arc.c revision 251629
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/* 22219089Spjd * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23228103Smm * Copyright 2011 Nexenta Systems, Inc. All rights reserved. 24249195Smm * Copyright (c) 2013 by Delphix. All rights reserved. 25251478Sdelphij * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. 26168404Spjd */ 27168404Spjd 28168404Spjd/* 29168404Spjd * DVA-based Adjustable Replacement Cache 30168404Spjd * 31168404Spjd * While much of the theory of operation used here is 32168404Spjd * based on the self-tuning, low overhead replacement cache 33168404Spjd * presented by Megiddo and Modha at FAST 2003, there are some 34168404Spjd * significant differences: 35168404Spjd * 36168404Spjd * 1. The Megiddo and Modha model assumes any page is evictable. 37168404Spjd * Pages in its cache cannot be "locked" into memory. This makes 38168404Spjd * the eviction algorithm simple: evict the last page in the list. 39168404Spjd * This also make the performance characteristics easy to reason 40168404Spjd * about. Our cache is not so simple. At any given moment, some 41168404Spjd * subset of the blocks in the cache are un-evictable because we 42168404Spjd * have handed out a reference to them. Blocks are only evictable 43168404Spjd * when there are no external references active. This makes 44168404Spjd * eviction far more problematic: we choose to evict the evictable 45168404Spjd * blocks that are the "lowest" in the list. 46168404Spjd * 47168404Spjd * There are times when it is not possible to evict the requested 48168404Spjd * space. In these circumstances we are unable to adjust the cache 49168404Spjd * size. To prevent the cache growing unbounded at these times we 50185029Spjd * implement a "cache throttle" that slows the flow of new data 51185029Spjd * into the cache until we can make space available. 52168404Spjd * 53168404Spjd * 2. The Megiddo and Modha model assumes a fixed cache size. 54168404Spjd * Pages are evicted when the cache is full and there is a cache 55168404Spjd * miss. Our model has a variable sized cache. It grows with 56185029Spjd * high use, but also tries to react to memory pressure from the 57168404Spjd * operating system: decreasing its size when system memory is 58168404Spjd * tight. 59168404Spjd * 60168404Spjd * 3. The Megiddo and Modha model assumes a fixed page size. All 61168404Spjd * elements of the cache are therefor exactly the same size. So 62168404Spjd * when adjusting the cache size following a cache miss, its simply 63168404Spjd * a matter of choosing a single page to evict. In our model, we 64168404Spjd * have variable sized cache blocks (rangeing from 512 bytes to 65168404Spjd * 128K bytes). We therefor choose a set of blocks to evict to make 66168404Spjd * space for a cache miss that approximates as closely as possible 67168404Spjd * the space used by the new block. 68168404Spjd * 69168404Spjd * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 70168404Spjd * by N. Megiddo & D. Modha, FAST 2003 71168404Spjd */ 72168404Spjd 73168404Spjd/* 74168404Spjd * The locking model: 75168404Spjd * 76168404Spjd * A new reference to a cache buffer can be obtained in two 77168404Spjd * ways: 1) via a hash table lookup using the DVA as a key, 78185029Spjd * or 2) via one of the ARC lists. The arc_read() interface 79168404Spjd * uses method 1, while the internal arc algorithms for 80168404Spjd * adjusting the cache use method 2. We therefor provide two 81168404Spjd * types of locks: 1) the hash table lock array, and 2) the 82168404Spjd * arc list locks. 83168404Spjd * 84168404Spjd * Buffers do not have their own mutexs, rather they rely on the 85168404Spjd * hash table mutexs for the bulk of their protection (i.e. most 86168404Spjd * fields in the arc_buf_hdr_t are protected by these mutexs). 87168404Spjd * 88168404Spjd * buf_hash_find() returns the appropriate mutex (held) when it 89168404Spjd * locates the requested buffer in the hash table. It returns 90168404Spjd * NULL for the mutex if the buffer was not in the table. 91168404Spjd * 92168404Spjd * buf_hash_remove() expects the appropriate hash mutex to be 93168404Spjd * already held before it is invoked. 94168404Spjd * 95168404Spjd * Each arc state also has a mutex which is used to protect the 96168404Spjd * buffer list associated with the state. When attempting to 97168404Spjd * obtain a hash table lock while holding an arc list lock you 98168404Spjd * must use: mutex_tryenter() to avoid deadlock. Also note that 99168404Spjd * the active state mutex must be held before the ghost state mutex. 100168404Spjd * 101168404Spjd * Arc buffers may have an associated eviction callback function. 102168404Spjd * This function will be invoked prior to removing the buffer (e.g. 103168404Spjd * in arc_do_user_evicts()). Note however that the data associated 104168404Spjd * with the buffer may be evicted prior to the callback. The callback 105168404Spjd * must be made with *no locks held* (to prevent deadlock). Additionally, 106168404Spjd * the users of callbacks must ensure that their private data is 107168404Spjd * protected from simultaneous callbacks from arc_buf_evict() 108168404Spjd * and arc_do_user_evicts(). 109168404Spjd * 110168404Spjd * Note that the majority of the performance stats are manipulated 111168404Spjd * with atomic operations. 112185029Spjd * 113185029Spjd * The L2ARC uses the l2arc_buflist_mtx global mutex for the following: 114185029Spjd * 115185029Spjd * - L2ARC buflist creation 116185029Spjd * - L2ARC buflist eviction 117185029Spjd * - L2ARC write completion, which walks L2ARC buflists 118185029Spjd * - ARC header destruction, as it removes from L2ARC buflists 119185029Spjd * - ARC header release, as it removes from L2ARC buflists 120168404Spjd */ 121168404Spjd 122168404Spjd#include <sys/spa.h> 123168404Spjd#include <sys/zio.h> 124251478Sdelphij#include <sys/zio_compress.h> 125168404Spjd#include <sys/zfs_context.h> 126168404Spjd#include <sys/arc.h> 127168404Spjd#include <sys/refcount.h> 128185029Spjd#include <sys/vdev.h> 129219089Spjd#include <sys/vdev_impl.h> 130168404Spjd#ifdef _KERNEL 131168404Spjd#include <sys/dnlc.h> 132168404Spjd#endif 133168404Spjd#include <sys/callb.h> 134168404Spjd#include <sys/kstat.h> 135248572Ssmh#include <sys/trim_map.h> 136219089Spjd#include <zfs_fletcher.h> 137168404Spjd#include <sys/sdt.h> 138168404Spjd 139191902Skmacy#include <vm/vm_pageout.h> 140191902Skmacy 141240133Smm#ifdef illumos 142240133Smm#ifndef _KERNEL 143240133Smm/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 144240133Smmboolean_t arc_watch = B_FALSE; 145240133Smmint arc_procfd; 146240133Smm#endif 147240133Smm#endif /* illumos */ 148240133Smm 149168404Spjdstatic kmutex_t arc_reclaim_thr_lock; 150168404Spjdstatic kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 151168404Spjdstatic uint8_t arc_thread_exit; 152168404Spjd 153185029Spjdextern int zfs_write_limit_shift; 154185029Spjdextern uint64_t zfs_write_limit_max; 155185029Spjdextern kmutex_t zfs_write_limit_lock; 156185029Spjd 157168404Spjd#define ARC_REDUCE_DNLC_PERCENT 3 158168404Spjduint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 159168404Spjd 160168404Spjdtypedef enum arc_reclaim_strategy { 161168404Spjd ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 162168404Spjd ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 163168404Spjd} arc_reclaim_strategy_t; 164168404Spjd 165168404Spjd/* number of seconds before growing cache again */ 166168404Spjdstatic int arc_grow_retry = 60; 167168404Spjd 168208373Smm/* shift of arc_c for calculating both min and max arc_p */ 169208373Smmstatic int arc_p_min_shift = 4; 170208373Smm 171208373Smm/* log2(fraction of arc to reclaim) */ 172208373Smmstatic int arc_shrink_shift = 5; 173208373Smm 174168404Spjd/* 175168404Spjd * minimum lifespan of a prefetch block in clock ticks 176168404Spjd * (initialized in arc_init()) 177168404Spjd */ 178168404Spjdstatic int arc_min_prefetch_lifespan; 179168404Spjd 180208373Smmstatic int arc_dead; 181194043Skmacyextern int zfs_prefetch_disable; 182168404Spjd 183168404Spjd/* 184185029Spjd * The arc has filled available memory and has now warmed up. 185185029Spjd */ 186185029Spjdstatic boolean_t arc_warm; 187185029Spjd 188185029Spjd/* 189168404Spjd * These tunables are for performance analysis. 190168404Spjd */ 191185029Spjduint64_t zfs_arc_max; 192185029Spjduint64_t zfs_arc_min; 193185029Spjduint64_t zfs_arc_meta_limit = 0; 194208373Smmint zfs_arc_grow_retry = 0; 195208373Smmint zfs_arc_shrink_shift = 0; 196208373Smmint zfs_arc_p_min_shift = 0; 197242845Sdelphijint zfs_disable_dup_eviction = 0; 198185029Spjd 199185029SpjdTUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max); 200185029SpjdTUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 201185029SpjdTUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 202168473SpjdSYSCTL_DECL(_vfs_zfs); 203217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 204168473Spjd "Maximum ARC size"); 205217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 206168473Spjd "Minimum ARC size"); 207168404Spjd 208168404Spjd/* 209185029Spjd * Note that buffers can be in one of 6 states: 210168404Spjd * ARC_anon - anonymous (discussed below) 211168404Spjd * ARC_mru - recently used, currently cached 212168404Spjd * ARC_mru_ghost - recentely used, no longer in cache 213168404Spjd * ARC_mfu - frequently used, currently cached 214168404Spjd * ARC_mfu_ghost - frequently used, no longer in cache 215185029Spjd * ARC_l2c_only - exists in L2ARC but not other states 216185029Spjd * When there are no active references to the buffer, they are 217185029Spjd * are linked onto a list in one of these arc states. These are 218185029Spjd * the only buffers that can be evicted or deleted. Within each 219185029Spjd * state there are multiple lists, one for meta-data and one for 220185029Spjd * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 221185029Spjd * etc.) is tracked separately so that it can be managed more 222185029Spjd * explicitly: favored over data, limited explicitly. 223168404Spjd * 224168404Spjd * Anonymous buffers are buffers that are not associated with 225168404Spjd * a DVA. These are buffers that hold dirty block copies 226168404Spjd * before they are written to stable storage. By definition, 227168404Spjd * they are "ref'd" and are considered part of arc_mru 228168404Spjd * that cannot be freed. Generally, they will aquire a DVA 229168404Spjd * as they are written and migrate onto the arc_mru list. 230185029Spjd * 231185029Spjd * The ARC_l2c_only state is for buffers that are in the second 232185029Spjd * level ARC but no longer in any of the ARC_m* lists. The second 233185029Spjd * level ARC itself may also contain buffers that are in any of 234185029Spjd * the ARC_m* states - meaning that a buffer can exist in two 235185029Spjd * places. The reason for the ARC_l2c_only state is to keep the 236185029Spjd * buffer header in the hash table, so that reads that hit the 237185029Spjd * second level ARC benefit from these fast lookups. 238168404Spjd */ 239168404Spjd 240205264Skmacy#define ARCS_LOCK_PAD CACHE_LINE_SIZE 241205231Skmacystruct arcs_lock { 242205231Skmacy kmutex_t arcs_lock; 243205231Skmacy#ifdef _KERNEL 244205231Skmacy unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 245205231Skmacy#endif 246205231Skmacy}; 247205231Skmacy 248205231Skmacy/* 249205231Skmacy * must be power of two for mask use to work 250205231Skmacy * 251205231Skmacy */ 252205231Skmacy#define ARC_BUFC_NUMDATALISTS 16 253205231Skmacy#define ARC_BUFC_NUMMETADATALISTS 16 254206796Spjd#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 255205231Skmacy 256168404Spjdtypedef struct arc_state { 257185029Spjd uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 258185029Spjd uint64_t arcs_size; /* total amount of data in this state */ 259205231Skmacy list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 260205264Skmacy struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 261168404Spjd} arc_state_t; 262168404Spjd 263206796Spjd#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 264205231Skmacy 265185029Spjd/* The 6 states: */ 266168404Spjdstatic arc_state_t ARC_anon; 267168404Spjdstatic arc_state_t ARC_mru; 268168404Spjdstatic arc_state_t ARC_mru_ghost; 269168404Spjdstatic arc_state_t ARC_mfu; 270168404Spjdstatic arc_state_t ARC_mfu_ghost; 271185029Spjdstatic arc_state_t ARC_l2c_only; 272168404Spjd 273168404Spjdtypedef struct arc_stats { 274168404Spjd kstat_named_t arcstat_hits; 275168404Spjd kstat_named_t arcstat_misses; 276168404Spjd kstat_named_t arcstat_demand_data_hits; 277168404Spjd kstat_named_t arcstat_demand_data_misses; 278168404Spjd kstat_named_t arcstat_demand_metadata_hits; 279168404Spjd kstat_named_t arcstat_demand_metadata_misses; 280168404Spjd kstat_named_t arcstat_prefetch_data_hits; 281168404Spjd kstat_named_t arcstat_prefetch_data_misses; 282168404Spjd kstat_named_t arcstat_prefetch_metadata_hits; 283168404Spjd kstat_named_t arcstat_prefetch_metadata_misses; 284168404Spjd kstat_named_t arcstat_mru_hits; 285168404Spjd kstat_named_t arcstat_mru_ghost_hits; 286168404Spjd kstat_named_t arcstat_mfu_hits; 287168404Spjd kstat_named_t arcstat_mfu_ghost_hits; 288205231Skmacy kstat_named_t arcstat_allocated; 289168404Spjd kstat_named_t arcstat_deleted; 290205231Skmacy kstat_named_t arcstat_stolen; 291168404Spjd kstat_named_t arcstat_recycle_miss; 292251629Sdelphij /* 293251629Sdelphij * Number of buffers that could not be evicted because the hash lock 294251629Sdelphij * was held by another thread. The lock may not necessarily be held 295251629Sdelphij * by something using the same buffer, since hash locks are shared 296251629Sdelphij * by multiple buffers. 297251629Sdelphij */ 298168404Spjd kstat_named_t arcstat_mutex_miss; 299251629Sdelphij /* 300251629Sdelphij * Number of buffers skipped because they have I/O in progress, are 301251629Sdelphij * indrect prefetch buffers that have not lived long enough, or are 302251629Sdelphij * not from the spa we're trying to evict from. 303251629Sdelphij */ 304168404Spjd kstat_named_t arcstat_evict_skip; 305208373Smm kstat_named_t arcstat_evict_l2_cached; 306208373Smm kstat_named_t arcstat_evict_l2_eligible; 307208373Smm kstat_named_t arcstat_evict_l2_ineligible; 308168404Spjd kstat_named_t arcstat_hash_elements; 309168404Spjd kstat_named_t arcstat_hash_elements_max; 310168404Spjd kstat_named_t arcstat_hash_collisions; 311168404Spjd kstat_named_t arcstat_hash_chains; 312168404Spjd kstat_named_t arcstat_hash_chain_max; 313168404Spjd kstat_named_t arcstat_p; 314168404Spjd kstat_named_t arcstat_c; 315168404Spjd kstat_named_t arcstat_c_min; 316168404Spjd kstat_named_t arcstat_c_max; 317168404Spjd kstat_named_t arcstat_size; 318185029Spjd kstat_named_t arcstat_hdr_size; 319208373Smm kstat_named_t arcstat_data_size; 320208373Smm kstat_named_t arcstat_other_size; 321185029Spjd kstat_named_t arcstat_l2_hits; 322185029Spjd kstat_named_t arcstat_l2_misses; 323185029Spjd kstat_named_t arcstat_l2_feeds; 324185029Spjd kstat_named_t arcstat_l2_rw_clash; 325208373Smm kstat_named_t arcstat_l2_read_bytes; 326208373Smm kstat_named_t arcstat_l2_write_bytes; 327185029Spjd kstat_named_t arcstat_l2_writes_sent; 328185029Spjd kstat_named_t arcstat_l2_writes_done; 329185029Spjd kstat_named_t arcstat_l2_writes_error; 330185029Spjd kstat_named_t arcstat_l2_writes_hdr_miss; 331185029Spjd kstat_named_t arcstat_l2_evict_lock_retry; 332185029Spjd kstat_named_t arcstat_l2_evict_reading; 333185029Spjd kstat_named_t arcstat_l2_free_on_write; 334185029Spjd kstat_named_t arcstat_l2_abort_lowmem; 335185029Spjd kstat_named_t arcstat_l2_cksum_bad; 336185029Spjd kstat_named_t arcstat_l2_io_error; 337185029Spjd kstat_named_t arcstat_l2_size; 338251478Sdelphij kstat_named_t arcstat_l2_asize; 339185029Spjd kstat_named_t arcstat_l2_hdr_size; 340251478Sdelphij kstat_named_t arcstat_l2_compress_successes; 341251478Sdelphij kstat_named_t arcstat_l2_compress_zeros; 342251478Sdelphij kstat_named_t arcstat_l2_compress_failures; 343205231Skmacy kstat_named_t arcstat_l2_write_trylock_fail; 344205231Skmacy kstat_named_t arcstat_l2_write_passed_headroom; 345205231Skmacy kstat_named_t arcstat_l2_write_spa_mismatch; 346206796Spjd kstat_named_t arcstat_l2_write_in_l2; 347205231Skmacy kstat_named_t arcstat_l2_write_hdr_io_in_progress; 348205231Skmacy kstat_named_t arcstat_l2_write_not_cacheable; 349205231Skmacy kstat_named_t arcstat_l2_write_full; 350205231Skmacy kstat_named_t arcstat_l2_write_buffer_iter; 351205231Skmacy kstat_named_t arcstat_l2_write_pios; 352205231Skmacy kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 353205231Skmacy kstat_named_t arcstat_l2_write_buffer_list_iter; 354205231Skmacy kstat_named_t arcstat_l2_write_buffer_list_null_iter; 355242845Sdelphij kstat_named_t arcstat_memory_throttle_count; 356242845Sdelphij kstat_named_t arcstat_duplicate_buffers; 357242845Sdelphij kstat_named_t arcstat_duplicate_buffers_size; 358242845Sdelphij kstat_named_t arcstat_duplicate_reads; 359168404Spjd} arc_stats_t; 360168404Spjd 361168404Spjdstatic arc_stats_t arc_stats = { 362168404Spjd { "hits", KSTAT_DATA_UINT64 }, 363168404Spjd { "misses", KSTAT_DATA_UINT64 }, 364168404Spjd { "demand_data_hits", KSTAT_DATA_UINT64 }, 365168404Spjd { "demand_data_misses", KSTAT_DATA_UINT64 }, 366168404Spjd { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 367168404Spjd { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 368168404Spjd { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 369168404Spjd { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 370168404Spjd { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 371168404Spjd { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 372168404Spjd { "mru_hits", KSTAT_DATA_UINT64 }, 373168404Spjd { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 374168404Spjd { "mfu_hits", KSTAT_DATA_UINT64 }, 375168404Spjd { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 376205231Skmacy { "allocated", KSTAT_DATA_UINT64 }, 377168404Spjd { "deleted", KSTAT_DATA_UINT64 }, 378205231Skmacy { "stolen", KSTAT_DATA_UINT64 }, 379168404Spjd { "recycle_miss", KSTAT_DATA_UINT64 }, 380168404Spjd { "mutex_miss", KSTAT_DATA_UINT64 }, 381168404Spjd { "evict_skip", KSTAT_DATA_UINT64 }, 382208373Smm { "evict_l2_cached", KSTAT_DATA_UINT64 }, 383208373Smm { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 384208373Smm { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 385168404Spjd { "hash_elements", KSTAT_DATA_UINT64 }, 386168404Spjd { "hash_elements_max", KSTAT_DATA_UINT64 }, 387168404Spjd { "hash_collisions", KSTAT_DATA_UINT64 }, 388168404Spjd { "hash_chains", KSTAT_DATA_UINT64 }, 389168404Spjd { "hash_chain_max", KSTAT_DATA_UINT64 }, 390168404Spjd { "p", KSTAT_DATA_UINT64 }, 391168404Spjd { "c", KSTAT_DATA_UINT64 }, 392168404Spjd { "c_min", KSTAT_DATA_UINT64 }, 393168404Spjd { "c_max", KSTAT_DATA_UINT64 }, 394185029Spjd { "size", KSTAT_DATA_UINT64 }, 395185029Spjd { "hdr_size", KSTAT_DATA_UINT64 }, 396208373Smm { "data_size", KSTAT_DATA_UINT64 }, 397208373Smm { "other_size", KSTAT_DATA_UINT64 }, 398185029Spjd { "l2_hits", KSTAT_DATA_UINT64 }, 399185029Spjd { "l2_misses", KSTAT_DATA_UINT64 }, 400185029Spjd { "l2_feeds", KSTAT_DATA_UINT64 }, 401185029Spjd { "l2_rw_clash", KSTAT_DATA_UINT64 }, 402208373Smm { "l2_read_bytes", KSTAT_DATA_UINT64 }, 403208373Smm { "l2_write_bytes", KSTAT_DATA_UINT64 }, 404185029Spjd { "l2_writes_sent", KSTAT_DATA_UINT64 }, 405185029Spjd { "l2_writes_done", KSTAT_DATA_UINT64 }, 406185029Spjd { "l2_writes_error", KSTAT_DATA_UINT64 }, 407185029Spjd { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 408185029Spjd { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 409185029Spjd { "l2_evict_reading", KSTAT_DATA_UINT64 }, 410185029Spjd { "l2_free_on_write", KSTAT_DATA_UINT64 }, 411185029Spjd { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 412185029Spjd { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 413185029Spjd { "l2_io_error", KSTAT_DATA_UINT64 }, 414185029Spjd { "l2_size", KSTAT_DATA_UINT64 }, 415251478Sdelphij { "l2_asize", KSTAT_DATA_UINT64 }, 416185029Spjd { "l2_hdr_size", KSTAT_DATA_UINT64 }, 417251478Sdelphij { "l2_compress_successes", KSTAT_DATA_UINT64 }, 418251478Sdelphij { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 419251478Sdelphij { "l2_compress_failures", KSTAT_DATA_UINT64 }, 420206796Spjd { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 421206796Spjd { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 422206796Spjd { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 423206796Spjd { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 424206796Spjd { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 425206796Spjd { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 426206796Spjd { "l2_write_full", KSTAT_DATA_UINT64 }, 427206796Spjd { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 428206796Spjd { "l2_write_pios", KSTAT_DATA_UINT64 }, 429206796Spjd { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 430206796Spjd { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 431242845Sdelphij { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 432242845Sdelphij { "memory_throttle_count", KSTAT_DATA_UINT64 }, 433242845Sdelphij { "duplicate_buffers", KSTAT_DATA_UINT64 }, 434242845Sdelphij { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 435242845Sdelphij { "duplicate_reads", KSTAT_DATA_UINT64 } 436168404Spjd}; 437168404Spjd 438168404Spjd#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 439168404Spjd 440168404Spjd#define ARCSTAT_INCR(stat, val) \ 441168404Spjd atomic_add_64(&arc_stats.stat.value.ui64, (val)); 442168404Spjd 443206796Spjd#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 444168404Spjd#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 445168404Spjd 446168404Spjd#define ARCSTAT_MAX(stat, val) { \ 447168404Spjd uint64_t m; \ 448168404Spjd while ((val) > (m = arc_stats.stat.value.ui64) && \ 449168404Spjd (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 450168404Spjd continue; \ 451168404Spjd} 452168404Spjd 453168404Spjd#define ARCSTAT_MAXSTAT(stat) \ 454168404Spjd ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 455168404Spjd 456168404Spjd/* 457168404Spjd * We define a macro to allow ARC hits/misses to be easily broken down by 458168404Spjd * two separate conditions, giving a total of four different subtypes for 459168404Spjd * each of hits and misses (so eight statistics total). 460168404Spjd */ 461168404Spjd#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 462168404Spjd if (cond1) { \ 463168404Spjd if (cond2) { \ 464168404Spjd ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 465168404Spjd } else { \ 466168404Spjd ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 467168404Spjd } \ 468168404Spjd } else { \ 469168404Spjd if (cond2) { \ 470168404Spjd ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 471168404Spjd } else { \ 472168404Spjd ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 473168404Spjd } \ 474168404Spjd } 475168404Spjd 476168404Spjdkstat_t *arc_ksp; 477206796Spjdstatic arc_state_t *arc_anon; 478168404Spjdstatic arc_state_t *arc_mru; 479168404Spjdstatic arc_state_t *arc_mru_ghost; 480168404Spjdstatic arc_state_t *arc_mfu; 481168404Spjdstatic arc_state_t *arc_mfu_ghost; 482185029Spjdstatic arc_state_t *arc_l2c_only; 483168404Spjd 484168404Spjd/* 485168404Spjd * There are several ARC variables that are critical to export as kstats -- 486168404Spjd * but we don't want to have to grovel around in the kstat whenever we wish to 487168404Spjd * manipulate them. For these variables, we therefore define them to be in 488168404Spjd * terms of the statistic variable. This assures that we are not introducing 489168404Spjd * the possibility of inconsistency by having shadow copies of the variables, 490168404Spjd * while still allowing the code to be readable. 491168404Spjd */ 492168404Spjd#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 493168404Spjd#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 494168404Spjd#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 495168404Spjd#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 496168404Spjd#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 497168404Spjd 498251478Sdelphij#define L2ARC_IS_VALID_COMPRESS(_c_) \ 499251478Sdelphij ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 500251478Sdelphij 501168404Spjdstatic int arc_no_grow; /* Don't try to grow cache size */ 502168404Spjdstatic uint64_t arc_tempreserve; 503209962Smmstatic uint64_t arc_loaned_bytes; 504185029Spjdstatic uint64_t arc_meta_used; 505185029Spjdstatic uint64_t arc_meta_limit; 506185029Spjdstatic uint64_t arc_meta_max = 0; 507229663SpjdSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RD, &arc_meta_used, 0, 508229663Spjd "ARC metadata used"); 509229663SpjdSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RW, &arc_meta_limit, 0, 510229663Spjd "ARC metadata limit"); 511168404Spjd 512185029Spjdtypedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 513185029Spjd 514168404Spjdtypedef struct arc_callback arc_callback_t; 515168404Spjd 516168404Spjdstruct arc_callback { 517168404Spjd void *acb_private; 518168404Spjd arc_done_func_t *acb_done; 519168404Spjd arc_buf_t *acb_buf; 520168404Spjd zio_t *acb_zio_dummy; 521168404Spjd arc_callback_t *acb_next; 522168404Spjd}; 523168404Spjd 524168404Spjdtypedef struct arc_write_callback arc_write_callback_t; 525168404Spjd 526168404Spjdstruct arc_write_callback { 527168404Spjd void *awcb_private; 528168404Spjd arc_done_func_t *awcb_ready; 529168404Spjd arc_done_func_t *awcb_done; 530168404Spjd arc_buf_t *awcb_buf; 531168404Spjd}; 532168404Spjd 533168404Spjdstruct arc_buf_hdr { 534168404Spjd /* protected by hash lock */ 535168404Spjd dva_t b_dva; 536168404Spjd uint64_t b_birth; 537168404Spjd uint64_t b_cksum0; 538168404Spjd 539168404Spjd kmutex_t b_freeze_lock; 540168404Spjd zio_cksum_t *b_freeze_cksum; 541219089Spjd void *b_thawed; 542168404Spjd 543168404Spjd arc_buf_hdr_t *b_hash_next; 544168404Spjd arc_buf_t *b_buf; 545168404Spjd uint32_t b_flags; 546168404Spjd uint32_t b_datacnt; 547168404Spjd 548168404Spjd arc_callback_t *b_acb; 549168404Spjd kcondvar_t b_cv; 550168404Spjd 551168404Spjd /* immutable */ 552168404Spjd arc_buf_contents_t b_type; 553168404Spjd uint64_t b_size; 554209962Smm uint64_t b_spa; 555168404Spjd 556168404Spjd /* protected by arc state mutex */ 557168404Spjd arc_state_t *b_state; 558168404Spjd list_node_t b_arc_node; 559168404Spjd 560168404Spjd /* updated atomically */ 561168404Spjd clock_t b_arc_access; 562168404Spjd 563168404Spjd /* self protecting */ 564168404Spjd refcount_t b_refcnt; 565185029Spjd 566185029Spjd l2arc_buf_hdr_t *b_l2hdr; 567185029Spjd list_node_t b_l2node; 568168404Spjd}; 569168404Spjd 570168404Spjdstatic arc_buf_t *arc_eviction_list; 571168404Spjdstatic kmutex_t arc_eviction_mtx; 572168404Spjdstatic arc_buf_hdr_t arc_eviction_hdr; 573168404Spjdstatic void arc_get_data_buf(arc_buf_t *buf); 574168404Spjdstatic void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock); 575185029Spjdstatic int arc_evict_needed(arc_buf_contents_t type); 576209962Smmstatic void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes); 577240133Smm#ifdef illumos 578240133Smmstatic void arc_buf_watch(arc_buf_t *buf); 579240133Smm#endif /* illumos */ 580168404Spjd 581209962Smmstatic boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab); 582208373Smm 583168404Spjd#define GHOST_STATE(state) \ 584185029Spjd ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 585185029Spjd (state) == arc_l2c_only) 586168404Spjd 587168404Spjd/* 588168404Spjd * Private ARC flags. These flags are private ARC only flags that will show up 589168404Spjd * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can 590168404Spjd * be passed in as arc_flags in things like arc_read. However, these flags 591168404Spjd * should never be passed and should only be set by ARC code. When adding new 592168404Spjd * public flags, make sure not to smash the private ones. 593168404Spjd */ 594168404Spjd 595168404Spjd#define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */ 596168404Spjd#define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */ 597168404Spjd#define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */ 598168404Spjd#define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */ 599168404Spjd#define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */ 600168404Spjd#define ARC_INDIRECT (1 << 14) /* this is an indirect block */ 601185029Spjd#define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */ 602185029Spjd#define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */ 603185029Spjd#define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */ 604185029Spjd#define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */ 605168404Spjd 606168404Spjd#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE) 607168404Spjd#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS) 608168404Spjd#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR) 609208373Smm#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH) 610168404Spjd#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ) 611168404Spjd#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE) 612185029Spjd#define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS) 613185029Spjd#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE) 614185029Spjd#define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \ 615185029Spjd (hdr)->b_l2hdr != NULL) 616185029Spjd#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING) 617185029Spjd#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED) 618185029Spjd#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD) 619168404Spjd 620168404Spjd/* 621185029Spjd * Other sizes 622185029Spjd */ 623185029Spjd 624185029Spjd#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 625185029Spjd#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 626185029Spjd 627185029Spjd/* 628168404Spjd * Hash table routines 629168404Spjd */ 630168404Spjd 631205253Skmacy#define HT_LOCK_PAD CACHE_LINE_SIZE 632168404Spjd 633168404Spjdstruct ht_lock { 634168404Spjd kmutex_t ht_lock; 635168404Spjd#ifdef _KERNEL 636168404Spjd unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 637168404Spjd#endif 638168404Spjd}; 639168404Spjd 640168404Spjd#define BUF_LOCKS 256 641168404Spjdtypedef struct buf_hash_table { 642168404Spjd uint64_t ht_mask; 643168404Spjd arc_buf_hdr_t **ht_table; 644205264Skmacy struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 645168404Spjd} buf_hash_table_t; 646168404Spjd 647168404Spjdstatic buf_hash_table_t buf_hash_table; 648168404Spjd 649168404Spjd#define BUF_HASH_INDEX(spa, dva, birth) \ 650168404Spjd (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 651168404Spjd#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 652168404Spjd#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 653219089Spjd#define HDR_LOCK(hdr) \ 654219089Spjd (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 655168404Spjd 656168404Spjduint64_t zfs_crc64_table[256]; 657168404Spjd 658185029Spjd/* 659185029Spjd * Level 2 ARC 660185029Spjd */ 661185029Spjd 662208373Smm#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 663251478Sdelphij#define L2ARC_HEADROOM 2 /* num of writes */ 664251478Sdelphij/* 665251478Sdelphij * If we discover during ARC scan any buffers to be compressed, we boost 666251478Sdelphij * our headroom for the next scanning cycle by this percentage multiple. 