arc.c revision 269732
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 24 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 25 * Copyright 2014 Nexenta Systems, Inc. All rights reserved. 26 */ 27 28/* 29 * DVA-based Adjustable Replacement Cache 30 * 31 * While much of the theory of operation used here is 32 * based on the self-tuning, low overhead replacement cache 33 * presented by Megiddo and Modha at FAST 2003, there are some 34 * significant differences: 35 * 36 * 1. The Megiddo and Modha model assumes any page is evictable. 37 * Pages in its cache cannot be "locked" into memory. This makes 38 * the eviction algorithm simple: evict the last page in the list. 39 * This also make the performance characteristics easy to reason 40 * about. Our cache is not so simple. At any given moment, some 41 * subset of the blocks in the cache are un-evictable because we 42 * have handed out a reference to them. Blocks are only evictable 43 * when there are no external references active. This makes 44 * eviction far more problematic: we choose to evict the evictable 45 * blocks that are the "lowest" in the list. 46 * 47 * There are times when it is not possible to evict the requested 48 * space. In these circumstances we are unable to adjust the cache 49 * size. To prevent the cache growing unbounded at these times we 50 * implement a "cache throttle" that slows the flow of new data 51 * into the cache until we can make space available. 52 * 53 * 2. The Megiddo and Modha model assumes a fixed cache size. 54 * Pages are evicted when the cache is full and there is a cache 55 * miss. Our model has a variable sized cache. It grows with 56 * high use, but also tries to react to memory pressure from the 57 * operating system: decreasing its size when system memory is 58 * tight. 59 * 60 * 3. The Megiddo and Modha model assumes a fixed page size. All 61 * elements of the cache are therefore exactly the same size. So 62 * when adjusting the cache size following a cache miss, its simply 63 * a matter of choosing a single page to evict. In our model, we 64 * have variable sized cache blocks (rangeing from 512 bytes to 65 * 128K bytes). We therefore choose a set of blocks to evict to make 66 * space for a cache miss that approximates as closely as possible 67 * the space used by the new block. 68 * 69 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 70 * by N. Megiddo & D. Modha, FAST 2003 71 */ 72 73/* 74 * The locking model: 75 * 76 * A new reference to a cache buffer can be obtained in two 77 * ways: 1) via a hash table lookup using the DVA as a key, 78 * or 2) via one of the ARC lists. The arc_read() interface 79 * uses method 1, while the internal arc algorithms for 80 * adjusting the cache use method 2. We therefore provide two 81 * types of locks: 1) the hash table lock array, and 2) the 82 * arc list locks. 83 * 84 * Buffers do not have their own mutexs, rather they rely on the 85 * hash table mutexs for the bulk of their protection (i.e. most 86 * fields in the arc_buf_hdr_t are protected by these mutexs). 87 * 88 * buf_hash_find() returns the appropriate mutex (held) when it 89 * locates the requested buffer in the hash table. It returns 90 * NULL for the mutex if the buffer was not in the table. 91 * 92 * buf_hash_remove() expects the appropriate hash mutex to be 93 * already held before it is invoked. 94 * 95 * Each arc state also has a mutex which is used to protect the 96 * buffer list associated with the state. When attempting to 97 * obtain a hash table lock while holding an arc list lock you 98 * must use: mutex_tryenter() to avoid deadlock. Also note that 99 * the active state mutex must be held before the ghost state mutex. 100 * 101 * Arc buffers may have an associated eviction callback function. 102 * This function will be invoked prior to removing the buffer (e.g. 103 * in arc_do_user_evicts()). Note however that the data associated 104 * with the buffer may be evicted prior to the callback. The callback 105 * must be made with *no locks held* (to prevent deadlock). Additionally, 106 * the users of callbacks must ensure that their private data is 107 * protected from simultaneous callbacks from arc_clear_callback() 108 * and arc_do_user_evicts(). 109 * 110 * Note that the majority of the performance stats are manipulated 111 * with atomic operations. 112 * 113 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following: 114 * 115 * - L2ARC buflist creation 116 * - L2ARC buflist eviction 117 * - L2ARC write completion, which walks L2ARC buflists 118 * - ARC header destruction, as it removes from L2ARC buflists 119 * - ARC header release, as it removes from L2ARC buflists 120 */ 121 122#include <sys/spa.h> 123#include <sys/zio.h> 124#include <sys/zio_compress.h> 125#include <sys/zfs_context.h> 126#include <sys/arc.h> 127#include <sys/refcount.h> 128#include <sys/vdev.h> 129#include <sys/vdev_impl.h> 130#include <sys/dsl_pool.h> 131#ifdef _KERNEL 132#include <sys/dnlc.h> 133#endif 134#include <sys/callb.h> 135#include <sys/kstat.h> 136#include <sys/trim_map.h> 137#include <zfs_fletcher.h> 138#include <sys/sdt.h> 139 140#include <vm/vm_pageout.h> 141 142#ifdef illumos 143#ifndef _KERNEL 144/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 145boolean_t arc_watch = B_FALSE; 146int arc_procfd; 147#endif 148#endif /* illumos */ 149 150static kmutex_t arc_reclaim_thr_lock; 151static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 152static uint8_t arc_thread_exit; 153 154#define ARC_REDUCE_DNLC_PERCENT 3 155uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 156 157typedef enum arc_reclaim_strategy { 158 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 159 ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 160} arc_reclaim_strategy_t; 161 162/* 163 * The number of iterations through arc_evict_*() before we 164 * drop & reacquire the lock. 165 */ 166int arc_evict_iterations = 100; 167 168/* number of seconds before growing cache again */ 169static int arc_grow_retry = 60; 170 171/* shift of arc_c for calculating both min and max arc_p */ 172static int arc_p_min_shift = 4; 173 174/* log2(fraction of arc to reclaim) */ 175static int arc_shrink_shift = 5; 176 177/* 178 * minimum lifespan of a prefetch block in clock ticks 179 * (initialized in arc_init()) 180 */ 181static int arc_min_prefetch_lifespan; 182 183/* 184 * If this percent of memory is free, don't throttle. 185 */ 186int arc_lotsfree_percent = 10; 187 188static int arc_dead; 189extern int zfs_prefetch_disable; 190 191/* 192 * The arc has filled available memory and has now warmed up. 193 */ 194static boolean_t arc_warm; 195 196/* 197 * These tunables are for performance analysis. 198 */ 199uint64_t zfs_arc_max; 200uint64_t zfs_arc_min; 201uint64_t zfs_arc_meta_limit = 0; 202int zfs_arc_grow_retry = 0; 203int zfs_arc_shrink_shift = 0; 204int zfs_arc_p_min_shift = 0; 205int zfs_disable_dup_eviction = 0; 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_phys_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 1650/* 1651 * Free up buf->b_data and if 'remove' is set, then pull the 1652 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 1653 */ 1654static void 1655arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 1656{ 1657 arc_buf_t **bufp; 1658 1659 /* free up data associated with the buf */ 1660 if (buf->b_data) { 1661 arc_state_t *state = buf->b_hdr->b_state; 1662 uint64_t size = buf->b_hdr->b_size; 1663 arc_buf_contents_t type = buf->b_hdr->b_type; 1664 1665 arc_cksum_verify(buf); 1666#ifdef illumos 1667 arc_buf_unwatch(buf); 1668#endif /* illumos */ 1669 1670 if (!recycle) { 1671 if (type == ARC_BUFC_METADATA) { 1672 arc_buf_data_free(buf, zio_buf_free); 1673 arc_space_return(size, ARC_SPACE_DATA); 1674 } else { 1675 ASSERT(type == ARC_BUFC_DATA); 1676 arc_buf_data_free(buf, zio_data_buf_free); 1677 ARCSTAT_INCR(arcstat_data_size, -size); 1678 atomic_add_64(&arc_size, -size); 1679 } 1680 } 1681 if (list_link_active(&buf->b_hdr->b_arc_node)) { 1682 uint64_t *cnt = &state->arcs_lsize[type]; 1683 1684 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1685 ASSERT(state != arc_anon); 1686 1687 ASSERT3U(*cnt, >=, size); 1688 atomic_add_64(cnt, -size); 1689 } 1690 ASSERT3U(state->arcs_size, >=, size); 1691 atomic_add_64(&state->arcs_size, -size); 1692 buf->b_data = NULL; 1693 1694 /* 1695 * If we're destroying a duplicate buffer make sure 1696 * that the appropriate statistics are updated. 1697 */ 1698 if (buf->b_hdr->b_datacnt > 1 && 1699 buf->b_hdr->b_type == ARC_BUFC_DATA) { 1700 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1701 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1702 } 1703 ASSERT(buf->b_hdr->b_datacnt > 0); 1704 buf->b_hdr->b_datacnt -= 1; 1705 } 1706 1707 /* only remove the buf if requested */ 1708 if (!remove) 1709 return; 1710 1711 /* remove the buf from the hdr list */ 1712 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1713 continue; 1714 *bufp = buf->b_next; 1715 buf->b_next = NULL; 1716 1717 ASSERT(buf->b_efunc == NULL); 1718 1719 /* clean up the buf */ 1720 buf->b_hdr = NULL; 1721 kmem_cache_free(buf_cache, buf); 1722} 1723 1724static void 1725arc_hdr_destroy(arc_buf_hdr_t *hdr) 1726{ 1727 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1728 ASSERT3P(hdr->b_state, ==, arc_anon); 1729 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1730 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1731 1732 if (l2hdr != NULL) { 1733 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1734 /* 1735 * To prevent arc_free() and l2arc_evict() from 1736 * attempting to free the same buffer at the same time, 1737 * a FREE_IN_PROGRESS flag is given to arc_free() to 1738 * give it priority. l2arc_evict() can't destroy this 1739 * header while we are waiting on l2arc_buflist_mtx. 1740 * 1741 * The hdr may be removed from l2ad_buflist before we 1742 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1743 */ 1744 if (!buflist_held) { 1745 mutex_enter(&l2arc_buflist_mtx); 1746 l2hdr = hdr->b_l2hdr; 1747 } 1748 1749 if (l2hdr != NULL) { 1750 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1751 hdr->b_size, 0); 1752 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1753 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1754 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1755 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 1756 -l2hdr->b_asize, 0, 0); 1757 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1758 if (hdr->b_state == arc_l2c_only) 1759 l2arc_hdr_stat_remove(); 1760 hdr->b_l2hdr = NULL; 1761 } 1762 1763 if (!buflist_held) 1764 mutex_exit(&l2arc_buflist_mtx); 1765 } 1766 1767 if (!BUF_EMPTY(hdr)) { 1768 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1769 buf_discard_identity(hdr); 1770 } 1771 while (hdr->b_buf) { 1772 arc_buf_t *buf = hdr->b_buf; 1773 1774 if (buf->b_efunc) { 1775 mutex_enter(&arc_eviction_mtx); 1776 mutex_enter(&buf->b_evict_lock); 1777 ASSERT(buf->b_hdr != NULL); 1778 arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1779 hdr->b_buf = buf->b_next; 1780 buf->b_hdr = &arc_eviction_hdr; 1781 buf->b_next = arc_eviction_list; 1782 arc_eviction_list = buf; 1783 mutex_exit(&buf->b_evict_lock); 1784 mutex_exit(&arc_eviction_mtx); 1785 } else { 1786 arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1787 } 1788 } 1789 if (hdr->b_freeze_cksum != NULL) { 1790 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1791 hdr->b_freeze_cksum = NULL; 1792 } 1793 if (hdr->b_thawed) { 1794 kmem_free(hdr->b_thawed, 1); 1795 hdr->b_thawed = NULL; 1796 } 1797 1798 ASSERT(!list_link_active(&hdr->b_arc_node)); 1799 ASSERT3P(hdr->b_hash_next, ==, NULL); 1800 ASSERT3P(hdr->b_acb, ==, NULL); 1801 kmem_cache_free(hdr_cache, hdr); 1802} 1803 1804void 1805arc_buf_free(arc_buf_t *buf, void *tag) 1806{ 1807 arc_buf_hdr_t *hdr = buf->b_hdr; 1808 int hashed = hdr->b_state != arc_anon; 1809 1810 ASSERT(buf->b_efunc == NULL); 1811 ASSERT(buf->b_data != NULL); 1812 1813 if (hashed) { 1814 kmutex_t *hash_lock = HDR_LOCK(hdr); 1815 1816 mutex_enter(hash_lock); 1817 hdr = buf->b_hdr; 1818 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1819 1820 (void) remove_reference(hdr, hash_lock, tag); 1821 if (hdr->b_datacnt > 1) { 1822 arc_buf_destroy(buf, FALSE, TRUE); 1823 } else { 1824 ASSERT(buf == hdr->b_buf); 1825 ASSERT(buf->b_efunc == NULL); 1826 hdr->b_flags |= ARC_BUF_AVAILABLE; 1827 } 1828 mutex_exit(hash_lock); 1829 } else if (HDR_IO_IN_PROGRESS(hdr)) { 1830 int destroy_hdr; 1831 /* 1832 * We are in the middle of an async write. Don't destroy 1833 * this buffer unless the write completes before we finish 1834 * decrementing the reference count. 1835 */ 1836 mutex_enter(&arc_eviction_mtx); 1837 (void) remove_reference(hdr, NULL, tag); 1838 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1839 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1840 mutex_exit(&arc_eviction_mtx); 1841 if (destroy_hdr) 1842 arc_hdr_destroy(hdr); 1843 } else { 1844 if (remove_reference(hdr, NULL, tag) > 0) 1845 arc_buf_destroy(buf, FALSE, TRUE); 1846 else 1847 arc_hdr_destroy(hdr); 1848 } 1849} 1850 1851boolean_t 1852arc_buf_remove_ref(arc_buf_t *buf, void* tag) 1853{ 1854 arc_buf_hdr_t *hdr = buf->b_hdr; 1855 kmutex_t *hash_lock = HDR_LOCK(hdr); 1856 boolean_t no_callback = (buf->b_efunc == NULL); 1857 1858 if (hdr->b_state == arc_anon) { 1859 ASSERT(hdr->b_datacnt == 1); 1860 arc_buf_free(buf, tag); 1861 return (no_callback); 1862 } 1863 1864 mutex_enter(hash_lock); 1865 hdr = buf->b_hdr; 1866 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1867 ASSERT(hdr->b_state != arc_anon); 1868 ASSERT(buf->b_data != NULL); 1869 1870 (void) remove_reference(hdr, hash_lock, tag); 1871 if (hdr->b_datacnt > 1) { 1872 if (no_callback) 1873 arc_buf_destroy(buf, FALSE, TRUE); 1874 } else if (no_callback) { 1875 ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1876 ASSERT(buf->b_efunc == NULL); 1877 hdr->b_flags |= ARC_BUF_AVAILABLE; 1878 } 1879 ASSERT(no_callback || hdr->b_datacnt > 1 || 1880 refcount_is_zero(&hdr->b_refcnt)); 1881 mutex_exit(hash_lock); 1882 return (no_callback); 1883} 1884 1885int 1886arc_buf_size(arc_buf_t *buf) 1887{ 1888 return (buf->b_hdr->b_size); 1889} 1890 1891/* 1892 * Called from the DMU to determine if the current buffer should be 1893 * evicted. In order to ensure proper locking, the eviction must be initiated 1894 * from the DMU. Return true if the buffer is associated with user data and 1895 * duplicate buffers still exist. 1896 */ 1897boolean_t 1898arc_buf_eviction_needed(arc_buf_t *buf) 1899{ 1900 arc_buf_hdr_t *hdr; 1901 boolean_t evict_needed = B_FALSE; 1902 1903 if (zfs_disable_dup_eviction) 1904 return (B_FALSE); 1905 1906 mutex_enter(&buf->b_evict_lock); 1907 hdr = buf->b_hdr; 1908 if (hdr == NULL) { 1909 /* 1910 * We are in arc_do_user_evicts(); let that function 1911 * perform the eviction. 1912 */ 1913 ASSERT(buf->b_data == NULL); 1914 mutex_exit(&buf->b_evict_lock); 1915 return (B_FALSE); 1916 } else if (buf->b_data == NULL) { 1917 /* 1918 * We have already been added to the arc eviction list; 1919 * recommend eviction. 1920 */ 1921 ASSERT3P(hdr, ==, &arc_eviction_hdr); 1922 mutex_exit(&buf->b_evict_lock); 1923 return (B_TRUE); 1924 } 1925 1926 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 1927 evict_needed = B_TRUE; 1928 1929 mutex_exit(&buf->b_evict_lock); 1930 return (evict_needed); 1931} 1932 1933/* 1934 * Evict buffers from list until we've removed the specified number of 1935 * bytes. Move the removed buffers to the appropriate evict state. 1936 * If the recycle flag is set, then attempt to "recycle" a buffer: 1937 * - look for a buffer to evict that is `bytes' long. 1938 * - return the data block from this buffer rather than freeing it. 1939 * This flag is used by callers that are trying to make space for a 1940 * new buffer in a full arc cache. 