arc.c revision 208373
174462Salfred/* 274462Salfred * CDDL HEADER START 3259118Shrs * 4259118Shrs * The contents of this file are subject to the terms of the 58870Srgrimes * Common Development and Distribution License (the "License"). 6259118Shrs * You may not use this file except in compliance with the License. 7259118Shrs * 8259118Shrs * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 98870Srgrimes * or http://www.opensolaris.org/os/licensing. 10259118Shrs * See the License for the specific language governing permissions 11259118Shrs * and limitations under the License. 12259118Shrs * 13259118Shrs * When distributing Covered Code, include this CDDL HEADER in each 14259118Shrs * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15259118Shrs * If applicable, add the following below this CDDL HEADER, with the 16259118Shrs * fields enclosed by brackets "[]" replaced with your own identifying 17259118Shrs * information: Portions Copyright [yyyy] [name of copyright owner] 18259118Shrs * 198870Srgrimes * CDDL HEADER END 20259118Shrs */ 21259118Shrs/* 22259118Shrs * Copyright 2009 Sun Microsystems, Inc. All rights reserved. 23259118Shrs * Use is subject to license terms. 24259118Shrs */ 25259118Shrs 26259118Shrs/* 27259118Shrs * DVA-based Adjustable Replacement Cache 28259118Shrs * 29259118Shrs * While much of the theory of operation used here is 30259118Shrs * based on the self-tuning, low overhead replacement cache 31259118Shrs * presented by Megiddo and Modha at FAST 2003, there are some 321902Swollman * significant differences: 331902Swollman * 341902Swollman * 1. The Megiddo and Modha model assumes any page is evictable. 35136582Sobrien * Pages in its cache cannot be "locked" into memory. This makes 3692986Sobrien * the eviction algorithm simple: evict the last page in the list. 371902Swollman * This also make the performance characteristics easy to reason 3892986Sobrien * about. Our cache is not so simple. At any given moment, some 3992986Sobrien * subset of the blocks in the cache are un-evictable because we 401902Swollman * have handed out a reference to them. Blocks are only evictable 411902Swollman * when there are no external references active. This makes 421902Swollman * eviction far more problematic: we choose to evict the evictable 431902Swollman * blocks that are the "lowest" in the list. 441902Swollman * 451902Swollman * There are times when it is not possible to evict the requested 461902Swollman * space. In these circumstances we are unable to adjust the cache 471902Swollman * size. To prevent the cache growing unbounded at these times we 4874462Salfred * implement a "cache throttle" that slows the flow of new data 4974462Salfred * into the cache until we can make space available. 50101067Snectar * 511902Swollman * 2. The Megiddo and Modha model assumes a fixed cache size. 521902Swollman * Pages are evicted when the cache is full and there is a cache 5311669Sphk * miss. Our model has a variable sized cache. It grows with 5474462Salfred * high use, but also tries to react to memory pressure from the 551902Swollman * operating system: decreasing its size when system memory is 561902Swollman * tight. 5774462Salfred * 581902Swollman * 3. The Megiddo and Modha model assumes a fixed page size. All 591902Swollman * elements of the cache are therefor exactly the same size. So 601902Swollman * when adjusting the cache size following a cache miss, its simply 611902Swollman * a matter of choosing a single page to evict. In our model, we 621902Swollman * have variable sized cache blocks (rangeing from 512 bytes to 631902Swollman * 128K bytes). We therefor choose a set of blocks to evict to make 641902Swollman * space for a cache miss that approximates as closely as possible 651902Swollman * the space used by the new block. 661902Swollman * 67283833Srodrigc * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 68283833Srodrigc * by N. Megiddo & D. Modha, FAST 2003 69283833Srodrigc */ 70283833Srodrigc 71283833Srodrigc/* 72283833Srodrigc * The locking model: 73283833Srodrigc * 74283833Srodrigc * A new reference to a cache buffer can be obtained in two 75283833Srodrigc * ways: 1) via a hash table lookup using the DVA as a key, 761902Swollman * or 2) via one of the ARC lists. The arc_read() interface 7774462Salfred * uses method 1, while the internal arc algorithms for 7874462Salfred * adjusting the cache use method 2. We therefor provide two 7974462Salfred * types of locks: 1) the hash table lock array, and 2) the 8074462Salfred * arc list locks. 8174462Salfred * 821902Swollman * Buffers do not have their own mutexs, rather they rely on the 831902Swollman * hash table mutexs for the bulk of their protection (i.e. most 84101045Sdarrenr * fields in the arc_buf_hdr_t are protected by these mutexs). 851902Swollman * 861902Swollman * buf_hash_find() returns the appropriate mutex (held) when it 871902Swollman * locates the requested buffer in the hash table. It returns 88101147Snectar * NULL for the mutex if the buffer was not in the table. 89101045Sdarrenr * 901902Swollman * buf_hash_remove() expects the appropriate hash mutex to be 911902Swollman * already held before it is invoked. 921902Swollman * 931902Swollman * Each arc state also has a mutex which is used to protect the 941902Swollman * buffer list associated with the state. When attempting to 951902Swollman * obtain a hash table lock while holding an arc list lock you 961902Swollman * must use: mutex_tryenter() to avoid deadlock. Also note that 971902Swollman * the active state mutex must be held before the ghost state mutex. 981902Swollman * 991902Swollman * Arc buffers may have an associated eviction callback function. 1001902Swollman * This function will be invoked prior to removing the buffer (e.g. 1011902Swollman * in arc_do_user_evicts()). Note however that the data associated 1021902Swollman * with the buffer may be evicted prior to the callback. The callback 1031902Swollman * must be made with *no locks held* (to prevent deadlock). Additionally, 1041902Swollman * the users of callbacks must ensure that their private data is 10574462Salfred * protected from simultaneous callbacks from arc_buf_evict() 1061902Swollman * and arc_do_user_evicts(). 1071902Swollman * 10821062Speter * Note that the majority of the performance stats are manipulated 1091902Swollman * with atomic operations. 1101902Swollman * 1111902Swollman * The L2ARC uses the l2arc_buflist_mtx global mutex for the following: 1121902Swollman * 11374462Salfred * - L2ARC buflist creation 11474462Salfred * - L2ARC buflist eviction 11574462Salfred * - L2ARC write completion, which walks L2ARC buflists 1161902Swollman * - ARC header destruction, as it removes from L2ARC buflists 11774462Salfred * - ARC header release, as it removes from L2ARC buflists 1181902Swollman */ 1191902Swollman 1201902Swollman#include <sys/spa.h> 1211902Swollman#include <sys/zio.h> 12274462Salfred#include <sys/zio_checksum.h> 1231902Swollman#include <sys/zfs_context.h> 1241902Swollman#include <sys/arc.h> 1251902Swollman#include <sys/refcount.h> 1261902Swollman#include <sys/vdev.h> 1271902Swollman#ifdef _KERNEL 1281902Swollman#include <sys/dnlc.h> 1291902Swollman#endif 1301902Swollman#include <sys/callb.h> 1311902Swollman#include <sys/kstat.h> 1321902Swollman#include <sys/sdt.h> 1331902Swollman 1341902Swollman#include <vm/vm_pageout.h> 1351902Swollman 1361902Swollmanstatic kmutex_t arc_reclaim_thr_lock; 1371902Swollmanstatic kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 1381902Swollmanstatic uint8_t arc_thread_exit; 1391902Swollman 1401902Swollmanextern int zfs_write_limit_shift; 1411902Swollmanextern uint64_t zfs_write_limit_max; 1421902Swollmanextern kmutex_t zfs_write_limit_lock; 1431902Swollman 1441902Swollman#define ARC_REDUCE_DNLC_PERCENT 3 1451902Swollmanuint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 1461902Swollman 147283833Srodrigctypedef enum arc_reclaim_strategy { 1481902Swollman ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 14974462Salfred ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 15074462Salfred} arc_reclaim_strategy_t; 1511902Swollman 1521902Swollman/* number of seconds before growing cache again */ 1531902Swollmanstatic int arc_grow_retry = 60; 154101045Sdarrenr 1551902Swollman/* shift of arc_c for calculating both min and max arc_p */ 1561902Swollmanstatic int arc_p_min_shift = 4; 1571902Swollman 1581902Swollman/* log2(fraction of arc to reclaim) */ 15974462Salfredstatic int arc_shrink_shift = 5; 1601902Swollman 161/* 162 * minimum lifespan of a prefetch block in clock ticks 163 * (initialized in arc_init()) 164 */ 165static int arc_min_prefetch_lifespan; 166 167static int arc_dead; 168extern int zfs_prefetch_disable; 169 170/* 171 * The arc has filled available memory and has now warmed up. 172 */ 173static boolean_t arc_warm; 174 175/* 176 * These tunables are for performance analysis. 177 */ 178uint64_t zfs_arc_max; 179uint64_t zfs_arc_min; 180uint64_t zfs_arc_meta_limit = 0; 181int zfs_mdcomp_disable = 0; 182int zfs_arc_grow_retry = 0; 183int zfs_arc_shrink_shift = 0; 184int zfs_arc_p_min_shift = 0; 185 186TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max); 187TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 188TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 189TUNABLE_INT("vfs.zfs.mdcomp_disable", &zfs_mdcomp_disable); 190SYSCTL_DECL(_vfs_zfs); 191SYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 192 "Maximum ARC size"); 193SYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 194 "Minimum ARC size"); 195SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RDTUN, 196 &zfs_mdcomp_disable, 0, "Disable metadata compression"); 197 198#ifdef ZIO_USE_UMA 199extern kmem_cache_t *zio_buf_cache[]; 200extern kmem_cache_t *zio_data_buf_cache[]; 201#endif 202 203/* 204 * Note that buffers can be in one of 6 states: 205 * ARC_anon - anonymous (discussed below) 206 * ARC_mru - recently used, currently cached 207 * ARC_mru_ghost - recentely used, no longer in cache 208 * ARC_mfu - frequently used, currently cached 209 * ARC_mfu_ghost - frequently used, no longer in cache 210 * ARC_l2c_only - exists in L2ARC but not other states 211 * When there are no active references to the buffer, they are 212 * are linked onto a list in one of these arc states. These are 213 * the only buffers that can be evicted or deleted. Within each 214 * state there are multiple lists, one for meta-data and one for 215 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 216 * etc.) is tracked separately so that it can be managed more 217 * explicitly: favored over data, limited explicitly. 218 * 219 * Anonymous buffers are buffers that are not associated with 220 * a DVA. These are buffers that hold dirty block copies 221 * before they are written to stable storage. By definition, 222 * they are "ref'd" and are considered part of arc_mru 223 * that cannot be freed. Generally, they will aquire a DVA 224 * as they are written and migrate onto the arc_mru list. 225 * 226 * The ARC_l2c_only state is for buffers that are in the second 227 * level ARC but no longer in any of the ARC_m* lists. The second 228 * level ARC itself may also contain buffers that are in any of 229 * the ARC_m* states - meaning that a buffer can exist in two 230 * places. The reason for the ARC_l2c_only state is to keep the 231 * buffer header in the hash table, so that reads that hit the 232 * second level ARC benefit from these fast lookups. 233 */ 234 235#define ARCS_LOCK_PAD CACHE_LINE_SIZE 236struct arcs_lock { 237 kmutex_t arcs_lock; 238#ifdef _KERNEL 239 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 240#endif 241}; 242 243/* 244 * must be power of two for mask use to work 245 * 246 */ 247#define ARC_BUFC_NUMDATALISTS 16 248#define ARC_BUFC_NUMMETADATALISTS 16 249#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 250 251typedef struct arc_state { 252 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 253 uint64_t arcs_size; /* total amount of data in this state */ 254 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 255 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 256} arc_state_t; 257 258#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 259 260/* The 6 states: */ 261static arc_state_t ARC_anon; 262static arc_state_t ARC_mru; 263static arc_state_t ARC_mru_ghost; 264static arc_state_t ARC_mfu; 265static arc_state_t ARC_mfu_ghost; 266static arc_state_t ARC_l2c_only; 267 268typedef struct arc_stats { 269 kstat_named_t arcstat_hits; 270 kstat_named_t arcstat_misses; 271 kstat_named_t arcstat_demand_data_hits; 272 kstat_named_t arcstat_demand_data_misses; 273 kstat_named_t arcstat_demand_metadata_hits; 274 kstat_named_t arcstat_demand_metadata_misses; 275 kstat_named_t arcstat_prefetch_data_hits; 276 kstat_named_t arcstat_prefetch_data_misses; 277 kstat_named_t arcstat_prefetch_metadata_hits; 278 kstat_named_t arcstat_prefetch_metadata_misses; 279 kstat_named_t arcstat_mru_hits; 280 kstat_named_t arcstat_mru_ghost_hits; 281 kstat_named_t arcstat_mfu_hits; 282 kstat_named_t arcstat_mfu_ghost_hits; 283 kstat_named_t arcstat_allocated; 284 kstat_named_t arcstat_deleted; 285 kstat_named_t arcstat_stolen; 286 kstat_named_t arcstat_recycle_miss; 287 kstat_named_t arcstat_mutex_miss; 288 kstat_named_t arcstat_evict_skip; 289 kstat_named_t arcstat_evict_l2_cached; 290 kstat_named_t arcstat_evict_l2_eligible; 291 kstat_named_t arcstat_evict_l2_ineligible; 292 kstat_named_t arcstat_hash_elements; 293 kstat_named_t arcstat_hash_elements_max; 294 kstat_named_t arcstat_hash_collisions; 295 kstat_named_t arcstat_hash_chains; 296 kstat_named_t arcstat_hash_chain_max; 297 kstat_named_t arcstat_p; 298 kstat_named_t arcstat_c; 299 kstat_named_t arcstat_c_min; 300 kstat_named_t arcstat_c_max; 301 kstat_named_t arcstat_size; 302 kstat_named_t arcstat_hdr_size; 303 kstat_named_t arcstat_data_size; 304 kstat_named_t arcstat_other_size; 305 kstat_named_t arcstat_l2_hits; 306 kstat_named_t arcstat_l2_misses; 307 kstat_named_t arcstat_l2_feeds; 308 kstat_named_t arcstat_l2_rw_clash; 309 kstat_named_t arcstat_l2_read_bytes; 310 kstat_named_t arcstat_l2_write_bytes; 311 kstat_named_t arcstat_l2_writes_sent; 312 kstat_named_t arcstat_l2_writes_done; 313 kstat_named_t arcstat_l2_writes_error; 314 kstat_named_t arcstat_l2_writes_hdr_miss; 315 kstat_named_t arcstat_l2_evict_lock_retry; 316 kstat_named_t arcstat_l2_evict_reading; 317 kstat_named_t arcstat_l2_free_on_write; 318 kstat_named_t arcstat_l2_abort_lowmem; 319 kstat_named_t arcstat_l2_cksum_bad; 320 kstat_named_t arcstat_l2_io_error; 321 kstat_named_t arcstat_l2_size; 322 kstat_named_t arcstat_l2_hdr_size; 323 kstat_named_t arcstat_memory_throttle_count; 324 kstat_named_t arcstat_l2_write_trylock_fail; 325 kstat_named_t arcstat_l2_write_passed_headroom; 326 kstat_named_t arcstat_l2_write_spa_mismatch; 327 kstat_named_t arcstat_l2_write_in_l2; 328 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 329 kstat_named_t arcstat_l2_write_not_cacheable; 330 kstat_named_t arcstat_l2_write_full; 331 kstat_named_t arcstat_l2_write_buffer_iter; 332 kstat_named_t arcstat_l2_write_pios; 333 kstat_named_t arcstat_l2_write_bytes_written; 334 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 335 kstat_named_t arcstat_l2_write_buffer_list_iter; 336 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 337} arc_stats_t; 338 339static arc_stats_t arc_stats = { 340 { "hits", KSTAT_DATA_UINT64 }, 341 { "misses", KSTAT_DATA_UINT64 }, 342 { "demand_data_hits", KSTAT_DATA_UINT64 }, 343 { "demand_data_misses", KSTAT_DATA_UINT64 }, 344 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 345 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 346 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 347 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 348 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 349 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 350 { "mru_hits", KSTAT_DATA_UINT64 }, 351 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 352 { "mfu_hits", KSTAT_DATA_UINT64 }, 353 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 354 { "allocated", KSTAT_DATA_UINT64 }, 355 { "deleted", KSTAT_DATA_UINT64 }, 356 { "stolen", KSTAT_DATA_UINT64 }, 357 { "recycle_miss", KSTAT_DATA_UINT64 }, 358 { "mutex_miss", KSTAT_DATA_UINT64 }, 359 { "evict_skip", KSTAT_DATA_UINT64 }, 360 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 361 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 362 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 363 { "hash_elements", KSTAT_DATA_UINT64 }, 364 { "hash_elements_max", KSTAT_DATA_UINT64 }, 365 { "hash_collisions", KSTAT_DATA_UINT64 }, 366 { "hash_chains", KSTAT_DATA_UINT64 }, 367 { "hash_chain_max", KSTAT_DATA_UINT64 }, 368 { "p", KSTAT_DATA_UINT64 }, 369 { "c", KSTAT_DATA_UINT64 }, 370 { "c_min", KSTAT_DATA_UINT64 }, 371 { "c_max", KSTAT_DATA_UINT64 }, 372 { "size", KSTAT_DATA_UINT64 }, 373 { "hdr_size", KSTAT_DATA_UINT64 }, 374 { "data_size", KSTAT_DATA_UINT64 }, 375 { "other_size", KSTAT_DATA_UINT64 }, 376 { "l2_hits", KSTAT_DATA_UINT64 }, 377 { "l2_misses", KSTAT_DATA_UINT64 }, 378 { "l2_feeds", KSTAT_DATA_UINT64 }, 379 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 380 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 381 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 382 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 383 { "l2_writes_done", KSTAT_DATA_UINT64 }, 384 { "l2_writes_error", KSTAT_DATA_UINT64 }, 385 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 386 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 387 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 388 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 389 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 390 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 391 { "l2_io_error", KSTAT_DATA_UINT64 }, 392 { "l2_size", KSTAT_DATA_UINT64 }, 393 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 394 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 395 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 396 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 397 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 398 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 399 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 400 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 401 { "l2_write_full", KSTAT_DATA_UINT64 }, 402 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 403 { "l2_write_pios", KSTAT_DATA_UINT64 }, 404 { "l2_write_bytes_written", KSTAT_DATA_UINT64 }, 405 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 406 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 407 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 } 408}; 409 410#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 411 412#define ARCSTAT_INCR(stat, val) \ 413 atomic_add_64(&arc_stats.stat.value.ui64, (val)); 414 415#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 416#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 417 418#define ARCSTAT_MAX(stat, val) { \ 419 uint64_t m; \ 420 while ((val) > (m = arc_stats.stat.value.ui64) && \ 421 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 422 continue; \ 423} 424 425#define ARCSTAT_MAXSTAT(stat) \ 426 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 427 428/* 429 * We define a macro to allow ARC hits/misses to be easily broken down by 430 * two separate conditions, giving a total of four different subtypes for 431 * each of hits and misses (so eight statistics total). 