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. 23249643Smm * Copyright (c) 2013 by Delphix. All rights reserved. 24262116Savg * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 25258557Savg * 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 61252751Sdelphij * 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 65252751Sdelphij * 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 80252751Sdelphij * 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> 124252140Sdelphij#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> 130260764Savg#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> 136251419Ssmh#include <sys/trim_map.h> 137219089Spjd#include <zfs_fletcher.h> 138168404Spjd#include <sys/sdt.h> 139168404Spjd 140191902Skmacy#include <vm/vm_pageout.h> 141191902Skmacy 142243674Smm#ifdef illumos 143243674Smm#ifndef _KERNEL 144243674Smm/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 145243674Smmboolean_t arc_watch = B_FALSE; 146243674Smmint arc_procfd; 147243674Smm#endif 148243674Smm#endif /* illumos */ 149243674Smm 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 162260764Savg/* 163260764Savg * The number of iterations through arc_evict_*() before we 164260764Savg * drop & reacquire the lock. 165260764Savg */ 166260764Savgint arc_evict_iterations = 100; 167260764Savg 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 183260764Savg/* 184260764Savg * If this percent of memory is free, don't throttle. 185260764Savg */ 186260764Savgint arc_lotsfree_percent = 10; 187260764Savg 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; 205248547Smmint zfs_disable_dup_eviction = 0; 206185029Spjd 207185029SpjdTUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max); 208185029SpjdTUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 209185029SpjdTUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 210168473SpjdSYSCTL_DECL(_vfs_zfs); 211217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 212168473Spjd "Maximum ARC size"); 213217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 214168473Spjd "Minimum ARC size"); 215168404Spjd 216168404Spjd/* 217185029Spjd * Note that buffers can be in one of 6 states: 218168404Spjd * ARC_anon - anonymous (discussed below) 219168404Spjd * ARC_mru - recently used, currently cached 220168404Spjd * ARC_mru_ghost - recentely used, no longer in cache 221168404Spjd * ARC_mfu - frequently used, currently cached 222168404Spjd * ARC_mfu_ghost - frequently used, no longer in cache 223185029Spjd * ARC_l2c_only - exists in L2ARC but not other states 224185029Spjd * When there are no active references to the buffer, they are 225185029Spjd * are linked onto a list in one of these arc states. These are 226185029Spjd * the only buffers that can be evicted or deleted. Within each 227185029Spjd * state there are multiple lists, one for meta-data and one for 228185029Spjd * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 229185029Spjd * etc.) is tracked separately so that it can be managed more 230185029Spjd * explicitly: favored over data, limited explicitly. 231168404Spjd * 232168404Spjd * Anonymous buffers are buffers that are not associated with 233168404Spjd * a DVA. These are buffers that hold dirty block copies 234168404Spjd * before they are written to stable storage. By definition, 235168404Spjd * they are "ref'd" and are considered part of arc_mru 236168404Spjd * that cannot be freed. Generally, they will aquire a DVA 237168404Spjd * as they are written and migrate onto the arc_mru list. 238185029Spjd * 239185029Spjd * The ARC_l2c_only state is for buffers that are in the second 240185029Spjd * level ARC but no longer in any of the ARC_m* lists. The second 241185029Spjd * level ARC itself may also contain buffers that are in any of 242185029Spjd * the ARC_m* states - meaning that a buffer can exist in two 243185029Spjd * places. The reason for the ARC_l2c_only state is to keep the 244185029Spjd * buffer header in the hash table, so that reads that hit the 245185029Spjd * second level ARC benefit from these fast lookups. 246168404Spjd */ 247168404Spjd 248205264Skmacy#define ARCS_LOCK_PAD CACHE_LINE_SIZE 249205231Skmacystruct arcs_lock { 250205231Skmacy kmutex_t arcs_lock; 251205231Skmacy#ifdef _KERNEL 252205231Skmacy unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 253205231Skmacy#endif 254205231Skmacy}; 255205231Skmacy 256205231Skmacy/* 257205231Skmacy * must be power of two for mask use to work 258205231Skmacy * 259205231Skmacy */ 260205231Skmacy#define ARC_BUFC_NUMDATALISTS 16 261205231Skmacy#define ARC_BUFC_NUMMETADATALISTS 16 262206796Spjd#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 263205231Skmacy 264168404Spjdtypedef struct arc_state { 265185029Spjd uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 266185029Spjd uint64_t arcs_size; /* total amount of data in this state */ 267205231Skmacy list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 268205264Skmacy struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 269168404Spjd} arc_state_t; 270168404Spjd 271206796Spjd#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 272205231Skmacy 273185029Spjd/* The 6 states: */ 274168404Spjdstatic arc_state_t ARC_anon; 275168404Spjdstatic arc_state_t ARC_mru; 276168404Spjdstatic arc_state_t ARC_mru_ghost; 277168404Spjdstatic arc_state_t ARC_mfu; 278168404Spjdstatic arc_state_t ARC_mfu_ghost; 279185029Spjdstatic arc_state_t ARC_l2c_only; 280168404Spjd 281168404Spjdtypedef struct arc_stats { 282168404Spjd kstat_named_t arcstat_hits; 283168404Spjd kstat_named_t arcstat_misses; 284168404Spjd kstat_named_t arcstat_demand_data_hits; 285168404Spjd kstat_named_t arcstat_demand_data_misses; 286168404Spjd kstat_named_t arcstat_demand_metadata_hits; 287168404Spjd kstat_named_t arcstat_demand_metadata_misses; 288168404Spjd kstat_named_t arcstat_prefetch_data_hits; 289168404Spjd kstat_named_t arcstat_prefetch_data_misses; 290168404Spjd kstat_named_t arcstat_prefetch_metadata_hits; 291168404Spjd kstat_named_t arcstat_prefetch_metadata_misses; 292168404Spjd kstat_named_t arcstat_mru_hits; 293168404Spjd kstat_named_t arcstat_mru_ghost_hits; 294168404Spjd kstat_named_t arcstat_mfu_hits; 295168404Spjd kstat_named_t arcstat_mfu_ghost_hits; 296205231Skmacy kstat_named_t arcstat_allocated; 297168404Spjd kstat_named_t arcstat_deleted; 298205231Skmacy kstat_named_t arcstat_stolen; 299168404Spjd kstat_named_t arcstat_recycle_miss; 300252749Sdelphij /* 301252749Sdelphij * Number of buffers that could not be evicted because the hash lock 302252749Sdelphij * was held by another thread. The lock may not necessarily be held 303252749Sdelphij * by something using the same buffer, since hash locks are shared 304252749Sdelphij * by multiple buffers. 305252749Sdelphij */ 306168404Spjd kstat_named_t arcstat_mutex_miss; 307252749Sdelphij /* 308252749Sdelphij * Number of buffers skipped because they have I/O in progress, are 309252749Sdelphij * indrect prefetch buffers that have not lived long enough, or are 310252749Sdelphij * not from the spa we're trying to evict from. 311252749Sdelphij */ 312168404Spjd kstat_named_t arcstat_evict_skip; 313208373Smm kstat_named_t arcstat_evict_l2_cached; 314208373Smm kstat_named_t arcstat_evict_l2_eligible; 315208373Smm kstat_named_t arcstat_evict_l2_ineligible; 316168404Spjd kstat_named_t arcstat_hash_elements; 317168404Spjd kstat_named_t arcstat_hash_elements_max; 318168404Spjd kstat_named_t arcstat_hash_collisions; 319168404Spjd kstat_named_t arcstat_hash_chains; 320168404Spjd kstat_named_t arcstat_hash_chain_max; 321168404Spjd kstat_named_t arcstat_p; 322168404Spjd kstat_named_t arcstat_c; 323168404Spjd kstat_named_t arcstat_c_min; 324168404Spjd kstat_named_t arcstat_c_max; 325168404Spjd kstat_named_t arcstat_size; 326185029Spjd kstat_named_t arcstat_hdr_size; 327208373Smm kstat_named_t arcstat_data_size; 328208373Smm kstat_named_t arcstat_other_size; 329185029Spjd kstat_named_t arcstat_l2_hits; 330185029Spjd kstat_named_t arcstat_l2_misses; 331185029Spjd kstat_named_t arcstat_l2_feeds; 332185029Spjd kstat_named_t arcstat_l2_rw_clash; 333208373Smm kstat_named_t arcstat_l2_read_bytes; 334208373Smm kstat_named_t arcstat_l2_write_bytes; 335185029Spjd kstat_named_t arcstat_l2_writes_sent; 336185029Spjd kstat_named_t arcstat_l2_writes_done; 337185029Spjd kstat_named_t arcstat_l2_writes_error; 338185029Spjd kstat_named_t arcstat_l2_writes_hdr_miss; 339185029Spjd kstat_named_t arcstat_l2_evict_lock_retry; 340185029Spjd kstat_named_t arcstat_l2_evict_reading; 341185029Spjd kstat_named_t arcstat_l2_free_on_write; 342185029Spjd kstat_named_t arcstat_l2_abort_lowmem; 343185029Spjd kstat_named_t arcstat_l2_cksum_bad; 344185029Spjd kstat_named_t arcstat_l2_io_error; 345185029Spjd kstat_named_t arcstat_l2_size; 346252140Sdelphij kstat_named_t arcstat_l2_asize; 347185029Spjd kstat_named_t arcstat_l2_hdr_size; 348252140Sdelphij kstat_named_t arcstat_l2_compress_successes; 349252140Sdelphij kstat_named_t arcstat_l2_compress_zeros; 350252140Sdelphij kstat_named_t arcstat_l2_compress_failures; 351205231Skmacy kstat_named_t arcstat_l2_write_trylock_fail; 352205231Skmacy kstat_named_t arcstat_l2_write_passed_headroom; 353205231Skmacy kstat_named_t arcstat_l2_write_spa_mismatch; 354206796Spjd kstat_named_t arcstat_l2_write_in_l2; 355205231Skmacy kstat_named_t arcstat_l2_write_hdr_io_in_progress; 356205231Skmacy kstat_named_t arcstat_l2_write_not_cacheable; 357205231Skmacy kstat_named_t arcstat_l2_write_full; 358205231Skmacy kstat_named_t arcstat_l2_write_buffer_iter; 359205231Skmacy kstat_named_t arcstat_l2_write_pios; 360205231Skmacy kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 361205231Skmacy kstat_named_t arcstat_l2_write_buffer_list_iter; 362205231Skmacy kstat_named_t arcstat_l2_write_buffer_list_null_iter; 363248547Smm kstat_named_t arcstat_memory_throttle_count; 364248547Smm kstat_named_t arcstat_duplicate_buffers; 365248547Smm kstat_named_t arcstat_duplicate_buffers_size; 366248547Smm kstat_named_t arcstat_duplicate_reads; 367168404Spjd} arc_stats_t; 368168404Spjd 369168404Spjdstatic arc_stats_t arc_stats = { 370168404Spjd { "hits", KSTAT_DATA_UINT64 }, 371168404Spjd { "misses", KSTAT_DATA_UINT64 }, 372168404Spjd { "demand_data_hits", KSTAT_DATA_UINT64 }, 373168404Spjd { "demand_data_misses", KSTAT_DATA_UINT64 }, 374168404Spjd { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 375168404Spjd { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 376168404Spjd { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 377168404Spjd { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 378168404Spjd { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 379168404Spjd { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 380168404Spjd { "mru_hits", KSTAT_DATA_UINT64 }, 381168404Spjd { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 382168404Spjd { "mfu_hits", KSTAT_DATA_UINT64 }, 383168404Spjd { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 384205231Skmacy { "allocated", KSTAT_DATA_UINT64 }, 385168404Spjd { "deleted", KSTAT_DATA_UINT64 }, 386205231Skmacy { "stolen", KSTAT_DATA_UINT64 }, 387168404Spjd { "recycle_miss", KSTAT_DATA_UINT64 }, 388168404Spjd { "mutex_miss", KSTAT_DATA_UINT64 }, 389168404Spjd { "evict_skip", KSTAT_DATA_UINT64 }, 390208373Smm { "evict_l2_cached", KSTAT_DATA_UINT64 }, 391208373Smm { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 392208373Smm { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 393168404Spjd { "hash_elements", KSTAT_DATA_UINT64 }, 394168404Spjd { "hash_elements_max", KSTAT_DATA_UINT64 }, 395168404Spjd { "hash_collisions", KSTAT_DATA_UINT64 }, 396168404Spjd { "hash_chains", KSTAT_DATA_UINT64 }, 397168404Spjd { "hash_chain_max", KSTAT_DATA_UINT64 }, 398168404Spjd { "p", KSTAT_DATA_UINT64 }, 399168404Spjd { "c", KSTAT_DATA_UINT64 }, 400168404Spjd { "c_min", KSTAT_DATA_UINT64 }, 401168404Spjd { "c_max", KSTAT_DATA_UINT64 }, 402185029Spjd { "size", KSTAT_DATA_UINT64 }, 403185029Spjd { "hdr_size", KSTAT_DATA_UINT64 }, 404208373Smm { "data_size", KSTAT_DATA_UINT64 }, 405208373Smm { "other_size", KSTAT_DATA_UINT64 }, 406185029Spjd { "l2_hits", KSTAT_DATA_UINT64 }, 407185029Spjd { "l2_misses", KSTAT_DATA_UINT64 }, 408185029Spjd { "l2_feeds", KSTAT_DATA_UINT64 }, 409185029Spjd { "l2_rw_clash", KSTAT_DATA_UINT64 }, 410208373Smm { "l2_read_bytes", KSTAT_DATA_UINT64 }, 411208373Smm { "l2_write_bytes", KSTAT_DATA_UINT64 }, 412185029Spjd { "l2_writes_sent", KSTAT_DATA_UINT64 }, 413185029Spjd { "l2_writes_done", KSTAT_DATA_UINT64 }, 414185029Spjd { "l2_writes_error", KSTAT_DATA_UINT64 }, 415185029Spjd { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 416185029Spjd { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 417185029Spjd { "l2_evict_reading", KSTAT_DATA_UINT64 }, 418185029Spjd { "l2_free_on_write", KSTAT_DATA_UINT64 }, 419185029Spjd { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 420185029Spjd { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 421185029Spjd { "l2_io_error", KSTAT_DATA_UINT64 }, 422185029Spjd { "l2_size", KSTAT_DATA_UINT64 }, 423252140Sdelphij { "l2_asize", KSTAT_DATA_UINT64 }, 424185029Spjd { "l2_hdr_size", KSTAT_DATA_UINT64 }, 425252140Sdelphij { "l2_compress_successes", KSTAT_DATA_UINT64 }, 426252140Sdelphij { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 427252140Sdelphij { "l2_compress_failures", KSTAT_DATA_UINT64 }, 428206796Spjd { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 429206796Spjd { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 430206796Spjd { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 431206796Spjd { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 432206796Spjd { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 433206796Spjd { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 434206796Spjd { "l2_write_full", KSTAT_DATA_UINT64 }, 435206796Spjd { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 436206796Spjd { "l2_write_pios", KSTAT_DATA_UINT64 }, 437206796Spjd { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 438206796Spjd { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 439248547Smm { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 440248547Smm { "memory_throttle_count", KSTAT_DATA_UINT64 }, 441248547Smm { "duplicate_buffers", KSTAT_DATA_UINT64 }, 442248547Smm { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 443248547Smm { "duplicate_reads", KSTAT_DATA_UINT64 } 444168404Spjd}; 445168404Spjd 446168404Spjd#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 447168404Spjd 448168404Spjd#define ARCSTAT_INCR(stat, val) \ 449252751Sdelphij atomic_add_64(&arc_stats.stat.value.ui64, (val)) 450168404Spjd 451206796Spjd#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 452168404Spjd#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 453168404Spjd 454168404Spjd#define ARCSTAT_MAX(stat, val) { \ 455168404Spjd uint64_t m; \ 456168404Spjd while ((val) > (m = arc_stats.stat.value.ui64) && \ 457168404Spjd (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 458168404Spjd continue; \ 459168404Spjd} 460168404Spjd 461168404Spjd#define ARCSTAT_MAXSTAT(stat) \ 462168404Spjd ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 463168404Spjd 464168404Spjd/* 465168404Spjd * We define a macro to allow ARC hits/misses to be easily broken down by 466168404Spjd * two separate conditions, giving a total of four different subtypes for 467168404Spjd * each of hits and misses (so eight statistics total). 468168404Spjd */ 469168404Spjd#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 470168404Spjd if (cond1) { \ 471168404Spjd if (cond2) { \ 472168404Spjd ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 473168404Spjd } else { \ 474168404Spjd ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 475168404Spjd } \ 476168404Spjd } else { \ 477168404Spjd if (cond2) { \ 478168404Spjd ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 479168404Spjd } else { \ 480168404Spjd ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 481168404Spjd } \ 482168404Spjd } 483168404Spjd 484168404Spjdkstat_t *arc_ksp; 485206796Spjdstatic arc_state_t *arc_anon; 486168404Spjdstatic arc_state_t *arc_mru; 487168404Spjdstatic arc_state_t *arc_mru_ghost; 488168404Spjdstatic arc_state_t *arc_mfu; 489168404Spjdstatic arc_state_t *arc_mfu_ghost; 490185029Spjdstatic arc_state_t *arc_l2c_only; 491168404Spjd 492168404Spjd/* 493168404Spjd * There are several ARC variables that are critical to export as kstats -- 494168404Spjd * but we don't want to have to grovel around in the kstat whenever we wish to 495168404Spjd * manipulate them. For these variables, we therefore define them to be in 496168404Spjd * terms of the statistic variable. This assures that we are not introducing 497168404Spjd * the possibility of inconsistency by having shadow copies of the variables, 498168404Spjd * while still allowing the code to be readable. 499168404Spjd */ 500168404Spjd#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 501168404Spjd#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 502168404Spjd#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 503168404Spjd#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 504168404Spjd#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 505168404Spjd 506252140Sdelphij#define L2ARC_IS_VALID_COMPRESS(_c_) \ 507252140Sdelphij ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 508252140Sdelphij 509168404Spjdstatic int arc_no_grow; /* Don't try to grow cache size */ 510168404Spjdstatic uint64_t arc_tempreserve; 511209962Smmstatic uint64_t arc_loaned_bytes; 512185029Spjdstatic uint64_t arc_meta_used; 513185029Spjdstatic uint64_t arc_meta_limit; 514185029Spjdstatic uint64_t arc_meta_max = 0; 515258635SavgSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RD, &arc_meta_used, 0, 516258635Savg "ARC metadata used"); 517258635SavgSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RW, &arc_meta_limit, 0, 518258635Savg "ARC metadata limit"); 519168404Spjd 520185029Spjdtypedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 521185029Spjd 522168404Spjdtypedef struct arc_callback arc_callback_t; 523168404Spjd 524168404Spjdstruct arc_callback { 525168404Spjd void *acb_private; 526168404Spjd arc_done_func_t *acb_done; 527168404Spjd arc_buf_t *acb_buf; 528168404Spjd zio_t *acb_zio_dummy; 529168404Spjd arc_callback_t *acb_next; 530168404Spjd}; 531168404Spjd 532168404Spjdtypedef struct arc_write_callback arc_write_callback_t; 533168404Spjd 534168404Spjdstruct arc_write_callback { 535168404Spjd void *awcb_private; 536168404Spjd arc_done_func_t *awcb_ready; 537260764Savg arc_done_func_t *awcb_physdone; 538168404Spjd arc_done_func_t *awcb_done; 539168404Spjd arc_buf_t *awcb_buf; 540168404Spjd}; 541168404Spjd 542168404Spjdstruct arc_buf_hdr { 543168404Spjd /* protected by hash lock */ 544168404Spjd dva_t b_dva; 545168404Spjd uint64_t b_birth; 546168404Spjd uint64_t b_cksum0; 547168404Spjd 548168404Spjd kmutex_t b_freeze_lock; 549168404Spjd zio_cksum_t *b_freeze_cksum; 550219089Spjd void *b_thawed; 551168404Spjd 552168404Spjd arc_buf_hdr_t *b_hash_next; 553168404Spjd arc_buf_t *b_buf; 554168404Spjd uint32_t b_flags; 555168404Spjd uint32_t b_datacnt; 556168404Spjd 557168404Spjd arc_callback_t *b_acb; 558168404Spjd kcondvar_t b_cv; 559168404Spjd 560168404Spjd /* immutable */ 561168404Spjd arc_buf_contents_t b_type; 562168404Spjd uint64_t b_size; 563209962Smm uint64_t b_spa; 564168404Spjd 565168404Spjd /* protected by arc state mutex */ 566168404Spjd arc_state_t *b_state; 567168404Spjd list_node_t b_arc_node; 568168404Spjd 569168404Spjd /* updated atomically */ 570168404Spjd clock_t b_arc_access; 571168404Spjd 572168404Spjd /* self protecting */ 573168404Spjd refcount_t b_refcnt; 574185029Spjd 575185029Spjd l2arc_buf_hdr_t *b_l2hdr; 576185029Spjd list_node_t b_l2node; 577168404Spjd}; 578168404Spjd 579168404Spjdstatic arc_buf_t *arc_eviction_list; 580168404Spjdstatic kmutex_t arc_eviction_mtx; 581168404Spjdstatic arc_buf_hdr_t arc_eviction_hdr; 582168404Spjdstatic void arc_get_data_buf(arc_buf_t *buf); 583168404Spjdstatic void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock); 584185029Spjdstatic int arc_evict_needed(arc_buf_contents_t type); 585209962Smmstatic void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes); 586243674Smm#ifdef illumos 587243674Smmstatic void arc_buf_watch(arc_buf_t *buf); 588243674Smm#endif /* illumos */ 589168404Spjd 590209962Smmstatic boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab); 591208373Smm 592168404Spjd#define GHOST_STATE(state) \ 593185029Spjd ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 594185029Spjd (state) == arc_l2c_only) 595168404Spjd 596168404Spjd/* 597168404Spjd * Private ARC flags. These flags are private ARC only flags that will show up 598168404Spjd * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can 599168404Spjd * be passed in as arc_flags in things like arc_read. However, these flags 600168404Spjd * should never be passed and should only be set by ARC code. When adding new 601168404Spjd * public flags, make sure not to smash the private ones. 602168404Spjd */ 603168404Spjd 604168404Spjd#define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */ 605168404Spjd#define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */ 606168404Spjd#define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */ 607168404Spjd#define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */ 608168404Spjd#define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */ 609168404Spjd#define ARC_INDIRECT (1 << 14) /* this is an indirect block */ 610185029Spjd#define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */ 611185029Spjd#define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */ 612185029Spjd#define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */ 613185029Spjd#define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */ 614168404Spjd 615168404Spjd#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE) 616168404Spjd#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS) 617168404Spjd#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR) 618208373Smm#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH) 619168404Spjd#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ) 620168404Spjd#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE) 621185029Spjd#define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS) 622185029Spjd#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE) 623185029Spjd#define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \ 624185029Spjd (hdr)->b_l2hdr != NULL) 625185029Spjd#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING) 626185029Spjd#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED) 627185029Spjd#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD) 628168404Spjd 629168404Spjd/* 630185029Spjd * Other sizes 631185029Spjd */ 632185029Spjd 633185029Spjd#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 634185029Spjd#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 635185029Spjd 636185029Spjd/* 637168404Spjd * Hash table routines 638168404Spjd */ 639168404Spjd 640205253Skmacy#define HT_LOCK_PAD CACHE_LINE_SIZE 641168404Spjd 642168404Spjdstruct ht_lock { 643168404Spjd kmutex_t ht_lock; 644168404Spjd#ifdef _KERNEL 645168404Spjd unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 646168404Spjd#endif 647168404Spjd}; 648168404Spjd 649168404Spjd#define BUF_LOCKS 256 650168404Spjdtypedef struct buf_hash_table { 651168404Spjd uint64_t ht_mask; 652168404Spjd arc_buf_hdr_t **ht_table; 653205264Skmacy struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 654168404Spjd} buf_hash_table_t; 655168404Spjd 656168404Spjdstatic buf_hash_table_t buf_hash_table; 657168404Spjd 658168404Spjd#define BUF_HASH_INDEX(spa, dva, birth) \ 659168404Spjd (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 660168404Spjd#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 661168404Spjd#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 662219089Spjd#define HDR_LOCK(hdr) \ 663219089Spjd (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 664168404Spjd 665168404Spjduint64_t zfs_crc64_table[256]; 666168404Spjd 667185029Spjd/* 668185029Spjd * Level 2 ARC 669185029Spjd */ 670185029Spjd 671208373Smm#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 672252140Sdelphij#define L2ARC_HEADROOM 2 /* num of writes */ 673252140Sdelphij/* 674252140Sdelphij * If we discover during ARC scan any buffers to be compressed, we boost 675252140Sdelphij * our headroom for the next scanning cycle by this percentage multiple. 