1941 * 1942 * This function makes a "best effort". It skips over any buffers 1943 * it can't get a hash_lock on, and so may not catch all candidates. 1944 * It may also return without evicting as much space as requested. 1945 */ 1946static void * 1947arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 1948 arc_buf_contents_t type) 1949{ 1950 arc_state_t *evicted_state; 1951 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 1952 int64_t bytes_remaining; 1953 arc_buf_hdr_t *ab, *ab_prev = NULL; 1954 list_t *evicted_list, *list, *evicted_list_start, *list_start; 1955 kmutex_t *lock, *evicted_lock; 1956 kmutex_t *hash_lock; 1957 boolean_t have_lock; 1958 void *stolen = NULL; 1959 arc_buf_hdr_t marker = { 0 }; 1960 int count = 0; 1961 static int evict_metadata_offset, evict_data_offset; 1962 int i, idx, offset, list_count, lists; 1963 1964 ASSERT(state == arc_mru || state == arc_mfu); 1965 1966 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 1967 1968 if (type == ARC_BUFC_METADATA) { 1969 offset = 0; 1970 list_count = ARC_BUFC_NUMMETADATALISTS; 1971 list_start = &state->arcs_lists[0]; 1972 evicted_list_start = &evicted_state->arcs_lists[0]; 1973 idx = evict_metadata_offset; 1974 } else { 1975 offset = ARC_BUFC_NUMMETADATALISTS; 1976 list_start = &state->arcs_lists[offset]; 1977 evicted_list_start = &evicted_state->arcs_lists[offset]; 1978 list_count = ARC_BUFC_NUMDATALISTS; 1979 idx = evict_data_offset; 1980 } 1981 bytes_remaining = evicted_state->arcs_lsize[type]; 1982 lists = 0; 1983 1984evict_start: 1985 list = &list_start[idx]; 1986 evicted_list = &evicted_list_start[idx]; 1987 lock = ARCS_LOCK(state, (offset + idx)); 1988 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 1989 1990 mutex_enter(lock); 1991 mutex_enter(evicted_lock); 1992 1993 for (ab = list_tail(list); ab; ab = ab_prev) { 1994 ab_prev = list_prev(list, ab); 1995 bytes_remaining -= (ab->b_size * ab->b_datacnt); 1996 /* prefetch buffers have a minimum lifespan */ 1997 if (HDR_IO_IN_PROGRESS(ab) || 1998 (spa && ab->b_spa != spa) || 1999 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 2000 ddi_get_lbolt() - ab->b_arc_access < 2001 arc_min_prefetch_lifespan)) { 2002 skipped++; 2003 continue; 2004 } 2005 /* "lookahead" for better eviction candidate */ 2006 if (recycle && ab->b_size != bytes && 2007 ab_prev && ab_prev->b_size == bytes) 2008 continue; 2009 2010 /* ignore markers */ 2011 if (ab->b_spa == 0) 2012 continue; 2013 2014 /* 2015 * It may take a long time to evict all the bufs requested. 2016 * To avoid blocking all arc activity, periodically drop 2017 * the arcs_mtx and give other threads a chance to run 2018 * before reacquiring the lock. 2019 * 2020 * If we are looking for a buffer to recycle, we are in 2021 * the hot code path, so don't sleep. 2022 */ 2023 if (!recycle && count++ > arc_evict_iterations) { 2024 list_insert_after(list, ab, &marker); 2025 mutex_exit(evicted_lock); 2026 mutex_exit(lock); 2027 kpreempt(KPREEMPT_SYNC); 2028 mutex_enter(lock); 2029 mutex_enter(evicted_lock); 2030 ab_prev = list_prev(list, &marker); 2031 list_remove(list, &marker); 2032 count = 0; 2033 continue; 2034 } 2035 2036 hash_lock = HDR_LOCK(ab); 2037 have_lock = MUTEX_HELD(hash_lock); 2038 if (have_lock || mutex_tryenter(hash_lock)) { 2039 ASSERT0(refcount_count(&ab->b_refcnt)); 2040 ASSERT(ab->b_datacnt > 0); 2041 while (ab->b_buf) { 2042 arc_buf_t *buf = ab->b_buf; 2043 if (!mutex_tryenter(&buf->b_evict_lock)) { 2044 missed += 1; 2045 break; 2046 } 2047 if (buf->b_data) { 2048 bytes_evicted += ab->b_size; 2049 if (recycle && ab->b_type == type && 2050 ab->b_size == bytes && 2051 !HDR_L2_WRITING(ab)) { 2052 stolen = buf->b_data; 2053 recycle = FALSE; 2054 } 2055 } 2056 if (buf->b_efunc) { 2057 mutex_enter(&arc_eviction_mtx); 2058 arc_buf_destroy(buf, 2059 buf->b_data == stolen, FALSE); 2060 ab->b_buf = buf->b_next; 2061 buf->b_hdr = &arc_eviction_hdr; 2062 buf->b_next = arc_eviction_list; 2063 arc_eviction_list = buf; 2064 mutex_exit(&arc_eviction_mtx); 2065 mutex_exit(&buf->b_evict_lock); 2066 } else { 2067 mutex_exit(&buf->b_evict_lock); 2068 arc_buf_destroy(buf, 2069 buf->b_data == stolen, TRUE); 2070 } 2071 } 2072 2073 if (ab->b_l2hdr) { 2074 ARCSTAT_INCR(arcstat_evict_l2_cached, 2075 ab->b_size); 2076 } else { 2077 if (l2arc_write_eligible(ab->b_spa, ab)) { 2078 ARCSTAT_INCR(arcstat_evict_l2_eligible, 2079 ab->b_size); 2080 } else { 2081 ARCSTAT_INCR( 2082 arcstat_evict_l2_ineligible, 2083 ab->b_size); 2084 } 2085 } 2086 2087 if (ab->b_datacnt == 0) { 2088 arc_change_state(evicted_state, ab, hash_lock); 2089 ASSERT(HDR_IN_HASH_TABLE(ab)); 2090 ab->b_flags |= ARC_IN_HASH_TABLE; 2091 ab->b_flags &= ~ARC_BUF_AVAILABLE; 2092 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 2093 } 2094 if (!have_lock) 2095 mutex_exit(hash_lock); 2096 if (bytes >= 0 && bytes_evicted >= bytes) 2097 break; 2098 if (bytes_remaining > 0) { 2099 mutex_exit(evicted_lock); 2100 mutex_exit(lock); 2101 idx = ((idx + 1) & (list_count - 1)); 2102 lists++; 2103 goto evict_start; 2104 } 2105 } else { 2106 missed += 1; 2107 } 2108 } 2109 2110 mutex_exit(evicted_lock); 2111 mutex_exit(lock); 2112 2113 idx = ((idx + 1) & (list_count - 1)); 2114 lists++; 2115 2116 if (bytes_evicted < bytes) { 2117 if (lists < list_count) 2118 goto evict_start; 2119 else 2120 dprintf("only evicted %lld bytes from %x", 2121 (longlong_t)bytes_evicted, state); 2122 } 2123 if (type == ARC_BUFC_METADATA) 2124 evict_metadata_offset = idx; 2125 else 2126 evict_data_offset = idx; 2127 2128 if (skipped) 2129 ARCSTAT_INCR(arcstat_evict_skip, skipped); 2130 2131 if (missed) 2132 ARCSTAT_INCR(arcstat_mutex_miss, missed); 2133 2134 /* 2135 * Note: we have just evicted some data into the ghost state, 2136 * potentially putting the ghost size over the desired size. Rather 2137 * that evicting from the ghost list in this hot code path, leave 2138 * this chore to the arc_reclaim_thread(). 2139 */ 2140 2141 if (stolen) 2142 ARCSTAT_BUMP(arcstat_stolen); 2143 return (stolen); 2144} 2145 2146/* 2147 * Remove buffers from list until we've removed the specified number of 2148 * bytes. Destroy the buffers that are removed. 2149 */ 2150static void 2151arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2152{ 2153 arc_buf_hdr_t *ab, *ab_prev; 2154 arc_buf_hdr_t marker = { 0 }; 2155 list_t *list, *list_start; 2156 kmutex_t *hash_lock, *lock; 2157 uint64_t bytes_deleted = 0; 2158 uint64_t bufs_skipped = 0; 2159 int count = 0; 2160 static int evict_offset; 2161 int list_count, idx = evict_offset; 2162 int offset, lists = 0; 2163 2164 ASSERT(GHOST_STATE(state)); 2165 2166 /* 2167 * data lists come after metadata lists 2168 */ 2169 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2170 list_count = ARC_BUFC_NUMDATALISTS; 2171 offset = ARC_BUFC_NUMMETADATALISTS; 2172 2173evict_start: 2174 list = &list_start[idx]; 2175 lock = ARCS_LOCK(state, idx + offset); 2176 2177 mutex_enter(lock); 2178 for (ab = list_tail(list); ab; ab = ab_prev) { 2179 ab_prev = list_prev(list, ab); 2180 if (ab->b_type > ARC_BUFC_NUMTYPES) 2181 panic("invalid ab=%p", (void *)ab); 2182 if (spa && ab->b_spa != spa) 2183 continue; 2184 2185 /* ignore markers */ 2186 if (ab->b_spa == 0) 2187 continue; 2188 2189 hash_lock = HDR_LOCK(ab); 2190 /* caller may be trying to modify this buffer, skip it */ 2191 if (MUTEX_HELD(hash_lock)) 2192 continue; 2193 2194 /* 2195 * It may take a long time to evict all the bufs requested. 2196 * To avoid blocking all arc activity, periodically drop 2197 * the arcs_mtx and give other threads a chance to run 2198 * before reacquiring the lock. 2199 */ 2200 if (count++ > arc_evict_iterations) { 2201 list_insert_after(list, ab, &marker); 2202 mutex_exit(lock); 2203 kpreempt(KPREEMPT_SYNC); 2204 mutex_enter(lock); 2205 ab_prev = list_prev(list, &marker); 2206 list_remove(list, &marker); 2207 count = 0; 2208 continue; 2209 } 2210 if (mutex_tryenter(hash_lock)) { 2211 ASSERT(!HDR_IO_IN_PROGRESS(ab)); 2212 ASSERT(ab->b_buf == NULL); 2213 ARCSTAT_BUMP(arcstat_deleted); 2214 bytes_deleted += ab->b_size; 2215 2216 if (ab->b_l2hdr != NULL) { 2217 /* 2218 * This buffer is cached on the 2nd Level ARC; 2219 * don't destroy the header. 2220 */ 2221 arc_change_state(arc_l2c_only, ab, hash_lock); 2222 mutex_exit(hash_lock); 2223 } else { 2224 arc_change_state(arc_anon, ab, hash_lock); 2225 mutex_exit(hash_lock); 2226 arc_hdr_destroy(ab); 2227 } 2228 2229 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 2230 if (bytes >= 0 && bytes_deleted >= bytes) 2231 break; 2232 } else if (bytes < 0) { 2233 /* 2234 * Insert a list marker and then wait for the 2235 * hash lock to become available. Once its 2236 * available, restart from where we left off. 2237 */ 2238 list_insert_after(list, ab, &marker); 2239 mutex_exit(lock); 2240 mutex_enter(hash_lock); 2241 mutex_exit(hash_lock); 2242 mutex_enter(lock); 2243 ab_prev = list_prev(list, &marker); 2244 list_remove(list, &marker); 2245 } else { 2246 bufs_skipped += 1; 2247 } 2248 2249 } 2250 mutex_exit(lock); 2251 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2252 lists++; 2253 2254 if (lists < list_count) 2255 goto evict_start; 2256 2257 evict_offset = idx; 2258 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2259 (bytes < 0 || bytes_deleted < bytes)) { 2260 list_start = &state->arcs_lists[0]; 2261 list_count = ARC_BUFC_NUMMETADATALISTS; 2262 offset = lists = 0; 2263 goto evict_start; 2264 } 2265 2266 if (bufs_skipped) { 2267 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2268 ASSERT(bytes >= 0); 2269 } 2270 2271 if (bytes_deleted < bytes) 2272 dprintf("only deleted %lld bytes from %p", 2273 (longlong_t)bytes_deleted, state); 2274} 2275 2276static void 2277arc_adjust(void) 2278{ 2279 int64_t adjustment, delta; 2280 2281 /* 2282 * Adjust MRU size 2283 */ 2284 2285 adjustment = MIN((int64_t)(arc_size - arc_c), 2286 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2287 arc_p)); 2288 2289 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2290 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2291 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2292 adjustment -= delta; 2293 } 2294 2295 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2296 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2297 (void) arc_evict(arc_mru, 0, delta, FALSE, 2298 ARC_BUFC_METADATA); 2299 } 2300 2301 /* 2302 * Adjust MFU size 2303 */ 2304 2305 adjustment = arc_size - arc_c; 2306 2307 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2308 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2309 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2310 adjustment -= delta; 2311 } 2312 2313 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2314 int64_t delta = MIN(adjustment, 2315 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2316 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2317 ARC_BUFC_METADATA); 2318 } 2319 2320 /* 2321 * Adjust ghost lists 2322 */ 2323 2324 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2325 2326 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2327 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2328 arc_evict_ghost(arc_mru_ghost, 0, delta); 2329 } 2330 2331 adjustment = 2332 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2333 2334 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2335 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2336 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2337 } 2338} 2339 2340static void 2341arc_do_user_evicts(void) 2342{ 2343 static arc_buf_t *tmp_arc_eviction_list; 2344 2345 /* 2346 * Move list over to avoid LOR 2347 */ 2348restart: 2349 mutex_enter(&arc_eviction_mtx); 2350 tmp_arc_eviction_list = arc_eviction_list; 2351 arc_eviction_list = NULL; 2352 mutex_exit(&arc_eviction_mtx); 2353 2354 while (tmp_arc_eviction_list != NULL) { 2355 arc_buf_t *buf = tmp_arc_eviction_list; 2356 tmp_arc_eviction_list = buf->b_next; 2357 mutex_enter(&buf->b_evict_lock); 2358 buf->b_hdr = NULL; 2359 mutex_exit(&buf->b_evict_lock); 2360 2361 if (buf->b_efunc != NULL) 2362 VERIFY0(buf->b_efunc(buf->b_private)); 2363 2364 buf->b_efunc = NULL; 2365 buf->b_private = NULL; 2366 kmem_cache_free(buf_cache, buf); 2367 } 2368 2369 if (arc_eviction_list != NULL) 2370 goto restart; 2371} 2372 2373/* 2374 * Flush all *evictable* data from the cache for the given spa. 2375 * NOTE: this will not touch "active" (i.e. referenced) data. 2376 */ 2377void 2378arc_flush(spa_t *spa) 2379{ 2380 uint64_t guid = 0; 2381 2382 if (spa) 2383 guid = spa_load_guid(spa); 2384 2385 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2386 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2387 if (spa) 2388 break; 2389 } 2390 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2391 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2392 if (spa) 2393 break; 2394 } 2395 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2396 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2397 if (spa) 2398 break; 2399 } 2400 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2401 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2402 if (spa) 2403 break; 2404 } 2405 2406 arc_evict_ghost(arc_mru_ghost, guid, -1); 2407 arc_evict_ghost(arc_mfu_ghost, guid, -1); 2408 2409 mutex_enter(&arc_reclaim_thr_lock); 2410 arc_do_user_evicts(); 2411 mutex_exit(&arc_reclaim_thr_lock); 2412 ASSERT(spa || arc_eviction_list == NULL); 2413} 2414 2415void 2416arc_shrink(void) 2417{ 2418 if (arc_c > arc_c_min) { 2419 uint64_t to_free; 2420 2421#ifdef _KERNEL 2422 to_free = arc_c >> arc_shrink_shift; 2423#else 2424 to_free = arc_c >> arc_shrink_shift; 2425#endif 2426 if (arc_c > arc_c_min + to_free) 2427 atomic_add_64(&arc_c, -to_free); 2428 else 2429 arc_c = arc_c_min; 2430 2431 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2432 if (arc_c > arc_size) 2433 arc_c = MAX(arc_size, arc_c_min); 2434 if (arc_p > arc_c) 2435 arc_p = (arc_c >> 1); 2436 ASSERT(arc_c >= arc_c_min); 2437 ASSERT((int64_t)arc_p >= 0); 2438 } 2439 2440 if (arc_size > arc_c) 2441 arc_adjust(); 2442} 2443 2444static int needfree = 0; 2445 2446static int 2447arc_reclaim_needed(void) 2448{ 2449 2450#ifdef _KERNEL 2451 2452 if (needfree) 2453 return (1); 2454 2455 /* 2456 * Cooperate with pagedaemon when it's time for it to scan 2457 * and reclaim some pages. 2458 */ 2459 if (vm_paging_needed()) 2460 return (1); 2461 2462#ifdef sun 2463 /* 2464 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2465 */ 2466 extra = desfree; 2467 2468 /* 2469 * check that we're out of range of the pageout scanner. It starts to 2470 * schedule paging if freemem is less than lotsfree and needfree. 2471 * lotsfree is the high-water mark for pageout, and needfree is the 2472 * number of needed free pages. We add extra pages here to make sure 2473 * the scanner doesn't start up while we're freeing memory. 2474 */ 2475 if (freemem < lotsfree + needfree + extra) 2476 return (1); 2477 2478 /* 2479 * check to make sure that swapfs has enough space so that anon 2480 * reservations can still succeed. anon_resvmem() checks that the 2481 * availrmem is greater than swapfs_minfree, and the number of reserved 2482 * swap pages. We also add a bit of extra here just to prevent 2483 * circumstances from getting really dire. 2484 */ 2485 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2486 return (1); 2487 2488#if defined(__i386) 2489 /* 2490 * If we're on an i386 platform, it's possible that we'll exhaust the 2491 * kernel heap space before we ever run out of available physical 2492 * memory. Most checks of the size of the heap_area compare against 2493 * tune.t_minarmem, which is the minimum available real memory that we 2494 * can have in the system. However, this is generally fixed at 25 pages 2495 * which is so low that it's useless. In this comparison, we seek to 2496 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2497 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2498 * free) 2499 */ 2500 if (btop(vmem_size(heap_arena, VMEM_FREE)) < 2501 (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2)) 2502 return (1); 2503#endif 2504#else /* !sun */ 2505 if (kmem_used() > (kmem_size() * 3) / 4) 2506 return (1); 2507#endif /* sun */ 2508 2509#else 2510 if (spa_get_random(100) == 0) 2511 return (1); 2512#endif 2513 return (0); 2514} 2515 2516extern kmem_cache_t *zio_buf_cache[]; 2517extern kmem_cache_t *zio_data_buf_cache[]; 2518 2519static void 2520arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2521{ 2522 size_t i; 2523 kmem_cache_t *prev_cache = NULL; 2524 kmem_cache_t *prev_data_cache = NULL; 2525 2526#ifdef _KERNEL 2527 if (arc_meta_used >= arc_meta_limit) { 2528 /* 2529 * We are exceeding our meta-data cache limit. 