432 */ 433#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 434 if (cond1) { \ 435 if (cond2) { \ 436 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 437 } else { \ 438 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 439 } \ 440 } else { \ 441 if (cond2) { \ 442 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 443 } else { \ 444 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 445 } \ 446 } 447 448kstat_t *arc_ksp; 449static arc_state_t *arc_anon; 450static arc_state_t *arc_mru; 451static arc_state_t *arc_mru_ghost; 452static arc_state_t *arc_mfu; 453static arc_state_t *arc_mfu_ghost; 454static arc_state_t *arc_l2c_only; 455 456/* 457 * There are several ARC variables that are critical to export as kstats -- 458 * but we don't want to have to grovel around in the kstat whenever we wish to 459 * manipulate them. For these variables, we therefore define them to be in 460 * terms of the statistic variable. This assures that we are not introducing 461 * the possibility of inconsistency by having shadow copies of the variables, 462 * while still allowing the code to be readable. 463 */ 464#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 465#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 466#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 467#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 468#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 469 470static int arc_no_grow; /* Don't try to grow cache size */ 471static uint64_t arc_tempreserve; 472static uint64_t arc_meta_used; 473static uint64_t arc_meta_limit; 474static uint64_t arc_meta_max = 0; 475SYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RDTUN, 476 &arc_meta_used, 0, "ARC metadata used"); 477SYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RDTUN, 478 &arc_meta_limit, 0, "ARC metadata limit"); 479 480typedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 481 482typedef struct arc_callback arc_callback_t; 483 484struct arc_callback { 485 void *acb_private; 486 arc_done_func_t *acb_done; 487 arc_buf_t *acb_buf; 488 zio_t *acb_zio_dummy; 489 arc_callback_t *acb_next; 490}; 491 492typedef struct arc_write_callback arc_write_callback_t; 493 494struct arc_write_callback { 495 void *awcb_private; 496 arc_done_func_t *awcb_ready; 497 arc_done_func_t *awcb_done; 498 arc_buf_t *awcb_buf; 499}; 500 501struct arc_buf_hdr { 502 /* protected by hash lock */ 503 dva_t b_dva; 504 uint64_t b_birth; 505 uint64_t b_cksum0; 506 507 kmutex_t b_freeze_lock; 508 zio_cksum_t *b_freeze_cksum; 509 510 arc_buf_hdr_t *b_hash_next; 511 arc_buf_t *b_buf; 512 uint32_t b_flags; 513 uint32_t b_datacnt; 514 515 arc_callback_t *b_acb; 516 kcondvar_t b_cv; 517 518 /* immutable */ 519 arc_buf_contents_t b_type; 520 uint64_t b_size; 521 spa_t *b_spa; 522 523 /* protected by arc state mutex */ 524 arc_state_t *b_state; 525 list_node_t b_arc_node; 526 527 /* updated atomically */ 528 clock_t b_arc_access; 529 530 /* self protecting */ 531 refcount_t b_refcnt; 532 533 l2arc_buf_hdr_t *b_l2hdr; 534 list_node_t b_l2node; 535}; 536 537static arc_buf_t *arc_eviction_list; 538static kmutex_t arc_eviction_mtx; 539static arc_buf_hdr_t arc_eviction_hdr; 540static void arc_get_data_buf(arc_buf_t *buf); 541static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock); 542static int arc_evict_needed(arc_buf_contents_t type); 543static void arc_evict_ghost(arc_state_t *state, spa_t *spa, int64_t bytes); 544 545static boolean_t l2arc_write_eligible(spa_t *spa, arc_buf_hdr_t *ab); 546 547#define GHOST_STATE(state) \ 548 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 549 (state) == arc_l2c_only) 550 551/* 552 * Private ARC flags. These flags are private ARC only flags that will show up 553 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can 554 * be passed in as arc_flags in things like arc_read. However, these flags 555 * should never be passed and should only be set by ARC code. When adding new 556 * public flags, make sure not to smash the private ones. 557 */ 558 559#define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */ 560#define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */ 561#define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */ 562#define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */ 563#define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */ 564#define ARC_INDIRECT (1 << 14) /* this is an indirect block */ 565#define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */ 566#define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */ 567#define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */ 568#define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */ 569#define ARC_STORED (1 << 19) /* has been store()d to */ 570 571#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE) 572#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS) 573#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR) 574#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH) 575#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ) 576#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE) 577#define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS) 578#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE) 579#define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \ 580 (hdr)->b_l2hdr != NULL) 581#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING) 582#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED) 583#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD) 584 585/* 586 * Other sizes 587 */ 588 589#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 590#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 591 592/* 593 * Hash table routines 594 */ 595 596#define HT_LOCK_PAD CACHE_LINE_SIZE 597 598struct ht_lock { 599 kmutex_t ht_lock; 600#ifdef _KERNEL 601 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 602#endif 603}; 604 605#define BUF_LOCKS 256 606typedef struct buf_hash_table { 607 uint64_t ht_mask; 608 arc_buf_hdr_t **ht_table; 609 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 610} buf_hash_table_t; 611 612static buf_hash_table_t buf_hash_table; 613 614#define BUF_HASH_INDEX(spa, dva, birth) \ 615 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 616#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 617#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 618#define HDR_LOCK(buf) \ 619 (BUF_HASH_LOCK(BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth))) 620 621uint64_t zfs_crc64_table[256]; 622 623#ifdef ZIO_USE_UMA 624extern kmem_cache_t *zio_buf_cache[]; 625extern kmem_cache_t *zio_data_buf_cache[]; 626#endif 627 628/* 629 * Level 2 ARC 630 */ 631 632#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 633#define L2ARC_HEADROOM 2 /* num of writes */ 634#define L2ARC_FEED_SECS 1 /* caching interval secs */ 635#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 636 637#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 638#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 639 640/* 641 * L2ARC Performance Tunables 642 */ 643uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 644uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 645uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 646uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 647uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 648boolean_t l2arc_noprefetch = B_FALSE; /* don't cache prefetch bufs */ 649boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 650boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 651 652SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 653 &l2arc_write_max, 0, "max write size"); 654SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 655 &l2arc_write_boost, 0, "extra write during warmup"); 656SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 657 &l2arc_headroom, 0, "number of dev writes"); 658SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 659 &l2arc_feed_secs, 0, "interval seconds"); 660SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 661 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 662 663SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 664 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 665SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 666 &l2arc_feed_again, 0, "turbo warmup"); 667SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 668 &l2arc_norw, 0, "no reads during writes"); 669 670SYSCTL_QUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 671 &ARC_anon.arcs_size, 0, "size of anonymous state"); 672SYSCTL_QUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 673 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 674SYSCTL_QUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 675 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 676 677SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 678 &ARC_mru.arcs_size, 0, "size of mru state"); 679SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 680 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 681SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 682 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 683 684SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 685 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 686SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 687 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 688 "size of metadata in mru ghost state"); 689SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 690 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 691 "size of data in mru ghost state"); 692 693SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 694 &ARC_mfu.arcs_size, 0, "size of mfu state"); 695SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 696 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 697SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 698 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 699 700SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 701 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 702SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 703 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 704 "size of metadata in mfu ghost state"); 705SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 706 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 707 "size of data in mfu ghost state"); 708 709SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 710 &ARC_l2c_only.arcs_size, 0, "size of mru state"); 711 712/* 713 * L2ARC Internals 714 */ 715typedef struct l2arc_dev { 716 vdev_t *l2ad_vdev; /* vdev */ 717 spa_t *l2ad_spa; /* spa */ 718 uint64_t l2ad_hand; /* next write location */ 719 uint64_t l2ad_write; /* desired write size, bytes */ 720 uint64_t l2ad_boost; /* warmup write boost, bytes */ 721 uint64_t l2ad_start; /* first addr on device */ 722 uint64_t l2ad_end; /* last addr on device */ 723 uint64_t l2ad_evict; /* last addr eviction reached */ 724 boolean_t l2ad_first; /* first sweep through */ 725 boolean_t l2ad_writing; /* currently writing */ 726 list_t *l2ad_buflist; /* buffer list */ 727 list_node_t l2ad_node; /* device list node */ 728} l2arc_dev_t; 729 730static list_t L2ARC_dev_list; /* device list */ 731static list_t *l2arc_dev_list; /* device list pointer */ 732static kmutex_t l2arc_dev_mtx; /* device list mutex */ 733static l2arc_dev_t *l2arc_dev_last; /* last device used */ 734static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 735static list_t L2ARC_free_on_write; /* free after write buf list */ 736static list_t *l2arc_free_on_write; /* free after write list ptr */ 737static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 738static uint64_t l2arc_ndev; /* number of devices */ 739 740typedef struct l2arc_read_callback { 741 arc_buf_t *l2rcb_buf; /* read buffer */ 742 spa_t *l2rcb_spa; /* spa */ 743 blkptr_t l2rcb_bp; /* original blkptr */ 744 zbookmark_t l2rcb_zb; /* original bookmark */ 745 int l2rcb_flags; /* original flags */ 746} l2arc_read_callback_t; 747 748typedef struct l2arc_write_callback { 749 l2arc_dev_t *l2wcb_dev; /* device info */ 750 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 751} l2arc_write_callback_t; 752 753struct l2arc_buf_hdr { 754 /* protected by arc_buf_hdr mutex */ 755 l2arc_dev_t *b_dev; /* L2ARC device */ 756 uint64_t b_daddr; /* disk address, offset byte */ 757}; 758 759typedef struct l2arc_data_free { 760 /* protected by l2arc_free_on_write_mtx */ 761 void *l2df_data; 762 size_t l2df_size; 763 void (*l2df_func)(void *, size_t); 764 list_node_t l2df_list_node; 765} l2arc_data_free_t; 766 767static kmutex_t l2arc_feed_thr_lock; 768static kcondvar_t l2arc_feed_thr_cv; 769static uint8_t l2arc_thread_exit; 770 771static void l2arc_read_done(zio_t *zio); 772static void l2arc_hdr_stat_add(void); 773static void l2arc_hdr_stat_remove(void); 774 775static uint64_t 776buf_hash(spa_t *spa, const dva_t *dva, uint64_t birth) 777{ 778 uintptr_t spav = (uintptr_t)spa; 779 uint8_t *vdva = (uint8_t *)dva; 780 uint64_t crc = -1ULL; 781 int i; 782 783 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 784 785 for (i = 0; i < sizeof (dva_t); i++) 786 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 787 788 crc ^= (spav>>8) ^ birth; 789 790 return (crc); 791} 792 793#define BUF_EMPTY(buf) \ 794 ((buf)->b_dva.dva_word[0] == 0 && \ 795 (buf)->b_dva.dva_word[1] == 0 && \ 796 (buf)->b_birth == 0) 797 798#define BUF_EQUAL(spa, dva, birth, buf) \ 799 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 800 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 801 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 802 803static arc_buf_hdr_t * 804buf_hash_find(spa_t *spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp) 805{ 806 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 807 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 808 arc_buf_hdr_t *buf; 809 810 mutex_enter(hash_lock); 811 for (buf = buf_hash_table.ht_table[idx]; buf != NULL; 812 buf = buf->b_hash_next) { 813 if (BUF_EQUAL(spa, dva, birth, buf)) { 814 *lockp = hash_lock; 815 return (buf); 816 } 817 } 818 mutex_exit(hash_lock); 819 *lockp = NULL; 820 return (NULL); 821} 822 823/* 824 * Insert an entry into the hash table. If there is already an element 825 * equal to elem in the hash table, then the already existing element 826 * will be returned and the new element will not be inserted. 827 * Otherwise returns NULL. 828 */ 829static arc_buf_hdr_t * 830buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp) 831{ 832 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 833 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 834 arc_buf_hdr_t *fbuf; 835 uint32_t i; 836 837 ASSERT(!HDR_IN_HASH_TABLE(buf)); 838 *lockp = hash_lock; 839 mutex_enter(hash_lock); 840 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL; 841 fbuf = fbuf->b_hash_next, i++) { 842 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf)) 843 return (fbuf); 844 } 845 846 buf->b_hash_next = buf_hash_table.ht_table[idx]; 847 buf_hash_table.ht_table[idx] = buf; 848 buf->b_flags |= ARC_IN_HASH_TABLE; 849 850 /* collect some hash table performance data */ 851 if (i > 0) { 852 ARCSTAT_BUMP(arcstat_hash_collisions); 853 if (i == 1) 854 ARCSTAT_BUMP(arcstat_hash_chains); 855 856 ARCSTAT_MAX(arcstat_hash_chain_max, i); 857 } 858 859 ARCSTAT_BUMP(arcstat_hash_elements); 860 ARCSTAT_MAXSTAT(arcstat_hash_elements); 861 862 return (NULL); 863} 864 865static void 866buf_hash_remove(arc_buf_hdr_t *buf) 867{ 868 arc_buf_hdr_t *fbuf, **bufp; 869 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 870 871 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 872 ASSERT(HDR_IN_HASH_TABLE(buf)); 873 874 bufp = &buf_hash_table.ht_table[idx]; 875 while ((fbuf = *bufp) != buf) { 876 ASSERT(fbuf != NULL); 877 bufp = &fbuf->b_hash_next; 878 } 879 *bufp = buf->b_hash_next; 880 buf->b_hash_next = NULL; 881 buf->b_flags &= ~ARC_IN_HASH_TABLE; 882 883 /* collect some hash table performance data */ 884 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 885 886 if (buf_hash_table.ht_table[idx] && 887 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 888 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 889} 890 891/* 892 * Global data structures and functions for the buf kmem cache. 893 */ 894static kmem_cache_t *hdr_cache; 895static kmem_cache_t *buf_cache; 896 897static void 898buf_fini(void) 899{ 900 int i; 901 902 kmem_free(buf_hash_table.ht_table, 903 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 904 for (i = 0; i < BUF_LOCKS; i++) 905 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 906 kmem_cache_destroy(hdr_cache); 907 kmem_cache_destroy(buf_cache); 908} 909 910/* 911 * Constructor callback - called when the cache is empty 912 * and a new buf is requested. 913 */ 914/* ARGSUSED */ 915static int 916hdr_cons(void *vbuf, void *unused, int kmflag) 917{ 918 arc_buf_hdr_t *buf = vbuf; 919 920 bzero(buf, sizeof (arc_buf_hdr_t)); 921 refcount_create(&buf->b_refcnt); 922 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL); 923 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 924 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 925 926 return (0); 927} 928 929/* ARGSUSED */ 930static int 931buf_cons(void *vbuf, void *unused, int kmflag) 932{ 933 arc_buf_t *buf = vbuf; 934 935 bzero(buf, sizeof (arc_buf_t)); 936 rw_init(&buf->b_lock, NULL, RW_DEFAULT, NULL); 937 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 938 939 return (0); 940} 941 942/* 943 * Destructor callback - called when a cached buf is 944 * no longer required. 945 */ 946/* ARGSUSED */ 947static void 948hdr_dest(void *vbuf, void *unused) 949{ 950 arc_buf_hdr_t *buf = vbuf; 951 952 refcount_destroy(&buf->b_refcnt); 953 cv_destroy(&buf->b_cv); 954 mutex_destroy(&buf->b_freeze_lock); 955 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 956} 957 958/* ARGSUSED */ 959static void 960buf_dest(void *vbuf, void *unused) 961{ 962 arc_buf_t *buf = vbuf; 963 964 rw_destroy(&buf->b_lock); 965 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 966} 967 968/* 969 * Reclaim callback -- invoked when memory is low. 970 */ 971/* ARGSUSED */ 972static void 973hdr_recl(void *unused) 974{ 975 dprintf("hdr_recl called\n"); 976 /* 977 * umem calls the reclaim func when we destroy the buf cache, 978 * which is after we do arc_fini(). 979 */ 980 if (!arc_dead) 981 cv_signal(&arc_reclaim_thr_cv); 982} 983 984static void 985buf_init(void) 986{ 987 uint64_t *ct; 988 uint64_t hsize = 1ULL << 12; 989 int i, j; 990 991 /* 992 * The hash table is big enough to fill all of physical memory 993 * with an average 64K block size. The table will take up 994 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers). 