667251478Sdelphij */ 668251478Sdelphij#define L2ARC_HEADROOM_BOOST 200 669208373Smm#define L2ARC_FEED_SECS 1 /* caching interval secs */ 670208373Smm#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 671185029Spjd 672185029Spjd#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 673185029Spjd#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 674185029Spjd 675185029Spjd/* 676185029Spjd * L2ARC Performance Tunables 677185029Spjd */ 678185029Spjduint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 679185029Spjduint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 680185029Spjduint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 681251478Sdelphijuint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 682185029Spjduint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 683208373Smmuint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 684219089Spjdboolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 685208373Smmboolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 686208373Smmboolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 687185029Spjd 688217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 689205231Skmacy &l2arc_write_max, 0, "max write size"); 690217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 691205231Skmacy &l2arc_write_boost, 0, "extra write during warmup"); 692217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 693205231Skmacy &l2arc_headroom, 0, "number of dev writes"); 694217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 695205231Skmacy &l2arc_feed_secs, 0, "interval seconds"); 696217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 697208373Smm &l2arc_feed_min_ms, 0, "min interval milliseconds"); 698205231Skmacy 699205231SkmacySYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 700205231Skmacy &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 701208373SmmSYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 702208373Smm &l2arc_feed_again, 0, "turbo warmup"); 703208373SmmSYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 704208373Smm &l2arc_norw, 0, "no reads during writes"); 705205231Skmacy 706217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 707205231Skmacy &ARC_anon.arcs_size, 0, "size of anonymous state"); 708217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 709205231Skmacy &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 710217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 711205231Skmacy &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 712205231Skmacy 713217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 714205231Skmacy &ARC_mru.arcs_size, 0, "size of mru state"); 715217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 716205231Skmacy &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 717217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 718205231Skmacy &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 719205231Skmacy 720217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 721205231Skmacy &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 722217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 723205231Skmacy &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 724205231Skmacy "size of metadata in mru ghost state"); 725217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 726205231Skmacy &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 727205231Skmacy "size of data in mru ghost state"); 728205231Skmacy 729217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 730205231Skmacy &ARC_mfu.arcs_size, 0, "size of mfu state"); 731217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 732205231Skmacy &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 733217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 734205231Skmacy &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 735205231Skmacy 736217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 737205231Skmacy &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 738217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 739205231Skmacy &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 740205231Skmacy "size of metadata in mfu ghost state"); 741217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 742205231Skmacy &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 743205231Skmacy "size of data in mfu ghost state"); 744205231Skmacy 745217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 746205231Skmacy &ARC_l2c_only.arcs_size, 0, "size of mru state"); 747205231Skmacy 748185029Spjd/* 749185029Spjd * L2ARC Internals 750185029Spjd */ 751185029Spjdtypedef struct l2arc_dev { 752185029Spjd vdev_t *l2ad_vdev; /* vdev */ 753185029Spjd spa_t *l2ad_spa; /* spa */ 754185029Spjd uint64_t l2ad_hand; /* next write location */ 755185029Spjd uint64_t l2ad_start; /* first addr on device */ 756185029Spjd uint64_t l2ad_end; /* last addr on device */ 757185029Spjd uint64_t l2ad_evict; /* last addr eviction reached */ 758185029Spjd boolean_t l2ad_first; /* first sweep through */ 759208373Smm boolean_t l2ad_writing; /* currently writing */ 760185029Spjd list_t *l2ad_buflist; /* buffer list */ 761185029Spjd list_node_t l2ad_node; /* device list node */ 762185029Spjd} l2arc_dev_t; 763185029Spjd 764185029Spjdstatic list_t L2ARC_dev_list; /* device list */ 765185029Spjdstatic list_t *l2arc_dev_list; /* device list pointer */ 766185029Spjdstatic kmutex_t l2arc_dev_mtx; /* device list mutex */ 767185029Spjdstatic l2arc_dev_t *l2arc_dev_last; /* last device used */ 768185029Spjdstatic kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 769185029Spjdstatic list_t L2ARC_free_on_write; /* free after write buf list */ 770185029Spjdstatic list_t *l2arc_free_on_write; /* free after write list ptr */ 771185029Spjdstatic kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 772185029Spjdstatic uint64_t l2arc_ndev; /* number of devices */ 773185029Spjd 774185029Spjdtypedef struct l2arc_read_callback { 775251478Sdelphij arc_buf_t *l2rcb_buf; /* read buffer */ 776251478Sdelphij spa_t *l2rcb_spa; /* spa */ 777251478Sdelphij blkptr_t l2rcb_bp; /* original blkptr */ 778251478Sdelphij zbookmark_t l2rcb_zb; /* original bookmark */ 779251478Sdelphij int l2rcb_flags; /* original flags */ 780251478Sdelphij enum zio_compress l2rcb_compress; /* applied compress */ 781185029Spjd} l2arc_read_callback_t; 782185029Spjd 783185029Spjdtypedef struct l2arc_write_callback { 784185029Spjd l2arc_dev_t *l2wcb_dev; /* device info */ 785185029Spjd arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 786185029Spjd} l2arc_write_callback_t; 787185029Spjd 788185029Spjdstruct l2arc_buf_hdr { 789185029Spjd /* protected by arc_buf_hdr mutex */ 790251478Sdelphij l2arc_dev_t *b_dev; /* L2ARC device */ 791251478Sdelphij uint64_t b_daddr; /* disk address, offset byte */ 792251478Sdelphij /* compression applied to buffer data */ 793251478Sdelphij enum zio_compress b_compress; 794251478Sdelphij /* real alloc'd buffer size depending on b_compress applied */ 795251478Sdelphij int b_asize; 796251478Sdelphij /* temporary buffer holder for in-flight compressed data */ 797251478Sdelphij void *b_tmp_cdata; 798185029Spjd}; 799185029Spjd 800185029Spjdtypedef struct l2arc_data_free { 801185029Spjd /* protected by l2arc_free_on_write_mtx */ 802185029Spjd void *l2df_data; 803185029Spjd size_t l2df_size; 804185029Spjd void (*l2df_func)(void *, size_t); 805185029Spjd list_node_t l2df_list_node; 806185029Spjd} l2arc_data_free_t; 807185029Spjd 808185029Spjdstatic kmutex_t l2arc_feed_thr_lock; 809185029Spjdstatic kcondvar_t l2arc_feed_thr_cv; 810185029Spjdstatic uint8_t l2arc_thread_exit; 811185029Spjd 812185029Spjdstatic void l2arc_read_done(zio_t *zio); 813185029Spjdstatic void l2arc_hdr_stat_add(void); 814185029Spjdstatic void l2arc_hdr_stat_remove(void); 815185029Spjd 816251478Sdelphijstatic boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr); 817251478Sdelphijstatic void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, 818251478Sdelphij enum zio_compress c); 819251478Sdelphijstatic void l2arc_release_cdata_buf(arc_buf_hdr_t *ab); 820251478Sdelphij 821168404Spjdstatic uint64_t 822209962Smmbuf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 823168404Spjd{ 824168404Spjd uint8_t *vdva = (uint8_t *)dva; 825168404Spjd uint64_t crc = -1ULL; 826168404Spjd int i; 827168404Spjd 828168404Spjd ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 829168404Spjd 830168404Spjd for (i = 0; i < sizeof (dva_t); i++) 831168404Spjd crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 832168404Spjd 833209962Smm crc ^= (spa>>8) ^ birth; 834168404Spjd 835168404Spjd return (crc); 836168404Spjd} 837168404Spjd 838168404Spjd#define BUF_EMPTY(buf) \ 839168404Spjd ((buf)->b_dva.dva_word[0] == 0 && \ 840168404Spjd (buf)->b_dva.dva_word[1] == 0 && \ 841168404Spjd (buf)->b_birth == 0) 842168404Spjd 843168404Spjd#define BUF_EQUAL(spa, dva, birth, buf) \ 844168404Spjd ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 845168404Spjd ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 846168404Spjd ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 847168404Spjd 848219089Spjdstatic void 849219089Spjdbuf_discard_identity(arc_buf_hdr_t *hdr) 850219089Spjd{ 851219089Spjd hdr->b_dva.dva_word[0] = 0; 852219089Spjd hdr->b_dva.dva_word[1] = 0; 853219089Spjd hdr->b_birth = 0; 854219089Spjd hdr->b_cksum0 = 0; 855219089Spjd} 856219089Spjd 857168404Spjdstatic arc_buf_hdr_t * 858209962Smmbuf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp) 859168404Spjd{ 860168404Spjd uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 861168404Spjd kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 862168404Spjd arc_buf_hdr_t *buf; 863168404Spjd 864168404Spjd mutex_enter(hash_lock); 865168404Spjd for (buf = buf_hash_table.ht_table[idx]; buf != NULL; 866168404Spjd buf = buf->b_hash_next) { 867168404Spjd if (BUF_EQUAL(spa, dva, birth, buf)) { 868168404Spjd *lockp = hash_lock; 869168404Spjd return (buf); 870168404Spjd } 871168404Spjd } 872168404Spjd mutex_exit(hash_lock); 873168404Spjd *lockp = NULL; 874168404Spjd return (NULL); 875168404Spjd} 876168404Spjd 877168404Spjd/* 878168404Spjd * Insert an entry into the hash table. If there is already an element 879168404Spjd * equal to elem in the hash table, then the already existing element 880168404Spjd * will be returned and the new element will not be inserted. 881168404Spjd * Otherwise returns NULL. 882168404Spjd */ 883168404Spjdstatic arc_buf_hdr_t * 884168404Spjdbuf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp) 885168404Spjd{ 886168404Spjd uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 887168404Spjd kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 888168404Spjd arc_buf_hdr_t *fbuf; 889168404Spjd uint32_t i; 890168404Spjd 891168404Spjd ASSERT(!HDR_IN_HASH_TABLE(buf)); 892168404Spjd *lockp = hash_lock; 893168404Spjd mutex_enter(hash_lock); 894168404Spjd for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL; 895168404Spjd fbuf = fbuf->b_hash_next, i++) { 896168404Spjd if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf)) 897168404Spjd return (fbuf); 898168404Spjd } 899168404Spjd 900168404Spjd buf->b_hash_next = buf_hash_table.ht_table[idx]; 901168404Spjd buf_hash_table.ht_table[idx] = buf; 902168404Spjd buf->b_flags |= ARC_IN_HASH_TABLE; 903168404Spjd 904168404Spjd /* collect some hash table performance data */ 905168404Spjd if (i > 0) { 906168404Spjd ARCSTAT_BUMP(arcstat_hash_collisions); 907168404Spjd if (i == 1) 908168404Spjd ARCSTAT_BUMP(arcstat_hash_chains); 909168404Spjd 910168404Spjd ARCSTAT_MAX(arcstat_hash_chain_max, i); 911168404Spjd } 912168404Spjd 913168404Spjd ARCSTAT_BUMP(arcstat_hash_elements); 914168404Spjd ARCSTAT_MAXSTAT(arcstat_hash_elements); 915168404Spjd 916168404Spjd return (NULL); 917168404Spjd} 918168404Spjd 919168404Spjdstatic void 920168404Spjdbuf_hash_remove(arc_buf_hdr_t *buf) 921168404Spjd{ 922168404Spjd arc_buf_hdr_t *fbuf, **bufp; 923168404Spjd uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 924168404Spjd 925168404Spjd ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 926168404Spjd ASSERT(HDR_IN_HASH_TABLE(buf)); 927168404Spjd 928168404Spjd bufp = &buf_hash_table.ht_table[idx]; 929168404Spjd while ((fbuf = *bufp) != buf) { 930168404Spjd ASSERT(fbuf != NULL); 931168404Spjd bufp = &fbuf->b_hash_next; 932168404Spjd } 933168404Spjd *bufp = buf->b_hash_next; 934168404Spjd buf->b_hash_next = NULL; 935168404Spjd buf->b_flags &= ~ARC_IN_HASH_TABLE; 936168404Spjd 937168404Spjd /* collect some hash table performance data */ 938168404Spjd ARCSTAT_BUMPDOWN(arcstat_hash_elements); 939168404Spjd 940168404Spjd if (buf_hash_table.ht_table[idx] && 941168404Spjd buf_hash_table.ht_table[idx]->b_hash_next == NULL) 942168404Spjd ARCSTAT_BUMPDOWN(arcstat_hash_chains); 943168404Spjd} 944168404Spjd 945168404Spjd/* 946168404Spjd * Global data structures and functions for the buf kmem cache. 947168404Spjd */ 948168404Spjdstatic kmem_cache_t *hdr_cache; 949168404Spjdstatic kmem_cache_t *buf_cache; 950168404Spjd 951168404Spjdstatic void 952168404Spjdbuf_fini(void) 953168404Spjd{ 954168404Spjd int i; 955168404Spjd 956168404Spjd kmem_free(buf_hash_table.ht_table, 957168404Spjd (buf_hash_table.ht_mask + 1) * sizeof (void *)); 958168404Spjd for (i = 0; i < BUF_LOCKS; i++) 959168404Spjd mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 960168404Spjd kmem_cache_destroy(hdr_cache); 961168404Spjd kmem_cache_destroy(buf_cache); 962168404Spjd} 963168404Spjd 964168404Spjd/* 965168404Spjd * Constructor callback - called when the cache is empty 966168404Spjd * and a new buf is requested. 967168404Spjd */ 968168404Spjd/* ARGSUSED */ 969168404Spjdstatic int 970168404Spjdhdr_cons(void *vbuf, void *unused, int kmflag) 971168404Spjd{ 972168404Spjd arc_buf_hdr_t *buf = vbuf; 973168404Spjd 974168404Spjd bzero(buf, sizeof (arc_buf_hdr_t)); 975168404Spjd refcount_create(&buf->b_refcnt); 976168404Spjd cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL); 977185029Spjd mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 978208373Smm arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 979185029Spjd 980168404Spjd return (0); 981168404Spjd} 982168404Spjd 983185029Spjd/* ARGSUSED */ 984185029Spjdstatic int 985185029Spjdbuf_cons(void *vbuf, void *unused, int kmflag) 986185029Spjd{ 987185029Spjd arc_buf_t *buf = vbuf; 988185029Spjd 989185029Spjd bzero(buf, sizeof (arc_buf_t)); 990219089Spjd mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 991208373Smm arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 992208373Smm 993185029Spjd return (0); 994185029Spjd} 995185029Spjd 996168404Spjd/* 997168404Spjd * Destructor callback - called when a cached buf is 998168404Spjd * no longer required. 999168404Spjd */ 1000168404Spjd/* ARGSUSED */ 1001168404Spjdstatic void 1002168404Spjdhdr_dest(void *vbuf, void *unused) 1003168404Spjd{ 1004168404Spjd arc_buf_hdr_t *buf = vbuf; 1005168404Spjd 1006219089Spjd ASSERT(BUF_EMPTY(buf)); 1007168404Spjd refcount_destroy(&buf->b_refcnt); 1008168404Spjd cv_destroy(&buf->b_cv); 1009185029Spjd mutex_destroy(&buf->b_freeze_lock); 1010208373Smm arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1011168404Spjd} 1012168404Spjd 1013185029Spjd/* ARGSUSED */ 1014185029Spjdstatic void 1015185029Spjdbuf_dest(void *vbuf, void *unused) 1016185029Spjd{ 1017185029Spjd arc_buf_t *buf = vbuf; 1018185029Spjd 1019219089Spjd mutex_destroy(&buf->b_evict_lock); 1020208373Smm arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1021185029Spjd} 1022185029Spjd 1023168404Spjd/* 1024168404Spjd * Reclaim callback -- invoked when memory is low. 1025168404Spjd */ 1026168404Spjd/* ARGSUSED */ 1027168404Spjdstatic void 1028168404Spjdhdr_recl(void *unused) 1029168404Spjd{ 1030168404Spjd dprintf("hdr_recl called\n"); 1031168404Spjd /* 1032168404Spjd * umem calls the reclaim func when we destroy the buf cache, 1033168404Spjd * which is after we do arc_fini(). 1034168404Spjd */ 1035168404Spjd if (!arc_dead) 1036168404Spjd cv_signal(&arc_reclaim_thr_cv); 1037168404Spjd} 1038168404Spjd 1039168404Spjdstatic void 1040168404Spjdbuf_init(void) 1041168404Spjd{ 1042168404Spjd uint64_t *ct; 1043168404Spjd uint64_t hsize = 1ULL << 12; 1044168404Spjd int i, j; 1045168404Spjd 1046168404Spjd /* 1047168404Spjd * The hash table is big enough to fill all of physical memory 1048168404Spjd * with an average 64K block size. The table will take up 1049168404Spjd * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers). 1050168404Spjd */ 1051168696Spjd while (hsize * 65536 < (uint64_t)physmem * PAGESIZE) 1052168404Spjd hsize <<= 1; 1053168404Spjdretry: 1054168404Spjd buf_hash_table.ht_mask = hsize - 1; 1055168404Spjd buf_hash_table.ht_table = 1056168404Spjd kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1057168404Spjd if (buf_hash_table.ht_table == NULL) { 1058168404Spjd ASSERT(hsize > (1ULL << 8)); 1059168404Spjd hsize >>= 1; 1060168404Spjd goto retry; 1061168404Spjd } 1062168404Spjd 1063168404Spjd hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1064168404Spjd 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1065168404Spjd buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1066185029Spjd 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1067168404Spjd 1068168404Spjd for (i = 0; i < 256; i++) 1069168404Spjd for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1070168404Spjd *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1071168404Spjd 1072168404Spjd for (i = 0; i < BUF_LOCKS; i++) { 1073168404Spjd mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1074168404Spjd NULL, MUTEX_DEFAULT, NULL); 1075168404Spjd } 1076168404Spjd} 1077168404Spjd 1078168404Spjd#define ARC_MINTIME (hz>>4) /* 62 ms */ 1079168404Spjd 1080168404Spjdstatic void 1081168404Spjdarc_cksum_verify(arc_buf_t *buf) 1082168404Spjd{ 1083168404Spjd zio_cksum_t zc; 1084168404Spjd 1085168404Spjd if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1086168404Spjd return; 1087168404Spjd 1088168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1089168404Spjd if (buf->b_hdr->b_freeze_cksum == NULL || 1090168404Spjd (buf->b_hdr->b_flags & ARC_IO_ERROR)) { 1091168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1092168404Spjd return; 1093168404Spjd } 1094168404Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1095168404Spjd if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1096168404Spjd panic("buffer modified while frozen!"); 1097168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1098168404Spjd} 1099168404Spjd 1100185029Spjdstatic int 1101185029Spjdarc_cksum_equal(arc_buf_t *buf) 1102185029Spjd{ 1103185029Spjd zio_cksum_t zc; 1104185029Spjd int equal; 1105185029Spjd 1106185029Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1107185029Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1108185029Spjd equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1109185029Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1110185029Spjd 1111185029Spjd return (equal); 1112185029Spjd} 1113185029Spjd 1114168404Spjdstatic void 1115185029Spjdarc_cksum_compute(arc_buf_t *buf, boolean_t force) 1116168404Spjd{ 1117185029Spjd if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1118168404Spjd return; 1119168404Spjd 1120168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1121168404Spjd if (buf->b_hdr->b_freeze_cksum != NULL) { 1122168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1123168404Spjd return; 1124168404Spjd } 1125168404Spjd buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1126168404Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1127168404Spjd buf->b_hdr->b_freeze_cksum); 1128168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1129240133Smm#ifdef illumos 1130240133Smm arc_buf_watch(buf); 1131240133Smm#endif /* illumos */ 1132168404Spjd} 1133168404Spjd 1134240133Smm#ifdef illumos 1135240133Smm#ifndef _KERNEL 1136240133Smmtypedef struct procctl { 1137240133Smm long cmd; 1138240133Smm prwatch_t prwatch; 1139240133Smm} procctl_t; 1140240133Smm#endif 1141240133Smm 1142240133Smm/* ARGSUSED */ 1143240133Smmstatic void 1144240133Smmarc_buf_unwatch(arc_buf_t *buf) 1145240133Smm{ 1146240133Smm#ifndef _KERNEL 1147240133Smm if (arc_watch) { 1148240133Smm int result; 1149240133Smm procctl_t ctl; 1150240133Smm ctl.cmd = PCWATCH; 1151240133Smm ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1152240133Smm ctl.prwatch.pr_size = 0; 1153240133Smm ctl.prwatch.pr_wflags = 0; 1154240133Smm result = write(arc_procfd, &ctl, sizeof (ctl)); 1155240133Smm ASSERT3U(result, ==, sizeof (ctl)); 1156240133Smm } 1157240133Smm#endif 1158240133Smm} 1159240133Smm 1160240133Smm/* ARGSUSED */ 1161240133Smmstatic void 1162240133Smmarc_buf_watch(arc_buf_t *buf) 1163240133Smm{ 1164240133Smm#ifndef _KERNEL 1165240133Smm if (arc_watch) { 1166240133Smm int result; 1167240133Smm procctl_t ctl; 1168240133Smm ctl.cmd = PCWATCH; 1169240133Smm ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1170240133Smm ctl.prwatch.pr_size = buf->b_hdr->b_size; 1171240133Smm ctl.prwatch.pr_wflags = WA_WRITE; 1172240133Smm result = write(arc_procfd, &ctl, sizeof (ctl)); 1173240133Smm ASSERT3U(result, ==, sizeof (ctl)); 1174240133Smm } 1175240133Smm#endif 1176240133Smm} 1177240133Smm#endif /* illumos */ 1178240133Smm 1179168404Spjdvoid 1180168404Spjdarc_buf_thaw(arc_buf_t *buf) 1181168404Spjd{ 1182185029Spjd if (zfs_flags & ZFS_DEBUG_MODIFY) { 1183185029Spjd if (buf->b_hdr->b_state != arc_anon) 1184185029Spjd panic("modifying non-anon buffer!"); 1185185029Spjd if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS) 1186185029Spjd panic("modifying buffer while i/o in progress!"); 1187185029Spjd arc_cksum_verify(buf); 1188185029Spjd } 1189168404Spjd 1190168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1191168404Spjd if (buf->b_hdr->b_freeze_cksum != NULL) { 1192168404Spjd kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1193168404Spjd buf->b_hdr->b_freeze_cksum = NULL; 1194168404Spjd } 1195219089Spjd 1196219089Spjd if (zfs_flags & ZFS_DEBUG_MODIFY) { 1197219089Spjd if (buf->b_hdr->b_thawed) 1198219089Spjd kmem_free(buf->b_hdr->b_thawed, 1); 1199219089Spjd buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1200219089Spjd } 1201219089Spjd 1202168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1203240133Smm 1204240133Smm#ifdef illumos 1205240133Smm arc_buf_unwatch(buf); 1206240133Smm#endif /* illumos */ 1207168404Spjd} 1208168404Spjd 1209168404Spjdvoid 1210168404Spjdarc_buf_freeze(arc_buf_t *buf) 1211168404Spjd{ 1212219089Spjd kmutex_t *hash_lock; 1213219089Spjd 1214168404Spjd if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1215168404Spjd return; 1216168404Spjd 1217219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 1218219089Spjd mutex_enter(hash_lock); 1219219089Spjd 1220168404Spjd ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1221168404Spjd buf->b_hdr->b_state == arc_anon); 1222185029Spjd arc_cksum_compute(buf, B_FALSE); 1223219089Spjd mutex_exit(hash_lock); 1224240133Smm 1225168404Spjd} 1226168404Spjd 1227168404Spjdstatic void 1228205231Skmacyget_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock) 1229205231Skmacy{ 1230205231Skmacy uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth); 1231205231Skmacy 1232206796Spjd if (ab->b_type == ARC_BUFC_METADATA) 1233206796Spjd buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1234205231Skmacy else { 1235206796Spjd buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1236205231Skmacy buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1237205231Skmacy } 1238205231Skmacy 1239205231Skmacy *list = &state->arcs_lists[buf_hashid]; 1240205231Skmacy *lock = ARCS_LOCK(state, buf_hashid); 1241205231Skmacy} 1242205231Skmacy 1243205231Skmacy 1244205231Skmacystatic void 1245168404Spjdadd_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1246168404Spjd{ 1247168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 1248168404Spjd 1249168404Spjd if ((refcount_add(&ab->b_refcnt, tag) == 1) && 1250168404Spjd (ab->b_state != arc_anon)) { 1251206796Spjd uint64_t delta = ab->b_size * ab->b_datacnt; 1252206796Spjd uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type]; 1253205231Skmacy list_t *list; 1254205231Skmacy kmutex_t *lock; 1255168404Spjd 1256205231Skmacy get_buf_info(ab, ab->b_state, &list, &lock); 1257205231Skmacy ASSERT(!MUTEX_HELD(lock)); 1258205231Skmacy mutex_enter(lock); 1259168404Spjd ASSERT(list_link_active(&ab->b_arc_node)); 1260185029Spjd list_remove(list, ab); 1261168404Spjd if (GHOST_STATE(ab->b_state)) { 1262240415Smm ASSERT0(ab->b_datacnt); 1263168404Spjd ASSERT3P(ab->b_buf, ==, NULL); 1264168404Spjd delta = ab->b_size; 1265168404Spjd } 1266168404Spjd ASSERT(delta > 0); 1267185029Spjd ASSERT3U(*size, >=, delta); 1268185029Spjd atomic_add_64(size, -delta); 1269206794Spjd mutex_exit(lock); 1270185029Spjd /* remove the prefetch flag if we get a reference */ 1271168404Spjd if (ab->b_flags & ARC_PREFETCH) 1272168404Spjd ab->b_flags &= ~ARC_PREFETCH; 1273168404Spjd } 1274168404Spjd} 1275168404Spjd 1276168404Spjdstatic int 1277168404Spjdremove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1278168404Spjd{ 1279168404Spjd int cnt; 1280168404Spjd arc_state_t *state = ab->b_state; 1281168404Spjd 1282168404Spjd ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1283168404Spjd ASSERT(!GHOST_STATE(state)); 1284168404Spjd 1285168404Spjd if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) && 1286168404Spjd (state != arc_anon)) { 1287185029Spjd uint64_t *size = &state->arcs_lsize[ab->b_type]; 1288205231Skmacy list_t *list; 1289205231Skmacy kmutex_t *lock; 1290185029Spjd 1291205231Skmacy get_buf_info(ab, state, &list, &lock); 1292205231Skmacy ASSERT(!MUTEX_HELD(lock)); 1293205231Skmacy mutex_enter(lock); 1294168404Spjd ASSERT(!list_link_active(&ab->b_arc_node)); 1295205231Skmacy list_insert_head(list, ab); 1296168404Spjd ASSERT(ab->b_datacnt > 0); 1297185029Spjd atomic_add_64(size, ab->b_size * ab->b_datacnt); 1298206794Spjd mutex_exit(lock); 1299168404Spjd } 1300168404Spjd return (cnt); 1301168404Spjd} 1302168404Spjd 1303168404Spjd/* 1304168404Spjd * Move the supplied buffer to the indicated state. The mutex 1305168404Spjd * for the buffer must be held by the caller. 