arc.c revision 268657
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_buf_evict() 108 * and arc_do_user_evicts(). 109 * 110 * Note that the majority of the performance stats are manipulated 111 * with atomic operations. 112 * 113 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following: 114 * 115 * - L2ARC buflist creation 116 * - L2ARC buflist eviction 117 * - L2ARC write completion, which walks L2ARC buflists 118 * - ARC header destruction, as it removes from L2ARC buflists 119 * - ARC header release, as it removes from L2ARC buflists 120 */ 121 122#include <sys/spa.h> 123#include <sys/zio.h> 124#include <sys/zio_compress.h> 125#include <sys/zfs_context.h> 126#include <sys/arc.h> 127#include <sys/refcount.h> 128#include <sys/vdev.h> 129#include <sys/vdev_impl.h> 130#include <sys/dsl_pool.h> 131#ifdef _KERNEL 132#include <sys/dnlc.h> 133#endif 134#include <sys/callb.h> 135#include <sys/kstat.h> 136#include <sys/trim_map.h> 137#include <zfs_fletcher.h> 138#include <sys/sdt.h> 139 140#include <vm/vm_pageout.h> 141 142#ifdef illumos 143#ifndef _KERNEL 144/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 145boolean_t arc_watch = B_FALSE; 146int arc_procfd; 147#endif 148#endif /* illumos */ 149 150static kmutex_t arc_reclaim_thr_lock; 151static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 152static uint8_t arc_thread_exit; 153 154#define ARC_REDUCE_DNLC_PERCENT 3 155uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 156 157typedef enum arc_reclaim_strategy { 158 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 159 ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 160} arc_reclaim_strategy_t; 161 162/* 163 * The number of iterations through arc_evict_*() before we 164 * drop & reacquire the lock. 165 */ 166int arc_evict_iterations = 100; 167 168/* number of seconds before growing cache again */ 169static int arc_grow_retry = 60; 170 171/* shift of arc_c for calculating both min and max arc_p */ 172static int arc_p_min_shift = 4; 173 174/* log2(fraction of arc to reclaim) */ 175static int arc_shrink_shift = 5; 176 177/* 178 * minimum lifespan of a prefetch block in clock ticks 179 * (initialized in arc_init()) 180 */ 181static int arc_min_prefetch_lifespan; 182 183/* 184 * If this percent of memory is free, don't throttle. 185 */ 186int arc_lotsfree_percent = 10; 187 188static int arc_dead; 189extern int zfs_prefetch_disable; 190 191/* 192 * The arc has filled available memory and has now warmed up. 193 */ 194static boolean_t arc_warm; 195 196/* 197 * These tunables are for performance analysis. 198 */ 199uint64_t zfs_arc_max; 200uint64_t zfs_arc_min; 201uint64_t zfs_arc_meta_limit = 0; 202int zfs_arc_grow_retry = 0; 203int zfs_arc_shrink_shift = 0; 204int zfs_arc_p_min_shift = 0; 205int zfs_disable_dup_eviction = 0; 206 207TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max); 208TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 209TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 210SYSCTL_DECL(_vfs_zfs); 211SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 212 "Maximum ARC size"); 213SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 214 "Minimum ARC size"); 215 216/* 217 * Note that buffers can be in one of 6 states: 218 * ARC_anon - anonymous (discussed below) 219 * ARC_mru - recently used, currently cached 220 * ARC_mru_ghost - recentely used, no longer in cache 221 * ARC_mfu - frequently used, currently cached 222 * ARC_mfu_ghost - frequently used, no longer in cache 223 * ARC_l2c_only - exists in L2ARC but not other states 224 * When there are no active references to the buffer, they are 225 * are linked onto a list in one of these arc states. These are 226 * the only buffers that can be evicted or deleted. Within each 227 * state there are multiple lists, one for meta-data and one for 228 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 229 * etc.) is tracked separately so that it can be managed more 230 * explicitly: favored over data, limited explicitly. 231 * 232 * Anonymous buffers are buffers that are not associated with 233 * a DVA. These are buffers that hold dirty block copies 234 * before they are written to stable storage. By definition, 235 * they are "ref'd" and are considered part of arc_mru 236 * that cannot be freed. Generally, they will aquire a DVA 237 * as they are written and migrate onto the arc_mru list. 238 * 239 * The ARC_l2c_only state is for buffers that are in the second 240 * level ARC but no longer in any of the ARC_m* lists. The second 241 * level ARC itself may also contain buffers that are in any of 242 * the ARC_m* states - meaning that a buffer can exist in two 243 * places. The reason for the ARC_l2c_only state is to keep the 244 * buffer header in the hash table, so that reads that hit the 245 * second level ARC benefit from these fast lookups. 246 */ 247 248#define ARCS_LOCK_PAD CACHE_LINE_SIZE 249struct arcs_lock { 250 kmutex_t arcs_lock; 251#ifdef _KERNEL 252 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 253#endif 254}; 255 256/* 257 * must be power of two for mask use to work 258 * 259 */ 260#define ARC_BUFC_NUMDATALISTS 16 261#define ARC_BUFC_NUMMETADATALISTS 16 262#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 263 264typedef struct arc_state { 265 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 266 uint64_t arcs_size; /* total amount of data in this state */ 267 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 268 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 269} arc_state_t; 270 271#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 272 273/* The 6 states: */ 274static arc_state_t ARC_anon; 275static arc_state_t ARC_mru; 276static arc_state_t ARC_mru_ghost; 277static arc_state_t ARC_mfu; 278static arc_state_t ARC_mfu_ghost; 279static arc_state_t ARC_l2c_only; 280 281typedef struct arc_stats { 282 kstat_named_t arcstat_hits; 283 kstat_named_t arcstat_misses; 284 kstat_named_t arcstat_demand_data_hits; 285 kstat_named_t arcstat_demand_data_misses; 286 kstat_named_t arcstat_demand_metadata_hits; 287 kstat_named_t arcstat_demand_metadata_misses; 288 kstat_named_t arcstat_prefetch_data_hits; 289 kstat_named_t arcstat_prefetch_data_misses; 290 kstat_named_t arcstat_prefetch_metadata_hits; 291 kstat_named_t arcstat_prefetch_metadata_misses; 292 kstat_named_t arcstat_mru_hits; 293 kstat_named_t arcstat_mru_ghost_hits; 294 kstat_named_t arcstat_mfu_hits; 295 kstat_named_t arcstat_mfu_ghost_hits; 296 kstat_named_t arcstat_allocated; 297 kstat_named_t arcstat_deleted; 298 kstat_named_t arcstat_stolen; 299 kstat_named_t arcstat_recycle_miss; 300 /* 301 * Number of buffers that could not be evicted because the hash lock 302 * was held by another thread. The lock may not necessarily be held 303 * by something using the same buffer, since hash locks are shared 304 * by multiple buffers. 305 */ 306 kstat_named_t arcstat_mutex_miss; 307 /* 308 * Number of buffers skipped because they have I/O in progress, are 309 * indrect prefetch buffers that have not lived long enough, or are 310 * not from the spa we're trying to evict from. 311 */ 312 kstat_named_t arcstat_evict_skip; 313 kstat_named_t arcstat_evict_l2_cached; 314 kstat_named_t arcstat_evict_l2_eligible; 315 kstat_named_t arcstat_evict_l2_ineligible; 316 kstat_named_t arcstat_hash_elements; 317 kstat_named_t arcstat_hash_elements_max; 318 kstat_named_t arcstat_hash_collisions; 319 kstat_named_t arcstat_hash_chains; 320 kstat_named_t arcstat_hash_chain_max; 321 kstat_named_t arcstat_p; 322 kstat_named_t arcstat_c; 323 kstat_named_t arcstat_c_min; 324 kstat_named_t arcstat_c_max; 325 kstat_named_t arcstat_size; 326 kstat_named_t arcstat_hdr_size; 327 kstat_named_t arcstat_data_size; 328 kstat_named_t arcstat_other_size; 329 kstat_named_t arcstat_l2_hits; 330 kstat_named_t arcstat_l2_misses; 331 kstat_named_t arcstat_l2_feeds; 332 kstat_named_t arcstat_l2_rw_clash; 333 kstat_named_t arcstat_l2_read_bytes; 334 kstat_named_t arcstat_l2_write_bytes; 335 kstat_named_t arcstat_l2_writes_sent; 336 kstat_named_t arcstat_l2_writes_done; 337 kstat_named_t arcstat_l2_writes_error; 338 kstat_named_t arcstat_l2_writes_hdr_miss; 339 kstat_named_t arcstat_l2_evict_lock_retry; 340 kstat_named_t arcstat_l2_evict_reading; 341 kstat_named_t arcstat_l2_free_on_write; 342 kstat_named_t arcstat_l2_abort_lowmem; 343 kstat_named_t arcstat_l2_cksum_bad; 344 kstat_named_t arcstat_l2_io_error; 345 kstat_named_t arcstat_l2_size; 346 kstat_named_t arcstat_l2_asize; 347 kstat_named_t arcstat_l2_hdr_size; 348 kstat_named_t arcstat_l2_compress_successes; 349 kstat_named_t arcstat_l2_compress_zeros; 350 kstat_named_t arcstat_l2_compress_failures; 351 kstat_named_t arcstat_l2_write_trylock_fail; 352 kstat_named_t arcstat_l2_write_passed_headroom; 353 kstat_named_t arcstat_l2_write_spa_mismatch; 354 kstat_named_t arcstat_l2_write_in_l2; 355 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 356 kstat_named_t arcstat_l2_write_not_cacheable; 357 kstat_named_t arcstat_l2_write_full; 358 kstat_named_t arcstat_l2_write_buffer_iter; 359 kstat_named_t arcstat_l2_write_pios; 360 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 361 kstat_named_t arcstat_l2_write_buffer_list_iter; 362 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 363 kstat_named_t arcstat_memory_throttle_count; 364 kstat_named_t arcstat_duplicate_buffers; 365 kstat_named_t arcstat_duplicate_buffers_size; 366 kstat_named_t arcstat_duplicate_reads; 367} arc_stats_t; 368 369static arc_stats_t arc_stats = { 370 { "hits", KSTAT_DATA_UINT64 }, 371 { "misses", KSTAT_DATA_UINT64 }, 372 { "demand_data_hits", KSTAT_DATA_UINT64 }, 373 { "demand_data_misses", KSTAT_DATA_UINT64 }, 374 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 375 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 376 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 377 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 378 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 379 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 380 { "mru_hits", KSTAT_DATA_UINT64 }, 381 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 382 { "mfu_hits", KSTAT_DATA_UINT64 }, 383 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 384 { "allocated", KSTAT_DATA_UINT64 }, 385 { "deleted", KSTAT_DATA_UINT64 }, 386 { "stolen", KSTAT_DATA_UINT64 }, 387 { "recycle_miss", KSTAT_DATA_UINT64 }, 388 { "mutex_miss", KSTAT_DATA_UINT64 }, 389 { "evict_skip", KSTAT_DATA_UINT64 }, 390 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 391 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 392 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 393 { "hash_elements", KSTAT_DATA_UINT64 }, 394 { "hash_elements_max", KSTAT_DATA_UINT64 }, 395 { "hash_collisions", KSTAT_DATA_UINT64 }, 396 { "hash_chains", KSTAT_DATA_UINT64 }, 397 { "hash_chain_max", KSTAT_DATA_UINT64 }, 398 { "p", KSTAT_DATA_UINT64 }, 399 { "c", KSTAT_DATA_UINT64 }, 400 { "c_min", KSTAT_DATA_UINT64 }, 401 { "c_max", KSTAT_DATA_UINT64 }, 402 { "size", KSTAT_DATA_UINT64 }, 403 { "hdr_size", KSTAT_DATA_UINT64 }, 404 { "data_size", KSTAT_DATA_UINT64 }, 405 { "other_size", KSTAT_DATA_UINT64 }, 406 { "l2_hits", KSTAT_DATA_UINT64 }, 407 { "l2_misses", KSTAT_DATA_UINT64 }, 408 { "l2_feeds", KSTAT_DATA_UINT64 }, 409 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 410 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 411 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 412 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 413 { "l2_writes_done", KSTAT_DATA_UINT64 }, 414 { "l2_writes_error", KSTAT_DATA_UINT64 }, 415 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 416 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 417 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 418 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 419 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 420 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 421 { "l2_io_error", KSTAT_DATA_UINT64 }, 422 { "l2_size", KSTAT_DATA_UINT64 }, 423 { "l2_asize", KSTAT_DATA_UINT64 }, 424 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 425 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 426 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 427 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 428 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 429 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 430 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 431 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 432 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 433 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 434 { "l2_write_full", KSTAT_DATA_UINT64 }, 435 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 436 { "l2_write_pios", KSTAT_DATA_UINT64 }, 437 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 438 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 439 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 440 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 441 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 442 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 443 { "duplicate_reads", KSTAT_DATA_UINT64 } 444}; 445 446#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 447 448#define ARCSTAT_INCR(stat, val) \ 449 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 450 451#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 452#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 453 454#define ARCSTAT_MAX(stat, val) { \ 455 uint64_t m; \ 456 while ((val) > (m = arc_stats.stat.value.ui64) && \ 457 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 458 continue; \ 459} 460 461#define ARCSTAT_MAXSTAT(stat) \ 462 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 463 464/* 465 * We define a macro to allow ARC hits/misses to be easily broken down by 466 * two separate conditions, giving a total of four different subtypes for 467 * each of hits and misses (so eight statistics total). 468 */ 469#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 470 if (cond1) { \ 471 if (cond2) { \ 472 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 473 } else { \ 474 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 475 } \ 476 } else { \ 477 if (cond2) { \ 478 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 479 } else { \ 480 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 481 } \ 482 } 483 484kstat_t *arc_ksp; 485static arc_state_t *arc_anon; 486static arc_state_t *arc_mru; 487static arc_state_t *arc_mru_ghost; 488static arc_state_t *arc_mfu; 489static arc_state_t *arc_mfu_ghost; 490static arc_state_t *arc_l2c_only; 491 492/* 493 * There are several ARC variables that are critical to export as kstats -- 494 * but we don't want to have to grovel around in the kstat whenever we wish to 495 * manipulate them. For these variables, we therefore define them to be in 496 * terms of the statistic variable. This assures that we are not introducing 497 * the possibility of inconsistency by having shadow copies of the variables, 498 * while still allowing the code to be readable. 499 */ 500#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 501#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 502#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 503#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 504#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 505 506#define L2ARC_IS_VALID_COMPRESS(_c_) \ 507 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 508 509static int arc_no_grow; /* Don't try to grow cache size */ 510static uint64_t arc_tempreserve; 511static uint64_t arc_loaned_bytes; 512static uint64_t arc_meta_used; 513static uint64_t arc_meta_limit; 514static uint64_t arc_meta_max = 0; 515SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RD, &arc_meta_used, 0, 516 "ARC metadata used"); 517SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RW, &arc_meta_limit, 0, 518 "ARC metadata limit"); 519 520typedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 521 522typedef struct arc_callback arc_callback_t; 523 524struct arc_callback { 525 void *acb_private; 526 arc_done_func_t *acb_done; 527 arc_buf_t *acb_buf; 528 zio_t *acb_zio_dummy; 529 arc_callback_t *acb_next; 530}; 531 532typedef struct arc_write_callback arc_write_callback_t; 533 534struct arc_write_callback { 535 void *awcb_private; 536 arc_done_func_t *awcb_ready; 537 arc_done_func_t *awcb_physdone; 538 arc_done_func_t *awcb_done; 539 arc_buf_t *awcb_buf; 540}; 541 542struct arc_buf_hdr { 543 /* protected by hash lock */ 544 dva_t b_dva; 545 uint64_t b_birth; 546 uint64_t b_cksum0; 547 548 kmutex_t b_freeze_lock; 549 zio_cksum_t *b_freeze_cksum; 550 void *b_thawed; 551 552 arc_buf_hdr_t *b_hash_next; 553 arc_buf_t *b_buf; 554 uint32_t b_flags; 555 uint32_t b_datacnt; 556 557 arc_callback_t *b_acb; 558 kcondvar_t b_cv; 559 560 /* immutable */ 561 arc_buf_contents_t b_type; 562 uint64_t b_size; 563 uint64_t b_spa; 564 565 /* protected by arc state mutex */ 566 arc_state_t *b_state; 567 list_node_t b_arc_node; 568 569 /* updated atomically */ 570 clock_t b_arc_access; 571 572 /* self protecting */ 573 refcount_t b_refcnt; 574 575 l2arc_buf_hdr_t *b_l2hdr; 576 list_node_t b_l2node; 577}; 578 579static arc_buf_t *arc_eviction_list; 580static kmutex_t arc_eviction_mtx; 581static arc_buf_hdr_t arc_eviction_hdr; 582static void arc_get_data_buf(arc_buf_t *buf); 583static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock); 584static int arc_evict_needed(arc_buf_contents_t type); 585static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes); 586#ifdef illumos 587static void arc_buf_watch(arc_buf_t *buf); 588#endif /* illumos */ 589 590static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab); 591 592#define GHOST_STATE(state) \ 593 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 594 (state) == arc_l2c_only) 595 596/* 597 * Private ARC flags. These flags are private ARC only flags that will show up 598 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can 599 * be passed in as arc_flags in things like arc_read. However, these flags 600 * should never be passed and should only be set by ARC code. When adding new 601 * public flags, make sure not to smash the private ones. 602 */ 603 604#define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */ 605#define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */ 606#define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */ 607#define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */ 608#define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */ 609#define ARC_INDIRECT (1 << 14) /* this is an indirect block */ 610#define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */ 611#define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */ 612#define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */ 613#define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */ 614 615#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE) 616#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS) 617#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR) 618#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH) 619#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ) 620#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE) 621#define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS) 622#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE) 623#define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \ 624 (hdr)->b_l2hdr != NULL) 625#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING) 626#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED) 627#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD) 628 629/* 630 * Other sizes 631 */ 632 633#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 634#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 635 636/* 637 * Hash table routines 638 */ 639 640#define HT_LOCK_PAD CACHE_LINE_SIZE 641 642struct ht_lock { 643 kmutex_t ht_lock; 644#ifdef _KERNEL 645 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 646#endif 647}; 648 649#define BUF_LOCKS 256 650typedef struct buf_hash_table { 651 uint64_t ht_mask; 652 arc_buf_hdr_t **ht_table; 653 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 654} buf_hash_table_t; 655 656static buf_hash_table_t buf_hash_table; 657 658#define BUF_HASH_INDEX(spa, dva, birth) \ 659 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 660#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 661#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 662#define HDR_LOCK(hdr) \ 663 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 664 665uint64_t zfs_crc64_table[256]; 666 667/* 668 * Level 2 ARC 669 */ 670 671#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 672#define L2ARC_HEADROOM 2 /* num of writes */ 673/* 674 * If we discover during ARC scan any buffers to be compressed, we boost 675 * our headroom for the next scanning cycle by this percentage multiple. 