995 */ 996 while (hsize * 65536 < (uint64_t)physmem * PAGESIZE) 997 hsize <<= 1; 998retry: 999 buf_hash_table.ht_mask = hsize - 1; 1000 buf_hash_table.ht_table = 1001 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1002 if (buf_hash_table.ht_table == NULL) { 1003 ASSERT(hsize > (1ULL << 8)); 1004 hsize >>= 1; 1005 goto retry; 1006 } 1007 1008 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1009 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1010 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1011 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1012 1013 for (i = 0; i < 256; i++) 1014 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1015 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1016 1017 for (i = 0; i < BUF_LOCKS; i++) { 1018 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1019 NULL, MUTEX_DEFAULT, NULL); 1020 } 1021} 1022 1023#define ARC_MINTIME (hz>>4) /* 62 ms */ 1024 1025static void 1026arc_cksum_verify(arc_buf_t *buf) 1027{ 1028 zio_cksum_t zc; 1029 1030 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1031 return; 1032 1033 mutex_enter(&buf->b_hdr->b_freeze_lock); 1034 if (buf->b_hdr->b_freeze_cksum == NULL || 1035 (buf->b_hdr->b_flags & ARC_IO_ERROR)) { 1036 mutex_exit(&buf->b_hdr->b_freeze_lock); 1037 return; 1038 } 1039 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1040 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1041 panic("buffer modified while frozen!"); 1042 mutex_exit(&buf->b_hdr->b_freeze_lock); 1043} 1044 1045static int 1046arc_cksum_equal(arc_buf_t *buf) 1047{ 1048 zio_cksum_t zc; 1049 int equal; 1050 1051 mutex_enter(&buf->b_hdr->b_freeze_lock); 1052 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1053 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1054 mutex_exit(&buf->b_hdr->b_freeze_lock); 1055 1056 return (equal); 1057} 1058 1059static void 1060arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1061{ 1062 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1063 return; 1064 1065 mutex_enter(&buf->b_hdr->b_freeze_lock); 1066 if (buf->b_hdr->b_freeze_cksum != NULL) { 1067 mutex_exit(&buf->b_hdr->b_freeze_lock); 1068 return; 1069 } 1070 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1071 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1072 buf->b_hdr->b_freeze_cksum); 1073 mutex_exit(&buf->b_hdr->b_freeze_lock); 1074} 1075 1076void 1077arc_buf_thaw(arc_buf_t *buf) 1078{ 1079 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1080 if (buf->b_hdr->b_state != arc_anon) 1081 panic("modifying non-anon buffer!"); 1082 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS) 1083 panic("modifying buffer while i/o in progress!"); 1084 arc_cksum_verify(buf); 1085 } 1086 1087 mutex_enter(&buf->b_hdr->b_freeze_lock); 1088 if (buf->b_hdr->b_freeze_cksum != NULL) { 1089 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1090 buf->b_hdr->b_freeze_cksum = NULL; 1091 } 1092 mutex_exit(&buf->b_hdr->b_freeze_lock); 1093} 1094 1095void 1096arc_buf_freeze(arc_buf_t *buf) 1097{ 1098 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1099 return; 1100 1101 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1102 buf->b_hdr->b_state == arc_anon); 1103 arc_cksum_compute(buf, B_FALSE); 1104} 1105 1106static void 1107get_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock) 1108{ 1109 uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth); 1110 1111 if (ab->b_type == ARC_BUFC_METADATA) 1112 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1113 else { 1114 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1115 buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1116 } 1117 1118 *list = &state->arcs_lists[buf_hashid]; 1119 *lock = ARCS_LOCK(state, buf_hashid); 1120} 1121 1122 1123static void 1124add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1125{ 1126 1127 ASSERT(MUTEX_HELD(hash_lock)); 1128 1129 if ((refcount_add(&ab->b_refcnt, tag) == 1) && 1130 (ab->b_state != arc_anon)) { 1131 uint64_t delta = ab->b_size * ab->b_datacnt; 1132 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type]; 1133 list_t *list; 1134 kmutex_t *lock; 1135 1136 get_buf_info(ab, ab->b_state, &list, &lock); 1137 ASSERT(!MUTEX_HELD(lock)); 1138 mutex_enter(lock); 1139 ASSERT(list_link_active(&ab->b_arc_node)); 1140 list_remove(list, ab); 1141 if (GHOST_STATE(ab->b_state)) { 1142 ASSERT3U(ab->b_datacnt, ==, 0); 1143 ASSERT3P(ab->b_buf, ==, NULL); 1144 delta = ab->b_size; 1145 } 1146 ASSERT(delta > 0); 1147 ASSERT3U(*size, >=, delta); 1148 atomic_add_64(size, -delta); 1149 mutex_exit(lock); 1150 /* remove the prefetch flag if we get a reference */ 1151 if (ab->b_flags & ARC_PREFETCH) 1152 ab->b_flags &= ~ARC_PREFETCH; 1153 } 1154} 1155 1156static int 1157remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1158{ 1159 int cnt; 1160 arc_state_t *state = ab->b_state; 1161 1162 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1163 ASSERT(!GHOST_STATE(state)); 1164 1165 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) && 1166 (state != arc_anon)) { 1167 uint64_t *size = &state->arcs_lsize[ab->b_type]; 1168 list_t *list; 1169 kmutex_t *lock; 1170 1171 get_buf_info(ab, state, &list, &lock); 1172 ASSERT(!MUTEX_HELD(lock)); 1173 mutex_enter(lock); 1174 ASSERT(!list_link_active(&ab->b_arc_node)); 1175 list_insert_head(list, ab); 1176 ASSERT(ab->b_datacnt > 0); 1177 atomic_add_64(size, ab->b_size * ab->b_datacnt); 1178 mutex_exit(lock); 1179 } 1180 return (cnt); 1181} 1182 1183/* 1184 * Move the supplied buffer to the indicated state. The mutex 1185 * for the buffer must be held by the caller. 1186 */ 1187static void 1188arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock) 1189{ 1190 arc_state_t *old_state = ab->b_state; 1191 int64_t refcnt = refcount_count(&ab->b_refcnt); 1192 uint64_t from_delta, to_delta; 1193 list_t *list; 1194 kmutex_t *lock; 1195 1196 ASSERT(MUTEX_HELD(hash_lock)); 1197 ASSERT(new_state != old_state); 1198 ASSERT(refcnt == 0 || ab->b_datacnt > 0); 1199 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state)); 1200 1201 from_delta = to_delta = ab->b_datacnt * ab->b_size; 1202 1203 /* 1204 * If this buffer is evictable, transfer it from the 1205 * old state list to the new state list. 1206 */ 1207 if (refcnt == 0) { 1208 if (old_state != arc_anon) { 1209 int use_mutex; 1210 uint64_t *size = &old_state->arcs_lsize[ab->b_type]; 1211 1212 get_buf_info(ab, old_state, &list, &lock); 1213 use_mutex = !MUTEX_HELD(lock); 1214 if (use_mutex) 1215 mutex_enter(lock); 1216 1217 ASSERT(list_link_active(&ab->b_arc_node)); 1218 list_remove(list, ab); 1219 1220 /* 1221 * If prefetching out of the ghost cache, 1222 * we will have a non-null datacnt. 1223 */ 1224 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) { 1225 /* ghost elements have a ghost size */ 1226 ASSERT(ab->b_buf == NULL); 1227 from_delta = ab->b_size; 1228 } 1229 ASSERT3U(*size, >=, from_delta); 1230 atomic_add_64(size, -from_delta); 1231 1232 if (use_mutex) 1233 mutex_exit(lock); 1234 } 1235 if (new_state != arc_anon) { 1236 int use_mutex; 1237 uint64_t *size = &new_state->arcs_lsize[ab->b_type]; 1238 1239 get_buf_info(ab, new_state, &list, &lock); 1240 use_mutex = !MUTEX_HELD(lock); 1241 if (use_mutex) 1242 mutex_enter(lock); 1243 1244 list_insert_head(list, ab); 1245 1246 /* ghost elements have a ghost size */ 1247 if (GHOST_STATE(new_state)) { 1248 ASSERT(ab->b_datacnt == 0); 1249 ASSERT(ab->b_buf == NULL); 1250 to_delta = ab->b_size; 1251 } 1252 atomic_add_64(size, to_delta); 1253 1254 if (use_mutex) 1255 mutex_exit(lock); 1256 } 1257 } 1258 1259 ASSERT(!BUF_EMPTY(ab)); 1260 if (new_state == arc_anon) { 1261 buf_hash_remove(ab); 1262 } 1263 1264 /* adjust state sizes */ 1265 if (to_delta) 1266 atomic_add_64(&new_state->arcs_size, to_delta); 1267 if (from_delta) { 1268 ASSERT3U(old_state->arcs_size, >=, from_delta); 1269 atomic_add_64(&old_state->arcs_size, -from_delta); 1270 } 1271 ab->b_state = new_state; 1272 1273 /* adjust l2arc hdr stats */ 1274 if (new_state == arc_l2c_only) 1275 l2arc_hdr_stat_add(); 1276 else if (old_state == arc_l2c_only) 1277 l2arc_hdr_stat_remove(); 1278} 1279 1280void 1281arc_space_consume(uint64_t space, arc_space_type_t type) 1282{ 1283 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1284 1285 switch (type) { 1286 case ARC_SPACE_DATA: 1287 ARCSTAT_INCR(arcstat_data_size, space); 1288 break; 1289 case ARC_SPACE_OTHER: 1290 ARCSTAT_INCR(arcstat_other_size, space); 1291 break; 1292 case ARC_SPACE_HDRS: 1293 ARCSTAT_INCR(arcstat_hdr_size, space); 1294 break; 1295 case ARC_SPACE_L2HDRS: 1296 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1297 break; 1298 } 1299 1300 atomic_add_64(&arc_meta_used, space); 1301 atomic_add_64(&arc_size, space); 1302} 1303 1304void 1305arc_space_return(uint64_t space, arc_space_type_t type) 1306{ 1307 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1308 1309 switch (type) { 1310 case ARC_SPACE_DATA: 1311 ARCSTAT_INCR(arcstat_data_size, -space); 1312 break; 1313 case ARC_SPACE_OTHER: 1314 ARCSTAT_INCR(arcstat_other_size, -space); 1315 break; 1316 case ARC_SPACE_HDRS: 1317 ARCSTAT_INCR(arcstat_hdr_size, -space); 1318 break; 1319 case ARC_SPACE_L2HDRS: 1320 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1321 break; 1322 } 1323 1324 ASSERT(arc_meta_used >= space); 1325 if (arc_meta_max < arc_meta_used) 1326 arc_meta_max = arc_meta_used; 1327 atomic_add_64(&arc_meta_used, -space); 1328 ASSERT(arc_size >= space); 1329 atomic_add_64(&arc_size, -space); 1330} 1331 1332void * 1333arc_data_buf_alloc(uint64_t size) 1334{ 1335 if (arc_evict_needed(ARC_BUFC_DATA)) 1336 cv_signal(&arc_reclaim_thr_cv); 1337 atomic_add_64(&arc_size, size); 1338 return (zio_data_buf_alloc(size)); 1339} 1340 1341void 1342arc_data_buf_free(void *buf, uint64_t size) 1343{ 1344 zio_data_buf_free(buf, size); 1345 ASSERT(arc_size >= size); 1346 atomic_add_64(&arc_size, -size); 1347} 1348 1349arc_buf_t * 1350arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1351{ 1352 arc_buf_hdr_t *hdr; 1353 arc_buf_t *buf; 1354 1355 ASSERT3U(size, >, 0); 1356 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1357 ASSERT(BUF_EMPTY(hdr)); 1358 hdr->b_size = size; 1359 hdr->b_type = type; 1360 hdr->b_spa = spa; 1361 hdr->b_state = arc_anon; 1362 hdr->b_arc_access = 0; 1363 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1364 buf->b_hdr = hdr; 1365 buf->b_data = NULL; 1366 buf->b_efunc = NULL; 1367 buf->b_private = NULL; 1368 buf->b_next = NULL; 1369 hdr->b_buf = buf; 1370 arc_get_data_buf(buf); 1371 hdr->b_datacnt = 1; 1372 hdr->b_flags = 0; 1373 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1374 (void) refcount_add(&hdr->b_refcnt, tag); 1375 1376 return (buf); 1377} 1378 1379static arc_buf_t * 1380arc_buf_clone(arc_buf_t *from) 1381{ 1382 arc_buf_t *buf; 1383 arc_buf_hdr_t *hdr = from->b_hdr; 1384 uint64_t size = hdr->b_size; 1385 1386 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1387 buf->b_hdr = hdr; 1388 buf->b_data = NULL; 1389 buf->b_efunc = NULL; 1390 buf->b_private = NULL; 1391 buf->b_next = hdr->b_buf; 1392 hdr->b_buf = buf; 1393 arc_get_data_buf(buf); 1394 bcopy(from->b_data, buf->b_data, size); 1395 hdr->b_datacnt += 1; 1396 return (buf); 1397} 1398 1399void 1400arc_buf_add_ref(arc_buf_t *buf, void* tag) 1401{ 1402 arc_buf_hdr_t *hdr; 1403 kmutex_t *hash_lock; 1404 1405 /* 1406 * Check to see if this buffer is evicted. Callers 1407 * must verify b_data != NULL to know if the add_ref 1408 * was successful. 1409 */ 1410 rw_enter(&buf->b_lock, RW_READER); 1411 if (buf->b_data == NULL) { 1412 rw_exit(&buf->b_lock); 1413 return; 1414 } 1415 hdr = buf->b_hdr; 1416 ASSERT(hdr != NULL); 1417 hash_lock = HDR_LOCK(hdr); 1418 mutex_enter(hash_lock); 1419 rw_exit(&buf->b_lock); 1420 1421 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1422 add_reference(hdr, hash_lock, tag); 1423 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1424 arc_access(hdr, hash_lock); 1425 mutex_exit(hash_lock); 1426 ARCSTAT_BUMP(arcstat_hits); 1427 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 1428 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1429 data, metadata, hits); 1430} 1431 1432/* 1433 * Free the arc data buffer. If it is an l2arc write in progress, 1434 * the buffer is placed on l2arc_free_on_write to be freed later. 1435 */ 1436static void 1437arc_buf_data_free(arc_buf_hdr_t *hdr, void (*free_func)(void *, size_t), 1438 void *data, size_t size) 1439{ 1440 if (HDR_L2_WRITING(hdr)) { 1441 l2arc_data_free_t *df; 1442 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1443 df->l2df_data = data; 1444 df->l2df_size = size; 1445 df->l2df_func = free_func; 1446 mutex_enter(&l2arc_free_on_write_mtx); 1447 list_insert_head(l2arc_free_on_write, df); 1448 mutex_exit(&l2arc_free_on_write_mtx); 1449 ARCSTAT_BUMP(arcstat_l2_free_on_write); 1450 } else { 1451 free_func(data, size); 1452 } 1453} 1454 1455static void 1456arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all) 1457{ 1458 arc_buf_t **bufp; 1459 1460 /* free up data associated with the buf */ 1461 if (buf->b_data) { 1462 arc_state_t *state = buf->b_hdr->b_state; 1463 uint64_t size = buf->b_hdr->b_size; 1464 arc_buf_contents_t type = buf->b_hdr->b_type; 1465 1466 arc_cksum_verify(buf); 1467 if (!recycle) { 1468 if (type == ARC_BUFC_METADATA) { 1469 arc_buf_data_free(buf->b_hdr, zio_buf_free, 1470 buf->b_data, size); 1471 arc_space_return(size, ARC_SPACE_DATA); 1472 } else { 1473 ASSERT(type == ARC_BUFC_DATA); 1474 arc_buf_data_free(buf->b_hdr, 1475 zio_data_buf_free, buf->b_data, size); 1476 ARCSTAT_INCR(arcstat_data_size, -size); 1477 atomic_add_64(&arc_size, -size); 1478 } 1479 } 1480 if (list_link_active(&buf->b_hdr->b_arc_node)) { 1481 uint64_t *cnt = &state->arcs_lsize[type]; 1482 1483 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1484 ASSERT(state != arc_anon); 1485 1486 ASSERT3U(*cnt, >=, size); 1487 atomic_add_64(cnt, -size); 1488 } 1489 ASSERT3U(state->arcs_size, >=, size); 1490 atomic_add_64(&state->arcs_size, -size); 1491 buf->b_data = NULL; 1492 ASSERT(buf->b_hdr->b_datacnt > 0); 1493 buf->b_hdr->b_datacnt -= 1; 1494 } 1495 1496 /* only remove the buf if requested */ 1497 if (!all) 1498 return; 1499 1500 /* remove the buf from the hdr list */ 1501 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1502 continue; 1503 *bufp = buf->b_next; 1504 1505 ASSERT(buf->b_efunc == NULL); 1506 1507 /* clean up the buf */ 1508 buf->b_hdr = NULL; 1509 kmem_cache_free(buf_cache, buf); 1510} 1511 1512static void 1513arc_hdr_destroy(arc_buf_hdr_t *hdr) 1514{ 1515 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1516 ASSERT3P(hdr->b_state, ==, arc_anon); 1517 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1518 ASSERT(!(hdr->b_flags & ARC_STORED)); 1519 1520 if (hdr->b_l2hdr != NULL) { 1521 if (!MUTEX_HELD(&l2arc_buflist_mtx)) { 1522 /* 1523 * To prevent arc_free() and l2arc_evict() from 1524 * attempting to free the same buffer at the same time, 1525 * a FREE_IN_PROGRESS flag is given to arc_free() to 1526 * give it priority. l2arc_evict() can't destroy this 1527 * header while we are waiting on l2arc_buflist_mtx. 1528 * 1529 * The hdr may be removed from l2ad_buflist before we 1530 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1531 */ 1532 mutex_enter(&l2arc_buflist_mtx); 1533 if (hdr->b_l2hdr != NULL) { 1534 list_remove(hdr->b_l2hdr->b_dev->l2ad_buflist, 1535 hdr); 1536 } 1537 mutex_exit(&l2arc_buflist_mtx); 1538 } else { 1539 list_remove(hdr->b_l2hdr->b_dev->l2ad_buflist, hdr); 1540 } 1541 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1542 kmem_free(hdr->b_l2hdr, sizeof (l2arc_buf_hdr_t)); 1543 if (hdr->b_state == arc_l2c_only) 1544 l2arc_hdr_stat_remove(); 1545 hdr->b_l2hdr = NULL; 1546 } 1547 1548 if (!BUF_EMPTY(hdr)) { 1549 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1550 bzero(&hdr->b_dva, sizeof (dva_t)); 1551 hdr->b_birth = 0; 1552 hdr->b_cksum0 = 0; 1553 } 1554 while (hdr->b_buf) { 1555 arc_buf_t *buf = hdr->b_buf; 1556 1557 if (buf->b_efunc) { 1558 mutex_enter(&arc_eviction_mtx); 1559 rw_enter(&buf->b_lock, RW_WRITER); 1560 ASSERT(buf->b_hdr != NULL); 1561 arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1562 hdr->b_buf = buf->b_next; 1563 buf->b_hdr = &arc_eviction_hdr; 1564 buf->b_next = arc_eviction_list; 1565 arc_eviction_list = buf; 1566 rw_exit(&buf->b_lock); 1567 mutex_exit(&arc_eviction_mtx); 1568 } else { 1569 arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1570 } 1571 } 1572 if (hdr->b_freeze_cksum != NULL) { 1573 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1574 hdr->b_freeze_cksum = NULL; 1575 } 1576 1577 ASSERT(!list_link_active(&hdr->b_arc_node)); 1578 ASSERT3P(hdr->b_hash_next, ==, NULL); 1579 ASSERT3P(hdr->b_acb, ==, NULL); 1580 kmem_cache_free(hdr_cache, hdr); 1581} 1582 1583void 1584arc_buf_free(arc_buf_t *buf, void *tag) 1585{ 1586 arc_buf_hdr_t *hdr = buf->b_hdr; 1587 int hashed = hdr->b_state != arc_anon; 1588 1589 ASSERT(buf->b_efunc == NULL); 1590 ASSERT(buf->b_data != NULL); 1591 1592 if (hashed) { 1593 kmutex_t *hash_lock = HDR_LOCK(hdr); 1594 1595 mutex_enter(hash_lock); 1596 (void) remove_reference(hdr, hash_lock, tag); 1597 if (hdr->b_datacnt > 1) 1598 arc_buf_destroy(buf, FALSE, TRUE); 1599 else 1600 hdr->b_flags |= ARC_BUF_AVAILABLE; 1601 mutex_exit(hash_lock); 1602 } else if (HDR_IO_IN_PROGRESS(hdr)) { 1603 int destroy_hdr; 1604 /* 1605 * We are in the middle of an async write. Don't destroy 1606 * this buffer unless the write completes before we finish 1607 * decrementing the reference count. 1608 */ 1609 mutex_enter(&arc_eviction_mtx); 1610 (void) remove_reference(hdr, NULL, tag); 1611 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1612 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1613 mutex_exit(&arc_eviction_mtx); 1614 if (destroy_hdr) 1615 arc_hdr_destroy(hdr); 1616 } else { 1617 if (remove_reference(hdr, NULL, tag) > 0) { 1618 ASSERT(HDR_IO_ERROR(hdr)); 1619 arc_buf_destroy(buf, FALSE, TRUE); 1620 } else { 1621 arc_hdr_destroy(hdr); 1622 } 1623 } 1624} 1625 1626int 1627arc_buf_remove_ref(arc_buf_t *buf, void* tag) 1628{ 1629 arc_buf_hdr_t *hdr = buf->b_hdr; 1630 kmutex_t *hash_lock = HDR_LOCK(hdr); 1631 int no_callback = (buf->b_efunc == NULL); 1632 1633 if (hdr->b_state == arc_anon) { 1634 arc_buf_free(buf, tag); 1635 return (no_callback); 1636 } 1637 1638 mutex_enter(hash_lock); 1639 ASSERT(hdr->b_state != arc_anon); 1640 ASSERT(buf->b_data != NULL); 1641 1642 (void) remove_reference(hdr, hash_lock, tag); 1643 if (hdr->b_datacnt > 1) { 1644 if (no_callback) 1645 arc_buf_destroy(buf, FALSE, TRUE); 1646 } else if (no_callback) { 1647 ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1648 hdr->b_flags |= ARC_BUF_AVAILABLE; 1649 } 1650 ASSERT(no_callback || hdr->b_datacnt > 1 || 1651 refcount_is_zero(&hdr->b_refcnt)); 1652 mutex_exit(hash_lock); 1653 return (no_callback); 1654} 1655 1656int 1657arc_buf_size(arc_buf_t *buf) 1658{ 1659 return (buf->b_hdr->b_size); 1660} 1661 1662/* 1663 * Evict buffers from list until we've removed the specified number of 1664 * bytes. Move the removed buffers to the appropriate evict state. 1665 * If the recycle flag is set, then attempt to "recycle" a buffer: 1666 * - look for a buffer to evict that is `bytes' long. 1667 * - return the data block from this buffer rather than freeing it. 1668 * This flag is used by callers that are trying to make space for a 1669 * new buffer in a full arc cache. 1670 * 1671 * This function makes a "best effort". It skips over any buffers 1672 * it can't get a hash_lock on, and so may not catch all candidates. 1673 * It may also return without evicting as much space as requested. 