arc.c revision 268649
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) 2013 by Delphix. All rights reserved. 24 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 25 * Copyright 2013 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_buf_evict() 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; 206 207TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max); 208TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 209TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 210SYSCTL_DECL(_vfs_zfs); 211SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 212 "Maximum ARC size"); 213SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 214 "Minimum ARC size"); 215 216/* 217 * Note that buffers can be in one of 6 states: 218 * ARC_anon - anonymous (discussed below) 219 * ARC_mru - recently used, currently cached 220 * ARC_mru_ghost - recentely used, no longer in cache 221 * ARC_mfu - frequently used, currently cached 222 * ARC_mfu_ghost - frequently used, no longer in cache 223 * ARC_l2c_only - exists in L2ARC but not other states 224 * When there are no active references to the buffer, they are 225 * are linked onto a list in one of these arc states. These are 226 * the only buffers that can be evicted or deleted. Within each 227 * state there are multiple lists, one for meta-data and one for 228 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 229 * etc.) is tracked separately so that it can be managed more 230 * explicitly: favored over data, limited explicitly. 231 * 232 * Anonymous buffers are buffers that are not associated with 233 * a DVA. These are buffers that hold dirty block copies 234 * before they are written to stable storage. By definition, 235 * they are "ref'd" and are considered part of arc_mru 236 * that cannot be freed. Generally, they will aquire a DVA 237 * as they are written and migrate onto the arc_mru list. 238 * 239 * The ARC_l2c_only state is for buffers that are in the second 240 * level ARC but no longer in any of the ARC_m* lists. The second 241 * level ARC itself may also contain buffers that are in any of 242 * the ARC_m* states - meaning that a buffer can exist in two 243 * places. The reason for the ARC_l2c_only state is to keep the 244 * buffer header in the hash table, so that reads that hit the 245 * second level ARC benefit from these fast lookups. 246 */ 247 248#define ARCS_LOCK_PAD CACHE_LINE_SIZE 249struct arcs_lock { 250 kmutex_t arcs_lock; 251#ifdef _KERNEL 252 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 253#endif 254}; 255 256/* 257 * must be power of two for mask use to work 258 * 259 */ 260#define ARC_BUFC_NUMDATALISTS 16 261#define ARC_BUFC_NUMMETADATALISTS 16 262#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 263 264typedef struct arc_state { 265 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 266 uint64_t arcs_size; /* total amount of data in this state */ 267 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 268 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 269} arc_state_t; 270 271#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 272 273/* The 6 states: */ 274static arc_state_t ARC_anon; 275static arc_state_t ARC_mru; 276static arc_state_t ARC_mru_ghost; 277static arc_state_t ARC_mfu; 278static arc_state_t ARC_mfu_ghost; 279static arc_state_t ARC_l2c_only; 280 281typedef struct arc_stats { 282 kstat_named_t arcstat_hits; 283 kstat_named_t arcstat_misses; 284 kstat_named_t arcstat_demand_data_hits; 285 kstat_named_t arcstat_demand_data_misses; 286 kstat_named_t arcstat_demand_metadata_hits; 287 kstat_named_t arcstat_demand_metadata_misses; 288 kstat_named_t arcstat_prefetch_data_hits; 289 kstat_named_t arcstat_prefetch_data_misses; 290 kstat_named_t arcstat_prefetch_metadata_hits; 291 kstat_named_t arcstat_prefetch_metadata_misses; 292 kstat_named_t arcstat_mru_hits; 293 kstat_named_t arcstat_mru_ghost_hits; 294 kstat_named_t arcstat_mfu_hits; 295 kstat_named_t arcstat_mfu_ghost_hits; 296 kstat_named_t arcstat_allocated; 297 kstat_named_t arcstat_deleted; 298 kstat_named_t arcstat_stolen; 299 kstat_named_t arcstat_recycle_miss; 300 /* 301 * Number of buffers that could not be evicted because the hash lock 302 * was held by another thread. The lock may not necessarily be held 303 * by something using the same buffer, since hash locks are shared 304 * by multiple buffers. 305 */ 306 kstat_named_t arcstat_mutex_miss; 307 /* 308 * Number of buffers skipped because they have I/O in progress, are 309 * indrect prefetch buffers that have not lived long enough, or are 310 * not from the spa we're trying to evict from. 311 */ 312 kstat_named_t arcstat_evict_skip; 313 kstat_named_t arcstat_evict_l2_cached; 314 kstat_named_t arcstat_evict_l2_eligible; 315 kstat_named_t arcstat_evict_l2_ineligible; 316 kstat_named_t arcstat_hash_elements; 317 kstat_named_t arcstat_hash_elements_max; 318 kstat_named_t arcstat_hash_collisions; 319 kstat_named_t arcstat_hash_chains; 320 kstat_named_t arcstat_hash_chain_max; 321 kstat_named_t arcstat_p; 322 kstat_named_t arcstat_c; 323 kstat_named_t arcstat_c_min; 324 kstat_named_t arcstat_c_max; 325 kstat_named_t arcstat_size; 326 kstat_named_t arcstat_hdr_size; 327 kstat_named_t arcstat_data_size; 328 kstat_named_t arcstat_other_size; 329 kstat_named_t arcstat_l2_hits; 330 kstat_named_t arcstat_l2_misses; 331 kstat_named_t arcstat_l2_feeds; 332 kstat_named_t arcstat_l2_rw_clash; 333 kstat_named_t arcstat_l2_read_bytes; 334 kstat_named_t arcstat_l2_write_bytes; 335 kstat_named_t arcstat_l2_writes_sent; 336 kstat_named_t arcstat_l2_writes_done; 337 kstat_named_t arcstat_l2_writes_error; 338 kstat_named_t arcstat_l2_writes_hdr_miss; 339 kstat_named_t arcstat_l2_evict_lock_retry; 340 kstat_named_t arcstat_l2_evict_reading; 341 kstat_named_t arcstat_l2_free_on_write; 342 kstat_named_t arcstat_l2_abort_lowmem; 343 kstat_named_t arcstat_l2_cksum_bad; 344 kstat_named_t arcstat_l2_io_error; 345 kstat_named_t arcstat_l2_size; 346 kstat_named_t arcstat_l2_asize; 347 kstat_named_t arcstat_l2_hdr_size; 348 kstat_named_t arcstat_l2_compress_successes; 349 kstat_named_t arcstat_l2_compress_zeros; 350 kstat_named_t arcstat_l2_compress_failures; 351 kstat_named_t arcstat_l2_write_trylock_fail; 352 kstat_named_t arcstat_l2_write_passed_headroom; 353 kstat_named_t arcstat_l2_write_spa_mismatch; 354 kstat_named_t arcstat_l2_write_in_l2; 355 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 356 kstat_named_t arcstat_l2_write_not_cacheable; 357 kstat_named_t arcstat_l2_write_full; 358 kstat_named_t arcstat_l2_write_buffer_iter; 359 kstat_named_t arcstat_l2_write_pios; 360 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 361 kstat_named_t arcstat_l2_write_buffer_list_iter; 362 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 363 kstat_named_t arcstat_memory_throttle_count; 364 kstat_named_t arcstat_duplicate_buffers; 365 kstat_named_t arcstat_duplicate_buffers_size; 366 kstat_named_t arcstat_duplicate_reads; 367} arc_stats_t; 368 369static arc_stats_t arc_stats = { 370 { "hits", KSTAT_DATA_UINT64 }, 371 { "misses", KSTAT_DATA_UINT64 }, 372 { "demand_data_hits", KSTAT_DATA_UINT64 }, 373 { "demand_data_misses", KSTAT_DATA_UINT64 }, 374 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 375 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 376 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 377 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 378 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 379 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 380 { "mru_hits", KSTAT_DATA_UINT64 }, 381 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 382 { "mfu_hits", KSTAT_DATA_UINT64 }, 383 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 384 { "allocated", KSTAT_DATA_UINT64 }, 385 { "deleted", KSTAT_DATA_UINT64 }, 386 { "stolen", KSTAT_DATA_UINT64 }, 387 { "recycle_miss", KSTAT_DATA_UINT64 }, 388 { "mutex_miss", KSTAT_DATA_UINT64 }, 389 { "evict_skip", KSTAT_DATA_UINT64 }, 390 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 391 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 392 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 393 { "hash_elements", KSTAT_DATA_UINT64 }, 394 { "hash_elements_max", KSTAT_DATA_UINT64 }, 395 { "hash_collisions", KSTAT_DATA_UINT64 }, 396 { "hash_chains", KSTAT_DATA_UINT64 }, 397 { "hash_chain_max", KSTAT_DATA_UINT64 }, 398 { "p", KSTAT_DATA_UINT64 }, 399 { "c", KSTAT_DATA_UINT64 }, 400 { "c_min", KSTAT_DATA_UINT64 }, 401 { "c_max", KSTAT_DATA_UINT64 }, 402 { "size", KSTAT_DATA_UINT64 }, 403 { "hdr_size", KSTAT_DATA_UINT64 }, 404 { "data_size", KSTAT_DATA_UINT64 }, 405 { "other_size", KSTAT_DATA_UINT64 }, 406 { "l2_hits", KSTAT_DATA_UINT64 }, 407 { "l2_misses", KSTAT_DATA_UINT64 }, 408 { "l2_feeds", KSTAT_DATA_UINT64 }, 409 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 410 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 411 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 412 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 413 { "l2_writes_done", KSTAT_DATA_UINT64 }, 414 { "l2_writes_error", KSTAT_DATA_UINT64 }, 415 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 416 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 417 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 418 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 419 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 420 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 421 { "l2_io_error", KSTAT_DATA_UINT64 }, 422 { "l2_size", KSTAT_DATA_UINT64 }, 423 { "l2_asize", KSTAT_DATA_UINT64 }, 424 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 425 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 426 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 427 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 428 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 429 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 430 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 431 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 432 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 433 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 434 { "l2_write_full", KSTAT_DATA_UINT64 }, 435 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 436 { "l2_write_pios", KSTAT_DATA_UINT64 }, 437 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 438 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 439 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 440 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 441 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 442 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 443 { "duplicate_reads", KSTAT_DATA_UINT64 } 444}; 445 446#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 447 448#define ARCSTAT_INCR(stat, val) \ 449 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 450 451#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 452#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 453 454#define ARCSTAT_MAX(stat, val) { \ 455 uint64_t m; \ 456 while ((val) > (m = arc_stats.stat.value.ui64) && \ 457 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 458 continue; \ 459} 460 461#define ARCSTAT_MAXSTAT(stat) \ 462 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 463 464/* 465 * We define a macro to allow ARC hits/misses to be easily broken down by 466 * two separate conditions, giving a total of four different subtypes for 467 * each of hits and misses (so eight statistics total). 468 */ 469#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 470 if (cond1) { \ 471 if (cond2) { \ 472 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 473 } else { \ 474 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 475 } \ 476 } else { \ 477 if (cond2) { \ 478 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 479 } else { \ 480 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 481 } \ 482 } 483 484kstat_t *arc_ksp; 485static arc_state_t *arc_anon; 486static arc_state_t *arc_mru; 487static arc_state_t *arc_mru_ghost; 488static arc_state_t *arc_mfu; 489static arc_state_t *arc_mfu_ghost; 490static arc_state_t *arc_l2c_only; 491 492/* 493 * There are several ARC variables that are critical to export as kstats -- 494 * but we don't want to have to grovel around in the kstat whenever we wish to 495 * manipulate them. For these variables, we therefore define them to be in 496 * terms of the statistic variable. This assures that we are not introducing 497 * the possibility of inconsistency by having shadow copies of the variables, 498 * while still allowing the code to be readable. 499 */ 500#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 501#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 502#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 503#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 504#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 505 506#define L2ARC_IS_VALID_COMPRESS(_c_) \ 507 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 508 509static int arc_no_grow; /* Don't try to grow cache size */ 510static uint64_t arc_tempreserve; 511static uint64_t arc_loaned_bytes; 512static uint64_t arc_meta_used; 513static uint64_t arc_meta_limit; 514static uint64_t arc_meta_max = 0; 515SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RD, &arc_meta_used, 0, 516 "ARC metadata used"); 517SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RW, &arc_meta_limit, 0, 518 "ARC metadata limit"); 519 520typedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 521 522typedef struct arc_callback arc_callback_t; 523 524struct arc_callback { 525 void *acb_private; 526 arc_done_func_t *acb_done; 527 arc_buf_t *acb_buf; 528 zio_t *acb_zio_dummy; 529 arc_callback_t *acb_next; 530}; 531 532typedef struct arc_write_callback arc_write_callback_t; 533 534struct arc_write_callback { 535 void *awcb_private; 536 arc_done_func_t *awcb_ready; 537 arc_done_func_t *awcb_physdone; 538 arc_done_func_t *awcb_done; 539 arc_buf_t *awcb_buf; 540}; 541 542struct arc_buf_hdr { 543 /* protected by hash lock */ 544 dva_t b_dva; 545 uint64_t b_birth; 546 uint64_t b_cksum0; 547 548 kmutex_t b_freeze_lock; 549 zio_cksum_t *b_freeze_cksum; 550 void *b_thawed; 551 552 arc_buf_hdr_t *b_hash_next; 553 arc_buf_t *b_buf; 554 uint32_t b_flags; 555 uint32_t b_datacnt; 556 557 arc_callback_t *b_acb; 558 kcondvar_t b_cv; 559 560 /* immutable */ 561 arc_buf_contents_t b_type; 562 uint64_t b_size; 563 uint64_t b_spa; 564 565 /* protected by arc state mutex */ 566 arc_state_t *b_state; 567 list_node_t b_arc_node; 568 569 /* updated atomically */ 570 clock_t b_arc_access; 571 572 /* self protecting */ 573 refcount_t b_refcnt; 574 575 l2arc_buf_hdr_t *b_l2hdr; 576 list_node_t b_l2node; 577}; 578 579static arc_buf_t *arc_eviction_list; 580static kmutex_t arc_eviction_mtx; 581static arc_buf_hdr_t arc_eviction_hdr; 582static void arc_get_data_buf(arc_buf_t *buf); 583static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock); 584static int arc_evict_needed(arc_buf_contents_t type); 585static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes); 586#ifdef illumos 587static void arc_buf_watch(arc_buf_t *buf); 588#endif /* illumos */ 589 590static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab); 591 592#define GHOST_STATE(state) \ 593 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 594 (state) == arc_l2c_only) 595 596/* 597 * Private ARC flags. These flags are private ARC only flags that will show up 598 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can 599 * be passed in as arc_flags in things like arc_read. However, these flags 600 * should never be passed and should only be set by ARC code. When adding new 601 * public flags, make sure not to smash the private ones. 602 */ 603 604#define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */ 605#define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */ 606#define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */ 607#define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */ 608#define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */ 609#define ARC_INDIRECT (1 << 14) /* this is an indirect block */ 610#define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */ 611#define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */ 612#define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */ 613#define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */ 614 615#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE) 616#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS) 617#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR) 618#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH) 619#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ) 620#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE) 621#define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS) 622#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE) 623#define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \ 624 (hdr)->b_l2hdr != NULL) 625#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING) 626#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED) 627#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD) 628 629/* 630 * Other sizes 631 */ 632 633#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 634#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 635 636/* 637 * Hash table routines 638 */ 639 640#define HT_LOCK_PAD CACHE_LINE_SIZE 641 642struct ht_lock { 643 kmutex_t ht_lock; 644#ifdef _KERNEL 645 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 646#endif 647}; 648 649#define BUF_LOCKS 256 650typedef struct buf_hash_table { 651 uint64_t ht_mask; 652 arc_buf_hdr_t **ht_table; 653 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 654} buf_hash_table_t; 655 656static buf_hash_table_t buf_hash_table; 657 658#define BUF_HASH_INDEX(spa, dva, birth) \ 659 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 660#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 661#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 662#define HDR_LOCK(hdr) \ 663 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 664 665uint64_t zfs_crc64_table[256]; 666 667/* 668 * Level 2 ARC 669 */ 670 671#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 672#define L2ARC_HEADROOM 2 /* num of writes */ 673/* 674 * If we discover during ARC scan any buffers to be compressed, we boost 675 * our headroom for the next scanning cycle by this percentage multiple. 