 5334 l2hdr->b_tmp_cdata = NULL; 5335} 5336 5337/* 5338 * This thread feeds the L2ARC at regular intervals. This is the beating 5339 * heart of the L2ARC. 5340 */ 5341static void 5342l2arc_feed_thread(void *dummy __unused) 5343{ 5344 callb_cpr_t cpr; 5345 l2arc_dev_t *dev; 5346 spa_t *spa; 5347 uint64_t size, wrote; 5348 clock_t begin, next = ddi_get_lbolt(); 5349 boolean_t headroom_boost = B_FALSE; 5350 5351 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5352 5353 mutex_enter(&l2arc_feed_thr_lock); 5354 5355 while (l2arc_thread_exit == 0) { 5356 CALLB_CPR_SAFE_BEGIN(&cpr); 5357 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5358 next - ddi_get_lbolt()); 5359 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5360 next = ddi_get_lbolt() + hz; 5361 5362 /* 5363 * Quick check for L2ARC devices. 5364 */ 5365 mutex_enter(&l2arc_dev_mtx); 5366 if (l2arc_ndev == 0) { 5367 mutex_exit(&l2arc_dev_mtx); 5368 continue; 5369 } 5370 mutex_exit(&l2arc_dev_mtx); 5371 begin = ddi_get_lbolt(); 5372 5373 /* 5374 * This selects the next l2arc device to write to, and in 5375 * doing so the next spa to feed from: dev->l2ad_spa. This 5376 * will return NULL if there are now no l2arc devices or if 5377 * they are all faulted. 5378 * 5379 * If a device is returned, its spa's config lock is also 5380 * held to prevent device removal. l2arc_dev_get_next() 5381 * will grab and release l2arc_dev_mtx. 5382 */ 5383 if ((dev = l2arc_dev_get_next()) == NULL) 5384 continue; 5385 5386 spa = dev->l2ad_spa; 5387 ASSERT(spa != NULL); 5388 5389 /* 5390 * If the pool is read-only then force the feed thread to 5391 * sleep a little longer. 5392 */ 5393 if (!spa_writeable(spa)) { 5394 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5395 spa_config_exit(spa, SCL_L2ARC, dev); 5396 continue; 5397 } 5398 5399 /* 5400 * Avoid contributing to memory pressure. 5401 */ 5402 if (arc_reclaim_needed()) { 5403 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5404 spa_config_exit(spa, SCL_L2ARC, dev); 5405 continue; 5406 } 5407 5408 ARCSTAT_BUMP(arcstat_l2_feeds); 5409 5410 size = l2arc_write_size(); 5411 5412 /* 5413 * Evict L2ARC buffers that will be overwritten. 5414 */ 5415 l2arc_evict(dev, size, B_FALSE); 5416 5417 /* 5418 * Write ARC buffers. 5419 */ 5420 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5421 5422 /* 5423 * Calculate interval between writes. 5424 */ 5425 next = l2arc_write_interval(begin, size, wrote); 5426 spa_config_exit(spa, SCL_L2ARC, dev); 5427 } 5428 5429 l2arc_thread_exit = 0; 5430 cv_broadcast(&l2arc_feed_thr_cv); 5431 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5432 thread_exit(); 5433} 5434 5435boolean_t 5436l2arc_vdev_present(vdev_t *vd) 5437{ 5438 l2arc_dev_t *dev; 5439 5440 mutex_enter(&l2arc_dev_mtx); 5441 for (dev = list_head(l2arc_dev_list); dev != NULL; 5442 dev = list_next(l2arc_dev_list, dev)) { 5443 if (dev->l2ad_vdev == vd) 5444 break; 5445 } 5446 mutex_exit(&l2arc_dev_mtx); 5447 5448 return (dev != NULL); 5449} 5450 5451/* 5452 * Add a vdev for use by the L2ARC. By this point the spa has already 5453 * validated the vdev and opened it. 5454 */ 5455void 5456l2arc_add_vdev(spa_t *spa, vdev_t *vd) 5457{ 5458 l2arc_dev_t *adddev; 5459 5460 ASSERT(!l2arc_vdev_present(vd)); 5461 5462 vdev_ashift_optimize(vd); 5463 5464 /* 5465 * Create a new l2arc device entry. 