1674 */ 1675static void * 1676arc_evict(arc_state_t *state, spa_t *spa, int64_t bytes, boolean_t recycle, 1677 arc_buf_contents_t type) 1678{ 1679 arc_state_t *evicted_state; 1680 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 1681 int64_t bytes_remaining; 1682 arc_buf_hdr_t *ab, *ab_prev = NULL; 1683 list_t *evicted_list, *list, *evicted_list_start, *list_start; 1684 kmutex_t *lock, *evicted_lock; 1685 kmutex_t *hash_lock; 1686 boolean_t have_lock; 1687 void *stolen = NULL; 1688 static int evict_metadata_offset, evict_data_offset; 1689 int i, idx, offset, list_count, count; 1690 1691 ASSERT(state == arc_mru || state == arc_mfu); 1692 1693 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 1694 1695 if (type == ARC_BUFC_METADATA) { 1696 offset = 0; 1697 list_count = ARC_BUFC_NUMMETADATALISTS; 1698 list_start = &state->arcs_lists[0]; 1699 evicted_list_start = &evicted_state->arcs_lists[0]; 1700 idx = evict_metadata_offset; 1701 } else { 1702 offset = ARC_BUFC_NUMMETADATALISTS; 1703 list_start = &state->arcs_lists[offset]; 1704 evicted_list_start = &evicted_state->arcs_lists[offset]; 1705 list_count = ARC_BUFC_NUMDATALISTS; 1706 idx = evict_data_offset; 1707 } 1708 bytes_remaining = evicted_state->arcs_lsize[type]; 1709 count = 0; 1710 1711evict_start: 1712 list = &list_start[idx]; 1713 evicted_list = &evicted_list_start[idx]; 1714 lock = ARCS_LOCK(state, (offset + idx)); 1715 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 1716 1717 mutex_enter(lock); 1718 mutex_enter(evicted_lock); 1719 1720 for (ab = list_tail(list); ab; ab = ab_prev) { 1721 ab_prev = list_prev(list, ab); 1722 bytes_remaining -= (ab->b_size * ab->b_datacnt); 1723 /* prefetch buffers have a minimum lifespan */ 1724 if (HDR_IO_IN_PROGRESS(ab) || 1725 (spa && ab->b_spa != spa) || 1726 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 1727 LBOLT - ab->b_arc_access < arc_min_prefetch_lifespan)) { 1728 skipped++; 1729 continue; 1730 } 1731 /* "lookahead" for better eviction candidate */ 1732 if (recycle && ab->b_size != bytes && 1733 ab_prev && ab_prev->b_size == bytes) 1734 continue; 1735 hash_lock = HDR_LOCK(ab); 1736 have_lock = MUTEX_HELD(hash_lock); 1737 if (have_lock || mutex_tryenter(hash_lock)) { 1738 ASSERT3U(refcount_count(&ab->b_refcnt), ==, 0); 1739 ASSERT(ab->b_datacnt > 0); 1740 while (ab->b_buf) { 1741 arc_buf_t *buf = ab->b_buf; 1742 if (!rw_tryenter(&buf->b_lock, RW_WRITER)) { 1743 missed += 1; 1744 break; 1745 } 1746 if (buf->b_data) { 1747 bytes_evicted += ab->b_size; 1748 if (recycle && ab->b_type == type && 1749 ab->b_size == bytes && 1750 !HDR_L2_WRITING(ab)) { 1751 stolen = buf->b_data; 1752 recycle = FALSE; 1753 } 1754 } 1755 if (buf->b_efunc) { 1756 mutex_enter(&arc_eviction_mtx); 1757 arc_buf_destroy(buf, 1758 buf->b_data == stolen, FALSE); 1759 ab->b_buf = buf->b_next; 1760 buf->b_hdr = &arc_eviction_hdr; 1761 buf->b_next = arc_eviction_list; 1762 arc_eviction_list = buf; 1763 mutex_exit(&arc_eviction_mtx); 1764 rw_exit(&buf->b_lock); 1765 } else { 1766 rw_exit(&buf->b_lock); 1767 arc_buf_destroy(buf, 1768 buf->b_data == stolen, TRUE); 1769 } 1770 } 1771 1772 if (ab->b_l2hdr) { 1773 ARCSTAT_INCR(arcstat_evict_l2_cached, 1774 ab->b_size); 1775 } else { 1776 if (l2arc_write_eligible(ab->b_spa, ab)) { 1777 ARCSTAT_INCR(arcstat_evict_l2_eligible, 1778 ab->b_size); 1779 } else { 1780 ARCSTAT_INCR( 1781 arcstat_evict_l2_ineligible, 1782 ab->b_size); 1783 } 1784 } 1785 1786 if (ab->b_datacnt == 0) { 1787 arc_change_state(evicted_state, ab, hash_lock); 1788 ASSERT(HDR_IN_HASH_TABLE(ab)); 1789 ab->b_flags |= ARC_IN_HASH_TABLE; 1790 ab->b_flags &= ~ARC_BUF_AVAILABLE; 1791 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 1792 } 1793 if (!have_lock) 1794 mutex_exit(hash_lock); 1795 if (bytes >= 0 && bytes_evicted >= bytes) 1796 break; 1797 if (bytes_remaining > 0) { 1798 mutex_exit(evicted_lock); 1799 mutex_exit(lock); 1800 idx = ((idx + 1) & (list_count - 1)); 1801 count++; 1802 goto evict_start; 1803 } 1804 } else { 1805 missed += 1; 1806 } 1807 } 1808 1809 mutex_exit(evicted_lock); 1810 mutex_exit(lock); 1811 1812 idx = ((idx + 1) & (list_count - 1)); 1813 count++; 1814 1815 if (bytes_evicted < bytes) { 1816 if (count < list_count) 1817 goto evict_start; 1818 else 1819 dprintf("only evicted %lld bytes from %x", 1820 (longlong_t)bytes_evicted, state); 1821 } 1822 if (type == ARC_BUFC_METADATA) 1823 evict_metadata_offset = idx; 1824 else 1825 evict_data_offset = idx; 1826 1827 if (skipped) 1828 ARCSTAT_INCR(arcstat_evict_skip, skipped); 1829 1830 if (missed) 1831 ARCSTAT_INCR(arcstat_mutex_miss, missed); 1832 1833 /* 1834 * We have just evicted some date into the ghost state, make 1835 * sure we also adjust the ghost state size if necessary. 1836 */ 1837 if (arc_no_grow && 1838 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) { 1839 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size + 1840 arc_mru_ghost->arcs_size - arc_c; 1841 1842 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) { 1843 int64_t todelete = 1844 MIN(arc_mru_ghost->arcs_lsize[type], mru_over); 1845 arc_evict_ghost(arc_mru_ghost, NULL, todelete); 1846 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) { 1847 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type], 1848 arc_mru_ghost->arcs_size + 1849 arc_mfu_ghost->arcs_size - arc_c); 1850 arc_evict_ghost(arc_mfu_ghost, NULL, todelete); 1851 } 1852 } 1853 if (stolen) 1854 ARCSTAT_BUMP(arcstat_stolen); 1855 1856 return (stolen); 1857} 1858 1859/* 1860 * Remove buffers from list until we've removed the specified number of 1861 * bytes. Destroy the buffers that are removed. 1862 */ 1863static void 1864arc_evict_ghost(arc_state_t *state, spa_t *spa, int64_t bytes) 1865{ 1866 arc_buf_hdr_t *ab, *ab_prev; 1867 list_t *list, *list_start; 1868 kmutex_t *hash_lock, *lock; 1869 uint64_t bytes_deleted = 0; 1870 uint64_t bufs_skipped = 0; 1871 static int evict_offset; 1872 int list_count, idx = evict_offset; 1873 int offset, count = 0; 1874 1875 ASSERT(GHOST_STATE(state)); 1876 1877 /* 1878 * data lists come after metadata lists 1879 */ 1880 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 1881 list_count = ARC_BUFC_NUMDATALISTS; 1882 offset = ARC_BUFC_NUMMETADATALISTS; 1883 1884evict_start: 1885 list = &list_start[idx]; 1886 lock = ARCS_LOCK(state, idx + offset); 1887 1888 mutex_enter(lock); 1889 for (ab = list_tail(list); ab; ab = ab_prev) { 1890 ab_prev = list_prev(list, ab); 1891 if (spa && ab->b_spa != spa) 1892 continue; 1893 hash_lock = HDR_LOCK(ab); 1894 if (mutex_tryenter(hash_lock)) { 1895 ASSERT(!HDR_IO_IN_PROGRESS(ab)); 1896 ASSERT(ab->b_buf == NULL); 1897 ARCSTAT_BUMP(arcstat_deleted); 1898 bytes_deleted += ab->b_size; 1899 1900 if (ab->b_l2hdr != NULL) { 1901 /* 1902 * This buffer is cached on the 2nd Level ARC; 1903 * don't destroy the header. 1904 */ 1905 arc_change_state(arc_l2c_only, ab, hash_lock); 1906 mutex_exit(hash_lock); 1907 } else { 1908 arc_change_state(arc_anon, ab, hash_lock); 1909 mutex_exit(hash_lock); 1910 arc_hdr_destroy(ab); 1911 } 1912 1913 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 1914 if (bytes >= 0 && bytes_deleted >= bytes) 1915 break; 1916 } else { 1917 if (bytes < 0) { 1918 /* 1919 * we're draining the ARC, retry 1920 */ 1921 mutex_exit(lock); 1922 mutex_enter(hash_lock); 1923 mutex_exit(hash_lock); 1924 goto evict_start; 1925 } 1926 bufs_skipped += 1; 1927 } 1928 } 1929 mutex_exit(lock); 1930 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 1931 count++; 1932 1933 if (count < list_count) 1934 goto evict_start; 1935 1936 evict_offset = idx; 1937 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 1938 (bytes < 0 || bytes_deleted < bytes)) { 1939 list_start = &state->arcs_lists[0]; 1940 list_count = ARC_BUFC_NUMMETADATALISTS; 1941 offset = count = 0; 1942 goto evict_start; 1943 } 1944 1945 if (bufs_skipped) { 1946 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 1947 ASSERT(bytes >= 0); 1948 } 1949 1950 if (bytes_deleted < bytes) 1951 dprintf("only deleted %lld bytes from %p", 1952 (longlong_t)bytes_deleted, state); 1953} 1954 1955static void 1956arc_adjust(void) 1957{ 1958 int64_t adjustment, delta; 1959 1960 /* 1961 * Adjust MRU size 1962 */ 1963 1964 adjustment = MIN(arc_size - arc_c, 1965 arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - arc_p); 1966 1967 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 1968 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 1969 (void) arc_evict(arc_mru, NULL, delta, FALSE, ARC_BUFC_DATA); 1970 adjustment -= delta; 1971 } 1972 1973 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 1974 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 1975 (void) arc_evict(arc_mru, NULL, delta, FALSE, 1976 ARC_BUFC_METADATA); 1977 } 1978 1979 /* 1980 * Adjust MFU size 1981 */ 1982 1983 adjustment = arc_size - arc_c; 1984 1985 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 1986 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 1987 (void) arc_evict(arc_mfu, NULL, delta, FALSE, ARC_BUFC_DATA); 1988 adjustment -= delta; 1989 } 1990 1991 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 1992 int64_t delta = MIN(adjustment, 1993 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 1994 (void) arc_evict(arc_mfu, NULL, delta, FALSE, 1995 ARC_BUFC_METADATA); 1996 } 1997 1998 /* 1999 * Adjust ghost lists 2000 */ 2001 2002 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2003 2004 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2005 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2006 arc_evict_ghost(arc_mru_ghost, NULL, delta); 2007 } 2008 2009 adjustment = 2010 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2011 2012 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2013 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2014 arc_evict_ghost(arc_mfu_ghost, NULL, delta); 2015 } 2016} 2017 2018static void 2019arc_do_user_evicts(void) 2020{ 2021 static arc_buf_t *tmp_arc_eviction_list; 2022 2023 /* 2024 * Move list over to avoid LOR 2025 */ 2026restart: 2027 mutex_enter(&arc_eviction_mtx); 2028 tmp_arc_eviction_list = arc_eviction_list; 2029 arc_eviction_list = NULL; 2030 mutex_exit(&arc_eviction_mtx); 2031 2032 while (tmp_arc_eviction_list != NULL) { 2033 arc_buf_t *buf = tmp_arc_eviction_list; 2034 tmp_arc_eviction_list = buf->b_next; 2035 rw_enter(&buf->b_lock, RW_WRITER); 2036 buf->b_hdr = NULL; 2037 rw_exit(&buf->b_lock); 2038 2039 if (buf->b_efunc != NULL) 2040 VERIFY(buf->b_efunc(buf) == 0); 2041 2042 buf->b_efunc = NULL; 2043 buf->b_private = NULL; 2044 kmem_cache_free(buf_cache, buf); 2045 } 2046 2047 if (arc_eviction_list != NULL) 2048 goto restart; 2049} 2050 2051/* 2052 * Flush all *evictable* data from the cache for the given spa. 2053 * NOTE: this will not touch "active" (i.e. referenced) data. 2054 */ 2055void 2056arc_flush(spa_t *spa) 2057{ 2058 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2059 (void) arc_evict(arc_mru, spa, -1, FALSE, ARC_BUFC_DATA); 2060 if (spa) 2061 break; 2062 } 2063 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2064 (void) arc_evict(arc_mru, spa, -1, FALSE, ARC_BUFC_METADATA); 2065 if (spa) 2066 break; 2067 } 2068 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2069 (void) arc_evict(arc_mfu, spa, -1, FALSE, ARC_BUFC_DATA); 2070 if (spa) 2071 break; 2072 } 2073 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2074 (void) arc_evict(arc_mfu, spa, -1, FALSE, ARC_BUFC_METADATA); 2075 if (spa) 2076 break; 2077 } 2078 2079 arc_evict_ghost(arc_mru_ghost, spa, -1); 2080 arc_evict_ghost(arc_mfu_ghost, spa, -1); 2081 2082 mutex_enter(&arc_reclaim_thr_lock); 2083 arc_do_user_evicts(); 2084 mutex_exit(&arc_reclaim_thr_lock); 2085 ASSERT(spa || arc_eviction_list == NULL); 2086} 2087 2088void 2089arc_shrink(void) 2090{ 2091 if (arc_c > arc_c_min) { 2092 uint64_t to_free; 2093 2094#ifdef _KERNEL 2095 to_free = arc_c >> arc_shrink_shift; 2096#else 2097 to_free = arc_c >> arc_shrink_shift; 2098#endif 2099 if (arc_c > arc_c_min + to_free) 2100 atomic_add_64(&arc_c, -to_free); 2101 else 2102 arc_c = arc_c_min; 2103 2104 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2105 if (arc_c > arc_size) 2106 arc_c = MAX(arc_size, arc_c_min); 2107 if (arc_p > arc_c) 2108 arc_p = (arc_c >> 1); 2109 ASSERT(arc_c >= arc_c_min); 2110 ASSERT((int64_t)arc_p >= 0); 2111 } 2112 2113 if (arc_size > arc_c) 2114 arc_adjust(); 2115} 2116 2117static int needfree = 0; 2118 2119static int 2120arc_reclaim_needed(void) 2121{ 2122#if 0 2123 uint64_t extra; 2124#endif 2125 2126#ifdef _KERNEL 2127 if (needfree) 2128 return (1); 2129 if (arc_size > arc_c_max) 2130 return (1); 2131 if (arc_size <= arc_c_min) 2132 return (0); 2133 2134 /* 2135 * If pages are needed or we're within 2048 pages 2136 * of needing to page need to reclaim 2137 */ 2138 if (vm_pages_needed || (vm_paging_target() > -2048)) 2139 return (1); 2140 2141#if 0 2142 /* 2143 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2144 */ 2145 extra = desfree; 2146 2147 /* 2148 * check that we're out of range of the pageout scanner. It starts to 2149 * schedule paging if freemem is less than lotsfree and needfree. 2150 * lotsfree is the high-water mark for pageout, and needfree is the 2151 * number of needed free pages. We add extra pages here to make sure 2152 * the scanner doesn't start up while we're freeing memory. 2153 */ 2154 if (freemem < lotsfree + needfree + extra) 2155 return (1); 2156 2157 /* 2158 * check to make sure that swapfs has enough space so that anon 2159 * reservations can still succeed. anon_resvmem() checks that the 2160 * availrmem is greater than swapfs_minfree, and the number of reserved 2161 * swap pages. We also add a bit of extra here just to prevent 2162 * circumstances from getting really dire. 2163 */ 2164 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2165 return (1); 2166 2167#if defined(__i386) 2168 /* 2169 * If we're on an i386 platform, it's possible that we'll exhaust the 2170 * kernel heap space before we ever run out of available physical 2171 * memory. Most checks of the size of the heap_area compare against 2172 * tune.t_minarmem, which is the minimum available real memory that we 2173 * can have in the system. However, this is generally fixed at 25 pages 2174 * which is so low that it's useless. In this comparison, we seek to 2175 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2176 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2177 * free) 2178 */ 2179 if (btop(vmem_size(heap_arena, VMEM_FREE)) < 2180 (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2)) 2181 return (1); 2182#endif 2183#else 2184 if (kmem_used() > (kmem_size() * 3) / 4) 2185 return (1); 2186#endif 2187 2188#else 2189 if (spa_get_random(100) == 0) 2190 return (1); 2191#endif 2192 return (0); 2193} 2194 2195static void 2196arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2197{ 2198#ifdef ZIO_USE_UMA 2199 size_t i; 2200 kmem_cache_t *prev_cache = NULL; 2201 kmem_cache_t *prev_data_cache = NULL; 2202#endif 2203 2204#ifdef _KERNEL 2205 if (arc_meta_used >= arc_meta_limit) { 2206 /* 2207 * We are exceeding our meta-data cache limit. 2208 * Purge some DNLC entries to release holds on meta-data. 2209 */ 2210 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2211 } 2212#if defined(__i386) 2213 /* 2214 * Reclaim unused memory from all kmem caches. 2215 */ 2216 kmem_reap(); 2217#endif 2218#endif 2219 2220 /* 2221 * An aggressive reclamation will shrink the cache size as well as 2222 * reap free buffers from the arc kmem caches. 2223 */ 2224 if (strat == ARC_RECLAIM_AGGR) 2225 arc_shrink(); 2226 2227#ifdef ZIO_USE_UMA 2228 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2229 if (zio_buf_cache[i] != prev_cache) { 2230 prev_cache = zio_buf_cache[i]; 2231 kmem_cache_reap_now(zio_buf_cache[i]); 2232 } 2233 if (zio_data_buf_cache[i] != prev_data_cache) { 2234 prev_data_cache = zio_data_buf_cache[i]; 2235 kmem_cache_reap_now(zio_data_buf_cache[i]); 2236 } 2237 } 2238#endif 2239 kmem_cache_reap_now(buf_cache); 2240 kmem_cache_reap_now(hdr_cache); 2241} 2242 2243static void 2244arc_reclaim_thread(void *dummy __unused) 2245{ 2246 clock_t growtime = 0; 2247 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2248 callb_cpr_t cpr; 2249 2250 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2251 2252 mutex_enter(&arc_reclaim_thr_lock); 2253 while (arc_thread_exit == 0) { 2254 if (arc_reclaim_needed()) { 2255 2256 if (arc_no_grow) { 2257 if (last_reclaim == ARC_RECLAIM_CONS) { 2258 last_reclaim = ARC_RECLAIM_AGGR; 2259 } else { 2260 last_reclaim = ARC_RECLAIM_CONS; 2261 } 2262 } else { 2263 arc_no_grow = TRUE; 2264 last_reclaim = ARC_RECLAIM_AGGR; 2265 membar_producer(); 2266 } 2267 2268 /* reset the growth delay for every reclaim */ 2269 growtime = LBOLT + (arc_grow_retry * hz); 2270 2271 if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2272 /* 2273 * If needfree is TRUE our vm_lowmem hook 2274 * was called and in that case we must free some 2275 * memory, so switch to aggressive mode. 2276 */ 2277 arc_no_grow = TRUE; 2278 last_reclaim = ARC_RECLAIM_AGGR; 2279 } 2280 arc_kmem_reap_now(last_reclaim); 2281 arc_warm = B_TRUE; 2282 2283 } else if (arc_no_grow && LBOLT >= growtime) { 2284 arc_no_grow = FALSE; 2285 } 2286 2287 if (needfree || 2288 (2 * arc_c < arc_size + 2289 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size)) 2290 arc_adjust(); 2291 2292 if (arc_eviction_list != NULL) 2293 arc_do_user_evicts(); 2294 2295 if (arc_reclaim_needed()) { 2296 needfree = 0; 2297#ifdef _KERNEL 2298 wakeup(&needfree); 2299#endif 2300 } 2301 2302 /* block until needed, or one second, whichever is shorter */ 2303 CALLB_CPR_SAFE_BEGIN(&cpr); 2304 (void) cv_timedwait(&arc_reclaim_thr_cv, 2305 &arc_reclaim_thr_lock, hz); 2306 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2307 } 2308 2309 arc_thread_exit = 0; 2310 cv_broadcast(&arc_reclaim_thr_cv); 2311 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2312 thread_exit(); 2313} 2314 2315/* 2316 * Adapt arc info given the number of bytes we are trying to add and 2317 * the state that we are comming from. This function is only called 2318 * when we are adding new content to the cache. 2319 */ 2320static void 2321arc_adapt(int bytes, arc_state_t *state) 2322{ 2323 int mult; 2324 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2325 2326 if (state == arc_l2c_only) 2327 return; 2328 2329 ASSERT(bytes > 0); 2330 /* 2331 * Adapt the target size of the MRU list: 2332 * - if we just hit in the MRU ghost list, then increase 2333 * the target size of the MRU list. 2334 * - if we just hit in the MFU ghost list, then increase 2335 * the target size of the MFU list by decreasing the 2336 * target size of the MRU list. 2337 */ 2338 if (state == arc_mru_ghost) { 2339 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2340 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2341 2342 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2343 } else if (state == arc_mfu_ghost) { 2344 uint64_t delta; 2345 2346 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2347 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2348 2349 delta = MIN(bytes * mult, arc_p); 2350 arc_p = MAX(arc_p_min, arc_p - delta); 2351 } 2352 ASSERT((int64_t)arc_p >= 0); 2353 2354 if (arc_reclaim_needed()) { 2355 cv_signal(&arc_reclaim_thr_cv); 2356 return; 2357 } 2358 2359 if (arc_no_grow) 2360 return; 2361 2362 if (arc_c >= arc_c_max) 2363 return; 2364 2365 /* 2366 * If we're within (2 * maxblocksize) bytes of the target 2367 * cache size, increment the target cache size 2368 */ 2369 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2370 atomic_add_64(&arc_c, (int64_t)bytes); 2371 if (arc_c > arc_c_max) 2372 arc_c = arc_c_max; 2373 else if (state == arc_anon) 2374 atomic_add_64(&arc_p, (int64_t)bytes); 2375 if (arc_p > arc_c) 2376 arc_p = arc_c; 2377 } 2378 ASSERT((int64_t)arc_p >= 0); 2379} 2380 2381/* 2382 * Check if the cache has reached its limits and eviction is required 2383 * prior to insert. 2384 */ 2385static int 2386arc_evict_needed(arc_buf_contents_t type) 2387{ 2388 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2389 return (1); 2390 2391#if 0 2392#ifdef _KERNEL 2393 /* 2394 * If zio data pages are being allocated out of a separate heap segment, 2395 * then enforce that the size of available vmem for this area remains 2396 * above about 1/32nd free. 2397 */ 2398 if (type == ARC_BUFC_DATA && zio_arena != NULL && 2399 vmem_size(zio_arena, VMEM_FREE) < 2400 (vmem_size(zio_arena, VMEM_ALLOC) >> 5)) 2401 return (1); 2402#endif 2403#endif 2404 2405 if (arc_reclaim_needed()) 2406 return (1); 2407 2408 return (arc_size > arc_c); 2409} 2410 2411/* 2412 * The buffer, supplied as the first argument, needs a data block. 2413 * So, if we are at cache max, determine which cache should be victimized. 2414 * We have the following cases: 2415 * 2416 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2417 * In this situation if we're out of space, but the resident size of the MFU is 2418 * under the limit, victimize the MFU cache to satisfy this insertion request. 2419 * 2420 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2421 * Here, we've used up all of the available space for the MRU, so we need to 2422 * evict from our own cache instead. Evict from the set of resident MRU 2423 * entries. 