676 */ 677#define L2ARC_HEADROOM_BOOST 200 678#define L2ARC_FEED_SECS 1 /* caching interval secs */ 679#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 680 681#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 682#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 683 684/* L2ARC Performance Tunables */ 685uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 686uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 687uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 688uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 689uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 690uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 691boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 692boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 693boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 694 695SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 696 &l2arc_write_max, 0, "max write size"); 697SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 698 &l2arc_write_boost, 0, "extra write during warmup"); 699SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 700 &l2arc_headroom, 0, "number of dev writes"); 701SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 702 &l2arc_feed_secs, 0, "interval seconds"); 703SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 704 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 705 706SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 707 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 708SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 709 &l2arc_feed_again, 0, "turbo warmup"); 710SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 711 &l2arc_norw, 0, "no reads during writes"); 712 713SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 714 &ARC_anon.arcs_size, 0, "size of anonymous state"); 715SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 716 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 717SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 718 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 719 720SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 721 &ARC_mru.arcs_size, 0, "size of mru state"); 722SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 723 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 724SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 725 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 726 727SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 728 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 729SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 730 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 731 "size of metadata in mru ghost state"); 732SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 733 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 734 "size of data in mru ghost state"); 735 736SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 737 &ARC_mfu.arcs_size, 0, "size of mfu state"); 738SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 739 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 740SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 741 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 742 743SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 744 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 745SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 746 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 747 "size of metadata in mfu ghost state"); 748SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 749 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 750 "size of data in mfu ghost state"); 751 752SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 753 &ARC_l2c_only.arcs_size, 0, "size of mru state"); 754 755/* 756 * L2ARC Internals 757 */ 758typedef struct l2arc_dev { 759 vdev_t *l2ad_vdev; /* vdev */ 760 spa_t *l2ad_spa; /* spa */ 761 uint64_t l2ad_hand; /* next write location */ 762 uint64_t l2ad_start; /* first addr on device */ 763 uint64_t l2ad_end; /* last addr on device */ 764 uint64_t l2ad_evict; /* last addr eviction reached */ 765 boolean_t l2ad_first; /* first sweep through */ 766 boolean_t l2ad_writing; /* currently writing */ 767 list_t *l2ad_buflist; /* buffer list */ 768 list_node_t l2ad_node; /* device list node */ 769} l2arc_dev_t; 770 771static list_t L2ARC_dev_list; /* device list */ 772static list_t *l2arc_dev_list; /* device list pointer */ 773static kmutex_t l2arc_dev_mtx; /* device list mutex */ 774static l2arc_dev_t *l2arc_dev_last; /* last device used */ 775static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 776static list_t L2ARC_free_on_write; /* free after write buf list */ 777static list_t *l2arc_free_on_write; /* free after write list ptr */ 778static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 779static uint64_t l2arc_ndev; /* number of devices */ 780 781typedef struct l2arc_read_callback { 782 arc_buf_t *l2rcb_buf; /* read buffer */ 783 spa_t *l2rcb_spa; /* spa */ 784 blkptr_t l2rcb_bp; /* original blkptr */ 785 zbookmark_t l2rcb_zb; /* original bookmark */ 786 int l2rcb_flags; /* original flags */ 787 enum zio_compress l2rcb_compress; /* applied compress */ 788} l2arc_read_callback_t; 789 790typedef struct l2arc_write_callback { 791 l2arc_dev_t *l2wcb_dev; /* device info */ 792 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 793} l2arc_write_callback_t; 794 795struct l2arc_buf_hdr { 796 /* protected by arc_buf_hdr mutex */ 797 l2arc_dev_t *b_dev; /* L2ARC device */ 798 uint64_t b_daddr; /* disk address, offset byte */ 799 /* compression applied to buffer data */ 800 enum zio_compress b_compress; 801 /* real alloc'd buffer size depending on b_compress applied */ 802 int b_asize; 803 /* temporary buffer holder for in-flight compressed data */ 804 void *b_tmp_cdata; 805}; 806 807typedef struct l2arc_data_free { 808 /* protected by l2arc_free_on_write_mtx */ 809 void *l2df_data; 810 size_t l2df_size; 811 void (*l2df_func)(void *, size_t); 812 list_node_t l2df_list_node; 813} l2arc_data_free_t; 814 815static kmutex_t l2arc_feed_thr_lock; 816static kcondvar_t l2arc_feed_thr_cv; 817static uint8_t l2arc_thread_exit; 818 819static void l2arc_read_done(zio_t *zio); 820static void l2arc_hdr_stat_add(void); 821static void l2arc_hdr_stat_remove(void); 822 823static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr); 824static void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, 825 enum zio_compress c); 826static void l2arc_release_cdata_buf(arc_buf_hdr_t *ab); 827 828static uint64_t 829buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 830{ 831 uint8_t *vdva = (uint8_t *)dva; 832 uint64_t crc = -1ULL; 833 int i; 834 835 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 836 837 for (i = 0; i < sizeof (dva_t); i++) 838 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 839 840 crc ^= (spa>>8) ^ birth; 841 842 return (crc); 843} 844 845#define BUF_EMPTY(buf) \ 846 ((buf)->b_dva.dva_word[0] == 0 && \ 847 (buf)->b_dva.dva_word[1] == 0 && \ 848 (buf)->b_cksum0 == 0) 849 850#define BUF_EQUAL(spa, dva, birth, buf) \ 851 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 852 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 853 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 854 855static void 856buf_discard_identity(arc_buf_hdr_t *hdr) 857{ 858 hdr->b_dva.dva_word[0] = 0; 859 hdr->b_dva.dva_word[1] = 0; 860 hdr->b_birth = 0; 861 hdr->b_cksum0 = 0; 862} 863 864static arc_buf_hdr_t * 865buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 866{ 867 const dva_t *dva = BP_IDENTITY(bp); 868 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 869 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 870 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 871 arc_buf_hdr_t *buf; 872 873 mutex_enter(hash_lock); 874 for (buf = buf_hash_table.ht_table[idx]; buf != NULL; 875 buf = buf->b_hash_next) { 876 if (BUF_EQUAL(spa, dva, birth, buf)) { 877 *lockp = hash_lock; 878 return (buf); 879 } 880 } 881 mutex_exit(hash_lock); 882 *lockp = NULL; 883 return (NULL); 884} 885 886/* 887 * Insert an entry into the hash table. If there is already an element 888 * equal to elem in the hash table, then the already existing element 889 * will be returned and the new element will not be inserted. 890 * Otherwise returns NULL. 891 */ 892static arc_buf_hdr_t * 893buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp) 894{ 895 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 896 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 897 arc_buf_hdr_t *fbuf; 898 uint32_t i; 899 900 ASSERT(!DVA_IS_EMPTY(&buf->b_dva)); 901 ASSERT(buf->b_birth != 0); 902 ASSERT(!HDR_IN_HASH_TABLE(buf)); 903 *lockp = hash_lock; 904 mutex_enter(hash_lock); 905 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL; 906 fbuf = fbuf->b_hash_next, i++) { 907 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf)) 908 return (fbuf); 909 } 910 911 buf->b_hash_next = buf_hash_table.ht_table[idx]; 912 buf_hash_table.ht_table[idx] = buf; 913 buf->b_flags |= ARC_IN_HASH_TABLE; 914 915 /* collect some hash table performance data */ 916 if (i > 0) { 917 ARCSTAT_BUMP(arcstat_hash_collisions); 918 if (i == 1) 919 ARCSTAT_BUMP(arcstat_hash_chains); 920 921 ARCSTAT_MAX(arcstat_hash_chain_max, i); 922 } 923 924 ARCSTAT_BUMP(arcstat_hash_elements); 925 ARCSTAT_MAXSTAT(arcstat_hash_elements); 926 927 return (NULL); 928} 929 930static void 931buf_hash_remove(arc_buf_hdr_t *buf) 932{ 933 arc_buf_hdr_t *fbuf, **bufp; 934 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 935 936 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 937 ASSERT(HDR_IN_HASH_TABLE(buf)); 938 939 bufp = &buf_hash_table.ht_table[idx]; 940 while ((fbuf = *bufp) != buf) { 941 ASSERT(fbuf != NULL); 942 bufp = &fbuf->b_hash_next; 943 } 944 *bufp = buf->b_hash_next; 945 buf->b_hash_next = NULL; 946 buf->b_flags &= ~ARC_IN_HASH_TABLE; 947 948 /* collect some hash table performance data */ 949 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 950 951 if (buf_hash_table.ht_table[idx] && 952 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 953 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 954} 955 956/* 957 * Global data structures and functions for the buf kmem cache. 958 */ 959static kmem_cache_t *hdr_cache; 960static kmem_cache_t *buf_cache; 961 962static void 963buf_fini(void) 964{ 965 int i; 966 967 kmem_free(buf_hash_table.ht_table, 968 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 969 for (i = 0; i < BUF_LOCKS; i++) 970 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 971 kmem_cache_destroy(hdr_cache); 972 kmem_cache_destroy(buf_cache); 973} 974 975/* 976 * Constructor callback - called when the cache is empty 977 * and a new buf is requested. 978 */ 979/* ARGSUSED */ 980static int 981hdr_cons(void *vbuf, void *unused, int kmflag) 982{ 983 arc_buf_hdr_t *buf = vbuf; 984 985 bzero(buf, sizeof (arc_buf_hdr_t)); 986 refcount_create(&buf->b_refcnt); 987 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL); 988 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 989 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 990 991 return (0); 992} 993 994/* ARGSUSED */ 995static int 996buf_cons(void *vbuf, void *unused, int kmflag) 997{ 998 arc_buf_t *buf = vbuf; 999 1000 bzero(buf, sizeof (arc_buf_t)); 1001 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1002 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1003 1004 return (0); 1005} 1006 1007/* 1008 * Destructor callback - called when a cached buf is 1009 * no longer required. 1010 */ 1011/* ARGSUSED */ 1012static void 1013hdr_dest(void *vbuf, void *unused) 1014{ 1015 arc_buf_hdr_t *buf = vbuf; 1016 1017 ASSERT(BUF_EMPTY(buf)); 1018 refcount_destroy(&buf->b_refcnt); 1019 cv_destroy(&buf->b_cv); 1020 mutex_destroy(&buf->b_freeze_lock); 1021 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1022} 1023 1024/* ARGSUSED */ 1025static void 1026buf_dest(void *vbuf, void *unused) 1027{ 1028 arc_buf_t *buf = vbuf; 1029 1030 mutex_destroy(&buf->b_evict_lock); 1031 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1032} 1033 1034/* 1035 * Reclaim callback -- invoked when memory is low. 1036 */ 1037/* ARGSUSED */ 1038static void 1039hdr_recl(void *unused) 1040{ 1041 dprintf("hdr_recl called\n"); 1042 /* 1043 * umem calls the reclaim func when we destroy the buf cache, 1044 * which is after we do arc_fini(). 1045 */ 1046 if (!arc_dead) 1047 cv_signal(&arc_reclaim_thr_cv); 1048} 1049 1050static void 1051buf_init(void) 1052{ 1053 uint64_t *ct; 1054 uint64_t hsize = 1ULL << 12; 1055 int i, j; 1056 1057 /* 1058 * The hash table is big enough to fill all of physical memory 1059 * with an average 64K block size. The table will take up 1060 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers). 1061 */ 1062 while (hsize * 65536 < (uint64_t)physmem * PAGESIZE) 1063 hsize <<= 1; 1064retry: 1065 buf_hash_table.ht_mask = hsize - 1; 1066 buf_hash_table.ht_table = 1067 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1068 if (buf_hash_table.ht_table == NULL) { 1069 ASSERT(hsize > (1ULL << 8)); 1070 hsize >>= 1; 1071 goto retry; 1072 } 1073 1074 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1075 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1076 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1077 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1078 1079 for (i = 0; i < 256; i++) 1080 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1081 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1082 1083 for (i = 0; i < BUF_LOCKS; i++) { 1084 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1085 NULL, MUTEX_DEFAULT, NULL); 1086 } 1087} 1088 1089#define ARC_MINTIME (hz>>4) /* 62 ms */ 1090 1091static void 1092arc_cksum_verify(arc_buf_t *buf) 1093{ 1094 zio_cksum_t zc; 1095 1096 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1097 return; 1098 1099 mutex_enter(&buf->b_hdr->b_freeze_lock); 1100 if (buf->b_hdr->b_freeze_cksum == NULL || 1101 (buf->b_hdr->b_flags & ARC_IO_ERROR)) { 1102 mutex_exit(&buf->b_hdr->b_freeze_lock); 1103 return; 1104 } 1105 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1106 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1107 panic("buffer modified while frozen!"); 1108 mutex_exit(&buf->b_hdr->b_freeze_lock); 1109} 1110 1111static int 1112arc_cksum_equal(arc_buf_t *buf) 1113{ 1114 zio_cksum_t zc; 1115 int equal; 1116 1117 mutex_enter(&buf->b_hdr->b_freeze_lock); 1118 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1119 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1120 mutex_exit(&buf->b_hdr->b_freeze_lock); 1121 1122 return (equal); 1123} 1124 1125static void 1126arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1127{ 1128 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1129 return; 1130 1131 mutex_enter(&buf->b_hdr->b_freeze_lock); 1132 if (buf->b_hdr->b_freeze_cksum != NULL) { 1133 mutex_exit(&buf->b_hdr->b_freeze_lock); 1134 return; 1135 } 1136 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1137 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1138 buf->b_hdr->b_freeze_cksum); 1139 mutex_exit(&buf->b_hdr->b_freeze_lock); 1140#ifdef illumos 1141 arc_buf_watch(buf); 1142#endif /* illumos */ 1143} 1144 1145#ifdef illumos 1146#ifndef _KERNEL 1147typedef struct procctl { 1148 long cmd; 1149 prwatch_t prwatch; 1150} procctl_t; 1151#endif 1152 1153/* ARGSUSED */ 1154static void 1155arc_buf_unwatch(arc_buf_t *buf) 1156{ 1157#ifndef _KERNEL 1158 if (arc_watch) { 1159 int result; 1160 procctl_t ctl; 1161 ctl.cmd = PCWATCH; 1162 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1163 ctl.prwatch.pr_size = 0; 1164 ctl.prwatch.pr_wflags = 0; 1165 result = write(arc_procfd, &ctl, sizeof (ctl)); 1166 ASSERT3U(result, ==, sizeof (ctl)); 1167 } 1168#endif 1169} 1170 1171/* ARGSUSED */ 1172static void 1173arc_buf_watch(arc_buf_t *buf) 1174{ 1175#ifndef _KERNEL 1176 if (arc_watch) { 1177 int result; 1178 procctl_t ctl; 1179 ctl.cmd = PCWATCH; 1180 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1181 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1182 ctl.prwatch.pr_wflags = WA_WRITE; 1183 result = write(arc_procfd, &ctl, sizeof (ctl)); 1184 ASSERT3U(result, ==, sizeof (ctl)); 1185 } 1186#endif 1187} 1188#endif /* illumos */ 1189 1190void 1191arc_buf_thaw(arc_buf_t *buf) 1192{ 1193 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1194 if (buf->b_hdr->b_state != arc_anon) 1195 panic("modifying non-anon buffer!"); 1196 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS) 1197 panic("modifying buffer while i/o in progress!"); 1198 arc_cksum_verify(buf); 1199 } 1200 1201 mutex_enter(&buf->b_hdr->b_freeze_lock); 1202 if (buf->b_hdr->b_freeze_cksum != NULL) { 1203 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1204 buf->b_hdr->b_freeze_cksum = NULL; 1205 } 1206 1207 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1208 if (buf->b_hdr->b_thawed) 1209 kmem_free(buf->b_hdr->b_thawed, 1); 1210 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1211 } 1212 1213 mutex_exit(&buf->b_hdr->b_freeze_lock); 1214 1215#ifdef illumos 1216 arc_buf_unwatch(buf); 1217#endif /* illumos */ 1218} 1219 1220void 1221arc_buf_freeze(arc_buf_t *buf) 1222{ 1223 kmutex_t *hash_lock; 1224 1225 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1226 return; 1227 1228 hash_lock = HDR_LOCK(buf->b_hdr); 1229 mutex_enter(hash_lock); 1230 1231 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1232 buf->b_hdr->b_state == arc_anon); 1233 arc_cksum_compute(buf, B_FALSE); 1234 mutex_exit(hash_lock); 1235 1236} 1237 1238static void 1239get_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock) 1240{ 1241 uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth); 1242 1243 if (ab->b_type == ARC_BUFC_METADATA) 1244 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1245 else { 1246 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1247 buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1248 } 1249 1250 *list = &state->arcs_lists[buf_hashid]; 1251 *lock = ARCS_LOCK(state, buf_hashid); 1252} 1253 1254 1255static void 1256add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1257{ 1258 ASSERT(MUTEX_HELD(hash_lock)); 1259 1260 if ((refcount_add(&ab->b_refcnt, tag) == 1) && 1261 (ab->b_state != arc_anon)) { 1262 uint64_t delta = ab->b_size * ab->b_datacnt; 1263 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type]; 1264 list_t *list; 1265 kmutex_t *lock; 1266 1267 get_buf_info(ab, ab->b_state, &list, &lock); 1268 ASSERT(!MUTEX_HELD(lock)); 1269 mutex_enter(lock); 1270 ASSERT(list_link_active(&ab->b_arc_node)); 1271 list_remove(list, ab); 1272 if (GHOST_STATE(ab->b_state)) { 1273 ASSERT0(ab->b_datacnt); 1274 ASSERT3P(ab->b_buf, ==, NULL); 1275 delta = ab->b_size; 1276 } 1277 ASSERT(delta > 0); 1278 ASSERT3U(*size, >=, delta); 1279 atomic_add_64(size, -delta); 1280 mutex_exit(lock); 1281 /* remove the prefetch flag if we get a reference */ 1282 if (ab->b_flags & ARC_PREFETCH) 1283 ab->b_flags &= ~ARC_PREFETCH; 1284 } 1285} 1286 1287static int 1288remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1289{ 1290 int cnt; 1291 arc_state_t *state = ab->b_state; 1292 1293 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1294 ASSERT(!GHOST_STATE(state)); 1295 1296 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) && 1297 (state != arc_anon)) { 1298 uint64_t *size = &state->arcs_lsize[ab->b_type]; 1299 list_t *list; 1300 kmutex_t *lock; 1301 1302 get_buf_info(ab, state, &list, &lock); 1303 ASSERT(!MUTEX_HELD(lock)); 1304 mutex_enter(lock); 1305 ASSERT(!list_link_active(&ab->b_arc_node)); 1306 list_insert_head(list, ab); 1307 ASSERT(ab->b_datacnt > 0); 1308 atomic_add_64(size, ab->b_size * ab->b_datacnt); 1309 mutex_exit(lock); 1310 } 1311 return (cnt); 1312} 1313 1314/* 1315 * Move the supplied buffer to the indicated state. The mutex 1316 * for the buffer must be held by the caller. 1317 */ 1318static void 1319arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock) 1320{ 1321 arc_state_t *old_state = ab->b_state; 1322 int64_t refcnt = refcount_count(&ab->b_refcnt); 1323 uint64_t from_delta, to_delta; 1324 list_t *list; 1325 kmutex_t *lock; 1326 1327 ASSERT(MUTEX_HELD(hash_lock)); 1328 ASSERT3P(new_state, !=, old_state); 1329 ASSERT(refcnt == 0 || ab->b_datacnt > 0); 1330 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state)); 1331 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon); 1332 1333 from_delta = to_delta = ab->b_datacnt * ab->b_size; 1334 1335 /* 1336 * If this buffer is evictable, transfer it from the 1337 * old state list to the new state list. 1338 */ 1339 if (refcnt == 0) { 1340 if (old_state != arc_anon) { 1341 int use_mutex; 1342 uint64_t *size = &old_state->arcs_lsize[ab->b_type]; 1343 1344 get_buf_info(ab, old_state, &list, &lock); 1345 use_mutex = !MUTEX_HELD(lock); 1346 if (use_mutex) 1347 mutex_enter(lock); 1348 1349 ASSERT(list_link_active(&ab->b_arc_node)); 1350 list_remove(list, ab); 1351 1352 /* 1353 * If prefetching out of the ghost cache, 1354 * we will have a non-zero datacnt. 1355 */ 1356 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) { 1357 /* ghost elements have a ghost size */ 1358 ASSERT(ab->b_buf == NULL); 1359 from_delta = ab->b_size; 1360 } 1361 ASSERT3U(*size, >=, from_delta); 1362 atomic_add_64(size, -from_delta); 1363 1364 if (use_mutex) 1365 mutex_exit(lock); 1366 } 1367 if (new_state != arc_anon) { 1368 int use_mutex; 1369 uint64_t *size = &new_state->arcs_lsize[ab->b_type]; 1370 1371 get_buf_info(ab, new_state, &list, &lock); 1372 use_mutex = !MUTEX_HELD(lock); 1373 if (use_mutex) 1374 mutex_enter(lock); 1375 1376 list_insert_head(list, ab); 1377 1378 /* ghost elements have a ghost size */ 1379 if (GHOST_STATE(new_state)) { 1380 ASSERT(ab->b_datacnt == 0); 1381 ASSERT(ab->b_buf == NULL); 1382 to_delta = ab->b_size; 1383 } 1384 atomic_add_64(size, to_delta); 1385 1386 if (use_mutex) 1387 mutex_exit(lock); 1388 } 1389 } 1390 1391 ASSERT(!BUF_EMPTY(ab)); 1392 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab)) 1393 buf_hash_remove(ab); 1394 1395 /* adjust state sizes */ 1396 if (to_delta) 1397 atomic_add_64(&new_state->arcs_size, to_delta); 1398 if (from_delta) { 1399 ASSERT3U(old_state->arcs_size, >=, from_delta); 1400 atomic_add_64(&old_state->arcs_size, -from_delta); 1401 } 1402 ab->b_state = new_state; 1403 1404 /* adjust l2arc hdr stats */ 1405 if (new_state == arc_l2c_only) 1406 l2arc_hdr_stat_add(); 1407 else if (old_state == arc_l2c_only) 1408 l2arc_hdr_stat_remove(); 1409} 1410 1411void 1412arc_space_consume(uint64_t space, arc_space_type_t type) 1413{ 1414 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1415 1416 switch (type) { 1417 case ARC_SPACE_DATA: 1418 ARCSTAT_INCR(arcstat_data_size, space); 1419 break; 1420 case ARC_SPACE_OTHER: 1421 ARCSTAT_INCR(arcstat_other_size, space); 1422 break; 1423 case ARC_SPACE_HDRS: 1424 ARCSTAT_INCR(arcstat_hdr_size, space); 1425 break; 1426 case ARC_SPACE_L2HDRS: 1427 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1428 break; 1429 } 1430 1431 atomic_add_64(&arc_meta_used, space); 1432 atomic_add_64(&arc_size, space); 1433} 1434 1435void 1436arc_space_return(uint64_t space, arc_space_type_t type) 1437{ 1438 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1439 1440 switch (type) { 1441 case ARC_SPACE_DATA: 1442 ARCSTAT_INCR(arcstat_data_size, -space); 1443 break; 1444 case ARC_SPACE_OTHER: 1445 ARCSTAT_INCR(arcstat_other_size, -space); 1446 break; 1447 case ARC_SPACE_HDRS: 1448 ARCSTAT_INCR(arcstat_hdr_size, -space); 1449 break; 1450 case ARC_SPACE_L2HDRS: 1451 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1452 break; 1453 } 1454 1455 ASSERT(arc_meta_used >= space); 1456 if (arc_meta_max < arc_meta_used) 1457 arc_meta_max = arc_meta_used; 1458 atomic_add_64(&arc_meta_used, -space); 1459 ASSERT(arc_size >= space); 1460 atomic_add_64(&arc_size, -space); 1461} 1462 1463void * 1464arc_data_buf_alloc(uint64_t size) 1465{ 1466 if (arc_evict_needed(ARC_BUFC_DATA)) 1467 cv_signal(&arc_reclaim_thr_cv); 1468 atomic_add_64(&arc_size, size); 1469 return (zio_data_buf_alloc(size)); 1470} 1471 1472void 1473arc_data_buf_free(void *buf, uint64_t size) 1474{ 1475 zio_data_buf_free(buf, size); 1476 ASSERT(arc_size >= size); 1477 atomic_add_64(&arc_size, -size); 1478} 1479 1480arc_buf_t * 1481arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1482{ 1483 arc_buf_hdr_t *hdr; 1484 arc_buf_t *buf; 1485 1486 ASSERT3U(size, >, 0); 1487 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1488 ASSERT(BUF_EMPTY(hdr)); 1489 hdr->b_size = size; 1490 hdr->b_type = type; 1491 hdr->b_spa = spa_load_guid(spa); 1492 hdr->b_state = arc_anon; 1493 hdr->b_arc_access = 0; 1494 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1495 buf->b_hdr = hdr; 1496 buf->b_data = NULL; 1497 buf->b_efunc = NULL; 1498 buf->b_private = NULL; 1499 buf->b_next = NULL; 1500 hdr->b_buf = buf; 1501 arc_get_data_buf(buf); 1502 hdr->b_datacnt = 1; 1503 hdr->b_flags = 0; 1504 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1505 (void) refcount_add(&hdr->b_refcnt, tag); 1506 1507 return (buf); 1508} 1509 1510static char *arc_onloan_tag = "onloan"; 1511 1512/* 1513 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1514 * flight data by arc_tempreserve_space() until they are "returned". Loaned 1515 * buffers must be returned to the arc before they can be used by the DMU or 1516 * freed. 