5466 */ 5467 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5468 adddev->l2ad_spa = spa; 5469 adddev->l2ad_vdev = vd; 5470 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5471 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5472 adddev->l2ad_hand = adddev->l2ad_start; 5473 adddev->l2ad_evict = adddev->l2ad_start; 5474 adddev->l2ad_first = B_TRUE; 5475 adddev->l2ad_writing = B_FALSE; 5476 5477 /* 5478 * This is a list of all ARC buffers that are still valid on the 5479 * device. 5480 */ 5481 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5482 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5483 offsetof(arc_buf_hdr_t, b_l2node)); 5484 5485 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5486 5487 /* 5488 * Add device to global list 5489 */ 5490 mutex_enter(&l2arc_dev_mtx); 5491 list_insert_head(l2arc_dev_list, adddev); 5492 atomic_inc_64(&l2arc_ndev); 5493 mutex_exit(&l2arc_dev_mtx); 5494} 5495 5496/* 5497 * Remove a vdev from the L2ARC. 5498 */ 5499void 5500l2arc_remove_vdev(vdev_t *vd) 5501{ 5502 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5503 5504 /* 5505 * Find the device by vdev 5506 */ 5507 mutex_enter(&l2arc_dev_mtx); 5508 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5509 nextdev = list_next(l2arc_dev_list, dev); 5510 if (vd == dev->l2ad_vdev) { 5511 remdev = dev; 5512 break; 5513 } 5514 } 5515 ASSERT(remdev != NULL); 5516 5517 /* 5518 * Remove device from global list 5519 */ 5520 list_remove(l2arc_dev_list, remdev); 5521 l2arc_dev_last = NULL; /* may have been invalidated */ 5522 atomic_dec_64(&l2arc_ndev); 5523 mutex_exit(&l2arc_dev_mtx); 5524 5525 /* 5526 * Clear all buflists and ARC references. L2ARC device flush. 5527 */ 5528 l2arc_evict(remdev, 0, B_TRUE); 5529 list_destroy(remdev->l2ad_buflist); 5530 kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5531 kmem_free(remdev, sizeof (l2arc_dev_t)); 5532} 5533 5534void 5535l2arc_init(void) 5536{ 5537 l2arc_thread_exit = 0; 5538 l2arc_ndev = 0; 5539 l2arc_writes_sent = 0; 5540 l2arc_writes_done = 0; 5541 5542 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5543 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5544 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5545 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5546 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5547 5548 l2arc_dev_list = &L2ARC_dev_list; 5549 l2arc_free_on_write = &L2ARC_free_on_write; 5550 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5551 offsetof(l2arc_dev_t, l2ad_node)); 5552 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5553 offsetof(l2arc_data_free_t, l2df_list_node)); 5554} 5555 5556void 5557l2arc_fini(void) 5558{ 5559 /* 5560 * This is called from dmu_fini(), which is called from spa_fini(); 5561 * Because of this, we can assume that all l2arc devices have 5562 * already been removed when the pools themselves were removed. 5563 */ 5564 5565 l2arc_do_free_on_write(); 5566 5567 mutex_destroy(&l2arc_feed_thr_lock); 5568 cv_destroy(&l2arc_feed_thr_cv); 5569 mutex_destroy(&l2arc_dev_mtx); 5570 mutex_destroy(&l2arc_buflist_mtx); 5571 mutex_destroy(&l2arc_free_on_write_mtx); 5572 5573 list_destroy(l2arc_dev_list); 5574 list_destroy(l2arc_free_on_write); 5575} 5576 5577void 5578l2arc_start(void) 5579{ 5580 if (!(spa_mode_global & FWRITE)) 5581 return; 5582 5583 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5584 TS_RUN, minclsyspri); 5585} 5586 5587void 5588l2arc_stop(void) 5589{ 5590 if (!(spa_mode_global & FWRITE)) 5591 return; 5592 5593 mutex_enter(&l2arc_feed_thr_lock); 5594 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5595 l2arc_thread_exit = 1; 5596 while (l2arc_thread_exit != 0) 5597 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5598 mutex_exit(&l2arc_feed_thr_lock); 5599} 5600