2424 * 2425 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2426 * c minus p represents the MFU space in the cache, since p is the size of the 2427 * cache that is dedicated to the MRU. In this situation there's still space on 2428 * the MFU side, so the MRU side needs to be victimized. 2429 * 2430 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2431 * MFU's resident set is consuming more space than it has been allotted. In 2432 * this situation, we must victimize our own cache, the MFU, for this insertion. 2433 */ 2434static void 2435arc_get_data_buf(arc_buf_t *buf) 2436{ 2437 arc_state_t *state = buf->b_hdr->b_state; 2438 uint64_t size = buf->b_hdr->b_size; 2439 arc_buf_contents_t type = buf->b_hdr->b_type; 2440 2441 arc_adapt(size, state); 2442 2443 /* 2444 * We have not yet reached cache maximum size, 2445 * just allocate a new buffer. 2446 */ 2447 if (!arc_evict_needed(type)) { 2448 if (type == ARC_BUFC_METADATA) { 2449 buf->b_data = zio_buf_alloc(size); 2450 arc_space_consume(size, ARC_SPACE_DATA); 2451 } else { 2452 ASSERT(type == ARC_BUFC_DATA); 2453 buf->b_data = zio_data_buf_alloc(size); 2454 ARCSTAT_INCR(arcstat_data_size, size); 2455 atomic_add_64(&arc_size, size); 2456 } 2457 goto out; 2458 } 2459 2460 /* 2461 * If we are prefetching from the mfu ghost list, this buffer 2462 * will end up on the mru list; so steal space from there. 2463 */ 2464 if (state == arc_mfu_ghost) 2465 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2466 else if (state == arc_mru_ghost) 2467 state = arc_mru; 2468 2469 if (state == arc_mru || state == arc_anon) { 2470 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2471 state = (arc_mfu->arcs_lsize[type] >= size && 2472 arc_p > mru_used) ? arc_mfu : arc_mru; 2473 } else { 2474 /* MFU cases */ 2475 uint64_t mfu_space = arc_c - arc_p; 2476 state = (arc_mru->arcs_lsize[type] >= size && 2477 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2478 } 2479 if ((buf->b_data = arc_evict(state, NULL, size, TRUE, type)) == NULL) { 2480 if (type == ARC_BUFC_METADATA) { 2481 buf->b_data = zio_buf_alloc(size); 2482 arc_space_consume(size, ARC_SPACE_DATA); 2483 } else { 2484 ASSERT(type == ARC_BUFC_DATA); 2485 buf->b_data = zio_data_buf_alloc(size); 2486 ARCSTAT_INCR(arcstat_data_size, size); 2487 atomic_add_64(&arc_size, size); 2488 } 2489 ARCSTAT_BUMP(arcstat_recycle_miss); 2490 } 2491 ASSERT(buf->b_data != NULL); 2492out: 2493 /* 2494 * Update the state size. Note that ghost states have a 2495 * "ghost size" and so don't need to be updated. 2496 */ 2497 if (!GHOST_STATE(buf->b_hdr->b_state)) { 2498 arc_buf_hdr_t *hdr = buf->b_hdr; 2499 2500 atomic_add_64(&hdr->b_state->arcs_size, size); 2501 if (list_link_active(&hdr->b_arc_node)) { 2502 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2503 atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2504 } 2505 /* 2506 * If we are growing the cache, and we are adding anonymous 2507 * data, and we have outgrown arc_p, update arc_p 2508 */ 2509 if (arc_size < arc_c && hdr->b_state == arc_anon && 2510 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2511 arc_p = MIN(arc_c, arc_p + size); 2512 } 2513 ARCSTAT_BUMP(arcstat_allocated); 2514} 2515 2516/* 2517 * This routine is called whenever a buffer is accessed. 2518 * NOTE: the hash lock is dropped in this function. 2519 */ 2520static void 2521arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2522{ 2523 ASSERT(MUTEX_HELD(hash_lock)); 2524 2525 if (buf->b_state == arc_anon) { 2526 /* 2527 * This buffer is not in the cache, and does not 2528 * appear in our "ghost" list. Add the new buffer 2529 * to the MRU state. 2530 */ 2531 2532 ASSERT(buf->b_arc_access == 0); 2533 buf->b_arc_access = LBOLT; 2534 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2535 arc_change_state(arc_mru, buf, hash_lock); 2536 2537 } else if (buf->b_state == arc_mru) { 2538 /* 2539 * If this buffer is here because of a prefetch, then either: 2540 * - clear the flag if this is a "referencing" read 2541 * (any subsequent access will bump this into the MFU state). 2542 * or 2543 * - move the buffer to the head of the list if this is 2544 * another prefetch (to make it less likely to be evicted). 2545 */ 2546 if ((buf->b_flags & ARC_PREFETCH) != 0) { 2547 if (refcount_count(&buf->b_refcnt) == 0) { 2548 ASSERT(list_link_active(&buf->b_arc_node)); 2549 } else { 2550 buf->b_flags &= ~ARC_PREFETCH; 2551 ARCSTAT_BUMP(arcstat_mru_hits); 2552 } 2553 buf->b_arc_access = LBOLT; 2554 return; 2555 } 2556 2557 /* 2558 * This buffer has been "accessed" only once so far, 2559 * but it is still in the cache. Move it to the MFU 2560 * state. 2561 */ 2562 if (LBOLT > buf->b_arc_access + ARC_MINTIME) { 2563 /* 2564 * More than 125ms have passed since we 2565 * instantiated this buffer. Move it to the 2566 * most frequently used state. 2567 */ 2568 buf->b_arc_access = LBOLT; 2569 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2570 arc_change_state(arc_mfu, buf, hash_lock); 2571 } 2572 ARCSTAT_BUMP(arcstat_mru_hits); 2573 } else if (buf->b_state == arc_mru_ghost) { 2574 arc_state_t *new_state; 2575 /* 2576 * This buffer has been "accessed" recently, but 2577 * was evicted from the cache. Move it to the 2578 * MFU state. 2579 */ 2580 2581 if (buf->b_flags & ARC_PREFETCH) { 2582 new_state = arc_mru; 2583 if (refcount_count(&buf->b_refcnt) > 0) 2584 buf->b_flags &= ~ARC_PREFETCH; 2585 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2586 } else { 2587 new_state = arc_mfu; 2588 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2589 } 2590 2591 buf->b_arc_access = LBOLT; 2592 arc_change_state(new_state, buf, hash_lock); 2593 2594 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 2595 } else if (buf->b_state == arc_mfu) { 2596 /* 2597 * This buffer has been accessed more than once and is 2598 * still in the cache. Keep it in the MFU state. 2599 * 2600 * NOTE: an add_reference() that occurred when we did 2601 * the arc_read() will have kicked this off the list. 2602 * If it was a prefetch, we will explicitly move it to 2603 * the head of the list now. 2604 */ 2605 if ((buf->b_flags & ARC_PREFETCH) != 0) { 2606 ASSERT(refcount_count(&buf->b_refcnt) == 0); 2607 ASSERT(list_link_active(&buf->b_arc_node)); 2608 } 2609 ARCSTAT_BUMP(arcstat_mfu_hits); 2610 buf->b_arc_access = LBOLT; 2611 } else if (buf->b_state == arc_mfu_ghost) { 2612 arc_state_t *new_state = arc_mfu; 2613 /* 2614 * This buffer has been accessed more than once but has 2615 * been evicted from the cache. Move it back to the 2616 * MFU state. 2617 */ 2618 2619 if (buf->b_flags & ARC_PREFETCH) { 2620 /* 2621 * This is a prefetch access... 2622 * move this block back to the MRU state. 2623 */ 2624 ASSERT3U(refcount_count(&buf->b_refcnt), ==, 0); 2625 new_state = arc_mru; 2626 } 2627 2628 buf->b_arc_access = LBOLT; 2629 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2630 arc_change_state(new_state, buf, hash_lock); 2631 2632 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 2633 } else if (buf->b_state == arc_l2c_only) { 2634 /* 2635 * This buffer is on the 2nd Level ARC. 2636 */ 2637 2638 buf->b_arc_access = LBOLT; 2639 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2640 arc_change_state(arc_mfu, buf, hash_lock); 2641 } else { 2642 ASSERT(!"invalid arc state"); 2643 } 2644} 2645 2646/* a generic arc_done_func_t which you can use */ 2647/* ARGSUSED */ 2648void 2649arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 2650{ 2651 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 2652 VERIFY(arc_buf_remove_ref(buf, arg) == 1); 2653} 2654 2655/* a generic arc_done_func_t */ 2656void 2657arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 2658{ 2659 arc_buf_t **bufp = arg; 2660 if (zio && zio->io_error) { 2661 VERIFY(arc_buf_remove_ref(buf, arg) == 1); 2662 *bufp = NULL; 2663 } else { 2664 *bufp = buf; 2665 } 2666} 2667 2668static void 2669arc_read_done(zio_t *zio) 2670{ 2671 arc_buf_hdr_t *hdr, *found; 2672 arc_buf_t *buf; 2673 arc_buf_t *abuf; /* buffer we're assigning to callback */ 2674 kmutex_t *hash_lock; 2675 arc_callback_t *callback_list, *acb; 2676 int freeable = FALSE; 2677 2678 buf = zio->io_private; 2679 hdr = buf->b_hdr; 2680 2681 /* 2682 * The hdr was inserted into hash-table and removed from lists 2683 * prior to starting I/O. We should find this header, since 2684 * it's in the hash table, and it should be legit since it's 2685 * not possible to evict it during the I/O. The only possible 2686 * reason for it not to be found is if we were freed during the 2687 * read. 2688 */ 2689 found = buf_hash_find(zio->io_spa, &hdr->b_dva, hdr->b_birth, 2690 &hash_lock); 2691 2692 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) || 2693 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 2694 (found == hdr && HDR_L2_READING(hdr))); 2695 2696 hdr->b_flags &= ~ARC_L2_EVICTED; 2697 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 2698 hdr->b_flags &= ~ARC_L2CACHE; 2699 2700 /* byteswap if necessary */ 2701 callback_list = hdr->b_acb; 2702 ASSERT(callback_list != NULL); 2703 if (BP_SHOULD_BYTESWAP(zio->io_bp)) { 2704 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 2705 byteswap_uint64_array : 2706 dmu_ot[BP_GET_TYPE(zio->io_bp)].ot_byteswap; 2707 func(buf->b_data, hdr->b_size); 2708 } 2709 2710 arc_cksum_compute(buf, B_FALSE); 2711 2712 /* create copies of the data buffer for the callers */ 2713 abuf = buf; 2714 for (acb = callback_list; acb; acb = acb->acb_next) { 2715 if (acb->acb_done) { 2716 if (abuf == NULL) 2717 abuf = arc_buf_clone(buf); 2718 acb->acb_buf = abuf; 2719 abuf = NULL; 2720 } 2721 } 2722 hdr->b_acb = NULL; 2723 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 2724 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 2725 if (abuf == buf) 2726 hdr->b_flags |= ARC_BUF_AVAILABLE; 2727 2728 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 2729 2730 if (zio->io_error != 0) { 2731 hdr->b_flags |= ARC_IO_ERROR; 2732 if (hdr->b_state != arc_anon) 2733 arc_change_state(arc_anon, hdr, hash_lock); 2734 if (HDR_IN_HASH_TABLE(hdr)) 2735 buf_hash_remove(hdr); 2736 freeable = refcount_is_zero(&hdr->b_refcnt); 2737 } 2738 2739 /* 2740 * Broadcast before we drop the hash_lock to avoid the possibility 2741 * that the hdr (and hence the cv) might be freed before we get to 2742 * the cv_broadcast(). 2743 */ 2744 cv_broadcast(&hdr->b_cv); 2745 2746 if (hash_lock) { 2747 /* 2748 * Only call arc_access on anonymous buffers. This is because 2749 * if we've issued an I/O for an evicted buffer, we've already 2750 * called arc_access (to prevent any simultaneous readers from 2751 * getting confused). 2752 */ 2753 if (zio->io_error == 0 && hdr->b_state == arc_anon) 2754 arc_access(hdr, hash_lock); 2755 mutex_exit(hash_lock); 2756 } else { 2757 /* 2758 * This block was freed while we waited for the read to 2759 * complete. It has been removed from the hash table and 2760 * moved to the anonymous state (so that it won't show up 2761 * in the cache). 2762 */ 2763 ASSERT3P(hdr->b_state, ==, arc_anon); 2764 freeable = refcount_is_zero(&hdr->b_refcnt); 2765 } 2766 2767 /* execute each callback and free its structure */ 2768 while ((acb = callback_list) != NULL) { 2769 if (acb->acb_done) 2770 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 2771 2772 if (acb->acb_zio_dummy != NULL) { 2773 acb->acb_zio_dummy->io_error = zio->io_error; 2774 zio_nowait(acb->acb_zio_dummy); 2775 } 2776 2777 callback_list = acb->acb_next; 2778 kmem_free(acb, sizeof (arc_callback_t)); 2779 } 2780 2781 if (freeable) 2782 arc_hdr_destroy(hdr); 2783} 2784 2785/* 2786 * "Read" the block block at the specified DVA (in bp) via the 2787 * cache. If the block is found in the cache, invoke the provided 2788 * callback immediately and return. Note that the `zio' parameter 2789 * in the callback will be NULL in this case, since no IO was 2790 * required. If the block is not in the cache pass the read request 2791 * on to the spa with a substitute callback function, so that the 2792 * requested block will be added to the cache. 2793 * 2794 * If a read request arrives for a block that has a read in-progress, 2795 * either wait for the in-progress read to complete (and return the 2796 * results); or, if this is a read with a "done" func, add a record 2797 * to the read to invoke the "done" func when the read completes, 2798 * and return; or just return. 2799 * 2800 * arc_read_done() will invoke all the requested "done" functions 2801 * for readers of this block. 2802 * 2803 * Normal callers should use arc_read and pass the arc buffer and offset 2804 * for the bp. But if you know you don't need locking, you can use 2805 * arc_read_bp. 2806 */ 2807int 2808arc_read(zio_t *pio, spa_t *spa, blkptr_t *bp, arc_buf_t *pbuf, 2809 arc_done_func_t *done, void *private, int priority, int zio_flags, 2810 uint32_t *arc_flags, const zbookmark_t *zb) 2811{ 2812 int err; 2813 2814 ASSERT(!refcount_is_zero(&pbuf->b_hdr->b_refcnt)); 2815 ASSERT3U((char *)bp - (char *)pbuf->b_data, <, pbuf->b_hdr->b_size); 2816 rw_enter(&pbuf->b_lock, RW_READER); 2817 2818 err = arc_read_nolock(pio, spa, bp, done, private, priority, 2819 zio_flags, arc_flags, zb); 2820 rw_exit(&pbuf->b_lock); 2821 return (err); 2822} 2823 2824int 2825arc_read_nolock(zio_t *pio, spa_t *spa, blkptr_t *bp, 2826 arc_done_func_t *done, void *private, int priority, int zio_flags, 2827 uint32_t *arc_flags, const zbookmark_t *zb) 2828{ 2829 arc_buf_hdr_t *hdr; 2830 arc_buf_t *buf; 2831 kmutex_t *hash_lock; 2832 zio_t *rzio; 2833 2834top: 2835 hdr = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_lock); 2836 if (hdr && hdr->b_datacnt > 0) { 2837 2838 *arc_flags |= ARC_CACHED; 2839 2840 if (HDR_IO_IN_PROGRESS(hdr)) { 2841 2842 if (*arc_flags & ARC_WAIT) { 2843 cv_wait(&hdr->b_cv, hash_lock); 2844 mutex_exit(hash_lock); 2845 goto top; 2846 } 2847 ASSERT(*arc_flags & ARC_NOWAIT); 2848 2849 if (done) { 2850 arc_callback_t *acb = NULL; 2851 2852 acb = kmem_zalloc(sizeof (arc_callback_t), 2853 KM_SLEEP); 2854 acb->acb_done = done; 2855 acb->acb_private = private; 2856 if (pio != NULL) 2857 acb->acb_zio_dummy = zio_null(pio, 2858 spa, NULL, NULL, zio_flags); 2859 2860 ASSERT(acb->acb_done != NULL); 2861 acb->acb_next = hdr->b_acb; 2862 hdr->b_acb = acb; 2863 add_reference(hdr, hash_lock, private); 2864 mutex_exit(hash_lock); 2865 return (0); 2866 } 2867 mutex_exit(hash_lock); 2868 return (0); 2869 } 2870 2871 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 2872 2873 if (done) { 2874 add_reference(hdr, hash_lock, private); 2875 /* 2876 * If this block is already in use, create a new 2877 * copy of the data so that we will be guaranteed 2878 * that arc_release() will always succeed. 2879 */ 2880 buf = hdr->b_buf; 2881 ASSERT(buf); 2882 ASSERT(buf->b_data); 2883 if (HDR_BUF_AVAILABLE(hdr)) { 2884 ASSERT(buf->b_efunc == NULL); 2885 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 2886 } else { 2887 buf = arc_buf_clone(buf); 2888 } 2889 } else if (*arc_flags & ARC_PREFETCH && 2890 refcount_count(&hdr->b_refcnt) == 0) { 2891 hdr->b_flags |= ARC_PREFETCH; 2892 } 2893 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 2894 arc_access(hdr, hash_lock); 2895 if (*arc_flags & ARC_L2CACHE) 2896 hdr->b_flags |= ARC_L2CACHE; 2897 mutex_exit(hash_lock); 2898 ARCSTAT_BUMP(arcstat_hits); 2899 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 2900 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 2901 data, metadata, hits); 2902 2903 if (done) 2904 done(NULL, buf, private); 2905 } else { 2906 uint64_t size = BP_GET_LSIZE(bp); 2907 arc_callback_t *acb; 2908 vdev_t *vd = NULL; 2909 uint64_t addr; 2910 boolean_t devw = B_FALSE; 2911 2912 if (hdr == NULL) { 2913 /* this block is not in the cache */ 2914 arc_buf_hdr_t *exists; 2915 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 2916 buf = arc_buf_alloc(spa, size, private, type); 2917 hdr = buf->b_hdr; 2918 hdr->b_dva = *BP_IDENTITY(bp); 2919 hdr->b_birth = bp->blk_birth; 2920 hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 2921 exists = buf_hash_insert(hdr, &hash_lock); 2922 if (exists) { 2923 /* somebody beat us to the hash insert */ 2924 mutex_exit(hash_lock); 2925 bzero(&hdr->b_dva, sizeof (dva_t)); 2926 hdr->b_birth = 0; 2927 hdr->b_cksum0 = 0; 2928 (void) arc_buf_remove_ref(buf, private); 2929 goto top; /* restart the IO request */ 2930 } 2931 /* if this is a prefetch, we don't have a reference */ 2932 if (*arc_flags & ARC_PREFETCH) { 2933 (void) remove_reference(hdr, hash_lock, 2934 private); 2935 hdr->b_flags |= ARC_PREFETCH; 2936 } 2937 if (*arc_flags & ARC_L2CACHE) 2938 hdr->b_flags |= ARC_L2CACHE; 2939 if (BP_GET_LEVEL(bp) > 0) 2940 hdr->b_flags |= ARC_INDIRECT; 2941 } else { 2942 /* this block is in the ghost cache */ 2943 ASSERT(GHOST_STATE(hdr->b_state)); 2944 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2945 ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 0); 2946 ASSERT(hdr->b_buf == NULL); 2947 2948 /* if this is a prefetch, we don't have a reference */ 2949 if (*arc_flags & ARC_PREFETCH) 2950 hdr->b_flags |= ARC_PREFETCH; 2951 else 2952 add_reference(hdr, hash_lock, private); 2953 if (*arc_flags & ARC_L2CACHE) 2954 hdr->b_flags |= ARC_L2CACHE; 2955 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2956 buf->b_hdr = hdr; 2957 buf->b_data = NULL; 2958 buf->b_efunc = NULL; 2959 buf->b_private = NULL; 2960 buf->b_next = NULL; 2961 hdr->b_buf = buf; 2962 arc_get_data_buf(buf); 2963 ASSERT(hdr->b_datacnt == 0); 2964 hdr->b_datacnt = 1; 2965 2966 } 2967 2968 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 2969 acb->acb_done = done; 2970 acb->acb_private = private; 2971 2972 ASSERT(hdr->b_acb == NULL); 2973 hdr->b_acb = acb; 2974 hdr->b_flags |= ARC_IO_IN_PROGRESS; 2975 2976 /* 2977 * If the buffer has been evicted, migrate it to a present state 2978 * before issuing the I/O. Once we drop the hash-table lock, 2979 * the header will be marked as I/O in progress and have an 2980 * attached buffer. At this point, anybody who finds this 2981 * buffer ought to notice that it's legit but has a pending I/O. 2982 */ 2983 2984 if (GHOST_STATE(hdr->b_state)) 2985 arc_access(hdr, hash_lock); 2986 2987 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL && 2988 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 2989 devw = hdr->b_l2hdr->b_dev->l2ad_writing; 2990 addr = hdr->b_l2hdr->b_daddr; 2991 /* 2992 * Lock out device removal. 2993 */ 2994 if (vdev_is_dead(vd) || 2995 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 2996 vd = NULL; 2997 } 2998 2999 mutex_exit(hash_lock); 3000 3001 ASSERT3U(hdr->b_size, ==, size); 3002 DTRACE_PROBE3(arc__miss, blkptr_t *, bp, uint64_t, size, 3003 zbookmark_t *, zb); 3004 ARCSTAT_BUMP(arcstat_misses); 3005 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3006 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3007 data, metadata, misses); 3008 3009 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3010 /* 3011 * Read from the L2ARC if the following are true: 3012 * 1. The L2ARC vdev was previously cached. 3013 * 2. This buffer still has L2ARC metadata. 3014 * 3. This buffer isn't currently writing to the L2ARC. 3015 * 4. The L2ARC entry wasn't evicted, which may 3016 * also have invalidated the vdev. 3017 * 5. This isn't prefetch and l2arc_noprefetch is set. 3018 */ 3019 if (hdr->b_l2hdr != NULL && 3020 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3021 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3022 l2arc_read_callback_t *cb; 3023 3024 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3025 ARCSTAT_BUMP(arcstat_l2_hits); 3026 3027 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3028 KM_SLEEP); 3029 cb->l2rcb_buf = buf; 3030 cb->l2rcb_spa = spa; 3031 cb->l2rcb_bp = *bp; 3032 cb->l2rcb_zb = *zb; 3033 cb->l2rcb_flags = zio_flags; 3034 3035 /* 3036 * l2arc read. The SCL_L2ARC lock will be 3037 * released by l2arc_read_done(). 