arc.c revision 258388
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 61251631Sdelphij * elements of the cache are therefore 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 65251631Sdelphij * 128K bytes). We therefore 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 80251631Sdelphij * adjusting the cache use method 2. We therefore 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) \ 441251631Sdelphij 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 675251631Sdelphij/* L2ARC Performance Tunables */ 676185029Spjduint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 677185029Spjduint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 678185029Spjduint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 679251478Sdelphijuint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 680185029Spjduint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 681208373Smmuint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 682219089Spjdboolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 683208373Smmboolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 684208373Smmboolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 685185029Spjd 686217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 687205231Skmacy &l2arc_write_max, 0, "max write size"); 688217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 689205231Skmacy &l2arc_write_boost, 0, "extra write during warmup"); 690217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 691205231Skmacy &l2arc_headroom, 0, "number of dev writes"); 692217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 693205231Skmacy &l2arc_feed_secs, 0, "interval seconds"); 694217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 695208373Smm &l2arc_feed_min_ms, 0, "min interval milliseconds"); 696205231Skmacy 697205231SkmacySYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 698205231Skmacy &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 699208373SmmSYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 700208373Smm &l2arc_feed_again, 0, "turbo warmup"); 701208373SmmSYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 702208373Smm &l2arc_norw, 0, "no reads during writes"); 703205231Skmacy 704217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 705205231Skmacy &ARC_anon.arcs_size, 0, "size of anonymous state"); 706217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 707205231Skmacy &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 708217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 709205231Skmacy &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 710205231Skmacy 711217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 712205231Skmacy &ARC_mru.arcs_size, 0, "size of mru state"); 713217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 714205231Skmacy &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 715217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 716205231Skmacy &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 717205231Skmacy 718217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 719205231Skmacy &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 720217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 721205231Skmacy &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 722205231Skmacy "size of metadata in mru ghost state"); 723217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 724205231Skmacy &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 725205231Skmacy "size of data in mru ghost state"); 726205231Skmacy 727217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 728205231Skmacy &ARC_mfu.arcs_size, 0, "size of mfu state"); 729217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 730205231Skmacy &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 731217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 732205231Skmacy &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 733205231Skmacy 734217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 735205231Skmacy &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 736217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 737205231Skmacy &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 738205231Skmacy "size of metadata in mfu ghost state"); 739217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 740205231Skmacy &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 741205231Skmacy "size of data in mfu ghost state"); 742205231Skmacy 743217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 744205231Skmacy &ARC_l2c_only.arcs_size, 0, "size of mru state"); 745205231Skmacy 746185029Spjd/* 747185029Spjd * L2ARC Internals 748185029Spjd */ 749185029Spjdtypedef struct l2arc_dev { 750185029Spjd vdev_t *l2ad_vdev; /* vdev */ 751185029Spjd spa_t *l2ad_spa; /* spa */ 752185029Spjd uint64_t l2ad_hand; /* next write location */ 753185029Spjd uint64_t l2ad_start; /* first addr on device */ 754185029Spjd uint64_t l2ad_end; /* last addr on device */ 755185029Spjd uint64_t l2ad_evict; /* last addr eviction reached */ 756185029Spjd boolean_t l2ad_first; /* first sweep through */ 757208373Smm boolean_t l2ad_writing; /* currently writing */ 758185029Spjd list_t *l2ad_buflist; /* buffer list */ 759185029Spjd list_node_t l2ad_node; /* device list node */ 760185029Spjd} l2arc_dev_t; 761185029Spjd 762185029Spjdstatic list_t L2ARC_dev_list; /* device list */ 763185029Spjdstatic list_t *l2arc_dev_list; /* device list pointer */ 764185029Spjdstatic kmutex_t l2arc_dev_mtx; /* device list mutex */ 765185029Spjdstatic l2arc_dev_t *l2arc_dev_last; /* last device used */ 766185029Spjdstatic kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 767185029Spjdstatic list_t L2ARC_free_on_write; /* free after write buf list */ 768185029Spjdstatic list_t *l2arc_free_on_write; /* free after write list ptr */ 769185029Spjdstatic kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 770185029Spjdstatic uint64_t l2arc_ndev; /* number of devices */ 771185029Spjd 772185029Spjdtypedef struct l2arc_read_callback { 773251478Sdelphij arc_buf_t *l2rcb_buf; /* read buffer */ 774251478Sdelphij spa_t *l2rcb_spa; /* spa */ 775251478Sdelphij blkptr_t l2rcb_bp; /* original blkptr */ 776251478Sdelphij zbookmark_t l2rcb_zb; /* original bookmark */ 777251478Sdelphij int l2rcb_flags; /* original flags */ 778251478Sdelphij enum zio_compress l2rcb_compress; /* applied compress */ 779185029Spjd} l2arc_read_callback_t; 780185029Spjd 781185029Spjdtypedef struct l2arc_write_callback { 782185029Spjd l2arc_dev_t *l2wcb_dev; /* device info */ 783185029Spjd arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 784185029Spjd} l2arc_write_callback_t; 785185029Spjd 786185029Spjdstruct l2arc_buf_hdr { 787185029Spjd /* protected by arc_buf_hdr mutex */ 788251478Sdelphij l2arc_dev_t *b_dev; /* L2ARC device */ 789251478Sdelphij uint64_t b_daddr; /* disk address, offset byte */ 790251478Sdelphij /* compression applied to buffer data */ 791251478Sdelphij enum zio_compress b_compress; 792251478Sdelphij /* real alloc'd buffer size depending on b_compress applied */ 793251478Sdelphij int b_asize; 794251478Sdelphij /* temporary buffer holder for in-flight compressed data */ 795251478Sdelphij void *b_tmp_cdata; 796185029Spjd}; 797185029Spjd 798185029Spjdtypedef struct l2arc_data_free { 799185029Spjd /* protected by l2arc_free_on_write_mtx */ 800185029Spjd void *l2df_data; 801185029Spjd size_t l2df_size; 802185029Spjd void (*l2df_func)(void *, size_t); 803185029Spjd list_node_t l2df_list_node; 804185029Spjd} l2arc_data_free_t; 805185029Spjd 806185029Spjdstatic kmutex_t l2arc_feed_thr_lock; 807185029Spjdstatic kcondvar_t l2arc_feed_thr_cv; 808185029Spjdstatic uint8_t l2arc_thread_exit; 809185029Spjd 810185029Spjdstatic void l2arc_read_done(zio_t *zio); 811185029Spjdstatic void l2arc_hdr_stat_add(void); 812185029Spjdstatic void l2arc_hdr_stat_remove(void); 813185029Spjd 814251478Sdelphijstatic boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr); 815251478Sdelphijstatic void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, 816251478Sdelphij enum zio_compress c); 817251478Sdelphijstatic void l2arc_release_cdata_buf(arc_buf_hdr_t *ab); 818251478Sdelphij 819168404Spjdstatic uint64_t 820209962Smmbuf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 821168404Spjd{ 822168404Spjd uint8_t *vdva = (uint8_t *)dva; 823168404Spjd uint64_t crc = -1ULL; 824168404Spjd int i; 825168404Spjd 826168404Spjd ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 827168404Spjd 828168404Spjd for (i = 0; i < sizeof (dva_t); i++) 829168404Spjd crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 830168404Spjd 831209962Smm crc ^= (spa>>8) ^ birth; 832168404Spjd 833168404Spjd return (crc); 834168404Spjd} 835168404Spjd 836168404Spjd#define BUF_EMPTY(buf) \ 837168404Spjd ((buf)->b_dva.dva_word[0] == 0 && \ 838168404Spjd (buf)->b_dva.dva_word[1] == 0 && \ 839168404Spjd (buf)->b_birth == 0) 840168404Spjd 841168404Spjd#define BUF_EQUAL(spa, dva, birth, buf) \ 842168404Spjd ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 843168404Spjd ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 844168404Spjd ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 845168404Spjd 846219089Spjdstatic void 847219089Spjdbuf_discard_identity(arc_buf_hdr_t *hdr) 848219089Spjd{ 849219089Spjd hdr->b_dva.dva_word[0] = 0; 850219089Spjd hdr->b_dva.dva_word[1] = 0; 851219089Spjd hdr->b_birth = 0; 852219089Spjd hdr->b_cksum0 = 0; 853219089Spjd} 854219089Spjd 855168404Spjdstatic arc_buf_hdr_t * 856209962Smmbuf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp) 857168404Spjd{ 858168404Spjd uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 859168404Spjd kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 860168404Spjd arc_buf_hdr_t *buf; 861168404Spjd 862168404Spjd mutex_enter(hash_lock); 863168404Spjd for (buf = buf_hash_table.ht_table[idx]; buf != NULL; 864168404Spjd buf = buf->b_hash_next) { 865168404Spjd if (BUF_EQUAL(spa, dva, birth, buf)) { 866168404Spjd *lockp = hash_lock; 867168404Spjd return (buf); 868168404Spjd } 869168404Spjd } 870168404Spjd mutex_exit(hash_lock); 871168404Spjd *lockp = NULL; 872168404Spjd return (NULL); 873168404Spjd} 874168404Spjd 875168404Spjd/* 876168404Spjd * Insert an entry into the hash table. If there is already an element 877168404Spjd * equal to elem in the hash table, then the already existing element 878168404Spjd * will be returned and the new element will not be inserted. 879168404Spjd * Otherwise returns NULL. 880168404Spjd */ 881168404Spjdstatic arc_buf_hdr_t * 882168404Spjdbuf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp) 883168404Spjd{ 884168404Spjd uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 885168404Spjd kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 886168404Spjd arc_buf_hdr_t *fbuf; 887168404Spjd uint32_t i; 888168404Spjd 889168404Spjd ASSERT(!HDR_IN_HASH_TABLE(buf)); 890168404Spjd *lockp = hash_lock; 891168404Spjd mutex_enter(hash_lock); 892168404Spjd for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL; 893168404Spjd fbuf = fbuf->b_hash_next, i++) { 894168404Spjd if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf)) 895168404Spjd return (fbuf); 896168404Spjd } 897168404Spjd 898168404Spjd buf->b_hash_next = buf_hash_table.ht_table[idx]; 899168404Spjd buf_hash_table.ht_table[idx] = buf; 900168404Spjd buf->b_flags |= ARC_IN_HASH_TABLE; 901168404Spjd 902168404Spjd /* collect some hash table performance data */ 903168404Spjd if (i > 0) { 904168404Spjd ARCSTAT_BUMP(arcstat_hash_collisions); 905168404Spjd if (i == 1) 906168404Spjd ARCSTAT_BUMP(arcstat_hash_chains); 907168404Spjd 908168404Spjd ARCSTAT_MAX(arcstat_hash_chain_max, i); 909168404Spjd } 910168404Spjd 911168404Spjd ARCSTAT_BUMP(arcstat_hash_elements); 912168404Spjd ARCSTAT_MAXSTAT(arcstat_hash_elements); 913168404Spjd 914168404Spjd return (NULL); 915168404Spjd} 916168404Spjd 917168404Spjdstatic void 918168404Spjdbuf_hash_remove(arc_buf_hdr_t *buf) 919168404Spjd{ 920168404Spjd arc_buf_hdr_t *fbuf, **bufp; 921168404Spjd uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 922168404Spjd 923168404Spjd ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 924168404Spjd ASSERT(HDR_IN_HASH_TABLE(buf)); 925168404Spjd 926168404Spjd bufp = &buf_hash_table.ht_table[idx]; 927168404Spjd while ((fbuf = *bufp) != buf) { 928168404Spjd ASSERT(fbuf != NULL); 929168404Spjd bufp = &fbuf->b_hash_next; 930168404Spjd } 931168404Spjd *bufp = buf->b_hash_next; 932168404Spjd buf->b_hash_next = NULL; 933168404Spjd buf->b_flags &= ~ARC_IN_HASH_TABLE; 934168404Spjd 935168404Spjd /* collect some hash table performance data */ 936168404Spjd ARCSTAT_BUMPDOWN(arcstat_hash_elements); 937168404Spjd 938168404Spjd if (buf_hash_table.ht_table[idx] && 939168404Spjd buf_hash_table.ht_table[idx]->b_hash_next == NULL) 940168404Spjd ARCSTAT_BUMPDOWN(arcstat_hash_chains); 941168404Spjd} 942168404Spjd 943168404Spjd/* 944168404Spjd * Global data structures and functions for the buf kmem cache. 945168404Spjd */ 946168404Spjdstatic kmem_cache_t *hdr_cache; 947168404Spjdstatic kmem_cache_t *buf_cache; 948168404Spjd 949168404Spjdstatic void 950168404Spjdbuf_fini(void) 951168404Spjd{ 952168404Spjd int i; 953168404Spjd 954168404Spjd kmem_free(buf_hash_table.ht_table, 955168404Spjd (buf_hash_table.ht_mask + 1) * sizeof (void *)); 956168404Spjd for (i = 0; i < BUF_LOCKS; i++) 957168404Spjd mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 958168404Spjd kmem_cache_destroy(hdr_cache); 959168404Spjd kmem_cache_destroy(buf_cache); 960168404Spjd} 961168404Spjd 962168404Spjd/* 963168404Spjd * Constructor callback - called when the cache is empty 964168404Spjd * and a new buf is requested. 965168404Spjd */ 966168404Spjd/* ARGSUSED */ 967168404Spjdstatic int 968168404Spjdhdr_cons(void *vbuf, void *unused, int kmflag) 969168404Spjd{ 970168404Spjd arc_buf_hdr_t *buf = vbuf; 971168404Spjd 972168404Spjd bzero(buf, sizeof (arc_buf_hdr_t)); 973168404Spjd refcount_create(&buf->b_refcnt); 974168404Spjd cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL); 975185029Spjd mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 976208373Smm arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 977185029Spjd 978168404Spjd return (0); 979168404Spjd} 980168404Spjd 981185029Spjd/* ARGSUSED */ 982185029Spjdstatic int 983185029Spjdbuf_cons(void *vbuf, void *unused, int kmflag) 984185029Spjd{ 985185029Spjd arc_buf_t *buf = vbuf; 986185029Spjd 987185029Spjd bzero(buf, sizeof (arc_buf_t)); 988219089Spjd mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 989208373Smm arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 990208373Smm 991185029Spjd return (0); 992185029Spjd} 993185029Spjd 994168404Spjd/* 995168404Spjd * Destructor callback - called when a cached buf is 996168404Spjd * no longer required. 997168404Spjd */ 998168404Spjd/* ARGSUSED */ 999168404Spjdstatic void 1000168404Spjdhdr_dest(void *vbuf, void *unused) 1001168404Spjd{ 1002168404Spjd arc_buf_hdr_t *buf = vbuf; 1003168404Spjd 1004219089Spjd ASSERT(BUF_EMPTY(buf)); 1005168404Spjd refcount_destroy(&buf->b_refcnt); 1006168404Spjd cv_destroy(&buf->b_cv); 1007185029Spjd mutex_destroy(&buf->b_freeze_lock); 1008208373Smm arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1009168404Spjd} 1010168404Spjd 1011185029Spjd/* ARGSUSED */ 1012185029Spjdstatic void 1013185029Spjdbuf_dest(void *vbuf, void *unused) 1014185029Spjd{ 1015185029Spjd arc_buf_t *buf = vbuf; 1016185029Spjd 1017219089Spjd mutex_destroy(&buf->b_evict_lock); 1018208373Smm arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1019185029Spjd} 1020185029Spjd 1021168404Spjd/* 1022168404Spjd * Reclaim callback -- invoked when memory is low. 1023168404Spjd */ 1024168404Spjd/* ARGSUSED */ 1025168404Spjdstatic void 1026168404Spjdhdr_recl(void *unused) 1027168404Spjd{ 1028168404Spjd dprintf("hdr_recl called\n"); 1029168404Spjd /* 1030168404Spjd * umem calls the reclaim func when we destroy the buf cache, 1031168404Spjd * which is after we do arc_fini(). 1032168404Spjd */ 1033168404Spjd if (!arc_dead) 1034168404Spjd cv_signal(&arc_reclaim_thr_cv); 1035168404Spjd} 1036168404Spjd 1037168404Spjdstatic void 1038168404Spjdbuf_init(void) 1039168404Spjd{ 1040168404Spjd uint64_t *ct; 1041168404Spjd uint64_t hsize = 1ULL << 12; 1042168404Spjd int i, j; 1043168404Spjd 1044168404Spjd /* 1045168404Spjd * The hash table is big enough to fill all of physical memory 1046168404Spjd * with an average 64K block size. The table will take up 1047168404Spjd * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers). 1048168404Spjd */ 1049168696Spjd while (hsize * 65536 < (uint64_t)physmem * PAGESIZE) 1050168404Spjd hsize <<= 1; 1051168404Spjdretry: 1052168404Spjd buf_hash_table.ht_mask = hsize - 1; 1053168404Spjd buf_hash_table.ht_table = 1054168404Spjd kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1055168404Spjd if (buf_hash_table.ht_table == NULL) { 1056168404Spjd ASSERT(hsize > (1ULL << 8)); 1057168404Spjd hsize >>= 1; 1058168404Spjd goto retry; 1059168404Spjd } 1060168404Spjd 1061168404Spjd hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1062168404Spjd 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1063168404Spjd buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1064185029Spjd 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1065168404Spjd 1066168404Spjd for (i = 0; i < 256; i++) 1067168404Spjd for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1068168404Spjd *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1069168404Spjd 1070168404Spjd for (i = 0; i < BUF_LOCKS; i++) { 1071168404Spjd mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1072168404Spjd NULL, MUTEX_DEFAULT, NULL); 1073168404Spjd } 1074168404Spjd} 1075168404Spjd 1076168404Spjd#define ARC_MINTIME (hz>>4) /* 62 ms */ 1077168404Spjd 1078168404Spjdstatic void 1079168404Spjdarc_cksum_verify(arc_buf_t *buf) 1080168404Spjd{ 1081168404Spjd zio_cksum_t zc; 1082168404Spjd 1083168404Spjd if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1084168404Spjd return; 1085168404Spjd 1086168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1087168404Spjd if (buf->b_hdr->b_freeze_cksum == NULL || 1088168404Spjd (buf->b_hdr->b_flags & ARC_IO_ERROR)) { 1089168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1090168404Spjd return; 1091168404Spjd } 1092168404Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1093168404Spjd if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1094168404Spjd panic("buffer modified while frozen!"); 1095168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1096168404Spjd} 1097168404Spjd 1098185029Spjdstatic int 1099185029Spjdarc_cksum_equal(arc_buf_t *buf) 1100185029Spjd{ 1101185029Spjd zio_cksum_t zc; 1102185029Spjd int equal; 1103185029Spjd 1104185029Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1105185029Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1106185029Spjd equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1107185029Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1108185029Spjd 1109185029Spjd return (equal); 1110185029Spjd} 1111185029Spjd 1112168404Spjdstatic void 1113185029Spjdarc_cksum_compute(arc_buf_t *buf, boolean_t force) 1114168404Spjd{ 1115185029Spjd if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1116168404Spjd return; 1117168404Spjd 1118168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1119168404Spjd if (buf->b_hdr->b_freeze_cksum != NULL) { 1120168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1121168404Spjd return; 1122168404Spjd } 1123168404Spjd buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1124168404Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1125168404Spjd buf->b_hdr->b_freeze_cksum); 1126168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1127240133Smm#ifdef illumos 1128240133Smm arc_buf_watch(buf); 1129240133Smm#endif /* illumos */ 1130168404Spjd} 1131168404Spjd 1132240133Smm#ifdef illumos 1133240133Smm#ifndef _KERNEL 1134240133Smmtypedef struct procctl { 1135240133Smm long cmd; 1136240133Smm prwatch_t prwatch; 1137240133Smm} procctl_t; 1138240133Smm#endif 1139240133Smm 1140240133Smm/* ARGSUSED */ 1141240133Smmstatic void 1142240133Smmarc_buf_unwatch(arc_buf_t *buf) 1143240133Smm{ 1144240133Smm#ifndef _KERNEL 1145240133Smm if (arc_watch) { 1146240133Smm int result; 1147240133Smm procctl_t ctl; 1148240133Smm ctl.cmd = PCWATCH; 1149240133Smm ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1150240133Smm ctl.prwatch.pr_size = 0; 1151240133Smm ctl.prwatch.pr_wflags = 0; 1152240133Smm result = write(arc_procfd, &ctl, sizeof (ctl)); 1153240133Smm ASSERT3U(result, ==, sizeof (ctl)); 1154240133Smm } 1155240133Smm#endif 1156240133Smm} 1157240133Smm 1158240133Smm/* ARGSUSED */ 1159240133Smmstatic void 1160240133Smmarc_buf_watch(arc_buf_t *buf) 1161240133Smm{ 1162240133Smm#ifndef _KERNEL 1163240133Smm if (arc_watch) { 1164240133Smm int result; 1165240133Smm procctl_t ctl; 1166240133Smm ctl.cmd = PCWATCH; 1167240133Smm ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1168240133Smm ctl.prwatch.pr_size = buf->b_hdr->b_size; 1169240133Smm ctl.prwatch.pr_wflags = WA_WRITE; 1170240133Smm result = write(arc_procfd, &ctl, sizeof (ctl)); 1171240133Smm ASSERT3U(result, ==, sizeof (ctl)); 1172240133Smm } 1173240133Smm#endif 1174240133Smm} 1175240133Smm#endif /* illumos */ 1176240133Smm 1177168404Spjdvoid 1178168404Spjdarc_buf_thaw(arc_buf_t *buf) 1179168404Spjd{ 1180185029Spjd if (zfs_flags & ZFS_DEBUG_MODIFY) { 1181185029Spjd if (buf->b_hdr->b_state != arc_anon) 1182185029Spjd panic("modifying non-anon buffer!"); 1183185029Spjd if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS) 1184185029Spjd panic("modifying buffer while i/o in progress!"); 1185185029Spjd arc_cksum_verify(buf); 1186185029Spjd } 1187168404Spjd 1188168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1189168404Spjd if (buf->b_hdr->b_freeze_cksum != NULL) { 1190168404Spjd kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1191168404Spjd buf->b_hdr->b_freeze_cksum = NULL; 1192168404Spjd } 1193219089Spjd 1194219089Spjd if (zfs_flags & ZFS_DEBUG_MODIFY) { 1195219089Spjd if (buf->b_hdr->b_thawed) 1196219089Spjd kmem_free(buf->b_hdr->b_thawed, 1); 1197219089Spjd buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1198219089Spjd } 1199219089Spjd 1200168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1201240133Smm 1202240133Smm#ifdef illumos 1203240133Smm arc_buf_unwatch(buf); 1204240133Smm#endif /* illumos */ 1205168404Spjd} 1206168404Spjd 1207168404Spjdvoid 1208168404Spjdarc_buf_freeze(arc_buf_t *buf) 1209168404Spjd{ 1210219089Spjd kmutex_t *hash_lock; 1211219089Spjd 1212168404Spjd if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1213168404Spjd return; 1214168404Spjd 1215219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 1216219089Spjd mutex_enter(hash_lock); 1217219089Spjd 1218168404Spjd ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1219168404Spjd buf->b_hdr->b_state == arc_anon); 1220185029Spjd arc_cksum_compute(buf, B_FALSE); 1221219089Spjd mutex_exit(hash_lock); 1222240133Smm 1223168404Spjd} 1224168404Spjd 1225168404Spjdstatic void 1226205231Skmacyget_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock) 1227205231Skmacy{ 1228205231Skmacy uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth); 1229205231Skmacy 1230206796Spjd if (ab->b_type == ARC_BUFC_METADATA) 1231206796Spjd buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1232205231Skmacy else { 1233206796Spjd buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1234205231Skmacy buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1235205231Skmacy } 1236205231Skmacy 1237205231Skmacy *list = &state->arcs_lists[buf_hashid]; 1238205231Skmacy *lock = ARCS_LOCK(state, buf_hashid); 1239205231Skmacy} 1240205231Skmacy 1241205231Skmacy 1242205231Skmacystatic void 1243168404Spjdadd_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1244168404Spjd{ 1245168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 1246168404Spjd 1247168404Spjd if ((refcount_add(&ab->b_refcnt, tag) == 1) && 1248168404Spjd (ab->b_state != arc_anon)) { 1249206796Spjd uint64_t delta = ab->b_size * ab->b_datacnt; 1250206796Spjd uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type]; 1251205231Skmacy list_t *list; 1252205231Skmacy kmutex_t *lock; 1253168404Spjd 1254205231Skmacy get_buf_info(ab, ab->b_state, &list, &lock); 1255205231Skmacy ASSERT(!MUTEX_HELD(lock)); 1256205231Skmacy mutex_enter(lock); 1257168404Spjd ASSERT(list_link_active(&ab->b_arc_node)); 1258185029Spjd list_remove(list, ab); 1259168404Spjd if (GHOST_STATE(ab->b_state)) { 1260240415Smm ASSERT0(ab->b_datacnt); 1261168404Spjd ASSERT3P(ab->b_buf, ==, NULL); 1262168404Spjd delta = ab->b_size; 1263168404Spjd } 1264168404Spjd ASSERT(delta > 0); 1265185029Spjd ASSERT3U(*size, >=, delta); 1266185029Spjd atomic_add_64(size, -delta); 1267206794Spjd mutex_exit(lock); 1268185029Spjd /* remove the prefetch flag if we get a reference */ 1269168404Spjd if (ab->b_flags & ARC_PREFETCH) 1270168404Spjd ab->b_flags &= ~ARC_PREFETCH; 1271168404Spjd } 1272168404Spjd} 1273168404Spjd 1274168404Spjdstatic int 1275168404Spjdremove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1276168404Spjd{ 1277168404Spjd int cnt; 1278168404Spjd arc_state_t *state = ab->b_state; 1279168404Spjd 1280168404Spjd ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1281168404Spjd ASSERT(!GHOST_STATE(state)); 1282168404Spjd 1283168404Spjd if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) && 1284168404Spjd (state != arc_anon)) { 1285185029Spjd uint64_t *size = &state->arcs_lsize[ab->b_type]; 1286205231Skmacy list_t *list; 1287205231Skmacy kmutex_t *lock; 1288185029Spjd 1289205231Skmacy get_buf_info(ab, state, &list, &lock); 1290205231Skmacy ASSERT(!MUTEX_HELD(lock)); 1291205231Skmacy mutex_enter(lock); 1292168404Spjd ASSERT(!list_link_active(&ab->b_arc_node)); 1293205231Skmacy list_insert_head(list, ab); 1294168404Spjd ASSERT(ab->b_datacnt > 0); 1295185029Spjd atomic_add_64(size, ab->b_size * ab->b_datacnt); 1296206794Spjd mutex_exit(lock); 1297168404Spjd } 1298168404Spjd return (cnt); 1299168404Spjd} 1300168404Spjd 1301168404Spjd/* 1302168404Spjd * Move the supplied buffer to the indicated state. The mutex 1303168404Spjd * for the buffer must be held by the caller. 