arc.c revision 268075
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. 23249195Smm * Copyright (c) 2013 by Delphix. All rights reserved. 24260835Sdelphij * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 25258389Savg * Copyright 2013 Nexenta Systems, Inc. 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> 130258632Savg#include <sys/dsl_pool.h> 131168404Spjd#ifdef _KERNEL 132168404Spjd#include <sys/dnlc.h> 133168404Spjd#endif 134168404Spjd#include <sys/callb.h> 135168404Spjd#include <sys/kstat.h> 136248572Ssmh#include <sys/trim_map.h> 137219089Spjd#include <zfs_fletcher.h> 138168404Spjd#include <sys/sdt.h> 139168404Spjd 140191902Skmacy#include <vm/vm_pageout.h> 141191902Skmacy 142240133Smm#ifdef illumos 143240133Smm#ifndef _KERNEL 144240133Smm/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 145240133Smmboolean_t arc_watch = B_FALSE; 146240133Smmint arc_procfd; 147240133Smm#endif 148240133Smm#endif /* illumos */ 149240133Smm 150168404Spjdstatic kmutex_t arc_reclaim_thr_lock; 151168404Spjdstatic kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 152168404Spjdstatic uint8_t arc_thread_exit; 153168404Spjd 154168404Spjd#define ARC_REDUCE_DNLC_PERCENT 3 155168404Spjduint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 156168404Spjd 157168404Spjdtypedef enum arc_reclaim_strategy { 158168404Spjd ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 159168404Spjd ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 160168404Spjd} arc_reclaim_strategy_t; 161168404Spjd 162258632Savg/* 163258632Savg * The number of iterations through arc_evict_*() before we 164258632Savg * drop & reacquire the lock. 165258632Savg */ 166258632Savgint arc_evict_iterations = 100; 167258632Savg 168168404Spjd/* number of seconds before growing cache again */ 169168404Spjdstatic int arc_grow_retry = 60; 170168404Spjd 171208373Smm/* shift of arc_c for calculating both min and max arc_p */ 172208373Smmstatic int arc_p_min_shift = 4; 173208373Smm 174208373Smm/* log2(fraction of arc to reclaim) */ 175208373Smmstatic int arc_shrink_shift = 5; 176208373Smm 177168404Spjd/* 178168404Spjd * minimum lifespan of a prefetch block in clock ticks 179168404Spjd * (initialized in arc_init()) 180168404Spjd */ 181168404Spjdstatic int arc_min_prefetch_lifespan; 182168404Spjd 183258632Savg/* 184258632Savg * If this percent of memory is free, don't throttle. 185258632Savg */ 186258632Savgint arc_lotsfree_percent = 10; 187258632Savg 188208373Smmstatic int arc_dead; 189194043Skmacyextern int zfs_prefetch_disable; 190168404Spjd 191168404Spjd/* 192185029Spjd * The arc has filled available memory and has now warmed up. 193185029Spjd */ 194185029Spjdstatic boolean_t arc_warm; 195185029Spjd 196185029Spjd/* 197168404Spjd * These tunables are for performance analysis. 198168404Spjd */ 199185029Spjduint64_t zfs_arc_max; 200185029Spjduint64_t zfs_arc_min; 201185029Spjduint64_t zfs_arc_meta_limit = 0; 202208373Smmint zfs_arc_grow_retry = 0; 203208373Smmint zfs_arc_shrink_shift = 0; 204208373Smmint zfs_arc_p_min_shift = 0; 205242845Sdelphijint zfs_disable_dup_eviction = 0; 206185029Spjd 207185029SpjdTUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 208168473SpjdSYSCTL_DECL(_vfs_zfs); 209217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 210168473Spjd "Maximum ARC size"); 211217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 212168473Spjd "Minimum ARC size"); 213168404Spjd 214168404Spjd/* 215185029Spjd * Note that buffers can be in one of 6 states: 216168404Spjd * ARC_anon - anonymous (discussed below) 217168404Spjd * ARC_mru - recently used, currently cached 218168404Spjd * ARC_mru_ghost - recentely used, no longer in cache 219168404Spjd * ARC_mfu - frequently used, currently cached 220168404Spjd * ARC_mfu_ghost - frequently used, no longer in cache 221185029Spjd * ARC_l2c_only - exists in L2ARC but not other states 222185029Spjd * When there are no active references to the buffer, they are 223185029Spjd * are linked onto a list in one of these arc states. These are 224185029Spjd * the only buffers that can be evicted or deleted. Within each 225185029Spjd * state there are multiple lists, one for meta-data and one for 226185029Spjd * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 227185029Spjd * etc.) is tracked separately so that it can be managed more 228185029Spjd * explicitly: favored over data, limited explicitly. 229168404Spjd * 230168404Spjd * Anonymous buffers are buffers that are not associated with 231168404Spjd * a DVA. These are buffers that hold dirty block copies 232168404Spjd * before they are written to stable storage. By definition, 233168404Spjd * they are "ref'd" and are considered part of arc_mru 234168404Spjd * that cannot be freed. Generally, they will aquire a DVA 235168404Spjd * as they are written and migrate onto the arc_mru list. 236185029Spjd * 237185029Spjd * The ARC_l2c_only state is for buffers that are in the second 238185029Spjd * level ARC but no longer in any of the ARC_m* lists. The second 239185029Spjd * level ARC itself may also contain buffers that are in any of 240185029Spjd * the ARC_m* states - meaning that a buffer can exist in two 241185029Spjd * places. The reason for the ARC_l2c_only state is to keep the 242185029Spjd * buffer header in the hash table, so that reads that hit the 243185029Spjd * second level ARC benefit from these fast lookups. 244168404Spjd */ 245168404Spjd 246205264Skmacy#define ARCS_LOCK_PAD CACHE_LINE_SIZE 247205231Skmacystruct arcs_lock { 248205231Skmacy kmutex_t arcs_lock; 249205231Skmacy#ifdef _KERNEL 250205231Skmacy unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 251205231Skmacy#endif 252205231Skmacy}; 253205231Skmacy 254205231Skmacy/* 255205231Skmacy * must be power of two for mask use to work 256205231Skmacy * 257205231Skmacy */ 258205231Skmacy#define ARC_BUFC_NUMDATALISTS 16 259205231Skmacy#define ARC_BUFC_NUMMETADATALISTS 16 260206796Spjd#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 261205231Skmacy 262168404Spjdtypedef struct arc_state { 263185029Spjd uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 264185029Spjd uint64_t arcs_size; /* total amount of data in this state */ 265205231Skmacy list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 266205264Skmacy struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 267168404Spjd} arc_state_t; 268168404Spjd 269206796Spjd#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 270205231Skmacy 271185029Spjd/* The 6 states: */ 272168404Spjdstatic arc_state_t ARC_anon; 273168404Spjdstatic arc_state_t ARC_mru; 274168404Spjdstatic arc_state_t ARC_mru_ghost; 275168404Spjdstatic arc_state_t ARC_mfu; 276168404Spjdstatic arc_state_t ARC_mfu_ghost; 277185029Spjdstatic arc_state_t ARC_l2c_only; 278168404Spjd 279168404Spjdtypedef struct arc_stats { 280168404Spjd kstat_named_t arcstat_hits; 281168404Spjd kstat_named_t arcstat_misses; 282168404Spjd kstat_named_t arcstat_demand_data_hits; 283168404Spjd kstat_named_t arcstat_demand_data_misses; 284168404Spjd kstat_named_t arcstat_demand_metadata_hits; 285168404Spjd kstat_named_t arcstat_demand_metadata_misses; 286168404Spjd kstat_named_t arcstat_prefetch_data_hits; 287168404Spjd kstat_named_t arcstat_prefetch_data_misses; 288168404Spjd kstat_named_t arcstat_prefetch_metadata_hits; 289168404Spjd kstat_named_t arcstat_prefetch_metadata_misses; 290168404Spjd kstat_named_t arcstat_mru_hits; 291168404Spjd kstat_named_t arcstat_mru_ghost_hits; 292168404Spjd kstat_named_t arcstat_mfu_hits; 293168404Spjd kstat_named_t arcstat_mfu_ghost_hits; 294205231Skmacy kstat_named_t arcstat_allocated; 295168404Spjd kstat_named_t arcstat_deleted; 296205231Skmacy kstat_named_t arcstat_stolen; 297168404Spjd kstat_named_t arcstat_recycle_miss; 298251629Sdelphij /* 299251629Sdelphij * Number of buffers that could not be evicted because the hash lock 300251629Sdelphij * was held by another thread. The lock may not necessarily be held 301251629Sdelphij * by something using the same buffer, since hash locks are shared 302251629Sdelphij * by multiple buffers. 303251629Sdelphij */ 304168404Spjd kstat_named_t arcstat_mutex_miss; 305251629Sdelphij /* 306251629Sdelphij * Number of buffers skipped because they have I/O in progress, are 307251629Sdelphij * indrect prefetch buffers that have not lived long enough, or are 308251629Sdelphij * not from the spa we're trying to evict from. 309251629Sdelphij */ 310168404Spjd kstat_named_t arcstat_evict_skip; 311208373Smm kstat_named_t arcstat_evict_l2_cached; 312208373Smm kstat_named_t arcstat_evict_l2_eligible; 313208373Smm kstat_named_t arcstat_evict_l2_ineligible; 314168404Spjd kstat_named_t arcstat_hash_elements; 315168404Spjd kstat_named_t arcstat_hash_elements_max; 316168404Spjd kstat_named_t arcstat_hash_collisions; 317168404Spjd kstat_named_t arcstat_hash_chains; 318168404Spjd kstat_named_t arcstat_hash_chain_max; 319168404Spjd kstat_named_t arcstat_p; 320168404Spjd kstat_named_t arcstat_c; 321168404Spjd kstat_named_t arcstat_c_min; 322168404Spjd kstat_named_t arcstat_c_max; 323168404Spjd kstat_named_t arcstat_size; 324185029Spjd kstat_named_t arcstat_hdr_size; 325208373Smm kstat_named_t arcstat_data_size; 326208373Smm kstat_named_t arcstat_other_size; 327185029Spjd kstat_named_t arcstat_l2_hits; 328185029Spjd kstat_named_t arcstat_l2_misses; 329185029Spjd kstat_named_t arcstat_l2_feeds; 330185029Spjd kstat_named_t arcstat_l2_rw_clash; 331208373Smm kstat_named_t arcstat_l2_read_bytes; 332208373Smm kstat_named_t arcstat_l2_write_bytes; 333185029Spjd kstat_named_t arcstat_l2_writes_sent; 334185029Spjd kstat_named_t arcstat_l2_writes_done; 335185029Spjd kstat_named_t arcstat_l2_writes_error; 336185029Spjd kstat_named_t arcstat_l2_writes_hdr_miss; 337185029Spjd kstat_named_t arcstat_l2_evict_lock_retry; 338185029Spjd kstat_named_t arcstat_l2_evict_reading; 339185029Spjd kstat_named_t arcstat_l2_free_on_write; 340185029Spjd kstat_named_t arcstat_l2_abort_lowmem; 341185029Spjd kstat_named_t arcstat_l2_cksum_bad; 342185029Spjd kstat_named_t arcstat_l2_io_error; 343185029Spjd kstat_named_t arcstat_l2_size; 344251478Sdelphij kstat_named_t arcstat_l2_asize; 345185029Spjd kstat_named_t arcstat_l2_hdr_size; 346251478Sdelphij kstat_named_t arcstat_l2_compress_successes; 347251478Sdelphij kstat_named_t arcstat_l2_compress_zeros; 348251478Sdelphij kstat_named_t arcstat_l2_compress_failures; 349205231Skmacy kstat_named_t arcstat_l2_write_trylock_fail; 350205231Skmacy kstat_named_t arcstat_l2_write_passed_headroom; 351205231Skmacy kstat_named_t arcstat_l2_write_spa_mismatch; 352206796Spjd kstat_named_t arcstat_l2_write_in_l2; 353205231Skmacy kstat_named_t arcstat_l2_write_hdr_io_in_progress; 354205231Skmacy kstat_named_t arcstat_l2_write_not_cacheable; 355205231Skmacy kstat_named_t arcstat_l2_write_full; 356205231Skmacy kstat_named_t arcstat_l2_write_buffer_iter; 357205231Skmacy kstat_named_t arcstat_l2_write_pios; 358205231Skmacy kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 359205231Skmacy kstat_named_t arcstat_l2_write_buffer_list_iter; 360205231Skmacy kstat_named_t arcstat_l2_write_buffer_list_null_iter; 361242845Sdelphij kstat_named_t arcstat_memory_throttle_count; 362242845Sdelphij kstat_named_t arcstat_duplicate_buffers; 363242845Sdelphij kstat_named_t arcstat_duplicate_buffers_size; 364242845Sdelphij kstat_named_t arcstat_duplicate_reads; 365168404Spjd} arc_stats_t; 366168404Spjd 367168404Spjdstatic arc_stats_t arc_stats = { 368168404Spjd { "hits", KSTAT_DATA_UINT64 }, 369168404Spjd { "misses", KSTAT_DATA_UINT64 }, 370168404Spjd { "demand_data_hits", KSTAT_DATA_UINT64 }, 371168404Spjd { "demand_data_misses", KSTAT_DATA_UINT64 }, 372168404Spjd { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 373168404Spjd { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 374168404Spjd { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 375168404Spjd { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 376168404Spjd { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 377168404Spjd { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 378168404Spjd { "mru_hits", KSTAT_DATA_UINT64 }, 379168404Spjd { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 380168404Spjd { "mfu_hits", KSTAT_DATA_UINT64 }, 381168404Spjd { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 382205231Skmacy { "allocated", KSTAT_DATA_UINT64 }, 383168404Spjd { "deleted", KSTAT_DATA_UINT64 }, 384205231Skmacy { "stolen", KSTAT_DATA_UINT64 }, 385168404Spjd { "recycle_miss", KSTAT_DATA_UINT64 }, 386168404Spjd { "mutex_miss", KSTAT_DATA_UINT64 }, 387168404Spjd { "evict_skip", KSTAT_DATA_UINT64 }, 388208373Smm { "evict_l2_cached", KSTAT_DATA_UINT64 }, 389208373Smm { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 390208373Smm { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 391168404Spjd { "hash_elements", KSTAT_DATA_UINT64 }, 392168404Spjd { "hash_elements_max", KSTAT_DATA_UINT64 }, 393168404Spjd { "hash_collisions", KSTAT_DATA_UINT64 }, 394168404Spjd { "hash_chains", KSTAT_DATA_UINT64 }, 395168404Spjd { "hash_chain_max", KSTAT_DATA_UINT64 }, 396168404Spjd { "p", KSTAT_DATA_UINT64 }, 397168404Spjd { "c", KSTAT_DATA_UINT64 }, 398168404Spjd { "c_min", KSTAT_DATA_UINT64 }, 399168404Spjd { "c_max", KSTAT_DATA_UINT64 }, 400185029Spjd { "size", KSTAT_DATA_UINT64 }, 401185029Spjd { "hdr_size", KSTAT_DATA_UINT64 }, 402208373Smm { "data_size", KSTAT_DATA_UINT64 }, 403208373Smm { "other_size", KSTAT_DATA_UINT64 }, 404185029Spjd { "l2_hits", KSTAT_DATA_UINT64 }, 405185029Spjd { "l2_misses", KSTAT_DATA_UINT64 }, 406185029Spjd { "l2_feeds", KSTAT_DATA_UINT64 }, 407185029Spjd { "l2_rw_clash", KSTAT_DATA_UINT64 }, 408208373Smm { "l2_read_bytes", KSTAT_DATA_UINT64 }, 409208373Smm { "l2_write_bytes", KSTAT_DATA_UINT64 }, 410185029Spjd { "l2_writes_sent", KSTAT_DATA_UINT64 }, 411185029Spjd { "l2_writes_done", KSTAT_DATA_UINT64 }, 412185029Spjd { "l2_writes_error", KSTAT_DATA_UINT64 }, 413185029Spjd { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 414185029Spjd { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 415185029Spjd { "l2_evict_reading", KSTAT_DATA_UINT64 }, 416185029Spjd { "l2_free_on_write", KSTAT_DATA_UINT64 }, 417185029Spjd { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 418185029Spjd { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 419185029Spjd { "l2_io_error", KSTAT_DATA_UINT64 }, 420185029Spjd { "l2_size", KSTAT_DATA_UINT64 }, 421251478Sdelphij { "l2_asize", KSTAT_DATA_UINT64 }, 422185029Spjd { "l2_hdr_size", KSTAT_DATA_UINT64 }, 423251478Sdelphij { "l2_compress_successes", KSTAT_DATA_UINT64 }, 424251478Sdelphij { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 425251478Sdelphij { "l2_compress_failures", KSTAT_DATA_UINT64 }, 426206796Spjd { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 427206796Spjd { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 428206796Spjd { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 429206796Spjd { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 430206796Spjd { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 431206796Spjd { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 432206796Spjd { "l2_write_full", KSTAT_DATA_UINT64 }, 433206796Spjd { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 434206796Spjd { "l2_write_pios", KSTAT_DATA_UINT64 }, 435206796Spjd { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 436206796Spjd { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 437242845Sdelphij { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 438242845Sdelphij { "memory_throttle_count", KSTAT_DATA_UINT64 }, 439242845Sdelphij { "duplicate_buffers", KSTAT_DATA_UINT64 }, 440242845Sdelphij { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 441242845Sdelphij { "duplicate_reads", KSTAT_DATA_UINT64 } 442168404Spjd}; 443168404Spjd 444168404Spjd#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 445168404Spjd 446168404Spjd#define ARCSTAT_INCR(stat, val) \ 447251631Sdelphij atomic_add_64(&arc_stats.stat.value.ui64, (val)) 448168404Spjd 449206796Spjd#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 450168404Spjd#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 451168404Spjd 452168404Spjd#define ARCSTAT_MAX(stat, val) { \ 453168404Spjd uint64_t m; \ 454168404Spjd while ((val) > (m = arc_stats.stat.value.ui64) && \ 455168404Spjd (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 456168404Spjd continue; \ 457168404Spjd} 458168404Spjd 459168404Spjd#define ARCSTAT_MAXSTAT(stat) \ 460168404Spjd ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 461168404Spjd 462168404Spjd/* 463168404Spjd * We define a macro to allow ARC hits/misses to be easily broken down by 464168404Spjd * two separate conditions, giving a total of four different subtypes for 465168404Spjd * each of hits and misses (so eight statistics total). 466168404Spjd */ 467168404Spjd#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 468168404Spjd if (cond1) { \ 469168404Spjd if (cond2) { \ 470168404Spjd ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 471168404Spjd } else { \ 472168404Spjd ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 473168404Spjd } \ 474168404Spjd } else { \ 475168404Spjd if (cond2) { \ 476168404Spjd ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 477168404Spjd } else { \ 478168404Spjd ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 479168404Spjd } \ 480168404Spjd } 481168404Spjd 482168404Spjdkstat_t *arc_ksp; 483206796Spjdstatic arc_state_t *arc_anon; 484168404Spjdstatic arc_state_t *arc_mru; 485168404Spjdstatic arc_state_t *arc_mru_ghost; 486168404Spjdstatic arc_state_t *arc_mfu; 487168404Spjdstatic arc_state_t *arc_mfu_ghost; 488185029Spjdstatic arc_state_t *arc_l2c_only; 489168404Spjd 490168404Spjd/* 491168404Spjd * There are several ARC variables that are critical to export as kstats -- 492168404Spjd * but we don't want to have to grovel around in the kstat whenever we wish to 493168404Spjd * manipulate them. For these variables, we therefore define them to be in 494168404Spjd * terms of the statistic variable. This assures that we are not introducing 495168404Spjd * the possibility of inconsistency by having shadow copies of the variables, 496168404Spjd * while still allowing the code to be readable. 497168404Spjd */ 498168404Spjd#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 499168404Spjd#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 500168404Spjd#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 501168404Spjd#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 502168404Spjd#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 503168404Spjd 504251478Sdelphij#define L2ARC_IS_VALID_COMPRESS(_c_) \ 505251478Sdelphij ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 506251478Sdelphij 507168404Spjdstatic int arc_no_grow; /* Don't try to grow cache size */ 508168404Spjdstatic uint64_t arc_tempreserve; 509209962Smmstatic uint64_t arc_loaned_bytes; 510185029Spjdstatic uint64_t arc_meta_used; 511185029Spjdstatic uint64_t arc_meta_limit; 512185029Spjdstatic uint64_t arc_meta_max = 0; 513229663SpjdSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RD, &arc_meta_used, 0, 514229663Spjd "ARC metadata used"); 515229663SpjdSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RW, &arc_meta_limit, 0, 516229663Spjd "ARC metadata limit"); 517168404Spjd 518185029Spjdtypedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 519185029Spjd 520168404Spjdtypedef struct arc_callback arc_callback_t; 521168404Spjd 522168404Spjdstruct arc_callback { 523168404Spjd void *acb_private; 524168404Spjd arc_done_func_t *acb_done; 525168404Spjd arc_buf_t *acb_buf; 526168404Spjd zio_t *acb_zio_dummy; 527168404Spjd arc_callback_t *acb_next; 528168404Spjd}; 529168404Spjd 530168404Spjdtypedef struct arc_write_callback arc_write_callback_t; 531168404Spjd 532168404Spjdstruct arc_write_callback { 533168404Spjd void *awcb_private; 534168404Spjd arc_done_func_t *awcb_ready; 535258632Savg arc_done_func_t *awcb_physdone; 536168404Spjd arc_done_func_t *awcb_done; 537168404Spjd arc_buf_t *awcb_buf; 538168404Spjd}; 539168404Spjd 540168404Spjdstruct arc_buf_hdr { 541168404Spjd /* protected by hash lock */ 542168404Spjd dva_t b_dva; 543168404Spjd uint64_t b_birth; 544168404Spjd uint64_t b_cksum0; 545168404Spjd 546168404Spjd kmutex_t b_freeze_lock; 547168404Spjd zio_cksum_t *b_freeze_cksum; 548219089Spjd void *b_thawed; 549168404Spjd 550168404Spjd arc_buf_hdr_t *b_hash_next; 551168404Spjd arc_buf_t *b_buf; 552168404Spjd uint32_t b_flags; 553168404Spjd uint32_t b_datacnt; 554168404Spjd 555168404Spjd arc_callback_t *b_acb; 556168404Spjd kcondvar_t b_cv; 557168404Spjd 558168404Spjd /* immutable */ 559168404Spjd arc_buf_contents_t b_type; 560168404Spjd uint64_t b_size; 561209962Smm uint64_t b_spa; 562168404Spjd 563168404Spjd /* protected by arc state mutex */ 564168404Spjd arc_state_t *b_state; 565168404Spjd list_node_t b_arc_node; 566168404Spjd 567168404Spjd /* updated atomically */ 568168404Spjd clock_t b_arc_access; 569168404Spjd 570168404Spjd /* self protecting */ 571168404Spjd refcount_t b_refcnt; 572185029Spjd 573185029Spjd l2arc_buf_hdr_t *b_l2hdr; 574185029Spjd list_node_t b_l2node; 575168404Spjd}; 576168404Spjd 577168404Spjdstatic arc_buf_t *arc_eviction_list; 578168404Spjdstatic kmutex_t arc_eviction_mtx; 579168404Spjdstatic arc_buf_hdr_t arc_eviction_hdr; 580168404Spjdstatic void arc_get_data_buf(arc_buf_t *buf); 581168404Spjdstatic void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock); 582185029Spjdstatic int arc_evict_needed(arc_buf_contents_t type); 583209962Smmstatic void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes); 584240133Smm#ifdef illumos 585240133Smmstatic void arc_buf_watch(arc_buf_t *buf); 586240133Smm#endif /* illumos */ 587168404Spjd 588209962Smmstatic boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab); 589208373Smm 590168404Spjd#define GHOST_STATE(state) \ 591185029Spjd ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 592185029Spjd (state) == arc_l2c_only) 593168404Spjd 594168404Spjd/* 595168404Spjd * Private ARC flags. These flags are private ARC only flags that will show up 596168404Spjd * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can 597168404Spjd * be passed in as arc_flags in things like arc_read. However, these flags 598168404Spjd * should never be passed and should only be set by ARC code. When adding new 599168404Spjd * public flags, make sure not to smash the private ones. 600168404Spjd */ 601168404Spjd 602168404Spjd#define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */ 603168404Spjd#define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */ 604168404Spjd#define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */ 605168404Spjd#define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */ 606168404Spjd#define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */ 607168404Spjd#define ARC_INDIRECT (1 << 14) /* this is an indirect block */ 608185029Spjd#define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */ 609185029Spjd#define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */ 610185029Spjd#define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */ 611185029Spjd#define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */ 612168404Spjd 613168404Spjd#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE) 614168404Spjd#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS) 615168404Spjd#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR) 616208373Smm#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH) 617168404Spjd#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ) 618168404Spjd#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE) 619185029Spjd#define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS) 620185029Spjd#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE) 621185029Spjd#define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \ 622185029Spjd (hdr)->b_l2hdr != NULL) 623185029Spjd#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING) 624185029Spjd#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED) 625185029Spjd#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD) 626168404Spjd 627168404Spjd/* 628185029Spjd * Other sizes 629185029Spjd */ 630185029Spjd 631185029Spjd#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 632185029Spjd#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 633185029Spjd 634185029Spjd/* 635168404Spjd * Hash table routines 636168404Spjd */ 637168404Spjd 638205253Skmacy#define HT_LOCK_PAD CACHE_LINE_SIZE 639168404Spjd 640168404Spjdstruct ht_lock { 641168404Spjd kmutex_t ht_lock; 642168404Spjd#ifdef _KERNEL 643168404Spjd unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 644168404Spjd#endif 645168404Spjd}; 646168404Spjd 647168404Spjd#define BUF_LOCKS 256 648168404Spjdtypedef struct buf_hash_table { 649168404Spjd uint64_t ht_mask; 650168404Spjd arc_buf_hdr_t **ht_table; 651205264Skmacy struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 652168404Spjd} buf_hash_table_t; 653168404Spjd 654168404Spjdstatic buf_hash_table_t buf_hash_table; 655168404Spjd 656168404Spjd#define BUF_HASH_INDEX(spa, dva, birth) \ 657168404Spjd (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 658168404Spjd#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 659168404Spjd#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 660219089Spjd#define HDR_LOCK(hdr) \ 661219089Spjd (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 662168404Spjd 663168404Spjduint64_t zfs_crc64_table[256]; 664168404Spjd 665185029Spjd/* 666185029Spjd * Level 2 ARC 667185029Spjd */ 668185029Spjd 669208373Smm#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 670251478Sdelphij#define L2ARC_HEADROOM 2 /* num of writes */ 671251478Sdelphij/* 672251478Sdelphij * If we discover during ARC scan any buffers to be compressed, we boost 673251478Sdelphij * our headroom for the next scanning cycle by this percentage multiple. 