676252140Sdelphij */ 677252140Sdelphij#define L2ARC_HEADROOM_BOOST 200 678208373Smm#define L2ARC_FEED_SECS 1 /* caching interval secs */ 679208373Smm#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 680185029Spjd 681185029Spjd#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 682185029Spjd#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 683185029Spjd 684252751Sdelphij/* L2ARC Performance Tunables */ 685185029Spjduint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 686185029Spjduint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 687185029Spjduint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 688252140Sdelphijuint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 689185029Spjduint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 690208373Smmuint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 691219089Spjdboolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 692208373Smmboolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 693208373Smmboolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 694185029Spjd 695217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 696205231Skmacy &l2arc_write_max, 0, "max write size"); 697217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 698205231Skmacy &l2arc_write_boost, 0, "extra write during warmup"); 699217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 700205231Skmacy &l2arc_headroom, 0, "number of dev writes"); 701217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 702205231Skmacy &l2arc_feed_secs, 0, "interval seconds"); 703217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 704208373Smm &l2arc_feed_min_ms, 0, "min interval milliseconds"); 705205231Skmacy 706205231SkmacySYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 707205231Skmacy &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 708208373SmmSYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 709208373Smm &l2arc_feed_again, 0, "turbo warmup"); 710208373SmmSYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 711208373Smm &l2arc_norw, 0, "no reads during writes"); 712205231Skmacy 713217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 714205231Skmacy &ARC_anon.arcs_size, 0, "size of anonymous state"); 715217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 716205231Skmacy &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 717217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 718205231Skmacy &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 719205231Skmacy 720217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 721205231Skmacy &ARC_mru.arcs_size, 0, "size of mru state"); 722217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 723205231Skmacy &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 724217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 725205231Skmacy &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 726205231Skmacy 727217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 728205231Skmacy &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 729217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 730205231Skmacy &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 731205231Skmacy "size of metadata in mru ghost state"); 732217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 733205231Skmacy &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 734205231Skmacy "size of data in mru ghost state"); 735205231Skmacy 736217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 737205231Skmacy &ARC_mfu.arcs_size, 0, "size of mfu state"); 738217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 739205231Skmacy &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 740217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 741205231Skmacy &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 742205231Skmacy 743217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 744205231Skmacy &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 745217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 746205231Skmacy &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 747205231Skmacy "size of metadata in mfu ghost state"); 748217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 749205231Skmacy &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 750205231Skmacy "size of data in mfu ghost state"); 751205231Skmacy 752217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 753205231Skmacy &ARC_l2c_only.arcs_size, 0, "size of mru state"); 754205231Skmacy 755185029Spjd/* 756185029Spjd * L2ARC Internals 757185029Spjd */ 758185029Spjdtypedef struct l2arc_dev { 759185029Spjd vdev_t *l2ad_vdev; /* vdev */ 760185029Spjd spa_t *l2ad_spa; /* spa */ 761185029Spjd uint64_t l2ad_hand; /* next write location */ 762185029Spjd uint64_t l2ad_start; /* first addr on device */ 763185029Spjd uint64_t l2ad_end; /* last addr on device */ 764185029Spjd uint64_t l2ad_evict; /* last addr eviction reached */ 765185029Spjd boolean_t l2ad_first; /* first sweep through */ 766208373Smm boolean_t l2ad_writing; /* currently writing */ 767185029Spjd list_t *l2ad_buflist; /* buffer list */ 768185029Spjd list_node_t l2ad_node; /* device list node */ 769185029Spjd} l2arc_dev_t; 770185029Spjd 771185029Spjdstatic list_t L2ARC_dev_list; /* device list */ 772185029Spjdstatic list_t *l2arc_dev_list; /* device list pointer */ 773185029Spjdstatic kmutex_t l2arc_dev_mtx; /* device list mutex */ 774185029Spjdstatic l2arc_dev_t *l2arc_dev_last; /* last device used */ 775185029Spjdstatic kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 776185029Spjdstatic list_t L2ARC_free_on_write; /* free after write buf list */ 777185029Spjdstatic list_t *l2arc_free_on_write; /* free after write list ptr */ 778185029Spjdstatic kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 779185029Spjdstatic uint64_t l2arc_ndev; /* number of devices */ 780185029Spjd 781185029Spjdtypedef struct l2arc_read_callback { 782252140Sdelphij arc_buf_t *l2rcb_buf; /* read buffer */ 783252140Sdelphij spa_t *l2rcb_spa; /* spa */ 784252140Sdelphij blkptr_t l2rcb_bp; /* original blkptr */ 785252140Sdelphij zbookmark_t l2rcb_zb; /* original bookmark */ 786252140Sdelphij int l2rcb_flags; /* original flags */ 787252140Sdelphij enum zio_compress l2rcb_compress; /* applied compress */ 788185029Spjd} l2arc_read_callback_t; 789185029Spjd 790185029Spjdtypedef struct l2arc_write_callback { 791185029Spjd l2arc_dev_t *l2wcb_dev; /* device info */ 792185029Spjd arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 793185029Spjd} l2arc_write_callback_t; 794185029Spjd 795185029Spjdstruct l2arc_buf_hdr { 796185029Spjd /* protected by arc_buf_hdr mutex */ 797252140Sdelphij l2arc_dev_t *b_dev; /* L2ARC device */ 798252140Sdelphij uint64_t b_daddr; /* disk address, offset byte */ 799252140Sdelphij /* compression applied to buffer data */ 800252140Sdelphij enum zio_compress b_compress; 801252140Sdelphij /* real alloc'd buffer size depending on b_compress applied */ 802252140Sdelphij int b_asize; 803252140Sdelphij /* temporary buffer holder for in-flight compressed data */ 804252140Sdelphij void *b_tmp_cdata; 805185029Spjd}; 806185029Spjd 807185029Spjdtypedef struct l2arc_data_free { 808185029Spjd /* protected by l2arc_free_on_write_mtx */ 809185029Spjd void *l2df_data; 810185029Spjd size_t l2df_size; 811185029Spjd void (*l2df_func)(void *, size_t); 812185029Spjd list_node_t l2df_list_node; 813185029Spjd} l2arc_data_free_t; 814185029Spjd 815185029Spjdstatic kmutex_t l2arc_feed_thr_lock; 816185029Spjdstatic kcondvar_t l2arc_feed_thr_cv; 817185029Spjdstatic uint8_t l2arc_thread_exit; 818185029Spjd 819185029Spjdstatic void l2arc_read_done(zio_t *zio); 820185029Spjdstatic void l2arc_hdr_stat_add(void); 821185029Spjdstatic void l2arc_hdr_stat_remove(void); 822185029Spjd 823252140Sdelphijstatic boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr); 824252140Sdelphijstatic void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, 825252140Sdelphij enum zio_compress c); 826252140Sdelphijstatic void l2arc_release_cdata_buf(arc_buf_hdr_t *ab); 827252140Sdelphij 828168404Spjdstatic uint64_t 829209962Smmbuf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 830168404Spjd{ 831168404Spjd uint8_t *vdva = (uint8_t *)dva; 832168404Spjd uint64_t crc = -1ULL; 833168404Spjd int i; 834168404Spjd 835168404Spjd ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 836168404Spjd 837168404Spjd for (i = 0; i < sizeof (dva_t); i++) 838168404Spjd crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 839168404Spjd 840209962Smm crc ^= (spa>>8) ^ birth; 841168404Spjd 842168404Spjd return (crc); 843168404Spjd} 844168404Spjd 845168404Spjd#define BUF_EMPTY(buf) \ 846168404Spjd ((buf)->b_dva.dva_word[0] == 0 && \ 847168404Spjd (buf)->b_dva.dva_word[1] == 0 && \ 848263398Sdelphij (buf)->b_cksum0 == 0) 849168404Spjd 850168404Spjd#define BUF_EQUAL(spa, dva, birth, buf) \ 851168404Spjd ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 852168404Spjd ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 853168404Spjd ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 854168404Spjd 855219089Spjdstatic void 856219089Spjdbuf_discard_identity(arc_buf_hdr_t *hdr) 857219089Spjd{ 858219089Spjd hdr->b_dva.dva_word[0] = 0; 859219089Spjd hdr->b_dva.dva_word[1] = 0; 860219089Spjd hdr->b_birth = 0; 861219089Spjd hdr->b_cksum0 = 0; 862219089Spjd} 863219089Spjd 864168404Spjdstatic arc_buf_hdr_t * 865209962Smmbuf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp) 866168404Spjd{ 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 898168404Spjd ASSERT(!HDR_IN_HASH_TABLE(buf)); 899168404Spjd *lockp = hash_lock; 900168404Spjd mutex_enter(hash_lock); 901168404Spjd for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL; 902168404Spjd fbuf = fbuf->b_hash_next, i++) { 903168404Spjd if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf)) 904168404Spjd return (fbuf); 905168404Spjd } 906168404Spjd 907168404Spjd buf->b_hash_next = buf_hash_table.ht_table[idx]; 908168404Spjd buf_hash_table.ht_table[idx] = buf; 909168404Spjd buf->b_flags |= ARC_IN_HASH_TABLE; 910168404Spjd 911168404Spjd /* collect some hash table performance data */ 912168404Spjd if (i > 0) { 913168404Spjd ARCSTAT_BUMP(arcstat_hash_collisions); 914168404Spjd if (i == 1) 915168404Spjd ARCSTAT_BUMP(arcstat_hash_chains); 916168404Spjd 917168404Spjd ARCSTAT_MAX(arcstat_hash_chain_max, i); 918168404Spjd } 919168404Spjd 920168404Spjd ARCSTAT_BUMP(arcstat_hash_elements); 921168404Spjd ARCSTAT_MAXSTAT(arcstat_hash_elements); 922168404Spjd 923168404Spjd return (NULL); 924168404Spjd} 925168404Spjd 926168404Spjdstatic void 927168404Spjdbuf_hash_remove(arc_buf_hdr_t *buf) 928168404Spjd{ 929168404Spjd arc_buf_hdr_t *fbuf, **bufp; 930168404Spjd uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 931168404Spjd 932168404Spjd ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 933168404Spjd ASSERT(HDR_IN_HASH_TABLE(buf)); 934168404Spjd 935168404Spjd bufp = &buf_hash_table.ht_table[idx]; 936168404Spjd while ((fbuf = *bufp) != buf) { 937168404Spjd ASSERT(fbuf != NULL); 938168404Spjd bufp = &fbuf->b_hash_next; 939168404Spjd } 940168404Spjd *bufp = buf->b_hash_next; 941168404Spjd buf->b_hash_next = NULL; 942168404Spjd buf->b_flags &= ~ARC_IN_HASH_TABLE; 943168404Spjd 944168404Spjd /* collect some hash table performance data */ 945168404Spjd ARCSTAT_BUMPDOWN(arcstat_hash_elements); 946168404Spjd 947168404Spjd if (buf_hash_table.ht_table[idx] && 948168404Spjd buf_hash_table.ht_table[idx]->b_hash_next == NULL) 949168404Spjd ARCSTAT_BUMPDOWN(arcstat_hash_chains); 950168404Spjd} 951168404Spjd 952168404Spjd/* 953168404Spjd * Global data structures and functions for the buf kmem cache. 954168404Spjd */ 955168404Spjdstatic kmem_cache_t *hdr_cache; 956168404Spjdstatic kmem_cache_t *buf_cache; 957168404Spjd 958168404Spjdstatic void 959168404Spjdbuf_fini(void) 960168404Spjd{ 961168404Spjd int i; 962168404Spjd 963168404Spjd kmem_free(buf_hash_table.ht_table, 964168404Spjd (buf_hash_table.ht_mask + 1) * sizeof (void *)); 965168404Spjd for (i = 0; i < BUF_LOCKS; i++) 966168404Spjd mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 967168404Spjd kmem_cache_destroy(hdr_cache); 968168404Spjd kmem_cache_destroy(buf_cache); 969168404Spjd} 970168404Spjd 971168404Spjd/* 972168404Spjd * Constructor callback - called when the cache is empty 973168404Spjd * and a new buf is requested. 974168404Spjd */ 975168404Spjd/* ARGSUSED */ 976168404Spjdstatic int 977168404Spjdhdr_cons(void *vbuf, void *unused, int kmflag) 978168404Spjd{ 979168404Spjd arc_buf_hdr_t *buf = vbuf; 980168404Spjd 981168404Spjd bzero(buf, sizeof (arc_buf_hdr_t)); 982168404Spjd refcount_create(&buf->b_refcnt); 983168404Spjd cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL); 984185029Spjd mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 985208373Smm arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 986185029Spjd 987168404Spjd return (0); 988168404Spjd} 989168404Spjd 990185029Spjd/* ARGSUSED */ 991185029Spjdstatic int 992185029Spjdbuf_cons(void *vbuf, void *unused, int kmflag) 993185029Spjd{ 994185029Spjd arc_buf_t *buf = vbuf; 995185029Spjd 996185029Spjd bzero(buf, sizeof (arc_buf_t)); 997219089Spjd mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 998208373Smm arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 999208373Smm 1000185029Spjd return (0); 1001185029Spjd} 1002185029Spjd 1003168404Spjd/* 1004168404Spjd * Destructor callback - called when a cached buf is 1005168404Spjd * no longer required. 1006168404Spjd */ 1007168404Spjd/* ARGSUSED */ 1008168404Spjdstatic void 1009168404Spjdhdr_dest(void *vbuf, void *unused) 1010168404Spjd{ 1011168404Spjd arc_buf_hdr_t *buf = vbuf; 1012168404Spjd 1013219089Spjd ASSERT(BUF_EMPTY(buf)); 1014168404Spjd refcount_destroy(&buf->b_refcnt); 1015168404Spjd cv_destroy(&buf->b_cv); 1016185029Spjd mutex_destroy(&buf->b_freeze_lock); 1017208373Smm arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1018168404Spjd} 1019168404Spjd 1020185029Spjd/* ARGSUSED */ 1021185029Spjdstatic void 1022185029Spjdbuf_dest(void *vbuf, void *unused) 1023185029Spjd{ 1024185029Spjd arc_buf_t *buf = vbuf; 1025185029Spjd 1026219089Spjd mutex_destroy(&buf->b_evict_lock); 1027208373Smm arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1028185029Spjd} 1029185029Spjd 1030168404Spjd/* 1031168404Spjd * Reclaim callback -- invoked when memory is low. 1032168404Spjd */ 1033168404Spjd/* ARGSUSED */ 1034168404Spjdstatic void 1035168404Spjdhdr_recl(void *unused) 1036168404Spjd{ 1037168404Spjd dprintf("hdr_recl called\n"); 1038168404Spjd /* 1039168404Spjd * umem calls the reclaim func when we destroy the buf cache, 1040168404Spjd * which is after we do arc_fini(). 1041168404Spjd */ 1042168404Spjd if (!arc_dead) 1043168404Spjd cv_signal(&arc_reclaim_thr_cv); 1044168404Spjd} 1045168404Spjd 1046168404Spjdstatic void 1047168404Spjdbuf_init(void) 1048168404Spjd{ 1049168404Spjd uint64_t *ct; 1050168404Spjd uint64_t hsize = 1ULL << 12; 1051168404Spjd int i, j; 1052168404Spjd 1053168404Spjd /* 1054168404Spjd * The hash table is big enough to fill all of physical memory 1055168404Spjd * with an average 64K block size. The table will take up 1056168404Spjd * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers). 1057168404Spjd */ 1058168696Spjd while (hsize * 65536 < (uint64_t)physmem * PAGESIZE) 1059168404Spjd hsize <<= 1; 1060168404Spjdretry: 1061168404Spjd buf_hash_table.ht_mask = hsize - 1; 1062168404Spjd buf_hash_table.ht_table = 1063168404Spjd kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1064168404Spjd if (buf_hash_table.ht_table == NULL) { 1065168404Spjd ASSERT(hsize > (1ULL << 8)); 1066168404Spjd hsize >>= 1; 1067168404Spjd goto retry; 1068168404Spjd } 1069168404Spjd 1070168404Spjd hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1071168404Spjd 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1072168404Spjd buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1073185029Spjd 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1074168404Spjd 1075168404Spjd for (i = 0; i < 256; i++) 1076168404Spjd for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1077168404Spjd *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1078168404Spjd 1079168404Spjd for (i = 0; i < BUF_LOCKS; i++) { 1080168404Spjd mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1081168404Spjd NULL, MUTEX_DEFAULT, NULL); 1082168404Spjd } 1083168404Spjd} 1084168404Spjd 1085168404Spjd#define ARC_MINTIME (hz>>4) /* 62 ms */ 1086168404Spjd 1087168404Spjdstatic void 1088168404Spjdarc_cksum_verify(arc_buf_t *buf) 1089168404Spjd{ 1090168404Spjd zio_cksum_t zc; 1091168404Spjd 1092168404Spjd if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1093168404Spjd return; 1094168404Spjd 1095168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1096168404Spjd if (buf->b_hdr->b_freeze_cksum == NULL || 1097168404Spjd (buf->b_hdr->b_flags & ARC_IO_ERROR)) { 1098168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1099168404Spjd return; 1100168404Spjd } 1101168404Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1102168404Spjd if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1103168404Spjd panic("buffer modified while frozen!"); 1104168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1105168404Spjd} 1106168404Spjd 1107185029Spjdstatic int 1108185029Spjdarc_cksum_equal(arc_buf_t *buf) 1109185029Spjd{ 1110185029Spjd zio_cksum_t zc; 1111185029Spjd int equal; 1112185029Spjd 1113185029Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1114185029Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1115185029Spjd equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1116185029Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1117185029Spjd 1118185029Spjd return (equal); 1119185029Spjd} 1120185029Spjd 1121168404Spjdstatic void 1122185029Spjdarc_cksum_compute(arc_buf_t *buf, boolean_t force) 1123168404Spjd{ 1124185029Spjd if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1125168404Spjd return; 1126168404Spjd 1127168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1128168404Spjd if (buf->b_hdr->b_freeze_cksum != NULL) { 1129168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1130168404Spjd return; 1131168404Spjd } 1132168404Spjd buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1133168404Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1134168404Spjd buf->b_hdr->b_freeze_cksum); 1135168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1136243674Smm#ifdef illumos 1137243674Smm arc_buf_watch(buf); 1138243674Smm#endif /* illumos */ 1139168404Spjd} 1140168404Spjd 1141243674Smm#ifdef illumos 1142243674Smm#ifndef _KERNEL 1143243674Smmtypedef struct procctl { 1144243674Smm long cmd; 1145243674Smm prwatch_t prwatch; 1146243674Smm} procctl_t; 1147243674Smm#endif 1148243674Smm 1149243674Smm/* ARGSUSED */ 1150243674Smmstatic void 1151243674Smmarc_buf_unwatch(arc_buf_t *buf) 1152243674Smm{ 1153243674Smm#ifndef _KERNEL 1154243674Smm if (arc_watch) { 1155243674Smm int result; 1156243674Smm procctl_t ctl; 1157243674Smm ctl.cmd = PCWATCH; 1158243674Smm ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1159243674Smm ctl.prwatch.pr_size = 0; 1160243674Smm ctl.prwatch.pr_wflags = 0; 1161243674Smm result = write(arc_procfd, &ctl, sizeof (ctl)); 1162243674Smm ASSERT3U(result, ==, sizeof (ctl)); 1163243674Smm } 1164243674Smm#endif 1165243674Smm} 1166243674Smm 1167243674Smm/* ARGSUSED */ 1168243674Smmstatic void 1169243674Smmarc_buf_watch(arc_buf_t *buf) 1170243674Smm{ 1171243674Smm#ifndef _KERNEL 1172243674Smm if (arc_watch) { 1173243674Smm int result; 1174243674Smm procctl_t ctl; 1175243674Smm ctl.cmd = PCWATCH; 1176243674Smm ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1177243674Smm ctl.prwatch.pr_size = buf->b_hdr->b_size; 1178243674Smm ctl.prwatch.pr_wflags = WA_WRITE; 1179243674Smm result = write(arc_procfd, &ctl, sizeof (ctl)); 1180243674Smm ASSERT3U(result, ==, sizeof (ctl)); 1181243674Smm } 1182243674Smm#endif 1183243674Smm} 1184243674Smm#endif /* illumos */ 1185243674Smm 1186168404Spjdvoid 1187168404Spjdarc_buf_thaw(arc_buf_t *buf) 1188168404Spjd{ 1189185029Spjd if (zfs_flags & ZFS_DEBUG_MODIFY) { 1190185029Spjd if (buf->b_hdr->b_state != arc_anon) 1191185029Spjd panic("modifying non-anon buffer!"); 1192185029Spjd if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS) 1193185029Spjd panic("modifying