1517 */ 1518arc_buf_t * 1519arc_loan_buf(spa_t *spa, int size) 1520{ 1521 arc_buf_t *buf; 1522 1523 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1524 1525 atomic_add_64(&arc_loaned_bytes, size); 1526 return (buf); 1527} 1528 1529/* 1530 * Return a loaned arc buffer to the arc. 1531 */ 1532void 1533arc_return_buf(arc_buf_t *buf, void *tag) 1534{ 1535 arc_buf_hdr_t *hdr = buf->b_hdr; 1536 1537 ASSERT(buf->b_data != NULL); 1538 (void) refcount_add(&hdr->b_refcnt, tag); 1539 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1540 1541 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1542} 1543 1544/* Detach an arc_buf from a dbuf (tag) */ 1545void 1546arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1547{ 1548 arc_buf_hdr_t *hdr; 1549 1550 ASSERT(buf->b_data != NULL); 1551 hdr = buf->b_hdr; 1552 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1553 (void) refcount_remove(&hdr->b_refcnt, tag); 1554 buf->b_efunc = NULL; 1555 buf->b_private = NULL; 1556 1557 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1558} 1559 1560static arc_buf_t * 1561arc_buf_clone(arc_buf_t *from) 1562{ 1563 arc_buf_t *buf; 1564 arc_buf_hdr_t *hdr = from->b_hdr; 1565 uint64_t size = hdr->b_size; 1566 1567 ASSERT(hdr->b_state != arc_anon); 1568 1569 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1570 buf->b_hdr = hdr; 1571 buf->b_data = NULL; 1572 buf->b_efunc = NULL; 1573 buf->b_private = NULL; 1574 buf->b_next = hdr->b_buf; 1575 hdr->b_buf = buf; 1576 arc_get_data_buf(buf); 1577 bcopy(from->b_data, buf->b_data, size); 1578 1579 /* 1580 * This buffer already exists in the arc so create a duplicate 1581 * copy for the caller. If the buffer is associated with user data 1582 * then track the size and number of duplicates. These stats will be 1583 * updated as duplicate buffers are created and destroyed. 1584 */ 1585 if (hdr->b_type == ARC_BUFC_DATA) { 1586 ARCSTAT_BUMP(arcstat_duplicate_buffers); 1587 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1588 } 1589 hdr->b_datacnt += 1; 1590 return (buf); 1591} 1592 1593void 1594arc_buf_add_ref(arc_buf_t *buf, void* tag) 1595{ 1596 arc_buf_hdr_t *hdr; 1597 kmutex_t *hash_lock; 1598 1599 /* 1600 * Check to see if this buffer is evicted. Callers 1601 * must verify b_data != NULL to know if the add_ref 1602 * was successful. 1603 */ 1604 mutex_enter(&buf->b_evict_lock); 1605 if (buf->b_data == NULL) { 1606 mutex_exit(&buf->b_evict_lock); 1607 return; 1608 } 1609 hash_lock = HDR_LOCK(buf->b_hdr); 1610 mutex_enter(hash_lock); 1611 hdr = buf->b_hdr; 1612 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1613 mutex_exit(&buf->b_evict_lock); 1614 1615 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1616 add_reference(hdr, hash_lock, tag); 1617 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1618 arc_access(hdr, hash_lock); 1619 mutex_exit(hash_lock); 1620 ARCSTAT_BUMP(arcstat_hits); 1621 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 1622 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1623 data, metadata, hits); 1624} 1625 1626/* 1627 * Free the arc data buffer. If it is an l2arc write in progress, 1628 * the buffer is placed on l2arc_free_on_write to be freed later. 1629 */ 1630static void 1631arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1632{ 1633 arc_buf_hdr_t *hdr = buf->b_hdr; 1634 1635 if (HDR_L2_WRITING(hdr)) { 1636 l2arc_data_free_t *df; 1637 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1638 df->l2df_data = buf->b_data; 1639 df->l2df_size = hdr->b_size; 1640 df->l2df_func = free_func; 1641 mutex_enter(&l2arc_free_on_write_mtx); 1642 list_insert_head(l2arc_free_on_write, df); 1643 mutex_exit(&l2arc_free_on_write_mtx); 1644 ARCSTAT_BUMP(arcstat_l2_free_on_write); 1645 } else { 1646 free_func(buf->b_data, hdr->b_size); 1647 } 1648} 1649 1650static void 1651arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all) 1652{ 1653 arc_buf_t **bufp; 1654 1655 /* free up data associated with the buf */ 1656 if (buf->b_data) { 1657 arc_state_t *state = buf->b_hdr->b_state; 1658 uint64_t size = buf->b_hdr->b_size; 1659 arc_buf_contents_t type = buf->b_hdr->b_type; 1660 1661 arc_cksum_verify(buf); 1662#ifdef illumos 1663 arc_buf_unwatch(buf); 1664#endif /* illumos */ 1665 1666 if (!recycle) { 1667 if (type == ARC_BUFC_METADATA) { 1668 arc_buf_data_free(buf, zio_buf_free); 1669 arc_space_return(size, ARC_SPACE_DATA); 1670 } else { 1671 ASSERT(type == ARC_BUFC_DATA); 1672 arc_buf_data_free(buf, zio_data_buf_free); 1673 ARCSTAT_INCR(arcstat_data_size, -size); 1674 atomic_add_64(&arc_size, -size); 1675 } 1676 } 1677 if (list_link_active(&buf->b_hdr->b_arc_node)) { 1678 uint64_t *cnt = &state->arcs_lsize[type]; 1679 1680 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1681 ASSERT(state != arc_anon); 1682 1683 ASSERT3U(*cnt, >=, size); 1684 atomic_add_64(cnt, -size); 1685 } 1686 ASSERT3U(state->arcs_size, >=, size); 1687 atomic_add_64(&state->arcs_size, -size); 1688 buf->b_data = NULL; 1689 1690 /* 1691 * If we're destroying a duplicate buffer make sure 1692 * that the appropriate statistics are updated. 1693 */ 1694 if (buf->b_hdr->b_datacnt > 1 && 1695 buf->b_hdr->b_type == ARC_BUFC_DATA) { 1696 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1697 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1698 } 1699 ASSERT(buf->b_hdr->b_datacnt > 0); 1700 buf->b_hdr->b_datacnt -= 1; 1701 } 1702 1703 /* only remove the buf if requested */ 1704 if (!all) 1705 return; 1706 1707 /* remove the buf from the hdr list */ 1708 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1709 continue; 1710 *bufp = buf->b_next; 1711 buf->b_next = NULL; 1712 1713 ASSERT(buf->b_efunc == NULL); 1714 1715 /* clean up the buf */ 1716 buf->b_hdr = NULL; 1717 kmem_cache_free(buf_cache, buf); 1718} 1719 1720static void 1721arc_hdr_destroy(arc_buf_hdr_t *hdr) 1722{ 1723 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1724 ASSERT3P(hdr->b_state, ==, arc_anon); 1725 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1726 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1727 1728 if (l2hdr != NULL) { 1729 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1730 /* 1731 * To prevent arc_free() and l2arc_evict() from 1732 * attempting to free the same buffer at the same time, 1733 * a FREE_IN_PROGRESS flag is given to arc_free() to 1734 * give it priority. l2arc_evict() can't destroy this 1735 * header while we are waiting on l2arc_buflist_mtx. 1736 * 1737 * The hdr may be removed from l2ad_buflist before we 1738 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1739 */ 1740 if (!buflist_held) { 1741 mutex_enter(&l2arc_buflist_mtx); 1742 l2hdr = hdr->b_l2hdr; 1743 } 1744 1745 if (l2hdr != NULL) { 1746 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1747 hdr->b_size, 0); 1748 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1749 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1750 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1751 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1752 if (hdr->b_state == arc_l2c_only) 1753 l2arc_hdr_stat_remove(); 1754 hdr->b_l2hdr = NULL; 1755 } 1756 1757 if (!buflist_held) 1758 mutex_exit(&l2arc_buflist_mtx); 1759 } 1760 1761 if (!BUF_EMPTY(hdr)) { 1762 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1763 buf_discard_identity(hdr); 1764 } 1765 while (hdr->b_buf) { 1766 arc_buf_t *buf = hdr->b_buf; 1767 1768 if (buf->b_efunc) { 1769 mutex_enter(&arc_eviction_mtx); 1770 mutex_enter(&buf->b_evict_lock); 1771 ASSERT(buf->b_hdr != NULL); 1772 arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1773 hdr->b_buf = buf->b_next; 1774 buf->b_hdr = &arc_eviction_hdr; 1775 buf->b_next = arc_eviction_list; 1776 arc_eviction_list = buf; 1777 mutex_exit(&buf->b_evict_lock); 1778 mutex_exit(&arc_eviction_mtx); 1779 } else { 1780 arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1781 } 1782 } 1783 if (hdr->b_freeze_cksum != NULL) { 1784 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1785 hdr->b_freeze_cksum = NULL; 1786 } 1787 if (hdr->b_thawed) { 1788 kmem_free(hdr->b_thawed, 1); 1789 hdr->b_thawed = NULL; 1790 } 1791 1792 ASSERT(!list_link_active(&hdr->b_arc_node)); 1793 ASSERT3P(hdr->b_hash_next, ==, NULL); 1794 ASSERT3P(hdr->b_acb, ==, NULL); 1795 kmem_cache_free(hdr_cache, hdr); 1796} 1797 1798void 1799arc_buf_free(arc_buf_t *buf, void *tag) 1800{ 1801 arc_buf_hdr_t *hdr = buf->b_hdr; 1802 int hashed = hdr->b_state != arc_anon; 1803 1804 ASSERT(buf->b_efunc == NULL); 1805 ASSERT(buf->b_data != NULL); 1806 1807 if (hashed) { 1808 kmutex_t *hash_lock = HDR_LOCK(hdr); 1809 1810 mutex_enter(hash_lock); 1811 hdr = buf->b_hdr; 1812 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1813 1814 (void) remove_reference(hdr, hash_lock, tag); 1815 if (hdr->b_datacnt > 1) { 1816 arc_buf_destroy(buf, FALSE, TRUE); 1817 } else { 1818 ASSERT(buf == hdr->b_buf); 1819 ASSERT(buf->b_efunc == NULL); 1820 hdr->b_flags |= ARC_BUF_AVAILABLE; 1821 } 1822 mutex_exit(hash_lock); 1823 } else if (HDR_IO_IN_PROGRESS(hdr)) { 1824 int destroy_hdr; 1825 /* 1826 * We are in the middle of an async write. Don't destroy 1827 * this buffer unless the write completes before we finish 1828 * decrementing the reference count. 1829 */ 1830 mutex_enter(&arc_eviction_mtx); 1831 (void) remove_reference(hdr, NULL, tag); 1832 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1833 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1834 mutex_exit(&arc_eviction_mtx); 1835 if (destroy_hdr) 1836 arc_hdr_destroy(hdr); 1837 } else { 1838 if (remove_reference(hdr, NULL, tag) > 0) 1839 arc_buf_destroy(buf, FALSE, TRUE); 1840 else 1841 arc_hdr_destroy(hdr); 1842 } 1843} 1844 1845boolean_t 1846arc_buf_remove_ref(arc_buf_t *buf, void* tag) 1847{ 1848 arc_buf_hdr_t *hdr = buf->b_hdr; 1849 kmutex_t *hash_lock = HDR_LOCK(hdr); 1850 boolean_t no_callback = (buf->b_efunc == NULL); 1851 1852 if (hdr->b_state == arc_anon) { 1853 ASSERT(hdr->b_datacnt == 1); 1854 arc_buf_free(buf, tag); 1855 return (no_callback); 1856 } 1857 1858 mutex_enter(hash_lock); 1859 hdr = buf->b_hdr; 1860 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1861 ASSERT(hdr->b_state != arc_anon); 1862 ASSERT(buf->b_data != NULL); 1863 1864 (void) remove_reference(hdr, hash_lock, tag); 1865 if (hdr->b_datacnt > 1) { 1866 if (no_callback) 1867 arc_buf_destroy(buf, FALSE, TRUE); 1868 } else if (no_callback) { 1869 ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1870 ASSERT(buf->b_efunc == NULL); 1871 hdr->b_flags |= ARC_BUF_AVAILABLE; 1872 } 1873 ASSERT(no_callback || hdr->b_datacnt > 1 || 1874 refcount_is_zero(&hdr->b_refcnt)); 1875 mutex_exit(hash_lock); 1876 return (no_callback); 1877} 1878 1879int 1880arc_buf_size(arc_buf_t *buf) 1881{ 1882 return (buf->b_hdr->b_size); 1883} 1884 1885/* 1886 * Called from the DMU to determine if the current buffer should be 1887 * evicted. In order to ensure proper locking, the eviction must be initiated 1888 * from the DMU. Return true if the buffer is associated with user data and 1889 * duplicate buffers still exist. 1890 */ 1891boolean_t 1892arc_buf_eviction_needed(arc_buf_t *buf) 1893{ 1894 arc_buf_hdr_t *hdr; 1895 boolean_t evict_needed = B_FALSE; 1896 1897 if (zfs_disable_dup_eviction) 1898 return (B_FALSE); 1899 1900 mutex_enter(&buf->b_evict_lock); 1901 hdr = buf->b_hdr; 1902 if (hdr == NULL) { 1903 /* 1904 * We are in arc_do_user_evicts(); let that function 1905 * perform the eviction. 1906 */ 1907 ASSERT(buf->b_data == NULL); 1908 mutex_exit(&buf->b_evict_lock); 1909 return (B_FALSE); 1910 } else if (buf->b_data == NULL) { 1911 /* 1912 * We have already been added to the arc eviction list; 1913 * recommend eviction. 1914 */ 1915 ASSERT3P(hdr, ==, &arc_eviction_hdr); 1916 mutex_exit(&buf->b_evict_lock); 1917 return (B_TRUE); 1918 } 1919 1920 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 1921 evict_needed = B_TRUE; 1922 1923 mutex_exit(&buf->b_evict_lock); 1924 return (evict_needed); 1925} 1926 1927/* 1928 * Evict buffers from list until we've removed the specified number of 1929 * bytes. Move the removed buffers to the appropriate evict state. 1930 * If the recycle flag is set, then attempt to "recycle" a buffer: 1931 * - look for a buffer to evict that is `bytes' long. 1932 * - return the data block from this buffer rather than freeing it. 1933 * This flag is used by callers that are trying to make space for a 1934 * new buffer in a full arc cache. 1935 * 1936 * This function makes a "best effort". It skips over any buffers 1937 * it can't get a hash_lock on, and so may not catch all candidates. 1938 * It may also return without evicting as much space as requested. 1939 */ 1940static void * 1941arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 1942 arc_buf_contents_t type) 1943{ 1944 arc_state_t *evicted_state; 1945 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 1946 int64_t bytes_remaining; 1947 arc_buf_hdr_t *ab, *ab_prev = NULL; 1948 list_t *evicted_list, *list, *evicted_list_start, *list_start; 1949 kmutex_t *lock, *evicted_lock; 1950 kmutex_t *hash_lock; 1951 boolean_t have_lock; 1952 void *stolen = NULL; 1953 arc_buf_hdr_t marker = { 0 }; 1954 int count = 0; 1955 static int evict_metadata_offset, evict_data_offset; 1956 int i, idx, offset, list_count, lists; 1957 1958 ASSERT(state == arc_mru || state == arc_mfu); 1959 1960 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 1961 1962 if (type == ARC_BUFC_METADATA) { 1963 offset = 0; 1964 list_count = ARC_BUFC_NUMMETADATALISTS; 1965 list_start = &state->arcs_lists[0]; 1966 evicted_list_start = &evicted_state->arcs_lists[0]; 1967 idx = evict_metadata_offset; 1968 } else { 1969 offset = ARC_BUFC_NUMMETADATALISTS; 1970 list_start = &state->arcs_lists[offset]; 1971 evicted_list_start = &evicted_state->arcs_lists[offset]; 1972 list_count = ARC_BUFC_NUMDATALISTS; 1973 idx = evict_data_offset; 1974 } 1975 bytes_remaining = evicted_state->arcs_lsize[type]; 1976 lists = 0; 1977 1978evict_start: 1979 list = &list_start[idx]; 1980 evicted_list = &evicted_list_start[idx]; 1981 lock = ARCS_LOCK(state, (offset + idx)); 1982 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 1983 1984 mutex_enter(lock); 1985 mutex_enter(evicted_lock); 1986 1987 for (ab = list_tail(list); ab; ab = ab_prev) { 1988 ab_prev = list_prev(list, ab); 1989 bytes_remaining -= (ab->b_size * ab->b_datacnt); 1990 /* prefetch buffers have a minimum lifespan */ 1991 if (HDR_IO_IN_PROGRESS(ab) || 1992 (spa && ab->b_spa != spa) || 1993 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 1994 ddi_get_lbolt() - ab->b_arc_access < 1995 arc_min_prefetch_lifespan)) { 1996 skipped++; 1997 continue; 1998 } 1999 /* "lookahead" for better eviction candidate */ 2000 if (recycle && ab->b_size != bytes && 2001 ab_prev && ab_prev->b_size == bytes) 2002 continue; 2003 2004 /* ignore markers */ 2005 if (ab->b_spa == 0) 2006 continue; 2007 2008 /* 2009 * It may take a long time to evict all the bufs requested. 2010 * To avoid blocking all arc activity, periodically drop 2011 * the arcs_mtx and give other threads a chance to run 2012 * before reacquiring the lock. 2013 * 2014 * If we are looking for a buffer to recycle, we are in 2015 * the hot code path, so don't sleep. 2016 */ 2017 if (!recycle && count++ > arc_evict_iterations) { 2018 list_insert_after(list, ab, &marker); 2019 mutex_exit(evicted_lock); 2020 mutex_exit(lock); 2021 kpreempt(KPREEMPT_SYNC); 2022 mutex_enter(lock); 2023 mutex_enter(evicted_lock); 2024 ab_prev = list_prev(list, &marker); 2025 list_remove(list, &marker); 2026 count = 0; 2027 continue; 2028 } 2029 2030 hash_lock = HDR_LOCK(ab); 2031 have_lock = MUTEX_HELD(hash_lock); 2032 if (have_lock || mutex_tryenter(hash_lock)) { 2033 ASSERT0(refcount_count(&ab->b_refcnt)); 2034 ASSERT(ab->b_datacnt > 0); 2035 while (ab->b_buf) { 2036 arc_buf_t *buf = ab->b_buf; 2037 if (!mutex_tryenter(&buf->b_evict_lock)) { 2038 missed += 1; 2039 break; 2040 } 2041 if (buf->b_data) { 2042 bytes_evicted += ab->b_size; 2043 if (recycle && ab->b_type == type && 2044 ab->b_size == bytes && 2045 !HDR_L2_WRITING(ab)) { 2046 stolen = buf->b_data; 2047 recycle = FALSE; 2048 } 2049 } 2050 if (buf->b_efunc) { 2051 mutex_enter(&arc_eviction_mtx); 2052 arc_buf_destroy(buf, 2053 buf->b_data == stolen, FALSE); 2054 ab->b_buf = buf->b_next; 2055 buf->b_hdr = &arc_eviction_hdr; 2056 buf->b_next = arc_eviction_list; 2057 arc_eviction_list = buf; 2058 mutex_exit(&arc_eviction_mtx); 2059 mutex_exit(&buf->b_evict_lock); 2060 } else { 2061 mutex_exit(&buf->b_evict_lock); 2062 arc_buf_destroy(buf, 2063 buf->b_data == stolen, TRUE); 2064 } 2065 } 2066 2067 if (ab->b_l2hdr) { 2068 ARCSTAT_INCR(arcstat_evict_l2_cached, 2069 ab->b_size); 2070 } else { 2071 if (l2arc_write_eligible(ab->b_spa, ab)) { 2072 ARCSTAT_INCR(arcstat_evict_l2_eligible, 2073 ab->b_size); 2074 } else { 2075 ARCSTAT_INCR( 2076 arcstat_evict_l2_ineligible, 2077 ab->b_size); 2078 } 2079 } 2080 2081 if (ab->b_datacnt == 0) { 2082 arc_change_state(evicted_state, ab, hash_lock); 2083 ASSERT(HDR_IN_HASH_TABLE(ab)); 2084 ab->b_flags |= ARC_IN_HASH_TABLE; 2085 ab->b_flags &= ~ARC_BUF_AVAILABLE; 2086 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 2087 } 2088 if (!have_lock) 2089 mutex_exit(hash_lock); 2090 if (bytes >= 0 && bytes_evicted >= bytes) 2091 break; 2092 if (bytes_remaining > 0) { 2093 mutex_exit(evicted_lock); 2094 mutex_exit(lock); 2095 idx = ((idx + 1) & (list_count - 1)); 2096 lists++; 2097 goto evict_start; 2098 } 2099 } else { 2100 missed += 1; 2101 } 2102 } 2103 2104 mutex_exit(evicted_lock); 2105 mutex_exit(lock); 2106 2107 idx = ((idx + 1) & (list_count - 1)); 2108 lists++; 2109 2110 if (bytes_evicted < bytes) { 2111 if (lists < list_count) 2112 goto evict_start; 2113 else 2114 dprintf("only evicted %lld bytes from %x", 2115 (longlong_t)bytes_evicted, state); 2116 } 2117 if (type == ARC_BUFC_METADATA) 2118 evict_metadata_offset = idx; 2119 else 2120 evict_data_offset = idx; 2121 2122 if (skipped) 2123 ARCSTAT_INCR(arcstat_evict_skip, skipped); 2124 2125 if (missed) 2126 ARCSTAT_INCR(arcstat_mutex_miss, missed); 2127 2128 /* 2129 * Note: we have just evicted some data into the ghost state, 2130 * potentially putting the ghost size over the desired size. Rather 2131 * that evicting from the ghost list in this hot code path, leave 2132 * this chore to the arc_reclaim_thread(). 2133 */ 2134 2135 if (stolen) 2136 ARCSTAT_BUMP(arcstat_stolen); 2137 return (stolen); 2138} 2139 2140/* 2141 * Remove buffers from list until we've removed the specified number of 2142 * bytes. Destroy the buffers that are removed. 