1304168404Spjd */ 1305168404Spjdstatic void 1306168404Spjdarc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock) 1307168404Spjd{ 1308168404Spjd arc_state_t *old_state = ab->b_state; 1309168404Spjd int64_t refcnt = refcount_count(&ab->b_refcnt); 1310168404Spjd uint64_t from_delta, to_delta; 1311205231Skmacy list_t *list; 1312205231Skmacy kmutex_t *lock; 1313168404Spjd 1314168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 1315168404Spjd ASSERT(new_state != old_state); 1316168404Spjd ASSERT(refcnt == 0 || ab->b_datacnt > 0); 1317168404Spjd ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state)); 1318219089Spjd ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon); 1319168404Spjd 1320168404Spjd from_delta = to_delta = ab->b_datacnt * ab->b_size; 1321168404Spjd 1322168404Spjd /* 1323168404Spjd * If this buffer is evictable, transfer it from the 1324168404Spjd * old state list to the new state list. 1325168404Spjd */ 1326168404Spjd if (refcnt == 0) { 1327168404Spjd if (old_state != arc_anon) { 1328205231Skmacy int use_mutex; 1329185029Spjd uint64_t *size = &old_state->arcs_lsize[ab->b_type]; 1330168404Spjd 1331205231Skmacy get_buf_info(ab, old_state, &list, &lock); 1332205231Skmacy use_mutex = !MUTEX_HELD(lock); 1333168404Spjd if (use_mutex) 1334205231Skmacy mutex_enter(lock); 1335168404Spjd 1336168404Spjd ASSERT(list_link_active(&ab->b_arc_node)); 1337205231Skmacy list_remove(list, ab); 1338168404Spjd 1339168404Spjd /* 1340168404Spjd * If prefetching out of the ghost cache, 1341219089Spjd * we will have a non-zero datacnt. 1342168404Spjd */ 1343168404Spjd if (GHOST_STATE(old_state) && ab->b_datacnt == 0) { 1344168404Spjd /* ghost elements have a ghost size */ 1345168404Spjd ASSERT(ab->b_buf == NULL); 1346168404Spjd from_delta = ab->b_size; 1347168404Spjd } 1348185029Spjd ASSERT3U(*size, >=, from_delta); 1349185029Spjd atomic_add_64(size, -from_delta); 1350168404Spjd 1351168404Spjd if (use_mutex) 1352205231Skmacy mutex_exit(lock); 1353168404Spjd } 1354168404Spjd if (new_state != arc_anon) { 1355206796Spjd int use_mutex; 1356185029Spjd uint64_t *size = &new_state->arcs_lsize[ab->b_type]; 1357168404Spjd 1358205231Skmacy get_buf_info(ab, new_state, &list, &lock); 1359205231Skmacy use_mutex = !MUTEX_HELD(lock); 1360168404Spjd if (use_mutex) 1361205231Skmacy mutex_enter(lock); 1362168404Spjd 1363205231Skmacy list_insert_head(list, ab); 1364168404Spjd 1365168404Spjd /* ghost elements have a ghost size */ 1366168404Spjd if (GHOST_STATE(new_state)) { 1367168404Spjd ASSERT(ab->b_datacnt == 0); 1368168404Spjd ASSERT(ab->b_buf == NULL); 1369168404Spjd to_delta = ab->b_size; 1370168404Spjd } 1371185029Spjd atomic_add_64(size, to_delta); 1372168404Spjd 1373168404Spjd if (use_mutex) 1374205231Skmacy mutex_exit(lock); 1375168404Spjd } 1376168404Spjd } 1377168404Spjd 1378168404Spjd ASSERT(!BUF_EMPTY(ab)); 1379219089Spjd if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab)) 1380168404Spjd buf_hash_remove(ab); 1381168404Spjd 1382168404Spjd /* adjust state sizes */ 1383168404Spjd if (to_delta) 1384168404Spjd atomic_add_64(&new_state->arcs_size, to_delta); 1385168404Spjd if (from_delta) { 1386168404Spjd ASSERT3U(old_state->arcs_size, >=, from_delta); 1387168404Spjd atomic_add_64(&old_state->arcs_size, -from_delta); 1388168404Spjd } 1389168404Spjd ab->b_state = new_state; 1390185029Spjd 1391185029Spjd /* adjust l2arc hdr stats */ 1392185029Spjd if (new_state == arc_l2c_only) 1393185029Spjd l2arc_hdr_stat_add(); 1394185029Spjd else if (old_state == arc_l2c_only) 1395185029Spjd l2arc_hdr_stat_remove(); 1396168404Spjd} 1397168404Spjd 1398185029Spjdvoid 1399208373Smmarc_space_consume(uint64_t space, arc_space_type_t type) 1400185029Spjd{ 1401208373Smm ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1402208373Smm 1403208373Smm switch (type) { 1404208373Smm case ARC_SPACE_DATA: 1405208373Smm ARCSTAT_INCR(arcstat_data_size, space); 1406208373Smm break; 1407208373Smm case ARC_SPACE_OTHER: 1408208373Smm ARCSTAT_INCR(arcstat_other_size, space); 1409208373Smm break; 1410208373Smm case ARC_SPACE_HDRS: 1411208373Smm ARCSTAT_INCR(arcstat_hdr_size, space); 1412208373Smm break; 1413208373Smm case ARC_SPACE_L2HDRS: 1414208373Smm ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1415208373Smm break; 1416208373Smm } 1417208373Smm 1418185029Spjd atomic_add_64(&arc_meta_used, space); 1419185029Spjd atomic_add_64(&arc_size, space); 1420185029Spjd} 1421185029Spjd 1422185029Spjdvoid 1423208373Smmarc_space_return(uint64_t space, arc_space_type_t type) 1424185029Spjd{ 1425208373Smm ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1426208373Smm 1427208373Smm switch (type) { 1428208373Smm case ARC_SPACE_DATA: 1429208373Smm ARCSTAT_INCR(arcstat_data_size, -space); 1430208373Smm break; 1431208373Smm case ARC_SPACE_OTHER: 1432208373Smm ARCSTAT_INCR(arcstat_other_size, -space); 1433208373Smm break; 1434208373Smm case ARC_SPACE_HDRS: 1435208373Smm ARCSTAT_INCR(arcstat_hdr_size, -space); 1436208373Smm break; 1437208373Smm case ARC_SPACE_L2HDRS: 1438208373Smm ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1439208373Smm break; 1440208373Smm } 1441208373Smm 1442185029Spjd ASSERT(arc_meta_used >= space); 1443185029Spjd if (arc_meta_max < arc_meta_used) 1444185029Spjd arc_meta_max = arc_meta_used; 1445185029Spjd atomic_add_64(&arc_meta_used, -space); 1446185029Spjd ASSERT(arc_size >= space); 1447185029Spjd atomic_add_64(&arc_size, -space); 1448185029Spjd} 1449185029Spjd 1450185029Spjdvoid * 1451185029Spjdarc_data_buf_alloc(uint64_t size) 1452185029Spjd{ 1453185029Spjd if (arc_evict_needed(ARC_BUFC_DATA)) 1454185029Spjd cv_signal(&arc_reclaim_thr_cv); 1455185029Spjd atomic_add_64(&arc_size, size); 1456185029Spjd return (zio_data_buf_alloc(size)); 1457185029Spjd} 1458185029Spjd 1459185029Spjdvoid 1460185029Spjdarc_data_buf_free(void *buf, uint64_t size) 1461185029Spjd{ 1462185029Spjd zio_data_buf_free(buf, size); 1463185029Spjd ASSERT(arc_size >= size); 1464185029Spjd atomic_add_64(&arc_size, -size); 1465185029Spjd} 1466185029Spjd 1467168404Spjdarc_buf_t * 1468168404Spjdarc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1469168404Spjd{ 1470168404Spjd arc_buf_hdr_t *hdr; 1471168404Spjd arc_buf_t *buf; 1472168404Spjd 1473168404Spjd ASSERT3U(size, >, 0); 1474185029Spjd hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1475168404Spjd ASSERT(BUF_EMPTY(hdr)); 1476168404Spjd hdr->b_size = size; 1477168404Spjd hdr->b_type = type; 1478228103Smm hdr->b_spa = spa_load_guid(spa); 1479168404Spjd hdr->b_state = arc_anon; 1480168404Spjd hdr->b_arc_access = 0; 1481185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1482168404Spjd buf->b_hdr = hdr; 1483168404Spjd buf->b_data = NULL; 1484168404Spjd buf->b_efunc = NULL; 1485168404Spjd buf->b_private = NULL; 1486168404Spjd buf->b_next = NULL; 1487168404Spjd hdr->b_buf = buf; 1488168404Spjd arc_get_data_buf(buf); 1489168404Spjd hdr->b_datacnt = 1; 1490168404Spjd hdr->b_flags = 0; 1491168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1492168404Spjd (void) refcount_add(&hdr->b_refcnt, tag); 1493168404Spjd 1494168404Spjd return (buf); 1495168404Spjd} 1496168404Spjd 1497209962Smmstatic char *arc_onloan_tag = "onloan"; 1498209962Smm 1499209962Smm/* 1500209962Smm * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1501209962Smm * flight data by arc_tempreserve_space() until they are "returned". Loaned 1502209962Smm * buffers must be returned to the arc before they can be used by the DMU or 1503209962Smm * freed. 1504209962Smm */ 1505209962Smmarc_buf_t * 1506209962Smmarc_loan_buf(spa_t *spa, int size) 1507209962Smm{ 1508209962Smm arc_buf_t *buf; 1509209962Smm 1510209962Smm buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1511209962Smm 1512209962Smm atomic_add_64(&arc_loaned_bytes, size); 1513209962Smm return (buf); 1514209962Smm} 1515209962Smm 1516209962Smm/* 1517209962Smm * Return a loaned arc buffer to the arc. 1518209962Smm */ 1519209962Smmvoid 1520209962Smmarc_return_buf(arc_buf_t *buf, void *tag) 1521209962Smm{ 1522209962Smm arc_buf_hdr_t *hdr = buf->b_hdr; 1523209962Smm 1524209962Smm ASSERT(buf->b_data != NULL); 1525219089Spjd (void) refcount_add(&hdr->b_refcnt, tag); 1526219089Spjd (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1527209962Smm 1528209962Smm atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1529209962Smm} 1530209962Smm 1531219089Spjd/* Detach an arc_buf from a dbuf (tag) */ 1532219089Spjdvoid 1533219089Spjdarc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1534219089Spjd{ 1535219089Spjd arc_buf_hdr_t *hdr; 1536219089Spjd 1537219089Spjd ASSERT(buf->b_data != NULL); 1538219089Spjd hdr = buf->b_hdr; 1539219089Spjd (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1540219089Spjd (void) refcount_remove(&hdr->b_refcnt, tag); 1541219089Spjd buf->b_efunc = NULL; 1542219089Spjd buf->b_private = NULL; 1543219089Spjd 1544219089Spjd atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1545219089Spjd} 1546219089Spjd 1547168404Spjdstatic arc_buf_t * 1548168404Spjdarc_buf_clone(arc_buf_t *from) 1549168404Spjd{ 1550168404Spjd arc_buf_t *buf; 1551168404Spjd arc_buf_hdr_t *hdr = from->b_hdr; 1552168404Spjd uint64_t size = hdr->b_size; 1553168404Spjd 1554219089Spjd ASSERT(hdr->b_state != arc_anon); 1555219089Spjd 1556185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1557168404Spjd buf->b_hdr = hdr; 1558168404Spjd buf->b_data = NULL; 1559168404Spjd buf->b_efunc = NULL; 1560168404Spjd buf->b_private = NULL; 1561168404Spjd buf->b_next = hdr->b_buf; 1562168404Spjd hdr->b_buf = buf; 1563168404Spjd arc_get_data_buf(buf); 1564168404Spjd bcopy(from->b_data, buf->b_data, size); 1565242845Sdelphij 1566242845Sdelphij /* 1567242845Sdelphij * This buffer already exists in the arc so create a duplicate 1568242845Sdelphij * copy for the caller. If the buffer is associated with user data 1569242845Sdelphij * then track the size and number of duplicates. These stats will be 1570242845Sdelphij * updated as duplicate buffers are created and destroyed. 1571242845Sdelphij */ 1572242845Sdelphij if (hdr->b_type == ARC_BUFC_DATA) { 1573242845Sdelphij ARCSTAT_BUMP(arcstat_duplicate_buffers); 1574242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1575242845Sdelphij } 1576168404Spjd hdr->b_datacnt += 1; 1577168404Spjd return (buf); 1578168404Spjd} 1579168404Spjd 1580168404Spjdvoid 1581168404Spjdarc_buf_add_ref(arc_buf_t *buf, void* tag) 1582168404Spjd{ 1583168404Spjd arc_buf_hdr_t *hdr; 1584168404Spjd kmutex_t *hash_lock; 1585168404Spjd 1586168404Spjd /* 1587185029Spjd * Check to see if this buffer is evicted. Callers 1588185029Spjd * must verify b_data != NULL to know if the add_ref 1589185029Spjd * was successful. 1590168404Spjd */ 1591219089Spjd mutex_enter(&buf->b_evict_lock); 1592185029Spjd if (buf->b_data == NULL) { 1593219089Spjd mutex_exit(&buf->b_evict_lock); 1594168404Spjd return; 1595168404Spjd } 1596219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 1597219089Spjd mutex_enter(hash_lock); 1598185029Spjd hdr = buf->b_hdr; 1599219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1600219089Spjd mutex_exit(&buf->b_evict_lock); 1601168404Spjd 1602168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1603168404Spjd add_reference(hdr, hash_lock, tag); 1604208373Smm DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1605168404Spjd arc_access(hdr, hash_lock); 1606168404Spjd mutex_exit(hash_lock); 1607168404Spjd ARCSTAT_BUMP(arcstat_hits); 1608168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 1609168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1610168404Spjd data, metadata, hits); 1611168404Spjd} 1612168404Spjd 1613185029Spjd/* 1614185029Spjd * Free the arc data buffer. If it is an l2arc write in progress, 1615185029Spjd * the buffer is placed on l2arc_free_on_write to be freed later. 1616185029Spjd */ 1617168404Spjdstatic void 1618240133Smmarc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1619185029Spjd{ 1620240133Smm arc_buf_hdr_t *hdr = buf->b_hdr; 1621240133Smm 1622185029Spjd if (HDR_L2_WRITING(hdr)) { 1623185029Spjd l2arc_data_free_t *df; 1624185029Spjd df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1625240133Smm df->l2df_data = buf->b_data; 1626240133Smm df->l2df_size = hdr->b_size; 1627185029Spjd df->l2df_func = free_func; 1628185029Spjd mutex_enter(&l2arc_free_on_write_mtx); 1629185029Spjd list_insert_head(l2arc_free_on_write, df); 1630185029Spjd mutex_exit(&l2arc_free_on_write_mtx); 1631185029Spjd ARCSTAT_BUMP(arcstat_l2_free_on_write); 1632185029Spjd } else { 1633240133Smm free_func(buf->b_data, hdr->b_size); 1634185029Spjd } 1635185029Spjd} 1636185029Spjd 1637185029Spjdstatic void 1638168404Spjdarc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all) 1639168404Spjd{ 1640168404Spjd arc_buf_t **bufp; 1641168404Spjd 1642168404Spjd /* free up data associated with the buf */ 1643168404Spjd if (buf->b_data) { 1644168404Spjd arc_state_t *state = buf->b_hdr->b_state; 1645168404Spjd uint64_t size = buf->b_hdr->b_size; 1646168404Spjd arc_buf_contents_t type = buf->b_hdr->b_type; 1647168404Spjd 1648168404Spjd arc_cksum_verify(buf); 1649240133Smm#ifdef illumos 1650240133Smm arc_buf_unwatch(buf); 1651240133Smm#endif /* illumos */ 1652219089Spjd 1653168404Spjd if (!recycle) { 1654168404Spjd if (type == ARC_BUFC_METADATA) { 1655240133Smm arc_buf_data_free(buf, zio_buf_free); 1656208373Smm arc_space_return(size, ARC_SPACE_DATA); 1657168404Spjd } else { 1658168404Spjd ASSERT(type == ARC_BUFC_DATA); 1659240133Smm arc_buf_data_free(buf, zio_data_buf_free); 1660208373Smm ARCSTAT_INCR(arcstat_data_size, -size); 1661185029Spjd atomic_add_64(&arc_size, -size); 1662168404Spjd } 1663168404Spjd } 1664168404Spjd if (list_link_active(&buf->b_hdr->b_arc_node)) { 1665185029Spjd uint64_t *cnt = &state->arcs_lsize[type]; 1666185029Spjd 1667168404Spjd ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1668168404Spjd ASSERT(state != arc_anon); 1669185029Spjd 1670185029Spjd ASSERT3U(*cnt, >=, size); 1671185029Spjd atomic_add_64(cnt, -size); 1672168404Spjd } 1673168404Spjd ASSERT3U(state->arcs_size, >=, size); 1674168404Spjd atomic_add_64(&state->arcs_size, -size); 1675168404Spjd buf->b_data = NULL; 1676242845Sdelphij 1677242845Sdelphij /* 1678242845Sdelphij * If we're destroying a duplicate buffer make sure 1679242845Sdelphij * that the appropriate statistics are updated. 1680242845Sdelphij */ 1681242845Sdelphij if (buf->b_hdr->b_datacnt > 1 && 1682242845Sdelphij buf->b_hdr->b_type == ARC_BUFC_DATA) { 1683242845Sdelphij ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1684242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1685242845Sdelphij } 1686168404Spjd ASSERT(buf->b_hdr->b_datacnt > 0); 1687168404Spjd buf->b_hdr->b_datacnt -= 1; 1688168404Spjd } 1689168404Spjd 1690168404Spjd /* only remove the buf if requested */ 1691168404Spjd if (!all) 1692168404Spjd return; 1693168404Spjd 1694168404Spjd /* remove the buf from the hdr list */ 1695168404Spjd for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1696168404Spjd continue; 1697168404Spjd *bufp = buf->b_next; 1698219089Spjd buf->b_next = NULL; 1699168404Spjd 1700168404Spjd ASSERT(buf->b_efunc == NULL); 1701168404Spjd 1702168404Spjd /* clean up the buf */ 1703168404Spjd buf->b_hdr = NULL; 1704168404Spjd kmem_cache_free(buf_cache, buf); 1705168404Spjd} 1706168404Spjd 1707168404Spjdstatic void 1708168404Spjdarc_hdr_destroy(arc_buf_hdr_t *hdr) 1709168404Spjd{ 1710168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1711168404Spjd ASSERT3P(hdr->b_state, ==, arc_anon); 1712168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1713219089Spjd l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1714168404Spjd 1715219089Spjd if (l2hdr != NULL) { 1716219089Spjd boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1717219089Spjd /* 1718219089Spjd * To prevent arc_free() and l2arc_evict() from 1719219089Spjd * attempting to free the same buffer at the same time, 1720219089Spjd * a FREE_IN_PROGRESS flag is given to arc_free() to 1721219089Spjd * give it priority. l2arc_evict() can't destroy this 1722219089Spjd * header while we are waiting on l2arc_buflist_mtx. 1723219089Spjd * 1724219089Spjd * The hdr may be removed from l2ad_buflist before we 1725219089Spjd * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1726219089Spjd */ 1727219089Spjd if (!buflist_held) { 1728185029Spjd mutex_enter(&l2arc_buflist_mtx); 1729219089Spjd l2hdr = hdr->b_l2hdr; 1730219089Spjd } 1731219089Spjd 1732219089Spjd if (l2hdr != NULL) { 1733248572Ssmh trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1734248574Ssmh hdr->b_size, 0); 1735219089Spjd list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1736219089Spjd ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1737251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1738219089Spjd kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1739219089Spjd if (hdr->b_state == arc_l2c_only) 1740219089Spjd l2arc_hdr_stat_remove(); 1741219089Spjd hdr->b_l2hdr = NULL; 1742219089Spjd } 1743219089Spjd 1744219089Spjd if (!buflist_held) 1745185029Spjd mutex_exit(&l2arc_buflist_mtx); 1746185029Spjd } 1747185029Spjd 1748168404Spjd if (!BUF_EMPTY(hdr)) { 1749168404Spjd ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1750219089Spjd buf_discard_identity(hdr); 1751168404Spjd } 1752168404Spjd while (hdr->b_buf) { 1753168404Spjd arc_buf_t *buf = hdr->b_buf; 1754168404Spjd 1755168404Spjd if (buf->b_efunc) { 1756168404Spjd mutex_enter(&arc_eviction_mtx); 1757219089Spjd mutex_enter(&buf->b_evict_lock); 1758168404Spjd ASSERT(buf->b_hdr != NULL); 1759168404Spjd arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1760168404Spjd hdr->b_buf = buf->b_next; 1761168404Spjd buf->b_hdr = &arc_eviction_hdr; 1762168404Spjd buf->b_next = arc_eviction_list; 1763168404Spjd arc_eviction_list = buf; 1764219089Spjd mutex_exit(&buf->b_evict_lock); 1765168404Spjd mutex_exit(&arc_eviction_mtx); 1766168404Spjd } else { 1767168404Spjd arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1768168404Spjd } 1769168404Spjd } 1770168404Spjd if (hdr->b_freeze_cksum != NULL) { 1771168404Spjd kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1772168404Spjd hdr->b_freeze_cksum = NULL; 1773168404Spjd } 1774219089Spjd if (hdr->b_thawed) { 1775219089Spjd kmem_free(hdr->b_thawed, 1); 1776219089Spjd hdr->b_thawed = NULL; 1777219089Spjd } 1778168404Spjd 1779168404Spjd ASSERT(!list_link_active(&hdr->b_arc_node)); 1780168404Spjd ASSERT3P(hdr->b_hash_next, ==, NULL); 1781168404Spjd ASSERT3P(hdr->b_acb, ==, NULL); 1782168404Spjd kmem_cache_free(hdr_cache, hdr); 1783168404Spjd} 1784168404Spjd 1785168404Spjdvoid 1786168404Spjdarc_buf_free(arc_buf_t *buf, void *tag) 1787168404Spjd{ 1788168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 1789168404Spjd int hashed = hdr->b_state != arc_anon; 1790168404Spjd 1791168404Spjd ASSERT(buf->b_efunc == NULL); 1792168404Spjd ASSERT(buf->b_data != NULL); 1793168404Spjd 1794168404Spjd if (hashed) { 1795168404Spjd kmutex_t *hash_lock = HDR_LOCK(hdr); 1796168404Spjd 1797168404Spjd mutex_enter(hash_lock); 1798219089Spjd hdr = buf->b_hdr; 1799219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1800219089Spjd 1801168404Spjd (void) remove_reference(hdr, hash_lock, tag); 1802219089Spjd if (hdr->b_datacnt > 1) { 1803168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1804219089Spjd } else { 1805219089Spjd ASSERT(buf == hdr->b_buf); 1806219089Spjd ASSERT(buf->b_efunc == NULL); 1807168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 1808219089Spjd } 1809168404Spjd mutex_exit(hash_lock); 1810168404Spjd } else if (HDR_IO_IN_PROGRESS(hdr)) { 1811168404Spjd int destroy_hdr; 1812168404Spjd /* 1813168404Spjd * We are in the middle of an async write. Don't destroy 1814168404Spjd * this buffer unless the write completes before we finish 1815168404Spjd * decrementing the reference count. 1816168404Spjd */ 1817168404Spjd mutex_enter(&arc_eviction_mtx); 1818168404Spjd (void) remove_reference(hdr, NULL, tag); 1819168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1820168404Spjd destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1821168404Spjd mutex_exit(&arc_eviction_mtx); 1822168404Spjd if (destroy_hdr) 1823168404Spjd arc_hdr_destroy(hdr); 1824168404Spjd } else { 1825219089Spjd if (remove_reference(hdr, NULL, tag) > 0) 1826168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1827219089Spjd else 1828168404Spjd arc_hdr_destroy(hdr); 1829168404Spjd } 1830168404Spjd} 1831168404Spjd 1832248571Smmboolean_t 1833168404Spjdarc_buf_remove_ref(arc_buf_t *buf, void* tag) 1834168404Spjd{ 1835168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 1836168404Spjd kmutex_t *hash_lock = HDR_LOCK(hdr); 1837248571Smm boolean_t no_callback = (buf->b_efunc == NULL); 1838168404Spjd 1839168404Spjd if (hdr->b_state == arc_anon) { 1840219089Spjd ASSERT(hdr->b_datacnt == 1); 1841168404Spjd arc_buf_free(buf, tag); 1842168404Spjd return (no_callback); 1843168404Spjd } 1844168404Spjd 1845168404Spjd mutex_enter(hash_lock); 1846219089Spjd hdr = buf->b_hdr; 1847219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1848168404Spjd ASSERT(hdr->b_state != arc_anon); 1849168404Spjd ASSERT(buf->b_data != NULL); 1850168404Spjd 1851168404Spjd (void) remove_reference(hdr, hash_lock, tag); 1852168404Spjd if (hdr->b_datacnt > 1) { 1853168404Spjd if (no_callback) 1854168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1855168404Spjd } else if (no_callback) { 1856168404Spjd ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1857219089Spjd ASSERT(buf->b_efunc == NULL); 1858168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 1859168404Spjd } 1860168404Spjd ASSERT(no_callback || hdr->b_datacnt > 1 || 1861168404Spjd refcount_is_zero(&hdr->b_refcnt)); 1862168404Spjd mutex_exit(hash_lock); 1863168404Spjd return (no_callback); 1864168404Spjd} 1865168404Spjd 1866168404Spjdint 1867168404Spjdarc_buf_size(arc_buf_t *buf) 1868168404Spjd{ 1869168404Spjd return (buf->b_hdr->b_size); 1870168404Spjd} 1871168404Spjd 1872168404Spjd/* 1873242845Sdelphij * Called from the DMU to determine if the current buffer should be 1874242845Sdelphij * evicted. In order to ensure proper locking, the eviction must be initiated 1875242845Sdelphij * from the DMU. Return true if the buffer is associated with user data and 1876242845Sdelphij * duplicate buffers still exist. 