buffer while i/o in progress!"); 1194185029Spjd arc_cksum_verify(buf); 1195185029Spjd } 1196168404Spjd 1197168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1198168404Spjd if (buf->b_hdr->b_freeze_cksum != NULL) { 1199168404Spjd kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1200168404Spjd buf->b_hdr->b_freeze_cksum = NULL; 1201168404Spjd } 1202219089Spjd 1203219089Spjd if (zfs_flags & ZFS_DEBUG_MODIFY) { 1204219089Spjd if (buf->b_hdr->b_thawed) 1205219089Spjd kmem_free(buf->b_hdr->b_thawed, 1); 1206219089Spjd buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1207219089Spjd } 1208219089Spjd 1209168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1210243674Smm 1211243674Smm#ifdef illumos 1212243674Smm arc_buf_unwatch(buf); 1213243674Smm#endif /* illumos */ 1214168404Spjd} 1215168404Spjd 1216168404Spjdvoid 1217168404Spjdarc_buf_freeze(arc_buf_t *buf) 1218168404Spjd{ 1219219089Spjd kmutex_t *hash_lock; 1220219089Spjd 1221168404Spjd if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1222168404Spjd return; 1223168404Spjd 1224219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 1225219089Spjd mutex_enter(hash_lock); 1226219089Spjd 1227168404Spjd ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1228168404Spjd buf->b_hdr->b_state == arc_anon); 1229185029Spjd arc_cksum_compute(buf, B_FALSE); 1230219089Spjd mutex_exit(hash_lock); 1231243674Smm 1232168404Spjd} 1233168404Spjd 1234168404Spjdstatic void 1235205231Skmacyget_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock) 1236205231Skmacy{ 1237205231Skmacy uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth); 1238205231Skmacy 1239206796Spjd if (ab->b_type == ARC_BUFC_METADATA) 1240206796Spjd buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1241205231Skmacy else { 1242206796Spjd buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1243205231Skmacy buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1244205231Skmacy } 1245205231Skmacy 1246205231Skmacy *list = &state->arcs_lists[buf_hashid]; 1247205231Skmacy *lock = ARCS_LOCK(state, buf_hashid); 1248205231Skmacy} 1249205231Skmacy 1250205231Skmacy 1251205231Skmacystatic void 1252168404Spjdadd_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1253168404Spjd{ 1254168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 1255168404Spjd 1256168404Spjd if ((refcount_add(&ab->b_refcnt, tag) == 1) && 1257168404Spjd (ab->b_state != arc_anon)) { 1258206796Spjd uint64_t delta = ab->b_size * ab->b_datacnt; 1259206796Spjd uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type]; 1260205231Skmacy list_t *list; 1261205231Skmacy kmutex_t *lock; 1262168404Spjd 1263205231Skmacy get_buf_info(ab, ab->b_state, &list, &lock); 1264205231Skmacy ASSERT(!MUTEX_HELD(lock)); 1265205231Skmacy mutex_enter(lock); 1266168404Spjd ASSERT(list_link_active(&ab->b_arc_node)); 1267185029Spjd list_remove(list, ab); 1268168404Spjd if (GHOST_STATE(ab->b_state)) { 1269243674Smm ASSERT0(ab->b_datacnt); 1270168404Spjd ASSERT3P(ab->b_buf, ==, NULL); 1271168404Spjd delta = ab->b_size; 1272168404Spjd } 1273168404Spjd ASSERT(delta > 0); 1274185029Spjd ASSERT3U(*size, >=, delta); 1275185029Spjd atomic_add_64(size, -delta); 1276206794Spjd mutex_exit(lock); 1277185029Spjd /* remove the prefetch flag if we get a reference */ 1278168404Spjd if (ab->b_flags & ARC_PREFETCH) 1279168404Spjd ab->b_flags &= ~ARC_PREFETCH; 1280168404Spjd } 1281168404Spjd} 1282168404Spjd 1283168404Spjdstatic int 1284168404Spjdremove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1285168404Spjd{ 1286168404Spjd int cnt; 1287168404Spjd arc_state_t *state = ab->b_state; 1288168404Spjd 1289168404Spjd ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1290168404Spjd ASSERT(!GHOST_STATE(state)); 1291168404Spjd 1292168404Spjd if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) && 1293168404Spjd (state != arc_anon)) { 1294185029Spjd uint64_t *size = &state->arcs_lsize[ab->b_type]; 1295205231Skmacy list_t *list; 1296205231Skmacy kmutex_t *lock; 1297185029Spjd 1298205231Skmacy get_buf_info(ab, state, &list, &lock); 1299205231Skmacy ASSERT(!MUTEX_HELD(lock)); 1300205231Skmacy mutex_enter(lock); 1301168404Spjd ASSERT(!list_link_active(&ab->b_arc_node)); 1302205231Skmacy list_insert_head(list, ab); 1303168404Spjd ASSERT(ab->b_datacnt > 0); 1304185029Spjd atomic_add_64(size, ab->b_size * ab->b_datacnt); 1305206794Spjd mutex_exit(lock); 1306168404Spjd } 1307168404Spjd return (cnt); 1308168404Spjd} 1309168404Spjd 1310168404Spjd/* 1311168404Spjd * Move the supplied buffer to the indicated state. The mutex 1312168404Spjd * for the buffer must be held by the caller. 1313168404Spjd */ 1314168404Spjdstatic void 1315168404Spjdarc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock) 1316168404Spjd{ 1317168404Spjd arc_state_t *old_state = ab->b_state; 1318168404Spjd int64_t refcnt = refcount_count(&ab->b_refcnt); 1319168404Spjd uint64_t from_delta, to_delta; 1320205231Skmacy list_t *list; 1321205231Skmacy kmutex_t *lock; 1322168404Spjd 1323168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 1324260764Savg ASSERT3P(new_state, !=, old_state); 1325168404Spjd ASSERT(refcnt == 0 || ab->b_datacnt > 0); 1326168404Spjd ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state)); 1327219089Spjd ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon); 1328168404Spjd 1329168404Spjd from_delta = to_delta = ab->b_datacnt * ab->b_size; 1330168404Spjd 1331168404Spjd /* 1332168404Spjd * If this buffer is evictable, transfer it from the 1333168404Spjd * old state list to the new state list. 1334168404Spjd */ 1335168404Spjd if (refcnt == 0) { 1336168404Spjd if (old_state != arc_anon) { 1337205231Skmacy int use_mutex; 1338185029Spjd uint64_t *size = &old_state->arcs_lsize[ab->b_type]; 1339168404Spjd 1340205231Skmacy get_buf_info(ab, old_state, &list, &lock); 1341205231Skmacy use_mutex = !MUTEX_HELD(lock); 1342168404Spjd if (use_mutex) 1343205231Skmacy mutex_enter(lock); 1344168404Spjd 1345168404Spjd ASSERT(list_link_active(&ab->b_arc_node)); 1346205231Skmacy list_remove(list, ab); 1347168404Spjd 1348168404Spjd /* 1349168404Spjd * If prefetching out of the ghost cache, 1350219089Spjd * we will have a non-zero datacnt. 1351168404Spjd */ 1352168404Spjd if (GHOST_STATE(old_state) && ab->b_datacnt == 0) { 1353168404Spjd /* ghost elements have a ghost size */ 1354168404Spjd ASSERT(ab->b_buf == NULL); 1355168404Spjd from_delta = ab->b_size; 1356168404Spjd } 1357185029Spjd ASSERT3U(*size, >=, from_delta); 1358185029Spjd atomic_add_64(size, -from_delta); 1359168404Spjd 1360168404Spjd if (use_mutex) 1361205231Skmacy mutex_exit(lock); 1362168404Spjd } 1363168404Spjd if (new_state != arc_anon) { 1364206796Spjd int use_mutex; 1365185029Spjd uint64_t *size = &new_state->arcs_lsize[ab->b_type]; 1366168404Spjd 1367205231Skmacy get_buf_info(ab, new_state, &list, &lock); 1368205231Skmacy use_mutex = !MUTEX_HELD(lock); 1369168404Spjd if (use_mutex) 1370205231Skmacy mutex_enter(lock); 1371168404Spjd 1372205231Skmacy list_insert_head(list, ab); 1373168404Spjd 1374168404Spjd /* ghost elements have a ghost size */ 1375168404Spjd if (GHOST_STATE(new_state)) { 1376168404Spjd ASSERT(ab->b_datacnt == 0); 1377168404Spjd ASSERT(ab->b_buf == NULL); 1378168404Spjd to_delta = ab->b_size; 1379168404Spjd } 1380185029Spjd atomic_add_64(size, to_delta); 1381168404Spjd 1382168404Spjd if (use_mutex) 1383205231Skmacy mutex_exit(lock); 1384168404Spjd } 1385168404Spjd } 1386168404Spjd 1387168404Spjd ASSERT(!BUF_EMPTY(ab)); 1388219089Spjd if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab)) 1389168404Spjd buf_hash_remove(ab); 1390168404Spjd 1391168404Spjd /* adjust state sizes */ 1392168404Spjd if (to_delta) 1393168404Spjd atomic_add_64(&new_state->arcs_size, to_delta); 1394168404Spjd if (from_delta) { 1395168404Spjd ASSERT3U(old_state->arcs_size, >=, from_delta); 1396168404Spjd atomic_add_64(&old_state->arcs_size, -from_delta); 1397168404Spjd } 1398168404Spjd ab->b_state = new_state; 1399185029Spjd 1400185029Spjd /* adjust l2arc hdr stats */ 1401185029Spjd if (new_state == arc_l2c_only) 1402185029Spjd l2arc_hdr_stat_add(); 1403185029Spjd else if (old_state == arc_l2c_only) 1404185029Spjd l2arc_hdr_stat_remove(); 1405168404Spjd} 1406168404Spjd 1407185029Spjdvoid 1408208373Smmarc_space_consume(uint64_t space, arc_space_type_t type) 1409185029Spjd{ 1410208373Smm ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1411208373Smm 1412208373Smm switch (type) { 1413208373Smm case ARC_SPACE_DATA: 1414208373Smm ARCSTAT_INCR(arcstat_data_size, space); 1415208373Smm break; 1416208373Smm case ARC_SPACE_OTHER: 1417208373Smm ARCSTAT_INCR(arcstat_other_size, space); 1418208373Smm break; 1419208373Smm case ARC_SPACE_HDRS: 1420208373Smm ARCSTAT_INCR(arcstat_hdr_size, space); 1421208373Smm break; 1422208373Smm case ARC_SPACE_L2HDRS: 1423208373Smm ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1424208373Smm break; 1425208373Smm } 1426208373Smm 1427185029Spjd atomic_add_64(&arc_meta_used, space); 1428185029Spjd atomic_add_64(&arc_size, space); 1429185029Spjd} 1430185029Spjd 1431185029Spjdvoid 1432208373Smmarc_space_return(uint64_t space, arc_space_type_t type) 1433185029Spjd{ 1434208373Smm ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1435208373Smm 1436208373Smm switch (type) { 1437208373Smm case ARC_SPACE_DATA: 1438208373Smm ARCSTAT_INCR(arcstat_data_size, -space); 1439208373Smm break; 1440208373Smm case ARC_SPACE_OTHER: 1441208373Smm ARCSTAT_INCR(arcstat_other_size, -space); 1442208373Smm break; 1443208373Smm case ARC_SPACE_HDRS: 1444208373Smm ARCSTAT_INCR(arcstat_hdr_size, -space); 1445208373Smm break; 1446208373Smm case ARC_SPACE_L2HDRS: 1447208373Smm ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1448208373Smm break; 1449208373Smm } 1450208373Smm 1451185029Spjd ASSERT(arc_meta_used >= space); 1452185029Spjd if (arc_meta_max < arc_meta_used) 1453185029Spjd arc_meta_max = arc_meta_used; 1454185029Spjd atomic_add_64(&arc_meta_used, -space); 1455185029Spjd ASSERT(arc_size >= space); 1456185029Spjd atomic_add_64(&arc_size, -space); 1457185029Spjd} 1458185029Spjd 1459185029Spjdvoid * 1460185029Spjdarc_data_buf_alloc(uint64_t size) 1461185029Spjd{ 1462185029Spjd if (arc_evict_needed(ARC_BUFC_DATA)) 1463185029Spjd cv_signal(&arc_reclaim_thr_cv); 1464185029Spjd atomic_add_64(&arc_size, size); 1465185029Spjd return (zio_data_buf_alloc(size)); 1466185029Spjd} 1467185029Spjd 1468185029Spjdvoid 1469185029Spjdarc_data_buf_free(void *buf, uint64_t size) 1470185029Spjd{ 1471185029Spjd zio_data_buf_free(buf, size); 1472185029Spjd ASSERT(arc_size >= size); 1473185029Spjd atomic_add_64(&arc_size, -size); 1474185029Spjd} 1475185029Spjd 1476168404Spjdarc_buf_t * 1477168404Spjdarc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1478168404Spjd{ 1479168404Spjd arc_buf_hdr_t *hdr; 1480168404Spjd arc_buf_t *buf; 1481168404Spjd 1482168404Spjd ASSERT3U(size, >, 0); 1483185029Spjd hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1484168404Spjd ASSERT(BUF_EMPTY(hdr)); 1485168404Spjd hdr->b_size = size; 1486168404Spjd hdr->b_type = type; 1487229578Smm hdr->b_spa = spa_load_guid(spa); 1488168404Spjd hdr->b_state = arc_anon; 1489168404Spjd hdr->b_arc_access = 0; 1490185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1491168404Spjd buf->b_hdr = hdr; 1492168404Spjd buf->b_data = NULL; 1493168404Spjd buf->b_efunc = NULL; 1494168404Spjd buf->b_private = NULL; 1495168404Spjd buf->b_next = NULL; 1496168404Spjd hdr->b_buf = buf; 1497168404Spjd arc_get_data_buf(buf); 1498168404Spjd hdr->b_datacnt = 1; 1499168404Spjd hdr->b_flags = 0; 1500168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1501168404Spjd (void) refcount_add(&hdr->b_refcnt, tag); 1502168404Spjd 1503168404Spjd return (buf); 1504168404Spjd} 1505168404Spjd 1506209962Smmstatic char *arc_onloan_tag = "onloan"; 1507209962Smm 1508209962Smm/* 1509209962Smm * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1510209962Smm * flight data by arc_tempreserve_space() until they are "returned". Loaned 1511209962Smm * buffers must be returned to the arc before they can be used by the DMU or 1512209962Smm * freed. 1513209962Smm */ 1514209962Smmarc_buf_t * 1515209962Smmarc_loan_buf(spa_t *spa, int size) 1516209962Smm{ 1517209962Smm arc_buf_t *buf; 1518209962Smm 1519209962Smm buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1520209962Smm 1521209962Smm atomic_add_64(&arc_loaned_bytes, size); 1522209962Smm return (buf); 1523209962Smm} 1524209962Smm 1525209962Smm/* 1526209962Smm * Return a loaned arc buffer to the arc. 1527209962Smm */ 1528209962Smmvoid 1529209962Smmarc_return_buf(arc_buf_t *buf, void *tag) 1530209962Smm{ 1531209962Smm arc_buf_hdr_t *hdr = buf->b_hdr; 1532209962Smm 1533209962Smm ASSERT(buf->b_data != NULL); 1534219089Spjd (void) refcount_add(&hdr->b_refcnt, tag); 1535219089Spjd (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1536209962Smm 1537209962Smm atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1538209962Smm} 1539209962Smm 1540219089Spjd/* Detach an arc_buf from a dbuf (tag) */ 1541219089Spjdvoid 1542219089Spjdarc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1543219089Spjd{ 1544219089Spjd arc_buf_hdr_t *hdr; 1545219089Spjd 1546219089Spjd ASSERT(buf->b_data != NULL); 1547219089Spjd hdr = buf->b_hdr; 1548219089Spjd (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1549219089Spjd (void) refcount_remove(&hdr->b_refcnt, tag); 1550219089Spjd buf->b_efunc = NULL; 1551219089Spjd buf->b_private = NULL; 1552219089Spjd 1553219089Spjd atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1554219089Spjd} 1555219089Spjd 1556168404Spjdstatic arc_buf_t * 1557168404Spjdarc_buf_clone(arc_buf_t *from) 1558168404Spjd{ 1559168404Spjd arc_buf_t *buf; 1560168404Spjd arc_buf_hdr_t *hdr = from->b_hdr; 1561168404Spjd uint64_t size = hdr->b_size; 1562168404Spjd 1563219089Spjd ASSERT(hdr->b_state != arc_anon); 1564219089Spjd 1565185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1566168404Spjd buf->b_hdr = hdr; 1567168404Spjd buf->b_data = NULL; 1568168404Spjd buf->b_efunc = NULL; 1569168404Spjd buf->b_private = NULL; 1570168404Spjd buf->b_next = hdr->b_buf; 1571168404Spjd hdr->b_buf = buf; 1572168404Spjd arc_get_data_buf(buf); 1573168404Spjd bcopy(from->b_data, buf->b_data, size); 1574248547Smm 1575248547Smm /* 1576248547Smm * This buffer already exists in the arc so create a duplicate 1577248547Smm * copy for the caller. If the buffer is associated with user data 1578248547Smm * then track the size and number of duplicates. These stats will be 1579248547Smm * updated as duplicate buffers are created and destroyed. 1580248547Smm */ 1581248547Smm if (hdr->b_type == ARC_BUFC_DATA) { 1582248547Smm ARCSTAT_BUMP(arcstat_duplicate_buffers); 1583248547Smm ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1584248547Smm } 1585168404Spjd hdr->b_datacnt += 1; 1586168404Spjd return (buf); 1587168404Spjd} 1588168404Spjd 1589168404Spjdvoid 1590168404Spjdarc_buf_add_ref(arc_buf_t *buf, void* tag) 1591168404Spjd{ 1592168404Spjd arc_buf_hdr_t *hdr; 1593168404Spjd kmutex_t *hash_lock; 1594168404Spjd 1595168404Spjd /* 1596185029Spjd * Check to see if this buffer is evicted. Callers 1597185029Spjd * must verify b_data != NULL to know if the add_ref 1598185029Spjd * was successful. 1599168404Spjd */ 1600219089Spjd mutex_enter(&buf->b_evict_lock); 1601185029Spjd if (buf->b_data == NULL) { 1602219089Spjd mutex_exit(&buf->b_evict_lock); 1603168404Spjd return; 1604168404Spjd } 1605219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 1606219089Spjd mutex_enter(hash_lock); 1607185029Spjd hdr = buf->b_hdr; 1608219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1609219089Spjd mutex_exit(&buf->b_evict_lock); 1610168404Spjd 1611168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1612168404Spjd add_reference(hdr, hash_lock, tag); 1613208373Smm DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1614168404Spjd arc_access(hdr, hash_lock); 1615168404Spjd mutex_exit(hash_lock); 1616168404Spjd ARCSTAT_BUMP(arcstat_hits); 1617168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 1618168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1619168404Spjd data, metadata, hits); 1620168404Spjd} 1621168404Spjd 1622185029Spjd/* 1623185029Spjd * Free the arc data buffer. If it is an l2arc write in progress, 1624185029Spjd * the buffer is placed on l2arc_free_on_write to be freed later. 1625185029Spjd */ 1626168404Spjdstatic void 1627243674Smmarc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1628185029Spjd{ 1629243674Smm arc_buf_hdr_t *hdr = buf->b_hdr; 1630243674Smm 1631185029Spjd if (HDR_L2_WRITING(hdr)) { 1632185029Spjd l2arc_data_free_t *df; 1633185029Spjd df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1634243674Smm df->l2df_data = buf->b_data; 1635243674Smm df->l2df_size = hdr->b_size; 1636185029Spjd df->l2df_func = free_func; 1637185029Spjd mutex_enter(&l2arc_free_on_write_mtx); 1638185029Spjd list_insert_head(l2arc_free_on_write, df); 1639185029Spjd mutex_exit(&l2arc_free_on_write_mtx); 1640185029Spjd ARCSTAT_BUMP(arcstat_l2_free_on_write); 1641185029Spjd } else { 1642243674Smm free_func(buf->b_data, hdr->b_size); 1643185029Spjd } 1644185029Spjd} 1645185029Spjd 1646185029Spjdstatic void 1647168404Spjdarc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all) 1648168404Spjd{ 1649168404Spjd arc_buf_t **bufp; 1650168404Spjd 1651168404Spjd /* free up data associated with the buf */ 1652168404Spjd if (buf->b_data) { 1653168404Spjd arc_state_t *state = buf->b_hdr->b_state; 1654168404Spjd uint64_t size = buf->b_hdr->b_size; 1655168404Spjd arc_buf_contents_t type = buf->b_hdr->b_type; 1656168404Spjd 1657168404Spjd arc_cksum_verify(buf); 1658243674Smm#ifdef illumos 1659243674Smm arc_buf_unwatch(buf); 1660243674Smm#endif /* illumos */ 1661219089Spjd 1662168404Spjd if (!recycle) { 1663168404Spjd if (type == ARC_BUFC_METADATA) { 1664243674Smm arc_buf_data_free(buf, zio_buf_free); 1665208373Smm arc_space_return(size, ARC_SPACE_DATA); 1666168404Spjd } else { 1667168404Spjd ASSERT(type == ARC_BUFC_DATA); 1668243674Smm arc_buf_data_free(buf, zio_data_buf_free); 1669208373Smm ARCSTAT_INCR(arcstat_data_size, -size); 1670185029Spjd atomic_add_64(&arc_size, -size); 1671168404Spjd } 1672168404Spjd } 1673168404Spjd if (list_link_active(&buf->b_hdr->b_arc_node)) { 1674185029Spjd uint64_t *cnt = &state->arcs_lsize[type]; 1675185029Spjd 1676168404Spjd ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1677168404Spjd ASSERT(state != arc_anon); 1678185029Spjd 1679185029Spjd ASSERT3U(*cnt, >=, size); 1680185029Spjd atomic_add_64(cnt, -size); 1681168404Spjd } 1682168404Spjd ASSERT3U(state->arcs_size, >=, size); 1683168404Spjd atomic_add_64(&state->arcs_size, -size); 1684168404Spjd buf->b_data = NULL; 1685248547Smm 1686248547Smm /* 1687248547Smm * If we're destroying a duplicate buffer make sure 1688248547Smm * that the appropriate statistics are updated. 1689248547Smm */ 1690248547Smm if (buf->b_hdr->b_datacnt > 1 && 1691248547Smm buf->b_hdr->b_type == ARC_BUFC_DATA) { 1692248547Smm ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1693248547Smm ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1694248547Smm } 1695168404Spjd ASSERT(buf->b_hdr->b_datacnt > 0); 1696168404Spjd buf->b_hdr->b_datacnt -= 1; 1697168404Spjd } 1698168404Spjd 1699168404Spjd /* only remove the buf if requested */ 1700168404Spjd if (!all) 1701168404Spjd return; 1702168404Spjd 1703168404Spjd /* remove the buf from the hdr list */ 1704168404Spjd for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1705168404Spjd continue; 1706168404Spjd *bufp = buf->b_next; 1707219089Spjd buf->b_next = NULL; 1708168404Spjd 1709168404Spjd ASSERT(buf->b_efunc == NULL); 1710168404Spjd 1711168404Spjd /* clean up the buf */ 1712168404Spjd buf->b_hdr = NULL; 1713168404Spjd kmem_cache_free(buf_cache, buf); 1714168404Spjd} 1715168404Spjd 1716168404Spjdstatic void 1717168404Spjdarc_hdr_destroy(arc_buf_hdr_t *hdr) 1718168404Spjd{ 1719168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1720168404Spjd ASSERT3P(hdr->b_state, ==, arc_anon); 1721168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1722219089Spjd l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1723168404Spjd 1724219089Spjd if (l2hdr != NULL) { 1725219089Spjd boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1726219089Spjd /* 1727219089Spjd * To prevent arc_free() and l2arc_evict() from 1728219089Spjd * attempting to free the same buffer at the same time, 1729219089Spjd * a FREE_IN_PROGRESS flag is given to arc_free() to 1730219089Spjd * give it priority. l2arc_evict() can't destroy this 1731219089Spjd * header while we are waiting on l2arc_buflist_mtx. 1732219089Spjd * 1733219089Spjd * The hdr may be removed from l2ad_buflist before we 1734219089Spjd * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1735219089Spjd */ 1736219089Spjd if (!buflist_held) { 1737185029Spjd mutex_enter(&l2arc_buflist_mtx); 1738219089Spjd l2hdr = hdr->b_l2hdr; 1739219089Spjd } 1740219089Spjd 1741219089Spjd if (l2hdr != NULL) { 1742251419Ssmh trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1743251419Ssmh hdr->b_size, 0); 1744219089Spjd list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1745219089Spjd ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1746252140Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1747219089Spjd kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1748219089Spjd if (hdr->b_state == arc_l2c_only) 1749219089Spjd l2arc_hdr_stat_remove(); 1750219089Spjd hdr->b_l2hdr = NULL; 1751219089Spjd } 1752219089Spjd 1753219089Spjd if (!buflist_held) 1754185029Spjd mutex_exit(&l2arc_buflist_mtx); 1755185029Spjd } 1756185029Spjd 1757168404Spjd if (!BUF_EMPTY(hdr)) { 1758168404Spjd ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1759219089Spjd buf_discard_identity(hdr); 1760168404Spjd } 1761168404Spjd while (hdr->b_buf) { 1762168404Spjd arc_buf_t *buf = hdr->b_buf; 1763168404Spjd 1764168404Spjd if (buf->b_efunc) { 1765168404Spjd mutex_enter(&arc_eviction_mtx); 1766219089Spjd mutex_enter(&buf->b_evict_lock); 1767168404Spjd ASSERT(buf->b_hdr != NULL); 1768168404Spjd arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1769168404Spjd hdr->b_buf = buf->b_next; 1770168404Spjd buf->b_hdr = &arc_eviction_hdr; 1771168404Spjd buf->b_next = arc_eviction_list; 1772168404Spjd arc_eviction_list = buf; 1773219089Spjd mutex_exit(&buf->b_evict_lock); 1774168404Spjd mutex_exit(&arc_eviction_mtx); 1775168404Spjd } else { 1776168404Spjd arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1777168404Spjd } 1778168404Spjd } 1779168404Spjd if (hdr->b_freeze_cksum != NULL) { 1780168404Spjd kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1781168404Spjd hdr->b_freeze_cksum = NULL; 1782168404Spjd } 1783219089Spjd if (hdr->b_thawed) { 1784219089Spjd kmem_free(hdr->b_thawed, 1); 1785219089Spjd hdr->b_thawed = NULL; 1786219089Spjd } 1787168404Spjd 1788168404Spjd ASSERT(!list_link_active(&hdr->b_arc_node)); 1789168404Spjd ASSERT3P(hdr->b_hash_next, ==, NULL); 1790168404Spjd ASSERT3P(hdr->b_acb, ==, NULL); 1791168404Spjd kmem_cache_free(hdr_cache, hdr); 1792168404Spjd} 1793168404Spjd 1794168404Spjdvoid 1795168404Spjdarc_buf_free(arc_buf_t *buf, void *tag) 1796168404Spjd{ 1797168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 1798168404Spjd int hashed = hdr->b_state != arc_anon; 1799168404Spjd 1800168404Spjd ASSERT(buf->b_efunc == NULL); 1801168404Spjd ASSERT(buf->b_data != NULL); 1802168404Spjd 1803168404Spjd if (hashed) { 1804168404Spjd kmutex_t *hash_lock = HDR_LOCK(hdr); 1805168404Spjd 1806168404Spjd mutex_enter(hash_lock); 1807219089Spjd hdr = buf->b_hdr; 1808219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1809219089Spjd 1810168404Spjd (void) remove_reference(hdr, hash_lock, tag); 1811219089Spjd if (hdr->b_datacnt > 1) { 1812168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1813219089Spjd } else { 1814219089Spjd ASSERT(buf == hdr->b_buf); 1815219089Spjd ASSERT(buf->b_efunc == NULL); 1816168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 1817219089Spjd } 1818168404Spjd mutex_exit(hash_lock); 1819168404Spjd } else if (HDR_IO_IN_PROGRESS(hdr)) { 1820168404Spjd int destroy_hdr; 1821168404Spjd /* 1822168404Spjd * We are in the middle of an async write. Don't destroy 1823168404Spjd * this buffer unless the write completes before we finish 1824168404Spjd * decrementing the reference count. 