676 */ 677#define L2ARC_HEADROOM_BOOST 200 678#define L2ARC_FEED_SECS 1 /* caching interval secs */ 679#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 680 681#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 682#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 683 684/* L2ARC Performance Tunables */ 685uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 686uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 687uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 688uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 689uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 690uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 691boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 692boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 693boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 694 695SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 696 &l2arc_write_max, 0, "max write size"); 697SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 698 &l2arc_write_boost, 0, "extra write during warmup"); 699SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 700 &l2arc_headroom, 0, "number of dev writes"); 701SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 702 &l2arc_feed_secs, 0, "interval seconds"); 703SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 704 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 705 706SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 707 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 708SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 709 &l2arc_feed_again, 0, "turbo warmup"); 710SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 711 &l2arc_norw, 0, "no reads during writes"); 712 713SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 714 &ARC_anon.arcs_size, 0, "size of anonymous state"); 715SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 716 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 717SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 718 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 719 720SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 721 &ARC_mru.arcs_size, 0, "size of mru state"); 722SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 723 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 724SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 725 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 726 727SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 728 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 729SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 730 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 731 "size of metadata in mru ghost state"); 732SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 733 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 734 "size of data in mru ghost state"); 735 736SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 737 &ARC_mfu.arcs_size, 0, "size of mfu state"); 738SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 739 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 740SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 741 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 742 743SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 744 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 745SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 746 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 747 "size of metadata in mfu ghost state"); 748SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 749 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 750 "size of data in mfu ghost state"); 751 752SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 753 &ARC_l2c_only.arcs_size, 0, "size of mru state"); 754 755/* 756 * L2ARC Internals 757 */ 758typedef struct l2arc_dev { 759 vdev_t *l2ad_vdev; /* vdev */ 760 spa_t *l2ad_spa; /* spa */ 761 uint64_t l2ad_hand; /* next write location */ 762 uint64_t l2ad_start; /* first addr on device */ 763 uint64_t l2ad_end; /* last addr on device */ 764 uint64_t l2ad_evict; /* last addr eviction reached */ 765 boolean_t l2ad_first; /* first sweep through */ 766 boolean_t l2ad_writing; /* currently writing */ 767 list_t *l2ad_buflist; /* buffer list */ 768 list_node_t l2ad_node; /* device list node */ 769} l2arc_dev_t; 770 771static list_t L2ARC_dev_list; /* device list */ 772static list_t *l2arc_dev_list; /* device list pointer */ 773static kmutex_t l2arc_dev_mtx; /* device list mutex */ 774static l2arc_dev_t *l2arc_dev_last; /* last device used */ 775static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 776static list_t L2ARC_free_on_write; /* free after write buf list */ 777static list_t *l2arc_free_on_write; /* free after write list ptr */ 778static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 779static uint64_t l2arc_ndev; /* number of devices */ 780 781typedef struct l2arc_read_callback { 782 arc_buf_t *l2rcb_buf; /* read buffer */ 783 spa_t *l2rcb_spa; /* spa */ 784 blkptr_t l2rcb_bp; /* original blkptr */ 785 zbookmark_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 1650static void 1651arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all) 1652{ 1653 arc_buf_t **bufp; 1654 1655 /* free up data associated with the buf */ 1656 if (buf->b_data) { 1657 arc_state_t *state = buf->b_hdr->b_state; 1658 uint64_t size = buf->b_hdr->b_size; 1659 arc_buf_contents_t type = buf->b_hdr->b_type; 1660 1661 arc_cksum_verify(buf); 1662#ifdef illumos 1663 arc_buf_unwatch(buf); 1664#endif /* illumos */ 1665 1666 if (!recycle) { 1667 if (type == ARC_BUFC_METADATA) { 1668 arc_buf_data_free(buf, zio_buf_free); 1669 arc_space_return(size, ARC_SPACE_DATA); 1670 } else { 1671 ASSERT(type == ARC_BUFC_DATA); 1672 arc_buf_data_free(buf, zio_data_buf_free); 1673 ARCSTAT_INCR(arcstat_data_size, -size); 1674 atomic_add_64(&arc_size, -size); 1675 } 1676 } 1677 if (list_link_active(&buf->b_hdr->b_arc_node)) { 1678 uint64_t *cnt = &state->arcs_lsize[type]; 1679 1680 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1681 ASSERT(state != arc_anon); 1682 1683 ASSERT3U(*cnt, >=, size); 1684 atomic_add_64(cnt, -size); 1685 } 1686 ASSERT3U(state->arcs_size, >=, size); 1687 atomic_add_64(&state->arcs_size, -size); 1688 buf->b_data = NULL; 1689 1690 /* 1691 * If we're destroying a duplicate buffer make sure 1692 * that the appropriate statistics are updated. 1693 */ 1694 if (buf->b_hdr->b_datacnt > 1 && 1695 buf->b_hdr->b_type == ARC_BUFC_DATA) { 1696 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1697 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1698 } 1699 ASSERT(buf->b_hdr->b_datacnt > 0); 1700 buf->b_hdr->b_datacnt -= 1; 1701 } 1702 1703 /* only remove the buf if requested */ 1704 if (!all) 1705 return; 1706 1707 /* remove the buf from the hdr list */ 1708 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1709 continue; 1710 *bufp = buf->b_next; 1711 buf->b_next = NULL; 1712 1713 ASSERT(buf->b_efunc == NULL); 1714 1715 /* clean up the buf */ 1716 buf->b_hdr = NULL; 1717 kmem_cache_free(buf_cache, buf); 1718} 1719 1720static void 1721arc_hdr_destroy(arc_buf_hdr_t *hdr) 1722{ 1723 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1724 ASSERT3P(hdr->b_state, ==, arc_anon); 1725 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1726 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1727 1728 if (l2hdr != NULL) { 1729 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1730 /* 1731 * To prevent arc_free() and l2arc_evict() from 1732 * attempting to free the same buffer at the same time, 1733 * a FREE_IN_PROGRESS flag is given to arc_free() to 1734 * give it priority. l2arc_evict() can't destroy this 1735 * header while we are waiting on l2arc_buflist_mtx. 1736 * 1737 * The hdr may be removed from l2ad_buflist before we 1738 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1739 */ 1740 if (!buflist_held) { 1741 mutex_enter(&l2arc_buflist_mtx); 1742 l2hdr = hdr->b_l2hdr; 1743 } 1744 1745 if (l2hdr != NULL) { 1746 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1747 hdr->b_size, 0); 1748 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1749 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1750 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1751 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 1752 -l2hdr->b_asize, 0, 0); 1753 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1754 if (hdr->b_state == arc_l2c_only) 1755 l2arc_hdr_stat_remove(); 1756 hdr->b_l2hdr = NULL; 1757 } 1758 1759 if (!buflist_held) 1760 mutex_exit(&l2arc_buflist_mtx); 1761 } 1762 1763 if (!BUF_EMPTY(hdr)) { 1764 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1765 buf_discard_identity(hdr); 1766 } 1767 while (hdr->b_buf) { 1768 arc_buf_t *buf = hdr->b_buf; 1769 1770 if (buf->b_efunc) { 1771 mutex_enter(&arc_eviction_mtx); 1772 mutex_enter(&buf->b_evict_lock); 1773 ASSERT(buf->b_hdr != NULL); 1774 arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1775 hdr->b_buf = buf->b_next; 1776 buf->b_hdr = &arc_eviction_hdr; 1777 buf->b_next = arc_eviction_list; 1778 arc_eviction_list = buf; 1779 mutex_exit(&buf->b_evict_lock); 1780 mutex_exit(&arc_eviction_mtx); 1781 } else { 1782 arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1783 } 1784 } 1785 if (hdr->b_freeze_cksum != NULL) { 1786 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1787 hdr->b_freeze_cksum = NULL; 1788 } 1789 if (hdr->b_thawed) { 1790 kmem_free(hdr->b_thawed, 1); 1791 hdr->b_thawed = NULL; 1792 } 1793 1794 ASSERT(!list_link_active(&hdr->b_arc_node)); 1795 ASSERT3P(hdr->b_hash_next, ==, NULL); 1796 ASSERT3P(hdr->b_acb, ==, NULL); 1797 kmem_cache_free(hdr_cache, hdr); 1798} 1799 1800void 1801arc_buf_free(arc_buf_t *buf, void *tag) 1802{ 1803 arc_buf_hdr_t *hdr = buf->b_hdr; 1804 int hashed = hdr->b_state != arc_anon; 1805 1806 ASSERT(buf->b_efunc == NULL); 1807 ASSERT(buf->b_data != NULL); 1808 1809 if (hashed) { 1810 kmutex_t *hash_lock = HDR_LOCK(hdr); 1811 1812 mutex_enter(hash_lock); 1813 hdr = buf->b_hdr; 1814 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1815 1816 (void) remove_reference(hdr, hash_lock, tag); 1817 if (hdr->b_datacnt > 1) { 1818 arc_buf_destroy(buf, FALSE, TRUE); 1819 } else { 1820 ASSERT(buf == hdr->b_buf); 1821 ASSERT(buf->b_efunc == NULL); 1822 hdr->b_flags |= ARC_BUF_AVAILABLE; 1823 } 1824 mutex_exit(hash_lock); 1825 } else if (HDR_IO_IN_PROGRESS(hdr)) { 1826 int destroy_hdr; 1827 /* 1828 * We are in the middle of an async write. Don't destroy 1829 * this buffer unless the write completes before we finish 1830 * decrementing the reference count. 1831 */ 1832 mutex_enter(&arc_eviction_mtx); 1833 (void) remove_reference(hdr, NULL, tag); 1834 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1835 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1836 mutex_exit(&arc_eviction_mtx); 1837 if (destroy_hdr) 1838 arc_hdr_destroy(hdr); 1839 } else { 1840 if (remove_reference(hdr, NULL, tag) > 0) 1841 arc_buf_destroy(buf, FALSE, TRUE); 1842 else 1843 arc_hdr_destroy(hdr); 1844 } 1845} 1846 1847boolean_t 1848arc_buf_remove_ref(arc_buf_t *buf, void* tag) 1849{ 1850 arc_buf_hdr_t *hdr = buf->b_hdr; 1851 kmutex_t *hash_lock = HDR_LOCK(hdr); 1852 boolean_t no_callback = (buf->b_efunc == NULL); 1853 1854 if (hdr->b_state == arc_anon) { 1855 ASSERT(hdr->b_datacnt == 1); 1856 arc_buf_free(buf, tag); 1857 return (no_callback); 1858 } 1859 1860 mutex_enter(hash_lock); 1861 hdr = buf->b_hdr; 1862 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1863 ASSERT(hdr->b_state != arc_anon); 1864 ASSERT(buf->b_data != NULL); 1865 1866 (void) remove_reference(hdr, hash_lock, tag); 1867 if (hdr->b_datacnt > 1) { 1868 if (no_callback) 1869 arc_buf_destroy(buf, FALSE, TRUE); 1870 } else if (no_callback) { 1871 ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1872 ASSERT(buf->b_efunc == NULL); 1873 hdr->b_flags |= ARC_BUF_AVAILABLE; 1874 } 1875 ASSERT(no_callback || hdr->b_datacnt > 1 || 1876 refcount_is_zero(&hdr->b_refcnt)); 1877 mutex_exit(hash_lock); 1878 return (no_callback); 1879} 1880 1881int 1882arc_buf_size(arc_buf_t *buf) 1883{ 1884 return (buf->b_hdr->b_size); 1885} 1886 1887/* 1888 * Called from the DMU to determine if the current buffer should be 1889 * evicted. In order to ensure proper locking, the eviction must be initiated 1890 * from the DMU. Return true if the buffer is associated with user data and 1891 * duplicate buffers still exist. 1892 */ 1893boolean_t 1894arc_buf_eviction_needed(arc_buf_t *buf) 1895{ 1896 arc_buf_hdr_t *hdr; 1897 boolean_t evict_needed = B_FALSE; 1898 1899 if (zfs_disable_dup_eviction) 1900 return (B_FALSE); 1901 1902 mutex_enter(&buf->b_evict_lock); 1903 hdr = buf->b_hdr; 1904 if (hdr == NULL) { 1905 /* 1906 * We are in arc_do_user_evicts(); let that function 1907 * perform the eviction. 1908 */ 1909 ASSERT(buf->b_data == NULL); 1910 mutex_exit(&buf->b_evict_lock); 1911 return (B_FALSE); 1912 } else if (buf->b_data == NULL) { 1913 /* 1914 * We have already been added to the arc eviction list; 1915 * recommend eviction. 1916 */ 1917 ASSERT3P(hdr, ==, &arc_eviction_hdr); 1918 mutex_exit(&buf->b_evict_lock); 1919 return (B_TRUE); 1920 } 1921 1922 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 1923 evict_needed = B_TRUE; 1924 1925 mutex_exit(&buf->b_evict_lock); 1926 return (evict_needed); 1927} 1928 1929/* 1930 * Evict buffers from list until we've removed the specified number of 1931 * bytes. Move the removed buffers to the appropriate evict state. 1932 * If the recycle flag is set, then attempt to "recycle" a buffer: 1933 * - look for a buffer to evict that is `bytes' long. 1934 * - return the data block from this buffer rather than freeing it. 1935 * This flag is used by callers that are trying to make space for a 1936 * new buffer in a full arc cache. 1937 * 1938 * This function makes a "best effort". It skips over any buffers 1939 * it can't get a hash_lock on, and so may not catch all candidates. 1940 * It may also return without evicting as much space as requested. 1941 */ 1942static void * 1943arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 1944 arc_buf_contents_t type) 1945{ 1946 arc_state_t *evicted_state; 1947 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 1948 int64_t bytes_remaining; 1949 arc_buf_hdr_t *ab, *ab_prev = NULL; 1950 list_t *evicted_list, *list, *evicted_list_start, *list_start; 1951 kmutex_t *lock, *evicted_lock; 1952 kmutex_t *hash_lock; 1953 boolean_t have_lock; 1954 void *stolen = NULL; 1955 arc_buf_hdr_t marker = { 0 }; 1956 int count = 0; 1957 static int evict_metadata_offset, evict_data_offset; 1958 int i, idx, offset, list_count, lists; 1959 1960 ASSERT(state == arc_mru || state == arc_mfu); 1961 1962 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 1963 1964 if (type == ARC_BUFC_METADATA) { 1965 offset = 0; 1966 list_count = ARC_BUFC_NUMMETADATALISTS; 1967 list_start = &state->arcs_lists[0]; 1968 evicted_list_start = &evicted_state->arcs_lists[0]; 1969 idx = evict_metadata_offset; 1970 } else { 1971 offset = ARC_BUFC_NUMMETADATALISTS; 1972 list_start = &state->arcs_lists[offset]; 1973 evicted_list_start = &evicted_state->arcs_lists[offset]; 1974 list_count = ARC_BUFC_NUMDATALISTS; 1975 idx = evict_data_offset; 1976 } 1977 bytes_remaining = evicted_state->arcs_lsize[type]; 1978 lists = 0; 1979 1980evict_start: 1981 list = &list_start[idx]; 1982 evicted_list = &evicted_list_start[idx]; 1983 lock = ARCS_LOCK(state, (offset + idx)); 1984 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 1985 1986 mutex_enter(lock); 1987 mutex_enter(evicted_lock); 1988 1989 for (ab = list_tail(list); ab; ab = ab_prev) { 1990 ab_prev = list_prev(list, ab); 1991 bytes_remaining -= (ab->b_size * ab->b_datacnt); 1992 /* prefetch buffers have a minimum lifespan */ 1993 if (HDR_IO_IN_PROGRESS(ab) || 1994 (spa && ab->b_spa != spa) || 1995 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 1996 ddi_get_lbolt() - ab->b_arc_access < 1997 arc_min_prefetch_lifespan)) { 1998 skipped++; 1999 continue; 2000 } 2001 /* "lookahead" for better eviction candidate */ 2002 if (recycle && ab->b_size != bytes && 2003 ab_prev && ab_prev->b_size == bytes) 2004 continue; 2005 2006 /* ignore markers */ 2007 if (ab->b_spa == 0) 2008 continue; 2009 2010 /* 2011 * It may take a long time to evict all the bufs requested. 