1877242845Sdelphij */ 1878242845Sdelphijboolean_t 1879242845Sdelphijarc_buf_eviction_needed(arc_buf_t *buf) 1880242845Sdelphij{ 1881242845Sdelphij arc_buf_hdr_t *hdr; 1882242845Sdelphij boolean_t evict_needed = B_FALSE; 1883242845Sdelphij 1884242845Sdelphij if (zfs_disable_dup_eviction) 1885242845Sdelphij return (B_FALSE); 1886242845Sdelphij 1887242845Sdelphij mutex_enter(&buf->b_evict_lock); 1888242845Sdelphij hdr = buf->b_hdr; 1889242845Sdelphij if (hdr == NULL) { 1890242845Sdelphij /* 1891242845Sdelphij * We are in arc_do_user_evicts(); let that function 1892242845Sdelphij * perform the eviction. 1893242845Sdelphij */ 1894242845Sdelphij ASSERT(buf->b_data == NULL); 1895242845Sdelphij mutex_exit(&buf->b_evict_lock); 1896242845Sdelphij return (B_FALSE); 1897242845Sdelphij } else if (buf->b_data == NULL) { 1898242845Sdelphij /* 1899242845Sdelphij * We have already been added to the arc eviction list; 1900242845Sdelphij * recommend eviction. 1901242845Sdelphij */ 1902242845Sdelphij ASSERT3P(hdr, ==, &arc_eviction_hdr); 1903242845Sdelphij mutex_exit(&buf->b_evict_lock); 1904242845Sdelphij return (B_TRUE); 1905242845Sdelphij } 1906242845Sdelphij 1907242845Sdelphij if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 1908242845Sdelphij evict_needed = B_TRUE; 1909242845Sdelphij 1910242845Sdelphij mutex_exit(&buf->b_evict_lock); 1911242845Sdelphij return (evict_needed); 1912242845Sdelphij} 1913242845Sdelphij 1914242845Sdelphij/* 1915168404Spjd * Evict buffers from list until we've removed the specified number of 1916168404Spjd * bytes. Move the removed buffers to the appropriate evict state. 1917168404Spjd * If the recycle flag is set, then attempt to "recycle" a buffer: 1918168404Spjd * - look for a buffer to evict that is `bytes' long. 1919168404Spjd * - return the data block from this buffer rather than freeing it. 1920168404Spjd * This flag is used by callers that are trying to make space for a 1921168404Spjd * new buffer in a full arc cache. 1922185029Spjd * 1923185029Spjd * This function makes a "best effort". It skips over any buffers 1924185029Spjd * it can't get a hash_lock on, and so may not catch all candidates. 1925185029Spjd * It may also return without evicting as much space as requested. 1926168404Spjd */ 1927168404Spjdstatic void * 1928209962Smmarc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 1929168404Spjd arc_buf_contents_t type) 1930168404Spjd{ 1931168404Spjd arc_state_t *evicted_state; 1932168404Spjd uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 1933205231Skmacy int64_t bytes_remaining; 1934168404Spjd arc_buf_hdr_t *ab, *ab_prev = NULL; 1935205231Skmacy list_t *evicted_list, *list, *evicted_list_start, *list_start; 1936205231Skmacy kmutex_t *lock, *evicted_lock; 1937168404Spjd kmutex_t *hash_lock; 1938168404Spjd boolean_t have_lock; 1939168404Spjd void *stolen = NULL; 1940205231Skmacy static int evict_metadata_offset, evict_data_offset; 1941205231Skmacy int i, idx, offset, list_count, count; 1942168404Spjd 1943168404Spjd ASSERT(state == arc_mru || state == arc_mfu); 1944168404Spjd 1945168404Spjd evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 1946206796Spjd 1947205231Skmacy if (type == ARC_BUFC_METADATA) { 1948205231Skmacy offset = 0; 1949205231Skmacy list_count = ARC_BUFC_NUMMETADATALISTS; 1950205231Skmacy list_start = &state->arcs_lists[0]; 1951205231Skmacy evicted_list_start = &evicted_state->arcs_lists[0]; 1952205231Skmacy idx = evict_metadata_offset; 1953205231Skmacy } else { 1954205231Skmacy offset = ARC_BUFC_NUMMETADATALISTS; 1955205231Skmacy list_start = &state->arcs_lists[offset]; 1956205231Skmacy evicted_list_start = &evicted_state->arcs_lists[offset]; 1957205231Skmacy list_count = ARC_BUFC_NUMDATALISTS; 1958205231Skmacy idx = evict_data_offset; 1959205231Skmacy } 1960205231Skmacy bytes_remaining = evicted_state->arcs_lsize[type]; 1961205231Skmacy count = 0; 1962206796Spjd 1963205231Skmacyevict_start: 1964205231Skmacy list = &list_start[idx]; 1965205231Skmacy evicted_list = &evicted_list_start[idx]; 1966205231Skmacy lock = ARCS_LOCK(state, (offset + idx)); 1967206796Spjd evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 1968168404Spjd 1969205231Skmacy mutex_enter(lock); 1970205231Skmacy mutex_enter(evicted_lock); 1971205231Skmacy 1972185029Spjd for (ab = list_tail(list); ab; ab = ab_prev) { 1973185029Spjd ab_prev = list_prev(list, ab); 1974205231Skmacy bytes_remaining -= (ab->b_size * ab->b_datacnt); 1975168404Spjd /* prefetch buffers have a minimum lifespan */ 1976168404Spjd if (HDR_IO_IN_PROGRESS(ab) || 1977185029Spjd (spa && ab->b_spa != spa) || 1978168404Spjd (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 1979219089Spjd ddi_get_lbolt() - ab->b_arc_access < 1980219089Spjd arc_min_prefetch_lifespan)) { 1981168404Spjd skipped++; 1982168404Spjd continue; 1983168404Spjd } 1984168404Spjd /* "lookahead" for better eviction candidate */ 1985168404Spjd if (recycle && ab->b_size != bytes && 1986168404Spjd ab_prev && ab_prev->b_size == bytes) 1987168404Spjd continue; 1988168404Spjd hash_lock = HDR_LOCK(ab); 1989168404Spjd have_lock = MUTEX_HELD(hash_lock); 1990168404Spjd if (have_lock || mutex_tryenter(hash_lock)) { 1991240415Smm ASSERT0(refcount_count(&ab->b_refcnt)); 1992168404Spjd ASSERT(ab->b_datacnt > 0); 1993168404Spjd while (ab->b_buf) { 1994168404Spjd arc_buf_t *buf = ab->b_buf; 1995219089Spjd if (!mutex_tryenter(&buf->b_evict_lock)) { 1996185029Spjd missed += 1; 1997185029Spjd break; 1998185029Spjd } 1999168404Spjd if (buf->b_data) { 2000168404Spjd bytes_evicted += ab->b_size; 2001168404Spjd if (recycle && ab->b_type == type && 2002185029Spjd ab->b_size == bytes && 2003185029Spjd !HDR_L2_WRITING(ab)) { 2004168404Spjd stolen = buf->b_data; 2005168404Spjd recycle = FALSE; 2006168404Spjd } 2007168404Spjd } 2008168404Spjd if (buf->b_efunc) { 2009168404Spjd mutex_enter(&arc_eviction_mtx); 2010168404Spjd arc_buf_destroy(buf, 2011168404Spjd buf->b_data == stolen, FALSE); 2012168404Spjd ab->b_buf = buf->b_next; 2013168404Spjd buf->b_hdr = &arc_eviction_hdr; 2014168404Spjd buf->b_next = arc_eviction_list; 2015168404Spjd arc_eviction_list = buf; 2016168404Spjd mutex_exit(&arc_eviction_mtx); 2017219089Spjd mutex_exit(&buf->b_evict_lock); 2018168404Spjd } else { 2019219089Spjd mutex_exit(&buf->b_evict_lock); 2020168404Spjd arc_buf_destroy(buf, 2021168404Spjd buf->b_data == stolen, TRUE); 2022168404Spjd } 2023168404Spjd } 2024208373Smm 2025208373Smm if (ab->b_l2hdr) { 2026208373Smm ARCSTAT_INCR(arcstat_evict_l2_cached, 2027208373Smm ab->b_size); 2028208373Smm } else { 2029208373Smm if (l2arc_write_eligible(ab->b_spa, ab)) { 2030208373Smm ARCSTAT_INCR(arcstat_evict_l2_eligible, 2031208373Smm ab->b_size); 2032208373Smm } else { 2033208373Smm ARCSTAT_INCR( 2034208373Smm arcstat_evict_l2_ineligible, 2035208373Smm ab->b_size); 2036208373Smm } 2037208373Smm } 2038208373Smm 2039185029Spjd if (ab->b_datacnt == 0) { 2040185029Spjd arc_change_state(evicted_state, ab, hash_lock); 2041185029Spjd ASSERT(HDR_IN_HASH_TABLE(ab)); 2042185029Spjd ab->b_flags |= ARC_IN_HASH_TABLE; 2043185029Spjd ab->b_flags &= ~ARC_BUF_AVAILABLE; 2044185029Spjd DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 2045185029Spjd } 2046168404Spjd if (!have_lock) 2047168404Spjd mutex_exit(hash_lock); 2048168404Spjd if (bytes >= 0 && bytes_evicted >= bytes) 2049168404Spjd break; 2050205231Skmacy if (bytes_remaining > 0) { 2051205231Skmacy mutex_exit(evicted_lock); 2052205231Skmacy mutex_exit(lock); 2053206796Spjd idx = ((idx + 1) & (list_count - 1)); 2054205231Skmacy count++; 2055205231Skmacy goto evict_start; 2056205231Skmacy } 2057168404Spjd } else { 2058168404Spjd missed += 1; 2059168404Spjd } 2060168404Spjd } 2061168404Spjd 2062205231Skmacy mutex_exit(evicted_lock); 2063205231Skmacy mutex_exit(lock); 2064206796Spjd 2065206796Spjd idx = ((idx + 1) & (list_count - 1)); 2066205231Skmacy count++; 2067168404Spjd 2068205231Skmacy if (bytes_evicted < bytes) { 2069205231Skmacy if (count < list_count) 2070205231Skmacy goto evict_start; 2071205231Skmacy else 2072205231Skmacy dprintf("only evicted %lld bytes from %x", 2073205231Skmacy (longlong_t)bytes_evicted, state); 2074205231Skmacy } 2075206796Spjd if (type == ARC_BUFC_METADATA) 2076205231Skmacy evict_metadata_offset = idx; 2077205231Skmacy else 2078205231Skmacy evict_data_offset = idx; 2079206796Spjd 2080168404Spjd if (skipped) 2081168404Spjd ARCSTAT_INCR(arcstat_evict_skip, skipped); 2082168404Spjd 2083168404Spjd if (missed) 2084168404Spjd ARCSTAT_INCR(arcstat_mutex_miss, missed); 2085168404Spjd 2086185029Spjd /* 2087248571Smm * We have just evicted some data into the ghost state, make 2088185029Spjd * sure we also adjust the ghost state size if necessary. 2089185029Spjd */ 2090185029Spjd if (arc_no_grow && 2091185029Spjd arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) { 2092185029Spjd int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size + 2093185029Spjd arc_mru_ghost->arcs_size - arc_c; 2094185029Spjd 2095185029Spjd if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) { 2096185029Spjd int64_t todelete = 2097185029Spjd MIN(arc_mru_ghost->arcs_lsize[type], mru_over); 2098209962Smm arc_evict_ghost(arc_mru_ghost, 0, todelete); 2099185029Spjd } else if (arc_mfu_ghost->arcs_lsize[type] > 0) { 2100185029Spjd int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type], 2101185029Spjd arc_mru_ghost->arcs_size + 2102185029Spjd arc_mfu_ghost->arcs_size - arc_c); 2103209962Smm arc_evict_ghost(arc_mfu_ghost, 0, todelete); 2104185029Spjd } 2105185029Spjd } 2106205231Skmacy if (stolen) 2107205231Skmacy ARCSTAT_BUMP(arcstat_stolen); 2108185029Spjd 2109168404Spjd return (stolen); 2110168404Spjd} 2111168404Spjd 2112168404Spjd/* 2113168404Spjd * Remove buffers from list until we've removed the specified number of 2114168404Spjd * bytes. Destroy the buffers that are removed. 2115168404Spjd */ 2116168404Spjdstatic void 2117209962Smmarc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2118168404Spjd{ 2119168404Spjd arc_buf_hdr_t *ab, *ab_prev; 2120219089Spjd arc_buf_hdr_t marker = { 0 }; 2121205231Skmacy list_t *list, *list_start; 2122205231Skmacy kmutex_t *hash_lock, *lock; 2123168404Spjd uint64_t bytes_deleted = 0; 2124168404Spjd uint64_t bufs_skipped = 0; 2125205231Skmacy static int evict_offset; 2126205231Skmacy int list_count, idx = evict_offset; 2127205231Skmacy int offset, count = 0; 2128168404Spjd 2129168404Spjd ASSERT(GHOST_STATE(state)); 2130205231Skmacy 2131205231Skmacy /* 2132205231Skmacy * data lists come after metadata lists 2133205231Skmacy */ 2134205231Skmacy list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2135205231Skmacy list_count = ARC_BUFC_NUMDATALISTS; 2136205231Skmacy offset = ARC_BUFC_NUMMETADATALISTS; 2137206796Spjd 2138205231Skmacyevict_start: 2139205231Skmacy list = &list_start[idx]; 2140205231Skmacy lock = ARCS_LOCK(state, idx + offset); 2141205231Skmacy 2142205231Skmacy mutex_enter(lock); 2143185029Spjd for (ab = list_tail(list); ab; ab = ab_prev) { 2144185029Spjd ab_prev = list_prev(list, ab); 2145185029Spjd if (spa && ab->b_spa != spa) 2146185029Spjd continue; 2147219089Spjd 2148219089Spjd /* ignore markers */ 2149219089Spjd if (ab->b_spa == 0) 2150219089Spjd continue; 2151219089Spjd 2152168404Spjd hash_lock = HDR_LOCK(ab); 2153219089Spjd /* caller may be trying to modify this buffer, skip it */ 2154219089Spjd if (MUTEX_HELD(hash_lock)) 2155219089Spjd continue; 2156168404Spjd if (mutex_tryenter(hash_lock)) { 2157168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(ab)); 2158168404Spjd ASSERT(ab->b_buf == NULL); 2159168404Spjd ARCSTAT_BUMP(arcstat_deleted); 2160168404Spjd bytes_deleted += ab->b_size; 2161185029Spjd 2162185029Spjd if (ab->b_l2hdr != NULL) { 2163185029Spjd /* 2164185029Spjd * This buffer is cached on the 2nd Level ARC; 2165185029Spjd * don't destroy the header. 2166185029Spjd */ 2167185029Spjd arc_change_state(arc_l2c_only, ab, hash_lock); 2168185029Spjd mutex_exit(hash_lock); 2169185029Spjd } else { 2170185029Spjd arc_change_state(arc_anon, ab, hash_lock); 2171185029Spjd mutex_exit(hash_lock); 2172185029Spjd arc_hdr_destroy(ab); 2173185029Spjd } 2174185029Spjd 2175168404Spjd DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 2176168404Spjd if (bytes >= 0 && bytes_deleted >= bytes) 2177168404Spjd break; 2178219089Spjd } else if (bytes < 0) { 2179219089Spjd /* 2180219089Spjd * Insert a list marker and then wait for the 2181219089Spjd * hash lock to become available. Once its 2182219089Spjd * available, restart from where we left off. 2183219089Spjd */ 2184219089Spjd list_insert_after(list, ab, &marker); 2185219089Spjd mutex_exit(lock); 2186219089Spjd mutex_enter(hash_lock); 2187219089Spjd mutex_exit(hash_lock); 2188219089Spjd mutex_enter(lock); 2189219089Spjd ab_prev = list_prev(list, &marker); 2190219089Spjd list_remove(list, &marker); 2191219089Spjd } else 2192168404Spjd bufs_skipped += 1; 2193168404Spjd } 2194205231Skmacy mutex_exit(lock); 2195206796Spjd idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2196205231Skmacy count++; 2197206796Spjd 2198205231Skmacy if (count < list_count) 2199205231Skmacy goto evict_start; 2200206796Spjd 2201205231Skmacy evict_offset = idx; 2202205231Skmacy if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2203185029Spjd (bytes < 0 || bytes_deleted < bytes)) { 2204205231Skmacy list_start = &state->arcs_lists[0]; 2205205231Skmacy list_count = ARC_BUFC_NUMMETADATALISTS; 2206205231Skmacy offset = count = 0; 2207205231Skmacy goto evict_start; 2208185029Spjd } 2209185029Spjd 2210168404Spjd if (bufs_skipped) { 2211168404Spjd ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2212168404Spjd ASSERT(bytes >= 0); 2213168404Spjd } 2214168404Spjd 2215168404Spjd if (bytes_deleted < bytes) 2216168404Spjd dprintf("only deleted %lld bytes from %p", 2217168404Spjd (longlong_t)bytes_deleted, state); 2218168404Spjd} 2219168404Spjd 2220168404Spjdstatic void 2221168404Spjdarc_adjust(void) 2222168404Spjd{ 2223208373Smm int64_t adjustment, delta; 2224168404Spjd 2225208373Smm /* 2226208373Smm * Adjust MRU size 2227208373Smm */ 2228168404Spjd 2229209275Smm adjustment = MIN((int64_t)(arc_size - arc_c), 2230209275Smm (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2231209275Smm arc_p)); 2232208373Smm 2233208373Smm if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2234208373Smm delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2235209962Smm (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2236208373Smm adjustment -= delta; 2237168404Spjd } 2238168404Spjd 2239208373Smm if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2240208373Smm delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2241209962Smm (void) arc_evict(arc_mru, 0, delta, FALSE, 2242185029Spjd ARC_BUFC_METADATA); 2243185029Spjd } 2244185029Spjd 2245208373Smm /* 2246208373Smm * Adjust MFU size 2247208373Smm */ 2248168404Spjd 2249208373Smm adjustment = arc_size - arc_c; 2250208373Smm 2251208373Smm if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2252208373Smm delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2253209962Smm (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2254208373Smm adjustment -= delta; 2255168404Spjd } 2256168404Spjd 2257208373Smm if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2258208373Smm int64_t delta = MIN(adjustment, 2259208373Smm arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2260209962Smm (void) arc_evict(arc_mfu, 0, delta, FALSE, 2261208373Smm ARC_BUFC_METADATA); 2262208373Smm } 2263168404Spjd 2264208373Smm /* 2265208373Smm * Adjust ghost lists 2266208373Smm */ 2267168404Spjd 2268208373Smm adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2269168404Spjd 2270208373Smm if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2271208373Smm delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2272209962Smm arc_evict_ghost(arc_mru_ghost, 0, delta); 2273208373Smm } 2274185029Spjd 2275208373Smm adjustment = 2276208373Smm arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2277208373Smm 2278208373Smm if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2279208373Smm delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2280209962Smm arc_evict_ghost(arc_mfu_ghost, 0, delta); 2281168404Spjd } 2282168404Spjd} 2283168404Spjd 2284168404Spjdstatic void 2285168404Spjdarc_do_user_evicts(void) 2286168404Spjd{ 2287191903Skmacy static arc_buf_t *tmp_arc_eviction_list; 2288191903Skmacy 2289191903Skmacy /* 2290191903Skmacy * Move list over to avoid LOR 2291191903Skmacy */ 2292206796Spjdrestart: 2293168404Spjd mutex_enter(&arc_eviction_mtx); 2294191903Skmacy tmp_arc_eviction_list = arc_eviction_list; 2295191903Skmacy arc_eviction_list = NULL; 2296191903Skmacy mutex_exit(&arc_eviction_mtx); 2297191903Skmacy 2298191903Skmacy while (tmp_arc_eviction_list != NULL) { 2299191903Skmacy arc_buf_t *buf = tmp_arc_eviction_list; 2300191903Skmacy tmp_arc_eviction_list = buf->b_next; 2301219089Spjd mutex_enter(&buf->b_evict_lock); 2302168404Spjd buf->b_hdr = NULL; 2303219089Spjd mutex_exit(&buf->b_evict_lock); 2304168404Spjd 2305168404Spjd if (buf->b_efunc != NULL) 2306168404Spjd VERIFY(buf->b_efunc(buf) == 0); 2307168404Spjd 2308168404Spjd buf->b_efunc = NULL; 2309168404Spjd buf->b_private = NULL; 2310168404Spjd kmem_cache_free(buf_cache, buf); 2311168404Spjd } 2312191903Skmacy 2313191903Skmacy if (arc_eviction_list != NULL) 2314191903Skmacy goto restart; 2315168404Spjd} 2316168404Spjd 2317168404Spjd/* 2318185029Spjd * Flush all *evictable* data from the cache for the given spa. 2319168404Spjd * NOTE: this will not touch "active" (i.e. referenced) data. 2320168404Spjd */ 2321168404Spjdvoid 2322185029Spjdarc_flush(spa_t *spa) 2323168404Spjd{ 2324209962Smm uint64_t guid = 0; 2325209962Smm 2326209962Smm if (spa) 2327228103Smm guid = spa_load_guid(spa); 2328209962Smm 2329205231Skmacy while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2330209962Smm (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2331185029Spjd if (spa) 2332185029Spjd break; 2333185029Spjd } 2334205231Skmacy while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2335209962Smm (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2336185029Spjd if (spa) 2337185029Spjd break; 2338185029Spjd } 2339205231Skmacy while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2340209962Smm (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2341185029Spjd if (spa) 2342185029Spjd break; 2343185029Spjd } 2344205231Skmacy while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2345209962Smm (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2346185029Spjd if (spa) 2347185029Spjd break; 2348185029Spjd } 2349168404Spjd 2350209962Smm arc_evict_ghost(arc_mru_ghost, guid, -1); 2351209962Smm arc_evict_ghost(arc_mfu_ghost, guid, -1); 2352168404Spjd 2353168404Spjd mutex_enter(&arc_reclaim_thr_lock); 2354168404Spjd arc_do_user_evicts(); 2355168404Spjd mutex_exit(&arc_reclaim_thr_lock); 2356185029Spjd ASSERT(spa || arc_eviction_list == NULL); 2357168404Spjd} 2358168404Spjd 2359168404Spjdvoid 2360168404Spjdarc_shrink(void) 2361168404Spjd{ 2362168404Spjd if (arc_c > arc_c_min) { 2363168404Spjd uint64_t to_free; 2364168404Spjd 2365168404Spjd#ifdef _KERNEL 2366168404Spjd to_free = arc_c >> arc_shrink_shift; 2367168404Spjd#else 2368168404Spjd to_free = arc_c >> arc_shrink_shift; 2369168404Spjd#endif 2370168404Spjd if (arc_c > arc_c_min + to_free) 2371168404Spjd atomic_add_64(&arc_c, -to_free); 2372168404Spjd else 2373168404Spjd arc_c = arc_c_min; 2374168404Spjd 2375168404Spjd atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2376168404Spjd if (arc_c > arc_size) 2377168404Spjd arc_c = MAX(arc_size, arc_c_min); 2378168404Spjd if (arc_p > arc_c) 2379168404Spjd arc_p = (arc_c >> 1); 2380168404Spjd ASSERT(arc_c >= arc_c_min); 2381168404Spjd ASSERT((int64_t)arc_p >= 0); 2382168404Spjd } 2383168404Spjd 2384168404Spjd if (arc_size > arc_c) 2385168404Spjd arc_adjust(); 2386168404Spjd} 2387168404Spjd 2388185029Spjdstatic int needfree = 0; 2389168404Spjd 2390168404Spjdstatic int 2391168404Spjdarc_reclaim_needed(void) 2392168404Spjd{ 2393168404Spjd 2394168404Spjd#ifdef _KERNEL 2395219089Spjd 2396197816Skmacy if (needfree) 2397197816Skmacy return (1); 2398168404Spjd 2399191902Skmacy /* 2400212780Savg * Cooperate with pagedaemon when it's time for it to scan 2401212780Savg * and reclaim some pages. 2402191902Skmacy */ 2403212783Savg if (vm_paging_needed()) 2404191902Skmacy return (1); 2405191902Skmacy 2406219089Spjd#ifdef sun 2407168404Spjd /* 2408185029Spjd * take 'desfree' extra pages, so we reclaim sooner, rather than later 2409185029Spjd */ 2410185029Spjd extra = desfree; 2411185029Spjd 2412185029Spjd /* 2413185029Spjd * check that we're out of range of the pageout scanner. It starts to 2414185029Spjd * schedule paging if freemem is less than lotsfree and needfree. 2415185029Spjd * lotsfree is the high-water mark for pageout, and needfree is the 2416185029Spjd * number of needed free pages. We add extra pages here to make sure 2417185029Spjd * the scanner doesn't start up while we're freeing memory. 2418185029Spjd */ 2419185029Spjd if (freemem < lotsfree + needfree + extra) 2420185029Spjd return (1); 2421185029Spjd 2422185029Spjd /* 2423168404Spjd * check to make sure that swapfs has enough space so that anon 2424185029Spjd * reservations can still succeed. anon_resvmem() checks that the 2425168404Spjd * availrmem is greater than swapfs_minfree, and the number of reserved 2426168404Spjd * swap pages. We also add a bit of extra here just to prevent 2427168404Spjd * circumstances from getting really dire. 2428168404Spjd */ 2429168404Spjd if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2430168404Spjd return (1); 2431168404Spjd 2432168404Spjd#if defined(__i386) 2433168404Spjd /* 2434168404Spjd * If we're on an i386 platform, it's possible that we'll exhaust the 2435168404Spjd * kernel heap space before we ever run out of available physical 2436168404Spjd * memory. Most checks of the size of the heap_area compare against 2437168404Spjd * tune.t_minarmem, which is the minimum available real memory that we 2438168404Spjd * can have in the system. However, this is generally fixed at 25 pages 2439168404Spjd * which is so low that it's useless. In this comparison, we seek to 2440168404Spjd * calculate the total heap-size, and reclaim if more than 3/4ths of the 2441185029Spjd * heap is allocated. (Or, in the calculation, if less than 1/4th is 2442168404Spjd * free) 2443168404Spjd */ 2444168404Spjd if (btop(vmem_size(heap_arena, VMEM_FREE)) < 2445168404Spjd (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2)) 2446168404Spjd return (1); 2447168404Spjd#endif 2448219089Spjd#else /* !sun */ 2449175633Spjd if (kmem_used() > (kmem_size() * 3) / 4) 2450168404Spjd return (1); 2451219089Spjd#endif /* sun */ 2452168404Spjd 2453168404Spjd#else 2454168404Spjd if (spa_get_random(100) == 0) 2455168404Spjd return (1); 2456168404Spjd#endif 2457168404Spjd return (0); 2458168404Spjd} 2459168404Spjd 2460208454Spjdextern kmem_cache_t *zio_buf_cache[]; 2461208454Spjdextern kmem_cache_t *zio_data_buf_cache[]; 2462208454Spjd 2463168404Spjdstatic void 2464168404Spjdarc_kmem_reap_now(arc_reclaim_strategy_t strat) 2465168404Spjd{ 2466168404Spjd size_t i; 2467168404Spjd kmem_cache_t *prev_cache = NULL; 2468168404Spjd kmem_cache_t *prev_data_cache = NULL; 2469168404Spjd 2470168404Spjd#ifdef _KERNEL 2471185029Spjd if (arc_meta_used >= arc_meta_limit) { 2472185029Spjd /* 2473185029Spjd * We are exceeding our meta-data cache limit. 2474185029Spjd * Purge some DNLC entries to release holds on meta-data. 2475185029Spjd */ 2476185029Spjd dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2477185029Spjd } 2478168404Spjd#if defined(__i386) 2479168404Spjd /* 2480168404Spjd * Reclaim unused memory from all kmem caches. 