2143 */ 2144static void 2145arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2146{ 2147 arc_buf_hdr_t *ab, *ab_prev; 2148 arc_buf_hdr_t marker = { 0 }; 2149 list_t *list, *list_start; 2150 kmutex_t *hash_lock, *lock; 2151 uint64_t bytes_deleted = 0; 2152 uint64_t bufs_skipped = 0; 2153 int count = 0; 2154 static int evict_offset; 2155 int list_count, idx = evict_offset; 2156 int offset, lists = 0; 2157 2158 ASSERT(GHOST_STATE(state)); 2159 2160 /* 2161 * data lists come after metadata lists 2162 */ 2163 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2164 list_count = ARC_BUFC_NUMDATALISTS; 2165 offset = ARC_BUFC_NUMMETADATALISTS; 2166 2167evict_start: 2168 list = &list_start[idx]; 2169 lock = ARCS_LOCK(state, idx + offset); 2170 2171 mutex_enter(lock); 2172 for (ab = list_tail(list); ab; ab = ab_prev) { 2173 ab_prev = list_prev(list, ab); 2174 if (ab->b_type > ARC_BUFC_NUMTYPES) 2175 panic("invalid ab=%p", (void *)ab); 2176 if (spa && ab->b_spa != spa) 2177 continue; 2178 2179 /* ignore markers */ 2180 if (ab->b_spa == 0) 2181 continue; 2182 2183 hash_lock = HDR_LOCK(ab); 2184 /* caller may be trying to modify this buffer, skip it */ 2185 if (MUTEX_HELD(hash_lock)) 2186 continue; 2187 2188 /* 2189 * It may take a long time to evict all the bufs requested. 2190 * To avoid blocking all arc activity, periodically drop 2191 * the arcs_mtx and give other threads a chance to run 2192 * before reacquiring the lock. 2193 */ 2194 if (count++ > arc_evict_iterations) { 2195 list_insert_after(list, ab, &marker); 2196 mutex_exit(lock); 2197 kpreempt(KPREEMPT_SYNC); 2198 mutex_enter(lock); 2199 ab_prev = list_prev(list, &marker); 2200 list_remove(list, &marker); 2201 count = 0; 2202 continue; 2203 } 2204 if (mutex_tryenter(hash_lock)) { 2205 ASSERT(!HDR_IO_IN_PROGRESS(ab)); 2206 ASSERT(ab->b_buf == NULL); 2207 ARCSTAT_BUMP(arcstat_deleted); 2208 bytes_deleted += ab->b_size; 2209 2210 if (ab->b_l2hdr != NULL) { 2211 /* 2212 * This buffer is cached on the 2nd Level ARC; 2213 * don't destroy the header. 2214 */ 2215 arc_change_state(arc_l2c_only, ab, hash_lock); 2216 mutex_exit(hash_lock); 2217 } else { 2218 arc_change_state(arc_anon, ab, hash_lock); 2219 mutex_exit(hash_lock); 2220 arc_hdr_destroy(ab); 2221 } 2222 2223 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 2224 if (bytes >= 0 && bytes_deleted >= bytes) 2225 break; 2226 } else if (bytes < 0) { 2227 /* 2228 * Insert a list marker and then wait for the 2229 * hash lock to become available. Once its 2230 * available, restart from where we left off. 2231 */ 2232 list_insert_after(list, ab, &marker); 2233 mutex_exit(lock); 2234 mutex_enter(hash_lock); 2235 mutex_exit(hash_lock); 2236 mutex_enter(lock); 2237 ab_prev = list_prev(list, &marker); 2238 list_remove(list, &marker); 2239 } else { 2240 bufs_skipped += 1; 2241 } 2242 2243 } 2244 mutex_exit(lock); 2245 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2246 lists++; 2247 2248 if (lists < list_count) 2249 goto evict_start; 2250 2251 evict_offset = idx; 2252 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2253 (bytes < 0 || bytes_deleted < bytes)) { 2254 list_start = &state->arcs_lists[0]; 2255 list_count = ARC_BUFC_NUMMETADATALISTS; 2256 offset = lists = 0; 2257 goto evict_start; 2258 } 2259 2260 if (bufs_skipped) { 2261 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2262 ASSERT(bytes >= 0); 2263 } 2264 2265 if (bytes_deleted < bytes) 2266 dprintf("only deleted %lld bytes from %p", 2267 (longlong_t)bytes_deleted, state); 2268} 2269 2270static void 2271arc_adjust(void) 2272{ 2273 int64_t adjustment, delta; 2274 2275 /* 2276 * Adjust MRU size 2277 */ 2278 2279 adjustment = MIN((int64_t)(arc_size - arc_c), 2280 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2281 arc_p)); 2282 2283 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2284 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2285 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2286 adjustment -= delta; 2287 } 2288 2289 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2290 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2291 (void) arc_evict(arc_mru, 0, delta, FALSE, 2292 ARC_BUFC_METADATA); 2293 } 2294 2295 /* 2296 * Adjust MFU size 2297 */ 2298 2299 adjustment = arc_size - arc_c; 2300 2301 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2302 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2303 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2304 adjustment -= delta; 2305 } 2306 2307 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2308 int64_t delta = MIN(adjustment, 2309 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2310 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2311 ARC_BUFC_METADATA); 2312 } 2313 2314 /* 2315 * Adjust ghost lists 2316 */ 2317 2318 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2319 2320 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2321 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2322 arc_evict_ghost(arc_mru_ghost, 0, delta); 2323 } 2324 2325 adjustment = 2326 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2327 2328 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2329 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2330 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2331 } 2332} 2333 2334static void 2335arc_do_user_evicts(void) 2336{ 2337 static arc_buf_t *tmp_arc_eviction_list; 2338 2339 /* 2340 * Move list over to avoid LOR 2341 */ 2342restart: 2343 mutex_enter(&arc_eviction_mtx); 2344 tmp_arc_eviction_list = arc_eviction_list; 2345 arc_eviction_list = NULL; 2346 mutex_exit(&arc_eviction_mtx); 2347 2348 while (tmp_arc_eviction_list != NULL) { 2349 arc_buf_t *buf = tmp_arc_eviction_list; 2350 tmp_arc_eviction_list = buf->b_next; 2351 mutex_enter(&buf->b_evict_lock); 2352 buf->b_hdr = NULL; 2353 mutex_exit(&buf->b_evict_lock); 2354 2355 if (buf->b_efunc != NULL) 2356 VERIFY(buf->b_efunc(buf) == 0); 2357 2358 buf->b_efunc = NULL; 2359 buf->b_private = NULL; 2360 kmem_cache_free(buf_cache, buf); 2361 } 2362 2363 if (arc_eviction_list != NULL) 2364 goto restart; 2365} 2366 2367/* 2368 * Flush all *evictable* data from the cache for the given spa. 2369 * NOTE: this will not touch "active" (i.e. referenced) data. 2370 */ 2371void 2372arc_flush(spa_t *spa) 2373{ 2374 uint64_t guid = 0; 2375 2376 if (spa) 2377 guid = spa_load_guid(spa); 2378 2379 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2380 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2381 if (spa) 2382 break; 2383 } 2384 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2385 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2386 if (spa) 2387 break; 2388 } 2389 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2390 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2391 if (spa) 2392 break; 2393 } 2394 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2395 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2396 if (spa) 2397 break; 2398 } 2399 2400 arc_evict_ghost(arc_mru_ghost, guid, -1); 2401 arc_evict_ghost(arc_mfu_ghost, guid, -1); 2402 2403 mutex_enter(&arc_reclaim_thr_lock); 2404 arc_do_user_evicts(); 2405 mutex_exit(&arc_reclaim_thr_lock); 2406 ASSERT(spa || arc_eviction_list == NULL); 2407} 2408 2409void 2410arc_shrink(void) 2411{ 2412 if (arc_c > arc_c_min) { 2413 uint64_t to_free; 2414 2415#ifdef _KERNEL 2416 to_free = arc_c >> arc_shrink_shift; 2417#else 2418 to_free = arc_c >> arc_shrink_shift; 2419#endif 2420 if (arc_c > arc_c_min + to_free) 2421 atomic_add_64(&arc_c, -to_free); 2422 else 2423 arc_c = arc_c_min; 2424 2425 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2426 if (arc_c > arc_size) 2427 arc_c = MAX(arc_size, arc_c_min); 2428 if (arc_p > arc_c) 2429 arc_p = (arc_c >> 1); 2430 ASSERT(arc_c >= arc_c_min); 2431 ASSERT((int64_t)arc_p >= 0); 2432 } 2433 2434 if (arc_size > arc_c) 2435 arc_adjust(); 2436} 2437 2438static int needfree = 0; 2439 2440static int 2441arc_reclaim_needed(void) 2442{ 2443 2444#ifdef _KERNEL 2445 2446 if (needfree) 2447 return (1); 2448 2449 /* 2450 * Cooperate with pagedaemon when it's time for it to scan 2451 * and reclaim some pages. 2452 */ 2453 if (vm_paging_needed()) 2454 return (1); 2455 2456#ifdef sun 2457 /* 2458 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2459 */ 2460 extra = desfree; 2461 2462 /* 2463 * check that we're out of range of the pageout scanner. It starts to 2464 * schedule paging if freemem is less than lotsfree and needfree. 2465 * lotsfree is the high-water mark for pageout, and needfree is the 2466 * number of needed free pages. We add extra pages here to make sure 2467 * the scanner doesn't start up while we're freeing memory. 2468 */ 2469 if (freemem < lotsfree + needfree + extra) 2470 return (1); 2471 2472 /* 2473 * check to make sure that swapfs has enough space so that anon 2474 * reservations can still succeed. anon_resvmem() checks that the 2475 * availrmem is greater than swapfs_minfree, and the number of reserved 2476 * swap pages. We also add a bit of extra here just to prevent 2477 * circumstances from getting really dire. 2478 */ 2479 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2480 return (1); 2481 2482#if defined(__i386) 2483 /* 2484 * If we're on an i386 platform, it's possible that we'll exhaust the 2485 * kernel heap space before we ever run out of available physical 2486 * memory. Most checks of the size of the heap_area compare against 2487 * tune.t_minarmem, which is the minimum available real memory that we 2488 * can have in the system. However, this is generally fixed at 25 pages 2489 * which is so low that it's useless. In this comparison, we seek to 2490 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2491 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2492 * free) 2493 */ 2494 if (btop(vmem_size(heap_arena, VMEM_FREE)) < 2495 (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2)) 2496 return (1); 2497#endif 2498#else /* !sun */ 2499 if (kmem_used() > (kmem_size() * 3) / 4) 2500 return (1); 2501#endif /* sun */ 2502 2503#else 2504 if (spa_get_random(100) == 0) 2505 return (1); 2506#endif 2507 return (0); 2508} 2509 2510extern kmem_cache_t *zio_buf_cache[]; 2511extern kmem_cache_t *zio_data_buf_cache[]; 2512 2513static void 2514arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2515{ 2516 size_t i; 2517 kmem_cache_t *prev_cache = NULL; 2518 kmem_cache_t *prev_data_cache = NULL; 2519 2520#ifdef _KERNEL 2521 if (arc_meta_used >= arc_meta_limit) { 2522 /* 2523 * We are exceeding our meta-data cache limit. 2524 * Purge some DNLC entries to release holds on meta-data. 2525 */ 2526 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2527 } 2528#if defined(__i386) 2529 /* 2530 * Reclaim unused memory from all kmem caches. 2531 */ 2532 kmem_reap(); 2533#endif 2534#endif 2535 2536 /* 2537 * An aggressive reclamation will shrink the cache size as well as 2538 * reap free buffers from the arc kmem caches. 2539 */ 2540 if (strat == ARC_RECLAIM_AGGR) 2541 arc_shrink(); 2542 2543 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2544 if (zio_buf_cache[i] != prev_cache) { 2545 prev_cache = zio_buf_cache[i]; 2546 kmem_cache_reap_now(zio_buf_cache[i]); 2547 } 2548 if (zio_data_buf_cache[i] != prev_data_cache) { 2549 prev_data_cache = zio_data_buf_cache[i]; 2550 kmem_cache_reap_now(zio_data_buf_cache[i]); 2551 } 2552 } 2553 kmem_cache_reap_now(buf_cache); 2554 kmem_cache_reap_now(hdr_cache); 2555} 2556 2557static void 2558arc_reclaim_thread(void *dummy __unused) 2559{ 2560 clock_t growtime = 0; 2561 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2562 callb_cpr_t cpr; 2563 2564 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2565 2566 mutex_enter(&arc_reclaim_thr_lock); 2567 while (arc_thread_exit == 0) { 2568 if (arc_reclaim_needed()) { 2569 2570 if (arc_no_grow) { 2571 if (last_reclaim == ARC_RECLAIM_CONS) { 2572 last_reclaim = ARC_RECLAIM_AGGR; 2573 } else { 2574 last_reclaim = ARC_RECLAIM_CONS; 2575 } 2576 } else { 2577 arc_no_grow = TRUE; 2578 last_reclaim = ARC_RECLAIM_AGGR; 2579 membar_producer(); 2580 } 2581 2582 /* reset the growth delay for every reclaim */ 2583 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2584 2585 if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2586 /* 2587 * If needfree is TRUE our vm_lowmem hook 2588 * was called and in that case we must free some 2589 * memory, so switch to aggressive mode. 2590 */ 2591 arc_no_grow = TRUE; 2592 last_reclaim = ARC_RECLAIM_AGGR; 2593 } 2594 arc_kmem_reap_now(last_reclaim); 2595 arc_warm = B_TRUE; 2596 2597 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2598 arc_no_grow = FALSE; 2599 } 2600 2601 arc_adjust(); 2602 2603 if (arc_eviction_list != NULL) 2604 arc_do_user_evicts(); 2605 2606#ifdef _KERNEL 2607 if (needfree) { 2608 needfree = 0; 2609 wakeup(&needfree); 2610 } 2611#endif 2612 2613 /* block until needed, or one second, whichever is shorter */ 2614 CALLB_CPR_SAFE_BEGIN(&cpr); 2615 (void) cv_timedwait(&arc_reclaim_thr_cv, 2616 &arc_reclaim_thr_lock, hz); 2617 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2618 } 2619 2620 arc_thread_exit = 0; 2621 cv_broadcast(&arc_reclaim_thr_cv); 2622 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2623 thread_exit(); 2624} 2625 2626/* 2627 * Adapt arc info given the number of bytes we are trying to add and 2628 * the state that we are comming from. This function is only called 2629 * when we are adding new content to the cache. 2630 */ 2631static void 2632arc_adapt(int bytes, arc_state_t *state) 2633{ 2634 int mult; 2635 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2636 2637 if (state == arc_l2c_only) 2638 return; 2639 2640 ASSERT(bytes > 0); 2641 /* 2642 * Adapt the target size of the MRU list: 2643 * - if we just hit in the MRU ghost list, then increase 2644 * the target size of the MRU list. 2645 * - if we just hit in the MFU ghost list, then increase 2646 * the target size of the MFU list by decreasing the 2647 * target size of the MRU list. 2648 */ 2649 if (state == arc_mru_ghost) { 2650 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2651 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2652 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2653 2654 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2655 } else if (state == arc_mfu_ghost) { 2656 uint64_t delta; 2657 2658 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2659 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2660 mult = MIN(mult, 10); 2661 2662 delta = MIN(bytes * mult, arc_p); 2663 arc_p = MAX(arc_p_min, arc_p - delta); 2664 } 2665 ASSERT((int64_t)arc_p >= 0); 2666 2667 if (arc_reclaim_needed()) { 2668 cv_signal(&arc_reclaim_thr_cv); 2669 return; 2670 } 2671 2672 if (arc_no_grow) 2673 return; 2674 2675 if (arc_c >= arc_c_max) 2676 return; 2677 2678 /* 2679 * If we're within (2 * maxblocksize) bytes of the target 2680 * cache size, increment the target cache size 2681 */ 2682 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2683 atomic_add_64(&arc_c, (int64_t)bytes); 2684 if (arc_c > arc_c_max) 2685 arc_c = arc_c_max; 2686 else if (state == arc_anon) 2687 atomic_add_64(&arc_p, (int64_t)bytes); 2688 if (arc_p > arc_c) 2689 arc_p = arc_c; 2690 } 2691 ASSERT((int64_t)arc_p >= 0); 2692} 2693 2694/* 2695 * Check if the cache has reached its limits and eviction is required 2696 * prior to insert. 2697 */ 2698static int 2699arc_evict_needed(arc_buf_contents_t type) 2700{ 2701 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2702 return (1); 2703 2704#ifdef sun 2705#ifdef _KERNEL 2706 /* 2707 * If zio data pages are being allocated out of a separate heap segment, 2708 * then enforce that the size of available vmem for this area remains 2709 * above about 1/32nd free. 2710 */ 2711 if (type == ARC_BUFC_DATA && zio_arena != NULL && 2712 vmem_size(zio_arena, VMEM_FREE) < 2713 (vmem_size(zio_arena, VMEM_ALLOC) >> 5)) 2714 return (1); 2715#endif 2716#endif /* sun */ 2717 2718 if (arc_reclaim_needed()) 2719 return (1); 2720 2721 return (arc_size > arc_c); 2722} 2723 2724/* 2725 * The buffer, supplied as the first argument, needs a data block. 2726 * So, if we are at cache max, determine which cache should be victimized. 2727 * We have the following cases: 2728 * 2729 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2730 * In this situation if we're out of space, but the resident size of the MFU is 2731 * under the limit, victimize the MFU cache to satisfy this insertion request. 2732 * 2733 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2734 * Here, we've used up all of the available space for the MRU, so we need to 2735 * evict from our own cache instead. Evict from the set of resident MRU 2736 * entries. 2737 * 2738 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2739 * c minus p represents the MFU space in the cache, since p is the size of the 2740 * cache that is dedicated to the MRU. In this situation there's still space on 2741 * the MFU side, so the MRU side needs to be victimized. 2742 * 2743 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2744 * MFU's resident set is consuming more space than it has been allotted. In 2745 * this situation, we must victimize our own cache, the MFU, for this insertion. 2746 */ 2747static void 2748arc_get_data_buf(arc_buf_t *buf) 2749{ 2750 arc_state_t *state = buf->b_hdr->b_state; 2751 uint64_t size = buf->b_hdr->b_size; 2752 arc_buf_contents_t type = buf->b_hdr->b_type; 2753 2754 arc_adapt(size, state); 2755 2756 /* 2757 * We have not yet reached cache maximum size, 2758 * just allocate a new buffer. 2759 */ 2760 if (!arc_evict_needed(type)) { 2761 if (type == ARC_BUFC_METADATA) { 2762 buf->b_data = zio_buf_alloc(size); 2763 arc_space_consume(size, ARC_SPACE_DATA); 2764 } else { 2765 ASSERT(type == ARC_BUFC_DATA); 2766 buf->b_data = zio_data_buf_alloc(size); 2767 ARCSTAT_INCR(arcstat_data_size, size); 2768 atomic_add_64(&arc_size, size); 2769 } 2770 goto out; 2771 } 2772 2773 /* 2774 * If we are prefetching from the mfu ghost list, this buffer 2775 * will end up on the mru list; so steal space from there. 2776 */ 2777 if (state == arc_mfu_ghost) 2778 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2779 else if (state == arc_mru_ghost) 2780 state = arc_mru; 2781 2782 if (state == arc_mru || state == arc_anon) { 2783 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2784 state = (arc_mfu->arcs_lsize[type] >= size && 2785 arc_p > mru_used) ? arc_mfu : arc_mru; 2786 } else { 2787 /* MFU cases */ 2788 uint64_t mfu_space = arc_c - arc_p; 2789 state = (arc_mru->arcs_lsize[type] >= size && 2790 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2791 } 2792 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2793 if (type == ARC_BUFC_METADATA) { 2794 buf->b_data = zio_buf_alloc(size); 2795 arc_space_consume(size, ARC_SPACE_DATA); 2796 } else { 2797 ASSERT(type == ARC_BUFC_DATA); 2798 buf->b_data = zio_data_buf_alloc(size); 2799 ARCSTAT_INCR(arcstat_data_size, size); 2800 atomic_add_64(&arc_size, size); 2801 } 2802 ARCSTAT_BUMP(arcstat_recycle_miss); 2803 } 2804 ASSERT(buf->b_data != NULL); 2805out: 2806 /* 2807 * Update the state size. Note that ghost states have a 2808 * "ghost size" and so don't need to be updated. 2809 */ 2810 if (!GHOST_STATE(buf->b_hdr->b_state)) { 2811 arc_buf_hdr_t *hdr = buf->b_hdr; 2812 2813 atomic_add_64(&hdr->b_state->arcs_size, size); 2814 if (list_link_active(&hdr->b_arc_node)) { 2815 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2816 atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2817 } 2818 /* 2819 * If we are growing the cache, and we are adding anonymous 2820 * data, and we have outgrown arc_p, update arc_p 2821 */ 2822 if (arc_size < arc_c && hdr->b_state == arc_anon && 2823 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2824 arc_p = MIN(arc_c, arc_p + size); 2825 } 2826 ARCSTAT_BUMP(arcstat_allocated); 2827} 2828 2829/* 2830 * This routine is called whenever a buffer is accessed. 2831 * NOTE: the hash lock is dropped in this function. 2832 */ 2833static void 2834arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2835{ 2836 clock_t now; 2837 2838 ASSERT(MUTEX_HELD(hash_lock)); 2839 2840 if (buf->b_state == arc_anon) { 2841 /* 2842 * This buffer is not in the cache, and does not 2843 * appear in our "ghost" list. Add the new buffer 2844 * to the MRU state. 2845 */ 2846 2847 ASSERT(buf->b_arc_access == 0); 2848 buf->b_arc_access = ddi_get_lbolt(); 2849 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2850 arc_change_state(arc_mru, buf, hash_lock); 2851 2852 } else if (buf->b_state == arc_mru) { 2853 now = ddi_get_lbolt(); 2854 2855 /* 2856 * If this buffer is here because of a prefetch, then either: 2857 * - clear the flag if this is a "referencing" read 2858 * (any subsequent access will bump this into the MFU state). 2859 * or 2860 * - move the buffer to the head of the list if this is 2861 * another prefetch (to make it less likely to be evicted). 2862 */ 2863 if ((buf->b_flags & ARC_PREFETCH) != 0) { 2864 if (refcount_count(&buf->b_refcnt) == 0) { 2865 ASSERT(list_link_active(&buf->b_arc_node)); 2866 } else { 2867 buf->b_flags &= ~ARC_PREFETCH; 2868 ARCSTAT_BUMP(arcstat_mru_hits); 2869 } 2870 buf->b_arc_access = now; 2871 return; 2872 } 2873 2874 /* 2875 * This buffer has been "accessed" only once so far, 2876 * but it is still in the cache. Move it to the MFU 2877 * state. 