3038 */ 3039 rzio = zio_read_phys(pio, vd, addr, size, 3040 buf->b_data, ZIO_CHECKSUM_OFF, 3041 l2arc_read_done, cb, priority, zio_flags | 3042 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL | 3043 ZIO_FLAG_DONT_PROPAGATE | 3044 ZIO_FLAG_DONT_RETRY, B_FALSE); 3045 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3046 zio_t *, rzio); 3047 ARCSTAT_INCR(arcstat_l2_read_bytes, size); 3048 3049 if (*arc_flags & ARC_NOWAIT) { 3050 zio_nowait(rzio); 3051 return (0); 3052 } 3053 3054 ASSERT(*arc_flags & ARC_WAIT); 3055 if (zio_wait(rzio) == 0) 3056 return (0); 3057 3058 /* l2arc read error; goto zio_read() */ 3059 } else { 3060 DTRACE_PROBE1(l2arc__miss, 3061 arc_buf_hdr_t *, hdr); 3062 ARCSTAT_BUMP(arcstat_l2_misses); 3063 if (HDR_L2_WRITING(hdr)) 3064 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3065 spa_config_exit(spa, SCL_L2ARC, vd); 3066 } 3067 } else { 3068 if (vd != NULL) 3069 spa_config_exit(spa, SCL_L2ARC, vd); 3070 if (l2arc_ndev != 0) { 3071 DTRACE_PROBE1(l2arc__miss, 3072 arc_buf_hdr_t *, hdr); 3073 ARCSTAT_BUMP(arcstat_l2_misses); 3074 } 3075 } 3076 3077 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3078 arc_read_done, buf, priority, zio_flags, zb); 3079 3080 if (*arc_flags & ARC_WAIT) 3081 return (zio_wait(rzio)); 3082 3083 ASSERT(*arc_flags & ARC_NOWAIT); 3084 zio_nowait(rzio); 3085 } 3086 return (0); 3087} 3088 3089/* 3090 * arc_read() variant to support pool traversal. If the block is already 3091 * in the ARC, make a copy of it; otherwise, the caller will do the I/O. 3092 * The idea is that we don't want pool traversal filling up memory, but 3093 * if the ARC already has the data anyway, we shouldn't pay for the I/O. 3094 */ 3095int 3096arc_tryread(spa_t *spa, blkptr_t *bp, void *data) 3097{ 3098 arc_buf_hdr_t *hdr; 3099 kmutex_t *hash_mtx; 3100 int rc = 0; 3101 3102 hdr = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_mtx); 3103 3104 if (hdr && hdr->b_datacnt > 0 && !HDR_IO_IN_PROGRESS(hdr)) { 3105 arc_buf_t *buf = hdr->b_buf; 3106 3107 ASSERT(buf); 3108 while (buf->b_data == NULL) { 3109 buf = buf->b_next; 3110 ASSERT(buf); 3111 } 3112 bcopy(buf->b_data, data, hdr->b_size); 3113 } else { 3114 rc = ENOENT; 3115 } 3116 3117 if (hash_mtx) 3118 mutex_exit(hash_mtx); 3119 3120 return (rc); 3121} 3122 3123void 3124arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3125{ 3126 ASSERT(buf->b_hdr != NULL); 3127 ASSERT(buf->b_hdr->b_state != arc_anon); 3128 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3129 buf->b_efunc = func; 3130 buf->b_private = private; 3131} 3132 3133/* 3134 * This is used by the DMU to let the ARC know that a buffer is 3135 * being evicted, so the ARC should clean up. If this arc buf 3136 * is not yet in the evicted state, it will be put there. 3137 */ 3138int 3139arc_buf_evict(arc_buf_t *buf) 3140{ 3141 arc_buf_hdr_t *hdr; 3142 kmutex_t *hash_lock; 3143 arc_buf_t **bufp; 3144 list_t *list, *evicted_list; 3145 kmutex_t *lock, *evicted_lock; 3146 3147 rw_enter(&buf->b_lock, RW_WRITER); 3148 hdr = buf->b_hdr; 3149 if (hdr == NULL) { 3150 /* 3151 * We are in arc_do_user_evicts(). 3152 */ 3153 ASSERT(buf->b_data == NULL); 3154 rw_exit(&buf->b_lock); 3155 return (0); 3156 } else if (buf->b_data == NULL) { 3157 arc_buf_t copy = *buf; /* structure assignment */ 3158 /* 3159 * We are on the eviction list; process this buffer now 3160 * but let arc_do_user_evicts() do the reaping. 3161 */ 3162 buf->b_efunc = NULL; 3163 rw_exit(&buf->b_lock); 3164 VERIFY(copy.b_efunc(©) == 0); 3165 return (1); 3166 } 3167 hash_lock = HDR_LOCK(hdr); 3168 mutex_enter(hash_lock); 3169 3170 ASSERT(buf->b_hdr == hdr); 3171 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3172 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3173 3174 /* 3175 * Pull this buffer off of the hdr 3176 */ 3177 bufp = &hdr->b_buf; 3178 while (*bufp != buf) 3179 bufp = &(*bufp)->b_next; 3180 *bufp = buf->b_next; 3181 3182 ASSERT(buf->b_data != NULL); 3183 arc_buf_destroy(buf, FALSE, FALSE); 3184 3185 if (hdr->b_datacnt == 0) { 3186 arc_state_t *old_state = hdr->b_state; 3187 arc_state_t *evicted_state; 3188 3189 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 3190 3191 evicted_state = 3192 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3193 3194 get_buf_info(hdr, old_state, &list, &lock); 3195 get_buf_info(hdr, evicted_state, &evicted_list, &evicted_lock); 3196 mutex_enter(lock); 3197 mutex_enter(evicted_lock); 3198 3199 arc_change_state(evicted_state, hdr, hash_lock); 3200 ASSERT(HDR_IN_HASH_TABLE(hdr)); 3201 hdr->b_flags |= ARC_IN_HASH_TABLE; 3202 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3203 3204 mutex_exit(evicted_lock); 3205 mutex_exit(lock); 3206 } 3207 mutex_exit(hash_lock); 3208 rw_exit(&buf->b_lock); 3209 3210 VERIFY(buf->b_efunc(buf) == 0); 3211 buf->b_efunc = NULL; 3212 buf->b_private = NULL; 3213 buf->b_hdr = NULL; 3214 kmem_cache_free(buf_cache, buf); 3215 return (1); 3216} 3217 3218/* 3219 * Release this buffer from the cache. This must be done 3220 * after a read and prior to modifying the buffer contents. 3221 * If the buffer has more than one reference, we must make 3222 * a new hdr for the buffer. 3223 */ 3224void 3225arc_release(arc_buf_t *buf, void *tag) 3226{ 3227 arc_buf_hdr_t *hdr; 3228 kmutex_t *hash_lock; 3229 l2arc_buf_hdr_t *l2hdr; 3230 uint64_t buf_size; 3231 boolean_t released = B_FALSE; 3232 3233 rw_enter(&buf->b_lock, RW_WRITER); 3234 hdr = buf->b_hdr; 3235 3236 /* this buffer is not on any list */ 3237 ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3238 ASSERT(!(hdr->b_flags & ARC_STORED)); 3239 3240 if (hdr->b_state == arc_anon) { 3241 /* this buffer is already released */ 3242 ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 1); 3243 ASSERT(BUF_EMPTY(hdr)); 3244 ASSERT(buf->b_efunc == NULL); 3245 arc_buf_thaw(buf); 3246 rw_exit(&buf->b_lock); 3247 released = B_TRUE; 3248 } else { 3249 hash_lock = HDR_LOCK(hdr); 3250 mutex_enter(hash_lock); 3251 } 3252 3253 l2hdr = hdr->b_l2hdr; 3254 if (l2hdr) { 3255 mutex_enter(&l2arc_buflist_mtx); 3256 hdr->b_l2hdr = NULL; 3257 buf_size = hdr->b_size; 3258 } 3259 3260 if (released) 3261 goto out; 3262 3263 /* 3264 * Do we have more than one buf? 3265 */ 3266 if (hdr->b_datacnt > 1) { 3267 arc_buf_hdr_t *nhdr; 3268 arc_buf_t **bufp; 3269 uint64_t blksz = hdr->b_size; 3270 spa_t *spa = hdr->b_spa; 3271 arc_buf_contents_t type = hdr->b_type; 3272 uint32_t flags = hdr->b_flags; 3273 3274 ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3275 /* 3276 * Pull the data off of this buf and attach it to 3277 * a new anonymous buf. 3278 */ 3279 (void) remove_reference(hdr, hash_lock, tag); 3280 bufp = &hdr->b_buf; 3281 while (*bufp != buf) 3282 bufp = &(*bufp)->b_next; 3283 *bufp = (*bufp)->b_next; 3284 buf->b_next = NULL; 3285 3286 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3287 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3288 if (refcount_is_zero(&hdr->b_refcnt)) { 3289 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3290 ASSERT3U(*size, >=, hdr->b_size); 3291 atomic_add_64(size, -hdr->b_size); 3292 } 3293 hdr->b_datacnt -= 1; 3294 arc_cksum_verify(buf); 3295 3296 mutex_exit(hash_lock); 3297 3298 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3299 nhdr->b_size = blksz; 3300 nhdr->b_spa = spa; 3301 nhdr->b_type = type; 3302 nhdr->b_buf = buf; 3303 nhdr->b_state = arc_anon; 3304 nhdr->b_arc_access = 0; 3305 nhdr->b_flags = flags & ARC_L2_WRITING; 3306 nhdr->b_l2hdr = NULL; 3307 nhdr->b_datacnt = 1; 3308 nhdr->b_freeze_cksum = NULL; 3309 (void) refcount_add(&nhdr->b_refcnt, tag); 3310 buf->b_hdr = nhdr; 3311 rw_exit(&buf->b_lock); 3312 atomic_add_64(&arc_anon->arcs_size, blksz); 3313 } else { 3314 rw_exit(&buf->b_lock); 3315 ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3316 ASSERT(!list_link_active(&hdr->b_arc_node)); 3317 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3318 arc_change_state(arc_anon, hdr, hash_lock); 3319 hdr->b_arc_access = 0; 3320 mutex_exit(hash_lock); 3321 3322 bzero(&hdr->b_dva, sizeof (dva_t)); 3323 hdr->b_birth = 0; 3324 hdr->b_cksum0 = 0; 3325 arc_buf_thaw(buf); 3326 } 3327 buf->b_efunc = NULL; 3328 buf->b_private = NULL; 3329 3330out: 3331 if (l2hdr) { 3332 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3333 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3334 ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3335 mutex_exit(&l2arc_buflist_mtx); 3336 } 3337} 3338 3339int 3340arc_released(arc_buf_t *buf) 3341{ 3342 int released; 3343 3344 rw_enter(&buf->b_lock, RW_READER); 3345 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3346 rw_exit(&buf->b_lock); 3347 return (released); 3348} 3349 3350int 3351arc_has_callback(arc_buf_t *buf) 3352{ 3353 int callback; 3354 3355 rw_enter(&buf->b_lock, RW_READER); 3356 callback = (buf->b_efunc != NULL); 3357 rw_exit(&buf->b_lock); 3358 return (callback); 3359} 3360 3361#ifdef ZFS_DEBUG 3362int 3363arc_referenced(arc_buf_t *buf) 3364{ 3365 int referenced; 3366 3367 rw_enter(&buf->b_lock, RW_READER); 3368 referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3369 rw_exit(&buf->b_lock); 3370 return (referenced); 3371} 3372#endif 3373 3374static void 3375arc_write_ready(zio_t *zio) 3376{ 3377 arc_write_callback_t *callback = zio->io_private; 3378 arc_buf_t *buf = callback->awcb_buf; 3379 arc_buf_hdr_t *hdr = buf->b_hdr; 3380 3381 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3382 callback->awcb_ready(zio, buf, callback->awcb_private); 3383 3384 /* 3385 * If the IO is already in progress, then this is a re-write 3386 * attempt, so we need to thaw and re-compute the cksum. 3387 * It is the responsibility of the callback to handle the 3388 * accounting for any re-write attempt. 3389 */ 3390 if (HDR_IO_IN_PROGRESS(hdr)) { 3391 mutex_enter(&hdr->b_freeze_lock); 3392 if (hdr->b_freeze_cksum != NULL) { 3393 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3394 hdr->b_freeze_cksum = NULL; 3395 } 3396 mutex_exit(&hdr->b_freeze_lock); 3397 } 3398 arc_cksum_compute(buf, B_FALSE); 3399 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3400} 3401 3402static void 3403arc_write_done(zio_t *zio) 3404{ 3405 arc_write_callback_t *callback = zio->io_private; 3406 arc_buf_t *buf = callback->awcb_buf; 3407 arc_buf_hdr_t *hdr = buf->b_hdr; 3408 3409 hdr->b_acb = NULL; 3410 3411 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3412 hdr->b_birth = zio->io_bp->blk_birth; 3413 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3414 /* 3415 * If the block to be written was all-zero, we may have 3416 * compressed it away. In this case no write was performed 3417 * so there will be no dva/birth-date/checksum. The buffer 3418 * must therefor remain anonymous (and uncached). 3419 */ 3420 if (!BUF_EMPTY(hdr)) { 3421 arc_buf_hdr_t *exists; 3422 kmutex_t *hash_lock; 3423 3424 arc_cksum_verify(buf); 3425 3426 exists = buf_hash_insert(hdr, &hash_lock); 3427 if (exists) { 3428 /* 3429 * This can only happen if we overwrite for 3430 * sync-to-convergence, because we remove 3431 * buffers from the hash table when we arc_free(). 3432 */ 3433 ASSERT(zio->io_flags & ZIO_FLAG_IO_REWRITE); 3434 ASSERT(DVA_EQUAL(BP_IDENTITY(&zio->io_bp_orig), 3435 BP_IDENTITY(zio->io_bp))); 3436 ASSERT3U(zio->io_bp_orig.blk_birth, ==, 3437 zio->io_bp->blk_birth); 3438 3439 ASSERT(refcount_is_zero(&exists->b_refcnt)); 3440 arc_change_state(arc_anon, exists, hash_lock); 3441 mutex_exit(hash_lock); 3442 arc_hdr_destroy(exists); 3443 exists = buf_hash_insert(hdr, &hash_lock); 3444 ASSERT3P(exists, ==, NULL); 3445 } 3446 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3447 /* if it's not anon, we are doing a scrub */ 3448 if (hdr->b_state == arc_anon) 3449 arc_access(hdr, hash_lock); 3450 mutex_exit(hash_lock); 3451 } else if (callback->awcb_done == NULL) { 3452 int destroy_hdr; 3453 /* 3454 * This is an anonymous buffer with no user callback, 3455 * destroy it if there are no active references. 3456 */ 3457 mutex_enter(&arc_eviction_mtx); 3458 destroy_hdr = refcount_is_zero(&hdr->b_refcnt); 3459 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3460 mutex_exit(&arc_eviction_mtx); 3461 if (destroy_hdr) 3462 arc_hdr_destroy(hdr); 3463 } else { 3464 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3465 } 3466 hdr->b_flags &= ~ARC_STORED; 3467 3468 if (callback->awcb_done) { 3469 ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3470 callback->awcb_done(zio, buf, callback->awcb_private); 3471 } 3472 3473 kmem_free(callback, sizeof (arc_write_callback_t)); 3474} 3475 3476static void 3477write_policy(spa_t *spa, const writeprops_t *wp, zio_prop_t *zp) 3478{ 3479 boolean_t ismd = (wp->wp_level > 0 || dmu_ot[wp->wp_type].ot_metadata); 3480 3481 /* Determine checksum setting */ 3482 if (ismd) { 3483 /* 3484 * Metadata always gets checksummed. If the data 3485 * checksum is multi-bit correctable, and it's not a 3486 * ZBT-style checksum, then it's suitable for metadata 3487 * as well. Otherwise, the metadata checksum defaults 3488 * to fletcher4. 3489 */ 3490 if (zio_checksum_table[wp->wp_oschecksum].ci_correctable && 3491 !zio_checksum_table[wp->wp_oschecksum].ci_zbt) 3492 zp->zp_checksum = wp->wp_oschecksum; 3493 else 3494 zp->zp_checksum = ZIO_CHECKSUM_FLETCHER_4; 3495 } else { 3496 zp->zp_checksum = zio_checksum_select(wp->wp_dnchecksum, 3497 wp->wp_oschecksum); 3498 } 3499 3500 /* Determine compression setting */ 3501 if (ismd) { 3502 /* 3503 * XXX -- we should design a compression algorithm 3504 * that specializes in arrays of bps. 3505 */ 3506 zp->zp_compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY : 3507 ZIO_COMPRESS_LZJB; 3508 } else { 3509 zp->zp_compress = zio_compress_select(wp->wp_dncompress, 3510 wp->wp_oscompress); 3511 } 3512 3513 zp->zp_type = wp->wp_type; 3514 zp->zp_level = wp->wp_level; 3515 zp->zp_ndvas = MIN(wp->wp_copies + ismd, spa_max_replication(spa)); 3516} 3517 3518zio_t * 3519arc_write(zio_t *pio, spa_t *spa, const writeprops_t *wp, 3520 boolean_t l2arc, uint64_t txg, blkptr_t *bp, arc_buf_t *buf, 3521 arc_done_func_t *ready, arc_done_func_t *done, void *private, int priority, 3522 int zio_flags, const zbookmark_t *zb) 3523{ 3524 arc_buf_hdr_t *hdr = buf->b_hdr; 3525 arc_write_callback_t *callback; 3526 zio_t *zio; 3527 zio_prop_t zp; 3528 3529 ASSERT(ready != NULL); 3530 ASSERT(!HDR_IO_ERROR(hdr)); 3531 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3532 ASSERT(hdr->b_acb == 0); 3533 if (l2arc) 3534 hdr->b_flags |= ARC_L2CACHE; 3535 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 3536 callback->awcb_ready = ready; 3537 callback->awcb_done = done; 3538 callback->awcb_private = private; 3539 callback->awcb_buf = buf; 3540 3541 write_policy(spa, wp, &zp); 3542 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, &zp, 3543 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb); 3544 3545 return (zio); 3546} 3547 3548int 3549arc_free(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 3550 zio_done_func_t *done, void *private, uint32_t arc_flags) 3551{ 3552 arc_buf_hdr_t *ab; 3553 kmutex_t *hash_lock; 3554 zio_t *zio; 3555 3556 /* 3557 * If this buffer is in the cache, release it, so it 3558 * can be re-used. 3559 */ 3560 ab = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_lock); 3561 if (ab != NULL) { 3562 /* 3563 * The checksum of blocks to free is not always 3564 * preserved (eg. on the deadlist). However, if it is 3565 * nonzero, it should match what we have in the cache. 3566 */ 3567 ASSERT(bp->blk_cksum.zc_word[0] == 0 || 3568 bp->blk_cksum.zc_word[0] == ab->b_cksum0 || 3569 bp->blk_fill == BLK_FILL_ALREADY_FREED); 3570 3571 if (ab->b_state != arc_anon) 3572 arc_change_state(arc_anon, ab, hash_lock); 3573 if (HDR_IO_IN_PROGRESS(ab)) { 3574 /* 3575 * This should only happen when we prefetch. 3576 */ 3577 ASSERT(ab->b_flags & ARC_PREFETCH); 3578 ASSERT3U(ab->b_datacnt, ==, 1); 3579 ab->b_flags |= ARC_FREED_IN_READ; 3580 if (HDR_IN_HASH_TABLE(ab)) 3581 buf_hash_remove(ab); 3582 ab->b_arc_access = 0; 3583 bzero(&ab->b_dva, sizeof (dva_t)); 3584 ab->b_birth = 0; 3585 ab->b_cksum0 = 0; 3586 ab->b_buf->b_efunc = NULL; 3587 ab->b_buf->b_private = NULL; 3588 mutex_exit(hash_lock); 3589 } else if (refcount_is_zero(&ab->b_refcnt)) { 3590 ab->b_flags |= ARC_FREE_IN_PROGRESS; 3591 mutex_exit(hash_lock); 3592 arc_hdr_destroy(ab); 3593 ARCSTAT_BUMP(arcstat_deleted); 3594 } else { 3595 /* 3596 * We still have an active reference on this 3597 * buffer. This can happen, e.g., from 3598 * dbuf_unoverride(). 3599 */ 3600 ASSERT(!HDR_IN_HASH_TABLE(ab)); 3601 ab->b_arc_access = 0; 3602 bzero(&ab->b_dva, sizeof (dva_t)); 3603 ab->b_birth = 0; 3604 ab->b_cksum0 = 0; 3605 ab->b_buf->b_efunc = NULL; 3606 ab->b_buf->b_private = NULL; 3607 mutex_exit(hash_lock); 3608 } 3609 } 3610 3611 zio = zio_free(pio, spa, txg, bp, done, private, ZIO_FLAG_MUSTSUCCEED); 3612 3613 if (arc_flags & ARC_WAIT) 3614 return (zio_wait(zio)); 3615 3616 ASSERT(arc_flags & ARC_NOWAIT); 3617 zio_nowait(zio); 3618 3619 return (0); 3620} 3621 3622static int 3623arc_memory_throttle(uint64_t reserve, uint64_t txg) 3624{ 3625#ifdef _KERNEL 3626 uint64_t inflight_data = arc_anon->arcs_size; 3627 uint64_t available_memory = ptoa((uintmax_t)cnt.v_free_count); 3628 static uint64_t page_load = 0; 3629 static uint64_t last_txg = 0; 3630 3631#if 0 3632#if defined(__i386) 3633 available_memory = 3634 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 3635#endif 3636#endif 3637 if (available_memory >= zfs_write_limit_max) 3638 return (0); 3639 3640 if (txg > last_txg) { 3641 last_txg = txg; 3642 page_load = 0; 3643 } 3644 /* 3645 * If we are in pageout, we know that memory is already tight, 3646 * the arc is already going to be evicting, so we just want to 3647 * continue to let page writes occur as quickly as possible. 3648 */ 3649 if (curproc == pageproc) { 3650 if (page_load > available_memory / 4) 3651 return (ERESTART); 3652 /* Note: reserve is inflated, so we deflate */ 3653 page_load += reserve / 8; 3654 return (0); 3655 } else if (page_load > 0 && arc_reclaim_needed()) { 3656 /* memory is low, delay before restarting */ 3657 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3658 return (EAGAIN); 3659 } 3660 page_load = 0; 3661 3662 if (arc_size > arc_c_min) { 3663 uint64_t evictable_memory = 3664 arc_mru->arcs_lsize[ARC_BUFC_DATA] + 3665 arc_mru->arcs_lsize[ARC_BUFC_METADATA] + 3666 arc_mfu->arcs_lsize[ARC_BUFC_DATA] + 3667 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]; 3668 available_memory += MIN(evictable_memory, arc_size - arc_c_min); 3669 } 3670 3671 if (inflight_data > available_memory / 4) { 3672 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3673 return (ERESTART); 3674 } 3675#endif 3676 return (0); 3677} 3678 3679void 3680arc_tempreserve_clear(uint64_t reserve) 3681{ 3682 atomic_add_64(&arc_tempreserve, -reserve); 3683 ASSERT((int64_t)arc_tempreserve >= 0); 3684} 3685 3686int 3687arc_tempreserve_space(uint64_t reserve, uint64_t txg) 3688{ 3689 int error; 3690 3691#ifdef ZFS_DEBUG 3692 /* 3693 * Once in a while, fail for no reason. Everything should cope. 3694 */ 3695 if (spa_get_random(10000) == 0) { 3696 dprintf("forcing random failure\n"); 3697 return (ERESTART); 3698 } 3699#endif 3700 if (reserve > arc_c/4 && !arc_no_grow) 3701 arc_c = MIN(arc_c_max, reserve * 4); 3702 if (reserve > arc_c) 3703 return (ENOMEM); 3704 3705 /* 3706 * Writes will, almost always, require additional memory allocations 3707 * in order to compress/encrypt/etc the data. We therefor need to 3708 * make sure that there is sufficient available memory for this. 3709 */ 3710 if (error = arc_memory_throttle(reserve, txg)) 3711 return (error); 3712 3713 /* 3714 * Throttle writes when the amount of dirty data in the cache 3715 * gets too large. We try to keep the cache less than half full 3716 * of dirty blocks so that our sync times don't grow too large. 