1306168404Spjd */ 1307168404Spjdstatic void 1308168404Spjdarc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock) 1309168404Spjd{ 1310168404Spjd arc_state_t *old_state = ab->b_state; 1311168404Spjd int64_t refcnt = refcount_count(&ab->b_refcnt); 1312168404Spjd uint64_t from_delta, to_delta; 1313205231Skmacy list_t *list; 1314205231Skmacy kmutex_t *lock; 1315168404Spjd 1316168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 1317168404Spjd ASSERT(new_state != old_state); 1318168404Spjd ASSERT(refcnt == 0 || ab->b_datacnt > 0); 1319168404Spjd ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state)); 1320219089Spjd ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon); 1321168404Spjd 1322168404Spjd from_delta = to_delta = ab->b_datacnt * ab->b_size; 1323168404Spjd 1324168404Spjd /* 1325168404Spjd * If this buffer is evictable, transfer it from the 1326168404Spjd * old state list to the new state list. 1327168404Spjd */ 1328168404Spjd if (refcnt == 0) { 1329168404Spjd if (old_state != arc_anon) { 1330205231Skmacy int use_mutex; 1331185029Spjd uint64_t *size = &old_state->arcs_lsize[ab->b_type]; 1332168404Spjd 1333205231Skmacy get_buf_info(ab, old_state, &list, &lock); 1334205231Skmacy use_mutex = !MUTEX_HELD(lock); 1335168404Spjd if (use_mutex) 1336205231Skmacy mutex_enter(lock); 1337168404Spjd 1338168404Spjd ASSERT(list_link_active(&ab->b_arc_node)); 1339205231Skmacy list_remove(list, ab); 1340168404Spjd 1341168404Spjd /* 1342168404Spjd * If prefetching out of the ghost cache, 1343219089Spjd * we will have a non-zero datacnt. 1344168404Spjd */ 1345168404Spjd if (GHOST_STATE(old_state) && ab->b_datacnt == 0) { 1346168404Spjd /* ghost elements have a ghost size */ 1347168404Spjd ASSERT(ab->b_buf == NULL); 1348168404Spjd from_delta = ab->b_size; 1349168404Spjd } 1350185029Spjd ASSERT3U(*size, >=, from_delta); 1351185029Spjd atomic_add_64(size, -from_delta); 1352168404Spjd 1353168404Spjd if (use_mutex) 1354205231Skmacy mutex_exit(lock); 1355168404Spjd } 1356168404Spjd if (new_state != arc_anon) { 1357206796Spjd int use_mutex; 1358185029Spjd uint64_t *size = &new_state->arcs_lsize[ab->b_type]; 1359168404Spjd 1360205231Skmacy get_buf_info(ab, new_state, &list, &lock); 1361205231Skmacy use_mutex = !MUTEX_HELD(lock); 1362168404Spjd if (use_mutex) 1363205231Skmacy mutex_enter(lock); 1364168404Spjd 1365205231Skmacy list_insert_head(list, ab); 1366168404Spjd 1367168404Spjd /* ghost elements have a ghost size */ 1368168404Spjd if (GHOST_STATE(new_state)) { 1369168404Spjd ASSERT(ab->b_datacnt == 0); 1370168404Spjd ASSERT(ab->b_buf == NULL); 1371168404Spjd to_delta = ab->b_size; 1372168404Spjd } 1373185029Spjd atomic_add_64(size, to_delta); 1374168404Spjd 1375168404Spjd if (use_mutex) 1376205231Skmacy mutex_exit(lock); 1377168404Spjd } 1378168404Spjd } 1379168404Spjd 1380168404Spjd ASSERT(!BUF_EMPTY(ab)); 1381219089Spjd if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab)) 1382168404Spjd buf_hash_remove(ab); 1383168404Spjd 1384168404Spjd /* adjust state sizes */ 1385168404Spjd if (to_delta) 1386168404Spjd atomic_add_64(&new_state->arcs_size, to_delta); 1387168404Spjd if (from_delta) { 1388168404Spjd ASSERT3U(old_state->arcs_size, >=, from_delta); 1389168404Spjd atomic_add_64(&old_state->arcs_size, -from_delta); 1390168404Spjd } 1391168404Spjd ab->b_state = new_state; 1392185029Spjd 1393185029Spjd /* adjust l2arc hdr stats */ 1394185029Spjd if (new_state == arc_l2c_only) 1395185029Spjd l2arc_hdr_stat_add(); 1396185029Spjd else if (old_state == arc_l2c_only) 1397185029Spjd l2arc_hdr_stat_remove(); 1398168404Spjd} 1399168404Spjd 1400185029Spjdvoid 1401208373Smmarc_space_consume(uint64_t space, arc_space_type_t type) 1402185029Spjd{ 1403208373Smm ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1404208373Smm 1405208373Smm switch (type) { 1406208373Smm case ARC_SPACE_DATA: 1407208373Smm ARCSTAT_INCR(arcstat_data_size, space); 1408208373Smm break; 1409208373Smm case ARC_SPACE_OTHER: 1410208373Smm ARCSTAT_INCR(arcstat_other_size, space); 1411208373Smm break; 1412208373Smm case ARC_SPACE_HDRS: 1413208373Smm ARCSTAT_INCR(arcstat_hdr_size, space); 1414208373Smm break; 1415208373Smm case ARC_SPACE_L2HDRS: 1416208373Smm ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1417208373Smm break; 1418208373Smm } 1419208373Smm 1420185029Spjd atomic_add_64(&arc_meta_used, space); 1421185029Spjd atomic_add_64(&arc_size, space); 1422185029Spjd} 1423185029Spjd 1424185029Spjdvoid 1425208373Smmarc_space_return(uint64_t space, arc_space_type_t type) 1426185029Spjd{ 1427208373Smm ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1428208373Smm 1429208373Smm switch (type) { 1430208373Smm case ARC_SPACE_DATA: 1431208373Smm ARCSTAT_INCR(arcstat_data_size, -space); 1432208373Smm break; 1433208373Smm case ARC_SPACE_OTHER: 1434208373Smm ARCSTAT_INCR(arcstat_other_size, -space); 1435208373Smm break; 1436208373Smm case ARC_SPACE_HDRS: 1437208373Smm ARCSTAT_INCR(arcstat_hdr_size, -space); 1438208373Smm break; 1439208373Smm case ARC_SPACE_L2HDRS: 1440208373Smm ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1441208373Smm break; 1442208373Smm } 1443208373Smm 1444185029Spjd ASSERT(arc_meta_used >= space); 1445185029Spjd if (arc_meta_max < arc_meta_used) 1446185029Spjd arc_meta_max = arc_meta_used; 1447185029Spjd atomic_add_64(&arc_meta_used, -space); 1448185029Spjd ASSERT(arc_size >= space); 1449185029Spjd atomic_add_64(&arc_size, -space); 1450185029Spjd} 1451185029Spjd 1452185029Spjdvoid * 1453185029Spjdarc_data_buf_alloc(uint64_t size) 1454185029Spjd{ 1455185029Spjd if (arc_evict_needed(ARC_BUFC_DATA)) 1456185029Spjd cv_signal(&arc_reclaim_thr_cv); 1457185029Spjd atomic_add_64(&arc_size, size); 1458185029Spjd return (zio_data_buf_alloc(size)); 1459185029Spjd} 1460185029Spjd 1461185029Spjdvoid 1462185029Spjdarc_data_buf_free(void *buf, uint64_t size) 1463185029Spjd{ 1464185029Spjd zio_data_buf_free(buf, size); 1465185029Spjd ASSERT(arc_size >= size); 1466185029Spjd atomic_add_64(&arc_size, -size); 1467185029Spjd} 1468185029Spjd 1469168404Spjdarc_buf_t * 1470168404Spjdarc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1471168404Spjd{ 1472168404Spjd arc_buf_hdr_t *hdr; 1473168404Spjd arc_buf_t *buf; 1474168404Spjd 1475168404Spjd ASSERT3U(size, >, 0); 1476185029Spjd hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1477168404Spjd ASSERT(BUF_EMPTY(hdr)); 1478168404Spjd hdr->b_size = size; 1479168404Spjd hdr->b_type = type; 1480228103Smm hdr->b_spa = spa_load_guid(spa); 1481168404Spjd hdr->b_state = arc_anon; 1482168404Spjd hdr->b_arc_access = 0; 1483185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1484168404Spjd buf->b_hdr = hdr; 1485168404Spjd buf->b_data = NULL; 1486168404Spjd buf->b_efunc = NULL; 1487168404Spjd buf->b_private = NULL; 1488168404Spjd buf->b_next = NULL; 1489168404Spjd hdr->b_buf = buf; 1490168404Spjd arc_get_data_buf(buf); 1491168404Spjd hdr->b_datacnt = 1; 1492168404Spjd hdr->b_flags = 0; 1493168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1494168404Spjd (void) refcount_add(&hdr->b_refcnt, tag); 1495168404Spjd 1496168404Spjd return (buf); 1497168404Spjd} 1498168404Spjd 1499209962Smmstatic char *arc_onloan_tag = "onloan"; 1500209962Smm 1501209962Smm/* 1502209962Smm * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1503209962Smm * flight data by arc_tempreserve_space() until they are "returned". Loaned 1504209962Smm * buffers must be returned to the arc before they can be used by the DMU or 1505209962Smm * freed. 1506209962Smm */ 1507209962Smmarc_buf_t * 1508209962Smmarc_loan_buf(spa_t *spa, int size) 1509209962Smm{ 1510209962Smm arc_buf_t *buf; 1511209962Smm 1512209962Smm buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1513209962Smm 1514209962Smm atomic_add_64(&arc_loaned_bytes, size); 1515209962Smm return (buf); 1516209962Smm} 1517209962Smm 1518209962Smm/* 1519209962Smm * Return a loaned arc buffer to the arc. 1520209962Smm */ 1521209962Smmvoid 1522209962Smmarc_return_buf(arc_buf_t *buf, void *tag) 1523209962Smm{ 1524209962Smm arc_buf_hdr_t *hdr = buf->b_hdr; 1525209962Smm 1526209962Smm ASSERT(buf->b_data != NULL); 1527219089Spjd (void) refcount_add(&hdr->b_refcnt, tag); 1528219089Spjd (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1529209962Smm 1530209962Smm atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1531209962Smm} 1532209962Smm 1533219089Spjd/* Detach an arc_buf from a dbuf (tag) */ 1534219089Spjdvoid 1535219089Spjdarc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1536219089Spjd{ 1537219089Spjd arc_buf_hdr_t *hdr; 1538219089Spjd 1539219089Spjd ASSERT(buf->b_data != NULL); 1540219089Spjd hdr = buf->b_hdr; 1541219089Spjd (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1542219089Spjd (void) refcount_remove(&hdr->b_refcnt, tag); 1543219089Spjd buf->b_efunc = NULL; 1544219089Spjd buf->b_private = NULL; 1545219089Spjd 1546219089Spjd atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1547219089Spjd} 1548219089Spjd 1549168404Spjdstatic arc_buf_t * 1550168404Spjdarc_buf_clone(arc_buf_t *from) 1551168404Spjd{ 1552168404Spjd arc_buf_t *buf; 1553168404Spjd arc_buf_hdr_t *hdr = from->b_hdr; 1554168404Spjd uint64_t size = hdr->b_size; 1555168404Spjd 1556219089Spjd ASSERT(hdr->b_state != arc_anon); 1557219089Spjd 1558185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1559168404Spjd buf->b_hdr = hdr; 1560168404Spjd buf->b_data = NULL; 1561168404Spjd buf->b_efunc = NULL; 1562168404Spjd buf->b_private = NULL; 1563168404Spjd buf->b_next = hdr->b_buf; 1564168404Spjd hdr->b_buf = buf; 1565168404Spjd arc_get_data_buf(buf); 1566168404Spjd bcopy(from->b_data, buf->b_data, size); 1567242845Sdelphij 1568242845Sdelphij /* 1569242845Sdelphij * This buffer already exists in the arc so create a duplicate 1570242845Sdelphij * copy for the caller. If the buffer is associated with user data 1571242845Sdelphij * then track the size and number of duplicates. These stats will be 1572242845Sdelphij * updated as duplicate buffers are created and destroyed. 1573242845Sdelphij */ 1574242845Sdelphij if (hdr->b_type == ARC_BUFC_DATA) { 1575242845Sdelphij ARCSTAT_BUMP(arcstat_duplicate_buffers); 1576242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1577242845Sdelphij } 1578168404Spjd hdr->b_datacnt += 1; 1579168404Spjd return (buf); 1580168404Spjd} 1581168404Spjd 1582168404Spjdvoid 1583168404Spjdarc_buf_add_ref(arc_buf_t *buf, void* tag) 1584168404Spjd{ 1585168404Spjd arc_buf_hdr_t *hdr; 1586168404Spjd kmutex_t *hash_lock; 1587168404Spjd 1588168404Spjd /* 1589185029Spjd * Check to see if this buffer is evicted. Callers 1590185029Spjd * must verify b_data != NULL to know if the add_ref 1591185029Spjd * was successful. 1592168404Spjd */ 1593219089Spjd mutex_enter(&buf->b_evict_lock); 1594185029Spjd if (buf->b_data == NULL) { 1595219089Spjd mutex_exit(&buf->b_evict_lock); 1596168404Spjd return; 1597168404Spjd } 1598219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 1599219089Spjd mutex_enter(hash_lock); 1600185029Spjd hdr = buf->b_hdr; 1601219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1602219089Spjd mutex_exit(&buf->b_evict_lock); 1603168404Spjd 1604168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1605168404Spjd add_reference(hdr, hash_lock, tag); 1606208373Smm DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1607168404Spjd arc_access(hdr, hash_lock); 1608168404Spjd mutex_exit(hash_lock); 1609168404Spjd ARCSTAT_BUMP(arcstat_hits); 1610168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 1611168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1612168404Spjd data, metadata, hits); 1613168404Spjd} 1614168404Spjd 1615185029Spjd/* 1616185029Spjd * Free the arc data buffer. If it is an l2arc write in progress, 1617185029Spjd * the buffer is placed on l2arc_free_on_write to be freed later. 1618185029Spjd */ 1619168404Spjdstatic void 1620240133Smmarc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1621185029Spjd{ 1622240133Smm arc_buf_hdr_t *hdr = buf->b_hdr; 1623240133Smm 1624185029Spjd if (HDR_L2_WRITING(hdr)) { 1625185029Spjd l2arc_data_free_t *df; 1626185029Spjd df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1627240133Smm df->l2df_data = buf->b_data; 1628240133Smm df->l2df_size = hdr->b_size; 1629185029Spjd df->l2df_func = free_func; 1630185029Spjd mutex_enter(&l2arc_free_on_write_mtx); 1631185029Spjd list_insert_head(l2arc_free_on_write, df); 1632185029Spjd mutex_exit(&l2arc_free_on_write_mtx); 1633185029Spjd ARCSTAT_BUMP(arcstat_l2_free_on_write); 1634185029Spjd } else { 1635240133Smm free_func(buf->b_data, hdr->b_size); 1636185029Spjd } 1637185029Spjd} 1638185029Spjd 1639185029Spjdstatic void 1640168404Spjdarc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all) 1641168404Spjd{ 1642168404Spjd arc_buf_t **bufp; 1643168404Spjd 1644168404Spjd /* free up data associated with the buf */ 1645168404Spjd if (buf->b_data) { 1646168404Spjd arc_state_t *state = buf->b_hdr->b_state; 1647168404Spjd uint64_t size = buf->b_hdr->b_size; 1648168404Spjd arc_buf_contents_t type = buf->b_hdr->b_type; 1649168404Spjd 1650168404Spjd arc_cksum_verify(buf); 1651240133Smm#ifdef illumos 1652240133Smm arc_buf_unwatch(buf); 1653240133Smm#endif /* illumos */ 1654219089Spjd 1655168404Spjd if (!recycle) { 1656168404Spjd if (type == ARC_BUFC_METADATA) { 1657240133Smm arc_buf_data_free(buf, zio_buf_free); 1658208373Smm arc_space_return(size, ARC_SPACE_DATA); 1659168404Spjd } else { 1660168404Spjd ASSERT(type == ARC_BUFC_DATA); 1661240133Smm arc_buf_data_free(buf, zio_data_buf_free); 1662208373Smm ARCSTAT_INCR(arcstat_data_size, -size); 1663185029Spjd atomic_add_64(&arc_size, -size); 1664168404Spjd } 1665168404Spjd } 1666168404Spjd if (list_link_active(&buf->b_hdr->b_arc_node)) { 1667185029Spjd uint64_t *cnt = &state->arcs_lsize[type]; 1668185029Spjd 1669168404Spjd ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1670168404Spjd ASSERT(state != arc_anon); 1671185029Spjd 1672185029Spjd ASSERT3U(*cnt, >=, size); 1673185029Spjd atomic_add_64(cnt, -size); 1674168404Spjd } 1675168404Spjd ASSERT3U(state->arcs_size, >=, size); 1676168404Spjd atomic_add_64(&state->arcs_size, -size); 1677168404Spjd buf->b_data = NULL; 1678242845Sdelphij 1679242845Sdelphij /* 1680242845Sdelphij * If we're destroying a duplicate buffer make sure 1681242845Sdelphij * that the appropriate statistics are updated. 1682242845Sdelphij */ 1683242845Sdelphij if (buf->b_hdr->b_datacnt > 1 && 1684242845Sdelphij buf->b_hdr->b_type == ARC_BUFC_DATA) { 1685242845Sdelphij ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1686242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1687242845Sdelphij } 1688168404Spjd ASSERT(buf->b_hdr->b_datacnt > 0); 1689168404Spjd buf->b_hdr->b_datacnt -= 1; 1690168404Spjd } 1691168404Spjd 1692168404Spjd /* only remove the buf if requested */ 1693168404Spjd if (!all) 1694168404Spjd return; 1695168404Spjd 1696168404Spjd /* remove the buf from the hdr list */ 1697168404Spjd for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1698168404Spjd continue; 1699168404Spjd *bufp = buf->b_next; 1700219089Spjd buf->b_next = NULL; 1701168404Spjd 1702168404Spjd ASSERT(buf->b_efunc == NULL); 1703168404Spjd 1704168404Spjd /* clean up the buf */ 1705168404Spjd buf->b_hdr = NULL; 1706168404Spjd kmem_cache_free(buf_cache, buf); 1707168404Spjd} 1708168404Spjd 1709168404Spjdstatic void 1710168404Spjdarc_hdr_destroy(arc_buf_hdr_t *hdr) 1711168404Spjd{ 1712168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1713168404Spjd ASSERT3P(hdr->b_state, ==, arc_anon); 1714168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1715219089Spjd l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1716168404Spjd 1717219089Spjd if (l2hdr != NULL) { 1718219089Spjd boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1719219089Spjd /* 1720219089Spjd * To prevent arc_free() and l2arc_evict() from 1721219089Spjd * attempting to free the same buffer at the same time, 1722219089Spjd * a FREE_IN_PROGRESS flag is given to arc_free() to 1723219089Spjd * give it priority. l2arc_evict() can't destroy this 1724219089Spjd * header while we are waiting on l2arc_buflist_mtx. 1725219089Spjd * 1726219089Spjd * The hdr may be removed from l2ad_buflist before we 1727219089Spjd * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1728219089Spjd */ 1729219089Spjd if (!buflist_held) { 1730185029Spjd mutex_enter(&l2arc_buflist_mtx); 1731219089Spjd l2hdr = hdr->b_l2hdr; 1732219089Spjd } 1733219089Spjd 1734219089Spjd if (l2hdr != NULL) { 1735248572Ssmh trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1736248574Ssmh hdr->b_size, 0); 1737219089Spjd list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1738219089Spjd ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1739251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1740219089Spjd kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1741219089Spjd if (hdr->b_state == arc_l2c_only) 1742219089Spjd l2arc_hdr_stat_remove(); 1743219089Spjd hdr->b_l2hdr = NULL; 1744219089Spjd } 1745219089Spjd 1746219089Spjd if (!buflist_held) 1747185029Spjd mutex_exit(&l2arc_buflist_mtx); 1748185029Spjd } 1749185029Spjd 1750168404Spjd if (!BUF_EMPTY(hdr)) { 1751168404Spjd ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1752219089Spjd buf_discard_identity(hdr); 1753168404Spjd } 1754168404Spjd while (hdr->b_buf) { 1755168404Spjd arc_buf_t *buf = hdr->b_buf; 1756168404Spjd 1757168404Spjd if (buf->b_efunc) { 1758168404Spjd mutex_enter(&arc_eviction_mtx); 1759219089Spjd mutex_enter(&buf->b_evict_lock); 1760168404Spjd ASSERT(buf->b_hdr != NULL); 1761168404Spjd arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1762168404Spjd hdr->b_buf = buf->b_next; 1763168404Spjd buf->b_hdr = &arc_eviction_hdr; 1764168404Spjd buf->b_next = arc_eviction_list; 1765168404Spjd arc_eviction_list = buf; 1766219089Spjd mutex_exit(&buf->b_evict_lock); 1767168404Spjd mutex_exit(&arc_eviction_mtx); 1768168404Spjd } else { 1769168404Spjd arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1770168404Spjd } 1771168404Spjd } 1772168404Spjd if (hdr->b_freeze_cksum != NULL) { 1773168404Spjd kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1774168404Spjd hdr->b_freeze_cksum = NULL; 1775168404Spjd } 1776219089Spjd if (hdr->b_thawed) { 1777219089Spjd kmem_free(hdr->b_thawed, 1); 1778219089Spjd hdr->b_thawed = NULL; 1779219089Spjd } 1780168404Spjd 1781168404Spjd ASSERT(!list_link_active(&hdr->b_arc_node)); 1782168404Spjd ASSERT3P(hdr->b_hash_next, ==, NULL); 1783168404Spjd ASSERT3P(hdr->b_acb, ==, NULL); 1784168404Spjd kmem_cache_free(hdr_cache, hdr); 1785168404Spjd} 1786168404Spjd 1787168404Spjdvoid 1788168404Spjdarc_buf_free(arc_buf_t *buf, void *tag) 1789168404Spjd{ 1790168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 1791168404Spjd int hashed = hdr->b_state != arc_anon; 1792168404Spjd 1793168404Spjd ASSERT(buf->b_efunc == NULL); 1794168404Spjd ASSERT(buf->b_data != NULL); 1795168404Spjd 1796168404Spjd if (hashed) { 1797168404Spjd kmutex_t *hash_lock = HDR_LOCK(hdr); 1798168404Spjd 1799168404Spjd mutex_enter(hash_lock); 1800219089Spjd hdr = buf->b_hdr; 1801219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1802219089Spjd 1803168404Spjd (void) remove_reference(hdr, hash_lock, tag); 1804219089Spjd if (hdr->b_datacnt > 1) { 1805168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1806219089Spjd } else { 1807219089Spjd ASSERT(buf == hdr->b_buf); 1808219089Spjd ASSERT(buf->b_efunc == NULL); 1809168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 1810219089Spjd } 1811168404Spjd mutex_exit(hash_lock); 1812168404Spjd } else if (HDR_IO_IN_PROGRESS(hdr)) { 1813168404Spjd int destroy_hdr; 1814168404Spjd /* 1815168404Spjd * We are in the middle of an async write. Don't destroy 1816168404Spjd * this buffer unless the write completes before we finish 1817168404Spjd * decrementing the reference count. 1818168404Spjd */ 1819168404Spjd mutex_enter(&arc_eviction_mtx); 1820168404Spjd (void) remove_reference(hdr, NULL, tag); 1821168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1822168404Spjd destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1823168404Spjd mutex_exit(&arc_eviction_mtx); 1824168404Spjd if (destroy_hdr) 1825168404Spjd arc_hdr_destroy(hdr); 1826168404Spjd } else { 1827219089Spjd if (remove_reference(hdr, NULL, tag) > 0) 1828168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1829219089Spjd else 1830168404Spjd arc_hdr_destroy(hdr); 1831168404Spjd } 1832168404Spjd} 1833168404Spjd 1834248571Smmboolean_t 1835168404Spjdarc_buf_remove_ref(arc_buf_t *buf, void* tag) 1836168404Spjd{ 1837168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 1838168404Spjd kmutex_t *hash_lock = HDR_LOCK(hdr); 1839248571Smm boolean_t no_callback = (buf->b_efunc == NULL); 1840168404Spjd 1841168404Spjd if (hdr->b_state == arc_anon) { 1842219089Spjd ASSERT(hdr->b_datacnt == 1); 1843168404Spjd arc_buf_free(buf, tag); 1844168404Spjd return (no_callback); 1845168404Spjd } 1846168404Spjd 1847168404Spjd mutex_enter(hash_lock); 1848219089Spjd hdr = buf->b_hdr; 1849219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1850168404Spjd ASSERT(hdr->b_state != arc_anon); 1851168404Spjd ASSERT(buf->b_data != NULL); 1852168404Spjd 1853168404Spjd (void) remove_reference(hdr, hash_lock, tag); 1854168404Spjd if (hdr->b_datacnt > 1) { 1855168404Spjd if (no_callback) 1856168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1857168404Spjd } else if (no_callback) { 1858168404Spjd ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1859219089Spjd ASSERT(buf->b_efunc == NULL); 1860168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 1861168404Spjd } 1862168404Spjd ASSERT(no_callback || hdr->b_datacnt > 1 || 1863168404Spjd refcount_is_zero(&hdr->b_refcnt)); 1864168404Spjd mutex_exit(hash_lock); 1865168404Spjd return (no_callback); 1866168404Spjd} 1867168404Spjd 1868168404Spjdint 1869168404Spjdarc_buf_size(arc_buf_t *buf) 1870168404Spjd{ 1871168404Spjd return (buf->b_hdr->b_size); 1872168404Spjd} 1873168404Spjd 1874168404Spjd/* 1875242845Sdelphij * Called from the DMU to determine if the current buffer should be 1876242845Sdelphij * evicted. In order to ensure proper locking, the eviction must be initiated 1877242845Sdelphij * from the DMU. Return true if the buffer is associated with user data and 1878242845Sdelphij * duplicate buffers still exist. 1879242845Sdelphij */ 1880242845Sdelphijboolean_t 1881242845Sdelphijarc_buf_eviction_needed(arc_buf_t *buf) 1882242845Sdelphij{ 1883242845Sdelphij arc_buf_hdr_t *hdr; 1884242845Sdelphij boolean_t evict_needed = B_FALSE; 1885242845Sdelphij 1886242845Sdelphij if (zfs_disable_dup_eviction) 1887242845Sdelphij return (B_FALSE); 1888242845Sdelphij 1889242845Sdelphij mutex_enter(&buf->b_evict_lock); 1890242845Sdelphij hdr = buf->b_hdr; 1891242845Sdelphij if (hdr == NULL) { 1892242845Sdelphij /* 1893242845Sdelphij * We are in arc_do_user_evicts(); let that function 1894242845Sdelphij * perform the eviction. 1895242845Sdelphij */ 1896242845Sdelphij ASSERT(buf->b_data == NULL); 1897242845Sdelphij mutex_exit(&buf->b_evict_lock); 1898242845Sdelphij return (B_FALSE); 1899242845Sdelphij } else if (buf->b_data == NULL) { 1900242845Sdelphij /* 1901242845Sdelphij * We have already been added to the arc eviction list; 1902242845Sdelphij * recommend eviction. 1903242845Sdelphij */ 1904242845Sdelphij ASSERT3P(hdr, ==, &arc_eviction_hdr); 1905242845Sdelphij mutex_exit(&buf->b_evict_lock); 1906242845Sdelphij return (B_TRUE); 1907242845Sdelphij } 1908242845Sdelphij 1909242845Sdelphij if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 1910242845Sdelphij evict_needed = B_TRUE; 1911242845Sdelphij 1912242845Sdelphij mutex_exit(&buf->b_evict_lock); 1913242845Sdelphij return (evict_needed); 1914242845Sdelphij} 1915242845Sdelphij 1916242845Sdelphij/* 1917168404Spjd * Evict buffers from list until we've removed the specified number of 1918168404Spjd * bytes. Move the removed buffers to the appropriate evict state. 1919168404Spjd * If the recycle flag is set, then attempt to "recycle" a buffer: 1920168404Spjd * - look for a buffer to evict that is `bytes' long. 1921168404Spjd * - return the data block from this buffer rather than freeing it. 1922168404Spjd * This flag is used by callers that are trying to make space for a 1923168404Spjd * new buffer in a full arc cache. 1924185029Spjd * 1925185029Spjd * This function makes a "best effort". It skips over any buffers 1926185029Spjd * it can't get a hash_lock on, and so may not catch all candidates. 1927185029Spjd * It may also return without evicting as much space as requested. 1928168404Spjd */ 1929168404Spjdstatic void * 1930209962Smmarc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 1931168404Spjd arc_buf_contents_t type) 1932168404Spjd{ 1933168404Spjd arc_state_t *evicted_state; 1934168404Spjd uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 1935205231Skmacy int64_t bytes_remaining; 1936168404Spjd arc_buf_hdr_t *ab, *ab_prev = NULL; 1937205231Skmacy list_t *evicted_list, *list, *evicted_list_start, *list_start; 1938205231Skmacy kmutex_t *lock, *evicted_lock; 1939168404Spjd kmutex_t *hash_lock; 1940168404Spjd boolean_t have_lock; 1941168404Spjd void *stolen = NULL; 1942205231Skmacy static int evict_metadata_offset, evict_data_offset; 1943205231Skmacy int i, idx, offset, list_count, count; 1944168404Spjd 1945168404Spjd ASSERT(state == arc_mru || state == arc_mfu); 1946168404Spjd 1947168404Spjd evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 1948206796Spjd 1949205231Skmacy if (type == ARC_BUFC_METADATA) { 1950205231Skmacy offset = 0; 1951205231Skmacy list_count = ARC_BUFC_NUMMETADATALISTS; 1952205231Skmacy list_start = &state->arcs_lists[0]; 1953205231Skmacy evicted_list_start = &evicted_state->arcs_lists[0]; 1954205231Skmacy idx = evict_metadata_offset; 1955205231Skmacy } else { 1956205231Skmacy offset = ARC_BUFC_NUMMETADATALISTS; 1957205231Skmacy list_start = &state->arcs_lists[offset]; 1958205231Skmacy evicted_list_start = &evicted_state->arcs_lists[offset]; 1959205231Skmacy list_count = ARC_BUFC_NUMDATALISTS; 1960205231Skmacy idx = evict_data_offset; 1961205231Skmacy } 1962205231Skmacy bytes_remaining = evicted_state->arcs_lsize[type]; 1963205231Skmacy count = 0; 1964206796Spjd 1965205231Skmacyevict_start: 1966205231Skmacy list = &list_start[idx]; 1967205231Skmacy evicted_list = &evicted_list_start[idx]; 1968205231Skmacy lock = ARCS_LOCK(state, (offset + idx)); 1969206796Spjd evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 1970168404Spjd 1971205231Skmacy mutex_enter(lock); 1972205231Skmacy mutex_enter(evicted_lock); 1973205231Skmacy 1974185029Spjd for (ab = list_tail(list); ab; ab = ab_prev) { 1975185029Spjd ab_prev = list_prev(list, ab); 1976205231Skmacy bytes_remaining -= (ab->b_size * ab->b_datacnt); 1977168404Spjd /* prefetch buffers have a minimum lifespan */ 1978168404Spjd if (HDR_IO_IN_PROGRESS(ab) || 1979185029Spjd (spa && ab->b_spa != spa) || 1980168404Spjd (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 1981219089Spjd ddi_get_lbolt() - ab->b_arc_access < 1982219089Spjd arc_min_prefetch_lifespan)) { 1983168404Spjd skipped++; 1984168404Spjd continue; 1985168404Spjd } 1986168404Spjd /* "lookahead" for better eviction candidate */ 1987168404Spjd if (recycle && ab->b_size != bytes && 1988168404Spjd ab_prev && ab_prev->b_size == bytes) 1989168404Spjd continue; 1990168404Spjd hash_lock = HDR_LOCK(ab); 1991168404Spjd have_lock = MUTEX_HELD(hash_lock); 1992168404Spjd if (have_lock || mutex_tryenter(hash_lock)) { 1993240415Smm ASSERT0(refcount_count(&ab->b_refcnt)); 1994168404Spjd ASSERT(ab->b_datacnt > 0); 1995168404Spjd while (ab->b_buf) { 1996168404Spjd arc_buf_t *buf = ab->b_buf; 1997219089Spjd if (!mutex_tryenter(&buf->b_evict_lock)) { 1998185029Spjd missed += 1; 1999185029Spjd break; 2000185029Spjd } 2001168404Spjd if (buf->b_data) { 2002168404Spjd bytes_evicted += ab->b_size; 2003168404Spjd if (recycle && ab->b_type == type && 2004185029Spjd ab->b_size == bytes && 2005185029Spjd !HDR_L2_WRITING(ab)) { 2006168404Spjd stolen = buf->b_data; 2007168404Spjd recycle = FALSE; 2008168404Spjd } 2009168404Spjd } 2010168404Spjd if (buf->b_efunc) { 2011168404Spjd mutex_enter(&arc_eviction_mtx); 2012168404Spjd arc_buf_destroy(buf, 2013168404Spjd buf->b_data == stolen, FALSE); 2014168404Spjd ab->b_buf = buf->b_next; 2015168404Spjd buf->b_hdr = &arc_eviction_hdr; 2016168404Spjd buf->b_next = arc_eviction_list; 2017168404Spjd arc_eviction_list = buf; 2018168404Spjd mutex_exit(&arc_eviction_mtx); 2019219089Spjd mutex_exit(&buf->b_evict_lock); 2020168404Spjd } else { 2021219089Spjd mutex_exit(&buf->b_evict_lock); 2022168404Spjd arc_buf_destroy(buf, 2023168404Spjd buf->b_data == stolen, TRUE); 2024168404Spjd } 2025168404Spjd } 2026208373Smm 2027208373Smm if (ab->b_l2hdr) { 2028208373Smm ARCSTAT_INCR(arcstat_evict_l2_cached, 2029208373Smm ab->b_size); 2030208373Smm } else { 2031208373Smm if (l2arc_write_eligible(ab->b_spa, ab)) { 2032208373Smm ARCSTAT_INCR(arcstat_evict_l2_eligible, 2033208373Smm ab->b_size); 2034208373Smm } else { 2035208373Smm ARCSTAT_INCR( 2036208373Smm arcstat_evict_l2_ineligible, 2037208373Smm ab->b_size); 2038208373Smm } 2039208373Smm } 2040208373Smm 2041185029Spjd if (ab->b_datacnt == 0) { 2042185029Spjd arc_change_state(evicted_state, ab, hash_lock); 2043185029Spjd ASSERT(HDR_IN_HASH_TABLE(ab)); 2044185029Spjd ab->b_flags |= ARC_IN_HASH_TABLE; 2045185029Spjd ab->b_flags &= ~ARC_BUF_AVAILABLE; 2046185029Spjd DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 2047185029Spjd } 2048168404Spjd if (!have_lock) 2049168404Spjd mutex_exit(hash_lock); 2050168404Spjd if (bytes >= 0 && bytes_evicted >= bytes) 2051168404Spjd break; 2052205231Skmacy if (bytes_remaining > 0) { 2053205231Skmacy mutex_exit(evicted_lock); 2054205231Skmacy mutex_exit(lock); 2055206796Spjd idx = ((idx + 1) & (list_count - 1)); 2056205231Skmacy count++; 2057205231Skmacy goto evict_start; 2058205231Skmacy } 2059168404Spjd } else { 2060168404Spjd missed += 1; 2061168404Spjd } 2062168404Spjd } 2063168404Spjd 2064205231Skmacy mutex_exit(evicted_lock); 2065205231Skmacy mutex_exit(lock); 2066206796Spjd 2067206796Spjd idx = ((idx + 1) & (list_count - 1)); 2068205231Skmacy count++; 2069168404Spjd 2070205231Skmacy if (bytes_evicted < bytes) { 2071205231Skmacy if (count < list_count) 2072205231Skmacy goto evict_start; 2073205231Skmacy else 2074205231Skmacy dprintf("only evicted %lld bytes from %x", 2075205231Skmacy (longlong_t)bytes_evicted, state); 2076205231Skmacy } 2077206796Spjd if (type == ARC_BUFC_METADATA) 2078205231Skmacy evict_metadata_offset = idx; 2079205231Skmacy else 2080205231Skmacy evict_data_offset = idx; 2081206796Spjd 2082168404Spjd if (skipped) 2083168404Spjd ARCSTAT_INCR(arcstat_evict_skip, skipped); 2084168404Spjd 2085168404Spjd if (missed) 2086168404Spjd ARCSTAT_INCR(arcstat_mutex_miss, missed); 2087168404Spjd 2088185029Spjd /* 2089248571Smm * We have just evicted some data into the ghost state, make 2090185029Spjd * sure we also adjust the ghost state size if necessary. 