2878 */ 2879 if (now > buf->b_arc_access + ARC_MINTIME) { 2880 /* 2881 * More than 125ms have passed since we 2882 * instantiated this buffer. Move it to the 2883 * most frequently used state. 2884 */ 2885 buf->b_arc_access = now; 2886 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2887 arc_change_state(arc_mfu, buf, hash_lock); 2888 } 2889 ARCSTAT_BUMP(arcstat_mru_hits); 2890 } else if (buf->b_state == arc_mru_ghost) { 2891 arc_state_t *new_state; 2892 /* 2893 * This buffer has been "accessed" recently, but 2894 * was evicted from the cache. Move it to the 2895 * MFU state. 2896 */ 2897 2898 if (buf->b_flags & ARC_PREFETCH) { 2899 new_state = arc_mru; 2900 if (refcount_count(&buf->b_refcnt) > 0) 2901 buf->b_flags &= ~ARC_PREFETCH; 2902 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2903 } else { 2904 new_state = arc_mfu; 2905 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2906 } 2907 2908 buf->b_arc_access = ddi_get_lbolt(); 2909 arc_change_state(new_state, buf, hash_lock); 2910 2911 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 2912 } else if (buf->b_state == arc_mfu) { 2913 /* 2914 * This buffer has been accessed more than once and is 2915 * still in the cache. Keep it in the MFU state. 2916 * 2917 * NOTE: an add_reference() that occurred when we did 2918 * the arc_read() will have kicked this off the list. 2919 * If it was a prefetch, we will explicitly move it to 2920 * the head of the list now. 2921 */ 2922 if ((buf->b_flags & ARC_PREFETCH) != 0) { 2923 ASSERT(refcount_count(&buf->b_refcnt) == 0); 2924 ASSERT(list_link_active(&buf->b_arc_node)); 2925 } 2926 ARCSTAT_BUMP(arcstat_mfu_hits); 2927 buf->b_arc_access = ddi_get_lbolt(); 2928 } else if (buf->b_state == arc_mfu_ghost) { 2929 arc_state_t *new_state = arc_mfu; 2930 /* 2931 * This buffer has been accessed more than once but has 2932 * been evicted from the cache. Move it back to the 2933 * MFU state. 2934 */ 2935 2936 if (buf->b_flags & ARC_PREFETCH) { 2937 /* 2938 * This is a prefetch access... 2939 * move this block back to the MRU state. 2940 */ 2941 ASSERT0(refcount_count(&buf->b_refcnt)); 2942 new_state = arc_mru; 2943 } 2944 2945 buf->b_arc_access = ddi_get_lbolt(); 2946 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2947 arc_change_state(new_state, buf, hash_lock); 2948 2949 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 2950 } else if (buf->b_state == arc_l2c_only) { 2951 /* 2952 * This buffer is on the 2nd Level ARC. 2953 */ 2954 2955 buf->b_arc_access = ddi_get_lbolt(); 2956 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2957 arc_change_state(arc_mfu, buf, hash_lock); 2958 } else { 2959 ASSERT(!"invalid arc state"); 2960 } 2961} 2962 2963/* a generic arc_done_func_t which you can use */ 2964/* ARGSUSED */ 2965void 2966arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 2967{ 2968 if (zio == NULL || zio->io_error == 0) 2969 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 2970 VERIFY(arc_buf_remove_ref(buf, arg)); 2971} 2972 2973/* a generic arc_done_func_t */ 2974void 2975arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 2976{ 2977 arc_buf_t **bufp = arg; 2978 if (zio && zio->io_error) { 2979 VERIFY(arc_buf_remove_ref(buf, arg)); 2980 *bufp = NULL; 2981 } else { 2982 *bufp = buf; 2983 ASSERT(buf->b_data); 2984 } 2985} 2986 2987static void 2988arc_read_done(zio_t *zio) 2989{ 2990 arc_buf_hdr_t *hdr; 2991 arc_buf_t *buf; 2992 arc_buf_t *abuf; /* buffer we're assigning to callback */ 2993 kmutex_t *hash_lock = NULL; 2994 arc_callback_t *callback_list, *acb; 2995 int freeable = FALSE; 2996 2997 buf = zio->io_private; 2998 hdr = buf->b_hdr; 2999 3000 /* 3001 * The hdr was inserted into hash-table and removed from lists 3002 * prior to starting I/O. We should find this header, since 3003 * it's in the hash table, and it should be legit since it's 3004 * not possible to evict it during the I/O. The only possible 3005 * reason for it not to be found is if we were freed during the 3006 * read. 3007 */ 3008 if (HDR_IN_HASH_TABLE(hdr)) { 3009 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3010 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3011 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3012 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3013 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3014 3015 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3016 &hash_lock); 3017 3018 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3019 hash_lock == NULL) || 3020 (found == hdr && 3021 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3022 (found == hdr && HDR_L2_READING(hdr))); 3023 } 3024 3025 hdr->b_flags &= ~ARC_L2_EVICTED; 3026 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 3027 hdr->b_flags &= ~ARC_L2CACHE; 3028 3029 /* byteswap if necessary */ 3030 callback_list = hdr->b_acb; 3031 ASSERT(callback_list != NULL); 3032 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3033 dmu_object_byteswap_t bswap = 3034 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3035 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3036 byteswap_uint64_array : 3037 dmu_ot_byteswap[bswap].ob_func; 3038 func(buf->b_data, hdr->b_size); 3039 } 3040 3041 arc_cksum_compute(buf, B_FALSE); 3042#ifdef illumos 3043 arc_buf_watch(buf); 3044#endif /* illumos */ 3045 3046 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3047 /* 3048 * Only call arc_access on anonymous buffers. This is because 3049 * if we've issued an I/O for an evicted buffer, we've already 3050 * called arc_access (to prevent any simultaneous readers from 3051 * getting confused). 3052 */ 3053 arc_access(hdr, hash_lock); 3054 } 3055 3056 /* create copies of the data buffer for the callers */ 3057 abuf = buf; 3058 for (acb = callback_list; acb; acb = acb->acb_next) { 3059 if (acb->acb_done) { 3060 if (abuf == NULL) { 3061 ARCSTAT_BUMP(arcstat_duplicate_reads); 3062 abuf = arc_buf_clone(buf); 3063 } 3064 acb->acb_buf = abuf; 3065 abuf = NULL; 3066 } 3067 } 3068 hdr->b_acb = NULL; 3069 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3070 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3071 if (abuf == buf) { 3072 ASSERT(buf->b_efunc == NULL); 3073 ASSERT(hdr->b_datacnt == 1); 3074 hdr->b_flags |= ARC_BUF_AVAILABLE; 3075 } 3076 3077 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3078 3079 if (zio->io_error != 0) { 3080 hdr->b_flags |= ARC_IO_ERROR; 3081 if (hdr->b_state != arc_anon) 3082 arc_change_state(arc_anon, hdr, hash_lock); 3083 if (HDR_IN_HASH_TABLE(hdr)) 3084 buf_hash_remove(hdr); 3085 freeable = refcount_is_zero(&hdr->b_refcnt); 3086 } 3087 3088 /* 3089 * Broadcast before we drop the hash_lock to avoid the possibility 3090 * that the hdr (and hence the cv) might be freed before we get to 3091 * the cv_broadcast(). 3092 */ 3093 cv_broadcast(&hdr->b_cv); 3094 3095 if (hash_lock) { 3096 mutex_exit(hash_lock); 3097 } else { 3098 /* 3099 * This block was freed while we waited for the read to 3100 * complete. It has been removed from the hash table and 3101 * moved to the anonymous state (so that it won't show up 3102 * in the cache). 3103 */ 3104 ASSERT3P(hdr->b_state, ==, arc_anon); 3105 freeable = refcount_is_zero(&hdr->b_refcnt); 3106 } 3107 3108 /* execute each callback and free its structure */ 3109 while ((acb = callback_list) != NULL) { 3110 if (acb->acb_done) 3111 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3112 3113 if (acb->acb_zio_dummy != NULL) { 3114 acb->acb_zio_dummy->io_error = zio->io_error; 3115 zio_nowait(acb->acb_zio_dummy); 3116 } 3117 3118 callback_list = acb->acb_next; 3119 kmem_free(acb, sizeof (arc_callback_t)); 3120 } 3121 3122 if (freeable) 3123 arc_hdr_destroy(hdr); 3124} 3125 3126/* 3127 * "Read" the block block at the specified DVA (in bp) via the 3128 * cache. If the block is found in the cache, invoke the provided 3129 * callback immediately and return. Note that the `zio' parameter 3130 * in the callback will be NULL in this case, since no IO was 3131 * required. If the block is not in the cache pass the read request 3132 * on to the spa with a substitute callback function, so that the 3133 * requested block will be added to the cache. 3134 * 3135 * If a read request arrives for a block that has a read in-progress, 3136 * either wait for the in-progress read to complete (and return the 3137 * results); or, if this is a read with a "done" func, add a record 3138 * to the read to invoke the "done" func when the read completes, 3139 * and return; or just return. 3140 * 3141 * arc_read_done() will invoke all the requested "done" functions 3142 * for readers of this block. 3143 */ 3144int 3145arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3146 void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags, 3147 const zbookmark_t *zb) 3148{ 3149 arc_buf_hdr_t *hdr = NULL; 3150 arc_buf_t *buf = NULL; 3151 kmutex_t *hash_lock = NULL; 3152 zio_t *rzio; 3153 uint64_t guid = spa_load_guid(spa); 3154 3155 ASSERT(!BP_IS_EMBEDDED(bp) || 3156 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3157 3158top: 3159 if (!BP_IS_EMBEDDED(bp)) { 3160 /* 3161 * Embedded BP's have no DVA and require no I/O to "read". 3162 * Create an anonymous arc buf to back it. 3163 */ 3164 hdr = buf_hash_find(guid, bp, &hash_lock); 3165 } 3166 3167 if (hdr != NULL && hdr->b_datacnt > 0) { 3168 3169 *arc_flags |= ARC_CACHED; 3170 3171 if (HDR_IO_IN_PROGRESS(hdr)) { 3172 3173 if (*arc_flags & ARC_WAIT) { 3174 cv_wait(&hdr->b_cv, hash_lock); 3175 mutex_exit(hash_lock); 3176 goto top; 3177 } 3178 ASSERT(*arc_flags & ARC_NOWAIT); 3179 3180 if (done) { 3181 arc_callback_t *acb = NULL; 3182 3183 acb = kmem_zalloc(sizeof (arc_callback_t), 3184 KM_SLEEP); 3185 acb->acb_done = done; 3186 acb->acb_private = private; 3187 if (pio != NULL) 3188 acb->acb_zio_dummy = zio_null(pio, 3189 spa, NULL, NULL, NULL, zio_flags); 3190 3191 ASSERT(acb->acb_done != NULL); 3192 acb->acb_next = hdr->b_acb; 3193 hdr->b_acb = acb; 3194 add_reference(hdr, hash_lock, private); 3195 mutex_exit(hash_lock); 3196 return (0); 3197 } 3198 mutex_exit(hash_lock); 3199 return (0); 3200 } 3201 3202 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3203 3204 if (done) { 3205 add_reference(hdr, hash_lock, private); 3206 /* 3207 * If this block is already in use, create a new 3208 * copy of the data so that we will be guaranteed 3209 * that arc_release() will always succeed. 3210 */ 3211 buf = hdr->b_buf; 3212 ASSERT(buf); 3213 ASSERT(buf->b_data); 3214 if (HDR_BUF_AVAILABLE(hdr)) { 3215 ASSERT(buf->b_efunc == NULL); 3216 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3217 } else { 3218 buf = arc_buf_clone(buf); 3219 } 3220 3221 } else if (*arc_flags & ARC_PREFETCH && 3222 refcount_count(&hdr->b_refcnt) == 0) { 3223 hdr->b_flags |= ARC_PREFETCH; 3224 } 3225 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3226 arc_access(hdr, hash_lock); 3227 if (*arc_flags & ARC_L2CACHE) 3228 hdr->b_flags |= ARC_L2CACHE; 3229 if (*arc_flags & ARC_L2COMPRESS) 3230 hdr->b_flags |= ARC_L2COMPRESS; 3231 mutex_exit(hash_lock); 3232 ARCSTAT_BUMP(arcstat_hits); 3233 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3234 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3235 data, metadata, hits); 3236 3237 if (done) 3238 done(NULL, buf, private); 3239 } else { 3240 uint64_t size = BP_GET_LSIZE(bp); 3241 arc_callback_t *acb; 3242 vdev_t *vd = NULL; 3243 uint64_t addr = 0; 3244 boolean_t devw = B_FALSE; 3245 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3246 uint64_t b_asize = 0; 3247 3248 if (hdr == NULL) { 3249 /* this block is not in the cache */ 3250 arc_buf_hdr_t *exists = NULL; 3251 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3252 buf = arc_buf_alloc(spa, size, private, type); 3253 hdr = buf->b_hdr; 3254 if (!BP_IS_EMBEDDED(bp)) { 3255 hdr->b_dva = *BP_IDENTITY(bp); 3256 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3257 hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3258 exists = buf_hash_insert(hdr, &hash_lock); 3259 } 3260 if (exists != NULL) { 3261 /* somebody beat us to the hash insert */ 3262 mutex_exit(hash_lock); 3263 buf_discard_identity(hdr); 3264 (void) arc_buf_remove_ref(buf, private); 3265 goto top; /* restart the IO request */ 3266 } 3267 /* if this is a prefetch, we don't have a reference */ 3268 if (*arc_flags & ARC_PREFETCH) { 3269 (void) remove_reference(hdr, hash_lock, 3270 private); 3271 hdr->b_flags |= ARC_PREFETCH; 3272 } 3273 if (*arc_flags & ARC_L2CACHE) 3274 hdr->b_flags |= ARC_L2CACHE; 3275 if (*arc_flags & ARC_L2COMPRESS) 3276 hdr->b_flags |= ARC_L2COMPRESS; 3277 if (BP_GET_LEVEL(bp) > 0) 3278 hdr->b_flags |= ARC_INDIRECT; 3279 } else { 3280 /* this block is in the ghost cache */ 3281 ASSERT(GHOST_STATE(hdr->b_state)); 3282 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3283 ASSERT0(refcount_count(&hdr->b_refcnt)); 3284 ASSERT(hdr->b_buf == NULL); 3285 3286 /* if this is a prefetch, we don't have a reference */ 3287 if (*arc_flags & ARC_PREFETCH) 3288 hdr->b_flags |= ARC_PREFETCH; 3289 else 3290 add_reference(hdr, hash_lock, private); 3291 if (*arc_flags & ARC_L2CACHE) 3292 hdr->b_flags |= ARC_L2CACHE; 3293 if (*arc_flags & ARC_L2COMPRESS) 3294 hdr->b_flags |= ARC_L2COMPRESS; 3295 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3296 buf->b_hdr = hdr; 3297 buf->b_data = NULL; 3298 buf->b_efunc = NULL; 3299 buf->b_private = NULL; 3300 buf->b_next = NULL; 3301 hdr->b_buf = buf; 3302 ASSERT(hdr->b_datacnt == 0); 3303 hdr->b_datacnt = 1; 3304 arc_get_data_buf(buf); 3305 arc_access(hdr, hash_lock); 3306 } 3307 3308 ASSERT(!GHOST_STATE(hdr->b_state)); 3309 3310 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3311 acb->acb_done = done; 3312 acb->acb_private = private; 3313 3314 ASSERT(hdr->b_acb == NULL); 3315 hdr->b_acb = acb; 3316 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3317 3318 if (hdr->b_l2hdr != NULL && 3319 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3320 devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3321 addr = hdr->b_l2hdr->b_daddr; 3322 b_compress = hdr->b_l2hdr->b_compress; 3323 b_asize = hdr->b_l2hdr->b_asize; 3324 /* 3325 * Lock out device removal. 3326 */ 3327 if (vdev_is_dead(vd) || 3328 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3329 vd = NULL; 3330 } 3331 3332 if (hash_lock != NULL) 3333 mutex_exit(hash_lock); 3334 3335 /* 3336 * At this point, we have a level 1 cache miss. Try again in 3337 * L2ARC if possible. 3338 */ 3339 ASSERT3U(hdr->b_size, ==, size); 3340 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3341 uint64_t, size, zbookmark_t *, zb); 3342 ARCSTAT_BUMP(arcstat_misses); 3343 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3344 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3345 data, metadata, misses); 3346#ifdef _KERNEL 3347 curthread->td_ru.ru_inblock++; 3348#endif 3349 3350 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3351 /* 3352 * Read from the L2ARC if the following are true: 3353 * 1. The L2ARC vdev was previously cached. 3354 * 2. This buffer still has L2ARC metadata. 3355 * 3. This buffer isn't currently writing to the L2ARC. 3356 * 4. The L2ARC entry wasn't evicted, which may 3357 * also have invalidated the vdev. 3358 * 5. This isn't prefetch and l2arc_noprefetch is set. 3359 */ 3360 if (hdr->b_l2hdr != NULL && 3361 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3362 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3363 l2arc_read_callback_t *cb; 3364 3365 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3366 ARCSTAT_BUMP(arcstat_l2_hits); 3367 3368 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3369 KM_SLEEP); 3370 cb->l2rcb_buf = buf; 3371 cb->l2rcb_spa = spa; 3372 cb->l2rcb_bp = *bp; 3373 cb->l2rcb_zb = *zb; 3374 cb->l2rcb_flags = zio_flags; 3375 cb->l2rcb_compress = b_compress; 3376 3377 ASSERT(addr >= VDEV_LABEL_START_SIZE && 3378 addr + size < vd->vdev_psize - 3379 VDEV_LABEL_END_SIZE); 3380 3381 /* 3382 * l2arc read. The SCL_L2ARC lock will be 3383 * released by l2arc_read_done(). 3384 * Issue a null zio if the underlying buffer 3385 * was squashed to zero size by compression. 3386 */ 3387 if (b_compress == ZIO_COMPRESS_EMPTY) { 3388 rzio = zio_null(pio, spa, vd, 3389 l2arc_read_done, cb, 3390 zio_flags | ZIO_FLAG_DONT_CACHE | 3391 ZIO_FLAG_CANFAIL | 3392 ZIO_FLAG_DONT_PROPAGATE | 3393 ZIO_FLAG_DONT_RETRY); 3394 } else { 3395 rzio = zio_read_phys(pio, vd, addr, 3396 b_asize, buf->b_data, 3397 ZIO_CHECKSUM_OFF, 3398 l2arc_read_done, cb, priority, 3399 zio_flags | ZIO_FLAG_DONT_CACHE | 3400 ZIO_FLAG_CANFAIL | 3401 ZIO_FLAG_DONT_PROPAGATE | 3402 ZIO_FLAG_DONT_RETRY, B_FALSE); 3403 } 3404 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3405 zio_t *, rzio); 3406 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3407 3408 if (*arc_flags & ARC_NOWAIT) { 3409 zio_nowait(rzio); 3410 return (0); 3411 } 3412 3413 ASSERT(*arc_flags & ARC_WAIT); 3414 if (zio_wait(rzio) == 0) 3415 return (0); 3416 3417 /* l2arc read error; goto zio_read() */ 3418 } else { 3419 DTRACE_PROBE1(l2arc__miss, 3420 arc_buf_hdr_t *, hdr); 3421 ARCSTAT_BUMP(arcstat_l2_misses); 3422 if (HDR_L2_WRITING(hdr)) 3423 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3424 spa_config_exit(spa, SCL_L2ARC, vd); 3425 } 3426 } else { 3427 if (vd != NULL) 3428 spa_config_exit(spa, SCL_L2ARC, vd); 3429 if (l2arc_ndev != 0) { 3430 DTRACE_PROBE1(l2arc__miss, 3431 arc_buf_hdr_t *, hdr); 3432 ARCSTAT_BUMP(arcstat_l2_misses); 3433 } 3434 } 3435 3436 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3437 arc_read_done, buf, priority, zio_flags, zb); 3438 3439 if (*arc_flags & ARC_WAIT) 3440 return (zio_wait(rzio)); 3441 3442 ASSERT(*arc_flags & ARC_NOWAIT); 3443 zio_nowait(rzio); 3444 } 3445 return (0); 3446} 3447 3448void 3449arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3450{ 3451 ASSERT(buf->b_hdr != NULL); 3452 ASSERT(buf->b_hdr->b_state != arc_anon); 3453 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3454 ASSERT(buf->b_efunc == NULL); 3455 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3456 3457 buf->b_efunc = func; 3458 buf->b_private = private; 3459} 3460 3461/* 3462 * Notify the arc that a block was freed, and thus will never be used again. 3463 */ 3464void 3465arc_freed(spa_t *spa, const blkptr_t *bp) 3466{ 3467 arc_buf_hdr_t *hdr; 3468 kmutex_t *hash_lock; 3469 uint64_t guid = spa_load_guid(spa); 3470 3471 ASSERT(!BP_IS_EMBEDDED(bp)); 3472 3473 hdr = buf_hash_find(guid, bp, &hash_lock); 3474 if (hdr == NULL) 3475 return; 3476 if (HDR_BUF_AVAILABLE(hdr)) { 3477 arc_buf_t *buf = hdr->b_buf; 3478 add_reference(hdr, hash_lock, FTAG); 3479 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3480 mutex_exit(hash_lock); 3481 3482 arc_release(buf, FTAG); 3483 (void) arc_buf_remove_ref(buf, FTAG); 3484 } else { 3485 mutex_exit(hash_lock); 3486 } 3487 3488} 3489 3490/* 3491 * This is used by the DMU to let the ARC know that a buffer is 3492 * being evicted, so the ARC should clean up. If this arc buf 3493 * is not yet in the evicted state, it will be put there. 3494 */ 3495int 3496arc_buf_evict(arc_buf_t *buf) 3497{ 3498 arc_buf_hdr_t *hdr; 3499 kmutex_t *hash_lock; 3500 arc_buf_t **bufp; 3501 list_t *list, *evicted_list; 3502 kmutex_t *lock, *evicted_lock; 3503 3504 mutex_enter(&buf->b_evict_lock); 3505 hdr = buf->b_hdr; 3506 if (hdr == NULL) { 3507 /* 3508 * We are in arc_do_user_evicts(). 3509 */ 3510 ASSERT(buf->b_data == NULL); 3511 mutex_exit(&buf->b_evict_lock); 3512 return (0); 3513 } else if (buf->b_data == NULL) { 3514 arc_buf_t copy = *buf; /* structure assignment */ 3515 /* 3516 * We are on the eviction list; process this buffer now 3517 * but let arc_do_user_evicts() do the reaping. 3518 */ 3519 buf->b_efunc = NULL; 3520 mutex_exit(&buf->b_evict_lock); 3521 VERIFY(copy.b_efunc(©) == 0); 3522 return (1); 3523 } 3524 hash_lock = HDR_LOCK(hdr); 3525 mutex_enter(hash_lock); 3526 hdr = buf->b_hdr; 3527 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3528 3529 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3530 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3531 3532 /* 3533 * Pull this buffer off of the hdr 3534 */ 3535 bufp = &hdr->b_buf; 3536 while (*bufp != buf) 3537 bufp = &(*bufp)->b_next; 3538 *bufp = buf->b_next; 3539 3540 ASSERT(buf->b_data != NULL); 3541 arc_buf_destroy(buf, FALSE, FALSE); 3542 3543 if (hdr->b_datacnt == 0) { 3544 arc_state_t *old_state = hdr->b_state; 3545 arc_state_t *evicted_state; 3546 3547 ASSERT(hdr->b_buf == NULL); 3548 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 3549 3550 evicted_state = 3551 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3552 3553 get_buf_info(hdr, old_state, &list, &lock); 3554 get_buf_info(hdr, evicted_state, &evicted_list, &evicted_lock); 3555 mutex_enter(lock); 3556 mutex_enter(evicted_lock); 3557 3558 arc_change_state(evicted_state, hdr, hash_lock); 3559 ASSERT(HDR_IN_HASH_TABLE(hdr)); 3560 hdr->b_flags |= ARC_IN_HASH_TABLE; 3561 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3562 3563 mutex_exit(evicted_lock); 3564 mutex_exit(lock); 3565 } 3566 mutex_exit(hash_lock); 3567 mutex_exit(&buf->b_evict_lock); 3568 3569 VERIFY(buf->b_efunc(buf) == 0); 3570 buf->b_efunc = NULL; 3571 buf->b_private = NULL; 3572 buf->b_hdr = NULL; 3573 buf->b_next = NULL; 3574 kmem_cache_free(buf_cache, buf); 3575 return (1); 3576} 3577 3578/* 3579 * Release this buffer from the cache, making it an anonymous buffer. This 3580 * must be done after a read and prior to modifying the buffer contents. 