2530 * Purge some DNLC entries to release holds on meta-data. 2531 */ 2532 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2533 } 2534#if defined(__i386) 2535 /* 2536 * Reclaim unused memory from all kmem caches. 2537 */ 2538 kmem_reap(); 2539#endif 2540#endif 2541 2542 /* 2543 * An aggressive reclamation will shrink the cache size as well as 2544 * reap free buffers from the arc kmem caches. 2545 */ 2546 if (strat == ARC_RECLAIM_AGGR) 2547 arc_shrink(); 2548 2549 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2550 if (zio_buf_cache[i] != prev_cache) { 2551 prev_cache = zio_buf_cache[i]; 2552 kmem_cache_reap_now(zio_buf_cache[i]); 2553 } 2554 if (zio_data_buf_cache[i] != prev_data_cache) { 2555 prev_data_cache = zio_data_buf_cache[i]; 2556 kmem_cache_reap_now(zio_data_buf_cache[i]); 2557 } 2558 } 2559 kmem_cache_reap_now(buf_cache); 2560 kmem_cache_reap_now(hdr_cache); 2561} 2562 2563static void 2564arc_reclaim_thread(void *dummy __unused) 2565{ 2566 clock_t growtime = 0; 2567 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2568 callb_cpr_t cpr; 2569 2570 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2571 2572 mutex_enter(&arc_reclaim_thr_lock); 2573 while (arc_thread_exit == 0) { 2574 if (arc_reclaim_needed()) { 2575 2576 if (arc_no_grow) { 2577 if (last_reclaim == ARC_RECLAIM_CONS) { 2578 last_reclaim = ARC_RECLAIM_AGGR; 2579 } else { 2580 last_reclaim = ARC_RECLAIM_CONS; 2581 } 2582 } else { 2583 arc_no_grow = TRUE; 2584 last_reclaim = ARC_RECLAIM_AGGR; 2585 membar_producer(); 2586 } 2587 2588 /* reset the growth delay for every reclaim */ 2589 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2590 2591 if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2592 /* 2593 * If needfree is TRUE our vm_lowmem hook 2594 * was called and in that case we must free some 2595 * memory, so switch to aggressive mode. 2596 */ 2597 arc_no_grow = TRUE; 2598 last_reclaim = ARC_RECLAIM_AGGR; 2599 } 2600 arc_kmem_reap_now(last_reclaim); 2601 arc_warm = B_TRUE; 2602 2603 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2604 arc_no_grow = FALSE; 2605 } 2606 2607 arc_adjust(); 2608 2609 if (arc_eviction_list != NULL) 2610 arc_do_user_evicts(); 2611 2612#ifdef _KERNEL 2613 if (needfree) { 2614 needfree = 0; 2615 wakeup(&needfree); 2616 } 2617#endif 2618 2619 /* block until needed, or one second, whichever is shorter */ 2620 CALLB_CPR_SAFE_BEGIN(&cpr); 2621 (void) cv_timedwait(&arc_reclaim_thr_cv, 2622 &arc_reclaim_thr_lock, hz); 2623 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2624 } 2625 2626 arc_thread_exit = 0; 2627 cv_broadcast(&arc_reclaim_thr_cv); 2628 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2629 thread_exit(); 2630} 2631 2632/* 2633 * Adapt arc info given the number of bytes we are trying to add and 2634 * the state that we are comming from. This function is only called 2635 * when we are adding new content to the cache. 2636 */ 2637static void 2638arc_adapt(int bytes, arc_state_t *state) 2639{ 2640 int mult; 2641 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2642 2643 if (state == arc_l2c_only) 2644 return; 2645 2646 ASSERT(bytes > 0); 2647 /* 2648 * Adapt the target size of the MRU list: 2649 * - if we just hit in the MRU ghost list, then increase 2650 * the target size of the MRU list. 2651 * - if we just hit in the MFU ghost list, then increase 2652 * the target size of the MFU list by decreasing the 2653 * target size of the MRU list. 2654 */ 2655 if (state == arc_mru_ghost) { 2656 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2657 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2658 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2659 2660 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2661 } else if (state == arc_mfu_ghost) { 2662 uint64_t delta; 2663 2664 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2665 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2666 mult = MIN(mult, 10); 2667 2668 delta = MIN(bytes * mult, arc_p); 2669 arc_p = MAX(arc_p_min, arc_p - delta); 2670 } 2671 ASSERT((int64_t)arc_p >= 0); 2672 2673 if (arc_reclaim_needed()) { 2674 cv_signal(&arc_reclaim_thr_cv); 2675 return; 2676 } 2677 2678 if (arc_no_grow) 2679 return; 2680 2681 if (arc_c >= arc_c_max) 2682 return; 2683 2684 /* 2685 * If we're within (2 * maxblocksize) bytes of the target 2686 * cache size, increment the target cache size 2687 */ 2688 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2689 atomic_add_64(&arc_c, (int64_t)bytes); 2690 if (arc_c > arc_c_max) 2691 arc_c = arc_c_max; 2692 else if (state == arc_anon) 2693 atomic_add_64(&arc_p, (int64_t)bytes); 2694 if (arc_p > arc_c) 2695 arc_p = arc_c; 2696 } 2697 ASSERT((int64_t)arc_p >= 0); 2698} 2699 2700/* 2701 * Check if the cache has reached its limits and eviction is required 2702 * prior to insert. 2703 */ 2704static int 2705arc_evict_needed(arc_buf_contents_t type) 2706{ 2707 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2708 return (1); 2709 2710#ifdef sun 2711#ifdef _KERNEL 2712 /* 2713 * If zio data pages are being allocated out of a separate heap segment, 2714 * then enforce that the size of available vmem for this area remains 2715 * above about 1/32nd free. 2716 */ 2717 if (type == ARC_BUFC_DATA && zio_arena != NULL && 2718 vmem_size(zio_arena, VMEM_FREE) < 2719 (vmem_size(zio_arena, VMEM_ALLOC) >> 5)) 2720 return (1); 2721#endif 2722#endif /* sun */ 2723 2724 if (arc_reclaim_needed()) 2725 return (1); 2726 2727 return (arc_size > arc_c); 2728} 2729 2730/* 2731 * The buffer, supplied as the first argument, needs a data block. 2732 * So, if we are at cache max, determine which cache should be victimized. 2733 * We have the following cases: 2734 * 2735 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2736 * In this situation if we're out of space, but the resident size of the MFU is 2737 * under the limit, victimize the MFU cache to satisfy this insertion request. 2738 * 2739 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2740 * Here, we've used up all of the available space for the MRU, so we need to 2741 * evict from our own cache instead. Evict from the set of resident MRU 2742 * entries. 2743 * 2744 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2745 * c minus p represents the MFU space in the cache, since p is the size of the 2746 * cache that is dedicated to the MRU. In this situation there's still space on 2747 * the MFU side, so the MRU side needs to be victimized. 2748 * 2749 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2750 * MFU's resident set is consuming more space than it has been allotted. In 2751 * this situation, we must victimize our own cache, the MFU, for this insertion. 2752 */ 2753static void 2754arc_get_data_buf(arc_buf_t *buf) 2755{ 2756 arc_state_t *state = buf->b_hdr->b_state; 2757 uint64_t size = buf->b_hdr->b_size; 2758 arc_buf_contents_t type = buf->b_hdr->b_type; 2759 2760 arc_adapt(size, state); 2761 2762 /* 2763 * We have not yet reached cache maximum size, 2764 * just allocate a new buffer. 2765 */ 2766 if (!arc_evict_needed(type)) { 2767 if (type == ARC_BUFC_METADATA) { 2768 buf->b_data = zio_buf_alloc(size); 2769 arc_space_consume(size, ARC_SPACE_DATA); 2770 } else { 2771 ASSERT(type == ARC_BUFC_DATA); 2772 buf->b_data = zio_data_buf_alloc(size); 2773 ARCSTAT_INCR(arcstat_data_size, size); 2774 atomic_add_64(&arc_size, size); 2775 } 2776 goto out; 2777 } 2778 2779 /* 2780 * If we are prefetching from the mfu ghost list, this buffer 2781 * will end up on the mru list; so steal space from there. 2782 */ 2783 if (state == arc_mfu_ghost) 2784 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2785 else if (state == arc_mru_ghost) 2786 state = arc_mru; 2787 2788 if (state == arc_mru || state == arc_anon) { 2789 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2790 state = (arc_mfu->arcs_lsize[type] >= size && 2791 arc_p > mru_used) ? arc_mfu : arc_mru; 2792 } else { 2793 /* MFU cases */ 2794 uint64_t mfu_space = arc_c - arc_p; 2795 state = (arc_mru->arcs_lsize[type] >= size && 2796 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2797 } 2798 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2799 if (type == ARC_BUFC_METADATA) { 2800 buf->b_data = zio_buf_alloc(size); 2801 arc_space_consume(size, ARC_SPACE_DATA); 2802 } else { 2803 ASSERT(type == ARC_BUFC_DATA); 2804 buf->b_data = zio_data_buf_alloc(size); 2805 ARCSTAT_INCR(arcstat_data_size, size); 2806 atomic_add_64(&arc_size, size); 2807 } 2808 ARCSTAT_BUMP(arcstat_recycle_miss); 2809 } 2810 ASSERT(buf->b_data != NULL); 2811out: 2812 /* 2813 * Update the state size. Note that ghost states have a 2814 * "ghost size" and so don't need to be updated. 2815 */ 2816 if (!GHOST_STATE(buf->b_hdr->b_state)) { 2817 arc_buf_hdr_t *hdr = buf->b_hdr; 2818 2819 atomic_add_64(&hdr->b_state->arcs_size, size); 2820 if (list_link_active(&hdr->b_arc_node)) { 2821 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2822 atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2823 } 2824 /* 2825 * If we are growing the cache, and we are adding anonymous 2826 * data, and we have outgrown arc_p, update arc_p 2827 */ 2828 if (arc_size < arc_c && hdr->b_state == arc_anon && 2829 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2830 arc_p = MIN(arc_c, arc_p + size); 2831 } 2832 ARCSTAT_BUMP(arcstat_allocated); 2833} 2834 2835/* 2836 * This routine is called whenever a buffer is accessed. 2837 * NOTE: the hash lock is dropped in this function. 2838 */ 2839static void 2840arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2841{ 2842 clock_t now; 2843 2844 ASSERT(MUTEX_HELD(hash_lock)); 2845 2846 if (buf->b_state == arc_anon) { 2847 /* 2848 * This buffer is not in the cache, and does not 2849 * appear in our "ghost" list. Add the new buffer 2850 * to the MRU state. 2851 */ 2852 2853 ASSERT(buf->b_arc_access == 0); 2854 buf->b_arc_access = ddi_get_lbolt(); 2855 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2856 arc_change_state(arc_mru, buf, hash_lock); 2857 2858 } else if (buf->b_state == arc_mru) { 2859 now = ddi_get_lbolt(); 2860 2861 /* 2862 * If this buffer is here because of a prefetch, then either: 2863 * - clear the flag if this is a "referencing" read 2864 * (any subsequent access will bump this into the MFU state). 2865 * or 2866 * - move the buffer to the head of the list if this is 2867 * another prefetch (to make it less likely to be evicted). 2868 */ 2869 if ((buf->b_flags & ARC_PREFETCH) != 0) { 2870 if (refcount_count(&buf->b_refcnt) == 0) { 2871 ASSERT(list_link_active(&buf->b_arc_node)); 2872 } else { 2873 buf->b_flags &= ~ARC_PREFETCH; 2874 ARCSTAT_BUMP(arcstat_mru_hits); 2875 } 2876 buf->b_arc_access = now; 2877 return; 2878 } 2879 2880 /* 2881 * This buffer has been "accessed" only once so far, 2882 * but it is still in the cache. Move it to the MFU 2883 * state. 2884 */ 2885 if (now > buf->b_arc_access + ARC_MINTIME) { 2886 /* 2887 * More than 125ms have passed since we 2888 * instantiated this buffer. Move it to the 2889 * most frequently used state. 2890 */ 2891 buf->b_arc_access = now; 2892 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2893 arc_change_state(arc_mfu, buf, hash_lock); 2894 } 2895 ARCSTAT_BUMP(arcstat_mru_hits); 2896 } else if (buf->b_state == arc_mru_ghost) { 2897 arc_state_t *new_state; 2898 /* 2899 * This buffer has been "accessed" recently, but 2900 * was evicted from the cache. Move it to the 2901 * MFU state. 2902 */ 2903 2904 if (buf->b_flags & ARC_PREFETCH) { 2905 new_state = arc_mru; 2906 if (refcount_count(&buf->b_refcnt) > 0) 2907 buf->b_flags &= ~ARC_PREFETCH; 2908 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2909 } else { 2910 new_state = arc_mfu; 2911 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2912 } 2913 2914 buf->b_arc_access = ddi_get_lbolt(); 2915 arc_change_state(new_state, buf, hash_lock); 2916 2917 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 2918 } else if (buf->b_state == arc_mfu) { 2919 /* 2920 * This buffer has been accessed more than once and is 2921 * still in the cache. Keep it in the MFU state. 2922 * 2923 * NOTE: an add_reference() that occurred when we did 2924 * the arc_read() will have kicked this off the list. 2925 * If it was a prefetch, we will explicitly move it to 2926 * the head of the list now. 2927 */ 2928 if ((buf->b_flags & ARC_PREFETCH) != 0) { 2929 ASSERT(refcount_count(&buf->b_refcnt) == 0); 2930 ASSERT(list_link_active(&buf->b_arc_node)); 2931 } 2932 ARCSTAT_BUMP(arcstat_mfu_hits); 2933 buf->b_arc_access = ddi_get_lbolt(); 2934 } else if (buf->b_state == arc_mfu_ghost) { 2935 arc_state_t *new_state = arc_mfu; 2936 /* 2937 * This buffer has been accessed more than once but has 2938 * been evicted from the cache. Move it back to the 2939 * MFU state. 2940 */ 2941 2942 if (buf->b_flags & ARC_PREFETCH) { 2943 /* 2944 * This is a prefetch access... 2945 * move this block back to the MRU state. 2946 */ 2947 ASSERT0(refcount_count(&buf->b_refcnt)); 2948 new_state = arc_mru; 2949 } 2950 2951 buf->b_arc_access = ddi_get_lbolt(); 2952 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2953 arc_change_state(new_state, buf, hash_lock); 2954 2955 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 2956 } else if (buf->b_state == arc_l2c_only) { 2957 /* 2958 * This buffer is on the 2nd Level ARC. 2959 */ 2960 2961 buf->b_arc_access = ddi_get_lbolt(); 2962 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2963 arc_change_state(arc_mfu, buf, hash_lock); 2964 } else { 2965 ASSERT(!"invalid arc state"); 2966 } 2967} 2968 2969/* a generic arc_done_func_t which you can use */ 2970/* ARGSUSED */ 2971void 2972arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 2973{ 2974 if (zio == NULL || zio->io_error == 0) 2975 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 2976 VERIFY(arc_buf_remove_ref(buf, arg)); 2977} 2978 2979/* a generic arc_done_func_t */ 2980void 2981arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 2982{ 2983 arc_buf_t **bufp = arg; 2984 if (zio && zio->io_error) { 2985 VERIFY(arc_buf_remove_ref(buf, arg)); 2986 *bufp = NULL; 2987 } else { 2988 *bufp = buf; 2989 ASSERT(buf->b_data); 2990 } 2991} 2992 2993static void 2994arc_read_done(zio_t *zio) 2995{ 2996 arc_buf_hdr_t *hdr; 2997 arc_buf_t *buf; 2998 arc_buf_t *abuf; /* buffer we're assigning to callback */ 2999 kmutex_t *hash_lock = NULL; 3000 arc_callback_t *callback_list, *acb; 3001 int freeable = FALSE; 3002 3003 buf = zio->io_private; 3004 hdr = buf->b_hdr; 3005 3006 /* 3007 * The hdr was inserted into hash-table and removed from lists 3008 * prior to starting I/O. We should find this header, since 3009 * it's in the hash table, and it should be legit since it's 3010 * not possible to evict it during the I/O. The only possible 3011 * reason for it not to be found is if we were freed during the 3012 * read. 3013 */ 3014 if (HDR_IN_HASH_TABLE(hdr)) { 3015 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3016 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3017 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3018 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3019 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3020 3021 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3022 &hash_lock); 3023 3024 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3025 hash_lock == NULL) || 3026 (found == hdr && 3027 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3028 (found == hdr && HDR_L2_READING(hdr))); 3029 } 3030 3031 hdr->b_flags &= ~ARC_L2_EVICTED; 3032 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 3033 hdr->b_flags &= ~ARC_L2CACHE; 3034 3035 /* byteswap if necessary */ 3036 callback_list = hdr->b_acb; 3037 ASSERT(callback_list != NULL); 3038 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3039 dmu_object_byteswap_t bswap = 3040 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3041 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3042 byteswap_uint64_array : 3043 dmu_ot_byteswap[bswap].ob_func; 3044 func(buf->b_data, hdr->b_size); 3045 } 3046 3047 arc_cksum_compute(buf, B_FALSE); 3048#ifdef illumos 3049 arc_buf_watch(buf); 3050#endif /* illumos */ 3051 3052 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3053 /* 3054 * Only call arc_access on anonymous buffers. This is because 3055 * if we've issued an I/O for an evicted buffer, we've already 3056 * called arc_access (to prevent any simultaneous readers from 3057 * getting confused). 