2481168404Spjd */ 2482168404Spjd kmem_reap(); 2483168404Spjd#endif 2484168404Spjd#endif 2485168404Spjd 2486168404Spjd /* 2487185029Spjd * An aggressive reclamation will shrink the cache size as well as 2488168404Spjd * reap free buffers from the arc kmem caches. 2489168404Spjd */ 2490168404Spjd if (strat == ARC_RECLAIM_AGGR) 2491168404Spjd arc_shrink(); 2492168404Spjd 2493168404Spjd for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2494168404Spjd if (zio_buf_cache[i] != prev_cache) { 2495168404Spjd prev_cache = zio_buf_cache[i]; 2496168404Spjd kmem_cache_reap_now(zio_buf_cache[i]); 2497168404Spjd } 2498168404Spjd if (zio_data_buf_cache[i] != prev_data_cache) { 2499168404Spjd prev_data_cache = zio_data_buf_cache[i]; 2500168404Spjd kmem_cache_reap_now(zio_data_buf_cache[i]); 2501168404Spjd } 2502168404Spjd } 2503168404Spjd kmem_cache_reap_now(buf_cache); 2504168404Spjd kmem_cache_reap_now(hdr_cache); 2505168404Spjd} 2506168404Spjd 2507168404Spjdstatic void 2508168404Spjdarc_reclaim_thread(void *dummy __unused) 2509168404Spjd{ 2510168404Spjd clock_t growtime = 0; 2511168404Spjd arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2512168404Spjd callb_cpr_t cpr; 2513168404Spjd 2514168404Spjd CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2515168404Spjd 2516168404Spjd mutex_enter(&arc_reclaim_thr_lock); 2517168404Spjd while (arc_thread_exit == 0) { 2518168404Spjd if (arc_reclaim_needed()) { 2519168404Spjd 2520168404Spjd if (arc_no_grow) { 2521168404Spjd if (last_reclaim == ARC_RECLAIM_CONS) { 2522168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2523168404Spjd } else { 2524168404Spjd last_reclaim = ARC_RECLAIM_CONS; 2525168404Spjd } 2526168404Spjd } else { 2527168404Spjd arc_no_grow = TRUE; 2528168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2529168404Spjd membar_producer(); 2530168404Spjd } 2531168404Spjd 2532168404Spjd /* reset the growth delay for every reclaim */ 2533219089Spjd growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2534168404Spjd 2535185029Spjd if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2536168404Spjd /* 2537185029Spjd * If needfree is TRUE our vm_lowmem hook 2538168404Spjd * was called and in that case we must free some 2539168404Spjd * memory, so switch to aggressive mode. 2540168404Spjd */ 2541168404Spjd arc_no_grow = TRUE; 2542168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2543168404Spjd } 2544168404Spjd arc_kmem_reap_now(last_reclaim); 2545185029Spjd arc_warm = B_TRUE; 2546185029Spjd 2547219089Spjd } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2548168404Spjd arc_no_grow = FALSE; 2549168404Spjd } 2550168404Spjd 2551209275Smm arc_adjust(); 2552168404Spjd 2553168404Spjd if (arc_eviction_list != NULL) 2554168404Spjd arc_do_user_evicts(); 2555168404Spjd 2556211762Savg#ifdef _KERNEL 2557211762Savg if (needfree) { 2558185029Spjd needfree = 0; 2559185029Spjd wakeup(&needfree); 2560211762Savg } 2561168404Spjd#endif 2562168404Spjd 2563168404Spjd /* block until needed, or one second, whichever is shorter */ 2564168404Spjd CALLB_CPR_SAFE_BEGIN(&cpr); 2565168404Spjd (void) cv_timedwait(&arc_reclaim_thr_cv, 2566168404Spjd &arc_reclaim_thr_lock, hz); 2567168404Spjd CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2568168404Spjd } 2569168404Spjd 2570168404Spjd arc_thread_exit = 0; 2571168404Spjd cv_broadcast(&arc_reclaim_thr_cv); 2572168404Spjd CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2573168404Spjd thread_exit(); 2574168404Spjd} 2575168404Spjd 2576168404Spjd/* 2577168404Spjd * Adapt arc info given the number of bytes we are trying to add and 2578168404Spjd * the state that we are comming from. This function is only called 2579168404Spjd * when we are adding new content to the cache. 2580168404Spjd */ 2581168404Spjdstatic void 2582168404Spjdarc_adapt(int bytes, arc_state_t *state) 2583168404Spjd{ 2584168404Spjd int mult; 2585208373Smm uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2586168404Spjd 2587185029Spjd if (state == arc_l2c_only) 2588185029Spjd return; 2589185029Spjd 2590168404Spjd ASSERT(bytes > 0); 2591168404Spjd /* 2592168404Spjd * Adapt the target size of the MRU list: 2593168404Spjd * - if we just hit in the MRU ghost list, then increase 2594168404Spjd * the target size of the MRU list. 2595168404Spjd * - if we just hit in the MFU ghost list, then increase 2596168404Spjd * the target size of the MFU list by decreasing the 2597168404Spjd * target size of the MRU list. 2598168404Spjd */ 2599168404Spjd if (state == arc_mru_ghost) { 2600168404Spjd mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2601168404Spjd 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2602209275Smm mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2603168404Spjd 2604208373Smm arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2605168404Spjd } else if (state == arc_mfu_ghost) { 2606208373Smm uint64_t delta; 2607208373Smm 2608168404Spjd mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2609168404Spjd 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2610209275Smm mult = MIN(mult, 10); 2611168404Spjd 2612208373Smm delta = MIN(bytes * mult, arc_p); 2613208373Smm arc_p = MAX(arc_p_min, arc_p - delta); 2614168404Spjd } 2615168404Spjd ASSERT((int64_t)arc_p >= 0); 2616168404Spjd 2617168404Spjd if (arc_reclaim_needed()) { 2618168404Spjd cv_signal(&arc_reclaim_thr_cv); 2619168404Spjd return; 2620168404Spjd } 2621168404Spjd 2622168404Spjd if (arc_no_grow) 2623168404Spjd return; 2624168404Spjd 2625168404Spjd if (arc_c >= arc_c_max) 2626168404Spjd return; 2627168404Spjd 2628168404Spjd /* 2629168404Spjd * If we're within (2 * maxblocksize) bytes of the target 2630168404Spjd * cache size, increment the target cache size 2631168404Spjd */ 2632168404Spjd if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2633168404Spjd atomic_add_64(&arc_c, (int64_t)bytes); 2634168404Spjd if (arc_c > arc_c_max) 2635168404Spjd arc_c = arc_c_max; 2636168404Spjd else if (state == arc_anon) 2637168404Spjd atomic_add_64(&arc_p, (int64_t)bytes); 2638168404Spjd if (arc_p > arc_c) 2639168404Spjd arc_p = arc_c; 2640168404Spjd } 2641168404Spjd ASSERT((int64_t)arc_p >= 0); 2642168404Spjd} 2643168404Spjd 2644168404Spjd/* 2645168404Spjd * Check if the cache has reached its limits and eviction is required 2646168404Spjd * prior to insert. 2647168404Spjd */ 2648168404Spjdstatic int 2649185029Spjdarc_evict_needed(arc_buf_contents_t type) 2650168404Spjd{ 2651185029Spjd if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2652185029Spjd return (1); 2653185029Spjd 2654219089Spjd#ifdef sun 2655185029Spjd#ifdef _KERNEL 2656185029Spjd /* 2657185029Spjd * If zio data pages are being allocated out of a separate heap segment, 2658185029Spjd * then enforce that the size of available vmem for this area remains 2659185029Spjd * above about 1/32nd free. 2660185029Spjd */ 2661185029Spjd if (type == ARC_BUFC_DATA && zio_arena != NULL && 2662185029Spjd vmem_size(zio_arena, VMEM_FREE) < 2663185029Spjd (vmem_size(zio_arena, VMEM_ALLOC) >> 5)) 2664185029Spjd return (1); 2665185029Spjd#endif 2666219089Spjd#endif /* sun */ 2667185029Spjd 2668168404Spjd if (arc_reclaim_needed()) 2669168404Spjd return (1); 2670168404Spjd 2671168404Spjd return (arc_size > arc_c); 2672168404Spjd} 2673168404Spjd 2674168404Spjd/* 2675168404Spjd * The buffer, supplied as the first argument, needs a data block. 2676168404Spjd * So, if we are at cache max, determine which cache should be victimized. 2677168404Spjd * We have the following cases: 2678168404Spjd * 2679168404Spjd * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2680168404Spjd * In this situation if we're out of space, but the resident size of the MFU is 2681168404Spjd * under the limit, victimize the MFU cache to satisfy this insertion request. 2682168404Spjd * 2683168404Spjd * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2684168404Spjd * Here, we've used up all of the available space for the MRU, so we need to 2685168404Spjd * evict from our own cache instead. Evict from the set of resident MRU 2686168404Spjd * entries. 2687168404Spjd * 2688168404Spjd * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2689168404Spjd * c minus p represents the MFU space in the cache, since p is the size of the 2690168404Spjd * cache that is dedicated to the MRU. In this situation there's still space on 2691168404Spjd * the MFU side, so the MRU side needs to be victimized. 2692168404Spjd * 2693168404Spjd * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2694168404Spjd * MFU's resident set is consuming more space than it has been allotted. In 2695168404Spjd * this situation, we must victimize our own cache, the MFU, for this insertion. 2696168404Spjd */ 2697168404Spjdstatic void 2698168404Spjdarc_get_data_buf(arc_buf_t *buf) 2699168404Spjd{ 2700168404Spjd arc_state_t *state = buf->b_hdr->b_state; 2701168404Spjd uint64_t size = buf->b_hdr->b_size; 2702168404Spjd arc_buf_contents_t type = buf->b_hdr->b_type; 2703168404Spjd 2704168404Spjd arc_adapt(size, state); 2705168404Spjd 2706168404Spjd /* 2707168404Spjd * We have not yet reached cache maximum size, 2708168404Spjd * just allocate a new buffer. 2709168404Spjd */ 2710185029Spjd if (!arc_evict_needed(type)) { 2711168404Spjd if (type == ARC_BUFC_METADATA) { 2712168404Spjd buf->b_data = zio_buf_alloc(size); 2713208373Smm arc_space_consume(size, ARC_SPACE_DATA); 2714168404Spjd } else { 2715168404Spjd ASSERT(type == ARC_BUFC_DATA); 2716168404Spjd buf->b_data = zio_data_buf_alloc(size); 2717208373Smm ARCSTAT_INCR(arcstat_data_size, size); 2718185029Spjd atomic_add_64(&arc_size, size); 2719168404Spjd } 2720168404Spjd goto out; 2721168404Spjd } 2722168404Spjd 2723168404Spjd /* 2724168404Spjd * If we are prefetching from the mfu ghost list, this buffer 2725168404Spjd * will end up on the mru list; so steal space from there. 2726168404Spjd */ 2727168404Spjd if (state == arc_mfu_ghost) 2728168404Spjd state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2729168404Spjd else if (state == arc_mru_ghost) 2730168404Spjd state = arc_mru; 2731168404Spjd 2732168404Spjd if (state == arc_mru || state == arc_anon) { 2733168404Spjd uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2734208373Smm state = (arc_mfu->arcs_lsize[type] >= size && 2735185029Spjd arc_p > mru_used) ? arc_mfu : arc_mru; 2736168404Spjd } else { 2737168404Spjd /* MFU cases */ 2738168404Spjd uint64_t mfu_space = arc_c - arc_p; 2739208373Smm state = (arc_mru->arcs_lsize[type] >= size && 2740185029Spjd mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2741168404Spjd } 2742209962Smm if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2743168404Spjd if (type == ARC_BUFC_METADATA) { 2744168404Spjd buf->b_data = zio_buf_alloc(size); 2745208373Smm arc_space_consume(size, ARC_SPACE_DATA); 2746168404Spjd } else { 2747168404Spjd ASSERT(type == ARC_BUFC_DATA); 2748168404Spjd buf->b_data = zio_data_buf_alloc(size); 2749208373Smm ARCSTAT_INCR(arcstat_data_size, size); 2750185029Spjd atomic_add_64(&arc_size, size); 2751168404Spjd } 2752168404Spjd ARCSTAT_BUMP(arcstat_recycle_miss); 2753168404Spjd } 2754168404Spjd ASSERT(buf->b_data != NULL); 2755168404Spjdout: 2756168404Spjd /* 2757168404Spjd * Update the state size. Note that ghost states have a 2758168404Spjd * "ghost size" and so don't need to be updated. 2759168404Spjd */ 2760168404Spjd if (!GHOST_STATE(buf->b_hdr->b_state)) { 2761168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 2762168404Spjd 2763168404Spjd atomic_add_64(&hdr->b_state->arcs_size, size); 2764168404Spjd if (list_link_active(&hdr->b_arc_node)) { 2765168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2766185029Spjd atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2767168404Spjd } 2768168404Spjd /* 2769168404Spjd * If we are growing the cache, and we are adding anonymous 2770168404Spjd * data, and we have outgrown arc_p, update arc_p 2771168404Spjd */ 2772168404Spjd if (arc_size < arc_c && hdr->b_state == arc_anon && 2773168404Spjd arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2774168404Spjd arc_p = MIN(arc_c, arc_p + size); 2775168404Spjd } 2776205231Skmacy ARCSTAT_BUMP(arcstat_allocated); 2777168404Spjd} 2778168404Spjd 2779168404Spjd/* 2780168404Spjd * This routine is called whenever a buffer is accessed. 2781168404Spjd * NOTE: the hash lock is dropped in this function. 2782168404Spjd */ 2783168404Spjdstatic void 2784168404Spjdarc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2785168404Spjd{ 2786219089Spjd clock_t now; 2787219089Spjd 2788168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 2789168404Spjd 2790168404Spjd if (buf->b_state == arc_anon) { 2791168404Spjd /* 2792168404Spjd * This buffer is not in the cache, and does not 2793168404Spjd * appear in our "ghost" list. Add the new buffer 2794168404Spjd * to the MRU state. 2795168404Spjd */ 2796168404Spjd 2797168404Spjd ASSERT(buf->b_arc_access == 0); 2798219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2799168404Spjd DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2800168404Spjd arc_change_state(arc_mru, buf, hash_lock); 2801168404Spjd 2802168404Spjd } else if (buf->b_state == arc_mru) { 2803219089Spjd now = ddi_get_lbolt(); 2804219089Spjd 2805168404Spjd /* 2806168404Spjd * If this buffer is here because of a prefetch, then either: 2807168404Spjd * - clear the flag if this is a "referencing" read 2808168404Spjd * (any subsequent access will bump this into the MFU state). 2809168404Spjd * or 2810168404Spjd * - move the buffer to the head of the list if this is 2811168404Spjd * another prefetch (to make it less likely to be evicted). 2812168404Spjd */ 2813168404Spjd if ((buf->b_flags & ARC_PREFETCH) != 0) { 2814168404Spjd if (refcount_count(&buf->b_refcnt) == 0) { 2815168404Spjd ASSERT(list_link_active(&buf->b_arc_node)); 2816168404Spjd } else { 2817168404Spjd buf->b_flags &= ~ARC_PREFETCH; 2818168404Spjd ARCSTAT_BUMP(arcstat_mru_hits); 2819168404Spjd } 2820219089Spjd buf->b_arc_access = now; 2821168404Spjd return; 2822168404Spjd } 2823168404Spjd 2824168404Spjd /* 2825168404Spjd * This buffer has been "accessed" only once so far, 2826168404Spjd * but it is still in the cache. Move it to the MFU 2827168404Spjd * state. 2828168404Spjd */ 2829219089Spjd if (now > buf->b_arc_access + ARC_MINTIME) { 2830168404Spjd /* 2831168404Spjd * More than 125ms have passed since we 2832168404Spjd * instantiated this buffer. Move it to the 2833168404Spjd * most frequently used state. 2834168404Spjd */ 2835219089Spjd buf->b_arc_access = now; 2836168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2837168404Spjd arc_change_state(arc_mfu, buf, hash_lock); 2838168404Spjd } 2839168404Spjd ARCSTAT_BUMP(arcstat_mru_hits); 2840168404Spjd } else if (buf->b_state == arc_mru_ghost) { 2841168404Spjd arc_state_t *new_state; 2842168404Spjd /* 2843168404Spjd * This buffer has been "accessed" recently, but 2844168404Spjd * was evicted from the cache. Move it to the 2845168404Spjd * MFU state. 2846168404Spjd */ 2847168404Spjd 2848168404Spjd if (buf->b_flags & ARC_PREFETCH) { 2849168404Spjd new_state = arc_mru; 2850168404Spjd if (refcount_count(&buf->b_refcnt) > 0) 2851168404Spjd buf->b_flags &= ~ARC_PREFETCH; 2852168404Spjd DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2853168404Spjd } else { 2854168404Spjd new_state = arc_mfu; 2855168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2856168404Spjd } 2857168404Spjd 2858219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2859168404Spjd arc_change_state(new_state, buf, hash_lock); 2860168404Spjd 2861168404Spjd ARCSTAT_BUMP(arcstat_mru_ghost_hits); 2862168404Spjd } else if (buf->b_state == arc_mfu) { 2863168404Spjd /* 2864168404Spjd * This buffer has been accessed more than once and is 2865168404Spjd * still in the cache. Keep it in the MFU state. 2866168404Spjd * 2867168404Spjd * NOTE: an add_reference() that occurred when we did 2868168404Spjd * the arc_read() will have kicked this off the list. 2869168404Spjd * If it was a prefetch, we will explicitly move it to 2870168404Spjd * the head of the list now. 2871168404Spjd */ 2872168404Spjd if ((buf->b_flags & ARC_PREFETCH) != 0) { 2873168404Spjd ASSERT(refcount_count(&buf->b_refcnt) == 0); 2874168404Spjd ASSERT(list_link_active(&buf->b_arc_node)); 2875168404Spjd } 2876168404Spjd ARCSTAT_BUMP(arcstat_mfu_hits); 2877219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2878168404Spjd } else if (buf->b_state == arc_mfu_ghost) { 2879168404Spjd arc_state_t *new_state = arc_mfu; 2880168404Spjd /* 2881168404Spjd * This buffer has been accessed more than once but has 2882168404Spjd * been evicted from the cache. Move it back to the 2883168404Spjd * MFU state. 2884168404Spjd */ 2885168404Spjd 2886168404Spjd if (buf->b_flags & ARC_PREFETCH) { 2887168404Spjd /* 2888168404Spjd * This is a prefetch access... 2889168404Spjd * move this block back to the MRU state. 2890168404Spjd */ 2891240415Smm ASSERT0(refcount_count(&buf->b_refcnt)); 2892168404Spjd new_state = arc_mru; 2893168404Spjd } 2894168404Spjd 2895219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2896168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2897168404Spjd arc_change_state(new_state, buf, hash_lock); 2898168404Spjd 2899168404Spjd ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 2900185029Spjd } else if (buf->b_state == arc_l2c_only) { 2901185029Spjd /* 2902185029Spjd * This buffer is on the 2nd Level ARC. 2903185029Spjd */ 2904185029Spjd 2905219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2906185029Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2907185029Spjd arc_change_state(arc_mfu, buf, hash_lock); 2908168404Spjd } else { 2909168404Spjd ASSERT(!"invalid arc state"); 2910168404Spjd } 2911168404Spjd} 2912168404Spjd 2913168404Spjd/* a generic arc_done_func_t which you can use */ 2914168404Spjd/* ARGSUSED */ 2915168404Spjdvoid 2916168404Spjdarc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 2917168404Spjd{ 2918219089Spjd if (zio == NULL || zio->io_error == 0) 2919219089Spjd bcopy(buf->b_data, arg, buf->b_hdr->b_size); 2920248571Smm VERIFY(arc_buf_remove_ref(buf, arg)); 2921168404Spjd} 2922168404Spjd 2923185029Spjd/* a generic arc_done_func_t */ 2924168404Spjdvoid 2925168404Spjdarc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 2926168404Spjd{ 2927168404Spjd arc_buf_t **bufp = arg; 2928168404Spjd if (zio && zio->io_error) { 2929248571Smm VERIFY(arc_buf_remove_ref(buf, arg)); 2930168404Spjd *bufp = NULL; 2931168404Spjd } else { 2932168404Spjd *bufp = buf; 2933219089Spjd ASSERT(buf->b_data); 2934168404Spjd } 2935168404Spjd} 2936168404Spjd 2937168404Spjdstatic void 2938168404Spjdarc_read_done(zio_t *zio) 2939168404Spjd{ 2940168404Spjd arc_buf_hdr_t *hdr, *found; 2941168404Spjd arc_buf_t *buf; 2942168404Spjd arc_buf_t *abuf; /* buffer we're assigning to callback */ 2943168404Spjd kmutex_t *hash_lock; 2944168404Spjd arc_callback_t *callback_list, *acb; 2945168404Spjd int freeable = FALSE; 2946168404Spjd 2947168404Spjd buf = zio->io_private; 2948168404Spjd hdr = buf->b_hdr; 2949168404Spjd 2950168404Spjd /* 2951168404Spjd * The hdr was inserted into hash-table and removed from lists 2952168404Spjd * prior to starting I/O. We should find this header, since 2953168404Spjd * it's in the hash table, and it should be legit since it's 2954168404Spjd * not possible to evict it during the I/O. The only possible 2955168404Spjd * reason for it not to be found is if we were freed during the 2956168404Spjd * read. 2957168404Spjd */ 2958209962Smm found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth, 2959168404Spjd &hash_lock); 2960168404Spjd 2961168404Spjd ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) || 2962185029Spjd (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 2963185029Spjd (found == hdr && HDR_L2_READING(hdr))); 2964168404Spjd 2965185029Spjd hdr->b_flags &= ~ARC_L2_EVICTED; 2966185029Spjd if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 2967185029Spjd hdr->b_flags &= ~ARC_L2CACHE; 2968206796Spjd 2969168404Spjd /* byteswap if necessary */ 2970168404Spjd callback_list = hdr->b_acb; 2971168404Spjd ASSERT(callback_list != NULL); 2972209101Smm if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 2973236884Smm dmu_object_byteswap_t bswap = 2974236884Smm DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 2975185029Spjd arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 2976185029Spjd byteswap_uint64_array : 2977236884Smm dmu_ot_byteswap[bswap].ob_func; 2978185029Spjd func(buf->b_data, hdr->b_size); 2979185029Spjd } 2980168404Spjd 2981185029Spjd arc_cksum_compute(buf, B_FALSE); 2982240133Smm#ifdef illumos 2983240133Smm arc_buf_watch(buf); 2984240133Smm#endif /* illumos */ 2985168404Spjd 2986219089Spjd if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 2987219089Spjd /* 2988219089Spjd * Only call arc_access on anonymous buffers. This is because 2989219089Spjd * if we've issued an I/O for an evicted buffer, we've already 2990219089Spjd * called arc_access (to prevent any simultaneous readers from 2991219089Spjd * getting confused). 2992219089Spjd */ 2993219089Spjd arc_access(hdr, hash_lock); 2994219089Spjd } 2995219089Spjd 2996168404Spjd /* create copies of the data buffer for the callers */ 2997168404Spjd abuf = buf; 2998168404Spjd for (acb = callback_list; acb; acb = acb->acb_next) { 2999168404Spjd if (acb->acb_done) { 3000242845Sdelphij if (abuf == NULL) { 3001242845Sdelphij ARCSTAT_BUMP(arcstat_duplicate_reads); 3002168404Spjd abuf = arc_buf_clone(buf); 3003242845Sdelphij } 3004168404Spjd acb->acb_buf = abuf; 3005168404Spjd abuf = NULL; 3006168404Spjd } 3007168404Spjd } 3008168404Spjd hdr->b_acb = NULL; 3009168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3010168404Spjd ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3011219089Spjd if (abuf == buf) { 3012219089Spjd ASSERT(buf->b_efunc == NULL); 3013219089Spjd ASSERT(hdr->b_datacnt == 1); 3014168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 3015219089Spjd } 3016168404Spjd 3017168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3018168404Spjd 3019168404Spjd if (zio->io_error != 0) { 3020168404Spjd hdr->b_flags |= ARC_IO_ERROR; 3021168404Spjd if (hdr->b_state != arc_anon) 3022168404Spjd arc_change_state(arc_anon, hdr, hash_lock); 3023168404Spjd if (HDR_IN_HASH_TABLE(hdr)) 3024168404Spjd buf_hash_remove(hdr); 3025168404Spjd freeable = refcount_is_zero(&hdr->b_refcnt); 3026168404Spjd } 3027168404Spjd 3028168404Spjd /* 3029168404Spjd * Broadcast before we drop the hash_lock to avoid the possibility 3030168404Spjd * that the hdr (and hence the cv) might be freed before we get to 3031168404Spjd * the cv_broadcast(). 3032168404Spjd */ 3033168404Spjd cv_broadcast(&hdr->b_cv); 3034168404Spjd 3035168404Spjd if (hash_lock) { 3036168404Spjd mutex_exit(hash_lock); 3037168404Spjd } else { 3038168404Spjd /* 3039168404Spjd * This block was freed while we waited for the read to 3040168404Spjd * complete. It has been removed from the hash table and 3041168404Spjd * moved to the anonymous state (so that it won't show up 3042168404Spjd * in the cache). 3043168404Spjd */ 3044168404Spjd ASSERT3P(hdr->b_state, ==, arc_anon); 3045168404Spjd freeable = refcount_is_zero(&hdr->b_refcnt); 3046168404Spjd } 3047168404Spjd 3048168404Spjd /* execute each callback and free its structure */ 3049168404Spjd while ((acb = callback_list) != NULL) { 3050168404Spjd if (acb->acb_done) 3051168404Spjd acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3052168404Spjd 3053168404Spjd if (acb->acb_zio_dummy != NULL) { 3054168404Spjd acb->acb_zio_dummy->io_error = zio->io_error; 3055168404Spjd zio_nowait(acb->acb_zio_dummy); 3056168404Spjd } 3057168404Spjd 3058168404Spjd callback_list = acb->acb_next; 3059168404Spjd kmem_free(acb, sizeof (arc_callback_t)); 3060168404Spjd } 3061168404Spjd 3062168404Spjd if (freeable) 3063168404Spjd arc_hdr_destroy(hdr); 3064168404Spjd} 3065168404Spjd 3066168404Spjd/* 3067168404Spjd * "Read" the block block at the specified DVA (in bp) via the 3068168404Spjd * cache. If the block is found in the cache, invoke the provided 3069168404Spjd * callback immediately and return. Note that the `zio' parameter 3070168404Spjd * in the callback will be NULL in this case, since no IO was 3071168404Spjd * required. If the block is not in the cache pass the read request 3072168404Spjd * on to the spa with a substitute callback function, so that the 3073168404Spjd * requested block will be added to the cache. 