674251478Sdelphij */ 675251478Sdelphij#define L2ARC_HEADROOM_BOOST 200 676208373Smm#define L2ARC_FEED_SECS 1 /* caching interval secs */ 677208373Smm#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 678185029Spjd 679185029Spjd#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 680185029Spjd#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 681185029Spjd 682251631Sdelphij/* L2ARC Performance Tunables */ 683185029Spjduint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 684185029Spjduint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 685185029Spjduint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 686251478Sdelphijuint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 687185029Spjduint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 688208373Smmuint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 689219089Spjdboolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 690208373Smmboolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 691208373Smmboolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 692185029Spjd 693217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 694205231Skmacy &l2arc_write_max, 0, "max write size"); 695217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 696205231Skmacy &l2arc_write_boost, 0, "extra write during warmup"); 697217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 698205231Skmacy &l2arc_headroom, 0, "number of dev writes"); 699217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 700205231Skmacy &l2arc_feed_secs, 0, "interval seconds"); 701217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 702208373Smm &l2arc_feed_min_ms, 0, "min interval milliseconds"); 703205231Skmacy 704205231SkmacySYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 705205231Skmacy &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 706208373SmmSYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 707208373Smm &l2arc_feed_again, 0, "turbo warmup"); 708208373SmmSYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 709208373Smm &l2arc_norw, 0, "no reads during writes"); 710205231Skmacy 711217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 712205231Skmacy &ARC_anon.arcs_size, 0, "size of anonymous state"); 713217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 714205231Skmacy &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 715217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 716205231Skmacy &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 717205231Skmacy 718217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 719205231Skmacy &ARC_mru.arcs_size, 0, "size of mru state"); 720217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 721205231Skmacy &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 722217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 723205231Skmacy &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 724205231Skmacy 725217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 726205231Skmacy &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 727217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 728205231Skmacy &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 729205231Skmacy "size of metadata in mru ghost state"); 730217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 731205231Skmacy &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 732205231Skmacy "size of data in mru ghost state"); 733205231Skmacy 734217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 735205231Skmacy &ARC_mfu.arcs_size, 0, "size of mfu state"); 736217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 737205231Skmacy &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 738217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 739205231Skmacy &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 740205231Skmacy 741217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 742205231Skmacy &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 743217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 744205231Skmacy &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 745205231Skmacy "size of metadata in mfu ghost state"); 746217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 747205231Skmacy &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 748205231Skmacy "size of data in mfu ghost state"); 749205231Skmacy 750217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 751205231Skmacy &ARC_l2c_only.arcs_size, 0, "size of mru state"); 752205231Skmacy 753185029Spjd/* 754185029Spjd * L2ARC Internals 755185029Spjd */ 756185029Spjdtypedef struct l2arc_dev { 757185029Spjd vdev_t *l2ad_vdev; /* vdev */ 758185029Spjd spa_t *l2ad_spa; /* spa */ 759185029Spjd uint64_t l2ad_hand; /* next write location */ 760185029Spjd uint64_t l2ad_start; /* first addr on device */ 761185029Spjd uint64_t l2ad_end; /* last addr on device */ 762185029Spjd uint64_t l2ad_evict; /* last addr eviction reached */ 763185029Spjd boolean_t l2ad_first; /* first sweep through */ 764208373Smm boolean_t l2ad_writing; /* currently writing */ 765185029Spjd list_t *l2ad_buflist; /* buffer list */ 766185029Spjd list_node_t l2ad_node; /* device list node */ 767185029Spjd} l2arc_dev_t; 768185029Spjd 769185029Spjdstatic list_t L2ARC_dev_list; /* device list */ 770185029Spjdstatic list_t *l2arc_dev_list; /* device list pointer */ 771185029Spjdstatic kmutex_t l2arc_dev_mtx; /* device list mutex */ 772185029Spjdstatic l2arc_dev_t *l2arc_dev_last; /* last device used */ 773185029Spjdstatic kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 774185029Spjdstatic list_t L2ARC_free_on_write; /* free after write buf list */ 775185029Spjdstatic list_t *l2arc_free_on_write; /* free after write list ptr */ 776185029Spjdstatic kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 777185029Spjdstatic uint64_t l2arc_ndev; /* number of devices */ 778185029Spjd 779185029Spjdtypedef struct l2arc_read_callback { 780251478Sdelphij arc_buf_t *l2rcb_buf; /* read buffer */ 781251478Sdelphij spa_t *l2rcb_spa; /* spa */ 782251478Sdelphij blkptr_t l2rcb_bp; /* original blkptr */ 783251478Sdelphij zbookmark_t l2rcb_zb; /* original bookmark */ 784251478Sdelphij int l2rcb_flags; /* original flags */ 785251478Sdelphij enum zio_compress l2rcb_compress; /* applied compress */ 786185029Spjd} l2arc_read_callback_t; 787185029Spjd 788185029Spjdtypedef struct l2arc_write_callback { 789185029Spjd l2arc_dev_t *l2wcb_dev; /* device info */ 790185029Spjd arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 791185029Spjd} l2arc_write_callback_t; 792185029Spjd 793185029Spjdstruct l2arc_buf_hdr { 794185029Spjd /* protected by arc_buf_hdr mutex */ 795251478Sdelphij l2arc_dev_t *b_dev; /* L2ARC device */ 796251478Sdelphij uint64_t b_daddr; /* disk address, offset byte */ 797251478Sdelphij /* compression applied to buffer data */ 798251478Sdelphij enum zio_compress b_compress; 799251478Sdelphij /* real alloc'd buffer size depending on b_compress applied */ 800251478Sdelphij int b_asize; 801251478Sdelphij /* temporary buffer holder for in-flight compressed data */ 802251478Sdelphij void *b_tmp_cdata; 803185029Spjd}; 804185029Spjd 805185029Spjdtypedef struct l2arc_data_free { 806185029Spjd /* protected by l2arc_free_on_write_mtx */ 807185029Spjd void *l2df_data; 808185029Spjd size_t l2df_size; 809185029Spjd void (*l2df_func)(void *, size_t); 810185029Spjd list_node_t l2df_list_node; 811185029Spjd} l2arc_data_free_t; 812185029Spjd 813185029Spjdstatic kmutex_t l2arc_feed_thr_lock; 814185029Spjdstatic kcondvar_t l2arc_feed_thr_cv; 815185029Spjdstatic uint8_t l2arc_thread_exit; 816185029Spjd 817185029Spjdstatic void l2arc_read_done(zio_t *zio); 818185029Spjdstatic void l2arc_hdr_stat_add(void); 819185029Spjdstatic void l2arc_hdr_stat_remove(void); 820185029Spjd 821251478Sdelphijstatic boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr); 822251478Sdelphijstatic void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, 823251478Sdelphij enum zio_compress c); 824251478Sdelphijstatic void l2arc_release_cdata_buf(arc_buf_hdr_t *ab); 825251478Sdelphij 826168404Spjdstatic uint64_t 827209962Smmbuf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 828168404Spjd{ 829168404Spjd uint8_t *vdva = (uint8_t *)dva; 830168404Spjd uint64_t crc = -1ULL; 831168404Spjd int i; 832168404Spjd 833168404Spjd ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 834168404Spjd 835168404Spjd for (i = 0; i < sizeof (dva_t); i++) 836168404Spjd crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 837168404Spjd 838209962Smm crc ^= (spa>>8) ^ birth; 839168404Spjd 840168404Spjd return (crc); 841168404Spjd} 842168404Spjd 843168404Spjd#define BUF_EMPTY(buf) \ 844168404Spjd ((buf)->b_dva.dva_word[0] == 0 && \ 845168404Spjd (buf)->b_dva.dva_word[1] == 0 && \ 846260150Sdelphij (buf)->b_cksum0 == 0) 847168404Spjd 848168404Spjd#define BUF_EQUAL(spa, dva, birth, buf) \ 849168404Spjd ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 850168404Spjd ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 851168404Spjd ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 852168404Spjd 853219089Spjdstatic void 854219089Spjdbuf_discard_identity(arc_buf_hdr_t *hdr) 855219089Spjd{ 856219089Spjd hdr->b_dva.dva_word[0] = 0; 857219089Spjd hdr->b_dva.dva_word[1] = 0; 858219089Spjd hdr->b_birth = 0; 859219089Spjd hdr->b_cksum0 = 0; 860219089Spjd} 861219089Spjd 862168404Spjdstatic arc_buf_hdr_t * 863268075Sdelphijbuf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 864168404Spjd{ 865268075Sdelphij const dva_t *dva = BP_IDENTITY(bp); 866268075Sdelphij uint64_t birth = BP_PHYSICAL_BIRTH(bp); 867168404Spjd uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 868168404Spjd kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 869168404Spjd arc_buf_hdr_t *buf; 870168404Spjd 871168404Spjd mutex_enter(hash_lock); 872168404Spjd for (buf = buf_hash_table.ht_table[idx]; buf != NULL; 873168404Spjd buf = buf->b_hash_next) { 874168404Spjd if (BUF_EQUAL(spa, dva, birth, buf)) { 875168404Spjd *lockp = hash_lock; 876168404Spjd return (buf); 877168404Spjd } 878168404Spjd } 879168404Spjd mutex_exit(hash_lock); 880168404Spjd *lockp = NULL; 881168404Spjd return (NULL); 882168404Spjd} 883168404Spjd 884168404Spjd/* 885168404Spjd * Insert an entry into the hash table. If there is already an element 886168404Spjd * equal to elem in the hash table, then the already existing element 887168404Spjd * will be returned and the new element will not be inserted. 888168404Spjd * Otherwise returns NULL. 889168404Spjd */ 890168404Spjdstatic arc_buf_hdr_t * 891168404Spjdbuf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp) 892168404Spjd{ 893168404Spjd uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 894168404Spjd kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 895168404Spjd arc_buf_hdr_t *fbuf; 896168404Spjd uint32_t i; 897168404Spjd 898268075Sdelphij ASSERT(!DVA_IS_EMPTY(&buf->b_dva)); 899268075Sdelphij ASSERT(buf->b_birth != 0); 900168404Spjd ASSERT(!HDR_IN_HASH_TABLE(buf)); 901168404Spjd *lockp = hash_lock; 902168404Spjd mutex_enter(hash_lock); 903168404Spjd for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL; 904168404Spjd fbuf = fbuf->b_hash_next, i++) { 905168404Spjd if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf)) 906168404Spjd return (fbuf); 907168404Spjd } 908168404Spjd 909168404Spjd buf->b_hash_next = buf_hash_table.ht_table[idx]; 910168404Spjd buf_hash_table.ht_table[idx] = buf; 911168404Spjd buf->b_flags |= ARC_IN_HASH_TABLE; 912168404Spjd 913168404Spjd /* collect some hash table performance data */ 914168404Spjd if (i > 0) { 915168404Spjd ARCSTAT_BUMP(arcstat_hash_collisions); 916168404Spjd if (i == 1) 917168404Spjd ARCSTAT_BUMP(arcstat_hash_chains); 918168404Spjd 919168404Spjd ARCSTAT_MAX(arcstat_hash_chain_max, i); 920168404Spjd } 921168404Spjd 922168404Spjd ARCSTAT_BUMP(arcstat_hash_elements); 923168404Spjd ARCSTAT_MAXSTAT(arcstat_hash_elements); 924168404Spjd 925168404Spjd return (NULL); 926168404Spjd} 927168404Spjd 928168404Spjdstatic void 929168404Spjdbuf_hash_remove(arc_buf_hdr_t *buf) 930168404Spjd{ 931168404Spjd arc_buf_hdr_t *fbuf, **bufp; 932168404Spjd uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 933168404Spjd 934168404Spjd ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 935168404Spjd ASSERT(HDR_IN_HASH_TABLE(buf)); 936168404Spjd 937168404Spjd bufp = &buf_hash_table.ht_table[idx]; 938168404Spjd while ((fbuf = *bufp) != buf) { 939168404Spjd ASSERT(fbuf != NULL); 940168404Spjd bufp = &fbuf->b_hash_next; 941168404Spjd } 942168404Spjd *bufp = buf->b_hash_next; 943168404Spjd buf->b_hash_next = NULL; 944168404Spjd buf->b_flags &= ~ARC_IN_HASH_TABLE; 945168404Spjd 946168404Spjd /* collect some hash table performance data */ 947168404Spjd ARCSTAT_BUMPDOWN(arcstat_hash_elements); 948168404Spjd 949168404Spjd if (buf_hash_table.ht_table[idx] && 950168404Spjd buf_hash_table.ht_table[idx]->b_hash_next == NULL) 951168404Spjd ARCSTAT_BUMPDOWN(arcstat_hash_chains); 952168404Spjd} 953168404Spjd 954168404Spjd/* 955168404Spjd * Global data structures and functions for the buf kmem cache. 956168404Spjd */ 957168404Spjdstatic kmem_cache_t *hdr_cache; 958168404Spjdstatic kmem_cache_t *buf_cache; 959168404Spjd 960168404Spjdstatic void 961168404Spjdbuf_fini(void) 962168404Spjd{ 963168404Spjd int i; 964168404Spjd 965168404Spjd kmem_free(buf_hash_table.ht_table, 966168404Spjd (buf_hash_table.ht_mask + 1) * sizeof (void *)); 967168404Spjd for (i = 0; i < BUF_LOCKS; i++) 968168404Spjd mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 969168404Spjd kmem_cache_destroy(hdr_cache); 970168404Spjd kmem_cache_destroy(buf_cache); 971168404Spjd} 972168404Spjd 973168404Spjd/* 974168404Spjd * Constructor callback - called when the cache is empty 975168404Spjd * and a new buf is requested. 976168404Spjd */ 977168404Spjd/* ARGSUSED */ 978168404Spjdstatic int 979168404Spjdhdr_cons(void *vbuf, void *unused, int kmflag) 980168404Spjd{ 981168404Spjd arc_buf_hdr_t *buf = vbuf; 982168404Spjd 983168404Spjd bzero(buf, sizeof (arc_buf_hdr_t)); 984168404Spjd refcount_create(&buf->b_refcnt); 985168404Spjd cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL); 986185029Spjd mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 987208373Smm arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 988185029Spjd 989168404Spjd return (0); 990168404Spjd} 991168404Spjd 992185029Spjd/* ARGSUSED */ 993185029Spjdstatic int 994185029Spjdbuf_cons(void *vbuf, void *unused, int kmflag) 995185029Spjd{ 996185029Spjd arc_buf_t *buf = vbuf; 997185029Spjd 998185029Spjd bzero(buf, sizeof (arc_buf_t)); 999219089Spjd mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1000208373Smm arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1001208373Smm 1002185029Spjd return (0); 1003185029Spjd} 1004185029Spjd 1005168404Spjd/* 1006168404Spjd * Destructor callback - called when a cached buf is 1007168404Spjd * no longer required. 1008168404Spjd */ 1009168404Spjd/* ARGSUSED */ 1010168404Spjdstatic void 1011168404Spjdhdr_dest(void *vbuf, void *unused) 1012168404Spjd{ 1013168404Spjd arc_buf_hdr_t *buf = vbuf; 1014168404Spjd 1015219089Spjd ASSERT(BUF_EMPTY(buf)); 1016168404Spjd refcount_destroy(&buf->b_refcnt); 1017168404Spjd cv_destroy(&buf->b_cv); 1018185029Spjd mutex_destroy(&buf->b_freeze_lock); 1019208373Smm arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1020168404Spjd} 1021168404Spjd 1022185029Spjd/* ARGSUSED */ 1023185029Spjdstatic void 1024185029Spjdbuf_dest(void *vbuf, void *unused) 1025185029Spjd{ 1026185029Spjd arc_buf_t *buf = vbuf; 1027185029Spjd 1028219089Spjd mutex_destroy(&buf->b_evict_lock); 1029208373Smm arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1030185029Spjd} 1031185029Spjd 1032168404Spjd/* 1033168404Spjd * Reclaim callback -- invoked when memory is low. 1034168404Spjd */ 1035168404Spjd/* ARGSUSED */ 1036168404Spjdstatic void 1037168404Spjdhdr_recl(void *unused) 1038168404Spjd{ 1039168404Spjd dprintf("hdr_recl called\n"); 1040168404Spjd /* 1041168404Spjd * umem calls the reclaim func when we destroy the buf cache, 1042168404Spjd * which is after we do arc_fini(). 1043168404Spjd */ 1044168404Spjd if (!arc_dead) 1045168404Spjd cv_signal(&arc_reclaim_thr_cv); 1046168404Spjd} 1047168404Spjd 1048168404Spjdstatic void 1049168404Spjdbuf_init(void) 1050168404Spjd{ 1051168404Spjd uint64_t *ct; 1052168404Spjd uint64_t hsize = 1ULL << 12; 1053168404Spjd int i, j; 1054168404Spjd 1055168404Spjd /* 1056168404Spjd * The hash table is big enough to fill all of physical memory 1057168404Spjd * with an average 64K block size. The table will take up 1058168404Spjd * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers). 1059168404Spjd */ 1060168696Spjd while (hsize * 65536 < (uint64_t)physmem * PAGESIZE) 1061168404Spjd hsize <<= 1; 1062168404Spjdretry: 1063168404Spjd buf_hash_table.ht_mask = hsize - 1; 1064168404Spjd buf_hash_table.ht_table = 1065168404Spjd kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1066168404Spjd if (buf_hash_table.ht_table == NULL) { 1067168404Spjd ASSERT(hsize > (1ULL << 8)); 1068168404Spjd hsize >>= 1; 1069168404Spjd goto retry; 1070168404Spjd } 1071168404Spjd 1072168404Spjd hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1073168404Spjd 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1074168404Spjd buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1075185029Spjd 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1076168404Spjd 1077168404Spjd for (i = 0; i < 256; i++) 1078168404Spjd for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1079168404Spjd *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1080168404Spjd 1081168404Spjd for (i = 0; i < BUF_LOCKS; i++) { 1082168404Spjd mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1083168404Spjd NULL, MUTEX_DEFAULT, NULL); 1084168404Spjd } 1085168404Spjd} 1086168404Spjd 1087168404Spjd#define ARC_MINTIME (hz>>4) /* 62 ms */ 1088168404Spjd 1089168404Spjdstatic void 1090168404Spjdarc_cksum_verify(arc_buf_t *buf) 1091168404Spjd{ 1092168404Spjd zio_cksum_t zc; 1093168404Spjd 1094168404Spjd if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1095168404Spjd return; 1096168404Spjd 1097168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1098168404Spjd if (buf->b_hdr->b_freeze_cksum == NULL || 1099168404Spjd (buf->b_hdr->b_flags & ARC_IO_ERROR)) { 1100168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1101168404Spjd return; 1102168404Spjd } 1103168404Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1104168404Spjd if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1105168404Spjd panic("buffer modified while frozen!"); 1106168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1107168404Spjd} 1108168404Spjd 1109185029Spjdstatic int 1110185029Spjdarc_cksum_equal(arc_buf_t *buf) 1111185029Spjd{ 1112185029Spjd zio_cksum_t zc; 1113185029Spjd int equal; 1114185029Spjd 1115185029Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1116185029Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1117185029Spjd equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1118185029Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1119185029Spjd 1120185029Spjd return (equal); 1121185029Spjd} 1122185029Spjd 1123168404Spjdstatic void 1124185029Spjdarc_cksum_compute(arc_buf_t *buf, boolean_t force) 1125168404Spjd{ 1126185029Spjd if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1127168404Spjd return; 1128168404Spjd 1129168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1130168404Spjd if (buf->b_hdr->b_freeze_cksum != NULL) { 1131168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1132168404Spjd return; 1133168404Spjd } 1134168404Spjd buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1135168404Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1136168404Spjd buf->b_hdr->b_freeze_cksum); 1137168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1138240133Smm#ifdef illumos 1139240133Smm arc_buf_watch(buf); 1140240133Smm#endif /* illumos */ 1141168404Spjd} 1142168404Spjd 1143240133Smm#ifdef illumos 1144240133Smm#ifndef _KERNEL 1145240133Smmtypedef struct procctl { 1146240133Smm long cmd; 1147240133Smm prwatch_t prwatch; 1148240133Smm} procctl_t; 1149240133Smm#endif 1150240133Smm 1151240133Smm/* ARGSUSED */ 1152240133Smmstatic void 1153240133Smmarc_buf_unwatch(arc_buf_t *buf) 1154240133Smm{ 1155240133Smm#ifndef _KERNEL 1156240133Smm if (arc_watch) { 1157240133Smm int result; 1158240133Smm procctl_t ctl; 1159240133Smm ctl.cmd = PCWATCH; 1160240133Smm ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1161240133Smm ctl.prwatch.pr_size = 0; 1162240133Smm ctl.prwatch.pr_wflags = 0; 1163240133Smm result = write(arc_procfd, &ctl, sizeof (ctl)); 1164240133Smm ASSERT3U(result, ==, sizeof (ctl)); 1165240133Smm } 1166240133Smm#endif 1167240133Smm} 1168240133Smm 1169240133Smm/* ARGSUSED */ 1170240133Smmstatic void 1171240133Smmarc_buf_watch(arc_buf_t *buf) 1172240133Smm{ 1173240133Smm#ifndef _KERNEL 1174240133Smm if (arc_watch) { 1175240133Smm int result; 1176240133Smm procctl_t ctl; 1177240133Smm ctl.cmd = PCWATCH; 1178240133Smm ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1179240133Smm ctl.prwatch.pr_size = buf->b_hdr->b_size; 1180240133Smm ctl.prwatch.pr_wflags = WA_WRITE; 1181240133Smm result = write(arc_procfd, &ctl, sizeof (ctl)); 1182240133Smm ASSERT3U(result, ==, sizeof (ctl)); 1183240133Smm } 1184240133Smm#endif 1185240133Smm} 1186240133Smm#endif /* illumos */ 1187240133Smm 1188168404Spjdvoid 1189168404Spjdarc_buf_thaw(arc_buf_t *buf) 1190168404Spjd{ 1191185029Spjd if (zfs_flags & ZFS_DEBUG_MODIFY) { 1192185029Spjd if (buf->b_hdr->b_state != arc_anon) 1193185029Spjd panic("modifying non-anon buffer!"); 1194185029Spjd if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS) 1195185029Spjd panic("modifying buffer while i/o in progress!"); 1196185029Spjd arc_cksum_verify(buf); 1197185029Spjd } 1198168404Spjd 1199168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1200168404Spjd if (buf->b_hdr->b_freeze_cksum != NULL) { 1201168404Spjd kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1202168404Spjd buf->b_hdr->b_freeze_cksum = NULL; 1203168404Spjd } 1204219089Spjd 1205219089Spjd if (zfs_flags & ZFS_DEBUG_MODIFY) { 1206219089Spjd if (buf->b_hdr->b_thawed) 1207219089Spjd kmem_free(buf->b_hdr->b_thawed, 1); 1208219089Spjd buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1209219089Spjd } 1210219089Spjd 1211168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1212240133Smm 1213240133Smm#ifdef illumos 1214240133Smm arc_buf_unwatch(buf); 1215240133Smm#endif /* illumos */ 1216168404Spjd} 1217168404Spjd 1218168404Spjdvoid 1219168404Spjdarc_buf_freeze(arc_buf_t *buf) 1220168404Spjd{ 1221219089Spjd kmutex_t *hash_lock; 1222219089Spjd 1223168404Spjd if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1224168404Spjd return; 1225168404Spjd 1226219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 1227219089Spjd mutex_enter(hash_lock); 1228219089Spjd 1229168404Spjd ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1230168404Spjd buf->b_hdr->b_state == arc_anon); 1231185029Spjd arc_cksum_compute(buf, B_FALSE); 1232219089Spjd mutex_exit(hash_lock); 1233240133Smm 1234168404Spjd} 1235168404Spjd 1236168404Spjdstatic void 1237205231Skmacyget_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock) 1238205231Skmacy{ 1239205231Skmacy uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth); 1240205231Skmacy 1241206796Spjd if (ab->b_type == ARC_BUFC_METADATA) 1242206796Spjd buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1243205231Skmacy else { 1244206796Spjd buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1245205231Skmacy buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1246205231Skmacy } 1247205231Skmacy 1248205231Skmacy *list = &state->arcs_lists[buf_hashid]; 1249205231Skmacy *lock = ARCS_LOCK(state, buf_hashid); 1250205231Skmacy} 1251205231Skmacy 1252205231Skmacy 1253205231Skmacystatic void 1254168404Spjdadd_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1255168404Spjd{ 1256168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 1257168404Spjd 1258168404Spjd if ((refcount_add(&ab->b_refcnt, tag) == 1) && 1259168404Spjd (ab->b_state != arc_anon)) { 1260206796Spjd uint64_t delta = ab->b_size * ab->b_datacnt; 1261206796Spjd uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type]; 1262205231Skmacy list_t *list; 1263205231Skmacy kmutex_t *lock; 1264168404Spjd 1265205231Skmacy get_buf_info(ab, ab->b_state, &list, &lock); 1266205231Skmacy ASSERT(!MUTEX_HELD(lock)); 1267205231Skmacy mutex_enter(lock); 1268168404Spjd ASSERT(list_link_active(&ab->b_arc_node)); 1269185029Spjd list_remove(list, ab); 1270168404Spjd if (GHOST_STATE(ab->b_state)) { 1271240415Smm ASSERT0(ab->b_datacnt); 1272168404Spjd ASSERT3P(ab->b_buf, ==, NULL); 1273168404Spjd delta = ab->b_size; 1274168404Spjd } 1275168404Spjd ASSERT(delta > 0); 1276185029Spjd ASSERT3U(*size, >=, delta); 1277185029Spjd atomic_add_64(size, -delta); 1278206794Spjd mutex_exit(lock); 1279185029Spjd /* remove the prefetch flag if we get a reference */ 1280168404Spjd if (ab->b_flags & ARC_PREFETCH) 1281168404Spjd ab->b_flags &= ~ARC_PREFETCH; 1282168404Spjd } 1283168404Spjd} 1284168404Spjd 1285168404Spjdstatic int 1286168404Spjdremove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1287168404Spjd{ 1288168404Spjd int cnt; 1289168404Spjd arc_state_t *state = ab->b_state; 1290168404Spjd 1291168404Spjd ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1292168404Spjd ASSERT(!GHOST_STATE(state)); 1293168404Spjd 1294168404Spjd if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) && 1295168404Spjd (state != arc_anon)) { 1296185029Spjd uint64_t *size = &state->arcs_lsize[ab->b_type]; 1297205231Skmacy list_t *list; 1298205231Skmacy kmutex_t *lock; 1299185029Spjd 1300205231Skmacy get_buf_info(ab, state, &list, &lock); 1301205231Skmacy ASSERT(!MUTEX_HELD(lock)); 1302205231Skmacy mutex_enter(lock); 1303168404Spjd ASSERT(!list_link_active(&ab->b_arc_node)); 1304205231Skmacy list_insert_head(list, ab); 1305168404Spjd ASSERT(ab->b_datacnt > 0); 1306185029Spjd atomic_add_64(size, ab->b_size * ab->b_datacnt); 1307206794Spjd mutex_exit(lock); 1308168404Spjd } 1309168404Spjd return (cnt); 1310168404Spjd} 1311168404Spjd 1312168404Spjd/* 1313168404Spjd * Move the supplied buffer to the indicated state. The mutex 1314168404Spjd * for the buffer must be held by the caller. 1315168404Spjd */ 1316168404Spjdstatic void 1317168404Spjdarc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock) 1318168404Spjd{ 1319168404Spjd arc_state_t *old_state = ab->b_state; 1320168404Spjd int64_t refcnt = refcount_count(&ab->b_refcnt); 1321168404Spjd uint64_t from_delta, to_delta; 1322205231Skmacy list_t *list; 1323205231Skmacy kmutex_t *lock; 1324168404Spjd 1325168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 1326258632Savg ASSERT3P(new_state, !=, old_state); 1327168404Spjd ASSERT(refcnt == 0 || ab->b_datacnt > 0); 1328168404Spjd ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state)); 1329219089Spjd ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon); 1330168404Spjd 1331168404Spjd from_delta = to_delta = ab->b_datacnt * ab->b_size; 1332168404Spjd 1333168404Spjd /* 1334168404Spjd * If this buffer is evictable, transfer it from the 1335168404Spjd * old state list to the new state list. 1336168404Spjd */ 1337168404Spjd if (refcnt == 0) { 1338168404Spjd if (old_state != arc_anon) { 1339205231Skmacy int use_mutex; 1340185029Spjd uint64_t *size = &old_state->arcs_lsize[ab->b_type]; 1341168404Spjd 1342205231Skmacy get_buf_info(ab, old_state, &list, &lock); 1343205231Skmacy use_mutex = !MUTEX_HELD(lock); 1344168404Spjd if (use_mutex) 1345205231Skmacy mutex_enter(lock); 1346168404Spjd 1347168404Spjd ASSERT(list_link_active(&ab->b_arc_node)); 1348205231Skmacy list_remove(list, ab); 1349168404Spjd 1350168404Spjd /* 1351168404Spjd * If prefetching out of the ghost cache, 1352219089Spjd * we will have a non-zero datacnt. 