3058 */ 3059 arc_access(hdr, hash_lock); 3060 } 3061 3062 /* create copies of the data buffer for the callers */ 3063 abuf = buf; 3064 for (acb = callback_list; acb; acb = acb->acb_next) { 3065 if (acb->acb_done) { 3066 if (abuf == NULL) { 3067 ARCSTAT_BUMP(arcstat_duplicate_reads); 3068 abuf = arc_buf_clone(buf); 3069 } 3070 acb->acb_buf = abuf; 3071 abuf = NULL; 3072 } 3073 } 3074 hdr->b_acb = NULL; 3075 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3076 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3077 if (abuf == buf) { 3078 ASSERT(buf->b_efunc == NULL); 3079 ASSERT(hdr->b_datacnt == 1); 3080 hdr->b_flags |= ARC_BUF_AVAILABLE; 3081 } 3082 3083 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3084 3085 if (zio->io_error != 0) { 3086 hdr->b_flags |= ARC_IO_ERROR; 3087 if (hdr->b_state != arc_anon) 3088 arc_change_state(arc_anon, hdr, hash_lock); 3089 if (HDR_IN_HASH_TABLE(hdr)) 3090 buf_hash_remove(hdr); 3091 freeable = refcount_is_zero(&hdr->b_refcnt); 3092 } 3093 3094 /* 3095 * Broadcast before we drop the hash_lock to avoid the possibility 3096 * that the hdr (and hence the cv) might be freed before we get to 3097 * the cv_broadcast(). 3098 */ 3099 cv_broadcast(&hdr->b_cv); 3100 3101 if (hash_lock) { 3102 mutex_exit(hash_lock); 3103 } else { 3104 /* 3105 * This block was freed while we waited for the read to 3106 * complete. It has been removed from the hash table and 3107 * moved to the anonymous state (so that it won't show up 3108 * in the cache). 3109 */ 3110 ASSERT3P(hdr->b_state, ==, arc_anon); 3111 freeable = refcount_is_zero(&hdr->b_refcnt); 3112 } 3113 3114 /* execute each callback and free its structure */ 3115 while ((acb = callback_list) != NULL) { 3116 if (acb->acb_done) 3117 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3118 3119 if (acb->acb_zio_dummy != NULL) { 3120 acb->acb_zio_dummy->io_error = zio->io_error; 3121 zio_nowait(acb->acb_zio_dummy); 3122 } 3123 3124 callback_list = acb->acb_next; 3125 kmem_free(acb, sizeof (arc_callback_t)); 3126 } 3127 3128 if (freeable) 3129 arc_hdr_destroy(hdr); 3130} 3131 3132/* 3133 * "Read" the block block at the specified DVA (in bp) via the 3134 * cache. If the block is found in the cache, invoke the provided 3135 * callback immediately and return. Note that the `zio' parameter 3136 * in the callback will be NULL in this case, since no IO was 3137 * required. If the block is not in the cache pass the read request 3138 * on to the spa with a substitute callback function, so that the 3139 * requested block will be added to the cache. 3140 * 3141 * If a read request arrives for a block that has a read in-progress, 3142 * either wait for the in-progress read to complete (and return the 3143 * results); or, if this is a read with a "done" func, add a record 3144 * to the read to invoke the "done" func when the read completes, 3145 * and return; or just return. 3146 * 3147 * arc_read_done() will invoke all the requested "done" functions 3148 * for readers of this block. 3149 */ 3150int 3151arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3152 void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags, 3153 const zbookmark_phys_t *zb) 3154{ 3155 arc_buf_hdr_t *hdr = NULL; 3156 arc_buf_t *buf = NULL; 3157 kmutex_t *hash_lock = NULL; 3158 zio_t *rzio; 3159 uint64_t guid = spa_load_guid(spa); 3160 3161 ASSERT(!BP_IS_EMBEDDED(bp) || 3162 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3163 3164top: 3165 if (!BP_IS_EMBEDDED(bp)) { 3166 /* 3167 * Embedded BP's have no DVA and require no I/O to "read". 3168 * Create an anonymous arc buf to back it. 3169 */ 3170 hdr = buf_hash_find(guid, bp, &hash_lock); 3171 } 3172 3173 if (hdr != NULL && hdr->b_datacnt > 0) { 3174 3175 *arc_flags |= ARC_CACHED; 3176 3177 if (HDR_IO_IN_PROGRESS(hdr)) { 3178 3179 if (*arc_flags & ARC_WAIT) { 3180 cv_wait(&hdr->b_cv, hash_lock); 3181 mutex_exit(hash_lock); 3182 goto top; 3183 } 3184 ASSERT(*arc_flags & ARC_NOWAIT); 3185 3186 if (done) { 3187 arc_callback_t *acb = NULL; 3188 3189 acb = kmem_zalloc(sizeof (arc_callback_t), 3190 KM_SLEEP); 3191 acb->acb_done = done; 3192 acb->acb_private = private; 3193 if (pio != NULL) 3194 acb->acb_zio_dummy = zio_null(pio, 3195 spa, NULL, NULL, NULL, zio_flags); 3196 3197 ASSERT(acb->acb_done != NULL); 3198 acb->acb_next = hdr->b_acb; 3199 hdr->b_acb = acb; 3200 add_reference(hdr, hash_lock, private); 3201 mutex_exit(hash_lock); 3202 return (0); 3203 } 3204 mutex_exit(hash_lock); 3205 return (0); 3206 } 3207 3208 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3209 3210 if (done) { 3211 add_reference(hdr, hash_lock, private); 3212 /* 3213 * If this block is already in use, create a new 3214 * copy of the data so that we will be guaranteed 3215 * that arc_release() will always succeed. 3216 */ 3217 buf = hdr->b_buf; 3218 ASSERT(buf); 3219 ASSERT(buf->b_data); 3220 if (HDR_BUF_AVAILABLE(hdr)) { 3221 ASSERT(buf->b_efunc == NULL); 3222 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3223 } else { 3224 buf = arc_buf_clone(buf); 3225 } 3226 3227 } else if (*arc_flags & ARC_PREFETCH && 3228 refcount_count(&hdr->b_refcnt) == 0) { 3229 hdr->b_flags |= ARC_PREFETCH; 3230 } 3231 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3232 arc_access(hdr, hash_lock); 3233 if (*arc_flags & ARC_L2CACHE) 3234 hdr->b_flags |= ARC_L2CACHE; 3235 if (*arc_flags & ARC_L2COMPRESS) 3236 hdr->b_flags |= ARC_L2COMPRESS; 3237 mutex_exit(hash_lock); 3238 ARCSTAT_BUMP(arcstat_hits); 3239 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3240 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3241 data, metadata, hits); 3242 3243 if (done) 3244 done(NULL, buf, private); 3245 } else { 3246 uint64_t size = BP_GET_LSIZE(bp); 3247 arc_callback_t *acb; 3248 vdev_t *vd = NULL; 3249 uint64_t addr = 0; 3250 boolean_t devw = B_FALSE; 3251 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3252 uint64_t b_asize = 0; 3253 3254 if (hdr == NULL) { 3255 /* this block is not in the cache */ 3256 arc_buf_hdr_t *exists = NULL; 3257 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3258 buf = arc_buf_alloc(spa, size, private, type); 3259 hdr = buf->b_hdr; 3260 if (!BP_IS_EMBEDDED(bp)) { 3261 hdr->b_dva = *BP_IDENTITY(bp); 3262 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3263 hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3264 exists = buf_hash_insert(hdr, &hash_lock); 3265 } 3266 if (exists != NULL) { 3267 /* somebody beat us to the hash insert */ 3268 mutex_exit(hash_lock); 3269 buf_discard_identity(hdr); 3270 (void) arc_buf_remove_ref(buf, private); 3271 goto top; /* restart the IO request */ 3272 } 3273 /* if this is a prefetch, we don't have a reference */ 3274 if (*arc_flags & ARC_PREFETCH) { 3275 (void) remove_reference(hdr, hash_lock, 3276 private); 3277 hdr->b_flags |= ARC_PREFETCH; 3278 } 3279 if (*arc_flags & ARC_L2CACHE) 3280 hdr->b_flags |= ARC_L2CACHE; 3281 if (*arc_flags & ARC_L2COMPRESS) 3282 hdr->b_flags |= ARC_L2COMPRESS; 3283 if (BP_GET_LEVEL(bp) > 0) 3284 hdr->b_flags |= ARC_INDIRECT; 3285 } else { 3286 /* this block is in the ghost cache */ 3287 ASSERT(GHOST_STATE(hdr->b_state)); 3288 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3289 ASSERT0(refcount_count(&hdr->b_refcnt)); 3290 ASSERT(hdr->b_buf == NULL); 3291 3292 /* if this is a prefetch, we don't have a reference */ 3293 if (*arc_flags & ARC_PREFETCH) 3294 hdr->b_flags |= ARC_PREFETCH; 3295 else 3296 add_reference(hdr, hash_lock, private); 3297 if (*arc_flags & ARC_L2CACHE) 3298 hdr->b_flags |= ARC_L2CACHE; 3299 if (*arc_flags & ARC_L2COMPRESS) 3300 hdr->b_flags |= ARC_L2COMPRESS; 3301 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3302 buf->b_hdr = hdr; 3303 buf->b_data = NULL; 3304 buf->b_efunc = NULL; 3305 buf->b_private = NULL; 3306 buf->b_next = NULL; 3307 hdr->b_buf = buf; 3308 ASSERT(hdr->b_datacnt == 0); 3309 hdr->b_datacnt = 1; 3310 arc_get_data_buf(buf); 3311 arc_access(hdr, hash_lock); 3312 } 3313 3314 ASSERT(!GHOST_STATE(hdr->b_state)); 3315 3316 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3317 acb->acb_done = done; 3318 acb->acb_private = private; 3319 3320 ASSERT(hdr->b_acb == NULL); 3321 hdr->b_acb = acb; 3322 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3323 3324 if (hdr->b_l2hdr != NULL && 3325 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3326 devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3327 addr = hdr->b_l2hdr->b_daddr; 3328 b_compress = hdr->b_l2hdr->b_compress; 3329 b_asize = hdr->b_l2hdr->b_asize; 3330 /* 3331 * Lock out device removal. 3332 */ 3333 if (vdev_is_dead(vd) || 3334 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3335 vd = NULL; 3336 } 3337 3338 if (hash_lock != NULL) 3339 mutex_exit(hash_lock); 3340 3341 /* 3342 * At this point, we have a level 1 cache miss. Try again in 3343 * L2ARC if possible. 3344 */ 3345 ASSERT3U(hdr->b_size, ==, size); 3346 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3347 uint64_t, size, zbookmark_phys_t *, zb); 3348 ARCSTAT_BUMP(arcstat_misses); 3349 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3350 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3351 data, metadata, misses); 3352#ifdef _KERNEL 3353 curthread->td_ru.ru_inblock++; 3354#endif 3355 3356 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3357 /* 3358 * Read from the L2ARC if the following are true: 3359 * 1. The L2ARC vdev was previously cached. 3360 * 2. This buffer still has L2ARC metadata. 3361 * 3. This buffer isn't currently writing to the L2ARC. 3362 * 4. The L2ARC entry wasn't evicted, which may 3363 * also have invalidated the vdev. 3364 * 5. This isn't prefetch and l2arc_noprefetch is set. 3365 */ 3366 if (hdr->b_l2hdr != NULL && 3367 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3368 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3369 l2arc_read_callback_t *cb; 3370 3371 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3372 ARCSTAT_BUMP(arcstat_l2_hits); 3373 3374 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3375 KM_SLEEP); 3376 cb->l2rcb_buf = buf; 3377 cb->l2rcb_spa = spa; 3378 cb->l2rcb_bp = *bp; 3379 cb->l2rcb_zb = *zb; 3380 cb->l2rcb_flags = zio_flags; 3381 cb->l2rcb_compress = b_compress; 3382 3383 ASSERT(addr >= VDEV_LABEL_START_SIZE && 3384 addr + size < vd->vdev_psize - 3385 VDEV_LABEL_END_SIZE); 3386 3387 /* 3388 * l2arc read. The SCL_L2ARC lock will be 3389 * released by l2arc_read_done(). 3390 * Issue a null zio if the underlying buffer 3391 * was squashed to zero size by compression. 3392 */ 3393 if (b_compress == ZIO_COMPRESS_EMPTY) { 3394 rzio = zio_null(pio, spa, vd, 3395 l2arc_read_done, cb, 3396 zio_flags | ZIO_FLAG_DONT_CACHE | 3397 ZIO_FLAG_CANFAIL | 3398 ZIO_FLAG_DONT_PROPAGATE | 3399 ZIO_FLAG_DONT_RETRY); 3400 } else { 3401 rzio = zio_read_phys(pio, vd, addr, 3402 b_asize, buf->b_data, 3403 ZIO_CHECKSUM_OFF, 3404 l2arc_read_done, cb, priority, 3405 zio_flags | ZIO_FLAG_DONT_CACHE | 3406 ZIO_FLAG_CANFAIL | 3407 ZIO_FLAG_DONT_PROPAGATE | 3408 ZIO_FLAG_DONT_RETRY, B_FALSE); 3409 } 3410 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3411 zio_t *, rzio); 3412 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3413 3414 if (*arc_flags & ARC_NOWAIT) { 3415 zio_nowait(rzio); 3416 return (0); 3417 } 3418 3419 ASSERT(*arc_flags & ARC_WAIT); 3420 if (zio_wait(rzio) == 0) 3421 return (0); 3422 3423 /* l2arc read error; goto zio_read() */ 3424 } else { 3425 DTRACE_PROBE1(l2arc__miss, 3426 arc_buf_hdr_t *, hdr); 3427 ARCSTAT_BUMP(arcstat_l2_misses); 3428 if (HDR_L2_WRITING(hdr)) 3429 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3430 spa_config_exit(spa, SCL_L2ARC, vd); 3431 } 3432 } else { 3433 if (vd != NULL) 3434 spa_config_exit(spa, SCL_L2ARC, vd); 3435 if (l2arc_ndev != 0) { 3436 DTRACE_PROBE1(l2arc__miss, 3437 arc_buf_hdr_t *, hdr); 3438 ARCSTAT_BUMP(arcstat_l2_misses); 3439 } 3440 } 3441 3442 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3443 arc_read_done, buf, priority, zio_flags, zb); 3444 3445 if (*arc_flags & ARC_WAIT) 3446 return (zio_wait(rzio)); 3447 3448 ASSERT(*arc_flags & ARC_NOWAIT); 3449 zio_nowait(rzio); 3450 } 3451 return (0); 3452} 3453 3454void 3455arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3456{ 3457 ASSERT(buf->b_hdr != NULL); 3458 ASSERT(buf->b_hdr->b_state != arc_anon); 3459 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3460 ASSERT(buf->b_efunc == NULL); 3461 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3462 3463 buf->b_efunc = func; 3464 buf->b_private = private; 3465} 3466 3467/* 3468 * Notify the arc that a block was freed, and thus will never be used again. 3469 */ 3470void 3471arc_freed(spa_t *spa, const blkptr_t *bp) 3472{ 3473 arc_buf_hdr_t *hdr; 3474 kmutex_t *hash_lock; 3475 uint64_t guid = spa_load_guid(spa); 3476 3477 ASSERT(!BP_IS_EMBEDDED(bp)); 3478 3479 hdr = buf_hash_find(guid, bp, &hash_lock); 3480 if (hdr == NULL) 3481 return; 3482 if (HDR_BUF_AVAILABLE(hdr)) { 3483 arc_buf_t *buf = hdr->b_buf; 3484 add_reference(hdr, hash_lock, FTAG); 3485 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3486 mutex_exit(hash_lock); 3487 3488 arc_release(buf, FTAG); 3489 (void) arc_buf_remove_ref(buf, FTAG); 3490 } else { 3491 mutex_exit(hash_lock); 3492 } 3493 3494} 3495 3496/* 3497 * Clear the user eviction callback set by arc_set_callback(), first calling 3498 * it if it exists. Because the presence of a callback keeps an arc_buf cached 3499 * clearing the callback may result in the arc_buf being destroyed. However, 3500 * it will not result in the *last* arc_buf being destroyed, hence the data 3501 * will remain cached in the ARC. We make a copy of the arc buffer here so 3502 * that we can process the callback without holding any locks. 3503 * 3504 * It's possible that the callback is already in the process of being cleared 3505 * by another thread. In this case we can not clear the callback. 3506 * 3507 * Returns B_TRUE if the callback was successfully called and cleared. 3508 */ 3509boolean_t 3510arc_clear_callback(arc_buf_t *buf) 3511{ 3512 arc_buf_hdr_t *hdr; 3513 kmutex_t *hash_lock; 3514 arc_evict_func_t *efunc = buf->b_efunc; 3515 void *private = buf->b_private; 3516 list_t *list, *evicted_list; 3517 kmutex_t *lock, *evicted_lock; 3518 3519 mutex_enter(&buf->b_evict_lock); 3520 hdr = buf->b_hdr; 3521 if (hdr == NULL) { 3522 /* 3523 * We are in arc_do_user_evicts(). 3524 */ 3525 ASSERT(buf->b_data == NULL); 3526 mutex_exit(&buf->b_evict_lock); 3527 return (B_FALSE); 3528 } else if (buf->b_data == NULL) { 3529 /* 3530 * We are on the eviction list; process this buffer now 3531 * but let arc_do_user_evicts() do the reaping. 