arc.c revision 193878
1168404Spjd/* 2168404Spjd * CDDL HEADER START 3168404Spjd * 4168404Spjd * The contents of this file are subject to the terms of the 5168404Spjd * Common Development and Distribution License (the "License"). 6168404Spjd * You may not use this file except in compliance with the License. 7168404Spjd * 8168404Spjd * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9168404Spjd * or http://www.opensolaris.org/os/licensing. 10168404Spjd * See the License for the specific language governing permissions 11168404Spjd * and limitations under the License. 12168404Spjd * 13168404Spjd * When distributing Covered Code, include this CDDL HEADER in each 14168404Spjd * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15168404Spjd * If applicable, add the following below this CDDL HEADER, with the 16168404Spjd * fields enclosed by brackets "[]" replaced with your own identifying 17168404Spjd * information: Portions Copyright [yyyy] [name of copyright owner] 18168404Spjd * 19168404Spjd * CDDL HEADER END 20168404Spjd */ 21168404Spjd/* 22185029Spjd * Copyright 2008 Sun Microsystems, Inc. All rights reserved. 23168404Spjd * Use is subject to license terms. 24168404Spjd */ 25168404Spjd 26168404Spjd/* 27168404Spjd * DVA-based Adjustable Replacement Cache 28168404Spjd * 29168404Spjd * While much of the theory of operation used here is 30168404Spjd * based on the self-tuning, low overhead replacement cache 31168404Spjd * presented by Megiddo and Modha at FAST 2003, there are some 32168404Spjd * significant differences: 33168404Spjd * 34168404Spjd * 1. The Megiddo and Modha model assumes any page is evictable. 35168404Spjd * Pages in its cache cannot be "locked" into memory. This makes 36168404Spjd * the eviction algorithm simple: evict the last page in the list. 37168404Spjd * This also make the performance characteristics easy to reason 38168404Spjd * about. Our cache is not so simple. At any given moment, some 39168404Spjd * subset of the blocks in the cache are un-evictable because we 40168404Spjd * have handed out a reference to them. Blocks are only evictable 41168404Spjd * when there are no external references active. This makes 42168404Spjd * eviction far more problematic: we choose to evict the evictable 43168404Spjd * blocks that are the "lowest" in the list. 44168404Spjd * 45168404Spjd * There are times when it is not possible to evict the requested 46168404Spjd * space. In these circumstances we are unable to adjust the cache 47168404Spjd * size. To prevent the cache growing unbounded at these times we 48185029Spjd * implement a "cache throttle" that slows the flow of new data 49185029Spjd * into the cache until we can make space available. 50168404Spjd * 51168404Spjd * 2. The Megiddo and Modha model assumes a fixed cache size. 52168404Spjd * Pages are evicted when the cache is full and there is a cache 53168404Spjd * miss. Our model has a variable sized cache. It grows with 54185029Spjd * high use, but also tries to react to memory pressure from the 55168404Spjd * operating system: decreasing its size when system memory is 56168404Spjd * tight. 57168404Spjd * 58168404Spjd * 3. The Megiddo and Modha model assumes a fixed page size. All 59168404Spjd * elements of the cache are therefor exactly the same size. So 60168404Spjd * when adjusting the cache size following a cache miss, its simply 61168404Spjd * a matter of choosing a single page to evict. In our model, we 62168404Spjd * have variable sized cache blocks (rangeing from 512 bytes to 63168404Spjd * 128K bytes). We therefor choose a set of blocks to evict to make 64168404Spjd * space for a cache miss that approximates as closely as possible 65168404Spjd * the space used by the new block. 66168404Spjd * 67168404Spjd * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 68168404Spjd * by N. Megiddo & D. Modha, FAST 2003 69168404Spjd */ 70168404Spjd 71168404Spjd/* 72168404Spjd * The locking model: 73168404Spjd * 74168404Spjd * A new reference to a cache buffer can be obtained in two 75168404Spjd * ways: 1) via a hash table lookup using the DVA as a key, 76185029Spjd * or 2) via one of the ARC lists. The arc_read() interface 77168404Spjd * uses method 1, while the internal arc algorithms for 78168404Spjd * adjusting the cache use method 2. We therefor provide two 79168404Spjd * types of locks: 1) the hash table lock array, and 2) the 80168404Spjd * arc list locks. 81168404Spjd * 82168404Spjd * Buffers do not have their own mutexs, rather they rely on the 83168404Spjd * hash table mutexs for the bulk of their protection (i.e. most 84168404Spjd * fields in the arc_buf_hdr_t are protected by these mutexs). 85168404Spjd * 86168404Spjd * buf_hash_find() returns the appropriate mutex (held) when it 87168404Spjd * locates the requested buffer in the hash table. It returns 88168404Spjd * NULL for the mutex if the buffer was not in the table. 89168404Spjd * 90168404Spjd * buf_hash_remove() expects the appropriate hash mutex to be 91168404Spjd * already held before it is invoked. 92168404Spjd * 93168404Spjd * Each arc state also has a mutex which is used to protect the 94168404Spjd * buffer list associated with the state. When attempting to 95168404Spjd * obtain a hash table lock while holding an arc list lock you 96168404Spjd * must use: mutex_tryenter() to avoid deadlock. Also note that 97168404Spjd * the active state mutex must be held before the ghost state mutex. 98168404Spjd * 99168404Spjd * Arc buffers may have an associated eviction callback function. 100168404Spjd * This function will be invoked prior to removing the buffer (e.g. 101168404Spjd * in arc_do_user_evicts()). Note however that the data associated 102168404Spjd * with the buffer may be evicted prior to the callback. The callback 103168404Spjd * must be made with *no locks held* (to prevent deadlock). Additionally, 104168404Spjd * the users of callbacks must ensure that their private data is 105168404Spjd * protected from simultaneous callbacks from arc_buf_evict() 106168404Spjd * and arc_do_user_evicts(). 107168404Spjd * 108168404Spjd * Note that the majority of the performance stats are manipulated 109168404Spjd * with atomic operations. 110185029Spjd * 111185029Spjd * The L2ARC uses the l2arc_buflist_mtx global mutex for the following: 112185029Spjd * 113185029Spjd * - L2ARC buflist creation 114185029Spjd * - L2ARC buflist eviction 115185029Spjd * - L2ARC write completion, which walks L2ARC buflists 116185029Spjd * - ARC header destruction, as it removes from L2ARC buflists 117185029Spjd * - ARC header release, as it removes from L2ARC buflists 118168404Spjd */ 119168404Spjd 120168404Spjd#include <sys/spa.h> 121168404Spjd#include <sys/zio.h> 122168404Spjd#include <sys/zio_checksum.h> 123168404Spjd#include <sys/zfs_context.h> 124168404Spjd#include <sys/arc.h> 125168404Spjd#include <sys/refcount.h> 126185029Spjd#include <sys/vdev.h> 127168404Spjd#ifdef _KERNEL 128168404Spjd#include <sys/dnlc.h> 129168404Spjd#endif 130168404Spjd#include <sys/callb.h> 131168404Spjd#include <sys/kstat.h> 132168404Spjd#include <sys/sdt.h> 133168404Spjd 134191902Skmacy#include <vm/vm_pageout.h> 135191902Skmacy 136168404Spjdstatic kmutex_t arc_reclaim_thr_lock; 137168404Spjdstatic kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 138168404Spjdstatic uint8_t arc_thread_exit; 139168404Spjd 140185029Spjdextern int zfs_write_limit_shift; 141185029Spjdextern uint64_t zfs_write_limit_max; 142185029Spjdextern kmutex_t zfs_write_limit_lock; 143185029Spjd 144168404Spjd#define ARC_REDUCE_DNLC_PERCENT 3 145168404Spjduint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 146168404Spjd 147168404Spjdtypedef enum arc_reclaim_strategy { 148168404Spjd ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 149168404Spjd ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 150168404Spjd} arc_reclaim_strategy_t; 151168404Spjd 152168404Spjd/* number of seconds before growing cache again */ 153168404Spjdstatic int arc_grow_retry = 60; 154168404Spjd 155168404Spjd/* 156168404Spjd * minimum lifespan of a prefetch block in clock ticks 157168404Spjd * (initialized in arc_init()) 158168404Spjd */ 159168404Spjdstatic int arc_min_prefetch_lifespan; 160168404Spjd 161193878Skmacyextern int zfs_prefetch_disable; 162193878Skmacyextern int zfs_prefetch_enable; 163168404Spjdstatic int arc_dead; 164168404Spjd 165168404Spjd/* 166185029Spjd * The arc has filled available memory and has now warmed up. 167185029Spjd */ 168185029Spjdstatic boolean_t arc_warm; 169185029Spjd 170185029Spjd/* 171168404Spjd * These tunables are for performance analysis. 172168404Spjd */ 173185029Spjduint64_t zfs_arc_max; 174185029Spjduint64_t zfs_arc_min; 175185029Spjduint64_t zfs_arc_meta_limit = 0; 176185029Spjdint zfs_mdcomp_disable = 0; 177185029Spjd 178185029SpjdTUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max); 179185029SpjdTUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 180185029SpjdTUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 181185029SpjdTUNABLE_INT("vfs.zfs.mdcomp_disable", &zfs_mdcomp_disable); 182168473SpjdSYSCTL_DECL(_vfs_zfs); 183185029SpjdSYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 184168473Spjd "Maximum ARC size"); 185185029SpjdSYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 186168473Spjd "Minimum ARC size"); 187185029SpjdSYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RDTUN, 188185029Spjd &zfs_mdcomp_disable, 0, "Disable metadata compression"); 189168404Spjd 190168404Spjd/* 191185029Spjd * Note that buffers can be in one of 6 states: 192168404Spjd * ARC_anon - anonymous (discussed below) 193168404Spjd * ARC_mru - recently used, currently cached 194168404Spjd * ARC_mru_ghost - recentely used, no longer in cache 195168404Spjd * ARC_mfu - frequently used, currently cached 196168404Spjd * ARC_mfu_ghost - frequently used, no longer in cache 197185029Spjd * ARC_l2c_only - exists in L2ARC but not other states 198185029Spjd * When there are no active references to the buffer, they are 199185029Spjd * are linked onto a list in one of these arc states. These are 200185029Spjd * the only buffers that can be evicted or deleted. Within each 201185029Spjd * state there are multiple lists, one for meta-data and one for 202185029Spjd * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 203185029Spjd * etc.) is tracked separately so that it can be managed more 204185029Spjd * explicitly: favored over data, limited explicitly. 205168404Spjd * 206168404Spjd * Anonymous buffers are buffers that are not associated with 207168404Spjd * a DVA. These are buffers that hold dirty block copies 208168404Spjd * before they are written to stable storage. By definition, 209168404Spjd * they are "ref'd" and are considered part of arc_mru 210168404Spjd * that cannot be freed. Generally, they will aquire a DVA 211168404Spjd * as they are written and migrate onto the arc_mru list. 212185029Spjd * 213185029Spjd * The ARC_l2c_only state is for buffers that are in the second 214185029Spjd * level ARC but no longer in any of the ARC_m* lists. The second 215185029Spjd * level ARC itself may also contain buffers that are in any of 216185029Spjd * the ARC_m* states - meaning that a buffer can exist in two 217185029Spjd * places. The reason for the ARC_l2c_only state is to keep the 218185029Spjd * buffer header in the hash table, so that reads that hit the 219185029Spjd * second level ARC benefit from these fast lookups. 220168404Spjd */ 221168404Spjd 222168404Spjdtypedef struct arc_state { 223185029Spjd list_t arcs_list[ARC_BUFC_NUMTYPES]; /* list of evictable buffers */ 224185029Spjd uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 225185029Spjd uint64_t arcs_size; /* total amount of data in this state */ 226168404Spjd kmutex_t arcs_mtx; 227168404Spjd} arc_state_t; 228168404Spjd 229185029Spjd/* The 6 states: */ 230168404Spjdstatic arc_state_t ARC_anon; 231168404Spjdstatic arc_state_t ARC_mru; 232168404Spjdstatic arc_state_t ARC_mru_ghost; 233168404Spjdstatic arc_state_t ARC_mfu; 234168404Spjdstatic arc_state_t ARC_mfu_ghost; 235185029Spjdstatic arc_state_t ARC_l2c_only; 236168404Spjd 237168404Spjdtypedef struct arc_stats { 238168404Spjd kstat_named_t arcstat_hits; 239168404Spjd kstat_named_t arcstat_misses; 240168404Spjd kstat_named_t arcstat_demand_data_hits; 241168404Spjd kstat_named_t arcstat_demand_data_misses; 242168404Spjd kstat_named_t arcstat_demand_metadata_hits; 243168404Spjd kstat_named_t arcstat_demand_metadata_misses; 244168404Spjd kstat_named_t arcstat_prefetch_data_hits; 245168404Spjd kstat_named_t arcstat_prefetch_data_misses; 246168404Spjd kstat_named_t arcstat_prefetch_metadata_hits; 247168404Spjd kstat_named_t arcstat_prefetch_metadata_misses; 248168404Spjd kstat_named_t arcstat_mru_hits; 249168404Spjd kstat_named_t arcstat_mru_ghost_hits; 250168404Spjd kstat_named_t arcstat_mfu_hits; 251168404Spjd kstat_named_t arcstat_mfu_ghost_hits; 252168404Spjd kstat_named_t arcstat_deleted; 253168404Spjd kstat_named_t arcstat_recycle_miss; 254168404Spjd kstat_named_t arcstat_mutex_miss; 255168404Spjd kstat_named_t arcstat_evict_skip; 256168404Spjd kstat_named_t arcstat_hash_elements; 257168404Spjd kstat_named_t arcstat_hash_elements_max; 258168404Spjd kstat_named_t arcstat_hash_collisions; 259168404Spjd kstat_named_t arcstat_hash_chains; 260168404Spjd kstat_named_t arcstat_hash_chain_max; 261168404Spjd kstat_named_t arcstat_p; 262168404Spjd kstat_named_t arcstat_c; 263168404Spjd kstat_named_t arcstat_c_min; 264168404Spjd kstat_named_t arcstat_c_max; 265168404Spjd kstat_named_t arcstat_size; 266185029Spjd kstat_named_t arcstat_hdr_size; 267185029Spjd kstat_named_t arcstat_l2_hits; 268185029Spjd kstat_named_t arcstat_l2_misses; 269185029Spjd kstat_named_t arcstat_l2_feeds; 270185029Spjd kstat_named_t arcstat_l2_rw_clash; 271185029Spjd kstat_named_t arcstat_l2_writes_sent; 272185029Spjd kstat_named_t arcstat_l2_writes_done; 273185029Spjd kstat_named_t arcstat_l2_writes_error; 274185029Spjd kstat_named_t arcstat_l2_writes_hdr_miss; 275185029Spjd kstat_named_t arcstat_l2_evict_lock_retry; 276185029Spjd kstat_named_t arcstat_l2_evict_reading; 277185029Spjd kstat_named_t arcstat_l2_free_on_write; 278185029Spjd kstat_named_t arcstat_l2_abort_lowmem; 279185029Spjd kstat_named_t arcstat_l2_cksum_bad; 280185029Spjd kstat_named_t arcstat_l2_io_error; 281185029Spjd kstat_named_t arcstat_l2_size; 282185029Spjd kstat_named_t arcstat_l2_hdr_size; 283185029Spjd kstat_named_t arcstat_memory_throttle_count; 284168404Spjd} arc_stats_t; 285168404Spjd 286168404Spjdstatic arc_stats_t arc_stats = { 287168404Spjd { "hits", KSTAT_DATA_UINT64 }, 288168404Spjd { "misses", KSTAT_DATA_UINT64 }, 289168404Spjd { "demand_data_hits", KSTAT_DATA_UINT64 }, 290168404Spjd { "demand_data_misses", KSTAT_DATA_UINT64 }, 291168404Spjd { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 292168404Spjd { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 293168404Spjd { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 294168404Spjd { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 295168404Spjd { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 296168404Spjd { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 297168404Spjd { "mru_hits", KSTAT_DATA_UINT64 }, 298168404Spjd { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 299168404Spjd { "mfu_hits", KSTAT_DATA_UINT64 }, 300168404Spjd { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 301168404Spjd { "deleted", KSTAT_DATA_UINT64 }, 302168404Spjd { "recycle_miss", KSTAT_DATA_UINT64 }, 303168404Spjd { "mutex_miss", KSTAT_DATA_UINT64 }, 304168404Spjd { "evict_skip", KSTAT_DATA_UINT64 }, 305168404Spjd { "hash_elements", KSTAT_DATA_UINT64 }, 306168404Spjd { "hash_elements_max", KSTAT_DATA_UINT64 }, 307168404Spjd { "hash_collisions", KSTAT_DATA_UINT64 }, 308168404Spjd { "hash_chains", KSTAT_DATA_UINT64 }, 309168404Spjd { "hash_chain_max", KSTAT_DATA_UINT64 }, 310168404Spjd { "p", KSTAT_DATA_UINT64 }, 311168404Spjd { "c", KSTAT_DATA_UINT64 }, 312168404Spjd { "c_min", KSTAT_DATA_UINT64 }, 313168404Spjd { "c_max", KSTAT_DATA_UINT64 }, 314185029Spjd { "size", KSTAT_DATA_UINT64 }, 315185029Spjd { "hdr_size", KSTAT_DATA_UINT64 }, 316185029Spjd { "l2_hits", KSTAT_DATA_UINT64 }, 317185029Spjd { "l2_misses", KSTAT_DATA_UINT64 }, 318185029Spjd { "l2_feeds", KSTAT_DATA_UINT64 }, 319185029Spjd { "l2_rw_clash", KSTAT_DATA_UINT64 }, 320185029Spjd { "l2_writes_sent", KSTAT_DATA_UINT64 }, 321185029Spjd { "l2_writes_done", KSTAT_DATA_UINT64 }, 322185029Spjd { "l2_writes_error", KSTAT_DATA_UINT64 }, 323185029Spjd { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 324185029Spjd { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 325185029Spjd { "l2_evict_reading", KSTAT_DATA_UINT64 }, 326185029Spjd { "l2_free_on_write", KSTAT_DATA_UINT64 }, 327185029Spjd { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 328185029Spjd { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 329185029Spjd { "l2_io_error", KSTAT_DATA_UINT64 }, 330185029Spjd { "l2_size", KSTAT_DATA_UINT64 }, 331185029Spjd { "l2_hdr_size", KSTAT_DATA_UINT64 }, 332185029Spjd { "memory_throttle_count", KSTAT_DATA_UINT64 } 333168404Spjd}; 334168404Spjd 335168404Spjd#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 336168404Spjd 337168404Spjd#define ARCSTAT_INCR(stat, val) \ 338168404Spjd atomic_add_64(&arc_stats.stat.value.ui64, (val)); 339168404Spjd 340168404Spjd#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 341168404Spjd#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 342168404Spjd 343168404Spjd#define ARCSTAT_MAX(stat, val) { \ 344168404Spjd uint64_t m; \ 345168404Spjd while ((val) > (m = arc_stats.stat.value.ui64) && \ 346168404Spjd (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 347168404Spjd continue; \ 348168404Spjd} 349168404Spjd 350168404Spjd#define ARCSTAT_MAXSTAT(stat) \ 351168404Spjd ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 352168404Spjd 353168404Spjd/* 354168404Spjd * We define a macro to allow ARC hits/misses to be easily broken down by 355168404Spjd * two separate conditions, giving a total of four different subtypes for 356168404Spjd * each of hits and misses (so eight statistics total). 357168404Spjd */ 358168404Spjd#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 359168404Spjd if (cond1) { \ 360168404Spjd if (cond2) { \ 361168404Spjd ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 362168404Spjd } else { \ 363168404Spjd ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 364168404Spjd } \ 365168404Spjd } else { \ 366168404Spjd if (cond2) { \ 367168404Spjd ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 368168404Spjd } else { \ 369168404Spjd ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 370168404Spjd } \ 371168404Spjd } 372168404Spjd 373168404Spjdkstat_t *arc_ksp; 374168404Spjdstatic arc_state_t *arc_anon; 375168404Spjdstatic arc_state_t *arc_mru; 376168404Spjdstatic arc_state_t *arc_mru_ghost; 377168404Spjdstatic arc_state_t *arc_mfu; 378168404Spjdstatic arc_state_t *arc_mfu_ghost; 379185029Spjdstatic arc_state_t *arc_l2c_only; 380168404Spjd 381168404Spjd/* 382168404Spjd * There are several ARC variables that are critical to export as kstats -- 383168404Spjd * but we don't want to have to grovel around in the kstat whenever we wish to 384168404Spjd * manipulate them. For these variables, we therefore define them to be in 385168404Spjd * terms of the statistic variable. This assures that we are not introducing 386168404Spjd * the possibility of inconsistency by having shadow copies of the variables, 387168404Spjd * while still allowing the code to be readable. 388168404Spjd */ 389168404Spjd#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 390168404Spjd#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 391168404Spjd#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 392168404Spjd#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 393168404Spjd#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 394168404Spjd 395168404Spjdstatic int arc_no_grow; /* Don't try to grow cache size */ 396168404Spjdstatic uint64_t arc_tempreserve; 397185029Spjdstatic uint64_t arc_meta_used; 398185029Spjdstatic uint64_t arc_meta_limit; 399185029Spjdstatic uint64_t arc_meta_max = 0; 400185029SpjdSYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RDTUN, 401185029Spjd &arc_meta_used, 0, "ARC metadata used"); 402185029SpjdSYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RDTUN, 403185029Spjd &arc_meta_limit, 0, "ARC metadata limit"); 404168404Spjd 405185029Spjdtypedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 406185029Spjd 407168404Spjdtypedef struct arc_callback arc_callback_t; 408168404Spjd 409168404Spjdstruct arc_callback { 410168404Spjd void *acb_private; 411168404Spjd arc_done_func_t *acb_done; 412168404Spjd arc_buf_t *acb_buf; 413168404Spjd zio_t *acb_zio_dummy; 414168404Spjd arc_callback_t *acb_next; 415168404Spjd}; 416168404Spjd 417168404Spjdtypedef struct arc_write_callback arc_write_callback_t; 418168404Spjd 419168404Spjdstruct arc_write_callback { 420168404Spjd void *awcb_private; 421168404Spjd arc_done_func_t *awcb_ready; 422168404Spjd arc_done_func_t *awcb_done; 423168404Spjd arc_buf_t *awcb_buf; 424168404Spjd}; 425168404Spjd 426168404Spjdstruct arc_buf_hdr { 427168404Spjd /* protected by hash lock */ 428168404Spjd dva_t b_dva; 429168404Spjd uint64_t b_birth; 430168404Spjd uint64_t b_cksum0; 431168404Spjd 432168404Spjd kmutex_t b_freeze_lock; 433168404Spjd zio_cksum_t *b_freeze_cksum; 434168404Spjd 435168404Spjd arc_buf_hdr_t *b_hash_next; 436168404Spjd arc_buf_t *b_buf; 437168404Spjd uint32_t b_flags; 438168404Spjd uint32_t b_datacnt; 439168404Spjd 440168404Spjd arc_callback_t *b_acb; 441168404Spjd kcondvar_t b_cv; 442168404Spjd 443168404Spjd /* immutable */ 444168404Spjd arc_buf_contents_t b_type; 445168404Spjd uint64_t b_size; 446168404Spjd spa_t *b_spa; 447168404Spjd 448168404Spjd /* protected by arc state mutex */ 449168404Spjd arc_state_t *b_state; 450168404Spjd list_node_t b_arc_node; 451168404Spjd 452168404Spjd /* updated atomically */ 453168404Spjd clock_t b_arc_access; 454168404Spjd 455168404Spjd /* self protecting */ 456168404Spjd refcount_t b_refcnt; 457185029Spjd 458185029Spjd l2arc_buf_hdr_t *b_l2hdr; 459185029Spjd list_node_t b_l2node; 460168404Spjd}; 461168404Spjd 462168404Spjdstatic arc_buf_t *arc_eviction_list; 463168404Spjdstatic kmutex_t arc_eviction_mtx; 464168404Spjdstatic arc_buf_hdr_t arc_eviction_hdr; 465168404Spjdstatic void arc_get_data_buf(arc_buf_t *buf); 466168404Spjdstatic void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock); 467185029Spjdstatic int arc_evict_needed(arc_buf_contents_t type); 468185029Spjdstatic void arc_evict_ghost(arc_state_t *state, spa_t *spa, int64_t bytes); 469168404Spjd 470168404Spjd#define GHOST_STATE(state) \ 471185029Spjd ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 472185029Spjd (state) == arc_l2c_only) 473168404Spjd 474168404Spjd/* 475168404Spjd * Private ARC flags. These flags are private ARC only flags that will show up 476168404Spjd * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can 477168404Spjd * be passed in as arc_flags in things like arc_read. However, these flags 478168404Spjd * should never be passed and should only be set by ARC code. When adding new 479168404Spjd * public flags, make sure not to smash the private ones. 480168404Spjd */ 481168404Spjd 482168404Spjd#define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */ 483168404Spjd#define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */ 484168404Spjd#define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */ 485168404Spjd#define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */ 486168404Spjd#define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */ 487168404Spjd#define ARC_INDIRECT (1 << 14) /* this is an indirect block */ 488185029Spjd#define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */ 489185029Spjd#define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */ 490185029Spjd#define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */ 491185029Spjd#define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */ 492185029Spjd#define ARC_STORED (1 << 19) /* has been store()d to */ 493168404Spjd 494168404Spjd#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE) 495168404Spjd#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS) 496168404Spjd#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR) 497168404Spjd#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ) 498168404Spjd#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE) 499185029Spjd#define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS) 500185029Spjd#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE) 501185029Spjd#define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \ 502185029Spjd (hdr)->b_l2hdr != NULL) 503185029Spjd#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING) 504185029Spjd#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED) 505185029Spjd#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD) 506168404Spjd 507168404Spjd/* 508185029Spjd * Other sizes 509185029Spjd */ 510185029Spjd 511185029Spjd#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 512185029Spjd#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 513185029Spjd 514185029Spjd/* 515168404Spjd * Hash table routines 516168404Spjd */ 517168404Spjd 518168404Spjd#define HT_LOCK_PAD 128 519168404Spjd 520168404Spjdstruct ht_lock { 521168404Spjd kmutex_t ht_lock; 522168404Spjd#ifdef _KERNEL 523168404Spjd unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 524168404Spjd#endif 525168404Spjd}; 526168404Spjd 527168404Spjd#define BUF_LOCKS 256 528168404Spjdtypedef struct buf_hash_table { 529168404Spjd uint64_t ht_mask; 530168404Spjd arc_buf_hdr_t **ht_table; 531168404Spjd struct ht_lock ht_locks[BUF_LOCKS]; 532168404Spjd} buf_hash_table_t; 533168404Spjd 534168404Spjdstatic buf_hash_table_t buf_hash_table; 535168404Spjd 536168404Spjd#define BUF_HASH_INDEX(spa, dva, birth) \ 537168404Spjd (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 538168404Spjd#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 539168404Spjd#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 540168404Spjd#define HDR_LOCK(buf) \ 541168404Spjd (BUF_HASH_LOCK(BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth))) 542168404Spjd 543168404Spjduint64_t zfs_crc64_table[256]; 544168404Spjd 545185029Spjd/* 546185029Spjd * Level 2 ARC 547185029Spjd */ 548185029Spjd 549185029Spjd#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 550185029Spjd#define L2ARC_HEADROOM 4 /* num of writes */ 551185029Spjd#define L2ARC_FEED_SECS 1 /* caching interval */ 552185029Spjd 553185029Spjd#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 554185029Spjd#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 555185029Spjd 556185029Spjd/* 557185029Spjd * L2ARC Performance Tunables 558185029Spjd */ 559185029Spjduint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 560185029Spjduint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 561185029Spjduint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 562185029Spjduint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 563185029Spjdboolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 564185029Spjd 565185029Spjd/* 566185029Spjd * L2ARC Internals 567185029Spjd */ 568185029Spjdtypedef struct l2arc_dev { 569185029Spjd vdev_t *l2ad_vdev; /* vdev */ 570185029Spjd spa_t *l2ad_spa; /* spa */ 571185029Spjd uint64_t l2ad_hand; /* next write location */ 572185029Spjd uint64_t l2ad_write; /* desired write size, bytes */ 573185029Spjd uint64_t l2ad_boost; /* warmup write boost, bytes */ 574185029Spjd uint64_t l2ad_start; /* first addr on device */ 575185029Spjd uint64_t l2ad_end; /* last addr on device */ 576185029Spjd uint64_t l2ad_evict; /* last addr eviction reached */ 577185029Spjd boolean_t l2ad_first; /* first sweep through */ 578185029Spjd list_t *l2ad_buflist; /* buffer list */ 579185029Spjd list_node_t l2ad_node; /* device list node */ 580185029Spjd} l2arc_dev_t; 581185029Spjd 582185029Spjdstatic list_t L2ARC_dev_list; /* device list */ 583185029Spjdstatic list_t *l2arc_dev_list; /* device list pointer */ 584185029Spjdstatic kmutex_t l2arc_dev_mtx; /* device list mutex */ 585185029Spjdstatic l2arc_dev_t *l2arc_dev_last; /* last device used */ 586185029Spjdstatic kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 587185029Spjdstatic list_t L2ARC_free_on_write; /* free after write buf list */ 588185029Spjdstatic list_t *l2arc_free_on_write; /* free after write list ptr */ 589185029Spjdstatic kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 590185029Spjdstatic uint64_t l2arc_ndev; /* number of devices */ 591185029Spjd 592185029Spjdtypedef struct l2arc_read_callback { 593185029Spjd arc_buf_t *l2rcb_buf; /* read buffer */ 594185029Spjd spa_t *l2rcb_spa; /* spa */ 595185029Spjd blkptr_t l2rcb_bp; /* original blkptr */ 596185029Spjd zbookmark_t l2rcb_zb; /* original bookmark */ 597185029Spjd int l2rcb_flags; /* original flags */ 598185029Spjd} l2arc_read_callback_t; 599185029Spjd 600185029Spjdtypedef struct l2arc_write_callback { 601185029Spjd l2arc_dev_t *l2wcb_dev; /* device info */ 602185029Spjd arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 603185029Spjd} l2arc_write_callback_t; 604185029Spjd 605185029Spjdstruct l2arc_buf_hdr { 606185029Spjd /* protected by arc_buf_hdr mutex */ 607185029Spjd l2arc_dev_t *b_dev; /* L2ARC device */ 608185029Spjd daddr_t b_daddr; /* disk address, offset byte */ 609185029Spjd}; 610185029Spjd 611185029Spjdtypedef struct l2arc_data_free { 612185029Spjd /* protected by l2arc_free_on_write_mtx */ 613185029Spjd void *l2df_data; 614185029Spjd size_t l2df_size; 615185029Spjd void (*l2df_func)(void *, size_t); 616185029Spjd list_node_t l2df_list_node; 617185029Spjd} l2arc_data_free_t; 618185029Spjd 619185029Spjdstatic kmutex_t l2arc_feed_thr_lock; 620185029Spjdstatic kcondvar_t l2arc_feed_thr_cv; 621185029Spjdstatic uint8_t l2arc_thread_exit; 622185029Spjd 623185029Spjdstatic void l2arc_read_done(zio_t *zio); 624185029Spjdstatic void l2arc_hdr_stat_add(void); 625185029Spjdstatic void l2arc_hdr_stat_remove(void); 626185029Spjd 627168404Spjdstatic uint64_t 628185029Spjdbuf_hash(spa_t *spa, const dva_t *dva, uint64_t birth) 629168404Spjd{ 630168404Spjd uintptr_t spav = (uintptr_t)spa; 631168404Spjd uint8_t *vdva = (uint8_t *)dva; 632168404Spjd uint64_t crc = -1ULL; 633168404Spjd int i; 634168404Spjd 635168404Spjd ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 636168404Spjd 637168404Spjd for (i = 0; i < sizeof (dva_t); i++) 638168404Spjd crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 639168404Spjd 640168404Spjd crc ^= (spav>>8) ^ birth; 641168404Spjd 642168404Spjd return (crc); 643168404Spjd} 644168404Spjd 645168404Spjd#define BUF_EMPTY(buf) \ 646168404Spjd ((buf)->b_dva.dva_word[0] == 0 && \ 647168404Spjd (buf)->b_dva.dva_word[1] == 0 && \ 648168404Spjd (buf)->b_birth == 0) 649168404Spjd 650168404Spjd#define BUF_EQUAL(spa, dva, birth, buf) \ 651168404Spjd ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 652168404Spjd ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 653168404Spjd ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 654168404Spjd 655168404Spjdstatic arc_buf_hdr_t * 656185029Spjdbuf_hash_find(spa_t *spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp) 657168404Spjd{ 658168404Spjd uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 659168404Spjd kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 660168404Spjd arc_buf_hdr_t *buf; 661168404Spjd 662168404Spjd mutex_enter(hash_lock); 663168404Spjd for (buf = buf_hash_table.ht_table[idx]; buf != NULL; 664168404Spjd buf = buf->b_hash_next) { 665168404Spjd if (BUF_EQUAL(spa, dva, birth, buf)) { 666168404Spjd *lockp = hash_lock; 667168404Spjd return (buf); 668168404Spjd } 669168404Spjd } 670168404Spjd mutex_exit(hash_lock); 671168404Spjd *lockp = NULL; 672168404Spjd return (NULL); 673168404Spjd} 674168404Spjd 675168404Spjd/* 676168404Spjd * Insert an entry into the hash table. If there is already an element 677168404Spjd * equal to elem in the hash table, then the already existing element 678168404Spjd * will be returned and the new element will not be inserted. 679168404Spjd * Otherwise returns NULL. 680168404Spjd */ 681168404Spjdstatic arc_buf_hdr_t * 682168404Spjdbuf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp) 683168404Spjd{ 684168404Spjd uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 685168404Spjd kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 686168404Spjd arc_buf_hdr_t *fbuf; 687168404Spjd uint32_t i; 688168404Spjd 689168404Spjd ASSERT(!HDR_IN_HASH_TABLE(buf)); 690168404Spjd *lockp = hash_lock; 691168404Spjd mutex_enter(hash_lock); 692168404Spjd for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL; 693168404Spjd fbuf = fbuf->b_hash_next, i++) { 694168404Spjd if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf)) 695168404Spjd return (fbuf); 696168404Spjd } 697168404Spjd 698168404Spjd buf->b_hash_next = buf_hash_table.ht_table[idx]; 699168404Spjd buf_hash_table.ht_table[idx] = buf; 700168404Spjd buf->b_flags |= ARC_IN_HASH_TABLE; 701168404Spjd 702168404Spjd /* collect some hash table performance data */ 703168404Spjd if (i > 0) { 704168404Spjd ARCSTAT_BUMP(arcstat_hash_collisions); 705168404Spjd if (i == 1) 706168404Spjd ARCSTAT_BUMP(arcstat_hash_chains); 707168404Spjd 708168404Spjd ARCSTAT_MAX(arcstat_hash_chain_max, i); 709168404Spjd } 710168404Spjd 711168404Spjd ARCSTAT_BUMP(arcstat_hash_elements); 712168404Spjd ARCSTAT_MAXSTAT(arcstat_hash_elements); 713168404Spjd 714168404Spjd return (NULL); 715168404Spjd} 716168404Spjd 717168404Spjdstatic void 718168404Spjdbuf_hash_remove(arc_buf_hdr_t *buf) 719168404Spjd{ 720168404Spjd arc_buf_hdr_t *fbuf, **bufp; 721168404Spjd uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 722168404Spjd 723168404Spjd ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 724168404Spjd ASSERT(HDR_IN_HASH_TABLE(buf)); 725168404Spjd 726168404Spjd bufp = &buf_hash_table.ht_table[idx]; 727168404Spjd while ((fbuf = *bufp) != buf) { 728168404Spjd ASSERT(fbuf != NULL); 729168404Spjd bufp = &fbuf->b_hash_next; 730168404Spjd } 731168404Spjd *bufp = buf->b_hash_next; 732168404Spjd buf->b_hash_next = NULL; 733168404Spjd buf->b_flags &= ~ARC_IN_HASH_TABLE; 734168404Spjd 735168404Spjd /* collect some hash table performance data */ 736168404Spjd ARCSTAT_BUMPDOWN(arcstat_hash_elements); 737168404Spjd 738168404Spjd if (buf_hash_table.ht_table[idx] && 739168404Spjd buf_hash_table.ht_table[idx]->b_hash_next == NULL) 740168404Spjd ARCSTAT_BUMPDOWN(arcstat_hash_chains); 741168404Spjd} 742168404Spjd 743168404Spjd/* 744168404Spjd * Global data structures and functions for the buf kmem cache. 745168404Spjd */ 746168404Spjdstatic kmem_cache_t *hdr_cache; 747168404Spjdstatic kmem_cache_t *buf_cache; 748168404Spjd 749168404Spjdstatic void 750168404Spjdbuf_fini(void) 751168404Spjd{ 752168404Spjd int i; 753168404Spjd 754168404Spjd kmem_free(buf_hash_table.ht_table, 755168404Spjd (buf_hash_table.ht_mask + 1) * sizeof (void *)); 756168404Spjd for (i = 0; i < BUF_LOCKS; i++) 757168404Spjd mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 758168404Spjd kmem_cache_destroy(hdr_cache); 759168404Spjd kmem_cache_destroy(buf_cache); 760168404Spjd} 761168404Spjd 762168404Spjd/* 763168404Spjd * Constructor callback - called when the cache is empty 764168404Spjd * and a new buf is requested. 765168404Spjd */ 766168404Spjd/* ARGSUSED */ 767168404Spjdstatic int 768168404Spjdhdr_cons(void *vbuf, void *unused, int kmflag) 769168404Spjd{ 770168404Spjd arc_buf_hdr_t *buf = vbuf; 771168404Spjd 772168404Spjd bzero(buf, sizeof (arc_buf_hdr_t)); 773168404Spjd refcount_create(&buf->b_refcnt); 774168404Spjd cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL); 775185029Spjd mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 776185029Spjd 777185029Spjd ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 778168404Spjd return (0); 779168404Spjd} 780168404Spjd 781185029Spjd/* ARGSUSED */ 782185029Spjdstatic int 783185029Spjdbuf_cons(void *vbuf, void *unused, int kmflag) 784185029Spjd{ 785185029Spjd arc_buf_t *buf = vbuf; 786185029Spjd 787185029Spjd bzero(buf, sizeof (arc_buf_t)); 788185029Spjd rw_init(&buf->b_lock, NULL, RW_DEFAULT, NULL); 789185029Spjd return (0); 790185029Spjd} 791185029Spjd 792168404Spjd/* 793168404Spjd * Destructor callback - called when a cached buf is 794168404Spjd * no longer required. 795168404Spjd */ 796168404Spjd/* ARGSUSED */ 797168404Spjdstatic void 798168404Spjdhdr_dest(void *vbuf, void *unused) 799168404Spjd{ 800168404Spjd arc_buf_hdr_t *buf = vbuf; 801168404Spjd 802168404Spjd refcount_destroy(&buf->b_refcnt); 803168404Spjd cv_destroy(&buf->b_cv); 804185029Spjd mutex_destroy(&buf->b_freeze_lock); 805185029Spjd 806185029Spjd ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 807168404Spjd} 808168404Spjd 809185029Spjd/* ARGSUSED */ 810185029Spjdstatic void 811185029Spjdbuf_dest(void *vbuf, void *unused) 812185029Spjd{ 813185029Spjd arc_buf_t *buf = vbuf; 814185029Spjd 815185029Spjd rw_destroy(&buf->b_lock); 816185029Spjd} 817185029Spjd 818168404Spjd/* 819168404Spjd * Reclaim callback -- invoked when memory is low. 820168404Spjd */ 821168404Spjd/* ARGSUSED */ 822168404Spjdstatic void 823168404Spjdhdr_recl(void *unused) 824168404Spjd{ 825168404Spjd dprintf("hdr_recl called\n"); 826168404Spjd /* 827168404Spjd * umem calls the reclaim func when we destroy the buf cache, 828168404Spjd * which is after we do arc_fini(). 829168404Spjd */ 830168404Spjd if (!arc_dead) 831168404Spjd cv_signal(&arc_reclaim_thr_cv); 832168404Spjd} 833168404Spjd 834168404Spjdstatic void 835168404Spjdbuf_init(void) 836168404Spjd{ 837168404Spjd uint64_t *ct; 838168404Spjd uint64_t hsize = 1ULL << 12; 839168404Spjd int i, j; 840168404Spjd 841168404Spjd /* 842168404Spjd * The hash table is big enough to fill all of physical memory 843168404Spjd * with an average 64K block size. The table will take up 844168404Spjd * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers). 845168404Spjd */ 846168696Spjd while (hsize * 65536 < (uint64_t)physmem * PAGESIZE) 847168404Spjd hsize <<= 1; 848168404Spjdretry: 849168404Spjd buf_hash_table.ht_mask = hsize - 1; 850168404Spjd buf_hash_table.ht_table = 851168404Spjd kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 852168404Spjd if (buf_hash_table.ht_table == NULL) { 853168404Spjd ASSERT(hsize > (1ULL << 8)); 854168404Spjd hsize >>= 1; 855168404Spjd goto retry; 856168404Spjd } 857168404Spjd 858168404Spjd hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 859168404Spjd 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 860168404Spjd buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 861185029Spjd 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 862168404Spjd 863168404Spjd for (i = 0; i < 256; i++) 864168404Spjd for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 865168404Spjd *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 866168404Spjd 867168404Spjd for (i = 0; i < BUF_LOCKS; i++) { 868168404Spjd mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 869168404Spjd NULL, MUTEX_DEFAULT, NULL); 870168404Spjd } 871168404Spjd} 872168404Spjd 873168404Spjd#define ARC_MINTIME (hz>>4) /* 62 ms */ 874168404Spjd 875168404Spjdstatic void 876168404Spjdarc_cksum_verify(arc_buf_t *buf) 877168404Spjd{ 878168404Spjd zio_cksum_t zc; 879168404Spjd 880168404Spjd if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 881168404Spjd return; 882168404Spjd 883168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 884168404Spjd if (buf->b_hdr->b_freeze_cksum == NULL || 885168404Spjd (buf->b_hdr->b_flags & ARC_IO_ERROR)) { 886168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 887168404Spjd return; 888168404Spjd } 889168404Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 890168404Spjd if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 891168404Spjd panic("buffer modified while frozen!"); 892168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 893168404Spjd} 894168404Spjd 895185029Spjdstatic int 896185029Spjdarc_cksum_equal(arc_buf_t *buf) 897185029Spjd{ 898185029Spjd zio_cksum_t zc; 899185029Spjd int equal; 900185029Spjd 901185029Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 902185029Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 903185029Spjd equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 904185029Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 905185029Spjd 906185029Spjd return (equal); 907185029Spjd} 908185029Spjd 909168404Spjdstatic void 910185029Spjdarc_cksum_compute(arc_buf_t *buf, boolean_t force) 911168404Spjd{ 912185029Spjd if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 913168404Spjd return; 914168404Spjd 915168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 916168404Spjd if (buf->b_hdr->b_freeze_cksum != NULL) { 917168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 918168404Spjd return; 919168404Spjd } 920168404Spjd buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 921168404Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 922168404Spjd buf->b_hdr->b_freeze_cksum); 923168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 924168404Spjd} 925168404Spjd 926168404Spjdvoid 927168404Spjdarc_buf_thaw(arc_buf_t *buf) 928168404Spjd{ 929185029Spjd if (zfs_flags & ZFS_DEBUG_MODIFY) { 930185029Spjd if (buf->b_hdr->b_state != arc_anon) 931185029Spjd panic("modifying non-anon buffer!"); 932185029Spjd if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS) 933185029Spjd panic("modifying buffer while i/o in progress!"); 934185029Spjd arc_cksum_verify(buf); 935185029Spjd } 936168404Spjd 937168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 938168404Spjd if (buf->b_hdr->b_freeze_cksum != NULL) { 939168404Spjd kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 940168404Spjd buf->b_hdr->b_freeze_cksum = NULL; 941168404Spjd } 942168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 943168404Spjd} 944168404Spjd 945168404Spjdvoid 946168404Spjdarc_buf_freeze(arc_buf_t *buf) 947168404Spjd{ 948168404Spjd if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 949168404Spjd return; 950168404Spjd 951168404Spjd ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 952168404Spjd buf->b_hdr->b_state == arc_anon); 953185029Spjd arc_cksum_compute(buf, B_FALSE); 954168404Spjd} 955168404Spjd 956168404Spjdstatic void 957168404Spjdadd_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 958168404Spjd{ 959168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 960168404Spjd 961168404Spjd if ((refcount_add(&ab->b_refcnt, tag) == 1) && 962168404Spjd (ab->b_state != arc_anon)) { 963168404Spjd uint64_t delta = ab->b_size * ab->b_datacnt; 964185029Spjd list_t *list = &ab->b_state->arcs_list[ab->b_type]; 965185029Spjd uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type]; 966168404Spjd 967168404Spjd ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx)); 968168404Spjd mutex_enter(&ab->b_state->arcs_mtx); 969168404Spjd ASSERT(list_link_active(&ab->b_arc_node)); 970185029Spjd list_remove(list, ab); 971168404Spjd if (GHOST_STATE(ab->b_state)) { 972168404Spjd ASSERT3U(ab->b_datacnt, ==, 0); 973168404Spjd ASSERT3P(ab->b_buf, ==, NULL); 974168404Spjd delta = ab->b_size; 975168404Spjd } 976168404Spjd ASSERT(delta > 0); 977185029Spjd ASSERT3U(*size, >=, delta); 978185029Spjd atomic_add_64(size, -delta); 979168404Spjd mutex_exit(&ab->b_state->arcs_mtx); 980185029Spjd /* remove the prefetch flag if we get a reference */ 981168404Spjd if (ab->b_flags & ARC_PREFETCH) 982168404Spjd ab->b_flags &= ~ARC_PREFETCH; 983168404Spjd } 984168404Spjd} 985168404Spjd 986168404Spjdstatic int 987168404Spjdremove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 988168404Spjd{ 989168404Spjd int cnt; 990168404Spjd arc_state_t *state = ab->b_state; 991168404Spjd 992168404Spjd ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 993168404Spjd ASSERT(!GHOST_STATE(state)); 994168404Spjd 995168404Spjd if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) && 996168404Spjd (state != arc_anon)) { 997185029Spjd uint64_t *size = &state->arcs_lsize[ab->b_type]; 998185029Spjd 999168404Spjd ASSERT(!MUTEX_HELD(&state->arcs_mtx)); 1000168404Spjd mutex_enter(&state->arcs_mtx); 1001168404Spjd ASSERT(!list_link_active(&ab->b_arc_node)); 1002185029Spjd list_insert_head(&state->arcs_list[ab->b_type], ab); 1003168404Spjd ASSERT(ab->b_datacnt > 0); 1004185029Spjd atomic_add_64(size, ab->b_size * ab->b_datacnt); 1005168404Spjd mutex_exit(&state->arcs_mtx); 1006168404Spjd } 1007168404Spjd return (cnt); 1008168404Spjd} 1009168404Spjd 1010168404Spjd/* 1011168404Spjd * Move the supplied buffer to the indicated state. The mutex 1012168404Spjd * for the buffer must be held by the caller. 1013168404Spjd */ 1014168404Spjdstatic void 1015168404Spjdarc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock) 1016168404Spjd{ 1017168404Spjd arc_state_t *old_state = ab->b_state; 1018168404Spjd int64_t refcnt = refcount_count(&ab->b_refcnt); 1019168404Spjd uint64_t from_delta, to_delta; 1020168404Spjd 1021168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 1022168404Spjd ASSERT(new_state != old_state); 1023168404Spjd ASSERT(refcnt == 0 || ab->b_datacnt > 0); 1024168404Spjd ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state)); 1025168404Spjd 1026168404Spjd from_delta = to_delta = ab->b_datacnt * ab->b_size; 1027168404Spjd 1028168404Spjd /* 1029168404Spjd * If this buffer is evictable, transfer it from the 1030168404Spjd * old state list to the new state list. 1031168404Spjd */ 1032168404Spjd if (refcnt == 0) { 1033168404Spjd if (old_state != arc_anon) { 1034168404Spjd int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx); 1035185029Spjd uint64_t *size = &old_state->arcs_lsize[ab->b_type]; 1036168404Spjd 1037168404Spjd if (use_mutex) 1038168404Spjd mutex_enter(&old_state->arcs_mtx); 1039168404Spjd 1040168404Spjd ASSERT(list_link_active(&ab->b_arc_node)); 1041185029Spjd list_remove(&old_state->arcs_list[ab->b_type], ab); 1042168404Spjd 1043168404Spjd /* 1044168404Spjd * If prefetching out of the ghost cache, 1045168404Spjd * we will have a non-null datacnt. 