3717 * Note: if two requests come in concurrently, we might let them 3718 * both succeed, when one of them should fail. Not a huge deal. 3719 */ 3720 if (reserve + arc_tempreserve + arc_anon->arcs_size > arc_c / 2 && 3721 arc_anon->arcs_size > arc_c / 4) { 3722 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 3723 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 3724 arc_tempreserve>>10, 3725 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 3726 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 3727 reserve>>10, arc_c>>10); 3728 return (ERESTART); 3729 } 3730 atomic_add_64(&arc_tempreserve, reserve); 3731 return (0); 3732} 3733 3734static kmutex_t arc_lowmem_lock; 3735#ifdef _KERNEL 3736static eventhandler_tag arc_event_lowmem = NULL; 3737 3738static void 3739arc_lowmem(void *arg __unused, int howto __unused) 3740{ 3741 3742 /* Serialize access via arc_lowmem_lock. */ 3743 mutex_enter(&arc_lowmem_lock); 3744 needfree = 1; 3745 cv_signal(&arc_reclaim_thr_cv); 3746 while (needfree) 3747 tsleep(&needfree, 0, "zfs:lowmem", hz / 5); 3748 mutex_exit(&arc_lowmem_lock); 3749} 3750#endif 3751 3752void 3753arc_init(void) 3754{ 3755 int prefetch_tunable_set = 0; 3756 int i; 3757 3758 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 3759 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 3760 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 3761 3762 /* Convert seconds to clock ticks */ 3763 arc_min_prefetch_lifespan = 1 * hz; 3764 3765 /* Start out with 1/8 of all memory */ 3766 arc_c = kmem_size() / 8; 3767#if 0 3768#ifdef _KERNEL 3769 /* 3770 * On architectures where the physical memory can be larger 3771 * than the addressable space (intel in 32-bit mode), we may 3772 * need to limit the cache to 1/8 of VM size. 3773 */ 3774 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 3775#endif 3776#endif 3777 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 3778 arc_c_min = MAX(arc_c / 4, 64<<18); 3779 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 3780 if (arc_c * 8 >= 1<<30) 3781 arc_c_max = (arc_c * 8) - (1<<30); 3782 else 3783 arc_c_max = arc_c_min; 3784 arc_c_max = MAX(arc_c * 5, arc_c_max); 3785#ifdef _KERNEL 3786 /* 3787 * Allow the tunables to override our calculations if they are 3788 * reasonable (ie. over 16MB) 3789 */ 3790 if (zfs_arc_max >= 64<<18 && zfs_arc_max < kmem_size()) 3791 arc_c_max = zfs_arc_max; 3792 if (zfs_arc_min >= 64<<18 && zfs_arc_min <= arc_c_max) 3793 arc_c_min = zfs_arc_min; 3794#endif 3795 arc_c = arc_c_max; 3796 arc_p = (arc_c >> 1); 3797 3798 /* limit meta-data to 1/4 of the arc capacity */ 3799 arc_meta_limit = arc_c_max / 4; 3800 3801 /* Allow the tunable to override if it is reasonable */ 3802 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 3803 arc_meta_limit = zfs_arc_meta_limit; 3804 3805 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 3806 arc_c_min = arc_meta_limit / 2; 3807 3808 if (zfs_arc_grow_retry > 0) 3809 arc_grow_retry = zfs_arc_grow_retry; 3810 3811 if (zfs_arc_shrink_shift > 0) 3812 arc_shrink_shift = zfs_arc_shrink_shift; 3813 3814 if (zfs_arc_p_min_shift > 0) 3815 arc_p_min_shift = zfs_arc_p_min_shift; 3816 3817 /* if kmem_flags are set, lets try to use less memory */ 3818 if (kmem_debugging()) 3819 arc_c = arc_c / 2; 3820 if (arc_c < arc_c_min) 3821 arc_c = arc_c_min; 3822 3823 zfs_arc_min = arc_c_min; 3824 zfs_arc_max = arc_c_max; 3825 3826 arc_anon = &ARC_anon; 3827 arc_mru = &ARC_mru; 3828 arc_mru_ghost = &ARC_mru_ghost; 3829 arc_mfu = &ARC_mfu; 3830 arc_mfu_ghost = &ARC_mfu_ghost; 3831 arc_l2c_only = &ARC_l2c_only; 3832 arc_size = 0; 3833 3834 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 3835 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 3836 NULL, MUTEX_DEFAULT, NULL); 3837 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 3838 NULL, MUTEX_DEFAULT, NULL); 3839 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 3840 NULL, MUTEX_DEFAULT, NULL); 3841 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 3842 NULL, MUTEX_DEFAULT, NULL); 3843 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 3844 NULL, MUTEX_DEFAULT, NULL); 3845 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 3846 NULL, MUTEX_DEFAULT, NULL); 3847 3848 list_create(&arc_mru->arcs_lists[i], 3849 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3850 list_create(&arc_mru_ghost->arcs_lists[i], 3851 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3852 list_create(&arc_mfu->arcs_lists[i], 3853 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3854 list_create(&arc_mfu_ghost->arcs_lists[i], 3855 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3856 list_create(&arc_mfu_ghost->arcs_lists[i], 3857 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3858 list_create(&arc_l2c_only->arcs_lists[i], 3859 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3860 } 3861 3862 buf_init(); 3863 3864 arc_thread_exit = 0; 3865 arc_eviction_list = NULL; 3866 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 3867 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 3868 3869 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 3870 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 3871 3872 if (arc_ksp != NULL) { 3873 arc_ksp->ks_data = &arc_stats; 3874 kstat_install(arc_ksp); 3875 } 3876 3877 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 3878 TS_RUN, minclsyspri); 3879 3880#ifdef _KERNEL 3881 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 3882 EVENTHANDLER_PRI_FIRST); 3883#endif 3884 3885 arc_dead = FALSE; 3886 arc_warm = B_FALSE; 3887 3888 if (zfs_write_limit_max == 0) 3889 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift; 3890 else 3891 zfs_write_limit_shift = 0; 3892 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL); 3893 3894#ifdef _KERNEL 3895 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 3896 prefetch_tunable_set = 1; 3897 3898#ifdef __i386__ 3899 if (prefetch_tunable_set == 0) { 3900 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 3901 "-- to enable,\n"); 3902 printf(" add \"vfs.zfs.prefetch_disable=0\" " 3903 "to /boot/loader.conf.\n"); 3904 zfs_prefetch_disable=1; 3905 } 3906#else 3907 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 3908 prefetch_tunable_set == 0) { 3909 printf("ZFS NOTICE: Prefetch is disabled by default if less " 3910 "than 4GB of RAM is present;\n" 3911 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 3912 "to /boot/loader.conf.\n"); 3913 zfs_prefetch_disable=1; 3914 } 3915#endif 3916 /* Warn about ZFS memory and address space requirements. */ 3917 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 3918 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 3919 "expect unstable behavior.\n"); 3920 } 3921 if (kmem_size() < 512 * (1 << 20)) { 3922 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 3923 "expect unstable behavior.\n"); 3924 printf(" Consider tuning vm.kmem_size and " 3925 "vm.kmem_size_max\n"); 3926 printf(" in /boot/loader.conf.\n"); 3927 } 3928#endif 3929} 3930 3931void 3932arc_fini(void) 3933{ 3934 int i; 3935 3936 mutex_enter(&arc_reclaim_thr_lock); 3937 arc_thread_exit = 1; 3938 cv_signal(&arc_reclaim_thr_cv); 3939 while (arc_thread_exit != 0) 3940 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 3941 mutex_exit(&arc_reclaim_thr_lock); 3942 3943 arc_flush(NULL); 3944 3945 arc_dead = TRUE; 3946 3947 if (arc_ksp != NULL) { 3948 kstat_delete(arc_ksp); 3949 arc_ksp = NULL; 3950 } 3951 3952 mutex_destroy(&arc_eviction_mtx); 3953 mutex_destroy(&arc_reclaim_thr_lock); 3954 cv_destroy(&arc_reclaim_thr_cv); 3955 3956 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 3957 list_destroy(&arc_mru->arcs_lists[i]); 3958 list_destroy(&arc_mru_ghost->arcs_lists[i]); 3959 list_destroy(&arc_mfu->arcs_lists[i]); 3960 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 3961 list_destroy(&arc_l2c_only->arcs_lists[i]); 3962 3963 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 3964 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 3965 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 3966 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 3967 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 3968 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 3969 } 3970 3971 mutex_destroy(&zfs_write_limit_lock); 3972 3973 buf_fini(); 3974 3975 mutex_destroy(&arc_lowmem_lock); 3976#ifdef _KERNEL 3977 if (arc_event_lowmem != NULL) 3978 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 3979#endif 3980} 3981 3982/* 3983 * Level 2 ARC 3984 * 3985 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 3986 * It uses dedicated storage devices to hold cached data, which are populated 3987 * using large infrequent writes. The main role of this cache is to boost 3988 * the performance of random read workloads. The intended L2ARC devices 3989 * include short-stroked disks, solid state disks, and other media with 3990 * substantially faster read latency than disk. 3991 * 3992 * +-----------------------+ 3993 * | ARC | 3994 * +-----------------------+ 3995 * | ^ ^ 3996 * | | | 3997 * l2arc_feed_thread() arc_read() 3998 * | | | 3999 * | l2arc read | 4000 * V | | 4001 * +---------------+ | 4002 * | L2ARC | | 4003 * +---------------+ | 4004 * | ^ | 4005 * l2arc_write() | | 4006 * | | | 4007 * V | | 4008 * +-------+ +-------+ 4009 * | vdev | | vdev | 4010 * | cache | | cache | 4011 * +-------+ +-------+ 4012 * +=========+ .-----. 4013 * : L2ARC : |-_____-| 4014 * : devices : | Disks | 4015 * +=========+ `-_____-' 4016 * 4017 * Read requests are satisfied from the following sources, in order: 4018 * 4019 * 1) ARC 4020 * 2) vdev cache of L2ARC devices 4021 * 3) L2ARC devices 4022 * 4) vdev cache of disks 4023 * 5) disks 4024 * 4025 * Some L2ARC device types exhibit extremely slow write performance. 4026 * To accommodate for this there are some significant differences between 4027 * the L2ARC and traditional cache design: 4028 * 4029 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4030 * the ARC behave as usual, freeing buffers and placing headers on ghost 4031 * lists. The ARC does not send buffers to the L2ARC during eviction as 4032 * this would add inflated write latencies for all ARC memory pressure. 4033 * 4034 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4035 * It does this by periodically scanning buffers from the eviction-end of 4036 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4037 * not already there. It scans until a headroom of buffers is satisfied, 4038 * which itself is a buffer for ARC eviction. The thread that does this is 4039 * l2arc_feed_thread(), illustrated below; example sizes are included to 4040 * provide a better sense of ratio than this diagram: 4041 * 4042 * head --> tail 4043 * +---------------------+----------+ 4044 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4045 * +---------------------+----------+ | o L2ARC eligible 4046 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4047 * +---------------------+----------+ | 4048 * 15.9 Gbytes ^ 32 Mbytes | 4049 * headroom | 4050 * l2arc_feed_thread() 4051 * | 4052 * l2arc write hand <--[oooo]--' 4053 * | 8 Mbyte 4054 * | write max 4055 * V 4056 * +==============================+ 4057 * L2ARC dev |####|#|###|###| |####| ... | 4058 * +==============================+ 4059 * 32 Gbytes 4060 * 4061 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4062 * evicted, then the L2ARC has cached a buffer much sooner than it probably 4063 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4064 * safe to say that this is an uncommon case, since buffers at the end of 4065 * the ARC lists have moved there due to inactivity. 4066 * 4067 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4068 * then the L2ARC simply misses copying some buffers. This serves as a 4069 * pressure valve to prevent heavy read workloads from both stalling the ARC 4070 * with waits and clogging the L2ARC with writes. This also helps prevent 4071 * the potential for the L2ARC to churn if it attempts to cache content too 4072 * quickly, such as during backups of the entire pool. 4073 * 4074 * 5. After system boot and before the ARC has filled main memory, there are 4075 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4076 * lists can remain mostly static. Instead of searching from tail of these 4077 * lists as pictured, the l2arc_feed_thread() will search from the list heads 4078 * for eligible buffers, greatly increasing its chance of finding them. 4079 * 4080 * The L2ARC device write speed is also boosted during this time so that 4081 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4082 * there are no L2ARC reads, and no fear of degrading read performance 4083 * through increased writes. 4084 * 4085 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4086 * the vdev queue can aggregate them into larger and fewer writes. Each 4087 * device is written to in a rotor fashion, sweeping writes through 4088 * available space then repeating. 4089 * 4090 * 7. The L2ARC does not store dirty content. It never needs to flush 4091 * write buffers back to disk based storage. 4092 * 4093 * 8. If an ARC buffer is written (and dirtied) which also exists in the 4094 * L2ARC, the now stale L2ARC buffer is immediately dropped. 4095 * 4096 * The performance of the L2ARC can be tweaked by a number of tunables, which 4097 * may be necessary for different workloads: 4098 * 4099 * l2arc_write_max max write bytes per interval 4100 * l2arc_write_boost extra write bytes during device warmup 4101 * l2arc_noprefetch skip caching prefetched buffers 4102 * l2arc_headroom number of max device writes to precache 4103 * l2arc_feed_secs seconds between L2ARC writing 4104 * 4105 * Tunables may be removed or added as future performance improvements are 4106 * integrated, and also may become zpool properties. 4107 * 4108 * There are three key functions that control how the L2ARC warms up: 4109 * 4110 * l2arc_write_eligible() check if a buffer is eligible to cache 4111 * l2arc_write_size() calculate how much to write 4112 * l2arc_write_interval() calculate sleep delay between writes 4113 * 4114 * These three functions determine what to write, how much, and how quickly 4115 * to send writes. 4116 */ 4117 4118static boolean_t 4119l2arc_write_eligible(spa_t *spa, arc_buf_hdr_t *ab) 4120{ 4121 /* 4122 * A buffer is *not* eligible for the L2ARC if it: 4123 * 1. belongs to a different spa. 4124 * 2. is already cached on the L2ARC. 4125 * 3. has an I/O in progress (it may be an incomplete read). 4126 * 4. is flagged not eligible (zfs property). 4127 */ 4128 if (ab->b_spa != spa) { 4129 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4130 return (B_FALSE); 4131 } 4132 if (ab->b_l2hdr != NULL) { 4133 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4134 return (B_FALSE); 4135 } 4136 if (HDR_IO_IN_PROGRESS(ab)) { 4137 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4138 return (B_FALSE); 4139 } 4140 if (!HDR_L2CACHE(ab)) { 4141 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4142 return (B_FALSE); 4143 } 4144 4145 return (B_TRUE); 4146} 4147 4148static uint64_t 4149l2arc_write_size(l2arc_dev_t *dev) 4150{ 4151 uint64_t size; 4152 4153 size = dev->l2ad_write; 4154 4155 if (arc_warm == B_FALSE) 4156 size += dev->l2ad_boost; 4157 4158 return (size); 4159 4160} 4161 4162static clock_t 4163l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4164{ 4165 clock_t interval, next; 4166 4167 /* 4168 * If the ARC lists are busy, increase our write rate; if the 4169 * lists are stale, idle back. This is achieved by checking 4170 * how much we previously wrote - if it was more than half of 4171 * what we wanted, schedule the next write much sooner. 4172 */ 4173 if (l2arc_feed_again && wrote > (wanted / 2)) 4174 interval = (hz * l2arc_feed_min_ms) / 1000; 4175 else 4176 interval = hz * l2arc_feed_secs; 4177 4178 next = MAX(LBOLT, MIN(LBOLT + interval, began + interval)); 4179 4180 return (next); 4181} 4182 4183static void 4184l2arc_hdr_stat_add(void) 4185{ 4186 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4187 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4188} 4189 4190static void 4191l2arc_hdr_stat_remove(void) 4192{ 4193 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4194 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4195} 4196 4197/* 4198 * Cycle through L2ARC devices. This is how L2ARC load balances. 4199 * If a device is returned, this also returns holding the spa config lock. 4200 */ 4201static l2arc_dev_t * 4202l2arc_dev_get_next(void) 4203{ 4204 l2arc_dev_t *first, *next = NULL; 4205 4206 /* 4207 * Lock out the removal of spas (spa_namespace_lock), then removal 4208 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4209 * both locks will be dropped and a spa config lock held instead. 4210 */ 4211 mutex_enter(&spa_namespace_lock); 4212 mutex_enter(&l2arc_dev_mtx); 4213 4214 /* if there are no vdevs, there is nothing to do */ 4215 if (l2arc_ndev == 0) 4216 goto out; 4217 4218 first = NULL; 4219 next = l2arc_dev_last; 4220 do { 4221 /* loop around the list looking for a non-faulted vdev */ 4222 if (next == NULL) { 4223 next = list_head(l2arc_dev_list); 4224 } else { 4225 next = list_next(l2arc_dev_list, next); 4226 if (next == NULL) 4227 next = list_head(l2arc_dev_list); 4228 } 4229 4230 /* if we have come back to the start, bail out */ 4231 if (first == NULL) 4232 first = next; 4233 else if (next == first) 4234 break; 4235 4236 } while (vdev_is_dead(next->l2ad_vdev)); 4237 4238 /* if we were unable to find any usable vdevs, return NULL */ 4239 if (vdev_is_dead(next->l2ad_vdev)) 4240 next = NULL; 4241 4242 l2arc_dev_last = next; 4243 4244out: 4245 mutex_exit(&l2arc_dev_mtx); 4246 4247 /* 4248 * Grab the config lock to prevent the 'next' device from being 4249 * removed while we are writing to it. 4250 */ 4251 if (next != NULL) 4252 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4253 mutex_exit(&spa_namespace_lock); 4254 4255 return (next); 4256} 4257 4258/* 4259 * Free buffers that were tagged for destruction. 4260 */ 4261static void 4262l2arc_do_free_on_write() 4263{ 4264 list_t *buflist; 4265 l2arc_data_free_t *df, *df_prev; 4266 4267 mutex_enter(&l2arc_free_on_write_mtx); 4268 buflist = l2arc_free_on_write; 4269 4270 for (df = list_tail(buflist); df; df = df_prev) { 4271 df_prev = list_prev(buflist, df); 4272 ASSERT(df->l2df_data != NULL); 4273 ASSERT(df->l2df_func != NULL); 4274 df->l2df_func(df->l2df_data, df->l2df_size); 4275 list_remove(buflist, df); 4276 kmem_free(df, sizeof (l2arc_data_free_t)); 4277 } 4278 4279 mutex_exit(&l2arc_free_on_write_mtx); 4280} 4281 4282/* 4283 * A write to a cache device has completed. Update all headers to allow 4284 * reads from these buffers to begin. 4285 */ 4286static void 4287l2arc_write_done(zio_t *zio) 4288{ 4289 l2arc_write_callback_t *cb; 4290 l2arc_dev_t *dev; 4291 list_t *buflist; 4292 arc_buf_hdr_t *head, *ab, *ab_prev; 4293 l2arc_buf_hdr_t *abl2; 4294 kmutex_t *hash_lock; 4295 4296 cb = zio->io_private; 4297 ASSERT(cb != NULL); 4298 dev = cb->l2wcb_dev; 4299 ASSERT(dev != NULL); 4300 head = cb->l2wcb_head; 4301 ASSERT(head != NULL); 4302 buflist = dev->l2ad_buflist; 4303 ASSERT(buflist != NULL); 4304 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4305 l2arc_write_callback_t *, cb); 4306 4307 if (zio->io_error != 0) 4308 ARCSTAT_BUMP(arcstat_l2_writes_error); 4309 4310 mutex_enter(&l2arc_buflist_mtx); 4311 4312 /* 4313 * All writes completed, or an error was hit. 4314 */ 4315 for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 4316 ab_prev = list_prev(buflist, ab); 4317 4318 hash_lock = HDR_LOCK(ab); 4319 if (!mutex_tryenter(hash_lock)) { 4320 /* 4321 * This buffer misses out. It may be in a stage 4322 * of eviction. Its ARC_L2_WRITING flag will be 4323 * left set, denying reads to this buffer. 