3532 */ 3533 buf->b_efunc = NULL; 3534 mutex_exit(&buf->b_evict_lock); 3535 VERIFY0(efunc(private)); 3536 return (B_TRUE); 3537 } 3538 hash_lock = HDR_LOCK(hdr); 3539 mutex_enter(hash_lock); 3540 hdr = buf->b_hdr; 3541 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3542 3543 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3544 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3545 3546 buf->b_efunc = NULL; 3547 buf->b_private = NULL; 3548 3549 if (hdr->b_datacnt > 1) { 3550 mutex_exit(&buf->b_evict_lock); 3551 arc_buf_destroy(buf, FALSE, TRUE); 3552 } else { 3553 ASSERT(buf == hdr->b_buf); 3554 hdr->b_flags |= ARC_BUF_AVAILABLE; 3555 mutex_exit(&buf->b_evict_lock); 3556 } 3557 3558 mutex_exit(hash_lock); 3559 VERIFY0(efunc(private)); 3560 return (B_TRUE); 3561} 3562 3563/* 3564 * Release this buffer from the cache, making it an anonymous buffer. This 3565 * must be done after a read and prior to modifying the buffer contents. 3566 * If the buffer has more than one reference, we must make 3567 * a new hdr for the buffer. 3568 */ 3569void 3570arc_release(arc_buf_t *buf, void *tag) 3571{ 3572 arc_buf_hdr_t *hdr; 3573 kmutex_t *hash_lock = NULL; 3574 l2arc_buf_hdr_t *l2hdr; 3575 uint64_t buf_size; 3576 3577 /* 3578 * It would be nice to assert that if it's DMU metadata (level > 3579 * 0 || it's the dnode file), then it must be syncing context. 3580 * But we don't know that information at this level. 3581 */ 3582 3583 mutex_enter(&buf->b_evict_lock); 3584 hdr = buf->b_hdr; 3585 3586 /* this buffer is not on any list */ 3587 ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3588 3589 if (hdr->b_state == arc_anon) { 3590 /* this buffer is already released */ 3591 ASSERT(buf->b_efunc == NULL); 3592 } else { 3593 hash_lock = HDR_LOCK(hdr); 3594 mutex_enter(hash_lock); 3595 hdr = buf->b_hdr; 3596 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3597 } 3598 3599 l2hdr = hdr->b_l2hdr; 3600 if (l2hdr) { 3601 mutex_enter(&l2arc_buflist_mtx); 3602 hdr->b_l2hdr = NULL; 3603 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3604 } 3605 buf_size = hdr->b_size; 3606 3607 /* 3608 * Do we have more than one buf? 3609 */ 3610 if (hdr->b_datacnt > 1) { 3611 arc_buf_hdr_t *nhdr; 3612 arc_buf_t **bufp; 3613 uint64_t blksz = hdr->b_size; 3614 uint64_t spa = hdr->b_spa; 3615 arc_buf_contents_t type = hdr->b_type; 3616 uint32_t flags = hdr->b_flags; 3617 3618 ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3619 /* 3620 * Pull the data off of this hdr and attach it to 3621 * a new anonymous hdr. 3622 */ 3623 (void) remove_reference(hdr, hash_lock, tag); 3624 bufp = &hdr->b_buf; 3625 while (*bufp != buf) 3626 bufp = &(*bufp)->b_next; 3627 *bufp = buf->b_next; 3628 buf->b_next = NULL; 3629 3630 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3631 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3632 if (refcount_is_zero(&hdr->b_refcnt)) { 3633 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3634 ASSERT3U(*size, >=, hdr->b_size); 3635 atomic_add_64(size, -hdr->b_size); 3636 } 3637 3638 /* 3639 * We're releasing a duplicate user data buffer, update 3640 * our statistics accordingly. 3641 */ 3642 if (hdr->b_type == ARC_BUFC_DATA) { 3643 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3644 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3645 -hdr->b_size); 3646 } 3647 hdr->b_datacnt -= 1; 3648 arc_cksum_verify(buf); 3649#ifdef illumos 3650 arc_buf_unwatch(buf); 3651#endif /* illumos */ 3652 3653 mutex_exit(hash_lock); 3654 3655 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3656 nhdr->b_size = blksz; 3657 nhdr->b_spa = spa; 3658 nhdr->b_type = type; 3659 nhdr->b_buf = buf; 3660 nhdr->b_state = arc_anon; 3661 nhdr->b_arc_access = 0; 3662 nhdr->b_flags = flags & ARC_L2_WRITING; 3663 nhdr->b_l2hdr = NULL; 3664 nhdr->b_datacnt = 1; 3665 nhdr->b_freeze_cksum = NULL; 3666 (void) refcount_add(&nhdr->b_refcnt, tag); 3667 buf->b_hdr = nhdr; 3668 mutex_exit(&buf->b_evict_lock); 3669 atomic_add_64(&arc_anon->arcs_size, blksz); 3670 } else { 3671 mutex_exit(&buf->b_evict_lock); 3672 ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3673 ASSERT(!list_link_active(&hdr->b_arc_node)); 3674 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3675 if (hdr->b_state != arc_anon) 3676 arc_change_state(arc_anon, hdr, hash_lock); 3677 hdr->b_arc_access = 0; 3678 if (hash_lock) 3679 mutex_exit(hash_lock); 3680 3681 buf_discard_identity(hdr); 3682 arc_buf_thaw(buf); 3683 } 3684 buf->b_efunc = NULL; 3685 buf->b_private = NULL; 3686 3687 if (l2hdr) { 3688 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3689 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 3690 -l2hdr->b_asize, 0, 0); 3691 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3692 hdr->b_size, 0); 3693 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3694 ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3695 mutex_exit(&l2arc_buflist_mtx); 3696 } 3697} 3698 3699int 3700arc_released(arc_buf_t *buf) 3701{ 3702 int released; 3703 3704 mutex_enter(&buf->b_evict_lock); 3705 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3706 mutex_exit(&buf->b_evict_lock); 3707 return (released); 3708} 3709 3710#ifdef ZFS_DEBUG 3711int 3712arc_referenced(arc_buf_t *buf) 3713{ 3714 int referenced; 3715 3716 mutex_enter(&buf->b_evict_lock); 3717 referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3718 mutex_exit(&buf->b_evict_lock); 3719 return (referenced); 3720} 3721#endif 3722 3723static void 3724arc_write_ready(zio_t *zio) 3725{ 3726 arc_write_callback_t *callback = zio->io_private; 3727 arc_buf_t *buf = callback->awcb_buf; 3728 arc_buf_hdr_t *hdr = buf->b_hdr; 3729 3730 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3731 callback->awcb_ready(zio, buf, callback->awcb_private); 3732 3733 /* 3734 * If the IO is already in progress, then this is a re-write 3735 * attempt, so we need to thaw and re-compute the cksum. 3736 * It is the responsibility of the callback to handle the 3737 * accounting for any re-write attempt. 3738 */ 3739 if (HDR_IO_IN_PROGRESS(hdr)) { 3740 mutex_enter(&hdr->b_freeze_lock); 3741 if (hdr->b_freeze_cksum != NULL) { 3742 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3743 hdr->b_freeze_cksum = NULL; 3744 } 3745 mutex_exit(&hdr->b_freeze_lock); 3746 } 3747 arc_cksum_compute(buf, B_FALSE); 3748 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3749} 3750 3751/* 3752 * The SPA calls this callback for each physical write that happens on behalf 3753 * of a logical write. See the comment in dbuf_write_physdone() for details. 3754 */ 3755static void 3756arc_write_physdone(zio_t *zio) 3757{ 3758 arc_write_callback_t *cb = zio->io_private; 3759 if (cb->awcb_physdone != NULL) 3760 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3761} 3762 3763static void 3764arc_write_done(zio_t *zio) 3765{ 3766 arc_write_callback_t *callback = zio->io_private; 3767 arc_buf_t *buf = callback->awcb_buf; 3768 arc_buf_hdr_t *hdr = buf->b_hdr; 3769 3770 ASSERT(hdr->b_acb == NULL); 3771 3772 if (zio->io_error == 0) { 3773 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 3774 buf_discard_identity(hdr); 3775 } else { 3776 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3777 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3778 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3779 } 3780 } else { 3781 ASSERT(BUF_EMPTY(hdr)); 3782 } 3783 3784 /* 3785 * If the block to be written was all-zero or compressed enough to be 3786 * embedded in the BP, no write was performed so there will be no 3787 * dva/birth/checksum. The buffer must therefore remain anonymous 3788 * (and uncached). 3789 */ 3790 if (!BUF_EMPTY(hdr)) { 3791 arc_buf_hdr_t *exists; 3792 kmutex_t *hash_lock; 3793 3794 ASSERT(zio->io_error == 0); 3795 3796 arc_cksum_verify(buf); 3797 3798 exists = buf_hash_insert(hdr, &hash_lock); 3799 if (exists) { 3800 /* 3801 * This can only happen if we overwrite for 3802 * sync-to-convergence, because we remove 3803 * buffers from the hash table when we arc_free(). 3804 */ 3805 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3806 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3807 panic("bad overwrite, hdr=%p exists=%p", 3808 (void *)hdr, (void *)exists); 3809 ASSERT(refcount_is_zero(&exists->b_refcnt)); 3810 arc_change_state(arc_anon, exists, hash_lock); 3811 mutex_exit(hash_lock); 3812 arc_hdr_destroy(exists); 3813 exists = buf_hash_insert(hdr, &hash_lock); 3814 ASSERT3P(exists, ==, NULL); 3815 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3816 /* nopwrite */ 3817 ASSERT(zio->io_prop.zp_nopwrite); 3818 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3819 panic("bad nopwrite, hdr=%p exists=%p", 3820 (void *)hdr, (void *)exists); 3821 } else { 3822 /* Dedup */ 3823 ASSERT(hdr->b_datacnt == 1); 3824 ASSERT(hdr->b_state == arc_anon); 3825 ASSERT(BP_GET_DEDUP(zio->io_bp)); 3826 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3827 } 3828 } 3829 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3830 /* if it's not anon, we are doing a scrub */ 3831 if (!exists && hdr->b_state == arc_anon) 3832 arc_access(hdr, hash_lock); 3833 mutex_exit(hash_lock); 3834 } else { 3835 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3836 } 3837 3838 ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3839 callback->awcb_done(zio, buf, callback->awcb_private); 3840 3841 kmem_free(callback, sizeof (arc_write_callback_t)); 3842} 3843 3844zio_t * 3845arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 3846 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 3847 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 3848 arc_done_func_t *done, void *private, zio_priority_t priority, 3849 int zio_flags, const zbookmark_phys_t *zb) 3850{ 3851 arc_buf_hdr_t *hdr = buf->b_hdr; 3852 arc_write_callback_t *callback; 3853 zio_t *zio; 3854 3855 ASSERT(ready != NULL); 3856 ASSERT(done != NULL); 3857 ASSERT(!HDR_IO_ERROR(hdr)); 3858 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3859 ASSERT(hdr->b_acb == NULL); 3860 if (l2arc) 3861 hdr->b_flags |= ARC_L2CACHE; 3862 if (l2arc_compress) 3863 hdr->b_flags |= ARC_L2COMPRESS; 3864 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 3865 callback->awcb_ready = ready; 3866 callback->awcb_physdone = physdone; 3867 callback->awcb_done = done; 3868 callback->awcb_private = private; 3869 callback->awcb_buf = buf; 3870 3871 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 3872 arc_write_ready, arc_write_physdone, arc_write_done, callback, 3873 priority, zio_flags, zb); 3874 3875 return (zio); 3876} 3877 3878static int 3879arc_memory_throttle(uint64_t reserve, uint64_t txg) 3880{ 3881#ifdef _KERNEL 3882 uint64_t available_memory = 3883 ptoa((uintmax_t)cnt.v_free_count + cnt.v_cache_count); 3884 static uint64_t page_load = 0; 3885 static uint64_t last_txg = 0; 3886 3887#ifdef sun 3888#if defined(__i386) 3889 available_memory = 3890 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 3891#endif 3892#endif /* sun */ 3893 3894 if (cnt.v_free_count + cnt.v_cache_count > 3895 (uint64_t)physmem * arc_lotsfree_percent / 100) 3896 return (0); 3897 3898 if (txg > last_txg) { 3899 last_txg = txg; 3900 page_load = 0; 3901 } 3902 /* 3903 * If we are in pageout, we know that memory is already tight, 3904 * the arc is already going to be evicting, so we just want to 3905 * continue to let page writes occur as quickly as possible. 3906 */ 3907 if (curproc == pageproc) { 3908 if (page_load > available_memory / 4) 3909 return (SET_ERROR(ERESTART)); 3910 /* Note: reserve is inflated, so we deflate */ 3911 page_load += reserve / 8; 3912 return (0); 3913 } else if (page_load > 0 && arc_reclaim_needed()) { 3914 /* memory is low, delay before restarting */ 3915 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3916 return (SET_ERROR(EAGAIN)); 3917 } 3918 page_load = 0; 3919#endif 3920 return (0); 3921} 3922 3923void 3924arc_tempreserve_clear(uint64_t reserve) 3925{ 3926 atomic_add_64(&arc_tempreserve, -reserve); 3927 ASSERT((int64_t)arc_tempreserve >= 0); 3928} 3929 3930int 3931arc_tempreserve_space(uint64_t reserve, uint64_t txg) 3932{ 3933 int error; 3934 uint64_t anon_size; 3935 3936 if (reserve > arc_c/4 && !arc_no_grow) 3937 arc_c = MIN(arc_c_max, reserve * 4); 3938 if (reserve > arc_c) 3939 return (SET_ERROR(ENOMEM)); 3940 3941 /* 3942 * Don't count loaned bufs as in flight dirty data to prevent long 3943 * network delays from blocking transactions that are ready to be 3944 * assigned to a txg. 3945 */ 3946 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 3947 3948 /* 3949 * Writes will, almost always, require additional memory allocations 3950 * in order to compress/encrypt/etc the data. We therefore need to 3951 * make sure that there is sufficient available memory for this. 3952 */ 3953 error = arc_memory_throttle(reserve, txg); 3954 if (error != 0) 3955 return (error); 3956 3957 /* 3958 * Throttle writes when the amount of dirty data in the cache 3959 * gets too large. We try to keep the cache less than half full 3960 * of dirty blocks so that our sync times don't grow too large. 3961 * Note: if two requests come in concurrently, we might let them 3962 * both succeed, when one of them should fail. Not a huge deal. 3963 */ 3964 3965 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 3966 anon_size > arc_c / 4) { 3967 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 3968 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 3969 arc_tempreserve>>10, 3970 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 3971 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 3972 reserve>>10, arc_c>>10); 3973 return (SET_ERROR(ERESTART)); 3974 } 3975 atomic_add_64(&arc_tempreserve, reserve); 3976 return (0); 3977} 3978 3979static kmutex_t arc_lowmem_lock; 3980#ifdef _KERNEL 3981static eventhandler_tag arc_event_lowmem = NULL; 3982 3983static void 3984arc_lowmem(void *arg __unused, int howto __unused) 3985{ 3986 3987 /* Serialize access via arc_lowmem_lock. */ 3988 mutex_enter(&arc_lowmem_lock); 3989 mutex_enter(&arc_reclaim_thr_lock); 3990 needfree = 1; 3991 cv_signal(&arc_reclaim_thr_cv); 3992 3993 /* 3994 * It is unsafe to block here in arbitrary threads, because we can come 3995 * here from ARC itself and may hold ARC locks and thus risk a deadlock 3996 * with ARC reclaim thread. 