3581 * If the buffer has more than one reference, we must make 3582 * a new hdr for the buffer. 3583 */ 3584void 3585arc_release(arc_buf_t *buf, void *tag) 3586{ 3587 arc_buf_hdr_t *hdr; 3588 kmutex_t *hash_lock = NULL; 3589 l2arc_buf_hdr_t *l2hdr; 3590 uint64_t buf_size; 3591 3592 /* 3593 * It would be nice to assert that if it's DMU metadata (level > 3594 * 0 || it's the dnode file), then it must be syncing context. 3595 * But we don't know that information at this level. 3596 */ 3597 3598 mutex_enter(&buf->b_evict_lock); 3599 hdr = buf->b_hdr; 3600 3601 /* this buffer is not on any list */ 3602 ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3603 3604 if (hdr->b_state == arc_anon) { 3605 /* this buffer is already released */ 3606 ASSERT(buf->b_efunc == NULL); 3607 } else { 3608 hash_lock = HDR_LOCK(hdr); 3609 mutex_enter(hash_lock); 3610 hdr = buf->b_hdr; 3611 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3612 } 3613 3614 l2hdr = hdr->b_l2hdr; 3615 if (l2hdr) { 3616 mutex_enter(&l2arc_buflist_mtx); 3617 hdr->b_l2hdr = NULL; 3618 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3619 } 3620 buf_size = hdr->b_size; 3621 3622 /* 3623 * Do we have more than one buf? 3624 */ 3625 if (hdr->b_datacnt > 1) { 3626 arc_buf_hdr_t *nhdr; 3627 arc_buf_t **bufp; 3628 uint64_t blksz = hdr->b_size; 3629 uint64_t spa = hdr->b_spa; 3630 arc_buf_contents_t type = hdr->b_type; 3631 uint32_t flags = hdr->b_flags; 3632 3633 ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3634 /* 3635 * Pull the data off of this hdr and attach it to 3636 * a new anonymous hdr. 3637 */ 3638 (void) remove_reference(hdr, hash_lock, tag); 3639 bufp = &hdr->b_buf; 3640 while (*bufp != buf) 3641 bufp = &(*bufp)->b_next; 3642 *bufp = buf->b_next; 3643 buf->b_next = NULL; 3644 3645 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3646 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3647 if (refcount_is_zero(&hdr->b_refcnt)) { 3648 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3649 ASSERT3U(*size, >=, hdr->b_size); 3650 atomic_add_64(size, -hdr->b_size); 3651 } 3652 3653 /* 3654 * We're releasing a duplicate user data buffer, update 3655 * our statistics accordingly. 3656 */ 3657 if (hdr->b_type == ARC_BUFC_DATA) { 3658 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3659 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3660 -hdr->b_size); 3661 } 3662 hdr->b_datacnt -= 1; 3663 arc_cksum_verify(buf); 3664#ifdef illumos 3665 arc_buf_unwatch(buf); 3666#endif /* illumos */ 3667 3668 mutex_exit(hash_lock); 3669 3670 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3671 nhdr->b_size = blksz; 3672 nhdr->b_spa = spa; 3673 nhdr->b_type = type; 3674 nhdr->b_buf = buf; 3675 nhdr->b_state = arc_anon; 3676 nhdr->b_arc_access = 0; 3677 nhdr->b_flags = flags & ARC_L2_WRITING; 3678 nhdr->b_l2hdr = NULL; 3679 nhdr->b_datacnt = 1; 3680 nhdr->b_freeze_cksum = NULL; 3681 (void) refcount_add(&nhdr->b_refcnt, tag); 3682 buf->b_hdr = nhdr; 3683 mutex_exit(&buf->b_evict_lock); 3684 atomic_add_64(&arc_anon->arcs_size, blksz); 3685 } else { 3686 mutex_exit(&buf->b_evict_lock); 3687 ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3688 ASSERT(!list_link_active(&hdr->b_arc_node)); 3689 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3690 if (hdr->b_state != arc_anon) 3691 arc_change_state(arc_anon, hdr, hash_lock); 3692 hdr->b_arc_access = 0; 3693 if (hash_lock) 3694 mutex_exit(hash_lock); 3695 3696 buf_discard_identity(hdr); 3697 arc_buf_thaw(buf); 3698 } 3699 buf->b_efunc = NULL; 3700 buf->b_private = NULL; 3701 3702 if (l2hdr) { 3703 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3704 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3705 hdr->b_size, 0); 3706 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3707 ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3708 mutex_exit(&l2arc_buflist_mtx); 3709 } 3710} 3711 3712int 3713arc_released(arc_buf_t *buf) 3714{ 3715 int released; 3716 3717 mutex_enter(&buf->b_evict_lock); 3718 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3719 mutex_exit(&buf->b_evict_lock); 3720 return (released); 3721} 3722 3723int 3724arc_has_callback(arc_buf_t *buf) 3725{ 3726 int callback; 3727 3728 mutex_enter(&buf->b_evict_lock); 3729 callback = (buf->b_efunc != NULL); 3730 mutex_exit(&buf->b_evict_lock); 3731 return (callback); 3732} 3733 3734#ifdef ZFS_DEBUG 3735int 3736arc_referenced(arc_buf_t *buf) 3737{ 3738 int referenced; 3739 3740 mutex_enter(&buf->b_evict_lock); 3741 referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3742 mutex_exit(&buf->b_evict_lock); 3743 return (referenced); 3744} 3745#endif 3746 3747static void 3748arc_write_ready(zio_t *zio) 3749{ 3750 arc_write_callback_t *callback = zio->io_private; 3751 arc_buf_t *buf = callback->awcb_buf; 3752 arc_buf_hdr_t *hdr = buf->b_hdr; 3753 3754 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3755 callback->awcb_ready(zio, buf, callback->awcb_private); 3756 3757 /* 3758 * If the IO is already in progress, then this is a re-write 3759 * attempt, so we need to thaw and re-compute the cksum. 3760 * It is the responsibility of the callback to handle the 3761 * accounting for any re-write attempt. 3762 */ 3763 if (HDR_IO_IN_PROGRESS(hdr)) { 3764 mutex_enter(&hdr->b_freeze_lock); 3765 if (hdr->b_freeze_cksum != NULL) { 3766 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3767 hdr->b_freeze_cksum = NULL; 3768 } 3769 mutex_exit(&hdr->b_freeze_lock); 3770 } 3771 arc_cksum_compute(buf, B_FALSE); 3772 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3773} 3774 3775/* 3776 * The SPA calls this callback for each physical write that happens on behalf 3777 * of a logical write. See the comment in dbuf_write_physdone() for details. 3778 */ 3779static void 3780arc_write_physdone(zio_t *zio) 3781{ 3782 arc_write_callback_t *cb = zio->io_private; 3783 if (cb->awcb_physdone != NULL) 3784 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3785} 3786 3787static void 3788arc_write_done(zio_t *zio) 3789{ 3790 arc_write_callback_t *callback = zio->io_private; 3791 arc_buf_t *buf = callback->awcb_buf; 3792 arc_buf_hdr_t *hdr = buf->b_hdr; 3793 3794 ASSERT(hdr->b_acb == NULL); 3795 3796 if (zio->io_error == 0) { 3797 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 3798 buf_discard_identity(hdr); 3799 } else { 3800 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3801 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3802 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3803 } 3804 } else { 3805 ASSERT(BUF_EMPTY(hdr)); 3806 } 3807 3808 /* 3809 * If the block to be written was all-zero or compressed enough to be 3810 * embedded in the BP, no write was performed so there will be no 3811 * dva/birth/checksum. The buffer must therefore remain anonymous 3812 * (and uncached). 3813 */ 3814 if (!BUF_EMPTY(hdr)) { 3815 arc_buf_hdr_t *exists; 3816 kmutex_t *hash_lock; 3817 3818 ASSERT(zio->io_error == 0); 3819 3820 arc_cksum_verify(buf); 3821 3822 exists = buf_hash_insert(hdr, &hash_lock); 3823 if (exists) { 3824 /* 3825 * This can only happen if we overwrite for 3826 * sync-to-convergence, because we remove 3827 * buffers from the hash table when we arc_free(). 3828 */ 3829 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3830 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3831 panic("bad overwrite, hdr=%p exists=%p", 3832 (void *)hdr, (void *)exists); 3833 ASSERT(refcount_is_zero(&exists->b_refcnt)); 3834 arc_change_state(arc_anon, exists, hash_lock); 3835 mutex_exit(hash_lock); 3836 arc_hdr_destroy(exists); 3837 exists = buf_hash_insert(hdr, &hash_lock); 3838 ASSERT3P(exists, ==, NULL); 3839 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3840 /* nopwrite */ 3841 ASSERT(zio->io_prop.zp_nopwrite); 3842 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3843 panic("bad nopwrite, hdr=%p exists=%p", 3844 (void *)hdr, (void *)exists); 3845 } else { 3846 /* Dedup */ 3847 ASSERT(hdr->b_datacnt == 1); 3848 ASSERT(hdr->b_state == arc_anon); 3849 ASSERT(BP_GET_DEDUP(zio->io_bp)); 3850 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3851 } 3852 } 3853 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3854 /* if it's not anon, we are doing a scrub */ 3855 if (!exists && hdr->b_state == arc_anon) 3856 arc_access(hdr, hash_lock); 3857 mutex_exit(hash_lock); 3858 } else { 3859 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3860 } 3861 3862 ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3863 callback->awcb_done(zio, buf, callback->awcb_private); 3864 3865 kmem_free(callback, sizeof (arc_write_callback_t)); 3866} 3867 3868zio_t * 3869arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 3870 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 3871 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 3872 arc_done_func_t *done, void *private, zio_priority_t priority, 3873 int zio_flags, const zbookmark_t *zb) 3874{ 3875 arc_buf_hdr_t *hdr = buf->b_hdr; 3876 arc_write_callback_t *callback; 3877 zio_t *zio; 3878 3879 ASSERT(ready != NULL); 3880 ASSERT(done != NULL); 3881 ASSERT(!HDR_IO_ERROR(hdr)); 3882 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3883 ASSERT(hdr->b_acb == NULL); 3884 if (l2arc) 3885 hdr->b_flags |= ARC_L2CACHE; 3886 if (l2arc_compress) 3887 hdr->b_flags |= ARC_L2COMPRESS; 3888 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 3889 callback->awcb_ready = ready; 3890 callback->awcb_physdone = physdone; 3891 callback->awcb_done = done; 3892 callback->awcb_private = private; 3893 callback->awcb_buf = buf; 3894 3895 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 3896 arc_write_ready, arc_write_physdone, arc_write_done, callback, 3897 priority, zio_flags, zb); 3898 3899 return (zio); 3900} 3901 3902static int 3903arc_memory_throttle(uint64_t reserve, uint64_t txg) 3904{ 3905#ifdef _KERNEL 3906 uint64_t available_memory = 3907 ptoa((uintmax_t)cnt.v_free_count + cnt.v_cache_count); 3908 static uint64_t page_load = 0; 3909 static uint64_t last_txg = 0; 3910 3911#ifdef sun 3912#if defined(__i386) 3913 available_memory = 3914 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 3915#endif 3916#endif /* sun */ 3917 3918 if (cnt.v_free_count + cnt.v_cache_count > 3919 (uint64_t)physmem * arc_lotsfree_percent / 100) 3920 return (0); 3921 3922 if (txg > last_txg) { 3923 last_txg = txg; 3924 page_load = 0; 3925 } 3926 /* 3927 * If we are in pageout, we know that memory is already tight, 3928 * the arc is already going to be evicting, so we just want to 3929 * continue to let page writes occur as quickly as possible. 3930 */ 3931 if (curproc == pageproc) { 3932 if (page_load > available_memory / 4) 3933 return (SET_ERROR(ERESTART)); 3934 /* Note: reserve is inflated, so we deflate */ 3935 page_load += reserve / 8; 3936 return (0); 3937 } else if (page_load > 0 && arc_reclaim_needed()) { 3938 /* memory is low, delay before restarting */ 3939 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3940 return (SET_ERROR(EAGAIN)); 3941 } 3942 page_load = 0; 3943#endif 3944 return (0); 3945} 3946 3947void 3948arc_tempreserve_clear(uint64_t reserve) 3949{ 3950 atomic_add_64(&arc_tempreserve, -reserve); 3951 ASSERT((int64_t)arc_tempreserve >= 0); 3952} 3953 3954int 3955arc_tempreserve_space(uint64_t reserve, uint64_t txg) 3956{ 3957 int error; 3958 uint64_t anon_size; 3959 3960 if (reserve > arc_c/4 && !arc_no_grow) 3961 arc_c = MIN(arc_c_max, reserve * 4); 3962 if (reserve > arc_c) 3963 return (SET_ERROR(ENOMEM)); 3964 3965 /* 3966 * Don't count loaned bufs as in flight dirty data to prevent long 3967 * network delays from blocking transactions that are ready to be 3968 * assigned to a txg. 3969 */ 3970 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 3971 3972 /* 3973 * Writes will, almost always, require additional memory allocations 3974 * in order to compress/encrypt/etc the data. We therefore need to 3975 * make sure that there is sufficient available memory for this. 3976 */ 3977 error = arc_memory_throttle(reserve, txg); 3978 if (error != 0) 3979 return (error); 3980 3981 /* 3982 * Throttle writes when the amount of dirty data in the cache 3983 * gets too large. We try to keep the cache less than half full 3984 * of dirty blocks so that our sync times don't grow too large. 3985 * Note: if two requests come in concurrently, we might let them 3986 * both succeed, when one of them should fail. Not a huge deal. 3987 */ 3988 3989 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 3990 anon_size > arc_c / 4) { 3991 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 3992 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 3993 arc_tempreserve>>10, 3994 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 3995 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 3996 reserve>>10, arc_c>>10); 3997 return (SET_ERROR(ERESTART)); 3998 } 3999 atomic_add_64(&arc_tempreserve, reserve); 4000 return (0); 4001} 4002 4003static kmutex_t arc_lowmem_lock; 4004#ifdef _KERNEL 4005static eventhandler_tag arc_event_lowmem = NULL; 4006 4007static void 4008arc_lowmem(void *arg __unused, int howto __unused) 4009{ 4010 4011 /* Serialize access via arc_lowmem_lock. */ 4012 mutex_enter(&arc_lowmem_lock); 4013 mutex_enter(&arc_reclaim_thr_lock); 4014 needfree = 1; 4015 cv_signal(&arc_reclaim_thr_cv); 4016 4017 /* 4018 * It is unsafe to block here in arbitrary threads, because we can come 4019 * here from ARC itself and may hold ARC locks and thus risk a deadlock 4020 * with ARC reclaim thread. 4021 */ 4022 if (curproc == pageproc) { 4023 while (needfree) 4024 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4025 } 4026 mutex_exit(&arc_reclaim_thr_lock); 4027 mutex_exit(&arc_lowmem_lock); 4028} 4029#endif 4030 4031void 4032arc_init(void) 4033{ 4034 int i, prefetch_tunable_set = 0; 4035 4036 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4037 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4038 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4039 4040 /* Convert seconds to clock ticks */ 4041 arc_min_prefetch_lifespan = 1 * hz; 4042 4043 /* Start out with 1/8 of all memory */ 4044 arc_c = kmem_size() / 8; 4045 4046#ifdef sun 4047#ifdef _KERNEL 4048 /* 4049 * On architectures where the physical memory can be larger 4050 * than the addressable space (intel in 32-bit mode), we may 4051 * need to limit the cache to 1/8 of VM size. 4052 */ 4053 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4054#endif 4055#endif /* sun */ 4056 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4057 arc_c_min = MAX(arc_c / 4, 64<<18); 4058 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4059 if (arc_c * 8 >= 1<<30) 4060 arc_c_max = (arc_c * 8) - (1<<30); 4061 else 4062 arc_c_max = arc_c_min; 4063 arc_c_max = MAX(arc_c * 5, arc_c_max); 4064 4065#ifdef _KERNEL 4066 /* 4067 * Allow the tunables to override our calculations if they are 4068 * reasonable (ie. over 16MB) 4069 */ 4070 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size()) 4071 arc_c_max = zfs_arc_max; 4072 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max) 4073 arc_c_min = zfs_arc_min; 4074#endif 4075 4076 arc_c = arc_c_max; 4077 arc_p = (arc_c >> 1); 4078 4079 /* limit meta-data to 1/4 of the arc capacity */ 4080 arc_meta_limit = arc_c_max / 4; 4081 4082 /* Allow the tunable to override if it is reasonable */ 4083 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4084 arc_meta_limit = zfs_arc_meta_limit; 4085 4086 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4087 arc_c_min = arc_meta_limit / 2; 4088 4089 if (zfs_arc_grow_retry > 0) 4090 arc_grow_retry = zfs_arc_grow_retry; 4091 4092 if (zfs_arc_shrink_shift > 0) 4093 arc_shrink_shift = zfs_arc_shrink_shift; 4094 4095 if (zfs_arc_p_min_shift > 0) 4096 arc_p_min_shift = zfs_arc_p_min_shift; 4097 4098 /* if kmem_flags are set, lets try to use less memory */ 4099 if (kmem_debugging()) 4100 arc_c = arc_c / 2; 4101 if (arc_c < arc_c_min) 4102 arc_c = arc_c_min; 4103 4104 zfs_arc_min = arc_c_min; 4105 zfs_arc_max = arc_c_max; 4106 4107 arc_anon = &ARC_anon; 4108 arc_mru = &ARC_mru; 4109 arc_mru_ghost = &ARC_mru_ghost; 4110 arc_mfu = &ARC_mfu; 4111 arc_mfu_ghost = &ARC_mfu_ghost; 4112 arc_l2c_only = &ARC_l2c_only; 4113 arc_size = 0; 4114 4115 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4116 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4117 NULL, MUTEX_DEFAULT, NULL); 4118 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4119 NULL, MUTEX_DEFAULT, NULL); 4120 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4121 NULL, MUTEX_DEFAULT, NULL); 4122 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4123 NULL, MUTEX_DEFAULT, NULL); 4124 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4125 NULL, MUTEX_DEFAULT, NULL); 4126 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4127 NULL, MUTEX_DEFAULT, NULL); 4128 4129 list_create(&arc_mru->arcs_lists[i], 4130 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4131 list_create(&arc_mru_ghost->arcs_lists[i], 4132 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4133 list_create(&arc_mfu->arcs_lists[i], 4134 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4135 list_create(&arc_mfu_ghost->arcs_lists[i], 4136 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4137 list_create(&arc_mfu_ghost->arcs_lists[i], 4138 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4139 list_create(&arc_l2c_only->arcs_lists[i], 4140 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4141 } 4142 4143 buf_init(); 4144 4145 arc_thread_exit = 0; 4146 arc_eviction_list = NULL; 4147 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4148 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4149 4150 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4151 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4152 4153 if (arc_ksp != NULL) { 4154 arc_ksp->ks_data = &arc_stats; 4155 kstat_install(arc_ksp); 4156 } 4157 4158 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4159 TS_RUN, minclsyspri); 4160 4161#ifdef _KERNEL 4162 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4163 EVENTHANDLER_PRI_FIRST); 4164#endif 4165 4166 arc_dead = FALSE; 4167 arc_warm = B_FALSE; 4168 4169 /* 4170 * Calculate maximum amount of dirty data per pool. 4171 * 4172 * If it has been set by /etc/system, take that. 4173 * Otherwise, use a percentage of physical memory defined by 4174 * zfs_dirty_data_max_percent (default 10%) with a cap at 4175 * zfs_dirty_data_max_max (default 4GB). 4176 */ 4177 if (zfs_dirty_data_max == 0) { 4178 zfs_dirty_data_max = ptob(physmem) * 4179 zfs_dirty_data_max_percent / 100; 4180 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4181 zfs_dirty_data_max_max); 4182 } 4183 4184#ifdef _KERNEL 4185 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4186 prefetch_tunable_set = 1; 4187 4188#ifdef __i386__ 4189 if (prefetch_tunable_set == 0) { 4190 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4191 "-- to enable,\n"); 4192 printf(" add \"vfs.zfs.prefetch_disable=0\" " 4193 "to /boot/loader.conf.\n"); 4194 zfs_prefetch_disable = 1; 4195 } 4196#else 4197 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4198 prefetch_tunable_set == 0) { 4199 printf("ZFS NOTICE: Prefetch is disabled by default if less " 4200 "than 4GB of RAM is present;\n" 4201 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4202 "to /boot/loader.conf.\n"); 4203 zfs_prefetch_disable = 1; 4204 } 4205#endif 4206 /* Warn about ZFS memory and address space requirements. */ 4207 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4208 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4209 "expect unstable behavior.\n"); 4210 } 4211 if (kmem_size() < 512 * (1 << 20)) { 4212 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4213 "expect unstable behavior.\n"); 4214 printf(" Consider tuning vm.kmem_size and " 4215 "vm.kmem_size_max\n"); 4216 printf(" in /boot/loader.conf.