1825168404Spjd */ 1826168404Spjd mutex_enter(&arc_eviction_mtx); 1827168404Spjd (void) remove_reference(hdr, NULL, tag); 1828168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1829168404Spjd destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1830168404Spjd mutex_exit(&arc_eviction_mtx); 1831168404Spjd if (destroy_hdr) 1832168404Spjd arc_hdr_destroy(hdr); 1833168404Spjd } else { 1834219089Spjd if (remove_reference(hdr, NULL, tag) > 0) 1835168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1836219089Spjd else 1837168404Spjd arc_hdr_destroy(hdr); 1838168404Spjd } 1839168404Spjd} 1840168404Spjd 1841249643Smmboolean_t 1842168404Spjdarc_buf_remove_ref(arc_buf_t *buf, void* tag) 1843168404Spjd{ 1844168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 1845168404Spjd kmutex_t *hash_lock = HDR_LOCK(hdr); 1846249643Smm boolean_t no_callback = (buf->b_efunc == NULL); 1847168404Spjd 1848168404Spjd if (hdr->b_state == arc_anon) { 1849219089Spjd ASSERT(hdr->b_datacnt == 1); 1850168404Spjd arc_buf_free(buf, tag); 1851168404Spjd return (no_callback); 1852168404Spjd } 1853168404Spjd 1854168404Spjd mutex_enter(hash_lock); 1855219089Spjd hdr = buf->b_hdr; 1856219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1857168404Spjd ASSERT(hdr->b_state != arc_anon); 1858168404Spjd ASSERT(buf->b_data != NULL); 1859168404Spjd 1860168404Spjd (void) remove_reference(hdr, hash_lock, tag); 1861168404Spjd if (hdr->b_datacnt > 1) { 1862168404Spjd if (no_callback) 1863168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1864168404Spjd } else if (no_callback) { 1865168404Spjd ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1866219089Spjd ASSERT(buf->b_efunc == NULL); 1867168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 1868168404Spjd } 1869168404Spjd ASSERT(no_callback || hdr->b_datacnt > 1 || 1870168404Spjd refcount_is_zero(&hdr->b_refcnt)); 1871168404Spjd mutex_exit(hash_lock); 1872168404Spjd return (no_callback); 1873168404Spjd} 1874168404Spjd 1875168404Spjdint 1876168404Spjdarc_buf_size(arc_buf_t *buf) 1877168404Spjd{ 1878168404Spjd return (buf->b_hdr->b_size); 1879168404Spjd} 1880168404Spjd 1881168404Spjd/* 1882248547Smm * Called from the DMU to determine if the current buffer should be 1883248547Smm * evicted. In order to ensure proper locking, the eviction must be initiated 1884248547Smm * from the DMU. Return true if the buffer is associated with user data and 1885248547Smm * duplicate buffers still exist. 1886248547Smm */ 1887248547Smmboolean_t 1888248547Smmarc_buf_eviction_needed(arc_buf_t *buf) 1889248547Smm{ 1890248547Smm arc_buf_hdr_t *hdr; 1891248547Smm boolean_t evict_needed = B_FALSE; 1892248547Smm 1893248547Smm if (zfs_disable_dup_eviction) 1894248547Smm return (B_FALSE); 1895248547Smm 1896248547Smm mutex_enter(&buf->b_evict_lock); 1897248547Smm hdr = buf->b_hdr; 1898248547Smm if (hdr == NULL) { 1899248547Smm /* 1900248547Smm * We are in arc_do_user_evicts(); let that function 1901248547Smm * perform the eviction. 1902248547Smm */ 1903248547Smm ASSERT(buf->b_data == NULL); 1904248547Smm mutex_exit(&buf->b_evict_lock); 1905248547Smm return (B_FALSE); 1906248547Smm } else if (buf->b_data == NULL) { 1907248547Smm /* 1908248547Smm * We have already been added to the arc eviction list; 1909248547Smm * recommend eviction. 1910248547Smm */ 1911248547Smm ASSERT3P(hdr, ==, &arc_eviction_hdr); 1912248547Smm mutex_exit(&buf->b_evict_lock); 1913248547Smm return (B_TRUE); 1914248547Smm } 1915248547Smm 1916248547Smm if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 1917248547Smm evict_needed = B_TRUE; 1918248547Smm 1919248547Smm mutex_exit(&buf->b_evict_lock); 1920248547Smm return (evict_needed); 1921248547Smm} 1922248547Smm 1923248547Smm/* 1924168404Spjd * Evict buffers from list until we've removed the specified number of 1925168404Spjd * bytes. Move the removed buffers to the appropriate evict state. 1926168404Spjd * If the recycle flag is set, then attempt to "recycle" a buffer: 1927168404Spjd * - look for a buffer to evict that is `bytes' long. 1928168404Spjd * - return the data block from this buffer rather than freeing it. 1929168404Spjd * This flag is used by callers that are trying to make space for a 1930168404Spjd * new buffer in a full arc cache. 1931185029Spjd * 1932185029Spjd * This function makes a "best effort". It skips over any buffers 1933185029Spjd * it can't get a hash_lock on, and so may not catch all candidates. 1934185029Spjd * It may also return without evicting as much space as requested. 1935168404Spjd */ 1936168404Spjdstatic void * 1937209962Smmarc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 1938168404Spjd arc_buf_contents_t type) 1939168404Spjd{ 1940168404Spjd arc_state_t *evicted_state; 1941168404Spjd uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 1942205231Skmacy int64_t bytes_remaining; 1943168404Spjd arc_buf_hdr_t *ab, *ab_prev = NULL; 1944205231Skmacy list_t *evicted_list, *list, *evicted_list_start, *list_start; 1945205231Skmacy kmutex_t *lock, *evicted_lock; 1946168404Spjd kmutex_t *hash_lock; 1947168404Spjd boolean_t have_lock; 1948168404Spjd void *stolen = NULL; 1949260764Savg arc_buf_hdr_t marker = { 0 }; 1950260764Savg int count = 0; 1951205231Skmacy static int evict_metadata_offset, evict_data_offset; 1952260764Savg int i, idx, offset, list_count, lists; 1953168404Spjd 1954168404Spjd ASSERT(state == arc_mru || state == arc_mfu); 1955168404Spjd 1956168404Spjd evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 1957206796Spjd 1958205231Skmacy if (type == ARC_BUFC_METADATA) { 1959205231Skmacy offset = 0; 1960205231Skmacy list_count = ARC_BUFC_NUMMETADATALISTS; 1961205231Skmacy list_start = &state->arcs_lists[0]; 1962205231Skmacy evicted_list_start = &evicted_state->arcs_lists[0]; 1963205231Skmacy idx = evict_metadata_offset; 1964205231Skmacy } else { 1965205231Skmacy offset = ARC_BUFC_NUMMETADATALISTS; 1966205231Skmacy list_start = &state->arcs_lists[offset]; 1967205231Skmacy evicted_list_start = &evicted_state->arcs_lists[offset]; 1968205231Skmacy list_count = ARC_BUFC_NUMDATALISTS; 1969205231Skmacy idx = evict_data_offset; 1970205231Skmacy } 1971205231Skmacy bytes_remaining = evicted_state->arcs_lsize[type]; 1972260764Savg lists = 0; 1973206796Spjd 1974205231Skmacyevict_start: 1975205231Skmacy list = &list_start[idx]; 1976205231Skmacy evicted_list = &evicted_list_start[idx]; 1977205231Skmacy lock = ARCS_LOCK(state, (offset + idx)); 1978206796Spjd evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 1979168404Spjd 1980205231Skmacy mutex_enter(lock); 1981205231Skmacy mutex_enter(evicted_lock); 1982205231Skmacy 1983185029Spjd for (ab = list_tail(list); ab; ab = ab_prev) { 1984185029Spjd ab_prev = list_prev(list, ab); 1985205231Skmacy bytes_remaining -= (ab->b_size * ab->b_datacnt); 1986168404Spjd /* prefetch buffers have a minimum lifespan */ 1987168404Spjd if (HDR_IO_IN_PROGRESS(ab) || 1988185029Spjd (spa && ab->b_spa != spa) || 1989168404Spjd (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 1990219089Spjd ddi_get_lbolt() - ab->b_arc_access < 1991219089Spjd arc_min_prefetch_lifespan)) { 1992168404Spjd skipped++; 1993168404Spjd continue; 1994168404Spjd } 1995168404Spjd /* "lookahead" for better eviction candidate */ 1996168404Spjd if (recycle && ab->b_size != bytes && 1997168404Spjd ab_prev && ab_prev->b_size == bytes) 1998168404Spjd continue; 1999260764Savg 2000260764Savg /* ignore markers */ 2001260764Savg if (ab->b_spa == 0) 2002260764Savg continue; 2003260764Savg 2004260764Savg /* 2005260764Savg * It may take a long time to evict all the bufs requested. 2006260764Savg * To avoid blocking all arc activity, periodically drop 2007260764Savg * the arcs_mtx and give other threads a chance to run 2008260764Savg * before reacquiring the lock. 2009260764Savg * 2010260764Savg * If we are looking for a buffer to recycle, we are in 2011260764Savg * the hot code path, so don't sleep. 2012260764Savg */ 2013260764Savg if (!recycle && count++ > arc_evict_iterations) { 2014260764Savg list_insert_after(list, ab, &marker); 2015260764Savg mutex_exit(evicted_lock); 2016260764Savg mutex_exit(lock); 2017260764Savg kpreempt(KPREEMPT_SYNC); 2018260764Savg mutex_enter(lock); 2019260764Savg mutex_enter(evicted_lock); 2020260764Savg ab_prev = list_prev(list, &marker); 2021260764Savg list_remove(list, &marker); 2022260764Savg count = 0; 2023260764Savg continue; 2024260764Savg } 2025260764Savg 2026168404Spjd hash_lock = HDR_LOCK(ab); 2027168404Spjd have_lock = MUTEX_HELD(hash_lock); 2028168404Spjd if (have_lock || mutex_tryenter(hash_lock)) { 2029243674Smm ASSERT0(refcount_count(&ab->b_refcnt)); 2030168404Spjd ASSERT(ab->b_datacnt > 0); 2031168404Spjd while (ab->b_buf) { 2032168404Spjd arc_buf_t *buf = ab->b_buf; 2033219089Spjd if (!mutex_tryenter(&buf->b_evict_lock)) { 2034185029Spjd missed += 1; 2035185029Spjd break; 2036185029Spjd } 2037168404Spjd if (buf->b_data) { 2038168404Spjd bytes_evicted += ab->b_size; 2039168404Spjd if (recycle && ab->b_type == type && 2040185029Spjd ab->b_size == bytes && 2041185029Spjd !HDR_L2_WRITING(ab)) { 2042168404Spjd stolen = buf->b_data; 2043168404Spjd recycle = FALSE; 2044168404Spjd } 2045168404Spjd } 2046168404Spjd if (buf->b_efunc) { 2047168404Spjd mutex_enter(&arc_eviction_mtx); 2048168404Spjd arc_buf_destroy(buf, 2049168404Spjd buf->b_data == stolen, FALSE); 2050168404Spjd ab->b_buf = buf->b_next; 2051168404Spjd buf->b_hdr = &arc_eviction_hdr; 2052168404Spjd buf->b_next = arc_eviction_list; 2053168404Spjd arc_eviction_list = buf; 2054168404Spjd mutex_exit(&arc_eviction_mtx); 2055219089Spjd mutex_exit(&buf->b_evict_lock); 2056168404Spjd } else { 2057219089Spjd mutex_exit(&buf->b_evict_lock); 2058168404Spjd arc_buf_destroy(buf, 2059168404Spjd buf->b_data == stolen, TRUE); 2060168404Spjd } 2061168404Spjd } 2062208373Smm 2063208373Smm if (ab->b_l2hdr) { 2064208373Smm ARCSTAT_INCR(arcstat_evict_l2_cached, 2065208373Smm ab->b_size); 2066208373Smm } else { 2067208373Smm if (l2arc_write_eligible(ab->b_spa, ab)) { 2068208373Smm ARCSTAT_INCR(arcstat_evict_l2_eligible, 2069208373Smm ab->b_size); 2070208373Smm } else { 2071208373Smm ARCSTAT_INCR( 2072208373Smm arcstat_evict_l2_ineligible, 2073208373Smm ab->b_size); 2074208373Smm } 2075208373Smm } 2076208373Smm 2077185029Spjd if (ab->b_datacnt == 0) { 2078185029Spjd arc_change_state(evicted_state, ab, hash_lock); 2079185029Spjd ASSERT(HDR_IN_HASH_TABLE(ab)); 2080185029Spjd ab->b_flags |= ARC_IN_HASH_TABLE; 2081185029Spjd ab->b_flags &= ~ARC_BUF_AVAILABLE; 2082185029Spjd DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 2083185029Spjd } 2084168404Spjd if (!have_lock) 2085168404Spjd mutex_exit(hash_lock); 2086168404Spjd if (bytes >= 0 && bytes_evicted >= bytes) 2087168404Spjd break; 2088205231Skmacy if (bytes_remaining > 0) { 2089205231Skmacy mutex_exit(evicted_lock); 2090205231Skmacy mutex_exit(lock); 2091206796Spjd idx = ((idx + 1) & (list_count - 1)); 2092260764Savg lists++; 2093205231Skmacy goto evict_start; 2094205231Skmacy } 2095168404Spjd } else { 2096168404Spjd missed += 1; 2097168404Spjd } 2098168404Spjd } 2099168404Spjd 2100205231Skmacy mutex_exit(evicted_lock); 2101205231Skmacy mutex_exit(lock); 2102206796Spjd 2103206796Spjd idx = ((idx + 1) & (list_count - 1)); 2104260764Savg lists++; 2105168404Spjd 2106205231Skmacy if (bytes_evicted < bytes) { 2107260764Savg if (lists < list_count) 2108205231Skmacy goto evict_start; 2109205231Skmacy else 2110205231Skmacy dprintf("only evicted %lld bytes from %x", 2111205231Skmacy (longlong_t)bytes_evicted, state); 2112205231Skmacy } 2113206796Spjd if (type == ARC_BUFC_METADATA) 2114205231Skmacy evict_metadata_offset = idx; 2115205231Skmacy else 2116205231Skmacy evict_data_offset = idx; 2117206796Spjd 2118168404Spjd if (skipped) 2119168404Spjd ARCSTAT_INCR(arcstat_evict_skip, skipped); 2120168404Spjd 2121168404Spjd if (missed) 2122168404Spjd ARCSTAT_INCR(arcstat_mutex_miss, missed); 2123168404Spjd 2124185029Spjd /* 2125260764Savg * Note: we have just evicted some data into the ghost state, 2126260764Savg * potentially putting the ghost size over the desired size. Rather 2127260764Savg * that evicting from the ghost list in this hot code path, leave 2128260764Savg * this chore to the arc_reclaim_thread(). 2129185029Spjd */ 2130185029Spjd 2131205231Skmacy if (stolen) 2132205231Skmacy ARCSTAT_BUMP(arcstat_stolen); 2133168404Spjd return (stolen); 2134168404Spjd} 2135168404Spjd 2136168404Spjd/* 2137168404Spjd * Remove buffers from list until we've removed the specified number of 2138168404Spjd * bytes. Destroy the buffers that are removed. 2139168404Spjd */ 2140168404Spjdstatic void 2141209962Smmarc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2142168404Spjd{ 2143168404Spjd arc_buf_hdr_t *ab, *ab_prev; 2144219089Spjd arc_buf_hdr_t marker = { 0 }; 2145205231Skmacy list_t *list, *list_start; 2146205231Skmacy kmutex_t *hash_lock, *lock; 2147168404Spjd uint64_t bytes_deleted = 0; 2148168404Spjd uint64_t bufs_skipped = 0; 2149260764Savg int count = 0; 2150205231Skmacy static int evict_offset; 2151205231Skmacy int list_count, idx = evict_offset; 2152260764Savg int offset, lists = 0; 2153168404Spjd 2154168404Spjd ASSERT(GHOST_STATE(state)); 2155205231Skmacy 2156205231Skmacy /* 2157205231Skmacy * data lists come after metadata lists 2158205231Skmacy */ 2159205231Skmacy list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2160205231Skmacy list_count = ARC_BUFC_NUMDATALISTS; 2161205231Skmacy offset = ARC_BUFC_NUMMETADATALISTS; 2162206796Spjd 2163205231Skmacyevict_start: 2164205231Skmacy list = &list_start[idx]; 2165205231Skmacy lock = ARCS_LOCK(state, idx + offset); 2166205231Skmacy 2167205231Skmacy mutex_enter(lock); 2168185029Spjd for (ab = list_tail(list); ab; ab = ab_prev) { 2169185029Spjd ab_prev = list_prev(list, ab); 2170260764Savg if (ab->b_type > ARC_BUFC_NUMTYPES) 2171260764Savg panic("invalid ab=%p", (void *)ab); 2172185029Spjd if (spa && ab->b_spa != spa) 2173185029Spjd continue; 2174219089Spjd 2175219089Spjd /* ignore markers */ 2176219089Spjd if (ab->b_spa == 0) 2177219089Spjd continue; 2178219089Spjd 2179168404Spjd hash_lock = HDR_LOCK(ab); 2180219089Spjd /* caller may be trying to modify this buffer, skip it */ 2181219089Spjd if (MUTEX_HELD(hash_lock)) 2182219089Spjd continue; 2183260764Savg 2184260764Savg /* 2185260764Savg * It may take a long time to evict all the bufs requested. 2186260764Savg * To avoid blocking all arc activity, periodically drop 2187260764Savg * the arcs_mtx and give other threads a chance to run 2188260764Savg * before reacquiring the lock. 2189260764Savg */ 2190260764Savg if (count++ > arc_evict_iterations) { 2191260764Savg list_insert_after(list, ab, &marker); 2192260764Savg mutex_exit(lock); 2193260764Savg kpreempt(KPREEMPT_SYNC); 2194260764Savg mutex_enter(lock); 2195260764Savg ab_prev = list_prev(list, &marker); 2196260764Savg list_remove(list, &marker); 2197260764Savg count = 0; 2198260764Savg continue; 2199260764Savg } 2200168404Spjd if (mutex_tryenter(hash_lock)) { 2201168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(ab)); 2202168404Spjd ASSERT(ab->b_buf == NULL); 2203168404Spjd ARCSTAT_BUMP(arcstat_deleted); 2204168404Spjd bytes_deleted += ab->b_size; 2205185029Spjd 2206185029Spjd if (ab->b_l2hdr != NULL) { 2207185029Spjd /* 2208185029Spjd * This buffer is cached on the 2nd Level ARC; 2209185029Spjd * don't destroy the header. 2210185029Spjd */ 2211185029Spjd arc_change_state(arc_l2c_only, ab, hash_lock); 2212185029Spjd mutex_exit(hash_lock); 2213185029Spjd } else { 2214185029Spjd arc_change_state(arc_anon, ab, hash_lock); 2215185029Spjd mutex_exit(hash_lock); 2216185029Spjd arc_hdr_destroy(ab); 2217185029Spjd } 2218185029Spjd 2219168404Spjd DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 2220168404Spjd if (bytes >= 0 && bytes_deleted >= bytes) 2221168404Spjd break; 2222219089Spjd } else if (bytes < 0) { 2223219089Spjd /* 2224219089Spjd * Insert a list marker and then wait for the 2225219089Spjd * hash lock to become available. Once its 2226219089Spjd * available, restart from where we left off. 2227219089Spjd */ 2228219089Spjd list_insert_after(list, ab, &marker); 2229219089Spjd mutex_exit(lock); 2230219089Spjd mutex_enter(hash_lock); 2231219089Spjd mutex_exit(hash_lock); 2232219089Spjd mutex_enter(lock); 2233219089Spjd ab_prev = list_prev(list, &marker); 2234219089Spjd list_remove(list, &marker); 2235260764Savg } else { 2236168404Spjd bufs_skipped += 1; 2237260764Savg } 2238260764Savg 2239168404Spjd } 2240205231Skmacy mutex_exit(lock); 2241206796Spjd idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2242260764Savg lists++; 2243206796Spjd 2244260764Savg if (lists < list_count) 2245205231Skmacy goto evict_start; 2246206796Spjd 2247205231Skmacy evict_offset = idx; 2248205231Skmacy if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2249185029Spjd (bytes < 0 || bytes_deleted < bytes)) { 2250205231Skmacy list_start = &state->arcs_lists[0]; 2251205231Skmacy list_count = ARC_BUFC_NUMMETADATALISTS; 2252260764Savg offset = lists = 0; 2253205231Skmacy goto evict_start; 2254185029Spjd } 2255185029Spjd 2256168404Spjd if (bufs_skipped) { 2257168404Spjd ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2258168404Spjd ASSERT(bytes >= 0); 2259168404Spjd } 2260168404Spjd 2261168404Spjd if (bytes_deleted < bytes) 2262168404Spjd dprintf("only deleted %lld bytes from %p", 2263168404Spjd (longlong_t)bytes_deleted, state); 2264168404Spjd} 2265168404Spjd 2266168404Spjdstatic void 2267168404Spjdarc_adjust(void) 2268168404Spjd{ 2269208373Smm int64_t adjustment, delta; 2270168404Spjd 2271208373Smm /* 2272208373Smm * Adjust MRU size 2273208373Smm */ 2274168404Spjd 2275209275Smm adjustment = MIN((int64_t)(arc_size - arc_c), 2276209275Smm (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2277209275Smm arc_p)); 2278208373Smm 2279208373Smm if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2280208373Smm delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2281209962Smm (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2282208373Smm adjustment -= delta; 2283168404Spjd } 2284168404Spjd 2285208373Smm if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2286208373Smm delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2287209962Smm (void) arc_evict(arc_mru, 0, delta, FALSE, 2288185029Spjd ARC_BUFC_METADATA); 2289185029Spjd } 2290185029Spjd 2291208373Smm /* 2292208373Smm * Adjust MFU size 2293208373Smm */ 2294168404Spjd 2295208373Smm adjustment = arc_size - arc_c; 2296208373Smm 2297208373Smm if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2298208373Smm delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2299209962Smm (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2300208373Smm adjustment -= delta; 2301168404Spjd } 2302168404Spjd 2303208373Smm if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2304208373Smm int64_t delta = MIN(adjustment, 2305208373Smm arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2306209962Smm (void) arc_evict(arc_mfu, 0, delta, FALSE, 2307208373Smm ARC_BUFC_METADATA); 2308208373Smm } 2309168404Spjd 2310208373Smm /* 2311208373Smm * Adjust ghost lists 2312208373Smm */ 2313168404Spjd 2314208373Smm adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2315168404Spjd 2316208373Smm if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2317208373Smm delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2318209962Smm arc_evict_ghost(arc_mru_ghost, 0, delta); 2319208373Smm } 2320185029Spjd 2321208373Smm adjustment = 2322208373Smm arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2323208373Smm 2324208373Smm if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2325208373Smm delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2326209962Smm arc_evict_ghost(arc_mfu_ghost, 0, delta); 2327168404Spjd } 2328168404Spjd} 2329168404Spjd 2330168404Spjdstatic void 2331168404Spjdarc_do_user_evicts(void) 2332168404Spjd{ 2333191903Skmacy static arc_buf_t *tmp_arc_eviction_list; 2334191903Skmacy 2335191903Skmacy /* 2336191903Skmacy * Move list over to avoid LOR 2337191903Skmacy */ 2338206796Spjdrestart: 2339168404Spjd mutex_enter(&arc_eviction_mtx); 2340191903Skmacy tmp_arc_eviction_list = arc_eviction_list; 2341191903Skmacy arc_eviction_list = NULL; 2342191903Skmacy mutex_exit(&arc_eviction_mtx); 2343191903Skmacy 2344191903Skmacy while (tmp_arc_eviction_list != NULL) { 2345191903Skmacy arc_buf_t *buf = tmp_arc_eviction_list; 2346191903Skmacy tmp_arc_eviction_list = buf->b_next; 2347219089Spjd mutex_enter(&buf->b_evict_lock); 2348168404Spjd buf->b_hdr = NULL; 2349219089Spjd mutex_exit(&buf->b_evict_lock); 2350168404Spjd 2351168404Spjd if (buf->b_efunc != NULL) 2352168404Spjd VERIFY(buf->b_efunc(buf) == 0); 2353168404Spjd 2354168404Spjd buf->b_efunc = NULL; 2355168404Spjd buf->b_private = NULL; 2356168404Spjd kmem_cache_free(buf_cache, buf); 2357168404Spjd } 2358191903Skmacy 2359191903Skmacy if (arc_eviction_list != NULL) 2360191903Skmacy goto restart; 2361168404Spjd} 2362168404Spjd 2363168404Spjd/* 2364185029Spjd * Flush all *evictable* data from the cache for the given spa. 2365168404Spjd * NOTE: this will not touch "active" (i.e. referenced) data. 2366168404Spjd */ 2367168404Spjdvoid 2368185029Spjdarc_flush(spa_t *spa) 2369168404Spjd{ 2370209962Smm uint64_t guid = 0; 2371209962Smm 2372209962Smm if (spa) 2373229578Smm guid = spa_load_guid(spa); 2374209962Smm 2375205231Skmacy while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2376209962Smm (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2377185029Spjd if (spa) 2378185029Spjd break; 2379185029Spjd } 2380205231Skmacy while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2381209962Smm (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2382185029Spjd if (spa) 2383185029Spjd break; 2384185029Spjd } 2385205231Skmacy while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2386209962Smm (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2387185029Spjd if (spa) 2388185029Spjd break; 2389185029Spjd } 2390205231Skmacy while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2391209962Smm (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2392185029Spjd if (spa) 2393185029Spjd break; 2394185029Spjd } 2395168404Spjd 2396209962Smm arc_evict_ghost(arc_mru_ghost, guid, -1); 2397209962Smm arc_evict_ghost(arc_mfu_ghost, guid, -1); 2398168404Spjd 2399168404Spjd mutex_enter(&arc_reclaim_thr_lock); 2400168404Spjd arc_do_user_evicts(); 2401168404Spjd mutex_exit(&arc_reclaim_thr_lock); 2402185029Spjd ASSERT(spa || arc_eviction_list == NULL); 2403168404Spjd} 2404168404Spjd 2405168404Spjdvoid 2406168404Spjdarc_shrink(void) 2407168404Spjd{ 2408168404Spjd if (arc_c > arc_c_min) { 2409168404Spjd uint64_t to_free; 2410168404Spjd 2411168404Spjd#ifdef _KERNEL 2412168404Spjd to_free = arc_c >> arc_shrink_shift; 2413168404Spjd#else 2414168404Spjd to_free = arc_c >> arc_shrink_shift; 2415168404Spjd#endif 2416168404Spjd if (arc_c > arc_c_min + to_free) 2417168404Spjd atomic_add_64(&arc_c, -to_free); 2418168404Spjd else 2419168404Spjd arc_c = arc_c_min; 2420168404Spjd 2421168404Spjd atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2422168404Spjd if (arc_c > arc_size) 2423168404Spjd arc_c = MAX(arc_size, arc_c_min); 2424168404Spjd if (arc_p > arc_c) 2425168404Spjd arc_p = (arc_c >> 1); 2426168404Spjd ASSERT(arc_c >= arc_c_min); 2427168404Spjd ASSERT((int64_t)arc_p >= 0); 2428168404Spjd } 2429168404Spjd 2430168404Spjd if (arc_size > arc_c) 2431168404Spjd arc_adjust(); 2432168404Spjd} 2433168404Spjd 2434185029Spjdstatic int needfree = 0; 2435168404Spjd 2436168404Spjdstatic int 2437168404Spjdarc_reclaim_needed(void) 2438168404Spjd{ 2439168404Spjd 2440168404Spjd#ifdef _KERNEL 2441219089Spjd 2442197816Skmacy if (needfree) 2443197816Skmacy return (1); 2444168404Spjd 2445191902Skmacy /* 2446212780Savg * Cooperate with pagedaemon when it's time for it to scan 2447212780Savg * and reclaim some pages. 