3997 */ 3998 if (curproc == pageproc) { 3999 while (needfree) 4000 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4001 } 4002 mutex_exit(&arc_reclaim_thr_lock); 4003 mutex_exit(&arc_lowmem_lock); 4004} 4005#endif 4006 4007void 4008arc_init(void) 4009{ 4010 int i, prefetch_tunable_set = 0; 4011 4012 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4013 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4014 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4015 4016 /* Convert seconds to clock ticks */ 4017 arc_min_prefetch_lifespan = 1 * hz; 4018 4019 /* Start out with 1/8 of all memory */ 4020 arc_c = kmem_size() / 8; 4021 4022#ifdef sun 4023#ifdef _KERNEL 4024 /* 4025 * On architectures where the physical memory can be larger 4026 * than the addressable space (intel in 32-bit mode), we may 4027 * need to limit the cache to 1/8 of VM size. 4028 */ 4029 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4030#endif 4031#endif /* sun */ 4032 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4033 arc_c_min = MAX(arc_c / 4, 64<<18); 4034 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4035 if (arc_c * 8 >= 1<<30) 4036 arc_c_max = (arc_c * 8) - (1<<30); 4037 else 4038 arc_c_max = arc_c_min; 4039 arc_c_max = MAX(arc_c * 5, arc_c_max); 4040 4041#ifdef _KERNEL 4042 /* 4043 * Allow the tunables to override our calculations if they are 4044 * reasonable (ie. over 16MB) 4045 */ 4046 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size()) 4047 arc_c_max = zfs_arc_max; 4048 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max) 4049 arc_c_min = zfs_arc_min; 4050#endif 4051 4052 arc_c = arc_c_max; 4053 arc_p = (arc_c >> 1); 4054 4055 /* limit meta-data to 1/4 of the arc capacity */ 4056 arc_meta_limit = arc_c_max / 4; 4057 4058 /* Allow the tunable to override if it is reasonable */ 4059 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4060 arc_meta_limit = zfs_arc_meta_limit; 4061 4062 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4063 arc_c_min = arc_meta_limit / 2; 4064 4065 if (zfs_arc_grow_retry > 0) 4066 arc_grow_retry = zfs_arc_grow_retry; 4067 4068 if (zfs_arc_shrink_shift > 0) 4069 arc_shrink_shift = zfs_arc_shrink_shift; 4070 4071 if (zfs_arc_p_min_shift > 0) 4072 arc_p_min_shift = zfs_arc_p_min_shift; 4073 4074 /* if kmem_flags are set, lets try to use less memory */ 4075 if (kmem_debugging()) 4076 arc_c = arc_c / 2; 4077 if (arc_c < arc_c_min) 4078 arc_c = arc_c_min; 4079 4080 zfs_arc_min = arc_c_min; 4081 zfs_arc_max = arc_c_max; 4082 4083 arc_anon = &ARC_anon; 4084 arc_mru = &ARC_mru; 4085 arc_mru_ghost = &ARC_mru_ghost; 4086 arc_mfu = &ARC_mfu; 4087 arc_mfu_ghost = &ARC_mfu_ghost; 4088 arc_l2c_only = &ARC_l2c_only; 4089 arc_size = 0; 4090 4091 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4092 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4093 NULL, MUTEX_DEFAULT, NULL); 4094 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4095 NULL, MUTEX_DEFAULT, NULL); 4096 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4097 NULL, MUTEX_DEFAULT, NULL); 4098 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4099 NULL, MUTEX_DEFAULT, NULL); 4100 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4101 NULL, MUTEX_DEFAULT, NULL); 4102 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4103 NULL, MUTEX_DEFAULT, NULL); 4104 4105 list_create(&arc_mru->arcs_lists[i], 4106 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4107 list_create(&arc_mru_ghost->arcs_lists[i], 4108 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4109 list_create(&arc_mfu->arcs_lists[i], 4110 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4111 list_create(&arc_mfu_ghost->arcs_lists[i], 4112 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4113 list_create(&arc_mfu_ghost->arcs_lists[i], 4114 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4115 list_create(&arc_l2c_only->arcs_lists[i], 4116 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4117 } 4118 4119 buf_init(); 4120 4121 arc_thread_exit = 0; 4122 arc_eviction_list = NULL; 4123 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4124 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4125 4126 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4127 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4128 4129 if (arc_ksp != NULL) { 4130 arc_ksp->ks_data = &arc_stats; 4131 kstat_install(arc_ksp); 4132 } 4133 4134 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4135 TS_RUN, minclsyspri); 4136 4137#ifdef _KERNEL 4138 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4139 EVENTHANDLER_PRI_FIRST); 4140#endif 4141 4142 arc_dead = FALSE; 4143 arc_warm = B_FALSE; 4144 4145 /* 4146 * Calculate maximum amount of dirty data per pool. 4147 * 4148 * If it has been set by /etc/system, take that. 4149 * Otherwise, use a percentage of physical memory defined by 4150 * zfs_dirty_data_max_percent (default 10%) with a cap at 4151 * zfs_dirty_data_max_max (default 4GB). 4152 */ 4153 if (zfs_dirty_data_max == 0) { 4154 zfs_dirty_data_max = ptob(physmem) * 4155 zfs_dirty_data_max_percent / 100; 4156 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4157 zfs_dirty_data_max_max); 4158 } 4159 4160#ifdef _KERNEL 4161 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4162 prefetch_tunable_set = 1; 4163 4164#ifdef __i386__ 4165 if (prefetch_tunable_set == 0) { 4166 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4167 "-- to enable,\n"); 4168 printf(" add \"vfs.zfs.prefetch_disable=0\" " 4169 "to /boot/loader.conf.\n"); 4170 zfs_prefetch_disable = 1; 4171 } 4172#else 4173 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4174 prefetch_tunable_set == 0) { 4175 printf("ZFS NOTICE: Prefetch is disabled by default if less " 4176 "than 4GB of RAM is present;\n" 4177 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4178 "to /boot/loader.conf.\n"); 4179 zfs_prefetch_disable = 1; 4180 } 4181#endif 4182 /* Warn about ZFS memory and address space requirements. */ 4183 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4184 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4185 "expect unstable behavior.\n"); 4186 } 4187 if (kmem_size() < 512 * (1 << 20)) { 4188 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4189 "expect unstable behavior.\n"); 4190 printf(" Consider tuning vm.kmem_size and " 4191 "vm.kmem_size_max\n"); 4192 printf(" in /boot/loader.conf.\n"); 4193 } 4194#endif 4195} 4196 4197void 4198arc_fini(void) 4199{ 4200 int i; 4201 4202 mutex_enter(&arc_reclaim_thr_lock); 4203 arc_thread_exit = 1; 4204 cv_signal(&arc_reclaim_thr_cv); 4205 while (arc_thread_exit != 0) 4206 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4207 mutex_exit(&arc_reclaim_thr_lock); 4208 4209 arc_flush(NULL); 4210 4211 arc_dead = TRUE; 4212 4213 if (arc_ksp != NULL) { 4214 kstat_delete(arc_ksp); 4215 arc_ksp = NULL; 4216 } 4217 4218 mutex_destroy(&arc_eviction_mtx); 4219 mutex_destroy(&arc_reclaim_thr_lock); 4220 cv_destroy(&arc_reclaim_thr_cv); 4221 4222 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4223 list_destroy(&arc_mru->arcs_lists[i]); 4224 list_destroy(&arc_mru_ghost->arcs_lists[i]); 4225 list_destroy(&arc_mfu->arcs_lists[i]); 4226 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4227 list_destroy(&arc_l2c_only->arcs_lists[i]); 4228 4229 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4230 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4231 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4232 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4233 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4234 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4235 } 4236 4237 buf_fini(); 4238 4239 ASSERT(arc_loaned_bytes == 0); 4240 4241 mutex_destroy(&arc_lowmem_lock); 4242#ifdef _KERNEL 4243 if (arc_event_lowmem != NULL) 4244 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4245#endif 4246} 4247 4248/* 4249 * Level 2 ARC 4250 * 4251 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4252 * It uses dedicated storage devices to hold cached data, which are populated 4253 * using large infrequent writes. The main role of this cache is to boost 4254 * the performance of random read workloads. The intended L2ARC devices 4255 * include short-stroked disks, solid state disks, and other media with 4256 * substantially faster read latency than disk. 4257 * 4258 * +-----------------------+ 4259 * | ARC | 4260 * +-----------------------+ 4261 * | ^ ^ 4262 * | | | 4263 * l2arc_feed_thread() arc_read() 4264 * | | | 4265 * | l2arc read | 4266 * V | | 4267 * +---------------+ | 4268 * | L2ARC | | 4269 * +---------------+ | 4270 * | ^ | 4271 * l2arc_write() | | 4272 * | | | 4273 * V | | 4274 * +-------+ +-------+ 4275 * | vdev | | vdev | 4276 * | cache | | cache | 4277 * +-------+ +-------+ 4278 * +=========+ .-----. 4279 * : L2ARC : |-_____-| 4280 * : devices : | Disks | 4281 * +=========+ `-_____-' 4282 * 4283 * Read requests are satisfied from the following sources, in order: 4284 * 4285 * 1) ARC 4286 * 2) vdev cache of L2ARC devices 4287 * 3) L2ARC devices 4288 * 4) vdev cache of disks 4289 * 5) disks 4290 * 4291 * Some L2ARC device types exhibit extremely slow write performance. 4292 * To accommodate for this there are some significant differences between 4293 * the L2ARC and traditional cache design: 4294 * 4295 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4296 * the ARC behave as usual, freeing buffers and placing headers on ghost 4297 * lists. The ARC does not send buffers to the L2ARC during eviction as 4298 * this would add inflated write latencies for all ARC memory pressure. 4299 * 4300 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4301 * It does this by periodically scanning buffers from the eviction-end of 4302 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4303 * not already there. It scans until a headroom of buffers is satisfied, 4304 * which itself is a buffer for ARC eviction. If a compressible buffer is 4305 * found during scanning and selected for writing to an L2ARC device, we 4306 * temporarily boost scanning headroom during the next scan cycle to make 4307 * sure we adapt to compression effects (which might significantly reduce 4308 * the data volume we write to L2ARC). The thread that does this is 4309 * l2arc_feed_thread(), illustrated below; example sizes are included to 4310 * provide a better sense of ratio than this diagram: 4311 * 4312 * head --> tail 4313 * +---------------------+----------+ 4314 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4315 * +---------------------+----------+ | o L2ARC eligible 4316 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4317 * +---------------------+----------+ | 4318 * 15.9 Gbytes ^ 32 Mbytes | 4319 * headroom | 4320 * l2arc_feed_thread() 4321 * | 4322 * l2arc write hand <--[oooo]--' 4323 * | 8 Mbyte 4324 * | write max 4325 * V 4326 * +==============================+ 4327 * L2ARC dev |####|#|###|###| |####| ... | 4328 * +==============================+ 4329 * 32 Gbytes 4330 * 4331 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4332 * evicted, then the L2ARC has cached a buffer much sooner than it probably 4333 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4334 * safe to say that this is an uncommon case, since buffers at the end of 4335 * the ARC lists have moved there due to inactivity. 4336 * 4337 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4338 * then the L2ARC simply misses copying some buffers. This serves as a 4339 * pressure valve to prevent heavy read workloads from both stalling the ARC 4340 * with waits and clogging the L2ARC with writes. This also helps prevent 4341 * the potential for the L2ARC to churn if it attempts to cache content too 4342 * quickly, such as during backups of the entire pool. 4343 * 4344 * 5. After system boot and before the ARC has filled main memory, there are 4345 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4346 * lists can remain mostly static. Instead of searching from tail of these 4347 * lists as pictured, the l2arc_feed_thread() will search from the list heads 4348 * for eligible buffers, greatly increasing its chance of finding them. 4349 * 4350 * The L2ARC device write speed is also boosted during this time so that 4351 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4352 * there are no L2ARC reads, and no fear of degrading read performance 4353 * through increased writes. 4354 * 4355 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4356 * the vdev queue can aggregate them into larger and fewer writes. Each 4357 * device is written to in a rotor fashion, sweeping writes through 4358 * available space then repeating. 4359 * 4360 * 7. The L2ARC does not store dirty content. It never needs to flush 4361 * write buffers back to disk based storage. 4362 * 4363 * 8. If an ARC buffer is written (and dirtied) which also exists in the 4364 * L2ARC, the now stale L2ARC buffer is immediately dropped. 4365 * 4366 * The performance of the L2ARC can be tweaked by a number of tunables, which 4367 * may be necessary for different workloads: 4368 * 4369 * l2arc_write_max max write bytes per interval 4370 * l2arc_write_boost extra write bytes during device warmup 4371 * l2arc_noprefetch skip caching prefetched buffers 4372 * l2arc_headroom number of max device writes to precache 4373 * l2arc_headroom_boost when we find compressed buffers during ARC 4374 * scanning, we multiply headroom by this 4375 * percentage factor for the next scan cycle, 4376 * since more compressed buffers are likely to 4377 * be present 4378 * l2arc_feed_secs seconds between L2ARC writing 4379 * 4380 * Tunables may be removed or added as future performance improvements are 4381 * integrated, and also may become zpool properties. 4382 * 4383 * There are three key functions that control how the L2ARC warms up: 4384 * 4385 * l2arc_write_eligible() check if a buffer is eligible to cache 4386 * l2arc_write_size() calculate how much to write 4387 * l2arc_write_interval() calculate sleep delay between writes 4388 * 4389 * These three functions determine what to write, how much, and how quickly 4390 * to send writes. 4391 */ 4392 4393static boolean_t 4394l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab) 4395{ 4396 /* 4397 * A buffer is *not* eligible for the L2ARC if it: 4398 * 1. belongs to a different spa. 4399 * 2. is already cached on the L2ARC. 4400 * 3. has an I/O in progress (it may be an incomplete read). 4401 * 4. is flagged not eligible (zfs property). 4402 */ 4403 if (ab->b_spa != spa_guid) { 4404 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4405 return (B_FALSE); 4406 } 4407 if (ab->b_l2hdr != NULL) { 4408 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4409 return (B_FALSE); 4410 } 4411 if (HDR_IO_IN_PROGRESS(ab)) { 4412 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4413 return (B_FALSE); 4414 } 4415 if (!HDR_L2CACHE(ab)) { 4416 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4417 return (B_FALSE); 4418 } 4419 4420 return (B_TRUE); 4421} 4422 4423static uint64_t 4424l2arc_write_size(void) 4425{ 4426 uint64_t size; 4427 4428 /* 4429 * Make sure our globals have meaningful values in case the user 4430 * altered them. 4431 */ 4432 size = l2arc_write_max; 4433 if (size == 0) { 4434 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4435 "be greater than zero, resetting it to the default (%d)", 4436 L2ARC_WRITE_SIZE); 4437 size = l2arc_write_max = L2ARC_WRITE_SIZE; 4438 } 4439 4440 if (arc_warm == B_FALSE) 4441 size += l2arc_write_boost; 4442 4443 return (size); 4444 4445} 4446 4447static clock_t 4448l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4449{ 4450 clock_t interval, next, now; 4451 4452 /* 4453 * If the ARC lists are busy, increase our write rate; if the 4454 * lists are stale, idle back. This is achieved by checking 4455 * how much we previously wrote - if it was more than half of 4456 * what we wanted, schedule the next write much sooner. 4457 */ 4458 if (l2arc_feed_again && wrote > (wanted / 2)) 4459 interval = (hz * l2arc_feed_min_ms) / 1000; 4460 else 4461 interval = hz * l2arc_feed_secs; 4462 4463 now = ddi_get_lbolt(); 4464 next = MAX(now, MIN(now + interval, began + interval)); 4465 4466 return (next); 4467} 4468 4469static void 4470l2arc_hdr_stat_add(void) 4471{ 4472 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4473 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4474} 4475 4476static void 4477l2arc_hdr_stat_remove(void) 4478{ 4479 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4480 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4481} 4482 4483/* 4484 * Cycle through L2ARC devices. This is how L2ARC load balances. 4485 * If a device is returned, this also returns holding the spa config lock. 4486 */ 4487static l2arc_dev_t * 4488l2arc_dev_get_next(void) 4489{ 4490 l2arc_dev_t *first, *next = NULL; 4491 4492 /* 4493 * Lock out the removal of spas (spa_namespace_lock), then removal 4494 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4495 * both locks will be dropped and a spa config lock held instead. 