1353168404Spjd */ 1354168404Spjd if (GHOST_STATE(old_state) && ab->b_datacnt == 0) { 1355168404Spjd /* ghost elements have a ghost size */ 1356168404Spjd ASSERT(ab->b_buf == NULL); 1357168404Spjd from_delta = ab->b_size; 1358168404Spjd } 1359185029Spjd ASSERT3U(*size, >=, from_delta); 1360185029Spjd atomic_add_64(size, -from_delta); 1361168404Spjd 1362168404Spjd if (use_mutex) 1363205231Skmacy mutex_exit(lock); 1364168404Spjd } 1365168404Spjd if (new_state != arc_anon) { 1366206796Spjd int use_mutex; 1367185029Spjd uint64_t *size = &new_state->arcs_lsize[ab->b_type]; 1368168404Spjd 1369205231Skmacy get_buf_info(ab, new_state, &list, &lock); 1370205231Skmacy use_mutex = !MUTEX_HELD(lock); 1371168404Spjd if (use_mutex) 1372205231Skmacy mutex_enter(lock); 1373168404Spjd 1374205231Skmacy list_insert_head(list, ab); 1375168404Spjd 1376168404Spjd /* ghost elements have a ghost size */ 1377168404Spjd if (GHOST_STATE(new_state)) { 1378168404Spjd ASSERT(ab->b_datacnt == 0); 1379168404Spjd ASSERT(ab->b_buf == NULL); 1380168404Spjd to_delta = ab->b_size; 1381168404Spjd } 1382185029Spjd atomic_add_64(size, to_delta); 1383168404Spjd 1384168404Spjd if (use_mutex) 1385205231Skmacy mutex_exit(lock); 1386168404Spjd } 1387168404Spjd } 1388168404Spjd 1389168404Spjd ASSERT(!BUF_EMPTY(ab)); 1390219089Spjd if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab)) 1391168404Spjd buf_hash_remove(ab); 1392168404Spjd 1393168404Spjd /* adjust state sizes */ 1394168404Spjd if (to_delta) 1395168404Spjd atomic_add_64(&new_state->arcs_size, to_delta); 1396168404Spjd if (from_delta) { 1397168404Spjd ASSERT3U(old_state->arcs_size, >=, from_delta); 1398168404Spjd atomic_add_64(&old_state->arcs_size, -from_delta); 1399168404Spjd } 1400168404Spjd ab->b_state = new_state; 1401185029Spjd 1402185029Spjd /* adjust l2arc hdr stats */ 1403185029Spjd if (new_state == arc_l2c_only) 1404185029Spjd l2arc_hdr_stat_add(); 1405185029Spjd else if (old_state == arc_l2c_only) 1406185029Spjd l2arc_hdr_stat_remove(); 1407168404Spjd} 1408168404Spjd 1409185029Spjdvoid 1410208373Smmarc_space_consume(uint64_t space, arc_space_type_t type) 1411185029Spjd{ 1412208373Smm ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1413208373Smm 1414208373Smm switch (type) { 1415208373Smm case ARC_SPACE_DATA: 1416208373Smm ARCSTAT_INCR(arcstat_data_size, space); 1417208373Smm break; 1418208373Smm case ARC_SPACE_OTHER: 1419208373Smm ARCSTAT_INCR(arcstat_other_size, space); 1420208373Smm break; 1421208373Smm case ARC_SPACE_HDRS: 1422208373Smm ARCSTAT_INCR(arcstat_hdr_size, space); 1423208373Smm break; 1424208373Smm case ARC_SPACE_L2HDRS: 1425208373Smm ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1426208373Smm break; 1427208373Smm } 1428208373Smm 1429185029Spjd atomic_add_64(&arc_meta_used, space); 1430185029Spjd atomic_add_64(&arc_size, space); 1431185029Spjd} 1432185029Spjd 1433185029Spjdvoid 1434208373Smmarc_space_return(uint64_t space, arc_space_type_t type) 1435185029Spjd{ 1436208373Smm ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1437208373Smm 1438208373Smm switch (type) { 1439208373Smm case ARC_SPACE_DATA: 1440208373Smm ARCSTAT_INCR(arcstat_data_size, -space); 1441208373Smm break; 1442208373Smm case ARC_SPACE_OTHER: 1443208373Smm ARCSTAT_INCR(arcstat_other_size, -space); 1444208373Smm break; 1445208373Smm case ARC_SPACE_HDRS: 1446208373Smm ARCSTAT_INCR(arcstat_hdr_size, -space); 1447208373Smm break; 1448208373Smm case ARC_SPACE_L2HDRS: 1449208373Smm ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1450208373Smm break; 1451208373Smm } 1452208373Smm 1453185029Spjd ASSERT(arc_meta_used >= space); 1454185029Spjd if (arc_meta_max < arc_meta_used) 1455185029Spjd arc_meta_max = arc_meta_used; 1456185029Spjd atomic_add_64(&arc_meta_used, -space); 1457185029Spjd ASSERT(arc_size >= space); 1458185029Spjd atomic_add_64(&arc_size, -space); 1459185029Spjd} 1460185029Spjd 1461185029Spjdvoid * 1462185029Spjdarc_data_buf_alloc(uint64_t size) 1463185029Spjd{ 1464185029Spjd if (arc_evict_needed(ARC_BUFC_DATA)) 1465185029Spjd cv_signal(&arc_reclaim_thr_cv); 1466185029Spjd atomic_add_64(&arc_size, size); 1467185029Spjd return (zio_data_buf_alloc(size)); 1468185029Spjd} 1469185029Spjd 1470185029Spjdvoid 1471185029Spjdarc_data_buf_free(void *buf, uint64_t size) 1472185029Spjd{ 1473185029Spjd zio_data_buf_free(buf, size); 1474185029Spjd ASSERT(arc_size >= size); 1475185029Spjd atomic_add_64(&arc_size, -size); 1476185029Spjd} 1477185029Spjd 1478168404Spjdarc_buf_t * 1479168404Spjdarc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1480168404Spjd{ 1481168404Spjd arc_buf_hdr_t *hdr; 1482168404Spjd arc_buf_t *buf; 1483168404Spjd 1484168404Spjd ASSERT3U(size, >, 0); 1485185029Spjd hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1486168404Spjd ASSERT(BUF_EMPTY(hdr)); 1487168404Spjd hdr->b_size = size; 1488168404Spjd hdr->b_type = type; 1489228103Smm hdr->b_spa = spa_load_guid(spa); 1490168404Spjd hdr->b_state = arc_anon; 1491168404Spjd hdr->b_arc_access = 0; 1492185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1493168404Spjd buf->b_hdr = hdr; 1494168404Spjd buf->b_data = NULL; 1495168404Spjd buf->b_efunc = NULL; 1496168404Spjd buf->b_private = NULL; 1497168404Spjd buf->b_next = NULL; 1498168404Spjd hdr->b_buf = buf; 1499168404Spjd arc_get_data_buf(buf); 1500168404Spjd hdr->b_datacnt = 1; 1501168404Spjd hdr->b_flags = 0; 1502168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1503168404Spjd (void) refcount_add(&hdr->b_refcnt, tag); 1504168404Spjd 1505168404Spjd return (buf); 1506168404Spjd} 1507168404Spjd 1508209962Smmstatic char *arc_onloan_tag = "onloan"; 1509209962Smm 1510209962Smm/* 1511209962Smm * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1512209962Smm * flight data by arc_tempreserve_space() until they are "returned". Loaned 1513209962Smm * buffers must be returned to the arc before they can be used by the DMU or 1514209962Smm * freed. 1515209962Smm */ 1516209962Smmarc_buf_t * 1517209962Smmarc_loan_buf(spa_t *spa, int size) 1518209962Smm{ 1519209962Smm arc_buf_t *buf; 1520209962Smm 1521209962Smm buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1522209962Smm 1523209962Smm atomic_add_64(&arc_loaned_bytes, size); 1524209962Smm return (buf); 1525209962Smm} 1526209962Smm 1527209962Smm/* 1528209962Smm * Return a loaned arc buffer to the arc. 1529209962Smm */ 1530209962Smmvoid 1531209962Smmarc_return_buf(arc_buf_t *buf, void *tag) 1532209962Smm{ 1533209962Smm arc_buf_hdr_t *hdr = buf->b_hdr; 1534209962Smm 1535209962Smm ASSERT(buf->b_data != NULL); 1536219089Spjd (void) refcount_add(&hdr->b_refcnt, tag); 1537219089Spjd (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1538209962Smm 1539209962Smm atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1540209962Smm} 1541209962Smm 1542219089Spjd/* Detach an arc_buf from a dbuf (tag) */ 1543219089Spjdvoid 1544219089Spjdarc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1545219089Spjd{ 1546219089Spjd arc_buf_hdr_t *hdr; 1547219089Spjd 1548219089Spjd ASSERT(buf->b_data != NULL); 1549219089Spjd hdr = buf->b_hdr; 1550219089Spjd (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1551219089Spjd (void) refcount_remove(&hdr->b_refcnt, tag); 1552219089Spjd buf->b_efunc = NULL; 1553219089Spjd buf->b_private = NULL; 1554219089Spjd 1555219089Spjd atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1556219089Spjd} 1557219089Spjd 1558168404Spjdstatic arc_buf_t * 1559168404Spjdarc_buf_clone(arc_buf_t *from) 1560168404Spjd{ 1561168404Spjd arc_buf_t *buf; 1562168404Spjd arc_buf_hdr_t *hdr = from->b_hdr; 1563168404Spjd uint64_t size = hdr->b_size; 1564168404Spjd 1565219089Spjd ASSERT(hdr->b_state != arc_anon); 1566219089Spjd 1567185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1568168404Spjd buf->b_hdr = hdr; 1569168404Spjd buf->b_data = NULL; 1570168404Spjd buf->b_efunc = NULL; 1571168404Spjd buf->b_private = NULL; 1572168404Spjd buf->b_next = hdr->b_buf; 1573168404Spjd hdr->b_buf = buf; 1574168404Spjd arc_get_data_buf(buf); 1575168404Spjd bcopy(from->b_data, buf->b_data, size); 1576242845Sdelphij 1577242845Sdelphij /* 1578242845Sdelphij * This buffer already exists in the arc so create a duplicate 1579242845Sdelphij * copy for the caller. If the buffer is associated with user data 1580242845Sdelphij * then track the size and number of duplicates. These stats will be 1581242845Sdelphij * updated as duplicate buffers are created and destroyed. 1582242845Sdelphij */ 1583242845Sdelphij if (hdr->b_type == ARC_BUFC_DATA) { 1584242845Sdelphij ARCSTAT_BUMP(arcstat_duplicate_buffers); 1585242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1586242845Sdelphij } 1587168404Spjd hdr->b_datacnt += 1; 1588168404Spjd return (buf); 1589168404Spjd} 1590168404Spjd 1591168404Spjdvoid 1592168404Spjdarc_buf_add_ref(arc_buf_t *buf, void* tag) 1593168404Spjd{ 1594168404Spjd arc_buf_hdr_t *hdr; 1595168404Spjd kmutex_t *hash_lock; 1596168404Spjd 1597168404Spjd /* 1598185029Spjd * Check to see if this buffer is evicted. Callers 1599185029Spjd * must verify b_data != NULL to know if the add_ref 1600185029Spjd * was successful. 1601168404Spjd */ 1602219089Spjd mutex_enter(&buf->b_evict_lock); 1603185029Spjd if (buf->b_data == NULL) { 1604219089Spjd mutex_exit(&buf->b_evict_lock); 1605168404Spjd return; 1606168404Spjd } 1607219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 1608219089Spjd mutex_enter(hash_lock); 1609185029Spjd hdr = buf->b_hdr; 1610219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1611219089Spjd mutex_exit(&buf->b_evict_lock); 1612168404Spjd 1613168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1614168404Spjd add_reference(hdr, hash_lock, tag); 1615208373Smm DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1616168404Spjd arc_access(hdr, hash_lock); 1617168404Spjd mutex_exit(hash_lock); 1618168404Spjd ARCSTAT_BUMP(arcstat_hits); 1619168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 1620168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1621168404Spjd data, metadata, hits); 1622168404Spjd} 1623168404Spjd 1624185029Spjd/* 1625185029Spjd * Free the arc data buffer. If it is an l2arc write in progress, 1626185029Spjd * the buffer is placed on l2arc_free_on_write to be freed later. 1627185029Spjd */ 1628168404Spjdstatic void 1629240133Smmarc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1630185029Spjd{ 1631240133Smm arc_buf_hdr_t *hdr = buf->b_hdr; 1632240133Smm 1633185029Spjd if (HDR_L2_WRITING(hdr)) { 1634185029Spjd l2arc_data_free_t *df; 1635185029Spjd df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1636240133Smm df->l2df_data = buf->b_data; 1637240133Smm df->l2df_size = hdr->b_size; 1638185029Spjd df->l2df_func = free_func; 1639185029Spjd mutex_enter(&l2arc_free_on_write_mtx); 1640185029Spjd list_insert_head(l2arc_free_on_write, df); 1641185029Spjd mutex_exit(&l2arc_free_on_write_mtx); 1642185029Spjd ARCSTAT_BUMP(arcstat_l2_free_on_write); 1643185029Spjd } else { 1644240133Smm free_func(buf->b_data, hdr->b_size); 1645185029Spjd } 1646185029Spjd} 1647185029Spjd 1648185029Spjdstatic void 1649168404Spjdarc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all) 1650168404Spjd{ 1651168404Spjd arc_buf_t **bufp; 1652168404Spjd 1653168404Spjd /* free up data associated with the buf */ 1654168404Spjd if (buf->b_data) { 1655168404Spjd arc_state_t *state = buf->b_hdr->b_state; 1656168404Spjd uint64_t size = buf->b_hdr->b_size; 1657168404Spjd arc_buf_contents_t type = buf->b_hdr->b_type; 1658168404Spjd 1659168404Spjd arc_cksum_verify(buf); 1660240133Smm#ifdef illumos 1661240133Smm arc_buf_unwatch(buf); 1662240133Smm#endif /* illumos */ 1663219089Spjd 1664168404Spjd if (!recycle) { 1665168404Spjd if (type == ARC_BUFC_METADATA) { 1666240133Smm arc_buf_data_free(buf, zio_buf_free); 1667208373Smm arc_space_return(size, ARC_SPACE_DATA); 1668168404Spjd } else { 1669168404Spjd ASSERT(type == ARC_BUFC_DATA); 1670240133Smm arc_buf_data_free(buf, zio_data_buf_free); 1671208373Smm ARCSTAT_INCR(arcstat_data_size, -size); 1672185029Spjd atomic_add_64(&arc_size, -size); 1673168404Spjd } 1674168404Spjd } 1675168404Spjd if (list_link_active(&buf->b_hdr->b_arc_node)) { 1676185029Spjd uint64_t *cnt = &state->arcs_lsize[type]; 1677185029Spjd 1678168404Spjd ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1679168404Spjd ASSERT(state != arc_anon); 1680185029Spjd 1681185029Spjd ASSERT3U(*cnt, >=, size); 1682185029Spjd atomic_add_64(cnt, -size); 1683168404Spjd } 1684168404Spjd ASSERT3U(state->arcs_size, >=, size); 1685168404Spjd atomic_add_64(&state->arcs_size, -size); 1686168404Spjd buf->b_data = NULL; 1687242845Sdelphij 1688242845Sdelphij /* 1689242845Sdelphij * If we're destroying a duplicate buffer make sure 1690242845Sdelphij * that the appropriate statistics are updated. 1691242845Sdelphij */ 1692242845Sdelphij if (buf->b_hdr->b_datacnt > 1 && 1693242845Sdelphij buf->b_hdr->b_type == ARC_BUFC_DATA) { 1694242845Sdelphij ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1695242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1696242845Sdelphij } 1697168404Spjd ASSERT(buf->b_hdr->b_datacnt > 0); 1698168404Spjd buf->b_hdr->b_datacnt -= 1; 1699168404Spjd } 1700168404Spjd 1701168404Spjd /* only remove the buf if requested */ 1702168404Spjd if (!all) 1703168404Spjd return; 1704168404Spjd 1705168404Spjd /* remove the buf from the hdr list */ 1706168404Spjd for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1707168404Spjd continue; 1708168404Spjd *bufp = buf->b_next; 1709219089Spjd buf->b_next = NULL; 1710168404Spjd 1711168404Spjd ASSERT(buf->b_efunc == NULL); 1712168404Spjd 1713168404Spjd /* clean up the buf */ 1714168404Spjd buf->b_hdr = NULL; 1715168404Spjd kmem_cache_free(buf_cache, buf); 1716168404Spjd} 1717168404Spjd 1718168404Spjdstatic void 1719168404Spjdarc_hdr_destroy(arc_buf_hdr_t *hdr) 1720168404Spjd{ 1721168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1722168404Spjd ASSERT3P(hdr->b_state, ==, arc_anon); 1723168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1724219089Spjd l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1725168404Spjd 1726219089Spjd if (l2hdr != NULL) { 1727219089Spjd boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1728219089Spjd /* 1729219089Spjd * To prevent arc_free() and l2arc_evict() from 1730219089Spjd * attempting to free the same buffer at the same time, 1731219089Spjd * a FREE_IN_PROGRESS flag is given to arc_free() to 1732219089Spjd * give it priority. l2arc_evict() can't destroy this 1733219089Spjd * header while we are waiting on l2arc_buflist_mtx. 1734219089Spjd * 1735219089Spjd * The hdr may be removed from l2ad_buflist before we 1736219089Spjd * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1737219089Spjd */ 1738219089Spjd if (!buflist_held) { 1739185029Spjd mutex_enter(&l2arc_buflist_mtx); 1740219089Spjd l2hdr = hdr->b_l2hdr; 1741219089Spjd } 1742219089Spjd 1743219089Spjd if (l2hdr != NULL) { 1744248572Ssmh trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1745248574Ssmh hdr->b_size, 0); 1746219089Spjd list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1747219089Spjd ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1748251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1749219089Spjd kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1750219089Spjd if (hdr->b_state == arc_l2c_only) 1751219089Spjd l2arc_hdr_stat_remove(); 1752219089Spjd hdr->b_l2hdr = NULL; 1753219089Spjd } 1754219089Spjd 1755219089Spjd if (!buflist_held) 1756185029Spjd mutex_exit(&l2arc_buflist_mtx); 1757185029Spjd } 1758185029Spjd 1759168404Spjd if (!BUF_EMPTY(hdr)) { 1760168404Spjd ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1761219089Spjd buf_discard_identity(hdr); 1762168404Spjd } 1763168404Spjd while (hdr->b_buf) { 1764168404Spjd arc_buf_t *buf = hdr->b_buf; 1765168404Spjd 1766168404Spjd if (buf->b_efunc) { 1767168404Spjd mutex_enter(&arc_eviction_mtx); 1768219089Spjd mutex_enter(&buf->b_evict_lock); 1769168404Spjd ASSERT(buf->b_hdr != NULL); 1770168404Spjd arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1771168404Spjd hdr->b_buf = buf->b_next; 1772168404Spjd buf->b_hdr = &arc_eviction_hdr; 1773168404Spjd buf->b_next = arc_eviction_list; 1774168404Spjd arc_eviction_list = buf; 1775219089Spjd mutex_exit(&buf->b_evict_lock); 1776168404Spjd mutex_exit(&arc_eviction_mtx); 1777168404Spjd } else { 1778168404Spjd arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1779168404Spjd } 1780168404Spjd } 1781168404Spjd if (hdr->b_freeze_cksum != NULL) { 1782168404Spjd kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1783168404Spjd hdr->b_freeze_cksum = NULL; 1784168404Spjd } 1785219089Spjd if (hdr->b_thawed) { 1786219089Spjd kmem_free(hdr->b_thawed, 1); 1787219089Spjd hdr->b_thawed = NULL; 1788219089Spjd } 1789168404Spjd 1790168404Spjd ASSERT(!list_link_active(&hdr->b_arc_node)); 1791168404Spjd ASSERT3P(hdr->b_hash_next, ==, NULL); 1792168404Spjd ASSERT3P(hdr->b_acb, ==, NULL); 1793168404Spjd kmem_cache_free(hdr_cache, hdr); 1794168404Spjd} 1795168404Spjd 1796168404Spjdvoid 1797168404Spjdarc_buf_free(arc_buf_t *buf, void *tag) 1798168404Spjd{ 1799168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 1800168404Spjd int hashed = hdr->b_state != arc_anon; 1801168404Spjd 1802168404Spjd ASSERT(buf->b_efunc == NULL); 1803168404Spjd ASSERT(buf->b_data != NULL); 1804168404Spjd 1805168404Spjd if (hashed) { 1806168404Spjd kmutex_t *hash_lock = HDR_LOCK(hdr); 1807168404Spjd 1808168404Spjd mutex_enter(hash_lock); 1809219089Spjd hdr = buf->b_hdr; 1810219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1811219089Spjd 1812168404Spjd (void) remove_reference(hdr, hash_lock, tag); 1813219089Spjd if (hdr->b_datacnt > 1) { 1814168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1815219089Spjd } else { 1816219089Spjd ASSERT(buf == hdr->b_buf); 1817219089Spjd ASSERT(buf->b_efunc == NULL); 1818168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 1819219089Spjd } 1820168404Spjd mutex_exit(hash_lock); 1821168404Spjd } else if (HDR_IO_IN_PROGRESS(hdr)) { 1822168404Spjd int destroy_hdr; 1823168404Spjd /* 1824168404Spjd * We are in the middle of an async write. Don't destroy 1825168404Spjd * this buffer unless the write completes before we finish 1826168404Spjd * decrementing the reference count. 1827168404Spjd */ 1828168404Spjd mutex_enter(&arc_eviction_mtx); 1829168404Spjd (void) remove_reference(hdr, NULL, tag); 1830168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1831168404Spjd destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1832168404Spjd mutex_exit(&arc_eviction_mtx); 1833168404Spjd if (destroy_hdr) 1834168404Spjd arc_hdr_destroy(hdr); 1835168404Spjd } else { 1836219089Spjd if (remove_reference(hdr, NULL, tag) > 0) 1837168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1838219089Spjd else 1839168404Spjd arc_hdr_destroy(hdr); 1840168404Spjd } 1841168404Spjd} 1842168404Spjd 1843248571Smmboolean_t 1844168404Spjdarc_buf_remove_ref(arc_buf_t *buf, void* tag) 1845168404Spjd{ 1846168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 1847168404Spjd kmutex_t *hash_lock = HDR_LOCK(hdr); 1848248571Smm boolean_t no_callback = (buf->b_efunc == NULL); 1849168404Spjd 1850168404Spjd if (hdr->b_state == arc_anon) { 1851219089Spjd ASSERT(hdr->b_datacnt == 1); 1852168404Spjd arc_buf_free(buf, tag); 1853168404Spjd return (no_callback); 1854168404Spjd } 1855168404Spjd 1856168404Spjd mutex_enter(hash_lock); 1857219089Spjd hdr = buf->b_hdr; 1858219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1859168404Spjd ASSERT(hdr->b_state != arc_anon); 1860168404Spjd ASSERT(buf->b_data != NULL); 1861168404Spjd 1862168404Spjd (void) remove_reference(hdr, hash_lock, tag); 1863168404Spjd if (hdr->b_datacnt > 1) { 1864168404Spjd if (no_callback) 1865168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1866168404Spjd } else if (no_callback) { 1867168404Spjd ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1868219089Spjd ASSERT(buf->b_efunc == NULL); 1869168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 1870168404Spjd } 1871168404Spjd ASSERT(no_callback || hdr->b_datacnt > 1 || 1872168404Spjd refcount_is_zero(&hdr->b_refcnt)); 1873168404Spjd mutex_exit(hash_lock); 1874168404Spjd return (no_callback); 1875168404Spjd} 1876168404Spjd 1877168404Spjdint 1878168404Spjdarc_buf_size(arc_buf_t *buf) 1879168404Spjd{ 1880168404Spjd return (buf->b_hdr->b_size); 1881168404Spjd} 1882168404Spjd 1883168404Spjd/* 1884242845Sdelphij * Called from the DMU to determine if the current buffer should be 1885242845Sdelphij * evicted. In order to ensure proper locking, the eviction must be initiated 1886242845Sdelphij * from the DMU. Return true if the buffer is associated with user data and 1887242845Sdelphij * duplicate buffers still exist. 1888242845Sdelphij */ 1889242845Sdelphijboolean_t 1890242845Sdelphijarc_buf_eviction_needed(arc_buf_t *buf) 1891242845Sdelphij{ 1892242845Sdelphij arc_buf_hdr_t *hdr; 1893242845Sdelphij boolean_t evict_needed = B_FALSE; 1894242845Sdelphij 1895242845Sdelphij if (zfs_disable_dup_eviction) 1896242845Sdelphij return (B_FALSE); 1897242845Sdelphij 1898242845Sdelphij mutex_enter(&buf->b_evict_lock); 1899242845Sdelphij hdr = buf->b_hdr; 1900242845Sdelphij if (hdr == NULL) { 1901242845Sdelphij /* 1902242845Sdelphij * We are in arc_do_user_evicts(); let that function 1903242845Sdelphij * perform the eviction. 1904242845Sdelphij */ 1905242845Sdelphij ASSERT(buf->b_data == NULL); 1906242845Sdelphij mutex_exit(&buf->b_evict_lock); 1907242845Sdelphij return (B_FALSE); 1908242845Sdelphij } else if (buf->b_data == NULL) { 1909242845Sdelphij /* 1910242845Sdelphij * We have already been added to the arc eviction list; 1911242845Sdelphij * recommend eviction. 1912242845Sdelphij */ 1913242845Sdelphij ASSERT3P(hdr, ==, &arc_eviction_hdr); 1914242845Sdelphij mutex_exit(&buf->b_evict_lock); 1915242845Sdelphij return (B_TRUE); 1916242845Sdelphij } 1917242845Sdelphij 1918242845Sdelphij if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 1919242845Sdelphij evict_needed = B_TRUE; 1920242845Sdelphij 1921242845Sdelphij mutex_exit(&buf->b_evict_lock); 1922242845Sdelphij return (evict_needed); 1923242845Sdelphij} 1924242845Sdelphij 1925242845Sdelphij/* 1926168404Spjd * Evict buffers from list until we've removed the specified number of 1927168404Spjd * bytes. Move the removed buffers to the appropriate evict state. 1928168404Spjd * If the recycle flag is set, then attempt to "recycle" a buffer: 1929168404Spjd * - look for a buffer to evict that is `bytes' long. 1930168404Spjd * - return the data block from this buffer rather than freeing it. 1931168404Spjd * This flag is used by callers that are trying to make space for a 1932168404Spjd * new buffer in a full arc cache. 1933185029Spjd * 1934185029Spjd * This function makes a "best effort". It skips over any buffers 1935185029Spjd * it can't get a hash_lock on, and so may not catch all candidates. 1936185029Spjd * It may also return without evicting as much space as requested. 1937168404Spjd */ 1938168404Spjdstatic void * 1939209962Smmarc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 1940168404Spjd arc_buf_contents_t type) 1941168404Spjd{ 1942168404Spjd arc_state_t *evicted_state; 1943168404Spjd uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 1944205231Skmacy int64_t bytes_remaining; 1945168404Spjd arc_buf_hdr_t *ab, *ab_prev = NULL; 1946205231Skmacy list_t *evicted_list, *list, *evicted_list_start, *list_start; 1947205231Skmacy kmutex_t *lock, *evicted_lock; 1948168404Spjd kmutex_t *hash_lock; 1949168404Spjd boolean_t have_lock; 1950168404Spjd void *stolen = NULL; 1951258632Savg arc_buf_hdr_t marker = { 0 }; 1952258632Savg int count = 0; 1953205231Skmacy static int evict_metadata_offset, evict_data_offset; 1954258632Savg int i, idx, offset, list_count, lists; 1955168404Spjd 1956168404Spjd ASSERT(state == arc_mru || state == arc_mfu); 1957168404Spjd 1958168404Spjd evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 1959206796Spjd 1960205231Skmacy if (type == ARC_BUFC_METADATA) { 1961205231Skmacy offset = 0; 1962205231Skmacy list_count = ARC_BUFC_NUMMETADATALISTS; 1963205231Skmacy list_start = &state->arcs_lists[0]; 1964205231Skmacy evicted_list_start = &evicted_state->arcs_lists[0]; 1965205231Skmacy idx = evict_metadata_offset; 1966205231Skmacy } else { 1967205231Skmacy offset = ARC_BUFC_NUMMETADATALISTS; 1968205231Skmacy list_start = &state->arcs_lists[offset]; 1969205231Skmacy evicted_list_start = &evicted_state->arcs_lists[offset]; 1970205231Skmacy list_count = ARC_BUFC_NUMDATALISTS; 1971205231Skmacy idx = evict_data_offset; 1972205231Skmacy } 1973205231Skmacy bytes_remaining = evicted_state->arcs_lsize[type]; 1974258632Savg lists = 0; 1975206796Spjd 1976205231Skmacyevict_start: 1977205231Skmacy list = &list_start[idx]; 1978205231Skmacy evicted_list = &evicted_list_start[idx]; 1979205231Skmacy lock = ARCS_LOCK(state, (offset + idx)); 1980206796Spjd evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 1981168404Spjd 1982205231Skmacy mutex_enter(lock); 1983205231Skmacy mutex_enter(evicted_lock); 1984205231Skmacy 1985185029Spjd for (ab = list_tail(list); ab; ab = ab_prev) { 1986185029Spjd ab_prev = list_prev(list, ab); 1987205231Skmacy bytes_remaining -= (ab->b_size * ab->b_datacnt); 1988168404Spjd /* prefetch buffers have a minimum lifespan */ 1989168404Spjd if (HDR_IO_IN_PROGRESS(ab) || 1990185029Spjd (spa && ab->b_spa != spa) || 1991168404Spjd (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 1992219089Spjd ddi_get_lbolt() - ab->b_arc_access < 1993219089Spjd arc_min_prefetch_lifespan)) { 1994168404Spjd skipped++; 1995168404Spjd continue; 1996168404Spjd } 1997168404Spjd /* "lookahead" for better eviction candidate */ 1998168404Spjd if (recycle && ab->b_size != bytes && 1999168404Spjd ab_prev && ab_prev->b_size == bytes) 2000168404Spjd continue; 2001258632Savg 2002258632Savg /* ignore markers */ 2003258632Savg if (ab->b_spa == 0) 2004258632Savg continue; 2005258632Savg 2006258632Savg /* 2007258632Savg * It may take a long time to evict all the bufs requested. 2008258632Savg * To avoid blocking all arc activity, periodically drop 2009258632Savg * the arcs_mtx and give other threads a chance to run 2010258632Savg * before reacquiring the lock. 2011258632Savg * 2012258632Savg * If we are looking for a buffer to recycle, we are in 2013258632Savg * the hot code path, so don't sleep. 2014258632Savg */ 2015258632Savg if (!recycle && count++ > arc_evict_iterations) { 2016258632Savg list_insert_after(list, ab, &marker); 2017258632Savg mutex_exit(evicted_lock); 2018258632Savg mutex_exit(lock); 2019258632Savg kpreempt(KPREEMPT_SYNC); 2020258632Savg mutex_enter(lock); 2021258632Savg mutex_enter(evicted_lock); 2022258632Savg ab_prev = list_prev(list, &marker); 2023258632Savg list_remove(list, &marker); 2024258632Savg count = 0; 2025258632Savg continue; 2026258632Savg } 2027258632Savg 2028168404Spjd hash_lock = HDR_LOCK(ab); 2029168404Spjd have_lock = MUTEX_HELD(hash_lock); 2030168404Spjd if (have_lock || mutex_tryenter(hash_lock)) { 2031240415Smm ASSERT0(refcount_count(&ab->b_refcnt)); 2032168404Spjd ASSERT(ab->b_datacnt > 0); 2033168404Spjd while (ab->b_buf) { 2034168404Spjd arc_buf_t *buf = ab->b_buf; 2035219089Spjd if (!mutex_tryenter(&buf->b_evict_lock)) { 2036185029Spjd missed += 1; 2037185029Spjd break; 2038185029Spjd } 2039168404Spjd if (buf->b_data) { 2040168404Spjd bytes_evicted += ab->b_size; 2041168404Spjd if (recycle && ab->b_type == type && 2042185029Spjd ab->b_size == bytes && 2043185029Spjd !HDR_L2_WRITING(ab)) { 2044168404Spjd stolen = buf->b_data; 2045168404Spjd recycle = FALSE; 2046168404Spjd } 2047168404Spjd } 2048168404Spjd if (buf->b_efunc) { 2049168404Spjd mutex_enter(&arc_eviction_mtx); 2050168404Spjd arc_buf_destroy(buf, 2051168404Spjd buf->b_data == stolen, FALSE); 2052168404Spjd ab->b_buf = buf->b_next; 2053168404Spjd buf->b_hdr = &arc_eviction_hdr; 2054168404Spjd buf->b_next = arc_eviction_list; 2055168404Spjd arc_eviction_list = buf; 2056168404Spjd mutex_exit(&arc_eviction_mtx); 2057219089Spjd mutex_exit(&buf->b_evict_lock); 2058168404Spjd } else { 2059219089Spjd mutex_exit(&buf->b_evict_lock); 2060168404Spjd arc_buf_destroy(buf, 2061168404Spjd buf->b_data == stolen, TRUE); 2062168404Spjd } 2063168404Spjd } 2064208373Smm 2065208373Smm if (ab->b_l2hdr) { 2066208373Smm ARCSTAT_INCR(arcstat_evict_l2_cached, 2067208373Smm ab->b_size); 2068208373Smm } else { 2069208373Smm if (l2arc_write_eligible(ab->b_spa, ab)) { 2070208373Smm ARCSTAT_INCR(arcstat_evict_l2_eligible, 2071208373Smm ab->b_size); 2072208373Smm } else { 2073208373Smm ARCSTAT_INCR( 2074208373Smm arcstat_evict_l2_ineligible, 2075208373Smm ab->b_size); 2076208373Smm } 2077208373Smm } 2078208373Smm 2079185029Spjd if (ab->b_datacnt == 0) { 2080185029Spjd arc_change_state(evicted_state, ab, hash_lock); 2081185029Spjd ASSERT(HDR_IN_HASH_TABLE(ab)); 2082185029Spjd ab->b_flags |= ARC_IN_HASH_TABLE; 2083185029Spjd ab->b_flags &= ~ARC_BUF_AVAILABLE; 2084185029Spjd DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 2085185029Spjd } 2086168404Spjd if (!have_lock) 2087168404Spjd mutex_exit(hash_lock); 2088168404Spjd if (bytes >= 0 && bytes_evicted >= bytes) 2089168404Spjd break; 2090205231Skmacy if (bytes_remaining > 0) { 2091205231Skmacy mutex_exit(evicted_lock); 2092205231Skmacy mutex_exit(lock); 2093206796Spjd idx = ((idx + 1) & (list_count - 1)); 2094258632Savg lists++; 2095205231Skmacy goto evict_start; 2096205231Skmacy } 2097168404Spjd } else { 2098168404Spjd missed += 1; 2099168404Spjd } 2100168404Spjd } 2101168404Spjd 2102205231Skmacy mutex_exit(evicted_lock); 2103205231Skmacy mutex_exit(lock); 2104206796Spjd 2105206796Spjd idx = ((idx + 1) & (list_count - 1)); 2106258632Savg lists++; 2107168404Spjd 2108205231Skmacy if (bytes_evicted < bytes) { 2109258632Savg if (lists < list_count) 2110205231Skmacy goto evict_start; 2111205231Skmacy else 2112205231Skmacy dprintf("only evicted %lld bytes from %x", 2113205231Skmacy (longlong_t)bytes_evicted, state); 2114205231Skmacy } 2115206796Spjd if (type == ARC_BUFC_METADATA) 2116205231Skmacy evict_metadata_offset = idx; 2117205231Skmacy else 2118205231Skmacy evict_data_offset = idx; 2119206796Spjd 2120168404Spjd if (skipped) 2121168404Spjd ARCSTAT_INCR(arcstat_evict_skip, skipped); 2122168404Spjd 2123168404Spjd if (missed) 2124168404Spjd ARCSTAT_INCR(arcstat_mutex_miss, missed); 2125168404Spjd 2126185029Spjd /* 2127258632Savg * Note: we have just evicted some data into the ghost state, 2128258632Savg * potentially putting the ghost size over the desired size. Rather 2129258632Savg * that evicting from the ghost list in this hot code path, leave 2130258632Savg * this chore to the arc_reclaim_thread(). 2131185029Spjd */ 2132185029Spjd 2133205231Skmacy if (stolen) 2134205231Skmacy ARCSTAT_BUMP(arcstat_stolen); 2135168404Spjd return (stolen); 2136168404Spjd} 2137168404Spjd 2138168404Spjd/* 2139168404Spjd * Remove buffers from list until we've removed the specified number of 2140168404Spjd * bytes. Destroy the buffers that are removed. 