\n"); 4217 } 4218#endif 4219} 4220 4221void 4222arc_fini(void) 4223{ 4224 int i; 4225 4226 mutex_enter(&arc_reclaim_thr_lock); 4227 arc_thread_exit = 1; 4228 cv_signal(&arc_reclaim_thr_cv); 4229 while (arc_thread_exit != 0) 4230 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4231 mutex_exit(&arc_reclaim_thr_lock); 4232 4233 arc_flush(NULL); 4234 4235 arc_dead = TRUE; 4236 4237 if (arc_ksp != NULL) { 4238 kstat_delete(arc_ksp); 4239 arc_ksp = NULL; 4240 } 4241 4242 mutex_destroy(&arc_eviction_mtx); 4243 mutex_destroy(&arc_reclaim_thr_lock); 4244 cv_destroy(&arc_reclaim_thr_cv); 4245 4246 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4247 list_destroy(&arc_mru->arcs_lists[i]); 4248 list_destroy(&arc_mru_ghost->arcs_lists[i]); 4249 list_destroy(&arc_mfu->arcs_lists[i]); 4250 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4251 list_destroy(&arc_l2c_only->arcs_lists[i]); 4252 4253 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4254 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4255 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4256 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4257 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4258 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4259 } 4260 4261 buf_fini(); 4262 4263 ASSERT(arc_loaned_bytes == 0); 4264 4265 mutex_destroy(&arc_lowmem_lock); 4266#ifdef _KERNEL 4267 if (arc_event_lowmem != NULL) 4268 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4269#endif 4270} 4271 4272/* 4273 * Level 2 ARC 4274 * 4275 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4276 * It uses dedicated storage devices to hold cached data, which are populated 4277 * using large infrequent writes. The main role of this cache is to boost 4278 * the performance of random read workloads. The intended L2ARC devices 4279 * include short-stroked disks, solid state disks, and other media with 4280 * substantially faster read latency than disk. 4281 * 4282 * +-----------------------+ 4283 * | ARC | 4284 * +-----------------------+ 4285 * | ^ ^ 4286 * | | | 4287 * l2arc_feed_thread() arc_read() 4288 * | | | 4289 * | l2arc read | 4290 * V | | 4291 * +---------------+ | 4292 * | L2ARC | | 4293 * +---------------+ | 4294 * | ^ | 4295 * l2arc_write() | | 4296 * | | | 4297 * V | | 4298 * +-------+ +-------+ 4299 * | vdev | | vdev | 4300 * | cache | | cache | 4301 * +-------+ +-------+ 4302 * +=========+ .-----. 4303 * : L2ARC : |-_____-| 4304 * : devices : | Disks | 4305 * +=========+ `-_____-' 4306 * 4307 * Read requests are satisfied from the following sources, in order: 4308 * 4309 * 1) ARC 4310 * 2) vdev cache of L2ARC devices 4311 * 3) L2ARC devices 4312 * 4) vdev cache of disks 4313 * 5) disks 4314 * 4315 * Some L2ARC device types exhibit extremely slow write performance. 4316 * To accommodate for this there are some significant differences between 4317 * the L2ARC and traditional cache design: 4318 * 4319 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4320 * the ARC behave as usual, freeing buffers and placing headers on ghost 4321 * lists. The ARC does not send buffers to the L2ARC during eviction as 4322 * this would add inflated write latencies for all ARC memory pressure. 4323 * 4324 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4325 * It does this by periodically scanning buffers from the eviction-end of 4326 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4327 * not already there. It scans until a headroom of buffers is satisfied, 4328 * which itself is a buffer for ARC eviction. If a compressible buffer is 4329 * found during scanning and selected for writing to an L2ARC device, we 4330 * temporarily boost scanning headroom during the next scan cycle to make 4331 * sure we adapt to compression effects (which might significantly reduce 4332 * the data volume we write to L2ARC). The thread that does this is 4333 * l2arc_feed_thread(), illustrated below; example sizes are included to 4334 * provide a better sense of ratio than this diagram: 4335 * 4336 * head --> tail 4337 * +---------------------+----------+ 4338 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4339 * +---------------------+----------+ | o L2ARC eligible 4340 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4341 * +---------------------+----------+ | 4342 * 15.9 Gbytes ^ 32 Mbytes | 4343 * headroom | 4344 * l2arc_feed_thread() 4345 * | 4346 * l2arc write hand <--[oooo]--' 4347 * | 8 Mbyte 4348 * | write max 4349 * V 4350 * +==============================+ 4351 * L2ARC dev |####|#|###|###| |####| ... | 4352 * +==============================+ 4353 * 32 Gbytes 4354 * 4355 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4356 * evicted, then the L2ARC has cached a buffer much sooner than it probably 4357 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4358 * safe to say that this is an uncommon case, since buffers at the end of 4359 * the ARC lists have moved there due to inactivity. 4360 * 4361 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4362 * then the L2ARC simply misses copying some buffers. This serves as a 4363 * pressure valve to prevent heavy read workloads from both stalling the ARC 4364 * with waits and clogging the L2ARC with writes. This also helps prevent 4365 * the potential for the L2ARC to churn if it attempts to cache content too 4366 * quickly, such as during backups of the entire pool. 4367 * 4368 * 5. After system boot and before the ARC has filled main memory, there are 4369 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4370 * lists can remain mostly static. Instead of searching from tail of these 4371 * lists as pictured, the l2arc_feed_thread() will search from the list heads 4372 * for eligible buffers, greatly increasing its chance of finding them. 4373 * 4374 * The L2ARC device write speed is also boosted during this time so that 4375 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4376 * there are no L2ARC reads, and no fear of degrading read performance 4377 * through increased writes. 4378 * 4379 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4380 * the vdev queue can aggregate them into larger and fewer writes. Each 4381 * device is written to in a rotor fashion, sweeping writes through 4382 * available space then repeating. 4383 * 4384 * 7. The L2ARC does not store dirty content. It never needs to flush 4385 * write buffers back to disk based storage. 4386 * 4387 * 8. If an ARC buffer is written (and dirtied) which also exists in the 4388 * L2ARC, the now stale L2ARC buffer is immediately dropped. 4389 * 4390 * The performance of the L2ARC can be tweaked by a number of tunables, which 4391 * may be necessary for different workloads: 4392 * 4393 * l2arc_write_max max write bytes per interval 4394 * l2arc_write_boost extra write bytes during device warmup 4395 * l2arc_noprefetch skip caching prefetched buffers 4396 * l2arc_headroom number of max device writes to precache 4397 * l2arc_headroom_boost when we find compressed buffers during ARC 4398 * scanning, we multiply headroom by this 4399 * percentage factor for the next scan cycle, 4400 * since more compressed buffers are likely to 4401 * be present 4402 * l2arc_feed_secs seconds between L2ARC writing 4403 * 4404 * Tunables may be removed or added as future performance improvements are 4405 * integrated, and also may become zpool properties. 4406 * 4407 * There are three key functions that control how the L2ARC warms up: 4408 * 4409 * l2arc_write_eligible() check if a buffer is eligible to cache 4410 * l2arc_write_size() calculate how much to write 4411 * l2arc_write_interval() calculate sleep delay between writes 4412 * 4413 * These three functions determine what to write, how much, and how quickly 4414 * to send writes. 4415 */ 4416 4417static boolean_t 4418l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab) 4419{ 4420 /* 4421 * A buffer is *not* eligible for the L2ARC if it: 4422 * 1. belongs to a different spa. 4423 * 2. is already cached on the L2ARC. 4424 * 3. has an I/O in progress (it may be an incomplete read). 4425 * 4. is flagged not eligible (zfs property). 4426 */ 4427 if (ab->b_spa != spa_guid) { 4428 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4429 return (B_FALSE); 4430 } 4431 if (ab->b_l2hdr != NULL) { 4432 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4433 return (B_FALSE); 4434 } 4435 if (HDR_IO_IN_PROGRESS(ab)) { 4436 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4437 return (B_FALSE); 4438 } 4439 if (!HDR_L2CACHE(ab)) { 4440 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4441 return (B_FALSE); 4442 } 4443 4444 return (B_TRUE); 4445} 4446 4447static uint64_t 4448l2arc_write_size(void) 4449{ 4450 uint64_t size; 4451 4452 /* 4453 * Make sure our globals have meaningful values in case the user 4454 * altered them. 4455 */ 4456 size = l2arc_write_max; 4457 if (size == 0) { 4458 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4459 "be greater than zero, resetting it to the default (%d)", 4460 L2ARC_WRITE_SIZE); 4461 size = l2arc_write_max = L2ARC_WRITE_SIZE; 4462 } 4463 4464 if (arc_warm == B_FALSE) 4465 size += l2arc_write_boost; 4466 4467 return (size); 4468 4469} 4470 4471static clock_t 4472l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4473{ 4474 clock_t interval, next, now; 4475 4476 /* 4477 * If the ARC lists are busy, increase our write rate; if the 4478 * lists are stale, idle back. This is achieved by checking 4479 * how much we previously wrote - if it was more than half of 4480 * what we wanted, schedule the next write much sooner. 4481 */ 4482 if (l2arc_feed_again && wrote > (wanted / 2)) 4483 interval = (hz * l2arc_feed_min_ms) / 1000; 4484 else 4485 interval = hz * l2arc_feed_secs; 4486 4487 now = ddi_get_lbolt(); 4488 next = MAX(now, MIN(now + interval, began + interval)); 4489 4490 return (next); 4491} 4492 4493static void 4494l2arc_hdr_stat_add(void) 4495{ 4496 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4497 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4498} 4499 4500static void 4501l2arc_hdr_stat_remove(void) 4502{ 4503 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4504 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4505} 4506 4507/* 4508 * Cycle through L2ARC devices. This is how L2ARC load balances. 4509 * If a device is returned, this also returns holding the spa config lock. 4510 */ 4511static l2arc_dev_t * 4512l2arc_dev_get_next(void) 4513{ 4514 l2arc_dev_t *first, *next = NULL; 4515 4516 /* 4517 * Lock out the removal of spas (spa_namespace_lock), then removal 4518 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4519 * both locks will be dropped and a spa config lock held instead. 4520 */ 4521 mutex_enter(&spa_namespace_lock); 4522 mutex_enter(&l2arc_dev_mtx); 4523 4524 /* if there are no vdevs, there is nothing to do */ 4525 if (l2arc_ndev == 0) 4526 goto out; 4527 4528 first = NULL; 4529 next = l2arc_dev_last; 4530 do { 4531 /* loop around the list looking for a non-faulted vdev */ 4532 if (next == NULL) { 4533 next = list_head(l2arc_dev_list); 4534 } else { 4535 next = list_next(l2arc_dev_list, next); 4536 if (next == NULL) 4537 next = list_head(l2arc_dev_list); 4538 } 4539 4540 /* if we have come back to the start, bail out */ 4541 if (first == NULL) 4542 first = next; 4543 else if (next == first) 4544 break; 4545 4546 } while (vdev_is_dead(next->l2ad_vdev)); 4547 4548 /* if we were unable to find any usable vdevs, return NULL */ 4549 if (vdev_is_dead(next->l2ad_vdev)) 4550 next = NULL; 4551 4552 l2arc_dev_last = next; 4553 4554out: 4555 mutex_exit(&l2arc_dev_mtx); 4556 4557 /* 4558 * Grab the config lock to prevent the 'next' device from being 4559 * removed while we are writing to it. 4560 */ 4561 if (next != NULL) 4562 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4563 mutex_exit(&spa_namespace_lock); 4564 4565 return (next); 4566} 4567 4568/* 4569 * Free buffers that were tagged for destruction. 4570 */ 4571static void 4572l2arc_do_free_on_write() 4573{ 4574 list_t *buflist; 4575 l2arc_data_free_t *df, *df_prev; 4576 4577 mutex_enter(&l2arc_free_on_write_mtx); 4578 buflist = l2arc_free_on_write; 4579 4580 for (df = list_tail(buflist); df; df = df_prev) { 4581 df_prev = list_prev(buflist, df); 4582 ASSERT(df->l2df_data != NULL); 4583 ASSERT(df->l2df_func != NULL); 4584 df->l2df_func(df->l2df_data, df->l2df_size); 4585 list_remove(buflist, df); 4586 kmem_free(df, sizeof (l2arc_data_free_t)); 4587 } 4588 4589 mutex_exit(&l2arc_free_on_write_mtx); 4590} 4591 4592/* 4593 * A write to a cache device has completed. Update all headers to allow 4594 * reads from these buffers to begin. 4595 */ 4596static void 4597l2arc_write_done(zio_t *zio) 4598{ 4599 l2arc_write_callback_t *cb; 4600 l2arc_dev_t *dev; 4601 list_t *buflist; 4602 arc_buf_hdr_t *head, *ab, *ab_prev; 4603 l2arc_buf_hdr_t *abl2; 4604 kmutex_t *hash_lock; 4605 4606 cb = zio->io_private; 4607 ASSERT(cb != NULL); 4608 dev = cb->l2wcb_dev; 4609 ASSERT(dev != NULL); 4610 head = cb->l2wcb_head; 4611 ASSERT(head != NULL); 4612 buflist = dev->l2ad_buflist; 4613 ASSERT(buflist != NULL); 4614 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4615 l2arc_write_callback_t *, cb); 4616 4617 if (zio->io_error != 0) 4618 ARCSTAT_BUMP(arcstat_l2_writes_error); 4619 4620 mutex_enter(&l2arc_buflist_mtx); 4621 4622 /* 4623 * All writes completed, or an error was hit. 4624 */ 4625 for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 4626 ab_prev = list_prev(buflist, ab); 4627 abl2 = ab->b_l2hdr; 4628 4629 /* 4630 * Release the temporary compressed buffer as soon as possible. 4631 */ 4632 if (abl2->b_compress != ZIO_COMPRESS_OFF) 4633 l2arc_release_cdata_buf(ab); 4634 4635 hash_lock = HDR_LOCK(ab); 4636 if (!mutex_tryenter(hash_lock)) { 4637 /* 4638 * This buffer misses out. It may be in a stage 4639 * of eviction. Its ARC_L2_WRITING flag will be 4640 * left set, denying reads to this buffer. 4641 */ 4642 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4643 continue; 4644 } 4645 4646 if (zio->io_error != 0) { 4647 /* 4648 * Error - drop L2ARC entry. 4649 */ 4650 list_remove(buflist, ab); 4651 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4652 ab->b_l2hdr = NULL; 4653 trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4654 ab->b_size, 0); 4655 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4656 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4657 } 4658 4659 /* 4660 * Allow ARC to begin reads to this L2ARC entry. 4661 */ 4662 ab->b_flags &= ~ARC_L2_WRITING; 4663 4664 mutex_exit(hash_lock); 4665 } 4666 4667 atomic_inc_64(&l2arc_writes_done); 4668 list_remove(buflist, head); 4669 kmem_cache_free(hdr_cache, head); 4670 mutex_exit(&l2arc_buflist_mtx); 4671 4672 l2arc_do_free_on_write(); 4673 4674 kmem_free(cb, sizeof (l2arc_write_callback_t)); 4675} 4676 4677/* 4678 * A read to a cache device completed. Validate buffer contents before 4679 * handing over to the regular ARC routines. 4680 */ 4681static void 4682l2arc_read_done(zio_t *zio) 4683{ 4684 l2arc_read_callback_t *cb; 4685 arc_buf_hdr_t *hdr; 4686 arc_buf_t *buf; 4687 kmutex_t *hash_lock; 4688 int equal; 4689 4690 ASSERT(zio->io_vd != NULL); 4691 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4692 4693 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4694 4695 cb = zio->io_private; 4696 ASSERT(cb != NULL); 4697 buf = cb->l2rcb_buf; 4698 ASSERT(buf != NULL); 4699 4700 hash_lock = HDR_LOCK(buf->b_hdr); 4701 mutex_enter(hash_lock); 4702 hdr = buf->b_hdr; 4703 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4704 4705 /* 4706 * If the buffer was compressed, decompress it first. 4707 */ 4708 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4709 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4710 ASSERT(zio->io_data != NULL); 4711 4712 /* 4713 * Check this survived the L2ARC journey. 4714 */ 4715 equal = arc_cksum_equal(buf); 4716 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4717 mutex_exit(hash_lock); 4718 zio->io_private = buf; 4719 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4720 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4721 arc_read_done(zio); 4722 } else { 4723 mutex_exit(hash_lock); 4724 /* 4725 * Buffer didn't survive caching. Increment stats and 4726 * reissue to the original storage device. 4727 */ 4728 if (zio->io_error != 0) { 4729 ARCSTAT_BUMP(arcstat_l2_io_error); 4730 } else { 4731 zio->io_error = SET_ERROR(EIO); 4732 } 4733 if (!equal) 4734 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4735 4736 /* 4737 * If there's no waiter, issue an async i/o to the primary 4738 * storage now. If there *is* a waiter, the caller must 4739 * issue the i/o in a context where it's OK to block. 4740 */ 4741 if (zio->io_waiter == NULL) { 4742 zio_t *pio = zio_unique_parent(zio); 4743 4744 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4745 4746 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4747 buf->b_data, zio->io_size, arc_read_done, buf, 4748 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4749 } 4750 } 4751 4752 kmem_free(cb, sizeof (l2arc_read_callback_t)); 4753} 4754 4755/* 4756 * This is the list priority from which the L2ARC will search for pages to 4757 * cache. This is used within loops (0..3) to cycle through lists in the 4758 * desired order. This order can have a significant effect on cache 4759 * performance. 4760 * 4761 * Currently the metadata lists are hit first, MFU then MRU, followed by 4762 * the data lists. This function returns a locked list, and also returns 4763 * the lock pointer. 4764 */ 4765static list_t * 4766l2arc_list_locked(int list_num, kmutex_t **lock) 4767{ 4768 list_t *list = NULL; 4769 int idx; 4770 4771 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4772 4773 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4774 idx = list_num; 4775 list = &arc_mfu->arcs_lists[idx]; 4776 *lock = ARCS_LOCK(arc_mfu, idx); 4777 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4778 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4779 list = &arc_mru->arcs_lists[idx]; 4780 *lock = ARCS_LOCK(arc_mru, idx); 4781 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4782 ARC_BUFC_NUMDATALISTS)) { 4783 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4784 list = &arc_mfu->arcs_lists[idx]; 4785 *lock = ARCS_LOCK(arc_mfu, idx); 4786 } else { 4787 idx = list_num - ARC_BUFC_NUMLISTS; 4788 list = &arc_mru->arcs_lists[idx]; 4789 *lock = ARCS_LOCK(arc_mru, idx); 4790 } 4791 4792 ASSERT(!(MUTEX_HELD(*lock))); 4793 mutex_enter(*lock); 4794 return (list); 4795} 4796 4797/* 4798 * Evict buffers from the device write hand to the distance specified in 4799 * bytes. This distance may span populated buffers, it may span nothing. 4800 * This is clearing a region on the L2ARC device ready for writing. 4801 * If the 'all' boolean is set, every buffer is evicted. 4802 */ 4803static void 4804l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4805{ 4806 list_t *buflist; 4807 l2arc_buf_hdr_t *abl2; 4808 arc_buf_hdr_t *ab, *ab_prev; 4809 kmutex_t *hash_lock; 4810 uint64_t taddr; 4811 4812 buflist = dev->l2ad_buflist; 4813 4814 if (buflist == NULL) 4815 return; 4816 4817 if (!all && dev->l2ad_first) { 4818 /* 4819 * This is the first sweep through the device. There is 4820 * nothing to evict. 4821 */ 4822 return; 4823 } 4824 4825 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4826 /* 4827 * When nearing the end of the device, evict to the end 4828 * before the device write hand jumps to the start. 4829 */ 4830 taddr = dev->l2ad_end; 4831 } else { 4832 taddr = dev->l2ad_hand + distance; 4833 } 4834 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4835 uint64_t, taddr, boolean_t, all); 4836 4837top: 4838 mutex_enter(&l2arc_buflist_mtx); 4839 for (ab = list_tail(buflist); ab; ab = ab_prev) { 4840 ab_prev = list_prev(buflist, ab); 4841 4842 hash_lock = HDR_LOCK(ab); 4843 if (!mutex_tryenter(hash_lock)) { 4844 /* 4845 * Missed the hash lock. Retry. 4846 */ 4847 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4848 mutex_exit(&l2arc_buflist_mtx); 4849 mutex_enter(hash_lock); 4850 mutex_exit(hash_lock); 4851 goto top; 4852 } 4853 4854 if (HDR_L2_WRITE_HEAD(ab)) { 4855 /* 4856 * We hit a write head node. Leave it for 4857 * l2arc_write_done(). 4858 */ 4859 list_remove(buflist, ab); 4860 mutex_exit(hash_lock); 4861 continue; 4862 } 4863 4864 if (!all && ab->b_l2hdr != NULL && 4865 (ab->b_l2hdr->b_daddr > taddr || 4866 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4867 /* 4868 * We've evicted to the target address, 4869 * or the end of the device. 4870 */ 4871 mutex_exit(hash_lock); 4872 break; 4873 } 4874 4875 if (HDR_FREE_IN_PROGRESS(ab)) { 4876 /* 4877 * Already on the path to destruction. 4878 */ 4879 mutex_exit(hash_lock); 4880 continue; 4881 } 4882 4883 if (ab->b_state == arc_l2c_only) { 4884 ASSERT(!HDR_L2_READING(ab)); 4885 /* 4886 * This doesn't exist in the ARC. Destroy. 4887 * arc_hdr_destroy() will call list_remove() 4888 * and decrement arcstat_l2_size. 4889 */ 4890 arc_change_state(arc_anon, ab, hash_lock); 4891 arc_hdr_destroy(ab); 4892 } else { 4893 /* 4894 * Invalidate issued or about to be issued 4895 * reads, since we may be about to write 4896 * over this location. 4897 */ 4898 if (HDR_L2_READING(ab)) { 4899 ARCSTAT_BUMP(arcstat_l2_evict_reading); 4900 ab->b_flags |= ARC_L2_EVICTED; 4901 } 4902 4903 /* 4904 * Tell ARC this no longer exists in L2ARC. 4905 */ 4906 if (ab->b_l2hdr != NULL) { 4907 abl2 = ab->b_l2hdr; 4908 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4909 ab->b_l2hdr = NULL; 4910 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4911 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4912 } 4913 list_remove(buflist, ab); 4914 4915 /* 4916 * This may have been leftover after a 4917 * failed write. 