4496 */ 4497 mutex_enter(&spa_namespace_lock); 4498 mutex_enter(&l2arc_dev_mtx); 4499 4500 /* if there are no vdevs, there is nothing to do */ 4501 if (l2arc_ndev == 0) 4502 goto out; 4503 4504 first = NULL; 4505 next = l2arc_dev_last; 4506 do { 4507 /* loop around the list looking for a non-faulted vdev */ 4508 if (next == NULL) { 4509 next = list_head(l2arc_dev_list); 4510 } else { 4511 next = list_next(l2arc_dev_list, next); 4512 if (next == NULL) 4513 next = list_head(l2arc_dev_list); 4514 } 4515 4516 /* if we have come back to the start, bail out */ 4517 if (first == NULL) 4518 first = next; 4519 else if (next == first) 4520 break; 4521 4522 } while (vdev_is_dead(next->l2ad_vdev)); 4523 4524 /* if we were unable to find any usable vdevs, return NULL */ 4525 if (vdev_is_dead(next->l2ad_vdev)) 4526 next = NULL; 4527 4528 l2arc_dev_last = next; 4529 4530out: 4531 mutex_exit(&l2arc_dev_mtx); 4532 4533 /* 4534 * Grab the config lock to prevent the 'next' device from being 4535 * removed while we are writing to it. 4536 */ 4537 if (next != NULL) 4538 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4539 mutex_exit(&spa_namespace_lock); 4540 4541 return (next); 4542} 4543 4544/* 4545 * Free buffers that were tagged for destruction. 4546 */ 4547static void 4548l2arc_do_free_on_write() 4549{ 4550 list_t *buflist; 4551 l2arc_data_free_t *df, *df_prev; 4552 4553 mutex_enter(&l2arc_free_on_write_mtx); 4554 buflist = l2arc_free_on_write; 4555 4556 for (df = list_tail(buflist); df; df = df_prev) { 4557 df_prev = list_prev(buflist, df); 4558 ASSERT(df->l2df_data != NULL); 4559 ASSERT(df->l2df_func != NULL); 4560 df->l2df_func(df->l2df_data, df->l2df_size); 4561 list_remove(buflist, df); 4562 kmem_free(df, sizeof (l2arc_data_free_t)); 4563 } 4564 4565 mutex_exit(&l2arc_free_on_write_mtx); 4566} 4567 4568/* 4569 * A write to a cache device has completed. Update all headers to allow 4570 * reads from these buffers to begin. 4571 */ 4572static void 4573l2arc_write_done(zio_t *zio) 4574{ 4575 l2arc_write_callback_t *cb; 4576 l2arc_dev_t *dev; 4577 list_t *buflist; 4578 arc_buf_hdr_t *head, *ab, *ab_prev; 4579 l2arc_buf_hdr_t *abl2; 4580 kmutex_t *hash_lock; 4581 int64_t bytes_dropped = 0; 4582 4583 cb = zio->io_private; 4584 ASSERT(cb != NULL); 4585 dev = cb->l2wcb_dev; 4586 ASSERT(dev != NULL); 4587 head = cb->l2wcb_head; 4588 ASSERT(head != NULL); 4589 buflist = dev->l2ad_buflist; 4590 ASSERT(buflist != NULL); 4591 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4592 l2arc_write_callback_t *, cb); 4593 4594 if (zio->io_error != 0) 4595 ARCSTAT_BUMP(arcstat_l2_writes_error); 4596 4597 mutex_enter(&l2arc_buflist_mtx); 4598 4599 /* 4600 * All writes completed, or an error was hit. 4601 */ 4602 for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 4603 ab_prev = list_prev(buflist, ab); 4604 abl2 = ab->b_l2hdr; 4605 4606 /* 4607 * Release the temporary compressed buffer as soon as possible. 4608 */ 4609 if (abl2->b_compress != ZIO_COMPRESS_OFF) 4610 l2arc_release_cdata_buf(ab); 4611 4612 hash_lock = HDR_LOCK(ab); 4613 if (!mutex_tryenter(hash_lock)) { 4614 /* 4615 * This buffer misses out. It may be in a stage 4616 * of eviction. Its ARC_L2_WRITING flag will be 4617 * left set, denying reads to this buffer. 4618 */ 4619 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4620 continue; 4621 } 4622 4623 if (zio->io_error != 0) { 4624 /* 4625 * Error - drop L2ARC entry. 4626 */ 4627 list_remove(buflist, ab); 4628 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4629 bytes_dropped += abl2->b_asize; 4630 ab->b_l2hdr = NULL; 4631 trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4632 ab->b_size, 0); 4633 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4634 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4635 } 4636 4637 /* 4638 * Allow ARC to begin reads to this L2ARC entry. 4639 */ 4640 ab->b_flags &= ~ARC_L2_WRITING; 4641 4642 mutex_exit(hash_lock); 4643 } 4644 4645 atomic_inc_64(&l2arc_writes_done); 4646 list_remove(buflist, head); 4647 kmem_cache_free(hdr_cache, head); 4648 mutex_exit(&l2arc_buflist_mtx); 4649 4650 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 4651 4652 l2arc_do_free_on_write(); 4653 4654 kmem_free(cb, sizeof (l2arc_write_callback_t)); 4655} 4656 4657/* 4658 * A read to a cache device completed. Validate buffer contents before 4659 * handing over to the regular ARC routines. 4660 */ 4661static void 4662l2arc_read_done(zio_t *zio) 4663{ 4664 l2arc_read_callback_t *cb; 4665 arc_buf_hdr_t *hdr; 4666 arc_buf_t *buf; 4667 kmutex_t *hash_lock; 4668 int equal; 4669 4670 ASSERT(zio->io_vd != NULL); 4671 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4672 4673 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4674 4675 cb = zio->io_private; 4676 ASSERT(cb != NULL); 4677 buf = cb->l2rcb_buf; 4678 ASSERT(buf != NULL); 4679 4680 hash_lock = HDR_LOCK(buf->b_hdr); 4681 mutex_enter(hash_lock); 4682 hdr = buf->b_hdr; 4683 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4684 4685 /* 4686 * If the buffer was compressed, decompress it first. 4687 */ 4688 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4689 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4690 ASSERT(zio->io_data != NULL); 4691 4692 /* 4693 * Check this survived the L2ARC journey. 4694 */ 4695 equal = arc_cksum_equal(buf); 4696 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4697 mutex_exit(hash_lock); 4698 zio->io_private = buf; 4699 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4700 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4701 arc_read_done(zio); 4702 } else { 4703 mutex_exit(hash_lock); 4704 /* 4705 * Buffer didn't survive caching. Increment stats and 4706 * reissue to the original storage device. 4707 */ 4708 if (zio->io_error != 0) { 4709 ARCSTAT_BUMP(arcstat_l2_io_error); 4710 } else { 4711 zio->io_error = SET_ERROR(EIO); 4712 } 4713 if (!equal) 4714 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4715 4716 /* 4717 * If there's no waiter, issue an async i/o to the primary 4718 * storage now. If there *is* a waiter, the caller must 4719 * issue the i/o in a context where it's OK to block. 4720 */ 4721 if (zio->io_waiter == NULL) { 4722 zio_t *pio = zio_unique_parent(zio); 4723 4724 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4725 4726 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4727 buf->b_data, zio->io_size, arc_read_done, buf, 4728 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4729 } 4730 } 4731 4732 kmem_free(cb, sizeof (l2arc_read_callback_t)); 4733} 4734 4735/* 4736 * This is the list priority from which the L2ARC will search for pages to 4737 * cache. This is used within loops (0..3) to cycle through lists in the 4738 * desired order. This order can have a significant effect on cache 4739 * performance. 4740 * 4741 * Currently the metadata lists are hit first, MFU then MRU, followed by 4742 * the data lists. This function returns a locked list, and also returns 4743 * the lock pointer. 4744 */ 4745static list_t * 4746l2arc_list_locked(int list_num, kmutex_t **lock) 4747{ 4748 list_t *list = NULL; 4749 int idx; 4750 4751 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4752 4753 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4754 idx = list_num; 4755 list = &arc_mfu->arcs_lists[idx]; 4756 *lock = ARCS_LOCK(arc_mfu, idx); 4757 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4758 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4759 list = &arc_mru->arcs_lists[idx]; 4760 *lock = ARCS_LOCK(arc_mru, idx); 4761 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4762 ARC_BUFC_NUMDATALISTS)) { 4763 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4764 list = &arc_mfu->arcs_lists[idx]; 4765 *lock = ARCS_LOCK(arc_mfu, idx); 4766 } else { 4767 idx = list_num - ARC_BUFC_NUMLISTS; 4768 list = &arc_mru->arcs_lists[idx]; 4769 *lock = ARCS_LOCK(arc_mru, idx); 4770 } 4771 4772 ASSERT(!(MUTEX_HELD(*lock))); 4773 mutex_enter(*lock); 4774 return (list); 4775} 4776 4777/* 4778 * Evict buffers from the device write hand to the distance specified in 4779 * bytes. This distance may span populated buffers, it may span nothing. 4780 * This is clearing a region on the L2ARC device ready for writing. 4781 * If the 'all' boolean is set, every buffer is evicted. 4782 */ 4783static void 4784l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4785{ 4786 list_t *buflist; 4787 l2arc_buf_hdr_t *abl2; 4788 arc_buf_hdr_t *ab, *ab_prev; 4789 kmutex_t *hash_lock; 4790 uint64_t taddr; 4791 int64_t bytes_evicted = 0; 4792 4793 buflist = dev->l2ad_buflist; 4794 4795 if (buflist == NULL) 4796 return; 4797 4798 if (!all && dev->l2ad_first) { 4799 /* 4800 * This is the first sweep through the device. There is 4801 * nothing to evict. 4802 */ 4803 return; 4804 } 4805 4806 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4807 /* 4808 * When nearing the end of the device, evict to the end 4809 * before the device write hand jumps to the start. 4810 */ 4811 taddr = dev->l2ad_end; 4812 } else { 4813 taddr = dev->l2ad_hand + distance; 4814 } 4815 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4816 uint64_t, taddr, boolean_t, all); 4817 4818top: 4819 mutex_enter(&l2arc_buflist_mtx); 4820 for (ab = list_tail(buflist); ab; ab = ab_prev) { 4821 ab_prev = list_prev(buflist, ab); 4822 4823 hash_lock = HDR_LOCK(ab); 4824 if (!mutex_tryenter(hash_lock)) { 4825 /* 4826 * Missed the hash lock. Retry. 4827 */ 4828 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4829 mutex_exit(&l2arc_buflist_mtx); 4830 mutex_enter(hash_lock); 4831 mutex_exit(hash_lock); 4832 goto top; 4833 } 4834 4835 if (HDR_L2_WRITE_HEAD(ab)) { 4836 /* 4837 * We hit a write head node. Leave it for 4838 * l2arc_write_done(). 4839 */ 4840 list_remove(buflist, ab); 4841 mutex_exit(hash_lock); 4842 continue; 4843 } 4844 4845 if (!all && ab->b_l2hdr != NULL && 4846 (ab->b_l2hdr->b_daddr > taddr || 4847 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4848 /* 4849 * We've evicted to the target address, 4850 * or the end of the device. 4851 */ 4852 mutex_exit(hash_lock); 4853 break; 4854 } 4855 4856 if (HDR_FREE_IN_PROGRESS(ab)) { 4857 /* 4858 * Already on the path to destruction. 4859 */ 4860 mutex_exit(hash_lock); 4861 continue; 4862 } 4863 4864 if (ab->b_state == arc_l2c_only) { 4865 ASSERT(!HDR_L2_READING(ab)); 4866 /* 4867 * This doesn't exist in the ARC. Destroy. 4868 * arc_hdr_destroy() will call list_remove() 4869 * and decrement arcstat_l2_size. 4870 */ 4871 arc_change_state(arc_anon, ab, hash_lock); 4872 arc_hdr_destroy(ab); 4873 } else { 4874 /* 4875 * Invalidate issued or about to be issued 4876 * reads, since we may be about to write 4877 * over this location. 4878 */ 4879 if (HDR_L2_READING(ab)) { 4880 ARCSTAT_BUMP(arcstat_l2_evict_reading); 4881 ab->b_flags |= ARC_L2_EVICTED; 4882 } 4883 4884 /* 4885 * Tell ARC this no longer exists in L2ARC. 4886 */ 4887 if (ab->b_l2hdr != NULL) { 4888 abl2 = ab->b_l2hdr; 4889 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4890 bytes_evicted += abl2->b_asize; 4891 ab->b_l2hdr = NULL; 4892 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4893 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4894 } 4895 list_remove(buflist, ab); 4896 4897 /* 4898 * This may have been leftover after a 4899 * failed write. 4900 */ 4901 ab->b_flags &= ~ARC_L2_WRITING; 4902 } 4903 mutex_exit(hash_lock); 4904 } 4905 mutex_exit(&l2arc_buflist_mtx); 4906 4907 vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); 4908 dev->l2ad_evict = taddr; 4909} 4910 4911/* 4912 * Find and write ARC buffers to the L2ARC device. 4913 * 4914 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 4915 * for reading until they have completed writing. 4916 * The headroom_boost is an in-out parameter used to maintain headroom boost 4917 * state between calls to this function. 4918 * 4919 * Returns the number of bytes actually written (which may be smaller than 4920 * the delta by which the device hand has changed due to alignment). 4921 */ 4922static uint64_t 4923l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 4924 boolean_t *headroom_boost) 4925{ 4926 arc_buf_hdr_t *ab, *ab_prev, *head; 4927 list_t *list; 4928 uint64_t write_asize, write_psize, write_sz, headroom, 4929 buf_compress_minsz; 4930 void *buf_data; 4931 kmutex_t *list_lock; 4932 boolean_t full; 4933 l2arc_write_callback_t *cb; 4934 zio_t *pio, *wzio; 4935 uint64_t guid = spa_load_guid(spa); 4936 const boolean_t do_headroom_boost = *headroom_boost; 4937 int try; 4938 4939 ASSERT(dev->l2ad_vdev != NULL); 4940 4941 /* Lower the flag now, we might want to raise it again later. */ 4942 *headroom_boost = B_FALSE; 4943 4944 pio = NULL; 4945 write_sz = write_asize = write_psize = 0; 4946 full = B_FALSE; 4947 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 4948 head->b_flags |= ARC_L2_WRITE_HEAD; 4949 4950 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 4951 /* 4952 * We will want to try to compress buffers that are at least 2x the 4953 * device sector size. 4954 */ 4955 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 4956 4957 /* 4958 * Copy buffers for L2ARC writing. 4959 */ 4960 mutex_enter(&l2arc_buflist_mtx); 4961 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 4962 uint64_t passed_sz = 0; 4963 4964 list = l2arc_list_locked(try, &list_lock); 4965 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 4966 4967 /* 4968 * L2ARC fast warmup. 4969 * 4970 * Until the ARC is warm and starts to evict, read from the 4971 * head of the ARC lists rather than the tail. 4972 */ 4973 if (arc_warm == B_FALSE) 4974 ab = list_head(list); 4975 else 4976 ab = list_tail(list); 4977 if (ab == NULL) 4978 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 4979 4980 headroom = target_sz * l2arc_headroom; 4981 if (do_headroom_boost) 4982 headroom = (headroom * l2arc_headroom_boost) / 100; 4983 4984 for (; ab; ab = ab_prev) { 4985 l2arc_buf_hdr_t *l2hdr; 4986 kmutex_t *hash_lock; 4987 uint64_t buf_sz; 4988 4989 if (arc_warm == B_FALSE) 4990 ab_prev = list_next(list, ab); 4991 else 4992 ab_prev = list_prev(list, ab); 4993 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size); 4994 4995 hash_lock = HDR_LOCK(ab); 4996 if (!mutex_tryenter(hash_lock)) { 4997 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 4998 /* 4999 * Skip this buffer rather than waiting. 5000 */ 5001 continue; 5002 } 5003 5004 passed_sz += ab->b_size; 5005 if (passed_sz > headroom) { 5006 /* 5007 * Searched too far. 5008 */ 5009 mutex_exit(hash_lock); 5010 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5011 break; 5012 } 5013 5014 if (!l2arc_write_eligible(guid, ab)) { 5015 mutex_exit(hash_lock); 5016 continue; 5017 } 5018 5019 if ((write_sz + ab->b_size) > target_sz) { 5020 full = B_TRUE; 5021 mutex_exit(hash_lock); 5022 ARCSTAT_BUMP(arcstat_l2_write_full); 5023 break; 5024 } 5025 5026 if (pio == NULL) { 5027 /* 5028 * Insert a dummy header on the buflist so 5029 * l2arc_write_done() can find where the 5030 * write buffers begin without searching. 5031 */ 5032 list_insert_head(dev->l2ad_buflist, head); 5033 5034 cb = kmem_alloc( 5035 sizeof (l2arc_write_callback_t), KM_SLEEP); 5036 cb->l2wcb_dev = dev; 5037 cb->l2wcb_head = head; 5038 pio = zio_root(spa, l2arc_write_done, cb, 5039 ZIO_FLAG_CANFAIL); 5040 ARCSTAT_BUMP(arcstat_l2_write_pios); 5041 } 5042 5043 /* 5044 * Create and add a new L2ARC header. 5045 */ 5046 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5047 l2hdr->b_dev = dev; 5048 ab->b_flags |= ARC_L2_WRITING; 5049 5050 /* 5051 * Temporarily stash the data buffer in b_tmp_cdata. 5052 * The subsequent write step will pick it up from 5053 * there. This is because can't access ab->b_buf 5054 * without holding the hash_lock, which we in turn 5055 * can't access without holding the ARC list locks 5056 * (which we want to avoid during compression/writing). 5057 */ 5058 l2hdr->b_compress = ZIO_COMPRESS_OFF; 5059 l2hdr->b_asize = ab->b_size; 5060 l2hdr->b_tmp_cdata = ab->b_buf->b_data; 5061 5062 buf_sz = ab->b_size; 5063 ab->b_l2hdr = l2hdr; 5064 5065 list_insert_head(dev->l2ad_buflist, ab); 5066 5067 /* 5068 * Compute and store the buffer cksum before 5069 * writing. On debug the cksum is verified first. 