4324 */ 4325 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4326 continue; 4327 } 4328 4329 if (zio->io_error != 0) { 4330 /* 4331 * Error - drop L2ARC entry. 4332 */ 4333 list_remove(buflist, ab); 4334 abl2 = ab->b_l2hdr; 4335 ab->b_l2hdr = NULL; 4336 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4337 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4338 } 4339 4340 /* 4341 * Allow ARC to begin reads to this L2ARC entry. 4342 */ 4343 ab->b_flags &= ~ARC_L2_WRITING; 4344 4345 mutex_exit(hash_lock); 4346 } 4347 4348 atomic_inc_64(&l2arc_writes_done); 4349 list_remove(buflist, head); 4350 kmem_cache_free(hdr_cache, head); 4351 mutex_exit(&l2arc_buflist_mtx); 4352 4353 l2arc_do_free_on_write(); 4354 4355 kmem_free(cb, sizeof (l2arc_write_callback_t)); 4356} 4357 4358/* 4359 * A read to a cache device completed. Validate buffer contents before 4360 * handing over to the regular ARC routines. 4361 */ 4362static void 4363l2arc_read_done(zio_t *zio) 4364{ 4365 l2arc_read_callback_t *cb; 4366 arc_buf_hdr_t *hdr; 4367 arc_buf_t *buf; 4368 kmutex_t *hash_lock; 4369 int equal; 4370 4371 ASSERT(zio->io_vd != NULL); 4372 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4373 4374 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4375 4376 cb = zio->io_private; 4377 ASSERT(cb != NULL); 4378 buf = cb->l2rcb_buf; 4379 ASSERT(buf != NULL); 4380 hdr = buf->b_hdr; 4381 ASSERT(hdr != NULL); 4382 4383 hash_lock = HDR_LOCK(hdr); 4384 mutex_enter(hash_lock); 4385 4386 /* 4387 * Check this survived the L2ARC journey. 4388 */ 4389 equal = arc_cksum_equal(buf); 4390 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4391 mutex_exit(hash_lock); 4392 zio->io_private = buf; 4393 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4394 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4395 arc_read_done(zio); 4396 } else { 4397 mutex_exit(hash_lock); 4398 /* 4399 * Buffer didn't survive caching. Increment stats and 4400 * reissue to the original storage device. 4401 */ 4402 if (zio->io_error != 0) { 4403 ARCSTAT_BUMP(arcstat_l2_io_error); 4404 } else { 4405 zio->io_error = EIO; 4406 } 4407 if (!equal) 4408 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4409 4410 /* 4411 * If there's no waiter, issue an async i/o to the primary 4412 * storage now. If there *is* a waiter, the caller must 4413 * issue the i/o in a context where it's OK to block. 4414 */ 4415 if (zio->io_waiter == NULL) 4416 zio_nowait(zio_read(zio->io_parent, 4417 cb->l2rcb_spa, &cb->l2rcb_bp, 4418 buf->b_data, zio->io_size, arc_read_done, buf, 4419 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4420 } 4421 4422 kmem_free(cb, sizeof (l2arc_read_callback_t)); 4423} 4424 4425/* 4426 * This is the list priority from which the L2ARC will search for pages to 4427 * cache. This is used within loops (0..3) to cycle through lists in the 4428 * desired order. This order can have a significant effect on cache 4429 * performance. 4430 * 4431 * Currently the metadata lists are hit first, MFU then MRU, followed by 4432 * the data lists. This function returns a locked list, and also returns 4433 * the lock pointer. 4434 */ 4435static list_t * 4436l2arc_list_locked(int list_num, kmutex_t **lock) 4437{ 4438 list_t *list; 4439 int idx; 4440 4441 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4442 4443 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4444 idx = list_num; 4445 list = &arc_mfu->arcs_lists[idx]; 4446 *lock = ARCS_LOCK(arc_mfu, idx); 4447 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4448 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4449 list = &arc_mru->arcs_lists[idx]; 4450 *lock = ARCS_LOCK(arc_mru, idx); 4451 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4452 ARC_BUFC_NUMDATALISTS)) { 4453 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4454 list = &arc_mfu->arcs_lists[idx]; 4455 *lock = ARCS_LOCK(arc_mfu, idx); 4456 } else { 4457 idx = list_num - ARC_BUFC_NUMLISTS; 4458 list = &arc_mru->arcs_lists[idx]; 4459 *lock = ARCS_LOCK(arc_mru, idx); 4460 } 4461 4462 ASSERT(!(MUTEX_HELD(*lock))); 4463 mutex_enter(*lock); 4464 return (list); 4465} 4466 4467/* 4468 * Evict buffers from the device write hand to the distance specified in 4469 * bytes. This distance may span populated buffers, it may span nothing. 4470 * This is clearing a region on the L2ARC device ready for writing. 4471 * If the 'all' boolean is set, every buffer is evicted. 4472 */ 4473static void 4474l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4475{ 4476 list_t *buflist; 4477 l2arc_buf_hdr_t *abl2; 4478 arc_buf_hdr_t *ab, *ab_prev; 4479 kmutex_t *hash_lock; 4480 uint64_t taddr; 4481 4482 buflist = dev->l2ad_buflist; 4483 4484 if (buflist == NULL) 4485 return; 4486 4487 if (!all && dev->l2ad_first) { 4488 /* 4489 * This is the first sweep through the device. There is 4490 * nothing to evict. 4491 */ 4492 return; 4493 } 4494 4495 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4496 /* 4497 * When nearing the end of the device, evict to the end 4498 * before the device write hand jumps to the start. 4499 */ 4500 taddr = dev->l2ad_end; 4501 } else { 4502 taddr = dev->l2ad_hand + distance; 4503 } 4504 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4505 uint64_t, taddr, boolean_t, all); 4506 4507top: 4508 mutex_enter(&l2arc_buflist_mtx); 4509 for (ab = list_tail(buflist); ab; ab = ab_prev) { 4510 ab_prev = list_prev(buflist, ab); 4511 4512 hash_lock = HDR_LOCK(ab); 4513 if (!mutex_tryenter(hash_lock)) { 4514 /* 4515 * Missed the hash lock. Retry. 4516 */ 4517 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4518 mutex_exit(&l2arc_buflist_mtx); 4519 mutex_enter(hash_lock); 4520 mutex_exit(hash_lock); 4521 goto top; 4522 } 4523 4524 if (HDR_L2_WRITE_HEAD(ab)) { 4525 /* 4526 * We hit a write head node. Leave it for 4527 * l2arc_write_done(). 4528 */ 4529 list_remove(buflist, ab); 4530 mutex_exit(hash_lock); 4531 continue; 4532 } 4533 4534 if (!all && ab->b_l2hdr != NULL && 4535 (ab->b_l2hdr->b_daddr > taddr || 4536 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4537 /* 4538 * We've evicted to the target address, 4539 * or the end of the device. 4540 */ 4541 mutex_exit(hash_lock); 4542 break; 4543 } 4544 4545 if (HDR_FREE_IN_PROGRESS(ab)) { 4546 /* 4547 * Already on the path to destruction. 4548 */ 4549 mutex_exit(hash_lock); 4550 continue; 4551 } 4552 4553 if (ab->b_state == arc_l2c_only) { 4554 ASSERT(!HDR_L2_READING(ab)); 4555 /* 4556 * This doesn't exist in the ARC. Destroy. 4557 * arc_hdr_destroy() will call list_remove() 4558 * and decrement arcstat_l2_size. 4559 */ 4560 arc_change_state(arc_anon, ab, hash_lock); 4561 arc_hdr_destroy(ab); 4562 } else { 4563 /* 4564 * Invalidate issued or about to be issued 4565 * reads, since we may be about to write 4566 * over this location. 4567 */ 4568 if (HDR_L2_READING(ab)) { 4569 ARCSTAT_BUMP(arcstat_l2_evict_reading); 4570 ab->b_flags |= ARC_L2_EVICTED; 4571 } 4572 4573 /* 4574 * Tell ARC this no longer exists in L2ARC. 4575 */ 4576 if (ab->b_l2hdr != NULL) { 4577 abl2 = ab->b_l2hdr; 4578 ab->b_l2hdr = NULL; 4579 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4580 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4581 } 4582 list_remove(buflist, ab); 4583 4584 /* 4585 * This may have been leftover after a 4586 * failed write. 4587 */ 4588 ab->b_flags &= ~ARC_L2_WRITING; 4589 } 4590 mutex_exit(hash_lock); 4591 } 4592 mutex_exit(&l2arc_buflist_mtx); 4593 4594 spa_l2cache_space_update(dev->l2ad_vdev, 0, -(taddr - dev->l2ad_evict)); 4595 dev->l2ad_evict = taddr; 4596} 4597 4598/* 4599 * Find and write ARC buffers to the L2ARC device. 4600 * 4601 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 4602 * for reading until they have completed writing. 4603 */ 4604static uint64_t 4605l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) 4606{ 4607 arc_buf_hdr_t *ab, *ab_prev, *head; 4608 l2arc_buf_hdr_t *hdrl2; 4609 list_t *list; 4610 uint64_t passed_sz, write_sz, buf_sz, headroom; 4611 void *buf_data; 4612 kmutex_t *hash_lock, *list_lock; 4613 boolean_t have_lock, full; 4614 l2arc_write_callback_t *cb; 4615 zio_t *pio, *wzio; 4616 int try; 4617 4618 ASSERT(dev->l2ad_vdev != NULL); 4619 4620 pio = NULL; 4621 write_sz = 0; 4622 full = B_FALSE; 4623 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 4624 head->b_flags |= ARC_L2_WRITE_HEAD; 4625 4626 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 4627 /* 4628 * Copy buffers for L2ARC writing. 4629 */ 4630 mutex_enter(&l2arc_buflist_mtx); 4631 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 4632 list = l2arc_list_locked(try, &list_lock); 4633 passed_sz = 0; 4634 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 4635 4636 /* 4637 * L2ARC fast warmup. 4638 * 4639 * Until the ARC is warm and starts to evict, read from the 4640 * head of the ARC lists rather than the tail. 4641 */ 4642 headroom = target_sz * l2arc_headroom; 4643 if (arc_warm == B_FALSE) 4644 ab = list_head(list); 4645 else 4646 ab = list_tail(list); 4647 if (ab == NULL) 4648 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 4649 4650 for (; ab; ab = ab_prev) { 4651 if (arc_warm == B_FALSE) 4652 ab_prev = list_next(list, ab); 4653 else 4654 ab_prev = list_prev(list, ab); 4655 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size); 4656 4657 hash_lock = HDR_LOCK(ab); 4658 have_lock = MUTEX_HELD(hash_lock); 4659 if (!have_lock && !mutex_tryenter(hash_lock)) { 4660 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 4661 /* 4662 * Skip this buffer rather than waiting. 4663 */ 4664 continue; 4665 } 4666 4667 passed_sz += ab->b_size; 4668 if (passed_sz > headroom) { 4669 /* 4670 * Searched too far. 4671 */ 4672 mutex_exit(hash_lock); 4673 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 4674 break; 4675 } 4676 4677 if (!l2arc_write_eligible(spa, ab)) { 4678 mutex_exit(hash_lock); 4679 continue; 4680 } 4681 4682 if ((write_sz + ab->b_size) > target_sz) { 4683 full = B_TRUE; 4684 mutex_exit(hash_lock); 4685 ARCSTAT_BUMP(arcstat_l2_write_full); 4686 break; 4687 } 4688 4689 if (pio == NULL) { 4690 /* 4691 * Insert a dummy header on the buflist so 4692 * l2arc_write_done() can find where the 4693 * write buffers begin without searching. 4694 */ 4695 list_insert_head(dev->l2ad_buflist, head); 4696 4697 cb = kmem_alloc( 4698 sizeof (l2arc_write_callback_t), KM_SLEEP); 4699 cb->l2wcb_dev = dev; 4700 cb->l2wcb_head = head; 4701 pio = zio_root(spa, l2arc_write_done, cb, 4702 ZIO_FLAG_CANFAIL); 4703 ARCSTAT_BUMP(arcstat_l2_write_pios); 4704 } 4705 4706 ARCSTAT_INCR(arcstat_l2_write_bytes_written, ab->b_size); 4707 /* 4708 * Create and add a new L2ARC header. 4709 */ 4710 hdrl2 = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 4711 hdrl2->b_dev = dev; 4712 hdrl2->b_daddr = dev->l2ad_hand; 4713 4714 ab->b_flags |= ARC_L2_WRITING; 4715 ab->b_l2hdr = hdrl2; 4716 list_insert_head(dev->l2ad_buflist, ab); 4717 buf_data = ab->b_buf->b_data; 4718 buf_sz = ab->b_size; 4719 4720 /* 4721 * Compute and store the buffer cksum before 4722 * writing. On debug the cksum is verified first. 4723 */ 4724 arc_cksum_verify(ab->b_buf); 4725 arc_cksum_compute(ab->b_buf, B_TRUE); 4726 4727 mutex_exit(hash_lock); 4728 4729 wzio = zio_write_phys(pio, dev->l2ad_vdev, 4730 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 4731 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 4732 ZIO_FLAG_CANFAIL, B_FALSE); 4733 4734 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 4735 zio_t *, wzio); 4736 (void) zio_nowait(wzio); 4737 4738 /* 4739 * Keep the clock hand suitably device-aligned. 4740 */ 4741 buf_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 4742 4743 write_sz += buf_sz; 4744 dev->l2ad_hand += buf_sz; 4745 } 4746 4747 mutex_exit(list_lock); 4748 4749 if (full == B_TRUE) 4750 break; 4751 } 4752 mutex_exit(&l2arc_buflist_mtx); 4753 4754 if (pio == NULL) { 4755 ASSERT3U(write_sz, ==, 0); 4756 kmem_cache_free(hdr_cache, head); 4757 return (0); 4758 } 4759 4760 ASSERT3U(write_sz, <=, target_sz); 4761 ARCSTAT_BUMP(arcstat_l2_writes_sent); 4762 ARCSTAT_INCR(arcstat_l2_write_bytes, write_sz); 4763 ARCSTAT_INCR(arcstat_l2_size, write_sz); 4764 spa_l2cache_space_update(dev->l2ad_vdev, 0, write_sz); 4765 4766 /* 4767 * Bump device hand to the device start if it is approaching the end. 4768 * l2arc_evict() will already have evicted ahead for this case. 4769 */ 4770 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 4771 spa_l2cache_space_update(dev->l2ad_vdev, 0, 4772 dev->l2ad_end - dev->l2ad_hand); 4773 dev->l2ad_hand = dev->l2ad_start; 4774 dev->l2ad_evict = dev->l2ad_start; 4775 dev->l2ad_first = B_FALSE; 4776 } 4777 4778 dev->l2ad_writing = B_TRUE; 4779 (void) zio_wait(pio); 4780 dev->l2ad_writing = B_FALSE; 4781 4782 return (write_sz); 4783} 4784 4785/* 4786 * This thread feeds the L2ARC at regular intervals. This is the beating 4787 * heart of the L2ARC. 4788 */ 4789static void 4790l2arc_feed_thread(void *dummy __unused) 4791{ 4792 callb_cpr_t cpr; 4793 l2arc_dev_t *dev; 4794 spa_t *spa; 4795 uint64_t size, wrote; 4796 clock_t begin, next = LBOLT; 4797 4798 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 4799 4800 mutex_enter(&l2arc_feed_thr_lock); 4801 4802 while (l2arc_thread_exit == 0) { 4803 CALLB_CPR_SAFE_BEGIN(&cpr); 4804 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 4805 next - LBOLT); 4806 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 4807 next = LBOLT + hz; 4808 4809 /* 4810 * Quick check for L2ARC devices. 4811 */ 4812 mutex_enter(&l2arc_dev_mtx); 4813 if (l2arc_ndev == 0) { 4814 mutex_exit(&l2arc_dev_mtx); 4815 continue; 4816 } 4817 mutex_exit(&l2arc_dev_mtx); 4818 begin = LBOLT; 4819 4820 /* 4821 * This selects the next l2arc device to write to, and in 4822 * doing so the next spa to feed from: dev->l2ad_spa. This 4823 * will return NULL if there are now no l2arc devices or if 4824 * they are all faulted. 4825 * 4826 * If a device is returned, its spa's config lock is also 4827 * held to prevent device removal. l2arc_dev_get_next() 4828 * will grab and release l2arc_dev_mtx. 4829 */ 4830 if ((dev = l2arc_dev_get_next()) == NULL) 4831 continue; 4832 4833 spa = dev->l2ad_spa; 4834 ASSERT(spa != NULL); 4835 4836 /* 4837 * Avoid contributing to memory pressure. 4838 */ 4839 if (arc_reclaim_needed()) { 4840 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 4841 spa_config_exit(spa, SCL_L2ARC, dev); 4842 continue; 4843 } 4844 4845 ARCSTAT_BUMP(arcstat_l2_feeds); 4846 4847 size = l2arc_write_size(dev); 4848 4849 /* 4850 * Evict L2ARC buffers that will be overwritten. 4851 */ 4852 l2arc_evict(dev, size, B_FALSE); 4853 4854 /* 4855 * Write ARC buffers. 4856 */ 4857 wrote = l2arc_write_buffers(spa, dev, size); 4858 4859 /* 4860 * Calculate interval between writes. 4861 */ 4862 next = l2arc_write_interval(begin, size, wrote); 4863 spa_config_exit(spa, SCL_L2ARC, dev); 4864 } 4865 4866 l2arc_thread_exit = 0; 4867 cv_broadcast(&l2arc_feed_thr_cv); 4868 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 4869 thread_exit(); 4870} 4871 4872boolean_t 4873l2arc_vdev_present(vdev_t *vd) 4874{ 4875 l2arc_dev_t *dev; 4876 4877 mutex_enter(&l2arc_dev_mtx); 4878 for (dev = list_head(l2arc_dev_list); dev != NULL; 4879 dev = list_next(l2arc_dev_list, dev)) { 4880 if (dev->l2ad_vdev == vd) 4881 break; 4882 } 4883 mutex_exit(&l2arc_dev_mtx); 4884 4885 return (dev != NULL); 4886} 4887 4888/* 4889 * Add a vdev for use by the L2ARC. By this point the spa has already 4890 * validated the vdev and opened it. 4891 */ 4892void 4893l2arc_add_vdev(spa_t *spa, vdev_t *vd, uint64_t start, uint64_t end) 4894{ 4895 l2arc_dev_t *adddev; 4896 4897 ASSERT(!l2arc_vdev_present(vd)); 4898 4899 /* 4900 * Create a new l2arc device entry. 4901 */ 4902 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 4903 adddev->l2ad_spa = spa; 4904 adddev->l2ad_vdev = vd; 4905 adddev->l2ad_write = l2arc_write_max; 4906 adddev->l2ad_boost = l2arc_write_boost; 4907 adddev->l2ad_start = start; 4908 adddev->l2ad_end = end; 4909 adddev->l2ad_hand = adddev->l2ad_start; 4910 adddev->l2ad_evict = adddev->l2ad_start; 4911 adddev->l2ad_first = B_TRUE; 4912 adddev->l2ad_writing = B_FALSE; 4913 ASSERT3U(adddev->l2ad_write, >, 0); 4914 4915 /* 4916 * This is a list of all ARC buffers that are still valid on the 4917 * device. 4918 */ 4919 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 4920 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 4921 offsetof(arc_buf_hdr_t, b_l2node)); 4922 4923 spa_l2cache_space_update(vd, adddev->l2ad_end - adddev->l2ad_hand, 0); 4924 4925 /* 4926 * Add device to global list 4927 */ 4928 mutex_enter(&l2arc_dev_mtx); 4929 list_insert_head(l2arc_dev_list, adddev); 4930 atomic_inc_64(&l2arc_ndev); 4931 mutex_exit(&l2arc_dev_mtx); 4932} 4933 4934/* 4935 * Remove a vdev from the L2ARC. 4936 */ 4937void 4938l2arc_remove_vdev(vdev_t *vd) 4939{ 4940 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 4941 4942 /* 4943 * Find the device by vdev 4944 */ 4945 mutex_enter(&l2arc_dev_mtx); 4946 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 4947 nextdev = list_next(l2arc_dev_list, dev); 4948 if (vd == dev->l2ad_vdev) { 4949 remdev = dev; 4950 break; 4951 } 4952 } 4953 ASSERT(remdev != NULL); 4954 4955 /* 4956 * Remove device from global list 4957 */ 4958 list_remove(l2arc_dev_list, remdev); 4959 l2arc_dev_last = NULL; /* may have been invalidated */ 4960 atomic_dec_64(&l2arc_ndev); 4961 mutex_exit(&l2arc_dev_mtx); 4962 4963 /* 4964 * Clear all buflists and ARC references. L2ARC device flush. 4965 */ 4966 l2arc_evict(remdev, 0, B_TRUE); 4967 list_destroy(remdev->l2ad_buflist); 4968 kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 4969 kmem_free(remdev, sizeof (l2arc_dev_t)); 4970} 4971 4972void 4973l2arc_init(void) 4974{ 4975 l2arc_thread_exit = 0; 4976 l2arc_ndev = 0; 4977 l2arc_writes_sent = 0; 4978 l2arc_writes_done = 0; 4979 4980 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4981 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 4982 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 4983 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 4984 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 4985 4986 l2arc_dev_list = &L2ARC_dev_list; 4987 l2arc_free_on_write = &L2ARC_free_on_write; 4988 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 4989 offsetof(l2arc_dev_t, l2ad_node)); 4990 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 4991 offsetof(l2arc_data_free_t, l2df_list_node)); 4992} 4993 4994void 4995l2arc_fini(void) 4996{ 4997 /* 4998 * This is called from dmu_fini(), which is called from spa_fini(); 4999 * Because of this, we can assume that all l2arc devices have 5000 * already been removed when the pools themselves were removed. 5001 */ 5002 5003 l2arc_do_free_on_write(); 5004 5005 mutex_destroy(&l2arc_feed_thr_lock); 5006 cv_destroy(&l2arc_feed_thr_cv); 5007 mutex_destroy(&l2arc_dev_mtx); 5008 mutex_destroy(&l2arc_buflist_mtx); 5009 mutex_destroy(&l2arc_free_on_write_mtx); 5010 5011 list_destroy(l2arc_dev_list); 5012 list_destroy(l2arc_free_on_write); 5013} 5014 5015void 5016l2arc_start(void) 5017{ 5018 if (!(spa_mode & FWRITE)) 5019 return; 5020 5021 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5022 TS_RUN, minclsyspri); 5023} 5024 5025void 5026l2arc_stop(void) 5027{ 5028 if (!(spa_mode & FWRITE)) 5029 return; 5030 5031 mutex_enter(&l2arc_feed_thr_lock); 5032 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5033 l2arc_thread_exit = 1; 5034 while (l2arc_thread_exit != 0) 5035 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5036 mutex_exit(&l2arc_feed_thr_lock); 5037} 5038