3074168404Spjd * 3075168404Spjd * If a read request arrives for a block that has a read in-progress, 3076168404Spjd * either wait for the in-progress read to complete (and return the 3077168404Spjd * results); or, if this is a read with a "done" func, add a record 3078168404Spjd * to the read to invoke the "done" func when the read completes, 3079168404Spjd * and return; or just return. 3080168404Spjd * 3081168404Spjd * arc_read_done() will invoke all the requested "done" functions 3082168404Spjd * for readers of this block. 3083168404Spjd */ 3084168404Spjdint 3085246666Smmarc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3086246666Smm void *private, int priority, int zio_flags, uint32_t *arc_flags, 3087246666Smm const zbookmark_t *zb) 3088168404Spjd{ 3089168404Spjd arc_buf_hdr_t *hdr; 3090247187Smm arc_buf_t *buf = NULL; 3091168404Spjd kmutex_t *hash_lock; 3092185029Spjd zio_t *rzio; 3093228103Smm uint64_t guid = spa_load_guid(spa); 3094168404Spjd 3095168404Spjdtop: 3096219089Spjd hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp), 3097219089Spjd &hash_lock); 3098168404Spjd if (hdr && hdr->b_datacnt > 0) { 3099168404Spjd 3100168404Spjd *arc_flags |= ARC_CACHED; 3101168404Spjd 3102168404Spjd if (HDR_IO_IN_PROGRESS(hdr)) { 3103168404Spjd 3104168404Spjd if (*arc_flags & ARC_WAIT) { 3105168404Spjd cv_wait(&hdr->b_cv, hash_lock); 3106168404Spjd mutex_exit(hash_lock); 3107168404Spjd goto top; 3108168404Spjd } 3109168404Spjd ASSERT(*arc_flags & ARC_NOWAIT); 3110168404Spjd 3111168404Spjd if (done) { 3112168404Spjd arc_callback_t *acb = NULL; 3113168404Spjd 3114168404Spjd acb = kmem_zalloc(sizeof (arc_callback_t), 3115168404Spjd KM_SLEEP); 3116168404Spjd acb->acb_done = done; 3117168404Spjd acb->acb_private = private; 3118168404Spjd if (pio != NULL) 3119168404Spjd acb->acb_zio_dummy = zio_null(pio, 3120209962Smm spa, NULL, NULL, NULL, zio_flags); 3121168404Spjd 3122168404Spjd ASSERT(acb->acb_done != NULL); 3123168404Spjd acb->acb_next = hdr->b_acb; 3124168404Spjd hdr->b_acb = acb; 3125168404Spjd add_reference(hdr, hash_lock, private); 3126168404Spjd mutex_exit(hash_lock); 3127168404Spjd return (0); 3128168404Spjd } 3129168404Spjd mutex_exit(hash_lock); 3130168404Spjd return (0); 3131168404Spjd } 3132168404Spjd 3133168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3134168404Spjd 3135168404Spjd if (done) { 3136168404Spjd add_reference(hdr, hash_lock, private); 3137168404Spjd /* 3138168404Spjd * If this block is already in use, create a new 3139168404Spjd * copy of the data so that we will be guaranteed 3140168404Spjd * that arc_release() will always succeed. 3141168404Spjd */ 3142168404Spjd buf = hdr->b_buf; 3143168404Spjd ASSERT(buf); 3144168404Spjd ASSERT(buf->b_data); 3145168404Spjd if (HDR_BUF_AVAILABLE(hdr)) { 3146168404Spjd ASSERT(buf->b_efunc == NULL); 3147168404Spjd hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3148168404Spjd } else { 3149168404Spjd buf = arc_buf_clone(buf); 3150168404Spjd } 3151219089Spjd 3152168404Spjd } else if (*arc_flags & ARC_PREFETCH && 3153168404Spjd refcount_count(&hdr->b_refcnt) == 0) { 3154168404Spjd hdr->b_flags |= ARC_PREFETCH; 3155168404Spjd } 3156168404Spjd DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3157168404Spjd arc_access(hdr, hash_lock); 3158185029Spjd if (*arc_flags & ARC_L2CACHE) 3159185029Spjd hdr->b_flags |= ARC_L2CACHE; 3160251478Sdelphij if (*arc_flags & ARC_L2COMPRESS) 3161251478Sdelphij hdr->b_flags |= ARC_L2COMPRESS; 3162168404Spjd mutex_exit(hash_lock); 3163168404Spjd ARCSTAT_BUMP(arcstat_hits); 3164168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3165168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3166168404Spjd data, metadata, hits); 3167168404Spjd 3168168404Spjd if (done) 3169168404Spjd done(NULL, buf, private); 3170168404Spjd } else { 3171168404Spjd uint64_t size = BP_GET_LSIZE(bp); 3172168404Spjd arc_callback_t *acb; 3173185029Spjd vdev_t *vd = NULL; 3174247187Smm uint64_t addr = 0; 3175208373Smm boolean_t devw = B_FALSE; 3176168404Spjd 3177168404Spjd if (hdr == NULL) { 3178168404Spjd /* this block is not in the cache */ 3179168404Spjd arc_buf_hdr_t *exists; 3180168404Spjd arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3181168404Spjd buf = arc_buf_alloc(spa, size, private, type); 3182168404Spjd hdr = buf->b_hdr; 3183168404Spjd hdr->b_dva = *BP_IDENTITY(bp); 3184219089Spjd hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3185168404Spjd hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3186168404Spjd exists = buf_hash_insert(hdr, &hash_lock); 3187168404Spjd if (exists) { 3188168404Spjd /* somebody beat us to the hash insert */ 3189168404Spjd mutex_exit(hash_lock); 3190219089Spjd buf_discard_identity(hdr); 3191168404Spjd (void) arc_buf_remove_ref(buf, private); 3192168404Spjd goto top; /* restart the IO request */ 3193168404Spjd } 3194168404Spjd /* if this is a prefetch, we don't have a reference */ 3195168404Spjd if (*arc_flags & ARC_PREFETCH) { 3196168404Spjd (void) remove_reference(hdr, hash_lock, 3197168404Spjd private); 3198168404Spjd hdr->b_flags |= ARC_PREFETCH; 3199168404Spjd } 3200185029Spjd if (*arc_flags & ARC_L2CACHE) 3201185029Spjd hdr->b_flags |= ARC_L2CACHE; 3202251478Sdelphij if (*arc_flags & ARC_L2COMPRESS) 3203251478Sdelphij hdr->b_flags |= ARC_L2COMPRESS; 3204168404Spjd if (BP_GET_LEVEL(bp) > 0) 3205168404Spjd hdr->b_flags |= ARC_INDIRECT; 3206168404Spjd } else { 3207168404Spjd /* this block is in the ghost cache */ 3208168404Spjd ASSERT(GHOST_STATE(hdr->b_state)); 3209168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3210240415Smm ASSERT0(refcount_count(&hdr->b_refcnt)); 3211168404Spjd ASSERT(hdr->b_buf == NULL); 3212168404Spjd 3213168404Spjd /* if this is a prefetch, we don't have a reference */ 3214168404Spjd if (*arc_flags & ARC_PREFETCH) 3215168404Spjd hdr->b_flags |= ARC_PREFETCH; 3216168404Spjd else 3217168404Spjd add_reference(hdr, hash_lock, private); 3218185029Spjd if (*arc_flags & ARC_L2CACHE) 3219185029Spjd hdr->b_flags |= ARC_L2CACHE; 3220251478Sdelphij if (*arc_flags & ARC_L2COMPRESS) 3221251478Sdelphij hdr->b_flags |= ARC_L2COMPRESS; 3222185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3223168404Spjd buf->b_hdr = hdr; 3224168404Spjd buf->b_data = NULL; 3225168404Spjd buf->b_efunc = NULL; 3226168404Spjd buf->b_private = NULL; 3227168404Spjd buf->b_next = NULL; 3228168404Spjd hdr->b_buf = buf; 3229168404Spjd ASSERT(hdr->b_datacnt == 0); 3230168404Spjd hdr->b_datacnt = 1; 3231219089Spjd arc_get_data_buf(buf); 3232219089Spjd arc_access(hdr, hash_lock); 3233168404Spjd } 3234168404Spjd 3235219089Spjd ASSERT(!GHOST_STATE(hdr->b_state)); 3236219089Spjd 3237168404Spjd acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3238168404Spjd acb->acb_done = done; 3239168404Spjd acb->acb_private = private; 3240168404Spjd 3241168404Spjd ASSERT(hdr->b_acb == NULL); 3242168404Spjd hdr->b_acb = acb; 3243168404Spjd hdr->b_flags |= ARC_IO_IN_PROGRESS; 3244168404Spjd 3245185029Spjd if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL && 3246185029Spjd (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3247208373Smm devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3248185029Spjd addr = hdr->b_l2hdr->b_daddr; 3249185029Spjd /* 3250185029Spjd * Lock out device removal. 3251185029Spjd */ 3252185029Spjd if (vdev_is_dead(vd) || 3253185029Spjd !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3254185029Spjd vd = NULL; 3255185029Spjd } 3256185029Spjd 3257168404Spjd mutex_exit(hash_lock); 3258168404Spjd 3259251629Sdelphij /* 3260251629Sdelphij * At this point, we have a level 1 cache miss. Try again in 3261251629Sdelphij * L2ARC if possible. 3262251629Sdelphij */ 3263168404Spjd ASSERT3U(hdr->b_size, ==, size); 3264219089Spjd DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3265219089Spjd uint64_t, size, zbookmark_t *, zb); 3266168404Spjd ARCSTAT_BUMP(arcstat_misses); 3267168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3268168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3269168404Spjd data, metadata, misses); 3270228392Spjd#ifdef _KERNEL 3271228392Spjd curthread->td_ru.ru_inblock++; 3272228392Spjd#endif 3273168404Spjd 3274208373Smm if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3275185029Spjd /* 3276185029Spjd * Read from the L2ARC if the following are true: 3277185029Spjd * 1. The L2ARC vdev was previously cached. 3278185029Spjd * 2. This buffer still has L2ARC metadata. 3279185029Spjd * 3. This buffer isn't currently writing to the L2ARC. 3280185029Spjd * 4. The L2ARC entry wasn't evicted, which may 3281185029Spjd * also have invalidated the vdev. 3282208373Smm * 5. This isn't prefetch and l2arc_noprefetch is set. 3283185029Spjd */ 3284185029Spjd if (hdr->b_l2hdr != NULL && 3285208373Smm !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3286208373Smm !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3287185029Spjd l2arc_read_callback_t *cb; 3288185029Spjd 3289185029Spjd DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3290185029Spjd ARCSTAT_BUMP(arcstat_l2_hits); 3291185029Spjd 3292185029Spjd cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3293185029Spjd KM_SLEEP); 3294185029Spjd cb->l2rcb_buf = buf; 3295185029Spjd cb->l2rcb_spa = spa; 3296185029Spjd cb->l2rcb_bp = *bp; 3297185029Spjd cb->l2rcb_zb = *zb; 3298185029Spjd cb->l2rcb_flags = zio_flags; 3299251478Sdelphij cb->l2rcb_compress = hdr->b_l2hdr->b_compress; 3300185029Spjd 3301247187Smm ASSERT(addr >= VDEV_LABEL_START_SIZE && 3302247187Smm addr + size < vd->vdev_psize - 3303247187Smm VDEV_LABEL_END_SIZE); 3304247187Smm 3305185029Spjd /* 3306185029Spjd * l2arc read. The SCL_L2ARC lock will be 3307185029Spjd * released by l2arc_read_done(). 3308251478Sdelphij * Issue a null zio if the underlying buffer 3309251478Sdelphij * was squashed to zero size by compression. 3310185029Spjd */ 3311251478Sdelphij if (hdr->b_l2hdr->b_compress == 3312251478Sdelphij ZIO_COMPRESS_EMPTY) { 3313251478Sdelphij rzio = zio_null(pio, spa, vd, 3314251478Sdelphij l2arc_read_done, cb, 3315251478Sdelphij zio_flags | ZIO_FLAG_DONT_CACHE | 3316251478Sdelphij ZIO_FLAG_CANFAIL | 3317251478Sdelphij ZIO_FLAG_DONT_PROPAGATE | 3318251478Sdelphij ZIO_FLAG_DONT_RETRY); 3319251478Sdelphij } else { 3320251478Sdelphij rzio = zio_read_phys(pio, vd, addr, 3321251478Sdelphij hdr->b_l2hdr->b_asize, 3322251478Sdelphij buf->b_data, ZIO_CHECKSUM_OFF, 3323251478Sdelphij l2arc_read_done, cb, priority, 3324251478Sdelphij zio_flags | ZIO_FLAG_DONT_CACHE | 3325251478Sdelphij ZIO_FLAG_CANFAIL | 3326251478Sdelphij ZIO_FLAG_DONT_PROPAGATE | 3327251478Sdelphij ZIO_FLAG_DONT_RETRY, B_FALSE); 3328251478Sdelphij } 3329185029Spjd DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3330185029Spjd zio_t *, rzio); 3331251478Sdelphij ARCSTAT_INCR(arcstat_l2_read_bytes, 3332251478Sdelphij hdr->b_l2hdr->b_asize); 3333185029Spjd 3334185029Spjd if (*arc_flags & ARC_NOWAIT) { 3335185029Spjd zio_nowait(rzio); 3336185029Spjd return (0); 3337185029Spjd } 3338185029Spjd 3339185029Spjd ASSERT(*arc_flags & ARC_WAIT); 3340185029Spjd if (zio_wait(rzio) == 0) 3341185029Spjd return (0); 3342185029Spjd 3343185029Spjd /* l2arc read error; goto zio_read() */ 3344185029Spjd } else { 3345185029Spjd DTRACE_PROBE1(l2arc__miss, 3346185029Spjd arc_buf_hdr_t *, hdr); 3347185029Spjd ARCSTAT_BUMP(arcstat_l2_misses); 3348185029Spjd if (HDR_L2_WRITING(hdr)) 3349185029Spjd ARCSTAT_BUMP(arcstat_l2_rw_clash); 3350185029Spjd spa_config_exit(spa, SCL_L2ARC, vd); 3351185029Spjd } 3352208373Smm } else { 3353208373Smm if (vd != NULL) 3354208373Smm spa_config_exit(spa, SCL_L2ARC, vd); 3355208373Smm if (l2arc_ndev != 0) { 3356208373Smm DTRACE_PROBE1(l2arc__miss, 3357208373Smm arc_buf_hdr_t *, hdr); 3358208373Smm ARCSTAT_BUMP(arcstat_l2_misses); 3359208373Smm } 3360185029Spjd } 3361185029Spjd 3362168404Spjd rzio = zio_read(pio, spa, bp, buf->b_data, size, 3363185029Spjd arc_read_done, buf, priority, zio_flags, zb); 3364168404Spjd 3365168404Spjd if (*arc_flags & ARC_WAIT) 3366168404Spjd return (zio_wait(rzio)); 3367168404Spjd 3368168404Spjd ASSERT(*arc_flags & ARC_NOWAIT); 3369168404Spjd zio_nowait(rzio); 3370168404Spjd } 3371168404Spjd return (0); 3372168404Spjd} 3373168404Spjd 3374168404Spjdvoid 3375168404Spjdarc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3376168404Spjd{ 3377168404Spjd ASSERT(buf->b_hdr != NULL); 3378168404Spjd ASSERT(buf->b_hdr->b_state != arc_anon); 3379168404Spjd ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3380219089Spjd ASSERT(buf->b_efunc == NULL); 3381219089Spjd ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3382219089Spjd 3383168404Spjd buf->b_efunc = func; 3384168404Spjd buf->b_private = private; 3385168404Spjd} 3386168404Spjd 3387168404Spjd/* 3388251520Sdelphij * Notify the arc that a block was freed, and thus will never be used again. 3389251520Sdelphij */ 3390251520Sdelphijvoid 3391251520Sdelphijarc_freed(spa_t *spa, const blkptr_t *bp) 3392251520Sdelphij{ 3393251520Sdelphij arc_buf_hdr_t *hdr; 3394251520Sdelphij kmutex_t *hash_lock; 3395251520Sdelphij uint64_t guid = spa_load_guid(spa); 3396251520Sdelphij 3397251520Sdelphij hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp), 3398251520Sdelphij &hash_lock); 3399251520Sdelphij if (hdr == NULL) 3400251520Sdelphij return; 3401251520Sdelphij if (HDR_BUF_AVAILABLE(hdr)) { 3402251520Sdelphij arc_buf_t *buf = hdr->b_buf; 3403251520Sdelphij add_reference(hdr, hash_lock, FTAG); 3404251520Sdelphij hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3405251520Sdelphij mutex_exit(hash_lock); 3406251520Sdelphij 3407251520Sdelphij arc_release(buf, FTAG); 3408251520Sdelphij (void) arc_buf_remove_ref(buf, FTAG); 3409251520Sdelphij } else { 3410251520Sdelphij mutex_exit(hash_lock); 3411251520Sdelphij } 3412251520Sdelphij 3413251520Sdelphij} 3414251520Sdelphij 3415251520Sdelphij/* 3416168404Spjd * This is used by the DMU to let the ARC know that a buffer is 3417168404Spjd * being evicted, so the ARC should clean up. If this arc buf 3418168404Spjd * is not yet in the evicted state, it will be put there. 3419168404Spjd */ 3420168404Spjdint 3421168404Spjdarc_buf_evict(arc_buf_t *buf) 3422168404Spjd{ 3423168404Spjd arc_buf_hdr_t *hdr; 3424168404Spjd kmutex_t *hash_lock; 3425168404Spjd arc_buf_t **bufp; 3426205231Skmacy list_t *list, *evicted_list; 3427205231Skmacy kmutex_t *lock, *evicted_lock; 3428206796Spjd 3429219089Spjd mutex_enter(&buf->b_evict_lock); 3430168404Spjd hdr = buf->b_hdr; 3431168404Spjd if (hdr == NULL) { 3432168404Spjd /* 3433168404Spjd * We are in arc_do_user_evicts(). 3434168404Spjd */ 3435168404Spjd ASSERT(buf->b_data == NULL); 3436219089Spjd mutex_exit(&buf->b_evict_lock); 3437168404Spjd return (0); 3438185029Spjd } else if (buf->b_data == NULL) { 3439185029Spjd arc_buf_t copy = *buf; /* structure assignment */ 3440185029Spjd /* 3441185029Spjd * We are on the eviction list; process this buffer now 3442185029Spjd * but let arc_do_user_evicts() do the reaping. 3443185029Spjd */ 3444185029Spjd buf->b_efunc = NULL; 3445219089Spjd mutex_exit(&buf->b_evict_lock); 3446185029Spjd VERIFY(copy.b_efunc(©) == 0); 3447185029Spjd return (1); 3448168404Spjd } 3449168404Spjd hash_lock = HDR_LOCK(hdr); 3450168404Spjd mutex_enter(hash_lock); 3451219089Spjd hdr = buf->b_hdr; 3452219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3453168404Spjd 3454168404Spjd ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3455168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3456168404Spjd 3457168404Spjd /* 3458168404Spjd * Pull this buffer off of the hdr 3459168404Spjd */ 3460168404Spjd bufp = &hdr->b_buf; 3461168404Spjd while (*bufp != buf) 3462168404Spjd bufp = &(*bufp)->b_next; 3463168404Spjd *bufp = buf->b_next; 3464168404Spjd 3465168404Spjd ASSERT(buf->b_data != NULL); 3466168404Spjd arc_buf_destroy(buf, FALSE, FALSE); 3467168404Spjd 3468168404Spjd if (hdr->b_datacnt == 0) { 3469168404Spjd arc_state_t *old_state = hdr->b_state; 3470168404Spjd arc_state_t *evicted_state; 3471168404Spjd 3472219089Spjd ASSERT(hdr->b_buf == NULL); 3473168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 3474168404Spjd 3475168404Spjd evicted_state = 3476168404Spjd (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3477168404Spjd 3478205231Skmacy get_buf_info(hdr, old_state, &list, &lock); 3479205231Skmacy get_buf_info(hdr, evicted_state, &evicted_list, &evicted_lock); 3480205231Skmacy mutex_enter(lock); 3481205231Skmacy mutex_enter(evicted_lock); 3482168404Spjd 3483168404Spjd arc_change_state(evicted_state, hdr, hash_lock); 3484168404Spjd ASSERT(HDR_IN_HASH_TABLE(hdr)); 3485185029Spjd hdr->b_flags |= ARC_IN_HASH_TABLE; 3486185029Spjd hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3487168404Spjd 3488205231Skmacy mutex_exit(evicted_lock); 3489205231Skmacy mutex_exit(lock); 3490168404Spjd } 3491168404Spjd mutex_exit(hash_lock); 3492219089Spjd mutex_exit(&buf->b_evict_lock); 3493168404Spjd 3494168404Spjd VERIFY(buf->b_efunc(buf) == 0); 3495168404Spjd buf->b_efunc = NULL; 3496168404Spjd buf->b_private = NULL; 3497168404Spjd buf->b_hdr = NULL; 3498219089Spjd buf->b_next = NULL; 3499168404Spjd kmem_cache_free(buf_cache, buf); 3500168404Spjd return (1); 3501168404Spjd} 3502168404Spjd 3503168404Spjd/* 3504251629Sdelphij * Release this buffer from the cache, making it an anonymous buffer. This 3505251629Sdelphij * must be done after a read and prior to modifying the buffer contents. 3506168404Spjd * If the buffer has more than one reference, we must make 3507185029Spjd * a new hdr for the buffer. 3508168404Spjd */ 3509168404Spjdvoid 3510168404Spjdarc_release(arc_buf_t *buf, void *tag) 3511168404Spjd{ 3512185029Spjd arc_buf_hdr_t *hdr; 3513219089Spjd kmutex_t *hash_lock = NULL; 3514185029Spjd l2arc_buf_hdr_t *l2hdr; 3515185029Spjd uint64_t buf_size; 3516168404Spjd 3517219089Spjd /* 3518219089Spjd * It would be nice to assert that if it's DMU metadata (level > 3519219089Spjd * 0 || it's the dnode file), then it must be syncing context. 3520219089Spjd * But we don't know that information at this level. 3521219089Spjd */ 3522219089Spjd 3523219089Spjd mutex_enter(&buf->b_evict_lock); 3524185029Spjd hdr = buf->b_hdr; 3525185029Spjd 3526168404Spjd /* this buffer is not on any list */ 3527168404Spjd ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3528168404Spjd 3529168404Spjd if (hdr->b_state == arc_anon) { 3530168404Spjd /* this buffer is already released */ 3531168404Spjd ASSERT(buf->b_efunc == NULL); 3532208373Smm } else { 3533208373Smm hash_lock = HDR_LOCK(hdr); 3534208373Smm mutex_enter(hash_lock); 3535219089Spjd hdr = buf->b_hdr; 3536219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3537168404Spjd } 3538168404Spjd 3539185029Spjd l2hdr = hdr->b_l2hdr; 3540185029Spjd if (l2hdr) { 3541185029Spjd mutex_enter(&l2arc_buflist_mtx); 3542185029Spjd hdr->b_l2hdr = NULL; 3543258388Savg list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3544185029Spjd } 3545247187Smm buf_size = hdr->b_size; 3546185029Spjd 3547168404Spjd /* 3548168404Spjd * Do we have more than one buf? 3549168404Spjd */ 3550185029Spjd if (hdr->b_datacnt > 1) { 3551168404Spjd arc_buf_hdr_t *nhdr; 3552168404Spjd arc_buf_t **bufp; 3553168404Spjd uint64_t blksz = hdr->b_size; 3554209962Smm uint64_t spa = hdr->b_spa; 3555168404Spjd arc_buf_contents_t type = hdr->b_type; 3556185029Spjd uint32_t flags = hdr->b_flags; 3557168404Spjd 3558185029Spjd ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3559168404Spjd /* 3560219089Spjd * Pull the data off of this hdr and attach it to 3561219089Spjd * a new anonymous hdr. 3562168404Spjd */ 3563168404Spjd (void) remove_reference(hdr, hash_lock, tag); 3564168404Spjd bufp = &hdr->b_buf; 3565168404Spjd while (*bufp != buf) 3566168404Spjd bufp = &(*bufp)->b_next; 3567219089Spjd *bufp = buf->b_next; 3568168404Spjd buf->b_next = NULL; 3569168404Spjd 3570168404Spjd ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3571168404Spjd atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3572168404Spjd if (refcount_is_zero(&hdr->b_refcnt)) { 3573185029Spjd uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3574185029Spjd ASSERT3U(*size, >=, hdr->b_size); 3575185029Spjd atomic_add_64(size, -hdr->b_size); 3576168404Spjd } 3577242845Sdelphij 3578242845Sdelphij /* 3579242845Sdelphij * We're releasing a duplicate user data buffer, update 3580242845Sdelphij * our statistics accordingly. 3581242845Sdelphij */ 3582242845Sdelphij if (hdr->b_type == ARC_BUFC_DATA) { 3583242845Sdelphij ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3584242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3585242845Sdelphij -hdr->b_size); 3586242845Sdelphij } 3587168404Spjd hdr->b_datacnt -= 1; 3588168404Spjd arc_cksum_verify(buf); 3589240133Smm#ifdef illumos 3590240133Smm arc_buf_unwatch(buf); 3591240133Smm#endif /* illumos */ 3592168404Spjd 3593168404Spjd mutex_exit(hash_lock); 3594168404Spjd 3595185029Spjd nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3596168404Spjd nhdr->b_size = blksz; 3597168404Spjd nhdr->b_spa = spa; 3598168404Spjd nhdr->b_type = type; 3599168404Spjd nhdr->b_buf = buf; 3600168404Spjd nhdr->b_state = arc_anon; 3601168404Spjd nhdr->b_arc_access = 0; 3602185029Spjd nhdr->b_flags = flags & ARC_L2_WRITING; 3603185029Spjd nhdr->b_l2hdr = NULL; 3604168404Spjd nhdr->b_datacnt = 1; 3605168404Spjd nhdr->b_freeze_cksum = NULL; 3606168404Spjd (void) refcount_add(&nhdr->b_refcnt, tag); 3607168404Spjd buf->b_hdr = nhdr; 3608219089Spjd mutex_exit(&buf->b_evict_lock); 3609168404Spjd atomic_add_64(&arc_anon->arcs_size, blksz); 3610168404Spjd } else { 3611219089Spjd mutex_exit(&buf->b_evict_lock); 3612168404Spjd ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3613168404Spjd ASSERT(!list_link_active(&hdr->b_arc_node)); 3614168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3615219089Spjd if (hdr->b_state != arc_anon) 3616219089Spjd arc_change_state(arc_anon, hdr, hash_lock); 3617168404Spjd hdr->b_arc_access = 0; 3618219089Spjd if (hash_lock) 3619219089Spjd mutex_exit(hash_lock); 3620185029Spjd 3621219089Spjd buf_discard_identity(hdr); 3622168404Spjd arc_buf_thaw(buf); 3623168404Spjd } 3624168404Spjd buf->b_efunc = NULL; 3625168404Spjd buf->b_private = NULL; 3626185029Spjd 3627185029Spjd if (l2hdr) { 3628251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3629248572Ssmh trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3630248574Ssmh hdr->b_size, 0); 3631185029Spjd kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3632185029Spjd ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3633185029Spjd mutex_exit(&l2arc_buflist_mtx); 3634185029Spjd } 3635168404Spjd} 3636168404Spjd 3637168404Spjdint 3638168404Spjdarc_released(arc_buf_t *buf) 3639168404Spjd{ 3640185029Spjd int released; 3641185029Spjd 3642219089Spjd mutex_enter(&buf->b_evict_lock); 3643185029Spjd released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3644219089Spjd mutex_exit(&buf->b_evict_lock); 3645185029Spjd return (released); 3646168404Spjd} 3647168404Spjd 3648168404Spjdint 3649168404Spjdarc_has_callback(arc_buf_t *buf) 3650168404Spjd{ 3651185029Spjd int callback; 3652185029Spjd 3653219089Spjd mutex_enter(&buf->b_evict_lock); 3654185029Spjd callback = (buf->b_efunc != NULL); 3655219089Spjd mutex_exit(&buf->b_evict_lock); 3656185029Spjd return (callback); 3657168404Spjd} 3658168404Spjd 3659168404Spjd#ifdef ZFS_DEBUG 3660168404Spjdint 3661168404Spjdarc_referenced(arc_buf_t *buf) 3662168404Spjd{ 3663185029Spjd int referenced; 3664185029Spjd 3665219089Spjd mutex_enter(&buf->b_evict_lock); 3666185029Spjd referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3667219089Spjd mutex_exit(&buf->b_evict_lock); 3668185029Spjd return (referenced); 3669168404Spjd} 3670168404Spjd#endif 3671168404Spjd 3672168404Spjdstatic void 3673168404Spjdarc_write_ready(zio_t *zio) 3674168404Spjd{ 3675168404Spjd arc_write_callback_t *callback = zio->io_private; 3676168404Spjd arc_buf_t *buf = callback->awcb_buf; 3677185029Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3678168404Spjd 3679185029Spjd ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3680185029Spjd callback->awcb_ready(zio, buf, callback->awcb_private); 3681185029Spjd 3682185029Spjd /* 3683185029Spjd * If the IO is already in progress, then this is a re-write 3684185029Spjd * attempt, so we need to thaw and re-compute the cksum. 