2448191902Skmacy */ 2449212783Savg if (vm_paging_needed()) 2450191902Skmacy return (1); 2451191902Skmacy 2452219089Spjd#ifdef sun 2453168404Spjd /* 2454185029Spjd * take 'desfree' extra pages, so we reclaim sooner, rather than later 2455185029Spjd */ 2456185029Spjd extra = desfree; 2457185029Spjd 2458185029Spjd /* 2459185029Spjd * check that we're out of range of the pageout scanner. It starts to 2460185029Spjd * schedule paging if freemem is less than lotsfree and needfree. 2461185029Spjd * lotsfree is the high-water mark for pageout, and needfree is the 2462185029Spjd * number of needed free pages. We add extra pages here to make sure 2463185029Spjd * the scanner doesn't start up while we're freeing memory. 2464185029Spjd */ 2465185029Spjd if (freemem < lotsfree + needfree + extra) 2466185029Spjd return (1); 2467185029Spjd 2468185029Spjd /* 2469168404Spjd * check to make sure that swapfs has enough space so that anon 2470185029Spjd * reservations can still succeed. anon_resvmem() checks that the 2471168404Spjd * availrmem is greater than swapfs_minfree, and the number of reserved 2472168404Spjd * swap pages. We also add a bit of extra here just to prevent 2473168404Spjd * circumstances from getting really dire. 2474168404Spjd */ 2475168404Spjd if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2476168404Spjd return (1); 2477168404Spjd 2478168404Spjd#if defined(__i386) 2479168404Spjd /* 2480168404Spjd * If we're on an i386 platform, it's possible that we'll exhaust the 2481168404Spjd * kernel heap space before we ever run out of available physical 2482168404Spjd * memory. Most checks of the size of the heap_area compare against 2483168404Spjd * tune.t_minarmem, which is the minimum available real memory that we 2484168404Spjd * can have in the system. However, this is generally fixed at 25 pages 2485168404Spjd * which is so low that it's useless. In this comparison, we seek to 2486168404Spjd * calculate the total heap-size, and reclaim if more than 3/4ths of the 2487185029Spjd * heap is allocated. (Or, in the calculation, if less than 1/4th is 2488168404Spjd * free) 2489168404Spjd */ 2490168404Spjd if (btop(vmem_size(heap_arena, VMEM_FREE)) < 2491168404Spjd (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2)) 2492168404Spjd return (1); 2493168404Spjd#endif 2494219089Spjd#else /* !sun */ 2495175633Spjd if (kmem_used() > (kmem_size() * 3) / 4) 2496168404Spjd return (1); 2497219089Spjd#endif /* sun */ 2498168404Spjd 2499168404Spjd#else 2500168404Spjd if (spa_get_random(100) == 0) 2501168404Spjd return (1); 2502168404Spjd#endif 2503168404Spjd return (0); 2504168404Spjd} 2505168404Spjd 2506208454Spjdextern kmem_cache_t *zio_buf_cache[]; 2507208454Spjdextern kmem_cache_t *zio_data_buf_cache[]; 2508208454Spjd 2509168404Spjdstatic void 2510168404Spjdarc_kmem_reap_now(arc_reclaim_strategy_t strat) 2511168404Spjd{ 2512168404Spjd size_t i; 2513168404Spjd kmem_cache_t *prev_cache = NULL; 2514168404Spjd kmem_cache_t *prev_data_cache = NULL; 2515168404Spjd 2516168404Spjd#ifdef _KERNEL 2517185029Spjd if (arc_meta_used >= arc_meta_limit) { 2518185029Spjd /* 2519185029Spjd * We are exceeding our meta-data cache limit. 2520185029Spjd * Purge some DNLC entries to release holds on meta-data. 2521185029Spjd */ 2522185029Spjd dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2523185029Spjd } 2524168404Spjd#if defined(__i386) 2525168404Spjd /* 2526168404Spjd * Reclaim unused memory from all kmem caches. 2527168404Spjd */ 2528168404Spjd kmem_reap(); 2529168404Spjd#endif 2530168404Spjd#endif 2531168404Spjd 2532168404Spjd /* 2533185029Spjd * An aggressive reclamation will shrink the cache size as well as 2534168404Spjd * reap free buffers from the arc kmem caches. 2535168404Spjd */ 2536168404Spjd if (strat == ARC_RECLAIM_AGGR) 2537168404Spjd arc_shrink(); 2538168404Spjd 2539168404Spjd for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2540168404Spjd if (zio_buf_cache[i] != prev_cache) { 2541168404Spjd prev_cache = zio_buf_cache[i]; 2542168404Spjd kmem_cache_reap_now(zio_buf_cache[i]); 2543168404Spjd } 2544168404Spjd if (zio_data_buf_cache[i] != prev_data_cache) { 2545168404Spjd prev_data_cache = zio_data_buf_cache[i]; 2546168404Spjd kmem_cache_reap_now(zio_data_buf_cache[i]); 2547168404Spjd } 2548168404Spjd } 2549168404Spjd kmem_cache_reap_now(buf_cache); 2550168404Spjd kmem_cache_reap_now(hdr_cache); 2551168404Spjd} 2552168404Spjd 2553168404Spjdstatic void 2554168404Spjdarc_reclaim_thread(void *dummy __unused) 2555168404Spjd{ 2556168404Spjd clock_t growtime = 0; 2557168404Spjd arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2558168404Spjd callb_cpr_t cpr; 2559168404Spjd 2560168404Spjd CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2561168404Spjd 2562168404Spjd mutex_enter(&arc_reclaim_thr_lock); 2563168404Spjd while (arc_thread_exit == 0) { 2564168404Spjd if (arc_reclaim_needed()) { 2565168404Spjd 2566168404Spjd if (arc_no_grow) { 2567168404Spjd if (last_reclaim == ARC_RECLAIM_CONS) { 2568168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2569168404Spjd } else { 2570168404Spjd last_reclaim = ARC_RECLAIM_CONS; 2571168404Spjd } 2572168404Spjd } else { 2573168404Spjd arc_no_grow = TRUE; 2574168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2575168404Spjd membar_producer(); 2576168404Spjd } 2577168404Spjd 2578168404Spjd /* reset the growth delay for every reclaim */ 2579219089Spjd growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2580168404Spjd 2581185029Spjd if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2582168404Spjd /* 2583185029Spjd * If needfree is TRUE our vm_lowmem hook 2584168404Spjd * was called and in that case we must free some 2585168404Spjd * memory, so switch to aggressive mode. 2586168404Spjd */ 2587168404Spjd arc_no_grow = TRUE; 2588168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2589168404Spjd } 2590168404Spjd arc_kmem_reap_now(last_reclaim); 2591185029Spjd arc_warm = B_TRUE; 2592185029Spjd 2593219089Spjd } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2594168404Spjd arc_no_grow = FALSE; 2595168404Spjd } 2596168404Spjd 2597209275Smm arc_adjust(); 2598168404Spjd 2599168404Spjd if (arc_eviction_list != NULL) 2600168404Spjd arc_do_user_evicts(); 2601168404Spjd 2602211762Savg#ifdef _KERNEL 2603211762Savg if (needfree) { 2604185029Spjd needfree = 0; 2605185029Spjd wakeup(&needfree); 2606211762Savg } 2607168404Spjd#endif 2608168404Spjd 2609168404Spjd /* block until needed, or one second, whichever is shorter */ 2610168404Spjd CALLB_CPR_SAFE_BEGIN(&cpr); 2611168404Spjd (void) cv_timedwait(&arc_reclaim_thr_cv, 2612168404Spjd &arc_reclaim_thr_lock, hz); 2613168404Spjd CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2614168404Spjd } 2615168404Spjd 2616168404Spjd arc_thread_exit = 0; 2617168404Spjd cv_broadcast(&arc_reclaim_thr_cv); 2618168404Spjd CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2619168404Spjd thread_exit(); 2620168404Spjd} 2621168404Spjd 2622168404Spjd/* 2623168404Spjd * Adapt arc info given the number of bytes we are trying to add and 2624168404Spjd * the state that we are comming from. This function is only called 2625168404Spjd * when we are adding new content to the cache. 2626168404Spjd */ 2627168404Spjdstatic void 2628168404Spjdarc_adapt(int bytes, arc_state_t *state) 2629168404Spjd{ 2630168404Spjd int mult; 2631208373Smm uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2632168404Spjd 2633185029Spjd if (state == arc_l2c_only) 2634185029Spjd return; 2635185029Spjd 2636168404Spjd ASSERT(bytes > 0); 2637168404Spjd /* 2638168404Spjd * Adapt the target size of the MRU list: 2639168404Spjd * - if we just hit in the MRU ghost list, then increase 2640168404Spjd * the target size of the MRU list. 2641168404Spjd * - if we just hit in the MFU ghost list, then increase 2642168404Spjd * the target size of the MFU list by decreasing the 2643168404Spjd * target size of the MRU list. 2644168404Spjd */ 2645168404Spjd if (state == arc_mru_ghost) { 2646168404Spjd mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2647168404Spjd 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2648209275Smm mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2649168404Spjd 2650208373Smm arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2651168404Spjd } else if (state == arc_mfu_ghost) { 2652208373Smm uint64_t delta; 2653208373Smm 2654168404Spjd mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2655168404Spjd 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2656209275Smm mult = MIN(mult, 10); 2657168404Spjd 2658208373Smm delta = MIN(bytes * mult, arc_p); 2659208373Smm arc_p = MAX(arc_p_min, arc_p - delta); 2660168404Spjd } 2661168404Spjd ASSERT((int64_t)arc_p >= 0); 2662168404Spjd 2663168404Spjd if (arc_reclaim_needed()) { 2664168404Spjd cv_signal(&arc_reclaim_thr_cv); 2665168404Spjd return; 2666168404Spjd } 2667168404Spjd 2668168404Spjd if (arc_no_grow) 2669168404Spjd return; 2670168404Spjd 2671168404Spjd if (arc_c >= arc_c_max) 2672168404Spjd return; 2673168404Spjd 2674168404Spjd /* 2675168404Spjd * If we're within (2 * maxblocksize) bytes of the target 2676168404Spjd * cache size, increment the target cache size 2677168404Spjd */ 2678168404Spjd if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2679168404Spjd atomic_add_64(&arc_c, (int64_t)bytes); 2680168404Spjd if (arc_c > arc_c_max) 2681168404Spjd arc_c = arc_c_max; 2682168404Spjd else if (state == arc_anon) 2683168404Spjd atomic_add_64(&arc_p, (int64_t)bytes); 2684168404Spjd if (arc_p > arc_c) 2685168404Spjd arc_p = arc_c; 2686168404Spjd } 2687168404Spjd ASSERT((int64_t)arc_p >= 0); 2688168404Spjd} 2689168404Spjd 2690168404Spjd/* 2691168404Spjd * Check if the cache has reached its limits and eviction is required 2692168404Spjd * prior to insert. 2693168404Spjd */ 2694168404Spjdstatic int 2695185029Spjdarc_evict_needed(arc_buf_contents_t type) 2696168404Spjd{ 2697185029Spjd if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2698185029Spjd return (1); 2699185029Spjd 2700219089Spjd#ifdef sun 2701185029Spjd#ifdef _KERNEL 2702185029Spjd /* 2703185029Spjd * If zio data pages are being allocated out of a separate heap segment, 2704185029Spjd * then enforce that the size of available vmem for this area remains 2705185029Spjd * above about 1/32nd free. 2706185029Spjd */ 2707185029Spjd if (type == ARC_BUFC_DATA && zio_arena != NULL && 2708185029Spjd vmem_size(zio_arena, VMEM_FREE) < 2709185029Spjd (vmem_size(zio_arena, VMEM_ALLOC) >> 5)) 2710185029Spjd return (1); 2711185029Spjd#endif 2712219089Spjd#endif /* sun */ 2713185029Spjd 2714168404Spjd if (arc_reclaim_needed()) 2715168404Spjd return (1); 2716168404Spjd 2717168404Spjd return (arc_size > arc_c); 2718168404Spjd} 2719168404Spjd 2720168404Spjd/* 2721168404Spjd * The buffer, supplied as the first argument, needs a data block. 2722168404Spjd * So, if we are at cache max, determine which cache should be victimized. 2723168404Spjd * We have the following cases: 2724168404Spjd * 2725168404Spjd * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2726168404Spjd * In this situation if we're out of space, but the resident size of the MFU is 2727168404Spjd * under the limit, victimize the MFU cache to satisfy this insertion request. 2728168404Spjd * 2729168404Spjd * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2730168404Spjd * Here, we've used up all of the available space for the MRU, so we need to 2731168404Spjd * evict from our own cache instead. Evict from the set of resident MRU 2732168404Spjd * entries. 2733168404Spjd * 2734168404Spjd * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2735168404Spjd * c minus p represents the MFU space in the cache, since p is the size of the 2736168404Spjd * cache that is dedicated to the MRU. In this situation there's still space on 2737168404Spjd * the MFU side, so the MRU side needs to be victimized. 2738168404Spjd * 2739168404Spjd * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2740168404Spjd * MFU's resident set is consuming more space than it has been allotted. In 2741168404Spjd * this situation, we must victimize our own cache, the MFU, for this insertion. 2742168404Spjd */ 2743168404Spjdstatic void 2744168404Spjdarc_get_data_buf(arc_buf_t *buf) 2745168404Spjd{ 2746168404Spjd arc_state_t *state = buf->b_hdr->b_state; 2747168404Spjd uint64_t size = buf->b_hdr->b_size; 2748168404Spjd arc_buf_contents_t type = buf->b_hdr->b_type; 2749168404Spjd 2750168404Spjd arc_adapt(size, state); 2751168404Spjd 2752168404Spjd /* 2753168404Spjd * We have not yet reached cache maximum size, 2754168404Spjd * just allocate a new buffer. 2755168404Spjd */ 2756185029Spjd if (!arc_evict_needed(type)) { 2757168404Spjd if (type == ARC_BUFC_METADATA) { 2758168404Spjd buf->b_data = zio_buf_alloc(size); 2759208373Smm arc_space_consume(size, ARC_SPACE_DATA); 2760168404Spjd } else { 2761168404Spjd ASSERT(type == ARC_BUFC_DATA); 2762168404Spjd buf->b_data = zio_data_buf_alloc(size); 2763208373Smm ARCSTAT_INCR(arcstat_data_size, size); 2764185029Spjd atomic_add_64(&arc_size, size); 2765168404Spjd } 2766168404Spjd goto out; 2767168404Spjd } 2768168404Spjd 2769168404Spjd /* 2770168404Spjd * If we are prefetching from the mfu ghost list, this buffer 2771168404Spjd * will end up on the mru list; so steal space from there. 2772168404Spjd */ 2773168404Spjd if (state == arc_mfu_ghost) 2774168404Spjd state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2775168404Spjd else if (state == arc_mru_ghost) 2776168404Spjd state = arc_mru; 2777168404Spjd 2778168404Spjd if (state == arc_mru || state == arc_anon) { 2779168404Spjd uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2780208373Smm state = (arc_mfu->arcs_lsize[type] >= size && 2781185029Spjd arc_p > mru_used) ? arc_mfu : arc_mru; 2782168404Spjd } else { 2783168404Spjd /* MFU cases */ 2784168404Spjd uint64_t mfu_space = arc_c - arc_p; 2785208373Smm state = (arc_mru->arcs_lsize[type] >= size && 2786185029Spjd mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2787168404Spjd } 2788209962Smm if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2789168404Spjd if (type == ARC_BUFC_METADATA) { 2790168404Spjd buf->b_data = zio_buf_alloc(size); 2791208373Smm arc_space_consume(size, ARC_SPACE_DATA); 2792168404Spjd } else { 2793168404Spjd ASSERT(type == ARC_BUFC_DATA); 2794168404Spjd buf->b_data = zio_data_buf_alloc(size); 2795208373Smm ARCSTAT_INCR(arcstat_data_size, size); 2796185029Spjd atomic_add_64(&arc_size, size); 2797168404Spjd } 2798168404Spjd ARCSTAT_BUMP(arcstat_recycle_miss); 2799168404Spjd } 2800168404Spjd ASSERT(buf->b_data != NULL); 2801168404Spjdout: 2802168404Spjd /* 2803168404Spjd * Update the state size. Note that ghost states have a 2804168404Spjd * "ghost size" and so don't need to be updated. 2805168404Spjd */ 2806168404Spjd if (!GHOST_STATE(buf->b_hdr->b_state)) { 2807168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 2808168404Spjd 2809168404Spjd atomic_add_64(&hdr->b_state->arcs_size, size); 2810168404Spjd if (list_link_active(&hdr->b_arc_node)) { 2811168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2812185029Spjd atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2813168404Spjd } 2814168404Spjd /* 2815168404Spjd * If we are growing the cache, and we are adding anonymous 2816168404Spjd * data, and we have outgrown arc_p, update arc_p 2817168404Spjd */ 2818168404Spjd if (arc_size < arc_c && hdr->b_state == arc_anon && 2819168404Spjd arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2820168404Spjd arc_p = MIN(arc_c, arc_p + size); 2821168404Spjd } 2822205231Skmacy ARCSTAT_BUMP(arcstat_allocated); 2823168404Spjd} 2824168404Spjd 2825168404Spjd/* 2826168404Spjd * This routine is called whenever a buffer is accessed. 2827168404Spjd * NOTE: the hash lock is dropped in this function. 2828168404Spjd */ 2829168404Spjdstatic void 2830168404Spjdarc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2831168404Spjd{ 2832219089Spjd clock_t now; 2833219089Spjd 2834168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 2835168404Spjd 2836168404Spjd if (buf->b_state == arc_anon) { 2837168404Spjd /* 2838168404Spjd * This buffer is not in the cache, and does not 2839168404Spjd * appear in our "ghost" list. Add the new buffer 2840168404Spjd * to the MRU state. 2841168404Spjd */ 2842168404Spjd 2843168404Spjd ASSERT(buf->b_arc_access == 0); 2844219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2845168404Spjd DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2846168404Spjd arc_change_state(arc_mru, buf, hash_lock); 2847168404Spjd 2848168404Spjd } else if (buf->b_state == arc_mru) { 2849219089Spjd now = ddi_get_lbolt(); 2850219089Spjd 2851168404Spjd /* 2852168404Spjd * If this buffer is here because of a prefetch, then either: 2853168404Spjd * - clear the flag if this is a "referencing" read 2854168404Spjd * (any subsequent access will bump this into the MFU state). 2855168404Spjd * or 2856168404Spjd * - move the buffer to the head of the list if this is 2857168404Spjd * another prefetch (to make it less likely to be evicted). 2858168404Spjd */ 2859168404Spjd if ((buf->b_flags & ARC_PREFETCH) != 0) { 2860168404Spjd if (refcount_count(&buf->b_refcnt) == 0) { 2861168404Spjd ASSERT(list_link_active(&buf->b_arc_node)); 2862168404Spjd } else { 2863168404Spjd buf->b_flags &= ~ARC_PREFETCH; 2864168404Spjd ARCSTAT_BUMP(arcstat_mru_hits); 2865168404Spjd } 2866219089Spjd buf->b_arc_access = now; 2867168404Spjd return; 2868168404Spjd } 2869168404Spjd 2870168404Spjd /* 2871168404Spjd * This buffer has been "accessed" only once so far, 2872168404Spjd * but it is still in the cache. Move it to the MFU 2873168404Spjd * state. 2874168404Spjd */ 2875219089Spjd if (now > buf->b_arc_access + ARC_MINTIME) { 2876168404Spjd /* 2877168404Spjd * More than 125ms have passed since we 2878168404Spjd * instantiated this buffer. Move it to the 2879168404Spjd * most frequently used state. 2880168404Spjd */ 2881219089Spjd buf->b_arc_access = now; 2882168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2883168404Spjd arc_change_state(arc_mfu, buf, hash_lock); 2884168404Spjd } 2885168404Spjd ARCSTAT_BUMP(arcstat_mru_hits); 2886168404Spjd } else if (buf->b_state == arc_mru_ghost) { 2887168404Spjd arc_state_t *new_state; 2888168404Spjd /* 2889168404Spjd * This buffer has been "accessed" recently, but 2890168404Spjd * was evicted from the cache. Move it to the 2891168404Spjd * MFU state. 2892168404Spjd */ 2893168404Spjd 2894168404Spjd if (buf->b_flags & ARC_PREFETCH) { 2895168404Spjd new_state = arc_mru; 2896168404Spjd if (refcount_count(&buf->b_refcnt) > 0) 2897168404Spjd buf->b_flags &= ~ARC_PREFETCH; 2898168404Spjd DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2899168404Spjd } else { 2900168404Spjd new_state = arc_mfu; 2901168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2902168404Spjd } 2903168404Spjd 2904219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2905168404Spjd arc_change_state(new_state, buf, hash_lock); 2906168404Spjd 2907168404Spjd ARCSTAT_BUMP(arcstat_mru_ghost_hits); 2908168404Spjd } else if (buf->b_state == arc_mfu) { 2909168404Spjd /* 2910168404Spjd * This buffer has been accessed more than once and is 2911168404Spjd * still in the cache. Keep it in the MFU state. 