2012 * To avoid blocking all arc activity, periodically drop 2013 * the arcs_mtx and give other threads a chance to run 2014 * before reacquiring the lock. 2015 * 2016 * If we are looking for a buffer to recycle, we are in 2017 * the hot code path, so don't sleep. 2018 */ 2019 if (!recycle && count++ > arc_evict_iterations) { 2020 list_insert_after(list, ab, &marker); 2021 mutex_exit(evicted_lock); 2022 mutex_exit(lock); 2023 kpreempt(KPREEMPT_SYNC); 2024 mutex_enter(lock); 2025 mutex_enter(evicted_lock); 2026 ab_prev = list_prev(list, &marker); 2027 list_remove(list, &marker); 2028 count = 0; 2029 continue; 2030 } 2031 2032 hash_lock = HDR_LOCK(ab); 2033 have_lock = MUTEX_HELD(hash_lock); 2034 if (have_lock || mutex_tryenter(hash_lock)) { 2035 ASSERT0(refcount_count(&ab->b_refcnt)); 2036 ASSERT(ab->b_datacnt > 0); 2037 while (ab->b_buf) { 2038 arc_buf_t *buf = ab->b_buf; 2039 if (!mutex_tryenter(&buf->b_evict_lock)) { 2040 missed += 1; 2041 break; 2042 } 2043 if (buf->b_data) { 2044 bytes_evicted += ab->b_size; 2045 if (recycle && ab->b_type == type && 2046 ab->b_size == bytes && 2047 !HDR_L2_WRITING(ab)) { 2048 stolen = buf->b_data; 2049 recycle = FALSE; 2050 } 2051 } 2052 if (buf->b_efunc) { 2053 mutex_enter(&arc_eviction_mtx); 2054 arc_buf_destroy(buf, 2055 buf->b_data == stolen, FALSE); 2056 ab->b_buf = buf->b_next; 2057 buf->b_hdr = &arc_eviction_hdr; 2058 buf->b_next = arc_eviction_list; 2059 arc_eviction_list = buf; 2060 mutex_exit(&arc_eviction_mtx); 2061 mutex_exit(&buf->b_evict_lock); 2062 } else { 2063 mutex_exit(&buf->b_evict_lock); 2064 arc_buf_destroy(buf, 2065 buf->b_data == stolen, TRUE); 2066 } 2067 } 2068 2069 if (ab->b_l2hdr) { 2070 ARCSTAT_INCR(arcstat_evict_l2_cached, 2071 ab->b_size); 2072 } else { 2073 if (l2arc_write_eligible(ab->b_spa, ab)) { 2074 ARCSTAT_INCR(arcstat_evict_l2_eligible, 2075 ab->b_size); 2076 } else { 2077 ARCSTAT_INCR( 2078 arcstat_evict_l2_ineligible, 2079 ab->b_size); 2080 } 2081 } 2082 2083 if (ab->b_datacnt == 0) { 2084 arc_change_state(evicted_state, ab, hash_lock); 2085 ASSERT(HDR_IN_HASH_TABLE(ab)); 2086 ab->b_flags |= ARC_IN_HASH_TABLE; 2087 ab->b_flags &= ~ARC_BUF_AVAILABLE; 2088 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 2089 } 2090 if (!have_lock) 2091 mutex_exit(hash_lock); 2092 if (bytes >= 0 && bytes_evicted >= bytes) 2093 break; 2094 if (bytes_remaining > 0) { 2095 mutex_exit(evicted_lock); 2096 mutex_exit(lock); 2097 idx = ((idx + 1) & (list_count - 1)); 2098 lists++; 2099 goto evict_start; 2100 } 2101 } else { 2102 missed += 1; 2103 } 2104 } 2105 2106 mutex_exit(evicted_lock); 2107 mutex_exit(lock); 2108 2109 idx = ((idx + 1) & (list_count - 1)); 2110 lists++; 2111 2112 if (bytes_evicted < bytes) { 2113 if (lists < list_count) 2114 goto evict_start; 2115 else 2116 dprintf("only evicted %lld bytes from %x", 2117 (longlong_t)bytes_evicted, state); 2118 } 2119 if (type == ARC_BUFC_METADATA) 2120 evict_metadata_offset = idx; 2121 else 2122 evict_data_offset = idx; 2123 2124 if (skipped) 2125 ARCSTAT_INCR(arcstat_evict_skip, skipped); 2126 2127 if (missed) 2128 ARCSTAT_INCR(arcstat_mutex_miss, missed); 2129 2130 /* 2131 * Note: we have just evicted some data into the ghost state, 2132 * potentially putting the ghost size over the desired size. Rather 2133 * that evicting from the ghost list in this hot code path, leave 2134 * this chore to the arc_reclaim_thread(). 2135 */ 2136 2137 if (stolen) 2138 ARCSTAT_BUMP(arcstat_stolen); 2139 return (stolen); 2140} 2141 2142/* 2143 * Remove buffers from list until we've removed the specified number of 2144 * bytes. Destroy the buffers that are removed. 2145 */ 2146static void 2147arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2148{ 2149 arc_buf_hdr_t *ab, *ab_prev; 2150 arc_buf_hdr_t marker = { 0 }; 2151 list_t *list, *list_start; 2152 kmutex_t *hash_lock, *lock; 2153 uint64_t bytes_deleted = 0; 2154 uint64_t bufs_skipped = 0; 2155 int count = 0; 2156 static int evict_offset; 2157 int list_count, idx = evict_offset; 2158 int offset, lists = 0; 2159 2160 ASSERT(GHOST_STATE(state)); 2161 2162 /* 2163 * data lists come after metadata lists 2164 */ 2165 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2166 list_count = ARC_BUFC_NUMDATALISTS; 2167 offset = ARC_BUFC_NUMMETADATALISTS; 2168 2169evict_start: 2170 list = &list_start[idx]; 2171 lock = ARCS_LOCK(state, idx + offset); 2172 2173 mutex_enter(lock); 2174 for (ab = list_tail(list); ab; ab = ab_prev) { 2175 ab_prev = list_prev(list, ab); 2176 if (ab->b_type > ARC_BUFC_NUMTYPES) 2177 panic("invalid ab=%p", (void *)ab); 2178 if (spa && ab->b_spa != spa) 2179 continue; 2180 2181 /* ignore markers */ 2182 if (ab->b_spa == 0) 2183 continue; 2184 2185 hash_lock = HDR_LOCK(ab); 2186 /* caller may be trying to modify this buffer, skip it */ 2187 if (MUTEX_HELD(hash_lock)) 2188 continue; 2189 2190 /* 2191 * It may take a long time to evict all the bufs requested. 2192 * To avoid blocking all arc activity, periodically drop 2193 * the arcs_mtx and give other threads a chance to run 2194 * before reacquiring the lock. 2195 */ 2196 if (count++ > arc_evict_iterations) { 2197 list_insert_after(list, ab, &marker); 2198 mutex_exit(lock); 2199 kpreempt(KPREEMPT_SYNC); 2200 mutex_enter(lock); 2201 ab_prev = list_prev(list, &marker); 2202 list_remove(list, &marker); 2203 count = 0; 2204 continue; 2205 } 2206 if (mutex_tryenter(hash_lock)) { 2207 ASSERT(!HDR_IO_IN_PROGRESS(ab)); 2208 ASSERT(ab->b_buf == NULL); 2209 ARCSTAT_BUMP(arcstat_deleted); 2210 bytes_deleted += ab->b_size; 2211 2212 if (ab->b_l2hdr != NULL) { 2213 /* 2214 * This buffer is cached on the 2nd Level ARC; 2215 * don't destroy the header. 2216 */ 2217 arc_change_state(arc_l2c_only, ab, hash_lock); 2218 mutex_exit(hash_lock); 2219 } else { 2220 arc_change_state(arc_anon, ab, hash_lock); 2221 mutex_exit(hash_lock); 2222 arc_hdr_destroy(ab); 2223 } 2224 2225 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 2226 if (bytes >= 0 && bytes_deleted >= bytes) 2227 break; 2228 } else if (bytes < 0) { 2229 /* 2230 * Insert a list marker and then wait for the 2231 * hash lock to become available. Once its 2232 * available, restart from where we left off. 2233 */ 2234 list_insert_after(list, ab, &marker); 2235 mutex_exit(lock); 2236 mutex_enter(hash_lock); 2237 mutex_exit(hash_lock); 2238 mutex_enter(lock); 2239 ab_prev = list_prev(list, &marker); 2240 list_remove(list, &marker); 2241 } else { 2242 bufs_skipped += 1; 2243 } 2244 2245 } 2246 mutex_exit(lock); 2247 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2248 lists++; 2249 2250 if (lists < list_count) 2251 goto evict_start; 2252 2253 evict_offset = idx; 2254 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2255 (bytes < 0 || bytes_deleted < bytes)) { 2256 list_start = &state->arcs_lists[0]; 2257 list_count = ARC_BUFC_NUMMETADATALISTS; 2258 offset = lists = 0; 2259 goto evict_start; 2260 } 2261 2262 if (bufs_skipped) { 2263 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2264 ASSERT(bytes >= 0); 2265 } 2266 2267 if (bytes_deleted < bytes) 2268 dprintf("only deleted %lld bytes from %p", 2269 (longlong_t)bytes_deleted, state); 2270} 2271 2272static void 2273arc_adjust(void) 2274{ 2275 int64_t adjustment, delta; 2276 2277 /* 2278 * Adjust MRU size 2279 */ 2280 2281 adjustment = MIN((int64_t)(arc_size - arc_c), 2282 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2283 arc_p)); 2284 2285 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2286 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2287 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2288 adjustment -= delta; 2289 } 2290 2291 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2292 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2293 (void) arc_evict(arc_mru, 0, delta, FALSE, 2294 ARC_BUFC_METADATA); 2295 } 2296 2297 /* 2298 * Adjust MFU size 2299 */ 2300 2301 adjustment = arc_size - arc_c; 2302 2303 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2304 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2305 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2306 adjustment -= delta; 2307 } 2308 2309 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2310 int64_t delta = MIN(adjustment, 2311 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2312 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2313 ARC_BUFC_METADATA); 2314 } 2315 2316 /* 2317 * Adjust ghost lists 2318 */ 2319 2320 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2321 2322 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2323 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2324 arc_evict_ghost(arc_mru_ghost, 0, delta); 2325 } 2326 2327 adjustment = 2328 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2329 2330 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2331 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2332 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2333 } 2334} 2335 2336static void 2337arc_do_user_evicts(void) 2338{ 2339 static arc_buf_t *tmp_arc_eviction_list; 2340 2341 /* 2342 * Move list over to avoid LOR 2343 */ 2344restart: 2345 mutex_enter(&arc_eviction_mtx); 2346 tmp_arc_eviction_list = arc_eviction_list; 2347 arc_eviction_list = NULL; 2348 mutex_exit(&arc_eviction_mtx); 2349 2350 while (tmp_arc_eviction_list != NULL) { 2351 arc_buf_t *buf = tmp_arc_eviction_list; 2352 tmp_arc_eviction_list = buf->b_next; 2353 mutex_enter(&buf->b_evict_lock); 2354 buf->b_hdr = NULL; 2355 mutex_exit(&buf->b_evict_lock); 2356 2357 if (buf->b_efunc != NULL) 2358 VERIFY(buf->b_efunc(buf) == 0); 2359 2360 buf->b_efunc = NULL; 2361 buf->b_private = NULL; 2362 kmem_cache_free(buf_cache, buf); 2363 } 2364 2365 if (arc_eviction_list != NULL) 2366 goto restart; 2367} 2368 2369/* 2370 * Flush all *evictable* data from the cache for the given spa. 2371 * NOTE: this will not touch "active" (i.e. referenced) data. 2372 */ 2373void 2374arc_flush(spa_t *spa) 2375{ 2376 uint64_t guid = 0; 2377 2378 if (spa) 2379 guid = spa_load_guid(spa); 2380 2381 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2382 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2383 if (spa) 2384 break; 2385 } 2386 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2387 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2388 if (spa) 2389 break; 2390 } 2391 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2392 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2393 if (spa) 2394 break; 2395 } 2396 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2397 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2398 if (spa) 2399 break; 2400 } 2401 2402 arc_evict_ghost(arc_mru_ghost, guid, -1); 2403 arc_evict_ghost(arc_mfu_ghost, guid, -1); 2404 2405 mutex_enter(&arc_reclaim_thr_lock); 2406 arc_do_user_evicts(); 2407 mutex_exit(&arc_reclaim_thr_lock); 2408 ASSERT(spa || arc_eviction_list == NULL); 2409} 2410 2411void 2412arc_shrink(void) 2413{ 2414 if (arc_c > arc_c_min) { 2415 uint64_t to_free; 2416 2417#ifdef _KERNEL 2418 to_free = arc_c >> arc_shrink_shift; 2419#else 2420 to_free = arc_c >> arc_shrink_shift; 2421#endif 2422 if (arc_c > arc_c_min + to_free) 2423 atomic_add_64(&arc_c, -to_free); 2424 else 2425 arc_c = arc_c_min; 2426 2427 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2428 if (arc_c > arc_size) 2429 arc_c = MAX(arc_size, arc_c_min); 2430 if (arc_p > arc_c) 2431 arc_p = (arc_c >> 1); 2432 ASSERT(arc_c >= arc_c_min); 2433 ASSERT((int64_t)arc_p >= 0); 2434 } 2435 2436 if (arc_size > arc_c) 2437 arc_adjust(); 2438} 2439 2440static int needfree = 0; 2441 2442static int 2443arc_reclaim_needed(void) 2444{ 2445 2446#ifdef _KERNEL 2447 2448 if (needfree) 2449 return (1); 2450 2451 /* 2452 * Cooperate with pagedaemon when it's time for it to scan 2453 * and reclaim some pages. 2454 */ 2455 if (vm_paging_needed()) 2456 return (1); 2457 2458#ifdef sun 2459 /* 2460 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2461 */ 2462 extra = desfree; 2463 2464 /* 2465 * check that we're out of range of the pageout scanner. It starts to 2466 * schedule paging if freemem is less than lotsfree and needfree. 2467 * lotsfree is the high-water mark for pageout, and needfree is the 2468 * number of needed free pages. We add extra pages here to make sure 2469 * the scanner doesn't start up while we're freeing memory. 2470 */ 2471 if (freemem < lotsfree + needfree + extra) 2472 return (1); 2473 2474 /* 2475 * check to make sure that swapfs has enough space so that anon 2476 * reservations can still succeed. anon_resvmem() checks that the 2477 * availrmem is greater than swapfs_minfree, and the number of reserved 2478 * swap pages. We also add a bit of extra here just to prevent 2479 * circumstances from getting really dire. 2480 */ 2481 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2482 return (1); 2483 2484#if defined(__i386) 2485 /* 2486 * If we're on an i386 platform, it's possible that we'll exhaust the 2487 * kernel heap space before we ever run out of available physical 2488 * memory. Most checks of the size of the heap_area compare against 2489 * tune.t_minarmem, which is the minimum available real memory that we 2490 * can have in the system. However, this is generally fixed at 25 pages 2491 * which is so low that it's useless. In this comparison, we seek to 2492 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2493 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2494 * free) 2495 */ 2496 if (btop(vmem_size(heap_arena, VMEM_FREE)) < 2497 (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2)) 2498 return (1); 2499#endif 2500#else /* !sun */ 2501 if (kmem_used() > (kmem_size() * 3) / 4) 2502 return (1); 2503#endif /* sun */ 2504 2505#else 2506 if (spa_get_random(100) == 0) 2507 return (1); 2508#endif 2509 return (0); 2510} 2511 2512extern kmem_cache_t *zio_buf_cache[]; 2513extern kmem_cache_t *zio_data_buf_cache[]; 2514 2515static void 2516arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2517{ 2518 size_t i; 2519 kmem_cache_t *prev_cache = NULL; 2520 kmem_cache_t *prev_data_cache = NULL; 2521 2522#ifdef _KERNEL 2523 if (arc_meta_used >= arc_meta_limit) { 2524 /* 2525 * We are exceeding our meta-data cache limit. 2526 * Purge some DNLC entries to release holds on meta-data. 2527 */ 2528 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2529 } 2530#if defined(__i386) 2531 /* 2532 * Reclaim unused memory from all kmem caches. 2533 */ 2534 kmem_reap(); 2535#endif 2536#endif 2537 2538 /* 2539 * An aggressive reclamation will shrink the cache size as well as 2540 * reap free buffers from the arc kmem caches. 2541 */ 2542 if (strat == ARC_RECLAIM_AGGR) 2543 arc_shrink(); 2544 2545 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2546 if (zio_buf_cache[i] != prev_cache) { 2547 prev_cache = zio_buf_cache[i]; 2548 kmem_cache_reap_now(zio_buf_cache[i]); 2549 } 2550 if (zio_data_buf_cache[i] != prev_data_cache) { 2551 prev_data_cache = zio_data_buf_cache[i]; 2552 kmem_cache_reap_now(zio_data_buf_cache[i]); 2553 } 2554 } 2555 kmem_cache_reap_now(buf_cache); 2556 kmem_cache_reap_now(hdr_cache); 2557} 2558 2559static void 2560arc_reclaim_thread(void *dummy __unused) 2561{ 2562 clock_t growtime = 0; 2563 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2564 callb_cpr_t cpr; 2565 2566 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2567 2568 mutex_enter(&arc_reclaim_thr_lock); 2569 while (arc_thread_exit == 0) { 2570 if (arc_reclaim_needed()) { 2571 2572 if (arc_no_grow) { 2573 if (last_reclaim == ARC_RECLAIM_CONS) { 2574 last_reclaim = ARC_RECLAIM_AGGR; 2575 } else { 2576 last_reclaim = ARC_RECLAIM_CONS; 2577 } 2578 } else { 2579 arc_no_grow = TRUE; 2580 last_reclaim = ARC_RECLAIM_AGGR; 2581 membar_producer(); 2582 } 2583 2584 /* reset the growth delay for every reclaim */ 2585 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2586 2587 if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2588 /* 2589 * If needfree is TRUE our vm_lowmem hook 2590 * was called and in that case we must free some 2591 * memory, so switch to aggressive mode. 2592 */ 2593 arc_no_grow = TRUE; 2594 last_reclaim = ARC_RECLAIM_AGGR; 2595 } 2596 arc_kmem_reap_now(last_reclaim); 2597 arc_warm = B_TRUE; 2598 2599 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2600 arc_no_grow = FALSE; 2601 } 2602 2603 arc_adjust(); 2604 2605 if (arc_eviction_list != NULL) 2606 arc_do_user_evicts(); 2607 2608#ifdef _KERNEL 2609 if (needfree) { 2610 needfree = 0; 2611 wakeup(&needfree); 2612 } 2613#endif 2614 2615 /* block until needed, or one second, whichever is shorter */ 2616 CALLB_CPR_SAFE_BEGIN(&cpr); 2617 (void) cv_timedwait(&arc_reclaim_thr_cv, 2618 &arc_reclaim_thr_lock, hz); 2619 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2620 } 2621 2622 arc_thread_exit = 0; 2623 cv_broadcast(&arc_reclaim_thr_cv); 2624 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2625 thread_exit(); 2626} 2627 2628/* 2629 * Adapt arc info given the number of bytes we are trying to add and 2630 * the state that we are comming from. This function is only called 2631 * when we are adding new content to the cache. 