3685185029Spjd * It is the responsibility of the callback to handle the 3686185029Spjd * accounting for any re-write attempt. 3687185029Spjd */ 3688185029Spjd if (HDR_IO_IN_PROGRESS(hdr)) { 3689185029Spjd mutex_enter(&hdr->b_freeze_lock); 3690185029Spjd if (hdr->b_freeze_cksum != NULL) { 3691185029Spjd kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3692185029Spjd hdr->b_freeze_cksum = NULL; 3693185029Spjd } 3694185029Spjd mutex_exit(&hdr->b_freeze_lock); 3695168404Spjd } 3696185029Spjd arc_cksum_compute(buf, B_FALSE); 3697185029Spjd hdr->b_flags |= ARC_IO_IN_PROGRESS; 3698168404Spjd} 3699168404Spjd 3700168404Spjdstatic void 3701168404Spjdarc_write_done(zio_t *zio) 3702168404Spjd{ 3703168404Spjd arc_write_callback_t *callback = zio->io_private; 3704168404Spjd arc_buf_t *buf = callback->awcb_buf; 3705168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3706168404Spjd 3707219089Spjd ASSERT(hdr->b_acb == NULL); 3708168404Spjd 3709219089Spjd if (zio->io_error == 0) { 3710219089Spjd hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3711219089Spjd hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3712219089Spjd hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3713219089Spjd } else { 3714219089Spjd ASSERT(BUF_EMPTY(hdr)); 3715219089Spjd } 3716219089Spjd 3717168404Spjd /* 3718168404Spjd * If the block to be written was all-zero, we may have 3719168404Spjd * compressed it away. In this case no write was performed 3720219089Spjd * so there will be no dva/birth/checksum. The buffer must 3721219089Spjd * therefore remain anonymous (and uncached). 3722168404Spjd */ 3723168404Spjd if (!BUF_EMPTY(hdr)) { 3724168404Spjd arc_buf_hdr_t *exists; 3725168404Spjd kmutex_t *hash_lock; 3726168404Spjd 3727219089Spjd ASSERT(zio->io_error == 0); 3728219089Spjd 3729168404Spjd arc_cksum_verify(buf); 3730168404Spjd 3731168404Spjd exists = buf_hash_insert(hdr, &hash_lock); 3732168404Spjd if (exists) { 3733168404Spjd /* 3734168404Spjd * This can only happen if we overwrite for 3735168404Spjd * sync-to-convergence, because we remove 3736168404Spjd * buffers from the hash table when we arc_free(). 3737168404Spjd */ 3738219089Spjd if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3739219089Spjd if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3740219089Spjd panic("bad overwrite, hdr=%p exists=%p", 3741219089Spjd (void *)hdr, (void *)exists); 3742219089Spjd ASSERT(refcount_is_zero(&exists->b_refcnt)); 3743219089Spjd arc_change_state(arc_anon, exists, hash_lock); 3744219089Spjd mutex_exit(hash_lock); 3745219089Spjd arc_hdr_destroy(exists); 3746219089Spjd exists = buf_hash_insert(hdr, &hash_lock); 3747219089Spjd ASSERT3P(exists, ==, NULL); 3748243524Smm } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3749243524Smm /* nopwrite */ 3750243524Smm ASSERT(zio->io_prop.zp_nopwrite); 3751243524Smm if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3752243524Smm panic("bad nopwrite, hdr=%p exists=%p", 3753243524Smm (void *)hdr, (void *)exists); 3754219089Spjd } else { 3755219089Spjd /* Dedup */ 3756219089Spjd ASSERT(hdr->b_datacnt == 1); 3757219089Spjd ASSERT(hdr->b_state == arc_anon); 3758219089Spjd ASSERT(BP_GET_DEDUP(zio->io_bp)); 3759219089Spjd ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3760219089Spjd } 3761168404Spjd } 3762168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3763185029Spjd /* if it's not anon, we are doing a scrub */ 3764219089Spjd if (!exists && hdr->b_state == arc_anon) 3765185029Spjd arc_access(hdr, hash_lock); 3766168404Spjd mutex_exit(hash_lock); 3767168404Spjd } else { 3768168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3769168404Spjd } 3770168404Spjd 3771219089Spjd ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3772219089Spjd callback->awcb_done(zio, buf, callback->awcb_private); 3773168404Spjd 3774168404Spjd kmem_free(callback, sizeof (arc_write_callback_t)); 3775168404Spjd} 3776168404Spjd 3777168404Spjdzio_t * 3778219089Spjdarc_write(zio_t *pio, spa_t *spa, uint64_t txg, 3779251478Sdelphij blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 3780251478Sdelphij const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *done, 3781251478Sdelphij void *private, int priority, int zio_flags, const zbookmark_t *zb) 3782168404Spjd{ 3783168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3784168404Spjd arc_write_callback_t *callback; 3785185029Spjd zio_t *zio; 3786168404Spjd 3787185029Spjd ASSERT(ready != NULL); 3788219089Spjd ASSERT(done != NULL); 3789168404Spjd ASSERT(!HDR_IO_ERROR(hdr)); 3790168404Spjd ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3791219089Spjd ASSERT(hdr->b_acb == NULL); 3792185029Spjd if (l2arc) 3793185029Spjd hdr->b_flags |= ARC_L2CACHE; 3794251478Sdelphij if (l2arc_compress) 3795251478Sdelphij hdr->b_flags |= ARC_L2COMPRESS; 3796168404Spjd callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 3797168404Spjd callback->awcb_ready = ready; 3798168404Spjd callback->awcb_done = done; 3799168404Spjd callback->awcb_private = private; 3800168404Spjd callback->awcb_buf = buf; 3801168404Spjd 3802219089Spjd zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 3803185029Spjd arc_write_ready, arc_write_done, callback, priority, zio_flags, zb); 3804185029Spjd 3805168404Spjd return (zio); 3806168404Spjd} 3807168404Spjd 3808185029Spjdstatic int 3809209962Smmarc_memory_throttle(uint64_t reserve, uint64_t inflight_data, uint64_t txg) 3810185029Spjd{ 3811185029Spjd#ifdef _KERNEL 3812219089Spjd uint64_t available_memory = 3813219089Spjd ptoa((uintmax_t)cnt.v_free_count + cnt.v_cache_count); 3814185029Spjd static uint64_t page_load = 0; 3815185029Spjd static uint64_t last_txg = 0; 3816185029Spjd 3817219089Spjd#ifdef sun 3818185029Spjd#if defined(__i386) 3819185029Spjd available_memory = 3820185029Spjd MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 3821185029Spjd#endif 3822219089Spjd#endif /* sun */ 3823185029Spjd if (available_memory >= zfs_write_limit_max) 3824185029Spjd return (0); 3825185029Spjd 3826185029Spjd if (txg > last_txg) { 3827185029Spjd last_txg = txg; 3828185029Spjd page_load = 0; 3829185029Spjd } 3830185029Spjd /* 3831185029Spjd * If we are in pageout, we know that memory is already tight, 3832185029Spjd * the arc is already going to be evicting, so we just want to 3833185029Spjd * continue to let page writes occur as quickly as possible. 3834185029Spjd */ 3835185029Spjd if (curproc == pageproc) { 3836185029Spjd if (page_load > available_memory / 4) 3837249195Smm return (SET_ERROR(ERESTART)); 3838185029Spjd /* Note: reserve is inflated, so we deflate */ 3839185029Spjd page_load += reserve / 8; 3840185029Spjd return (0); 3841185029Spjd } else if (page_load > 0 && arc_reclaim_needed()) { 3842185029Spjd /* memory is low, delay before restarting */ 3843185029Spjd ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3844249195Smm return (SET_ERROR(EAGAIN)); 3845185029Spjd } 3846185029Spjd page_load = 0; 3847185029Spjd 3848185029Spjd if (arc_size > arc_c_min) { 3849185029Spjd uint64_t evictable_memory = 3850185029Spjd arc_mru->arcs_lsize[ARC_BUFC_DATA] + 3851185029Spjd arc_mru->arcs_lsize[ARC_BUFC_METADATA] + 3852185029Spjd arc_mfu->arcs_lsize[ARC_BUFC_DATA] + 3853185029Spjd arc_mfu->arcs_lsize[ARC_BUFC_METADATA]; 3854185029Spjd available_memory += MIN(evictable_memory, arc_size - arc_c_min); 3855185029Spjd } 3856185029Spjd 3857185029Spjd if (inflight_data > available_memory / 4) { 3858185029Spjd ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3859249195Smm return (SET_ERROR(ERESTART)); 3860185029Spjd } 3861185029Spjd#endif 3862185029Spjd return (0); 3863185029Spjd} 3864185029Spjd 3865168404Spjdvoid 3866185029Spjdarc_tempreserve_clear(uint64_t reserve) 3867168404Spjd{ 3868185029Spjd atomic_add_64(&arc_tempreserve, -reserve); 3869168404Spjd ASSERT((int64_t)arc_tempreserve >= 0); 3870168404Spjd} 3871168404Spjd 3872168404Spjdint 3873185029Spjdarc_tempreserve_space(uint64_t reserve, uint64_t txg) 3874168404Spjd{ 3875185029Spjd int error; 3876209962Smm uint64_t anon_size; 3877185029Spjd 3878168404Spjd#ifdef ZFS_DEBUG 3879168404Spjd /* 3880168404Spjd * Once in a while, fail for no reason. Everything should cope. 3881168404Spjd */ 3882168404Spjd if (spa_get_random(10000) == 0) { 3883168404Spjd dprintf("forcing random failure\n"); 3884249195Smm return (SET_ERROR(ERESTART)); 3885168404Spjd } 3886168404Spjd#endif 3887185029Spjd if (reserve > arc_c/4 && !arc_no_grow) 3888185029Spjd arc_c = MIN(arc_c_max, reserve * 4); 3889185029Spjd if (reserve > arc_c) 3890249195Smm return (SET_ERROR(ENOMEM)); 3891168404Spjd 3892168404Spjd /* 3893209962Smm * Don't count loaned bufs as in flight dirty data to prevent long 3894209962Smm * network delays from blocking transactions that are ready to be 3895209962Smm * assigned to a txg. 3896209962Smm */ 3897209962Smm anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 3898209962Smm 3899209962Smm /* 3900185029Spjd * Writes will, almost always, require additional memory allocations 3901251631Sdelphij * in order to compress/encrypt/etc the data. We therefore need to 3902185029Spjd * make sure that there is sufficient available memory for this. 3903185029Spjd */ 3904209962Smm if (error = arc_memory_throttle(reserve, anon_size, txg)) 3905185029Spjd return (error); 3906185029Spjd 3907185029Spjd /* 3908168404Spjd * Throttle writes when the amount of dirty data in the cache 3909168404Spjd * gets too large. We try to keep the cache less than half full 3910168404Spjd * of dirty blocks so that our sync times don't grow too large. 3911168404Spjd * Note: if two requests come in concurrently, we might let them 3912168404Spjd * both succeed, when one of them should fail. Not a huge deal. 3913168404Spjd */ 3914209962Smm 3915209962Smm if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 3916209962Smm anon_size > arc_c / 4) { 3917185029Spjd dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 3918185029Spjd "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 3919185029Spjd arc_tempreserve>>10, 3920185029Spjd arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 3921185029Spjd arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 3922185029Spjd reserve>>10, arc_c>>10); 3923249195Smm return (SET_ERROR(ERESTART)); 3924168404Spjd } 3925185029Spjd atomic_add_64(&arc_tempreserve, reserve); 3926168404Spjd return (0); 3927168404Spjd} 3928168404Spjd 3929168582Spjdstatic kmutex_t arc_lowmem_lock; 3930168404Spjd#ifdef _KERNEL 3931168566Spjdstatic eventhandler_tag arc_event_lowmem = NULL; 3932168404Spjd 3933168404Spjdstatic void 3934168566Spjdarc_lowmem(void *arg __unused, int howto __unused) 3935168404Spjd{ 3936168404Spjd 3937168566Spjd /* Serialize access via arc_lowmem_lock. */ 3938168566Spjd mutex_enter(&arc_lowmem_lock); 3939219089Spjd mutex_enter(&arc_reclaim_thr_lock); 3940185029Spjd needfree = 1; 3941168404Spjd cv_signal(&arc_reclaim_thr_cv); 3942241773Savg 3943241773Savg /* 3944241773Savg * It is unsafe to block here in arbitrary threads, because we can come 3945241773Savg * here from ARC itself and may hold ARC locks and thus risk a deadlock 3946241773Savg * with ARC reclaim thread. 3947241773Savg */ 3948241773Savg if (curproc == pageproc) { 3949241773Savg while (needfree) 3950241773Savg msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 3951241773Savg } 3952219089Spjd mutex_exit(&arc_reclaim_thr_lock); 3953168566Spjd mutex_exit(&arc_lowmem_lock); 3954168404Spjd} 3955168404Spjd#endif 3956168404Spjd 3957168404Spjdvoid 3958168404Spjdarc_init(void) 3959168404Spjd{ 3960219089Spjd int i, prefetch_tunable_set = 0; 3961205231Skmacy 3962168404Spjd mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 3963168404Spjd cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 3964168566Spjd mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 3965168404Spjd 3966168404Spjd /* Convert seconds to clock ticks */ 3967168404Spjd arc_min_prefetch_lifespan = 1 * hz; 3968168404Spjd 3969168404Spjd /* Start out with 1/8 of all memory */ 3970168566Spjd arc_c = kmem_size() / 8; 3971219089Spjd 3972219089Spjd#ifdef sun 3973192360Skmacy#ifdef _KERNEL 3974192360Skmacy /* 3975192360Skmacy * On architectures where the physical memory can be larger 3976192360Skmacy * than the addressable space (intel in 32-bit mode), we may 3977192360Skmacy * need to limit the cache to 1/8 of VM size. 3978192360Skmacy */ 3979192360Skmacy arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 3980192360Skmacy#endif 3981219089Spjd#endif /* sun */ 3982168566Spjd /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 3983168566Spjd arc_c_min = MAX(arc_c / 4, 64<<18); 3984168566Spjd /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 3985168404Spjd if (arc_c * 8 >= 1<<30) 3986168404Spjd arc_c_max = (arc_c * 8) - (1<<30); 3987168404Spjd else 3988168404Spjd arc_c_max = arc_c_min; 3989175633Spjd arc_c_max = MAX(arc_c * 5, arc_c_max); 3990219089Spjd 3991168481Spjd#ifdef _KERNEL 3992168404Spjd /* 3993168404Spjd * Allow the tunables to override our calculations if they are 3994168566Spjd * reasonable (ie. over 16MB) 3995168404Spjd */ 3996219089Spjd if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size()) 3997168404Spjd arc_c_max = zfs_arc_max; 3998219089Spjd if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max) 3999168404Spjd arc_c_min = zfs_arc_min; 4000168481Spjd#endif 4001219089Spjd 4002168404Spjd arc_c = arc_c_max; 4003168404Spjd arc_p = (arc_c >> 1); 4004168404Spjd 4005185029Spjd /* limit meta-data to 1/4 of the arc capacity */ 4006185029Spjd arc_meta_limit = arc_c_max / 4; 4007185029Spjd 4008185029Spjd /* Allow the tunable to override if it is reasonable */ 4009185029Spjd if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4010185029Spjd arc_meta_limit = zfs_arc_meta_limit; 4011185029Spjd 4012185029Spjd if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4013185029Spjd arc_c_min = arc_meta_limit / 2; 4014185029Spjd 4015208373Smm if (zfs_arc_grow_retry > 0) 4016208373Smm arc_grow_retry = zfs_arc_grow_retry; 4017208373Smm 4018208373Smm if (zfs_arc_shrink_shift > 0) 4019208373Smm arc_shrink_shift = zfs_arc_shrink_shift; 4020208373Smm 4021208373Smm if (zfs_arc_p_min_shift > 0) 4022208373Smm arc_p_min_shift = zfs_arc_p_min_shift; 4023208373Smm 4024168404Spjd /* if kmem_flags are set, lets try to use less memory */ 4025168404Spjd if (kmem_debugging()) 4026168404Spjd arc_c = arc_c / 2; 4027168404Spjd if (arc_c < arc_c_min) 4028168404Spjd arc_c = arc_c_min; 4029168404Spjd 4030168473Spjd zfs_arc_min = arc_c_min; 4031168473Spjd zfs_arc_max = arc_c_max; 4032168473Spjd 4033168404Spjd arc_anon = &ARC_anon; 4034168404Spjd arc_mru = &ARC_mru; 4035168404Spjd arc_mru_ghost = &ARC_mru_ghost; 4036168404Spjd arc_mfu = &ARC_mfu; 4037168404Spjd arc_mfu_ghost = &ARC_mfu_ghost; 4038185029Spjd arc_l2c_only = &ARC_l2c_only; 4039168404Spjd arc_size = 0; 4040168404Spjd 4041205231Skmacy for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4042205231Skmacy mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4043205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4044205231Skmacy mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4045205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4046205231Skmacy mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4047205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4048205231Skmacy mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4049205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4050205231Skmacy mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4051205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4052205231Skmacy mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4053205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4054206796Spjd 4055205231Skmacy list_create(&arc_mru->arcs_lists[i], 4056205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4057205231Skmacy list_create(&arc_mru_ghost->arcs_lists[i], 4058205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4059205231Skmacy list_create(&arc_mfu->arcs_lists[i], 4060205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4061205231Skmacy list_create(&arc_mfu_ghost->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_l2c_only->arcs_lists[i], 4066205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4067205231Skmacy } 4068168404Spjd 4069168404Spjd buf_init(); 4070168404Spjd 4071168404Spjd arc_thread_exit = 0; 4072168404Spjd arc_eviction_list = NULL; 4073168404Spjd mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4074168404Spjd bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4075168404Spjd 4076168404Spjd arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4077168404Spjd sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4078168404Spjd 4079168404Spjd if (arc_ksp != NULL) { 4080168404Spjd arc_ksp->ks_data = &arc_stats; 4081168404Spjd kstat_install(arc_ksp); 4082168404Spjd } 4083168404Spjd 4084168404Spjd (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4085168404Spjd TS_RUN, minclsyspri); 4086168404Spjd 4087168404Spjd#ifdef _KERNEL 4088168566Spjd arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4089168404Spjd EVENTHANDLER_PRI_FIRST); 4090168404Spjd#endif 4091168404Spjd 4092168404Spjd arc_dead = FALSE; 4093185029Spjd arc_warm = B_FALSE; 4094168566Spjd 4095185029Spjd if (zfs_write_limit_max == 0) 4096185029Spjd zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift; 4097185029Spjd else 4098185029Spjd zfs_write_limit_shift = 0; 4099185029Spjd mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL); 4100185029Spjd 4101168566Spjd#ifdef _KERNEL 4102194043Skmacy if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4103193953Skmacy prefetch_tunable_set = 1; 4104206796Spjd 4105193878Skmacy#ifdef __i386__ 4106193953Skmacy if (prefetch_tunable_set == 0) { 4107196863Strasz printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4108196863Strasz "-- to enable,\n"); 4109196863Strasz printf(" add \"vfs.zfs.prefetch_disable=0\" " 4110196863Strasz "to /boot/loader.conf.\n"); 4111219089Spjd zfs_prefetch_disable = 1; 4112193878Skmacy } 4113206796Spjd#else 4114193878Skmacy if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4115193953Skmacy prefetch_tunable_set == 0) { 4116196863Strasz printf("ZFS NOTICE: Prefetch is disabled by default if less " 4117196941Strasz "than 4GB of RAM is present;\n" 4118196863Strasz " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4119196863Strasz "to /boot/loader.conf.\n"); 4120219089Spjd zfs_prefetch_disable = 1; 4121193878Skmacy } 4122206796Spjd#endif 4123175633Spjd /* Warn about ZFS memory and address space requirements. */ 4124168696Spjd if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4125168987Sbmah printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4126168987Sbmah "expect unstable behavior.\n"); 4127175633Spjd } 4128175633Spjd if (kmem_size() < 512 * (1 << 20)) { 4129173419Spjd printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4130168987Sbmah "expect unstable behavior.\n"); 4131185029Spjd printf(" Consider tuning vm.kmem_size and " 4132173419Spjd "vm.kmem_size_max\n"); 4133185029Spjd printf(" in /boot/loader.conf.\n"); 4134168566Spjd } 4135168566Spjd#endif 4136168404Spjd} 4137168404Spjd 4138168404Spjdvoid 4139168404Spjdarc_fini(void) 4140168404Spjd{ 4141205231Skmacy int i; 4142206796Spjd 4143168404Spjd mutex_enter(&arc_reclaim_thr_lock); 4144168404Spjd arc_thread_exit = 1; 4145168404Spjd cv_signal(&arc_reclaim_thr_cv); 4146168404Spjd while (arc_thread_exit != 0) 4147168404Spjd cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4148168404Spjd mutex_exit(&arc_reclaim_thr_lock); 4149168404Spjd 4150185029Spjd arc_flush(NULL); 4151168404Spjd 4152168404Spjd arc_dead = TRUE; 4153168404Spjd 4154168404Spjd if (arc_ksp != NULL) { 4155168404Spjd kstat_delete(arc_ksp); 4156168404Spjd arc_ksp = NULL; 4157168404Spjd } 4158168404Spjd 4159168404Spjd mutex_destroy(&arc_eviction_mtx); 4160168404Spjd mutex_destroy(&arc_reclaim_thr_lock); 4161168404Spjd cv_destroy(&arc_reclaim_thr_cv); 4162168404Spjd 4163205231Skmacy for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4164205231Skmacy list_destroy(&arc_mru->arcs_lists[i]); 4165205231Skmacy list_destroy(&arc_mru_ghost->arcs_lists[i]); 4166205231Skmacy list_destroy(&arc_mfu->arcs_lists[i]); 4167205231Skmacy list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4168206795Spjd list_destroy(&arc_l2c_only->arcs_lists[i]); 4169168404Spjd 4170205231Skmacy mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4171205231Skmacy mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4172205231Skmacy mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4173205231Skmacy mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4174205231Skmacy mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4175206795Spjd mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4176205231Skmacy } 4177206796Spjd 4178185029Spjd mutex_destroy(&zfs_write_limit_lock); 4179185029Spjd 4180168404Spjd buf_fini(); 4181168404Spjd 4182209962Smm ASSERT(arc_loaned_bytes == 0); 4183209962Smm 4184168582Spjd mutex_destroy(&arc_lowmem_lock); 4185168404Spjd#ifdef _KERNEL 4186168566Spjd if (arc_event_lowmem != NULL) 4187168566Spjd EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4188168404Spjd#endif 4189168404Spjd} 4190185029Spjd 4191185029Spjd/* 4192185029Spjd * Level 2 ARC 4193185029Spjd * 4194185029Spjd * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4195185029Spjd * It uses dedicated storage devices to hold cached data, which are populated 4196185029Spjd * using large infrequent writes. The main role of this cache is to boost 4197185029Spjd * the performance of random read workloads. The intended L2ARC devices 4198185029Spjd * include short-stroked disks, solid state disks, and other media with 4199185029Spjd * substantially faster read latency than disk. 4200185029Spjd * 4201185029Spjd * +-----------------------+ 4202185029Spjd * | ARC | 4203185029Spjd * +-----------------------+ 4204185029Spjd * | ^ ^ 4205185029Spjd * | | | 4206185029Spjd * l2arc_feed_thread() arc_read() 4207185029Spjd * | | | 4208185029Spjd * | l2arc read | 4209185029Spjd * V | | 4210185029Spjd * +---------------+ | 4211185029Spjd * | L2ARC | | 4212185029Spjd * +---------------+ | 4213185029Spjd * | ^ | 4214185029Spjd * l2arc_write() | | 4215185029Spjd * | | | 4216185029Spjd * V | | 4217185029Spjd * +-------+ +-------+ 4218185029Spjd * | vdev | | vdev | 4219185029Spjd * | cache | | cache | 4220185029Spjd * +-------+ +-------+ 4221185029Spjd * +=========+ .-----. 4222185029Spjd * : L2ARC : |-_____-| 4223185029Spjd * : devices : | Disks | 4224185029Spjd * +=========+ `-_____-' 4225185029Spjd * 4226185029Spjd * Read requests are satisfied from the following sources, in order: 4227185029Spjd * 4228185029Spjd * 1) ARC 4229185029Spjd * 2) vdev cache of L2ARC devices 4230185029Spjd * 3) L2ARC devices 4231185029Spjd * 4) vdev cache of disks 4232185029Spjd * 5) disks 4233185029Spjd * 4234185029Spjd * Some L2ARC device types exhibit extremely slow write performance. 4235185029Spjd * To accommodate for this there are some significant differences between 4236185029Spjd * the L2ARC and traditional cache design: 4237185029Spjd * 4238185029Spjd * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4239185029Spjd * the ARC behave as usual, freeing buffers and placing headers on ghost 4240185029Spjd * lists. The ARC does not send buffers to the L2ARC during eviction as 4241185029Spjd * this would add inflated write latencies for all ARC memory pressure. 