2141168404Spjd */ 2142168404Spjdstatic void 2143209962Smmarc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2144168404Spjd{ 2145168404Spjd arc_buf_hdr_t *ab, *ab_prev; 2146219089Spjd arc_buf_hdr_t marker = { 0 }; 2147205231Skmacy list_t *list, *list_start; 2148205231Skmacy kmutex_t *hash_lock, *lock; 2149168404Spjd uint64_t bytes_deleted = 0; 2150168404Spjd uint64_t bufs_skipped = 0; 2151258632Savg int count = 0; 2152205231Skmacy static int evict_offset; 2153205231Skmacy int list_count, idx = evict_offset; 2154258632Savg int offset, lists = 0; 2155168404Spjd 2156168404Spjd ASSERT(GHOST_STATE(state)); 2157205231Skmacy 2158205231Skmacy /* 2159205231Skmacy * data lists come after metadata lists 2160205231Skmacy */ 2161205231Skmacy list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2162205231Skmacy list_count = ARC_BUFC_NUMDATALISTS; 2163205231Skmacy offset = ARC_BUFC_NUMMETADATALISTS; 2164206796Spjd 2165205231Skmacyevict_start: 2166205231Skmacy list = &list_start[idx]; 2167205231Skmacy lock = ARCS_LOCK(state, idx + offset); 2168205231Skmacy 2169205231Skmacy mutex_enter(lock); 2170185029Spjd for (ab = list_tail(list); ab; ab = ab_prev) { 2171185029Spjd ab_prev = list_prev(list, ab); 2172258632Savg if (ab->b_type > ARC_BUFC_NUMTYPES) 2173258632Savg panic("invalid ab=%p", (void *)ab); 2174185029Spjd if (spa && ab->b_spa != spa) 2175185029Spjd continue; 2176219089Spjd 2177219089Spjd /* ignore markers */ 2178219089Spjd if (ab->b_spa == 0) 2179219089Spjd continue; 2180219089Spjd 2181168404Spjd hash_lock = HDR_LOCK(ab); 2182219089Spjd /* caller may be trying to modify this buffer, skip it */ 2183219089Spjd if (MUTEX_HELD(hash_lock)) 2184219089Spjd continue; 2185258632Savg 2186258632Savg /* 2187258632Savg * It may take a long time to evict all the bufs requested. 2188258632Savg * To avoid blocking all arc activity, periodically drop 2189258632Savg * the arcs_mtx and give other threads a chance to run 2190258632Savg * before reacquiring the lock. 2191258632Savg */ 2192258632Savg if (count++ > arc_evict_iterations) { 2193258632Savg list_insert_after(list, ab, &marker); 2194258632Savg mutex_exit(lock); 2195258632Savg kpreempt(KPREEMPT_SYNC); 2196258632Savg mutex_enter(lock); 2197258632Savg ab_prev = list_prev(list, &marker); 2198258632Savg list_remove(list, &marker); 2199258632Savg count = 0; 2200258632Savg continue; 2201258632Savg } 2202168404Spjd if (mutex_tryenter(hash_lock)) { 2203168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(ab)); 2204168404Spjd ASSERT(ab->b_buf == NULL); 2205168404Spjd ARCSTAT_BUMP(arcstat_deleted); 2206168404Spjd bytes_deleted += ab->b_size; 2207185029Spjd 2208185029Spjd if (ab->b_l2hdr != NULL) { 2209185029Spjd /* 2210185029Spjd * This buffer is cached on the 2nd Level ARC; 2211185029Spjd * don't destroy the header. 2212185029Spjd */ 2213185029Spjd arc_change_state(arc_l2c_only, ab, hash_lock); 2214185029Spjd mutex_exit(hash_lock); 2215185029Spjd } else { 2216185029Spjd arc_change_state(arc_anon, ab, hash_lock); 2217185029Spjd mutex_exit(hash_lock); 2218185029Spjd arc_hdr_destroy(ab); 2219185029Spjd } 2220185029Spjd 2221168404Spjd DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 2222168404Spjd if (bytes >= 0 && bytes_deleted >= bytes) 2223168404Spjd break; 2224219089Spjd } else if (bytes < 0) { 2225219089Spjd /* 2226219089Spjd * Insert a list marker and then wait for the 2227219089Spjd * hash lock to become available. Once its 2228219089Spjd * available, restart from where we left off. 2229219089Spjd */ 2230219089Spjd list_insert_after(list, ab, &marker); 2231219089Spjd mutex_exit(lock); 2232219089Spjd mutex_enter(hash_lock); 2233219089Spjd mutex_exit(hash_lock); 2234219089Spjd mutex_enter(lock); 2235219089Spjd ab_prev = list_prev(list, &marker); 2236219089Spjd list_remove(list, &marker); 2237258632Savg } else { 2238168404Spjd bufs_skipped += 1; 2239258632Savg } 2240258632Savg 2241168404Spjd } 2242205231Skmacy mutex_exit(lock); 2243206796Spjd idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2244258632Savg lists++; 2245206796Spjd 2246258632Savg if (lists < list_count) 2247205231Skmacy goto evict_start; 2248206796Spjd 2249205231Skmacy evict_offset = idx; 2250205231Skmacy if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2251185029Spjd (bytes < 0 || bytes_deleted < bytes)) { 2252205231Skmacy list_start = &state->arcs_lists[0]; 2253205231Skmacy list_count = ARC_BUFC_NUMMETADATALISTS; 2254258632Savg offset = lists = 0; 2255205231Skmacy goto evict_start; 2256185029Spjd } 2257185029Spjd 2258168404Spjd if (bufs_skipped) { 2259168404Spjd ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2260168404Spjd ASSERT(bytes >= 0); 2261168404Spjd } 2262168404Spjd 2263168404Spjd if (bytes_deleted < bytes) 2264168404Spjd dprintf("only deleted %lld bytes from %p", 2265168404Spjd (longlong_t)bytes_deleted, state); 2266168404Spjd} 2267168404Spjd 2268168404Spjdstatic void 2269168404Spjdarc_adjust(void) 2270168404Spjd{ 2271208373Smm int64_t adjustment, delta; 2272168404Spjd 2273208373Smm /* 2274208373Smm * Adjust MRU size 2275208373Smm */ 2276168404Spjd 2277209275Smm adjustment = MIN((int64_t)(arc_size - arc_c), 2278209275Smm (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2279209275Smm arc_p)); 2280208373Smm 2281208373Smm if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2282208373Smm delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2283209962Smm (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2284208373Smm adjustment -= delta; 2285168404Spjd } 2286168404Spjd 2287208373Smm if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2288208373Smm delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2289209962Smm (void) arc_evict(arc_mru, 0, delta, FALSE, 2290185029Spjd ARC_BUFC_METADATA); 2291185029Spjd } 2292185029Spjd 2293208373Smm /* 2294208373Smm * Adjust MFU size 2295208373Smm */ 2296168404Spjd 2297208373Smm adjustment = arc_size - arc_c; 2298208373Smm 2299208373Smm if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2300208373Smm delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2301209962Smm (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2302208373Smm adjustment -= delta; 2303168404Spjd } 2304168404Spjd 2305208373Smm if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2306208373Smm int64_t delta = MIN(adjustment, 2307208373Smm arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2308209962Smm (void) arc_evict(arc_mfu, 0, delta, FALSE, 2309208373Smm ARC_BUFC_METADATA); 2310208373Smm } 2311168404Spjd 2312208373Smm /* 2313208373Smm * Adjust ghost lists 2314208373Smm */ 2315168404Spjd 2316208373Smm adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2317168404Spjd 2318208373Smm if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2319208373Smm delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2320209962Smm arc_evict_ghost(arc_mru_ghost, 0, delta); 2321208373Smm } 2322185029Spjd 2323208373Smm adjustment = 2324208373Smm arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2325208373Smm 2326208373Smm if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2327208373Smm delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2328209962Smm arc_evict_ghost(arc_mfu_ghost, 0, delta); 2329168404Spjd } 2330168404Spjd} 2331168404Spjd 2332168404Spjdstatic void 2333168404Spjdarc_do_user_evicts(void) 2334168404Spjd{ 2335191903Skmacy static arc_buf_t *tmp_arc_eviction_list; 2336191903Skmacy 2337191903Skmacy /* 2338191903Skmacy * Move list over to avoid LOR 2339191903Skmacy */ 2340206796Spjdrestart: 2341168404Spjd mutex_enter(&arc_eviction_mtx); 2342191903Skmacy tmp_arc_eviction_list = arc_eviction_list; 2343191903Skmacy arc_eviction_list = NULL; 2344191903Skmacy mutex_exit(&arc_eviction_mtx); 2345191903Skmacy 2346191903Skmacy while (tmp_arc_eviction_list != NULL) { 2347191903Skmacy arc_buf_t *buf = tmp_arc_eviction_list; 2348191903Skmacy tmp_arc_eviction_list = buf->b_next; 2349219089Spjd mutex_enter(&buf->b_evict_lock); 2350168404Spjd buf->b_hdr = NULL; 2351219089Spjd mutex_exit(&buf->b_evict_lock); 2352168404Spjd 2353168404Spjd if (buf->b_efunc != NULL) 2354168404Spjd VERIFY(buf->b_efunc(buf) == 0); 2355168404Spjd 2356168404Spjd buf->b_efunc = NULL; 2357168404Spjd buf->b_private = NULL; 2358168404Spjd kmem_cache_free(buf_cache, buf); 2359168404Spjd } 2360191903Skmacy 2361191903Skmacy if (arc_eviction_list != NULL) 2362191903Skmacy goto restart; 2363168404Spjd} 2364168404Spjd 2365168404Spjd/* 2366185029Spjd * Flush all *evictable* data from the cache for the given spa. 2367168404Spjd * NOTE: this will not touch "active" (i.e. referenced) data. 2368168404Spjd */ 2369168404Spjdvoid 2370185029Spjdarc_flush(spa_t *spa) 2371168404Spjd{ 2372209962Smm uint64_t guid = 0; 2373209962Smm 2374209962Smm if (spa) 2375228103Smm guid = spa_load_guid(spa); 2376209962Smm 2377205231Skmacy while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2378209962Smm (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2379185029Spjd if (spa) 2380185029Spjd break; 2381185029Spjd } 2382205231Skmacy while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2383209962Smm (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2384185029Spjd if (spa) 2385185029Spjd break; 2386185029Spjd } 2387205231Skmacy while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2388209962Smm (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2389185029Spjd if (spa) 2390185029Spjd break; 2391185029Spjd } 2392205231Skmacy while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2393209962Smm (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2394185029Spjd if (spa) 2395185029Spjd break; 2396185029Spjd } 2397168404Spjd 2398209962Smm arc_evict_ghost(arc_mru_ghost, guid, -1); 2399209962Smm arc_evict_ghost(arc_mfu_ghost, guid, -1); 2400168404Spjd 2401168404Spjd mutex_enter(&arc_reclaim_thr_lock); 2402168404Spjd arc_do_user_evicts(); 2403168404Spjd mutex_exit(&arc_reclaim_thr_lock); 2404185029Spjd ASSERT(spa || arc_eviction_list == NULL); 2405168404Spjd} 2406168404Spjd 2407168404Spjdvoid 2408168404Spjdarc_shrink(void) 2409168404Spjd{ 2410168404Spjd if (arc_c > arc_c_min) { 2411168404Spjd uint64_t to_free; 2412168404Spjd 2413168404Spjd#ifdef _KERNEL 2414168404Spjd to_free = arc_c >> arc_shrink_shift; 2415168404Spjd#else 2416168404Spjd to_free = arc_c >> arc_shrink_shift; 2417168404Spjd#endif 2418168404Spjd if (arc_c > arc_c_min + to_free) 2419168404Spjd atomic_add_64(&arc_c, -to_free); 2420168404Spjd else 2421168404Spjd arc_c = arc_c_min; 2422168404Spjd 2423168404Spjd atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2424168404Spjd if (arc_c > arc_size) 2425168404Spjd arc_c = MAX(arc_size, arc_c_min); 2426168404Spjd if (arc_p > arc_c) 2427168404Spjd arc_p = (arc_c >> 1); 2428168404Spjd ASSERT(arc_c >= arc_c_min); 2429168404Spjd ASSERT((int64_t)arc_p >= 0); 2430168404Spjd } 2431168404Spjd 2432168404Spjd if (arc_size > arc_c) 2433168404Spjd arc_adjust(); 2434168404Spjd} 2435168404Spjd 2436185029Spjdstatic int needfree = 0; 2437168404Spjd 2438168404Spjdstatic int 2439168404Spjdarc_reclaim_needed(void) 2440168404Spjd{ 2441168404Spjd 2442168404Spjd#ifdef _KERNEL 2443219089Spjd 2444197816Skmacy if (needfree) 2445197816Skmacy return (1); 2446168404Spjd 2447191902Skmacy /* 2448212780Savg * Cooperate with pagedaemon when it's time for it to scan 2449212780Savg * and reclaim some pages. 2450191902Skmacy */ 2451212783Savg if (vm_paging_needed()) 2452191902Skmacy return (1); 2453191902Skmacy 2454219089Spjd#ifdef sun 2455168404Spjd /* 2456185029Spjd * take 'desfree' extra pages, so we reclaim sooner, rather than later 2457185029Spjd */ 2458185029Spjd extra = desfree; 2459185029Spjd 2460185029Spjd /* 2461185029Spjd * check that we're out of range of the pageout scanner. It starts to 2462185029Spjd * schedule paging if freemem is less than lotsfree and needfree. 2463185029Spjd * lotsfree is the high-water mark for pageout, and needfree is the 2464185029Spjd * number of needed free pages. We add extra pages here to make sure 2465185029Spjd * the scanner doesn't start up while we're freeing memory. 2466185029Spjd */ 2467185029Spjd if (freemem < lotsfree + needfree + extra) 2468185029Spjd return (1); 2469185029Spjd 2470185029Spjd /* 2471168404Spjd * check to make sure that swapfs has enough space so that anon 2472185029Spjd * reservations can still succeed. anon_resvmem() checks that the 2473168404Spjd * availrmem is greater than swapfs_minfree, and the number of reserved 2474168404Spjd * swap pages. We also add a bit of extra here just to prevent 2475168404Spjd * circumstances from getting really dire. 2476168404Spjd */ 2477168404Spjd if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2478168404Spjd return (1); 2479168404Spjd 2480168404Spjd#if defined(__i386) 2481168404Spjd /* 2482168404Spjd * If we're on an i386 platform, it's possible that we'll exhaust the 2483168404Spjd * kernel heap space before we ever run out of available physical 2484168404Spjd * memory. Most checks of the size of the heap_area compare against 2485168404Spjd * tune.t_minarmem, which is the minimum available real memory that we 2486168404Spjd * can have in the system. However, this is generally fixed at 25 pages 2487168404Spjd * which is so low that it's useless. In this comparison, we seek to 2488168404Spjd * calculate the total heap-size, and reclaim if more than 3/4ths of the 2489185029Spjd * heap is allocated. (Or, in the calculation, if less than 1/4th is 2490168404Spjd * free) 2491168404Spjd */ 2492168404Spjd if (btop(vmem_size(heap_arena, VMEM_FREE)) < 2493168404Spjd (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2)) 2494168404Spjd return (1); 2495168404Spjd#endif 2496219089Spjd#else /* !sun */ 2497175633Spjd if (kmem_used() > (kmem_size() * 3) / 4) 2498168404Spjd return (1); 2499219089Spjd#endif /* sun */ 2500168404Spjd 2501168404Spjd#else 2502168404Spjd if (spa_get_random(100) == 0) 2503168404Spjd return (1); 2504168404Spjd#endif 2505168404Spjd return (0); 2506168404Spjd} 2507168404Spjd 2508208454Spjdextern kmem_cache_t *zio_buf_cache[]; 2509208454Spjdextern kmem_cache_t *zio_data_buf_cache[]; 2510208454Spjd 2511168404Spjdstatic void 2512168404Spjdarc_kmem_reap_now(arc_reclaim_strategy_t strat) 2513168404Spjd{ 2514168404Spjd size_t i; 2515168404Spjd kmem_cache_t *prev_cache = NULL; 2516168404Spjd kmem_cache_t *prev_data_cache = NULL; 2517168404Spjd 2518168404Spjd#ifdef _KERNEL 2519185029Spjd if (arc_meta_used >= arc_meta_limit) { 2520185029Spjd /* 2521185029Spjd * We are exceeding our meta-data cache limit. 2522185029Spjd * Purge some DNLC entries to release holds on meta-data. 2523185029Spjd */ 2524185029Spjd dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2525185029Spjd } 2526168404Spjd#if defined(__i386) 2527168404Spjd /* 2528168404Spjd * Reclaim unused memory from all kmem caches. 2529168404Spjd */ 2530168404Spjd kmem_reap(); 2531168404Spjd#endif 2532168404Spjd#endif 2533168404Spjd 2534168404Spjd /* 2535185029Spjd * An aggressive reclamation will shrink the cache size as well as 2536168404Spjd * reap free buffers from the arc kmem caches. 2537168404Spjd */ 2538168404Spjd if (strat == ARC_RECLAIM_AGGR) 2539168404Spjd arc_shrink(); 2540168404Spjd 2541168404Spjd for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2542168404Spjd if (zio_buf_cache[i] != prev_cache) { 2543168404Spjd prev_cache = zio_buf_cache[i]; 2544168404Spjd kmem_cache_reap_now(zio_buf_cache[i]); 2545168404Spjd } 2546168404Spjd if (zio_data_buf_cache[i] != prev_data_cache) { 2547168404Spjd prev_data_cache = zio_data_buf_cache[i]; 2548168404Spjd kmem_cache_reap_now(zio_data_buf_cache[i]); 2549168404Spjd } 2550168404Spjd } 2551168404Spjd kmem_cache_reap_now(buf_cache); 2552168404Spjd kmem_cache_reap_now(hdr_cache); 2553168404Spjd} 2554168404Spjd 2555168404Spjdstatic void 2556168404Spjdarc_reclaim_thread(void *dummy __unused) 2557168404Spjd{ 2558168404Spjd clock_t growtime = 0; 2559168404Spjd arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2560168404Spjd callb_cpr_t cpr; 2561168404Spjd 2562168404Spjd CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2563168404Spjd 2564168404Spjd mutex_enter(&arc_reclaim_thr_lock); 2565168404Spjd while (arc_thread_exit == 0) { 2566168404Spjd if (arc_reclaim_needed()) { 2567168404Spjd 2568168404Spjd if (arc_no_grow) { 2569168404Spjd if (last_reclaim == ARC_RECLAIM_CONS) { 2570168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2571168404Spjd } else { 2572168404Spjd last_reclaim = ARC_RECLAIM_CONS; 2573168404Spjd } 2574168404Spjd } else { 2575168404Spjd arc_no_grow = TRUE; 2576168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2577168404Spjd membar_producer(); 2578168404Spjd } 2579168404Spjd 2580168404Spjd /* reset the growth delay for every reclaim */ 2581219089Spjd growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2582168404Spjd 2583185029Spjd if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2584168404Spjd /* 2585185029Spjd * If needfree is TRUE our vm_lowmem hook 2586168404Spjd * was called and in that case we must free some 2587168404Spjd * memory, so switch to aggressive mode. 2588168404Spjd */ 2589168404Spjd arc_no_grow = TRUE; 2590168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2591168404Spjd } 2592168404Spjd arc_kmem_reap_now(last_reclaim); 2593185029Spjd arc_warm = B_TRUE; 2594185029Spjd 2595219089Spjd } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2596168404Spjd arc_no_grow = FALSE; 2597168404Spjd } 2598168404Spjd 2599209275Smm arc_adjust(); 2600168404Spjd 2601168404Spjd if (arc_eviction_list != NULL) 2602168404Spjd arc_do_user_evicts(); 2603168404Spjd 2604211762Savg#ifdef _KERNEL 2605211762Savg if (needfree) { 2606185029Spjd needfree = 0; 2607185029Spjd wakeup(&needfree); 2608211762Savg } 2609168404Spjd#endif 2610168404Spjd 2611168404Spjd /* block until needed, or one second, whichever is shorter */ 2612168404Spjd CALLB_CPR_SAFE_BEGIN(&cpr); 2613168404Spjd (void) cv_timedwait(&arc_reclaim_thr_cv, 2614168404Spjd &arc_reclaim_thr_lock, hz); 2615168404Spjd CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2616168404Spjd } 2617168404Spjd 2618168404Spjd arc_thread_exit = 0; 2619168404Spjd cv_broadcast(&arc_reclaim_thr_cv); 2620168404Spjd CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2621168404Spjd thread_exit(); 2622168404Spjd} 2623168404Spjd 2624168404Spjd/* 2625168404Spjd * Adapt arc info given the number of bytes we are trying to add and 2626168404Spjd * the state that we are comming from. This function is only called 2627168404Spjd * when we are adding new content to the cache. 2628168404Spjd */ 2629168404Spjdstatic void 2630168404Spjdarc_adapt(int bytes, arc_state_t *state) 2631168404Spjd{ 2632168404Spjd int mult; 2633208373Smm uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2634168404Spjd 2635185029Spjd if (state == arc_l2c_only) 2636185029Spjd return; 2637185029Spjd 2638168404Spjd ASSERT(bytes > 0); 2639168404Spjd /* 2640168404Spjd * Adapt the target size of the MRU list: 2641168404Spjd * - if we just hit in the MRU ghost list, then increase 2642168404Spjd * the target size of the MRU list. 2643168404Spjd * - if we just hit in the MFU ghost list, then increase 2644168404Spjd * the target size of the MFU list by decreasing the 2645168404Spjd * target size of the MRU list. 2646168404Spjd */ 2647168404Spjd if (state == arc_mru_ghost) { 2648168404Spjd mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2649168404Spjd 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2650209275Smm mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2651168404Spjd 2652208373Smm arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2653168404Spjd } else if (state == arc_mfu_ghost) { 2654208373Smm uint64_t delta; 2655208373Smm 2656168404Spjd mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2657168404Spjd 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2658209275Smm mult = MIN(mult, 10); 2659168404Spjd 2660208373Smm delta = MIN(bytes * mult, arc_p); 2661208373Smm arc_p = MAX(arc_p_min, arc_p - delta); 2662168404Spjd } 2663168404Spjd ASSERT((int64_t)arc_p >= 0); 2664168404Spjd 2665168404Spjd if (arc_reclaim_needed()) { 2666168404Spjd cv_signal(&arc_reclaim_thr_cv); 2667168404Spjd return; 2668168404Spjd } 2669168404Spjd 2670168404Spjd if (arc_no_grow) 2671168404Spjd return; 2672168404Spjd 2673168404Spjd if (arc_c >= arc_c_max) 2674168404Spjd return; 2675168404Spjd 2676168404Spjd /* 2677168404Spjd * If we're within (2 * maxblocksize) bytes of the target 2678168404Spjd * cache size, increment the target cache size 2679168404Spjd */ 2680168404Spjd if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2681168404Spjd atomic_add_64(&arc_c, (int64_t)bytes); 2682168404Spjd if (arc_c > arc_c_max) 2683168404Spjd arc_c = arc_c_max; 2684168404Spjd else if (state == arc_anon) 2685168404Spjd atomic_add_64(&arc_p, (int64_t)bytes); 2686168404Spjd if (arc_p > arc_c) 2687168404Spjd arc_p = arc_c; 2688168404Spjd } 2689168404Spjd ASSERT((int64_t)arc_p >= 0); 2690168404Spjd} 2691168404Spjd 2692168404Spjd/* 2693168404Spjd * Check if the cache has reached its limits and eviction is required 2694168404Spjd * prior to insert. 2695168404Spjd */ 2696168404Spjdstatic int 2697185029Spjdarc_evict_needed(arc_buf_contents_t type) 2698168404Spjd{ 2699185029Spjd if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2700185029Spjd return (1); 2701185029Spjd 2702219089Spjd#ifdef sun 2703185029Spjd#ifdef _KERNEL 2704185029Spjd /* 2705185029Spjd * If zio data pages are being allocated out of a separate heap segment, 2706185029Spjd * then enforce that the size of available vmem for this area remains 2707185029Spjd * above about 1/32nd free. 2708185029Spjd */ 2709185029Spjd if (type == ARC_BUFC_DATA && zio_arena != NULL && 2710185029Spjd vmem_size(zio_arena, VMEM_FREE) < 2711185029Spjd (vmem_size(zio_arena, VMEM_ALLOC) >> 5)) 2712185029Spjd return (1); 2713185029Spjd#endif 2714219089Spjd#endif /* sun */ 2715185029Spjd 2716168404Spjd if (arc_reclaim_needed()) 2717168404Spjd return (1); 2718168404Spjd 2719168404Spjd return (arc_size > arc_c); 2720168404Spjd} 2721168404Spjd 2722168404Spjd/* 2723168404Spjd * The buffer, supplied as the first argument, needs a data block. 2724168404Spjd * So, if we are at cache max, determine which cache should be victimized. 2725168404Spjd * We have the following cases: 2726168404Spjd * 2727168404Spjd * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2728168404Spjd * In this situation if we're out of space, but the resident size of the MFU is 2729168404Spjd * under the limit, victimize the MFU cache to satisfy this insertion request. 2730168404Spjd * 2731168404Spjd * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2732168404Spjd * Here, we've used up all of the available space for the MRU, so we need to 2733168404Spjd * evict from our own cache instead. Evict from the set of resident MRU 2734168404Spjd * entries. 2735168404Spjd * 2736168404Spjd * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2737168404Spjd * c minus p represents the MFU space in the cache, since p is the size of the 2738168404Spjd * cache that is dedicated to the MRU. In this situation there's still space on 2739168404Spjd * the MFU side, so the MRU side needs to be victimized. 2740168404Spjd * 2741168404Spjd * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2742168404Spjd * MFU's resident set is consuming more space than it has been allotted. In 2743168404Spjd * this situation, we must victimize our own cache, the MFU, for this insertion. 2744168404Spjd */ 2745168404Spjdstatic void 2746168404Spjdarc_get_data_buf(arc_buf_t *buf) 2747168404Spjd{ 2748168404Spjd arc_state_t *state = buf->b_hdr->b_state; 2749168404Spjd uint64_t size = buf->b_hdr->b_size; 2750168404Spjd arc_buf_contents_t type = buf->b_hdr->b_type; 2751168404Spjd 2752168404Spjd arc_adapt(size, state); 2753168404Spjd 2754168404Spjd /* 2755168404Spjd * We have not yet reached cache maximum size, 2756168404Spjd * just allocate a new buffer. 2757168404Spjd */ 2758185029Spjd if (!arc_evict_needed(type)) { 2759168404Spjd if (type == ARC_BUFC_METADATA) { 2760168404Spjd buf->b_data = zio_buf_alloc(size); 2761208373Smm arc_space_consume(size, ARC_SPACE_DATA); 2762168404Spjd } else { 2763168404Spjd ASSERT(type == ARC_BUFC_DATA); 2764168404Spjd buf->b_data = zio_data_buf_alloc(size); 2765208373Smm ARCSTAT_INCR(arcstat_data_size, size); 2766185029Spjd atomic_add_64(&arc_size, size); 2767168404Spjd } 2768168404Spjd goto out; 2769168404Spjd } 2770168404Spjd 2771168404Spjd /* 2772168404Spjd * If we are prefetching from the mfu ghost list, this buffer 2773168404Spjd * will end up on the mru list; so steal space from there. 2774168404Spjd */ 2775168404Spjd if (state == arc_mfu_ghost) 2776168404Spjd state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2777168404Spjd else if (state == arc_mru_ghost) 2778168404Spjd state = arc_mru; 2779168404Spjd 2780168404Spjd if (state == arc_mru || state == arc_anon) { 2781168404Spjd uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2782208373Smm state = (arc_mfu->arcs_lsize[type] >= size && 2783185029Spjd arc_p > mru_used) ? arc_mfu : arc_mru; 2784168404Spjd } else { 2785168404Spjd /* MFU cases */ 2786168404Spjd uint64_t mfu_space = arc_c - arc_p; 2787208373Smm state = (arc_mru->arcs_lsize[type] >= size && 2788185029Spjd mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2789168404Spjd } 2790209962Smm if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2791168404Spjd if (type == ARC_BUFC_METADATA) { 2792168404Spjd buf->b_data = zio_buf_alloc(size); 2793208373Smm arc_space_consume(size, ARC_SPACE_DATA); 2794168404Spjd } else { 2795168404Spjd ASSERT(type == ARC_BUFC_DATA); 2796168404Spjd buf->b_data = zio_data_buf_alloc(size); 2797208373Smm ARCSTAT_INCR(arcstat_data_size, size); 2798185029Spjd atomic_add_64(&arc_size, size); 2799168404Spjd } 2800168404Spjd ARCSTAT_BUMP(arcstat_recycle_miss); 2801168404Spjd } 2802168404Spjd ASSERT(buf->b_data != NULL); 2803168404Spjdout: 2804168404Spjd /* 2805168404Spjd * Update the state size. Note that ghost states have a 2806168404Spjd * "ghost size" and so don't need to be updated. 2807168404Spjd */ 2808168404Spjd if (!GHOST_STATE(buf->b_hdr->b_state)) { 2809168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 2810168404Spjd 2811168404Spjd atomic_add_64(&hdr->b_state->arcs_size, size); 2812168404Spjd if (list_link_active(&hdr->b_arc_node)) { 2813168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2814185029Spjd atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2815168404Spjd } 2816168404Spjd /* 2817168404Spjd * If we are growing the cache, and we are adding anonymous 2818168404Spjd * data, and we have outgrown arc_p, update arc_p 2819168404Spjd */ 2820168404Spjd if (arc_size < arc_c && hdr->b_state == arc_anon && 2821168404Spjd arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2822168404Spjd arc_p = MIN(arc_c, arc_p + size); 2823168404Spjd } 2824205231Skmacy ARCSTAT_BUMP(arcstat_allocated); 2825168404Spjd} 2826168404Spjd 2827168404Spjd/* 2828168404Spjd * This routine is called whenever a buffer is accessed. 2829168404Spjd * NOTE: the hash lock is dropped in this function. 2830168404Spjd */ 2831168404Spjdstatic void 2832168404Spjdarc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2833168404Spjd{ 2834219089Spjd clock_t now; 2835219089Spjd 2836168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 2837168404Spjd 2838168404Spjd if (buf->b_state == arc_anon) { 2839168404Spjd /* 2840168404Spjd * This buffer is not in the cache, and does not 2841168404Spjd * appear in our "ghost" list. Add the new buffer 2842168404Spjd * to the MRU state. 2843168404Spjd */ 2844168404Spjd 2845168404Spjd ASSERT(buf->b_arc_access == 0); 2846219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2847168404Spjd DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2848168404Spjd arc_change_state(arc_mru, buf, hash_lock); 2849168404Spjd 2850168404Spjd } else if (buf->b_state == arc_mru) { 2851219089Spjd now = ddi_get_lbolt(); 2852219089Spjd 2853168404Spjd /* 2854168404Spjd * If this buffer is here because of a prefetch, then either: 2855168404Spjd * - clear the flag if this is a "referencing" read 2856168404Spjd * (any subsequent access will bump this into the MFU state). 2857168404Spjd * or 2858168404Spjd * - move the buffer to the head of the list if this is 2859168404Spjd * another prefetch (to make it less likely to be evicted). 