1046168404Spjd */ 1047168404Spjd if (GHOST_STATE(old_state) && ab->b_datacnt == 0) { 1048168404Spjd /* ghost elements have a ghost size */ 1049168404Spjd ASSERT(ab->b_buf == NULL); 1050168404Spjd from_delta = ab->b_size; 1051168404Spjd } 1052185029Spjd ASSERT3U(*size, >=, from_delta); 1053185029Spjd atomic_add_64(size, -from_delta); 1054168404Spjd 1055168404Spjd if (use_mutex) 1056168404Spjd mutex_exit(&old_state->arcs_mtx); 1057168404Spjd } 1058168404Spjd if (new_state != arc_anon) { 1059168404Spjd int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx); 1060185029Spjd uint64_t *size = &new_state->arcs_lsize[ab->b_type]; 1061168404Spjd 1062168404Spjd if (use_mutex) 1063168404Spjd mutex_enter(&new_state->arcs_mtx); 1064168404Spjd 1065185029Spjd list_insert_head(&new_state->arcs_list[ab->b_type], ab); 1066168404Spjd 1067168404Spjd /* ghost elements have a ghost size */ 1068168404Spjd if (GHOST_STATE(new_state)) { 1069168404Spjd ASSERT(ab->b_datacnt == 0); 1070168404Spjd ASSERT(ab->b_buf == NULL); 1071168404Spjd to_delta = ab->b_size; 1072168404Spjd } 1073185029Spjd atomic_add_64(size, to_delta); 1074168404Spjd 1075168404Spjd if (use_mutex) 1076168404Spjd mutex_exit(&new_state->arcs_mtx); 1077168404Spjd } 1078168404Spjd } 1079168404Spjd 1080168404Spjd ASSERT(!BUF_EMPTY(ab)); 1081185029Spjd if (new_state == arc_anon) { 1082168404Spjd buf_hash_remove(ab); 1083168404Spjd } 1084168404Spjd 1085168404Spjd /* adjust state sizes */ 1086168404Spjd if (to_delta) 1087168404Spjd atomic_add_64(&new_state->arcs_size, to_delta); 1088168404Spjd if (from_delta) { 1089168404Spjd ASSERT3U(old_state->arcs_size, >=, from_delta); 1090168404Spjd atomic_add_64(&old_state->arcs_size, -from_delta); 1091168404Spjd } 1092168404Spjd ab->b_state = new_state; 1093185029Spjd 1094185029Spjd /* adjust l2arc hdr stats */ 1095185029Spjd if (new_state == arc_l2c_only) 1096185029Spjd l2arc_hdr_stat_add(); 1097185029Spjd else if (old_state == arc_l2c_only) 1098185029Spjd l2arc_hdr_stat_remove(); 1099168404Spjd} 1100168404Spjd 1101185029Spjdvoid 1102185029Spjdarc_space_consume(uint64_t space) 1103185029Spjd{ 1104185029Spjd atomic_add_64(&arc_meta_used, space); 1105185029Spjd atomic_add_64(&arc_size, space); 1106185029Spjd} 1107185029Spjd 1108185029Spjdvoid 1109185029Spjdarc_space_return(uint64_t space) 1110185029Spjd{ 1111185029Spjd ASSERT(arc_meta_used >= space); 1112185029Spjd if (arc_meta_max < arc_meta_used) 1113185029Spjd arc_meta_max = arc_meta_used; 1114185029Spjd atomic_add_64(&arc_meta_used, -space); 1115185029Spjd ASSERT(arc_size >= space); 1116185029Spjd atomic_add_64(&arc_size, -space); 1117185029Spjd} 1118185029Spjd 1119185029Spjdvoid * 1120185029Spjdarc_data_buf_alloc(uint64_t size) 1121185029Spjd{ 1122185029Spjd if (arc_evict_needed(ARC_BUFC_DATA)) 1123185029Spjd cv_signal(&arc_reclaim_thr_cv); 1124185029Spjd atomic_add_64(&arc_size, size); 1125185029Spjd return (zio_data_buf_alloc(size)); 1126185029Spjd} 1127185029Spjd 1128185029Spjdvoid 1129185029Spjdarc_data_buf_free(void *buf, uint64_t size) 1130185029Spjd{ 1131185029Spjd zio_data_buf_free(buf, size); 1132185029Spjd ASSERT(arc_size >= size); 1133185029Spjd atomic_add_64(&arc_size, -size); 1134185029Spjd} 1135185029Spjd 1136168404Spjdarc_buf_t * 1137168404Spjdarc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1138168404Spjd{ 1139168404Spjd arc_buf_hdr_t *hdr; 1140168404Spjd arc_buf_t *buf; 1141168404Spjd 1142168404Spjd ASSERT3U(size, >, 0); 1143185029Spjd hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1144168404Spjd ASSERT(BUF_EMPTY(hdr)); 1145168404Spjd hdr->b_size = size; 1146168404Spjd hdr->b_type = type; 1147168404Spjd hdr->b_spa = spa; 1148168404Spjd hdr->b_state = arc_anon; 1149168404Spjd hdr->b_arc_access = 0; 1150185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1151168404Spjd buf->b_hdr = hdr; 1152168404Spjd buf->b_data = NULL; 1153168404Spjd buf->b_efunc = NULL; 1154168404Spjd buf->b_private = NULL; 1155168404Spjd buf->b_next = NULL; 1156168404Spjd hdr->b_buf = buf; 1157168404Spjd arc_get_data_buf(buf); 1158168404Spjd hdr->b_datacnt = 1; 1159168404Spjd hdr->b_flags = 0; 1160168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1161168404Spjd (void) refcount_add(&hdr->b_refcnt, tag); 1162168404Spjd 1163168404Spjd return (buf); 1164168404Spjd} 1165168404Spjd 1166168404Spjdstatic arc_buf_t * 1167168404Spjdarc_buf_clone(arc_buf_t *from) 1168168404Spjd{ 1169168404Spjd arc_buf_t *buf; 1170168404Spjd arc_buf_hdr_t *hdr = from->b_hdr; 1171168404Spjd uint64_t size = hdr->b_size; 1172168404Spjd 1173185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1174168404Spjd buf->b_hdr = hdr; 1175168404Spjd buf->b_data = NULL; 1176168404Spjd buf->b_efunc = NULL; 1177168404Spjd buf->b_private = NULL; 1178168404Spjd buf->b_next = hdr->b_buf; 1179168404Spjd hdr->b_buf = buf; 1180168404Spjd arc_get_data_buf(buf); 1181168404Spjd bcopy(from->b_data, buf->b_data, size); 1182168404Spjd hdr->b_datacnt += 1; 1183168404Spjd return (buf); 1184168404Spjd} 1185168404Spjd 1186168404Spjdvoid 1187168404Spjdarc_buf_add_ref(arc_buf_t *buf, void* tag) 1188168404Spjd{ 1189168404Spjd arc_buf_hdr_t *hdr; 1190168404Spjd kmutex_t *hash_lock; 1191168404Spjd 1192168404Spjd /* 1193185029Spjd * Check to see if this buffer is evicted. Callers 1194185029Spjd * must verify b_data != NULL to know if the add_ref 1195185029Spjd * was successful. 1196168404Spjd */ 1197185029Spjd rw_enter(&buf->b_lock, RW_READER); 1198185029Spjd if (buf->b_data == NULL) { 1199185029Spjd rw_exit(&buf->b_lock); 1200168404Spjd return; 1201168404Spjd } 1202185029Spjd hdr = buf->b_hdr; 1203185029Spjd ASSERT(hdr != NULL); 1204168404Spjd hash_lock = HDR_LOCK(hdr); 1205168404Spjd mutex_enter(hash_lock); 1206185029Spjd rw_exit(&buf->b_lock); 1207168404Spjd 1208168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1209168404Spjd add_reference(hdr, hash_lock, tag); 1210168404Spjd arc_access(hdr, hash_lock); 1211168404Spjd mutex_exit(hash_lock); 1212168404Spjd ARCSTAT_BUMP(arcstat_hits); 1213168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 1214168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1215168404Spjd data, metadata, hits); 1216168404Spjd} 1217168404Spjd 1218185029Spjd/* 1219185029Spjd * Free the arc data buffer. If it is an l2arc write in progress, 1220185029Spjd * the buffer is placed on l2arc_free_on_write to be freed later. 1221185029Spjd */ 1222168404Spjdstatic void 1223185029Spjdarc_buf_data_free(arc_buf_hdr_t *hdr, void (*free_func)(void *, size_t), 1224185029Spjd void *data, size_t size) 1225185029Spjd{ 1226185029Spjd if (HDR_L2_WRITING(hdr)) { 1227185029Spjd l2arc_data_free_t *df; 1228185029Spjd df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1229185029Spjd df->l2df_data = data; 1230185029Spjd df->l2df_size = size; 1231185029Spjd df->l2df_func = free_func; 1232185029Spjd mutex_enter(&l2arc_free_on_write_mtx); 1233185029Spjd list_insert_head(l2arc_free_on_write, df); 1234185029Spjd mutex_exit(&l2arc_free_on_write_mtx); 1235185029Spjd ARCSTAT_BUMP(arcstat_l2_free_on_write); 1236185029Spjd } else { 1237185029Spjd free_func(data, size); 1238185029Spjd } 1239185029Spjd} 1240185029Spjd 1241185029Spjdstatic void 1242168404Spjdarc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all) 1243168404Spjd{ 1244168404Spjd arc_buf_t **bufp; 1245168404Spjd 1246168404Spjd /* free up data associated with the buf */ 1247168404Spjd if (buf->b_data) { 1248168404Spjd arc_state_t *state = buf->b_hdr->b_state; 1249168404Spjd uint64_t size = buf->b_hdr->b_size; 1250168404Spjd arc_buf_contents_t type = buf->b_hdr->b_type; 1251168404Spjd 1252168404Spjd arc_cksum_verify(buf); 1253168404Spjd if (!recycle) { 1254168404Spjd if (type == ARC_BUFC_METADATA) { 1255185029Spjd arc_buf_data_free(buf->b_hdr, zio_buf_free, 1256185029Spjd buf->b_data, size); 1257185029Spjd arc_space_return(size); 1258168404Spjd } else { 1259168404Spjd ASSERT(type == ARC_BUFC_DATA); 1260185029Spjd arc_buf_data_free(buf->b_hdr, 1261185029Spjd zio_data_buf_free, buf->b_data, size); 1262185029Spjd atomic_add_64(&arc_size, -size); 1263168404Spjd } 1264168404Spjd } 1265168404Spjd if (list_link_active(&buf->b_hdr->b_arc_node)) { 1266185029Spjd uint64_t *cnt = &state->arcs_lsize[type]; 1267185029Spjd 1268168404Spjd ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1269168404Spjd ASSERT(state != arc_anon); 1270185029Spjd 1271185029Spjd ASSERT3U(*cnt, >=, size); 1272185029Spjd atomic_add_64(cnt, -size); 1273168404Spjd } 1274168404Spjd ASSERT3U(state->arcs_size, >=, size); 1275168404Spjd atomic_add_64(&state->arcs_size, -size); 1276168404Spjd buf->b_data = NULL; 1277168404Spjd ASSERT(buf->b_hdr->b_datacnt > 0); 1278168404Spjd buf->b_hdr->b_datacnt -= 1; 1279168404Spjd } 1280168404Spjd 1281168404Spjd /* only remove the buf if requested */ 1282168404Spjd if (!all) 1283168404Spjd return; 1284168404Spjd 1285168404Spjd /* remove the buf from the hdr list */ 1286168404Spjd for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1287168404Spjd continue; 1288168404Spjd *bufp = buf->b_next; 1289168404Spjd 1290168404Spjd ASSERT(buf->b_efunc == NULL); 1291168404Spjd 1292168404Spjd /* clean up the buf */ 1293168404Spjd buf->b_hdr = NULL; 1294168404Spjd kmem_cache_free(buf_cache, buf); 1295168404Spjd} 1296168404Spjd 1297168404Spjdstatic void 1298168404Spjdarc_hdr_destroy(arc_buf_hdr_t *hdr) 1299168404Spjd{ 1300168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1301168404Spjd ASSERT3P(hdr->b_state, ==, arc_anon); 1302168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1303185029Spjd ASSERT(!(hdr->b_flags & ARC_STORED)); 1304168404Spjd 1305185029Spjd if (hdr->b_l2hdr != NULL) { 1306185029Spjd if (!MUTEX_HELD(&l2arc_buflist_mtx)) { 1307185029Spjd /* 1308185029Spjd * To prevent arc_free() and l2arc_evict() from 1309185029Spjd * attempting to free the same buffer at the same time, 1310185029Spjd * a FREE_IN_PROGRESS flag is given to arc_free() to 1311185029Spjd * give it priority. l2arc_evict() can't destroy this 1312185029Spjd * header while we are waiting on l2arc_buflist_mtx. 1313185029Spjd * 1314185029Spjd * The hdr may be removed from l2ad_buflist before we 1315185029Spjd * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1316185029Spjd */ 1317185029Spjd mutex_enter(&l2arc_buflist_mtx); 1318185029Spjd if (hdr->b_l2hdr != NULL) { 1319185029Spjd list_remove(hdr->b_l2hdr->b_dev->l2ad_buflist, 1320185029Spjd hdr); 1321185029Spjd } 1322185029Spjd mutex_exit(&l2arc_buflist_mtx); 1323185029Spjd } else { 1324185029Spjd list_remove(hdr->b_l2hdr->b_dev->l2ad_buflist, hdr); 1325185029Spjd } 1326185029Spjd ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1327185029Spjd kmem_free(hdr->b_l2hdr, sizeof (l2arc_buf_hdr_t)); 1328185029Spjd if (hdr->b_state == arc_l2c_only) 1329185029Spjd l2arc_hdr_stat_remove(); 1330185029Spjd hdr->b_l2hdr = NULL; 1331185029Spjd } 1332185029Spjd 1333168404Spjd if (!BUF_EMPTY(hdr)) { 1334168404Spjd ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1335168404Spjd bzero(&hdr->b_dva, sizeof (dva_t)); 1336168404Spjd hdr->b_birth = 0; 1337168404Spjd hdr->b_cksum0 = 0; 1338168404Spjd } 1339168404Spjd while (hdr->b_buf) { 1340168404Spjd arc_buf_t *buf = hdr->b_buf; 1341168404Spjd 1342168404Spjd if (buf->b_efunc) { 1343168404Spjd mutex_enter(&arc_eviction_mtx); 1344185029Spjd rw_enter(&buf->b_lock, RW_WRITER); 1345168404Spjd ASSERT(buf->b_hdr != NULL); 1346168404Spjd arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1347168404Spjd hdr->b_buf = buf->b_next; 1348168404Spjd buf->b_hdr = &arc_eviction_hdr; 1349168404Spjd buf->b_next = arc_eviction_list; 1350168404Spjd arc_eviction_list = buf; 1351185029Spjd rw_exit(&buf->b_lock); 1352168404Spjd mutex_exit(&arc_eviction_mtx); 1353168404Spjd } else { 1354168404Spjd arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1355168404Spjd } 1356168404Spjd } 1357168404Spjd if (hdr->b_freeze_cksum != NULL) { 1358168404Spjd kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1359168404Spjd hdr->b_freeze_cksum = NULL; 1360168404Spjd } 1361168404Spjd 1362168404Spjd ASSERT(!list_link_active(&hdr->b_arc_node)); 1363168404Spjd ASSERT3P(hdr->b_hash_next, ==, NULL); 1364168404Spjd ASSERT3P(hdr->b_acb, ==, NULL); 1365168404Spjd kmem_cache_free(hdr_cache, hdr); 1366168404Spjd} 1367168404Spjd 1368168404Spjdvoid 1369168404Spjdarc_buf_free(arc_buf_t *buf, void *tag) 1370168404Spjd{ 1371168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 1372168404Spjd int hashed = hdr->b_state != arc_anon; 1373168404Spjd 1374168404Spjd ASSERT(buf->b_efunc == NULL); 1375168404Spjd ASSERT(buf->b_data != NULL); 1376168404Spjd 1377168404Spjd if (hashed) { 1378168404Spjd kmutex_t *hash_lock = HDR_LOCK(hdr); 1379168404Spjd 1380168404Spjd mutex_enter(hash_lock); 1381168404Spjd (void) remove_reference(hdr, hash_lock, tag); 1382168404Spjd if (hdr->b_datacnt > 1) 1383168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1384168404Spjd else 1385168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 1386168404Spjd mutex_exit(hash_lock); 1387168404Spjd } else if (HDR_IO_IN_PROGRESS(hdr)) { 1388168404Spjd int destroy_hdr; 1389168404Spjd /* 1390168404Spjd * We are in the middle of an async write. Don't destroy 1391168404Spjd * this buffer unless the write completes before we finish 1392168404Spjd * decrementing the reference count. 1393168404Spjd */ 1394168404Spjd mutex_enter(&arc_eviction_mtx); 1395168404Spjd (void) remove_reference(hdr, NULL, tag); 1396168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1397168404Spjd destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1398168404Spjd mutex_exit(&arc_eviction_mtx); 1399168404Spjd if (destroy_hdr) 1400168404Spjd arc_hdr_destroy(hdr); 1401168404Spjd } else { 1402168404Spjd if (remove_reference(hdr, NULL, tag) > 0) { 1403168404Spjd ASSERT(HDR_IO_ERROR(hdr)); 1404168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1405168404Spjd } else { 1406168404Spjd arc_hdr_destroy(hdr); 1407168404Spjd } 1408168404Spjd } 1409168404Spjd} 1410168404Spjd 1411168404Spjdint 1412168404Spjdarc_buf_remove_ref(arc_buf_t *buf, void* tag) 1413168404Spjd{ 1414168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 1415168404Spjd kmutex_t *hash_lock = HDR_LOCK(hdr); 1416168404Spjd int no_callback = (buf->b_efunc == NULL); 1417168404Spjd 1418168404Spjd if (hdr->b_state == arc_anon) { 1419168404Spjd arc_buf_free(buf, tag); 1420168404Spjd return (no_callback); 1421168404Spjd } 1422168404Spjd 1423168404Spjd mutex_enter(hash_lock); 1424168404Spjd ASSERT(hdr->b_state != arc_anon); 1425168404Spjd ASSERT(buf->b_data != NULL); 1426168404Spjd 1427168404Spjd (void) remove_reference(hdr, hash_lock, tag); 1428168404Spjd if (hdr->b_datacnt > 1) { 1429168404Spjd if (no_callback) 1430168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1431168404Spjd } else if (no_callback) { 1432168404Spjd ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1433168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 1434168404Spjd } 1435168404Spjd ASSERT(no_callback || hdr->b_datacnt > 1 || 1436168404Spjd refcount_is_zero(&hdr->b_refcnt)); 1437168404Spjd mutex_exit(hash_lock); 1438168404Spjd return (no_callback); 1439168404Spjd} 1440168404Spjd 1441168404Spjdint 1442168404Spjdarc_buf_size(arc_buf_t *buf) 1443168404Spjd{ 1444168404Spjd return (buf->b_hdr->b_size); 1445168404Spjd} 1446168404Spjd 1447168404Spjd/* 1448168404Spjd * Evict buffers from list until we've removed the specified number of 1449168404Spjd * bytes. Move the removed buffers to the appropriate evict state. 1450168404Spjd * If the recycle flag is set, then attempt to "recycle" a buffer: 1451168404Spjd * - look for a buffer to evict that is `bytes' long. 1452168404Spjd * - return the data block from this buffer rather than freeing it. 1453168404Spjd * This flag is used by callers that are trying to make space for a 1454168404Spjd * new buffer in a full arc cache. 1455185029Spjd * 1456185029Spjd * This function makes a "best effort". It skips over any buffers 1457185029Spjd * it can't get a hash_lock on, and so may not catch all candidates. 1458185029Spjd * It may also return without evicting as much space as requested. 1459168404Spjd */ 1460168404Spjdstatic void * 1461185029Spjdarc_evict(arc_state_t *state, spa_t *spa, int64_t bytes, boolean_t recycle, 1462168404Spjd arc_buf_contents_t type) 1463168404Spjd{ 1464168404Spjd arc_state_t *evicted_state; 1465168404Spjd uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 1466168404Spjd arc_buf_hdr_t *ab, *ab_prev = NULL; 1467185029Spjd list_t *list = &state->arcs_list[type]; 1468168404Spjd kmutex_t *hash_lock; 1469168404Spjd boolean_t have_lock; 1470168404Spjd void *stolen = NULL; 1471168404Spjd 1472168404Spjd ASSERT(state == arc_mru || state == arc_mfu); 1473168404Spjd 1474168404Spjd evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 1475168404Spjd 1476168404Spjd mutex_enter(&state->arcs_mtx); 1477168404Spjd mutex_enter(&evicted_state->arcs_mtx); 1478168404Spjd 1479185029Spjd for (ab = list_tail(list); ab; ab = ab_prev) { 1480185029Spjd ab_prev = list_prev(list, ab); 1481168404Spjd /* prefetch buffers have a minimum lifespan */ 1482168404Spjd if (HDR_IO_IN_PROGRESS(ab) || 1483185029Spjd (spa && ab->b_spa != spa) || 1484168404Spjd (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 1485174049Sjb LBOLT - ab->b_arc_access < arc_min_prefetch_lifespan)) { 1486168404Spjd skipped++; 1487168404Spjd continue; 1488168404Spjd } 1489168404Spjd /* "lookahead" for better eviction candidate */ 1490168404Spjd if (recycle && ab->b_size != bytes && 1491168404Spjd ab_prev && ab_prev->b_size == bytes) 1492168404Spjd continue; 1493168404Spjd hash_lock = HDR_LOCK(ab); 1494168404Spjd have_lock = MUTEX_HELD(hash_lock); 1495168404Spjd if (have_lock || mutex_tryenter(hash_lock)) { 1496168404Spjd ASSERT3U(refcount_count(&ab->b_refcnt), ==, 0); 1497168404Spjd ASSERT(ab->b_datacnt > 0); 1498168404Spjd while (ab->b_buf) { 1499168404Spjd arc_buf_t *buf = ab->b_buf; 1500185029Spjd if (!rw_tryenter(&buf->b_lock, RW_WRITER)) { 1501185029Spjd missed += 1; 1502185029Spjd break; 1503185029Spjd } 1504168404Spjd if (buf->b_data) { 1505168404Spjd bytes_evicted += ab->b_size; 1506168404Spjd if (recycle && ab->b_type == type && 1507185029Spjd ab->b_size == bytes && 1508185029Spjd !HDR_L2_WRITING(ab)) { 1509168404Spjd stolen = buf->b_data; 1510168404Spjd recycle = FALSE; 1511168404Spjd } 1512168404Spjd } 1513168404Spjd if (buf->b_efunc) { 1514168404Spjd mutex_enter(&arc_eviction_mtx); 1515168404Spjd arc_buf_destroy(buf, 1516168404Spjd buf->b_data == stolen, FALSE); 1517168404Spjd ab->b_buf = buf->b_next; 1518168404Spjd buf->b_hdr = &arc_eviction_hdr; 1519168404Spjd buf->b_next = arc_eviction_list; 1520168404Spjd arc_eviction_list = buf; 1521168404Spjd mutex_exit(&arc_eviction_mtx); 1522185029Spjd rw_exit(&buf->b_lock); 1523168404Spjd } else { 1524185029Spjd rw_exit(&buf->b_lock); 1525168404Spjd arc_buf_destroy(buf, 1526168404Spjd buf->b_data == stolen, TRUE); 1527168404Spjd } 1528168404Spjd } 1529185029Spjd if (ab->b_datacnt == 0) { 1530185029Spjd arc_change_state(evicted_state, ab, hash_lock); 1531185029Spjd ASSERT(HDR_IN_HASH_TABLE(ab)); 1532185029Spjd ab->b_flags |= ARC_IN_HASH_TABLE; 1533185029Spjd ab->b_flags &= ~ARC_BUF_AVAILABLE; 1534185029Spjd DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 1535185029Spjd } 1536168404Spjd if (!have_lock) 1537168404Spjd mutex_exit(hash_lock); 1538168404Spjd if (bytes >= 0 && bytes_evicted >= bytes) 1539168404Spjd break; 1540168404Spjd } else { 1541168404Spjd missed += 1; 1542168404Spjd } 1543168404Spjd } 1544168404Spjd 1545168404Spjd mutex_exit(&evicted_state->arcs_mtx); 1546168404Spjd mutex_exit(&state->arcs_mtx); 1547168404Spjd 1548168404Spjd if (bytes_evicted < bytes) 1549168404Spjd dprintf("only evicted %lld bytes from %x", 1550168404Spjd (longlong_t)bytes_evicted, state); 1551168404Spjd 1552168404Spjd if (skipped) 1553168404Spjd ARCSTAT_INCR(arcstat_evict_skip, skipped); 1554168404Spjd 1555168404Spjd if (missed) 1556168404Spjd ARCSTAT_INCR(arcstat_mutex_miss, missed); 1557168404Spjd 1558185029Spjd /* 1559185029Spjd * We have just evicted some date into the ghost state, make 1560185029Spjd * sure we also adjust the ghost state size if necessary. 1561185029Spjd */ 1562185029Spjd if (arc_no_grow && 1563185029Spjd arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) { 1564185029Spjd int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size + 1565185029Spjd arc_mru_ghost->arcs_size - arc_c; 1566185029Spjd 1567185029Spjd if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) { 1568185029Spjd int64_t todelete = 1569185029Spjd MIN(arc_mru_ghost->arcs_lsize[type], mru_over); 1570185029Spjd arc_evict_ghost(arc_mru_ghost, NULL, todelete); 1571185029Spjd } else if (arc_mfu_ghost->arcs_lsize[type] > 0) { 1572185029Spjd int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type], 1573185029Spjd arc_mru_ghost->arcs_size + 1574185029Spjd arc_mfu_ghost->arcs_size - arc_c); 1575185029Spjd arc_evict_ghost(arc_mfu_ghost, NULL, todelete); 1576185029Spjd } 1577185029Spjd } 1578185029Spjd 1579168404Spjd return (stolen); 1580168404Spjd} 1581168404Spjd 1582168404Spjd/* 1583168404Spjd * Remove buffers from list until we've removed the specified number of 1584168404Spjd * bytes. Destroy the buffers that are removed. 1585168404Spjd */ 1586168404Spjdstatic void 1587185029Spjdarc_evict_ghost(arc_state_t *state, spa_t *spa, int64_t bytes) 1588168404Spjd{ 1589168404Spjd arc_buf_hdr_t *ab, *ab_prev; 1590185029Spjd list_t *list = &state->arcs_list[ARC_BUFC_DATA]; 1591168404Spjd kmutex_t *hash_lock; 1592168404Spjd uint64_t bytes_deleted = 0; 1593168404Spjd uint64_t bufs_skipped = 0; 1594168404Spjd 1595168404Spjd ASSERT(GHOST_STATE(state)); 1596168404Spjdtop: 1597168404Spjd mutex_enter(&state->arcs_mtx); 1598185029Spjd for (ab = list_tail(list); ab; ab = ab_prev) { 1599185029Spjd ab_prev = list_prev(list, ab); 1600185029Spjd if (spa && ab->b_spa != spa) 1601185029Spjd continue; 1602168404Spjd hash_lock = HDR_LOCK(ab); 1603168404Spjd if (mutex_tryenter(hash_lock)) { 1604168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(ab)); 1605168404Spjd ASSERT(ab->b_buf == NULL); 1606168404Spjd ARCSTAT_BUMP(arcstat_deleted); 1607168404Spjd bytes_deleted += ab->b_size; 1608185029Spjd 1609185029Spjd if (ab->b_l2hdr != NULL) { 1610185029Spjd /* 1611185029Spjd * This buffer is cached on the 2nd Level ARC; 1612185029Spjd * don't destroy the header. 