2091185029Spjd */ 2092185029Spjd if (arc_no_grow && 2093185029Spjd arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) { 2094185029Spjd int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size + 2095185029Spjd arc_mru_ghost->arcs_size - arc_c; 2096185029Spjd 2097185029Spjd if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) { 2098185029Spjd int64_t todelete = 2099185029Spjd MIN(arc_mru_ghost->arcs_lsize[type], mru_over); 2100209962Smm arc_evict_ghost(arc_mru_ghost, 0, todelete); 2101185029Spjd } else if (arc_mfu_ghost->arcs_lsize[type] > 0) { 2102185029Spjd int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type], 2103185029Spjd arc_mru_ghost->arcs_size + 2104185029Spjd arc_mfu_ghost->arcs_size - arc_c); 2105209962Smm arc_evict_ghost(arc_mfu_ghost, 0, todelete); 2106185029Spjd } 2107185029Spjd } 2108205231Skmacy if (stolen) 2109205231Skmacy ARCSTAT_BUMP(arcstat_stolen); 2110185029Spjd 2111168404Spjd return (stolen); 2112168404Spjd} 2113168404Spjd 2114168404Spjd/* 2115168404Spjd * Remove buffers from list until we've removed the specified number of 2116168404Spjd * bytes. Destroy the buffers that are removed. 2117168404Spjd */ 2118168404Spjdstatic void 2119209962Smmarc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2120168404Spjd{ 2121168404Spjd arc_buf_hdr_t *ab, *ab_prev; 2122219089Spjd arc_buf_hdr_t marker = { 0 }; 2123205231Skmacy list_t *list, *list_start; 2124205231Skmacy kmutex_t *hash_lock, *lock; 2125168404Spjd uint64_t bytes_deleted = 0; 2126168404Spjd uint64_t bufs_skipped = 0; 2127205231Skmacy static int evict_offset; 2128205231Skmacy int list_count, idx = evict_offset; 2129205231Skmacy int offset, count = 0; 2130168404Spjd 2131168404Spjd ASSERT(GHOST_STATE(state)); 2132205231Skmacy 2133205231Skmacy /* 2134205231Skmacy * data lists come after metadata lists 2135205231Skmacy */ 2136205231Skmacy list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2137205231Skmacy list_count = ARC_BUFC_NUMDATALISTS; 2138205231Skmacy offset = ARC_BUFC_NUMMETADATALISTS; 2139206796Spjd 2140205231Skmacyevict_start: 2141205231Skmacy list = &list_start[idx]; 2142205231Skmacy lock = ARCS_LOCK(state, idx + offset); 2143205231Skmacy 2144205231Skmacy mutex_enter(lock); 2145185029Spjd for (ab = list_tail(list); ab; ab = ab_prev) { 2146185029Spjd ab_prev = list_prev(list, ab); 2147185029Spjd if (spa && ab->b_spa != spa) 2148185029Spjd continue; 2149219089Spjd 2150219089Spjd /* ignore markers */ 2151219089Spjd if (ab->b_spa == 0) 2152219089Spjd continue; 2153219089Spjd 2154168404Spjd hash_lock = HDR_LOCK(ab); 2155219089Spjd /* caller may be trying to modify this buffer, skip it */ 2156219089Spjd if (MUTEX_HELD(hash_lock)) 2157219089Spjd continue; 2158168404Spjd if (mutex_tryenter(hash_lock)) { 2159168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(ab)); 2160168404Spjd ASSERT(ab->b_buf == NULL); 2161168404Spjd ARCSTAT_BUMP(arcstat_deleted); 2162168404Spjd bytes_deleted += ab->b_size; 2163185029Spjd 2164185029Spjd if (ab->b_l2hdr != NULL) { 2165185029Spjd /* 2166185029Spjd * This buffer is cached on the 2nd Level ARC; 2167185029Spjd * don't destroy the header. 2168185029Spjd */ 2169185029Spjd arc_change_state(arc_l2c_only, ab, hash_lock); 2170185029Spjd mutex_exit(hash_lock); 2171185029Spjd } else { 2172185029Spjd arc_change_state(arc_anon, ab, hash_lock); 2173185029Spjd mutex_exit(hash_lock); 2174185029Spjd arc_hdr_destroy(ab); 2175185029Spjd } 2176185029Spjd 2177168404Spjd DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 2178168404Spjd if (bytes >= 0 && bytes_deleted >= bytes) 2179168404Spjd break; 2180219089Spjd } else if (bytes < 0) { 2181219089Spjd /* 2182219089Spjd * Insert a list marker and then wait for the 2183219089Spjd * hash lock to become available. Once its 2184219089Spjd * available, restart from where we left off. 2185219089Spjd */ 2186219089Spjd list_insert_after(list, ab, &marker); 2187219089Spjd mutex_exit(lock); 2188219089Spjd mutex_enter(hash_lock); 2189219089Spjd mutex_exit(hash_lock); 2190219089Spjd mutex_enter(lock); 2191219089Spjd ab_prev = list_prev(list, &marker); 2192219089Spjd list_remove(list, &marker); 2193219089Spjd } else 2194168404Spjd bufs_skipped += 1; 2195168404Spjd } 2196205231Skmacy mutex_exit(lock); 2197206796Spjd idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2198205231Skmacy count++; 2199206796Spjd 2200205231Skmacy if (count < list_count) 2201205231Skmacy goto evict_start; 2202206796Spjd 2203205231Skmacy evict_offset = idx; 2204205231Skmacy if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2205185029Spjd (bytes < 0 || bytes_deleted < bytes)) { 2206205231Skmacy list_start = &state->arcs_lists[0]; 2207205231Skmacy list_count = ARC_BUFC_NUMMETADATALISTS; 2208205231Skmacy offset = count = 0; 2209205231Skmacy goto evict_start; 2210185029Spjd } 2211185029Spjd 2212168404Spjd if (bufs_skipped) { 2213168404Spjd ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2214168404Spjd ASSERT(bytes >= 0); 2215168404Spjd } 2216168404Spjd 2217168404Spjd if (bytes_deleted < bytes) 2218168404Spjd dprintf("only deleted %lld bytes from %p", 2219168404Spjd (longlong_t)bytes_deleted, state); 2220168404Spjd} 2221168404Spjd 2222168404Spjdstatic void 2223168404Spjdarc_adjust(void) 2224168404Spjd{ 2225208373Smm int64_t adjustment, delta; 2226168404Spjd 2227208373Smm /* 2228208373Smm * Adjust MRU size 2229208373Smm */ 2230168404Spjd 2231209275Smm adjustment = MIN((int64_t)(arc_size - arc_c), 2232209275Smm (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2233209275Smm arc_p)); 2234208373Smm 2235208373Smm if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2236208373Smm delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2237209962Smm (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2238208373Smm adjustment -= delta; 2239168404Spjd } 2240168404Spjd 2241208373Smm if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2242208373Smm delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2243209962Smm (void) arc_evict(arc_mru, 0, delta, FALSE, 2244185029Spjd ARC_BUFC_METADATA); 2245185029Spjd } 2246185029Spjd 2247208373Smm /* 2248208373Smm * Adjust MFU size 2249208373Smm */ 2250168404Spjd 2251208373Smm adjustment = arc_size - arc_c; 2252208373Smm 2253208373Smm if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2254208373Smm delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2255209962Smm (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2256208373Smm adjustment -= delta; 2257168404Spjd } 2258168404Spjd 2259208373Smm if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2260208373Smm int64_t delta = MIN(adjustment, 2261208373Smm arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2262209962Smm (void) arc_evict(arc_mfu, 0, delta, FALSE, 2263208373Smm ARC_BUFC_METADATA); 2264208373Smm } 2265168404Spjd 2266208373Smm /* 2267208373Smm * Adjust ghost lists 2268208373Smm */ 2269168404Spjd 2270208373Smm adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2271168404Spjd 2272208373Smm if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2273208373Smm delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2274209962Smm arc_evict_ghost(arc_mru_ghost, 0, delta); 2275208373Smm } 2276185029Spjd 2277208373Smm adjustment = 2278208373Smm arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2279208373Smm 2280208373Smm if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2281208373Smm delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2282209962Smm arc_evict_ghost(arc_mfu_ghost, 0, delta); 2283168404Spjd } 2284168404Spjd} 2285168404Spjd 2286168404Spjdstatic void 2287168404Spjdarc_do_user_evicts(void) 2288168404Spjd{ 2289191903Skmacy static arc_buf_t *tmp_arc_eviction_list; 2290191903Skmacy 2291191903Skmacy /* 2292191903Skmacy * Move list over to avoid LOR 2293191903Skmacy */ 2294206796Spjdrestart: 2295168404Spjd mutex_enter(&arc_eviction_mtx); 2296191903Skmacy tmp_arc_eviction_list = arc_eviction_list; 2297191903Skmacy arc_eviction_list = NULL; 2298191903Skmacy mutex_exit(&arc_eviction_mtx); 2299191903Skmacy 2300191903Skmacy while (tmp_arc_eviction_list != NULL) { 2301191903Skmacy arc_buf_t *buf = tmp_arc_eviction_list; 2302191903Skmacy tmp_arc_eviction_list = buf->b_next; 2303219089Spjd mutex_enter(&buf->b_evict_lock); 2304168404Spjd buf->b_hdr = NULL; 2305219089Spjd mutex_exit(&buf->b_evict_lock); 2306168404Spjd 2307168404Spjd if (buf->b_efunc != NULL) 2308168404Spjd VERIFY(buf->b_efunc(buf) == 0); 2309168404Spjd 2310168404Spjd buf->b_efunc = NULL; 2311168404Spjd buf->b_private = NULL; 2312168404Spjd kmem_cache_free(buf_cache, buf); 2313168404Spjd } 2314191903Skmacy 2315191903Skmacy if (arc_eviction_list != NULL) 2316191903Skmacy goto restart; 2317168404Spjd} 2318168404Spjd 2319168404Spjd/* 2320185029Spjd * Flush all *evictable* data from the cache for the given spa. 2321168404Spjd * NOTE: this will not touch "active" (i.e. referenced) data. 2322168404Spjd */ 2323168404Spjdvoid 2324185029Spjdarc_flush(spa_t *spa) 2325168404Spjd{ 2326209962Smm uint64_t guid = 0; 2327209962Smm 2328209962Smm if (spa) 2329228103Smm guid = spa_load_guid(spa); 2330209962Smm 2331205231Skmacy while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2332209962Smm (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2333185029Spjd if (spa) 2334185029Spjd break; 2335185029Spjd } 2336205231Skmacy while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2337209962Smm (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2338185029Spjd if (spa) 2339185029Spjd break; 2340185029Spjd } 2341205231Skmacy while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2342209962Smm (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2343185029Spjd if (spa) 2344185029Spjd break; 2345185029Spjd } 2346205231Skmacy while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2347209962Smm (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2348185029Spjd if (spa) 2349185029Spjd break; 2350185029Spjd } 2351168404Spjd 2352209962Smm arc_evict_ghost(arc_mru_ghost, guid, -1); 2353209962Smm arc_evict_ghost(arc_mfu_ghost, guid, -1); 2354168404Spjd 2355168404Spjd mutex_enter(&arc_reclaim_thr_lock); 2356168404Spjd arc_do_user_evicts(); 2357168404Spjd mutex_exit(&arc_reclaim_thr_lock); 2358185029Spjd ASSERT(spa || arc_eviction_list == NULL); 2359168404Spjd} 2360168404Spjd 2361168404Spjdvoid 2362168404Spjdarc_shrink(void) 2363168404Spjd{ 2364168404Spjd if (arc_c > arc_c_min) { 2365168404Spjd uint64_t to_free; 2366168404Spjd 2367168404Spjd#ifdef _KERNEL 2368168404Spjd to_free = arc_c >> arc_shrink_shift; 2369168404Spjd#else 2370168404Spjd to_free = arc_c >> arc_shrink_shift; 2371168404Spjd#endif 2372168404Spjd if (arc_c > arc_c_min + to_free) 2373168404Spjd atomic_add_64(&arc_c, -to_free); 2374168404Spjd else 2375168404Spjd arc_c = arc_c_min; 2376168404Spjd 2377168404Spjd atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2378168404Spjd if (arc_c > arc_size) 2379168404Spjd arc_c = MAX(arc_size, arc_c_min); 2380168404Spjd if (arc_p > arc_c) 2381168404Spjd arc_p = (arc_c >> 1); 2382168404Spjd ASSERT(arc_c >= arc_c_min); 2383168404Spjd ASSERT((int64_t)arc_p >= 0); 2384168404Spjd } 2385168404Spjd 2386168404Spjd if (arc_size > arc_c) 2387168404Spjd arc_adjust(); 2388168404Spjd} 2389168404Spjd 2390185029Spjdstatic int needfree = 0; 2391168404Spjd 2392168404Spjdstatic int 2393168404Spjdarc_reclaim_needed(void) 2394168404Spjd{ 2395168404Spjd 2396168404Spjd#ifdef _KERNEL 2397219089Spjd 2398197816Skmacy if (needfree) 2399197816Skmacy return (1); 2400168404Spjd 2401191902Skmacy /* 2402212780Savg * Cooperate with pagedaemon when it's time for it to scan 2403212780Savg * and reclaim some pages. 2404191902Skmacy */ 2405212783Savg if (vm_paging_needed()) 2406191902Skmacy return (1); 2407191902Skmacy 2408219089Spjd#ifdef sun 2409168404Spjd /* 2410185029Spjd * take 'desfree' extra pages, so we reclaim sooner, rather than later 2411185029Spjd */ 2412185029Spjd extra = desfree; 2413185029Spjd 2414185029Spjd /* 2415185029Spjd * check that we're out of range of the pageout scanner. It starts to 2416185029Spjd * schedule paging if freemem is less than lotsfree and needfree. 2417185029Spjd * lotsfree is the high-water mark for pageout, and needfree is the 2418185029Spjd * number of needed free pages. We add extra pages here to make sure 2419185029Spjd * the scanner doesn't start up while we're freeing memory. 2420185029Spjd */ 2421185029Spjd if (freemem < lotsfree + needfree + extra) 2422185029Spjd return (1); 2423185029Spjd 2424185029Spjd /* 2425168404Spjd * check to make sure that swapfs has enough space so that anon 2426185029Spjd * reservations can still succeed. anon_resvmem() checks that the 2427168404Spjd * availrmem is greater than swapfs_minfree, and the number of reserved 2428168404Spjd * swap pages. We also add a bit of extra here just to prevent 2429168404Spjd * circumstances from getting really dire. 2430168404Spjd */ 2431168404Spjd if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2432168404Spjd return (1); 2433168404Spjd 2434168404Spjd#if defined(__i386) 2435168404Spjd /* 2436168404Spjd * If we're on an i386 platform, it's possible that we'll exhaust the 2437168404Spjd * kernel heap space before we ever run out of available physical 2438168404Spjd * memory. Most checks of the size of the heap_area compare against 2439168404Spjd * tune.t_minarmem, which is the minimum available real memory that we 2440168404Spjd * can have in the system. However, this is generally fixed at 25 pages 2441168404Spjd * which is so low that it's useless. In this comparison, we seek to 2442168404Spjd * calculate the total heap-size, and reclaim if more than 3/4ths of the 2443185029Spjd * heap is allocated. (Or, in the calculation, if less than 1/4th is 2444168404Spjd * free) 2445168404Spjd */ 2446168404Spjd if (btop(vmem_size(heap_arena, VMEM_FREE)) < 2447168404Spjd (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2)) 2448168404Spjd return (1); 2449168404Spjd#endif 2450219089Spjd#else /* !sun */ 2451175633Spjd if (kmem_used() > (kmem_size() * 3) / 4) 2452168404Spjd return (1); 2453219089Spjd#endif /* sun */ 2454168404Spjd 2455168404Spjd#else 2456168404Spjd if (spa_get_random(100) == 0) 2457168404Spjd return (1); 2458168404Spjd#endif 2459168404Spjd return (0); 2460168404Spjd} 2461168404Spjd 2462208454Spjdextern kmem_cache_t *zio_buf_cache[]; 2463208454Spjdextern kmem_cache_t *zio_data_buf_cache[]; 2464208454Spjd 2465168404Spjdstatic void 2466168404Spjdarc_kmem_reap_now(arc_reclaim_strategy_t strat) 2467168404Spjd{ 2468168404Spjd size_t i; 2469168404Spjd kmem_cache_t *prev_cache = NULL; 2470168404Spjd kmem_cache_t *prev_data_cache = NULL; 2471168404Spjd 2472168404Spjd#ifdef _KERNEL 2473185029Spjd if (arc_meta_used >= arc_meta_limit) { 2474185029Spjd /* 2475185029Spjd * We are exceeding our meta-data cache limit. 2476185029Spjd * Purge some DNLC entries to release holds on meta-data. 2477185029Spjd */ 2478185029Spjd dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2479185029Spjd } 2480168404Spjd#if defined(__i386) 2481168404Spjd /* 2482168404Spjd * Reclaim unused memory from all kmem caches. 2483168404Spjd */ 2484168404Spjd kmem_reap(); 2485168404Spjd#endif 2486168404Spjd#endif 2487168404Spjd 2488168404Spjd /* 2489185029Spjd * An aggressive reclamation will shrink the cache size as well as 2490168404Spjd * reap free buffers from the arc kmem caches. 2491168404Spjd */ 2492168404Spjd if (strat == ARC_RECLAIM_AGGR) 2493168404Spjd arc_shrink(); 2494168404Spjd 2495168404Spjd for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2496168404Spjd if (zio_buf_cache[i] != prev_cache) { 2497168404Spjd prev_cache = zio_buf_cache[i]; 2498168404Spjd kmem_cache_reap_now(zio_buf_cache[i]); 2499168404Spjd } 2500168404Spjd if (zio_data_buf_cache[i] != prev_data_cache) { 2501168404Spjd prev_data_cache = zio_data_buf_cache[i]; 2502168404Spjd kmem_cache_reap_now(zio_data_buf_cache[i]); 2503168404Spjd } 2504168404Spjd } 2505168404Spjd kmem_cache_reap_now(buf_cache); 2506168404Spjd kmem_cache_reap_now(hdr_cache); 2507168404Spjd} 2508168404Spjd 2509168404Spjdstatic void 2510168404Spjdarc_reclaim_thread(void *dummy __unused) 2511168404Spjd{ 2512168404Spjd clock_t growtime = 0; 2513168404Spjd arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2514168404Spjd callb_cpr_t cpr; 2515168404Spjd 2516168404Spjd CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2517168404Spjd 2518168404Spjd mutex_enter(&arc_reclaim_thr_lock); 2519168404Spjd while (arc_thread_exit == 0) { 2520168404Spjd if (arc_reclaim_needed()) { 2521168404Spjd 2522168404Spjd if (arc_no_grow) { 2523168404Spjd if (last_reclaim == ARC_RECLAIM_CONS) { 2524168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2525168404Spjd } else { 2526168404Spjd last_reclaim = ARC_RECLAIM_CONS; 2527168404Spjd } 2528168404Spjd } else { 2529168404Spjd arc_no_grow = TRUE; 2530168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2531168404Spjd membar_producer(); 2532168404Spjd } 2533168404Spjd 2534168404Spjd /* reset the growth delay for every reclaim */ 2535219089Spjd growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2536168404Spjd 2537185029Spjd if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2538168404Spjd /* 2539185029Spjd * If needfree is TRUE our vm_lowmem hook 2540168404Spjd * was called and in that case we must free some 2541168404Spjd * memory, so switch to aggressive mode. 2542168404Spjd */ 2543168404Spjd arc_no_grow = TRUE; 2544168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2545168404Spjd } 2546168404Spjd arc_kmem_reap_now(last_reclaim); 2547185029Spjd arc_warm = B_TRUE; 2548185029Spjd 2549219089Spjd } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2550168404Spjd arc_no_grow = FALSE; 2551168404Spjd } 2552168404Spjd 2553209275Smm arc_adjust(); 2554168404Spjd 2555168404Spjd if (arc_eviction_list != NULL) 2556168404Spjd arc_do_user_evicts(); 2557168404Spjd 2558211762Savg#ifdef _KERNEL 2559211762Savg if (needfree) { 2560185029Spjd needfree = 0; 2561185029Spjd wakeup(&needfree); 2562211762Savg } 2563168404Spjd#endif 2564168404Spjd 2565168404Spjd /* block until needed, or one second, whichever is shorter */ 2566168404Spjd CALLB_CPR_SAFE_BEGIN(&cpr); 2567168404Spjd (void) cv_timedwait(&arc_reclaim_thr_cv, 2568168404Spjd &arc_reclaim_thr_lock, hz); 2569168404Spjd CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2570168404Spjd } 2571168404Spjd 2572168404Spjd arc_thread_exit = 0; 2573168404Spjd cv_broadcast(&arc_reclaim_thr_cv); 2574168404Spjd CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2575168404Spjd thread_exit(); 2576168404Spjd} 2577168404Spjd 2578168404Spjd/* 2579168404Spjd * Adapt arc info given the number of bytes we are trying to add and 2580168404Spjd * the state that we are comming from. This function is only called 2581168404Spjd * when we are adding new content to the cache. 2582168404Spjd */ 2583168404Spjdstatic void 2584168404Spjdarc_adapt(int bytes, arc_state_t *state) 2585168404Spjd{ 2586168404Spjd int mult; 2587208373Smm uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2588168404Spjd 2589185029Spjd if (state == arc_l2c_only) 2590185029Spjd return; 2591185029Spjd 2592168404Spjd ASSERT(bytes > 0); 2593168404Spjd /* 2594168404Spjd * Adapt the target size of the MRU list: 2595168404Spjd * - if we just hit in the MRU ghost list, then increase 2596168404Spjd * the target size of the MRU list. 2597168404Spjd * - if we just hit in the MFU ghost list, then increase 2598168404Spjd * the target size of the MFU list by decreasing the 2599168404Spjd * target size of the MRU list. 2600168404Spjd */ 2601168404Spjd if (state == arc_mru_ghost) { 2602168404Spjd mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2603168404Spjd 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2604209275Smm mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2605168404Spjd 2606208373Smm arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2607168404Spjd } else if (state == arc_mfu_ghost) { 2608208373Smm uint64_t delta; 2609208373Smm 2610168404Spjd mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2611168404Spjd 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2612209275Smm mult = MIN(mult, 10); 2613168404Spjd 2614208373Smm delta = MIN(bytes * mult, arc_p); 2615208373Smm arc_p = MAX(arc_p_min, arc_p - delta); 2616168404Spjd } 2617168404Spjd ASSERT((int64_t)arc_p >= 0); 2618168404Spjd 2619168404Spjd if (arc_reclaim_needed()) { 2620168404Spjd cv_signal(&arc_reclaim_thr_cv); 2621168404Spjd return; 2622168404Spjd } 2623168404Spjd 2624168404Spjd if (arc_no_grow) 2625168404Spjd return; 2626168404Spjd 2627168404Spjd if (arc_c >= arc_c_max) 2628168404Spjd return; 2629168404Spjd 2630168404Spjd /* 2631168404Spjd * If we're within (2 * maxblocksize) bytes of the target 2632168404Spjd * cache size, increment the target cache size 2633168404Spjd */ 2634168404Spjd if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2635168404Spjd atomic_add_64(&arc_c, (int64_t)bytes); 2636168404Spjd if (arc_c > arc_c_max) 2637168404Spjd arc_c = arc_c_max; 2638168404Spjd else if (state == arc_anon) 2639168404Spjd atomic_add_64(&arc_p, (int64_t)bytes); 2640168404Spjd if (arc_p > arc_c) 2641168404Spjd arc_p = arc_c; 2642168404Spjd } 2643168404Spjd ASSERT((int64_t)arc_p >= 0); 2644168404Spjd} 2645168404Spjd 2646168404Spjd/* 2647168404Spjd * Check if the cache has reached its limits and eviction is required 2648168404Spjd * prior to insert. 2649168404Spjd */ 2650168404Spjdstatic int 2651185029Spjdarc_evict_needed(arc_buf_contents_t type) 2652168404Spjd{ 2653185029Spjd if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2654185029Spjd return (1); 2655185029Spjd 2656219089Spjd#ifdef sun 2657185029Spjd#ifdef _KERNEL 2658185029Spjd /* 2659185029Spjd * If zio data pages are being allocated out of a separate heap segment, 2660185029Spjd * then enforce that the size of available vmem for this area remains 2661185029Spjd * above about 1/32nd free. 2662185029Spjd */ 2663185029Spjd if (type == ARC_BUFC_DATA && zio_arena != NULL && 2664185029Spjd vmem_size(zio_arena, VMEM_FREE) < 2665185029Spjd (vmem_size(zio_arena, VMEM_ALLOC) >> 5)) 2666185029Spjd return (1); 2667185029Spjd#endif 2668219089Spjd#endif /* sun */ 2669185029Spjd 2670168404Spjd if (arc_reclaim_needed()) 2671168404Spjd return (1); 2672168404Spjd 2673168404Spjd return (arc_size > arc_c); 2674168404Spjd} 2675168404Spjd 2676168404Spjd/* 2677168404Spjd * The buffer, supplied as the first argument, needs a data block. 2678168404Spjd * So, if we are at cache max, determine which cache should be victimized. 2679168404Spjd * We have the following cases: 2680168404Spjd * 2681168404Spjd * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2682168404Spjd * In this situation if we're out of space, but the resident size of the MFU is 2683168404Spjd * under the limit, victimize the MFU cache to satisfy this insertion request. 2684168404Spjd * 2685168404Spjd * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2686168404Spjd * Here, we've used up all of the available space for the MRU, so we need to 2687168404Spjd * evict from our own cache instead. Evict from the set of resident MRU 2688168404Spjd * entries. 2689168404Spjd * 2690168404Spjd * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2691168404Spjd * c minus p represents the MFU space in the cache, since p is the size of the 2692168404Spjd * cache that is dedicated to the MRU. In this situation there's still space on 2693168404Spjd * the MFU side, so the MRU side needs to be victimized. 2694168404Spjd * 2695168404Spjd * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2696168404Spjd * MFU's resident set is consuming more space than it has been allotted. In 2697168404Spjd * this situation, we must victimize our own cache, the MFU, for this insertion. 2698168404Spjd */ 2699168404Spjdstatic void 2700168404Spjdarc_get_data_buf(arc_buf_t *buf) 2701168404Spjd{ 2702168404Spjd arc_state_t *state = buf->b_hdr->b_state; 2703168404Spjd uint64_t size = buf->b_hdr->b_size; 2704168404Spjd arc_buf_contents_t type = buf->b_hdr->b_type; 2705168404Spjd 2706168404Spjd arc_adapt(size, state); 2707168404Spjd 2708168404Spjd /* 2709168404Spjd * We have not yet reached cache maximum size, 2710168404Spjd * just allocate a new buffer. 2711168404Spjd */ 2712185029Spjd if (!arc_evict_needed(type)) { 2713168404Spjd if (type == ARC_BUFC_METADATA) { 2714168404Spjd buf->b_data = zio_buf_alloc(size); 2715208373Smm arc_space_consume(size, ARC_SPACE_DATA); 2716168404Spjd } else { 2717168404Spjd ASSERT(type == ARC_BUFC_DATA); 2718168404Spjd buf->b_data = zio_data_buf_alloc(size); 2719208373Smm ARCSTAT_INCR(arcstat_data_size, size); 2720185029Spjd atomic_add_64(&arc_size, size); 2721168404Spjd } 2722168404Spjd goto out; 2723168404Spjd } 2724168404Spjd 2725168404Spjd /* 2726168404Spjd * If we are prefetching from the mfu ghost list, this buffer 2727168404Spjd * will end up on the mru list; so steal space from there. 2728168404Spjd */ 2729168404Spjd if (state == arc_mfu_ghost) 2730168404Spjd state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2731168404Spjd else if (state == arc_mru_ghost) 2732168404Spjd state = arc_mru; 2733168404Spjd 2734168404Spjd if (state == arc_mru || state == arc_anon) { 2735168404Spjd uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2736208373Smm state = (arc_mfu->arcs_lsize[type] >= size && 2737185029Spjd arc_p > mru_used) ? arc_mfu : arc_mru; 2738168404Spjd } else { 2739168404Spjd /* MFU cases */ 2740168404Spjd uint64_t mfu_space = arc_c - arc_p; 2741208373Smm state = (arc_mru->arcs_lsize[type] >= size && 2742185029Spjd mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2743168404Spjd } 2744209962Smm if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2745168404Spjd if (type == ARC_BUFC_METADATA) { 2746168404Spjd buf->b_data = zio_buf_alloc(size); 2747208373Smm arc_space_consume(size, ARC_SPACE_DATA); 2748168404Spjd } else { 2749168404Spjd ASSERT(type == ARC_BUFC_DATA); 2750168404Spjd buf->b_data = zio_data_buf_alloc(size); 2751208373Smm ARCSTAT_INCR(arcstat_data_size, size); 2752185029Spjd atomic_add_64(&arc_size, size); 2753168404Spjd } 2754168404Spjd ARCSTAT_BUMP(arcstat_recycle_miss); 2755168404Spjd } 2756168404Spjd ASSERT(buf->b_data != NULL); 2757168404Spjdout: 2758168404Spjd /* 2759168404Spjd * Update the state size. Note that ghost states have a 2760168404Spjd * "ghost size" and so don't need to be updated. 