2860168404Spjd */ 2861168404Spjd if ((buf->b_flags & ARC_PREFETCH) != 0) { 2862168404Spjd if (refcount_count(&buf->b_refcnt) == 0) { 2863168404Spjd ASSERT(list_link_active(&buf->b_arc_node)); 2864168404Spjd } else { 2865168404Spjd buf->b_flags &= ~ARC_PREFETCH; 2866168404Spjd ARCSTAT_BUMP(arcstat_mru_hits); 2867168404Spjd } 2868219089Spjd buf->b_arc_access = now; 2869168404Spjd return; 2870168404Spjd } 2871168404Spjd 2872168404Spjd /* 2873168404Spjd * This buffer has been "accessed" only once so far, 2874168404Spjd * but it is still in the cache. Move it to the MFU 2875168404Spjd * state. 2876168404Spjd */ 2877219089Spjd if (now > buf->b_arc_access + ARC_MINTIME) { 2878168404Spjd /* 2879168404Spjd * More than 125ms have passed since we 2880168404Spjd * instantiated this buffer. Move it to the 2881168404Spjd * most frequently used state. 2882168404Spjd */ 2883219089Spjd buf->b_arc_access = now; 2884168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2885168404Spjd arc_change_state(arc_mfu, buf, hash_lock); 2886168404Spjd } 2887168404Spjd ARCSTAT_BUMP(arcstat_mru_hits); 2888168404Spjd } else if (buf->b_state == arc_mru_ghost) { 2889168404Spjd arc_state_t *new_state; 2890168404Spjd /* 2891168404Spjd * This buffer has been "accessed" recently, but 2892168404Spjd * was evicted from the cache. Move it to the 2893168404Spjd * MFU state. 2894168404Spjd */ 2895168404Spjd 2896168404Spjd if (buf->b_flags & ARC_PREFETCH) { 2897168404Spjd new_state = arc_mru; 2898168404Spjd if (refcount_count(&buf->b_refcnt) > 0) 2899168404Spjd buf->b_flags &= ~ARC_PREFETCH; 2900168404Spjd DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2901168404Spjd } else { 2902168404Spjd new_state = arc_mfu; 2903168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2904168404Spjd } 2905168404Spjd 2906219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2907168404Spjd arc_change_state(new_state, buf, hash_lock); 2908168404Spjd 2909168404Spjd ARCSTAT_BUMP(arcstat_mru_ghost_hits); 2910168404Spjd } else if (buf->b_state == arc_mfu) { 2911168404Spjd /* 2912168404Spjd * This buffer has been accessed more than once and is 2913168404Spjd * still in the cache. Keep it in the MFU state. 2914168404Spjd * 2915168404Spjd * NOTE: an add_reference() that occurred when we did 2916168404Spjd * the arc_read() will have kicked this off the list. 2917168404Spjd * If it was a prefetch, we will explicitly move it to 2918168404Spjd * the head of the list now. 2919168404Spjd */ 2920168404Spjd if ((buf->b_flags & ARC_PREFETCH) != 0) { 2921168404Spjd ASSERT(refcount_count(&buf->b_refcnt) == 0); 2922168404Spjd ASSERT(list_link_active(&buf->b_arc_node)); 2923168404Spjd } 2924168404Spjd ARCSTAT_BUMP(arcstat_mfu_hits); 2925219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2926168404Spjd } else if (buf->b_state == arc_mfu_ghost) { 2927168404Spjd arc_state_t *new_state = arc_mfu; 2928168404Spjd /* 2929168404Spjd * This buffer has been accessed more than once but has 2930168404Spjd * been evicted from the cache. Move it back to the 2931168404Spjd * MFU state. 2932168404Spjd */ 2933168404Spjd 2934168404Spjd if (buf->b_flags & ARC_PREFETCH) { 2935168404Spjd /* 2936168404Spjd * This is a prefetch access... 2937168404Spjd * move this block back to the MRU state. 2938168404Spjd */ 2939240415Smm ASSERT0(refcount_count(&buf->b_refcnt)); 2940168404Spjd new_state = arc_mru; 2941168404Spjd } 2942168404Spjd 2943219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2944168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2945168404Spjd arc_change_state(new_state, buf, hash_lock); 2946168404Spjd 2947168404Spjd ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 2948185029Spjd } else if (buf->b_state == arc_l2c_only) { 2949185029Spjd /* 2950185029Spjd * This buffer is on the 2nd Level ARC. 2951185029Spjd */ 2952185029Spjd 2953219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2954185029Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2955185029Spjd arc_change_state(arc_mfu, buf, hash_lock); 2956168404Spjd } else { 2957168404Spjd ASSERT(!"invalid arc state"); 2958168404Spjd } 2959168404Spjd} 2960168404Spjd 2961168404Spjd/* a generic arc_done_func_t which you can use */ 2962168404Spjd/* ARGSUSED */ 2963168404Spjdvoid 2964168404Spjdarc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 2965168404Spjd{ 2966219089Spjd if (zio == NULL || zio->io_error == 0) 2967219089Spjd bcopy(buf->b_data, arg, buf->b_hdr->b_size); 2968248571Smm VERIFY(arc_buf_remove_ref(buf, arg)); 2969168404Spjd} 2970168404Spjd 2971185029Spjd/* a generic arc_done_func_t */ 2972168404Spjdvoid 2973168404Spjdarc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 2974168404Spjd{ 2975168404Spjd arc_buf_t **bufp = arg; 2976168404Spjd if (zio && zio->io_error) { 2977248571Smm VERIFY(arc_buf_remove_ref(buf, arg)); 2978168404Spjd *bufp = NULL; 2979168404Spjd } else { 2980168404Spjd *bufp = buf; 2981219089Spjd ASSERT(buf->b_data); 2982168404Spjd } 2983168404Spjd} 2984168404Spjd 2985168404Spjdstatic void 2986168404Spjdarc_read_done(zio_t *zio) 2987168404Spjd{ 2988268075Sdelphij arc_buf_hdr_t *hdr; 2989168404Spjd arc_buf_t *buf; 2990168404Spjd arc_buf_t *abuf; /* buffer we're assigning to callback */ 2991268075Sdelphij kmutex_t *hash_lock = NULL; 2992168404Spjd arc_callback_t *callback_list, *acb; 2993168404Spjd int freeable = FALSE; 2994168404Spjd 2995168404Spjd buf = zio->io_private; 2996168404Spjd hdr = buf->b_hdr; 2997168404Spjd 2998168404Spjd /* 2999168404Spjd * The hdr was inserted into hash-table and removed from lists 3000168404Spjd * prior to starting I/O. We should find this header, since 3001168404Spjd * it's in the hash table, and it should be legit since it's 3002168404Spjd * not possible to evict it during the I/O. The only possible 3003168404Spjd * reason for it not to be found is if we were freed during the 3004168404Spjd * read. 3005168404Spjd */ 3006268075Sdelphij if (HDR_IN_HASH_TABLE(hdr)) { 3007268075Sdelphij ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3008268075Sdelphij ASSERT3U(hdr->b_dva.dva_word[0], ==, 3009268075Sdelphij BP_IDENTITY(zio->io_bp)->dva_word[0]); 3010268075Sdelphij ASSERT3U(hdr->b_dva.dva_word[1], ==, 3011268075Sdelphij BP_IDENTITY(zio->io_bp)->dva_word[1]); 3012168404Spjd 3013268075Sdelphij arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3014268075Sdelphij &hash_lock); 3015168404Spjd 3016268075Sdelphij ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3017268075Sdelphij hash_lock == NULL) || 3018268075Sdelphij (found == hdr && 3019268075Sdelphij DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3020268075Sdelphij (found == hdr && HDR_L2_READING(hdr))); 3021268075Sdelphij } 3022268075Sdelphij 3023185029Spjd hdr->b_flags &= ~ARC_L2_EVICTED; 3024185029Spjd if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 3025185029Spjd hdr->b_flags &= ~ARC_L2CACHE; 3026206796Spjd 3027168404Spjd /* byteswap if necessary */ 3028168404Spjd callback_list = hdr->b_acb; 3029168404Spjd ASSERT(callback_list != NULL); 3030209101Smm if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3031236884Smm dmu_object_byteswap_t bswap = 3032236884Smm DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3033185029Spjd arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3034185029Spjd byteswap_uint64_array : 3035236884Smm dmu_ot_byteswap[bswap].ob_func; 3036185029Spjd func(buf->b_data, hdr->b_size); 3037185029Spjd } 3038168404Spjd 3039185029Spjd arc_cksum_compute(buf, B_FALSE); 3040240133Smm#ifdef illumos 3041240133Smm arc_buf_watch(buf); 3042240133Smm#endif /* illumos */ 3043168404Spjd 3044219089Spjd if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3045219089Spjd /* 3046219089Spjd * Only call arc_access on anonymous buffers. This is because 3047219089Spjd * if we've issued an I/O for an evicted buffer, we've already 3048219089Spjd * called arc_access (to prevent any simultaneous readers from 3049219089Spjd * getting confused). 3050219089Spjd */ 3051219089Spjd arc_access(hdr, hash_lock); 3052219089Spjd } 3053219089Spjd 3054168404Spjd /* create copies of the data buffer for the callers */ 3055168404Spjd abuf = buf; 3056168404Spjd for (acb = callback_list; acb; acb = acb->acb_next) { 3057168404Spjd if (acb->acb_done) { 3058242845Sdelphij if (abuf == NULL) { 3059242845Sdelphij ARCSTAT_BUMP(arcstat_duplicate_reads); 3060168404Spjd abuf = arc_buf_clone(buf); 3061242845Sdelphij } 3062168404Spjd acb->acb_buf = abuf; 3063168404Spjd abuf = NULL; 3064168404Spjd } 3065168404Spjd } 3066168404Spjd hdr->b_acb = NULL; 3067168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3068168404Spjd ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3069219089Spjd if (abuf == buf) { 3070219089Spjd ASSERT(buf->b_efunc == NULL); 3071219089Spjd ASSERT(hdr->b_datacnt == 1); 3072168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 3073219089Spjd } 3074168404Spjd 3075168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3076168404Spjd 3077168404Spjd if (zio->io_error != 0) { 3078168404Spjd hdr->b_flags |= ARC_IO_ERROR; 3079168404Spjd if (hdr->b_state != arc_anon) 3080168404Spjd arc_change_state(arc_anon, hdr, hash_lock); 3081168404Spjd if (HDR_IN_HASH_TABLE(hdr)) 3082168404Spjd buf_hash_remove(hdr); 3083168404Spjd freeable = refcount_is_zero(&hdr->b_refcnt); 3084168404Spjd } 3085168404Spjd 3086168404Spjd /* 3087168404Spjd * Broadcast before we drop the hash_lock to avoid the possibility 3088168404Spjd * that the hdr (and hence the cv) might be freed before we get to 3089168404Spjd * the cv_broadcast(). 3090168404Spjd */ 3091168404Spjd cv_broadcast(&hdr->b_cv); 3092168404Spjd 3093168404Spjd if (hash_lock) { 3094168404Spjd mutex_exit(hash_lock); 3095168404Spjd } else { 3096168404Spjd /* 3097168404Spjd * This block was freed while we waited for the read to 3098168404Spjd * complete. It has been removed from the hash table and 3099168404Spjd * moved to the anonymous state (so that it won't show up 3100168404Spjd * in the cache). 3101168404Spjd */ 3102168404Spjd ASSERT3P(hdr->b_state, ==, arc_anon); 3103168404Spjd freeable = refcount_is_zero(&hdr->b_refcnt); 3104168404Spjd } 3105168404Spjd 3106168404Spjd /* execute each callback and free its structure */ 3107168404Spjd while ((acb = callback_list) != NULL) { 3108168404Spjd if (acb->acb_done) 3109168404Spjd acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3110168404Spjd 3111168404Spjd if (acb->acb_zio_dummy != NULL) { 3112168404Spjd acb->acb_zio_dummy->io_error = zio->io_error; 3113168404Spjd zio_nowait(acb->acb_zio_dummy); 3114168404Spjd } 3115168404Spjd 3116168404Spjd callback_list = acb->acb_next; 3117168404Spjd kmem_free(acb, sizeof (arc_callback_t)); 3118168404Spjd } 3119168404Spjd 3120168404Spjd if (freeable) 3121168404Spjd arc_hdr_destroy(hdr); 3122168404Spjd} 3123168404Spjd 3124168404Spjd/* 3125168404Spjd * "Read" the block block at the specified DVA (in bp) via the 3126168404Spjd * cache. If the block is found in the cache, invoke the provided 3127168404Spjd * callback immediately and return. Note that the `zio' parameter 3128168404Spjd * in the callback will be NULL in this case, since no IO was 3129168404Spjd * required. If the block is not in the cache pass the read request 3130168404Spjd * on to the spa with a substitute callback function, so that the 3131168404Spjd * requested block will be added to the cache. 3132168404Spjd * 3133168404Spjd * If a read request arrives for a block that has a read in-progress, 3134168404Spjd * either wait for the in-progress read to complete (and return the 3135168404Spjd * results); or, if this is a read with a "done" func, add a record 3136168404Spjd * to the read to invoke the "done" func when the read completes, 3137168404Spjd * and return; or just return. 3138168404Spjd * 3139168404Spjd * arc_read_done() will invoke all the requested "done" functions 3140168404Spjd * for readers of this block. 3141168404Spjd */ 3142168404Spjdint 3143246666Smmarc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3144258632Savg void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags, 3145246666Smm const zbookmark_t *zb) 3146168404Spjd{ 3147268075Sdelphij arc_buf_hdr_t *hdr = NULL; 3148247187Smm arc_buf_t *buf = NULL; 3149268075Sdelphij kmutex_t *hash_lock = NULL; 3150185029Spjd zio_t *rzio; 3151228103Smm uint64_t guid = spa_load_guid(spa); 3152168404Spjd 3153268075Sdelphij ASSERT(!BP_IS_EMBEDDED(bp) || 3154268075Sdelphij BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3155268075Sdelphij 3156168404Spjdtop: 3157268075Sdelphij if (!BP_IS_EMBEDDED(bp)) { 3158268075Sdelphij /* 3159268075Sdelphij * Embedded BP's have no DVA and require no I/O to "read". 3160268075Sdelphij * Create an anonymous arc buf to back it. 3161268075Sdelphij */ 3162268075Sdelphij hdr = buf_hash_find(guid, bp, &hash_lock); 3163268075Sdelphij } 3164168404Spjd 3165268075Sdelphij if (hdr != NULL && hdr->b_datacnt > 0) { 3166268075Sdelphij 3167168404Spjd *arc_flags |= ARC_CACHED; 3168168404Spjd 3169168404Spjd if (HDR_IO_IN_PROGRESS(hdr)) { 3170168404Spjd 3171168404Spjd if (*arc_flags & ARC_WAIT) { 3172168404Spjd cv_wait(&hdr->b_cv, hash_lock); 3173168404Spjd mutex_exit(hash_lock); 3174168404Spjd goto top; 3175168404Spjd } 3176168404Spjd ASSERT(*arc_flags & ARC_NOWAIT); 3177168404Spjd 3178168404Spjd if (done) { 3179168404Spjd arc_callback_t *acb = NULL; 3180168404Spjd 3181168404Spjd acb = kmem_zalloc(sizeof (arc_callback_t), 3182168404Spjd KM_SLEEP); 3183168404Spjd acb->acb_done = done; 3184168404Spjd acb->acb_private = private; 3185168404Spjd if (pio != NULL) 3186168404Spjd acb->acb_zio_dummy = zio_null(pio, 3187209962Smm spa, NULL, NULL, NULL, zio_flags); 3188168404Spjd 3189168404Spjd ASSERT(acb->acb_done != NULL); 3190168404Spjd acb->acb_next = hdr->b_acb; 3191168404Spjd hdr->b_acb = acb; 3192168404Spjd add_reference(hdr, hash_lock, private); 3193168404Spjd mutex_exit(hash_lock); 3194168404Spjd return (0); 3195168404Spjd } 3196168404Spjd mutex_exit(hash_lock); 3197168404Spjd return (0); 3198168404Spjd } 3199168404Spjd 3200168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3201168404Spjd 3202168404Spjd if (done) { 3203168404Spjd add_reference(hdr, hash_lock, private); 3204168404Spjd /* 3205168404Spjd * If this block is already in use, create a new 3206168404Spjd * copy of the data so that we will be guaranteed 3207168404Spjd * that arc_release() will always succeed. 3208168404Spjd */ 3209168404Spjd buf = hdr->b_buf; 3210168404Spjd ASSERT(buf); 3211168404Spjd ASSERT(buf->b_data); 3212168404Spjd if (HDR_BUF_AVAILABLE(hdr)) { 3213168404Spjd ASSERT(buf->b_efunc == NULL); 3214168404Spjd hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3215168404Spjd } else { 3216168404Spjd buf = arc_buf_clone(buf); 3217168404Spjd } 3218219089Spjd 3219168404Spjd } else if (*arc_flags & ARC_PREFETCH && 3220168404Spjd refcount_count(&hdr->b_refcnt) == 0) { 3221168404Spjd hdr->b_flags |= ARC_PREFETCH; 3222168404Spjd } 3223168404Spjd DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3224168404Spjd arc_access(hdr, hash_lock); 3225185029Spjd if (*arc_flags & ARC_L2CACHE) 3226185029Spjd hdr->b_flags |= ARC_L2CACHE; 3227251478Sdelphij if (*arc_flags & ARC_L2COMPRESS) 3228251478Sdelphij hdr->b_flags |= ARC_L2COMPRESS; 3229168404Spjd mutex_exit(hash_lock); 3230168404Spjd ARCSTAT_BUMP(arcstat_hits); 3231168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3232168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3233168404Spjd data, metadata, hits); 3234168404Spjd 3235168404Spjd if (done) 3236168404Spjd done(NULL, buf, private); 3237168404Spjd } else { 3238168404Spjd uint64_t size = BP_GET_LSIZE(bp); 3239268075Sdelphij arc_callback_t *acb; 3240185029Spjd vdev_t *vd = NULL; 3241247187Smm uint64_t addr = 0; 3242208373Smm boolean_t devw = B_FALSE; 3243258389Savg enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3244258389Savg uint64_t b_asize = 0; 3245168404Spjd 3246168404Spjd if (hdr == NULL) { 3247168404Spjd /* this block is not in the cache */ 3248268075Sdelphij arc_buf_hdr_t *exists = NULL; 3249168404Spjd arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3250168404Spjd buf = arc_buf_alloc(spa, size, private, type); 3251168404Spjd hdr = buf->b_hdr; 3252268075Sdelphij if (!BP_IS_EMBEDDED(bp)) { 3253268075Sdelphij hdr->b_dva = *BP_IDENTITY(bp); 3254268075Sdelphij hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3255268075Sdelphij hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3256268075Sdelphij exists = buf_hash_insert(hdr, &hash_lock); 3257268075Sdelphij } 3258268075Sdelphij if (exists != NULL) { 3259168404Spjd /* somebody beat us to the hash insert */ 3260168404Spjd mutex_exit(hash_lock); 3261219089Spjd buf_discard_identity(hdr); 3262168404Spjd (void) arc_buf_remove_ref(buf, private); 3263168404Spjd goto top; /* restart the IO request */ 3264168404Spjd } 3265168404Spjd /* if this is a prefetch, we don't have a reference */ 3266168404Spjd if (*arc_flags & ARC_PREFETCH) { 3267168404Spjd (void) remove_reference(hdr, hash_lock, 3268168404Spjd private); 3269168404Spjd hdr->b_flags |= ARC_PREFETCH; 3270168404Spjd } 3271185029Spjd if (*arc_flags & ARC_L2CACHE) 3272185029Spjd hdr->b_flags |= ARC_L2CACHE; 3273251478Sdelphij if (*arc_flags & ARC_L2COMPRESS) 3274251478Sdelphij hdr->b_flags |= ARC_L2COMPRESS; 3275168404Spjd if (BP_GET_LEVEL(bp) > 0) 3276168404Spjd hdr->b_flags |= ARC_INDIRECT; 3277168404Spjd } else { 3278168404Spjd /* this block is in the ghost cache */ 3279168404Spjd ASSERT(GHOST_STATE(hdr->b_state)); 3280168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3281240415Smm ASSERT0(refcount_count(&hdr->b_refcnt)); 3282168404Spjd ASSERT(hdr->b_buf == NULL); 3283168404Spjd 3284168404Spjd /* if this is a prefetch, we don't have a reference */ 3285168404Spjd if (*arc_flags & ARC_PREFETCH) 3286168404Spjd hdr->b_flags |= ARC_PREFETCH; 3287168404Spjd else 3288168404Spjd add_reference(hdr, hash_lock, private); 3289185029Spjd if (*arc_flags & ARC_L2CACHE) 3290185029Spjd hdr->b_flags |= ARC_L2CACHE; 3291251478Sdelphij if (*arc_flags & ARC_L2COMPRESS) 3292251478Sdelphij hdr->b_flags |= ARC_L2COMPRESS; 3293185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3294168404Spjd buf->b_hdr = hdr; 3295168404Spjd buf->b_data = NULL; 3296168404Spjd buf->b_efunc = NULL; 3297168404Spjd buf->b_private = NULL; 3298168404Spjd buf->b_next = NULL; 3299168404Spjd hdr->b_buf = buf; 3300168404Spjd ASSERT(hdr->b_datacnt == 0); 3301168404Spjd hdr->b_datacnt = 1; 3302219089Spjd arc_get_data_buf(buf); 3303219089Spjd arc_access(hdr, hash_lock); 3304168404Spjd } 3305168404Spjd 3306219089Spjd ASSERT(!GHOST_STATE(hdr->b_state)); 3307219089Spjd 3308168404Spjd acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3309168404Spjd acb->acb_done = done; 3310168404Spjd acb->acb_private = private; 3311168404Spjd 3312168404Spjd ASSERT(hdr->b_acb == NULL); 3313168404Spjd hdr->b_acb = acb; 3314168404Spjd hdr->b_flags |= ARC_IO_IN_PROGRESS; 3315168404Spjd 3316258389Savg if (hdr->b_l2hdr != NULL && 3317185029Spjd (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3318208373Smm devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3319185029Spjd addr = hdr->b_l2hdr->b_daddr; 3320258389Savg b_compress = hdr->b_l2hdr->b_compress; 3321258389Savg b_asize = hdr->b_l2hdr->b_asize; 3322185029Spjd /* 3323185029Spjd * Lock out device removal. 3324185029Spjd */ 3325185029Spjd if (vdev_is_dead(vd) || 3326185029Spjd !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3327185029Spjd vd = NULL; 3328185029Spjd } 3329185029Spjd 3330268075Sdelphij if (hash_lock != NULL) 3331268075Sdelphij mutex_exit(hash_lock); 3332168404Spjd 3333251629Sdelphij /* 3334251629Sdelphij * At this point, we have a level 1 cache miss. Try again in 3335251629Sdelphij * L2ARC if possible. 3336251629Sdelphij */ 3337168404Spjd ASSERT3U(hdr->b_size, ==, size); 3338219089Spjd DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3339219089Spjd uint64_t, size, zbookmark_t *, zb); 3340168404Spjd ARCSTAT_BUMP(arcstat_misses); 3341168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3342168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3343168404Spjd data, metadata, misses); 3344228392Spjd#ifdef _KERNEL 3345228392Spjd curthread->td_ru.ru_inblock++; 3346228392Spjd#endif 3347168404Spjd 3348208373Smm if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3349185029Spjd /* 3350185029Spjd * Read from the L2ARC if the following are true: 3351185029Spjd * 1. The L2ARC vdev was previously cached. 3352185029Spjd * 2. This buffer still has L2ARC metadata. 3353185029Spjd * 3. This buffer isn't currently writing to the L2ARC. 3354185029Spjd * 4. The L2ARC entry wasn't evicted, which may 3355185029Spjd * also have invalidated the vdev. 3356208373Smm * 5. This isn't prefetch and l2arc_noprefetch is set. 3357185029Spjd */ 3358185029Spjd if (hdr->b_l2hdr != NULL && 3359208373Smm !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3360208373Smm !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3361185029Spjd l2arc_read_callback_t *cb; 3362185029Spjd 3363185029Spjd DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3364185029Spjd ARCSTAT_BUMP(arcstat_l2_hits); 3365185029Spjd 3366185029Spjd cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3367185029Spjd KM_SLEEP); 3368185029Spjd cb->l2rcb_buf = buf; 3369185029Spjd cb->l2rcb_spa = spa; 3370185029Spjd cb->l2rcb_bp = *bp; 3371185029Spjd cb->l2rcb_zb = *zb; 3372185029Spjd cb->l2rcb_flags = zio_flags; 3373258389Savg cb->l2rcb_compress = b_compress; 3374185029Spjd 3375247187Smm ASSERT(addr >= VDEV_LABEL_START_SIZE && 3376247187Smm addr + size < vd->vdev_psize - 3377247187Smm VDEV_LABEL_END_SIZE); 3378247187Smm 3379185029Spjd /* 3380185029Spjd * l2arc read. The SCL_L2ARC lock will be 3381185029Spjd * released by l2arc_read_done(). 3382251478Sdelphij * Issue a null zio if the underlying buffer 3383251478Sdelphij * was squashed to zero size by compression. 3384185029Spjd */ 3385258389Savg if (b_compress == ZIO_COMPRESS_EMPTY) { 3386251478Sdelphij rzio = zio_null(pio, spa, vd, 3387251478Sdelphij l2arc_read_done, cb, 3388251478Sdelphij zio_flags | ZIO_FLAG_DONT_CACHE | 3389251478Sdelphij ZIO_FLAG_CANFAIL | 3390251478Sdelphij ZIO_FLAG_DONT_PROPAGATE | 3391251478Sdelphij ZIO_FLAG_DONT_RETRY); 3392251478Sdelphij } else { 3393251478Sdelphij rzio = zio_read_phys(pio, vd, addr, 3394258389Savg b_asize, buf->b_data, 3395258389Savg ZIO_CHECKSUM_OFF, 3396251478Sdelphij l2arc_read_done, cb, priority, 3397251478Sdelphij zio_flags | ZIO_FLAG_DONT_CACHE | 3398251478Sdelphij ZIO_FLAG_CANFAIL | 3399251478Sdelphij ZIO_FLAG_DONT_PROPAGATE | 3400251478Sdelphij ZIO_FLAG_DONT_RETRY, B_FALSE); 3401251478Sdelphij } 3402185029Spjd DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3403185029Spjd zio_t *, rzio); 3404258389Savg ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3405185029Spjd 3406185029Spjd if (*arc_flags & ARC_NOWAIT) { 3407185029Spjd zio_nowait(rzio); 3408185029Spjd return (0); 3409185029Spjd } 3410185029Spjd 3411185029Spjd ASSERT(*arc_flags & ARC_WAIT); 3412185029Spjd if (zio_wait(rzio) == 0) 3413185029Spjd return (0); 3414185029Spjd 3415185029Spjd /* l2arc read error; goto zio_read() */ 3416185029Spjd } else { 3417185029Spjd DTRACE_PROBE1(l2arc__miss, 3418185029Spjd arc_buf_hdr_t *, hdr); 3419185029Spjd ARCSTAT_BUMP(arcstat_l2_misses); 3420185029Spjd if (HDR_L2_WRITING(hdr)) 3421185029Spjd ARCSTAT_BUMP(arcstat_l2_rw_clash); 3422185029Spjd spa_config_exit(spa, SCL_L2ARC, vd); 3423185029Spjd } 3424208373Smm } else { 3425208373Smm if (vd != NULL) 3426208373Smm spa_config_exit(spa, SCL_L2ARC, vd); 3427208373Smm if (l2arc_ndev != 0) { 3428208373Smm DTRACE_PROBE1(l2arc__miss, 3429208373Smm arc_buf_hdr_t *, hdr); 3430208373Smm ARCSTAT_BUMP(arcstat_l2_misses); 3431208373Smm } 3432185029Spjd } 3433185029Spjd 3434168404Spjd rzio = zio_read(pio, spa, bp, buf->b_data, size, 3435185029Spjd arc_read_done, buf, priority, zio_flags, zb); 3436168404Spjd 3437168404Spjd if (*arc_flags & ARC_WAIT) 3438168404Spjd return (zio_wait(rzio)); 3439168404Spjd 3440168404Spjd ASSERT(*arc_flags & ARC_NOWAIT); 3441168404Spjd zio_nowait(rzio); 3442168404Spjd } 3443168404Spjd return (0); 3444168404Spjd} 3445168404Spjd 3446168404Spjdvoid 3447168404Spjdarc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3448168404Spjd{ 3449168404Spjd ASSERT(buf->b_hdr != NULL); 3450168404Spjd ASSERT(buf->b_hdr->b_state != arc_anon); 3451168404Spjd ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3452219089Spjd ASSERT(buf->b_efunc == NULL); 3453219089Spjd ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3454219089Spjd 3455168404Spjd buf->b_efunc = func; 3456168404Spjd buf->b_private = private; 3457168404Spjd} 3458168404Spjd 3459168404Spjd/* 3460251520Sdelphij * Notify the arc that a block was freed, and thus will never be used again. 3461251520Sdelphij */ 3462251520Sdelphijvoid 3463251520Sdelphijarc_freed(spa_t *spa, const blkptr_t *bp) 3464251520Sdelphij{ 3465251520Sdelphij arc_buf_hdr_t *hdr; 3466251520Sdelphij kmutex_t *hash_lock; 3467251520Sdelphij uint64_t guid = spa_load_guid(spa); 3468251520Sdelphij 3469268075Sdelphij ASSERT(!BP_IS_EMBEDDED(bp)); 3470268075Sdelphij 3471268075Sdelphij hdr = buf_hash_find(guid, bp, &hash_lock); 3472251520Sdelphij if (hdr == NULL) 3473251520Sdelphij return; 3474251520Sdelphij if (HDR_BUF_AVAILABLE(hdr)) { 3475251520Sdelphij arc_buf_t *buf = hdr->b_buf; 3476251520Sdelphij add_reference(hdr, hash_lock, FTAG); 3477251520Sdelphij hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3478251520Sdelphij mutex_exit(hash_lock); 3479251520Sdelphij 3480251520Sdelphij arc_release(buf, FTAG); 3481251520Sdelphij (void) arc_buf_remove_ref(buf, FTAG); 3482251520Sdelphij } else { 3483251520Sdelphij mutex_exit(hash_lock); 3484251520Sdelphij } 3485251520Sdelphij 3486251520Sdelphij} 3487251520Sdelphij 3488251520Sdelphij/* 3489168404Spjd * This is used by the DMU to let the ARC know that a buffer is 3490168404Spjd * being evicted, so the ARC should clean up. If this arc buf 3491168404Spjd * is not yet in the evicted state, it will be put there. 3492168404Spjd */ 3493168404Spjdint 3494168404Spjdarc_buf_evict(arc_buf_t *buf) 3495168404Spjd{ 3496168404Spjd arc_buf_hdr_t *hdr; 3497168404Spjd kmutex_t *hash_lock; 3498168404Spjd arc_buf_t **bufp; 3499205231Skmacy list_t *list, *evicted_list; 3500205231Skmacy kmutex_t *lock, *evicted_lock; 3501206796Spjd 3502219089Spjd mutex_enter(&buf->b_evict_lock); 3503168404Spjd hdr = buf->b_hdr; 3504168404Spjd if (hdr == NULL) { 3505168404Spjd /* 3506168404Spjd * We are in arc_do_user_evicts(). 3507168404Spjd */ 3508168404Spjd ASSERT(buf->b_data == NULL); 3509219089Spjd mutex_exit(&buf->b_evict_lock); 3510168404Spjd return (0); 3511185029Spjd } else if (buf->b_data == NULL) { 3512185029Spjd arc_buf_t copy = *buf; /* structure assignment */ 3513185029Spjd /* 3514185029Spjd * We are on the eviction list; process this buffer now 3515185029Spjd * but let arc_do_user_evicts() do the reaping. 3516185029Spjd */ 3517185029Spjd buf->b_efunc = NULL; 3518219089Spjd mutex_exit(&buf->b_evict_lock); 3519185029Spjd VERIFY(copy.b_efunc(©) == 0); 3520185029Spjd return (1); 3521168404Spjd } 3522168404Spjd hash_lock = HDR_LOCK(hdr); 3523168404Spjd mutex_enter(hash_lock); 3524219089Spjd hdr = buf->b_hdr; 3525219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3526168404Spjd 3527168404Spjd ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3528168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3529168404Spjd 3530168404Spjd /* 3531168404Spjd * Pull this buffer off of the hdr 3532168404Spjd */ 3533168404Spjd bufp = &hdr->b_buf; 3534168404Spjd while (*bufp != buf) 3535168404Spjd bufp = &(*bufp)->b_next; 3536168404Spjd *bufp = buf->b_next; 3537168404Spjd 3538168404Spjd ASSERT(buf->b_data != NULL); 3539168404Spjd arc_buf_destroy(buf, FALSE, FALSE); 3540168404Spjd 3541168404Spjd if (hdr->b_datacnt == 0) { 3542168404Spjd arc_state_t *old_state = hdr->b_state; 3543168404Spjd arc_state_t *evicted_state; 3544168404Spjd 3545219089Spjd ASSERT(hdr->b_buf == NULL); 3546168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 3547168404Spjd 3548168404Spjd evicted_state = 3549168404Spjd (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3550168404Spjd 3551205231Skmacy get_buf_info(hdr, old_state, &list, &lock); 3552205231Skmacy get_buf_info(hdr, evicted_state, &evicted_list, &evicted_lock); 3553205231Skmacy mutex_enter(lock); 3554205231Skmacy mutex_enter(evicted_lock); 3555168404Spjd 3556168404Spjd arc_change_state(evicted_state, hdr, hash_lock); 3557168404Spjd ASSERT(HDR_IN_HASH_TABLE(hdr)); 3558185029Spjd hdr->b_flags |= ARC_IN_HASH_TABLE; 3559185029Spjd hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3560168404Spjd 3561205231Skmacy mutex_exit(evicted_lock); 3562205231Skmacy mutex_exit(lock); 3563168404Spjd } 3564168404Spjd mutex_exit(hash_lock); 3565219089Spjd mutex_exit(&buf->b_evict_lock); 3566168404Spjd 3567168404Spjd VERIFY(buf->b_efunc(buf) == 0); 3568168404Spjd buf->b_efunc = NULL; 3569168404Spjd buf->b_private = NULL; 3570168404Spjd buf->b_hdr = NULL; 3571219089Spjd buf->b_next = NULL; 3572168404Spjd kmem_cache_free(buf_cache, buf); 3573168404Spjd return (1); 3574168404Spjd} 3575168404Spjd 3576168404Spjd/* 3577251629Sdelphij * Release this buffer from the cache, making it an anonymous buffer. This 3578251629Sdelphij * must be done after a read and prior to modifying the buffer contents. 