2632 */ 2633static void 2634arc_adapt(int bytes, arc_state_t *state) 2635{ 2636 int mult; 2637 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2638 2639 if (state == arc_l2c_only) 2640 return; 2641 2642 ASSERT(bytes > 0); 2643 /* 2644 * Adapt the target size of the MRU list: 2645 * - if we just hit in the MRU ghost list, then increase 2646 * the target size of the MRU list. 2647 * - if we just hit in the MFU ghost list, then increase 2648 * the target size of the MFU list by decreasing the 2649 * target size of the MRU list. 2650 */ 2651 if (state == arc_mru_ghost) { 2652 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2653 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2654 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2655 2656 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2657 } else if (state == arc_mfu_ghost) { 2658 uint64_t delta; 2659 2660 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2661 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2662 mult = MIN(mult, 10); 2663 2664 delta = MIN(bytes * mult, arc_p); 2665 arc_p = MAX(arc_p_min, arc_p - delta); 2666 } 2667 ASSERT((int64_t)arc_p >= 0); 2668 2669 if (arc_reclaim_needed()) { 2670 cv_signal(&arc_reclaim_thr_cv); 2671 return; 2672 } 2673 2674 if (arc_no_grow) 2675 return; 2676 2677 if (arc_c >= arc_c_max) 2678 return; 2679 2680 /* 2681 * If we're within (2 * maxblocksize) bytes of the target 2682 * cache size, increment the target cache size 2683 */ 2684 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2685 atomic_add_64(&arc_c, (int64_t)bytes); 2686 if (arc_c > arc_c_max) 2687 arc_c = arc_c_max; 2688 else if (state == arc_anon) 2689 atomic_add_64(&arc_p, (int64_t)bytes); 2690 if (arc_p > arc_c) 2691 arc_p = arc_c; 2692 } 2693 ASSERT((int64_t)arc_p >= 0); 2694} 2695 2696/* 2697 * Check if the cache has reached its limits and eviction is required 2698 * prior to insert. 2699 */ 2700static int 2701arc_evict_needed(arc_buf_contents_t type) 2702{ 2703 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2704 return (1); 2705 2706#ifdef sun 2707#ifdef _KERNEL 2708 /* 2709 * If zio data pages are being allocated out of a separate heap segment, 2710 * then enforce that the size of available vmem for this area remains 2711 * above about 1/32nd free. 2712 */ 2713 if (type == ARC_BUFC_DATA && zio_arena != NULL && 2714 vmem_size(zio_arena, VMEM_FREE) < 2715 (vmem_size(zio_arena, VMEM_ALLOC) >> 5)) 2716 return (1); 2717#endif 2718#endif /* sun */ 2719 2720 if (arc_reclaim_needed()) 2721 return (1); 2722 2723 return (arc_size > arc_c); 2724} 2725 2726/* 2727 * The buffer, supplied as the first argument, needs a data block. 2728 * So, if we are at cache max, determine which cache should be victimized. 2729 * We have the following cases: 2730 * 2731 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2732 * In this situation if we're out of space, but the resident size of the MFU is 2733 * under the limit, victimize the MFU cache to satisfy this insertion request. 2734 * 2735 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2736 * Here, we've used up all of the available space for the MRU, so we need to 2737 * evict from our own cache instead. Evict from the set of resident MRU 2738 * entries. 2739 * 2740 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2741 * c minus p represents the MFU space in the cache, since p is the size of the 2742 * cache that is dedicated to the MRU. In this situation there's still space on 2743 * the MFU side, so the MRU side needs to be victimized. 2744 * 2745 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2746 * MFU's resident set is consuming more space than it has been allotted. In 2747 * this situation, we must victimize our own cache, the MFU, for this insertion. 2748 */ 2749static void 2750arc_get_data_buf(arc_buf_t *buf) 2751{ 2752 arc_state_t *state = buf->b_hdr->b_state; 2753 uint64_t size = buf->b_hdr->b_size; 2754 arc_buf_contents_t type = buf->b_hdr->b_type; 2755 2756 arc_adapt(size, state); 2757 2758 /* 2759 * We have not yet reached cache maximum size, 2760 * just allocate a new buffer. 2761 */ 2762 if (!arc_evict_needed(type)) { 2763 if (type == ARC_BUFC_METADATA) { 2764 buf->b_data = zio_buf_alloc(size); 2765 arc_space_consume(size, ARC_SPACE_DATA); 2766 } else { 2767 ASSERT(type == ARC_BUFC_DATA); 2768 buf->b_data = zio_data_buf_alloc(size); 2769 ARCSTAT_INCR(arcstat_data_size, size); 2770 atomic_add_64(&arc_size, size); 2771 } 2772 goto out; 2773 } 2774 2775 /* 2776 * If we are prefetching from the mfu ghost list, this buffer 2777 * will end up on the mru list; so steal space from there. 2778 */ 2779 if (state == arc_mfu_ghost) 2780 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2781 else if (state == arc_mru_ghost) 2782 state = arc_mru; 2783 2784 if (state == arc_mru || state == arc_anon) { 2785 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2786 state = (arc_mfu->arcs_lsize[type] >= size && 2787 arc_p > mru_used) ? arc_mfu : arc_mru; 2788 } else { 2789 /* MFU cases */ 2790 uint64_t mfu_space = arc_c - arc_p; 2791 state = (arc_mru->arcs_lsize[type] >= size && 2792 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2793 } 2794 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2795 if (type == ARC_BUFC_METADATA) { 2796 buf->b_data = zio_buf_alloc(size); 2797 arc_space_consume(size, ARC_SPACE_DATA); 2798 } else { 2799 ASSERT(type == ARC_BUFC_DATA); 2800 buf->b_data = zio_data_buf_alloc(size); 2801 ARCSTAT_INCR(arcstat_data_size, size); 2802 atomic_add_64(&arc_size, size); 2803 } 2804 ARCSTAT_BUMP(arcstat_recycle_miss); 2805 } 2806 ASSERT(buf->b_data != NULL); 2807out: 2808 /* 2809 * Update the state size. Note that ghost states have a 2810 * "ghost size" and so don't need to be updated. 2811 */ 2812 if (!GHOST_STATE(buf->b_hdr->b_state)) { 2813 arc_buf_hdr_t *hdr = buf->b_hdr; 2814 2815 atomic_add_64(&hdr->b_state->arcs_size, size); 2816 if (list_link_active(&hdr->b_arc_node)) { 2817 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2818 atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2819 } 2820 /* 2821 * If we are growing the cache, and we are adding anonymous 2822 * data, and we have outgrown arc_p, update arc_p 2823 */ 2824 if (arc_size < arc_c && hdr->b_state == arc_anon && 2825 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2826 arc_p = MIN(arc_c, arc_p + size); 2827 } 2828 ARCSTAT_BUMP(arcstat_allocated); 2829} 2830 2831/* 2832 * This routine is called whenever a buffer is accessed. 2833 * NOTE: the hash lock is dropped in this function. 2834 */ 2835static void 2836arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2837{ 2838 clock_t now; 2839 2840 ASSERT(MUTEX_HELD(hash_lock)); 2841 2842 if (buf->b_state == arc_anon) { 2843 /* 2844 * This buffer is not in the cache, and does not 2845 * appear in our "ghost" list. Add the new buffer 2846 * to the MRU state. 2847 */ 2848 2849 ASSERT(buf->b_arc_access == 0); 2850 buf->b_arc_access = ddi_get_lbolt(); 2851 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2852 arc_change_state(arc_mru, buf, hash_lock); 2853 2854 } else if (buf->b_state == arc_mru) { 2855 now = ddi_get_lbolt(); 2856 2857 /* 2858 * If this buffer is here because of a prefetch, then either: 2859 * - clear the flag if this is a "referencing" read 2860 * (any subsequent access will bump this into the MFU state). 2861 * or 2862 * - move the buffer to the head of the list if this is 2863 * another prefetch (to make it less likely to be evicted). 2864 */ 2865 if ((buf->b_flags & ARC_PREFETCH) != 0) { 2866 if (refcount_count(&buf->b_refcnt) == 0) { 2867 ASSERT(list_link_active(&buf->b_arc_node)); 2868 } else { 2869 buf->b_flags &= ~ARC_PREFETCH; 2870 ARCSTAT_BUMP(arcstat_mru_hits); 2871 } 2872 buf->b_arc_access = now; 2873 return; 2874 } 2875 2876 /* 2877 * This buffer has been "accessed" only once so far, 2878 * but it is still in the cache. Move it to the MFU 2879 * state. 2880 */ 2881 if (now > buf->b_arc_access + ARC_MINTIME) { 2882 /* 2883 * More than 125ms have passed since we 2884 * instantiated this buffer. Move it to the 2885 * most frequently used state. 2886 */ 2887 buf->b_arc_access = now; 2888 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2889 arc_change_state(arc_mfu, buf, hash_lock); 2890 } 2891 ARCSTAT_BUMP(arcstat_mru_hits); 2892 } else if (buf->b_state == arc_mru_ghost) { 2893 arc_state_t *new_state; 2894 /* 2895 * This buffer has been "accessed" recently, but 2896 * was evicted from the cache. Move it to the 2897 * MFU state. 2898 */ 2899 2900 if (buf->b_flags & ARC_PREFETCH) { 2901 new_state = arc_mru; 2902 if (refcount_count(&buf->b_refcnt) > 0) 2903 buf->b_flags &= ~ARC_PREFETCH; 2904 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2905 } else { 2906 new_state = arc_mfu; 2907 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2908 } 2909 2910 buf->b_arc_access = ddi_get_lbolt(); 2911 arc_change_state(new_state, buf, hash_lock); 2912 2913 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 2914 } else if (buf->b_state == arc_mfu) { 2915 /* 2916 * This buffer has been accessed more than once and is 2917 * still in the cache. Keep it in the MFU state. 2918 * 2919 * NOTE: an add_reference() that occurred when we did 2920 * the arc_read() will have kicked this off the list. 2921 * If it was a prefetch, we will explicitly move it to 2922 * the head of the list now. 2923 */ 2924 if ((buf->b_flags & ARC_PREFETCH) != 0) { 2925 ASSERT(refcount_count(&buf->b_refcnt) == 0); 2926 ASSERT(list_link_active(&buf->b_arc_node)); 2927 } 2928 ARCSTAT_BUMP(arcstat_mfu_hits); 2929 buf->b_arc_access = ddi_get_lbolt(); 2930 } else if (buf->b_state == arc_mfu_ghost) { 2931 arc_state_t *new_state = arc_mfu; 2932 /* 2933 * This buffer has been accessed more than once but has 2934 * been evicted from the cache. Move it back to the 2935 * MFU state. 2936 */ 2937 2938 if (buf->b_flags & ARC_PREFETCH) { 2939 /* 2940 * This is a prefetch access... 2941 * move this block back to the MRU state. 2942 */ 2943 ASSERT0(refcount_count(&buf->b_refcnt)); 2944 new_state = arc_mru; 2945 } 2946 2947 buf->b_arc_access = ddi_get_lbolt(); 2948 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2949 arc_change_state(new_state, buf, hash_lock); 2950 2951 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 2952 } else if (buf->b_state == arc_l2c_only) { 2953 /* 2954 * This buffer is on the 2nd Level ARC. 2955 */ 2956 2957 buf->b_arc_access = ddi_get_lbolt(); 2958 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2959 arc_change_state(arc_mfu, buf, hash_lock); 2960 } else { 2961 ASSERT(!"invalid arc state"); 2962 } 2963} 2964 2965/* a generic arc_done_func_t which you can use */ 2966/* ARGSUSED */ 2967void 2968arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 2969{ 2970 if (zio == NULL || zio->io_error == 0) 2971 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 2972 VERIFY(arc_buf_remove_ref(buf, arg)); 2973} 2974 2975/* a generic arc_done_func_t */ 2976void 2977arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 2978{ 2979 arc_buf_t **bufp = arg; 2980 if (zio && zio->io_error) { 2981 VERIFY(arc_buf_remove_ref(buf, arg)); 2982 *bufp = NULL; 2983 } else { 2984 *bufp = buf; 2985 ASSERT(buf->b_data); 2986 } 2987} 2988 2989static void 2990arc_read_done(zio_t *zio) 2991{ 2992 arc_buf_hdr_t *hdr; 2993 arc_buf_t *buf; 2994 arc_buf_t *abuf; /* buffer we're assigning to callback */ 2995 kmutex_t *hash_lock = NULL; 2996 arc_callback_t *callback_list, *acb; 2997 int freeable = FALSE; 2998 2999 buf = zio->io_private; 3000 hdr = buf->b_hdr; 3001 3002 /* 3003 * The hdr was inserted into hash-table and removed from lists 3004 * prior to starting I/O. We should find this header, since 3005 * it's in the hash table, and it should be legit since it's 3006 * not possible to evict it during the I/O. The only possible 3007 * reason for it not to be found is if we were freed during the 3008 * read. 3009 */ 3010 if (HDR_IN_HASH_TABLE(hdr)) { 3011 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3012 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3013 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3014 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3015 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3016 3017 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3018 &hash_lock); 3019 3020 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3021 hash_lock == NULL) || 3022 (found == hdr && 3023 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3024 (found == hdr && HDR_L2_READING(hdr))); 3025 } 3026 3027 hdr->b_flags &= ~ARC_L2_EVICTED; 3028 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 3029 hdr->b_flags &= ~ARC_L2CACHE; 3030 3031 /* byteswap if necessary */ 3032 callback_list = hdr->b_acb; 3033 ASSERT(callback_list != NULL); 3034 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3035 dmu_object_byteswap_t bswap = 3036 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3037 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3038 byteswap_uint64_array : 3039 dmu_ot_byteswap[bswap].ob_func; 3040 func(buf->b_data, hdr->b_size); 3041 } 3042 3043 arc_cksum_compute(buf, B_FALSE); 3044#ifdef illumos 3045 arc_buf_watch(buf); 3046#endif /* illumos */ 3047 3048 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3049 /* 3050 * Only call arc_access on anonymous buffers. This is because 3051 * if we've issued an I/O for an evicted buffer, we've already 3052 * called arc_access (to prevent any simultaneous readers from 3053 * getting confused). 3054 */ 3055 arc_access(hdr, hash_lock); 3056 } 3057 3058 /* create copies of the data buffer for the callers */ 3059 abuf = buf; 3060 for (acb = callback_list; acb; acb = acb->acb_next) { 3061 if (acb->acb_done) { 3062 if (abuf == NULL) { 3063 ARCSTAT_BUMP(arcstat_duplicate_reads); 3064 abuf = arc_buf_clone(buf); 3065 } 3066 acb->acb_buf = abuf; 3067 abuf = NULL; 3068 } 3069 } 3070 hdr->b_acb = NULL; 3071 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3072 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3073 if (abuf == buf) { 3074 ASSERT(buf->b_efunc == NULL); 3075 ASSERT(hdr->b_datacnt == 1); 3076 hdr->b_flags |= ARC_BUF_AVAILABLE; 3077 } 3078 3079 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3080 3081 if (zio->io_error != 0) { 3082 hdr->b_flags |= ARC_IO_ERROR; 3083 if (hdr->b_state != arc_anon) 3084 arc_change_state(arc_anon, hdr, hash_lock); 3085 if (HDR_IN_HASH_TABLE(hdr)) 3086 buf_hash_remove(hdr); 3087 freeable = refcount_is_zero(&hdr->b_refcnt); 3088 } 3089 3090 /* 3091 * Broadcast before we drop the hash_lock to avoid the possibility 3092 * that the hdr (and hence the cv) might be freed before we get to 3093 * the cv_broadcast(). 3094 */ 3095 cv_broadcast(&hdr->b_cv); 3096 3097 if (hash_lock) { 3098 mutex_exit(hash_lock); 3099 } else { 3100 /* 3101 * This block was freed while we waited for the read to 3102 * complete. It has been removed from the hash table and 3103 * moved to the anonymous state (so that it won't show up 3104 * in the cache). 3105 */ 3106 ASSERT3P(hdr->b_state, ==, arc_anon); 3107 freeable = refcount_is_zero(&hdr->b_refcnt); 3108 } 3109 3110 /* execute each callback and free its structure */ 3111 while ((acb = callback_list) != NULL) { 3112 if (acb->acb_done) 3113 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3114 3115 if (acb->acb_zio_dummy != NULL) { 3116 acb->acb_zio_dummy->io_error = zio->io_error; 3117 zio_nowait(acb->acb_zio_dummy); 3118 } 3119 3120 callback_list = acb->acb_next; 3121 kmem_free(acb, sizeof (arc_callback_t)); 3122 } 3123 3124 if (freeable) 3125 arc_hdr_destroy(hdr); 3126} 3127 3128/* 3129 * "Read" the block block at the specified DVA (in bp) via the 3130 * cache. If the block is found in the cache, invoke the provided 3131 * callback immediately and return. Note that the `zio' parameter 3132 * in the callback will be NULL in this case, since no IO was 3133 * required. If the block is not in the cache pass the read request 3134 * on to the spa with a substitute callback function, so that the 3135 * requested block will be added to the cache. 3136 * 3137 * If a read request arrives for a block that has a read in-progress, 3138 * either wait for the in-progress read to complete (and return the 3139 * results); or, if this is a read with a "done" func, add a record 3140 * to the read to invoke the "done" func when the read completes, 3141 * and return; or just return. 3142 * 3143 * arc_read_done() will invoke all the requested "done" functions 3144 * for readers of this block. 3145 */ 3146int 3147arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3148 void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags, 3149 const zbookmark_phys_t *zb) 3150{ 3151 arc_buf_hdr_t *hdr = NULL; 3152 arc_buf_t *buf = NULL; 3153 kmutex_t *hash_lock = NULL; 3154 zio_t *rzio; 3155 uint64_t guid = spa_load_guid(spa); 3156 3157 ASSERT(!BP_IS_EMBEDDED(bp) || 3158 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3159 3160top: 3161 if (!BP_IS_EMBEDDED(bp)) { 3162 /* 3163 * Embedded BP's have no DVA and require no I/O to "read". 3164 * Create an anonymous arc buf to back it. 3165 */ 3166 hdr = buf_hash_find(guid, bp, &hash_lock); 3167 } 3168 3169 if (hdr != NULL && hdr->b_datacnt > 0) { 3170 3171 *arc_flags |= ARC_CACHED; 3172 3173 if (HDR_IO_IN_PROGRESS(hdr)) { 3174 3175 if (*arc_flags & ARC_WAIT) { 3176 cv_wait(&hdr->b_cv, hash_lock); 3177 mutex_exit(hash_lock); 3178 goto top; 3179 } 3180 ASSERT(*arc_flags & ARC_NOWAIT); 3181 3182 if (done) { 3183 arc_callback_t *acb = NULL; 3184 3185 acb = kmem_zalloc(sizeof (arc_callback_t), 3186 KM_SLEEP); 3187 acb->acb_done = done; 3188 acb->acb_private = private; 3189 if (pio != NULL) 3190 acb->acb_zio_dummy = zio_null(pio, 3191 spa, NULL, NULL, NULL, zio_flags); 3192 3193 ASSERT(acb->acb_done != NULL); 3194 acb->acb_next = hdr->b_acb; 3195 hdr->b_acb = acb; 3196 add_reference(hdr, hash_lock, private); 3197 mutex_exit(hash_lock); 3198 return (0); 3199 } 3200 mutex_exit(hash_lock); 3201 return (0); 3202 } 3203 3204 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3205 3206 if (done) { 3207 add_reference(hdr, hash_lock, private); 3208 /* 3209 * If this block is already in use, create a new 3210 * copy of the data so that we will be guaranteed 3211 * that arc_release() will always succeed. 