2912168404Spjd * 2913168404Spjd * NOTE: an add_reference() that occurred when we did 2914168404Spjd * the arc_read() will have kicked this off the list. 2915168404Spjd * If it was a prefetch, we will explicitly move it to 2916168404Spjd * the head of the list now. 2917168404Spjd */ 2918168404Spjd if ((buf->b_flags & ARC_PREFETCH) != 0) { 2919168404Spjd ASSERT(refcount_count(&buf->b_refcnt) == 0); 2920168404Spjd ASSERT(list_link_active(&buf->b_arc_node)); 2921168404Spjd } 2922168404Spjd ARCSTAT_BUMP(arcstat_mfu_hits); 2923219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2924168404Spjd } else if (buf->b_state == arc_mfu_ghost) { 2925168404Spjd arc_state_t *new_state = arc_mfu; 2926168404Spjd /* 2927168404Spjd * This buffer has been accessed more than once but has 2928168404Spjd * been evicted from the cache. Move it back to the 2929168404Spjd * MFU state. 2930168404Spjd */ 2931168404Spjd 2932168404Spjd if (buf->b_flags & ARC_PREFETCH) { 2933168404Spjd /* 2934168404Spjd * This is a prefetch access... 2935168404Spjd * move this block back to the MRU state. 2936168404Spjd */ 2937243674Smm ASSERT0(refcount_count(&buf->b_refcnt)); 2938168404Spjd new_state = arc_mru; 2939168404Spjd } 2940168404Spjd 2941219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2942168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2943168404Spjd arc_change_state(new_state, buf, hash_lock); 2944168404Spjd 2945168404Spjd ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 2946185029Spjd } else if (buf->b_state == arc_l2c_only) { 2947185029Spjd /* 2948185029Spjd * This buffer is on the 2nd Level ARC. 2949185029Spjd */ 2950185029Spjd 2951219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2952185029Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2953185029Spjd arc_change_state(arc_mfu, buf, hash_lock); 2954168404Spjd } else { 2955168404Spjd ASSERT(!"invalid arc state"); 2956168404Spjd } 2957168404Spjd} 2958168404Spjd 2959168404Spjd/* a generic arc_done_func_t which you can use */ 2960168404Spjd/* ARGSUSED */ 2961168404Spjdvoid 2962168404Spjdarc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 2963168404Spjd{ 2964219089Spjd if (zio == NULL || zio->io_error == 0) 2965219089Spjd bcopy(buf->b_data, arg, buf->b_hdr->b_size); 2966249643Smm VERIFY(arc_buf_remove_ref(buf, arg)); 2967168404Spjd} 2968168404Spjd 2969185029Spjd/* a generic arc_done_func_t */ 2970168404Spjdvoid 2971168404Spjdarc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 2972168404Spjd{ 2973168404Spjd arc_buf_t **bufp = arg; 2974168404Spjd if (zio && zio->io_error) { 2975249643Smm VERIFY(arc_buf_remove_ref(buf, arg)); 2976168404Spjd *bufp = NULL; 2977168404Spjd } else { 2978168404Spjd *bufp = buf; 2979219089Spjd ASSERT(buf->b_data); 2980168404Spjd } 2981168404Spjd} 2982168404Spjd 2983168404Spjdstatic void 2984168404Spjdarc_read_done(zio_t *zio) 2985168404Spjd{ 2986168404Spjd arc_buf_hdr_t *hdr, *found; 2987168404Spjd arc_buf_t *buf; 2988168404Spjd arc_buf_t *abuf; /* buffer we're assigning to callback */ 2989168404Spjd kmutex_t *hash_lock; 2990168404Spjd arc_callback_t *callback_list, *acb; 2991168404Spjd int freeable = FALSE; 2992168404Spjd 2993168404Spjd buf = zio->io_private; 2994168404Spjd hdr = buf->b_hdr; 2995168404Spjd 2996168404Spjd /* 2997168404Spjd * The hdr was inserted into hash-table and removed from lists 2998168404Spjd * prior to starting I/O. We should find this header, since 2999168404Spjd * it's in the hash table, and it should be legit since it's 3000168404Spjd * not possible to evict it during the I/O. The only possible 3001168404Spjd * reason for it not to be found is if we were freed during the 3002168404Spjd * read. 3003168404Spjd */ 3004209962Smm found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth, 3005168404Spjd &hash_lock); 3006168404Spjd 3007168404Spjd ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) || 3008185029Spjd (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3009185029Spjd (found == hdr && HDR_L2_READING(hdr))); 3010168404Spjd 3011185029Spjd hdr->b_flags &= ~ARC_L2_EVICTED; 3012185029Spjd if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 3013185029Spjd hdr->b_flags &= ~ARC_L2CACHE; 3014206796Spjd 3015168404Spjd /* byteswap if necessary */ 3016168404Spjd callback_list = hdr->b_acb; 3017168404Spjd ASSERT(callback_list != NULL); 3018209101Smm if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3019243674Smm dmu_object_byteswap_t bswap = 3020243674Smm DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3021185029Spjd arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3022185029Spjd byteswap_uint64_array : 3023243674Smm dmu_ot_byteswap[bswap].ob_func; 3024185029Spjd func(buf->b_data, hdr->b_size); 3025185029Spjd } 3026168404Spjd 3027185029Spjd arc_cksum_compute(buf, B_FALSE); 3028243674Smm#ifdef illumos 3029243674Smm arc_buf_watch(buf); 3030243674Smm#endif /* illumos */ 3031168404Spjd 3032219089Spjd if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3033219089Spjd /* 3034219089Spjd * Only call arc_access on anonymous buffers. This is because 3035219089Spjd * if we've issued an I/O for an evicted buffer, we've already 3036219089Spjd * called arc_access (to prevent any simultaneous readers from 3037219089Spjd * getting confused). 3038219089Spjd */ 3039219089Spjd arc_access(hdr, hash_lock); 3040219089Spjd } 3041219089Spjd 3042168404Spjd /* create copies of the data buffer for the callers */ 3043168404Spjd abuf = buf; 3044168404Spjd for (acb = callback_list; acb; acb = acb->acb_next) { 3045168404Spjd if (acb->acb_done) { 3046248547Smm if (abuf == NULL) { 3047248547Smm ARCSTAT_BUMP(arcstat_duplicate_reads); 3048168404Spjd abuf = arc_buf_clone(buf); 3049248547Smm } 3050168404Spjd acb->acb_buf = abuf; 3051168404Spjd abuf = NULL; 3052168404Spjd } 3053168404Spjd } 3054168404Spjd hdr->b_acb = NULL; 3055168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3056168404Spjd ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3057219089Spjd if (abuf == buf) { 3058219089Spjd ASSERT(buf->b_efunc == NULL); 3059219089Spjd ASSERT(hdr->b_datacnt == 1); 3060168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 3061219089Spjd } 3062168404Spjd 3063168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3064168404Spjd 3065168404Spjd if (zio->io_error != 0) { 3066168404Spjd hdr->b_flags |= ARC_IO_ERROR; 3067168404Spjd if (hdr->b_state != arc_anon) 3068168404Spjd arc_change_state(arc_anon, hdr, hash_lock); 3069168404Spjd if (HDR_IN_HASH_TABLE(hdr)) 3070168404Spjd buf_hash_remove(hdr); 3071168404Spjd freeable = refcount_is_zero(&hdr->b_refcnt); 3072168404Spjd } 3073168404Spjd 3074168404Spjd /* 3075168404Spjd * Broadcast before we drop the hash_lock to avoid the possibility 3076168404Spjd * that the hdr (and hence the cv) might be freed before we get to 3077168404Spjd * the cv_broadcast(). 3078168404Spjd */ 3079168404Spjd cv_broadcast(&hdr->b_cv); 3080168404Spjd 3081168404Spjd if (hash_lock) { 3082168404Spjd mutex_exit(hash_lock); 3083168404Spjd } else { 3084168404Spjd /* 3085168404Spjd * This block was freed while we waited for the read to 3086168404Spjd * complete. It has been removed from the hash table and 3087168404Spjd * moved to the anonymous state (so that it won't show up 3088168404Spjd * in the cache). 3089168404Spjd */ 3090168404Spjd ASSERT3P(hdr->b_state, ==, arc_anon); 3091168404Spjd freeable = refcount_is_zero(&hdr->b_refcnt); 3092168404Spjd } 3093168404Spjd 3094168404Spjd /* execute each callback and free its structure */ 3095168404Spjd while ((acb = callback_list) != NULL) { 3096168404Spjd if (acb->acb_done) 3097168404Spjd acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3098168404Spjd 3099168404Spjd if (acb->acb_zio_dummy != NULL) { 3100168404Spjd acb->acb_zio_dummy->io_error = zio->io_error; 3101168404Spjd zio_nowait(acb->acb_zio_dummy); 3102168404Spjd } 3103168404Spjd 3104168404Spjd callback_list = acb->acb_next; 3105168404Spjd kmem_free(acb, sizeof (arc_callback_t)); 3106168404Spjd } 3107168404Spjd 3108168404Spjd if (freeable) 3109168404Spjd arc_hdr_destroy(hdr); 3110168404Spjd} 3111168404Spjd 3112168404Spjd/* 3113168404Spjd * "Read" the block block at the specified DVA (in bp) via the 3114168404Spjd * cache. If the block is found in the cache, invoke the provided 3115168404Spjd * callback immediately and return. Note that the `zio' parameter 3116168404Spjd * in the callback will be NULL in this case, since no IO was 3117168404Spjd * required. If the block is not in the cache pass the read request 3118168404Spjd * on to the spa with a substitute callback function, so that the 3119168404Spjd * requested block will be added to the cache. 3120168404Spjd * 3121168404Spjd * If a read request arrives for a block that has a read in-progress, 3122168404Spjd * either wait for the in-progress read to complete (and return the 3123168404Spjd * results); or, if this is a read with a "done" func, add a record 3124168404Spjd * to the read to invoke the "done" func when the read completes, 3125168404Spjd * and return; or just return. 3126168404Spjd * 3127168404Spjd * arc_read_done() will invoke all the requested "done" functions 3128168404Spjd * for readers of this block. 3129168404Spjd */ 3130168404Spjdint 3131247406Smmarc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3132260764Savg void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags, 3133247406Smm const zbookmark_t *zb) 3134168404Spjd{ 3135168404Spjd arc_buf_hdr_t *hdr; 3136248369Smm arc_buf_t *buf = NULL; 3137168404Spjd kmutex_t *hash_lock; 3138185029Spjd zio_t *rzio; 3139229578Smm uint64_t guid = spa_load_guid(spa); 3140168404Spjd 3141168404Spjdtop: 3142219089Spjd hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp), 3143219089Spjd &hash_lock); 3144168404Spjd if (hdr && hdr->b_datacnt > 0) { 3145168404Spjd 3146168404Spjd *arc_flags |= ARC_CACHED; 3147168404Spjd 3148168404Spjd if (HDR_IO_IN_PROGRESS(hdr)) { 3149168404Spjd 3150168404Spjd if (*arc_flags & ARC_WAIT) { 3151168404Spjd cv_wait(&hdr->b_cv, hash_lock); 3152168404Spjd mutex_exit(hash_lock); 3153168404Spjd goto top; 3154168404Spjd } 3155168404Spjd ASSERT(*arc_flags & ARC_NOWAIT); 3156168404Spjd 3157168404Spjd if (done) { 3158168404Spjd arc_callback_t *acb = NULL; 3159168404Spjd 3160168404Spjd acb = kmem_zalloc(sizeof (arc_callback_t), 3161168404Spjd KM_SLEEP); 3162168404Spjd acb->acb_done = done; 3163168404Spjd acb->acb_private = private; 3164168404Spjd if (pio != NULL) 3165168404Spjd acb->acb_zio_dummy = zio_null(pio, 3166209962Smm spa, NULL, NULL, NULL, zio_flags); 3167168404Spjd 3168168404Spjd ASSERT(acb->acb_done != NULL); 3169168404Spjd acb->acb_next = hdr->b_acb; 3170168404Spjd hdr->b_acb = acb; 3171168404Spjd add_reference(hdr, hash_lock, private); 3172168404Spjd mutex_exit(hash_lock); 3173168404Spjd return (0); 3174168404Spjd } 3175168404Spjd mutex_exit(hash_lock); 3176168404Spjd return (0); 3177168404Spjd } 3178168404Spjd 3179168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3180168404Spjd 3181168404Spjd if (done) { 3182168404Spjd add_reference(hdr, hash_lock, private); 3183168404Spjd /* 3184168404Spjd * If this block is already in use, create a new 3185168404Spjd * copy of the data so that we will be guaranteed 3186168404Spjd * that arc_release() will always succeed. 3187168404Spjd */ 3188168404Spjd buf = hdr->b_buf; 3189168404Spjd ASSERT(buf); 3190168404Spjd ASSERT(buf->b_data); 3191168404Spjd if (HDR_BUF_AVAILABLE(hdr)) { 3192168404Spjd ASSERT(buf->b_efunc == NULL); 3193168404Spjd hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3194168404Spjd } else { 3195168404Spjd buf = arc_buf_clone(buf); 3196168404Spjd } 3197219089Spjd 3198168404Spjd } else if (*arc_flags & ARC_PREFETCH && 3199168404Spjd refcount_count(&hdr->b_refcnt) == 0) { 3200168404Spjd hdr->b_flags |= ARC_PREFETCH; 3201168404Spjd } 3202168404Spjd DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3203168404Spjd arc_access(hdr, hash_lock); 3204185029Spjd if (*arc_flags & ARC_L2CACHE) 3205185029Spjd hdr->b_flags |= ARC_L2CACHE; 3206252140Sdelphij if (*arc_flags & ARC_L2COMPRESS) 3207252140Sdelphij hdr->b_flags |= ARC_L2COMPRESS; 3208168404Spjd mutex_exit(hash_lock); 3209168404Spjd ARCSTAT_BUMP(arcstat_hits); 3210168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3211168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3212168404Spjd data, metadata, hits); 3213168404Spjd 3214168404Spjd if (done) 3215168404Spjd done(NULL, buf, private); 3216168404Spjd } else { 3217168404Spjd uint64_t size = BP_GET_LSIZE(bp); 3218168404Spjd arc_callback_t *acb; 3219185029Spjd vdev_t *vd = NULL; 3220248369Smm uint64_t addr = 0; 3221208373Smm boolean_t devw = B_FALSE; 3222258557Savg enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3223258557Savg uint64_t b_asize = 0; 3224168404Spjd 3225168404Spjd if (hdr == NULL) { 3226168404Spjd /* this block is not in the cache */ 3227168404Spjd arc_buf_hdr_t *exists; 3228168404Spjd arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3229168404Spjd buf = arc_buf_alloc(spa, size, private, type); 3230168404Spjd hdr = buf->b_hdr; 3231168404Spjd hdr->b_dva = *BP_IDENTITY(bp); 3232219089Spjd hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3233168404Spjd hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3234168404Spjd exists = buf_hash_insert(hdr, &hash_lock); 3235168404Spjd if (exists) { 3236168404Spjd /* somebody beat us to the hash insert */ 3237168404Spjd mutex_exit(hash_lock); 3238219089Spjd buf_discard_identity(hdr); 3239168404Spjd (void) arc_buf_remove_ref(buf, private); 3240168404Spjd goto top; /* restart the IO request */ 3241168404Spjd } 3242168404Spjd /* if this is a prefetch, we don't have a reference */ 3243168404Spjd if (*arc_flags & ARC_PREFETCH) { 3244168404Spjd (void) remove_reference(hdr, hash_lock, 3245168404Spjd private); 3246168404Spjd hdr->b_flags |= ARC_PREFETCH; 3247168404Spjd } 3248185029Spjd if (*arc_flags & ARC_L2CACHE) 3249185029Spjd hdr->b_flags |= ARC_L2CACHE; 3250252140Sdelphij if (*arc_flags & ARC_L2COMPRESS) 3251252140Sdelphij hdr->b_flags |= ARC_L2COMPRESS; 3252168404Spjd if (BP_GET_LEVEL(bp) > 0) 3253168404Spjd hdr->b_flags |= ARC_INDIRECT; 3254168404Spjd } else { 3255168404Spjd /* this block is in the ghost cache */ 3256168404Spjd ASSERT(GHOST_STATE(hdr->b_state)); 3257168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3258243674Smm ASSERT0(refcount_count(&hdr->b_refcnt)); 3259168404Spjd ASSERT(hdr->b_buf == NULL); 3260168404Spjd 3261168404Spjd /* if this is a prefetch, we don't have a reference */ 3262168404Spjd if (*arc_flags & ARC_PREFETCH) 3263168404Spjd hdr->b_flags |= ARC_PREFETCH; 3264168404Spjd else 3265168404Spjd add_reference(hdr, hash_lock, private); 3266185029Spjd if (*arc_flags & ARC_L2CACHE) 3267185029Spjd hdr->b_flags |= ARC_L2CACHE; 3268252140Sdelphij if (*arc_flags & ARC_L2COMPRESS) 3269252140Sdelphij hdr->b_flags |= ARC_L2COMPRESS; 3270185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3271168404Spjd buf->b_hdr = hdr; 3272168404Spjd buf->b_data = NULL; 3273168404Spjd buf->b_efunc = NULL; 3274168404Spjd buf->b_private = NULL; 3275168404Spjd buf->b_next = NULL; 3276168404Spjd hdr->b_buf = buf; 3277168404Spjd ASSERT(hdr->b_datacnt == 0); 3278168404Spjd hdr->b_datacnt = 1; 3279219089Spjd arc_get_data_buf(buf); 3280219089Spjd arc_access(hdr, hash_lock); 3281168404Spjd } 3282168404Spjd 3283219089Spjd ASSERT(!GHOST_STATE(hdr->b_state)); 3284219089Spjd 3285168404Spjd acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3286168404Spjd acb->acb_done = done; 3287168404Spjd acb->acb_private = private; 3288168404Spjd 3289168404Spjd ASSERT(hdr->b_acb == NULL); 3290168404Spjd hdr->b_acb = acb; 3291168404Spjd hdr->b_flags |= ARC_IO_IN_PROGRESS; 3292168404Spjd 3293258557Savg if (hdr->b_l2hdr != NULL && 3294185029Spjd (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3295208373Smm devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3296185029Spjd addr = hdr->b_l2hdr->b_daddr; 3297258557Savg b_compress = hdr->b_l2hdr->b_compress; 3298258557Savg b_asize = hdr->b_l2hdr->b_asize; 3299185029Spjd /* 3300185029Spjd * Lock out device removal. 3301185029Spjd */ 3302185029Spjd if (vdev_is_dead(vd) || 3303185029Spjd !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3304185029Spjd vd = NULL; 3305185029Spjd } 3306185029Spjd 3307168404Spjd mutex_exit(hash_lock); 3308168404Spjd 3309252749Sdelphij /* 3310252749Sdelphij * At this point, we have a level 1 cache miss. Try again in 3311252749Sdelphij * L2ARC if possible. 3312252749Sdelphij */ 3313168404Spjd ASSERT3U(hdr->b_size, ==, size); 3314219089Spjd DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3315219089Spjd uint64_t, size, zbookmark_t *, zb); 3316168404Spjd ARCSTAT_BUMP(arcstat_misses); 3317168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3318168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3319168404Spjd data, metadata, misses); 3320229568Smm#ifdef _KERNEL 3321229568Smm curthread->td_ru.ru_inblock++; 3322229568Smm#endif 3323168404Spjd 3324208373Smm if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3325185029Spjd /* 3326185029Spjd * Read from the L2ARC if the following are true: 3327185029Spjd * 1. The L2ARC vdev was previously cached. 3328185029Spjd * 2. This buffer still has L2ARC metadata. 3329185029Spjd * 3. This buffer isn't currently writing to the L2ARC. 3330185029Spjd * 4. The L2ARC entry wasn't evicted, which may 3331185029Spjd * also have invalidated the vdev. 3332208373Smm * 5. This isn't prefetch and l2arc_noprefetch is set. 3333185029Spjd */ 3334185029Spjd if (hdr->b_l2hdr != NULL && 3335208373Smm !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3336208373Smm !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3337185029Spjd l2arc_read_callback_t *cb; 3338185029Spjd 3339185029Spjd DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3340185029Spjd ARCSTAT_BUMP(arcstat_l2_hits); 3341185029Spjd 3342185029Spjd cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3343185029Spjd KM_SLEEP); 3344185029Spjd cb->l2rcb_buf = buf; 3345185029Spjd cb->l2rcb_spa = spa; 3346185029Spjd cb->l2rcb_bp = *bp; 3347185029Spjd cb->l2rcb_zb = *zb; 3348185029Spjd cb->l2rcb_flags = zio_flags; 3349258557Savg cb->l2rcb_compress = b_compress; 3350185029Spjd 3351248369Smm ASSERT(addr >= VDEV_LABEL_START_SIZE && 3352248369Smm addr + size < vd->vdev_psize - 3353248369Smm VDEV_LABEL_END_SIZE); 3354248369Smm 3355185029Spjd /* 3356185029Spjd * l2arc read. The SCL_L2ARC lock will be 3357185029Spjd * released by l2arc_read_done(). 3358252140Sdelphij * Issue a null zio if the underlying buffer 3359252140Sdelphij * was squashed to zero size by compression. 3360185029Spjd */ 3361258557Savg if (b_compress == ZIO_COMPRESS_EMPTY) { 3362252140Sdelphij rzio = zio_null(pio, spa, vd, 3363252140Sdelphij l2arc_read_done, cb, 3364252140Sdelphij zio_flags | ZIO_FLAG_DONT_CACHE | 3365252140Sdelphij ZIO_FLAG_CANFAIL | 3366252140Sdelphij ZIO_FLAG_DONT_PROPAGATE | 3367252140Sdelphij ZIO_FLAG_DONT_RETRY); 3368252140Sdelphij } else { 3369252140Sdelphij rzio = zio_read_phys(pio, vd, addr, 3370258557Savg b_asize, buf->b_data, 3371258557Savg ZIO_CHECKSUM_OFF, 3372252140Sdelphij l2arc_read_done, cb, priority, 3373252140Sdelphij zio_flags | ZIO_FLAG_DONT_CACHE | 3374252140Sdelphij ZIO_FLAG_CANFAIL | 3375252140Sdelphij ZIO_FLAG_DONT_PROPAGATE | 3376252140Sdelphij ZIO_FLAG_DONT_RETRY, B_FALSE); 3377252140Sdelphij } 3378185029Spjd DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3379185029Spjd zio_t *, rzio); 3380258557Savg ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3381185029Spjd 3382185029Spjd if (*arc_flags & ARC_NOWAIT) { 3383185029Spjd zio_nowait(rzio); 3384185029Spjd return (0); 3385185029Spjd } 3386185029Spjd 3387185029Spjd ASSERT(*arc_flags & ARC_WAIT); 3388185029Spjd if (zio_wait(rzio) == 0) 3389185029Spjd return (0); 3390185029Spjd 3391185029Spjd /* l2arc read error; goto zio_read() */ 3392185029Spjd } else { 3393185029Spjd DTRACE_PROBE1(l2arc__miss, 3394185029Spjd arc_buf_hdr_t *, hdr); 3395185029Spjd ARCSTAT_BUMP(arcstat_l2_misses); 3396185029Spjd if (HDR_L2_WRITING(hdr)) 3397185029Spjd ARCSTAT_BUMP(arcstat_l2_rw_clash); 3398185029Spjd spa_config_exit(spa, SCL_L2ARC, vd); 3399185029Spjd } 3400208373Smm } else { 3401208373Smm if (vd != NULL) 3402208373Smm spa_config_exit(spa, SCL_L2ARC, vd); 3403208373Smm if (l2arc_ndev != 0) { 3404208373Smm DTRACE_PROBE1(l2arc__miss, 3405208373Smm arc_buf_hdr_t *, hdr); 3406208373Smm ARCSTAT_BUMP(arcstat_l2_misses); 3407208373Smm } 3408185029Spjd } 3409185029Spjd 3410168404Spjd rzio = zio_read(pio, spa, bp, buf->b_data, size, 3411185029Spjd arc_read_done, buf, priority, zio_flags, zb); 3412168404Spjd 3413168404Spjd if (*arc_flags & ARC_WAIT) 3414168404Spjd return (zio_wait(rzio)); 3415168404Spjd 3416168404Spjd ASSERT(*arc_flags & ARC_NOWAIT); 3417168404Spjd zio_nowait(rzio); 3418168404Spjd } 3419168404Spjd return (0); 3420168404Spjd} 3421168404Spjd 3422168404Spjdvoid 3423168404Spjdarc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3424168404Spjd{ 3425168404Spjd ASSERT(buf->b_hdr != NULL); 3426168404Spjd ASSERT(buf->b_hdr->b_state != arc_anon); 3427168404Spjd ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3428219089Spjd ASSERT(buf->b_efunc == NULL); 3429219089Spjd ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3430219089Spjd 3431168404Spjd buf->b_efunc = func; 3432168404Spjd buf->b_private = private; 3433168404Spjd} 3434168404Spjd 3435168404Spjd/* 3436252142Sdelphij * Notify the arc that a block was freed, and thus will never be used again. 3437252142Sdelphij */ 3438252142Sdelphijvoid 3439252142Sdelphijarc_freed(spa_t *spa, const blkptr_t *bp) 3440252142Sdelphij{ 3441252142Sdelphij arc_buf_hdr_t *hdr; 3442252142Sdelphij kmutex_t *hash_lock; 3443252142Sdelphij uint64_t guid = spa_load_guid(spa); 3444252142Sdelphij 3445252142Sdelphij hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp), 3446252142Sdelphij &hash_lock); 3447252142Sdelphij if (hdr == NULL) 3448252142Sdelphij return; 3449252142Sdelphij if (HDR_BUF_AVAILABLE(hdr)) { 3450252142Sdelphij arc_buf_t *buf = hdr->b_buf; 3451252142Sdelphij add_reference(hdr, hash_lock, FTAG); 3452252142Sdelphij hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3453252142Sdelphij mutex_exit(hash_lock); 3454252142Sdelphij 3455252142Sdelphij arc_release(buf, FTAG); 3456252142Sdelphij (void) arc_buf_remove_ref(buf, FTAG); 3457252142Sdelphij } else { 3458252142Sdelphij mutex_exit(hash_lock); 3459252142Sdelphij } 3460252142Sdelphij 3461252142Sdelphij} 3462252142Sdelphij 3463252142Sdelphij/* 3464168404Spjd * This is used by the DMU to let the ARC know that a buffer is 3465168404Spjd * being evicted, so the ARC should clean up. If this arc buf 3466168404Spjd * is not yet in the evicted state, it will be put there. 