3579168404Spjd * If the buffer has more than one reference, we must make 3580185029Spjd * a new hdr for the buffer. 3581168404Spjd */ 3582168404Spjdvoid 3583168404Spjdarc_release(arc_buf_t *buf, void *tag) 3584168404Spjd{ 3585185029Spjd arc_buf_hdr_t *hdr; 3586219089Spjd kmutex_t *hash_lock = NULL; 3587185029Spjd l2arc_buf_hdr_t *l2hdr; 3588185029Spjd uint64_t buf_size; 3589168404Spjd 3590219089Spjd /* 3591219089Spjd * It would be nice to assert that if it's DMU metadata (level > 3592219089Spjd * 0 || it's the dnode file), then it must be syncing context. 3593219089Spjd * But we don't know that information at this level. 3594219089Spjd */ 3595219089Spjd 3596219089Spjd mutex_enter(&buf->b_evict_lock); 3597185029Spjd hdr = buf->b_hdr; 3598185029Spjd 3599168404Spjd /* this buffer is not on any list */ 3600168404Spjd ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3601168404Spjd 3602168404Spjd if (hdr->b_state == arc_anon) { 3603168404Spjd /* this buffer is already released */ 3604168404Spjd ASSERT(buf->b_efunc == NULL); 3605208373Smm } else { 3606208373Smm hash_lock = HDR_LOCK(hdr); 3607208373Smm mutex_enter(hash_lock); 3608219089Spjd hdr = buf->b_hdr; 3609219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3610168404Spjd } 3611168404Spjd 3612185029Spjd l2hdr = hdr->b_l2hdr; 3613185029Spjd if (l2hdr) { 3614185029Spjd mutex_enter(&l2arc_buflist_mtx); 3615185029Spjd hdr->b_l2hdr = NULL; 3616258388Savg list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3617185029Spjd } 3618247187Smm buf_size = hdr->b_size; 3619185029Spjd 3620168404Spjd /* 3621168404Spjd * Do we have more than one buf? 3622168404Spjd */ 3623185029Spjd if (hdr->b_datacnt > 1) { 3624168404Spjd arc_buf_hdr_t *nhdr; 3625168404Spjd arc_buf_t **bufp; 3626168404Spjd uint64_t blksz = hdr->b_size; 3627209962Smm uint64_t spa = hdr->b_spa; 3628168404Spjd arc_buf_contents_t type = hdr->b_type; 3629185029Spjd uint32_t flags = hdr->b_flags; 3630168404Spjd 3631185029Spjd ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3632168404Spjd /* 3633219089Spjd * Pull the data off of this hdr and attach it to 3634219089Spjd * a new anonymous hdr. 3635168404Spjd */ 3636168404Spjd (void) remove_reference(hdr, hash_lock, tag); 3637168404Spjd bufp = &hdr->b_buf; 3638168404Spjd while (*bufp != buf) 3639168404Spjd bufp = &(*bufp)->b_next; 3640219089Spjd *bufp = buf->b_next; 3641168404Spjd buf->b_next = NULL; 3642168404Spjd 3643168404Spjd ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3644168404Spjd atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3645168404Spjd if (refcount_is_zero(&hdr->b_refcnt)) { 3646185029Spjd uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3647185029Spjd ASSERT3U(*size, >=, hdr->b_size); 3648185029Spjd atomic_add_64(size, -hdr->b_size); 3649168404Spjd } 3650242845Sdelphij 3651242845Sdelphij /* 3652242845Sdelphij * We're releasing a duplicate user data buffer, update 3653242845Sdelphij * our statistics accordingly. 3654242845Sdelphij */ 3655242845Sdelphij if (hdr->b_type == ARC_BUFC_DATA) { 3656242845Sdelphij ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3657242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3658242845Sdelphij -hdr->b_size); 3659242845Sdelphij } 3660168404Spjd hdr->b_datacnt -= 1; 3661168404Spjd arc_cksum_verify(buf); 3662240133Smm#ifdef illumos 3663240133Smm arc_buf_unwatch(buf); 3664240133Smm#endif /* illumos */ 3665168404Spjd 3666168404Spjd mutex_exit(hash_lock); 3667168404Spjd 3668185029Spjd nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3669168404Spjd nhdr->b_size = blksz; 3670168404Spjd nhdr->b_spa = spa; 3671168404Spjd nhdr->b_type = type; 3672168404Spjd nhdr->b_buf = buf; 3673168404Spjd nhdr->b_state = arc_anon; 3674168404Spjd nhdr->b_arc_access = 0; 3675185029Spjd nhdr->b_flags = flags & ARC_L2_WRITING; 3676185029Spjd nhdr->b_l2hdr = NULL; 3677168404Spjd nhdr->b_datacnt = 1; 3678168404Spjd nhdr->b_freeze_cksum = NULL; 3679168404Spjd (void) refcount_add(&nhdr->b_refcnt, tag); 3680168404Spjd buf->b_hdr = nhdr; 3681219089Spjd mutex_exit(&buf->b_evict_lock); 3682168404Spjd atomic_add_64(&arc_anon->arcs_size, blksz); 3683168404Spjd } else { 3684219089Spjd mutex_exit(&buf->b_evict_lock); 3685168404Spjd ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3686168404Spjd ASSERT(!list_link_active(&hdr->b_arc_node)); 3687168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3688219089Spjd if (hdr->b_state != arc_anon) 3689219089Spjd arc_change_state(arc_anon, hdr, hash_lock); 3690168404Spjd hdr->b_arc_access = 0; 3691219089Spjd if (hash_lock) 3692219089Spjd mutex_exit(hash_lock); 3693185029Spjd 3694219089Spjd buf_discard_identity(hdr); 3695168404Spjd arc_buf_thaw(buf); 3696168404Spjd } 3697168404Spjd buf->b_efunc = NULL; 3698168404Spjd buf->b_private = NULL; 3699185029Spjd 3700185029Spjd if (l2hdr) { 3701251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3702248572Ssmh trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3703248574Ssmh hdr->b_size, 0); 3704185029Spjd kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3705185029Spjd ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3706185029Spjd mutex_exit(&l2arc_buflist_mtx); 3707185029Spjd } 3708168404Spjd} 3709168404Spjd 3710168404Spjdint 3711168404Spjdarc_released(arc_buf_t *buf) 3712168404Spjd{ 3713185029Spjd int released; 3714185029Spjd 3715219089Spjd mutex_enter(&buf->b_evict_lock); 3716185029Spjd released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3717219089Spjd mutex_exit(&buf->b_evict_lock); 3718185029Spjd return (released); 3719168404Spjd} 3720168404Spjd 3721168404Spjdint 3722168404Spjdarc_has_callback(arc_buf_t *buf) 3723168404Spjd{ 3724185029Spjd int callback; 3725185029Spjd 3726219089Spjd mutex_enter(&buf->b_evict_lock); 3727185029Spjd callback = (buf->b_efunc != NULL); 3728219089Spjd mutex_exit(&buf->b_evict_lock); 3729185029Spjd return (callback); 3730168404Spjd} 3731168404Spjd 3732168404Spjd#ifdef ZFS_DEBUG 3733168404Spjdint 3734168404Spjdarc_referenced(arc_buf_t *buf) 3735168404Spjd{ 3736185029Spjd int referenced; 3737185029Spjd 3738219089Spjd mutex_enter(&buf->b_evict_lock); 3739185029Spjd referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3740219089Spjd mutex_exit(&buf->b_evict_lock); 3741185029Spjd return (referenced); 3742168404Spjd} 3743168404Spjd#endif 3744168404Spjd 3745168404Spjdstatic void 3746168404Spjdarc_write_ready(zio_t *zio) 3747168404Spjd{ 3748168404Spjd arc_write_callback_t *callback = zio->io_private; 3749168404Spjd arc_buf_t *buf = callback->awcb_buf; 3750185029Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3751168404Spjd 3752185029Spjd ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3753185029Spjd callback->awcb_ready(zio, buf, callback->awcb_private); 3754185029Spjd 3755185029Spjd /* 3756185029Spjd * If the IO is already in progress, then this is a re-write 3757185029Spjd * attempt, so we need to thaw and re-compute the cksum. 3758185029Spjd * It is the responsibility of the callback to handle the 3759185029Spjd * accounting for any re-write attempt. 3760185029Spjd */ 3761185029Spjd if (HDR_IO_IN_PROGRESS(hdr)) { 3762185029Spjd mutex_enter(&hdr->b_freeze_lock); 3763185029Spjd if (hdr->b_freeze_cksum != NULL) { 3764185029Spjd kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3765185029Spjd hdr->b_freeze_cksum = NULL; 3766185029Spjd } 3767185029Spjd mutex_exit(&hdr->b_freeze_lock); 3768168404Spjd } 3769185029Spjd arc_cksum_compute(buf, B_FALSE); 3770185029Spjd hdr->b_flags |= ARC_IO_IN_PROGRESS; 3771168404Spjd} 3772168404Spjd 3773258632Savg/* 3774258632Savg * The SPA calls this callback for each physical write that happens on behalf 3775258632Savg * of a logical write. See the comment in dbuf_write_physdone() for details. 3776258632Savg */ 3777168404Spjdstatic void 3778258632Savgarc_write_physdone(zio_t *zio) 3779258632Savg{ 3780258632Savg arc_write_callback_t *cb = zio->io_private; 3781258632Savg if (cb->awcb_physdone != NULL) 3782258632Savg cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3783258632Savg} 3784258632Savg 3785258632Savgstatic void 3786168404Spjdarc_write_done(zio_t *zio) 3787168404Spjd{ 3788168404Spjd arc_write_callback_t *callback = zio->io_private; 3789168404Spjd arc_buf_t *buf = callback->awcb_buf; 3790168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3791168404Spjd 3792219089Spjd ASSERT(hdr->b_acb == NULL); 3793168404Spjd 3794219089Spjd if (zio->io_error == 0) { 3795268075Sdelphij if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 3796260150Sdelphij buf_discard_identity(hdr); 3797260150Sdelphij } else { 3798260150Sdelphij hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3799260150Sdelphij hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3800260150Sdelphij hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3801260150Sdelphij } 3802219089Spjd } else { 3803219089Spjd ASSERT(BUF_EMPTY(hdr)); 3804219089Spjd } 3805219089Spjd 3806168404Spjd /* 3807268075Sdelphij * If the block to be written was all-zero or compressed enough to be 3808268075Sdelphij * embedded in the BP, no write was performed so there will be no 3809268075Sdelphij * dva/birth/checksum. The buffer must therefore remain anonymous 3810268075Sdelphij * (and uncached). 3811168404Spjd */ 3812168404Spjd if (!BUF_EMPTY(hdr)) { 3813168404Spjd arc_buf_hdr_t *exists; 3814168404Spjd kmutex_t *hash_lock; 3815168404Spjd 3816219089Spjd ASSERT(zio->io_error == 0); 3817219089Spjd 3818168404Spjd arc_cksum_verify(buf); 3819168404Spjd 3820168404Spjd exists = buf_hash_insert(hdr, &hash_lock); 3821168404Spjd if (exists) { 3822168404Spjd /* 3823168404Spjd * This can only happen if we overwrite for 3824168404Spjd * sync-to-convergence, because we remove 3825168404Spjd * buffers from the hash table when we arc_free(). 3826168404Spjd */ 3827219089Spjd if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3828219089Spjd if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3829219089Spjd panic("bad overwrite, hdr=%p exists=%p", 3830219089Spjd (void *)hdr, (void *)exists); 3831219089Spjd ASSERT(refcount_is_zero(&exists->b_refcnt)); 3832219089Spjd arc_change_state(arc_anon, exists, hash_lock); 3833219089Spjd mutex_exit(hash_lock); 3834219089Spjd arc_hdr_destroy(exists); 3835219089Spjd exists = buf_hash_insert(hdr, &hash_lock); 3836219089Spjd ASSERT3P(exists, ==, NULL); 3837243524Smm } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3838243524Smm /* nopwrite */ 3839243524Smm ASSERT(zio->io_prop.zp_nopwrite); 3840243524Smm if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3841243524Smm panic("bad nopwrite, hdr=%p exists=%p", 3842243524Smm (void *)hdr, (void *)exists); 3843219089Spjd } else { 3844219089Spjd /* Dedup */ 3845219089Spjd ASSERT(hdr->b_datacnt == 1); 3846219089Spjd ASSERT(hdr->b_state == arc_anon); 3847219089Spjd ASSERT(BP_GET_DEDUP(zio->io_bp)); 3848219089Spjd ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3849219089Spjd } 3850168404Spjd } 3851168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3852185029Spjd /* if it's not anon, we are doing a scrub */ 3853219089Spjd if (!exists && hdr->b_state == arc_anon) 3854185029Spjd arc_access(hdr, hash_lock); 3855168404Spjd mutex_exit(hash_lock); 3856168404Spjd } else { 3857168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3858168404Spjd } 3859168404Spjd 3860219089Spjd ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3861219089Spjd callback->awcb_done(zio, buf, callback->awcb_private); 3862168404Spjd 3863168404Spjd kmem_free(callback, sizeof (arc_write_callback_t)); 3864168404Spjd} 3865168404Spjd 3866168404Spjdzio_t * 3867219089Spjdarc_write(zio_t *pio, spa_t *spa, uint64_t txg, 3868251478Sdelphij blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 3869258632Savg const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 3870258632Savg arc_done_func_t *done, void *private, zio_priority_t priority, 3871258632Savg int zio_flags, const zbookmark_t *zb) 3872168404Spjd{ 3873168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3874168404Spjd arc_write_callback_t *callback; 3875185029Spjd zio_t *zio; 3876168404Spjd 3877185029Spjd ASSERT(ready != NULL); 3878219089Spjd ASSERT(done != NULL); 3879168404Spjd ASSERT(!HDR_IO_ERROR(hdr)); 3880168404Spjd ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3881219089Spjd ASSERT(hdr->b_acb == NULL); 3882185029Spjd if (l2arc) 3883185029Spjd hdr->b_flags |= ARC_L2CACHE; 3884251478Sdelphij if (l2arc_compress) 3885251478Sdelphij hdr->b_flags |= ARC_L2COMPRESS; 3886168404Spjd callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 3887168404Spjd callback->awcb_ready = ready; 3888258632Savg callback->awcb_physdone = physdone; 3889168404Spjd callback->awcb_done = done; 3890168404Spjd callback->awcb_private = private; 3891168404Spjd callback->awcb_buf = buf; 3892168404Spjd 3893219089Spjd zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 3894258632Savg arc_write_ready, arc_write_physdone, arc_write_done, callback, 3895258632Savg priority, zio_flags, zb); 3896185029Spjd 3897168404Spjd return (zio); 3898168404Spjd} 3899168404Spjd 3900185029Spjdstatic int 3901258632Savgarc_memory_throttle(uint64_t reserve, uint64_t txg) 3902185029Spjd{ 3903185029Spjd#ifdef _KERNEL 3904219089Spjd uint64_t available_memory = 3905263620Sbdrewery ptoa((uintmax_t)vm_cnt.v_free_count + vm_cnt.v_cache_count); 3906185029Spjd static uint64_t page_load = 0; 3907185029Spjd static uint64_t last_txg = 0; 3908185029Spjd 3909219089Spjd#ifdef sun 3910185029Spjd#if defined(__i386) 3911185029Spjd available_memory = 3912185029Spjd MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 3913185029Spjd#endif 3914219089Spjd#endif /* sun */ 3915258632Savg 3916263620Sbdrewery if (vm_cnt.v_free_count + vm_cnt.v_cache_count > 3917258632Savg (uint64_t)physmem * arc_lotsfree_percent / 100) 3918185029Spjd return (0); 3919185029Spjd 3920185029Spjd if (txg > last_txg) { 3921185029Spjd last_txg = txg; 3922185029Spjd page_load = 0; 3923185029Spjd } 3924185029Spjd /* 3925185029Spjd * If we are in pageout, we know that memory is already tight, 3926185029Spjd * the arc is already going to be evicting, so we just want to 3927185029Spjd * continue to let page writes occur as quickly as possible. 3928185029Spjd */ 3929185029Spjd if (curproc == pageproc) { 3930185029Spjd if (page_load > available_memory / 4) 3931249195Smm return (SET_ERROR(ERESTART)); 3932185029Spjd /* Note: reserve is inflated, so we deflate */ 3933185029Spjd page_load += reserve / 8; 3934185029Spjd return (0); 3935185029Spjd } else if (page_load > 0 && arc_reclaim_needed()) { 3936185029Spjd /* memory is low, delay before restarting */ 3937185029Spjd ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3938249195Smm return (SET_ERROR(EAGAIN)); 3939185029Spjd } 3940185029Spjd page_load = 0; 3941185029Spjd#endif 3942185029Spjd return (0); 3943185029Spjd} 3944185029Spjd 3945168404Spjdvoid 3946185029Spjdarc_tempreserve_clear(uint64_t reserve) 3947168404Spjd{ 3948185029Spjd atomic_add_64(&arc_tempreserve, -reserve); 3949168404Spjd ASSERT((int64_t)arc_tempreserve >= 0); 3950168404Spjd} 3951168404Spjd 3952168404Spjdint 3953185029Spjdarc_tempreserve_space(uint64_t reserve, uint64_t txg) 3954168404Spjd{ 3955185029Spjd int error; 3956209962Smm uint64_t anon_size; 3957185029Spjd 3958185029Spjd if (reserve > arc_c/4 && !arc_no_grow) 3959185029Spjd arc_c = MIN(arc_c_max, reserve * 4); 3960185029Spjd if (reserve > arc_c) 3961249195Smm return (SET_ERROR(ENOMEM)); 3962168404Spjd 3963168404Spjd /* 3964209962Smm * Don't count loaned bufs as in flight dirty data to prevent long 3965209962Smm * network delays from blocking transactions that are ready to be 3966209962Smm * assigned to a txg. 3967209962Smm */ 3968209962Smm anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 3969209962Smm 3970209962Smm /* 3971185029Spjd * Writes will, almost always, require additional memory allocations 3972251631Sdelphij * in order to compress/encrypt/etc the data. We therefore need to 3973185029Spjd * make sure that there is sufficient available memory for this. 3974185029Spjd */ 3975258632Savg error = arc_memory_throttle(reserve, txg); 3976258632Savg if (error != 0) 3977185029Spjd return (error); 3978185029Spjd 3979185029Spjd /* 3980168404Spjd * Throttle writes when the amount of dirty data in the cache 3981168404Spjd * gets too large. We try to keep the cache less than half full 3982168404Spjd * of dirty blocks so that our sync times don't grow too large. 3983168404Spjd * Note: if two requests come in concurrently, we might let them 3984168404Spjd * both succeed, when one of them should fail. Not a huge deal. 3985168404Spjd */ 3986209962Smm 3987209962Smm if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 3988209962Smm anon_size > arc_c / 4) { 3989185029Spjd dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 3990185029Spjd "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 3991185029Spjd arc_tempreserve>>10, 3992185029Spjd arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 3993185029Spjd arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 3994185029Spjd reserve>>10, arc_c>>10); 3995249195Smm return (SET_ERROR(ERESTART)); 3996168404Spjd } 3997185029Spjd atomic_add_64(&arc_tempreserve, reserve); 3998168404Spjd return (0); 3999168404Spjd} 4000168404Spjd 4001168582Spjdstatic kmutex_t arc_lowmem_lock; 4002168404Spjd#ifdef _KERNEL 4003168566Spjdstatic eventhandler_tag arc_event_lowmem = NULL; 4004168404Spjd 4005168404Spjdstatic void 4006168566Spjdarc_lowmem(void *arg __unused, int howto __unused) 4007168404Spjd{ 4008168404Spjd 4009168566Spjd /* Serialize access via arc_lowmem_lock. */ 4010168566Spjd mutex_enter(&arc_lowmem_lock); 4011219089Spjd mutex_enter(&arc_reclaim_thr_lock); 4012185029Spjd needfree = 1; 4013168404Spjd cv_signal(&arc_reclaim_thr_cv); 4014241773Savg 4015241773Savg /* 4016241773Savg * It is unsafe to block here in arbitrary threads, because we can come 4017241773Savg * here from ARC itself and may hold ARC locks and thus risk a deadlock 4018241773Savg * with ARC reclaim thread. 4019241773Savg */ 4020241773Savg if (curproc == pageproc) { 4021241773Savg while (needfree) 4022241773Savg msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4023241773Savg } 4024219089Spjd mutex_exit(&arc_reclaim_thr_lock); 4025168566Spjd mutex_exit(&arc_lowmem_lock); 4026168404Spjd} 4027168404Spjd#endif 4028168404Spjd 4029168404Spjdvoid 4030168404Spjdarc_init(void) 4031168404Spjd{ 4032219089Spjd int i, prefetch_tunable_set = 0; 4033205231Skmacy 4034168404Spjd mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4035168404Spjd cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4036168566Spjd mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4037168404Spjd 4038168404Spjd /* Convert seconds to clock ticks */ 4039168404Spjd arc_min_prefetch_lifespan = 1 * hz; 4040168404Spjd 4041168404Spjd /* Start out with 1/8 of all memory */ 4042168566Spjd arc_c = kmem_size() / 8; 4043219089Spjd 4044219089Spjd#ifdef sun 4045192360Skmacy#ifdef _KERNEL 4046192360Skmacy /* 4047192360Skmacy * On architectures where the physical memory can be larger 4048192360Skmacy * than the addressable space (intel in 32-bit mode), we may 4049192360Skmacy * need to limit the cache to 1/8 of VM size. 4050192360Skmacy */ 4051192360Skmacy arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4052192360Skmacy#endif 4053219089Spjd#endif /* sun */ 4054168566Spjd /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4055168566Spjd arc_c_min = MAX(arc_c / 4, 64<<18); 4056168566Spjd /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4057168404Spjd if (arc_c * 8 >= 1<<30) 4058168404Spjd arc_c_max = (arc_c * 8) - (1<<30); 4059168404Spjd else 4060168404Spjd arc_c_max = arc_c_min; 4061175633Spjd arc_c_max = MAX(arc_c * 5, arc_c_max); 4062219089Spjd 4063168481Spjd#ifdef _KERNEL 4064168404Spjd /* 4065168404Spjd * Allow the tunables to override our calculations if they are 4066168566Spjd * reasonable (ie. over 16MB) 4067168404Spjd */ 4068219089Spjd if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size()) 4069168404Spjd arc_c_max = zfs_arc_max; 4070219089Spjd if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max) 4071168404Spjd arc_c_min = zfs_arc_min; 4072168481Spjd#endif 4073219089Spjd 4074168404Spjd arc_c = arc_c_max; 4075168404Spjd arc_p = (arc_c >> 1); 4076168404Spjd 4077185029Spjd /* limit meta-data to 1/4 of the arc capacity */ 4078185029Spjd arc_meta_limit = arc_c_max / 4; 4079185029Spjd 4080185029Spjd /* Allow the tunable to override if it is reasonable */ 4081185029Spjd if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4082185029Spjd arc_meta_limit = zfs_arc_meta_limit; 4083185029Spjd 4084185029Spjd if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4085185029Spjd arc_c_min = arc_meta_limit / 2; 4086185029Spjd 4087208373Smm if (zfs_arc_grow_retry > 0) 4088208373Smm arc_grow_retry = zfs_arc_grow_retry; 4089208373Smm 4090208373Smm if (zfs_arc_shrink_shift > 0) 4091208373Smm arc_shrink_shift = zfs_arc_shrink_shift; 4092208373Smm 4093208373Smm if (zfs_arc_p_min_shift > 0) 4094208373Smm arc_p_min_shift = zfs_arc_p_min_shift; 4095208373Smm 4096168404Spjd /* if kmem_flags are set, lets try to use less memory */ 4097168404Spjd if (kmem_debugging()) 4098168404Spjd arc_c = arc_c / 2; 4099168404Spjd if (arc_c < arc_c_min) 4100168404Spjd arc_c = arc_c_min; 4101168404Spjd 4102168473Spjd zfs_arc_min = arc_c_min; 4103168473Spjd zfs_arc_max = arc_c_max; 4104168473Spjd 4105168404Spjd arc_anon = &ARC_anon; 4106168404Spjd arc_mru = &ARC_mru; 4107168404Spjd arc_mru_ghost = &ARC_mru_ghost; 4108168404Spjd arc_mfu = &ARC_mfu; 4109168404Spjd arc_mfu_ghost = &ARC_mfu_ghost; 4110185029Spjd arc_l2c_only = &ARC_l2c_only; 4111168404Spjd arc_size = 0; 4112168404Spjd 4113205231Skmacy for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4114205231Skmacy mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4115205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4116205231Skmacy mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4117205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4118205231Skmacy mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4119205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4120205231Skmacy mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4121205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4122205231Skmacy mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4123205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4124205231Skmacy mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4125205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4126206796Spjd 4127205231Skmacy list_create(&arc_mru->arcs_lists[i], 4128205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4129205231Skmacy list_create(&arc_mru_ghost->arcs_lists[i], 4130205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4131205231Skmacy list_create(&arc_mfu->arcs_lists[i], 4132205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4133205231Skmacy list_create(&arc_mfu_ghost->arcs_lists[i], 4134205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4135205231Skmacy list_create(&arc_mfu_ghost->arcs_lists[i], 4136205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4137205231Skmacy list_create(&arc_l2c_only->arcs_lists[i], 4138205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4139205231Skmacy } 4140168404Spjd 4141168404Spjd buf_init(); 4142168404Spjd 4143168404Spjd arc_thread_exit = 0; 4144168404Spjd arc_eviction_list = NULL; 4145168404Spjd mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4146168404Spjd bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4147168404Spjd 4148168404Spjd arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4149168404Spjd sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4150168404Spjd 4151168404Spjd if (arc_ksp != NULL) { 4152168404Spjd arc_ksp->ks_data = &arc_stats; 4153168404Spjd kstat_install(arc_ksp); 4154168404Spjd } 4155168404Spjd 4156168404Spjd (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4157168404Spjd TS_RUN, minclsyspri); 4158168404Spjd 4159168404Spjd#ifdef _KERNEL 4160168566Spjd arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4161168404Spjd EVENTHANDLER_PRI_FIRST); 4162168404Spjd#endif 4163168404Spjd 4164168404Spjd arc_dead = FALSE; 4165185029Spjd arc_warm = B_FALSE; 4166168566Spjd 4167258632Savg /* 4168258632Savg * Calculate maximum amount of dirty data per pool. 4169258632Savg * 4170258632Savg * If it has been set by /etc/system, take that. 4171258632Savg * Otherwise, use a percentage of physical memory defined by 4172258632Savg * zfs_dirty_data_max_percent (default 10%) with a cap at 4173258632Savg * zfs_dirty_data_max_max (default 4GB). 4174258632Savg */ 4175258632Savg if (zfs_dirty_data_max == 0) { 4176258632Savg zfs_dirty_data_max = ptob(physmem) * 4177258632Savg zfs_dirty_data_max_percent / 100; 4178258632Savg zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4179258632Savg zfs_dirty_data_max_max); 4180258632Savg } 4181185029Spjd 4182168566Spjd#ifdef _KERNEL 4183194043Skmacy if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4184193953Skmacy prefetch_tunable_set = 1; 4185206796Spjd 4186193878Skmacy#ifdef __i386__ 4187193953Skmacy if (prefetch_tunable_set == 0) { 4188196863Strasz printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4189196863Strasz "-- to enable,\n"); 4190196863Strasz printf(" add \"vfs.zfs.prefetch_disable=0\" " 4191196863Strasz "to /boot/loader.conf.\n"); 4192219089Spjd zfs_prefetch_disable = 1; 4193193878Skmacy } 4194206796Spjd#else 4195193878Skmacy if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4196193953Skmacy prefetch_tunable_set == 0) { 4197196863Strasz printf("ZFS NOTICE: Prefetch is disabled by default if less " 4198196941Strasz "than 4GB of RAM is present;\n" 4199196863Strasz " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4200196863Strasz "to /boot/loader.conf.\n"); 4201219089Spjd zfs_prefetch_disable = 1; 4202193878Skmacy } 4203206796Spjd#endif 4204175633Spjd /* Warn about ZFS memory and address space requirements. */ 4205168696Spjd if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4206168987Sbmah printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4207168987Sbmah "expect unstable behavior.\n"); 4208175633Spjd } 4209175633Spjd if (kmem_size() < 512 * (1 << 20)) { 4210173419Spjd printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4211168987Sbmah "expect unstable behavior.\n"); 4212185029Spjd printf(" Consider tuning vm.kmem_size and " 4213173419Spjd "vm.kmem_size_max\n"); 4214185029Spjd printf(" in /boot/loader.conf.