4242185029Spjd * 4243185029Spjd * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4244185029Spjd * It does this by periodically scanning buffers from the eviction-end of 4245185029Spjd * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4246251478Sdelphij * not already there. It scans until a headroom of buffers is satisfied, 4247251478Sdelphij * which itself is a buffer for ARC eviction. If a compressible buffer is 4248251478Sdelphij * found during scanning and selected for writing to an L2ARC device, we 4249251478Sdelphij * temporarily boost scanning headroom during the next scan cycle to make 4250251478Sdelphij * sure we adapt to compression effects (which might significantly reduce 4251251478Sdelphij * the data volume we write to L2ARC). The thread that does this is 4252185029Spjd * l2arc_feed_thread(), illustrated below; example sizes are included to 4253185029Spjd * provide a better sense of ratio than this diagram: 4254185029Spjd * 4255185029Spjd * head --> tail 4256185029Spjd * +---------------------+----------+ 4257185029Spjd * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4258185029Spjd * +---------------------+----------+ | o L2ARC eligible 4259185029Spjd * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4260185029Spjd * +---------------------+----------+ | 4261185029Spjd * 15.9 Gbytes ^ 32 Mbytes | 4262185029Spjd * headroom | 4263185029Spjd * l2arc_feed_thread() 4264185029Spjd * | 4265185029Spjd * l2arc write hand <--[oooo]--' 4266185029Spjd * | 8 Mbyte 4267185029Spjd * | write max 4268185029Spjd * V 4269185029Spjd * +==============================+ 4270185029Spjd * L2ARC dev |####|#|###|###| |####| ... | 4271185029Spjd * +==============================+ 4272185029Spjd * 32 Gbytes 4273185029Spjd * 4274185029Spjd * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4275185029Spjd * evicted, then the L2ARC has cached a buffer much sooner than it probably 4276185029Spjd * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4277185029Spjd * safe to say that this is an uncommon case, since buffers at the end of 4278185029Spjd * the ARC lists have moved there due to inactivity. 4279185029Spjd * 4280185029Spjd * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4281185029Spjd * then the L2ARC simply misses copying some buffers. This serves as a 4282185029Spjd * pressure valve to prevent heavy read workloads from both stalling the ARC 4283185029Spjd * with waits and clogging the L2ARC with writes. This also helps prevent 4284185029Spjd * the potential for the L2ARC to churn if it attempts to cache content too 4285185029Spjd * quickly, such as during backups of the entire pool. 4286185029Spjd * 4287185029Spjd * 5. After system boot and before the ARC has filled main memory, there are 4288185029Spjd * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4289185029Spjd * lists can remain mostly static. Instead of searching from tail of these 4290185029Spjd * lists as pictured, the l2arc_feed_thread() will search from the list heads 4291185029Spjd * for eligible buffers, greatly increasing its chance of finding them. 4292185029Spjd * 4293185029Spjd * The L2ARC device write speed is also boosted during this time so that 4294185029Spjd * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4295185029Spjd * there are no L2ARC reads, and no fear of degrading read performance 4296185029Spjd * through increased writes. 4297185029Spjd * 4298185029Spjd * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4299185029Spjd * the vdev queue can aggregate them into larger and fewer writes. Each 4300185029Spjd * device is written to in a rotor fashion, sweeping writes through 4301185029Spjd * available space then repeating. 4302185029Spjd * 4303185029Spjd * 7. The L2ARC does not store dirty content. It never needs to flush 4304185029Spjd * write buffers back to disk based storage. 4305185029Spjd * 4306185029Spjd * 8. If an ARC buffer is written (and dirtied) which also exists in the 4307185029Spjd * L2ARC, the now stale L2ARC buffer is immediately dropped. 4308185029Spjd * 4309185029Spjd * The performance of the L2ARC can be tweaked by a number of tunables, which 4310185029Spjd * may be necessary for different workloads: 4311185029Spjd * 4312185029Spjd * l2arc_write_max max write bytes per interval 4313185029Spjd * l2arc_write_boost extra write bytes during device warmup 4314185029Spjd * l2arc_noprefetch skip caching prefetched buffers 4315185029Spjd * l2arc_headroom number of max device writes to precache 4316251478Sdelphij * l2arc_headroom_boost when we find compressed buffers during ARC 4317251478Sdelphij * scanning, we multiply headroom by this 4318251478Sdelphij * percentage factor for the next scan cycle, 4319251478Sdelphij * since more compressed buffers are likely to 4320251478Sdelphij * be present 4321185029Spjd * l2arc_feed_secs seconds between L2ARC writing 4322185029Spjd * 4323185029Spjd * Tunables may be removed or added as future performance improvements are 4324185029Spjd * integrated, and also may become zpool properties. 4325208373Smm * 4326208373Smm * There are three key functions that control how the L2ARC warms up: 4327208373Smm * 4328208373Smm * l2arc_write_eligible() check if a buffer is eligible to cache 4329208373Smm * l2arc_write_size() calculate how much to write 4330208373Smm * l2arc_write_interval() calculate sleep delay between writes 4331208373Smm * 4332208373Smm * These three functions determine what to write, how much, and how quickly 4333208373Smm * to send writes. 4334185029Spjd */ 4335185029Spjd 4336208373Smmstatic boolean_t 4337209962Smml2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab) 4338208373Smm{ 4339208373Smm /* 4340208373Smm * A buffer is *not* eligible for the L2ARC if it: 4341208373Smm * 1. belongs to a different spa. 4342208373Smm * 2. is already cached on the L2ARC. 4343208373Smm * 3. has an I/O in progress (it may be an incomplete read). 4344208373Smm * 4. is flagged not eligible (zfs property). 4345208373Smm */ 4346209962Smm if (ab->b_spa != spa_guid) { 4347208373Smm ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4348208373Smm return (B_FALSE); 4349208373Smm } 4350208373Smm if (ab->b_l2hdr != NULL) { 4351208373Smm ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4352208373Smm return (B_FALSE); 4353208373Smm } 4354208373Smm if (HDR_IO_IN_PROGRESS(ab)) { 4355208373Smm ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4356208373Smm return (B_FALSE); 4357208373Smm } 4358208373Smm if (!HDR_L2CACHE(ab)) { 4359208373Smm ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4360208373Smm return (B_FALSE); 4361208373Smm } 4362208373Smm 4363208373Smm return (B_TRUE); 4364208373Smm} 4365208373Smm 4366208373Smmstatic uint64_t 4367251478Sdelphijl2arc_write_size(void) 4368208373Smm{ 4369208373Smm uint64_t size; 4370208373Smm 4371251478Sdelphij /* 4372251478Sdelphij * Make sure our globals have meaningful values in case the user 4373251478Sdelphij * altered them. 4374251478Sdelphij */ 4375251478Sdelphij size = l2arc_write_max; 4376251478Sdelphij if (size == 0) { 4377251478Sdelphij cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4378251478Sdelphij "be greater than zero, resetting it to the default (%d)", 4379251478Sdelphij L2ARC_WRITE_SIZE); 4380251478Sdelphij size = l2arc_write_max = L2ARC_WRITE_SIZE; 4381251478Sdelphij } 4382208373Smm 4383208373Smm if (arc_warm == B_FALSE) 4384251478Sdelphij size += l2arc_write_boost; 4385208373Smm 4386208373Smm return (size); 4387208373Smm 4388208373Smm} 4389208373Smm 4390208373Smmstatic clock_t 4391208373Smml2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4392208373Smm{ 4393219089Spjd clock_t interval, next, now; 4394208373Smm 4395208373Smm /* 4396208373Smm * If the ARC lists are busy, increase our write rate; if the 4397208373Smm * lists are stale, idle back. This is achieved by checking 4398208373Smm * how much we previously wrote - if it was more than half of 4399208373Smm * what we wanted, schedule the next write much sooner. 4400208373Smm */ 4401208373Smm if (l2arc_feed_again && wrote > (wanted / 2)) 4402208373Smm interval = (hz * l2arc_feed_min_ms) / 1000; 4403208373Smm else 4404208373Smm interval = hz * l2arc_feed_secs; 4405208373Smm 4406219089Spjd now = ddi_get_lbolt(); 4407219089Spjd next = MAX(now, MIN(now + interval, began + interval)); 4408208373Smm 4409208373Smm return (next); 4410208373Smm} 4411208373Smm 4412185029Spjdstatic void 4413185029Spjdl2arc_hdr_stat_add(void) 4414185029Spjd{ 4415185029Spjd ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4416185029Spjd ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4417185029Spjd} 4418185029Spjd 4419185029Spjdstatic void 4420185029Spjdl2arc_hdr_stat_remove(void) 4421185029Spjd{ 4422185029Spjd ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4423185029Spjd ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4424185029Spjd} 4425185029Spjd 4426185029Spjd/* 4427185029Spjd * Cycle through L2ARC devices. This is how L2ARC load balances. 4428185029Spjd * If a device is returned, this also returns holding the spa config lock. 4429185029Spjd */ 4430185029Spjdstatic l2arc_dev_t * 4431185029Spjdl2arc_dev_get_next(void) 4432185029Spjd{ 4433185029Spjd l2arc_dev_t *first, *next = NULL; 4434185029Spjd 4435185029Spjd /* 4436185029Spjd * Lock out the removal of spas (spa_namespace_lock), then removal 4437185029Spjd * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4438185029Spjd * both locks will be dropped and a spa config lock held instead. 4439185029Spjd */ 4440185029Spjd mutex_enter(&spa_namespace_lock); 4441185029Spjd mutex_enter(&l2arc_dev_mtx); 4442185029Spjd 4443185029Spjd /* if there are no vdevs, there is nothing to do */ 4444185029Spjd if (l2arc_ndev == 0) 4445185029Spjd goto out; 4446185029Spjd 4447185029Spjd first = NULL; 4448185029Spjd next = l2arc_dev_last; 4449185029Spjd do { 4450185029Spjd /* loop around the list looking for a non-faulted vdev */ 4451185029Spjd if (next == NULL) { 4452185029Spjd next = list_head(l2arc_dev_list); 4453185029Spjd } else { 4454185029Spjd next = list_next(l2arc_dev_list, next); 4455185029Spjd if (next == NULL) 4456185029Spjd next = list_head(l2arc_dev_list); 4457185029Spjd } 4458185029Spjd 4459185029Spjd /* if we have come back to the start, bail out */ 4460185029Spjd if (first == NULL) 4461185029Spjd first = next; 4462185029Spjd else if (next == first) 4463185029Spjd break; 4464185029Spjd 4465185029Spjd } while (vdev_is_dead(next->l2ad_vdev)); 4466185029Spjd 4467185029Spjd /* if we were unable to find any usable vdevs, return NULL */ 4468185029Spjd if (vdev_is_dead(next->l2ad_vdev)) 4469185029Spjd next = NULL; 4470185029Spjd 4471185029Spjd l2arc_dev_last = next; 4472185029Spjd 4473185029Spjdout: 4474185029Spjd mutex_exit(&l2arc_dev_mtx); 4475185029Spjd 4476185029Spjd /* 4477185029Spjd * Grab the config lock to prevent the 'next' device from being 4478185029Spjd * removed while we are writing to it. 4479185029Spjd */ 4480185029Spjd if (next != NULL) 4481185029Spjd spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4482185029Spjd mutex_exit(&spa_namespace_lock); 4483185029Spjd 4484185029Spjd return (next); 4485185029Spjd} 4486185029Spjd 4487185029Spjd/* 4488185029Spjd * Free buffers that were tagged for destruction. 4489185029Spjd */ 4490185029Spjdstatic void 4491185029Spjdl2arc_do_free_on_write() 4492185029Spjd{ 4493185029Spjd list_t *buflist; 4494185029Spjd l2arc_data_free_t *df, *df_prev; 4495185029Spjd 4496185029Spjd mutex_enter(&l2arc_free_on_write_mtx); 4497185029Spjd buflist = l2arc_free_on_write; 4498185029Spjd 4499185029Spjd for (df = list_tail(buflist); df; df = df_prev) { 4500185029Spjd df_prev = list_prev(buflist, df); 4501185029Spjd ASSERT(df->l2df_data != NULL); 4502185029Spjd ASSERT(df->l2df_func != NULL); 4503185029Spjd df->l2df_func(df->l2df_data, df->l2df_size); 4504185029Spjd list_remove(buflist, df); 4505185029Spjd kmem_free(df, sizeof (l2arc_data_free_t)); 4506185029Spjd } 4507185029Spjd 4508185029Spjd mutex_exit(&l2arc_free_on_write_mtx); 4509185029Spjd} 4510185029Spjd 4511185029Spjd/* 4512185029Spjd * A write to a cache device has completed. Update all headers to allow 4513185029Spjd * reads from these buffers to begin. 4514185029Spjd */ 4515185029Spjdstatic void 4516185029Spjdl2arc_write_done(zio_t *zio) 4517185029Spjd{ 4518185029Spjd l2arc_write_callback_t *cb; 4519185029Spjd l2arc_dev_t *dev; 4520185029Spjd list_t *buflist; 4521185029Spjd arc_buf_hdr_t *head, *ab, *ab_prev; 4522185029Spjd l2arc_buf_hdr_t *abl2; 4523185029Spjd kmutex_t *hash_lock; 4524185029Spjd 4525185029Spjd cb = zio->io_private; 4526185029Spjd ASSERT(cb != NULL); 4527185029Spjd dev = cb->l2wcb_dev; 4528185029Spjd ASSERT(dev != NULL); 4529185029Spjd head = cb->l2wcb_head; 4530185029Spjd ASSERT(head != NULL); 4531185029Spjd buflist = dev->l2ad_buflist; 4532185029Spjd ASSERT(buflist != NULL); 4533185029Spjd DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4534185029Spjd l2arc_write_callback_t *, cb); 4535185029Spjd 4536185029Spjd if (zio->io_error != 0) 4537185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_error); 4538185029Spjd 4539185029Spjd mutex_enter(&l2arc_buflist_mtx); 4540185029Spjd 4541185029Spjd /* 4542185029Spjd * All writes completed, or an error was hit. 4543185029Spjd */ 4544185029Spjd for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 4545185029Spjd ab_prev = list_prev(buflist, ab); 4546185029Spjd 4547185029Spjd hash_lock = HDR_LOCK(ab); 4548185029Spjd if (!mutex_tryenter(hash_lock)) { 4549185029Spjd /* 4550185029Spjd * This buffer misses out. It may be in a stage 4551185029Spjd * of eviction. Its ARC_L2_WRITING flag will be 4552185029Spjd * left set, denying reads to this buffer. 4553185029Spjd */ 4554185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4555185029Spjd continue; 4556185029Spjd } 4557185029Spjd 4558251478Sdelphij abl2 = ab->b_l2hdr; 4559251478Sdelphij 4560251478Sdelphij /* 4561251478Sdelphij * Release the temporary compressed buffer as soon as possible. 4562251478Sdelphij */ 4563251478Sdelphij if (abl2->b_compress != ZIO_COMPRESS_OFF) 4564251478Sdelphij l2arc_release_cdata_buf(ab); 4565251478Sdelphij 4566185029Spjd if (zio->io_error != 0) { 4567185029Spjd /* 4568185029Spjd * Error - drop L2ARC entry. 4569185029Spjd */ 4570185029Spjd list_remove(buflist, ab); 4571251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4572185029Spjd ab->b_l2hdr = NULL; 4573248572Ssmh trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4574248574Ssmh ab->b_size, 0); 4575185029Spjd kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4576185029Spjd ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4577185029Spjd } 4578185029Spjd 4579185029Spjd /* 4580185029Spjd * Allow ARC to begin reads to this L2ARC entry. 4581185029Spjd */ 4582185029Spjd ab->b_flags &= ~ARC_L2_WRITING; 4583185029Spjd 4584185029Spjd mutex_exit(hash_lock); 4585185029Spjd } 4586185029Spjd 4587185029Spjd atomic_inc_64(&l2arc_writes_done); 4588185029Spjd list_remove(buflist, head); 4589185029Spjd kmem_cache_free(hdr_cache, head); 4590185029Spjd mutex_exit(&l2arc_buflist_mtx); 4591185029Spjd 4592185029Spjd l2arc_do_free_on_write(); 4593185029Spjd 4594185029Spjd kmem_free(cb, sizeof (l2arc_write_callback_t)); 4595185029Spjd} 4596185029Spjd 4597185029Spjd/* 4598185029Spjd * A read to a cache device completed. Validate buffer contents before 4599185029Spjd * handing over to the regular ARC routines. 4600185029Spjd */ 4601185029Spjdstatic void 4602185029Spjdl2arc_read_done(zio_t *zio) 4603185029Spjd{ 4604185029Spjd l2arc_read_callback_t *cb; 4605185029Spjd arc_buf_hdr_t *hdr; 4606185029Spjd arc_buf_t *buf; 4607185029Spjd kmutex_t *hash_lock; 4608185029Spjd int equal; 4609185029Spjd 4610185029Spjd ASSERT(zio->io_vd != NULL); 4611185029Spjd ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4612185029Spjd 4613185029Spjd spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4614185029Spjd 4615185029Spjd cb = zio->io_private; 4616185029Spjd ASSERT(cb != NULL); 4617185029Spjd buf = cb->l2rcb_buf; 4618185029Spjd ASSERT(buf != NULL); 4619185029Spjd 4620219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 4621185029Spjd mutex_enter(hash_lock); 4622219089Spjd hdr = buf->b_hdr; 4623219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4624185029Spjd 4625185029Spjd /* 4626251478Sdelphij * If the buffer was compressed, decompress it first. 4627251478Sdelphij */ 4628251478Sdelphij if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4629251478Sdelphij l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4630251478Sdelphij ASSERT(zio->io_data != NULL); 4631251478Sdelphij 4632251478Sdelphij /* 4633185029Spjd * Check this survived the L2ARC journey. 4634185029Spjd */ 4635185029Spjd equal = arc_cksum_equal(buf); 4636185029Spjd if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4637185029Spjd mutex_exit(hash_lock); 4638185029Spjd zio->io_private = buf; 4639185029Spjd zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4640185029Spjd zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4641185029Spjd arc_read_done(zio); 4642185029Spjd } else { 4643185029Spjd mutex_exit(hash_lock); 4644185029Spjd /* 4645185029Spjd * Buffer didn't survive caching. Increment stats and 4646185029Spjd * reissue to the original storage device. 4647185029Spjd */ 4648185029Spjd if (zio->io_error != 0) { 4649185029Spjd ARCSTAT_BUMP(arcstat_l2_io_error); 4650185029Spjd } else { 4651249195Smm zio->io_error = SET_ERROR(EIO); 4652185029Spjd } 4653185029Spjd if (!equal) 4654185029Spjd ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4655185029Spjd 4656185029Spjd /* 4657185029Spjd * If there's no waiter, issue an async i/o to the primary 4658185029Spjd * storage now. If there *is* a waiter, the caller must 4659185029Spjd * issue the i/o in a context where it's OK to block. 4660185029Spjd */ 4661209962Smm if (zio->io_waiter == NULL) { 4662209962Smm zio_t *pio = zio_unique_parent(zio); 4663209962Smm 4664209962Smm ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4665209962Smm 4666209962Smm zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4667185029Spjd buf->b_data, zio->io_size, arc_read_done, buf, 4668185029Spjd zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4669209962Smm } 4670185029Spjd } 4671185029Spjd 4672185029Spjd kmem_free(cb, sizeof (l2arc_read_callback_t)); 4673185029Spjd} 4674185029Spjd 4675185029Spjd/* 4676185029Spjd * This is the list priority from which the L2ARC will search for pages to 4677185029Spjd * cache. This is used within loops (0..3) to cycle through lists in the 4678185029Spjd * desired order. This order can have a significant effect on cache 4679185029Spjd * performance. 4680185029Spjd * 4681185029Spjd * Currently the metadata lists are hit first, MFU then MRU, followed by 4682185029Spjd * the data lists. This function returns a locked list, and also returns 4683185029Spjd * the lock pointer. 4684185029Spjd */ 4685185029Spjdstatic list_t * 4686185029Spjdl2arc_list_locked(int list_num, kmutex_t **lock) 4687185029Spjd{ 4688247187Smm list_t *list = NULL; 4689205231Skmacy int idx; 4690185029Spjd 4691206796Spjd ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4692206796Spjd 4693205231Skmacy if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4694205231Skmacy idx = list_num; 4695205231Skmacy list = &arc_mfu->arcs_lists[idx]; 4696205231Skmacy *lock = ARCS_LOCK(arc_mfu, idx); 4697206796Spjd } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4698205231Skmacy idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4699205231Skmacy list = &arc_mru->arcs_lists[idx]; 4700205231Skmacy *lock = ARCS_LOCK(arc_mru, idx); 4701206796Spjd } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4702205231Skmacy ARC_BUFC_NUMDATALISTS)) { 4703205231Skmacy idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4704205231Skmacy list = &arc_mfu->arcs_lists[idx]; 4705205231Skmacy *lock = ARCS_LOCK(arc_mfu, idx); 4706205231Skmacy } else { 4707205231Skmacy idx = list_num - ARC_BUFC_NUMLISTS; 4708205231Skmacy list = &arc_mru->arcs_lists[idx]; 4709205231Skmacy *lock = ARCS_LOCK(arc_mru, idx); 4710185029Spjd } 4711185029Spjd 4712185029Spjd ASSERT(!(MUTEX_HELD(*lock))); 4713185029Spjd mutex_enter(*lock); 4714185029Spjd return (list); 4715185029Spjd} 4716185029Spjd 4717185029Spjd/* 4718185029Spjd * Evict buffers from the device write hand to the distance specified in 4719185029Spjd * bytes. This distance may span populated buffers, it may span nothing. 4720185029Spjd * This is clearing a region on the L2ARC device ready for writing. 4721185029Spjd * If the 'all' boolean is set, every buffer is evicted. 4722185029Spjd */ 4723185029Spjdstatic void 4724185029Spjdl2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4725185029Spjd{ 4726185029Spjd list_t *buflist; 4727185029Spjd l2arc_buf_hdr_t *abl2; 4728185029Spjd arc_buf_hdr_t *ab, *ab_prev; 4729185029Spjd kmutex_t *hash_lock; 4730185029Spjd uint64_t taddr; 4731185029Spjd 4732185029Spjd buflist = dev->l2ad_buflist; 4733185029Spjd 4734185029Spjd if (buflist == NULL) 4735185029Spjd return; 4736185029Spjd 4737185029Spjd if (!all && dev->l2ad_first) { 4738185029Spjd /* 4739185029Spjd * This is the first sweep through the device. There is 4740185029Spjd * nothing to evict. 4741185029Spjd */ 4742185029Spjd return; 4743185029Spjd } 4744185029Spjd 4745185029Spjd if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4746185029Spjd /* 4747185029Spjd * When nearing the end of the device, evict to the end 4748185029Spjd * before the device write hand jumps to the start. 4749185029Spjd */ 4750185029Spjd taddr = dev->l2ad_end; 4751185029Spjd } else { 4752185029Spjd taddr = dev->l2ad_hand + distance; 4753185029Spjd } 4754185029Spjd DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4755185029Spjd uint64_t, taddr, boolean_t, all); 4756185029Spjd 4757185029Spjdtop: 4758185029Spjd mutex_enter(&l2arc_buflist_mtx); 4759185029Spjd for (ab = list_tail(buflist); ab; ab = ab_prev) { 4760185029Spjd ab_prev = list_prev(buflist, ab); 4761185029Spjd 4762185029Spjd hash_lock = HDR_LOCK(ab); 4763185029Spjd if (!mutex_tryenter(hash_lock)) { 4764185029Spjd /* 4765185029Spjd * Missed the hash lock. Retry. 4766185029Spjd */ 4767185029Spjd ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4768185029Spjd mutex_exit(&l2arc_buflist_mtx); 4769185029Spjd mutex_enter(hash_lock); 4770185029Spjd mutex_exit(hash_lock); 4771185029Spjd goto top; 4772185029Spjd } 4773185029Spjd 4774185029Spjd if (HDR_L2_WRITE_HEAD(ab)) { 4775185029Spjd /* 4776185029Spjd * We hit a write head node. Leave it for 4777185029Spjd * l2arc_write_done(). 4778185029Spjd */ 4779185029Spjd list_remove(buflist, ab); 4780185029Spjd mutex_exit(hash_lock); 4781185029Spjd continue; 4782185029Spjd } 4783185029Spjd 4784185029Spjd if (!all && ab->b_l2hdr != NULL && 4785185029Spjd (ab->b_l2hdr->b_daddr > taddr || 4786185029Spjd ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4787185029Spjd /* 4788185029Spjd * We've evicted to the target address, 4789185029Spjd * or the end of the device. 4790185029Spjd */ 4791185029Spjd mutex_exit(hash_lock); 4792185029Spjd break; 4793185029Spjd } 4794185029Spjd 4795185029Spjd if (HDR_FREE_IN_PROGRESS(ab)) { 4796185029Spjd /* 4797185029Spjd * Already on the path to destruction. 4798185029Spjd */ 4799185029Spjd mutex_exit(hash_lock); 4800185029Spjd continue; 4801185029Spjd } 4802185029Spjd 4803185029Spjd if (ab->b_state == arc_l2c_only) { 4804185029Spjd ASSERT(!HDR_L2_READING(ab)); 4805185029Spjd /* 4806185029Spjd * This doesn't exist in the ARC. Destroy. 4807185029Spjd * arc_hdr_destroy() will call list_remove() 4808185029Spjd * and decrement arcstat_l2_size. 4809185029Spjd */ 4810185029Spjd arc_change_state(arc_anon, ab, hash_lock); 4811185029Spjd arc_hdr_destroy(ab); 4812185029Spjd } else { 4813185029Spjd /* 4814185029Spjd * Invalidate issued or about to be issued 4815185029Spjd * reads, since we may be about to write 4816185029Spjd * over this location. 