1613185029Spjd */ 1614185029Spjd arc_change_state(arc_l2c_only, ab, hash_lock); 1615185029Spjd mutex_exit(hash_lock); 1616185029Spjd } else { 1617185029Spjd arc_change_state(arc_anon, ab, hash_lock); 1618185029Spjd mutex_exit(hash_lock); 1619185029Spjd arc_hdr_destroy(ab); 1620185029Spjd } 1621185029Spjd 1622168404Spjd DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 1623168404Spjd if (bytes >= 0 && bytes_deleted >= bytes) 1624168404Spjd break; 1625168404Spjd } else { 1626168404Spjd if (bytes < 0) { 1627168404Spjd mutex_exit(&state->arcs_mtx); 1628168404Spjd mutex_enter(hash_lock); 1629168404Spjd mutex_exit(hash_lock); 1630168404Spjd goto top; 1631168404Spjd } 1632168404Spjd bufs_skipped += 1; 1633168404Spjd } 1634168404Spjd } 1635168404Spjd mutex_exit(&state->arcs_mtx); 1636168404Spjd 1637185029Spjd if (list == &state->arcs_list[ARC_BUFC_DATA] && 1638185029Spjd (bytes < 0 || bytes_deleted < bytes)) { 1639185029Spjd list = &state->arcs_list[ARC_BUFC_METADATA]; 1640185029Spjd goto top; 1641185029Spjd } 1642185029Spjd 1643168404Spjd if (bufs_skipped) { 1644168404Spjd ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 1645168404Spjd ASSERT(bytes >= 0); 1646168404Spjd } 1647168404Spjd 1648168404Spjd if (bytes_deleted < bytes) 1649168404Spjd dprintf("only deleted %lld bytes from %p", 1650168404Spjd (longlong_t)bytes_deleted, state); 1651168404Spjd} 1652168404Spjd 1653168404Spjdstatic void 1654168404Spjdarc_adjust(void) 1655168404Spjd{ 1656168404Spjd int64_t top_sz, mru_over, arc_over, todelete; 1657168404Spjd 1658185029Spjd top_sz = arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used; 1659168404Spjd 1660185029Spjd if (top_sz > arc_p && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 1661185029Spjd int64_t toevict = 1662185029Spjd MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], top_sz - arc_p); 1663185029Spjd (void) arc_evict(arc_mru, NULL, toevict, FALSE, ARC_BUFC_DATA); 1664168404Spjd top_sz = arc_anon->arcs_size + arc_mru->arcs_size; 1665168404Spjd } 1666168404Spjd 1667185029Spjd if (top_sz > arc_p && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 1668185029Spjd int64_t toevict = 1669185029Spjd MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], top_sz - arc_p); 1670185029Spjd (void) arc_evict(arc_mru, NULL, toevict, FALSE, 1671185029Spjd ARC_BUFC_METADATA); 1672185029Spjd top_sz = arc_anon->arcs_size + arc_mru->arcs_size; 1673185029Spjd } 1674185029Spjd 1675168404Spjd mru_over = top_sz + arc_mru_ghost->arcs_size - arc_c; 1676168404Spjd 1677168404Spjd if (mru_over > 0) { 1678185029Spjd if (arc_mru_ghost->arcs_size > 0) { 1679185029Spjd todelete = MIN(arc_mru_ghost->arcs_size, mru_over); 1680185029Spjd arc_evict_ghost(arc_mru_ghost, NULL, todelete); 1681168404Spjd } 1682168404Spjd } 1683168404Spjd 1684168404Spjd if ((arc_over = arc_size - arc_c) > 0) { 1685168404Spjd int64_t tbl_over; 1686168404Spjd 1687185029Spjd if (arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 1688185029Spjd int64_t toevict = 1689185029Spjd MIN(arc_mfu->arcs_lsize[ARC_BUFC_DATA], arc_over); 1690185029Spjd (void) arc_evict(arc_mfu, NULL, toevict, FALSE, 1691185029Spjd ARC_BUFC_DATA); 1692185029Spjd arc_over = arc_size - arc_c; 1693168404Spjd } 1694168404Spjd 1695185029Spjd if (arc_over > 0 && 1696185029Spjd arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 1697185029Spjd int64_t toevict = 1698185029Spjd MIN(arc_mfu->arcs_lsize[ARC_BUFC_METADATA], 1699185029Spjd arc_over); 1700185029Spjd (void) arc_evict(arc_mfu, NULL, toevict, FALSE, 1701185029Spjd ARC_BUFC_METADATA); 1702185029Spjd } 1703168404Spjd 1704185029Spjd tbl_over = arc_size + arc_mru_ghost->arcs_size + 1705185029Spjd arc_mfu_ghost->arcs_size - arc_c * 2; 1706185029Spjd 1707185029Spjd if (tbl_over > 0 && arc_mfu_ghost->arcs_size > 0) { 1708185029Spjd todelete = MIN(arc_mfu_ghost->arcs_size, tbl_over); 1709185029Spjd arc_evict_ghost(arc_mfu_ghost, NULL, todelete); 1710168404Spjd } 1711168404Spjd } 1712168404Spjd} 1713168404Spjd 1714168404Spjdstatic void 1715168404Spjdarc_do_user_evicts(void) 1716168404Spjd{ 1717191903Skmacy static arc_buf_t *tmp_arc_eviction_list; 1718191903Skmacy 1719191903Skmacy /* 1720191903Skmacy * Move list over to avoid LOR 1721191903Skmacy */ 1722191903Skmacyrestart: 1723168404Spjd mutex_enter(&arc_eviction_mtx); 1724191903Skmacy tmp_arc_eviction_list = arc_eviction_list; 1725191903Skmacy arc_eviction_list = NULL; 1726191903Skmacy mutex_exit(&arc_eviction_mtx); 1727191903Skmacy 1728191903Skmacy while (tmp_arc_eviction_list != NULL) { 1729191903Skmacy arc_buf_t *buf = tmp_arc_eviction_list; 1730191903Skmacy tmp_arc_eviction_list = buf->b_next; 1731185029Spjd rw_enter(&buf->b_lock, RW_WRITER); 1732168404Spjd buf->b_hdr = NULL; 1733185029Spjd rw_exit(&buf->b_lock); 1734168404Spjd 1735168404Spjd if (buf->b_efunc != NULL) 1736168404Spjd VERIFY(buf->b_efunc(buf) == 0); 1737168404Spjd 1738168404Spjd buf->b_efunc = NULL; 1739168404Spjd buf->b_private = NULL; 1740168404Spjd kmem_cache_free(buf_cache, buf); 1741168404Spjd } 1742191903Skmacy 1743191903Skmacy if (arc_eviction_list != NULL) 1744191903Skmacy goto restart; 1745168404Spjd} 1746168404Spjd 1747168404Spjd/* 1748185029Spjd * Flush all *evictable* data from the cache for the given spa. 1749168404Spjd * NOTE: this will not touch "active" (i.e. referenced) data. 1750168404Spjd */ 1751168404Spjdvoid 1752185029Spjdarc_flush(spa_t *spa) 1753168404Spjd{ 1754185029Spjd while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) { 1755185029Spjd (void) arc_evict(arc_mru, spa, -1, FALSE, ARC_BUFC_DATA); 1756185029Spjd if (spa) 1757185029Spjd break; 1758185029Spjd } 1759185029Spjd while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) { 1760185029Spjd (void) arc_evict(arc_mru, spa, -1, FALSE, ARC_BUFC_METADATA); 1761185029Spjd if (spa) 1762185029Spjd break; 1763185029Spjd } 1764185029Spjd while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) { 1765185029Spjd (void) arc_evict(arc_mfu, spa, -1, FALSE, ARC_BUFC_DATA); 1766185029Spjd if (spa) 1767185029Spjd break; 1768185029Spjd } 1769185029Spjd while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) { 1770185029Spjd (void) arc_evict(arc_mfu, spa, -1, FALSE, ARC_BUFC_METADATA); 1771185029Spjd if (spa) 1772185029Spjd break; 1773185029Spjd } 1774168404Spjd 1775185029Spjd arc_evict_ghost(arc_mru_ghost, spa, -1); 1776185029Spjd arc_evict_ghost(arc_mfu_ghost, spa, -1); 1777168404Spjd 1778168404Spjd mutex_enter(&arc_reclaim_thr_lock); 1779168404Spjd arc_do_user_evicts(); 1780168404Spjd mutex_exit(&arc_reclaim_thr_lock); 1781185029Spjd ASSERT(spa || arc_eviction_list == NULL); 1782168404Spjd} 1783168404Spjd 1784168404Spjdint arc_shrink_shift = 5; /* log2(fraction of arc to reclaim) */ 1785168404Spjd 1786168404Spjdvoid 1787168404Spjdarc_shrink(void) 1788168404Spjd{ 1789168404Spjd if (arc_c > arc_c_min) { 1790168404Spjd uint64_t to_free; 1791168404Spjd 1792168404Spjd#ifdef _KERNEL 1793168404Spjd to_free = arc_c >> arc_shrink_shift; 1794168404Spjd#else 1795168404Spjd to_free = arc_c >> arc_shrink_shift; 1796168404Spjd#endif 1797168404Spjd if (arc_c > arc_c_min + to_free) 1798168404Spjd atomic_add_64(&arc_c, -to_free); 1799168404Spjd else 1800168404Spjd arc_c = arc_c_min; 1801168404Spjd 1802168404Spjd atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 1803168404Spjd if (arc_c > arc_size) 1804168404Spjd arc_c = MAX(arc_size, arc_c_min); 1805168404Spjd if (arc_p > arc_c) 1806168404Spjd arc_p = (arc_c >> 1); 1807168404Spjd ASSERT(arc_c >= arc_c_min); 1808168404Spjd ASSERT((int64_t)arc_p >= 0); 1809168404Spjd } 1810168404Spjd 1811168404Spjd if (arc_size > arc_c) 1812168404Spjd arc_adjust(); 1813168404Spjd} 1814168404Spjd 1815185029Spjdstatic int needfree = 0; 1816168404Spjd 1817168404Spjdstatic int 1818168404Spjdarc_reclaim_needed(void) 1819168404Spjd{ 1820168404Spjd#if 0 1821168404Spjd uint64_t extra; 1822168404Spjd#endif 1823168404Spjd 1824168404Spjd#ifdef _KERNEL 1825168404Spjd 1826191902Skmacy /* 1827191902Skmacy * If pages are needed or we're within 2048 pages 1828191902Skmacy * of needing to page need to reclaim 1829191902Skmacy */ 1830191902Skmacy if (vm_pages_needed || (vm_paging_target() > -2048)) 1831191902Skmacy return (1); 1832191902Skmacy 1833185029Spjd if (needfree) 1834168404Spjd return (1); 1835168404Spjd 1836168404Spjd#if 0 1837168404Spjd /* 1838185029Spjd * take 'desfree' extra pages, so we reclaim sooner, rather than later 1839185029Spjd */ 1840185029Spjd extra = desfree; 1841185029Spjd 1842185029Spjd /* 1843185029Spjd * check that we're out of range of the pageout scanner. It starts to 1844185029Spjd * schedule paging if freemem is less than lotsfree and needfree. 1845185029Spjd * lotsfree is the high-water mark for pageout, and needfree is the 1846185029Spjd * number of needed free pages. We add extra pages here to make sure 1847185029Spjd * the scanner doesn't start up while we're freeing memory. 1848185029Spjd */ 1849185029Spjd if (freemem < lotsfree + needfree + extra) 1850185029Spjd return (1); 1851185029Spjd 1852185029Spjd /* 1853168404Spjd * check to make sure that swapfs has enough space so that anon 1854185029Spjd * reservations can still succeed. anon_resvmem() checks that the 1855168404Spjd * availrmem is greater than swapfs_minfree, and the number of reserved 1856168404Spjd * swap pages. We also add a bit of extra here just to prevent 1857168404Spjd * circumstances from getting really dire. 1858168404Spjd */ 1859168404Spjd if (availrmem < swapfs_minfree + swapfs_reserve + extra) 1860168404Spjd return (1); 1861168404Spjd 1862168404Spjd#if defined(__i386) 1863168404Spjd /* 1864168404Spjd * If we're on an i386 platform, it's possible that we'll exhaust the 1865168404Spjd * kernel heap space before we ever run out of available physical 1866168404Spjd * memory. Most checks of the size of the heap_area compare against 1867168404Spjd * tune.t_minarmem, which is the minimum available real memory that we 1868168404Spjd * can have in the system. However, this is generally fixed at 25 pages 1869168404Spjd * which is so low that it's useless. In this comparison, we seek to 1870168404Spjd * calculate the total heap-size, and reclaim if more than 3/4ths of the 1871185029Spjd * heap is allocated. (Or, in the calculation, if less than 1/4th is 1872168404Spjd * free) 1873168404Spjd */ 1874168404Spjd if (btop(vmem_size(heap_arena, VMEM_FREE)) < 1875168404Spjd (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2)) 1876168404Spjd return (1); 1877168404Spjd#endif 1878168404Spjd#else 1879175633Spjd if (kmem_used() > (kmem_size() * 3) / 4) 1880168404Spjd return (1); 1881168404Spjd#endif 1882168404Spjd 1883168404Spjd#else 1884168404Spjd if (spa_get_random(100) == 0) 1885168404Spjd return (1); 1886168404Spjd#endif 1887168404Spjd return (0); 1888168404Spjd} 1889168404Spjd 1890168404Spjdstatic void 1891168404Spjdarc_kmem_reap_now(arc_reclaim_strategy_t strat) 1892168404Spjd{ 1893168404Spjd#ifdef ZIO_USE_UMA 1894168404Spjd size_t i; 1895168404Spjd kmem_cache_t *prev_cache = NULL; 1896168404Spjd kmem_cache_t *prev_data_cache = NULL; 1897168404Spjd extern kmem_cache_t *zio_buf_cache[]; 1898168404Spjd extern kmem_cache_t *zio_data_buf_cache[]; 1899168404Spjd#endif 1900168404Spjd 1901168404Spjd#ifdef _KERNEL 1902185029Spjd if (arc_meta_used >= arc_meta_limit) { 1903185029Spjd /* 1904185029Spjd * We are exceeding our meta-data cache limit. 1905185029Spjd * Purge some DNLC entries to release holds on meta-data. 1906185029Spjd */ 1907185029Spjd dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 1908185029Spjd } 1909168404Spjd#if defined(__i386) 1910168404Spjd /* 1911168404Spjd * Reclaim unused memory from all kmem caches. 1912168404Spjd */ 1913168404Spjd kmem_reap(); 1914168404Spjd#endif 1915168404Spjd#endif 1916168404Spjd 1917168404Spjd /* 1918185029Spjd * An aggressive reclamation will shrink the cache size as well as 1919168404Spjd * reap free buffers from the arc kmem caches. 1920168404Spjd */ 1921168404Spjd if (strat == ARC_RECLAIM_AGGR) 1922168404Spjd arc_shrink(); 1923168404Spjd 1924168404Spjd#ifdef ZIO_USE_UMA 1925168404Spjd for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 1926168404Spjd if (zio_buf_cache[i] != prev_cache) { 1927168404Spjd prev_cache = zio_buf_cache[i]; 1928168404Spjd kmem_cache_reap_now(zio_buf_cache[i]); 1929168404Spjd } 1930168404Spjd if (zio_data_buf_cache[i] != prev_data_cache) { 1931168404Spjd prev_data_cache = zio_data_buf_cache[i]; 1932168404Spjd kmem_cache_reap_now(zio_data_buf_cache[i]); 1933168404Spjd } 1934168404Spjd } 1935168404Spjd#endif 1936168404Spjd kmem_cache_reap_now(buf_cache); 1937168404Spjd kmem_cache_reap_now(hdr_cache); 1938168404Spjd} 1939168404Spjd 1940168404Spjdstatic void 1941168404Spjdarc_reclaim_thread(void *dummy __unused) 1942168404Spjd{ 1943168404Spjd clock_t growtime = 0; 1944168404Spjd arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 1945168404Spjd callb_cpr_t cpr; 1946168404Spjd 1947168404Spjd CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 1948168404Spjd 1949168404Spjd mutex_enter(&arc_reclaim_thr_lock); 1950168404Spjd while (arc_thread_exit == 0) { 1951168404Spjd if (arc_reclaim_needed()) { 1952168404Spjd 1953168404Spjd if (arc_no_grow) { 1954168404Spjd if (last_reclaim == ARC_RECLAIM_CONS) { 1955168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 1956168404Spjd } else { 1957168404Spjd last_reclaim = ARC_RECLAIM_CONS; 1958168404Spjd } 1959168404Spjd } else { 1960168404Spjd arc_no_grow = TRUE; 1961168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 1962168404Spjd membar_producer(); 1963168404Spjd } 1964168404Spjd 1965168404Spjd /* reset the growth delay for every reclaim */ 1966174049Sjb growtime = LBOLT + (arc_grow_retry * hz); 1967168404Spjd 1968185029Spjd if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 1969168404Spjd /* 1970185029Spjd * If needfree is TRUE our vm_lowmem hook 1971168404Spjd * was called and in that case we must free some 1972168404Spjd * memory, so switch to aggressive mode. 1973168404Spjd */ 1974168404Spjd arc_no_grow = TRUE; 1975168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 1976168404Spjd } 1977168404Spjd arc_kmem_reap_now(last_reclaim); 1978185029Spjd arc_warm = B_TRUE; 1979185029Spjd 1980185029Spjd } else if (arc_no_grow && LBOLT >= growtime) { 1981168404Spjd arc_no_grow = FALSE; 1982168404Spjd } 1983168404Spjd 1984185029Spjd if (needfree || 1985168404Spjd (2 * arc_c < arc_size + 1986168404Spjd arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size)) 1987168404Spjd arc_adjust(); 1988168404Spjd 1989168404Spjd if (arc_eviction_list != NULL) 1990168404Spjd arc_do_user_evicts(); 1991168404Spjd 1992168404Spjd if (arc_reclaim_needed()) { 1993185029Spjd needfree = 0; 1994168404Spjd#ifdef _KERNEL 1995185029Spjd wakeup(&needfree); 1996168404Spjd#endif 1997168404Spjd } 1998168404Spjd 1999168404Spjd /* block until needed, or one second, whichever is shorter */ 2000168404Spjd CALLB_CPR_SAFE_BEGIN(&cpr); 2001168404Spjd (void) cv_timedwait(&arc_reclaim_thr_cv, 2002168404Spjd &arc_reclaim_thr_lock, hz); 2003168404Spjd CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2004168404Spjd } 2005168404Spjd 2006168404Spjd arc_thread_exit = 0; 2007168404Spjd cv_broadcast(&arc_reclaim_thr_cv); 2008168404Spjd CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2009168404Spjd thread_exit(); 2010168404Spjd} 2011168404Spjd 2012168404Spjd/* 2013168404Spjd * Adapt arc info given the number of bytes we are trying to add and 2014168404Spjd * the state that we are comming from. This function is only called 2015168404Spjd * when we are adding new content to the cache. 2016168404Spjd */ 2017168404Spjdstatic void 2018168404Spjdarc_adapt(int bytes, arc_state_t *state) 2019168404Spjd{ 2020168404Spjd int mult; 2021168404Spjd 2022185029Spjd if (state == arc_l2c_only) 2023185029Spjd return; 2024185029Spjd 2025168404Spjd ASSERT(bytes > 0); 2026168404Spjd /* 2027168404Spjd * Adapt the target size of the MRU list: 2028168404Spjd * - if we just hit in the MRU ghost list, then increase 2029168404Spjd * the target size of the MRU list. 2030168404Spjd * - if we just hit in the MFU ghost list, then increase 2031168404Spjd * the target size of the MFU list by decreasing the 2032168404Spjd * target size of the MRU list. 2033168404Spjd */ 2034168404Spjd if (state == arc_mru_ghost) { 2035168404Spjd mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2036168404Spjd 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2037168404Spjd 2038168404Spjd arc_p = MIN(arc_c, arc_p + bytes * mult); 2039168404Spjd } else if (state == arc_mfu_ghost) { 2040168404Spjd mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2041168404Spjd 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2042168404Spjd 2043168404Spjd arc_p = MAX(0, (int64_t)arc_p - bytes * mult); 2044168404Spjd } 2045168404Spjd ASSERT((int64_t)arc_p >= 0); 2046168404Spjd 2047168404Spjd if (arc_reclaim_needed()) { 2048168404Spjd cv_signal(&arc_reclaim_thr_cv); 2049168404Spjd return; 2050168404Spjd } 2051168404Spjd 2052168404Spjd if (arc_no_grow) 2053168404Spjd return; 2054168404Spjd 2055168404Spjd if (arc_c >= arc_c_max) 2056168404Spjd return; 2057168404Spjd 2058168404Spjd /* 2059168404Spjd * If we're within (2 * maxblocksize) bytes of the target 2060168404Spjd * cache size, increment the target cache size 2061168404Spjd */ 2062168404Spjd if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2063168404Spjd atomic_add_64(&arc_c, (int64_t)bytes); 2064168404Spjd if (arc_c > arc_c_max) 2065168404Spjd arc_c = arc_c_max; 2066168404Spjd else if (state == arc_anon) 2067168404Spjd atomic_add_64(&arc_p, (int64_t)bytes); 2068168404Spjd if (arc_p > arc_c) 2069168404Spjd arc_p = arc_c; 2070168404Spjd } 2071168404Spjd ASSERT((int64_t)arc_p >= 0); 2072168404Spjd} 2073168404Spjd 2074168404Spjd/* 2075168404Spjd * Check if the cache has reached its limits and eviction is required 2076168404Spjd * prior to insert. 2077168404Spjd */ 2078168404Spjdstatic int 2079185029Spjdarc_evict_needed(arc_buf_contents_t type) 2080168404Spjd{ 2081185029Spjd if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2082185029Spjd return (1); 2083185029Spjd 2084185029Spjd#if 0 2085185029Spjd#ifdef _KERNEL 2086185029Spjd /* 2087185029Spjd * If zio data pages are being allocated out of a separate heap segment, 2088185029Spjd * then enforce that the size of available vmem for this area remains 2089185029Spjd * above about 1/32nd free. 2090185029Spjd */ 2091185029Spjd if (type == ARC_BUFC_DATA && zio_arena != NULL && 2092185029Spjd vmem_size(zio_arena, VMEM_FREE) < 2093185029Spjd (vmem_size(zio_arena, VMEM_ALLOC) >> 5)) 2094185029Spjd return (1); 2095185029Spjd#endif 2096185029Spjd#endif 2097185029Spjd 2098168404Spjd if (arc_reclaim_needed()) 2099168404Spjd return (1); 2100168404Spjd 2101168404Spjd return (arc_size > arc_c); 2102168404Spjd} 2103168404Spjd 2104168404Spjd/* 2105168404Spjd * The buffer, supplied as the first argument, needs a data block. 2106168404Spjd * So, if we are at cache max, determine which cache should be victimized. 2107168404Spjd * We have the following cases: 2108168404Spjd * 2109168404Spjd * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2110168404Spjd * In this situation if we're out of space, but the resident size of the MFU is 2111168404Spjd * under the limit, victimize the MFU cache to satisfy this insertion request. 2112168404Spjd * 2113168404Spjd * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2114168404Spjd * Here, we've used up all of the available space for the MRU, so we need to 2115168404Spjd * evict from our own cache instead. Evict from the set of resident MRU 2116168404Spjd * entries. 2117168404Spjd * 2118168404Spjd * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2119168404Spjd * c minus p represents the MFU space in the cache, since p is the size of the 2120168404Spjd * cache that is dedicated to the MRU. In this situation there's still space on 2121168404Spjd * the MFU side, so the MRU side needs to be victimized. 2122168404Spjd * 2123168404Spjd * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2124168404Spjd * MFU's resident set is consuming more space than it has been allotted. In 2125168404Spjd * this situation, we must victimize our own cache, the MFU, for this insertion. 2126168404Spjd */ 2127168404Spjdstatic void 2128168404Spjdarc_get_data_buf(arc_buf_t *buf) 2129168404Spjd{ 2130168404Spjd arc_state_t *state = buf->b_hdr->b_state; 2131168404Spjd uint64_t size = buf->b_hdr->b_size; 2132168404Spjd arc_buf_contents_t type = buf->b_hdr->b_type; 2133168404Spjd 2134168404Spjd arc_adapt(size, state); 2135168404Spjd 2136168404Spjd /* 2137168404Spjd * We have not yet reached cache maximum size, 2138168404Spjd * just allocate a new buffer. 2139168404Spjd */ 2140185029Spjd if (!arc_evict_needed(type)) { 2141168404Spjd if (type == ARC_BUFC_METADATA) { 2142168404Spjd buf->b_data = zio_buf_alloc(size); 2143185029Spjd arc_space_consume(size); 2144168404Spjd } else { 2145168404Spjd ASSERT(type == ARC_BUFC_DATA); 2146168404Spjd buf->b_data = zio_data_buf_alloc(size); 2147185029Spjd atomic_add_64(&arc_size, size); 2148168404Spjd } 2149168404Spjd goto out; 2150168404Spjd } 2151168404Spjd 2152168404Spjd /* 2153168404Spjd * If we are prefetching from the mfu ghost list, this buffer 2154168404Spjd * will end up on the mru list; so steal space from there. 2155168404Spjd */ 2156168404Spjd if (state == arc_mfu_ghost) 2157168404Spjd state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2158168404Spjd else if (state == arc_mru_ghost) 2159168404Spjd state = arc_mru; 2160168404Spjd 2161168404Spjd if (state == arc_mru || state == arc_anon) { 2162168404Spjd uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2163185029Spjd state = (arc_mfu->arcs_lsize[type] > 0 && 2164185029Spjd arc_p > mru_used) ? arc_mfu : arc_mru; 2165168404Spjd } else { 2166168404Spjd /* MFU cases */ 2167168404Spjd uint64_t mfu_space = arc_c - arc_p; 2168185029Spjd state = (arc_mru->arcs_lsize[type] > 0 && 2169185029Spjd mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2170168404Spjd } 2171185029Spjd if ((buf->b_data = arc_evict(state, NULL, size, TRUE, type)) == NULL) { 2172168404Spjd if (type == ARC_BUFC_METADATA) { 2173168404Spjd buf->b_data = zio_buf_alloc(size); 2174185029Spjd arc_space_consume(size); 2175168404Spjd } else { 2176168404Spjd ASSERT(type == ARC_BUFC_DATA); 2177168404Spjd buf->b_data = zio_data_buf_alloc(size); 2178185029Spjd atomic_add_64(&arc_size, size); 2179168404Spjd } 2180168404Spjd ARCSTAT_BUMP(arcstat_recycle_miss); 2181168404Spjd } 2182168404Spjd ASSERT(buf->b_data != NULL); 2183168404Spjdout: 2184168404Spjd /* 2185168404Spjd * Update the state size. Note that ghost states have a 2186168404Spjd * "ghost size" and so don't need to be updated. 2187168404Spjd */ 2188168404Spjd if (!GHOST_STATE(buf->b_hdr->b_state)) { 2189168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 2190168404Spjd 2191168404Spjd atomic_add_64(&hdr->b_state->arcs_size, size); 2192168404Spjd if (list_link_active(&hdr->b_arc_node)) { 2193168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2194185029Spjd atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2195168404Spjd } 2196168404Spjd /* 2197168404Spjd * If we are growing the cache, and we are adding anonymous 2198168404Spjd * data, and we have outgrown arc_p, update arc_p 2199168404Spjd */ 2200168404Spjd if (arc_size < arc_c && hdr->b_state == arc_anon && 2201168404Spjd arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2202168404Spjd arc_p = MIN(arc_c, arc_p + size); 2203168404Spjd } 2204168404Spjd} 2205168404Spjd 2206168404Spjd/* 2207168404Spjd * This routine is called whenever a buffer is accessed. 2208168404Spjd * NOTE: the hash lock is dropped in this function. 2209168404Spjd */ 2210168404Spjdstatic void 2211168404Spjdarc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2212168404Spjd{ 2213168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 2214168404Spjd 2215168404Spjd if (buf->b_state == arc_anon) { 2216168404Spjd /* 2217168404Spjd * This buffer is not in the cache, and does not 2218168404Spjd * appear in our "ghost" list. Add the new buffer 2219168404Spjd * to the MRU state. 2220168404Spjd */ 2221168404Spjd 2222168404Spjd ASSERT(buf->b_arc_access == 0); 2223174049Sjb buf->b_arc_access = LBOLT; 2224168404Spjd DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2225168404Spjd arc_change_state(arc_mru, buf, hash_lock); 2226168404Spjd 2227168404Spjd } else if (buf->b_state == arc_mru) { 2228168404Spjd /* 2229168404Spjd * If this buffer is here because of a prefetch, then either: 2230168404Spjd * - clear the flag if this is a "referencing" read 2231168404Spjd * (any subsequent access will bump this into the MFU state). 2232168404Spjd * or 2233168404Spjd * - move the buffer to the head of the list if this is 2234168404Spjd * another prefetch (to make it less likely to be evicted). 2235168404Spjd */ 2236168404Spjd if ((buf->b_flags & ARC_PREFETCH) != 0) { 2237168404Spjd if (refcount_count(&buf->b_refcnt) == 0) { 2238168404Spjd ASSERT(list_link_active(&buf->b_arc_node)); 2239168404Spjd } else { 2240168404Spjd buf->b_flags &= ~ARC_PREFETCH; 2241168404Spjd ARCSTAT_BUMP(arcstat_mru_hits); 2242168404Spjd } 2243174049Sjb buf->b_arc_access = LBOLT; 2244168404Spjd return; 2245168404Spjd } 2246168404Spjd 2247168404Spjd /* 2248168404Spjd * This buffer has been "accessed" only once so far, 2249168404Spjd * but it is still in the cache. Move it to the MFU 2250168404Spjd * state. 2251168404Spjd */ 2252174049Sjb if (LBOLT > buf->b_arc_access + ARC_MINTIME) { 2253168404Spjd /* 2254168404Spjd * More than 125ms have passed since we 2255168404Spjd * instantiated this buffer. Move it to the 2256168404Spjd * most frequently used state. 