3212 */ 3213 buf = hdr->b_buf; 3214 ASSERT(buf); 3215 ASSERT(buf->b_data); 3216 if (HDR_BUF_AVAILABLE(hdr)) { 3217 ASSERT(buf->b_efunc == NULL); 3218 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3219 } else { 3220 buf = arc_buf_clone(buf); 3221 } 3222 3223 } else if (*arc_flags & ARC_PREFETCH && 3224 refcount_count(&hdr->b_refcnt) == 0) { 3225 hdr->b_flags |= ARC_PREFETCH; 3226 } 3227 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3228 arc_access(hdr, hash_lock); 3229 if (*arc_flags & ARC_L2CACHE) 3230 hdr->b_flags |= ARC_L2CACHE; 3231 if (*arc_flags & ARC_L2COMPRESS) 3232 hdr->b_flags |= ARC_L2COMPRESS; 3233 mutex_exit(hash_lock); 3234 ARCSTAT_BUMP(arcstat_hits); 3235 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3236 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3237 data, metadata, hits); 3238 3239 if (done) 3240 done(NULL, buf, private); 3241 } else { 3242 uint64_t size = BP_GET_LSIZE(bp); 3243 arc_callback_t *acb; 3244 vdev_t *vd = NULL; 3245 uint64_t addr = 0; 3246 boolean_t devw = B_FALSE; 3247 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3248 uint64_t b_asize = 0; 3249 3250 if (hdr == NULL) { 3251 /* this block is not in the cache */ 3252 arc_buf_hdr_t *exists = NULL; 3253 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3254 buf = arc_buf_alloc(spa, size, private, type); 3255 hdr = buf->b_hdr; 3256 if (!BP_IS_EMBEDDED(bp)) { 3257 hdr->b_dva = *BP_IDENTITY(bp); 3258 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3259 hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3260 exists = buf_hash_insert(hdr, &hash_lock); 3261 } 3262 if (exists != NULL) { 3263 /* somebody beat us to the hash insert */ 3264 mutex_exit(hash_lock); 3265 buf_discard_identity(hdr); 3266 (void) arc_buf_remove_ref(buf, private); 3267 goto top; /* restart the IO request */ 3268 } 3269 /* if this is a prefetch, we don't have a reference */ 3270 if (*arc_flags & ARC_PREFETCH) { 3271 (void) remove_reference(hdr, hash_lock, 3272 private); 3273 hdr->b_flags |= ARC_PREFETCH; 3274 } 3275 if (*arc_flags & ARC_L2CACHE) 3276 hdr->b_flags |= ARC_L2CACHE; 3277 if (*arc_flags & ARC_L2COMPRESS) 3278 hdr->b_flags |= ARC_L2COMPRESS; 3279 if (BP_GET_LEVEL(bp) > 0) 3280 hdr->b_flags |= ARC_INDIRECT; 3281 } else { 3282 /* this block is in the ghost cache */ 3283 ASSERT(GHOST_STATE(hdr->b_state)); 3284 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3285 ASSERT0(refcount_count(&hdr->b_refcnt)); 3286 ASSERT(hdr->b_buf == NULL); 3287 3288 /* if this is a prefetch, we don't have a reference */ 3289 if (*arc_flags & ARC_PREFETCH) 3290 hdr->b_flags |= ARC_PREFETCH; 3291 else 3292 add_reference(hdr, hash_lock, private); 3293 if (*arc_flags & ARC_L2CACHE) 3294 hdr->b_flags |= ARC_L2CACHE; 3295 if (*arc_flags & ARC_L2COMPRESS) 3296 hdr->b_flags |= ARC_L2COMPRESS; 3297 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3298 buf->b_hdr = hdr; 3299 buf->b_data = NULL; 3300 buf->b_efunc = NULL; 3301 buf->b_private = NULL; 3302 buf->b_next = NULL; 3303 hdr->b_buf = buf; 3304 ASSERT(hdr->b_datacnt == 0); 3305 hdr->b_datacnt = 1; 3306 arc_get_data_buf(buf); 3307 arc_access(hdr, hash_lock); 3308 } 3309 3310 ASSERT(!GHOST_STATE(hdr->b_state)); 3311 3312 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3313 acb->acb_done = done; 3314 acb->acb_private = private; 3315 3316 ASSERT(hdr->b_acb == NULL); 3317 hdr->b_acb = acb; 3318 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3319 3320 if (hdr->b_l2hdr != NULL && 3321 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3322 devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3323 addr = hdr->b_l2hdr->b_daddr; 3324 b_compress = hdr->b_l2hdr->b_compress; 3325 b_asize = hdr->b_l2hdr->b_asize; 3326 /* 3327 * Lock out device removal. 3328 */ 3329 if (vdev_is_dead(vd) || 3330 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3331 vd = NULL; 3332 } 3333 3334 if (hash_lock != NULL) 3335 mutex_exit(hash_lock); 3336 3337 /* 3338 * At this point, we have a level 1 cache miss. Try again in 3339 * L2ARC if possible. 3340 */ 3341 ASSERT3U(hdr->b_size, ==, size); 3342 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3343 uint64_t, size, zbookmark_phys_t *, zb); 3344 ARCSTAT_BUMP(arcstat_misses); 3345 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3346 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3347 data, metadata, misses); 3348#ifdef _KERNEL 3349 curthread->td_ru.ru_inblock++; 3350#endif 3351 3352 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3353 /* 3354 * Read from the L2ARC if the following are true: 3355 * 1. The L2ARC vdev was previously cached. 3356 * 2. This buffer still has L2ARC metadata. 3357 * 3. This buffer isn't currently writing to the L2ARC. 3358 * 4. The L2ARC entry wasn't evicted, which may 3359 * also have invalidated the vdev. 3360 * 5. This isn't prefetch and l2arc_noprefetch is set. 3361 */ 3362 if (hdr->b_l2hdr != NULL && 3363 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3364 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3365 l2arc_read_callback_t *cb; 3366 3367 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3368 ARCSTAT_BUMP(arcstat_l2_hits); 3369 3370 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3371 KM_SLEEP); 3372 cb->l2rcb_buf = buf; 3373 cb->l2rcb_spa = spa; 3374 cb->l2rcb_bp = *bp; 3375 cb->l2rcb_zb = *zb; 3376 cb->l2rcb_flags = zio_flags; 3377 cb->l2rcb_compress = b_compress; 3378 3379 ASSERT(addr >= VDEV_LABEL_START_SIZE && 3380 addr + size < vd->vdev_psize - 3381 VDEV_LABEL_END_SIZE); 3382 3383 /* 3384 * l2arc read. The SCL_L2ARC lock will be 3385 * released by l2arc_read_done(). 3386 * Issue a null zio if the underlying buffer 3387 * was squashed to zero size by compression. 3388 */ 3389 if (b_compress == ZIO_COMPRESS_EMPTY) { 3390 rzio = zio_null(pio, spa, vd, 3391 l2arc_read_done, cb, 3392 zio_flags | ZIO_FLAG_DONT_CACHE | 3393 ZIO_FLAG_CANFAIL | 3394 ZIO_FLAG_DONT_PROPAGATE | 3395 ZIO_FLAG_DONT_RETRY); 3396 } else { 3397 rzio = zio_read_phys(pio, vd, addr, 3398 b_asize, buf->b_data, 3399 ZIO_CHECKSUM_OFF, 3400 l2arc_read_done, cb, priority, 3401 zio_flags | ZIO_FLAG_DONT_CACHE | 3402 ZIO_FLAG_CANFAIL | 3403 ZIO_FLAG_DONT_PROPAGATE | 3404 ZIO_FLAG_DONT_RETRY, B_FALSE); 3405 } 3406 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3407 zio_t *, rzio); 3408 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3409 3410 if (*arc_flags & ARC_NOWAIT) { 3411 zio_nowait(rzio); 3412 return (0); 3413 } 3414 3415 ASSERT(*arc_flags & ARC_WAIT); 3416 if (zio_wait(rzio) == 0) 3417 return (0); 3418 3419 /* l2arc read error; goto zio_read() */ 3420 } else { 3421 DTRACE_PROBE1(l2arc__miss, 3422 arc_buf_hdr_t *, hdr); 3423 ARCSTAT_BUMP(arcstat_l2_misses); 3424 if (HDR_L2_WRITING(hdr)) 3425 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3426 spa_config_exit(spa, SCL_L2ARC, vd); 3427 } 3428 } else { 3429 if (vd != NULL) 3430 spa_config_exit(spa, SCL_L2ARC, vd); 3431 if (l2arc_ndev != 0) { 3432 DTRACE_PROBE1(l2arc__miss, 3433 arc_buf_hdr_t *, hdr); 3434 ARCSTAT_BUMP(arcstat_l2_misses); 3435 } 3436 } 3437 3438 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3439 arc_read_done, buf, priority, zio_flags, zb); 3440 3441 if (*arc_flags & ARC_WAIT) 3442 return (zio_wait(rzio)); 3443 3444 ASSERT(*arc_flags & ARC_NOWAIT); 3445 zio_nowait(rzio); 3446 } 3447 return (0); 3448} 3449 3450void 3451arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3452{ 3453 ASSERT(buf->b_hdr != NULL); 3454 ASSERT(buf->b_hdr->b_state != arc_anon); 3455 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3456 ASSERT(buf->b_efunc == NULL); 3457 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3458 3459 buf->b_efunc = func; 3460 buf->b_private = private; 3461} 3462 3463/* 3464 * Notify the arc that a block was freed, and thus will never be used again. 3465 */ 3466void 3467arc_freed(spa_t *spa, const blkptr_t *bp) 3468{ 3469 arc_buf_hdr_t *hdr; 3470 kmutex_t *hash_lock; 3471 uint64_t guid = spa_load_guid(spa); 3472 3473 ASSERT(!BP_IS_EMBEDDED(bp)); 3474 3475 hdr = buf_hash_find(guid, bp, &hash_lock); 3476 if (hdr == NULL) 3477 return; 3478 if (HDR_BUF_AVAILABLE(hdr)) { 3479 arc_buf_t *buf = hdr->b_buf; 3480 add_reference(hdr, hash_lock, FTAG); 3481 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3482 mutex_exit(hash_lock); 3483 3484 arc_release(buf, FTAG); 3485 (void) arc_buf_remove_ref(buf, FTAG); 3486 } else { 3487 mutex_exit(hash_lock); 3488 } 3489 3490} 3491 3492/* 3493 * This is used by the DMU to let the ARC know that a buffer is 3494 * being evicted, so the ARC should clean up. If this arc buf 3495 * is not yet in the evicted state, it will be put there. 3496 */ 3497int 3498arc_buf_evict(arc_buf_t *buf) 3499{ 3500 arc_buf_hdr_t *hdr; 3501 kmutex_t *hash_lock; 3502 arc_buf_t **bufp; 3503 list_t *list, *evicted_list; 3504 kmutex_t *lock, *evicted_lock; 3505 3506 mutex_enter(&buf->b_evict_lock); 3507 hdr = buf->b_hdr; 3508 if (hdr == NULL) { 3509 /* 3510 * We are in arc_do_user_evicts(). 3511 */ 3512 ASSERT(buf->b_data == NULL); 3513 mutex_exit(&buf->b_evict_lock); 3514 return (0); 3515 } else if (buf->b_data == NULL) { 3516 arc_buf_t copy = *buf; /* structure assignment */ 3517 /* 3518 * We are on the eviction list; process this buffer now 3519 * but let arc_do_user_evicts() do the reaping. 3520 */ 3521 buf->b_efunc = NULL; 3522 mutex_exit(&buf->b_evict_lock); 3523 VERIFY(copy.b_efunc(©) == 0); 3524 return (1); 3525 } 3526 hash_lock = HDR_LOCK(hdr); 3527 mutex_enter(hash_lock); 3528 hdr = buf->b_hdr; 3529 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3530 3531 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3532 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3533 3534 /* 3535 * Pull this buffer off of the hdr 3536 */ 3537 bufp = &hdr->b_buf; 3538 while (*bufp != buf) 3539 bufp = &(*bufp)->b_next; 3540 *bufp = buf->b_next; 3541 3542 ASSERT(buf->b_data != NULL); 3543 arc_buf_destroy(buf, FALSE, FALSE); 3544 3545 if (hdr->b_datacnt == 0) { 3546 arc_state_t *old_state = hdr->b_state; 3547 arc_state_t *evicted_state; 3548 3549 ASSERT(hdr->b_buf == NULL); 3550 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 3551 3552 evicted_state = 3553 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3554 3555 get_buf_info(hdr, old_state, &list, &lock); 3556 get_buf_info(hdr, evicted_state, &evicted_list, &evicted_lock); 3557 mutex_enter(lock); 3558 mutex_enter(evicted_lock); 3559 3560 arc_change_state(evicted_state, hdr, hash_lock); 3561 ASSERT(HDR_IN_HASH_TABLE(hdr)); 3562 hdr->b_flags |= ARC_IN_HASH_TABLE; 3563 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3564 3565 mutex_exit(evicted_lock); 3566 mutex_exit(lock); 3567 } 3568 mutex_exit(hash_lock); 3569 mutex_exit(&buf->b_evict_lock); 3570 3571 VERIFY(buf->b_efunc(buf) == 0); 3572 buf->b_efunc = NULL; 3573 buf->b_private = NULL; 3574 buf->b_hdr = NULL; 3575 buf->b_next = NULL; 3576 kmem_cache_free(buf_cache, buf); 3577 return (1); 3578} 3579 3580/* 3581 * Release this buffer from the cache, making it an anonymous buffer. This 3582 * must be done after a read and prior to modifying the buffer contents. 3583 * If the buffer has more than one reference, we must make 3584 * a new hdr for the buffer. 3585 */ 3586void 3587arc_release(arc_buf_t *buf, void *tag) 3588{ 3589 arc_buf_hdr_t *hdr; 3590 kmutex_t *hash_lock = NULL; 3591 l2arc_buf_hdr_t *l2hdr; 3592 uint64_t buf_size; 3593 3594 /* 3595 * It would be nice to assert that if it's DMU metadata (level > 3596 * 0 || it's the dnode file), then it must be syncing context. 3597 * But we don't know that information at this level. 3598 */ 3599 3600 mutex_enter(&buf->b_evict_lock); 3601 hdr = buf->b_hdr; 3602 3603 /* this buffer is not on any list */ 3604 ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3605 3606 if (hdr->b_state == arc_anon) { 3607 /* this buffer is already released */ 3608 ASSERT(buf->b_efunc == NULL); 3609 } else { 3610 hash_lock = HDR_LOCK(hdr); 3611 mutex_enter(hash_lock); 3612 hdr = buf->b_hdr; 3613 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3614 } 3615 3616 l2hdr = hdr->b_l2hdr; 3617 if (l2hdr) { 3618 mutex_enter(&l2arc_buflist_mtx); 3619 hdr->b_l2hdr = NULL; 3620 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3621 } 3622 buf_size = hdr->b_size; 3623 3624 /* 3625 * Do we have more than one buf? 3626 */ 3627 if (hdr->b_datacnt > 1) { 3628 arc_buf_hdr_t *nhdr; 3629 arc_buf_t **bufp; 3630 uint64_t blksz = hdr->b_size; 3631 uint64_t spa = hdr->b_spa; 3632 arc_buf_contents_t type = hdr->b_type; 3633 uint32_t flags = hdr->b_flags; 3634 3635 ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3636 /* 3637 * Pull the data off of this hdr and attach it to 3638 * a new anonymous hdr. 3639 */ 3640 (void) remove_reference(hdr, hash_lock, tag); 3641 bufp = &hdr->b_buf; 3642 while (*bufp != buf) 3643 bufp = &(*bufp)->b_next; 3644 *bufp = buf->b_next; 3645 buf->b_next = NULL; 3646 3647 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3648 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3649 if (refcount_is_zero(&hdr->b_refcnt)) { 3650 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3651 ASSERT3U(*size, >=, hdr->b_size); 3652 atomic_add_64(size, -hdr->b_size); 3653 } 3654 3655 /* 3656 * We're releasing a duplicate user data buffer, update 3657 * our statistics accordingly. 3658 */ 3659 if (hdr->b_type == ARC_BUFC_DATA) { 3660 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3661 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3662 -hdr->b_size); 3663 } 3664 hdr->b_datacnt -= 1; 3665 arc_cksum_verify(buf); 3666#ifdef illumos 3667 arc_buf_unwatch(buf); 3668#endif /* illumos */ 3669 3670 mutex_exit(hash_lock); 3671 3672 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3673 nhdr->b_size = blksz; 3674 nhdr->b_spa = spa; 3675 nhdr->b_type = type; 3676 nhdr->b_buf = buf; 3677 nhdr->b_state = arc_anon; 3678 nhdr->b_arc_access = 0; 3679 nhdr->b_flags = flags & ARC_L2_WRITING; 3680 nhdr->b_l2hdr = NULL; 3681 nhdr->b_datacnt = 1; 3682 nhdr->b_freeze_cksum = NULL; 3683 (void) refcount_add(&nhdr->b_refcnt, tag); 3684 buf->b_hdr = nhdr; 3685 mutex_exit(&buf->b_evict_lock); 3686 atomic_add_64(&arc_anon->arcs_size, blksz); 3687 } else { 3688 mutex_exit(&buf->b_evict_lock); 3689 ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3690 ASSERT(!list_link_active(&hdr->b_arc_node)); 3691 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3692 if (hdr->b_state != arc_anon) 3693 arc_change_state(arc_anon, hdr, hash_lock); 3694 hdr->b_arc_access = 0; 3695 if (hash_lock) 3696 mutex_exit(hash_lock); 3697 3698 buf_discard_identity(hdr); 3699 arc_buf_thaw(buf); 3700 } 3701 buf->b_efunc = NULL; 3702 buf->b_private = NULL; 3703 3704 if (l2hdr) { 3705 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3706 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 3707 -l2hdr->b_asize, 0, 0); 3708 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3709 hdr->b_size, 0); 3710 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3711 ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3712 mutex_exit(&l2arc_buflist_mtx); 3713 } 3714} 3715 3716int 3717arc_released(arc_buf_t *buf) 3718{ 3719 int released; 3720 3721 mutex_enter(&buf->b_evict_lock); 3722 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3723 mutex_exit(&buf->b_evict_lock); 3724 return (released); 3725} 3726 3727int 3728arc_has_callback(arc_buf_t *buf) 3729{ 3730 int callback; 3731 3732 mutex_enter(&buf->b_evict_lock); 3733 callback = (buf->b_efunc != NULL); 3734 mutex_exit(&buf->b_evict_lock); 3735 return (callback); 3736} 3737 3738#ifdef ZFS_DEBUG 3739int 3740arc_referenced(arc_buf_t *buf) 3741{ 3742 int referenced; 3743 3744 mutex_enter(&buf->b_evict_lock); 3745 referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3746 mutex_exit(&buf->b_evict_lock); 3747 return (referenced); 3748} 3749#endif 3750 3751static void 3752arc_write_ready(zio_t *zio) 3753{ 3754 arc_write_callback_t *callback = zio->io_private; 3755 arc_buf_t *buf = callback->awcb_buf; 3756 arc_buf_hdr_t *hdr = buf->b_hdr; 3757 3758 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3759 callback->awcb_ready(zio, buf, callback->awcb_private); 3760 3761 /* 3762 * If the IO is already in progress, then this is a re-write 3763 * attempt, so we need to thaw and re-compute the cksum. 3764 * It is the responsibility of the callback to handle the 3765 * accounting for any re-write attempt. 3766 */ 3767 if (HDR_IO_IN_PROGRESS(hdr)) { 3768 mutex_enter(&hdr->b_freeze_lock); 3769 if (hdr->b_freeze_cksum != NULL) { 3770 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3771 hdr->b_freeze_cksum = NULL; 3772 } 3773 mutex_exit(&hdr->b_freeze_lock); 3774 } 3775 arc_cksum_compute(buf, B_FALSE); 3776 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3777} 3778 3779/* 3780 * The SPA calls this callback for each physical write that happens on behalf 3781 * of a logical write. See the comment in dbuf_write_physdone() for details. 3782 */ 3783static void 3784arc_write_physdone(zio_t *zio) 3785{ 3786 arc_write_callback_t *cb = zio->io_private; 3787 if (cb->awcb_physdone != NULL) 3788 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3789} 3790 3791static void 3792arc_write_done(zio_t *zio) 3793{ 3794 arc_write_callback_t *callback = zio->io_private; 3795 arc_buf_t *buf = callback->awcb_buf; 3796 arc_buf_hdr_t *hdr = buf->b_hdr; 3797 3798 ASSERT(hdr->b_acb == NULL); 3799 3800 if (zio->io_error == 0) { 3801 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 3802 buf_discard_identity(hdr); 3803 } else { 3804 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3805 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3806 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3807 } 3808 } else { 3809 ASSERT(BUF_EMPTY(hdr)); 3810 } 3811 3812 /* 3813 * If the block to be written was all-zero or compressed enough to be 3814 * embedded in the BP, no write was performed so there will be no 3815 * dva/birth/checksum. The buffer must therefore remain anonymous 3816 * (and uncached). 3817 */ 3818 if (!BUF_EMPTY(hdr)) { 3819 arc_buf_hdr_t *exists; 3820 kmutex_t *hash_lock; 3821 3822 ASSERT(zio->io_error == 0); 3823 3824 arc_cksum_verify(buf); 3825 3826 exists = buf_hash_insert(hdr, &hash_lock); 3827 if (exists) { 3828 /* 3829 * This can only happen if we overwrite for 3830 * sync-to-convergence, because we remove 3831 * buffers from the hash table when we arc_free(). 