2761168404Spjd */ 2762168404Spjd if (!GHOST_STATE(buf->b_hdr->b_state)) { 2763168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 2764168404Spjd 2765168404Spjd atomic_add_64(&hdr->b_state->arcs_size, size); 2766168404Spjd if (list_link_active(&hdr->b_arc_node)) { 2767168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2768185029Spjd atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2769168404Spjd } 2770168404Spjd /* 2771168404Spjd * If we are growing the cache, and we are adding anonymous 2772168404Spjd * data, and we have outgrown arc_p, update arc_p 2773168404Spjd */ 2774168404Spjd if (arc_size < arc_c && hdr->b_state == arc_anon && 2775168404Spjd arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2776168404Spjd arc_p = MIN(arc_c, arc_p + size); 2777168404Spjd } 2778205231Skmacy ARCSTAT_BUMP(arcstat_allocated); 2779168404Spjd} 2780168404Spjd 2781168404Spjd/* 2782168404Spjd * This routine is called whenever a buffer is accessed. 2783168404Spjd * NOTE: the hash lock is dropped in this function. 2784168404Spjd */ 2785168404Spjdstatic void 2786168404Spjdarc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2787168404Spjd{ 2788219089Spjd clock_t now; 2789219089Spjd 2790168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 2791168404Spjd 2792168404Spjd if (buf->b_state == arc_anon) { 2793168404Spjd /* 2794168404Spjd * This buffer is not in the cache, and does not 2795168404Spjd * appear in our "ghost" list. Add the new buffer 2796168404Spjd * to the MRU state. 2797168404Spjd */ 2798168404Spjd 2799168404Spjd ASSERT(buf->b_arc_access == 0); 2800219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2801168404Spjd DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2802168404Spjd arc_change_state(arc_mru, buf, hash_lock); 2803168404Spjd 2804168404Spjd } else if (buf->b_state == arc_mru) { 2805219089Spjd now = ddi_get_lbolt(); 2806219089Spjd 2807168404Spjd /* 2808168404Spjd * If this buffer is here because of a prefetch, then either: 2809168404Spjd * - clear the flag if this is a "referencing" read 2810168404Spjd * (any subsequent access will bump this into the MFU state). 2811168404Spjd * or 2812168404Spjd * - move the buffer to the head of the list if this is 2813168404Spjd * another prefetch (to make it less likely to be evicted). 2814168404Spjd */ 2815168404Spjd if ((buf->b_flags & ARC_PREFETCH) != 0) { 2816168404Spjd if (refcount_count(&buf->b_refcnt) == 0) { 2817168404Spjd ASSERT(list_link_active(&buf->b_arc_node)); 2818168404Spjd } else { 2819168404Spjd buf->b_flags &= ~ARC_PREFETCH; 2820168404Spjd ARCSTAT_BUMP(arcstat_mru_hits); 2821168404Spjd } 2822219089Spjd buf->b_arc_access = now; 2823168404Spjd return; 2824168404Spjd } 2825168404Spjd 2826168404Spjd /* 2827168404Spjd * This buffer has been "accessed" only once so far, 2828168404Spjd * but it is still in the cache. Move it to the MFU 2829168404Spjd * state. 2830168404Spjd */ 2831219089Spjd if (now > buf->b_arc_access + ARC_MINTIME) { 2832168404Spjd /* 2833168404Spjd * More than 125ms have passed since we 2834168404Spjd * instantiated this buffer. Move it to the 2835168404Spjd * most frequently used state. 2836168404Spjd */ 2837219089Spjd buf->b_arc_access = now; 2838168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2839168404Spjd arc_change_state(arc_mfu, buf, hash_lock); 2840168404Spjd } 2841168404Spjd ARCSTAT_BUMP(arcstat_mru_hits); 2842168404Spjd } else if (buf->b_state == arc_mru_ghost) { 2843168404Spjd arc_state_t *new_state; 2844168404Spjd /* 2845168404Spjd * This buffer has been "accessed" recently, but 2846168404Spjd * was evicted from the cache. Move it to the 2847168404Spjd * MFU state. 2848168404Spjd */ 2849168404Spjd 2850168404Spjd if (buf->b_flags & ARC_PREFETCH) { 2851168404Spjd new_state = arc_mru; 2852168404Spjd if (refcount_count(&buf->b_refcnt) > 0) 2853168404Spjd buf->b_flags &= ~ARC_PREFETCH; 2854168404Spjd DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2855168404Spjd } else { 2856168404Spjd new_state = arc_mfu; 2857168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2858168404Spjd } 2859168404Spjd 2860219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2861168404Spjd arc_change_state(new_state, buf, hash_lock); 2862168404Spjd 2863168404Spjd ARCSTAT_BUMP(arcstat_mru_ghost_hits); 2864168404Spjd } else if (buf->b_state == arc_mfu) { 2865168404Spjd /* 2866168404Spjd * This buffer has been accessed more than once and is 2867168404Spjd * still in the cache. Keep it in the MFU state. 2868168404Spjd * 2869168404Spjd * NOTE: an add_reference() that occurred when we did 2870168404Spjd * the arc_read() will have kicked this off the list. 2871168404Spjd * If it was a prefetch, we will explicitly move it to 2872168404Spjd * the head of the list now. 2873168404Spjd */ 2874168404Spjd if ((buf->b_flags & ARC_PREFETCH) != 0) { 2875168404Spjd ASSERT(refcount_count(&buf->b_refcnt) == 0); 2876168404Spjd ASSERT(list_link_active(&buf->b_arc_node)); 2877168404Spjd } 2878168404Spjd ARCSTAT_BUMP(arcstat_mfu_hits); 2879219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2880168404Spjd } else if (buf->b_state == arc_mfu_ghost) { 2881168404Spjd arc_state_t *new_state = arc_mfu; 2882168404Spjd /* 2883168404Spjd * This buffer has been accessed more than once but has 2884168404Spjd * been evicted from the cache. Move it back to the 2885168404Spjd * MFU state. 2886168404Spjd */ 2887168404Spjd 2888168404Spjd if (buf->b_flags & ARC_PREFETCH) { 2889168404Spjd /* 2890168404Spjd * This is a prefetch access... 2891168404Spjd * move this block back to the MRU state. 2892168404Spjd */ 2893240415Smm ASSERT0(refcount_count(&buf->b_refcnt)); 2894168404Spjd new_state = arc_mru; 2895168404Spjd } 2896168404Spjd 2897219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2898168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2899168404Spjd arc_change_state(new_state, buf, hash_lock); 2900168404Spjd 2901168404Spjd ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 2902185029Spjd } else if (buf->b_state == arc_l2c_only) { 2903185029Spjd /* 2904185029Spjd * This buffer is on the 2nd Level ARC. 2905185029Spjd */ 2906185029Spjd 2907219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2908185029Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2909185029Spjd arc_change_state(arc_mfu, buf, hash_lock); 2910168404Spjd } else { 2911168404Spjd ASSERT(!"invalid arc state"); 2912168404Spjd } 2913168404Spjd} 2914168404Spjd 2915168404Spjd/* a generic arc_done_func_t which you can use */ 2916168404Spjd/* ARGSUSED */ 2917168404Spjdvoid 2918168404Spjdarc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 2919168404Spjd{ 2920219089Spjd if (zio == NULL || zio->io_error == 0) 2921219089Spjd bcopy(buf->b_data, arg, buf->b_hdr->b_size); 2922248571Smm VERIFY(arc_buf_remove_ref(buf, arg)); 2923168404Spjd} 2924168404Spjd 2925185029Spjd/* a generic arc_done_func_t */ 2926168404Spjdvoid 2927168404Spjdarc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 2928168404Spjd{ 2929168404Spjd arc_buf_t **bufp = arg; 2930168404Spjd if (zio && zio->io_error) { 2931248571Smm VERIFY(arc_buf_remove_ref(buf, arg)); 2932168404Spjd *bufp = NULL; 2933168404Spjd } else { 2934168404Spjd *bufp = buf; 2935219089Spjd ASSERT(buf->b_data); 2936168404Spjd } 2937168404Spjd} 2938168404Spjd 2939168404Spjdstatic void 2940168404Spjdarc_read_done(zio_t *zio) 2941168404Spjd{ 2942168404Spjd arc_buf_hdr_t *hdr, *found; 2943168404Spjd arc_buf_t *buf; 2944168404Spjd arc_buf_t *abuf; /* buffer we're assigning to callback */ 2945168404Spjd kmutex_t *hash_lock; 2946168404Spjd arc_callback_t *callback_list, *acb; 2947168404Spjd int freeable = FALSE; 2948168404Spjd 2949168404Spjd buf = zio->io_private; 2950168404Spjd hdr = buf->b_hdr; 2951168404Spjd 2952168404Spjd /* 2953168404Spjd * The hdr was inserted into hash-table and removed from lists 2954168404Spjd * prior to starting I/O. We should find this header, since 2955168404Spjd * it's in the hash table, and it should be legit since it's 2956168404Spjd * not possible to evict it during the I/O. The only possible 2957168404Spjd * reason for it not to be found is if we were freed during the 2958168404Spjd * read. 2959168404Spjd */ 2960209962Smm found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth, 2961168404Spjd &hash_lock); 2962168404Spjd 2963168404Spjd ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) || 2964185029Spjd (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 2965185029Spjd (found == hdr && HDR_L2_READING(hdr))); 2966168404Spjd 2967185029Spjd hdr->b_flags &= ~ARC_L2_EVICTED; 2968185029Spjd if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 2969185029Spjd hdr->b_flags &= ~ARC_L2CACHE; 2970206796Spjd 2971168404Spjd /* byteswap if necessary */ 2972168404Spjd callback_list = hdr->b_acb; 2973168404Spjd ASSERT(callback_list != NULL); 2974209101Smm if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 2975236884Smm dmu_object_byteswap_t bswap = 2976236884Smm DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 2977185029Spjd arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 2978185029Spjd byteswap_uint64_array : 2979236884Smm dmu_ot_byteswap[bswap].ob_func; 2980185029Spjd func(buf->b_data, hdr->b_size); 2981185029Spjd } 2982168404Spjd 2983185029Spjd arc_cksum_compute(buf, B_FALSE); 2984240133Smm#ifdef illumos 2985240133Smm arc_buf_watch(buf); 2986240133Smm#endif /* illumos */ 2987168404Spjd 2988219089Spjd if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 2989219089Spjd /* 2990219089Spjd * Only call arc_access on anonymous buffers. This is because 2991219089Spjd * if we've issued an I/O for an evicted buffer, we've already 2992219089Spjd * called arc_access (to prevent any simultaneous readers from 2993219089Spjd * getting confused). 2994219089Spjd */ 2995219089Spjd arc_access(hdr, hash_lock); 2996219089Spjd } 2997219089Spjd 2998168404Spjd /* create copies of the data buffer for the callers */ 2999168404Spjd abuf = buf; 3000168404Spjd for (acb = callback_list; acb; acb = acb->acb_next) { 3001168404Spjd if (acb->acb_done) { 3002242845Sdelphij if (abuf == NULL) { 3003242845Sdelphij ARCSTAT_BUMP(arcstat_duplicate_reads); 3004168404Spjd abuf = arc_buf_clone(buf); 3005242845Sdelphij } 3006168404Spjd acb->acb_buf = abuf; 3007168404Spjd abuf = NULL; 3008168404Spjd } 3009168404Spjd } 3010168404Spjd hdr->b_acb = NULL; 3011168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3012168404Spjd ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3013219089Spjd if (abuf == buf) { 3014219089Spjd ASSERT(buf->b_efunc == NULL); 3015219089Spjd ASSERT(hdr->b_datacnt == 1); 3016168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 3017219089Spjd } 3018168404Spjd 3019168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3020168404Spjd 3021168404Spjd if (zio->io_error != 0) { 3022168404Spjd hdr->b_flags |= ARC_IO_ERROR; 3023168404Spjd if (hdr->b_state != arc_anon) 3024168404Spjd arc_change_state(arc_anon, hdr, hash_lock); 3025168404Spjd if (HDR_IN_HASH_TABLE(hdr)) 3026168404Spjd buf_hash_remove(hdr); 3027168404Spjd freeable = refcount_is_zero(&hdr->b_refcnt); 3028168404Spjd } 3029168404Spjd 3030168404Spjd /* 3031168404Spjd * Broadcast before we drop the hash_lock to avoid the possibility 3032168404Spjd * that the hdr (and hence the cv) might be freed before we get to 3033168404Spjd * the cv_broadcast(). 3034168404Spjd */ 3035168404Spjd cv_broadcast(&hdr->b_cv); 3036168404Spjd 3037168404Spjd if (hash_lock) { 3038168404Spjd mutex_exit(hash_lock); 3039168404Spjd } else { 3040168404Spjd /* 3041168404Spjd * This block was freed while we waited for the read to 3042168404Spjd * complete. It has been removed from the hash table and 3043168404Spjd * moved to the anonymous state (so that it won't show up 3044168404Spjd * in the cache). 3045168404Spjd */ 3046168404Spjd ASSERT3P(hdr->b_state, ==, arc_anon); 3047168404Spjd freeable = refcount_is_zero(&hdr->b_refcnt); 3048168404Spjd } 3049168404Spjd 3050168404Spjd /* execute each callback and free its structure */ 3051168404Spjd while ((acb = callback_list) != NULL) { 3052168404Spjd if (acb->acb_done) 3053168404Spjd acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3054168404Spjd 3055168404Spjd if (acb->acb_zio_dummy != NULL) { 3056168404Spjd acb->acb_zio_dummy->io_error = zio->io_error; 3057168404Spjd zio_nowait(acb->acb_zio_dummy); 3058168404Spjd } 3059168404Spjd 3060168404Spjd callback_list = acb->acb_next; 3061168404Spjd kmem_free(acb, sizeof (arc_callback_t)); 3062168404Spjd } 3063168404Spjd 3064168404Spjd if (freeable) 3065168404Spjd arc_hdr_destroy(hdr); 3066168404Spjd} 3067168404Spjd 3068168404Spjd/* 3069168404Spjd * "Read" the block block at the specified DVA (in bp) via the 3070168404Spjd * cache. If the block is found in the cache, invoke the provided 3071168404Spjd * callback immediately and return. Note that the `zio' parameter 3072168404Spjd * in the callback will be NULL in this case, since no IO was 3073168404Spjd * required. If the block is not in the cache pass the read request 3074168404Spjd * on to the spa with a substitute callback function, so that the 3075168404Spjd * requested block will be added to the cache. 3076168404Spjd * 3077168404Spjd * If a read request arrives for a block that has a read in-progress, 3078168404Spjd * either wait for the in-progress read to complete (and return the 3079168404Spjd * results); or, if this is a read with a "done" func, add a record 3080168404Spjd * to the read to invoke the "done" func when the read completes, 3081168404Spjd * and return; or just return. 3082168404Spjd * 3083168404Spjd * arc_read_done() will invoke all the requested "done" functions 3084168404Spjd * for readers of this block. 3085168404Spjd */ 3086168404Spjdint 3087246666Smmarc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3088246666Smm void *private, int priority, int zio_flags, uint32_t *arc_flags, 3089246666Smm const zbookmark_t *zb) 3090168404Spjd{ 3091168404Spjd arc_buf_hdr_t *hdr; 3092247187Smm arc_buf_t *buf = NULL; 3093168404Spjd kmutex_t *hash_lock; 3094185029Spjd zio_t *rzio; 3095228103Smm uint64_t guid = spa_load_guid(spa); 3096168404Spjd 3097168404Spjdtop: 3098219089Spjd hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp), 3099219089Spjd &hash_lock); 3100168404Spjd if (hdr && hdr->b_datacnt > 0) { 3101168404Spjd 3102168404Spjd *arc_flags |= ARC_CACHED; 3103168404Spjd 3104168404Spjd if (HDR_IO_IN_PROGRESS(hdr)) { 3105168404Spjd 3106168404Spjd if (*arc_flags & ARC_WAIT) { 3107168404Spjd cv_wait(&hdr->b_cv, hash_lock); 3108168404Spjd mutex_exit(hash_lock); 3109168404Spjd goto top; 3110168404Spjd } 3111168404Spjd ASSERT(*arc_flags & ARC_NOWAIT); 3112168404Spjd 3113168404Spjd if (done) { 3114168404Spjd arc_callback_t *acb = NULL; 3115168404Spjd 3116168404Spjd acb = kmem_zalloc(sizeof (arc_callback_t), 3117168404Spjd KM_SLEEP); 3118168404Spjd acb->acb_done = done; 3119168404Spjd acb->acb_private = private; 3120168404Spjd if (pio != NULL) 3121168404Spjd acb->acb_zio_dummy = zio_null(pio, 3122209962Smm spa, NULL, NULL, NULL, zio_flags); 3123168404Spjd 3124168404Spjd ASSERT(acb->acb_done != NULL); 3125168404Spjd acb->acb_next = hdr->b_acb; 3126168404Spjd hdr->b_acb = acb; 3127168404Spjd add_reference(hdr, hash_lock, private); 3128168404Spjd mutex_exit(hash_lock); 3129168404Spjd return (0); 3130168404Spjd } 3131168404Spjd mutex_exit(hash_lock); 3132168404Spjd return (0); 3133168404Spjd } 3134168404Spjd 3135168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3136168404Spjd 3137168404Spjd if (done) { 3138168404Spjd add_reference(hdr, hash_lock, private); 3139168404Spjd /* 3140168404Spjd * If this block is already in use, create a new 3141168404Spjd * copy of the data so that we will be guaranteed 3142168404Spjd * that arc_release() will always succeed. 3143168404Spjd */ 3144168404Spjd buf = hdr->b_buf; 3145168404Spjd ASSERT(buf); 3146168404Spjd ASSERT(buf->b_data); 3147168404Spjd if (HDR_BUF_AVAILABLE(hdr)) { 3148168404Spjd ASSERT(buf->b_efunc == NULL); 3149168404Spjd hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3150168404Spjd } else { 3151168404Spjd buf = arc_buf_clone(buf); 3152168404Spjd } 3153219089Spjd 3154168404Spjd } else if (*arc_flags & ARC_PREFETCH && 3155168404Spjd refcount_count(&hdr->b_refcnt) == 0) { 3156168404Spjd hdr->b_flags |= ARC_PREFETCH; 3157168404Spjd } 3158168404Spjd DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3159168404Spjd arc_access(hdr, hash_lock); 3160185029Spjd if (*arc_flags & ARC_L2CACHE) 3161185029Spjd hdr->b_flags |= ARC_L2CACHE; 3162251478Sdelphij if (*arc_flags & ARC_L2COMPRESS) 3163251478Sdelphij hdr->b_flags |= ARC_L2COMPRESS; 3164168404Spjd mutex_exit(hash_lock); 3165168404Spjd ARCSTAT_BUMP(arcstat_hits); 3166168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3167168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3168168404Spjd data, metadata, hits); 3169168404Spjd 3170168404Spjd if (done) 3171168404Spjd done(NULL, buf, private); 3172168404Spjd } else { 3173168404Spjd uint64_t size = BP_GET_LSIZE(bp); 3174168404Spjd arc_callback_t *acb; 3175185029Spjd vdev_t *vd = NULL; 3176247187Smm uint64_t addr = 0; 3177208373Smm boolean_t devw = B_FALSE; 3178168404Spjd 3179168404Spjd if (hdr == NULL) { 3180168404Spjd /* this block is not in the cache */ 3181168404Spjd arc_buf_hdr_t *exists; 3182168404Spjd arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3183168404Spjd buf = arc_buf_alloc(spa, size, private, type); 3184168404Spjd hdr = buf->b_hdr; 3185168404Spjd hdr->b_dva = *BP_IDENTITY(bp); 3186219089Spjd hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3187168404Spjd hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3188168404Spjd exists = buf_hash_insert(hdr, &hash_lock); 3189168404Spjd if (exists) { 3190168404Spjd /* somebody beat us to the hash insert */ 3191168404Spjd mutex_exit(hash_lock); 3192219089Spjd buf_discard_identity(hdr); 3193168404Spjd (void) arc_buf_remove_ref(buf, private); 3194168404Spjd goto top; /* restart the IO request */ 3195168404Spjd } 3196168404Spjd /* if this is a prefetch, we don't have a reference */ 3197168404Spjd if (*arc_flags & ARC_PREFETCH) { 3198168404Spjd (void) remove_reference(hdr, hash_lock, 3199168404Spjd private); 3200168404Spjd hdr->b_flags |= ARC_PREFETCH; 3201168404Spjd } 3202185029Spjd if (*arc_flags & ARC_L2CACHE) 3203185029Spjd hdr->b_flags |= ARC_L2CACHE; 3204251478Sdelphij if (*arc_flags & ARC_L2COMPRESS) 3205251478Sdelphij hdr->b_flags |= ARC_L2COMPRESS; 3206168404Spjd if (BP_GET_LEVEL(bp) > 0) 3207168404Spjd hdr->b_flags |= ARC_INDIRECT; 3208168404Spjd } else { 3209168404Spjd /* this block is in the ghost cache */ 3210168404Spjd ASSERT(GHOST_STATE(hdr->b_state)); 3211168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3212240415Smm ASSERT0(refcount_count(&hdr->b_refcnt)); 3213168404Spjd ASSERT(hdr->b_buf == NULL); 3214168404Spjd 3215168404Spjd /* if this is a prefetch, we don't have a reference */ 3216168404Spjd if (*arc_flags & ARC_PREFETCH) 3217168404Spjd hdr->b_flags |= ARC_PREFETCH; 3218168404Spjd else 3219168404Spjd add_reference(hdr, hash_lock, private); 3220185029Spjd if (*arc_flags & ARC_L2CACHE) 3221185029Spjd hdr->b_flags |= ARC_L2CACHE; 3222251478Sdelphij if (*arc_flags & ARC_L2COMPRESS) 3223251478Sdelphij hdr->b_flags |= ARC_L2COMPRESS; 3224185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3225168404Spjd buf->b_hdr = hdr; 3226168404Spjd buf->b_data = NULL; 3227168404Spjd buf->b_efunc = NULL; 3228168404Spjd buf->b_private = NULL; 3229168404Spjd buf->b_next = NULL; 3230168404Spjd hdr->b_buf = buf; 3231168404Spjd ASSERT(hdr->b_datacnt == 0); 3232168404Spjd hdr->b_datacnt = 1; 3233219089Spjd arc_get_data_buf(buf); 3234219089Spjd arc_access(hdr, hash_lock); 3235168404Spjd } 3236168404Spjd 3237219089Spjd ASSERT(!GHOST_STATE(hdr->b_state)); 3238219089Spjd 3239168404Spjd acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3240168404Spjd acb->acb_done = done; 3241168404Spjd acb->acb_private = private; 3242168404Spjd 3243168404Spjd ASSERT(hdr->b_acb == NULL); 3244168404Spjd hdr->b_acb = acb; 3245168404Spjd hdr->b_flags |= ARC_IO_IN_PROGRESS; 3246168404Spjd 3247185029Spjd if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL && 3248185029Spjd (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3249208373Smm devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3250185029Spjd addr = hdr->b_l2hdr->b_daddr; 3251185029Spjd /* 3252185029Spjd * Lock out device removal. 3253185029Spjd */ 3254185029Spjd if (vdev_is_dead(vd) || 3255185029Spjd !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3256185029Spjd vd = NULL; 3257185029Spjd } 3258185029Spjd 3259168404Spjd mutex_exit(hash_lock); 3260168404Spjd 3261251629Sdelphij /* 3262251629Sdelphij * At this point, we have a level 1 cache miss. Try again in 3263251629Sdelphij * L2ARC if possible. 3264251629Sdelphij */ 3265168404Spjd ASSERT3U(hdr->b_size, ==, size); 3266219089Spjd DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3267219089Spjd uint64_t, size, zbookmark_t *, zb); 3268168404Spjd ARCSTAT_BUMP(arcstat_misses); 3269168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3270168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3271168404Spjd data, metadata, misses); 3272228392Spjd#ifdef _KERNEL 3273228392Spjd curthread->td_ru.ru_inblock++; 3274228392Spjd#endif 3275168404Spjd 3276208373Smm if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3277185029Spjd /* 3278185029Spjd * Read from the L2ARC if the following are true: 3279185029Spjd * 1. The L2ARC vdev was previously cached. 3280185029Spjd * 2. This buffer still has L2ARC metadata. 3281185029Spjd * 3. This buffer isn't currently writing to the L2ARC. 3282185029Spjd * 4. The L2ARC entry wasn't evicted, which may 3283185029Spjd * also have invalidated the vdev. 3284208373Smm * 5. This isn't prefetch and l2arc_noprefetch is set. 3285185029Spjd */ 3286185029Spjd if (hdr->b_l2hdr != NULL && 3287208373Smm !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3288208373Smm !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3289185029Spjd l2arc_read_callback_t *cb; 3290185029Spjd 3291185029Spjd DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3292185029Spjd ARCSTAT_BUMP(arcstat_l2_hits); 3293185029Spjd 3294185029Spjd cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3295185029Spjd KM_SLEEP); 3296185029Spjd cb->l2rcb_buf = buf; 3297185029Spjd cb->l2rcb_spa = spa; 3298185029Spjd cb->l2rcb_bp = *bp; 3299185029Spjd cb->l2rcb_zb = *zb; 3300185029Spjd cb->l2rcb_flags = zio_flags; 3301251478Sdelphij cb->l2rcb_compress = hdr->b_l2hdr->b_compress; 3302185029Spjd 3303247187Smm ASSERT(addr >= VDEV_LABEL_START_SIZE && 3304247187Smm addr + size < vd->vdev_psize - 3305247187Smm VDEV_LABEL_END_SIZE); 3306247187Smm 3307185029Spjd /* 3308185029Spjd * l2arc read. The SCL_L2ARC lock will be 3309185029Spjd * released by l2arc_read_done(). 3310251478Sdelphij * Issue a null zio if the underlying buffer 3311251478Sdelphij * was squashed to zero size by compression. 3312185029Spjd */ 3313251478Sdelphij if (hdr->b_l2hdr->b_compress == 3314251478Sdelphij ZIO_COMPRESS_EMPTY) { 3315251478Sdelphij rzio = zio_null(pio, spa, vd, 3316251478Sdelphij l2arc_read_done, cb, 3317251478Sdelphij zio_flags | ZIO_FLAG_DONT_CACHE | 3318251478Sdelphij ZIO_FLAG_CANFAIL | 3319251478Sdelphij ZIO_FLAG_DONT_PROPAGATE | 3320251478Sdelphij ZIO_FLAG_DONT_RETRY); 3321251478Sdelphij } else { 3322251478Sdelphij rzio = zio_read_phys(pio, vd, addr, 3323251478Sdelphij hdr->b_l2hdr->b_asize, 3324251478Sdelphij buf->b_data, ZIO_CHECKSUM_OFF, 3325251478Sdelphij l2arc_read_done, cb, priority, 3326251478Sdelphij zio_flags | ZIO_FLAG_DONT_CACHE | 3327251478Sdelphij ZIO_FLAG_CANFAIL | 3328251478Sdelphij ZIO_FLAG_DONT_PROPAGATE | 3329251478Sdelphij ZIO_FLAG_DONT_RETRY, B_FALSE); 3330251478Sdelphij } 3331185029Spjd DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3332185029Spjd zio_t *, rzio); 3333251478Sdelphij ARCSTAT_INCR(arcstat_l2_read_bytes, 3334251478Sdelphij hdr->b_l2hdr->b_asize); 3335185029Spjd 3336185029Spjd if (*arc_flags & ARC_NOWAIT) { 3337185029Spjd zio_nowait(rzio); 3338185029Spjd return (0); 3339185029Spjd } 3340185029Spjd 3341185029Spjd ASSERT(*arc_flags & ARC_WAIT); 3342185029Spjd if (zio_wait(rzio) == 0) 3343185029Spjd return (0); 3344185029Spjd 3345185029Spjd /* l2arc read error; goto zio_read() */ 3346185029Spjd } else { 3347185029Spjd DTRACE_PROBE1(l2arc__miss, 3348185029Spjd arc_buf_hdr_t *, hdr); 3349185029Spjd ARCSTAT_BUMP(arcstat_l2_misses); 3350185029Spjd if (HDR_L2_WRITING(hdr)) 3351185029Spjd ARCSTAT_BUMP(arcstat_l2_rw_clash); 3352185029Spjd spa_config_exit(spa, SCL_L2ARC, vd); 3353185029Spjd } 3354208373Smm } else { 3355208373Smm if (vd != NULL) 3356208373Smm spa_config_exit(spa, SCL_L2ARC, vd); 3357208373Smm if (l2arc_ndev != 0) { 3358208373Smm DTRACE_PROBE1(l2arc__miss, 3359208373Smm arc_buf_hdr_t *, hdr); 3360208373Smm ARCSTAT_BUMP(arcstat_l2_misses); 3361208373Smm } 3362185029Spjd } 3363185029Spjd 3364168404Spjd rzio = zio_read(pio, spa, bp, buf->b_data, size, 3365185029Spjd arc_read_done, buf, priority, zio_flags, zb); 3366168404Spjd 3367168404Spjd if (*arc_flags & ARC_WAIT) 3368168404Spjd return (zio_wait(rzio)); 3369168404Spjd 3370168404Spjd ASSERT(*arc_flags & ARC_NOWAIT); 3371168404Spjd zio_nowait(rzio); 3372168404Spjd } 3373168404Spjd return (0); 3374168404Spjd} 3375168404Spjd 3376168404Spjdvoid 3377168404Spjdarc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3378168404Spjd{ 3379168404Spjd ASSERT(buf->b_hdr != NULL); 3380168404Spjd ASSERT(buf->b_hdr->b_state != arc_anon); 3381168404Spjd ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3382219089Spjd ASSERT(buf->b_efunc == NULL); 3383219089Spjd ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3384219089Spjd 3385168404Spjd buf->b_efunc = func; 3386168404Spjd buf->b_private = private; 3387168404Spjd} 3388168404Spjd 3389168404Spjd/* 3390251520Sdelphij * Notify the arc that a block was freed, and thus will never be used again. 3391251520Sdelphij */ 3392251520Sdelphijvoid 3393251520Sdelphijarc_freed(spa_t *spa, const blkptr_t *bp) 3394251520Sdelphij{ 3395251520Sdelphij arc_buf_hdr_t *hdr; 3396251520Sdelphij kmutex_t *hash_lock; 3397251520Sdelphij uint64_t guid = spa_load_guid(spa); 3398251520Sdelphij 3399251520Sdelphij hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp), 3400251520Sdelphij &hash_lock); 3401251520Sdelphij if (hdr == NULL) 3402251520Sdelphij return; 3403251520Sdelphij if (HDR_BUF_AVAILABLE(hdr)) { 3404251520Sdelphij arc_buf_t *buf = hdr->b_buf; 3405251520Sdelphij add_reference(hdr, hash_lock, FTAG); 3406251520Sdelphij hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3407251520Sdelphij mutex_exit(hash_lock); 3408251520Sdelphij 3409251520Sdelphij arc_release(buf, FTAG); 3410251520Sdelphij (void) arc_buf_remove_ref(buf, FTAG); 3411251520Sdelphij } else { 3412251520Sdelphij mutex_exit(hash_lock); 3413251520Sdelphij } 3414251520Sdelphij 3415251520Sdelphij} 3416251520Sdelphij 3417251520Sdelphij/* 3418168404Spjd * This is used by the DMU to let the ARC know that a buffer is 3419168404Spjd * being evicted, so the ARC should clean up. If this arc buf 3420168404Spjd * is not yet in the evicted state, it will be put there. 3421168404Spjd */ 3422168404Spjdint 3423168404Spjdarc_buf_evict(arc_buf_t *buf) 3424168404Spjd{ 3425168404Spjd arc_buf_hdr_t *hdr; 3426168404Spjd kmutex_t *hash_lock; 3427168404Spjd arc_buf_t **bufp; 3428205231Skmacy list_t *list, *evicted_list; 3429205231Skmacy kmutex_t *lock, *evicted_lock; 3430206796Spjd 3431219089Spjd mutex_enter(&buf->b_evict_lock); 3432168404Spjd hdr = buf->b_hdr; 3433168404Spjd if (hdr == NULL) { 3434168404Spjd /* 3435168404Spjd * We are in arc_do_user_evicts(). 3436168404Spjd */ 3437168404Spjd ASSERT(buf->b_data == NULL); 3438219089Spjd mutex_exit(&buf->b_evict_lock); 3439168404Spjd return (0); 3440185029Spjd } else if (buf->b_data == NULL) { 3441185029Spjd arc_buf_t copy = *buf; /* structure assignment */ 3442185029Spjd /* 3443185029Spjd * We are on the eviction list; process this buffer now 3444185029Spjd * but let arc_do_user_evicts() do the reaping. 