2257168404Spjd */ 2258174049Sjb buf->b_arc_access = LBOLT; 2259168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2260168404Spjd arc_change_state(arc_mfu, buf, hash_lock); 2261168404Spjd } 2262168404Spjd ARCSTAT_BUMP(arcstat_mru_hits); 2263168404Spjd } else if (buf->b_state == arc_mru_ghost) { 2264168404Spjd arc_state_t *new_state; 2265168404Spjd /* 2266168404Spjd * This buffer has been "accessed" recently, but 2267168404Spjd * was evicted from the cache. Move it to the 2268168404Spjd * MFU state. 2269168404Spjd */ 2270168404Spjd 2271168404Spjd if (buf->b_flags & ARC_PREFETCH) { 2272168404Spjd new_state = arc_mru; 2273168404Spjd if (refcount_count(&buf->b_refcnt) > 0) 2274168404Spjd buf->b_flags &= ~ARC_PREFETCH; 2275168404Spjd DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2276168404Spjd } else { 2277168404Spjd new_state = arc_mfu; 2278168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2279168404Spjd } 2280168404Spjd 2281174049Sjb buf->b_arc_access = LBOLT; 2282168404Spjd arc_change_state(new_state, buf, hash_lock); 2283168404Spjd 2284168404Spjd ARCSTAT_BUMP(arcstat_mru_ghost_hits); 2285168404Spjd } else if (buf->b_state == arc_mfu) { 2286168404Spjd /* 2287168404Spjd * This buffer has been accessed more than once and is 2288168404Spjd * still in the cache. Keep it in the MFU state. 2289168404Spjd * 2290168404Spjd * NOTE: an add_reference() that occurred when we did 2291168404Spjd * the arc_read() will have kicked this off the list. 2292168404Spjd * If it was a prefetch, we will explicitly move it to 2293168404Spjd * the head of the list now. 2294168404Spjd */ 2295168404Spjd if ((buf->b_flags & ARC_PREFETCH) != 0) { 2296168404Spjd ASSERT(refcount_count(&buf->b_refcnt) == 0); 2297168404Spjd ASSERT(list_link_active(&buf->b_arc_node)); 2298168404Spjd } 2299168404Spjd ARCSTAT_BUMP(arcstat_mfu_hits); 2300174049Sjb buf->b_arc_access = LBOLT; 2301168404Spjd } else if (buf->b_state == arc_mfu_ghost) { 2302168404Spjd arc_state_t *new_state = arc_mfu; 2303168404Spjd /* 2304168404Spjd * This buffer has been accessed more than once but has 2305168404Spjd * been evicted from the cache. Move it back to the 2306168404Spjd * MFU state. 2307168404Spjd */ 2308168404Spjd 2309168404Spjd if (buf->b_flags & ARC_PREFETCH) { 2310168404Spjd /* 2311168404Spjd * This is a prefetch access... 2312168404Spjd * move this block back to the MRU state. 2313168404Spjd */ 2314168404Spjd ASSERT3U(refcount_count(&buf->b_refcnt), ==, 0); 2315168404Spjd new_state = arc_mru; 2316168404Spjd } 2317168404Spjd 2318174049Sjb buf->b_arc_access = LBOLT; 2319168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2320168404Spjd arc_change_state(new_state, buf, hash_lock); 2321168404Spjd 2322168404Spjd ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 2323185029Spjd } else if (buf->b_state == arc_l2c_only) { 2324185029Spjd /* 2325185029Spjd * This buffer is on the 2nd Level ARC. 2326185029Spjd */ 2327185029Spjd 2328185029Spjd buf->b_arc_access = LBOLT; 2329185029Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2330185029Spjd arc_change_state(arc_mfu, buf, hash_lock); 2331168404Spjd } else { 2332168404Spjd ASSERT(!"invalid arc state"); 2333168404Spjd } 2334168404Spjd} 2335168404Spjd 2336168404Spjd/* a generic arc_done_func_t which you can use */ 2337168404Spjd/* ARGSUSED */ 2338168404Spjdvoid 2339168404Spjdarc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 2340168404Spjd{ 2341168404Spjd bcopy(buf->b_data, arg, buf->b_hdr->b_size); 2342168404Spjd VERIFY(arc_buf_remove_ref(buf, arg) == 1); 2343168404Spjd} 2344168404Spjd 2345185029Spjd/* a generic arc_done_func_t */ 2346168404Spjdvoid 2347168404Spjdarc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 2348168404Spjd{ 2349168404Spjd arc_buf_t **bufp = arg; 2350168404Spjd if (zio && zio->io_error) { 2351168404Spjd VERIFY(arc_buf_remove_ref(buf, arg) == 1); 2352168404Spjd *bufp = NULL; 2353168404Spjd } else { 2354168404Spjd *bufp = buf; 2355168404Spjd } 2356168404Spjd} 2357168404Spjd 2358168404Spjdstatic void 2359168404Spjdarc_read_done(zio_t *zio) 2360168404Spjd{ 2361168404Spjd arc_buf_hdr_t *hdr, *found; 2362168404Spjd arc_buf_t *buf; 2363168404Spjd arc_buf_t *abuf; /* buffer we're assigning to callback */ 2364168404Spjd kmutex_t *hash_lock; 2365168404Spjd arc_callback_t *callback_list, *acb; 2366168404Spjd int freeable = FALSE; 2367168404Spjd 2368168404Spjd buf = zio->io_private; 2369168404Spjd hdr = buf->b_hdr; 2370168404Spjd 2371168404Spjd /* 2372168404Spjd * The hdr was inserted into hash-table and removed from lists 2373168404Spjd * prior to starting I/O. We should find this header, since 2374168404Spjd * it's in the hash table, and it should be legit since it's 2375168404Spjd * not possible to evict it during the I/O. The only possible 2376168404Spjd * reason for it not to be found is if we were freed during the 2377168404Spjd * read. 2378168404Spjd */ 2379168404Spjd found = buf_hash_find(zio->io_spa, &hdr->b_dva, hdr->b_birth, 2380168404Spjd &hash_lock); 2381168404Spjd 2382168404Spjd ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) || 2383185029Spjd (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 2384185029Spjd (found == hdr && HDR_L2_READING(hdr))); 2385168404Spjd 2386185029Spjd hdr->b_flags &= ~ARC_L2_EVICTED; 2387185029Spjd if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 2388185029Spjd hdr->b_flags &= ~ARC_L2CACHE; 2389185029Spjd 2390168404Spjd /* byteswap if necessary */ 2391168404Spjd callback_list = hdr->b_acb; 2392168404Spjd ASSERT(callback_list != NULL); 2393185029Spjd if (BP_SHOULD_BYTESWAP(zio->io_bp)) { 2394185029Spjd arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 2395185029Spjd byteswap_uint64_array : 2396185029Spjd dmu_ot[BP_GET_TYPE(zio->io_bp)].ot_byteswap; 2397185029Spjd func(buf->b_data, hdr->b_size); 2398185029Spjd } 2399168404Spjd 2400185029Spjd arc_cksum_compute(buf, B_FALSE); 2401168404Spjd 2402168404Spjd /* create copies of the data buffer for the callers */ 2403168404Spjd abuf = buf; 2404168404Spjd for (acb = callback_list; acb; acb = acb->acb_next) { 2405168404Spjd if (acb->acb_done) { 2406168404Spjd if (abuf == NULL) 2407168404Spjd abuf = arc_buf_clone(buf); 2408168404Spjd acb->acb_buf = abuf; 2409168404Spjd abuf = NULL; 2410168404Spjd } 2411168404Spjd } 2412168404Spjd hdr->b_acb = NULL; 2413168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 2414168404Spjd ASSERT(!HDR_BUF_AVAILABLE(hdr)); 2415168404Spjd if (abuf == buf) 2416168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 2417168404Spjd 2418168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 2419168404Spjd 2420168404Spjd if (zio->io_error != 0) { 2421168404Spjd hdr->b_flags |= ARC_IO_ERROR; 2422168404Spjd if (hdr->b_state != arc_anon) 2423168404Spjd arc_change_state(arc_anon, hdr, hash_lock); 2424168404Spjd if (HDR_IN_HASH_TABLE(hdr)) 2425168404Spjd buf_hash_remove(hdr); 2426168404Spjd freeable = refcount_is_zero(&hdr->b_refcnt); 2427168404Spjd } 2428168404Spjd 2429168404Spjd /* 2430168404Spjd * Broadcast before we drop the hash_lock to avoid the possibility 2431168404Spjd * that the hdr (and hence the cv) might be freed before we get to 2432168404Spjd * the cv_broadcast(). 2433168404Spjd */ 2434168404Spjd cv_broadcast(&hdr->b_cv); 2435168404Spjd 2436168404Spjd if (hash_lock) { 2437168404Spjd /* 2438168404Spjd * Only call arc_access on anonymous buffers. This is because 2439168404Spjd * if we've issued an I/O for an evicted buffer, we've already 2440168404Spjd * called arc_access (to prevent any simultaneous readers from 2441168404Spjd * getting confused). 2442168404Spjd */ 2443168404Spjd if (zio->io_error == 0 && hdr->b_state == arc_anon) 2444168404Spjd arc_access(hdr, hash_lock); 2445168404Spjd mutex_exit(hash_lock); 2446168404Spjd } else { 2447168404Spjd /* 2448168404Spjd * This block was freed while we waited for the read to 2449168404Spjd * complete. It has been removed from the hash table and 2450168404Spjd * moved to the anonymous state (so that it won't show up 2451168404Spjd * in the cache). 2452168404Spjd */ 2453168404Spjd ASSERT3P(hdr->b_state, ==, arc_anon); 2454168404Spjd freeable = refcount_is_zero(&hdr->b_refcnt); 2455168404Spjd } 2456168404Spjd 2457168404Spjd /* execute each callback and free its structure */ 2458168404Spjd while ((acb = callback_list) != NULL) { 2459168404Spjd if (acb->acb_done) 2460168404Spjd acb->acb_done(zio, acb->acb_buf, acb->acb_private); 2461168404Spjd 2462168404Spjd if (acb->acb_zio_dummy != NULL) { 2463168404Spjd acb->acb_zio_dummy->io_error = zio->io_error; 2464168404Spjd zio_nowait(acb->acb_zio_dummy); 2465168404Spjd } 2466168404Spjd 2467168404Spjd callback_list = acb->acb_next; 2468168404Spjd kmem_free(acb, sizeof (arc_callback_t)); 2469168404Spjd } 2470168404Spjd 2471168404Spjd if (freeable) 2472168404Spjd arc_hdr_destroy(hdr); 2473168404Spjd} 2474168404Spjd 2475168404Spjd/* 2476168404Spjd * "Read" the block block at the specified DVA (in bp) via the 2477168404Spjd * cache. If the block is found in the cache, invoke the provided 2478168404Spjd * callback immediately and return. Note that the `zio' parameter 2479168404Spjd * in the callback will be NULL in this case, since no IO was 2480168404Spjd * required. If the block is not in the cache pass the read request 2481168404Spjd * on to the spa with a substitute callback function, so that the 2482168404Spjd * requested block will be added to the cache. 2483168404Spjd * 2484168404Spjd * If a read request arrives for a block that has a read in-progress, 2485168404Spjd * either wait for the in-progress read to complete (and return the 2486168404Spjd * results); or, if this is a read with a "done" func, add a record 2487168404Spjd * to the read to invoke the "done" func when the read completes, 2488168404Spjd * and return; or just return. 2489168404Spjd * 2490168404Spjd * arc_read_done() will invoke all the requested "done" functions 2491168404Spjd * for readers of this block. 2492185029Spjd * 2493185029Spjd * Normal callers should use arc_read and pass the arc buffer and offset 2494185029Spjd * for the bp. But if you know you don't need locking, you can use 2495185029Spjd * arc_read_bp. 2496168404Spjd */ 2497168404Spjdint 2498185029Spjdarc_read(zio_t *pio, spa_t *spa, blkptr_t *bp, arc_buf_t *pbuf, 2499185029Spjd arc_done_func_t *done, void *private, int priority, int zio_flags, 2500185029Spjd uint32_t *arc_flags, const zbookmark_t *zb) 2501168404Spjd{ 2502185029Spjd int err; 2503185029Spjd arc_buf_hdr_t *hdr = pbuf->b_hdr; 2504185029Spjd 2505185029Spjd ASSERT(!refcount_is_zero(&pbuf->b_hdr->b_refcnt)); 2506185029Spjd ASSERT3U((char *)bp - (char *)pbuf->b_data, <, pbuf->b_hdr->b_size); 2507185029Spjd rw_enter(&pbuf->b_lock, RW_READER); 2508185029Spjd 2509185029Spjd err = arc_read_nolock(pio, spa, bp, done, private, priority, 2510185029Spjd zio_flags, arc_flags, zb); 2511185029Spjd 2512185029Spjd ASSERT3P(hdr, ==, pbuf->b_hdr); 2513185029Spjd rw_exit(&pbuf->b_lock); 2514185029Spjd return (err); 2515185029Spjd} 2516185029Spjd 2517185029Spjdint 2518185029Spjdarc_read_nolock(zio_t *pio, spa_t *spa, blkptr_t *bp, 2519185029Spjd arc_done_func_t *done, void *private, int priority, int zio_flags, 2520185029Spjd uint32_t *arc_flags, const zbookmark_t *zb) 2521185029Spjd{ 2522168404Spjd arc_buf_hdr_t *hdr; 2523168404Spjd arc_buf_t *buf; 2524168404Spjd kmutex_t *hash_lock; 2525185029Spjd zio_t *rzio; 2526168404Spjd 2527168404Spjdtop: 2528168404Spjd hdr = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_lock); 2529168404Spjd if (hdr && hdr->b_datacnt > 0) { 2530168404Spjd 2531168404Spjd *arc_flags |= ARC_CACHED; 2532168404Spjd 2533168404Spjd if (HDR_IO_IN_PROGRESS(hdr)) { 2534168404Spjd 2535168404Spjd if (*arc_flags & ARC_WAIT) { 2536168404Spjd cv_wait(&hdr->b_cv, hash_lock); 2537168404Spjd mutex_exit(hash_lock); 2538168404Spjd goto top; 2539168404Spjd } 2540168404Spjd ASSERT(*arc_flags & ARC_NOWAIT); 2541168404Spjd 2542168404Spjd if (done) { 2543168404Spjd arc_callback_t *acb = NULL; 2544168404Spjd 2545168404Spjd acb = kmem_zalloc(sizeof (arc_callback_t), 2546168404Spjd KM_SLEEP); 2547168404Spjd acb->acb_done = done; 2548168404Spjd acb->acb_private = private; 2549168404Spjd if (pio != NULL) 2550168404Spjd acb->acb_zio_dummy = zio_null(pio, 2551185029Spjd spa, NULL, NULL, zio_flags); 2552168404Spjd 2553168404Spjd ASSERT(acb->acb_done != NULL); 2554168404Spjd acb->acb_next = hdr->b_acb; 2555168404Spjd hdr->b_acb = acb; 2556168404Spjd add_reference(hdr, hash_lock, private); 2557168404Spjd mutex_exit(hash_lock); 2558168404Spjd return (0); 2559168404Spjd } 2560168404Spjd mutex_exit(hash_lock); 2561168404Spjd return (0); 2562168404Spjd } 2563168404Spjd 2564168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 2565168404Spjd 2566168404Spjd if (done) { 2567168404Spjd add_reference(hdr, hash_lock, private); 2568168404Spjd /* 2569168404Spjd * If this block is already in use, create a new 2570168404Spjd * copy of the data so that we will be guaranteed 2571168404Spjd * that arc_release() will always succeed. 2572168404Spjd */ 2573168404Spjd buf = hdr->b_buf; 2574168404Spjd ASSERT(buf); 2575168404Spjd ASSERT(buf->b_data); 2576168404Spjd if (HDR_BUF_AVAILABLE(hdr)) { 2577168404Spjd ASSERT(buf->b_efunc == NULL); 2578168404Spjd hdr->b_flags &= ~ARC_BUF_AVAILABLE; 2579168404Spjd } else { 2580168404Spjd buf = arc_buf_clone(buf); 2581168404Spjd } 2582168404Spjd } else if (*arc_flags & ARC_PREFETCH && 2583168404Spjd refcount_count(&hdr->b_refcnt) == 0) { 2584168404Spjd hdr->b_flags |= ARC_PREFETCH; 2585168404Spjd } 2586168404Spjd DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 2587168404Spjd arc_access(hdr, hash_lock); 2588185029Spjd if (*arc_flags & ARC_L2CACHE) 2589185029Spjd hdr->b_flags |= ARC_L2CACHE; 2590168404Spjd mutex_exit(hash_lock); 2591168404Spjd ARCSTAT_BUMP(arcstat_hits); 2592168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 2593168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 2594168404Spjd data, metadata, hits); 2595168404Spjd 2596168404Spjd if (done) 2597168404Spjd done(NULL, buf, private); 2598168404Spjd } else { 2599168404Spjd uint64_t size = BP_GET_LSIZE(bp); 2600168404Spjd arc_callback_t *acb; 2601185029Spjd vdev_t *vd = NULL; 2602185029Spjd daddr_t addr; 2603168404Spjd 2604168404Spjd if (hdr == NULL) { 2605168404Spjd /* this block is not in the cache */ 2606168404Spjd arc_buf_hdr_t *exists; 2607168404Spjd arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 2608168404Spjd buf = arc_buf_alloc(spa, size, private, type); 2609168404Spjd hdr = buf->b_hdr; 2610168404Spjd hdr->b_dva = *BP_IDENTITY(bp); 2611168404Spjd hdr->b_birth = bp->blk_birth; 2612168404Spjd hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 2613168404Spjd exists = buf_hash_insert(hdr, &hash_lock); 2614168404Spjd if (exists) { 2615168404Spjd /* somebody beat us to the hash insert */ 2616168404Spjd mutex_exit(hash_lock); 2617168404Spjd bzero(&hdr->b_dva, sizeof (dva_t)); 2618168404Spjd hdr->b_birth = 0; 2619168404Spjd hdr->b_cksum0 = 0; 2620168404Spjd (void) arc_buf_remove_ref(buf, private); 2621168404Spjd goto top; /* restart the IO request */ 2622168404Spjd } 2623168404Spjd /* if this is a prefetch, we don't have a reference */ 2624168404Spjd if (*arc_flags & ARC_PREFETCH) { 2625168404Spjd (void) remove_reference(hdr, hash_lock, 2626168404Spjd private); 2627168404Spjd hdr->b_flags |= ARC_PREFETCH; 2628168404Spjd } 2629185029Spjd if (*arc_flags & ARC_L2CACHE) 2630185029Spjd hdr->b_flags |= ARC_L2CACHE; 2631168404Spjd if (BP_GET_LEVEL(bp) > 0) 2632168404Spjd hdr->b_flags |= ARC_INDIRECT; 2633168404Spjd } else { 2634168404Spjd /* this block is in the ghost cache */ 2635168404Spjd ASSERT(GHOST_STATE(hdr->b_state)); 2636168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2637168404Spjd ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 0); 2638168404Spjd ASSERT(hdr->b_buf == NULL); 2639168404Spjd 2640168404Spjd /* if this is a prefetch, we don't have a reference */ 2641168404Spjd if (*arc_flags & ARC_PREFETCH) 2642168404Spjd hdr->b_flags |= ARC_PREFETCH; 2643168404Spjd else 2644168404Spjd add_reference(hdr, hash_lock, private); 2645185029Spjd if (*arc_flags & ARC_L2CACHE) 2646185029Spjd hdr->b_flags |= ARC_L2CACHE; 2647185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2648168404Spjd buf->b_hdr = hdr; 2649168404Spjd buf->b_data = NULL; 2650168404Spjd buf->b_efunc = NULL; 2651168404Spjd buf->b_private = NULL; 2652168404Spjd buf->b_next = NULL; 2653168404Spjd hdr->b_buf = buf; 2654168404Spjd arc_get_data_buf(buf); 2655168404Spjd ASSERT(hdr->b_datacnt == 0); 2656168404Spjd hdr->b_datacnt = 1; 2657168404Spjd 2658168404Spjd } 2659168404Spjd 2660168404Spjd acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 2661168404Spjd acb->acb_done = done; 2662168404Spjd acb->acb_private = private; 2663168404Spjd 2664168404Spjd ASSERT(hdr->b_acb == NULL); 2665168404Spjd hdr->b_acb = acb; 2666168404Spjd hdr->b_flags |= ARC_IO_IN_PROGRESS; 2667168404Spjd 2668168404Spjd /* 2669168404Spjd * If the buffer has been evicted, migrate it to a present state 2670168404Spjd * before issuing the I/O. Once we drop the hash-table lock, 2671168404Spjd * the header will be marked as I/O in progress and have an 2672168404Spjd * attached buffer. At this point, anybody who finds this 2673168404Spjd * buffer ought to notice that it's legit but has a pending I/O. 2674168404Spjd */ 2675168404Spjd 2676168404Spjd if (GHOST_STATE(hdr->b_state)) 2677168404Spjd arc_access(hdr, hash_lock); 2678185029Spjd 2679185029Spjd if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL && 2680185029Spjd (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 2681185029Spjd addr = hdr->b_l2hdr->b_daddr; 2682185029Spjd /* 2683185029Spjd * Lock out device removal. 2684185029Spjd */ 2685185029Spjd if (vdev_is_dead(vd) || 2686185029Spjd !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 2687185029Spjd vd = NULL; 2688185029Spjd } 2689185029Spjd 2690168404Spjd mutex_exit(hash_lock); 2691168404Spjd 2692168404Spjd ASSERT3U(hdr->b_size, ==, size); 2693168404Spjd DTRACE_PROBE3(arc__miss, blkptr_t *, bp, uint64_t, size, 2694168404Spjd zbookmark_t *, zb); 2695168404Spjd ARCSTAT_BUMP(arcstat_misses); 2696168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 2697168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 2698168404Spjd data, metadata, misses); 2699168404Spjd 2700185029Spjd if (vd != NULL) { 2701185029Spjd /* 2702185029Spjd * Read from the L2ARC if the following are true: 2703185029Spjd * 1. The L2ARC vdev was previously cached. 2704185029Spjd * 2. This buffer still has L2ARC metadata. 2705185029Spjd * 3. This buffer isn't currently writing to the L2ARC. 2706185029Spjd * 4. The L2ARC entry wasn't evicted, which may 2707185029Spjd * also have invalidated the vdev. 2708185029Spjd */ 2709185029Spjd if (hdr->b_l2hdr != NULL && 2710185029Spjd !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr)) { 2711185029Spjd l2arc_read_callback_t *cb; 2712185029Spjd 2713185029Spjd DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 2714185029Spjd ARCSTAT_BUMP(arcstat_l2_hits); 2715185029Spjd 2716185029Spjd cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 2717185029Spjd KM_SLEEP); 2718185029Spjd cb->l2rcb_buf = buf; 2719185029Spjd cb->l2rcb_spa = spa; 2720185029Spjd cb->l2rcb_bp = *bp; 2721185029Spjd cb->l2rcb_zb = *zb; 2722185029Spjd cb->l2rcb_flags = zio_flags; 2723185029Spjd 2724185029Spjd /* 2725185029Spjd * l2arc read. The SCL_L2ARC lock will be 2726185029Spjd * released by l2arc_read_done(). 2727185029Spjd */ 2728185029Spjd rzio = zio_read_phys(pio, vd, addr, size, 2729185029Spjd buf->b_data, ZIO_CHECKSUM_OFF, 2730185029Spjd l2arc_read_done, cb, priority, zio_flags | 2731185029Spjd ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL | 2732185029Spjd ZIO_FLAG_DONT_PROPAGATE | 2733185029Spjd ZIO_FLAG_DONT_RETRY, B_FALSE); 2734185029Spjd DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 2735185029Spjd zio_t *, rzio); 2736185029Spjd 2737185029Spjd if (*arc_flags & ARC_NOWAIT) { 2738185029Spjd zio_nowait(rzio); 2739185029Spjd return (0); 2740185029Spjd } 2741185029Spjd 2742185029Spjd ASSERT(*arc_flags & ARC_WAIT); 2743185029Spjd if (zio_wait(rzio) == 0) 2744185029Spjd return (0); 2745185029Spjd 2746185029Spjd /* l2arc read error; goto zio_read() */ 2747185029Spjd } else { 2748185029Spjd DTRACE_PROBE1(l2arc__miss, 2749185029Spjd arc_buf_hdr_t *, hdr); 2750185029Spjd ARCSTAT_BUMP(arcstat_l2_misses); 2751185029Spjd if (HDR_L2_WRITING(hdr)) 2752185029Spjd ARCSTAT_BUMP(arcstat_l2_rw_clash); 2753185029Spjd spa_config_exit(spa, SCL_L2ARC, vd); 2754185029Spjd } 2755185029Spjd } 2756185029Spjd 2757168404Spjd rzio = zio_read(pio, spa, bp, buf->b_data, size, 2758185029Spjd arc_read_done, buf, priority, zio_flags, zb); 2759168404Spjd 2760168404Spjd if (*arc_flags & ARC_WAIT) 2761168404Spjd return (zio_wait(rzio)); 2762168404Spjd 2763168404Spjd ASSERT(*arc_flags & ARC_NOWAIT); 2764168404Spjd zio_nowait(rzio); 2765168404Spjd } 2766168404Spjd return (0); 2767168404Spjd} 2768168404Spjd 2769168404Spjd/* 2770168404Spjd * arc_read() variant to support pool traversal. If the block is already 2771168404Spjd * in the ARC, make a copy of it; otherwise, the caller will do the I/O. 2772168404Spjd * The idea is that we don't want pool traversal filling up memory, but 2773168404Spjd * if the ARC already has the data anyway, we shouldn't pay for the I/O. 2774168404Spjd */ 2775168404Spjdint 2776168404Spjdarc_tryread(spa_t *spa, blkptr_t *bp, void *data) 2777168404Spjd{ 2778168404Spjd arc_buf_hdr_t *hdr; 2779168404Spjd kmutex_t *hash_mtx; 2780168404Spjd int rc = 0; 2781168404Spjd 2782168404Spjd hdr = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_mtx); 2783168404Spjd 2784168404Spjd if (hdr && hdr->b_datacnt > 0 && !HDR_IO_IN_PROGRESS(hdr)) { 2785168404Spjd arc_buf_t *buf = hdr->b_buf; 2786168404Spjd 2787168404Spjd ASSERT(buf); 2788168404Spjd while (buf->b_data == NULL) { 2789168404Spjd buf = buf->b_next; 2790168404Spjd ASSERT(buf); 2791168404Spjd } 2792168404Spjd bcopy(buf->b_data, data, hdr->b_size); 2793168404Spjd } else { 2794168404Spjd rc = ENOENT; 2795168404Spjd } 2796168404Spjd 2797168404Spjd if (hash_mtx) 2798168404Spjd mutex_exit(hash_mtx); 2799168404Spjd 2800168404Spjd return (rc); 2801168404Spjd} 2802168404Spjd 2803168404Spjdvoid 2804168404Spjdarc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 2805168404Spjd{ 2806168404Spjd ASSERT(buf->b_hdr != NULL); 2807168404Spjd ASSERT(buf->b_hdr->b_state != arc_anon); 2808168404Spjd ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 2809168404Spjd buf->b_efunc = func; 2810168404Spjd buf->b_private = private; 2811168404Spjd} 2812168404Spjd 2813168404Spjd/* 2814168404Spjd * This is used by the DMU to let the ARC know that a buffer is 2815168404Spjd * being evicted, so the ARC should clean up. If this arc buf 2816168404Spjd * is not yet in the evicted state, it will be put there. 2817168404Spjd */ 2818168404Spjdint 2819168404Spjdarc_buf_evict(arc_buf_t *buf) 2820168404Spjd{ 2821168404Spjd arc_buf_hdr_t *hdr; 2822168404Spjd kmutex_t *hash_lock; 2823168404Spjd arc_buf_t **bufp; 2824168404Spjd 2825185029Spjd rw_enter(&buf->b_lock, RW_WRITER); 2826168404Spjd hdr = buf->b_hdr; 2827168404Spjd if (hdr == NULL) { 2828168404Spjd /* 2829168404Spjd * We are in arc_do_user_evicts(). 2830168404Spjd */ 2831168404Spjd ASSERT(buf->b_data == NULL); 2832185029Spjd rw_exit(&buf->b_lock); 2833168404Spjd return (0); 2834185029Spjd } else if (buf->b_data == NULL) { 2835185029Spjd arc_buf_t copy = *buf; /* structure assignment */ 2836185029Spjd /* 2837185029Spjd * We are on the eviction list; process this buffer now 2838185029Spjd * but let arc_do_user_evicts() do the reaping. 