4918 */ 4919 ab->b_flags &= ~ARC_L2_WRITING; 4920 } 4921 mutex_exit(hash_lock); 4922 } 4923 mutex_exit(&l2arc_buflist_mtx); 4924 4925 vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0); 4926 dev->l2ad_evict = taddr; 4927} 4928 4929/* 4930 * Find and write ARC buffers to the L2ARC device. 4931 * 4932 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 4933 * for reading until they have completed writing. 4934 * The headroom_boost is an in-out parameter used to maintain headroom boost 4935 * state between calls to this function. 4936 * 4937 * Returns the number of bytes actually written (which may be smaller than 4938 * the delta by which the device hand has changed due to alignment). 4939 */ 4940static uint64_t 4941l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 4942 boolean_t *headroom_boost) 4943{ 4944 arc_buf_hdr_t *ab, *ab_prev, *head; 4945 list_t *list; 4946 uint64_t write_asize, write_psize, write_sz, headroom, 4947 buf_compress_minsz; 4948 void *buf_data; 4949 kmutex_t *list_lock; 4950 boolean_t full; 4951 l2arc_write_callback_t *cb; 4952 zio_t *pio, *wzio; 4953 uint64_t guid = spa_load_guid(spa); 4954 const boolean_t do_headroom_boost = *headroom_boost; 4955 int try; 4956 4957 ASSERT(dev->l2ad_vdev != NULL); 4958 4959 /* Lower the flag now, we might want to raise it again later. */ 4960 *headroom_boost = B_FALSE; 4961 4962 pio = NULL; 4963 write_sz = write_asize = write_psize = 0; 4964 full = B_FALSE; 4965 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 4966 head->b_flags |= ARC_L2_WRITE_HEAD; 4967 4968 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 4969 /* 4970 * We will want to try to compress buffers that are at least 2x the 4971 * device sector size. 4972 */ 4973 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 4974 4975 /* 4976 * Copy buffers for L2ARC writing. 4977 */ 4978 mutex_enter(&l2arc_buflist_mtx); 4979 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 4980 uint64_t passed_sz = 0; 4981 4982 list = l2arc_list_locked(try, &list_lock); 4983 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 4984 4985 /* 4986 * L2ARC fast warmup. 4987 * 4988 * Until the ARC is warm and starts to evict, read from the 4989 * head of the ARC lists rather than the tail. 4990 */ 4991 if (arc_warm == B_FALSE) 4992 ab = list_head(list); 4993 else 4994 ab = list_tail(list); 4995 if (ab == NULL) 4996 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 4997 4998 headroom = target_sz * l2arc_headroom; 4999 if (do_headroom_boost) 5000 headroom = (headroom * l2arc_headroom_boost) / 100; 5001 5002 for (; ab; ab = ab_prev) { 5003 l2arc_buf_hdr_t *l2hdr; 5004 kmutex_t *hash_lock; 5005 uint64_t buf_sz; 5006 5007 if (arc_warm == B_FALSE) 5008 ab_prev = list_next(list, ab); 5009 else 5010 ab_prev = list_prev(list, ab); 5011 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size); 5012 5013 hash_lock = HDR_LOCK(ab); 5014 if (!mutex_tryenter(hash_lock)) { 5015 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5016 /* 5017 * Skip this buffer rather than waiting. 5018 */ 5019 continue; 5020 } 5021 5022 passed_sz += ab->b_size; 5023 if (passed_sz > headroom) { 5024 /* 5025 * Searched too far. 5026 */ 5027 mutex_exit(hash_lock); 5028 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5029 break; 5030 } 5031 5032 if (!l2arc_write_eligible(guid, ab)) { 5033 mutex_exit(hash_lock); 5034 continue; 5035 } 5036 5037 if ((write_sz + ab->b_size) > target_sz) { 5038 full = B_TRUE; 5039 mutex_exit(hash_lock); 5040 ARCSTAT_BUMP(arcstat_l2_write_full); 5041 break; 5042 } 5043 5044 if (pio == NULL) { 5045 /* 5046 * Insert a dummy header on the buflist so 5047 * l2arc_write_done() can find where the 5048 * write buffers begin without searching. 5049 */ 5050 list_insert_head(dev->l2ad_buflist, head); 5051 5052 cb = kmem_alloc( 5053 sizeof (l2arc_write_callback_t), KM_SLEEP); 5054 cb->l2wcb_dev = dev; 5055 cb->l2wcb_head = head; 5056 pio = zio_root(spa, l2arc_write_done, cb, 5057 ZIO_FLAG_CANFAIL); 5058 ARCSTAT_BUMP(arcstat_l2_write_pios); 5059 } 5060 5061 /* 5062 * Create and add a new L2ARC header. 5063 */ 5064 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5065 l2hdr->b_dev = dev; 5066 ab->b_flags |= ARC_L2_WRITING; 5067 5068 /* 5069 * Temporarily stash the data buffer in b_tmp_cdata. 5070 * The subsequent write step will pick it up from 5071 * there. This is because can't access ab->b_buf 5072 * without holding the hash_lock, which we in turn 5073 * can't access without holding the ARC list locks 5074 * (which we want to avoid during compression/writing). 5075 */ 5076 l2hdr->b_compress = ZIO_COMPRESS_OFF; 5077 l2hdr->b_asize = ab->b_size; 5078 l2hdr->b_tmp_cdata = ab->b_buf->b_data; 5079 5080 buf_sz = ab->b_size; 5081 ab->b_l2hdr = l2hdr; 5082 5083 list_insert_head(dev->l2ad_buflist, ab); 5084 5085 /* 5086 * Compute and store the buffer cksum before 5087 * writing. On debug the cksum is verified first. 5088 */ 5089 arc_cksum_verify(ab->b_buf); 5090 arc_cksum_compute(ab->b_buf, B_TRUE); 5091 5092 mutex_exit(hash_lock); 5093 5094 write_sz += buf_sz; 5095 } 5096 5097 mutex_exit(list_lock); 5098 5099 if (full == B_TRUE) 5100 break; 5101 } 5102 5103 /* No buffers selected for writing? */ 5104 if (pio == NULL) { 5105 ASSERT0(write_sz); 5106 mutex_exit(&l2arc_buflist_mtx); 5107 kmem_cache_free(hdr_cache, head); 5108 return (0); 5109 } 5110 5111 /* 5112 * Now start writing the buffers. We're starting at the write head 5113 * and work backwards, retracing the course of the buffer selector 5114 * loop above. 5115 */ 5116 for (ab = list_prev(dev->l2ad_buflist, head); ab; 5117 ab = list_prev(dev->l2ad_buflist, ab)) { 5118 l2arc_buf_hdr_t *l2hdr; 5119 uint64_t buf_sz; 5120 5121 /* 5122 * We shouldn't need to lock the buffer here, since we flagged 5123 * it as ARC_L2_WRITING in the previous step, but we must take 5124 * care to only access its L2 cache parameters. In particular, 5125 * ab->b_buf may be invalid by now due to ARC eviction. 5126 */ 5127 l2hdr = ab->b_l2hdr; 5128 l2hdr->b_daddr = dev->l2ad_hand; 5129 5130 if ((ab->b_flags & ARC_L2COMPRESS) && 5131 l2hdr->b_asize >= buf_compress_minsz) { 5132 if (l2arc_compress_buf(l2hdr)) { 5133 /* 5134 * If compression succeeded, enable headroom 5135 * boost on the next scan cycle. 5136 */ 5137 *headroom_boost = B_TRUE; 5138 } 5139 } 5140 5141 /* 5142 * Pick up the buffer data we had previously stashed away 5143 * (and now potentially also compressed). 5144 */ 5145 buf_data = l2hdr->b_tmp_cdata; 5146 buf_sz = l2hdr->b_asize; 5147 5148 /* Compression may have squashed the buffer to zero length. */ 5149 if (buf_sz != 0) { 5150 uint64_t buf_p_sz; 5151 5152 wzio = zio_write_phys(pio, dev->l2ad_vdev, 5153 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5154 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5155 ZIO_FLAG_CANFAIL, B_FALSE); 5156 5157 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5158 zio_t *, wzio); 5159 (void) zio_nowait(wzio); 5160 5161 write_asize += buf_sz; 5162 /* 5163 * Keep the clock hand suitably device-aligned. 5164 */ 5165 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5166 write_psize += buf_p_sz; 5167 dev->l2ad_hand += buf_p_sz; 5168 } 5169 } 5170 5171 mutex_exit(&l2arc_buflist_mtx); 5172 5173 ASSERT3U(write_asize, <=, target_sz); 5174 ARCSTAT_BUMP(arcstat_l2_writes_sent); 5175 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5176 ARCSTAT_INCR(arcstat_l2_size, write_sz); 5177 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 5178 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); 5179 5180 /* 5181 * Bump device hand to the device start if it is approaching the end. 5182 * l2arc_evict() will already have evicted ahead for this case. 5183 */ 5184 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5185 vdev_space_update(dev->l2ad_vdev, 5186 dev->l2ad_end - dev->l2ad_hand, 0, 0); 5187 dev->l2ad_hand = dev->l2ad_start; 5188 dev->l2ad_evict = dev->l2ad_start; 5189 dev->l2ad_first = B_FALSE; 5190 } 5191 5192 dev->l2ad_writing = B_TRUE; 5193 (void) zio_wait(pio); 5194 dev->l2ad_writing = B_FALSE; 5195 5196 return (write_asize); 5197} 5198 5199/* 5200 * Compresses an L2ARC buffer. 5201 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5202 * size in l2hdr->b_asize. This routine tries to compress the data and 5203 * depending on the compression result there are three possible outcomes: 5204 * *) The buffer was incompressible. The original l2hdr contents were left 5205 * untouched and are ready for writing to an L2 device. 5206 * *) The buffer was all-zeros, so there is no need to write it to an L2 5207 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5208 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5209 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5210 * data buffer which holds the compressed data to be written, and b_asize 5211 * tells us how much data there is. b_compress is set to the appropriate 5212 * compression algorithm. Once writing is done, invoke 5213 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5214 * 5215 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5216 * buffer was incompressible). 5217 */ 5218static boolean_t 5219l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5220{ 5221 void *cdata; 5222 size_t csize, len, rounded; 5223 5224 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5225 ASSERT(l2hdr->b_tmp_cdata != NULL); 5226 5227 len = l2hdr->b_asize; 5228 cdata = zio_data_buf_alloc(len); 5229 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5230 cdata, l2hdr->b_asize, (size_t)(1ULL << l2hdr->b_dev->l2ad_vdev->vdev_ashift)); 5231 5232 rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE); 5233 if (rounded > csize) { 5234 bzero((char *)cdata + csize, rounded - csize); 5235 csize = rounded; 5236 } 5237 5238 if (csize == 0) { 5239 /* zero block, indicate that there's nothing to write */ 5240 zio_data_buf_free(cdata, len); 5241 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5242 l2hdr->b_asize = 0; 5243 l2hdr->b_tmp_cdata = NULL; 5244 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5245 return (B_TRUE); 5246 } else if (csize > 0 && csize < len) { 5247 /* 5248 * Compression succeeded, we'll keep the cdata around for 5249 * writing and release it afterwards. 5250 */ 5251 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5252 l2hdr->b_asize = csize; 5253 l2hdr->b_tmp_cdata = cdata; 5254 ARCSTAT_BUMP(arcstat_l2_compress_successes); 5255 return (B_TRUE); 5256 } else { 5257 /* 5258 * Compression failed, release the compressed buffer. 5259 * l2hdr will be left unmodified. 5260 */ 5261 zio_data_buf_free(cdata, len); 5262 ARCSTAT_BUMP(arcstat_l2_compress_failures); 5263 return (B_FALSE); 5264 } 5265} 5266 5267/* 5268 * Decompresses a zio read back from an l2arc device. On success, the 5269 * underlying zio's io_data buffer is overwritten by the uncompressed 5270 * version. On decompression error (corrupt compressed stream), the 5271 * zio->io_error value is set to signal an I/O error. 5272 * 5273 * Please note that the compressed data stream is not checksummed, so 5274 * if the underlying device is experiencing data corruption, we may feed 5275 * corrupt data to the decompressor, so the decompressor needs to be 5276 * able to handle this situation (LZ4 does). 5277 */ 5278static void 5279l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5280{ 5281 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5282 5283 if (zio->io_error != 0) { 5284 /* 5285 * An io error has occured, just restore the original io 5286 * size in preparation for a main pool read. 5287 */ 5288 zio->io_orig_size = zio->io_size = hdr->b_size; 5289 return; 5290 } 5291 5292 if (c == ZIO_COMPRESS_EMPTY) { 5293 /* 5294 * An empty buffer results in a null zio, which means we 5295 * need to fill its io_data after we're done restoring the 5296 * buffer's contents. 5297 */ 5298 ASSERT(hdr->b_buf != NULL); 5299 bzero(hdr->b_buf->b_data, hdr->b_size); 5300 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5301 } else { 5302 ASSERT(zio->io_data != NULL); 5303 /* 5304 * We copy the compressed data from the start of the arc buffer 5305 * (the zio_read will have pulled in only what we need, the 5306 * rest is garbage which we will overwrite at decompression) 5307 * and then decompress back to the ARC data buffer. This way we 5308 * can minimize copying by simply decompressing back over the 5309 * original compressed data (rather than decompressing to an 5310 * aux buffer and then copying back the uncompressed buffer, 5311 * which is likely to be much larger). 5312 */ 5313 uint64_t csize; 5314 void *cdata; 5315 5316 csize = zio->io_size; 5317 cdata = zio_data_buf_alloc(csize); 5318 bcopy(zio->io_data, cdata, csize); 5319 if (zio_decompress_data(c, cdata, zio->io_data, csize, 5320 hdr->b_size) != 0) 5321 zio->io_error = EIO; 5322 zio_data_buf_free(cdata, csize); 5323 } 5324 5325 /* Restore the expected uncompressed IO size. */ 5326 zio->io_orig_size = zio->io_size = hdr->b_size; 5327} 5328 5329/* 5330 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5331 * This buffer serves as a temporary holder of compressed data while 5332 * the buffer entry is being written to an l2arc device. Once that is 5333 * done, we can dispose of it. 5334 */ 5335static void 5336l2arc_release_cdata_buf(arc_buf_hdr_t *ab) 5337{ 5338 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr; 5339 5340 if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) { 5341 /* 5342 * If the data was compressed, then we've allocated a 5343 * temporary buffer for it, so now we need to release it. 5344 */ 5345 ASSERT(l2hdr->b_tmp_cdata != NULL); 5346 zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size); 5347 } 5348 l2hdr->b_tmp_cdata = NULL; 5349} 5350 5351/* 5352 * This thread feeds the L2ARC at regular intervals. This is the beating 5353 * heart of the L2ARC. 5354 */ 5355static void 5356l2arc_feed_thread(void *dummy __unused) 5357{ 5358 callb_cpr_t cpr; 5359 l2arc_dev_t *dev; 5360 spa_t *spa; 5361 uint64_t size, wrote; 5362 clock_t begin, next = ddi_get_lbolt(); 5363 boolean_t headroom_boost = B_FALSE; 5364 5365 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5366 5367 mutex_enter(&l2arc_feed_thr_lock); 5368 5369 while (l2arc_thread_exit == 0) { 5370 CALLB_CPR_SAFE_BEGIN(&cpr); 5371 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5372 next - ddi_get_lbolt()); 5373 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5374 next = ddi_get_lbolt() + hz; 5375 5376 /* 5377 * Quick check for L2ARC devices. 5378 */ 5379 mutex_enter(&l2arc_dev_mtx); 5380 if (l2arc_ndev == 0) { 5381 mutex_exit(&l2arc_dev_mtx); 5382 continue; 5383 } 5384 mutex_exit(&l2arc_dev_mtx); 5385 begin = ddi_get_lbolt(); 5386 5387 /* 5388 * This selects the next l2arc device to write to, and in 5389 * doing so the next spa to feed from: dev->l2ad_spa. This 5390 * will return NULL if there are now no l2arc devices or if 5391 * they are all faulted. 5392 * 5393 * If a device is returned, its spa's config lock is also 5394 * held to prevent device removal. l2arc_dev_get_next() 5395 * will grab and release l2arc_dev_mtx. 5396 */ 5397 if ((dev = l2arc_dev_get_next()) == NULL) 5398 continue; 5399 5400 spa = dev->l2ad_spa; 5401 ASSERT(spa != NULL); 5402 5403 /* 5404 * If the pool is read-only then force the feed thread to 5405 * sleep a little longer. 5406 */ 5407 if (!spa_writeable(spa)) { 5408 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5409 spa_config_exit(spa, SCL_L2ARC, dev); 5410 continue; 5411 } 5412 5413 /* 5414 * Avoid contributing to memory pressure. 5415 */ 5416 if (arc_reclaim_needed()) { 5417 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5418 spa_config_exit(spa, SCL_L2ARC, dev); 5419 continue; 5420 } 5421 5422 ARCSTAT_BUMP(arcstat_l2_feeds); 5423 5424 size = l2arc_write_size(); 5425 5426 /* 5427 * Evict L2ARC buffers that will be overwritten. 5428 */ 5429 l2arc_evict(dev, size, B_FALSE); 5430 5431 /* 5432 * Write ARC buffers. 5433 */ 5434 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5435 5436 /* 5437 * Calculate interval between writes. 5438 */ 5439 next = l2arc_write_interval(begin, size, wrote); 5440 spa_config_exit(spa, SCL_L2ARC, dev); 5441 } 5442 5443 l2arc_thread_exit = 0; 5444 cv_broadcast(&l2arc_feed_thr_cv); 5445 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5446 thread_exit(); 5447} 5448 5449boolean_t 5450l2arc_vdev_present(vdev_t *vd) 5451{ 5452 l2arc_dev_t *dev; 5453 5454 mutex_enter(&l2arc_dev_mtx); 5455 for (dev = list_head(l2arc_dev_list); dev != NULL; 5456 dev = list_next(l2arc_dev_list, dev)) { 5457 if (dev->l2ad_vdev == vd) 5458 break; 5459 } 5460 mutex_exit(&l2arc_dev_mtx); 5461 5462 return (dev != NULL); 5463} 5464 5465/* 5466 * Add a vdev for use by the L2ARC. By this point the spa has already 5467 * validated the vdev and opened it. 5468 */ 5469void 5470l2arc_add_vdev(spa_t *spa, vdev_t *vd) 5471{ 5472 l2arc_dev_t *adddev; 5473 5474 ASSERT(!l2arc_vdev_present(vd)); 5475 5476 vdev_ashift_optimize(vd); 5477 5478 /* 5479 * Create a new l2arc device entry. 5480 */ 5481 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5482 adddev->l2ad_spa = spa; 5483 adddev->l2ad_vdev = vd; 5484 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5485 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5486 adddev->l2ad_hand = adddev->l2ad_start; 5487 adddev->l2ad_evict = adddev->l2ad_start; 5488 adddev->l2ad_first = B_TRUE; 5489 adddev->l2ad_writing = B_FALSE; 5490 5491 /* 5492 * This is a list of all ARC buffers that are still valid on the 5493 * device. 5494 */ 5495 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5496 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5497 offsetof(arc_buf_hdr_t, b_l2node)); 5498 5499 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5500 5501 /* 5502 * Add device to global list 5503 */ 5504 mutex_enter(&l2arc_dev_mtx); 5505 list_insert_head(l2arc_dev_list, adddev); 5506 atomic_inc_64(&l2arc_ndev); 5507 mutex_exit(&l2arc_dev_mtx); 5508} 5509 5510/* 5511 * Remove a vdev from the L2ARC. 5512 */ 5513void 5514l2arc_remove_vdev(vdev_t *vd) 5515{ 5516 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5517 5518 /* 5519 * Find the device by vdev 5520 */ 5521 mutex_enter(&l2arc_dev_mtx); 5522 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5523 nextdev = list_next(l2arc_dev_list, dev); 5524 if (vd == dev->l2ad_vdev) { 5525 remdev = dev; 5526 break; 5527 } 5528 } 5529 ASSERT(remdev != NULL); 5530 5531 /* 5532 * Remove device from global list 5533 */ 5534 list_remove(l2arc_dev_list, remdev); 5535 l2arc_dev_last = NULL; /* may have been invalidated */ 5536 atomic_dec_64(&l2arc_ndev); 5537 mutex_exit(&l2arc_dev_mtx); 5538 5539 /* 5540 * Clear all buflists and ARC references. L2ARC device flush. 5541 */ 5542 l2arc_evict(remdev, 0, B_TRUE); 5543 list_destroy(remdev->l2ad_buflist); 5544 kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5545 kmem_free(remdev, sizeof (l2arc_dev_t)); 5546} 5547 5548void 5549l2arc_init(void) 5550{ 5551 l2arc_thread_exit = 0; 5552 l2arc_ndev = 0; 5553 l2arc_writes_sent = 0; 5554 l2arc_writes_done = 0; 5555 5556 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5557 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5558 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5559 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5560 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5561 5562 l2arc_dev_list = &L2ARC_dev_list; 5563 l2arc_free_on_write = &L2ARC_free_on_write; 5564 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5565 offsetof(l2arc_dev_t, l2ad_node)); 5566 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5567 offsetof(l2arc_data_free_t, l2df_list_node)); 5568} 5569 5570void 5571l2arc_fini(void) 5572{ 5573 /* 5574 * This is called from dmu_fini(), which is called from spa_fini(); 5575 * Because of this, we can assume that all l2arc devices have 5576 * already been removed when the pools themselves were removed. 5577 */ 5578 5579 l2arc_do_free_on_write(); 5580 5581 mutex_destroy(&l2arc_feed_thr_lock); 5582 cv_destroy(&l2arc_feed_thr_cv); 5583 mutex_destroy(&l2arc_dev_mtx); 5584 mutex_destroy(&l2arc_buflist_mtx); 5585 mutex_destroy(&l2arc_free_on_write_mtx); 5586 5587 list_destroy(l2arc_dev_list); 5588 list_destroy(l2arc_free_on_write); 5589} 5590 5591void 5592l2arc_start(void) 5593{ 5594 if (!(spa_mode_global & FWRITE)) 5595 return; 5596 5597 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5598 TS_RUN, minclsyspri); 5599} 5600 5601void 5602l2arc_stop(void) 5603{ 5604 if (!(spa_mode_global & FWRITE)) 5605 return; 5606 5607 mutex_enter(&l2arc_feed_thr_lock); 5608 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5609 l2arc_thread_exit = 1; 5610 while (l2arc_thread_exit != 0) 5611 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5612 mutex_exit(&l2arc_feed_thr_lock); 5613} 5614