3832 */ 3833 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3834 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3835 panic("bad overwrite, hdr=%p exists=%p", 3836 (void *)hdr, (void *)exists); 3837 ASSERT(refcount_is_zero(&exists->b_refcnt)); 3838 arc_change_state(arc_anon, exists, hash_lock); 3839 mutex_exit(hash_lock); 3840 arc_hdr_destroy(exists); 3841 exists = buf_hash_insert(hdr, &hash_lock); 3842 ASSERT3P(exists, ==, NULL); 3843 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3844 /* nopwrite */ 3845 ASSERT(zio->io_prop.zp_nopwrite); 3846 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3847 panic("bad nopwrite, hdr=%p exists=%p", 3848 (void *)hdr, (void *)exists); 3849 } else { 3850 /* Dedup */ 3851 ASSERT(hdr->b_datacnt == 1); 3852 ASSERT(hdr->b_state == arc_anon); 3853 ASSERT(BP_GET_DEDUP(zio->io_bp)); 3854 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3855 } 3856 } 3857 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3858 /* if it's not anon, we are doing a scrub */ 3859 if (!exists && hdr->b_state == arc_anon) 3860 arc_access(hdr, hash_lock); 3861 mutex_exit(hash_lock); 3862 } else { 3863 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3864 } 3865 3866 ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3867 callback->awcb_done(zio, buf, callback->awcb_private); 3868 3869 kmem_free(callback, sizeof (arc_write_callback_t)); 3870} 3871 3872zio_t * 3873arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 3874 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 3875 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 3876 arc_done_func_t *done, void *private, zio_priority_t priority, 3877 int zio_flags, const zbookmark_phys_t *zb) 3878{ 3879 arc_buf_hdr_t *hdr = buf->b_hdr; 3880 arc_write_callback_t *callback; 3881 zio_t *zio; 3882 3883 ASSERT(ready != NULL); 3884 ASSERT(done != NULL); 3885 ASSERT(!HDR_IO_ERROR(hdr)); 3886 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3887 ASSERT(hdr->b_acb == NULL); 3888 if (l2arc) 3889 hdr->b_flags |= ARC_L2CACHE; 3890 if (l2arc_compress) 3891 hdr->b_flags |= ARC_L2COMPRESS; 3892 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 3893 callback->awcb_ready = ready; 3894 callback->awcb_physdone = physdone; 3895 callback->awcb_done = done; 3896 callback->awcb_private = private; 3897 callback->awcb_buf = buf; 3898 3899 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 3900 arc_write_ready, arc_write_physdone, arc_write_done, callback, 3901 priority, zio_flags, zb); 3902 3903 return (zio); 3904} 3905 3906static int 3907arc_memory_throttle(uint64_t reserve, uint64_t txg) 3908{ 3909#ifdef _KERNEL 3910 uint64_t available_memory = 3911 ptoa((uintmax_t)cnt.v_free_count + cnt.v_cache_count); 3912 static uint64_t page_load = 0; 3913 static uint64_t last_txg = 0; 3914 3915#ifdef sun 3916#if defined(__i386) 3917 available_memory = 3918 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 3919#endif 3920#endif /* sun */ 3921 3922 if (cnt.v_free_count + cnt.v_cache_count > 3923 (uint64_t)physmem * arc_lotsfree_percent / 100) 3924 return (0); 3925 3926 if (txg > last_txg) { 3927 last_txg = txg; 3928 page_load = 0; 3929 } 3930 /* 3931 * If we are in pageout, we know that memory is already tight, 3932 * the arc is already going to be evicting, so we just want to 3933 * continue to let page writes occur as quickly as possible. 3934 */ 3935 if (curproc == pageproc) { 3936 if (page_load > available_memory / 4) 3937 return (SET_ERROR(ERESTART)); 3938 /* Note: reserve is inflated, so we deflate */ 3939 page_load += reserve / 8; 3940 return (0); 3941 } else if (page_load > 0 && arc_reclaim_needed()) { 3942 /* memory is low, delay before restarting */ 3943 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3944 return (SET_ERROR(EAGAIN)); 3945 } 3946 page_load = 0; 3947#endif 3948 return (0); 3949} 3950 3951void 3952arc_tempreserve_clear(uint64_t reserve) 3953{ 3954 atomic_add_64(&arc_tempreserve, -reserve); 3955 ASSERT((int64_t)arc_tempreserve >= 0); 3956} 3957 3958int 3959arc_tempreserve_space(uint64_t reserve, uint64_t txg) 3960{ 3961 int error; 3962 uint64_t anon_size; 3963 3964 if (reserve > arc_c/4 && !arc_no_grow) 3965 arc_c = MIN(arc_c_max, reserve * 4); 3966 if (reserve > arc_c) 3967 return (SET_ERROR(ENOMEM)); 3968 3969 /* 3970 * Don't count loaned bufs as in flight dirty data to prevent long 3971 * network delays from blocking transactions that are ready to be 3972 * assigned to a txg. 3973 */ 3974 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 3975 3976 /* 3977 * Writes will, almost always, require additional memory allocations 3978 * in order to compress/encrypt/etc the data. We therefore need to 3979 * make sure that there is sufficient available memory for this. 3980 */ 3981 error = arc_memory_throttle(reserve, txg); 3982 if (error != 0) 3983 return (error); 3984 3985 /* 3986 * Throttle writes when the amount of dirty data in the cache 3987 * gets too large. We try to keep the cache less than half full 3988 * of dirty blocks so that our sync times don't grow too large. 3989 * Note: if two requests come in concurrently, we might let them 3990 * both succeed, when one of them should fail. Not a huge deal. 3991 */ 3992 3993 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 3994 anon_size > arc_c / 4) { 3995 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 3996 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 3997 arc_tempreserve>>10, 3998 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 3999 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 4000 reserve>>10, arc_c>>10); 4001 return (SET_ERROR(ERESTART)); 4002 } 4003 atomic_add_64(&arc_tempreserve, reserve); 4004 return (0); 4005} 4006 4007static kmutex_t arc_lowmem_lock; 4008#ifdef _KERNEL 4009static eventhandler_tag arc_event_lowmem = NULL; 4010 4011static void 4012arc_lowmem(void *arg __unused, int howto __unused) 4013{ 4014 4015 /* Serialize access via arc_lowmem_lock. */ 4016 mutex_enter(&arc_lowmem_lock); 4017 mutex_enter(&arc_reclaim_thr_lock); 4018 needfree = 1; 4019 cv_signal(&arc_reclaim_thr_cv); 4020 4021 /* 4022 * It is unsafe to block here in arbitrary threads, because we can come 4023 * here from ARC itself and may hold ARC locks and thus risk a deadlock 4024 * with ARC reclaim thread. 4025 */ 4026 if (curproc == pageproc) { 4027 while (needfree) 4028 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4029 } 4030 mutex_exit(&arc_reclaim_thr_lock); 4031 mutex_exit(&arc_lowmem_lock); 4032} 4033#endif 4034 4035void 4036arc_init(void) 4037{ 4038 int i, prefetch_tunable_set = 0; 4039 4040 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4041 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4042 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4043 4044 /* Convert seconds to clock ticks */ 4045 arc_min_prefetch_lifespan = 1 * hz; 4046 4047 /* Start out with 1/8 of all memory */ 4048 arc_c = kmem_size() / 8; 4049 4050#ifdef sun 4051#ifdef _KERNEL 4052 /* 4053 * On architectures where the physical memory can be larger 4054 * than the addressable space (intel in 32-bit mode), we may 4055 * need to limit the cache to 1/8 of VM size. 4056 */ 4057 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4058#endif 4059#endif /* sun */ 4060 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4061 arc_c_min = MAX(arc_c / 4, 64<<18); 4062 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4063 if (arc_c * 8 >= 1<<30) 4064 arc_c_max = (arc_c * 8) - (1<<30); 4065 else 4066 arc_c_max = arc_c_min; 4067 arc_c_max = MAX(arc_c * 5, arc_c_max); 4068 4069#ifdef _KERNEL 4070 /* 4071 * Allow the tunables to override our calculations if they are 4072 * reasonable (ie. over 16MB) 4073 */ 4074 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size()) 4075 arc_c_max = zfs_arc_max; 4076 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max) 4077 arc_c_min = zfs_arc_min; 4078#endif 4079 4080 arc_c = arc_c_max; 4081 arc_p = (arc_c >> 1); 4082 4083 /* limit meta-data to 1/4 of the arc capacity */ 4084 arc_meta_limit = arc_c_max / 4; 4085 4086 /* Allow the tunable to override if it is reasonable */ 4087 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4088 arc_meta_limit = zfs_arc_meta_limit; 4089 4090 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4091 arc_c_min = arc_meta_limit / 2; 4092 4093 if (zfs_arc_grow_retry > 0) 4094 arc_grow_retry = zfs_arc_grow_retry; 4095 4096 if (zfs_arc_shrink_shift > 0) 4097 arc_shrink_shift = zfs_arc_shrink_shift; 4098 4099 if (zfs_arc_p_min_shift > 0) 4100 arc_p_min_shift = zfs_arc_p_min_shift; 4101 4102 /* if kmem_flags are set, lets try to use less memory */ 4103 if (kmem_debugging()) 4104 arc_c = arc_c / 2; 4105 if (arc_c < arc_c_min) 4106 arc_c = arc_c_min; 4107 4108 zfs_arc_min = arc_c_min; 4109 zfs_arc_max = arc_c_max; 4110 4111 arc_anon = &ARC_anon; 4112 arc_mru = &ARC_mru; 4113 arc_mru_ghost = &ARC_mru_ghost; 4114 arc_mfu = &ARC_mfu; 4115 arc_mfu_ghost = &ARC_mfu_ghost; 4116 arc_l2c_only = &ARC_l2c_only; 4117 arc_size = 0; 4118 4119 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4120 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4121 NULL, MUTEX_DEFAULT, NULL); 4122 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4123 NULL, MUTEX_DEFAULT, NULL); 4124 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4125 NULL, MUTEX_DEFAULT, NULL); 4126 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4127 NULL, MUTEX_DEFAULT, NULL); 4128 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4129 NULL, MUTEX_DEFAULT, NULL); 4130 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4131 NULL, MUTEX_DEFAULT, NULL); 4132 4133 list_create(&arc_mru->arcs_lists[i], 4134 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4135 list_create(&arc_mru_ghost->arcs_lists[i], 4136 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4137 list_create(&arc_mfu->arcs_lists[i], 4138 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4139 list_create(&arc_mfu_ghost->arcs_lists[i], 4140 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4141 list_create(&arc_mfu_ghost->arcs_lists[i], 4142 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4143 list_create(&arc_l2c_only->arcs_lists[i], 4144 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4145 } 4146 4147 buf_init(); 4148 4149 arc_thread_exit = 0; 4150 arc_eviction_list = NULL; 4151 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4152 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4153 4154 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4155 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4156 4157 if (arc_ksp != NULL) { 4158 arc_ksp->ks_data = &arc_stats; 4159 kstat_install(arc_ksp); 4160 } 4161 4162 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4163 TS_RUN, minclsyspri); 4164 4165#ifdef _KERNEL 4166 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4167 EVENTHANDLER_PRI_FIRST); 4168#endif 4169 4170 arc_dead = FALSE; 4171 arc_warm = B_FALSE; 4172 4173 /* 4174 * Calculate maximum amount of dirty data per pool. 4175 * 4176 * If it has been set by /etc/system, take that. 4177 * Otherwise, use a percentage of physical memory defined by 4178 * zfs_dirty_data_max_percent (default 10%) with a cap at 4179 * zfs_dirty_data_max_max (default 4GB). 4180 */ 4181 if (zfs_dirty_data_max == 0) { 4182 zfs_dirty_data_max = ptob(physmem) * 4183 zfs_dirty_data_max_percent / 100; 4184 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4185 zfs_dirty_data_max_max); 4186 } 4187 4188#ifdef _KERNEL 4189 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4190 prefetch_tunable_set = 1; 4191 4192#ifdef __i386__ 4193 if (prefetch_tunable_set == 0) { 4194 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4195 "-- to enable,\n"); 4196 printf(" add \"vfs.zfs.prefetch_disable=0\" " 4197 "to /boot/loader.conf.\n"); 4198 zfs_prefetch_disable = 1; 4199 } 4200#else 4201 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4202 prefetch_tunable_set == 0) { 4203 printf("ZFS NOTICE: Prefetch is disabled by default if less " 4204 "than 4GB of RAM is present;\n" 4205 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4206 "to /boot/loader.conf.\n"); 4207 zfs_prefetch_disable = 1; 4208 } 4209#endif 4210 /* Warn about ZFS memory and address space requirements. */ 4211 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4212 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4213 "expect unstable behavior.\n"); 4214 } 4215 if (kmem_size() < 512 * (1 << 20)) { 4216 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4217 "expect unstable behavior.\n"); 4218 printf(" Consider tuning vm.kmem_size and " 4219 "vm.kmem_size_max\n"); 4220 printf(" in /boot/loader.conf.\n"); 4221 } 4222#endif 4223} 4224 4225void 4226arc_fini(void) 4227{ 4228 int i; 4229 4230 mutex_enter(&arc_reclaim_thr_lock); 4231 arc_thread_exit = 1; 4232 cv_signal(&arc_reclaim_thr_cv); 4233 while (arc_thread_exit != 0) 4234 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4235 mutex_exit(&arc_reclaim_thr_lock); 4236 4237 arc_flush(NULL); 4238 4239 arc_dead = TRUE; 4240 4241 if (arc_ksp != NULL) { 4242 kstat_delete(arc_ksp); 4243 arc_ksp = NULL; 4244 } 4245 4246 mutex_destroy(&arc_eviction_mtx); 4247 mutex_destroy(&arc_reclaim_thr_lock); 4248 cv_destroy(&arc_reclaim_thr_cv); 4249 4250 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4251 list_destroy(&arc_mru->arcs_lists[i]); 4252 list_destroy(&arc_mru_ghost->arcs_lists[i]); 4253 list_destroy(&arc_mfu->arcs_lists[i]); 4254 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4255 list_destroy(&arc_l2c_only->arcs_lists[i]); 4256 4257 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4258 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4259 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4260 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4261 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4262 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4263 } 4264 4265 buf_fini(); 4266 4267 ASSERT(arc_loaned_bytes == 0); 4268 4269 mutex_destroy(&arc_lowmem_lock); 4270#ifdef _KERNEL 4271 if (arc_event_lowmem != NULL) 4272 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4273#endif 4274} 4275 4276/* 4277 * Level 2 ARC 4278 * 4279 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4280 * It uses dedicated storage devices to hold cached data, which are populated 4281 * using large infrequent writes. The main role of this cache is to boost 4282 * the performance of random read workloads. The intended L2ARC devices 4283 * include short-stroked disks, solid state disks, and other media with 4284 * substantially faster read latency than disk. 4285 * 4286 * +-----------------------+ 4287 * | ARC | 4288 * +-----------------------+ 4289 * | ^ ^ 4290 * | | | 4291 * l2arc_feed_thread() arc_read() 4292 * | | | 4293 * | l2arc read | 4294 * V | | 4295 * +---------------+ | 4296 * | L2ARC | | 4297 * +---------------+ | 4298 * | ^ | 4299 * l2arc_write() | | 4300 * | | | 4301 * V | | 4302 * +-------+ +-------+ 4303 * | vdev | | vdev | 4304 * | cache | | cache | 4305 * +-------+ +-------+ 4306 * +=========+ .-----. 4307 * : L2ARC : |-_____-| 4308 * : devices : | Disks | 4309 * +=========+ `-_____-' 4310 * 4311 * Read requests are satisfied from the following sources, in order: 4312 * 4313 * 1) ARC 4314 * 2) vdev cache of L2ARC devices 4315 * 3) L2ARC devices 4316 * 4) vdev cache of disks 4317 * 5) disks 4318 * 4319 * Some L2ARC device types exhibit extremely slow write performance. 4320 * To accommodate for this there are some significant differences between 4321 * the L2ARC and traditional cache design: 4322 * 4323 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4324 * the ARC behave as usual, freeing buffers and placing headers on ghost 4325 * lists. The ARC does not send buffers to the L2ARC during eviction as 4326 * this would add inflated write latencies for all ARC memory pressure. 4327 * 4328 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4329 * It does this by periodically scanning buffers from the eviction-end of 4330 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4331 * not already there. It scans until a headroom of buffers is satisfied, 4332 * which itself is a buffer for ARC eviction. If a compressible buffer is 4333 * found during scanning and selected for writing to an L2ARC device, we 4334 * temporarily boost scanning headroom during the next scan cycle to make 4335 * sure we adapt to compression effects (which might significantly reduce 4336 * the data volume we write to L2ARC). The thread that does this is 4337 * l2arc_feed_thread(), illustrated below; example sizes are included to 4338 * provide a better sense of ratio than this diagram: 4339 * 4340 * head --> tail 4341 * +---------------------+----------+ 4342 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4343 * +---------------------+----------+ | o L2ARC eligible 4344 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4345 * +---------------------+----------+ | 4346 * 15.9 Gbytes ^ 32 Mbytes | 4347 * headroom | 4348 * l2arc_feed_thread() 4349 * | 4350 * l2arc write hand <--[oooo]--' 4351 * | 8 Mbyte 4352 * | write max 4353 * V 4354 * +==============================+ 4355 * L2ARC dev |####|#|###|###| |####| ... | 4356 * +==============================+ 4357 * 32 Gbytes 4358 * 4359 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4360 * evicted, then the L2ARC has cached a buffer much sooner than it probably 4361 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4362 * safe to say that this is an uncommon case, since buffers at the end of 4363 * the ARC lists have moved there due to inactivity. 4364 * 4365 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4366 * then the L2ARC simply misses copying some buffers. This serves as a 4367 * pressure valve to prevent heavy read workloads from both stalling the ARC 4368 * with waits and clogging the L2ARC with writes. This also helps prevent 4369 * the potential for the L2ARC to churn if it attempts to cache content too 4370 * quickly, such as during backups of the entire pool. 4371 * 4372 * 5. After system boot and before the ARC has filled main memory, there are 4373 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4374 * lists can remain mostly static. Instead of searching from tail of these 4375 * lists as pictured, the l2arc_feed_thread() will search from the list heads 4376 * for eligible buffers, greatly increasing its chance of finding them. 4377 * 4378 * The L2ARC device write speed is also boosted during this time so that 4379 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4380 * there are no L2ARC reads, and no fear of degrading read performance 4381 * through increased writes. 