3445185029Spjd */ 3446185029Spjd buf->b_efunc = NULL; 3447219089Spjd mutex_exit(&buf->b_evict_lock); 3448185029Spjd VERIFY(copy.b_efunc(©) == 0); 3449185029Spjd return (1); 3450168404Spjd } 3451168404Spjd hash_lock = HDR_LOCK(hdr); 3452168404Spjd mutex_enter(hash_lock); 3453219089Spjd hdr = buf->b_hdr; 3454219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3455168404Spjd 3456168404Spjd ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3457168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3458168404Spjd 3459168404Spjd /* 3460168404Spjd * Pull this buffer off of the hdr 3461168404Spjd */ 3462168404Spjd bufp = &hdr->b_buf; 3463168404Spjd while (*bufp != buf) 3464168404Spjd bufp = &(*bufp)->b_next; 3465168404Spjd *bufp = buf->b_next; 3466168404Spjd 3467168404Spjd ASSERT(buf->b_data != NULL); 3468168404Spjd arc_buf_destroy(buf, FALSE, FALSE); 3469168404Spjd 3470168404Spjd if (hdr->b_datacnt == 0) { 3471168404Spjd arc_state_t *old_state = hdr->b_state; 3472168404Spjd arc_state_t *evicted_state; 3473168404Spjd 3474219089Spjd ASSERT(hdr->b_buf == NULL); 3475168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 3476168404Spjd 3477168404Spjd evicted_state = 3478168404Spjd (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3479168404Spjd 3480205231Skmacy get_buf_info(hdr, old_state, &list, &lock); 3481205231Skmacy get_buf_info(hdr, evicted_state, &evicted_list, &evicted_lock); 3482205231Skmacy mutex_enter(lock); 3483205231Skmacy mutex_enter(evicted_lock); 3484168404Spjd 3485168404Spjd arc_change_state(evicted_state, hdr, hash_lock); 3486168404Spjd ASSERT(HDR_IN_HASH_TABLE(hdr)); 3487185029Spjd hdr->b_flags |= ARC_IN_HASH_TABLE; 3488185029Spjd hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3489168404Spjd 3490205231Skmacy mutex_exit(evicted_lock); 3491205231Skmacy mutex_exit(lock); 3492168404Spjd } 3493168404Spjd mutex_exit(hash_lock); 3494219089Spjd mutex_exit(&buf->b_evict_lock); 3495168404Spjd 3496168404Spjd VERIFY(buf->b_efunc(buf) == 0); 3497168404Spjd buf->b_efunc = NULL; 3498168404Spjd buf->b_private = NULL; 3499168404Spjd buf->b_hdr = NULL; 3500219089Spjd buf->b_next = NULL; 3501168404Spjd kmem_cache_free(buf_cache, buf); 3502168404Spjd return (1); 3503168404Spjd} 3504168404Spjd 3505168404Spjd/* 3506251629Sdelphij * Release this buffer from the cache, making it an anonymous buffer. This 3507251629Sdelphij * must be done after a read and prior to modifying the buffer contents. 3508168404Spjd * If the buffer has more than one reference, we must make 3509185029Spjd * a new hdr for the buffer. 3510168404Spjd */ 3511168404Spjdvoid 3512168404Spjdarc_release(arc_buf_t *buf, void *tag) 3513168404Spjd{ 3514185029Spjd arc_buf_hdr_t *hdr; 3515219089Spjd kmutex_t *hash_lock = NULL; 3516185029Spjd l2arc_buf_hdr_t *l2hdr; 3517185029Spjd uint64_t buf_size; 3518168404Spjd 3519219089Spjd /* 3520219089Spjd * It would be nice to assert that if it's DMU metadata (level > 3521219089Spjd * 0 || it's the dnode file), then it must be syncing context. 3522219089Spjd * But we don't know that information at this level. 3523219089Spjd */ 3524219089Spjd 3525219089Spjd mutex_enter(&buf->b_evict_lock); 3526185029Spjd hdr = buf->b_hdr; 3527185029Spjd 3528168404Spjd /* this buffer is not on any list */ 3529168404Spjd ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3530168404Spjd 3531168404Spjd if (hdr->b_state == arc_anon) { 3532168404Spjd /* this buffer is already released */ 3533168404Spjd ASSERT(buf->b_efunc == NULL); 3534208373Smm } else { 3535208373Smm hash_lock = HDR_LOCK(hdr); 3536208373Smm mutex_enter(hash_lock); 3537219089Spjd hdr = buf->b_hdr; 3538219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3539168404Spjd } 3540168404Spjd 3541185029Spjd l2hdr = hdr->b_l2hdr; 3542185029Spjd if (l2hdr) { 3543185029Spjd mutex_enter(&l2arc_buflist_mtx); 3544185029Spjd hdr->b_l2hdr = NULL; 3545185029Spjd } 3546247187Smm buf_size = hdr->b_size; 3547185029Spjd 3548168404Spjd /* 3549168404Spjd * Do we have more than one buf? 3550168404Spjd */ 3551185029Spjd if (hdr->b_datacnt > 1) { 3552168404Spjd arc_buf_hdr_t *nhdr; 3553168404Spjd arc_buf_t **bufp; 3554168404Spjd uint64_t blksz = hdr->b_size; 3555209962Smm uint64_t spa = hdr->b_spa; 3556168404Spjd arc_buf_contents_t type = hdr->b_type; 3557185029Spjd uint32_t flags = hdr->b_flags; 3558168404Spjd 3559185029Spjd ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3560168404Spjd /* 3561219089Spjd * Pull the data off of this hdr and attach it to 3562219089Spjd * a new anonymous hdr. 3563168404Spjd */ 3564168404Spjd (void) remove_reference(hdr, hash_lock, tag); 3565168404Spjd bufp = &hdr->b_buf; 3566168404Spjd while (*bufp != buf) 3567168404Spjd bufp = &(*bufp)->b_next; 3568219089Spjd *bufp = buf->b_next; 3569168404Spjd buf->b_next = NULL; 3570168404Spjd 3571168404Spjd ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3572168404Spjd atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3573168404Spjd if (refcount_is_zero(&hdr->b_refcnt)) { 3574185029Spjd uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3575185029Spjd ASSERT3U(*size, >=, hdr->b_size); 3576185029Spjd atomic_add_64(size, -hdr->b_size); 3577168404Spjd } 3578242845Sdelphij 3579242845Sdelphij /* 3580242845Sdelphij * We're releasing a duplicate user data buffer, update 3581242845Sdelphij * our statistics accordingly. 3582242845Sdelphij */ 3583242845Sdelphij if (hdr->b_type == ARC_BUFC_DATA) { 3584242845Sdelphij ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3585242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3586242845Sdelphij -hdr->b_size); 3587242845Sdelphij } 3588168404Spjd hdr->b_datacnt -= 1; 3589168404Spjd arc_cksum_verify(buf); 3590240133Smm#ifdef illumos 3591240133Smm arc_buf_unwatch(buf); 3592240133Smm#endif /* illumos */ 3593168404Spjd 3594168404Spjd mutex_exit(hash_lock); 3595168404Spjd 3596185029Spjd nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3597168404Spjd nhdr->b_size = blksz; 3598168404Spjd nhdr->b_spa = spa; 3599168404Spjd nhdr->b_type = type; 3600168404Spjd nhdr->b_buf = buf; 3601168404Spjd nhdr->b_state = arc_anon; 3602168404Spjd nhdr->b_arc_access = 0; 3603185029Spjd nhdr->b_flags = flags & ARC_L2_WRITING; 3604185029Spjd nhdr->b_l2hdr = NULL; 3605168404Spjd nhdr->b_datacnt = 1; 3606168404Spjd nhdr->b_freeze_cksum = NULL; 3607168404Spjd (void) refcount_add(&nhdr->b_refcnt, tag); 3608168404Spjd buf->b_hdr = nhdr; 3609219089Spjd mutex_exit(&buf->b_evict_lock); 3610168404Spjd atomic_add_64(&arc_anon->arcs_size, blksz); 3611168404Spjd } else { 3612219089Spjd mutex_exit(&buf->b_evict_lock); 3613168404Spjd ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3614168404Spjd ASSERT(!list_link_active(&hdr->b_arc_node)); 3615168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3616219089Spjd if (hdr->b_state != arc_anon) 3617219089Spjd arc_change_state(arc_anon, hdr, hash_lock); 3618168404Spjd hdr->b_arc_access = 0; 3619219089Spjd if (hash_lock) 3620219089Spjd mutex_exit(hash_lock); 3621185029Spjd 3622219089Spjd buf_discard_identity(hdr); 3623168404Spjd arc_buf_thaw(buf); 3624168404Spjd } 3625168404Spjd buf->b_efunc = NULL; 3626168404Spjd buf->b_private = NULL; 3627185029Spjd 3628185029Spjd if (l2hdr) { 3629251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3630248572Ssmh trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3631248574Ssmh hdr->b_size, 0); 3632185029Spjd list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3633185029Spjd kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3634185029Spjd ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3635185029Spjd mutex_exit(&l2arc_buflist_mtx); 3636185029Spjd } 3637168404Spjd} 3638168404Spjd 3639168404Spjdint 3640168404Spjdarc_released(arc_buf_t *buf) 3641168404Spjd{ 3642185029Spjd int released; 3643185029Spjd 3644219089Spjd mutex_enter(&buf->b_evict_lock); 3645185029Spjd released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3646219089Spjd mutex_exit(&buf->b_evict_lock); 3647185029Spjd return (released); 3648168404Spjd} 3649168404Spjd 3650168404Spjdint 3651168404Spjdarc_has_callback(arc_buf_t *buf) 3652168404Spjd{ 3653185029Spjd int callback; 3654185029Spjd 3655219089Spjd mutex_enter(&buf->b_evict_lock); 3656185029Spjd callback = (buf->b_efunc != NULL); 3657219089Spjd mutex_exit(&buf->b_evict_lock); 3658185029Spjd return (callback); 3659168404Spjd} 3660168404Spjd 3661168404Spjd#ifdef ZFS_DEBUG 3662168404Spjdint 3663168404Spjdarc_referenced(arc_buf_t *buf) 3664168404Spjd{ 3665185029Spjd int referenced; 3666185029Spjd 3667219089Spjd mutex_enter(&buf->b_evict_lock); 3668185029Spjd referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3669219089Spjd mutex_exit(&buf->b_evict_lock); 3670185029Spjd return (referenced); 3671168404Spjd} 3672168404Spjd#endif 3673168404Spjd 3674168404Spjdstatic void 3675168404Spjdarc_write_ready(zio_t *zio) 3676168404Spjd{ 3677168404Spjd arc_write_callback_t *callback = zio->io_private; 3678168404Spjd arc_buf_t *buf = callback->awcb_buf; 3679185029Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3680168404Spjd 3681185029Spjd ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3682185029Spjd callback->awcb_ready(zio, buf, callback->awcb_private); 3683185029Spjd 3684185029Spjd /* 3685185029Spjd * If the IO is already in progress, then this is a re-write 3686185029Spjd * attempt, so we need to thaw and re-compute the cksum. 3687185029Spjd * It is the responsibility of the callback to handle the 3688185029Spjd * accounting for any re-write attempt. 3689185029Spjd */ 3690185029Spjd if (HDR_IO_IN_PROGRESS(hdr)) { 3691185029Spjd mutex_enter(&hdr->b_freeze_lock); 3692185029Spjd if (hdr->b_freeze_cksum != NULL) { 3693185029Spjd kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3694185029Spjd hdr->b_freeze_cksum = NULL; 3695185029Spjd } 3696185029Spjd mutex_exit(&hdr->b_freeze_lock); 3697168404Spjd } 3698185029Spjd arc_cksum_compute(buf, B_FALSE); 3699185029Spjd hdr->b_flags |= ARC_IO_IN_PROGRESS; 3700168404Spjd} 3701168404Spjd 3702168404Spjdstatic void 3703168404Spjdarc_write_done(zio_t *zio) 3704168404Spjd{ 3705168404Spjd arc_write_callback_t *callback = zio->io_private; 3706168404Spjd arc_buf_t *buf = callback->awcb_buf; 3707168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3708168404Spjd 3709219089Spjd ASSERT(hdr->b_acb == NULL); 3710168404Spjd 3711219089Spjd if (zio->io_error == 0) { 3712219089Spjd hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3713219089Spjd hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3714219089Spjd hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3715219089Spjd } else { 3716219089Spjd ASSERT(BUF_EMPTY(hdr)); 3717219089Spjd } 3718219089Spjd 3719168404Spjd /* 3720168404Spjd * If the block to be written was all-zero, we may have 3721168404Spjd * compressed it away. In this case no write was performed 3722219089Spjd * so there will be no dva/birth/checksum. The buffer must 3723219089Spjd * therefore remain anonymous (and uncached). 3724168404Spjd */ 3725168404Spjd if (!BUF_EMPTY(hdr)) { 3726168404Spjd arc_buf_hdr_t *exists; 3727168404Spjd kmutex_t *hash_lock; 3728168404Spjd 3729219089Spjd ASSERT(zio->io_error == 0); 3730219089Spjd 3731168404Spjd arc_cksum_verify(buf); 3732168404Spjd 3733168404Spjd exists = buf_hash_insert(hdr, &hash_lock); 3734168404Spjd if (exists) { 3735168404Spjd /* 3736168404Spjd * This can only happen if we overwrite for 3737168404Spjd * sync-to-convergence, because we remove 3738168404Spjd * buffers from the hash table when we arc_free(). 3739168404Spjd */ 3740219089Spjd if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3741219089Spjd if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3742219089Spjd panic("bad overwrite, hdr=%p exists=%p", 3743219089Spjd (void *)hdr, (void *)exists); 3744219089Spjd ASSERT(refcount_is_zero(&exists->b_refcnt)); 3745219089Spjd arc_change_state(arc_anon, exists, hash_lock); 3746219089Spjd mutex_exit(hash_lock); 3747219089Spjd arc_hdr_destroy(exists); 3748219089Spjd exists = buf_hash_insert(hdr, &hash_lock); 3749219089Spjd ASSERT3P(exists, ==, NULL); 3750243524Smm } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3751243524Smm /* nopwrite */ 3752243524Smm ASSERT(zio->io_prop.zp_nopwrite); 3753243524Smm if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3754243524Smm panic("bad nopwrite, hdr=%p exists=%p", 3755243524Smm (void *)hdr, (void *)exists); 3756219089Spjd } else { 3757219089Spjd /* Dedup */ 3758219089Spjd ASSERT(hdr->b_datacnt == 1); 3759219089Spjd ASSERT(hdr->b_state == arc_anon); 3760219089Spjd ASSERT(BP_GET_DEDUP(zio->io_bp)); 3761219089Spjd ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3762219089Spjd } 3763168404Spjd } 3764168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3765185029Spjd /* if it's not anon, we are doing a scrub */ 3766219089Spjd if (!exists && hdr->b_state == arc_anon) 3767185029Spjd arc_access(hdr, hash_lock); 3768168404Spjd mutex_exit(hash_lock); 3769168404Spjd } else { 3770168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3771168404Spjd } 3772168404Spjd 3773219089Spjd ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3774219089Spjd callback->awcb_done(zio, buf, callback->awcb_private); 3775168404Spjd 3776168404Spjd kmem_free(callback, sizeof (arc_write_callback_t)); 3777168404Spjd} 3778168404Spjd 3779168404Spjdzio_t * 3780219089Spjdarc_write(zio_t *pio, spa_t *spa, uint64_t txg, 3781251478Sdelphij blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 3782251478Sdelphij const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *done, 3783251478Sdelphij void *private, int priority, int zio_flags, const zbookmark_t *zb) 3784168404Spjd{ 3785168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3786168404Spjd arc_write_callback_t *callback; 3787185029Spjd zio_t *zio; 3788168404Spjd 3789185029Spjd ASSERT(ready != NULL); 3790219089Spjd ASSERT(done != NULL); 3791168404Spjd ASSERT(!HDR_IO_ERROR(hdr)); 3792168404Spjd ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3793219089Spjd ASSERT(hdr->b_acb == NULL); 3794185029Spjd if (l2arc) 3795185029Spjd hdr->b_flags |= ARC_L2CACHE; 3796251478Sdelphij if (l2arc_compress) 3797251478Sdelphij hdr->b_flags |= ARC_L2COMPRESS; 3798168404Spjd callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 3799168404Spjd callback->awcb_ready = ready; 3800168404Spjd callback->awcb_done = done; 3801168404Spjd callback->awcb_private = private; 3802168404Spjd callback->awcb_buf = buf; 3803168404Spjd 3804219089Spjd zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 3805185029Spjd arc_write_ready, arc_write_done, callback, priority, zio_flags, zb); 3806185029Spjd 3807168404Spjd return (zio); 3808168404Spjd} 3809168404Spjd 3810185029Spjdstatic int 3811209962Smmarc_memory_throttle(uint64_t reserve, uint64_t inflight_data, uint64_t txg) 3812185029Spjd{ 3813185029Spjd#ifdef _KERNEL 3814219089Spjd uint64_t available_memory = 3815219089Spjd ptoa((uintmax_t)cnt.v_free_count + cnt.v_cache_count); 3816185029Spjd static uint64_t page_load = 0; 3817185029Spjd static uint64_t last_txg = 0; 3818185029Spjd 3819219089Spjd#ifdef sun 3820185029Spjd#if defined(__i386) 3821185029Spjd available_memory = 3822185029Spjd MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 3823185029Spjd#endif 3824219089Spjd#endif /* sun */ 3825185029Spjd if (available_memory >= zfs_write_limit_max) 3826185029Spjd return (0); 3827185029Spjd 3828185029Spjd if (txg > last_txg) { 3829185029Spjd last_txg = txg; 3830185029Spjd page_load = 0; 3831185029Spjd } 3832185029Spjd /* 3833185029Spjd * If we are in pageout, we know that memory is already tight, 3834185029Spjd * the arc is already going to be evicting, so we just want to 3835185029Spjd * continue to let page writes occur as quickly as possible. 3836185029Spjd */ 3837185029Spjd if (curproc == pageproc) { 3838185029Spjd if (page_load > available_memory / 4) 3839249195Smm return (SET_ERROR(ERESTART)); 3840185029Spjd /* Note: reserve is inflated, so we deflate */ 3841185029Spjd page_load += reserve / 8; 3842185029Spjd return (0); 3843185029Spjd } else if (page_load > 0 && arc_reclaim_needed()) { 3844185029Spjd /* memory is low, delay before restarting */ 3845185029Spjd ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3846249195Smm return (SET_ERROR(EAGAIN)); 3847185029Spjd } 3848185029Spjd page_load = 0; 3849185029Spjd 3850185029Spjd if (arc_size > arc_c_min) { 3851185029Spjd uint64_t evictable_memory = 3852185029Spjd arc_mru->arcs_lsize[ARC_BUFC_DATA] + 3853185029Spjd arc_mru->arcs_lsize[ARC_BUFC_METADATA] + 3854185029Spjd arc_mfu->arcs_lsize[ARC_BUFC_DATA] + 3855185029Spjd arc_mfu->arcs_lsize[ARC_BUFC_METADATA]; 3856185029Spjd available_memory += MIN(evictable_memory, arc_size - arc_c_min); 3857185029Spjd } 3858185029Spjd 3859185029Spjd if (inflight_data > available_memory / 4) { 3860185029Spjd ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3861249195Smm return (SET_ERROR(ERESTART)); 3862185029Spjd } 3863185029Spjd#endif 3864185029Spjd return (0); 3865185029Spjd} 3866185029Spjd 3867168404Spjdvoid 3868185029Spjdarc_tempreserve_clear(uint64_t reserve) 3869168404Spjd{ 3870185029Spjd atomic_add_64(&arc_tempreserve, -reserve); 3871168404Spjd ASSERT((int64_t)arc_tempreserve >= 0); 3872168404Spjd} 3873168404Spjd 3874168404Spjdint 3875185029Spjdarc_tempreserve_space(uint64_t reserve, uint64_t txg) 3876168404Spjd{ 3877185029Spjd int error; 3878209962Smm uint64_t anon_size; 3879185029Spjd 3880168404Spjd#ifdef ZFS_DEBUG 3881168404Spjd /* 3882168404Spjd * Once in a while, fail for no reason. Everything should cope. 3883168404Spjd */ 3884168404Spjd if (spa_get_random(10000) == 0) { 3885168404Spjd dprintf("forcing random failure\n"); 3886249195Smm return (SET_ERROR(ERESTART)); 3887168404Spjd } 3888168404Spjd#endif 3889185029Spjd if (reserve > arc_c/4 && !arc_no_grow) 3890185029Spjd arc_c = MIN(arc_c_max, reserve * 4); 3891185029Spjd if (reserve > arc_c) 3892249195Smm return (SET_ERROR(ENOMEM)); 3893168404Spjd 3894168404Spjd /* 3895209962Smm * Don't count loaned bufs as in flight dirty data to prevent long 3896209962Smm * network delays from blocking transactions that are ready to be 3897209962Smm * assigned to a txg. 3898209962Smm */ 3899209962Smm anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 3900209962Smm 3901209962Smm /* 3902185029Spjd * Writes will, almost always, require additional memory allocations 3903185029Spjd * in order to compress/encrypt/etc the data. We therefor need to 3904185029Spjd * make sure that there is sufficient available memory for this. 3905185029Spjd */ 3906209962Smm if (error = arc_memory_throttle(reserve, anon_size, txg)) 3907185029Spjd return (error); 3908185029Spjd 3909185029Spjd /* 3910168404Spjd * Throttle writes when the amount of dirty data in the cache 3911168404Spjd * gets too large. We try to keep the cache less than half full 3912168404Spjd * of dirty blocks so that our sync times don't grow too large. 3913168404Spjd * Note: if two requests come in concurrently, we might let them 3914168404Spjd * both succeed, when one of them should fail. Not a huge deal. 3915168404Spjd */ 3916209962Smm 3917209962Smm if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 3918209962Smm anon_size > arc_c / 4) { 3919185029Spjd dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 3920185029Spjd "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 3921185029Spjd arc_tempreserve>>10, 3922185029Spjd arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 3923185029Spjd arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 3924185029Spjd reserve>>10, arc_c>>10); 3925249195Smm return (SET_ERROR(ERESTART)); 3926168404Spjd } 3927185029Spjd atomic_add_64(&arc_tempreserve, reserve); 3928168404Spjd return (0); 3929168404Spjd} 3930168404Spjd 3931168582Spjdstatic kmutex_t arc_lowmem_lock; 3932168404Spjd#ifdef _KERNEL 3933168566Spjdstatic eventhandler_tag arc_event_lowmem = NULL; 3934168404Spjd 3935168404Spjdstatic void 3936168566Spjdarc_lowmem(void *arg __unused, int howto __unused) 3937168404Spjd{ 3938168404Spjd 3939168566Spjd /* Serialize access via arc_lowmem_lock. */ 3940168566Spjd mutex_enter(&arc_lowmem_lock); 3941219089Spjd mutex_enter(&arc_reclaim_thr_lock); 3942185029Spjd needfree = 1; 3943168404Spjd cv_signal(&arc_reclaim_thr_cv); 3944241773Savg 3945241773Savg /* 3946241773Savg * It is unsafe to block here in arbitrary threads, because we can come 3947241773Savg * here from ARC itself and may hold ARC locks and thus risk a deadlock 3948241773Savg * with ARC reclaim thread. 3949241773Savg */ 3950241773Savg if (curproc == pageproc) { 3951241773Savg while (needfree) 3952241773Savg msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 3953241773Savg } 3954219089Spjd mutex_exit(&arc_reclaim_thr_lock); 3955168566Spjd mutex_exit(&arc_lowmem_lock); 3956168404Spjd} 3957168404Spjd#endif 3958168404Spjd 3959168404Spjdvoid 3960168404Spjdarc_init(void) 3961168404Spjd{ 3962219089Spjd int i, prefetch_tunable_set = 0; 3963205231Skmacy 3964168404Spjd mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 3965168404Spjd cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 3966168566Spjd mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 3967168404Spjd 3968168404Spjd /* Convert seconds to clock ticks */ 3969168404Spjd arc_min_prefetch_lifespan = 1 * hz; 3970168404Spjd 3971168404Spjd /* Start out with 1/8 of all memory */ 3972168566Spjd arc_c = kmem_size() / 8; 3973219089Spjd 3974219089Spjd#ifdef sun 3975192360Skmacy#ifdef _KERNEL 3976192360Skmacy /* 3977192360Skmacy * On architectures where the physical memory can be larger 3978192360Skmacy * than the addressable space (intel in 32-bit mode), we may 3979192360Skmacy * need to limit the cache to 1/8 of VM size. 3980192360Skmacy */ 3981192360Skmacy arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 3982192360Skmacy#endif 3983219089Spjd#endif /* sun */ 3984168566Spjd /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 3985168566Spjd arc_c_min = MAX(arc_c / 4, 64<<18); 3986168566Spjd /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 3987168404Spjd if (arc_c * 8 >= 1<<30) 3988168404Spjd arc_c_max = (arc_c * 8) - (1<<30); 3989168404Spjd else 3990168404Spjd arc_c_max = arc_c_min; 3991175633Spjd arc_c_max = MAX(arc_c * 5, arc_c_max); 3992219089Spjd 3993168481Spjd#ifdef _KERNEL 3994168404Spjd /* 3995168404Spjd * Allow the tunables to override our calculations if they are 3996168566Spjd * reasonable (ie. over 16MB) 3997168404Spjd */ 3998219089Spjd if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size()) 3999168404Spjd arc_c_max = zfs_arc_max; 4000219089Spjd if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max) 4001168404Spjd arc_c_min = zfs_arc_min; 4002168481Spjd#endif 4003219089Spjd 4004168404Spjd arc_c = arc_c_max; 4005168404Spjd arc_p = (arc_c >> 1); 4006168404Spjd 4007185029Spjd /* limit meta-data to 1/4 of the arc capacity */ 4008185029Spjd arc_meta_limit = arc_c_max / 4; 4009185029Spjd 4010185029Spjd /* Allow the tunable to override if it is reasonable */ 4011185029Spjd if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4012185029Spjd arc_meta_limit = zfs_arc_meta_limit; 4013185029Spjd 4014185029Spjd if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4015185029Spjd arc_c_min = arc_meta_limit / 2; 4016185029Spjd 4017208373Smm if (zfs_arc_grow_retry > 0) 4018208373Smm arc_grow_retry = zfs_arc_grow_retry; 4019208373Smm 4020208373Smm if (zfs_arc_shrink_shift > 0) 4021208373Smm arc_shrink_shift = zfs_arc_shrink_shift; 4022208373Smm 4023208373Smm if (zfs_arc_p_min_shift > 0) 4024208373Smm arc_p_min_shift = zfs_arc_p_min_shift; 4025208373Smm 4026168404Spjd /* if kmem_flags are set, lets try to use less memory */ 4027168404Spjd if (kmem_debugging()) 4028168404Spjd arc_c = arc_c / 2; 4029168404Spjd if (arc_c < arc_c_min) 4030168404Spjd arc_c = arc_c_min; 4031168404Spjd 4032168473Spjd zfs_arc_min = arc_c_min; 4033168473Spjd zfs_arc_max = arc_c_max; 4034168473Spjd 4035168404Spjd arc_anon = &ARC_anon; 4036168404Spjd arc_mru = &ARC_mru; 4037168404Spjd arc_mru_ghost = &ARC_mru_ghost; 4038168404Spjd arc_mfu = &ARC_mfu; 4039168404Spjd arc_mfu_ghost = &ARC_mfu_ghost; 4040185029Spjd arc_l2c_only = &ARC_l2c_only; 4041168404Spjd arc_size = 0; 4042168404Spjd 4043205231Skmacy for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4044205231Skmacy mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4045205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4046205231Skmacy mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4047205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4048205231Skmacy mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4049205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4050205231Skmacy mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4051205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4052205231Skmacy mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4053205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4054205231Skmacy mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4055205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4056206796Spjd 4057205231Skmacy list_create(&arc_mru->arcs_lists[i], 4058205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4059205231Skmacy list_create(&arc_mru_ghost->arcs_lists[i], 4060205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4061205231Skmacy list_create(&arc_mfu->arcs_lists[i], 4062205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4063205231Skmacy list_create(&arc_mfu_ghost->arcs_lists[i], 4064205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4065205231Skmacy list_create(&arc_mfu_ghost->arcs_lists[i], 4066205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4067205231Skmacy list_create(&arc_l2c_only->arcs_lists[i], 4068205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4069205231Skmacy } 4070168404Spjd 4071168404Spjd buf_init(); 4072168404Spjd 4073168404Spjd arc_thread_exit = 0; 4074168404Spjd arc_eviction_list = NULL; 4075168404Spjd mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4076168404Spjd bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4077168404Spjd 4078168404Spjd arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4079168404Spjd sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4080168404Spjd 4081168404Spjd if (arc_ksp != NULL) { 4082168404Spjd arc_ksp->ks_data = &arc_stats; 4083168404Spjd kstat_install(arc_ksp); 4084168404Spjd } 4085168404Spjd 4086168404Spjd (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4087168404Spjd TS_RUN, minclsyspri); 4088168404Spjd 4089168404Spjd#ifdef _KERNEL 4090168566Spjd arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4091168404Spjd EVENTHANDLER_PRI_FIRST); 4092168404Spjd#endif 4093168404Spjd 4094168404Spjd arc_dead = FALSE; 4095185029Spjd arc_warm = B_FALSE; 4096168566Spjd 4097185029Spjd if (zfs_write_limit_max == 0) 4098185029Spjd zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift; 4099185029Spjd else 4100185029Spjd zfs_write_limit_shift = 0; 4101185029Spjd mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL); 4102185029Spjd 4103168566Spjd#ifdef _KERNEL 4104194043Skmacy if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4105193953Skmacy prefetch_tunable_set = 1; 4106206796Spjd 4107193878Skmacy#ifdef __i386__ 4108193953Skmacy if (prefetch_tunable_set == 0) { 4109196863Strasz printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4110196863Strasz "-- to enable,\n"); 4111196863Strasz printf(" add \"vfs.zfs.prefetch_disable=0\" " 4112196863Strasz "to /boot/loader.conf.\n"); 4113219089Spjd zfs_prefetch_disable = 1; 4114193878Skmacy } 4115206796Spjd#else 4116193878Skmacy if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4117193953Skmacy prefetch_tunable_set == 0) { 4118196863Strasz printf("ZFS NOTICE: Prefetch is disabled by default if less " 4119196941Strasz "than 4GB of RAM is present;\n" 4120196863Strasz " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4121196863Strasz "to /boot/loader.conf.\n"); 4122219089Spjd zfs_prefetch_disable = 1; 4123193878Skmacy } 4124206796Spjd#endif 4125175633Spjd /* Warn about ZFS memory and address space requirements. */ 4126168696Spjd if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4127168987Sbmah printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4128168987Sbmah "expect unstable behavior.\n"); 4129175633Spjd } 4130175633Spjd if (kmem_size() < 512 * (1 << 20)) { 4131173419Spjd printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4132168987Sbmah "expect unstable behavior.\n"); 4133185029Spjd printf(" Consider tuning vm.kmem_size and " 4134173419Spjd "vm.kmem_size_max\n"); 4135185029Spjd printf(" in /boot/loader.conf.