3467168404Spjd */ 3468168404Spjdint 3469168404Spjdarc_buf_evict(arc_buf_t *buf) 3470168404Spjd{ 3471168404Spjd arc_buf_hdr_t *hdr; 3472168404Spjd kmutex_t *hash_lock; 3473168404Spjd arc_buf_t **bufp; 3474205231Skmacy list_t *list, *evicted_list; 3475205231Skmacy kmutex_t *lock, *evicted_lock; 3476206796Spjd 3477219089Spjd mutex_enter(&buf->b_evict_lock); 3478168404Spjd hdr = buf->b_hdr; 3479168404Spjd if (hdr == NULL) { 3480168404Spjd /* 3481168404Spjd * We are in arc_do_user_evicts(). 3482168404Spjd */ 3483168404Spjd ASSERT(buf->b_data == NULL); 3484219089Spjd mutex_exit(&buf->b_evict_lock); 3485168404Spjd return (0); 3486185029Spjd } else if (buf->b_data == NULL) { 3487185029Spjd arc_buf_t copy = *buf; /* structure assignment */ 3488185029Spjd /* 3489185029Spjd * We are on the eviction list; process this buffer now 3490185029Spjd * but let arc_do_user_evicts() do the reaping. 3491185029Spjd */ 3492185029Spjd buf->b_efunc = NULL; 3493219089Spjd mutex_exit(&buf->b_evict_lock); 3494185029Spjd VERIFY(copy.b_efunc(©) == 0); 3495185029Spjd return (1); 3496168404Spjd } 3497168404Spjd hash_lock = HDR_LOCK(hdr); 3498168404Spjd mutex_enter(hash_lock); 3499219089Spjd hdr = buf->b_hdr; 3500219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3501168404Spjd 3502168404Spjd ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3503168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3504168404Spjd 3505168404Spjd /* 3506168404Spjd * Pull this buffer off of the hdr 3507168404Spjd */ 3508168404Spjd bufp = &hdr->b_buf; 3509168404Spjd while (*bufp != buf) 3510168404Spjd bufp = &(*bufp)->b_next; 3511168404Spjd *bufp = buf->b_next; 3512168404Spjd 3513168404Spjd ASSERT(buf->b_data != NULL); 3514168404Spjd arc_buf_destroy(buf, FALSE, FALSE); 3515168404Spjd 3516168404Spjd if (hdr->b_datacnt == 0) { 3517168404Spjd arc_state_t *old_state = hdr->b_state; 3518168404Spjd arc_state_t *evicted_state; 3519168404Spjd 3520219089Spjd ASSERT(hdr->b_buf == NULL); 3521168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 3522168404Spjd 3523168404Spjd evicted_state = 3524168404Spjd (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3525168404Spjd 3526205231Skmacy get_buf_info(hdr, old_state, &list, &lock); 3527205231Skmacy get_buf_info(hdr, evicted_state, &evicted_list, &evicted_lock); 3528205231Skmacy mutex_enter(lock); 3529205231Skmacy mutex_enter(evicted_lock); 3530168404Spjd 3531168404Spjd arc_change_state(evicted_state, hdr, hash_lock); 3532168404Spjd ASSERT(HDR_IN_HASH_TABLE(hdr)); 3533185029Spjd hdr->b_flags |= ARC_IN_HASH_TABLE; 3534185029Spjd hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3535168404Spjd 3536205231Skmacy mutex_exit(evicted_lock); 3537205231Skmacy mutex_exit(lock); 3538168404Spjd } 3539168404Spjd mutex_exit(hash_lock); 3540219089Spjd mutex_exit(&buf->b_evict_lock); 3541168404Spjd 3542168404Spjd VERIFY(buf->b_efunc(buf) == 0); 3543168404Spjd buf->b_efunc = NULL; 3544168404Spjd buf->b_private = NULL; 3545168404Spjd buf->b_hdr = NULL; 3546219089Spjd buf->b_next = NULL; 3547168404Spjd kmem_cache_free(buf_cache, buf); 3548168404Spjd return (1); 3549168404Spjd} 3550168404Spjd 3551168404Spjd/* 3552252749Sdelphij * Release this buffer from the cache, making it an anonymous buffer. This 3553252749Sdelphij * must be done after a read and prior to modifying the buffer contents. 3554168404Spjd * If the buffer has more than one reference, we must make 3555185029Spjd * a new hdr for the buffer. 3556168404Spjd */ 3557168404Spjdvoid 3558168404Spjdarc_release(arc_buf_t *buf, void *tag) 3559168404Spjd{ 3560185029Spjd arc_buf_hdr_t *hdr; 3561219089Spjd kmutex_t *hash_lock = NULL; 3562185029Spjd l2arc_buf_hdr_t *l2hdr; 3563185029Spjd uint64_t buf_size; 3564168404Spjd 3565219089Spjd /* 3566219089Spjd * It would be nice to assert that if it's DMU metadata (level > 3567219089Spjd * 0 || it's the dnode file), then it must be syncing context. 3568219089Spjd * But we don't know that information at this level. 3569219089Spjd */ 3570219089Spjd 3571219089Spjd mutex_enter(&buf->b_evict_lock); 3572185029Spjd hdr = buf->b_hdr; 3573185029Spjd 3574168404Spjd /* this buffer is not on any list */ 3575168404Spjd ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3576168404Spjd 3577168404Spjd if (hdr->b_state == arc_anon) { 3578168404Spjd /* this buffer is already released */ 3579168404Spjd ASSERT(buf->b_efunc == NULL); 3580208373Smm } else { 3581208373Smm hash_lock = HDR_LOCK(hdr); 3582208373Smm mutex_enter(hash_lock); 3583219089Spjd hdr = buf->b_hdr; 3584219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3585168404Spjd } 3586168404Spjd 3587185029Spjd l2hdr = hdr->b_l2hdr; 3588185029Spjd if (l2hdr) { 3589185029Spjd mutex_enter(&l2arc_buflist_mtx); 3590185029Spjd hdr->b_l2hdr = NULL; 3591258560Savg list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3592185029Spjd } 3593248369Smm buf_size = hdr->b_size; 3594185029Spjd 3595168404Spjd /* 3596168404Spjd * Do we have more than one buf? 3597168404Spjd */ 3598185029Spjd if (hdr->b_datacnt > 1) { 3599168404Spjd arc_buf_hdr_t *nhdr; 3600168404Spjd arc_buf_t **bufp; 3601168404Spjd uint64_t blksz = hdr->b_size; 3602209962Smm uint64_t spa = hdr->b_spa; 3603168404Spjd arc_buf_contents_t type = hdr->b_type; 3604185029Spjd uint32_t flags = hdr->b_flags; 3605168404Spjd 3606185029Spjd ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3607168404Spjd /* 3608219089Spjd * Pull the data off of this hdr and attach it to 3609219089Spjd * a new anonymous hdr. 3610168404Spjd */ 3611168404Spjd (void) remove_reference(hdr, hash_lock, tag); 3612168404Spjd bufp = &hdr->b_buf; 3613168404Spjd while (*bufp != buf) 3614168404Spjd bufp = &(*bufp)->b_next; 3615219089Spjd *bufp = buf->b_next; 3616168404Spjd buf->b_next = NULL; 3617168404Spjd 3618168404Spjd ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3619168404Spjd atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3620168404Spjd if (refcount_is_zero(&hdr->b_refcnt)) { 3621185029Spjd uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3622185029Spjd ASSERT3U(*size, >=, hdr->b_size); 3623185029Spjd atomic_add_64(size, -hdr->b_size); 3624168404Spjd } 3625248547Smm 3626248547Smm /* 3627248547Smm * We're releasing a duplicate user data buffer, update 3628248547Smm * our statistics accordingly. 3629248547Smm */ 3630248547Smm if (hdr->b_type == ARC_BUFC_DATA) { 3631248547Smm ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3632248547Smm ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3633248547Smm -hdr->b_size); 3634248547Smm } 3635168404Spjd hdr->b_datacnt -= 1; 3636168404Spjd arc_cksum_verify(buf); 3637243674Smm#ifdef illumos 3638243674Smm arc_buf_unwatch(buf); 3639243674Smm#endif /* illumos */ 3640168404Spjd 3641168404Spjd mutex_exit(hash_lock); 3642168404Spjd 3643185029Spjd nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3644168404Spjd nhdr->b_size = blksz; 3645168404Spjd nhdr->b_spa = spa; 3646168404Spjd nhdr->b_type = type; 3647168404Spjd nhdr->b_buf = buf; 3648168404Spjd nhdr->b_state = arc_anon; 3649168404Spjd nhdr->b_arc_access = 0; 3650185029Spjd nhdr->b_flags = flags & ARC_L2_WRITING; 3651185029Spjd nhdr->b_l2hdr = NULL; 3652168404Spjd nhdr->b_datacnt = 1; 3653168404Spjd nhdr->b_freeze_cksum = NULL; 3654168404Spjd (void) refcount_add(&nhdr->b_refcnt, tag); 3655168404Spjd buf->b_hdr = nhdr; 3656219089Spjd mutex_exit(&buf->b_evict_lock); 3657168404Spjd atomic_add_64(&arc_anon->arcs_size, blksz); 3658168404Spjd } else { 3659219089Spjd mutex_exit(&buf->b_evict_lock); 3660168404Spjd ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3661168404Spjd ASSERT(!list_link_active(&hdr->b_arc_node)); 3662168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3663219089Spjd if (hdr->b_state != arc_anon) 3664219089Spjd arc_change_state(arc_anon, hdr, hash_lock); 3665168404Spjd hdr->b_arc_access = 0; 3666219089Spjd if (hash_lock) 3667219089Spjd mutex_exit(hash_lock); 3668185029Spjd 3669219089Spjd buf_discard_identity(hdr); 3670168404Spjd arc_buf_thaw(buf); 3671168404Spjd } 3672168404Spjd buf->b_efunc = NULL; 3673168404Spjd buf->b_private = NULL; 3674185029Spjd 3675185029Spjd if (l2hdr) { 3676252140Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3677251419Ssmh trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3678251419Ssmh hdr->b_size, 0); 3679185029Spjd kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3680185029Spjd ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3681185029Spjd mutex_exit(&l2arc_buflist_mtx); 3682185029Spjd } 3683168404Spjd} 3684168404Spjd 3685168404Spjdint 3686168404Spjdarc_released(arc_buf_t *buf) 3687168404Spjd{ 3688185029Spjd int released; 3689185029Spjd 3690219089Spjd mutex_enter(&buf->b_evict_lock); 3691185029Spjd released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3692219089Spjd mutex_exit(&buf->b_evict_lock); 3693185029Spjd return (released); 3694168404Spjd} 3695168404Spjd 3696168404Spjdint 3697168404Spjdarc_has_callback(arc_buf_t *buf) 3698168404Spjd{ 3699185029Spjd int callback; 3700185029Spjd 3701219089Spjd mutex_enter(&buf->b_evict_lock); 3702185029Spjd callback = (buf->b_efunc != NULL); 3703219089Spjd mutex_exit(&buf->b_evict_lock); 3704185029Spjd return (callback); 3705168404Spjd} 3706168404Spjd 3707168404Spjd#ifdef ZFS_DEBUG 3708168404Spjdint 3709168404Spjdarc_referenced(arc_buf_t *buf) 3710168404Spjd{ 3711185029Spjd int referenced; 3712185029Spjd 3713219089Spjd mutex_enter(&buf->b_evict_lock); 3714185029Spjd referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3715219089Spjd mutex_exit(&buf->b_evict_lock); 3716185029Spjd return (referenced); 3717168404Spjd} 3718168404Spjd#endif 3719168404Spjd 3720168404Spjdstatic void 3721168404Spjdarc_write_ready(zio_t *zio) 3722168404Spjd{ 3723168404Spjd arc_write_callback_t *callback = zio->io_private; 3724168404Spjd arc_buf_t *buf = callback->awcb_buf; 3725185029Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3726168404Spjd 3727185029Spjd ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3728185029Spjd callback->awcb_ready(zio, buf, callback->awcb_private); 3729185029Spjd 3730185029Spjd /* 3731185029Spjd * If the IO is already in progress, then this is a re-write 3732185029Spjd * attempt, so we need to thaw and re-compute the cksum. 3733185029Spjd * It is the responsibility of the callback to handle the 3734185029Spjd * accounting for any re-write attempt. 3735185029Spjd */ 3736185029Spjd if (HDR_IO_IN_PROGRESS(hdr)) { 3737185029Spjd mutex_enter(&hdr->b_freeze_lock); 3738185029Spjd if (hdr->b_freeze_cksum != NULL) { 3739185029Spjd kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3740185029Spjd hdr->b_freeze_cksum = NULL; 3741185029Spjd } 3742185029Spjd mutex_exit(&hdr->b_freeze_lock); 3743168404Spjd } 3744185029Spjd arc_cksum_compute(buf, B_FALSE); 3745185029Spjd hdr->b_flags |= ARC_IO_IN_PROGRESS; 3746168404Spjd} 3747168404Spjd 3748260764Savg/* 3749260764Savg * The SPA calls this callback for each physical write that happens on behalf 3750260764Savg * of a logical write. See the comment in dbuf_write_physdone() for details. 3751260764Savg */ 3752168404Spjdstatic void 3753260764Savgarc_write_physdone(zio_t *zio) 3754260764Savg{ 3755260764Savg arc_write_callback_t *cb = zio->io_private; 3756260764Savg if (cb->awcb_physdone != NULL) 3757260764Savg cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3758260764Savg} 3759260764Savg 3760260764Savgstatic void 3761168404Spjdarc_write_done(zio_t *zio) 3762168404Spjd{ 3763168404Spjd arc_write_callback_t *callback = zio->io_private; 3764168404Spjd arc_buf_t *buf = callback->awcb_buf; 3765168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3766168404Spjd 3767219089Spjd ASSERT(hdr->b_acb == NULL); 3768168404Spjd 3769219089Spjd if (zio->io_error == 0) { 3770263398Sdelphij if (BP_IS_HOLE(zio->io_bp)) { 3771263398Sdelphij buf_discard_identity(hdr); 3772263398Sdelphij } else { 3773263398Sdelphij hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3774263398Sdelphij hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3775263398Sdelphij hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3776263398Sdelphij } 3777219089Spjd } else { 3778219089Spjd ASSERT(BUF_EMPTY(hdr)); 3779219089Spjd } 3780219089Spjd 3781168404Spjd /* 3782168404Spjd * If the block to be written was all-zero, we may have 3783168404Spjd * compressed it away. In this case no write was performed 3784219089Spjd * so there will be no dva/birth/checksum. The buffer must 3785219089Spjd * therefore remain anonymous (and uncached). 3786168404Spjd */ 3787168404Spjd if (!BUF_EMPTY(hdr)) { 3788168404Spjd arc_buf_hdr_t *exists; 3789168404Spjd kmutex_t *hash_lock; 3790168404Spjd 3791219089Spjd ASSERT(zio->io_error == 0); 3792219089Spjd 3793168404Spjd arc_cksum_verify(buf); 3794168404Spjd 3795168404Spjd exists = buf_hash_insert(hdr, &hash_lock); 3796168404Spjd if (exists) { 3797168404Spjd /* 3798168404Spjd * This can only happen if we overwrite for 3799168404Spjd * sync-to-convergence, because we remove 3800168404Spjd * buffers from the hash table when we arc_free(). 3801168404Spjd */ 3802219089Spjd if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3803219089Spjd if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3804219089Spjd panic("bad overwrite, hdr=%p exists=%p", 3805219089Spjd (void *)hdr, (void *)exists); 3806219089Spjd ASSERT(refcount_is_zero(&exists->b_refcnt)); 3807219089Spjd arc_change_state(arc_anon, exists, hash_lock); 3808219089Spjd mutex_exit(hash_lock); 3809219089Spjd arc_hdr_destroy(exists); 3810219089Spjd exists = buf_hash_insert(hdr, &hash_lock); 3811219089Spjd ASSERT3P(exists, ==, NULL); 3812244087Smm } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3813244087Smm /* nopwrite */ 3814244087Smm ASSERT(zio->io_prop.zp_nopwrite); 3815244087Smm if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3816244087Smm panic("bad nopwrite, hdr=%p exists=%p", 3817244087Smm (void *)hdr, (void *)exists); 3818219089Spjd } else { 3819219089Spjd /* Dedup */ 3820219089Spjd ASSERT(hdr->b_datacnt == 1); 3821219089Spjd ASSERT(hdr->b_state == arc_anon); 3822219089Spjd ASSERT(BP_GET_DEDUP(zio->io_bp)); 3823219089Spjd ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3824219089Spjd } 3825168404Spjd } 3826168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3827185029Spjd /* if it's not anon, we are doing a scrub */ 3828219089Spjd if (!exists && hdr->b_state == arc_anon) 3829185029Spjd arc_access(hdr, hash_lock); 3830168404Spjd mutex_exit(hash_lock); 3831168404Spjd } else { 3832168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3833168404Spjd } 3834168404Spjd 3835219089Spjd ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3836219089Spjd callback->awcb_done(zio, buf, callback->awcb_private); 3837168404Spjd 3838168404Spjd kmem_free(callback, sizeof (arc_write_callback_t)); 3839168404Spjd} 3840168404Spjd 3841168404Spjdzio_t * 3842219089Spjdarc_write(zio_t *pio, spa_t *spa, uint64_t txg, 3843252140Sdelphij blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 3844260764Savg const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 3845260764Savg arc_done_func_t *done, void *private, zio_priority_t priority, 3846260764Savg int zio_flags, const zbookmark_t *zb) 3847168404Spjd{ 3848168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3849168404Spjd arc_write_callback_t *callback; 3850185029Spjd zio_t *zio; 3851168404Spjd 3852185029Spjd ASSERT(ready != NULL); 3853219089Spjd ASSERT(done != NULL); 3854168404Spjd ASSERT(!HDR_IO_ERROR(hdr)); 3855168404Spjd ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3856219089Spjd ASSERT(hdr->b_acb == NULL); 3857185029Spjd if (l2arc) 3858185029Spjd hdr->b_flags |= ARC_L2CACHE; 3859252140Sdelphij if (l2arc_compress) 3860252140Sdelphij hdr->b_flags |= ARC_L2COMPRESS; 3861168404Spjd callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 3862168404Spjd callback->awcb_ready = ready; 3863260764Savg callback->awcb_physdone = physdone; 3864168404Spjd callback->awcb_done = done; 3865168404Spjd callback->awcb_private = private; 3866168404Spjd callback->awcb_buf = buf; 3867168404Spjd 3868219089Spjd zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 3869260764Savg arc_write_ready, arc_write_physdone, arc_write_done, callback, 3870260764Savg priority, zio_flags, zb); 3871185029Spjd 3872168404Spjd return (zio); 3873168404Spjd} 3874168404Spjd 3875185029Spjdstatic int 3876260764Savgarc_memory_throttle(uint64_t reserve, uint64_t txg) 3877185029Spjd{ 3878185029Spjd#ifdef _KERNEL 3879219089Spjd uint64_t available_memory = 3880219089Spjd ptoa((uintmax_t)cnt.v_free_count + cnt.v_cache_count); 3881185029Spjd static uint64_t page_load = 0; 3882185029Spjd static uint64_t last_txg = 0; 3883185029Spjd 3884219089Spjd#ifdef sun 3885185029Spjd#if defined(__i386) 3886185029Spjd available_memory = 3887185029Spjd MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 3888185029Spjd#endif 3889219089Spjd#endif /* sun */ 3890260764Savg 3891260764Savg if (cnt.v_free_count + cnt.v_cache_count > 3892260764Savg (uint64_t)physmem * arc_lotsfree_percent / 100) 3893185029Spjd return (0); 3894185029Spjd 3895185029Spjd if (txg > last_txg) { 3896185029Spjd last_txg = txg; 3897185029Spjd page_load = 0; 3898185029Spjd } 3899185029Spjd /* 3900185029Spjd * If we are in pageout, we know that memory is already tight, 3901185029Spjd * the arc is already going to be evicting, so we just want to 3902185029Spjd * continue to let page writes occur as quickly as possible. 3903185029Spjd */ 3904185029Spjd if (curproc == pageproc) { 3905185029Spjd if (page_load > available_memory / 4) 3906249643Smm return (SET_ERROR(ERESTART)); 3907185029Spjd /* Note: reserve is inflated, so we deflate */ 3908185029Spjd page_load += reserve / 8; 3909185029Spjd return (0); 3910185029Spjd } else if (page_load > 0 && arc_reclaim_needed()) { 3911185029Spjd /* memory is low, delay before restarting */ 3912185029Spjd ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3913249643Smm return (SET_ERROR(EAGAIN)); 3914185029Spjd } 3915185029Spjd page_load = 0; 3916185029Spjd#endif 3917185029Spjd return (0); 3918185029Spjd} 3919185029Spjd 3920168404Spjdvoid 3921185029Spjdarc_tempreserve_clear(uint64_t reserve) 3922168404Spjd{ 3923185029Spjd atomic_add_64(&arc_tempreserve, -reserve); 3924168404Spjd ASSERT((int64_t)arc_tempreserve >= 0); 3925168404Spjd} 3926168404Spjd 3927168404Spjdint 3928185029Spjdarc_tempreserve_space(uint64_t reserve, uint64_t txg) 3929168404Spjd{ 3930185029Spjd int error; 3931209962Smm uint64_t anon_size; 3932185029Spjd 3933185029Spjd if (reserve > arc_c/4 && !arc_no_grow) 3934185029Spjd arc_c = MIN(arc_c_max, reserve * 4); 3935185029Spjd if (reserve > arc_c) 3936249643Smm return (SET_ERROR(ENOMEM)); 3937168404Spjd 3938168404Spjd /* 3939209962Smm * Don't count loaned bufs as in flight dirty data to prevent long 3940209962Smm * network delays from blocking transactions that are ready to be 3941209962Smm * assigned to a txg. 3942209962Smm */ 3943209962Smm anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 3944209962Smm 3945209962Smm /* 3946185029Spjd * Writes will, almost always, require additional memory allocations 3947252751Sdelphij * in order to compress/encrypt/etc the data. We therefore need to 3948185029Spjd * make sure that there is sufficient available memory for this. 3949185029Spjd */ 3950260764Savg error = arc_memory_throttle(reserve, txg); 3951260764Savg if (error != 0) 3952185029Spjd return (error); 3953185029Spjd 3954185029Spjd /* 3955168404Spjd * Throttle writes when the amount of dirty data in the cache 3956168404Spjd * gets too large. We try to keep the cache less than half full 3957168404Spjd * of dirty blocks so that our sync times don't grow too large. 3958168404Spjd * Note: if two requests come in concurrently, we might let them 3959168404Spjd * both succeed, when one of them should fail. Not a huge deal. 3960168404Spjd */ 3961209962Smm 3962209962Smm if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 3963209962Smm anon_size > arc_c / 4) { 3964185029Spjd dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 3965185029Spjd "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 3966185029Spjd arc_tempreserve>>10, 3967185029Spjd arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 3968185029Spjd arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 3969185029Spjd reserve>>10, arc_c>>10); 3970249643Smm return (SET_ERROR(ERESTART)); 3971168404Spjd } 3972185029Spjd atomic_add_64(&arc_tempreserve, reserve); 3973168404Spjd return (0); 3974168404Spjd} 3975168404Spjd 3976168582Spjdstatic kmutex_t arc_lowmem_lock; 3977168404Spjd#ifdef _KERNEL 3978168566Spjdstatic eventhandler_tag arc_event_lowmem = NULL; 3979168404Spjd 3980168404Spjdstatic void 3981168566Spjdarc_lowmem(void *arg __unused, int howto __unused) 3982168404Spjd{ 3983168404Spjd 3984168566Spjd /* Serialize access via arc_lowmem_lock. */ 3985168566Spjd mutex_enter(&arc_lowmem_lock); 3986219089Spjd mutex_enter(&arc_reclaim_thr_lock); 3987185029Spjd needfree = 1; 3988168404Spjd cv_signal(&arc_reclaim_thr_cv); 3989242858Savg 3990242858Savg /* 3991242858Savg * It is unsafe to block here in arbitrary threads, because we can come 3992242858Savg * here from ARC itself and may hold ARC locks and thus risk a deadlock 3993242858Savg * with ARC reclaim thread. 