4382 * 4383 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4384 * the vdev queue can aggregate them into larger and fewer writes. Each 4385 * device is written to in a rotor fashion, sweeping writes through 4386 * available space then repeating. 4387 * 4388 * 7. The L2ARC does not store dirty content. It never needs to flush 4389 * write buffers back to disk based storage. 4390 * 4391 * 8. If an ARC buffer is written (and dirtied) which also exists in the 4392 * L2ARC, the now stale L2ARC buffer is immediately dropped. 4393 * 4394 * The performance of the L2ARC can be tweaked by a number of tunables, which 4395 * may be necessary for different workloads: 4396 * 4397 * l2arc_write_max max write bytes per interval 4398 * l2arc_write_boost extra write bytes during device warmup 4399 * l2arc_noprefetch skip caching prefetched buffers 4400 * l2arc_headroom number of max device writes to precache 4401 * l2arc_headroom_boost when we find compressed buffers during ARC 4402 * scanning, we multiply headroom by this 4403 * percentage factor for the next scan cycle, 4404 * since more compressed buffers are likely to 4405 * be present 4406 * l2arc_feed_secs seconds between L2ARC writing 4407 * 4408 * Tunables may be removed or added as future performance improvements are 4409 * integrated, and also may become zpool properties. 4410 * 4411 * There are three key functions that control how the L2ARC warms up: 4412 * 4413 * l2arc_write_eligible() check if a buffer is eligible to cache 4414 * l2arc_write_size() calculate how much to write 4415 * l2arc_write_interval() calculate sleep delay between writes 4416 * 4417 * These three functions determine what to write, how much, and how quickly 4418 * to send writes. 4419 */ 4420 4421static boolean_t 4422l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab) 4423{ 4424 /* 4425 * A buffer is *not* eligible for the L2ARC if it: 4426 * 1. belongs to a different spa. 4427 * 2. is already cached on the L2ARC. 4428 * 3. has an I/O in progress (it may be an incomplete read). 4429 * 4. is flagged not eligible (zfs property). 4430 */ 4431 if (ab->b_spa != spa_guid) { 4432 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4433 return (B_FALSE); 4434 } 4435 if (ab->b_l2hdr != NULL) { 4436 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4437 return (B_FALSE); 4438 } 4439 if (HDR_IO_IN_PROGRESS(ab)) { 4440 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4441 return (B_FALSE); 4442 } 4443 if (!HDR_L2CACHE(ab)) { 4444 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4445 return (B_FALSE); 4446 } 4447 4448 return (B_TRUE); 4449} 4450 4451static uint64_t 4452l2arc_write_size(void) 4453{ 4454 uint64_t size; 4455 4456 /* 4457 * Make sure our globals have meaningful values in case the user 4458 * altered them. 4459 */ 4460 size = l2arc_write_max; 4461 if (size == 0) { 4462 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4463 "be greater than zero, resetting it to the default (%d)", 4464 L2ARC_WRITE_SIZE); 4465 size = l2arc_write_max = L2ARC_WRITE_SIZE; 4466 } 4467 4468 if (arc_warm == B_FALSE) 4469 size += l2arc_write_boost; 4470 4471 return (size); 4472 4473} 4474 4475static clock_t 4476l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4477{ 4478 clock_t interval, next, now; 4479 4480 /* 4481 * If the ARC lists are busy, increase our write rate; if the 4482 * lists are stale, idle back. This is achieved by checking 4483 * how much we previously wrote - if it was more than half of 4484 * what we wanted, schedule the next write much sooner. 4485 */ 4486 if (l2arc_feed_again && wrote > (wanted / 2)) 4487 interval = (hz * l2arc_feed_min_ms) / 1000; 4488 else 4489 interval = hz * l2arc_feed_secs; 4490 4491 now = ddi_get_lbolt(); 4492 next = MAX(now, MIN(now + interval, began + interval)); 4493 4494 return (next); 4495} 4496 4497static void 4498l2arc_hdr_stat_add(void) 4499{ 4500 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4501 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4502} 4503 4504static void 4505l2arc_hdr_stat_remove(void) 4506{ 4507 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4508 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4509} 4510 4511/* 4512 * Cycle through L2ARC devices. This is how L2ARC load balances. 4513 * If a device is returned, this also returns holding the spa config lock. 4514 */ 4515static l2arc_dev_t * 4516l2arc_dev_get_next(void) 4517{ 4518 l2arc_dev_t *first, *next = NULL; 4519 4520 /* 4521 * Lock out the removal of spas (spa_namespace_lock), then removal 4522 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4523 * both locks will be dropped and a spa config lock held instead. 4524 */ 4525 mutex_enter(&spa_namespace_lock); 4526 mutex_enter(&l2arc_dev_mtx); 4527 4528 /* if there are no vdevs, there is nothing to do */ 4529 if (l2arc_ndev == 0) 4530 goto out; 4531 4532 first = NULL; 4533 next = l2arc_dev_last; 4534 do { 4535 /* loop around the list looking for a non-faulted vdev */ 4536 if (next == NULL) { 4537 next = list_head(l2arc_dev_list); 4538 } else { 4539 next = list_next(l2arc_dev_list, next); 4540 if (next == NULL) 4541 next = list_head(l2arc_dev_list); 4542 } 4543 4544 /* if we have come back to the start, bail out */ 4545 if (first == NULL) 4546 first = next; 4547 else if (next == first) 4548 break; 4549 4550 } while (vdev_is_dead(next->l2ad_vdev)); 4551 4552 /* if we were unable to find any usable vdevs, return NULL */ 4553 if (vdev_is_dead(next->l2ad_vdev)) 4554 next = NULL; 4555 4556 l2arc_dev_last = next; 4557 4558out: 4559 mutex_exit(&l2arc_dev_mtx); 4560 4561 /* 4562 * Grab the config lock to prevent the 'next' device from being 4563 * removed while we are writing to it. 4564 */ 4565 if (next != NULL) 4566 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4567 mutex_exit(&spa_namespace_lock); 4568 4569 return (next); 4570} 4571 4572/* 4573 * Free buffers that were tagged for destruction. 4574 */ 4575static void 4576l2arc_do_free_on_write() 4577{ 4578 list_t *buflist; 4579 l2arc_data_free_t *df, *df_prev; 4580 4581 mutex_enter(&l2arc_free_on_write_mtx); 4582 buflist = l2arc_free_on_write; 4583 4584 for (df = list_tail(buflist); df; df = df_prev) { 4585 df_prev = list_prev(buflist, df); 4586 ASSERT(df->l2df_data != NULL); 4587 ASSERT(df->l2df_func != NULL); 4588 df->l2df_func(df->l2df_data, df->l2df_size); 4589 list_remove(buflist, df); 4590 kmem_free(df, sizeof (l2arc_data_free_t)); 4591 } 4592 4593 mutex_exit(&l2arc_free_on_write_mtx); 4594} 4595 4596/* 4597 * A write to a cache device has completed. Update all headers to allow 4598 * reads from these buffers to begin. 4599 */ 4600static void 4601l2arc_write_done(zio_t *zio) 4602{ 4603 l2arc_write_callback_t *cb; 4604 l2arc_dev_t *dev; 4605 list_t *buflist; 4606 arc_buf_hdr_t *head, *ab, *ab_prev; 4607 l2arc_buf_hdr_t *abl2; 4608 kmutex_t *hash_lock; 4609 int64_t bytes_dropped = 0; 4610 4611 cb = zio->io_private; 4612 ASSERT(cb != NULL); 4613 dev = cb->l2wcb_dev; 4614 ASSERT(dev != NULL); 4615 head = cb->l2wcb_head; 4616 ASSERT(head != NULL); 4617 buflist = dev->l2ad_buflist; 4618 ASSERT(buflist != NULL); 4619 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4620 l2arc_write_callback_t *, cb); 4621 4622 if (zio->io_error != 0) 4623 ARCSTAT_BUMP(arcstat_l2_writes_error); 4624 4625 mutex_enter(&l2arc_buflist_mtx); 4626 4627 /* 4628 * All writes completed, or an error was hit. 4629 */ 4630 for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 4631 ab_prev = list_prev(buflist, ab); 4632 abl2 = ab->b_l2hdr; 4633 4634 /* 4635 * Release the temporary compressed buffer as soon as possible. 4636 */ 4637 if (abl2->b_compress != ZIO_COMPRESS_OFF) 4638 l2arc_release_cdata_buf(ab); 4639 4640 hash_lock = HDR_LOCK(ab); 4641 if (!mutex_tryenter(hash_lock)) { 4642 /* 4643 * This buffer misses out. It may be in a stage 4644 * of eviction. Its ARC_L2_WRITING flag will be 4645 * left set, denying reads to this buffer. 4646 */ 4647 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4648 continue; 4649 } 4650 4651 if (zio->io_error != 0) { 4652 /* 4653 * Error - drop L2ARC entry. 4654 */ 4655 list_remove(buflist, ab); 4656 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4657 bytes_dropped += abl2->b_asize; 4658 ab->b_l2hdr = NULL; 4659 trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4660 ab->b_size, 0); 4661 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4662 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4663 } 4664 4665 /* 4666 * Allow ARC to begin reads to this L2ARC entry. 4667 */ 4668 ab->b_flags &= ~ARC_L2_WRITING; 4669 4670 mutex_exit(hash_lock); 4671 } 4672 4673 atomic_inc_64(&l2arc_writes_done); 4674 list_remove(buflist, head); 4675 kmem_cache_free(hdr_cache, head); 4676 mutex_exit(&l2arc_buflist_mtx); 4677 4678 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 4679 4680 l2arc_do_free_on_write(); 4681 4682 kmem_free(cb, sizeof (l2arc_write_callback_t)); 4683} 4684 4685/* 4686 * A read to a cache device completed. Validate buffer contents before 4687 * handing over to the regular ARC routines. 4688 */ 4689static void 4690l2arc_read_done(zio_t *zio) 4691{ 4692 l2arc_read_callback_t *cb; 4693 arc_buf_hdr_t *hdr; 4694 arc_buf_t *buf; 4695 kmutex_t *hash_lock; 4696 int equal; 4697 4698 ASSERT(zio->io_vd != NULL); 4699 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4700 4701 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4702 4703 cb = zio->io_private; 4704 ASSERT(cb != NULL); 4705 buf = cb->l2rcb_buf; 4706 ASSERT(buf != NULL); 4707 4708 hash_lock = HDR_LOCK(buf->b_hdr); 4709 mutex_enter(hash_lock); 4710 hdr = buf->b_hdr; 4711 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4712 4713 /* 4714 * If the buffer was compressed, decompress it first. 4715 */ 4716 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4717 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4718 ASSERT(zio->io_data != NULL); 4719 4720 /* 4721 * Check this survived the L2ARC journey. 4722 */ 4723 equal = arc_cksum_equal(buf); 4724 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4725 mutex_exit(hash_lock); 4726 zio->io_private = buf; 4727 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4728 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4729 arc_read_done(zio); 4730 } else { 4731 mutex_exit(hash_lock); 4732 /* 4733 * Buffer didn't survive caching. Increment stats and 4734 * reissue to the original storage device. 4735 */ 4736 if (zio->io_error != 0) { 4737 ARCSTAT_BUMP(arcstat_l2_io_error); 4738 } else { 4739 zio->io_error = SET_ERROR(EIO); 4740 } 4741 if (!equal) 4742 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4743 4744 /* 4745 * If there's no waiter, issue an async i/o to the primary 4746 * storage now. If there *is* a waiter, the caller must 4747 * issue the i/o in a context where it's OK to block. 4748 */ 4749 if (zio->io_waiter == NULL) { 4750 zio_t *pio = zio_unique_parent(zio); 4751 4752 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4753 4754 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4755 buf->b_data, zio->io_size, arc_read_done, buf, 4756 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4757 } 4758 } 4759 4760 kmem_free(cb, sizeof (l2arc_read_callback_t)); 4761} 4762 4763/* 4764 * This is the list priority from which the L2ARC will search for pages to 4765 * cache. This is used within loops (0..3) to cycle through lists in the 4766 * desired order. This order can have a significant effect on cache 4767 * performance. 4768 * 4769 * Currently the metadata lists are hit first, MFU then MRU, followed by 4770 * the data lists. This function returns a locked list, and also returns 4771 * the lock pointer. 4772 */ 4773static list_t * 4774l2arc_list_locked(int list_num, kmutex_t **lock) 4775{ 4776 list_t *list = NULL; 4777 int idx; 4778 4779 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4780 4781 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4782 idx = list_num; 4783 list = &arc_mfu->arcs_lists[idx]; 4784 *lock = ARCS_LOCK(arc_mfu, idx); 4785 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4786 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4787 list = &arc_mru->arcs_lists[idx]; 4788 *lock = ARCS_LOCK(arc_mru, idx); 4789 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4790 ARC_BUFC_NUMDATALISTS)) { 4791 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4792 list = &arc_mfu->arcs_lists[idx]; 4793 *lock = ARCS_LOCK(arc_mfu, idx); 4794 } else { 4795 idx = list_num - ARC_BUFC_NUMLISTS; 4796 list = &arc_mru->arcs_lists[idx]; 4797 *lock = ARCS_LOCK(arc_mru, idx); 4798 } 4799 4800 ASSERT(!(MUTEX_HELD(*lock))); 4801 mutex_enter(*lock); 4802 return (list); 4803} 4804 4805/* 4806 * Evict buffers from the device write hand to the distance specified in 4807 * bytes. This distance may span populated buffers, it may span nothing. 4808 * This is clearing a region on the L2ARC device ready for writing. 4809 * If the 'all' boolean is set, every buffer is evicted. 4810 */ 4811static void 4812l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4813{ 4814 list_t *buflist; 4815 l2arc_buf_hdr_t *abl2; 4816 arc_buf_hdr_t *ab, *ab_prev; 4817 kmutex_t *hash_lock; 4818 uint64_t taddr; 4819 int64_t bytes_evicted = 0; 4820 4821 buflist = dev->l2ad_buflist; 4822 4823 if (buflist == NULL) 4824 return; 4825 4826 if (!all && dev->l2ad_first) { 4827 /* 4828 * This is the first sweep through the device. There is 4829 * nothing to evict. 4830 */ 4831 return; 4832 } 4833 4834 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4835 /* 4836 * When nearing the end of the device, evict to the end 4837 * before the device write hand jumps to the start. 4838 */ 4839 taddr = dev->l2ad_end; 4840 } else { 4841 taddr = dev->l2ad_hand + distance; 4842 } 4843 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4844 uint64_t, taddr, boolean_t, all); 4845 4846top: 4847 mutex_enter(&l2arc_buflist_mtx); 4848 for (ab = list_tail(buflist); ab; ab = ab_prev) { 4849 ab_prev = list_prev(buflist, ab); 4850 4851 hash_lock = HDR_LOCK(ab); 4852 if (!mutex_tryenter(hash_lock)) { 4853 /* 4854 * Missed the hash lock. Retry. 4855 */ 4856 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4857 mutex_exit(&l2arc_buflist_mtx); 4858 mutex_enter(hash_lock); 4859 mutex_exit(hash_lock); 4860 goto top; 4861 } 4862 4863 if (HDR_L2_WRITE_HEAD(ab)) { 4864 /* 4865 * We hit a write head node. Leave it for 4866 * l2arc_write_done(). 4867 */ 4868 list_remove(buflist, ab); 4869 mutex_exit(hash_lock); 4870 continue; 4871 } 4872 4873 if (!all && ab->b_l2hdr != NULL && 4874 (ab->b_l2hdr->b_daddr > taddr || 4875 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4876 /* 4877 * We've evicted to the target address, 4878 * or the end of the device. 4879 */ 4880 mutex_exit(hash_lock); 4881 break; 4882 } 4883 4884 if (HDR_FREE_IN_PROGRESS(ab)) { 4885 /* 4886 * Already on the path to destruction. 4887 */ 4888 mutex_exit(hash_lock); 4889 continue; 4890 } 4891 4892 if (ab->b_state == arc_l2c_only) { 4893 ASSERT(!HDR_L2_READING(ab)); 4894 /* 4895 * This doesn't exist in the ARC. Destroy. 4896 * arc_hdr_destroy() will call list_remove() 4897 * and decrement arcstat_l2_size. 4898 */ 4899 arc_change_state(arc_anon, ab, hash_lock); 4900 arc_hdr_destroy(ab); 4901 } else { 4902 /* 4903 * Invalidate issued or about to be issued 4904 * reads, since we may be about to write 4905 * over this location. 4906 */ 4907 if (HDR_L2_READING(ab)) { 4908 ARCSTAT_BUMP(arcstat_l2_evict_reading); 4909 ab->b_flags |= ARC_L2_EVICTED; 4910 } 4911 4912 /* 4913 * Tell ARC this no longer exists in L2ARC. 4914 */ 4915 if (ab->b_l2hdr != NULL) { 4916 abl2 = ab->b_l2hdr; 4917 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4918 bytes_evicted += abl2->b_asize; 4919 ab->b_l2hdr = NULL; 4920 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4921 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4922 } 4923 list_remove(buflist, ab); 4924 4925 /* 4926 * This may have been leftover after a 4927 * failed write. 4928 */ 4929 ab->b_flags &= ~ARC_L2_WRITING; 4930 } 4931 mutex_exit(hash_lock); 4932 } 4933 mutex_exit(&l2arc_buflist_mtx); 4934 4935 vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); 4936 dev->l2ad_evict = taddr; 4937} 4938 4939/* 4940 * Find and write ARC buffers to the L2ARC device. 4941 * 4942 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 4943 * for reading until they have completed writing. 4944 * The headroom_boost is an in-out parameter used to maintain headroom boost 4945 * state between calls to this function. 4946 * 4947 * Returns the number of bytes actually written (which may be smaller than 4948 * the delta by which the device hand has changed due to alignment). 4949 */ 4950static uint64_t 4951l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 4952 boolean_t *headroom_boost) 4953{ 4954 arc_buf_hdr_t *ab, *ab_prev, *head; 4955 list_t *list; 4956 uint64_t write_asize, write_psize, write_sz, headroom, 4957 buf_compress_minsz; 4958 void *buf_data; 4959 kmutex_t *list_lock; 4960 boolean_t full; 4961 l2arc_write_callback_t *cb; 4962 zio_t *pio, *wzio; 4963 uint64_t guid = spa_load_guid(spa); 4964 const boolean_t do_headroom_boost = *headroom_boost; 4965 int try; 4966 4967 ASSERT(dev->l2ad_vdev != NULL); 4968 4969 /* Lower the flag now, we might want to raise it again later. */ 4970 *headroom_boost = B_FALSE; 4971 4972 pio = NULL; 4973 write_sz = write_asize = write_psize = 0; 4974 full = B_FALSE; 4975 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 4976 head->b_flags |= ARC_L2_WRITE_HEAD; 4977 4978 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 4979 /* 4980 * We will want to try to compress buffers that are at least 2x the 4981 * device sector size. 4982 */ 4983 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 4984 4985 /* 4986 * Copy buffers for L2ARC writing. 4987 */ 4988 mutex_enter(&l2arc_buflist_mtx); 4989 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 4990 uint64_t passed_sz = 0; 4991 4992 list = l2arc_list_locked(try, &list_lock); 4993 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 4994 4995 /* 4996 * L2ARC fast warmup. 