\n"); 4136168566Spjd } 4137168566Spjd#endif 4138168404Spjd} 4139168404Spjd 4140168404Spjdvoid 4141168404Spjdarc_fini(void) 4142168404Spjd{ 4143205231Skmacy int i; 4144206796Spjd 4145168404Spjd mutex_enter(&arc_reclaim_thr_lock); 4146168404Spjd arc_thread_exit = 1; 4147168404Spjd cv_signal(&arc_reclaim_thr_cv); 4148168404Spjd while (arc_thread_exit != 0) 4149168404Spjd cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4150168404Spjd mutex_exit(&arc_reclaim_thr_lock); 4151168404Spjd 4152185029Spjd arc_flush(NULL); 4153168404Spjd 4154168404Spjd arc_dead = TRUE; 4155168404Spjd 4156168404Spjd if (arc_ksp != NULL) { 4157168404Spjd kstat_delete(arc_ksp); 4158168404Spjd arc_ksp = NULL; 4159168404Spjd } 4160168404Spjd 4161168404Spjd mutex_destroy(&arc_eviction_mtx); 4162168404Spjd mutex_destroy(&arc_reclaim_thr_lock); 4163168404Spjd cv_destroy(&arc_reclaim_thr_cv); 4164168404Spjd 4165205231Skmacy for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4166205231Skmacy list_destroy(&arc_mru->arcs_lists[i]); 4167205231Skmacy list_destroy(&arc_mru_ghost->arcs_lists[i]); 4168205231Skmacy list_destroy(&arc_mfu->arcs_lists[i]); 4169205231Skmacy list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4170206795Spjd list_destroy(&arc_l2c_only->arcs_lists[i]); 4171168404Spjd 4172205231Skmacy mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4173205231Skmacy mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4174205231Skmacy mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4175205231Skmacy mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4176205231Skmacy mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4177206795Spjd mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4178205231Skmacy } 4179206796Spjd 4180185029Spjd mutex_destroy(&zfs_write_limit_lock); 4181185029Spjd 4182168404Spjd buf_fini(); 4183168404Spjd 4184209962Smm ASSERT(arc_loaned_bytes == 0); 4185209962Smm 4186168582Spjd mutex_destroy(&arc_lowmem_lock); 4187168404Spjd#ifdef _KERNEL 4188168566Spjd if (arc_event_lowmem != NULL) 4189168566Spjd EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4190168404Spjd#endif 4191168404Spjd} 4192185029Spjd 4193185029Spjd/* 4194185029Spjd * Level 2 ARC 4195185029Spjd * 4196185029Spjd * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4197185029Spjd * It uses dedicated storage devices to hold cached data, which are populated 4198185029Spjd * using large infrequent writes. The main role of this cache is to boost 4199185029Spjd * the performance of random read workloads. The intended L2ARC devices 4200185029Spjd * include short-stroked disks, solid state disks, and other media with 4201185029Spjd * substantially faster read latency than disk. 4202185029Spjd * 4203185029Spjd * +-----------------------+ 4204185029Spjd * | ARC | 4205185029Spjd * +-----------------------+ 4206185029Spjd * | ^ ^ 4207185029Spjd * | | | 4208185029Spjd * l2arc_feed_thread() arc_read() 4209185029Spjd * | | | 4210185029Spjd * | l2arc read | 4211185029Spjd * V | | 4212185029Spjd * +---------------+ | 4213185029Spjd * | L2ARC | | 4214185029Spjd * +---------------+ | 4215185029Spjd * | ^ | 4216185029Spjd * l2arc_write() | | 4217185029Spjd * | | | 4218185029Spjd * V | | 4219185029Spjd * +-------+ +-------+ 4220185029Spjd * | vdev | | vdev | 4221185029Spjd * | cache | | cache | 4222185029Spjd * +-------+ +-------+ 4223185029Spjd * +=========+ .-----. 4224185029Spjd * : L2ARC : |-_____-| 4225185029Spjd * : devices : | Disks | 4226185029Spjd * +=========+ `-_____-' 4227185029Spjd * 4228185029Spjd * Read requests are satisfied from the following sources, in order: 4229185029Spjd * 4230185029Spjd * 1) ARC 4231185029Spjd * 2) vdev cache of L2ARC devices 4232185029Spjd * 3) L2ARC devices 4233185029Spjd * 4) vdev cache of disks 4234185029Spjd * 5) disks 4235185029Spjd * 4236185029Spjd * Some L2ARC device types exhibit extremely slow write performance. 4237185029Spjd * To accommodate for this there are some significant differences between 4238185029Spjd * the L2ARC and traditional cache design: 4239185029Spjd * 4240185029Spjd * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4241185029Spjd * the ARC behave as usual, freeing buffers and placing headers on ghost 4242185029Spjd * lists. The ARC does not send buffers to the L2ARC during eviction as 4243185029Spjd * this would add inflated write latencies for all ARC memory pressure. 4244185029Spjd * 4245185029Spjd * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4246185029Spjd * It does this by periodically scanning buffers from the eviction-end of 4247185029Spjd * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4248251478Sdelphij * not already there. It scans until a headroom of buffers is satisfied, 4249251478Sdelphij * which itself is a buffer for ARC eviction. If a compressible buffer is 4250251478Sdelphij * found during scanning and selected for writing to an L2ARC device, we 4251251478Sdelphij * temporarily boost scanning headroom during the next scan cycle to make 4252251478Sdelphij * sure we adapt to compression effects (which might significantly reduce 4253251478Sdelphij * the data volume we write to L2ARC). The thread that does this is 4254185029Spjd * l2arc_feed_thread(), illustrated below; example sizes are included to 4255185029Spjd * provide a better sense of ratio than this diagram: 4256185029Spjd * 4257185029Spjd * head --> tail 4258185029Spjd * +---------------------+----------+ 4259185029Spjd * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4260185029Spjd * +---------------------+----------+ | o L2ARC eligible 4261185029Spjd * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4262185029Spjd * +---------------------+----------+ | 4263185029Spjd * 15.9 Gbytes ^ 32 Mbytes | 4264185029Spjd * headroom | 4265185029Spjd * l2arc_feed_thread() 4266185029Spjd * | 4267185029Spjd * l2arc write hand <--[oooo]--' 4268185029Spjd * | 8 Mbyte 4269185029Spjd * | write max 4270185029Spjd * V 4271185029Spjd * +==============================+ 4272185029Spjd * L2ARC dev |####|#|###|###| |####| ... | 4273185029Spjd * +==============================+ 4274185029Spjd * 32 Gbytes 4275185029Spjd * 4276185029Spjd * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4277185029Spjd * evicted, then the L2ARC has cached a buffer much sooner than it probably 4278185029Spjd * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4279185029Spjd * safe to say that this is an uncommon case, since buffers at the end of 4280185029Spjd * the ARC lists have moved there due to inactivity. 4281185029Spjd * 4282185029Spjd * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4283185029Spjd * then the L2ARC simply misses copying some buffers. This serves as a 4284185029Spjd * pressure valve to prevent heavy read workloads from both stalling the ARC 4285185029Spjd * with waits and clogging the L2ARC with writes. This also helps prevent 4286185029Spjd * the potential for the L2ARC to churn if it attempts to cache content too 4287185029Spjd * quickly, such as during backups of the entire pool. 4288185029Spjd * 4289185029Spjd * 5. After system boot and before the ARC has filled main memory, there are 4290185029Spjd * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4291185029Spjd * lists can remain mostly static. Instead of searching from tail of these 4292185029Spjd * lists as pictured, the l2arc_feed_thread() will search from the list heads 4293185029Spjd * for eligible buffers, greatly increasing its chance of finding them. 4294185029Spjd * 4295185029Spjd * The L2ARC device write speed is also boosted during this time so that 4296185029Spjd * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4297185029Spjd * there are no L2ARC reads, and no fear of degrading read performance 4298185029Spjd * through increased writes. 4299185029Spjd * 4300185029Spjd * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4301185029Spjd * the vdev queue can aggregate them into larger and fewer writes. Each 4302185029Spjd * device is written to in a rotor fashion, sweeping writes through 4303185029Spjd * available space then repeating. 4304185029Spjd * 4305185029Spjd * 7. The L2ARC does not store dirty content. It never needs to flush 4306185029Spjd * write buffers back to disk based storage. 4307185029Spjd * 4308185029Spjd * 8. If an ARC buffer is written (and dirtied) which also exists in the 4309185029Spjd * L2ARC, the now stale L2ARC buffer is immediately dropped. 4310185029Spjd * 4311185029Spjd * The performance of the L2ARC can be tweaked by a number of tunables, which 4312185029Spjd * may be necessary for different workloads: 4313185029Spjd * 4314185029Spjd * l2arc_write_max max write bytes per interval 4315185029Spjd * l2arc_write_boost extra write bytes during device warmup 4316185029Spjd * l2arc_noprefetch skip caching prefetched buffers 4317185029Spjd * l2arc_headroom number of max device writes to precache 4318251478Sdelphij * l2arc_headroom_boost when we find compressed buffers during ARC 4319251478Sdelphij * scanning, we multiply headroom by this 4320251478Sdelphij * percentage factor for the next scan cycle, 4321251478Sdelphij * since more compressed buffers are likely to 4322251478Sdelphij * be present 4323185029Spjd * l2arc_feed_secs seconds between L2ARC writing 4324185029Spjd * 4325185029Spjd * Tunables may be removed or added as future performance improvements are 4326185029Spjd * integrated, and also may become zpool properties. 4327208373Smm * 4328208373Smm * There are three key functions that control how the L2ARC warms up: 4329208373Smm * 4330208373Smm * l2arc_write_eligible() check if a buffer is eligible to cache 4331208373Smm * l2arc_write_size() calculate how much to write 4332208373Smm * l2arc_write_interval() calculate sleep delay between writes 4333208373Smm * 4334208373Smm * These three functions determine what to write, how much, and how quickly 4335208373Smm * to send writes. 4336185029Spjd */ 4337185029Spjd 4338208373Smmstatic boolean_t 4339209962Smml2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab) 4340208373Smm{ 4341208373Smm /* 4342208373Smm * A buffer is *not* eligible for the L2ARC if it: 4343208373Smm * 1. belongs to a different spa. 4344208373Smm * 2. is already cached on the L2ARC. 4345208373Smm * 3. has an I/O in progress (it may be an incomplete read). 4346208373Smm * 4. is flagged not eligible (zfs property). 4347208373Smm */ 4348209962Smm if (ab->b_spa != spa_guid) { 4349208373Smm ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4350208373Smm return (B_FALSE); 4351208373Smm } 4352208373Smm if (ab->b_l2hdr != NULL) { 4353208373Smm ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4354208373Smm return (B_FALSE); 4355208373Smm } 4356208373Smm if (HDR_IO_IN_PROGRESS(ab)) { 4357208373Smm ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4358208373Smm return (B_FALSE); 4359208373Smm } 4360208373Smm if (!HDR_L2CACHE(ab)) { 4361208373Smm ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4362208373Smm return (B_FALSE); 4363208373Smm } 4364208373Smm 4365208373Smm return (B_TRUE); 4366208373Smm} 4367208373Smm 4368208373Smmstatic uint64_t 4369251478Sdelphijl2arc_write_size(void) 4370208373Smm{ 4371208373Smm uint64_t size; 4372208373Smm 4373251478Sdelphij /* 4374251478Sdelphij * Make sure our globals have meaningful values in case the user 4375251478Sdelphij * altered them. 4376251478Sdelphij */ 4377251478Sdelphij size = l2arc_write_max; 4378251478Sdelphij if (size == 0) { 4379251478Sdelphij cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4380251478Sdelphij "be greater than zero, resetting it to the default (%d)", 4381251478Sdelphij L2ARC_WRITE_SIZE); 4382251478Sdelphij size = l2arc_write_max = L2ARC_WRITE_SIZE; 4383251478Sdelphij } 4384208373Smm 4385208373Smm if (arc_warm == B_FALSE) 4386251478Sdelphij size += l2arc_write_boost; 4387208373Smm 4388208373Smm return (size); 4389208373Smm 4390208373Smm} 4391208373Smm 4392208373Smmstatic clock_t 4393208373Smml2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4394208373Smm{ 4395219089Spjd clock_t interval, next, now; 4396208373Smm 4397208373Smm /* 4398208373Smm * If the ARC lists are busy, increase our write rate; if the 4399208373Smm * lists are stale, idle back. This is achieved by checking 4400208373Smm * how much we previously wrote - if it was more than half of 4401208373Smm * what we wanted, schedule the next write much sooner. 4402208373Smm */ 4403208373Smm if (l2arc_feed_again && wrote > (wanted / 2)) 4404208373Smm interval = (hz * l2arc_feed_min_ms) / 1000; 4405208373Smm else 4406208373Smm interval = hz * l2arc_feed_secs; 4407208373Smm 4408219089Spjd now = ddi_get_lbolt(); 4409219089Spjd next = MAX(now, MIN(now + interval, began + interval)); 4410208373Smm 4411208373Smm return (next); 4412208373Smm} 4413208373Smm 4414185029Spjdstatic void 4415185029Spjdl2arc_hdr_stat_add(void) 4416185029Spjd{ 4417185029Spjd ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4418185029Spjd ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4419185029Spjd} 4420185029Spjd 4421185029Spjdstatic void 4422185029Spjdl2arc_hdr_stat_remove(void) 4423185029Spjd{ 4424185029Spjd ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4425185029Spjd ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4426185029Spjd} 4427185029Spjd 4428185029Spjd/* 4429185029Spjd * Cycle through L2ARC devices. This is how L2ARC load balances. 4430185029Spjd * If a device is returned, this also returns holding the spa config lock. 4431185029Spjd */ 4432185029Spjdstatic l2arc_dev_t * 4433185029Spjdl2arc_dev_get_next(void) 4434185029Spjd{ 4435185029Spjd l2arc_dev_t *first, *next = NULL; 4436185029Spjd 4437185029Spjd /* 4438185029Spjd * Lock out the removal of spas (spa_namespace_lock), then removal 4439185029Spjd * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4440185029Spjd * both locks will be dropped and a spa config lock held instead. 4441185029Spjd */ 4442185029Spjd mutex_enter(&spa_namespace_lock); 4443185029Spjd mutex_enter(&l2arc_dev_mtx); 4444185029Spjd 4445185029Spjd /* if there are no vdevs, there is nothing to do */ 4446185029Spjd if (l2arc_ndev == 0) 4447185029Spjd goto out; 4448185029Spjd 4449185029Spjd first = NULL; 4450185029Spjd next = l2arc_dev_last; 4451185029Spjd do { 4452185029Spjd /* loop around the list looking for a non-faulted vdev */ 4453185029Spjd if (next == NULL) { 4454185029Spjd next = list_head(l2arc_dev_list); 4455185029Spjd } else { 4456185029Spjd next = list_next(l2arc_dev_list, next); 4457185029Spjd if (next == NULL) 4458185029Spjd next = list_head(l2arc_dev_list); 4459185029Spjd } 4460185029Spjd 4461185029Spjd /* if we have come back to the start, bail out */ 4462185029Spjd if (first == NULL) 4463185029Spjd first = next; 4464185029Spjd else if (next == first) 4465185029Spjd break; 4466185029Spjd 4467185029Spjd } while (vdev_is_dead(next->l2ad_vdev)); 4468185029Spjd 4469185029Spjd /* if we were unable to find any usable vdevs, return NULL */ 4470185029Spjd if (vdev_is_dead(next->l2ad_vdev)) 4471185029Spjd next = NULL; 4472185029Spjd 4473185029Spjd l2arc_dev_last = next; 4474185029Spjd 4475185029Spjdout: 4476185029Spjd mutex_exit(&l2arc_dev_mtx); 4477185029Spjd 4478185029Spjd /* 4479185029Spjd * Grab the config lock to prevent the 'next' device from being 4480185029Spjd * removed while we are writing to it. 4481185029Spjd */ 4482185029Spjd if (next != NULL) 4483185029Spjd spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4484185029Spjd mutex_exit(&spa_namespace_lock); 4485185029Spjd 4486185029Spjd return (next); 4487185029Spjd} 4488185029Spjd 4489185029Spjd/* 4490185029Spjd * Free buffers that were tagged for destruction. 4491185029Spjd */ 4492185029Spjdstatic void 4493185029Spjdl2arc_do_free_on_write() 4494185029Spjd{ 4495185029Spjd list_t *buflist; 4496185029Spjd l2arc_data_free_t *df, *df_prev; 4497185029Spjd 4498185029Spjd mutex_enter(&l2arc_free_on_write_mtx); 4499185029Spjd buflist = l2arc_free_on_write; 4500185029Spjd 4501185029Spjd for (df = list_tail(buflist); df; df = df_prev) { 4502185029Spjd df_prev = list_prev(buflist, df); 4503185029Spjd ASSERT(df->l2df_data != NULL); 4504185029Spjd ASSERT(df->l2df_func != NULL); 4505185029Spjd df->l2df_func(df->l2df_data, df->l2df_size); 4506185029Spjd list_remove(buflist, df); 4507185029Spjd kmem_free(df, sizeof (l2arc_data_free_t)); 4508185029Spjd } 4509185029Spjd 4510185029Spjd mutex_exit(&l2arc_free_on_write_mtx); 4511185029Spjd} 4512185029Spjd 4513185029Spjd/* 4514185029Spjd * A write to a cache device has completed. Update all headers to allow 4515185029Spjd * reads from these buffers to begin. 4516185029Spjd */ 4517185029Spjdstatic void 4518185029Spjdl2arc_write_done(zio_t *zio) 4519185029Spjd{ 4520185029Spjd l2arc_write_callback_t *cb; 4521185029Spjd l2arc_dev_t *dev; 4522185029Spjd list_t *buflist; 4523185029Spjd arc_buf_hdr_t *head, *ab, *ab_prev; 4524185029Spjd l2arc_buf_hdr_t *abl2; 4525185029Spjd kmutex_t *hash_lock; 4526185029Spjd 4527185029Spjd cb = zio->io_private; 4528185029Spjd ASSERT(cb != NULL); 4529185029Spjd dev = cb->l2wcb_dev; 4530185029Spjd ASSERT(dev != NULL); 4531185029Spjd head = cb->l2wcb_head; 4532185029Spjd ASSERT(head != NULL); 4533185029Spjd buflist = dev->l2ad_buflist; 4534185029Spjd ASSERT(buflist != NULL); 4535185029Spjd DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4536185029Spjd l2arc_write_callback_t *, cb); 4537185029Spjd 4538185029Spjd if (zio->io_error != 0) 4539185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_error); 4540185029Spjd 4541185029Spjd mutex_enter(&l2arc_buflist_mtx); 4542185029Spjd 4543185029Spjd /* 4544185029Spjd * All writes completed, or an error was hit. 4545185029Spjd */ 4546185029Spjd for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 4547185029Spjd ab_prev = list_prev(buflist, ab); 4548185029Spjd 4549185029Spjd hash_lock = HDR_LOCK(ab); 4550185029Spjd if (!mutex_tryenter(hash_lock)) { 4551185029Spjd /* 4552185029Spjd * This buffer misses out. It may be in a stage 4553185029Spjd * of eviction. Its ARC_L2_WRITING flag will be 4554185029Spjd * left set, denying reads to this buffer. 4555185029Spjd */ 4556185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4557185029Spjd continue; 4558185029Spjd } 4559185029Spjd 4560251478Sdelphij abl2 = ab->b_l2hdr; 4561251478Sdelphij 4562251478Sdelphij /* 4563251478Sdelphij * Release the temporary compressed buffer as soon as possible. 4564251478Sdelphij */ 4565251478Sdelphij if (abl2->b_compress != ZIO_COMPRESS_OFF) 4566251478Sdelphij l2arc_release_cdata_buf(ab); 4567251478Sdelphij 4568185029Spjd if (zio->io_error != 0) { 4569185029Spjd /* 4570185029Spjd * Error - drop L2ARC entry. 4571185029Spjd */ 4572185029Spjd list_remove(buflist, ab); 4573251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4574185029Spjd ab->b_l2hdr = NULL; 4575248572Ssmh trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4576248574Ssmh ab->b_size, 0); 4577185029Spjd kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4578185029Spjd ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4579185029Spjd } 4580185029Spjd 4581185029Spjd /* 4582185029Spjd * Allow ARC to begin reads to this L2ARC entry. 4583185029Spjd */ 4584185029Spjd ab->b_flags &= ~ARC_L2_WRITING; 4585185029Spjd 4586185029Spjd mutex_exit(hash_lock); 4587185029Spjd } 4588185029Spjd 4589185029Spjd atomic_inc_64(&l2arc_writes_done); 4590185029Spjd list_remove(buflist, head); 4591185029Spjd kmem_cache_free(hdr_cache, head); 4592185029Spjd mutex_exit(&l2arc_buflist_mtx); 4593185029Spjd 4594185029Spjd l2arc_do_free_on_write(); 4595185029Spjd 4596185029Spjd kmem_free(cb, sizeof (l2arc_write_callback_t)); 4597185029Spjd} 4598185029Spjd 4599185029Spjd/* 4600185029Spjd * A read to a cache device completed. Validate buffer contents before 4601185029Spjd * handing over to the regular ARC routines. 4602185029Spjd */ 4603185029Spjdstatic void 4604185029Spjdl2arc_read_done(zio_t *zio) 4605185029Spjd{ 4606185029Spjd l2arc_read_callback_t *cb; 4607185029Spjd arc_buf_hdr_t *hdr; 4608185029Spjd arc_buf_t *buf; 4609185029Spjd kmutex_t *hash_lock; 4610185029Spjd int equal; 4611185029Spjd 4612185029Spjd ASSERT(zio->io_vd != NULL); 4613185029Spjd ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4614185029Spjd 4615185029Spjd spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4616185029Spjd 4617185029Spjd cb = zio->io_private; 4618185029Spjd ASSERT(cb != NULL); 4619185029Spjd buf = cb->l2rcb_buf; 4620185029Spjd ASSERT(buf != NULL); 4621185029Spjd 4622219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 4623185029Spjd mutex_enter(hash_lock); 4624219089Spjd hdr = buf->b_hdr; 4625219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4626185029Spjd 4627185029Spjd /* 4628251478Sdelphij * If the buffer was compressed, decompress it first. 4629251478Sdelphij */ 4630251478Sdelphij if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4631251478Sdelphij l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4632251478Sdelphij ASSERT(zio->io_data != NULL); 4633251478Sdelphij 4634251478Sdelphij /* 4635185029Spjd * Check this survived the L2ARC journey. 4636185029Spjd */ 4637185029Spjd equal = arc_cksum_equal(buf); 4638185029Spjd if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4639185029Spjd mutex_exit(hash_lock); 4640185029Spjd zio->io_private = buf; 4641185029Spjd zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4642185029Spjd zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4643185029Spjd arc_read_done(zio); 4644185029Spjd } else { 4645185029Spjd mutex_exit(hash_lock); 4646185029Spjd /* 4647185029Spjd * Buffer didn't survive caching. Increment stats and 4648185029Spjd * reissue to the original storage device. 4649185029Spjd */ 4650185029Spjd if (zio->io_error != 0) { 4651185029Spjd ARCSTAT_BUMP(arcstat_l2_io_error); 4652185029Spjd } else { 4653249195Smm zio->io_error = SET_ERROR(EIO); 4654185029Spjd } 4655185029Spjd if (!equal) 4656185029Spjd ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4657185029Spjd 4658185029Spjd /* 4659185029Spjd * If there's no waiter, issue an async i/o to the primary 4660185029Spjd * storage now. If there *is* a waiter, the caller must 4661185029Spjd * issue the i/o in a context where it's OK to block. 4662185029Spjd */ 4663209962Smm if (zio->io_waiter == NULL) { 4664209962Smm zio_t *pio = zio_unique_parent(zio); 4665209962Smm 4666209962Smm ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4667209962Smm 4668209962Smm zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4669185029Spjd buf->b_data, zio->io_size, arc_read_done, buf, 4670185029Spjd zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4671209962Smm } 4672185029Spjd } 4673185029Spjd 4674185029Spjd kmem_free(cb, sizeof (l2arc_read_callback_t)); 4675185029Spjd} 4676185029Spjd 4677185029Spjd/* 4678185029Spjd * This is the list priority from which the L2ARC will search for pages to 4679185029Spjd * cache. This is used within loops (0..3) to cycle through lists in the 4680185029Spjd * desired order. This order can have a significant effect on cache 4681185029Spjd * performance. 4682185029Spjd * 4683185029Spjd * Currently the metadata lists are hit first, MFU then MRU, followed by 4684185029Spjd * the data lists. This function returns a locked list, and also returns 4685185029Spjd * the lock pointer. 4686185029Spjd */ 4687185029Spjdstatic list_t * 4688185029Spjdl2arc_list_locked(int list_num, kmutex_t **lock) 4689185029Spjd{ 4690247187Smm list_t *list = NULL; 4691205231Skmacy int idx; 4692185029Spjd 4693206796Spjd ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4694206796Spjd 4695205231Skmacy if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4696205231Skmacy idx = list_num; 4697205231Skmacy list = &arc_mfu->arcs_lists[idx]; 4698205231Skmacy *lock = ARCS_LOCK(arc_mfu, idx); 4699206796Spjd } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4700205231Skmacy idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4701205231Skmacy list = &arc_mru->arcs_lists[idx]; 4702205231Skmacy *lock = ARCS_LOCK(arc_mru, idx); 4703206796Spjd } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4704205231Skmacy ARC_BUFC_NUMDATALISTS)) { 4705205231Skmacy idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4706205231Skmacy list = &arc_mfu->arcs_lists[idx]; 4707205231Skmacy *lock = ARCS_LOCK(arc_mfu, idx); 4708205231Skmacy } else { 4709205231Skmacy idx = list_num - ARC_BUFC_NUMLISTS; 4710205231Skmacy list = &arc_mru->arcs_lists[idx]; 4711205231Skmacy *lock = ARCS_LOCK(arc_mru, idx); 4712185029Spjd } 4713185029Spjd 4714185029Spjd ASSERT(!(MUTEX_HELD(*lock))); 4715185029Spjd mutex_enter(*lock); 4716185029Spjd return (list); 4717185029Spjd} 4718185029Spjd 4719185029Spjd/* 4720185029Spjd * Evict buffers from the device write hand to the distance specified in 4721185029Spjd * bytes. This distance may span populated buffers, it may span nothing. 4722185029Spjd * This is clearing a region on the L2ARC device ready for writing. 4723185029Spjd * If the 'all' boolean is set, every buffer is evicted. 4724185029Spjd */ 4725185029Spjdstatic void 4726185029Spjdl2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4727185029Spjd{ 4728185029Spjd list_t *buflist; 4729185029Spjd l2arc_buf_hdr_t *abl2; 4730185029Spjd arc_buf_hdr_t *ab, *ab_prev; 4731185029Spjd kmutex_t *hash_lock; 4732185029Spjd uint64_t taddr; 4733185029Spjd 4734185029Spjd buflist = dev->l2ad_buflist; 4735185029Spjd 4736185029Spjd if (buflist == NULL) 4737185029Spjd return; 4738185029Spjd 4739185029Spjd if (!all && dev->l2ad_first) { 4740185029Spjd /* 4741185029Spjd * This is the first sweep through the device. There is 4742185029Spjd * nothing to evict. 4743185029Spjd */ 4744185029Spjd return; 4745185029Spjd } 4746185029Spjd 4747185029Spjd if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4748185029Spjd /* 4749185029Spjd * When nearing the end of the device, evict to the end 4750185029Spjd * before the device write hand jumps to the start. 4751185029Spjd */ 4752185029Spjd taddr = dev->l2ad_end; 4753185029Spjd } else { 4754185029Spjd taddr = dev->l2ad_hand + distance; 4755185029Spjd } 4756185029Spjd DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4757185029Spjd uint64_t, taddr, boolean_t, all); 4758185029Spjd 4759185029Spjdtop: 4760185029Spjd mutex_enter(&l2arc_buflist_mtx); 4761185029Spjd for (ab = list_tail(buflist); ab; ab = ab_prev) { 4762185029Spjd ab_prev = list_prev(buflist, ab); 4763185029Spjd 4764185029Spjd hash_lock = HDR_LOCK(ab); 4765185029Spjd if (!mutex_tryenter(hash_lock)) { 4766185029Spjd /* 4767185029Spjd * Missed the hash lock. Retry. 4768185029Spjd */ 4769185029Spjd ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4770185029Spjd mutex_exit(&l2arc_buflist_mtx); 4771185029Spjd mutex_enter(hash_lock); 4772185029Spjd mutex_exit(hash_lock); 4773185029Spjd goto top; 4774185029Spjd } 4775185029Spjd 4776185029Spjd if (HDR_L2_WRITE_HEAD(ab)) { 4777185029Spjd /* 4778185029Spjd * We hit a write head node. Leave it for 4779185029Spjd * l2arc_write_done(). 4780185029Spjd */ 4781185029Spjd list_remove(buflist, ab); 4782185029Spjd mutex_exit(hash_lock); 4783185029Spjd continue; 4784185029Spjd } 4785185029Spjd 4786185029Spjd if (!all && ab->b_l2hdr != NULL && 4787185029Spjd (ab->b_l2hdr->b_daddr > taddr || 4788185029Spjd ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4789185029Spjd /* 4790185029Spjd * We've evicted to the target address, 4791185029Spjd * or the end of the device. 4792185029Spjd */ 4793185029Spjd mutex_exit(hash_lock); 4794185029Spjd break; 4795185029Spjd } 4796185029Spjd 4797185029Spjd if (HDR_FREE_IN_PROGRESS(ab)) { 4798185029Spjd /* 4799185029Spjd * Already on the path to destruction. 4800185029Spjd */ 4801185029Spjd mutex_exit(hash_lock); 4802185029Spjd continue; 4803185029Spjd } 4804185029Spjd 4805185029Spjd if (ab->b_state == arc_l2c_only) { 4806185029Spjd ASSERT(!HDR_L2_READING(ab)); 4807185029Spjd /* 4808185029Spjd * This doesn't exist in the ARC. Destroy. 4809185029Spjd * arc_hdr_destroy() will call list_remove() 4810185029Spjd * and decrement arcstat_l2_size. 4811185029Spjd */ 4812185029Spjd arc_change_state(arc_anon, ab, hash_lock); 4813185029Spjd arc_hdr_destroy(ab); 4814185029Spjd } else { 4815185029Spjd /* 4816185029Spjd * Invalidate issued or about to be issued 4817185029Spjd * reads, since we may be about to write 4818185029Spjd * over this location. 4819185029Spjd */ 4820185029Spjd if (HDR_L2_READING(ab)) { 4821185029Spjd ARCSTAT_BUMP(arcstat_l2_evict_reading); 4822185029Spjd ab->b_flags |= ARC_L2_EVICTED; 4823185029Spjd } 4824185029Spjd 4825185029Spjd /* 4826185029Spjd * Tell ARC this no longer exists in L2ARC. 4827185029Spjd */ 4828185029Spjd if (ab->b_l2hdr != NULL) { 4829185029Spjd abl2 = ab->b_l2hdr; 4830251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4831185029Spjd ab->b_l2hdr = NULL; 4832185029Spjd kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4833185029Spjd ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4834185029Spjd } 4835185029Spjd list_remove(buflist, ab); 4836185029Spjd 4837185029Spjd /* 4838185029Spjd * This may have been leftover after a 4839185029Spjd * failed write. 