\n"); 4215168566Spjd } 4216168566Spjd#endif 4217168404Spjd} 4218168404Spjd 4219168404Spjdvoid 4220168404Spjdarc_fini(void) 4221168404Spjd{ 4222205231Skmacy int i; 4223206796Spjd 4224168404Spjd mutex_enter(&arc_reclaim_thr_lock); 4225168404Spjd arc_thread_exit = 1; 4226168404Spjd cv_signal(&arc_reclaim_thr_cv); 4227168404Spjd while (arc_thread_exit != 0) 4228168404Spjd cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4229168404Spjd mutex_exit(&arc_reclaim_thr_lock); 4230168404Spjd 4231185029Spjd arc_flush(NULL); 4232168404Spjd 4233168404Spjd arc_dead = TRUE; 4234168404Spjd 4235168404Spjd if (arc_ksp != NULL) { 4236168404Spjd kstat_delete(arc_ksp); 4237168404Spjd arc_ksp = NULL; 4238168404Spjd } 4239168404Spjd 4240168404Spjd mutex_destroy(&arc_eviction_mtx); 4241168404Spjd mutex_destroy(&arc_reclaim_thr_lock); 4242168404Spjd cv_destroy(&arc_reclaim_thr_cv); 4243168404Spjd 4244205231Skmacy for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4245205231Skmacy list_destroy(&arc_mru->arcs_lists[i]); 4246205231Skmacy list_destroy(&arc_mru_ghost->arcs_lists[i]); 4247205231Skmacy list_destroy(&arc_mfu->arcs_lists[i]); 4248205231Skmacy list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4249206795Spjd list_destroy(&arc_l2c_only->arcs_lists[i]); 4250168404Spjd 4251205231Skmacy mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4252205231Skmacy mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4253205231Skmacy mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4254205231Skmacy mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4255205231Skmacy mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4256206795Spjd mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4257205231Skmacy } 4258206796Spjd 4259168404Spjd buf_fini(); 4260168404Spjd 4261209962Smm ASSERT(arc_loaned_bytes == 0); 4262209962Smm 4263168582Spjd mutex_destroy(&arc_lowmem_lock); 4264168404Spjd#ifdef _KERNEL 4265168566Spjd if (arc_event_lowmem != NULL) 4266168566Spjd EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4267168404Spjd#endif 4268168404Spjd} 4269185029Spjd 4270185029Spjd/* 4271185029Spjd * Level 2 ARC 4272185029Spjd * 4273185029Spjd * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4274185029Spjd * It uses dedicated storage devices to hold cached data, which are populated 4275185029Spjd * using large infrequent writes. The main role of this cache is to boost 4276185029Spjd * the performance of random read workloads. The intended L2ARC devices 4277185029Spjd * include short-stroked disks, solid state disks, and other media with 4278185029Spjd * substantially faster read latency than disk. 4279185029Spjd * 4280185029Spjd * +-----------------------+ 4281185029Spjd * | ARC | 4282185029Spjd * +-----------------------+ 4283185029Spjd * | ^ ^ 4284185029Spjd * | | | 4285185029Spjd * l2arc_feed_thread() arc_read() 4286185029Spjd * | | | 4287185029Spjd * | l2arc read | 4288185029Spjd * V | | 4289185029Spjd * +---------------+ | 4290185029Spjd * | L2ARC | | 4291185029Spjd * +---------------+ | 4292185029Spjd * | ^ | 4293185029Spjd * l2arc_write() | | 4294185029Spjd * | | | 4295185029Spjd * V | | 4296185029Spjd * +-------+ +-------+ 4297185029Spjd * | vdev | | vdev | 4298185029Spjd * | cache | | cache | 4299185029Spjd * +-------+ +-------+ 4300185029Spjd * +=========+ .-----. 4301185029Spjd * : L2ARC : |-_____-| 4302185029Spjd * : devices : | Disks | 4303185029Spjd * +=========+ `-_____-' 4304185029Spjd * 4305185029Spjd * Read requests are satisfied from the following sources, in order: 4306185029Spjd * 4307185029Spjd * 1) ARC 4308185029Spjd * 2) vdev cache of L2ARC devices 4309185029Spjd * 3) L2ARC devices 4310185029Spjd * 4) vdev cache of disks 4311185029Spjd * 5) disks 4312185029Spjd * 4313185029Spjd * Some L2ARC device types exhibit extremely slow write performance. 4314185029Spjd * To accommodate for this there are some significant differences between 4315185029Spjd * the L2ARC and traditional cache design: 4316185029Spjd * 4317185029Spjd * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4318185029Spjd * the ARC behave as usual, freeing buffers and placing headers on ghost 4319185029Spjd * lists. The ARC does not send buffers to the L2ARC during eviction as 4320185029Spjd * this would add inflated write latencies for all ARC memory pressure. 4321185029Spjd * 4322185029Spjd * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4323185029Spjd * It does this by periodically scanning buffers from the eviction-end of 4324185029Spjd * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4325251478Sdelphij * not already there. It scans until a headroom of buffers is satisfied, 4326251478Sdelphij * which itself is a buffer for ARC eviction. If a compressible buffer is 4327251478Sdelphij * found during scanning and selected for writing to an L2ARC device, we 4328251478Sdelphij * temporarily boost scanning headroom during the next scan cycle to make 4329251478Sdelphij * sure we adapt to compression effects (which might significantly reduce 4330251478Sdelphij * the data volume we write to L2ARC). The thread that does this is 4331185029Spjd * l2arc_feed_thread(), illustrated below; example sizes are included to 4332185029Spjd * provide a better sense of ratio than this diagram: 4333185029Spjd * 4334185029Spjd * head --> tail 4335185029Spjd * +---------------------+----------+ 4336185029Spjd * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4337185029Spjd * +---------------------+----------+ | o L2ARC eligible 4338185029Spjd * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4339185029Spjd * +---------------------+----------+ | 4340185029Spjd * 15.9 Gbytes ^ 32 Mbytes | 4341185029Spjd * headroom | 4342185029Spjd * l2arc_feed_thread() 4343185029Spjd * | 4344185029Spjd * l2arc write hand <--[oooo]--' 4345185029Spjd * | 8 Mbyte 4346185029Spjd * | write max 4347185029Spjd * V 4348185029Spjd * +==============================+ 4349185029Spjd * L2ARC dev |####|#|###|###| |####| ... | 4350185029Spjd * +==============================+ 4351185029Spjd * 32 Gbytes 4352185029Spjd * 4353185029Spjd * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4354185029Spjd * evicted, then the L2ARC has cached a buffer much sooner than it probably 4355185029Spjd * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4356185029Spjd * safe to say that this is an uncommon case, since buffers at the end of 4357185029Spjd * the ARC lists have moved there due to inactivity. 4358185029Spjd * 4359185029Spjd * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4360185029Spjd * then the L2ARC simply misses copying some buffers. This serves as a 4361185029Spjd * pressure valve to prevent heavy read workloads from both stalling the ARC 4362185029Spjd * with waits and clogging the L2ARC with writes. This also helps prevent 4363185029Spjd * the potential for the L2ARC to churn if it attempts to cache content too 4364185029Spjd * quickly, such as during backups of the entire pool. 4365185029Spjd * 4366185029Spjd * 5. After system boot and before the ARC has filled main memory, there are 4367185029Spjd * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4368185029Spjd * lists can remain mostly static. Instead of searching from tail of these 4369185029Spjd * lists as pictured, the l2arc_feed_thread() will search from the list heads 4370185029Spjd * for eligible buffers, greatly increasing its chance of finding them. 4371185029Spjd * 4372185029Spjd * The L2ARC device write speed is also boosted during this time so that 4373185029Spjd * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4374185029Spjd * there are no L2ARC reads, and no fear of degrading read performance 4375185029Spjd * through increased writes. 4376185029Spjd * 4377185029Spjd * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4378185029Spjd * the vdev queue can aggregate them into larger and fewer writes. Each 4379185029Spjd * device is written to in a rotor fashion, sweeping writes through 4380185029Spjd * available space then repeating. 4381185029Spjd * 4382185029Spjd * 7. The L2ARC does not store dirty content. It never needs to flush 4383185029Spjd * write buffers back to disk based storage. 4384185029Spjd * 4385185029Spjd * 8. If an ARC buffer is written (and dirtied) which also exists in the 4386185029Spjd * L2ARC, the now stale L2ARC buffer is immediately dropped. 4387185029Spjd * 4388185029Spjd * The performance of the L2ARC can be tweaked by a number of tunables, which 4389185029Spjd * may be necessary for different workloads: 4390185029Spjd * 4391185029Spjd * l2arc_write_max max write bytes per interval 4392185029Spjd * l2arc_write_boost extra write bytes during device warmup 4393185029Spjd * l2arc_noprefetch skip caching prefetched buffers 4394185029Spjd * l2arc_headroom number of max device writes to precache 4395251478Sdelphij * l2arc_headroom_boost when we find compressed buffers during ARC 4396251478Sdelphij * scanning, we multiply headroom by this 4397251478Sdelphij * percentage factor for the next scan cycle, 4398251478Sdelphij * since more compressed buffers are likely to 4399251478Sdelphij * be present 4400185029Spjd * l2arc_feed_secs seconds between L2ARC writing 4401185029Spjd * 4402185029Spjd * Tunables may be removed or added as future performance improvements are 4403185029Spjd * integrated, and also may become zpool properties. 4404208373Smm * 4405208373Smm * There are three key functions that control how the L2ARC warms up: 4406208373Smm * 4407208373Smm * l2arc_write_eligible() check if a buffer is eligible to cache 4408208373Smm * l2arc_write_size() calculate how much to write 4409208373Smm * l2arc_write_interval() calculate sleep delay between writes 4410208373Smm * 4411208373Smm * These three functions determine what to write, how much, and how quickly 4412208373Smm * to send writes. 4413185029Spjd */ 4414185029Spjd 4415208373Smmstatic boolean_t 4416209962Smml2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab) 4417208373Smm{ 4418208373Smm /* 4419208373Smm * A buffer is *not* eligible for the L2ARC if it: 4420208373Smm * 1. belongs to a different spa. 4421208373Smm * 2. is already cached on the L2ARC. 4422208373Smm * 3. has an I/O in progress (it may be an incomplete read). 4423208373Smm * 4. is flagged not eligible (zfs property). 4424208373Smm */ 4425209962Smm if (ab->b_spa != spa_guid) { 4426208373Smm ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4427208373Smm return (B_FALSE); 4428208373Smm } 4429208373Smm if (ab->b_l2hdr != NULL) { 4430208373Smm ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4431208373Smm return (B_FALSE); 4432208373Smm } 4433208373Smm if (HDR_IO_IN_PROGRESS(ab)) { 4434208373Smm ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4435208373Smm return (B_FALSE); 4436208373Smm } 4437208373Smm if (!HDR_L2CACHE(ab)) { 4438208373Smm ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4439208373Smm return (B_FALSE); 4440208373Smm } 4441208373Smm 4442208373Smm return (B_TRUE); 4443208373Smm} 4444208373Smm 4445208373Smmstatic uint64_t 4446251478Sdelphijl2arc_write_size(void) 4447208373Smm{ 4448208373Smm uint64_t size; 4449208373Smm 4450251478Sdelphij /* 4451251478Sdelphij * Make sure our globals have meaningful values in case the user 4452251478Sdelphij * altered them. 4453251478Sdelphij */ 4454251478Sdelphij size = l2arc_write_max; 4455251478Sdelphij if (size == 0) { 4456251478Sdelphij cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4457251478Sdelphij "be greater than zero, resetting it to the default (%d)", 4458251478Sdelphij L2ARC_WRITE_SIZE); 4459251478Sdelphij size = l2arc_write_max = L2ARC_WRITE_SIZE; 4460251478Sdelphij } 4461208373Smm 4462208373Smm if (arc_warm == B_FALSE) 4463251478Sdelphij size += l2arc_write_boost; 4464208373Smm 4465208373Smm return (size); 4466208373Smm 4467208373Smm} 4468208373Smm 4469208373Smmstatic clock_t 4470208373Smml2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4471208373Smm{ 4472219089Spjd clock_t interval, next, now; 4473208373Smm 4474208373Smm /* 4475208373Smm * If the ARC lists are busy, increase our write rate; if the 4476208373Smm * lists are stale, idle back. This is achieved by checking 4477208373Smm * how much we previously wrote - if it was more than half of 4478208373Smm * what we wanted, schedule the next write much sooner. 4479208373Smm */ 4480208373Smm if (l2arc_feed_again && wrote > (wanted / 2)) 4481208373Smm interval = (hz * l2arc_feed_min_ms) / 1000; 4482208373Smm else 4483208373Smm interval = hz * l2arc_feed_secs; 4484208373Smm 4485219089Spjd now = ddi_get_lbolt(); 4486219089Spjd next = MAX(now, MIN(now + interval, began + interval)); 4487208373Smm 4488208373Smm return (next); 4489208373Smm} 4490208373Smm 4491185029Spjdstatic void 4492185029Spjdl2arc_hdr_stat_add(void) 4493185029Spjd{ 4494185029Spjd ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4495185029Spjd ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4496185029Spjd} 4497185029Spjd 4498185029Spjdstatic void 4499185029Spjdl2arc_hdr_stat_remove(void) 4500185029Spjd{ 4501185029Spjd ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4502185029Spjd ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4503185029Spjd} 4504185029Spjd 4505185029Spjd/* 4506185029Spjd * Cycle through L2ARC devices. This is how L2ARC load balances. 4507185029Spjd * If a device is returned, this also returns holding the spa config lock. 4508185029Spjd */ 4509185029Spjdstatic l2arc_dev_t * 4510185029Spjdl2arc_dev_get_next(void) 4511185029Spjd{ 4512185029Spjd l2arc_dev_t *first, *next = NULL; 4513185029Spjd 4514185029Spjd /* 4515185029Spjd * Lock out the removal of spas (spa_namespace_lock), then removal 4516185029Spjd * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4517185029Spjd * both locks will be dropped and a spa config lock held instead. 4518185029Spjd */ 4519185029Spjd mutex_enter(&spa_namespace_lock); 4520185029Spjd mutex_enter(&l2arc_dev_mtx); 4521185029Spjd 4522185029Spjd /* if there are no vdevs, there is nothing to do */ 4523185029Spjd if (l2arc_ndev == 0) 4524185029Spjd goto out; 4525185029Spjd 4526185029Spjd first = NULL; 4527185029Spjd next = l2arc_dev_last; 4528185029Spjd do { 4529185029Spjd /* loop around the list looking for a non-faulted vdev */ 4530185029Spjd if (next == NULL) { 4531185029Spjd next = list_head(l2arc_dev_list); 4532185029Spjd } else { 4533185029Spjd next = list_next(l2arc_dev_list, next); 4534185029Spjd if (next == NULL) 4535185029Spjd next = list_head(l2arc_dev_list); 4536185029Spjd } 4537185029Spjd 4538185029Spjd /* if we have come back to the start, bail out */ 4539185029Spjd if (first == NULL) 4540185029Spjd first = next; 4541185029Spjd else if (next == first) 4542185029Spjd break; 4543185029Spjd 4544185029Spjd } while (vdev_is_dead(next->l2ad_vdev)); 4545185029Spjd 4546185029Spjd /* if we were unable to find any usable vdevs, return NULL */ 4547185029Spjd if (vdev_is_dead(next->l2ad_vdev)) 4548185029Spjd next = NULL; 4549185029Spjd 4550185029Spjd l2arc_dev_last = next; 4551185029Spjd 4552185029Spjdout: 4553185029Spjd mutex_exit(&l2arc_dev_mtx); 4554185029Spjd 4555185029Spjd /* 4556185029Spjd * Grab the config lock to prevent the 'next' device from being 4557185029Spjd * removed while we are writing to it. 4558185029Spjd */ 4559185029Spjd if (next != NULL) 4560185029Spjd spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4561185029Spjd mutex_exit(&spa_namespace_lock); 4562185029Spjd 4563185029Spjd return (next); 4564185029Spjd} 4565185029Spjd 4566185029Spjd/* 4567185029Spjd * Free buffers that were tagged for destruction. 4568185029Spjd */ 4569185029Spjdstatic void 4570185029Spjdl2arc_do_free_on_write() 4571185029Spjd{ 4572185029Spjd list_t *buflist; 4573185029Spjd l2arc_data_free_t *df, *df_prev; 4574185029Spjd 4575185029Spjd mutex_enter(&l2arc_free_on_write_mtx); 4576185029Spjd buflist = l2arc_free_on_write; 4577185029Spjd 4578185029Spjd for (df = list_tail(buflist); df; df = df_prev) { 4579185029Spjd df_prev = list_prev(buflist, df); 4580185029Spjd ASSERT(df->l2df_data != NULL); 4581185029Spjd ASSERT(df->l2df_func != NULL); 4582185029Spjd df->l2df_func(df->l2df_data, df->l2df_size); 4583185029Spjd list_remove(buflist, df); 4584185029Spjd kmem_free(df, sizeof (l2arc_data_free_t)); 4585185029Spjd } 4586185029Spjd 4587185029Spjd mutex_exit(&l2arc_free_on_write_mtx); 4588185029Spjd} 4589185029Spjd 4590185029Spjd/* 4591185029Spjd * A write to a cache device has completed. Update all headers to allow 4592185029Spjd * reads from these buffers to begin. 4593185029Spjd */ 4594185029Spjdstatic void 4595185029Spjdl2arc_write_done(zio_t *zio) 4596185029Spjd{ 4597185029Spjd l2arc_write_callback_t *cb; 4598185029Spjd l2arc_dev_t *dev; 4599185029Spjd list_t *buflist; 4600185029Spjd arc_buf_hdr_t *head, *ab, *ab_prev; 4601185029Spjd l2arc_buf_hdr_t *abl2; 4602185029Spjd kmutex_t *hash_lock; 4603185029Spjd 4604185029Spjd cb = zio->io_private; 4605185029Spjd ASSERT(cb != NULL); 4606185029Spjd dev = cb->l2wcb_dev; 4607185029Spjd ASSERT(dev != NULL); 4608185029Spjd head = cb->l2wcb_head; 4609185029Spjd ASSERT(head != NULL); 4610185029Spjd buflist = dev->l2ad_buflist; 4611185029Spjd ASSERT(buflist != NULL); 4612185029Spjd DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4613185029Spjd l2arc_write_callback_t *, cb); 4614185029Spjd 4615185029Spjd if (zio->io_error != 0) 4616185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_error); 4617185029Spjd 4618185029Spjd mutex_enter(&l2arc_buflist_mtx); 4619185029Spjd 4620185029Spjd /* 4621185029Spjd * All writes completed, or an error was hit. 4622185029Spjd */ 4623185029Spjd for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 4624185029Spjd ab_prev = list_prev(buflist, ab); 4625260835Sdelphij abl2 = ab->b_l2hdr; 4626185029Spjd 4627260835Sdelphij /* 4628260835Sdelphij * Release the temporary compressed buffer as soon as possible. 4629260835Sdelphij */ 4630260835Sdelphij if (abl2->b_compress != ZIO_COMPRESS_OFF) 4631260835Sdelphij l2arc_release_cdata_buf(ab); 4632260835Sdelphij 4633185029Spjd hash_lock = HDR_LOCK(ab); 4634185029Spjd if (!mutex_tryenter(hash_lock)) { 4635185029Spjd /* 4636185029Spjd * This buffer misses out. It may be in a stage 4637185029Spjd * of eviction. Its ARC_L2_WRITING flag will be 4638185029Spjd * left set, denying reads to this buffer. 4639185029Spjd */ 4640185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4641185029Spjd continue; 4642185029Spjd } 4643185029Spjd 4644185029Spjd if (zio->io_error != 0) { 4645185029Spjd /* 4646185029Spjd * Error - drop L2ARC entry. 4647185029Spjd */ 4648185029Spjd list_remove(buflist, ab); 4649251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4650185029Spjd ab->b_l2hdr = NULL; 4651248572Ssmh trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4652248574Ssmh ab->b_size, 0); 4653185029Spjd kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4654185029Spjd ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4655185029Spjd } 4656185029Spjd 4657185029Spjd /* 4658185029Spjd * Allow ARC to begin reads to this L2ARC entry. 4659185029Spjd */ 4660185029Spjd ab->b_flags &= ~ARC_L2_WRITING; 4661185029Spjd 4662185029Spjd mutex_exit(hash_lock); 4663185029Spjd } 4664185029Spjd 4665185029Spjd atomic_inc_64(&l2arc_writes_done); 4666185029Spjd list_remove(buflist, head); 4667185029Spjd kmem_cache_free(hdr_cache, head); 4668185029Spjd mutex_exit(&l2arc_buflist_mtx); 4669185029Spjd 4670185029Spjd l2arc_do_free_on_write(); 4671185029Spjd 4672185029Spjd kmem_free(cb, sizeof (l2arc_write_callback_t)); 4673185029Spjd} 4674185029Spjd 4675185029Spjd/* 4676185029Spjd * A read to a cache device completed. Validate buffer contents before 4677185029Spjd * handing over to the regular ARC routines. 4678185029Spjd */ 4679185029Spjdstatic void 4680185029Spjdl2arc_read_done(zio_t *zio) 4681185029Spjd{ 4682185029Spjd l2arc_read_callback_t *cb; 4683185029Spjd arc_buf_hdr_t *hdr; 4684185029Spjd arc_buf_t *buf; 4685185029Spjd kmutex_t *hash_lock; 4686185029Spjd int equal; 4687185029Spjd 4688185029Spjd ASSERT(zio->io_vd != NULL); 4689185029Spjd ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4690185029Spjd 4691185029Spjd spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4692185029Spjd 4693185029Spjd cb = zio->io_private; 4694185029Spjd ASSERT(cb != NULL); 4695185029Spjd buf = cb->l2rcb_buf; 4696185029Spjd ASSERT(buf != NULL); 4697185029Spjd 4698219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 4699185029Spjd mutex_enter(hash_lock); 4700219089Spjd hdr = buf->b_hdr; 4701219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4702185029Spjd 4703185029Spjd /* 4704251478Sdelphij * If the buffer was compressed, decompress it first. 4705251478Sdelphij */ 4706251478Sdelphij if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4707251478Sdelphij l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4708251478Sdelphij ASSERT(zio->io_data != NULL); 4709251478Sdelphij 4710251478Sdelphij /* 4711185029Spjd * Check this survived the L2ARC journey. 4712185029Spjd */ 4713185029Spjd equal = arc_cksum_equal(buf); 4714185029Spjd if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4715185029Spjd mutex_exit(hash_lock); 4716185029Spjd zio->io_private = buf; 4717185029Spjd zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4718185029Spjd zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4719185029Spjd arc_read_done(zio); 4720185029Spjd } else { 4721185029Spjd mutex_exit(hash_lock); 4722185029Spjd /* 4723185029Spjd * Buffer didn't survive caching. Increment stats and 4724185029Spjd * reissue to the original storage device. 4725185029Spjd */ 4726185029Spjd if (zio->io_error != 0) { 4727185029Spjd ARCSTAT_BUMP(arcstat_l2_io_error); 4728185029Spjd } else { 4729249195Smm zio->io_error = SET_ERROR(EIO); 4730185029Spjd } 4731185029Spjd if (!equal) 4732185029Spjd ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4733185029Spjd 4734185029Spjd /* 4735185029Spjd * If there's no waiter, issue an async i/o to the primary 4736185029Spjd * storage now. If there *is* a waiter, the caller must 4737185029Spjd * issue the i/o in a context where it's OK to block. 4738185029Spjd */ 4739209962Smm if (zio->io_waiter == NULL) { 4740209962Smm zio_t *pio = zio_unique_parent(zio); 4741209962Smm 4742209962Smm ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4743209962Smm 4744209962Smm zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4745185029Spjd buf->b_data, zio->io_size, arc_read_done, buf, 4746185029Spjd zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4747209962Smm } 4748185029Spjd } 4749185029Spjd 4750185029Spjd kmem_free(cb, sizeof (l2arc_read_callback_t)); 4751185029Spjd} 4752185029Spjd 4753185029Spjd/* 4754185029Spjd * This is the list priority from which the L2ARC will search for pages to 4755185029Spjd * cache. This is used within loops (0..3) to cycle through lists in the 4756185029Spjd * desired order. This order can have a significant effect on cache 4757185029Spjd * performance. 4758185029Spjd * 4759185029Spjd * Currently the metadata lists are hit first, MFU then MRU, followed by 4760185029Spjd * the data lists. This function returns a locked list, and also returns 4761185029Spjd * the lock pointer. 4762185029Spjd */ 4763185029Spjdstatic list_t * 4764185029Spjdl2arc_list_locked(int list_num, kmutex_t **lock) 4765185029Spjd{ 4766247187Smm list_t *list = NULL; 4767205231Skmacy int idx; 4768185029Spjd 4769206796Spjd ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4770206796Spjd 4771205231Skmacy if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4772205231Skmacy idx = list_num; 4773205231Skmacy list = &arc_mfu->arcs_lists[idx]; 4774205231Skmacy *lock = ARCS_LOCK(arc_mfu, idx); 4775206796Spjd } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4776205231Skmacy idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4777205231Skmacy list = &arc_mru->arcs_lists[idx]; 4778205231Skmacy *lock = ARCS_LOCK(arc_mru, idx); 4779206796Spjd } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4780205231Skmacy ARC_BUFC_NUMDATALISTS)) { 4781205231Skmacy idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4782205231Skmacy list = &arc_mfu->arcs_lists[idx]; 4783205231Skmacy *lock = ARCS_LOCK(arc_mfu, idx); 4784205231Skmacy } else { 4785205231Skmacy idx = list_num - ARC_BUFC_NUMLISTS; 4786205231Skmacy list = &arc_mru->arcs_lists[idx]; 4787205231Skmacy *lock = ARCS_LOCK(arc_mru, idx); 4788185029Spjd } 4789185029Spjd 4790185029Spjd ASSERT(!(MUTEX_HELD(*lock))); 4791185029Spjd mutex_enter(*lock); 4792185029Spjd return (list); 4793185029Spjd} 4794185029Spjd 4795185029Spjd/* 4796185029Spjd * Evict buffers from the device write hand to the distance specified in 4797185029Spjd * bytes. This distance may span populated buffers, it may span nothing. 4798185029Spjd * This is clearing a region on the L2ARC device ready for writing. 4799185029Spjd * If the 'all' boolean is set, every buffer is evicted. 4800185029Spjd */ 4801185029Spjdstatic void 4802185029Spjdl2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4803185029Spjd{ 4804185029Spjd list_t *buflist; 4805185029Spjd l2arc_buf_hdr_t *abl2; 4806185029Spjd arc_buf_hdr_t *ab, *ab_prev; 4807185029Spjd kmutex_t *hash_lock; 4808185029Spjd uint64_t taddr; 4809185029Spjd 4810185029Spjd buflist = dev->l2ad_buflist; 4811185029Spjd 4812185029Spjd if (buflist == NULL) 4813185029Spjd return; 4814185029Spjd 4815185029Spjd if (!all && dev->l2ad_first) { 4816185029Spjd /* 4817185029Spjd * This is the first sweep through the device. There is 4818185029Spjd * nothing to evict. 4819185029Spjd */ 4820185029Spjd return; 4821185029Spjd } 4822185029Spjd 4823185029Spjd if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4824185029Spjd /* 4825185029Spjd * When nearing the end of the device, evict to the end 4826185029Spjd * before the device write hand jumps to the start. 4827185029Spjd */ 4828185029Spjd taddr = dev->l2ad_end; 4829185029Spjd } else { 4830185029Spjd taddr = dev->l2ad_hand + distance; 4831185029Spjd } 4832185029Spjd DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4833185029Spjd uint64_t, taddr, boolean_t, all); 4834185029Spjd 4835185029Spjdtop: 4836185029Spjd mutex_enter(&l2arc_buflist_mtx); 4837185029Spjd for (ab = list_tail(buflist); ab; ab = ab_prev) { 4838185029Spjd ab_prev = list_prev(buflist, ab); 4839185029Spjd 4840185029Spjd hash_lock = HDR_LOCK(ab); 4841185029Spjd if (!mutex_tryenter(hash_lock)) { 4842185029Spjd /* 4843185029Spjd * Missed the hash lock. Retry. 4844185029Spjd */ 4845185029Spjd ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4846185029Spjd mutex_exit(&l2arc_buflist_mtx); 4847185029Spjd mutex_enter(hash_lock); 4848185029Spjd mutex_exit(hash_lock); 4849185029Spjd goto top; 4850185029Spjd } 4851185029Spjd 4852185029Spjd if (HDR_L2_WRITE_HEAD(ab)) { 4853185029Spjd /* 4854185029Spjd * We hit a write head node. Leave it for 4855185029Spjd * l2arc_write_done(). 4856185029Spjd */ 4857185029Spjd list_remove(buflist, ab); 4858185029Spjd mutex_exit(hash_lock); 4859185029Spjd continue; 4860185029Spjd } 4861185029Spjd 4862185029Spjd if (!all && ab->b_l2hdr != NULL && 4863185029Spjd (ab->b_l2hdr->b_daddr > taddr || 4864185029Spjd ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4865185029Spjd /* 4866185029Spjd * We've evicted to the target address, 4867185029Spjd * or the end of the device. 4868185029Spjd */ 4869185029Spjd mutex_exit(hash_lock); 4870185029Spjd break; 4871185029Spjd } 4872185029Spjd 4873185029Spjd if (HDR_FREE_IN_PROGRESS(ab)) { 4874185029Spjd /* 4875185029Spjd * Already on the path to destruction. 4876185029Spjd */ 4877185029Spjd mutex_exit(hash_lock); 4878185029Spjd continue; 4879185029Spjd } 4880185029Spjd 4881185029Spjd if (ab->b_state == arc_l2c_only) { 4882185029Spjd ASSERT(!HDR_L2_READING(ab)); 4883185029Spjd /* 4884185029Spjd * This doesn't exist in the ARC. Destroy. 4885185029Spjd * arc_hdr_destroy() will call list_remove() 4886185029Spjd * and decrement arcstat_l2_size. 4887185029Spjd */ 4888185029Spjd arc_change_state(arc_anon, ab, hash_lock); 4889185029Spjd arc_hdr_destroy(ab); 4890185029Spjd } else { 4891185029Spjd /* 4892185029Spjd * Invalidate issued or about to be issued 4893185029Spjd * reads, since we may be about to write 4894185029Spjd * over this location. 4895185029Spjd */ 4896185029Spjd if (HDR_L2_READING(ab)) { 4897185029Spjd ARCSTAT_BUMP(arcstat_l2_evict_reading); 4898185029Spjd ab->b_flags |= ARC_L2_EVICTED; 4899185029Spjd } 4900185029Spjd 4901185029Spjd /* 4902185029Spjd * Tell ARC this no longer exists in L2ARC. 4903185029Spjd */ 4904185029Spjd if (ab->b_l2hdr != NULL) { 4905185029Spjd abl2 = ab->b_l2hdr; 4906251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4907185029Spjd ab->b_l2hdr = NULL; 4908185029Spjd kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4909185029Spjd ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4910185029Spjd } 4911185029Spjd list_remove(buflist, ab); 4912185029Spjd 4913185029Spjd /* 4914185029Spjd * This may have been leftover after a 4915185029Spjd * failed write. 4916185029Spjd */ 4917185029Spjd ab->b_flags &= ~ARC_L2_WRITING; 4918185029Spjd } 4919185029Spjd mutex_exit(hash_lock); 4920185029Spjd } 4921185029Spjd mutex_exit(&l2arc_buflist_mtx); 4922185029Spjd 4923219089Spjd vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0); 4924185029Spjd dev->l2ad_evict = taddr; 4925185029Spjd} 4926185029Spjd 4927185029Spjd/* 4928185029Spjd * Find and write ARC buffers to the L2ARC device. 4929185029Spjd * 4930185029Spjd * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 4931185029Spjd * for reading until they have completed writing. 