2839185029Spjd */ 2840185029Spjd buf->b_efunc = NULL; 2841185029Spjd rw_exit(&buf->b_lock); 2842185029Spjd VERIFY(copy.b_efunc(©) == 0); 2843185029Spjd return (1); 2844168404Spjd } 2845168404Spjd hash_lock = HDR_LOCK(hdr); 2846168404Spjd mutex_enter(hash_lock); 2847168404Spjd 2848168404Spjd ASSERT(buf->b_hdr == hdr); 2849168404Spjd ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 2850168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 2851168404Spjd 2852168404Spjd /* 2853168404Spjd * Pull this buffer off of the hdr 2854168404Spjd */ 2855168404Spjd bufp = &hdr->b_buf; 2856168404Spjd while (*bufp != buf) 2857168404Spjd bufp = &(*bufp)->b_next; 2858168404Spjd *bufp = buf->b_next; 2859168404Spjd 2860168404Spjd ASSERT(buf->b_data != NULL); 2861168404Spjd arc_buf_destroy(buf, FALSE, FALSE); 2862168404Spjd 2863168404Spjd if (hdr->b_datacnt == 0) { 2864168404Spjd arc_state_t *old_state = hdr->b_state; 2865168404Spjd arc_state_t *evicted_state; 2866168404Spjd 2867168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2868168404Spjd 2869168404Spjd evicted_state = 2870168404Spjd (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2871168404Spjd 2872168404Spjd mutex_enter(&old_state->arcs_mtx); 2873168404Spjd mutex_enter(&evicted_state->arcs_mtx); 2874168404Spjd 2875168404Spjd arc_change_state(evicted_state, hdr, hash_lock); 2876168404Spjd ASSERT(HDR_IN_HASH_TABLE(hdr)); 2877185029Spjd hdr->b_flags |= ARC_IN_HASH_TABLE; 2878185029Spjd hdr->b_flags &= ~ARC_BUF_AVAILABLE; 2879168404Spjd 2880168404Spjd mutex_exit(&evicted_state->arcs_mtx); 2881168404Spjd mutex_exit(&old_state->arcs_mtx); 2882168404Spjd } 2883168404Spjd mutex_exit(hash_lock); 2884185029Spjd rw_exit(&buf->b_lock); 2885168404Spjd 2886168404Spjd VERIFY(buf->b_efunc(buf) == 0); 2887168404Spjd buf->b_efunc = NULL; 2888168404Spjd buf->b_private = NULL; 2889168404Spjd buf->b_hdr = NULL; 2890168404Spjd kmem_cache_free(buf_cache, buf); 2891168404Spjd return (1); 2892168404Spjd} 2893168404Spjd 2894168404Spjd/* 2895168404Spjd * Release this buffer from the cache. This must be done 2896168404Spjd * after a read and prior to modifying the buffer contents. 2897168404Spjd * If the buffer has more than one reference, we must make 2898185029Spjd * a new hdr for the buffer. 2899168404Spjd */ 2900168404Spjdvoid 2901168404Spjdarc_release(arc_buf_t *buf, void *tag) 2902168404Spjd{ 2903185029Spjd arc_buf_hdr_t *hdr; 2904185029Spjd kmutex_t *hash_lock; 2905185029Spjd l2arc_buf_hdr_t *l2hdr; 2906185029Spjd uint64_t buf_size; 2907168404Spjd 2908185029Spjd rw_enter(&buf->b_lock, RW_WRITER); 2909185029Spjd hdr = buf->b_hdr; 2910185029Spjd 2911168404Spjd /* this buffer is not on any list */ 2912168404Spjd ASSERT(refcount_count(&hdr->b_refcnt) > 0); 2913185029Spjd ASSERT(!(hdr->b_flags & ARC_STORED)); 2914168404Spjd 2915168404Spjd if (hdr->b_state == arc_anon) { 2916168404Spjd /* this buffer is already released */ 2917168404Spjd ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 1); 2918168404Spjd ASSERT(BUF_EMPTY(hdr)); 2919168404Spjd ASSERT(buf->b_efunc == NULL); 2920168404Spjd arc_buf_thaw(buf); 2921185029Spjd rw_exit(&buf->b_lock); 2922168404Spjd return; 2923168404Spjd } 2924168404Spjd 2925185029Spjd hash_lock = HDR_LOCK(hdr); 2926168404Spjd mutex_enter(hash_lock); 2927168404Spjd 2928185029Spjd l2hdr = hdr->b_l2hdr; 2929185029Spjd if (l2hdr) { 2930185029Spjd mutex_enter(&l2arc_buflist_mtx); 2931185029Spjd hdr->b_l2hdr = NULL; 2932185029Spjd buf_size = hdr->b_size; 2933185029Spjd } 2934185029Spjd 2935168404Spjd /* 2936168404Spjd * Do we have more than one buf? 2937168404Spjd */ 2938185029Spjd if (hdr->b_datacnt > 1) { 2939168404Spjd arc_buf_hdr_t *nhdr; 2940168404Spjd arc_buf_t **bufp; 2941168404Spjd uint64_t blksz = hdr->b_size; 2942168404Spjd spa_t *spa = hdr->b_spa; 2943168404Spjd arc_buf_contents_t type = hdr->b_type; 2944185029Spjd uint32_t flags = hdr->b_flags; 2945168404Spjd 2946185029Spjd ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 2947168404Spjd /* 2948168404Spjd * Pull the data off of this buf and attach it to 2949168404Spjd * a new anonymous buf. 2950168404Spjd */ 2951168404Spjd (void) remove_reference(hdr, hash_lock, tag); 2952168404Spjd bufp = &hdr->b_buf; 2953168404Spjd while (*bufp != buf) 2954168404Spjd bufp = &(*bufp)->b_next; 2955168404Spjd *bufp = (*bufp)->b_next; 2956168404Spjd buf->b_next = NULL; 2957168404Spjd 2958168404Spjd ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 2959168404Spjd atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 2960168404Spjd if (refcount_is_zero(&hdr->b_refcnt)) { 2961185029Spjd uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 2962185029Spjd ASSERT3U(*size, >=, hdr->b_size); 2963185029Spjd atomic_add_64(size, -hdr->b_size); 2964168404Spjd } 2965168404Spjd hdr->b_datacnt -= 1; 2966168404Spjd arc_cksum_verify(buf); 2967168404Spjd 2968168404Spjd mutex_exit(hash_lock); 2969168404Spjd 2970185029Spjd nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 2971168404Spjd nhdr->b_size = blksz; 2972168404Spjd nhdr->b_spa = spa; 2973168404Spjd nhdr->b_type = type; 2974168404Spjd nhdr->b_buf = buf; 2975168404Spjd nhdr->b_state = arc_anon; 2976168404Spjd nhdr->b_arc_access = 0; 2977185029Spjd nhdr->b_flags = flags & ARC_L2_WRITING; 2978185029Spjd nhdr->b_l2hdr = NULL; 2979168404Spjd nhdr->b_datacnt = 1; 2980168404Spjd nhdr->b_freeze_cksum = NULL; 2981168404Spjd (void) refcount_add(&nhdr->b_refcnt, tag); 2982168404Spjd buf->b_hdr = nhdr; 2983185029Spjd rw_exit(&buf->b_lock); 2984168404Spjd atomic_add_64(&arc_anon->arcs_size, blksz); 2985168404Spjd } else { 2986185029Spjd rw_exit(&buf->b_lock); 2987168404Spjd ASSERT(refcount_count(&hdr->b_refcnt) == 1); 2988168404Spjd ASSERT(!list_link_active(&hdr->b_arc_node)); 2989168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2990168404Spjd arc_change_state(arc_anon, hdr, hash_lock); 2991168404Spjd hdr->b_arc_access = 0; 2992168404Spjd mutex_exit(hash_lock); 2993185029Spjd 2994168404Spjd bzero(&hdr->b_dva, sizeof (dva_t)); 2995168404Spjd hdr->b_birth = 0; 2996168404Spjd hdr->b_cksum0 = 0; 2997168404Spjd arc_buf_thaw(buf); 2998168404Spjd } 2999168404Spjd buf->b_efunc = NULL; 3000168404Spjd buf->b_private = NULL; 3001185029Spjd 3002185029Spjd if (l2hdr) { 3003185029Spjd list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3004185029Spjd kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3005185029Spjd ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3006185029Spjd mutex_exit(&l2arc_buflist_mtx); 3007185029Spjd } 3008168404Spjd} 3009168404Spjd 3010168404Spjdint 3011168404Spjdarc_released(arc_buf_t *buf) 3012168404Spjd{ 3013185029Spjd int released; 3014185029Spjd 3015185029Spjd rw_enter(&buf->b_lock, RW_READER); 3016185029Spjd released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3017185029Spjd rw_exit(&buf->b_lock); 3018185029Spjd return (released); 3019168404Spjd} 3020168404Spjd 3021168404Spjdint 3022168404Spjdarc_has_callback(arc_buf_t *buf) 3023168404Spjd{ 3024185029Spjd int callback; 3025185029Spjd 3026185029Spjd rw_enter(&buf->b_lock, RW_READER); 3027185029Spjd callback = (buf->b_efunc != NULL); 3028185029Spjd rw_exit(&buf->b_lock); 3029185029Spjd return (callback); 3030168404Spjd} 3031168404Spjd 3032168404Spjd#ifdef ZFS_DEBUG 3033168404Spjdint 3034168404Spjdarc_referenced(arc_buf_t *buf) 3035168404Spjd{ 3036185029Spjd int referenced; 3037185029Spjd 3038185029Spjd rw_enter(&buf->b_lock, RW_READER); 3039185029Spjd referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3040185029Spjd rw_exit(&buf->b_lock); 3041185029Spjd return (referenced); 3042168404Spjd} 3043168404Spjd#endif 3044168404Spjd 3045168404Spjdstatic void 3046168404Spjdarc_write_ready(zio_t *zio) 3047168404Spjd{ 3048168404Spjd arc_write_callback_t *callback = zio->io_private; 3049168404Spjd arc_buf_t *buf = callback->awcb_buf; 3050185029Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3051168404Spjd 3052185029Spjd ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3053185029Spjd callback->awcb_ready(zio, buf, callback->awcb_private); 3054185029Spjd 3055185029Spjd /* 3056185029Spjd * If the IO is already in progress, then this is a re-write 3057185029Spjd * attempt, so we need to thaw and re-compute the cksum. 3058185029Spjd * It is the responsibility of the callback to handle the 3059185029Spjd * accounting for any re-write attempt. 3060185029Spjd */ 3061185029Spjd if (HDR_IO_IN_PROGRESS(hdr)) { 3062185029Spjd mutex_enter(&hdr->b_freeze_lock); 3063185029Spjd if (hdr->b_freeze_cksum != NULL) { 3064185029Spjd kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3065185029Spjd hdr->b_freeze_cksum = NULL; 3066185029Spjd } 3067185029Spjd mutex_exit(&hdr->b_freeze_lock); 3068168404Spjd } 3069185029Spjd arc_cksum_compute(buf, B_FALSE); 3070185029Spjd hdr->b_flags |= ARC_IO_IN_PROGRESS; 3071168404Spjd} 3072168404Spjd 3073168404Spjdstatic void 3074168404Spjdarc_write_done(zio_t *zio) 3075168404Spjd{ 3076168404Spjd arc_write_callback_t *callback = zio->io_private; 3077168404Spjd arc_buf_t *buf = callback->awcb_buf; 3078168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3079168404Spjd 3080168404Spjd hdr->b_acb = NULL; 3081168404Spjd 3082168404Spjd hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3083168404Spjd hdr->b_birth = zio->io_bp->blk_birth; 3084168404Spjd hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3085168404Spjd /* 3086168404Spjd * If the block to be written was all-zero, we may have 3087168404Spjd * compressed it away. In this case no write was performed 3088168404Spjd * so there will be no dva/birth-date/checksum. The buffer 3089168404Spjd * must therefor remain anonymous (and uncached). 3090168404Spjd */ 3091168404Spjd if (!BUF_EMPTY(hdr)) { 3092168404Spjd arc_buf_hdr_t *exists; 3093168404Spjd kmutex_t *hash_lock; 3094168404Spjd 3095168404Spjd arc_cksum_verify(buf); 3096168404Spjd 3097168404Spjd exists = buf_hash_insert(hdr, &hash_lock); 3098168404Spjd if (exists) { 3099168404Spjd /* 3100168404Spjd * This can only happen if we overwrite for 3101168404Spjd * sync-to-convergence, because we remove 3102168404Spjd * buffers from the hash table when we arc_free(). 3103168404Spjd */ 3104185029Spjd ASSERT(zio->io_flags & ZIO_FLAG_IO_REWRITE); 3105168404Spjd ASSERT(DVA_EQUAL(BP_IDENTITY(&zio->io_bp_orig), 3106168404Spjd BP_IDENTITY(zio->io_bp))); 3107168404Spjd ASSERT3U(zio->io_bp_orig.blk_birth, ==, 3108168404Spjd zio->io_bp->blk_birth); 3109168404Spjd 3110168404Spjd ASSERT(refcount_is_zero(&exists->b_refcnt)); 3111168404Spjd arc_change_state(arc_anon, exists, hash_lock); 3112168404Spjd mutex_exit(hash_lock); 3113168404Spjd arc_hdr_destroy(exists); 3114168404Spjd exists = buf_hash_insert(hdr, &hash_lock); 3115168404Spjd ASSERT3P(exists, ==, NULL); 3116168404Spjd } 3117168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3118185029Spjd /* if it's not anon, we are doing a scrub */ 3119185029Spjd if (hdr->b_state == arc_anon) 3120185029Spjd arc_access(hdr, hash_lock); 3121168404Spjd mutex_exit(hash_lock); 3122168404Spjd } else if (callback->awcb_done == NULL) { 3123168404Spjd int destroy_hdr; 3124168404Spjd /* 3125168404Spjd * This is an anonymous buffer with no user callback, 3126168404Spjd * destroy it if there are no active references. 3127168404Spjd */ 3128168404Spjd mutex_enter(&arc_eviction_mtx); 3129168404Spjd destroy_hdr = refcount_is_zero(&hdr->b_refcnt); 3130168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3131168404Spjd mutex_exit(&arc_eviction_mtx); 3132168404Spjd if (destroy_hdr) 3133168404Spjd arc_hdr_destroy(hdr); 3134168404Spjd } else { 3135168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3136168404Spjd } 3137185029Spjd hdr->b_flags &= ~ARC_STORED; 3138168404Spjd 3139168404Spjd if (callback->awcb_done) { 3140168404Spjd ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3141168404Spjd callback->awcb_done(zio, buf, callback->awcb_private); 3142168404Spjd } 3143168404Spjd 3144168404Spjd kmem_free(callback, sizeof (arc_write_callback_t)); 3145168404Spjd} 3146168404Spjd 3147185029Spjdstatic void 3148185029Spjdwrite_policy(spa_t *spa, const writeprops_t *wp, zio_prop_t *zp) 3149185029Spjd{ 3150185029Spjd boolean_t ismd = (wp->wp_level > 0 || dmu_ot[wp->wp_type].ot_metadata); 3151185029Spjd 3152185029Spjd /* Determine checksum setting */ 3153185029Spjd if (ismd) { 3154185029Spjd /* 3155185029Spjd * Metadata always gets checksummed. If the data 3156185029Spjd * checksum is multi-bit correctable, and it's not a 3157185029Spjd * ZBT-style checksum, then it's suitable for metadata 3158185029Spjd * as well. Otherwise, the metadata checksum defaults 3159185029Spjd * to fletcher4. 3160185029Spjd */ 3161185029Spjd if (zio_checksum_table[wp->wp_oschecksum].ci_correctable && 3162185029Spjd !zio_checksum_table[wp->wp_oschecksum].ci_zbt) 3163185029Spjd zp->zp_checksum = wp->wp_oschecksum; 3164185029Spjd else 3165185029Spjd zp->zp_checksum = ZIO_CHECKSUM_FLETCHER_4; 3166185029Spjd } else { 3167185029Spjd zp->zp_checksum = zio_checksum_select(wp->wp_dnchecksum, 3168185029Spjd wp->wp_oschecksum); 3169185029Spjd } 3170185029Spjd 3171185029Spjd /* Determine compression setting */ 3172185029Spjd if (ismd) { 3173185029Spjd /* 3174185029Spjd * XXX -- we should design a compression algorithm 3175185029Spjd * that specializes in arrays of bps. 3176185029Spjd */ 3177185029Spjd zp->zp_compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY : 3178185029Spjd ZIO_COMPRESS_LZJB; 3179185029Spjd } else { 3180185029Spjd zp->zp_compress = zio_compress_select(wp->wp_dncompress, 3181185029Spjd wp->wp_oscompress); 3182185029Spjd } 3183185029Spjd 3184185029Spjd zp->zp_type = wp->wp_type; 3185185029Spjd zp->zp_level = wp->wp_level; 3186185029Spjd zp->zp_ndvas = MIN(wp->wp_copies + ismd, spa_max_replication(spa)); 3187185029Spjd} 3188185029Spjd 3189168404Spjdzio_t * 3190185029Spjdarc_write(zio_t *pio, spa_t *spa, const writeprops_t *wp, 3191185029Spjd boolean_t l2arc, uint64_t txg, blkptr_t *bp, arc_buf_t *buf, 3192168404Spjd arc_done_func_t *ready, arc_done_func_t *done, void *private, int priority, 3193185029Spjd int zio_flags, const zbookmark_t *zb) 3194168404Spjd{ 3195168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3196168404Spjd arc_write_callback_t *callback; 3197185029Spjd zio_t *zio; 3198185029Spjd zio_prop_t zp; 3199168404Spjd 3200185029Spjd ASSERT(ready != NULL); 3201168404Spjd ASSERT(!HDR_IO_ERROR(hdr)); 3202168404Spjd ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3203168404Spjd ASSERT(hdr->b_acb == 0); 3204185029Spjd if (l2arc) 3205185029Spjd hdr->b_flags |= ARC_L2CACHE; 3206168404Spjd callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 3207168404Spjd callback->awcb_ready = ready; 3208168404Spjd callback->awcb_done = done; 3209168404Spjd callback->awcb_private = private; 3210168404Spjd callback->awcb_buf = buf; 3211168404Spjd 3212185029Spjd write_policy(spa, wp, &zp); 3213185029Spjd zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, &zp, 3214185029Spjd arc_write_ready, arc_write_done, callback, priority, zio_flags, zb); 3215185029Spjd 3216168404Spjd return (zio); 3217168404Spjd} 3218168404Spjd 3219168404Spjdint 3220168404Spjdarc_free(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 3221168404Spjd zio_done_func_t *done, void *private, uint32_t arc_flags) 3222168404Spjd{ 3223168404Spjd arc_buf_hdr_t *ab; 3224168404Spjd kmutex_t *hash_lock; 3225168404Spjd zio_t *zio; 3226168404Spjd 3227168404Spjd /* 3228168404Spjd * If this buffer is in the cache, release it, so it 3229168404Spjd * can be re-used. 3230168404Spjd */ 3231168404Spjd ab = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_lock); 3232168404Spjd if (ab != NULL) { 3233168404Spjd /* 3234168404Spjd * The checksum of blocks to free is not always 3235168404Spjd * preserved (eg. on the deadlist). However, if it is 3236168404Spjd * nonzero, it should match what we have in the cache. 3237168404Spjd */ 3238168404Spjd ASSERT(bp->blk_cksum.zc_word[0] == 0 || 3239185029Spjd bp->blk_cksum.zc_word[0] == ab->b_cksum0 || 3240185029Spjd bp->blk_fill == BLK_FILL_ALREADY_FREED); 3241185029Spjd 3242168404Spjd if (ab->b_state != arc_anon) 3243168404Spjd arc_change_state(arc_anon, ab, hash_lock); 3244168404Spjd if (HDR_IO_IN_PROGRESS(ab)) { 3245168404Spjd /* 3246168404Spjd * This should only happen when we prefetch. 3247168404Spjd */ 3248168404Spjd ASSERT(ab->b_flags & ARC_PREFETCH); 3249168404Spjd ASSERT3U(ab->b_datacnt, ==, 1); 3250168404Spjd ab->b_flags |= ARC_FREED_IN_READ; 3251168404Spjd if (HDR_IN_HASH_TABLE(ab)) 3252168404Spjd buf_hash_remove(ab); 3253168404Spjd ab->b_arc_access = 0; 3254168404Spjd bzero(&ab->b_dva, sizeof (dva_t)); 3255168404Spjd ab->b_birth = 0; 3256168404Spjd ab->b_cksum0 = 0; 3257168404Spjd ab->b_buf->b_efunc = NULL; 3258168404Spjd ab->b_buf->b_private = NULL; 3259168404Spjd mutex_exit(hash_lock); 3260168404Spjd } else if (refcount_is_zero(&ab->b_refcnt)) { 3261185029Spjd ab->b_flags |= ARC_FREE_IN_PROGRESS; 3262168404Spjd mutex_exit(hash_lock); 3263168404Spjd arc_hdr_destroy(ab); 3264168404Spjd ARCSTAT_BUMP(arcstat_deleted); 3265168404Spjd } else { 3266168404Spjd /* 3267168404Spjd * We still have an active reference on this 3268168404Spjd * buffer. This can happen, e.g., from 3269168404Spjd * dbuf_unoverride(). 3270168404Spjd */ 3271168404Spjd ASSERT(!HDR_IN_HASH_TABLE(ab)); 3272168404Spjd ab->b_arc_access = 0; 3273168404Spjd bzero(&ab->b_dva, sizeof (dva_t)); 3274168404Spjd ab->b_birth = 0; 3275168404Spjd ab->b_cksum0 = 0; 3276168404Spjd ab->b_buf->b_efunc = NULL; 3277168404Spjd ab->b_buf->b_private = NULL; 3278168404Spjd mutex_exit(hash_lock); 3279168404Spjd } 3280168404Spjd } 3281168404Spjd 3282185029Spjd zio = zio_free(pio, spa, txg, bp, done, private, ZIO_FLAG_MUSTSUCCEED); 3283168404Spjd 3284168404Spjd if (arc_flags & ARC_WAIT) 3285168404Spjd return (zio_wait(zio)); 3286168404Spjd 3287168404Spjd ASSERT(arc_flags & ARC_NOWAIT); 3288168404Spjd zio_nowait(zio); 3289168404Spjd 3290168404Spjd return (0); 3291168404Spjd} 3292168404Spjd 3293185029Spjdstatic int 3294185029Spjdarc_memory_throttle(uint64_t reserve, uint64_t txg) 3295185029Spjd{ 3296185029Spjd#ifdef _KERNEL 3297185029Spjd uint64_t inflight_data = arc_anon->arcs_size; 3298185029Spjd uint64_t available_memory = ptoa((uintmax_t)cnt.v_free_count); 3299185029Spjd static uint64_t page_load = 0; 3300185029Spjd static uint64_t last_txg = 0; 3301185029Spjd 3302185029Spjd#if 0 3303185029Spjd#if defined(__i386) 3304185029Spjd available_memory = 3305185029Spjd MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 3306185029Spjd#endif 3307185029Spjd#endif 3308185029Spjd if (available_memory >= zfs_write_limit_max) 3309185029Spjd return (0); 3310185029Spjd 3311185029Spjd if (txg > last_txg) { 3312185029Spjd last_txg = txg; 3313185029Spjd page_load = 0; 3314185029Spjd } 3315185029Spjd /* 3316185029Spjd * If we are in pageout, we know that memory is already tight, 3317185029Spjd * the arc is already going to be evicting, so we just want to 3318185029Spjd * continue to let page writes occur as quickly as possible. 3319185029Spjd */ 3320185029Spjd if (curproc == pageproc) { 3321185029Spjd if (page_load > available_memory / 4) 3322185029Spjd return (ERESTART); 3323185029Spjd /* Note: reserve is inflated, so we deflate */ 3324185029Spjd page_load += reserve / 8; 3325185029Spjd return (0); 3326185029Spjd } else if (page_load > 0 && arc_reclaim_needed()) { 3327185029Spjd /* memory is low, delay before restarting */ 3328185029Spjd ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3329185029Spjd return (EAGAIN); 3330185029Spjd } 3331185029Spjd page_load = 0; 3332185029Spjd 3333185029Spjd if (arc_size > arc_c_min) { 3334185029Spjd uint64_t evictable_memory = 3335185029Spjd arc_mru->arcs_lsize[ARC_BUFC_DATA] + 3336185029Spjd arc_mru->arcs_lsize[ARC_BUFC_METADATA] + 3337185029Spjd arc_mfu->arcs_lsize[ARC_BUFC_DATA] + 3338185029Spjd arc_mfu->arcs_lsize[ARC_BUFC_METADATA]; 3339185029Spjd available_memory += MIN(evictable_memory, arc_size - arc_c_min); 3340185029Spjd } 3341185029Spjd 3342185029Spjd if (inflight_data > available_memory / 4) { 3343185029Spjd ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3344185029Spjd return (ERESTART); 3345185029Spjd } 3346185029Spjd#endif 3347185029Spjd return (0); 3348185029Spjd} 3349185029Spjd 3350168404Spjdvoid 3351185029Spjdarc_tempreserve_clear(uint64_t reserve) 3352168404Spjd{ 3353185029Spjd atomic_add_64(&arc_tempreserve, -reserve); 3354168404Spjd ASSERT((int64_t)arc_tempreserve >= 0); 3355168404Spjd} 3356168404Spjd 3357168404Spjdint 3358185029Spjdarc_tempreserve_space(uint64_t reserve, uint64_t txg) 3359168404Spjd{ 3360185029Spjd int error; 3361185029Spjd 3362168404Spjd#ifdef ZFS_DEBUG 3363168404Spjd /* 3364168404Spjd * Once in a while, fail for no reason. Everything should cope. 3365168404Spjd */ 3366168404Spjd if (spa_get_random(10000) == 0) { 3367168404Spjd dprintf("forcing random failure\n"); 3368168404Spjd return (ERESTART); 3369168404Spjd } 3370168404Spjd#endif 3371185029Spjd if (reserve > arc_c/4 && !arc_no_grow) 3372185029Spjd arc_c = MIN(arc_c_max, reserve * 4); 3373185029Spjd if (reserve > arc_c) 3374168404Spjd return (ENOMEM); 3375168404Spjd 3376168404Spjd /* 3377185029Spjd * Writes will, almost always, require additional memory allocations 3378185029Spjd * in order to compress/encrypt/etc the data. We therefor need to 3379185029Spjd * make sure that there is sufficient available memory for this. 3380185029Spjd */ 3381185029Spjd if (error = arc_memory_throttle(reserve, txg)) 3382185029Spjd return (error); 3383185029Spjd 3384185029Spjd /* 3385168404Spjd * Throttle writes when the amount of dirty data in the cache 3386168404Spjd * gets too large. We try to keep the cache less than half full 3387168404Spjd * of dirty blocks so that our sync times don't grow too large. 