4840185029Spjd */ 4841185029Spjd ab->b_flags &= ~ARC_L2_WRITING; 4842185029Spjd } 4843185029Spjd mutex_exit(hash_lock); 4844185029Spjd } 4845185029Spjd mutex_exit(&l2arc_buflist_mtx); 4846185029Spjd 4847219089Spjd vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0); 4848185029Spjd dev->l2ad_evict = taddr; 4849185029Spjd} 4850185029Spjd 4851185029Spjd/* 4852185029Spjd * Find and write ARC buffers to the L2ARC device. 4853185029Spjd * 4854185029Spjd * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 4855185029Spjd * for reading until they have completed writing. 4856251478Sdelphij * The headroom_boost is an in-out parameter used to maintain headroom boost 4857251478Sdelphij * state between calls to this function. 4858251478Sdelphij * 4859251478Sdelphij * Returns the number of bytes actually written (which may be smaller than 4860251478Sdelphij * the delta by which the device hand has changed due to alignment). 4861185029Spjd */ 4862208373Smmstatic uint64_t 4863251478Sdelphijl2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 4864251478Sdelphij boolean_t *headroom_boost) 4865185029Spjd{ 4866185029Spjd arc_buf_hdr_t *ab, *ab_prev, *head; 4867185029Spjd list_t *list; 4868251478Sdelphij uint64_t write_asize, write_psize, write_sz, headroom, 4869251478Sdelphij buf_compress_minsz; 4870185029Spjd void *buf_data; 4871251478Sdelphij kmutex_t *list_lock; 4872251478Sdelphij boolean_t full; 4873185029Spjd l2arc_write_callback_t *cb; 4874185029Spjd zio_t *pio, *wzio; 4875228103Smm uint64_t guid = spa_load_guid(spa); 4876251478Sdelphij const boolean_t do_headroom_boost = *headroom_boost; 4877185029Spjd int try; 4878185029Spjd 4879185029Spjd ASSERT(dev->l2ad_vdev != NULL); 4880185029Spjd 4881251478Sdelphij /* Lower the flag now, we might want to raise it again later. */ 4882251478Sdelphij *headroom_boost = B_FALSE; 4883251478Sdelphij 4884185029Spjd pio = NULL; 4885251478Sdelphij write_sz = write_asize = write_psize = 0; 4886185029Spjd full = B_FALSE; 4887185029Spjd head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 4888185029Spjd head->b_flags |= ARC_L2_WRITE_HEAD; 4889185029Spjd 4890205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 4891185029Spjd /* 4892251478Sdelphij * We will want to try to compress buffers that are at least 2x the 4893251478Sdelphij * device sector size. 4894251478Sdelphij */ 4895251478Sdelphij buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 4896251478Sdelphij 4897251478Sdelphij /* 4898185029Spjd * Copy buffers for L2ARC writing. 4899185029Spjd */ 4900185029Spjd mutex_enter(&l2arc_buflist_mtx); 4901206796Spjd for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 4902251478Sdelphij uint64_t passed_sz = 0; 4903251478Sdelphij 4904185029Spjd list = l2arc_list_locked(try, &list_lock); 4905205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 4906185029Spjd 4907185029Spjd /* 4908185029Spjd * L2ARC fast warmup. 4909185029Spjd * 4910185029Spjd * Until the ARC is warm and starts to evict, read from the 4911185029Spjd * head of the ARC lists rather than the tail. 4912185029Spjd */ 4913185029Spjd if (arc_warm == B_FALSE) 4914185029Spjd ab = list_head(list); 4915185029Spjd else 4916185029Spjd ab = list_tail(list); 4917206796Spjd if (ab == NULL) 4918205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 4919185029Spjd 4920251478Sdelphij headroom = target_sz * l2arc_headroom; 4921251478Sdelphij if (do_headroom_boost) 4922251478Sdelphij headroom = (headroom * l2arc_headroom_boost) / 100; 4923251478Sdelphij 4924185029Spjd for (; ab; ab = ab_prev) { 4925251478Sdelphij l2arc_buf_hdr_t *l2hdr; 4926251478Sdelphij kmutex_t *hash_lock; 4927251478Sdelphij uint64_t buf_sz; 4928251478Sdelphij 4929185029Spjd if (arc_warm == B_FALSE) 4930185029Spjd ab_prev = list_next(list, ab); 4931185029Spjd else 4932185029Spjd ab_prev = list_prev(list, ab); 4933205231Skmacy ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size); 4934206796Spjd 4935185029Spjd hash_lock = HDR_LOCK(ab); 4936251478Sdelphij if (!mutex_tryenter(hash_lock)) { 4937205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 4938185029Spjd /* 4939185029Spjd * Skip this buffer rather than waiting. 4940185029Spjd */ 4941185029Spjd continue; 4942185029Spjd } 4943185029Spjd 4944185029Spjd passed_sz += ab->b_size; 4945185029Spjd if (passed_sz > headroom) { 4946185029Spjd /* 4947185029Spjd * Searched too far. 4948185029Spjd */ 4949185029Spjd mutex_exit(hash_lock); 4950205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 4951185029Spjd break; 4952185029Spjd } 4953185029Spjd 4954209962Smm if (!l2arc_write_eligible(guid, ab)) { 4955185029Spjd mutex_exit(hash_lock); 4956185029Spjd continue; 4957185029Spjd } 4958185029Spjd 4959185029Spjd if ((write_sz + ab->b_size) > target_sz) { 4960185029Spjd full = B_TRUE; 4961185029Spjd mutex_exit(hash_lock); 4962205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_full); 4963185029Spjd break; 4964185029Spjd } 4965185029Spjd 4966185029Spjd if (pio == NULL) { 4967185029Spjd /* 4968185029Spjd * Insert a dummy header on the buflist so 4969185029Spjd * l2arc_write_done() can find where the 4970185029Spjd * write buffers begin without searching. 4971185029Spjd */ 4972185029Spjd list_insert_head(dev->l2ad_buflist, head); 4973185029Spjd 4974185029Spjd cb = kmem_alloc( 4975185029Spjd sizeof (l2arc_write_callback_t), KM_SLEEP); 4976185029Spjd cb->l2wcb_dev = dev; 4977185029Spjd cb->l2wcb_head = head; 4978185029Spjd pio = zio_root(spa, l2arc_write_done, cb, 4979185029Spjd ZIO_FLAG_CANFAIL); 4980205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_pios); 4981185029Spjd } 4982185029Spjd 4983185029Spjd /* 4984185029Spjd * Create and add a new L2ARC header. 4985185029Spjd */ 4986251478Sdelphij l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 4987251478Sdelphij l2hdr->b_dev = dev; 4988251478Sdelphij ab->b_flags |= ARC_L2_WRITING; 4989185029Spjd 4990251478Sdelphij /* 4991251478Sdelphij * Temporarily stash the data buffer in b_tmp_cdata. 4992251478Sdelphij * The subsequent write step will pick it up from 4993251478Sdelphij * there. This is because can't access ab->b_buf 4994251478Sdelphij * without holding the hash_lock, which we in turn 4995251478Sdelphij * can't access without holding the ARC list locks 4996251478Sdelphij * (which we want to avoid during compression/writing). 4997251478Sdelphij */ 4998251478Sdelphij l2hdr->b_compress = ZIO_COMPRESS_OFF; 4999251478Sdelphij l2hdr->b_asize = ab->b_size; 5000251478Sdelphij l2hdr->b_tmp_cdata = ab->b_buf->b_data; 5001251478Sdelphij 5002185029Spjd buf_sz = ab->b_size; 5003251478Sdelphij ab->b_l2hdr = l2hdr; 5004185029Spjd 5005251478Sdelphij list_insert_head(dev->l2ad_buflist, ab); 5006251478Sdelphij 5007185029Spjd /* 5008185029Spjd * Compute and store the buffer cksum before 5009185029Spjd * writing. On debug the cksum is verified first. 5010185029Spjd */ 5011185029Spjd arc_cksum_verify(ab->b_buf); 5012185029Spjd arc_cksum_compute(ab->b_buf, B_TRUE); 5013185029Spjd 5014185029Spjd mutex_exit(hash_lock); 5015185029Spjd 5016251478Sdelphij write_sz += buf_sz; 5017251478Sdelphij } 5018251478Sdelphij 5019251478Sdelphij mutex_exit(list_lock); 5020251478Sdelphij 5021251478Sdelphij if (full == B_TRUE) 5022251478Sdelphij break; 5023251478Sdelphij } 5024251478Sdelphij 5025251478Sdelphij /* No buffers selected for writing? */ 5026251478Sdelphij if (pio == NULL) { 5027251478Sdelphij ASSERT0(write_sz); 5028251478Sdelphij mutex_exit(&l2arc_buflist_mtx); 5029251478Sdelphij kmem_cache_free(hdr_cache, head); 5030251478Sdelphij return (0); 5031251478Sdelphij } 5032251478Sdelphij 5033251478Sdelphij /* 5034251478Sdelphij * Now start writing the buffers. We're starting at the write head 5035251478Sdelphij * and work backwards, retracing the course of the buffer selector 5036251478Sdelphij * loop above. 5037251478Sdelphij */ 5038251478Sdelphij for (ab = list_prev(dev->l2ad_buflist, head); ab; 5039251478Sdelphij ab = list_prev(dev->l2ad_buflist, ab)) { 5040251478Sdelphij l2arc_buf_hdr_t *l2hdr; 5041251478Sdelphij uint64_t buf_sz; 5042251478Sdelphij 5043251478Sdelphij /* 5044251478Sdelphij * We shouldn't need to lock the buffer here, since we flagged 5045251478Sdelphij * it as ARC_L2_WRITING in the previous step, but we must take 5046251478Sdelphij * care to only access its L2 cache parameters. In particular, 5047251478Sdelphij * ab->b_buf may be invalid by now due to ARC eviction. 5048251478Sdelphij */ 5049251478Sdelphij l2hdr = ab->b_l2hdr; 5050251478Sdelphij l2hdr->b_daddr = dev->l2ad_hand; 5051251478Sdelphij 5052251478Sdelphij if ((ab->b_flags & ARC_L2COMPRESS) && 5053251478Sdelphij l2hdr->b_asize >= buf_compress_minsz) { 5054251478Sdelphij if (l2arc_compress_buf(l2hdr)) { 5055251478Sdelphij /* 5056251478Sdelphij * If compression succeeded, enable headroom 5057251478Sdelphij * boost on the next scan cycle. 5058251478Sdelphij */ 5059251478Sdelphij *headroom_boost = B_TRUE; 5060251478Sdelphij } 5061251478Sdelphij } 5062251478Sdelphij 5063251478Sdelphij /* 5064251478Sdelphij * Pick up the buffer data we had previously stashed away 5065251478Sdelphij * (and now potentially also compressed). 5066251478Sdelphij */ 5067251478Sdelphij buf_data = l2hdr->b_tmp_cdata; 5068251478Sdelphij buf_sz = l2hdr->b_asize; 5069251478Sdelphij 5070251478Sdelphij /* Compression may have squashed the buffer to zero length. */ 5071251478Sdelphij if (buf_sz != 0) { 5072251478Sdelphij uint64_t buf_p_sz; 5073251478Sdelphij 5074185029Spjd wzio = zio_write_phys(pio, dev->l2ad_vdev, 5075185029Spjd dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5076185029Spjd NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5077185029Spjd ZIO_FLAG_CANFAIL, B_FALSE); 5078185029Spjd 5079185029Spjd DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5080185029Spjd zio_t *, wzio); 5081185029Spjd (void) zio_nowait(wzio); 5082185029Spjd 5083251478Sdelphij write_asize += buf_sz; 5084185029Spjd /* 5085185029Spjd * Keep the clock hand suitably device-aligned. 5086185029Spjd */ 5087251478Sdelphij buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5088251478Sdelphij write_psize += buf_p_sz; 5089251478Sdelphij dev->l2ad_hand += buf_p_sz; 5090185029Spjd } 5091251478Sdelphij } 5092185029Spjd 5093185029Spjd mutex_exit(&l2arc_buflist_mtx); 5094185029Spjd 5095251478Sdelphij ASSERT3U(write_asize, <=, target_sz); 5096185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_sent); 5097251478Sdelphij ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5098185029Spjd ARCSTAT_INCR(arcstat_l2_size, write_sz); 5099251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, write_asize); 5100251478Sdelphij vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); 5101185029Spjd 5102185029Spjd /* 5103185029Spjd * Bump device hand to the device start if it is approaching the end. 5104185029Spjd * l2arc_evict() will already have evicted ahead for this case. 5105185029Spjd */ 5106185029Spjd if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5107219089Spjd vdev_space_update(dev->l2ad_vdev, 5108219089Spjd dev->l2ad_end - dev->l2ad_hand, 0, 0); 5109185029Spjd dev->l2ad_hand = dev->l2ad_start; 5110185029Spjd dev->l2ad_evict = dev->l2ad_start; 5111185029Spjd dev->l2ad_first = B_FALSE; 5112185029Spjd } 5113185029Spjd 5114208373Smm dev->l2ad_writing = B_TRUE; 5115185029Spjd (void) zio_wait(pio); 5116208373Smm dev->l2ad_writing = B_FALSE; 5117208373Smm 5118251478Sdelphij return (write_asize); 5119185029Spjd} 5120185029Spjd 5121185029Spjd/* 5122251478Sdelphij * Compresses an L2ARC buffer. 5123251478Sdelphij * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5124251478Sdelphij * size in l2hdr->b_asize. This routine tries to compress the data and 5125251478Sdelphij * depending on the compression result there are three possible outcomes: 5126251478Sdelphij * *) The buffer was incompressible. The original l2hdr contents were left 5127251478Sdelphij * untouched and are ready for writing to an L2 device. 5128251478Sdelphij * *) The buffer was all-zeros, so there is no need to write it to an L2 5129251478Sdelphij * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5130251478Sdelphij * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5131251478Sdelphij * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5132251478Sdelphij * data buffer which holds the compressed data to be written, and b_asize 5133251478Sdelphij * tells us how much data there is. b_compress is set to the appropriate 5134251478Sdelphij * compression algorithm. Once writing is done, invoke 5135251478Sdelphij * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5136251478Sdelphij * 5137251478Sdelphij * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5138251478Sdelphij * buffer was incompressible). 5139251478Sdelphij */ 5140251478Sdelphijstatic boolean_t 5141251478Sdelphijl2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5142251478Sdelphij{ 5143251478Sdelphij void *cdata; 5144251478Sdelphij size_t csize, len; 5145251478Sdelphij 5146251478Sdelphij ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5147251478Sdelphij ASSERT(l2hdr->b_tmp_cdata != NULL); 5148251478Sdelphij 5149251478Sdelphij len = l2hdr->b_asize; 5150251478Sdelphij cdata = zio_data_buf_alloc(len); 5151251478Sdelphij csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5152251478Sdelphij cdata, l2hdr->b_asize); 5153251478Sdelphij 5154251478Sdelphij if (csize == 0) { 5155251478Sdelphij /* zero block, indicate that there's nothing to write */ 5156251478Sdelphij zio_data_buf_free(cdata, len); 5157251478Sdelphij l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5158251478Sdelphij l2hdr->b_asize = 0; 5159251478Sdelphij l2hdr->b_tmp_cdata = NULL; 5160251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5161251478Sdelphij return (B_TRUE); 5162251478Sdelphij } else if (csize > 0 && csize < len) { 5163251478Sdelphij /* 5164251478Sdelphij * Compression succeeded, we'll keep the cdata around for 5165251478Sdelphij * writing and release it afterwards. 5166251478Sdelphij */ 5167251478Sdelphij l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5168251478Sdelphij l2hdr->b_asize = csize; 5169251478Sdelphij l2hdr->b_tmp_cdata = cdata; 5170251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_successes); 5171251478Sdelphij return (B_TRUE); 5172251478Sdelphij } else { 5173251478Sdelphij /* 5174251478Sdelphij * Compression failed, release the compressed buffer. 5175251478Sdelphij * l2hdr will be left unmodified. 5176251478Sdelphij */ 5177251478Sdelphij zio_data_buf_free(cdata, len); 5178251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_failures); 5179251478Sdelphij return (B_FALSE); 5180251478Sdelphij } 5181251478Sdelphij} 5182251478Sdelphij 5183251478Sdelphij/* 5184251478Sdelphij * Decompresses a zio read back from an l2arc device. On success, the 5185251478Sdelphij * underlying zio's io_data buffer is overwritten by the uncompressed 5186251478Sdelphij * version. On decompression error (corrupt compressed stream), the 5187251478Sdelphij * zio->io_error value is set to signal an I/O error. 5188251478Sdelphij * 5189251478Sdelphij * Please note that the compressed data stream is not checksummed, so 5190251478Sdelphij * if the underlying device is experiencing data corruption, we may feed 5191251478Sdelphij * corrupt data to the decompressor, so the decompressor needs to be 5192251478Sdelphij * able to handle this situation (LZ4 does). 5193251478Sdelphij */ 5194251478Sdelphijstatic void 5195251478Sdelphijl2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5196251478Sdelphij{ 5197251478Sdelphij ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5198251478Sdelphij 5199251478Sdelphij if (zio->io_error != 0) { 5200251478Sdelphij /* 5201251478Sdelphij * An io error has occured, just restore the original io 5202251478Sdelphij * size in preparation for a main pool read. 5203251478Sdelphij */ 5204251478Sdelphij zio->io_orig_size = zio->io_size = hdr->b_size; 5205251478Sdelphij return; 5206251478Sdelphij } 5207251478Sdelphij 5208251478Sdelphij if (c == ZIO_COMPRESS_EMPTY) { 5209251478Sdelphij /* 5210251478Sdelphij * An empty buffer results in a null zio, which means we 5211251478Sdelphij * need to fill its io_data after we're done restoring the 5212251478Sdelphij * buffer's contents. 5213251478Sdelphij */ 5214251478Sdelphij ASSERT(hdr->b_buf != NULL); 5215251478Sdelphij bzero(hdr->b_buf->b_data, hdr->b_size); 5216251478Sdelphij zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5217251478Sdelphij } else { 5218251478Sdelphij ASSERT(zio->io_data != NULL); 5219251478Sdelphij /* 5220251478Sdelphij * We copy the compressed data from the start of the arc buffer 5221251478Sdelphij * (the zio_read will have pulled in only what we need, the 5222251478Sdelphij * rest is garbage which we will overwrite at decompression) 5223251478Sdelphij * and then decompress back to the ARC data buffer. This way we 5224251478Sdelphij * can minimize copying by simply decompressing back over the 5225251478Sdelphij * original compressed data (rather than decompressing to an 5226251478Sdelphij * aux buffer and then copying back the uncompressed buffer, 5227251478Sdelphij * which is likely to be much larger). 5228251478Sdelphij */ 5229251478Sdelphij uint64_t csize; 5230251478Sdelphij void *cdata; 5231251478Sdelphij 5232251478Sdelphij csize = zio->io_size; 5233251478Sdelphij cdata = zio_data_buf_alloc(csize); 5234251478Sdelphij bcopy(zio->io_data, cdata, csize); 5235251478Sdelphij if (zio_decompress_data(c, cdata, zio->io_data, csize, 5236251478Sdelphij hdr->b_size) != 0) 5237251478Sdelphij zio->io_error = EIO; 5238251478Sdelphij zio_data_buf_free(cdata, csize); 5239251478Sdelphij } 5240251478Sdelphij 5241251478Sdelphij /* Restore the expected uncompressed IO size. */ 5242251478Sdelphij zio->io_orig_size = zio->io_size = hdr->b_size; 5243251478Sdelphij} 5244251478Sdelphij 5245251478Sdelphij/* 5246251478Sdelphij * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5247251478Sdelphij * This buffer serves as a temporary holder of compressed data while 5248251478Sdelphij * the buffer entry is being written to an l2arc device. Once that is 5249251478Sdelphij * done, we can dispose of it. 5250251478Sdelphij */ 5251251478Sdelphijstatic void 5252251478Sdelphijl2arc_release_cdata_buf(arc_buf_hdr_t *ab) 5253251478Sdelphij{ 5254251478Sdelphij l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr; 5255251478Sdelphij 5256251478Sdelphij if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) { 5257251478Sdelphij /* 5258251478Sdelphij * If the data was compressed, then we've allocated a 5259251478Sdelphij * temporary buffer for it, so now we need to release it. 5260251478Sdelphij */ 5261251478Sdelphij ASSERT(l2hdr->b_tmp_cdata != NULL); 5262251478Sdelphij zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size); 5263251478Sdelphij } 5264251478Sdelphij l2hdr->b_tmp_cdata = NULL; 5265251478Sdelphij} 5266251478Sdelphij 5267251478Sdelphij/* 5268185029Spjd * This thread feeds the L2ARC at regular intervals. This is the beating 5269185029Spjd * heart of the L2ARC. 5270185029Spjd */ 5271185029Spjdstatic void 5272185029Spjdl2arc_feed_thread(void *dummy __unused) 5273185029Spjd{ 5274185029Spjd callb_cpr_t cpr; 5275185029Spjd l2arc_dev_t *dev; 5276185029Spjd spa_t *spa; 5277208373Smm uint64_t size, wrote; 5278219089Spjd clock_t begin, next = ddi_get_lbolt(); 5279251478Sdelphij boolean_t headroom_boost = B_FALSE; 5280185029Spjd 5281185029Spjd CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5282185029Spjd 5283185029Spjd mutex_enter(&l2arc_feed_thr_lock); 5284185029Spjd 5285185029Spjd while (l2arc_thread_exit == 0) { 5286185029Spjd CALLB_CPR_SAFE_BEGIN(&cpr); 5287185029Spjd (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5288219089Spjd next - ddi_get_lbolt()); 5289185029Spjd CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5290219089Spjd next = ddi_get_lbolt() + hz; 5291185029Spjd 5292185029Spjd /* 5293185029Spjd * Quick check for L2ARC devices. 5294185029Spjd */ 5295185029Spjd mutex_enter(&l2arc_dev_mtx); 5296185029Spjd if (l2arc_ndev == 0) { 5297185029Spjd mutex_exit(&l2arc_dev_mtx); 5298185029Spjd continue; 5299185029Spjd } 5300185029Spjd mutex_exit(&l2arc_dev_mtx); 5301219089Spjd begin = ddi_get_lbolt(); 5302185029Spjd 5303185029Spjd /* 5304185029Spjd * This selects the next l2arc device to write to, and in 5305185029Spjd * doing so the next spa to feed from: dev->l2ad_spa. This 5306185029Spjd * will return NULL if there are now no l2arc devices or if 5307185029Spjd * they are all faulted. 5308185029Spjd * 5309185029Spjd * If a device is returned, its spa's config lock is also 5310185029Spjd * held to prevent device removal. l2arc_dev_get_next() 5311185029Spjd * will grab and release l2arc_dev_mtx. 5312185029Spjd */ 5313185029Spjd if ((dev = l2arc_dev_get_next()) == NULL) 5314185029Spjd continue; 5315185029Spjd 5316185029Spjd spa = dev->l2ad_spa; 5317185029Spjd ASSERT(spa != NULL); 5318185029Spjd 5319185029Spjd /* 5320219089Spjd * If the pool is read-only then force the feed thread to 5321219089Spjd * sleep a little longer. 5322219089Spjd */ 5323219089Spjd if (!spa_writeable(spa)) { 5324219089Spjd next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5325219089Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5326219089Spjd continue; 5327219089Spjd } 5328219089Spjd 5329219089Spjd /* 5330185029Spjd * Avoid contributing to memory pressure. 5331185029Spjd */ 5332185029Spjd if (arc_reclaim_needed()) { 5333185029Spjd ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5334185029Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5335185029Spjd continue; 5336185029Spjd } 5337185029Spjd 5338185029Spjd ARCSTAT_BUMP(arcstat_l2_feeds); 5339185029Spjd 5340251478Sdelphij size = l2arc_write_size(); 5341185029Spjd 5342185029Spjd /* 5343185029Spjd * Evict L2ARC buffers that will be overwritten. 5344185029Spjd */ 5345185029Spjd l2arc_evict(dev, size, B_FALSE); 5346185029Spjd 5347185029Spjd /* 5348185029Spjd * Write ARC buffers. 5349185029Spjd */ 5350251478Sdelphij wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5351208373Smm 5352208373Smm /* 5353208373Smm * Calculate interval between writes. 5354208373Smm */ 5355208373Smm next = l2arc_write_interval(begin, size, wrote); 5356185029Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5357185029Spjd } 5358185029Spjd 5359185029Spjd l2arc_thread_exit = 0; 5360185029Spjd cv_broadcast(&l2arc_feed_thr_cv); 5361185029Spjd CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5362185029Spjd thread_exit(); 5363185029Spjd} 5364185029Spjd 5365185029Spjdboolean_t 5366185029Spjdl2arc_vdev_present(vdev_t *vd) 5367185029Spjd{ 5368185029Spjd l2arc_dev_t *dev; 5369185029Spjd 5370185029Spjd mutex_enter(&l2arc_dev_mtx); 5371185029Spjd for (dev = list_head(l2arc_dev_list); dev != NULL; 5372185029Spjd dev = list_next(l2arc_dev_list, dev)) { 5373185029Spjd if (dev->l2ad_vdev == vd) 5374185029Spjd break; 5375185029Spjd } 5376185029Spjd mutex_exit(&l2arc_dev_mtx); 5377185029Spjd 5378185029Spjd return (dev != NULL); 5379185029Spjd} 5380185029Spjd 5381185029Spjd/* 5382185029Spjd * Add a vdev for use by the L2ARC. By this point the spa has already 5383185029Spjd * validated the vdev and opened it. 5384185029Spjd */ 5385185029Spjdvoid 5386219089Spjdl2arc_add_vdev(spa_t *spa, vdev_t *vd) 5387185029Spjd{ 5388185029Spjd l2arc_dev_t *adddev; 5389185029Spjd 5390185029Spjd ASSERT(!l2arc_vdev_present(vd)); 5391185029Spjd 5392185029Spjd /* 5393185029Spjd * Create a new l2arc device entry. 5394185029Spjd */ 5395185029Spjd adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5396185029Spjd adddev->l2ad_spa = spa; 5397185029Spjd adddev->l2ad_vdev = vd; 5398219089Spjd adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5399219089Spjd adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5400185029Spjd adddev->l2ad_hand = adddev->l2ad_start; 5401185029Spjd adddev->l2ad_evict = adddev->l2ad_start; 5402185029Spjd adddev->l2ad_first = B_TRUE; 5403208373Smm adddev->l2ad_writing = B_FALSE; 5404185029Spjd 5405185029Spjd /* 5406185029Spjd * This is a list of all ARC buffers that are still valid on the 5407185029Spjd * device. 5408185029Spjd */ 5409185029Spjd adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5410185029Spjd list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5411185029Spjd offsetof(arc_buf_hdr_t, b_l2node)); 5412185029Spjd 5413219089Spjd vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5414185029Spjd 5415185029Spjd /* 5416185029Spjd * Add device to global list 5417185029Spjd */ 5418185029Spjd mutex_enter(&l2arc_dev_mtx); 5419185029Spjd list_insert_head(l2arc_dev_list, adddev); 5420185029Spjd atomic_inc_64(&l2arc_ndev); 5421185029Spjd mutex_exit(&l2arc_dev_mtx); 5422185029Spjd} 5423185029Spjd 5424185029Spjd/* 5425185029Spjd * Remove a vdev from the L2ARC. 5426185029Spjd */ 5427185029Spjdvoid 5428185029Spjdl2arc_remove_vdev(vdev_t *vd) 5429185029Spjd{ 5430185029Spjd l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5431185029Spjd 5432185029Spjd /* 5433185029Spjd * Find the device by vdev 5434185029Spjd */ 5435185029Spjd mutex_enter(&l2arc_dev_mtx); 5436185029Spjd for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5437185029Spjd nextdev = list_next(l2arc_dev_list, dev); 5438185029Spjd if (vd == dev->l2ad_vdev) { 5439185029Spjd remdev = dev; 5440185029Spjd break; 5441185029Spjd } 5442185029Spjd } 5443185029Spjd ASSERT(remdev != NULL); 5444185029Spjd 5445185029Spjd /* 5446185029Spjd * Remove device from global list 5447185029Spjd */ 5448185029Spjd list_remove(l2arc_dev_list, remdev); 5449185029Spjd l2arc_dev_last = NULL; /* may have been invalidated */ 5450185029Spjd atomic_dec_64(&l2arc_ndev); 5451185029Spjd mutex_exit(&l2arc_dev_mtx); 5452185029Spjd 5453185029Spjd /* 5454185029Spjd * Clear all buflists and ARC references. L2ARC device flush. 5455185029Spjd */ 5456185029Spjd l2arc_evict(remdev, 0, B_TRUE); 5457185029Spjd list_destroy(remdev->l2ad_buflist); 5458185029Spjd kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5459185029Spjd kmem_free(remdev, sizeof (l2arc_dev_t)); 5460185029Spjd} 5461185029Spjd 5462185029Spjdvoid 5463185029Spjdl2arc_init(void) 5464185029Spjd{ 5465185029Spjd l2arc_thread_exit = 0; 5466185029Spjd l2arc_ndev = 0; 5467185029Spjd l2arc_writes_sent = 0; 5468185029Spjd l2arc_writes_done = 0; 5469185029Spjd 5470185029Spjd mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5471185029Spjd cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5472185029Spjd mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5473185029Spjd mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5474185029Spjd mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5475185029Spjd 5476185029Spjd l2arc_dev_list = &L2ARC_dev_list; 5477185029Spjd l2arc_free_on_write = &L2ARC_free_on_write; 5478185029Spjd list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5479185029Spjd offsetof(l2arc_dev_t, l2ad_node)); 5480185029Spjd list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5481185029Spjd offsetof(l2arc_data_free_t, l2df_list_node)); 5482185029Spjd} 5483185029Spjd 5484185029Spjdvoid 5485185029Spjdl2arc_fini(void) 5486185029Spjd{ 5487185029Spjd /* 5488185029Spjd * This is called from dmu_fini(), which is called from spa_fini(); 5489185029Spjd * Because of this, we can assume that all l2arc devices have 5490185029Spjd * already been removed when the pools themselves were removed. 5491185029Spjd */ 5492185029Spjd 5493185029Spjd l2arc_do_free_on_write(); 5494185029Spjd 5495185029Spjd mutex_destroy(&l2arc_feed_thr_lock); 5496185029Spjd cv_destroy(&l2arc_feed_thr_cv); 5497185029Spjd mutex_destroy(&l2arc_dev_mtx); 5498185029Spjd mutex_destroy(&l2arc_buflist_mtx); 5499185029Spjd mutex_destroy(&l2arc_free_on_write_mtx); 5500185029Spjd 5501185029Spjd list_destroy(l2arc_dev_list); 5502185029Spjd list_destroy(l2arc_free_on_write); 5503185029Spjd} 5504185029Spjd 5505185029Spjdvoid 5506185029Spjdl2arc_start(void) 5507185029Spjd{ 5508209962Smm if (!(spa_mode_global & FWRITE)) 5509185029Spjd return; 5510185029Spjd 5511185029Spjd (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5512185029Spjd TS_RUN, minclsyspri); 5513185029Spjd} 5514185029Spjd 5515185029Spjdvoid 5516185029Spjdl2arc_stop(void) 5517185029Spjd{ 5518209962Smm if (!(spa_mode_global & FWRITE)) 5519185029Spjd return; 5520185029Spjd 5521185029Spjd mutex_enter(&l2arc_feed_thr_lock); 5522185029Spjd cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5523185029Spjd l2arc_thread_exit = 1; 5524185029Spjd while (l2arc_thread_exit != 0) 5525185029Spjd cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5526185029Spjd mutex_exit(&l2arc_feed_thr_lock); 5527185029Spjd} 5528