5070 */ 5071 arc_cksum_verify(ab->b_buf); 5072 arc_cksum_compute(ab->b_buf, B_TRUE); 5073 5074 mutex_exit(hash_lock); 5075 5076 write_sz += buf_sz; 5077 } 5078 5079 mutex_exit(list_lock); 5080 5081 if (full == B_TRUE) 5082 break; 5083 } 5084 5085 /* No buffers selected for writing? */ 5086 if (pio == NULL) { 5087 ASSERT0(write_sz); 5088 mutex_exit(&l2arc_buflist_mtx); 5089 kmem_cache_free(hdr_cache, head); 5090 return (0); 5091 } 5092 5093 /* 5094 * Now start writing the buffers. We're starting at the write head 5095 * and work backwards, retracing the course of the buffer selector 5096 * loop above. 5097 */ 5098 for (ab = list_prev(dev->l2ad_buflist, head); ab; 5099 ab = list_prev(dev->l2ad_buflist, ab)) { 5100 l2arc_buf_hdr_t *l2hdr; 5101 uint64_t buf_sz; 5102 5103 /* 5104 * We shouldn't need to lock the buffer here, since we flagged 5105 * it as ARC_L2_WRITING in the previous step, but we must take 5106 * care to only access its L2 cache parameters. In particular, 5107 * ab->b_buf may be invalid by now due to ARC eviction. 5108 */ 5109 l2hdr = ab->b_l2hdr; 5110 l2hdr->b_daddr = dev->l2ad_hand; 5111 5112 if ((ab->b_flags & ARC_L2COMPRESS) && 5113 l2hdr->b_asize >= buf_compress_minsz) { 5114 if (l2arc_compress_buf(l2hdr)) { 5115 /* 5116 * If compression succeeded, enable headroom 5117 * boost on the next scan cycle. 5118 */ 5119 *headroom_boost = B_TRUE; 5120 } 5121 } 5122 5123 /* 5124 * Pick up the buffer data we had previously stashed away 5125 * (and now potentially also compressed). 5126 */ 5127 buf_data = l2hdr->b_tmp_cdata; 5128 buf_sz = l2hdr->b_asize; 5129 5130 /* Compression may have squashed the buffer to zero length. */ 5131 if (buf_sz != 0) { 5132 uint64_t buf_p_sz; 5133 5134 wzio = zio_write_phys(pio, dev->l2ad_vdev, 5135 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5136 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5137 ZIO_FLAG_CANFAIL, B_FALSE); 5138 5139 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5140 zio_t *, wzio); 5141 (void) zio_nowait(wzio); 5142 5143 write_asize += buf_sz; 5144 /* 5145 * Keep the clock hand suitably device-aligned. 5146 */ 5147 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5148 write_psize += buf_p_sz; 5149 dev->l2ad_hand += buf_p_sz; 5150 } 5151 } 5152 5153 mutex_exit(&l2arc_buflist_mtx); 5154 5155 ASSERT3U(write_asize, <=, target_sz); 5156 ARCSTAT_BUMP(arcstat_l2_writes_sent); 5157 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5158 ARCSTAT_INCR(arcstat_l2_size, write_sz); 5159 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 5160 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 5161 5162 /* 5163 * Bump device hand to the device start if it is approaching the end. 5164 * l2arc_evict() will already have evicted ahead for this case. 5165 */ 5166 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5167 dev->l2ad_hand = dev->l2ad_start; 5168 dev->l2ad_evict = dev->l2ad_start; 5169 dev->l2ad_first = B_FALSE; 5170 } 5171 5172 dev->l2ad_writing = B_TRUE; 5173 (void) zio_wait(pio); 5174 dev->l2ad_writing = B_FALSE; 5175 5176 return (write_asize); 5177} 5178 5179/* 5180 * Compresses an L2ARC buffer. 5181 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5182 * size in l2hdr->b_asize. This routine tries to compress the data and 5183 * depending on the compression result there are three possible outcomes: 5184 * *) The buffer was incompressible. The original l2hdr contents were left 5185 * untouched and are ready for writing to an L2 device. 5186 * *) The buffer was all-zeros, so there is no need to write it to an L2 5187 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5188 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5189 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5190 * data buffer which holds the compressed data to be written, and b_asize 5191 * tells us how much data there is. b_compress is set to the appropriate 5192 * compression algorithm. Once writing is done, invoke 5193 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5194 * 5195 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5196 * buffer was incompressible). 5197 */ 5198static boolean_t 5199l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5200{ 5201 void *cdata; 5202 size_t csize, len, rounded; 5203 5204 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5205 ASSERT(l2hdr->b_tmp_cdata != NULL); 5206 5207 len = l2hdr->b_asize; 5208 cdata = zio_data_buf_alloc(len); 5209 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5210 cdata, l2hdr->b_asize); 5211 5212 rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE); 5213 if (rounded > csize) { 5214 bzero((char *)cdata + csize, rounded - csize); 5215 csize = rounded; 5216 } 5217 5218 if (csize == 0) { 5219 /* zero block, indicate that there's nothing to write */ 5220 zio_data_buf_free(cdata, len); 5221 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5222 l2hdr->b_asize = 0; 5223 l2hdr->b_tmp_cdata = NULL; 5224 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5225 return (B_TRUE); 5226 } else if (csize > 0 && csize < len) { 5227 /* 5228 * Compression succeeded, we'll keep the cdata around for 5229 * writing and release it afterwards. 5230 */ 5231 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5232 l2hdr->b_asize = csize; 5233 l2hdr->b_tmp_cdata = cdata; 5234 ARCSTAT_BUMP(arcstat_l2_compress_successes); 5235 return (B_TRUE); 5236 } else { 5237 /* 5238 * Compression failed, release the compressed buffer. 5239 * l2hdr will be left unmodified. 5240 */ 5241 zio_data_buf_free(cdata, len); 5242 ARCSTAT_BUMP(arcstat_l2_compress_failures); 5243 return (B_FALSE); 5244 } 5245} 5246 5247/* 5248 * Decompresses a zio read back from an l2arc device. On success, the 5249 * underlying zio's io_data buffer is overwritten by the uncompressed 5250 * version. On decompression error (corrupt compressed stream), the 5251 * zio->io_error value is set to signal an I/O error. 5252 * 5253 * Please note that the compressed data stream is not checksummed, so 5254 * if the underlying device is experiencing data corruption, we may feed 5255 * corrupt data to the decompressor, so the decompressor needs to be 5256 * able to handle this situation (LZ4 does). 5257 */ 5258static void 5259l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5260{ 5261 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5262 5263 if (zio->io_error != 0) { 5264 /* 5265 * An io error has occured, just restore the original io 5266 * size in preparation for a main pool read. 5267 */ 5268 zio->io_orig_size = zio->io_size = hdr->b_size; 5269 return; 5270 } 5271 5272 if (c == ZIO_COMPRESS_EMPTY) { 5273 /* 5274 * An empty buffer results in a null zio, which means we 5275 * need to fill its io_data after we're done restoring the 5276 * buffer's contents. 5277 */ 5278 ASSERT(hdr->b_buf != NULL); 5279 bzero(hdr->b_buf->b_data, hdr->b_size); 5280 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5281 } else { 5282 ASSERT(zio->io_data != NULL); 5283 /* 5284 * We copy the compressed data from the start of the arc buffer 5285 * (the zio_read will have pulled in only what we need, the 5286 * rest is garbage which we will overwrite at decompression) 5287 * and then decompress back to the ARC data buffer. This way we 5288 * can minimize copying by simply decompressing back over the 5289 * original compressed data (rather than decompressing to an 5290 * aux buffer and then copying back the uncompressed buffer, 5291 * which is likely to be much larger). 5292 */ 5293 uint64_t csize; 5294 void *cdata; 5295 5296 csize = zio->io_size; 5297 cdata = zio_data_buf_alloc(csize); 5298 bcopy(zio->io_data, cdata, csize); 5299 if (zio_decompress_data(c, cdata, zio->io_data, csize, 5300 hdr->b_size) != 0) 5301 zio->io_error = EIO; 5302 zio_data_buf_free(cdata, csize); 5303 } 5304 5305 /* Restore the expected uncompressed IO size. */ 5306 zio->io_orig_size = zio->io_size = hdr->b_size; 5307} 5308 5309/* 5310 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5311 * This buffer serves as a temporary holder of compressed data while 5312 * the buffer entry is being written to an l2arc device. Once that is 5313 * done, we can dispose of it. 5314 */ 5315static void 5316l2arc_release_cdata_buf(arc_buf_hdr_t *ab) 5317{ 5318 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr; 5319 5320 if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) { 5321 /* 5322 * If the data was compressed, then we've allocated a 5323 * temporary buffer for it, so now we need to release it. 5324 */ 5325 ASSERT(l2hdr->b_tmp_cdata != NULL); 5326 zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size); 5327 } 5328 l2hdr->b_tmp_cdata = NULL; 5329} 5330 5331/* 5332 * This thread feeds the L2ARC at regular intervals. This is the beating 5333 * heart of the L2ARC. 5334 */ 5335static void 5336l2arc_feed_thread(void *dummy __unused) 5337{ 5338 callb_cpr_t cpr; 5339 l2arc_dev_t *dev; 5340 spa_t *spa; 5341 uint64_t size, wrote; 5342 clock_t begin, next = ddi_get_lbolt(); 5343 boolean_t headroom_boost = B_FALSE; 5344 5345 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5346 5347 mutex_enter(&l2arc_feed_thr_lock); 5348 5349 while (l2arc_thread_exit == 0) { 5350 CALLB_CPR_SAFE_BEGIN(&cpr); 5351 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5352 next - ddi_get_lbolt()); 5353 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5354 next = ddi_get_lbolt() + hz; 5355 5356 /* 5357 * Quick check for L2ARC devices. 5358 */ 5359 mutex_enter(&l2arc_dev_mtx); 5360 if (l2arc_ndev == 0) { 5361 mutex_exit(&l2arc_dev_mtx); 5362 continue; 5363 } 5364 mutex_exit(&l2arc_dev_mtx); 5365 begin = ddi_get_lbolt(); 5366 5367 /* 5368 * This selects the next l2arc device to write to, and in 5369 * doing so the next spa to feed from: dev->l2ad_spa. This 5370 * will return NULL if there are now no l2arc devices or if 5371 * they are all faulted. 5372 * 5373 * If a device is returned, its spa's config lock is also 5374 * held to prevent device removal. l2arc_dev_get_next() 5375 * will grab and release l2arc_dev_mtx. 5376 */ 5377 if ((dev = l2arc_dev_get_next()) == NULL) 5378 continue; 5379 5380 spa = dev->l2ad_spa; 5381 ASSERT(spa != NULL); 5382 5383 /* 5384 * If the pool is read-only then force the feed thread to 5385 * sleep a little longer. 5386 */ 5387 if (!spa_writeable(spa)) { 5388 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5389 spa_config_exit(spa, SCL_L2ARC, dev); 5390 continue; 5391 } 5392 5393 /* 5394 * Avoid contributing to memory pressure. 5395 */ 5396 if (arc_reclaim_needed()) { 5397 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5398 spa_config_exit(spa, SCL_L2ARC, dev); 5399 continue; 5400 } 5401 5402 ARCSTAT_BUMP(arcstat_l2_feeds); 5403 5404 size = l2arc_write_size(); 5405 5406 /* 5407 * Evict L2ARC buffers that will be overwritten. 5408 */ 5409 l2arc_evict(dev, size, B_FALSE); 5410 5411 /* 5412 * Write ARC buffers. 5413 */ 5414 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5415 5416 /* 5417 * Calculate interval between writes. 5418 */ 5419 next = l2arc_write_interval(begin, size, wrote); 5420 spa_config_exit(spa, SCL_L2ARC, dev); 5421 } 5422 5423 l2arc_thread_exit = 0; 5424 cv_broadcast(&l2arc_feed_thr_cv); 5425 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5426 thread_exit(); 5427} 5428 5429boolean_t 5430l2arc_vdev_present(vdev_t *vd) 5431{ 5432 l2arc_dev_t *dev; 5433 5434 mutex_enter(&l2arc_dev_mtx); 5435 for (dev = list_head(l2arc_dev_list); dev != NULL; 5436 dev = list_next(l2arc_dev_list, dev)) { 5437 if (dev->l2ad_vdev == vd) 5438 break; 5439 } 5440 mutex_exit(&l2arc_dev_mtx); 5441 5442 return (dev != NULL); 5443} 5444 5445/* 5446 * Add a vdev for use by the L2ARC. By this point the spa has already 5447 * validated the vdev and opened it. 5448 */ 5449void 5450l2arc_add_vdev(spa_t *spa, vdev_t *vd) 5451{ 5452 l2arc_dev_t *adddev; 5453 5454 ASSERT(!l2arc_vdev_present(vd)); 5455 5456 vdev_ashift_optimize(vd); 5457 5458 /* 5459 * Create a new l2arc device entry. 5460 */ 5461 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5462 adddev->l2ad_spa = spa; 5463 adddev->l2ad_vdev = vd; 5464 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5465 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5466 adddev->l2ad_hand = adddev->l2ad_start; 5467 adddev->l2ad_evict = adddev->l2ad_start; 5468 adddev->l2ad_first = B_TRUE; 5469 adddev->l2ad_writing = B_FALSE; 5470 5471 /* 5472 * This is a list of all ARC buffers that are still valid on the 5473 * device. 5474 */ 5475 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5476 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5477 offsetof(arc_buf_hdr_t, b_l2node)); 5478 5479 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5480 5481 /* 5482 * Add device to global list 5483 */ 5484 mutex_enter(&l2arc_dev_mtx); 5485 list_insert_head(l2arc_dev_list, adddev); 5486 atomic_inc_64(&l2arc_ndev); 5487 mutex_exit(&l2arc_dev_mtx); 5488} 5489 5490/* 5491 * Remove a vdev from the L2ARC. 5492 */ 5493void 5494l2arc_remove_vdev(vdev_t *vd) 5495{ 5496 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5497 5498 /* 5499 * Find the device by vdev 5500 */ 5501 mutex_enter(&l2arc_dev_mtx); 5502 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5503 nextdev = list_next(l2arc_dev_list, dev); 5504 if (vd == dev->l2ad_vdev) { 5505 remdev = dev; 5506 break; 5507 } 5508 } 5509 ASSERT(remdev != NULL); 5510 5511 /* 5512 * Remove device from global list 5513 */ 5514 list_remove(l2arc_dev_list, remdev); 5515 l2arc_dev_last = NULL; /* may have been invalidated */ 5516 atomic_dec_64(&l2arc_ndev); 5517 mutex_exit(&l2arc_dev_mtx); 5518 5519 /* 5520 * Clear all buflists and ARC references. L2ARC device flush. 5521 */ 5522 l2arc_evict(remdev, 0, B_TRUE); 5523 list_destroy(remdev->l2ad_buflist); 5524 kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5525 kmem_free(remdev, sizeof (l2arc_dev_t)); 5526} 5527 5528void 5529l2arc_init(void) 5530{ 5531 l2arc_thread_exit = 0; 5532 l2arc_ndev = 0; 5533 l2arc_writes_sent = 0; 5534 l2arc_writes_done = 0; 5535 5536 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5537 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5538 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5539 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5540 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5541 5542 l2arc_dev_list = &L2ARC_dev_list; 5543 l2arc_free_on_write = &L2ARC_free_on_write; 5544 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5545 offsetof(l2arc_dev_t, l2ad_node)); 5546 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5547 offsetof(l2arc_data_free_t, l2df_list_node)); 5548} 5549 5550void 5551l2arc_fini(void) 5552{ 5553 /* 5554 * This is called from dmu_fini(), which is called from spa_fini(); 5555 * Because of this, we can assume that all l2arc devices have 5556 * already been removed when the pools themselves were removed. 5557 */ 5558 5559 l2arc_do_free_on_write(); 5560 5561 mutex_destroy(&l2arc_feed_thr_lock); 5562 cv_destroy(&l2arc_feed_thr_cv); 5563 mutex_destroy(&l2arc_dev_mtx); 5564 mutex_destroy(&l2arc_buflist_mtx); 5565 mutex_destroy(&l2arc_free_on_write_mtx); 5566 5567 list_destroy(l2arc_dev_list); 5568 list_destroy(l2arc_free_on_write); 5569} 5570 5571void 5572l2arc_start(void) 5573{ 5574 if (!(spa_mode_global & FWRITE)) 5575 return; 5576 5577 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5578 TS_RUN, minclsyspri); 5579} 5580 5581void 5582l2arc_stop(void) 5583{ 5584 if (!(spa_mode_global & FWRITE)) 5585 return; 5586 5587 mutex_enter(&l2arc_feed_thr_lock); 5588 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5589 l2arc_thread_exit = 1; 5590 while (l2arc_thread_exit != 0) 5591 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5592 mutex_exit(&l2arc_feed_thr_lock); 5593} 5594