3388168404Spjd * Note: if two requests come in concurrently, we might let them 3389168404Spjd * both succeed, when one of them should fail. Not a huge deal. 3390168404Spjd */ 3391185029Spjd if (reserve + arc_tempreserve + arc_anon->arcs_size > arc_c / 2 && 3392185029Spjd arc_anon->arcs_size > arc_c / 4) { 3393185029Spjd dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 3394185029Spjd "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 3395185029Spjd arc_tempreserve>>10, 3396185029Spjd arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 3397185029Spjd arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 3398185029Spjd reserve>>10, arc_c>>10); 3399168404Spjd return (ERESTART); 3400168404Spjd } 3401185029Spjd atomic_add_64(&arc_tempreserve, reserve); 3402168404Spjd return (0); 3403168404Spjd} 3404168404Spjd 3405168582Spjdstatic kmutex_t arc_lowmem_lock; 3406168404Spjd#ifdef _KERNEL 3407168566Spjdstatic eventhandler_tag arc_event_lowmem = NULL; 3408168404Spjd 3409168404Spjdstatic void 3410168566Spjdarc_lowmem(void *arg __unused, int howto __unused) 3411168404Spjd{ 3412168404Spjd 3413168566Spjd /* Serialize access via arc_lowmem_lock. */ 3414168566Spjd mutex_enter(&arc_lowmem_lock); 3415185029Spjd needfree = 1; 3416168404Spjd cv_signal(&arc_reclaim_thr_cv); 3417185029Spjd while (needfree) 3418185029Spjd tsleep(&needfree, 0, "zfs:lowmem", hz / 5); 3419168566Spjd mutex_exit(&arc_lowmem_lock); 3420168404Spjd} 3421168404Spjd#endif 3422168404Spjd 3423168404Spjdvoid 3424168404Spjdarc_init(void) 3425168404Spjd{ 3426168404Spjd mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 3427168404Spjd cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 3428168566Spjd mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 3429168404Spjd 3430168404Spjd /* Convert seconds to clock ticks */ 3431168404Spjd arc_min_prefetch_lifespan = 1 * hz; 3432168404Spjd 3433168404Spjd /* Start out with 1/8 of all memory */ 3434168566Spjd arc_c = kmem_size() / 8; 3435192360Skmacy#if 0 3436192360Skmacy#ifdef _KERNEL 3437192360Skmacy /* 3438192360Skmacy * On architectures where the physical memory can be larger 3439192360Skmacy * than the addressable space (intel in 32-bit mode), we may 3440192360Skmacy * need to limit the cache to 1/8 of VM size. 3441192360Skmacy */ 3442192360Skmacy arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 3443192360Skmacy#endif 3444192360Skmacy#endif 3445168566Spjd /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 3446168566Spjd arc_c_min = MAX(arc_c / 4, 64<<18); 3447168566Spjd /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 3448168404Spjd if (arc_c * 8 >= 1<<30) 3449168404Spjd arc_c_max = (arc_c * 8) - (1<<30); 3450168404Spjd else 3451168404Spjd arc_c_max = arc_c_min; 3452175633Spjd arc_c_max = MAX(arc_c * 5, arc_c_max); 3453168481Spjd#ifdef _KERNEL 3454168404Spjd /* 3455168404Spjd * Allow the tunables to override our calculations if they are 3456168566Spjd * reasonable (ie. over 16MB) 3457168404Spjd */ 3458168566Spjd if (zfs_arc_max >= 64<<18 && zfs_arc_max < kmem_size()) 3459168404Spjd arc_c_max = zfs_arc_max; 3460168566Spjd if (zfs_arc_min >= 64<<18 && zfs_arc_min <= arc_c_max) 3461168404Spjd arc_c_min = zfs_arc_min; 3462168481Spjd#endif 3463168404Spjd arc_c = arc_c_max; 3464168404Spjd arc_p = (arc_c >> 1); 3465168404Spjd 3466185029Spjd /* limit meta-data to 1/4 of the arc capacity */ 3467185029Spjd arc_meta_limit = arc_c_max / 4; 3468185029Spjd 3469185029Spjd /* Allow the tunable to override if it is reasonable */ 3470185029Spjd if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 3471185029Spjd arc_meta_limit = zfs_arc_meta_limit; 3472185029Spjd 3473185029Spjd if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 3474185029Spjd arc_c_min = arc_meta_limit / 2; 3475185029Spjd 3476168404Spjd /* if kmem_flags are set, lets try to use less memory */ 3477168404Spjd if (kmem_debugging()) 3478168404Spjd arc_c = arc_c / 2; 3479168404Spjd if (arc_c < arc_c_min) 3480168404Spjd arc_c = arc_c_min; 3481168404Spjd 3482168473Spjd zfs_arc_min = arc_c_min; 3483168473Spjd zfs_arc_max = arc_c_max; 3484168473Spjd 3485168404Spjd arc_anon = &ARC_anon; 3486168404Spjd arc_mru = &ARC_mru; 3487168404Spjd arc_mru_ghost = &ARC_mru_ghost; 3488168404Spjd arc_mfu = &ARC_mfu; 3489168404Spjd arc_mfu_ghost = &ARC_mfu_ghost; 3490185029Spjd arc_l2c_only = &ARC_l2c_only; 3491168404Spjd arc_size = 0; 3492168404Spjd 3493168404Spjd mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); 3494168404Spjd mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); 3495168404Spjd mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); 3496168404Spjd mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); 3497168404Spjd mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); 3498185029Spjd mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); 3499168404Spjd 3500185029Spjd list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA], 3501185029Spjd sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3502185029Spjd list_create(&arc_mru->arcs_list[ARC_BUFC_DATA], 3503185029Spjd sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3504185029Spjd list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA], 3505185029Spjd sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3506185029Spjd list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA], 3507185029Spjd sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3508185029Spjd list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA], 3509185029Spjd sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3510185029Spjd list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA], 3511185029Spjd sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3512185029Spjd list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA], 3513185029Spjd sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3514185029Spjd list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA], 3515185029Spjd sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3516185029Spjd list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA], 3517185029Spjd sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3518185029Spjd list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA], 3519185029Spjd sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3520168404Spjd 3521168404Spjd buf_init(); 3522168404Spjd 3523168404Spjd arc_thread_exit = 0; 3524168404Spjd arc_eviction_list = NULL; 3525168404Spjd mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 3526168404Spjd bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 3527168404Spjd 3528168404Spjd arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 3529168404Spjd sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 3530168404Spjd 3531168404Spjd if (arc_ksp != NULL) { 3532168404Spjd arc_ksp->ks_data = &arc_stats; 3533168404Spjd kstat_install(arc_ksp); 3534168404Spjd } 3535168404Spjd 3536168404Spjd (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 3537168404Spjd TS_RUN, minclsyspri); 3538168404Spjd 3539168404Spjd#ifdef _KERNEL 3540168566Spjd arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 3541168404Spjd EVENTHANDLER_PRI_FIRST); 3542168404Spjd#endif 3543168404Spjd 3544168404Spjd arc_dead = FALSE; 3545185029Spjd arc_warm = B_FALSE; 3546168566Spjd 3547185029Spjd if (zfs_write_limit_max == 0) 3548185029Spjd zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift; 3549185029Spjd else 3550185029Spjd zfs_write_limit_shift = 0; 3551185029Spjd mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL); 3552185029Spjd 3553168566Spjd#ifdef _KERNEL 3554193878Skmacy#ifdef __i386__ 3555193878Skmacy if (zfs_prefetch_enable != 1) { 3556193878Skmacy printf("ZFS NOTICE: prefetch is disabled by default on i386" 3557193878Skmacy " - add enable to tunable to change.\n" ); 3558193878Skmacy zfs_prefetch_disable=1; 3559193878Skmacy } 3560193878Skmacy#endif 3561193878Skmacy if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 3562193878Skmacy (zfs_prefetch_enable != 1) && (zfs_prefetch_disable != 1)) { 3563193878Skmacy printf("ZFS NOTICE: system has less than 4GB and prefetch enable is not set" 3564193878Skmacy "... disabling.\n"); 3565193878Skmacy zfs_prefetch_disable=1; 3566193878Skmacy } 3567175633Spjd /* Warn about ZFS memory and address space requirements. */ 3568168696Spjd if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 3569168987Sbmah printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 3570168987Sbmah "expect unstable behavior.\n"); 3571175633Spjd } 3572175633Spjd if (kmem_size() < 512 * (1 << 20)) { 3573173419Spjd printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 3574168987Sbmah "expect unstable behavior.\n"); 3575185029Spjd printf(" Consider tuning vm.kmem_size and " 3576173419Spjd "vm.kmem_size_max\n"); 3577185029Spjd printf(" in /boot/loader.conf.\n"); 3578168566Spjd } 3579168566Spjd#endif 3580168404Spjd} 3581168404Spjd 3582168404Spjdvoid 3583168404Spjdarc_fini(void) 3584168404Spjd{ 3585185029Spjd 3586168404Spjd mutex_enter(&arc_reclaim_thr_lock); 3587168404Spjd arc_thread_exit = 1; 3588168404Spjd cv_signal(&arc_reclaim_thr_cv); 3589168404Spjd while (arc_thread_exit != 0) 3590168404Spjd cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 3591168404Spjd mutex_exit(&arc_reclaim_thr_lock); 3592168404Spjd 3593185029Spjd arc_flush(NULL); 3594168404Spjd 3595168404Spjd arc_dead = TRUE; 3596168404Spjd 3597168404Spjd if (arc_ksp != NULL) { 3598168404Spjd kstat_delete(arc_ksp); 3599168404Spjd arc_ksp = NULL; 3600168404Spjd } 3601168404Spjd 3602168404Spjd mutex_destroy(&arc_eviction_mtx); 3603168404Spjd mutex_destroy(&arc_reclaim_thr_lock); 3604168404Spjd cv_destroy(&arc_reclaim_thr_cv); 3605168404Spjd 3606185029Spjd list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]); 3607185029Spjd list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 3608185029Spjd list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]); 3609185029Spjd list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 3610185029Spjd list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]); 3611185029Spjd list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 3612185029Spjd list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]); 3613185029Spjd list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 3614168404Spjd 3615168404Spjd mutex_destroy(&arc_anon->arcs_mtx); 3616168404Spjd mutex_destroy(&arc_mru->arcs_mtx); 3617168404Spjd mutex_destroy(&arc_mru_ghost->arcs_mtx); 3618168404Spjd mutex_destroy(&arc_mfu->arcs_mtx); 3619168404Spjd mutex_destroy(&arc_mfu_ghost->arcs_mtx); 3620168404Spjd 3621185029Spjd mutex_destroy(&zfs_write_limit_lock); 3622185029Spjd 3623168404Spjd buf_fini(); 3624168404Spjd 3625168582Spjd mutex_destroy(&arc_lowmem_lock); 3626168404Spjd#ifdef _KERNEL 3627168566Spjd if (arc_event_lowmem != NULL) 3628168566Spjd EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 3629168404Spjd#endif 3630168404Spjd} 3631185029Spjd 3632185029Spjd/* 3633185029Spjd * Level 2 ARC 3634185029Spjd * 3635185029Spjd * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 3636185029Spjd * It uses dedicated storage devices to hold cached data, which are populated 3637185029Spjd * using large infrequent writes. The main role of this cache is to boost 3638185029Spjd * the performance of random read workloads. The intended L2ARC devices 3639185029Spjd * include short-stroked disks, solid state disks, and other media with 3640185029Spjd * substantially faster read latency than disk. 3641185029Spjd * 3642185029Spjd * +-----------------------+ 3643185029Spjd * | ARC | 3644185029Spjd * +-----------------------+ 3645185029Spjd * | ^ ^ 3646185029Spjd * | | | 3647185029Spjd * l2arc_feed_thread() arc_read() 3648185029Spjd * | | | 3649185029Spjd * | l2arc read | 3650185029Spjd * V | | 3651185029Spjd * +---------------+ | 3652185029Spjd * | L2ARC | | 3653185029Spjd * +---------------+ | 3654185029Spjd * | ^ | 3655185029Spjd * l2arc_write() | | 3656185029Spjd * | | | 3657185029Spjd * V | | 3658185029Spjd * +-------+ +-------+ 3659185029Spjd * | vdev | | vdev | 3660185029Spjd * | cache | | cache | 3661185029Spjd * +-------+ +-------+ 3662185029Spjd * +=========+ .-----. 3663185029Spjd * : L2ARC : |-_____-| 3664185029Spjd * : devices : | Disks | 3665185029Spjd * +=========+ `-_____-' 3666185029Spjd * 3667185029Spjd * Read requests are satisfied from the following sources, in order: 3668185029Spjd * 3669185029Spjd * 1) ARC 3670185029Spjd * 2) vdev cache of L2ARC devices 3671185029Spjd * 3) L2ARC devices 3672185029Spjd * 4) vdev cache of disks 3673185029Spjd * 5) disks 3674185029Spjd * 3675185029Spjd * Some L2ARC device types exhibit extremely slow write performance. 3676185029Spjd * To accommodate for this there are some significant differences between 3677185029Spjd * the L2ARC and traditional cache design: 3678185029Spjd * 3679185029Spjd * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 3680185029Spjd * the ARC behave as usual, freeing buffers and placing headers on ghost 3681185029Spjd * lists. The ARC does not send buffers to the L2ARC during eviction as 3682185029Spjd * this would add inflated write latencies for all ARC memory pressure. 3683185029Spjd * 3684185029Spjd * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 3685185029Spjd * It does this by periodically scanning buffers from the eviction-end of 3686185029Spjd * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 3687185029Spjd * not already there. It scans until a headroom of buffers is satisfied, 3688185029Spjd * which itself is a buffer for ARC eviction. The thread that does this is 3689185029Spjd * l2arc_feed_thread(), illustrated below; example sizes are included to 3690185029Spjd * provide a better sense of ratio than this diagram: 3691185029Spjd * 3692185029Spjd * head --> tail 3693185029Spjd * +---------------------+----------+ 3694185029Spjd * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 3695185029Spjd * +---------------------+----------+ | o L2ARC eligible 3696185029Spjd * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 3697185029Spjd * +---------------------+----------+ | 3698185029Spjd * 15.9 Gbytes ^ 32 Mbytes | 3699185029Spjd * headroom | 3700185029Spjd * l2arc_feed_thread() 3701185029Spjd * | 3702185029Spjd * l2arc write hand <--[oooo]--' 3703185029Spjd * | 8 Mbyte 3704185029Spjd * | write max 3705185029Spjd * V 3706185029Spjd * +==============================+ 3707185029Spjd * L2ARC dev |####|#|###|###| |####| ... | 3708185029Spjd * +==============================+ 3709185029Spjd * 32 Gbytes 3710185029Spjd * 3711185029Spjd * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 3712185029Spjd * evicted, then the L2ARC has cached a buffer much sooner than it probably 3713185029Spjd * needed to, potentially wasting L2ARC device bandwidth and storage. It is 3714185029Spjd * safe to say that this is an uncommon case, since buffers at the end of 3715185029Spjd * the ARC lists have moved there due to inactivity. 3716185029Spjd * 3717185029Spjd * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 3718185029Spjd * then the L2ARC simply misses copying some buffers. This serves as a 3719185029Spjd * pressure valve to prevent heavy read workloads from both stalling the ARC 3720185029Spjd * with waits and clogging the L2ARC with writes. This also helps prevent 3721185029Spjd * the potential for the L2ARC to churn if it attempts to cache content too 3722185029Spjd * quickly, such as during backups of the entire pool. 3723185029Spjd * 3724185029Spjd * 5. After system boot and before the ARC has filled main memory, there are 3725185029Spjd * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 3726185029Spjd * lists can remain mostly static. Instead of searching from tail of these 3727185029Spjd * lists as pictured, the l2arc_feed_thread() will search from the list heads 3728185029Spjd * for eligible buffers, greatly increasing its chance of finding them. 3729185029Spjd * 3730185029Spjd * The L2ARC device write speed is also boosted during this time so that 3731185029Spjd * the L2ARC warms up faster. Since there have been no ARC evictions yet, 3732185029Spjd * there are no L2ARC reads, and no fear of degrading read performance 3733185029Spjd * through increased writes. 3734185029Spjd * 3735185029Spjd * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 3736185029Spjd * the vdev queue can aggregate them into larger and fewer writes. Each 3737185029Spjd * device is written to in a rotor fashion, sweeping writes through 3738185029Spjd * available space then repeating. 3739185029Spjd * 3740185029Spjd * 7. The L2ARC does not store dirty content. It never needs to flush 3741185029Spjd * write buffers back to disk based storage. 3742185029Spjd * 3743185029Spjd * 8. If an ARC buffer is written (and dirtied) which also exists in the 3744185029Spjd * L2ARC, the now stale L2ARC buffer is immediately dropped. 3745185029Spjd * 3746185029Spjd * The performance of the L2ARC can be tweaked by a number of tunables, which 3747185029Spjd * may be necessary for different workloads: 3748185029Spjd * 3749185029Spjd * l2arc_write_max max write bytes per interval 3750185029Spjd * l2arc_write_boost extra write bytes during device warmup 3751185029Spjd * l2arc_noprefetch skip caching prefetched buffers 3752185029Spjd * l2arc_headroom number of max device writes to precache 3753185029Spjd * l2arc_feed_secs seconds between L2ARC writing 3754185029Spjd * 3755185029Spjd * Tunables may be removed or added as future performance improvements are 3756185029Spjd * integrated, and also may become zpool properties. 3757185029Spjd */ 3758185029Spjd 3759185029Spjdstatic void 3760185029Spjdl2arc_hdr_stat_add(void) 3761185029Spjd{ 3762185029Spjd ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 3763185029Spjd ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 3764185029Spjd} 3765185029Spjd 3766185029Spjdstatic void 3767185029Spjdl2arc_hdr_stat_remove(void) 3768185029Spjd{ 3769185029Spjd ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 3770185029Spjd ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 3771185029Spjd} 3772185029Spjd 3773185029Spjd/* 3774185029Spjd * Cycle through L2ARC devices. This is how L2ARC load balances. 3775185029Spjd * If a device is returned, this also returns holding the spa config lock. 3776185029Spjd */ 3777185029Spjdstatic l2arc_dev_t * 3778185029Spjdl2arc_dev_get_next(void) 3779185029Spjd{ 3780185029Spjd l2arc_dev_t *first, *next = NULL; 3781185029Spjd 3782185029Spjd /* 3783185029Spjd * Lock out the removal of spas (spa_namespace_lock), then removal 3784185029Spjd * of cache devices (l2arc_dev_mtx). Once a device has been selected, 3785185029Spjd * both locks will be dropped and a spa config lock held instead. 3786185029Spjd */ 3787185029Spjd mutex_enter(&spa_namespace_lock); 3788185029Spjd mutex_enter(&l2arc_dev_mtx); 3789185029Spjd 3790185029Spjd /* if there are no vdevs, there is nothing to do */ 3791185029Spjd if (l2arc_ndev == 0) 3792185029Spjd goto out; 3793185029Spjd 3794185029Spjd first = NULL; 3795185029Spjd next = l2arc_dev_last; 3796185029Spjd do { 3797185029Spjd /* loop around the list looking for a non-faulted vdev */ 3798185029Spjd if (next == NULL) { 3799185029Spjd next = list_head(l2arc_dev_list); 3800185029Spjd } else { 3801185029Spjd next = list_next(l2arc_dev_list, next); 3802185029Spjd if (next == NULL) 3803185029Spjd next = list_head(l2arc_dev_list); 3804185029Spjd } 3805185029Spjd 3806185029Spjd /* if we have come back to the start, bail out */ 3807185029Spjd if (first == NULL) 3808185029Spjd first = next; 3809185029Spjd else if (next == first) 3810185029Spjd break; 3811185029Spjd 3812185029Spjd } while (vdev_is_dead(next->l2ad_vdev)); 3813185029Spjd 3814185029Spjd /* if we were unable to find any usable vdevs, return NULL */ 3815185029Spjd if (vdev_is_dead(next->l2ad_vdev)) 3816185029Spjd next = NULL; 3817185029Spjd 3818185029Spjd l2arc_dev_last = next; 3819185029Spjd 3820185029Spjdout: 3821185029Spjd mutex_exit(&l2arc_dev_mtx); 3822185029Spjd 3823185029Spjd /* 3824185029Spjd * Grab the config lock to prevent the 'next' device from being 3825185029Spjd * removed while we are writing to it. 3826185029Spjd */ 3827185029Spjd if (next != NULL) 3828185029Spjd spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 3829185029Spjd mutex_exit(&spa_namespace_lock); 3830185029Spjd 3831185029Spjd return (next); 3832185029Spjd} 3833185029Spjd 3834185029Spjd/* 3835185029Spjd * Free buffers that were tagged for destruction. 3836185029Spjd */ 3837185029Spjdstatic void 3838185029Spjdl2arc_do_free_on_write() 3839185029Spjd{ 3840185029Spjd list_t *buflist; 3841185029Spjd l2arc_data_free_t *df, *df_prev; 3842185029Spjd 3843185029Spjd mutex_enter(&l2arc_free_on_write_mtx); 3844185029Spjd buflist = l2arc_free_on_write; 3845185029Spjd 3846185029Spjd for (df = list_tail(buflist); df; df = df_prev) { 3847185029Spjd df_prev = list_prev(buflist, df); 3848185029Spjd ASSERT(df->l2df_data != NULL); 3849185029Spjd ASSERT(df->l2df_func != NULL); 3850185029Spjd df->l2df_func(df->l2df_data, df->l2df_size); 3851185029Spjd list_remove(buflist, df); 3852185029Spjd kmem_free(df, sizeof (l2arc_data_free_t)); 3853185029Spjd } 3854185029Spjd 3855185029Spjd mutex_exit(&l2arc_free_on_write_mtx); 3856185029Spjd} 3857185029Spjd 3858185029Spjd/* 3859185029Spjd * A write to a cache device has completed. Update all headers to allow 3860185029Spjd * reads from these buffers to begin. 3861185029Spjd */ 3862185029Spjdstatic void 3863185029Spjdl2arc_write_done(zio_t *zio) 3864185029Spjd{ 3865185029Spjd l2arc_write_callback_t *cb; 3866185029Spjd l2arc_dev_t *dev; 3867185029Spjd list_t *buflist; 3868185029Spjd arc_buf_hdr_t *head, *ab, *ab_prev; 3869185029Spjd l2arc_buf_hdr_t *abl2; 3870185029Spjd kmutex_t *hash_lock; 3871185029Spjd 3872185029Spjd cb = zio->io_private; 3873185029Spjd ASSERT(cb != NULL); 3874185029Spjd dev = cb->l2wcb_dev; 3875185029Spjd ASSERT(dev != NULL); 3876185029Spjd head = cb->l2wcb_head; 3877185029Spjd ASSERT(head != NULL); 3878185029Spjd buflist = dev->l2ad_buflist; 3879185029Spjd ASSERT(buflist != NULL); 3880185029Spjd DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 3881185029Spjd l2arc_write_callback_t *, cb); 3882185029Spjd 3883185029Spjd if (zio->io_error != 0) 3884185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_error); 3885185029Spjd 3886185029Spjd mutex_enter(&l2arc_buflist_mtx); 3887185029Spjd 3888185029Spjd /* 3889185029Spjd * All writes completed, or an error was hit. 3890185029Spjd */ 3891185029Spjd for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 3892185029Spjd ab_prev = list_prev(buflist, ab); 3893185029Spjd 3894185029Spjd hash_lock = HDR_LOCK(ab); 3895185029Spjd if (!mutex_tryenter(hash_lock)) { 3896185029Spjd /* 3897185029Spjd * This buffer misses out. It may be in a stage 3898185029Spjd * of eviction. Its ARC_L2_WRITING flag will be 3899185029Spjd * left set, denying reads to this buffer. 3900185029Spjd */ 3901185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 3902185029Spjd continue; 3903185029Spjd } 3904185029Spjd 3905185029Spjd if (zio->io_error != 0) { 3906185029Spjd /* 3907185029Spjd * Error - drop L2ARC entry. 3908185029Spjd */ 3909185029Spjd list_remove(buflist, ab); 3910185029Spjd abl2 = ab->b_l2hdr; 3911185029Spjd ab->b_l2hdr = NULL; 3912185029Spjd kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 3913185029Spjd ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 3914185029Spjd } 3915185029Spjd 3916185029Spjd /* 3917185029Spjd * Allow ARC to begin reads to this L2ARC entry. 3918185029Spjd */ 3919185029Spjd ab->b_flags &= ~ARC_L2_WRITING; 3920185029Spjd 3921185029Spjd mutex_exit(hash_lock); 3922185029Spjd } 3923185029Spjd 3924185029Spjd atomic_inc_64(&l2arc_writes_done); 3925185029Spjd list_remove(buflist, head); 3926185029Spjd kmem_cache_free(hdr_cache, head); 3927185029Spjd mutex_exit(&l2arc_buflist_mtx); 3928185029Spjd 3929185029Spjd l2arc_do_free_on_write(); 3930185029Spjd 3931185029Spjd kmem_free(cb, sizeof (l2arc_write_callback_t)); 3932185029Spjd} 3933185029Spjd 3934185029Spjd/* 3935185029Spjd * A read to a cache device completed. Validate buffer contents before 3936185029Spjd * handing over to the regular ARC routines. 3937185029Spjd */ 3938185029Spjdstatic void 3939185029Spjdl2arc_read_done(zio_t *zio) 3940185029Spjd{ 3941185029Spjd l2arc_read_callback_t *cb; 3942185029Spjd arc_buf_hdr_t *hdr; 3943185029Spjd arc_buf_t *buf; 3944185029Spjd kmutex_t *hash_lock; 3945185029Spjd int equal; 3946185029Spjd 3947185029Spjd ASSERT(zio->io_vd != NULL); 3948185029Spjd ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 3949185029Spjd 3950185029Spjd spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 3951185029Spjd 3952185029Spjd cb = zio->io_private; 3953185029Spjd ASSERT(cb != NULL); 3954185029Spjd buf = cb->l2rcb_buf; 3955185029Spjd ASSERT(buf != NULL); 3956185029Spjd hdr = buf->b_hdr; 3957185029Spjd ASSERT(hdr != NULL); 3958185029Spjd 3959185029Spjd hash_lock = HDR_LOCK(hdr); 3960185029Spjd mutex_enter(hash_lock); 3961185029Spjd 3962185029Spjd /* 3963185029Spjd * Check this survived the L2ARC journey. 