4817185029Spjd */ 4818185029Spjd if (HDR_L2_READING(ab)) { 4819185029Spjd ARCSTAT_BUMP(arcstat_l2_evict_reading); 4820185029Spjd ab->b_flags |= ARC_L2_EVICTED; 4821185029Spjd } 4822185029Spjd 4823185029Spjd /* 4824185029Spjd * Tell ARC this no longer exists in L2ARC. 4825185029Spjd */ 4826185029Spjd if (ab->b_l2hdr != NULL) { 4827185029Spjd abl2 = ab->b_l2hdr; 4828251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4829185029Spjd ab->b_l2hdr = NULL; 4830185029Spjd kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4831185029Spjd ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4832185029Spjd } 4833185029Spjd list_remove(buflist, ab); 4834185029Spjd 4835185029Spjd /* 4836185029Spjd * This may have been leftover after a 4837185029Spjd * failed write. 4838185029Spjd */ 4839185029Spjd ab->b_flags &= ~ARC_L2_WRITING; 4840185029Spjd } 4841185029Spjd mutex_exit(hash_lock); 4842185029Spjd } 4843185029Spjd mutex_exit(&l2arc_buflist_mtx); 4844185029Spjd 4845219089Spjd vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0); 4846185029Spjd dev->l2ad_evict = taddr; 4847185029Spjd} 4848185029Spjd 4849185029Spjd/* 4850185029Spjd * Find and write ARC buffers to the L2ARC device. 4851185029Spjd * 4852185029Spjd * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 4853185029Spjd * for reading until they have completed writing. 4854251478Sdelphij * The headroom_boost is an in-out parameter used to maintain headroom boost 4855251478Sdelphij * state between calls to this function. 4856251478Sdelphij * 4857251478Sdelphij * Returns the number of bytes actually written (which may be smaller than 4858251478Sdelphij * the delta by which the device hand has changed due to alignment). 4859185029Spjd */ 4860208373Smmstatic uint64_t 4861251478Sdelphijl2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 4862251478Sdelphij boolean_t *headroom_boost) 4863185029Spjd{ 4864185029Spjd arc_buf_hdr_t *ab, *ab_prev, *head; 4865185029Spjd list_t *list; 4866251478Sdelphij uint64_t write_asize, write_psize, write_sz, headroom, 4867251478Sdelphij buf_compress_minsz; 4868185029Spjd void *buf_data; 4869251478Sdelphij kmutex_t *list_lock; 4870251478Sdelphij boolean_t full; 4871185029Spjd l2arc_write_callback_t *cb; 4872185029Spjd zio_t *pio, *wzio; 4873228103Smm uint64_t guid = spa_load_guid(spa); 4874251478Sdelphij const boolean_t do_headroom_boost = *headroom_boost; 4875185029Spjd int try; 4876185029Spjd 4877185029Spjd ASSERT(dev->l2ad_vdev != NULL); 4878185029Spjd 4879251478Sdelphij /* Lower the flag now, we might want to raise it again later. */ 4880251478Sdelphij *headroom_boost = B_FALSE; 4881251478Sdelphij 4882185029Spjd pio = NULL; 4883251478Sdelphij write_sz = write_asize = write_psize = 0; 4884185029Spjd full = B_FALSE; 4885185029Spjd head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 4886185029Spjd head->b_flags |= ARC_L2_WRITE_HEAD; 4887185029Spjd 4888205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 4889185029Spjd /* 4890251478Sdelphij * We will want to try to compress buffers that are at least 2x the 4891251478Sdelphij * device sector size. 4892251478Sdelphij */ 4893251478Sdelphij buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 4894251478Sdelphij 4895251478Sdelphij /* 4896185029Spjd * Copy buffers for L2ARC writing. 4897185029Spjd */ 4898185029Spjd mutex_enter(&l2arc_buflist_mtx); 4899206796Spjd for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 4900251478Sdelphij uint64_t passed_sz = 0; 4901251478Sdelphij 4902185029Spjd list = l2arc_list_locked(try, &list_lock); 4903205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 4904185029Spjd 4905185029Spjd /* 4906185029Spjd * L2ARC fast warmup. 4907185029Spjd * 4908185029Spjd * Until the ARC is warm and starts to evict, read from the 4909185029Spjd * head of the ARC lists rather than the tail. 4910185029Spjd */ 4911185029Spjd if (arc_warm == B_FALSE) 4912185029Spjd ab = list_head(list); 4913185029Spjd else 4914185029Spjd ab = list_tail(list); 4915206796Spjd if (ab == NULL) 4916205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 4917185029Spjd 4918251478Sdelphij headroom = target_sz * l2arc_headroom; 4919251478Sdelphij if (do_headroom_boost) 4920251478Sdelphij headroom = (headroom * l2arc_headroom_boost) / 100; 4921251478Sdelphij 4922185029Spjd for (; ab; ab = ab_prev) { 4923251478Sdelphij l2arc_buf_hdr_t *l2hdr; 4924251478Sdelphij kmutex_t *hash_lock; 4925251478Sdelphij uint64_t buf_sz; 4926251478Sdelphij 4927185029Spjd if (arc_warm == B_FALSE) 4928185029Spjd ab_prev = list_next(list, ab); 4929185029Spjd else 4930185029Spjd ab_prev = list_prev(list, ab); 4931205231Skmacy ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size); 4932206796Spjd 4933185029Spjd hash_lock = HDR_LOCK(ab); 4934251478Sdelphij if (!mutex_tryenter(hash_lock)) { 4935205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 4936185029Spjd /* 4937185029Spjd * Skip this buffer rather than waiting. 4938185029Spjd */ 4939185029Spjd continue; 4940185029Spjd } 4941185029Spjd 4942185029Spjd passed_sz += ab->b_size; 4943185029Spjd if (passed_sz > headroom) { 4944185029Spjd /* 4945185029Spjd * Searched too far. 4946185029Spjd */ 4947185029Spjd mutex_exit(hash_lock); 4948205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 4949185029Spjd break; 4950185029Spjd } 4951185029Spjd 4952209962Smm if (!l2arc_write_eligible(guid, ab)) { 4953185029Spjd mutex_exit(hash_lock); 4954185029Spjd continue; 4955185029Spjd } 4956185029Spjd 4957185029Spjd if ((write_sz + ab->b_size) > target_sz) { 4958185029Spjd full = B_TRUE; 4959185029Spjd mutex_exit(hash_lock); 4960205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_full); 4961185029Spjd break; 4962185029Spjd } 4963185029Spjd 4964185029Spjd if (pio == NULL) { 4965185029Spjd /* 4966185029Spjd * Insert a dummy header on the buflist so 4967185029Spjd * l2arc_write_done() can find where the 4968185029Spjd * write buffers begin without searching. 4969185029Spjd */ 4970185029Spjd list_insert_head(dev->l2ad_buflist, head); 4971185029Spjd 4972185029Spjd cb = kmem_alloc( 4973185029Spjd sizeof (l2arc_write_callback_t), KM_SLEEP); 4974185029Spjd cb->l2wcb_dev = dev; 4975185029Spjd cb->l2wcb_head = head; 4976185029Spjd pio = zio_root(spa, l2arc_write_done, cb, 4977185029Spjd ZIO_FLAG_CANFAIL); 4978205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_pios); 4979185029Spjd } 4980185029Spjd 4981185029Spjd /* 4982185029Spjd * Create and add a new L2ARC header. 4983185029Spjd */ 4984251478Sdelphij l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 4985251478Sdelphij l2hdr->b_dev = dev; 4986251478Sdelphij ab->b_flags |= ARC_L2_WRITING; 4987185029Spjd 4988251478Sdelphij /* 4989251478Sdelphij * Temporarily stash the data buffer in b_tmp_cdata. 4990251478Sdelphij * The subsequent write step will pick it up from 4991251478Sdelphij * there. This is because can't access ab->b_buf 4992251478Sdelphij * without holding the hash_lock, which we in turn 4993251478Sdelphij * can't access without holding the ARC list locks 4994251478Sdelphij * (which we want to avoid during compression/writing). 4995251478Sdelphij */ 4996251478Sdelphij l2hdr->b_compress = ZIO_COMPRESS_OFF; 4997251478Sdelphij l2hdr->b_asize = ab->b_size; 4998251478Sdelphij l2hdr->b_tmp_cdata = ab->b_buf->b_data; 4999251478Sdelphij 5000185029Spjd buf_sz = ab->b_size; 5001251478Sdelphij ab->b_l2hdr = l2hdr; 5002185029Spjd 5003251478Sdelphij list_insert_head(dev->l2ad_buflist, ab); 5004251478Sdelphij 5005185029Spjd /* 5006185029Spjd * Compute and store the buffer cksum before 5007185029Spjd * writing. On debug the cksum is verified first. 5008185029Spjd */ 5009185029Spjd arc_cksum_verify(ab->b_buf); 5010185029Spjd arc_cksum_compute(ab->b_buf, B_TRUE); 5011185029Spjd 5012185029Spjd mutex_exit(hash_lock); 5013185029Spjd 5014251478Sdelphij write_sz += buf_sz; 5015251478Sdelphij } 5016251478Sdelphij 5017251478Sdelphij mutex_exit(list_lock); 5018251478Sdelphij 5019251478Sdelphij if (full == B_TRUE) 5020251478Sdelphij break; 5021251478Sdelphij } 5022251478Sdelphij 5023251478Sdelphij /* No buffers selected for writing? */ 5024251478Sdelphij if (pio == NULL) { 5025251478Sdelphij ASSERT0(write_sz); 5026251478Sdelphij mutex_exit(&l2arc_buflist_mtx); 5027251478Sdelphij kmem_cache_free(hdr_cache, head); 5028251478Sdelphij return (0); 5029251478Sdelphij } 5030251478Sdelphij 5031251478Sdelphij /* 5032251478Sdelphij * Now start writing the buffers. We're starting at the write head 5033251478Sdelphij * and work backwards, retracing the course of the buffer selector 5034251478Sdelphij * loop above. 5035251478Sdelphij */ 5036251478Sdelphij for (ab = list_prev(dev->l2ad_buflist, head); ab; 5037251478Sdelphij ab = list_prev(dev->l2ad_buflist, ab)) { 5038251478Sdelphij l2arc_buf_hdr_t *l2hdr; 5039251478Sdelphij uint64_t buf_sz; 5040251478Sdelphij 5041251478Sdelphij /* 5042251478Sdelphij * We shouldn't need to lock the buffer here, since we flagged 5043251478Sdelphij * it as ARC_L2_WRITING in the previous step, but we must take 5044251478Sdelphij * care to only access its L2 cache parameters. In particular, 5045251478Sdelphij * ab->b_buf may be invalid by now due to ARC eviction. 5046251478Sdelphij */ 5047251478Sdelphij l2hdr = ab->b_l2hdr; 5048251478Sdelphij l2hdr->b_daddr = dev->l2ad_hand; 5049251478Sdelphij 5050251478Sdelphij if ((ab->b_flags & ARC_L2COMPRESS) && 5051251478Sdelphij l2hdr->b_asize >= buf_compress_minsz) { 5052251478Sdelphij if (l2arc_compress_buf(l2hdr)) { 5053251478Sdelphij /* 5054251478Sdelphij * If compression succeeded, enable headroom 5055251478Sdelphij * boost on the next scan cycle. 5056251478Sdelphij */ 5057251478Sdelphij *headroom_boost = B_TRUE; 5058251478Sdelphij } 5059251478Sdelphij } 5060251478Sdelphij 5061251478Sdelphij /* 5062251478Sdelphij * Pick up the buffer data we had previously stashed away 5063251478Sdelphij * (and now potentially also compressed). 5064251478Sdelphij */ 5065251478Sdelphij buf_data = l2hdr->b_tmp_cdata; 5066251478Sdelphij buf_sz = l2hdr->b_asize; 5067251478Sdelphij 5068251478Sdelphij /* Compression may have squashed the buffer to zero length. */ 5069251478Sdelphij if (buf_sz != 0) { 5070251478Sdelphij uint64_t buf_p_sz; 5071251478Sdelphij 5072185029Spjd wzio = zio_write_phys(pio, dev->l2ad_vdev, 5073185029Spjd dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5074185029Spjd NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5075185029Spjd ZIO_FLAG_CANFAIL, B_FALSE); 5076185029Spjd 5077185029Spjd DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5078185029Spjd zio_t *, wzio); 5079185029Spjd (void) zio_nowait(wzio); 5080185029Spjd 5081251478Sdelphij write_asize += buf_sz; 5082185029Spjd /* 5083185029Spjd * Keep the clock hand suitably device-aligned. 5084185029Spjd */ 5085251478Sdelphij buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5086251478Sdelphij write_psize += buf_p_sz; 5087251478Sdelphij dev->l2ad_hand += buf_p_sz; 5088185029Spjd } 5089251478Sdelphij } 5090185029Spjd 5091185029Spjd mutex_exit(&l2arc_buflist_mtx); 5092185029Spjd 5093251478Sdelphij ASSERT3U(write_asize, <=, target_sz); 5094185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_sent); 5095251478Sdelphij ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5096185029Spjd ARCSTAT_INCR(arcstat_l2_size, write_sz); 5097251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, write_asize); 5098251478Sdelphij vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); 5099185029Spjd 5100185029Spjd /* 5101185029Spjd * Bump device hand to the device start if it is approaching the end. 5102185029Spjd * l2arc_evict() will already have evicted ahead for this case. 5103185029Spjd */ 5104185029Spjd if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5105219089Spjd vdev_space_update(dev->l2ad_vdev, 5106219089Spjd dev->l2ad_end - dev->l2ad_hand, 0, 0); 5107185029Spjd dev->l2ad_hand = dev->l2ad_start; 5108185029Spjd dev->l2ad_evict = dev->l2ad_start; 5109185029Spjd dev->l2ad_first = B_FALSE; 5110185029Spjd } 5111185029Spjd 5112208373Smm dev->l2ad_writing = B_TRUE; 5113185029Spjd (void) zio_wait(pio); 5114208373Smm dev->l2ad_writing = B_FALSE; 5115208373Smm 5116251478Sdelphij return (write_asize); 5117185029Spjd} 5118185029Spjd 5119185029Spjd/* 5120251478Sdelphij * Compresses an L2ARC buffer. 5121251478Sdelphij * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5122251478Sdelphij * size in l2hdr->b_asize. This routine tries to compress the data and 5123251478Sdelphij * depending on the compression result there are three possible outcomes: 5124251478Sdelphij * *) The buffer was incompressible. The original l2hdr contents were left 5125251478Sdelphij * untouched and are ready for writing to an L2 device. 5126251478Sdelphij * *) The buffer was all-zeros, so there is no need to write it to an L2 5127251478Sdelphij * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5128251478Sdelphij * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5129251478Sdelphij * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5130251478Sdelphij * data buffer which holds the compressed data to be written, and b_asize 5131251478Sdelphij * tells us how much data there is. b_compress is set to the appropriate 5132251478Sdelphij * compression algorithm. Once writing is done, invoke 5133251478Sdelphij * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5134251478Sdelphij * 5135251478Sdelphij * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5136251478Sdelphij * buffer was incompressible). 5137251478Sdelphij */ 5138251478Sdelphijstatic boolean_t 5139251478Sdelphijl2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5140251478Sdelphij{ 5141251478Sdelphij void *cdata; 5142251478Sdelphij size_t csize, len; 5143251478Sdelphij 5144251478Sdelphij ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5145251478Sdelphij ASSERT(l2hdr->b_tmp_cdata != NULL); 5146251478Sdelphij 5147251478Sdelphij len = l2hdr->b_asize; 5148251478Sdelphij cdata = zio_data_buf_alloc(len); 5149251478Sdelphij csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5150256889Ssmh cdata, l2hdr->b_asize, (size_t)(1ULL << l2hdr->b_dev->l2ad_vdev->vdev_ashift)); 5151251478Sdelphij 5152251478Sdelphij if (csize == 0) { 5153251478Sdelphij /* zero block, indicate that there's nothing to write */ 5154251478Sdelphij zio_data_buf_free(cdata, len); 5155251478Sdelphij l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5156251478Sdelphij l2hdr->b_asize = 0; 5157251478Sdelphij l2hdr->b_tmp_cdata = NULL; 5158251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5159251478Sdelphij return (B_TRUE); 5160251478Sdelphij } else if (csize > 0 && csize < len) { 5161251478Sdelphij /* 5162251478Sdelphij * Compression succeeded, we'll keep the cdata around for 5163251478Sdelphij * writing and release it afterwards. 5164251478Sdelphij */ 5165251478Sdelphij l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5166251478Sdelphij l2hdr->b_asize = csize; 5167251478Sdelphij l2hdr->b_tmp_cdata = cdata; 5168251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_successes); 5169251478Sdelphij return (B_TRUE); 5170251478Sdelphij } else { 5171251478Sdelphij /* 5172251478Sdelphij * Compression failed, release the compressed buffer. 5173251478Sdelphij * l2hdr will be left unmodified. 5174251478Sdelphij */ 5175251478Sdelphij zio_data_buf_free(cdata, len); 5176251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_failures); 5177251478Sdelphij return (B_FALSE); 5178251478Sdelphij } 5179251478Sdelphij} 5180251478Sdelphij 5181251478Sdelphij/* 5182251478Sdelphij * Decompresses a zio read back from an l2arc device. On success, the 5183251478Sdelphij * underlying zio's io_data buffer is overwritten by the uncompressed 5184251478Sdelphij * version. On decompression error (corrupt compressed stream), the 5185251478Sdelphij * zio->io_error value is set to signal an I/O error. 5186251478Sdelphij * 5187251478Sdelphij * Please note that the compressed data stream is not checksummed, so 5188251478Sdelphij * if the underlying device is experiencing data corruption, we may feed 5189251478Sdelphij * corrupt data to the decompressor, so the decompressor needs to be 5190251478Sdelphij * able to handle this situation (LZ4 does). 5191251478Sdelphij */ 5192251478Sdelphijstatic void 5193251478Sdelphijl2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5194251478Sdelphij{ 5195251478Sdelphij ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5196251478Sdelphij 5197251478Sdelphij if (zio->io_error != 0) { 5198251478Sdelphij /* 5199251478Sdelphij * An io error has occured, just restore the original io 5200251478Sdelphij * size in preparation for a main pool read. 5201251478Sdelphij */ 5202251478Sdelphij zio->io_orig_size = zio->io_size = hdr->b_size; 5203251478Sdelphij return; 5204251478Sdelphij } 5205251478Sdelphij 5206251478Sdelphij if (c == ZIO_COMPRESS_EMPTY) { 5207251478Sdelphij /* 5208251478Sdelphij * An empty buffer results in a null zio, which means we 5209251478Sdelphij * need to fill its io_data after we're done restoring the 5210251478Sdelphij * buffer's contents. 5211251478Sdelphij */ 5212251478Sdelphij ASSERT(hdr->b_buf != NULL); 5213251478Sdelphij bzero(hdr->b_buf->b_data, hdr->b_size); 5214251478Sdelphij zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5215251478Sdelphij } else { 5216251478Sdelphij ASSERT(zio->io_data != NULL); 5217251478Sdelphij /* 5218251478Sdelphij * We copy the compressed data from the start of the arc buffer 5219251478Sdelphij * (the zio_read will have pulled in only what we need, the 5220251478Sdelphij * rest is garbage which we will overwrite at decompression) 5221251478Sdelphij * and then decompress back to the ARC data buffer. This way we 5222251478Sdelphij * can minimize copying by simply decompressing back over the 5223251478Sdelphij * original compressed data (rather than decompressing to an 5224251478Sdelphij * aux buffer and then copying back the uncompressed buffer, 5225251478Sdelphij * which is likely to be much larger). 5226251478Sdelphij */ 5227251478Sdelphij uint64_t csize; 5228251478Sdelphij void *cdata; 5229251478Sdelphij 5230251478Sdelphij csize = zio->io_size; 5231251478Sdelphij cdata = zio_data_buf_alloc(csize); 5232251478Sdelphij bcopy(zio->io_data, cdata, csize); 5233251478Sdelphij if (zio_decompress_data(c, cdata, zio->io_data, csize, 5234251478Sdelphij hdr->b_size) != 0) 5235251478Sdelphij zio->io_error = EIO; 5236251478Sdelphij zio_data_buf_free(cdata, csize); 5237251478Sdelphij } 5238251478Sdelphij 5239251478Sdelphij /* Restore the expected uncompressed IO size. */ 5240251478Sdelphij zio->io_orig_size = zio->io_size = hdr->b_size; 5241251478Sdelphij} 5242251478Sdelphij 5243251478Sdelphij/* 5244251478Sdelphij * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5245251478Sdelphij * This buffer serves as a temporary holder of compressed data while 5246251478Sdelphij * the buffer entry is being written to an l2arc device. Once that is 5247251478Sdelphij * done, we can dispose of it. 5248251478Sdelphij */ 5249251478Sdelphijstatic void 5250251478Sdelphijl2arc_release_cdata_buf(arc_buf_hdr_t *ab) 5251251478Sdelphij{ 5252251478Sdelphij l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr; 5253251478Sdelphij 5254251478Sdelphij if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) { 5255251478Sdelphij /* 5256251478Sdelphij * If the data was compressed, then we've allocated a 5257251478Sdelphij * temporary buffer for it, so now we need to release it. 5258251478Sdelphij */ 5259251478Sdelphij ASSERT(l2hdr->b_tmp_cdata != NULL); 5260251478Sdelphij zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size); 5261251478Sdelphij } 5262251478Sdelphij l2hdr->b_tmp_cdata = NULL; 5263251478Sdelphij} 5264251478Sdelphij 5265251478Sdelphij/* 5266185029Spjd * This thread feeds the L2ARC at regular intervals. This is the beating 5267185029Spjd * heart of the L2ARC. 5268185029Spjd */ 5269185029Spjdstatic void 5270185029Spjdl2arc_feed_thread(void *dummy __unused) 5271185029Spjd{ 5272185029Spjd callb_cpr_t cpr; 5273185029Spjd l2arc_dev_t *dev; 5274185029Spjd spa_t *spa; 5275208373Smm uint64_t size, wrote; 5276219089Spjd clock_t begin, next = ddi_get_lbolt(); 5277251478Sdelphij boolean_t headroom_boost = B_FALSE; 5278185029Spjd 5279185029Spjd CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5280185029Spjd 5281185029Spjd mutex_enter(&l2arc_feed_thr_lock); 5282185029Spjd 5283185029Spjd while (l2arc_thread_exit == 0) { 5284185029Spjd CALLB_CPR_SAFE_BEGIN(&cpr); 5285185029Spjd (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5286219089Spjd next - ddi_get_lbolt()); 5287185029Spjd CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5288219089Spjd next = ddi_get_lbolt() + hz; 5289185029Spjd 5290185029Spjd /* 5291185029Spjd * Quick check for L2ARC devices. 5292185029Spjd */ 5293185029Spjd mutex_enter(&l2arc_dev_mtx); 5294185029Spjd if (l2arc_ndev == 0) { 5295185029Spjd mutex_exit(&l2arc_dev_mtx); 5296185029Spjd continue; 5297185029Spjd } 5298185029Spjd mutex_exit(&l2arc_dev_mtx); 5299219089Spjd begin = ddi_get_lbolt(); 5300185029Spjd 5301185029Spjd /* 5302185029Spjd * This selects the next l2arc device to write to, and in 5303185029Spjd * doing so the next spa to feed from: dev->l2ad_spa. This 5304185029Spjd * will return NULL if there are now no l2arc devices or if 5305185029Spjd * they are all faulted. 5306185029Spjd * 5307185029Spjd * If a device is returned, its spa's config lock is also 5308185029Spjd * held to prevent device removal. l2arc_dev_get_next() 5309185029Spjd * will grab and release l2arc_dev_mtx. 5310185029Spjd */ 5311185029Spjd if ((dev = l2arc_dev_get_next()) == NULL) 5312185029Spjd continue; 5313185029Spjd 5314185029Spjd spa = dev->l2ad_spa; 5315185029Spjd ASSERT(spa != NULL); 5316185029Spjd 5317185029Spjd /* 5318219089Spjd * If the pool is read-only then force the feed thread to 5319219089Spjd * sleep a little longer. 5320219089Spjd */ 5321219089Spjd if (!spa_writeable(spa)) { 5322219089Spjd next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5323219089Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5324219089Spjd continue; 5325219089Spjd } 5326219089Spjd 5327219089Spjd /* 5328185029Spjd * Avoid contributing to memory pressure. 5329185029Spjd */ 5330185029Spjd if (arc_reclaim_needed()) { 5331185029Spjd ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5332185029Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5333185029Spjd continue; 5334185029Spjd } 5335185029Spjd 5336185029Spjd ARCSTAT_BUMP(arcstat_l2_feeds); 5337185029Spjd 5338251478Sdelphij size = l2arc_write_size(); 5339185029Spjd 5340185029Spjd /* 5341185029Spjd * Evict L2ARC buffers that will be overwritten. 5342185029Spjd */ 5343185029Spjd l2arc_evict(dev, size, B_FALSE); 5344185029Spjd 5345185029Spjd /* 5346185029Spjd * Write ARC buffers. 5347185029Spjd */ 5348251478Sdelphij wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5349208373Smm 5350208373Smm /* 5351208373Smm * Calculate interval between writes. 5352208373Smm */ 5353208373Smm next = l2arc_write_interval(begin, size, wrote); 5354185029Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5355185029Spjd } 5356185029Spjd 5357185029Spjd l2arc_thread_exit = 0; 5358185029Spjd cv_broadcast(&l2arc_feed_thr_cv); 5359185029Spjd CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5360185029Spjd thread_exit(); 5361185029Spjd} 5362185029Spjd 5363185029Spjdboolean_t 5364185029Spjdl2arc_vdev_present(vdev_t *vd) 5365185029Spjd{ 5366185029Spjd l2arc_dev_t *dev; 5367185029Spjd 5368185029Spjd mutex_enter(&l2arc_dev_mtx); 5369185029Spjd for (dev = list_head(l2arc_dev_list); dev != NULL; 5370185029Spjd dev = list_next(l2arc_dev_list, dev)) { 5371185029Spjd if (dev->l2ad_vdev == vd) 5372185029Spjd break; 5373185029Spjd } 5374185029Spjd mutex_exit(&l2arc_dev_mtx); 5375185029Spjd 5376185029Spjd return (dev != NULL); 5377185029Spjd} 5378185029Spjd 5379185029Spjd/* 5380185029Spjd * Add a vdev for use by the L2ARC. By this point the spa has already 5381185029Spjd * validated the vdev and opened it. 5382185029Spjd */ 5383185029Spjdvoid 5384219089Spjdl2arc_add_vdev(spa_t *spa, vdev_t *vd) 5385185029Spjd{ 5386185029Spjd l2arc_dev_t *adddev; 5387185029Spjd 5388185029Spjd ASSERT(!l2arc_vdev_present(vd)); 5389185029Spjd 5390255753Sgibbs vdev_ashift_optimize(vd); 5391255753Sgibbs 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