4997 * 4998 * Until the ARC is warm and starts to evict, read from the 4999 * head of the ARC lists rather than the tail. 5000 */ 5001 if (arc_warm == B_FALSE) 5002 ab = list_head(list); 5003 else 5004 ab = list_tail(list); 5005 if (ab == NULL) 5006 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 5007 5008 headroom = target_sz * l2arc_headroom; 5009 if (do_headroom_boost) 5010 headroom = (headroom * l2arc_headroom_boost) / 100; 5011 5012 for (; ab; ab = ab_prev) { 5013 l2arc_buf_hdr_t *l2hdr; 5014 kmutex_t *hash_lock; 5015 uint64_t buf_sz; 5016 5017 if (arc_warm == B_FALSE) 5018 ab_prev = list_next(list, ab); 5019 else 5020 ab_prev = list_prev(list, ab); 5021 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size); 5022 5023 hash_lock = HDR_LOCK(ab); 5024 if (!mutex_tryenter(hash_lock)) { 5025 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5026 /* 5027 * Skip this buffer rather than waiting. 5028 */ 5029 continue; 5030 } 5031 5032 passed_sz += ab->b_size; 5033 if (passed_sz > headroom) { 5034 /* 5035 * Searched too far. 5036 */ 5037 mutex_exit(hash_lock); 5038 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5039 break; 5040 } 5041 5042 if (!l2arc_write_eligible(guid, ab)) { 5043 mutex_exit(hash_lock); 5044 continue; 5045 } 5046 5047 if ((write_sz + ab->b_size) > target_sz) { 5048 full = B_TRUE; 5049 mutex_exit(hash_lock); 5050 ARCSTAT_BUMP(arcstat_l2_write_full); 5051 break; 5052 } 5053 5054 if (pio == NULL) { 5055 /* 5056 * Insert a dummy header on the buflist so 5057 * l2arc_write_done() can find where the 5058 * write buffers begin without searching. 5059 */ 5060 list_insert_head(dev->l2ad_buflist, head); 5061 5062 cb = kmem_alloc( 5063 sizeof (l2arc_write_callback_t), KM_SLEEP); 5064 cb->l2wcb_dev = dev; 5065 cb->l2wcb_head = head; 5066 pio = zio_root(spa, l2arc_write_done, cb, 5067 ZIO_FLAG_CANFAIL); 5068 ARCSTAT_BUMP(arcstat_l2_write_pios); 5069 } 5070 5071 /* 5072 * Create and add a new L2ARC header. 5073 */ 5074 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5075 l2hdr->b_dev = dev; 5076 ab->b_flags |= ARC_L2_WRITING; 5077 5078 /* 5079 * Temporarily stash the data buffer in b_tmp_cdata. 5080 * The subsequent write step will pick it up from 5081 * there. This is because can't access ab->b_buf 5082 * without holding the hash_lock, which we in turn 5083 * can't access without holding the ARC list locks 5084 * (which we want to avoid during compression/writing). 5085 */ 5086 l2hdr->b_compress = ZIO_COMPRESS_OFF; 5087 l2hdr->b_asize = ab->b_size; 5088 l2hdr->b_tmp_cdata = ab->b_buf->b_data; 5089 5090 buf_sz = ab->b_size; 5091 ab->b_l2hdr = l2hdr; 5092 5093 list_insert_head(dev->l2ad_buflist, ab); 5094 5095 /* 5096 * Compute and store the buffer cksum before 5097 * writing. On debug the cksum is verified first. 5098 */ 5099 arc_cksum_verify(ab->b_buf); 5100 arc_cksum_compute(ab->b_buf, B_TRUE); 5101 5102 mutex_exit(hash_lock); 5103 5104 write_sz += buf_sz; 5105 } 5106 5107 mutex_exit(list_lock); 5108 5109 if (full == B_TRUE) 5110 break; 5111 } 5112 5113 /* No buffers selected for writing? */ 5114 if (pio == NULL) { 5115 ASSERT0(write_sz); 5116 mutex_exit(&l2arc_buflist_mtx); 5117 kmem_cache_free(hdr_cache, head); 5118 return (0); 5119 } 5120 5121 /* 5122 * Now start writing the buffers. We're starting at the write head 5123 * and work backwards, retracing the course of the buffer selector 5124 * loop above. 5125 */ 5126 for (ab = list_prev(dev->l2ad_buflist, head); ab; 5127 ab = list_prev(dev->l2ad_buflist, ab)) { 5128 l2arc_buf_hdr_t *l2hdr; 5129 uint64_t buf_sz; 5130 5131 /* 5132 * We shouldn't need to lock the buffer here, since we flagged 5133 * it as ARC_L2_WRITING in the previous step, but we must take 5134 * care to only access its L2 cache parameters. In particular, 5135 * ab->b_buf may be invalid by now due to ARC eviction. 5136 */ 5137 l2hdr = ab->b_l2hdr; 5138 l2hdr->b_daddr = dev->l2ad_hand; 5139 5140 if ((ab->b_flags & ARC_L2COMPRESS) && 5141 l2hdr->b_asize >= buf_compress_minsz) { 5142 if (l2arc_compress_buf(l2hdr)) { 5143 /* 5144 * If compression succeeded, enable headroom 5145 * boost on the next scan cycle. 5146 */ 5147 *headroom_boost = B_TRUE; 5148 } 5149 } 5150 5151 /* 5152 * Pick up the buffer data we had previously stashed away 5153 * (and now potentially also compressed). 5154 */ 5155 buf_data = l2hdr->b_tmp_cdata; 5156 buf_sz = l2hdr->b_asize; 5157 5158 /* Compression may have squashed the buffer to zero length. */ 5159 if (buf_sz != 0) { 5160 uint64_t buf_p_sz; 5161 5162 wzio = zio_write_phys(pio, dev->l2ad_vdev, 5163 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5164 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5165 ZIO_FLAG_CANFAIL, B_FALSE); 5166 5167 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5168 zio_t *, wzio); 5169 (void) zio_nowait(wzio); 5170 5171 write_asize += buf_sz; 5172 /* 5173 * Keep the clock hand suitably device-aligned. 5174 */ 5175 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5176 write_psize += buf_p_sz; 5177 dev->l2ad_hand += buf_p_sz; 5178 } 5179 } 5180 5181 mutex_exit(&l2arc_buflist_mtx); 5182 5183 ASSERT3U(write_asize, <=, target_sz); 5184 ARCSTAT_BUMP(arcstat_l2_writes_sent); 5185 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5186 ARCSTAT_INCR(arcstat_l2_size, write_sz); 5187 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 5188 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 5189 5190 /* 5191 * Bump device hand to the device start if it is approaching the end. 5192 * l2arc_evict() will already have evicted ahead for this case. 5193 */ 5194 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5195 dev->l2ad_hand = dev->l2ad_start; 5196 dev->l2ad_evict = dev->l2ad_start; 5197 dev->l2ad_first = B_FALSE; 5198 } 5199 5200 dev->l2ad_writing = B_TRUE; 5201 (void) zio_wait(pio); 5202 dev->l2ad_writing = B_FALSE; 5203 5204 return (write_asize); 5205} 5206 5207/* 5208 * Compresses an L2ARC buffer. 5209 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5210 * size in l2hdr->b_asize. This routine tries to compress the data and 5211 * depending on the compression result there are three possible outcomes: 5212 * *) The buffer was incompressible. The original l2hdr contents were left 5213 * untouched and are ready for writing to an L2 device. 5214 * *) The buffer was all-zeros, so there is no need to write it to an L2 5215 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5216 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5217 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5218 * data buffer which holds the compressed data to be written, and b_asize 5219 * tells us how much data there is. b_compress is set to the appropriate 5220 * compression algorithm. Once writing is done, invoke 5221 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5222 * 5223 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5224 * buffer was incompressible). 5225 */ 5226static boolean_t 5227l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5228{ 5229 void *cdata; 5230 size_t csize, len, rounded; 5231 5232 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5233 ASSERT(l2hdr->b_tmp_cdata != NULL); 5234 5235 len = l2hdr->b_asize; 5236 cdata = zio_data_buf_alloc(len); 5237 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5238 cdata, l2hdr->b_asize, (size_t)(1ULL << l2hdr->b_dev->l2ad_vdev->vdev_ashift)); 5239 5240 rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE); 5241 if (rounded > csize) { 5242 bzero((char *)cdata + csize, rounded - csize); 5243 csize = rounded; 5244 } 5245 5246 if (csize == 0) { 5247 /* zero block, indicate that there's nothing to write */ 5248 zio_data_buf_free(cdata, len); 5249 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5250 l2hdr->b_asize = 0; 5251 l2hdr->b_tmp_cdata = NULL; 5252 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5253 return (B_TRUE); 5254 } else if (csize > 0 && csize < len) { 5255 /* 5256 * Compression succeeded, we'll keep the cdata around for 5257 * writing and release it afterwards. 5258 */ 5259 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5260 l2hdr->b_asize = csize; 5261 l2hdr->b_tmp_cdata = cdata; 5262 ARCSTAT_BUMP(arcstat_l2_compress_successes); 5263 return (B_TRUE); 5264 } else { 5265 /* 5266 * Compression failed, release the compressed buffer. 5267 * l2hdr will be left unmodified. 5268 */ 5269 zio_data_buf_free(cdata, len); 5270 ARCSTAT_BUMP(arcstat_l2_compress_failures); 5271 return (B_FALSE); 5272 } 5273} 5274 5275/* 5276 * Decompresses a zio read back from an l2arc device. On success, the 5277 * underlying zio's io_data buffer is overwritten by the uncompressed 5278 * version. On decompression error (corrupt compressed stream), the 5279 * zio->io_error value is set to signal an I/O error. 5280 * 5281 * Please note that the compressed data stream is not checksummed, so 5282 * if the underlying device is experiencing data corruption, we may feed 5283 * corrupt data to the decompressor, so the decompressor needs to be 5284 * able to handle this situation (LZ4 does). 5285 */ 5286static void 5287l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5288{ 5289 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5290 5291 if (zio->io_error != 0) { 5292 /* 5293 * An io error has occured, just restore the original io 5294 * size in preparation for a main pool read. 5295 */ 5296 zio->io_orig_size = zio->io_size = hdr->b_size; 5297 return; 5298 } 5299 5300 if (c == ZIO_COMPRESS_EMPTY) { 5301 /* 5302 * An empty buffer results in a null zio, which means we 5303 * need to fill its io_data after we're done restoring the 5304 * buffer's contents. 5305 */ 5306 ASSERT(hdr->b_buf != NULL); 5307 bzero(hdr->b_buf->b_data, hdr->b_size); 5308 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5309 } else { 5310 ASSERT(zio->io_data != NULL); 5311 /* 5312 * We copy the compressed data from the start of the arc buffer 5313 * (the zio_read will have pulled in only what we need, the 5314 * rest is garbage which we will overwrite at decompression) 5315 * and then decompress back to the ARC data buffer. This way we 5316 * can minimize copying by simply decompressing back over the 5317 * original compressed data (rather than decompressing to an 5318 * aux buffer and then copying back the uncompressed buffer, 5319 * which is likely to be much larger). 5320 */ 5321 uint64_t csize; 5322 void *cdata; 5323 5324 csize = zio->io_size; 5325 cdata = zio_data_buf_alloc(csize); 5326 bcopy(zio->io_data, cdata, csize); 5327 if (zio_decompress_data(c, cdata, zio->io_data, csize, 5328 hdr->b_size) != 0) 5329 zio->io_error = EIO; 5330 zio_data_buf_free(cdata, csize); 5331 } 5332 5333 /* Restore the expected uncompressed IO size. */ 5334 zio->io_orig_size = zio->io_size = hdr->b_size; 5335} 5336 5337/* 5338 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5339 * This buffer serves as a temporary holder of compressed data while 5340 * the buffer entry is being written to an l2arc device. Once that is 5341 * done, we can dispose of it. 5342 */ 5343static void 5344l2arc_release_cdata_buf(arc_buf_hdr_t *ab) 5345{ 5346 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr; 5347 5348 if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) { 5349 /* 5350 * If the data was compressed, then we've allocated a 5351 * temporary buffer for it, so now we need to release it. 5352 */ 5353 ASSERT(l2hdr->b_tmp_cdata != NULL); 5354 zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size); 5355 } 5356 l2hdr->b_tmp_cdata = NULL; 5357} 5358 5359/* 5360 * This thread feeds the L2ARC at regular intervals. This is the beating 5361 * heart of the L2ARC. 5362 */ 5363static void 5364l2arc_feed_thread(void *dummy __unused) 5365{ 5366 callb_cpr_t cpr; 5367 l2arc_dev_t *dev; 5368 spa_t *spa; 5369 uint64_t size, wrote; 5370 clock_t begin, next = ddi_get_lbolt(); 5371 boolean_t headroom_boost = B_FALSE; 5372 5373 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5374 5375 mutex_enter(&l2arc_feed_thr_lock); 5376 5377 while (l2arc_thread_exit == 0) { 5378 CALLB_CPR_SAFE_BEGIN(&cpr); 5379 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5380 next - ddi_get_lbolt()); 5381 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5382 next = ddi_get_lbolt() + hz; 5383 5384 /* 5385 * Quick check for L2ARC devices. 5386 */ 5387 mutex_enter(&l2arc_dev_mtx); 5388 if (l2arc_ndev == 0) { 5389 mutex_exit(&l2arc_dev_mtx); 5390 continue; 5391 } 5392 mutex_exit(&l2arc_dev_mtx); 5393 begin = ddi_get_lbolt(); 5394 5395 /* 5396 * This selects the next l2arc device to write to, and in 5397 * doing so the next spa to feed from: dev->l2ad_spa. This 5398 * will return NULL if there are now no l2arc devices or if 5399 * they are all faulted. 5400 * 5401 * If a device is returned, its spa's config lock is also 5402 * held to prevent device removal. l2arc_dev_get_next() 5403 * will grab and release l2arc_dev_mtx. 5404 */ 5405 if ((dev = l2arc_dev_get_next()) == NULL) 5406 continue; 5407 5408 spa = dev->l2ad_spa; 5409 ASSERT(spa != NULL); 5410 5411 /* 5412 * If the pool is read-only then force the feed thread to 5413 * sleep a little longer. 5414 */ 5415 if (!spa_writeable(spa)) { 5416 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5417 spa_config_exit(spa, SCL_L2ARC, dev); 5418 continue; 5419 } 5420 5421 /* 5422 * Avoid contributing to memory pressure. 5423 */ 5424 if (arc_reclaim_needed()) { 5425 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5426 spa_config_exit(spa, SCL_L2ARC, dev); 5427 continue; 5428 } 5429 5430 ARCSTAT_BUMP(arcstat_l2_feeds); 5431 5432 size = l2arc_write_size(); 5433 5434 /* 5435 * Evict L2ARC buffers that will be overwritten. 5436 */ 5437 l2arc_evict(dev, size, B_FALSE); 5438 5439 /* 5440 * Write ARC buffers. 5441 */ 5442 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5443 5444 /* 5445 * Calculate interval between writes. 5446 */ 5447 next = l2arc_write_interval(begin, size, wrote); 5448 spa_config_exit(spa, SCL_L2ARC, dev); 5449 } 5450 5451 l2arc_thread_exit = 0; 5452 cv_broadcast(&l2arc_feed_thr_cv); 5453 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5454 thread_exit(); 5455} 5456 5457boolean_t 5458l2arc_vdev_present(vdev_t *vd) 5459{ 5460 l2arc_dev_t *dev; 5461 5462 mutex_enter(&l2arc_dev_mtx); 5463 for (dev = list_head(l2arc_dev_list); dev != NULL; 5464 dev = list_next(l2arc_dev_list, dev)) { 5465 if (dev->l2ad_vdev == vd) 5466 break; 5467 } 5468 mutex_exit(&l2arc_dev_mtx); 5469 5470 return (dev != NULL); 5471} 5472 5473/* 5474 * Add a vdev for use by the L2ARC. By this point the spa has already 5475 * validated the vdev and opened it. 5476 */ 5477void 5478l2arc_add_vdev(spa_t *spa, vdev_t *vd) 5479{ 5480 l2arc_dev_t *adddev; 5481 5482 ASSERT(!l2arc_vdev_present(vd)); 5483 5484 vdev_ashift_optimize(vd); 5485 5486 /* 5487 * Create a new l2arc device entry. 5488 */ 5489 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5490 adddev->l2ad_spa = spa; 5491 adddev->l2ad_vdev = vd; 5492 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5493 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5494 adddev->l2ad_hand = adddev->l2ad_start; 5495 adddev->l2ad_evict = adddev->l2ad_start; 5496 adddev->l2ad_first = B_TRUE; 5497 adddev->l2ad_writing = B_FALSE; 5498 5499 /* 5500 * This is a list of all ARC buffers that are still valid on the 5501 * device. 5502 */ 5503 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5504 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5505 offsetof(arc_buf_hdr_t, b_l2node)); 5506 5507 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5508 5509 /* 5510 * Add device to global list 5511 */ 5512 mutex_enter(&l2arc_dev_mtx); 5513 list_insert_head(l2arc_dev_list, adddev); 5514 atomic_inc_64(&l2arc_ndev); 5515 mutex_exit(&l2arc_dev_mtx); 5516} 5517 5518/* 5519 * Remove a vdev from the L2ARC. 5520 */ 5521void 5522l2arc_remove_vdev(vdev_t *vd) 5523{ 5524 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5525 5526 /* 5527 * Find the device by vdev 5528 */ 5529 mutex_enter(&l2arc_dev_mtx); 5530 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5531 nextdev = list_next(l2arc_dev_list, dev); 5532 if (vd == dev->l2ad_vdev) { 5533 remdev = dev; 5534 break; 5535 } 5536 } 5537 ASSERT(remdev != NULL); 5538 5539 /* 5540 * Remove device from global list 5541 */ 5542 list_remove(l2arc_dev_list, remdev); 5543 l2arc_dev_last = NULL; /* may have been invalidated */ 5544 atomic_dec_64(&l2arc_ndev); 5545 mutex_exit(&l2arc_dev_mtx); 5546 5547 /* 5548 * Clear all buflists and ARC references. L2ARC device flush. 5549 */ 5550 l2arc_evict(remdev, 0, B_TRUE); 5551 list_destroy(remdev->l2ad_buflist); 5552 kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5553 kmem_free(remdev, sizeof (l2arc_dev_t)); 5554} 5555 5556void 5557l2arc_init(void) 5558{ 5559 l2arc_thread_exit = 0; 5560 l2arc_ndev = 0; 5561 l2arc_writes_sent = 0; 5562 l2arc_writes_done = 0; 5563 5564 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5565 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5566 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5567 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5568 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5569 5570 l2arc_dev_list = &L2ARC_dev_list; 5571 l2arc_free_on_write = &L2ARC_free_on_write; 5572 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5573 offsetof(l2arc_dev_t, l2ad_node)); 5574 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5575 offsetof(l2arc_data_free_t, l2df_list_node)); 5576} 5577 5578void 5579l2arc_fini(void) 5580{ 5581 /* 5582 * This is called from dmu_fini(), which is called from spa_fini(); 5583 * Because of this, we can assume that all l2arc devices have 5584 * already been removed when the pools themselves were removed. 5585 */ 5586 5587 l2arc_do_free_on_write(); 5588 5589 mutex_destroy(&l2arc_feed_thr_lock); 5590 cv_destroy(&l2arc_feed_thr_cv); 5591 mutex_destroy(&l2arc_dev_mtx); 5592 mutex_destroy(&l2arc_buflist_mtx); 5593 mutex_destroy(&l2arc_free_on_write_mtx); 5594 5595 list_destroy(l2arc_dev_list); 5596 list_destroy(l2arc_free_on_write); 5597} 5598 5599void 5600l2arc_start(void) 5601{ 5602 if (!(spa_mode_global & FWRITE)) 5603 return; 5604 5605 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5606 TS_RUN, minclsyspri); 5607} 5608 5609void 5610l2arc_stop(void) 5611{ 5612 if (!(spa_mode_global & FWRITE)) 5613 return; 5614 5615 mutex_enter(&l2arc_feed_thr_lock); 5616 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5617 l2arc_thread_exit = 1; 5618 while (l2arc_thread_exit != 0) 5619 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5620 mutex_exit(&l2arc_feed_thr_lock); 5621} 5622