3964185029Spjd */ 3965185029Spjd equal = arc_cksum_equal(buf); 3966185029Spjd if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 3967185029Spjd mutex_exit(hash_lock); 3968185029Spjd zio->io_private = buf; 3969185029Spjd zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 3970185029Spjd zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 3971185029Spjd arc_read_done(zio); 3972185029Spjd } else { 3973185029Spjd mutex_exit(hash_lock); 3974185029Spjd /* 3975185029Spjd * Buffer didn't survive caching. Increment stats and 3976185029Spjd * reissue to the original storage device. 3977185029Spjd */ 3978185029Spjd if (zio->io_error != 0) { 3979185029Spjd ARCSTAT_BUMP(arcstat_l2_io_error); 3980185029Spjd } else { 3981185029Spjd zio->io_error = EIO; 3982185029Spjd } 3983185029Spjd if (!equal) 3984185029Spjd ARCSTAT_BUMP(arcstat_l2_cksum_bad); 3985185029Spjd 3986185029Spjd /* 3987185029Spjd * If there's no waiter, issue an async i/o to the primary 3988185029Spjd * storage now. If there *is* a waiter, the caller must 3989185029Spjd * issue the i/o in a context where it's OK to block. 3990185029Spjd */ 3991185029Spjd if (zio->io_waiter == NULL) 3992185029Spjd zio_nowait(zio_read(zio->io_parent, 3993185029Spjd cb->l2rcb_spa, &cb->l2rcb_bp, 3994185029Spjd buf->b_data, zio->io_size, arc_read_done, buf, 3995185029Spjd zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 3996185029Spjd } 3997185029Spjd 3998185029Spjd kmem_free(cb, sizeof (l2arc_read_callback_t)); 3999185029Spjd} 4000185029Spjd 4001185029Spjd/* 4002185029Spjd * This is the list priority from which the L2ARC will search for pages to 4003185029Spjd * cache. This is used within loops (0..3) to cycle through lists in the 4004185029Spjd * desired order. This order can have a significant effect on cache 4005185029Spjd * performance. 4006185029Spjd * 4007185029Spjd * Currently the metadata lists are hit first, MFU then MRU, followed by 4008185029Spjd * the data lists. This function returns a locked list, and also returns 4009185029Spjd * the lock pointer. 4010185029Spjd */ 4011185029Spjdstatic list_t * 4012185029Spjdl2arc_list_locked(int list_num, kmutex_t **lock) 4013185029Spjd{ 4014185029Spjd list_t *list; 4015185029Spjd 4016185029Spjd ASSERT(list_num >= 0 && list_num <= 3); 4017185029Spjd 4018185029Spjd switch (list_num) { 4019185029Spjd case 0: 4020185029Spjd list = &arc_mfu->arcs_list[ARC_BUFC_METADATA]; 4021185029Spjd *lock = &arc_mfu->arcs_mtx; 4022185029Spjd break; 4023185029Spjd case 1: 4024185029Spjd list = &arc_mru->arcs_list[ARC_BUFC_METADATA]; 4025185029Spjd *lock = &arc_mru->arcs_mtx; 4026185029Spjd break; 4027185029Spjd case 2: 4028185029Spjd list = &arc_mfu->arcs_list[ARC_BUFC_DATA]; 4029185029Spjd *lock = &arc_mfu->arcs_mtx; 4030185029Spjd break; 4031185029Spjd case 3: 4032185029Spjd list = &arc_mru->arcs_list[ARC_BUFC_DATA]; 4033185029Spjd *lock = &arc_mru->arcs_mtx; 4034185029Spjd break; 4035185029Spjd } 4036185029Spjd 4037185029Spjd ASSERT(!(MUTEX_HELD(*lock))); 4038185029Spjd mutex_enter(*lock); 4039185029Spjd return (list); 4040185029Spjd} 4041185029Spjd 4042185029Spjd/* 4043185029Spjd * Evict buffers from the device write hand to the distance specified in 4044185029Spjd * bytes. This distance may span populated buffers, it may span nothing. 4045185029Spjd * This is clearing a region on the L2ARC device ready for writing. 4046185029Spjd * If the 'all' boolean is set, every buffer is evicted. 4047185029Spjd */ 4048185029Spjdstatic void 4049185029Spjdl2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4050185029Spjd{ 4051185029Spjd list_t *buflist; 4052185029Spjd l2arc_buf_hdr_t *abl2; 4053185029Spjd arc_buf_hdr_t *ab, *ab_prev; 4054185029Spjd kmutex_t *hash_lock; 4055185029Spjd uint64_t taddr; 4056185029Spjd 4057185029Spjd buflist = dev->l2ad_buflist; 4058185029Spjd 4059185029Spjd if (buflist == NULL) 4060185029Spjd return; 4061185029Spjd 4062185029Spjd if (!all && dev->l2ad_first) { 4063185029Spjd /* 4064185029Spjd * This is the first sweep through the device. There is 4065185029Spjd * nothing to evict. 4066185029Spjd */ 4067185029Spjd return; 4068185029Spjd } 4069185029Spjd 4070185029Spjd if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4071185029Spjd /* 4072185029Spjd * When nearing the end of the device, evict to the end 4073185029Spjd * before the device write hand jumps to the start. 4074185029Spjd */ 4075185029Spjd taddr = dev->l2ad_end; 4076185029Spjd } else { 4077185029Spjd taddr = dev->l2ad_hand + distance; 4078185029Spjd } 4079185029Spjd DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4080185029Spjd uint64_t, taddr, boolean_t, all); 4081185029Spjd 4082185029Spjdtop: 4083185029Spjd mutex_enter(&l2arc_buflist_mtx); 4084185029Spjd for (ab = list_tail(buflist); ab; ab = ab_prev) { 4085185029Spjd ab_prev = list_prev(buflist, ab); 4086185029Spjd 4087185029Spjd hash_lock = HDR_LOCK(ab); 4088185029Spjd if (!mutex_tryenter(hash_lock)) { 4089185029Spjd /* 4090185029Spjd * Missed the hash lock. Retry. 4091185029Spjd */ 4092185029Spjd ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4093185029Spjd mutex_exit(&l2arc_buflist_mtx); 4094185029Spjd mutex_enter(hash_lock); 4095185029Spjd mutex_exit(hash_lock); 4096185029Spjd goto top; 4097185029Spjd } 4098185029Spjd 4099185029Spjd if (HDR_L2_WRITE_HEAD(ab)) { 4100185029Spjd /* 4101185029Spjd * We hit a write head node. Leave it for 4102185029Spjd * l2arc_write_done(). 4103185029Spjd */ 4104185029Spjd list_remove(buflist, ab); 4105185029Spjd mutex_exit(hash_lock); 4106185029Spjd continue; 4107185029Spjd } 4108185029Spjd 4109185029Spjd if (!all && ab->b_l2hdr != NULL && 4110185029Spjd (ab->b_l2hdr->b_daddr > taddr || 4111185029Spjd ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4112185029Spjd /* 4113185029Spjd * We've evicted to the target address, 4114185029Spjd * or the end of the device. 4115185029Spjd */ 4116185029Spjd mutex_exit(hash_lock); 4117185029Spjd break; 4118185029Spjd } 4119185029Spjd 4120185029Spjd if (HDR_FREE_IN_PROGRESS(ab)) { 4121185029Spjd /* 4122185029Spjd * Already on the path to destruction. 4123185029Spjd */ 4124185029Spjd mutex_exit(hash_lock); 4125185029Spjd continue; 4126185029Spjd } 4127185029Spjd 4128185029Spjd if (ab->b_state == arc_l2c_only) { 4129185029Spjd ASSERT(!HDR_L2_READING(ab)); 4130185029Spjd /* 4131185029Spjd * This doesn't exist in the ARC. Destroy. 4132185029Spjd * arc_hdr_destroy() will call list_remove() 4133185029Spjd * and decrement arcstat_l2_size. 4134185029Spjd */ 4135185029Spjd arc_change_state(arc_anon, ab, hash_lock); 4136185029Spjd arc_hdr_destroy(ab); 4137185029Spjd } else { 4138185029Spjd /* 4139185029Spjd * Invalidate issued or about to be issued 4140185029Spjd * reads, since we may be about to write 4141185029Spjd * over this location. 4142185029Spjd */ 4143185029Spjd if (HDR_L2_READING(ab)) { 4144185029Spjd ARCSTAT_BUMP(arcstat_l2_evict_reading); 4145185029Spjd ab->b_flags |= ARC_L2_EVICTED; 4146185029Spjd } 4147185029Spjd 4148185029Spjd /* 4149185029Spjd * Tell ARC this no longer exists in L2ARC. 4150185029Spjd */ 4151185029Spjd if (ab->b_l2hdr != NULL) { 4152185029Spjd abl2 = ab->b_l2hdr; 4153185029Spjd ab->b_l2hdr = NULL; 4154185029Spjd kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4155185029Spjd ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4156185029Spjd } 4157185029Spjd list_remove(buflist, ab); 4158185029Spjd 4159185029Spjd /* 4160185029Spjd * This may have been leftover after a 4161185029Spjd * failed write. 4162185029Spjd */ 4163185029Spjd ab->b_flags &= ~ARC_L2_WRITING; 4164185029Spjd } 4165185029Spjd mutex_exit(hash_lock); 4166185029Spjd } 4167185029Spjd mutex_exit(&l2arc_buflist_mtx); 4168185029Spjd 4169185029Spjd spa_l2cache_space_update(dev->l2ad_vdev, 0, -(taddr - dev->l2ad_evict)); 4170185029Spjd dev->l2ad_evict = taddr; 4171185029Spjd} 4172185029Spjd 4173185029Spjd/* 4174185029Spjd * Find and write ARC buffers to the L2ARC device. 4175185029Spjd * 4176185029Spjd * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 4177185029Spjd * for reading until they have completed writing. 4178185029Spjd */ 4179185029Spjdstatic void 4180185029Spjdl2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) 4181185029Spjd{ 4182185029Spjd arc_buf_hdr_t *ab, *ab_prev, *head; 4183185029Spjd l2arc_buf_hdr_t *hdrl2; 4184185029Spjd list_t *list; 4185185029Spjd uint64_t passed_sz, write_sz, buf_sz, headroom; 4186185029Spjd void *buf_data; 4187185029Spjd kmutex_t *hash_lock, *list_lock; 4188185029Spjd boolean_t have_lock, full; 4189185029Spjd l2arc_write_callback_t *cb; 4190185029Spjd zio_t *pio, *wzio; 4191185029Spjd int try; 4192185029Spjd 4193185029Spjd ASSERT(dev->l2ad_vdev != NULL); 4194185029Spjd 4195185029Spjd pio = NULL; 4196185029Spjd write_sz = 0; 4197185029Spjd full = B_FALSE; 4198185029Spjd head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 4199185029Spjd head->b_flags |= ARC_L2_WRITE_HEAD; 4200185029Spjd 4201185029Spjd /* 4202185029Spjd * Copy buffers for L2ARC writing. 4203185029Spjd */ 4204185029Spjd mutex_enter(&l2arc_buflist_mtx); 4205185029Spjd for (try = 0; try <= 3; try++) { 4206185029Spjd list = l2arc_list_locked(try, &list_lock); 4207185029Spjd passed_sz = 0; 4208185029Spjd 4209185029Spjd /* 4210185029Spjd * L2ARC fast warmup. 4211185029Spjd * 4212185029Spjd * Until the ARC is warm and starts to evict, read from the 4213185029Spjd * head of the ARC lists rather than the tail. 4214185029Spjd */ 4215185029Spjd headroom = target_sz * l2arc_headroom; 4216185029Spjd if (arc_warm == B_FALSE) 4217185029Spjd ab = list_head(list); 4218185029Spjd else 4219185029Spjd ab = list_tail(list); 4220185029Spjd 4221185029Spjd for (; ab; ab = ab_prev) { 4222185029Spjd if (arc_warm == B_FALSE) 4223185029Spjd ab_prev = list_next(list, ab); 4224185029Spjd else 4225185029Spjd ab_prev = list_prev(list, ab); 4226185029Spjd 4227185029Spjd hash_lock = HDR_LOCK(ab); 4228185029Spjd have_lock = MUTEX_HELD(hash_lock); 4229185029Spjd if (!have_lock && !mutex_tryenter(hash_lock)) { 4230185029Spjd /* 4231185029Spjd * Skip this buffer rather than waiting. 4232185029Spjd */ 4233185029Spjd continue; 4234185029Spjd } 4235185029Spjd 4236185029Spjd passed_sz += ab->b_size; 4237185029Spjd if (passed_sz > headroom) { 4238185029Spjd /* 4239185029Spjd * Searched too far. 4240185029Spjd */ 4241185029Spjd mutex_exit(hash_lock); 4242185029Spjd break; 4243185029Spjd } 4244185029Spjd 4245185029Spjd if (ab->b_spa != spa) { 4246185029Spjd mutex_exit(hash_lock); 4247185029Spjd continue; 4248185029Spjd } 4249185029Spjd 4250185029Spjd if (ab->b_l2hdr != NULL) { 4251185029Spjd /* 4252185029Spjd * Already in L2ARC. 4253185029Spjd */ 4254185029Spjd mutex_exit(hash_lock); 4255185029Spjd continue; 4256185029Spjd } 4257185029Spjd 4258185029Spjd if (HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab)) { 4259185029Spjd mutex_exit(hash_lock); 4260185029Spjd continue; 4261185029Spjd } 4262185029Spjd 4263185029Spjd if ((write_sz + ab->b_size) > target_sz) { 4264185029Spjd full = B_TRUE; 4265185029Spjd mutex_exit(hash_lock); 4266185029Spjd break; 4267185029Spjd } 4268185029Spjd 4269185029Spjd if (ab->b_buf == NULL) { 4270185029Spjd DTRACE_PROBE1(l2arc__buf__null, void *, ab); 4271185029Spjd mutex_exit(hash_lock); 4272185029Spjd continue; 4273185029Spjd } 4274185029Spjd 4275185029Spjd if (pio == NULL) { 4276185029Spjd /* 4277185029Spjd * Insert a dummy header on the buflist so 4278185029Spjd * l2arc_write_done() can find where the 4279185029Spjd * write buffers begin without searching. 4280185029Spjd */ 4281185029Spjd list_insert_head(dev->l2ad_buflist, head); 4282185029Spjd 4283185029Spjd cb = kmem_alloc( 4284185029Spjd sizeof (l2arc_write_callback_t), KM_SLEEP); 4285185029Spjd cb->l2wcb_dev = dev; 4286185029Spjd cb->l2wcb_head = head; 4287185029Spjd pio = zio_root(spa, l2arc_write_done, cb, 4288185029Spjd ZIO_FLAG_CANFAIL); 4289185029Spjd } 4290185029Spjd 4291185029Spjd /* 4292185029Spjd * Create and add a new L2ARC header. 4293185029Spjd */ 4294185029Spjd hdrl2 = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 4295185029Spjd hdrl2->b_dev = dev; 4296185029Spjd hdrl2->b_daddr = dev->l2ad_hand; 4297185029Spjd 4298185029Spjd ab->b_flags |= ARC_L2_WRITING; 4299185029Spjd ab->b_l2hdr = hdrl2; 4300185029Spjd list_insert_head(dev->l2ad_buflist, ab); 4301185029Spjd buf_data = ab->b_buf->b_data; 4302185029Spjd buf_sz = ab->b_size; 4303185029Spjd 4304185029Spjd /* 4305185029Spjd * Compute and store the buffer cksum before 4306185029Spjd * writing. On debug the cksum is verified first. 4307185029Spjd */ 4308185029Spjd arc_cksum_verify(ab->b_buf); 4309185029Spjd arc_cksum_compute(ab->b_buf, B_TRUE); 4310185029Spjd 4311185029Spjd mutex_exit(hash_lock); 4312185029Spjd 4313185029Spjd wzio = zio_write_phys(pio, dev->l2ad_vdev, 4314185029Spjd dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 4315185029Spjd NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 4316185029Spjd ZIO_FLAG_CANFAIL, B_FALSE); 4317185029Spjd 4318185029Spjd DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 4319185029Spjd zio_t *, wzio); 4320185029Spjd (void) zio_nowait(wzio); 4321185029Spjd 4322185029Spjd /* 4323185029Spjd * Keep the clock hand suitably device-aligned. 4324185029Spjd */ 4325185029Spjd buf_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 4326185029Spjd 4327185029Spjd write_sz += buf_sz; 4328185029Spjd dev->l2ad_hand += buf_sz; 4329185029Spjd } 4330185029Spjd 4331185029Spjd mutex_exit(list_lock); 4332185029Spjd 4333185029Spjd if (full == B_TRUE) 4334185029Spjd break; 4335185029Spjd } 4336185029Spjd mutex_exit(&l2arc_buflist_mtx); 4337185029Spjd 4338185029Spjd if (pio == NULL) { 4339185029Spjd ASSERT3U(write_sz, ==, 0); 4340185029Spjd kmem_cache_free(hdr_cache, head); 4341185029Spjd return; 4342185029Spjd } 4343185029Spjd 4344185029Spjd ASSERT3U(write_sz, <=, target_sz); 4345185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_sent); 4346185029Spjd ARCSTAT_INCR(arcstat_l2_size, write_sz); 4347185029Spjd spa_l2cache_space_update(dev->l2ad_vdev, 0, write_sz); 4348185029Spjd 4349185029Spjd /* 4350185029Spjd * Bump device hand to the device start if it is approaching the end. 4351185029Spjd * l2arc_evict() will already have evicted ahead for this case. 4352185029Spjd */ 4353185029Spjd if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 4354185029Spjd spa_l2cache_space_update(dev->l2ad_vdev, 0, 4355185029Spjd dev->l2ad_end - dev->l2ad_hand); 4356185029Spjd dev->l2ad_hand = dev->l2ad_start; 4357185029Spjd dev->l2ad_evict = dev->l2ad_start; 4358185029Spjd dev->l2ad_first = B_FALSE; 4359185029Spjd } 4360185029Spjd 4361185029Spjd (void) zio_wait(pio); 4362185029Spjd} 4363185029Spjd 4364185029Spjd/* 4365185029Spjd * This thread feeds the L2ARC at regular intervals. This is the beating 4366185029Spjd * heart of the L2ARC. 4367185029Spjd */ 4368185029Spjdstatic void 4369185029Spjdl2arc_feed_thread(void *dummy __unused) 4370185029Spjd{ 4371185029Spjd callb_cpr_t cpr; 4372185029Spjd l2arc_dev_t *dev; 4373185029Spjd spa_t *spa; 4374185029Spjd uint64_t size; 4375185029Spjd 4376185029Spjd CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 4377185029Spjd 4378185029Spjd mutex_enter(&l2arc_feed_thr_lock); 4379185029Spjd 4380185029Spjd while (l2arc_thread_exit == 0) { 4381185029Spjd /* 4382185029Spjd * Pause for l2arc_feed_secs seconds between writes. 4383185029Spjd */ 4384185029Spjd CALLB_CPR_SAFE_BEGIN(&cpr); 4385185029Spjd (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 4386185029Spjd hz * l2arc_feed_secs); 4387185029Spjd CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 4388185029Spjd 4389185029Spjd /* 4390185029Spjd * Quick check for L2ARC devices. 4391185029Spjd */ 4392185029Spjd mutex_enter(&l2arc_dev_mtx); 4393185029Spjd if (l2arc_ndev == 0) { 4394185029Spjd mutex_exit(&l2arc_dev_mtx); 4395185029Spjd continue; 4396185029Spjd } 4397185029Spjd mutex_exit(&l2arc_dev_mtx); 4398185029Spjd 4399185029Spjd /* 4400185029Spjd * This selects the next l2arc device to write to, and in 4401185029Spjd * doing so the next spa to feed from: dev->l2ad_spa. This 4402185029Spjd * will return NULL if there are now no l2arc devices or if 4403185029Spjd * they are all faulted. 4404185029Spjd * 4405185029Spjd * If a device is returned, its spa's config lock is also 4406185029Spjd * held to prevent device removal. l2arc_dev_get_next() 4407185029Spjd * will grab and release l2arc_dev_mtx. 4408185029Spjd */ 4409185029Spjd if ((dev = l2arc_dev_get_next()) == NULL) 4410185029Spjd continue; 4411185029Spjd 4412185029Spjd spa = dev->l2ad_spa; 4413185029Spjd ASSERT(spa != NULL); 4414185029Spjd 4415185029Spjd /* 4416185029Spjd * Avoid contributing to memory pressure. 4417185029Spjd */ 4418185029Spjd if (arc_reclaim_needed()) { 4419185029Spjd ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 4420185029Spjd spa_config_exit(spa, SCL_L2ARC, dev); 4421185029Spjd continue; 4422185029Spjd } 4423185029Spjd 4424185029Spjd ARCSTAT_BUMP(arcstat_l2_feeds); 4425185029Spjd 4426185029Spjd size = dev->l2ad_write; 4427185029Spjd if (arc_warm == B_FALSE) 4428185029Spjd size += dev->l2ad_boost; 4429185029Spjd 4430185029Spjd /* 4431185029Spjd * Evict L2ARC buffers that will be overwritten. 4432185029Spjd */ 4433185029Spjd l2arc_evict(dev, size, B_FALSE); 4434185029Spjd 4435185029Spjd /* 4436185029Spjd * Write ARC buffers. 4437185029Spjd */ 4438185029Spjd l2arc_write_buffers(spa, dev, size); 4439185029Spjd spa_config_exit(spa, SCL_L2ARC, dev); 4440185029Spjd } 4441185029Spjd 4442185029Spjd l2arc_thread_exit = 0; 4443185029Spjd cv_broadcast(&l2arc_feed_thr_cv); 4444185029Spjd CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 4445185029Spjd thread_exit(); 4446185029Spjd} 4447185029Spjd 4448185029Spjdboolean_t 4449185029Spjdl2arc_vdev_present(vdev_t *vd) 4450185029Spjd{ 4451185029Spjd l2arc_dev_t *dev; 4452185029Spjd 4453185029Spjd mutex_enter(&l2arc_dev_mtx); 4454185029Spjd for (dev = list_head(l2arc_dev_list); dev != NULL; 4455185029Spjd dev = list_next(l2arc_dev_list, dev)) { 4456185029Spjd if (dev->l2ad_vdev == vd) 4457185029Spjd break; 4458185029Spjd } 4459185029Spjd mutex_exit(&l2arc_dev_mtx); 4460185029Spjd 4461185029Spjd return (dev != NULL); 4462185029Spjd} 4463185029Spjd 4464185029Spjd/* 4465185029Spjd * Add a vdev for use by the L2ARC. By this point the spa has already 4466185029Spjd * validated the vdev and opened it. 4467185029Spjd */ 4468185029Spjdvoid 4469185029Spjdl2arc_add_vdev(spa_t *spa, vdev_t *vd, uint64_t start, uint64_t end) 4470185029Spjd{ 4471185029Spjd l2arc_dev_t *adddev; 4472185029Spjd 4473185029Spjd ASSERT(!l2arc_vdev_present(vd)); 4474185029Spjd 4475185029Spjd /* 4476185029Spjd * Create a new l2arc device entry. 4477185029Spjd */ 4478185029Spjd adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 4479185029Spjd adddev->l2ad_spa = spa; 4480185029Spjd adddev->l2ad_vdev = vd; 4481185029Spjd adddev->l2ad_write = l2arc_write_max; 4482185029Spjd adddev->l2ad_boost = l2arc_write_boost; 4483185029Spjd adddev->l2ad_start = start; 4484185029Spjd adddev->l2ad_end = end; 4485185029Spjd adddev->l2ad_hand = adddev->l2ad_start; 4486185029Spjd adddev->l2ad_evict = adddev->l2ad_start; 4487185029Spjd adddev->l2ad_first = B_TRUE; 4488185029Spjd ASSERT3U(adddev->l2ad_write, >, 0); 4489185029Spjd 4490185029Spjd /* 4491185029Spjd * This is a list of all ARC buffers that are still valid on the 4492185029Spjd * device. 4493185029Spjd */ 4494185029Spjd adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 4495185029Spjd list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 4496185029Spjd offsetof(arc_buf_hdr_t, b_l2node)); 4497185029Spjd 4498185029Spjd spa_l2cache_space_update(vd, adddev->l2ad_end - adddev->l2ad_hand, 0); 4499185029Spjd 4500185029Spjd /* 4501185029Spjd * Add device to global list 4502185029Spjd */ 4503185029Spjd mutex_enter(&l2arc_dev_mtx); 4504185029Spjd list_insert_head(l2arc_dev_list, adddev); 4505185029Spjd atomic_inc_64(&l2arc_ndev); 4506185029Spjd mutex_exit(&l2arc_dev_mtx); 4507185029Spjd} 4508185029Spjd 4509185029Spjd/* 4510185029Spjd * Remove a vdev from the L2ARC. 4511185029Spjd */ 4512185029Spjdvoid 4513185029Spjdl2arc_remove_vdev(vdev_t *vd) 4514185029Spjd{ 4515185029Spjd l2arc_dev_t *dev, *nextdev, *remdev = NULL; 4516185029Spjd 4517185029Spjd /* 4518185029Spjd * Find the device by vdev 4519185029Spjd */ 4520185029Spjd mutex_enter(&l2arc_dev_mtx); 4521185029Spjd for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 4522185029Spjd nextdev = list_next(l2arc_dev_list, dev); 4523185029Spjd if (vd == dev->l2ad_vdev) { 4524185029Spjd remdev = dev; 4525185029Spjd break; 4526185029Spjd } 4527185029Spjd } 4528185029Spjd ASSERT(remdev != NULL); 4529185029Spjd 4530185029Spjd /* 4531185029Spjd * Remove device from global list 4532185029Spjd */ 4533185029Spjd list_remove(l2arc_dev_list, remdev); 4534185029Spjd l2arc_dev_last = NULL; /* may have been invalidated */ 4535185029Spjd atomic_dec_64(&l2arc_ndev); 4536185029Spjd mutex_exit(&l2arc_dev_mtx); 4537185029Spjd 4538185029Spjd /* 4539185029Spjd * Clear all buflists and ARC references. L2ARC device flush. 4540185029Spjd */ 4541185029Spjd l2arc_evict(remdev, 0, B_TRUE); 4542185029Spjd list_destroy(remdev->l2ad_buflist); 4543185029Spjd kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 4544185029Spjd kmem_free(remdev, sizeof (l2arc_dev_t)); 4545185029Spjd} 4546185029Spjd 4547185029Spjdvoid 4548185029Spjdl2arc_init(void) 4549185029Spjd{ 4550185029Spjd l2arc_thread_exit = 0; 4551185029Spjd l2arc_ndev = 0; 4552185029Spjd l2arc_writes_sent = 0; 4553185029Spjd l2arc_writes_done = 0; 4554185029Spjd 4555185029Spjd mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4556185029Spjd cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 4557185029Spjd mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 4558185029Spjd mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 4559185029Spjd mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 4560185029Spjd 4561185029Spjd l2arc_dev_list = &L2ARC_dev_list; 4562185029Spjd l2arc_free_on_write = &L2ARC_free_on_write; 4563185029Spjd list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 4564185029Spjd offsetof(l2arc_dev_t, l2ad_node)); 4565185029Spjd list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 4566185029Spjd offsetof(l2arc_data_free_t, l2df_list_node)); 4567185029Spjd} 4568185029Spjd 4569185029Spjdvoid 4570185029Spjdl2arc_fini(void) 4571185029Spjd{ 4572185029Spjd /* 4573185029Spjd * This is called from dmu_fini(), which is called from spa_fini(); 4574185029Spjd * Because of this, we can assume that all l2arc devices have 4575185029Spjd * already been removed when the pools themselves were removed. 4576185029Spjd */ 4577185029Spjd 4578185029Spjd l2arc_do_free_on_write(); 4579185029Spjd 4580185029Spjd mutex_destroy(&l2arc_feed_thr_lock); 4581185029Spjd cv_destroy(&l2arc_feed_thr_cv); 4582185029Spjd mutex_destroy(&l2arc_dev_mtx); 4583185029Spjd mutex_destroy(&l2arc_buflist_mtx); 4584185029Spjd mutex_destroy(&l2arc_free_on_write_mtx); 4585185029Spjd 4586185029Spjd list_destroy(l2arc_dev_list); 4587185029Spjd list_destroy(l2arc_free_on_write); 4588185029Spjd} 4589185029Spjd 4590185029Spjdvoid 4591185029Spjdl2arc_start(void) 4592185029Spjd{ 4593185029Spjd if (!(spa_mode & FWRITE)) 4594185029Spjd return; 4595185029Spjd 4596185029Spjd (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 4597185029Spjd TS_RUN, minclsyspri); 4598185029Spjd} 4599185029Spjd 4600185029Spjdvoid 4601185029Spjdl2arc_stop(void) 4602185029Spjd{ 4603185029Spjd if (!(spa_mode & FWRITE)) 4604185029Spjd return; 4605185029Spjd 4606185029Spjd mutex_enter(&l2arc_feed_thr_lock); 4607185029Spjd cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 4608185029Spjd l2arc_thread_exit = 1; 4609185029Spjd while (l2arc_thread_exit != 0) 4610185029Spjd cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 4611185029Spjd mutex_exit(&l2arc_feed_thr_lock); 4612185029Spjd} 4613