4932251478Sdelphij * The headroom_boost is an in-out parameter used to maintain headroom boost 4933251478Sdelphij * state between calls to this function. 4934251478Sdelphij * 4935251478Sdelphij * Returns the number of bytes actually written (which may be smaller than 4936251478Sdelphij * the delta by which the device hand has changed due to alignment). 4937185029Spjd */ 4938208373Smmstatic uint64_t 4939251478Sdelphijl2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 4940251478Sdelphij boolean_t *headroom_boost) 4941185029Spjd{ 4942185029Spjd arc_buf_hdr_t *ab, *ab_prev, *head; 4943185029Spjd list_t *list; 4944251478Sdelphij uint64_t write_asize, write_psize, write_sz, headroom, 4945251478Sdelphij buf_compress_minsz; 4946185029Spjd void *buf_data; 4947251478Sdelphij kmutex_t *list_lock; 4948251478Sdelphij boolean_t full; 4949185029Spjd l2arc_write_callback_t *cb; 4950185029Spjd zio_t *pio, *wzio; 4951228103Smm uint64_t guid = spa_load_guid(spa); 4952251478Sdelphij const boolean_t do_headroom_boost = *headroom_boost; 4953185029Spjd int try; 4954185029Spjd 4955185029Spjd ASSERT(dev->l2ad_vdev != NULL); 4956185029Spjd 4957251478Sdelphij /* Lower the flag now, we might want to raise it again later. */ 4958251478Sdelphij *headroom_boost = B_FALSE; 4959251478Sdelphij 4960185029Spjd pio = NULL; 4961251478Sdelphij write_sz = write_asize = write_psize = 0; 4962185029Spjd full = B_FALSE; 4963185029Spjd head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 4964185029Spjd head->b_flags |= ARC_L2_WRITE_HEAD; 4965185029Spjd 4966205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 4967185029Spjd /* 4968251478Sdelphij * We will want to try to compress buffers that are at least 2x the 4969251478Sdelphij * device sector size. 4970251478Sdelphij */ 4971251478Sdelphij buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 4972251478Sdelphij 4973251478Sdelphij /* 4974185029Spjd * Copy buffers for L2ARC writing. 4975185029Spjd */ 4976185029Spjd mutex_enter(&l2arc_buflist_mtx); 4977206796Spjd for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 4978251478Sdelphij uint64_t passed_sz = 0; 4979251478Sdelphij 4980185029Spjd list = l2arc_list_locked(try, &list_lock); 4981205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 4982185029Spjd 4983185029Spjd /* 4984185029Spjd * L2ARC fast warmup. 4985185029Spjd * 4986185029Spjd * Until the ARC is warm and starts to evict, read from the 4987185029Spjd * head of the ARC lists rather than the tail. 4988185029Spjd */ 4989185029Spjd if (arc_warm == B_FALSE) 4990185029Spjd ab = list_head(list); 4991185029Spjd else 4992185029Spjd ab = list_tail(list); 4993206796Spjd if (ab == NULL) 4994205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 4995185029Spjd 4996251478Sdelphij headroom = target_sz * l2arc_headroom; 4997251478Sdelphij if (do_headroom_boost) 4998251478Sdelphij headroom = (headroom * l2arc_headroom_boost) / 100; 4999251478Sdelphij 5000185029Spjd for (; ab; ab = ab_prev) { 5001251478Sdelphij l2arc_buf_hdr_t *l2hdr; 5002251478Sdelphij kmutex_t *hash_lock; 5003251478Sdelphij uint64_t buf_sz; 5004251478Sdelphij 5005185029Spjd if (arc_warm == B_FALSE) 5006185029Spjd ab_prev = list_next(list, ab); 5007185029Spjd else 5008185029Spjd ab_prev = list_prev(list, ab); 5009205231Skmacy ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size); 5010206796Spjd 5011185029Spjd hash_lock = HDR_LOCK(ab); 5012251478Sdelphij if (!mutex_tryenter(hash_lock)) { 5013205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5014185029Spjd /* 5015185029Spjd * Skip this buffer rather than waiting. 5016185029Spjd */ 5017185029Spjd continue; 5018185029Spjd } 5019185029Spjd 5020185029Spjd passed_sz += ab->b_size; 5021185029Spjd if (passed_sz > headroom) { 5022185029Spjd /* 5023185029Spjd * Searched too far. 5024185029Spjd */ 5025185029Spjd mutex_exit(hash_lock); 5026205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5027185029Spjd break; 5028185029Spjd } 5029185029Spjd 5030209962Smm if (!l2arc_write_eligible(guid, ab)) { 5031185029Spjd mutex_exit(hash_lock); 5032185029Spjd continue; 5033185029Spjd } 5034185029Spjd 5035185029Spjd if ((write_sz + ab->b_size) > target_sz) { 5036185029Spjd full = B_TRUE; 5037185029Spjd mutex_exit(hash_lock); 5038205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_full); 5039185029Spjd break; 5040185029Spjd } 5041185029Spjd 5042185029Spjd if (pio == NULL) { 5043185029Spjd /* 5044185029Spjd * Insert a dummy header on the buflist so 5045185029Spjd * l2arc_write_done() can find where the 5046185029Spjd * write buffers begin without searching. 5047185029Spjd */ 5048185029Spjd list_insert_head(dev->l2ad_buflist, head); 5049185029Spjd 5050185029Spjd cb = kmem_alloc( 5051185029Spjd sizeof (l2arc_write_callback_t), KM_SLEEP); 5052185029Spjd cb->l2wcb_dev = dev; 5053185029Spjd cb->l2wcb_head = head; 5054185029Spjd pio = zio_root(spa, l2arc_write_done, cb, 5055185029Spjd ZIO_FLAG_CANFAIL); 5056205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_pios); 5057185029Spjd } 5058185029Spjd 5059185029Spjd /* 5060185029Spjd * Create and add a new L2ARC header. 5061185029Spjd */ 5062251478Sdelphij l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5063251478Sdelphij l2hdr->b_dev = dev; 5064251478Sdelphij ab->b_flags |= ARC_L2_WRITING; 5065185029Spjd 5066251478Sdelphij /* 5067251478Sdelphij * Temporarily stash the data buffer in b_tmp_cdata. 5068251478Sdelphij * The subsequent write step will pick it up from 5069251478Sdelphij * there. This is because can't access ab->b_buf 5070251478Sdelphij * without holding the hash_lock, which we in turn 5071251478Sdelphij * can't access without holding the ARC list locks 5072251478Sdelphij * (which we want to avoid during compression/writing). 5073251478Sdelphij */ 5074251478Sdelphij l2hdr->b_compress = ZIO_COMPRESS_OFF; 5075251478Sdelphij l2hdr->b_asize = ab->b_size; 5076251478Sdelphij l2hdr->b_tmp_cdata = ab->b_buf->b_data; 5077251478Sdelphij 5078185029Spjd buf_sz = ab->b_size; 5079251478Sdelphij ab->b_l2hdr = l2hdr; 5080185029Spjd 5081251478Sdelphij list_insert_head(dev->l2ad_buflist, ab); 5082251478Sdelphij 5083185029Spjd /* 5084185029Spjd * Compute and store the buffer cksum before 5085185029Spjd * writing. On debug the cksum is verified first. 5086185029Spjd */ 5087185029Spjd arc_cksum_verify(ab->b_buf); 5088185029Spjd arc_cksum_compute(ab->b_buf, B_TRUE); 5089185029Spjd 5090185029Spjd mutex_exit(hash_lock); 5091185029Spjd 5092251478Sdelphij write_sz += buf_sz; 5093251478Sdelphij } 5094251478Sdelphij 5095251478Sdelphij mutex_exit(list_lock); 5096251478Sdelphij 5097251478Sdelphij if (full == B_TRUE) 5098251478Sdelphij break; 5099251478Sdelphij } 5100251478Sdelphij 5101251478Sdelphij /* No buffers selected for writing? */ 5102251478Sdelphij if (pio == NULL) { 5103251478Sdelphij ASSERT0(write_sz); 5104251478Sdelphij mutex_exit(&l2arc_buflist_mtx); 5105251478Sdelphij kmem_cache_free(hdr_cache, head); 5106251478Sdelphij return (0); 5107251478Sdelphij } 5108251478Sdelphij 5109251478Sdelphij /* 5110251478Sdelphij * Now start writing the buffers. We're starting at the write head 5111251478Sdelphij * and work backwards, retracing the course of the buffer selector 5112251478Sdelphij * loop above. 5113251478Sdelphij */ 5114251478Sdelphij for (ab = list_prev(dev->l2ad_buflist, head); ab; 5115251478Sdelphij ab = list_prev(dev->l2ad_buflist, ab)) { 5116251478Sdelphij l2arc_buf_hdr_t *l2hdr; 5117251478Sdelphij uint64_t buf_sz; 5118251478Sdelphij 5119251478Sdelphij /* 5120251478Sdelphij * We shouldn't need to lock the buffer here, since we flagged 5121251478Sdelphij * it as ARC_L2_WRITING in the previous step, but we must take 5122251478Sdelphij * care to only access its L2 cache parameters. In particular, 5123251478Sdelphij * ab->b_buf may be invalid by now due to ARC eviction. 5124251478Sdelphij */ 5125251478Sdelphij l2hdr = ab->b_l2hdr; 5126251478Sdelphij l2hdr->b_daddr = dev->l2ad_hand; 5127251478Sdelphij 5128251478Sdelphij if ((ab->b_flags & ARC_L2COMPRESS) && 5129251478Sdelphij l2hdr->b_asize >= buf_compress_minsz) { 5130251478Sdelphij if (l2arc_compress_buf(l2hdr)) { 5131251478Sdelphij /* 5132251478Sdelphij * If compression succeeded, enable headroom 5133251478Sdelphij * boost on the next scan cycle. 5134251478Sdelphij */ 5135251478Sdelphij *headroom_boost = B_TRUE; 5136251478Sdelphij } 5137251478Sdelphij } 5138251478Sdelphij 5139251478Sdelphij /* 5140251478Sdelphij * Pick up the buffer data we had previously stashed away 5141251478Sdelphij * (and now potentially also compressed). 5142251478Sdelphij */ 5143251478Sdelphij buf_data = l2hdr->b_tmp_cdata; 5144251478Sdelphij buf_sz = l2hdr->b_asize; 5145251478Sdelphij 5146251478Sdelphij /* Compression may have squashed the buffer to zero length. */ 5147251478Sdelphij if (buf_sz != 0) { 5148251478Sdelphij uint64_t buf_p_sz; 5149251478Sdelphij 5150185029Spjd wzio = zio_write_phys(pio, dev->l2ad_vdev, 5151185029Spjd dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5152185029Spjd NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5153185029Spjd ZIO_FLAG_CANFAIL, B_FALSE); 5154185029Spjd 5155185029Spjd DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5156185029Spjd zio_t *, wzio); 5157185029Spjd (void) zio_nowait(wzio); 5158185029Spjd 5159251478Sdelphij write_asize += buf_sz; 5160185029Spjd /* 5161185029Spjd * Keep the clock hand suitably device-aligned. 5162185029Spjd */ 5163251478Sdelphij buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5164251478Sdelphij write_psize += buf_p_sz; 5165251478Sdelphij dev->l2ad_hand += buf_p_sz; 5166185029Spjd } 5167251478Sdelphij } 5168185029Spjd 5169185029Spjd mutex_exit(&l2arc_buflist_mtx); 5170185029Spjd 5171251478Sdelphij ASSERT3U(write_asize, <=, target_sz); 5172185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_sent); 5173251478Sdelphij ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5174185029Spjd ARCSTAT_INCR(arcstat_l2_size, write_sz); 5175251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, write_asize); 5176251478Sdelphij vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); 5177185029Spjd 5178185029Spjd /* 5179185029Spjd * Bump device hand to the device start if it is approaching the end. 5180185029Spjd * l2arc_evict() will already have evicted ahead for this case. 5181185029Spjd */ 5182185029Spjd if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5183219089Spjd vdev_space_update(dev->l2ad_vdev, 5184219089Spjd dev->l2ad_end - dev->l2ad_hand, 0, 0); 5185185029Spjd dev->l2ad_hand = dev->l2ad_start; 5186185029Spjd dev->l2ad_evict = dev->l2ad_start; 5187185029Spjd dev->l2ad_first = B_FALSE; 5188185029Spjd } 5189185029Spjd 5190208373Smm dev->l2ad_writing = B_TRUE; 5191185029Spjd (void) zio_wait(pio); 5192208373Smm dev->l2ad_writing = B_FALSE; 5193208373Smm 5194251478Sdelphij return (write_asize); 5195185029Spjd} 5196185029Spjd 5197185029Spjd/* 5198251478Sdelphij * Compresses an L2ARC buffer. 5199251478Sdelphij * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5200251478Sdelphij * size in l2hdr->b_asize. This routine tries to compress the data and 5201251478Sdelphij * depending on the compression result there are three possible outcomes: 5202251478Sdelphij * *) The buffer was incompressible. The original l2hdr contents were left 5203251478Sdelphij * untouched and are ready for writing to an L2 device. 5204251478Sdelphij * *) The buffer was all-zeros, so there is no need to write it to an L2 5205251478Sdelphij * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5206251478Sdelphij * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5207251478Sdelphij * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5208251478Sdelphij * data buffer which holds the compressed data to be written, and b_asize 5209251478Sdelphij * tells us how much data there is. b_compress is set to the appropriate 5210251478Sdelphij * compression algorithm. Once writing is done, invoke 5211251478Sdelphij * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5212251478Sdelphij * 5213251478Sdelphij * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5214251478Sdelphij * buffer was incompressible). 5215251478Sdelphij */ 5216251478Sdelphijstatic boolean_t 5217251478Sdelphijl2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5218251478Sdelphij{ 5219251478Sdelphij void *cdata; 5220268075Sdelphij size_t csize, len, rounded; 5221251478Sdelphij 5222251478Sdelphij ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5223251478Sdelphij ASSERT(l2hdr->b_tmp_cdata != NULL); 5224251478Sdelphij 5225251478Sdelphij len = l2hdr->b_asize; 5226251478Sdelphij cdata = zio_data_buf_alloc(len); 5227251478Sdelphij csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5228256889Ssmh cdata, l2hdr->b_asize, (size_t)(1ULL << l2hdr->b_dev->l2ad_vdev->vdev_ashift)); 5229251478Sdelphij 5230268075Sdelphij rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE); 5231268075Sdelphij if (rounded > csize) { 5232268075Sdelphij bzero((char *)cdata + csize, rounded - csize); 5233268075Sdelphij csize = rounded; 5234268075Sdelphij } 5235268075Sdelphij 5236251478Sdelphij if (csize == 0) { 5237251478Sdelphij /* zero block, indicate that there's nothing to write */ 5238251478Sdelphij zio_data_buf_free(cdata, len); 5239251478Sdelphij l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5240251478Sdelphij l2hdr->b_asize = 0; 5241251478Sdelphij l2hdr->b_tmp_cdata = NULL; 5242251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5243251478Sdelphij return (B_TRUE); 5244251478Sdelphij } else if (csize > 0 && csize < len) { 5245251478Sdelphij /* 5246251478Sdelphij * Compression succeeded, we'll keep the cdata around for 5247251478Sdelphij * writing and release it afterwards. 5248251478Sdelphij */ 5249251478Sdelphij l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5250251478Sdelphij l2hdr->b_asize = csize; 5251251478Sdelphij l2hdr->b_tmp_cdata = cdata; 5252251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_successes); 5253251478Sdelphij return (B_TRUE); 5254251478Sdelphij } else { 5255251478Sdelphij /* 5256251478Sdelphij * Compression failed, release the compressed buffer. 5257251478Sdelphij * l2hdr will be left unmodified. 5258251478Sdelphij */ 5259251478Sdelphij zio_data_buf_free(cdata, len); 5260251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_failures); 5261251478Sdelphij return (B_FALSE); 5262251478Sdelphij } 5263251478Sdelphij} 5264251478Sdelphij 5265251478Sdelphij/* 5266251478Sdelphij * Decompresses a zio read back from an l2arc device. On success, the 5267251478Sdelphij * underlying zio's io_data buffer is overwritten by the uncompressed 5268251478Sdelphij * version. On decompression error (corrupt compressed stream), the 5269251478Sdelphij * zio->io_error value is set to signal an I/O error. 5270251478Sdelphij * 5271251478Sdelphij * Please note that the compressed data stream is not checksummed, so 5272251478Sdelphij * if the underlying device is experiencing data corruption, we may feed 5273251478Sdelphij * corrupt data to the decompressor, so the decompressor needs to be 5274251478Sdelphij * able to handle this situation (LZ4 does). 5275251478Sdelphij */ 5276251478Sdelphijstatic void 5277251478Sdelphijl2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5278251478Sdelphij{ 5279251478Sdelphij ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5280251478Sdelphij 5281251478Sdelphij if (zio->io_error != 0) { 5282251478Sdelphij /* 5283251478Sdelphij * An io error has occured, just restore the original io 5284251478Sdelphij * size in preparation for a main pool read. 5285251478Sdelphij */ 5286251478Sdelphij zio->io_orig_size = zio->io_size = hdr->b_size; 5287251478Sdelphij return; 5288251478Sdelphij } 5289251478Sdelphij 5290251478Sdelphij if (c == ZIO_COMPRESS_EMPTY) { 5291251478Sdelphij /* 5292251478Sdelphij * An empty buffer results in a null zio, which means we 5293251478Sdelphij * need to fill its io_data after we're done restoring the 5294251478Sdelphij * buffer's contents. 5295251478Sdelphij */ 5296251478Sdelphij ASSERT(hdr->b_buf != NULL); 5297251478Sdelphij bzero(hdr->b_buf->b_data, hdr->b_size); 5298251478Sdelphij zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5299251478Sdelphij } else { 5300251478Sdelphij ASSERT(zio->io_data != NULL); 5301251478Sdelphij /* 5302251478Sdelphij * We copy the compressed data from the start of the arc buffer 5303251478Sdelphij * (the zio_read will have pulled in only what we need, the 5304251478Sdelphij * rest is garbage which we will overwrite at decompression) 5305251478Sdelphij * and then decompress back to the ARC data buffer. This way we 5306251478Sdelphij * can minimize copying by simply decompressing back over the 5307251478Sdelphij * original compressed data (rather than decompressing to an 5308251478Sdelphij * aux buffer and then copying back the uncompressed buffer, 5309251478Sdelphij * which is likely to be much larger). 5310251478Sdelphij */ 5311251478Sdelphij uint64_t csize; 5312251478Sdelphij void *cdata; 5313251478Sdelphij 5314251478Sdelphij csize = zio->io_size; 5315251478Sdelphij cdata = zio_data_buf_alloc(csize); 5316251478Sdelphij bcopy(zio->io_data, cdata, csize); 5317251478Sdelphij if (zio_decompress_data(c, cdata, zio->io_data, csize, 5318251478Sdelphij hdr->b_size) != 0) 5319251478Sdelphij zio->io_error = EIO; 5320251478Sdelphij zio_data_buf_free(cdata, csize); 5321251478Sdelphij } 5322251478Sdelphij 5323251478Sdelphij /* Restore the expected uncompressed IO size. */ 5324251478Sdelphij zio->io_orig_size = zio->io_size = hdr->b_size; 5325251478Sdelphij} 5326251478Sdelphij 5327251478Sdelphij/* 5328251478Sdelphij * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5329251478Sdelphij * This buffer serves as a temporary holder of compressed data while 5330251478Sdelphij * the buffer entry is being written to an l2arc device. Once that is 5331251478Sdelphij * done, we can dispose of it. 5332251478Sdelphij */ 5333251478Sdelphijstatic void 5334251478Sdelphijl2arc_release_cdata_buf(arc_buf_hdr_t *ab) 5335251478Sdelphij{ 5336251478Sdelphij l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr; 5337251478Sdelphij 5338251478Sdelphij if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) { 5339251478Sdelphij /* 5340251478Sdelphij * If the data was compressed, then we've allocated a 5341251478Sdelphij * temporary buffer for it, so now we need to release it. 5342251478Sdelphij */ 5343251478Sdelphij ASSERT(l2hdr->b_tmp_cdata != NULL); 5344251478Sdelphij zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size); 5345251478Sdelphij } 5346251478Sdelphij l2hdr->b_tmp_cdata = NULL; 5347251478Sdelphij} 5348251478Sdelphij 5349251478Sdelphij/* 5350185029Spjd * This thread feeds the L2ARC at regular intervals. This is the beating 5351185029Spjd * heart of the L2ARC. 5352185029Spjd */ 5353185029Spjdstatic void 5354185029Spjdl2arc_feed_thread(void *dummy __unused) 5355185029Spjd{ 5356185029Spjd callb_cpr_t cpr; 5357185029Spjd l2arc_dev_t *dev; 5358185029Spjd spa_t *spa; 5359208373Smm uint64_t size, wrote; 5360219089Spjd clock_t begin, next = ddi_get_lbolt(); 5361251478Sdelphij boolean_t headroom_boost = B_FALSE; 5362185029Spjd 5363185029Spjd CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5364185029Spjd 5365185029Spjd mutex_enter(&l2arc_feed_thr_lock); 5366185029Spjd 5367185029Spjd while (l2arc_thread_exit == 0) { 5368185029Spjd CALLB_CPR_SAFE_BEGIN(&cpr); 5369185029Spjd (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5370219089Spjd next - ddi_get_lbolt()); 5371185029Spjd CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5372219089Spjd next = ddi_get_lbolt() + hz; 5373185029Spjd 5374185029Spjd /* 5375185029Spjd * Quick check for L2ARC devices. 5376185029Spjd */ 5377185029Spjd mutex_enter(&l2arc_dev_mtx); 5378185029Spjd if (l2arc_ndev == 0) { 5379185029Spjd mutex_exit(&l2arc_dev_mtx); 5380185029Spjd continue; 5381185029Spjd } 5382185029Spjd mutex_exit(&l2arc_dev_mtx); 5383219089Spjd begin = ddi_get_lbolt(); 5384185029Spjd 5385185029Spjd /* 5386185029Spjd * This selects the next l2arc device to write to, and in 5387185029Spjd * doing so the next spa to feed from: dev->l2ad_spa. This 5388185029Spjd * will return NULL if there are now no l2arc devices or if 5389185029Spjd * they are all faulted. 5390185029Spjd * 5391185029Spjd * If a device is returned, its spa's config lock is also 5392185029Spjd * held to prevent device removal. l2arc_dev_get_next() 5393185029Spjd * will grab and release l2arc_dev_mtx. 5394185029Spjd */ 5395185029Spjd if ((dev = l2arc_dev_get_next()) == NULL) 5396185029Spjd continue; 5397185029Spjd 5398185029Spjd spa = dev->l2ad_spa; 5399185029Spjd ASSERT(spa != NULL); 5400185029Spjd 5401185029Spjd /* 5402219089Spjd * If the pool is read-only then force the feed thread to 5403219089Spjd * sleep a little longer. 5404219089Spjd */ 5405219089Spjd if (!spa_writeable(spa)) { 5406219089Spjd next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5407219089Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5408219089Spjd continue; 5409219089Spjd } 5410219089Spjd 5411219089Spjd /* 5412185029Spjd * Avoid contributing to memory pressure. 5413185029Spjd */ 5414185029Spjd if (arc_reclaim_needed()) { 5415185029Spjd ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5416185029Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5417185029Spjd continue; 5418185029Spjd } 5419185029Spjd 5420185029Spjd ARCSTAT_BUMP(arcstat_l2_feeds); 5421185029Spjd 5422251478Sdelphij size = l2arc_write_size(); 5423185029Spjd 5424185029Spjd /* 5425185029Spjd * Evict L2ARC buffers that will be overwritten. 5426185029Spjd */ 5427185029Spjd l2arc_evict(dev, size, B_FALSE); 5428185029Spjd 5429185029Spjd /* 5430185029Spjd * Write ARC buffers. 5431185029Spjd */ 5432251478Sdelphij wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5433208373Smm 5434208373Smm /* 5435208373Smm * Calculate interval between writes. 5436208373Smm */ 5437208373Smm next = l2arc_write_interval(begin, size, wrote); 5438185029Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5439185029Spjd } 5440185029Spjd 5441185029Spjd l2arc_thread_exit = 0; 5442185029Spjd cv_broadcast(&l2arc_feed_thr_cv); 5443185029Spjd CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5444185029Spjd thread_exit(); 5445185029Spjd} 5446185029Spjd 5447185029Spjdboolean_t 5448185029Spjdl2arc_vdev_present(vdev_t *vd) 5449185029Spjd{ 5450185029Spjd l2arc_dev_t *dev; 5451185029Spjd 5452185029Spjd mutex_enter(&l2arc_dev_mtx); 5453185029Spjd for (dev = list_head(l2arc_dev_list); dev != NULL; 5454185029Spjd dev = list_next(l2arc_dev_list, dev)) { 5455185029Spjd if (dev->l2ad_vdev == vd) 5456185029Spjd break; 5457185029Spjd } 5458185029Spjd mutex_exit(&l2arc_dev_mtx); 5459185029Spjd 5460185029Spjd return (dev != NULL); 5461185029Spjd} 5462185029Spjd 5463185029Spjd/* 5464185029Spjd * Add a vdev for use by the L2ARC. By this point the spa has already 5465185029Spjd * validated the vdev and opened it. 5466185029Spjd */ 5467185029Spjdvoid 5468219089Spjdl2arc_add_vdev(spa_t *spa, vdev_t *vd) 5469185029Spjd{ 5470185029Spjd l2arc_dev_t *adddev; 5471185029Spjd 5472185029Spjd ASSERT(!l2arc_vdev_present(vd)); 5473185029Spjd 5474255753Sgibbs vdev_ashift_optimize(vd); 5475255753Sgibbs 5476185029Spjd /* 5477185029Spjd * Create a new l2arc device entry. 5478185029Spjd */ 5479185029Spjd adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5480185029Spjd adddev->l2ad_spa = spa; 5481185029Spjd adddev->l2ad_vdev = vd; 5482219089Spjd adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5483219089Spjd adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5484185029Spjd adddev->l2ad_hand = adddev->l2ad_start; 5485185029Spjd adddev->l2ad_evict = adddev->l2ad_start; 5486185029Spjd adddev->l2ad_first = B_TRUE; 5487208373Smm adddev->l2ad_writing = B_FALSE; 5488185029Spjd 5489185029Spjd /* 5490185029Spjd * This is a list of all ARC buffers that are still valid on the 5491185029Spjd * device. 5492185029Spjd */ 5493185029Spjd adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5494185029Spjd list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5495185029Spjd offsetof(arc_buf_hdr_t, b_l2node)); 5496185029Spjd 5497219089Spjd vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5498185029Spjd 5499185029Spjd /* 5500185029Spjd * Add device to global list 5501185029Spjd */ 5502185029Spjd mutex_enter(&l2arc_dev_mtx); 5503185029Spjd list_insert_head(l2arc_dev_list, adddev); 5504185029Spjd atomic_inc_64(&l2arc_ndev); 5505185029Spjd mutex_exit(&l2arc_dev_mtx); 5506185029Spjd} 5507185029Spjd 5508185029Spjd/* 5509185029Spjd * Remove a vdev from the L2ARC. 5510185029Spjd */ 5511185029Spjdvoid 5512185029Spjdl2arc_remove_vdev(vdev_t *vd) 5513185029Spjd{ 5514185029Spjd l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5515185029Spjd 5516185029Spjd /* 5517185029Spjd * Find the device by vdev 5518185029Spjd */ 5519185029Spjd mutex_enter(&l2arc_dev_mtx); 5520185029Spjd for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5521185029Spjd nextdev = list_next(l2arc_dev_list, dev); 5522185029Spjd if (vd == dev->l2ad_vdev) { 5523185029Spjd remdev = dev; 5524185029Spjd break; 5525185029Spjd } 5526185029Spjd } 5527185029Spjd ASSERT(remdev != NULL); 5528185029Spjd 5529185029Spjd /* 5530185029Spjd * Remove device from global list 5531185029Spjd */ 5532185029Spjd list_remove(l2arc_dev_list, remdev); 5533185029Spjd l2arc_dev_last = NULL; /* may have been invalidated */ 5534185029Spjd atomic_dec_64(&l2arc_ndev); 5535185029Spjd mutex_exit(&l2arc_dev_mtx); 5536185029Spjd 5537185029Spjd /* 5538185029Spjd * Clear all buflists and ARC references. L2ARC device flush. 5539185029Spjd */ 5540185029Spjd l2arc_evict(remdev, 0, B_TRUE); 5541185029Spjd list_destroy(remdev->l2ad_buflist); 5542185029Spjd kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5543185029Spjd kmem_free(remdev, sizeof (l2arc_dev_t)); 5544185029Spjd} 5545185029Spjd 5546185029Spjdvoid 5547185029Spjdl2arc_init(void) 5548185029Spjd{ 5549185029Spjd l2arc_thread_exit = 0; 5550185029Spjd l2arc_ndev = 0; 5551185029Spjd l2arc_writes_sent = 0; 5552185029Spjd l2arc_writes_done = 0; 5553185029Spjd 5554185029Spjd mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5555185029Spjd cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5556185029Spjd mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5557185029Spjd mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5558185029Spjd mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5559185029Spjd 5560185029Spjd l2arc_dev_list = &L2ARC_dev_list; 5561185029Spjd l2arc_free_on_write = &L2ARC_free_on_write; 5562185029Spjd list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5563185029Spjd offsetof(l2arc_dev_t, l2ad_node)); 5564185029Spjd list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5565185029Spjd offsetof(l2arc_data_free_t, l2df_list_node)); 5566185029Spjd} 5567185029Spjd 5568185029Spjdvoid 5569185029Spjdl2arc_fini(void) 5570185029Spjd{ 5571185029Spjd /* 5572185029Spjd * This is called from dmu_fini(), which is called from spa_fini(); 5573185029Spjd * Because of this, we can assume that all l2arc devices have 5574185029Spjd * already been removed when the pools themselves were removed. 5575185029Spjd */ 5576185029Spjd 5577185029Spjd l2arc_do_free_on_write(); 5578185029Spjd 5579185029Spjd mutex_destroy(&l2arc_feed_thr_lock); 5580185029Spjd cv_destroy(&l2arc_feed_thr_cv); 5581185029Spjd mutex_destroy(&l2arc_dev_mtx); 5582185029Spjd mutex_destroy(&l2arc_buflist_mtx); 5583185029Spjd mutex_destroy(&l2arc_free_on_write_mtx); 5584185029Spjd 5585185029Spjd list_destroy(l2arc_dev_list); 5586185029Spjd list_destroy(l2arc_free_on_write); 5587185029Spjd} 5588185029Spjd 5589185029Spjdvoid 5590185029Spjdl2arc_start(void) 5591185029Spjd{ 5592209962Smm if (!(spa_mode_global & FWRITE)) 5593185029Spjd return; 5594185029Spjd 5595185029Spjd (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5596185029Spjd TS_RUN, minclsyspri); 5597185029Spjd} 5598185029Spjd 5599185029Spjdvoid 5600185029Spjdl2arc_stop(void) 5601185029Spjd{ 5602209962Smm if (!(spa_mode_global & FWRITE)) 5603185029Spjd return; 5604185029Spjd 5605185029Spjd mutex_enter(&l2arc_feed_thr_lock); 5606185029Spjd cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5607185029Spjd l2arc_thread_exit = 1; 5608185029Spjd while (l2arc_thread_exit != 0) 5609185029Spjd cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5610185029Spjd mutex_exit(&l2arc_feed_thr_lock); 5611185029Spjd} 5612