3994242858Savg */ 3995242858Savg if (curproc == pageproc) { 3996242858Savg while (needfree) 3997242858Savg msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 3998242858Savg } 3999219089Spjd mutex_exit(&arc_reclaim_thr_lock); 4000168566Spjd mutex_exit(&arc_lowmem_lock); 4001168404Spjd} 4002168404Spjd#endif 4003168404Spjd 4004168404Spjdvoid 4005168404Spjdarc_init(void) 4006168404Spjd{ 4007219089Spjd int i, prefetch_tunable_set = 0; 4008205231Skmacy 4009168404Spjd mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4010168404Spjd cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4011168566Spjd mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4012168404Spjd 4013168404Spjd /* Convert seconds to clock ticks */ 4014168404Spjd arc_min_prefetch_lifespan = 1 * hz; 4015168404Spjd 4016168404Spjd /* Start out with 1/8 of all memory */ 4017168566Spjd arc_c = kmem_size() / 8; 4018219089Spjd 4019219089Spjd#ifdef sun 4020192360Skmacy#ifdef _KERNEL 4021192360Skmacy /* 4022192360Skmacy * On architectures where the physical memory can be larger 4023192360Skmacy * than the addressable space (intel in 32-bit mode), we may 4024192360Skmacy * need to limit the cache to 1/8 of VM size. 4025192360Skmacy */ 4026192360Skmacy arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4027192360Skmacy#endif 4028219089Spjd#endif /* sun */ 4029168566Spjd /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4030168566Spjd arc_c_min = MAX(arc_c / 4, 64<<18); 4031168566Spjd /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4032168404Spjd if (arc_c * 8 >= 1<<30) 4033168404Spjd arc_c_max = (arc_c * 8) - (1<<30); 4034168404Spjd else 4035168404Spjd arc_c_max = arc_c_min; 4036175633Spjd arc_c_max = MAX(arc_c * 5, arc_c_max); 4037219089Spjd 4038168481Spjd#ifdef _KERNEL 4039168404Spjd /* 4040168404Spjd * Allow the tunables to override our calculations if they are 4041168566Spjd * reasonable (ie. over 16MB) 4042168404Spjd */ 4043219089Spjd if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size()) 4044168404Spjd arc_c_max = zfs_arc_max; 4045219089Spjd if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max) 4046168404Spjd arc_c_min = zfs_arc_min; 4047168481Spjd#endif 4048219089Spjd 4049168404Spjd arc_c = arc_c_max; 4050168404Spjd arc_p = (arc_c >> 1); 4051168404Spjd 4052185029Spjd /* limit meta-data to 1/4 of the arc capacity */ 4053185029Spjd arc_meta_limit = arc_c_max / 4; 4054185029Spjd 4055185029Spjd /* Allow the tunable to override if it is reasonable */ 4056185029Spjd if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4057185029Spjd arc_meta_limit = zfs_arc_meta_limit; 4058185029Spjd 4059185029Spjd if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4060185029Spjd arc_c_min = arc_meta_limit / 2; 4061185029Spjd 4062208373Smm if (zfs_arc_grow_retry > 0) 4063208373Smm arc_grow_retry = zfs_arc_grow_retry; 4064208373Smm 4065208373Smm if (zfs_arc_shrink_shift > 0) 4066208373Smm arc_shrink_shift = zfs_arc_shrink_shift; 4067208373Smm 4068208373Smm if (zfs_arc_p_min_shift > 0) 4069208373Smm arc_p_min_shift = zfs_arc_p_min_shift; 4070208373Smm 4071168404Spjd /* if kmem_flags are set, lets try to use less memory */ 4072168404Spjd if (kmem_debugging()) 4073168404Spjd arc_c = arc_c / 2; 4074168404Spjd if (arc_c < arc_c_min) 4075168404Spjd arc_c = arc_c_min; 4076168404Spjd 4077168473Spjd zfs_arc_min = arc_c_min; 4078168473Spjd zfs_arc_max = arc_c_max; 4079168473Spjd 4080168404Spjd arc_anon = &ARC_anon; 4081168404Spjd arc_mru = &ARC_mru; 4082168404Spjd arc_mru_ghost = &ARC_mru_ghost; 4083168404Spjd arc_mfu = &ARC_mfu; 4084168404Spjd arc_mfu_ghost = &ARC_mfu_ghost; 4085185029Spjd arc_l2c_only = &ARC_l2c_only; 4086168404Spjd arc_size = 0; 4087168404Spjd 4088205231Skmacy for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4089205231Skmacy mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4090205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4091205231Skmacy mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4092205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4093205231Skmacy mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4094205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4095205231Skmacy mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4096205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4097205231Skmacy mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4098205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4099205231Skmacy mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4100205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4101206796Spjd 4102205231Skmacy list_create(&arc_mru->arcs_lists[i], 4103205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4104205231Skmacy list_create(&arc_mru_ghost->arcs_lists[i], 4105205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4106205231Skmacy list_create(&arc_mfu->arcs_lists[i], 4107205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4108205231Skmacy list_create(&arc_mfu_ghost->arcs_lists[i], 4109205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4110205231Skmacy list_create(&arc_mfu_ghost->arcs_lists[i], 4111205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4112205231Skmacy list_create(&arc_l2c_only->arcs_lists[i], 4113205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4114205231Skmacy } 4115168404Spjd 4116168404Spjd buf_init(); 4117168404Spjd 4118168404Spjd arc_thread_exit = 0; 4119168404Spjd arc_eviction_list = NULL; 4120168404Spjd mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4121168404Spjd bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4122168404Spjd 4123168404Spjd arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4124168404Spjd sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4125168404Spjd 4126168404Spjd if (arc_ksp != NULL) { 4127168404Spjd arc_ksp->ks_data = &arc_stats; 4128168404Spjd kstat_install(arc_ksp); 4129168404Spjd } 4130168404Spjd 4131168404Spjd (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4132168404Spjd TS_RUN, minclsyspri); 4133168404Spjd 4134168404Spjd#ifdef _KERNEL 4135168566Spjd arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4136168404Spjd EVENTHANDLER_PRI_FIRST); 4137168404Spjd#endif 4138168404Spjd 4139168404Spjd arc_dead = FALSE; 4140185029Spjd arc_warm = B_FALSE; 4141168566Spjd 4142260764Savg /* 4143260764Savg * Calculate maximum amount of dirty data per pool. 4144260764Savg * 4145260764Savg * If it has been set by /etc/system, take that. 4146260764Savg * Otherwise, use a percentage of physical memory defined by 4147260764Savg * zfs_dirty_data_max_percent (default 10%) with a cap at 4148260764Savg * zfs_dirty_data_max_max (default 4GB). 4149260764Savg */ 4150260764Savg if (zfs_dirty_data_max == 0) { 4151260764Savg zfs_dirty_data_max = ptob(physmem) * 4152260764Savg zfs_dirty_data_max_percent / 100; 4153260764Savg zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4154260764Savg zfs_dirty_data_max_max); 4155260764Savg } 4156185029Spjd 4157168566Spjd#ifdef _KERNEL 4158194043Skmacy if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4159193953Skmacy prefetch_tunable_set = 1; 4160206796Spjd 4161193878Skmacy#ifdef __i386__ 4162193953Skmacy if (prefetch_tunable_set == 0) { 4163196863Strasz printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4164196863Strasz "-- to enable,\n"); 4165196863Strasz printf(" add \"vfs.zfs.prefetch_disable=0\" " 4166196863Strasz "to /boot/loader.conf.\n"); 4167219089Spjd zfs_prefetch_disable = 1; 4168193878Skmacy } 4169206796Spjd#else 4170193878Skmacy if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4171193953Skmacy prefetch_tunable_set == 0) { 4172196863Strasz printf("ZFS NOTICE: Prefetch is disabled by default if less " 4173196941Strasz "than 4GB of RAM is present;\n" 4174196863Strasz " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4175196863Strasz "to /boot/loader.conf.\n"); 4176219089Spjd zfs_prefetch_disable = 1; 4177193878Skmacy } 4178206796Spjd#endif 4179175633Spjd /* Warn about ZFS memory and address space requirements. */ 4180168696Spjd if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4181168987Sbmah printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4182168987Sbmah "expect unstable behavior.\n"); 4183175633Spjd } 4184175633Spjd if (kmem_size() < 512 * (1 << 20)) { 4185173419Spjd printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4186168987Sbmah "expect unstable behavior.\n"); 4187185029Spjd printf(" Consider tuning vm.kmem_size and " 4188173419Spjd "vm.kmem_size_max\n"); 4189185029Spjd printf(" in /boot/loader.conf.\n"); 4190168566Spjd } 4191168566Spjd#endif 4192168404Spjd} 4193168404Spjd 4194168404Spjdvoid 4195168404Spjdarc_fini(void) 4196168404Spjd{ 4197205231Skmacy int i; 4198206796Spjd 4199168404Spjd mutex_enter(&arc_reclaim_thr_lock); 4200168404Spjd arc_thread_exit = 1; 4201168404Spjd cv_signal(&arc_reclaim_thr_cv); 4202168404Spjd while (arc_thread_exit != 0) 4203168404Spjd cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4204168404Spjd mutex_exit(&arc_reclaim_thr_lock); 4205168404Spjd 4206185029Spjd arc_flush(NULL); 4207168404Spjd 4208168404Spjd arc_dead = TRUE; 4209168404Spjd 4210168404Spjd if (arc_ksp != NULL) { 4211168404Spjd kstat_delete(arc_ksp); 4212168404Spjd arc_ksp = NULL; 4213168404Spjd } 4214168404Spjd 4215168404Spjd mutex_destroy(&arc_eviction_mtx); 4216168404Spjd mutex_destroy(&arc_reclaim_thr_lock); 4217168404Spjd cv_destroy(&arc_reclaim_thr_cv); 4218168404Spjd 4219205231Skmacy for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4220205231Skmacy list_destroy(&arc_mru->arcs_lists[i]); 4221205231Skmacy list_destroy(&arc_mru_ghost->arcs_lists[i]); 4222205231Skmacy list_destroy(&arc_mfu->arcs_lists[i]); 4223205231Skmacy list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4224206795Spjd list_destroy(&arc_l2c_only->arcs_lists[i]); 4225168404Spjd 4226205231Skmacy mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4227205231Skmacy mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4228205231Skmacy mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4229205231Skmacy mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4230205231Skmacy mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4231206795Spjd mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4232205231Skmacy } 4233206796Spjd 4234168404Spjd buf_fini(); 4235168404Spjd 4236209962Smm ASSERT(arc_loaned_bytes == 0); 4237209962Smm 4238168582Spjd mutex_destroy(&arc_lowmem_lock); 4239168404Spjd#ifdef _KERNEL 4240168566Spjd if (arc_event_lowmem != NULL) 4241168566Spjd EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4242168404Spjd#endif 4243168404Spjd} 4244185029Spjd 4245185029Spjd/* 4246185029Spjd * Level 2 ARC 4247185029Spjd * 4248185029Spjd * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4249185029Spjd * It uses dedicated storage devices to hold cached data, which are populated 4250185029Spjd * using large infrequent writes. The main role of this cache is to boost 4251185029Spjd * the performance of random read workloads. The intended L2ARC devices 4252185029Spjd * include short-stroked disks, solid state disks, and other media with 4253185029Spjd * substantially faster read latency than disk. 4254185029Spjd * 4255185029Spjd * +-----------------------+ 4256185029Spjd * | ARC | 4257185029Spjd * +-----------------------+ 4258185029Spjd * | ^ ^ 4259185029Spjd * | | | 4260185029Spjd * l2arc_feed_thread() arc_read() 4261185029Spjd * | | | 4262185029Spjd * | l2arc read | 4263185029Spjd * V | | 4264185029Spjd * +---------------+ | 4265185029Spjd * | L2ARC | | 4266185029Spjd * +---------------+ | 4267185029Spjd * | ^ | 4268185029Spjd * l2arc_write() | | 4269185029Spjd * | | | 4270185029Spjd * V | | 4271185029Spjd * +-------+ +-------+ 4272185029Spjd * | vdev | | vdev | 4273185029Spjd * | cache | | cache | 4274185029Spjd * +-------+ +-------+ 4275185029Spjd * +=========+ .-----. 4276185029Spjd * : L2ARC : |-_____-| 4277185029Spjd * : devices : | Disks | 4278185029Spjd * +=========+ `-_____-' 4279185029Spjd * 4280185029Spjd * Read requests are satisfied from the following sources, in order: 4281185029Spjd * 4282185029Spjd * 1) ARC 4283185029Spjd * 2) vdev cache of L2ARC devices 4284185029Spjd * 3) L2ARC devices 4285185029Spjd * 4) vdev cache of disks 4286185029Spjd * 5) disks 4287185029Spjd * 4288185029Spjd * Some L2ARC device types exhibit extremely slow write performance. 4289185029Spjd * To accommodate for this there are some significant differences between 4290185029Spjd * the L2ARC and traditional cache design: 4291185029Spjd * 4292185029Spjd * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4293185029Spjd * the ARC behave as usual, freeing buffers and placing headers on ghost 4294185029Spjd * lists. The ARC does not send buffers to the L2ARC during eviction as 4295185029Spjd * this would add inflated write latencies for all ARC memory pressure. 4296185029Spjd * 4297185029Spjd * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4298185029Spjd * It does this by periodically scanning buffers from the eviction-end of 4299185029Spjd * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4300252140Sdelphij * not already there. It scans until a headroom of buffers is satisfied, 4301252140Sdelphij * which itself is a buffer for ARC eviction. If a compressible buffer is 4302252140Sdelphij * found during scanning and selected for writing to an L2ARC device, we 4303252140Sdelphij * temporarily boost scanning headroom during the next scan cycle to make 4304252140Sdelphij * sure we adapt to compression effects (which might significantly reduce 4305252140Sdelphij * the data volume we write to L2ARC). The thread that does this is 4306185029Spjd * l2arc_feed_thread(), illustrated below; example sizes are included to 4307185029Spjd * provide a better sense of ratio than this diagram: 4308185029Spjd * 4309185029Spjd * head --> tail 4310185029Spjd * +---------------------+----------+ 4311185029Spjd * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4312185029Spjd * +---------------------+----------+ | o L2ARC eligible 4313185029Spjd * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4314185029Spjd * +---------------------+----------+ | 4315185029Spjd * 15.9 Gbytes ^ 32 Mbytes | 4316185029Spjd * headroom | 4317185029Spjd * l2arc_feed_thread() 4318185029Spjd * | 4319185029Spjd * l2arc write hand <--[oooo]--' 4320185029Spjd * | 8 Mbyte 4321185029Spjd * | write max 4322185029Spjd * V 4323185029Spjd * +==============================+ 4324185029Spjd * L2ARC dev |####|#|###|###| |####| ... | 4325185029Spjd * +==============================+ 4326185029Spjd * 32 Gbytes 4327185029Spjd * 4328185029Spjd * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4329185029Spjd * evicted, then the L2ARC has cached a buffer much sooner than it probably 4330185029Spjd * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4331185029Spjd * safe to say that this is an uncommon case, since buffers at the end of 4332185029Spjd * the ARC lists have moved there due to inactivity. 4333185029Spjd * 4334185029Spjd * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4335185029Spjd * then the L2ARC simply misses copying some buffers. This serves as a 4336185029Spjd * pressure valve to prevent heavy read workloads from both stalling the ARC 4337185029Spjd * with waits and clogging the L2ARC with writes. This also helps prevent 4338185029Spjd * the potential for the L2ARC to churn if it attempts to cache content too 4339185029Spjd * quickly, such as during backups of the entire pool. 4340185029Spjd * 4341185029Spjd * 5. After system boot and before the ARC has filled main memory, there are 4342185029Spjd * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4343185029Spjd * lists can remain mostly static. Instead of searching from tail of these 4344185029Spjd * lists as pictured, the l2arc_feed_thread() will search from the list heads 4345185029Spjd * for eligible buffers, greatly increasing its chance of finding them. 4346185029Spjd * 4347185029Spjd * The L2ARC device write speed is also boosted during this time so that 4348185029Spjd * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4349185029Spjd * there are no L2ARC reads, and no fear of degrading read performance 4350185029Spjd * through increased writes. 4351185029Spjd * 4352185029Spjd * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4353185029Spjd * the vdev queue can aggregate them into larger and fewer writes. Each 4354185029Spjd * device is written to in a rotor fashion, sweeping writes through 4355185029Spjd * available space then repeating. 4356185029Spjd * 4357185029Spjd * 7. The L2ARC does not store dirty content. It never needs to flush 4358185029Spjd * write buffers back to disk based storage. 4359185029Spjd * 4360185029Spjd * 8. If an ARC buffer is written (and dirtied) which also exists in the 4361185029Spjd * L2ARC, the now stale L2ARC buffer is immediately dropped. 4362185029Spjd * 4363185029Spjd * The performance of the L2ARC can be tweaked by a number of tunables, which 4364185029Spjd * may be necessary for different workloads: 4365185029Spjd * 4366185029Spjd * l2arc_write_max max write bytes per interval 4367185029Spjd * l2arc_write_boost extra write bytes during device warmup 4368185029Spjd * l2arc_noprefetch skip caching prefetched buffers 4369185029Spjd * l2arc_headroom number of max device writes to precache 4370252140Sdelphij * l2arc_headroom_boost when we find compressed buffers during ARC 4371252140Sdelphij * scanning, we multiply headroom by this 4372252140Sdelphij * percentage factor for the next scan cycle, 4373252140Sdelphij * since more compressed buffers are likely to 4374252140Sdelphij * be present 4375185029Spjd * l2arc_feed_secs seconds between L2ARC writing 4376185029Spjd * 4377185029Spjd * Tunables may be removed or added as future performance improvements are 4378185029Spjd * integrated, and also may become zpool properties. 4379208373Smm * 4380208373Smm * There are three key functions that control how the L2ARC warms up: 4381208373Smm * 4382208373Smm * l2arc_write_eligible() check if a buffer is eligible to cache 4383208373Smm * l2arc_write_size() calculate how much to write 4384208373Smm * l2arc_write_interval() calculate sleep delay between writes 4385208373Smm * 4386208373Smm * These three functions determine what to write, how much, and how quickly 4387208373Smm * to send writes. 4388185029Spjd */ 4389185029Spjd 4390208373Smmstatic boolean_t 4391209962Smml2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab) 4392208373Smm{ 4393208373Smm /* 4394208373Smm * A buffer is *not* eligible for the L2ARC if it: 4395208373Smm * 1. belongs to a different spa. 4396208373Smm * 2. is already cached on the L2ARC. 4397208373Smm * 3. has an I/O in progress (it may be an incomplete read). 4398208373Smm * 4. is flagged not eligible (zfs property). 4399208373Smm */ 4400209962Smm if (ab->b_spa != spa_guid) { 4401208373Smm ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4402208373Smm return (B_FALSE); 4403208373Smm } 4404208373Smm if (ab->b_l2hdr != NULL) { 4405208373Smm ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4406208373Smm return (B_FALSE); 4407208373Smm } 4408208373Smm if (HDR_IO_IN_PROGRESS(ab)) { 4409208373Smm ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4410208373Smm return (B_FALSE); 4411208373Smm } 4412208373Smm if (!HDR_L2CACHE(ab)) { 4413208373Smm ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4414208373Smm return (B_FALSE); 4415208373Smm } 4416208373Smm 4417208373Smm return (B_TRUE); 4418208373Smm} 4419208373Smm 4420208373Smmstatic uint64_t 4421252140Sdelphijl2arc_write_size(void) 4422208373Smm{ 4423208373Smm uint64_t size; 4424208373Smm 4425252140Sdelphij /* 4426252140Sdelphij * Make sure our globals have meaningful values in case the user 4427252140Sdelphij * altered them. 4428252140Sdelphij */ 4429252140Sdelphij size = l2arc_write_max; 4430252140Sdelphij if (size == 0) { 4431252140Sdelphij cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4432252140Sdelphij "be greater than zero, resetting it to the default (%d)", 4433252140Sdelphij L2ARC_WRITE_SIZE); 4434252140Sdelphij size = l2arc_write_max = L2ARC_WRITE_SIZE; 4435252140Sdelphij } 4436208373Smm 4437208373Smm if (arc_warm == B_FALSE) 4438252140Sdelphij size += l2arc_write_boost; 4439208373Smm 4440208373Smm return (size); 4441208373Smm 4442208373Smm} 4443208373Smm 4444208373Smmstatic clock_t 4445208373Smml2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4446208373Smm{ 4447219089Spjd clock_t interval, next, now; 4448208373Smm 4449208373Smm /* 4450208373Smm * If the ARC lists are busy, increase our write rate; if the 4451208373Smm * lists are stale, idle back. This is achieved by checking 4452208373Smm * how much we previously wrote - if it was more than half of 4453208373Smm * what we wanted, schedule the next write much sooner. 4454208373Smm */ 4455208373Smm if (l2arc_feed_again && wrote > (wanted / 2)) 4456208373Smm interval = (hz * l2arc_feed_min_ms) / 1000; 4457208373Smm else 4458208373Smm interval = hz * l2arc_feed_secs; 4459208373Smm 4460219089Spjd now = ddi_get_lbolt(); 4461219089Spjd next = MAX(now, MIN(now + interval, began + interval)); 4462208373Smm 4463208373Smm return (next); 4464208373Smm} 4465208373Smm 4466185029Spjdstatic void 4467185029Spjdl2arc_hdr_stat_add(void) 4468185029Spjd{ 4469185029Spjd ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4470185029Spjd ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4471185029Spjd} 4472185029Spjd 4473185029Spjdstatic void 4474185029Spjdl2arc_hdr_stat_remove(void) 4475185029Spjd{ 4476185029Spjd ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4477185029Spjd ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4478185029Spjd} 4479185029Spjd 4480185029Spjd/* 4481185029Spjd * Cycle through L2ARC devices. This is how L2ARC load balances. 4482185029Spjd * If a device is returned, this also returns holding the spa config lock. 4483185029Spjd */ 4484185029Spjdstatic l2arc_dev_t * 4485185029Spjdl2arc_dev_get_next(void) 4486185029Spjd{ 4487185029Spjd l2arc_dev_t *first, *next = NULL; 4488185029Spjd 4489185029Spjd /* 4490185029Spjd * Lock out the removal of spas (spa_namespace_lock), then removal 4491185029Spjd * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4492185029Spjd * both locks will be dropped and a spa config lock held instead. 4493185029Spjd */ 4494185029Spjd mutex_enter(&spa_namespace_lock); 4495185029Spjd mutex_enter(&l2arc_dev_mtx); 4496185029Spjd 4497185029Spjd /* if there are no vdevs, there is nothing to do */ 4498185029Spjd if (l2arc_ndev == 0) 4499185029Spjd goto out; 4500185029Spjd 4501185029Spjd first = NULL; 4502185029Spjd next = l2arc_dev_last; 4503185029Spjd do { 4504185029Spjd /* loop around the list looking for a non-faulted vdev */ 4505185029Spjd if (next == NULL) { 4506185029Spjd next = list_head(l2arc_dev_list); 4507185029Spjd } else { 4508185029Spjd next = list_next(l2arc_dev_list, next); 4509185029Spjd if (next == NULL) 4510185029Spjd next = list_head(l2arc_dev_list); 4511185029Spjd } 4512185029Spjd 4513185029Spjd /* if we have come back to the start, bail out */ 4514185029Spjd if (first == NULL) 4515185029Spjd first = next; 4516185029Spjd else if (next == first) 4517185029Spjd break; 4518185029Spjd 4519185029Spjd } while (vdev_is_dead(next->l2ad_vdev)); 4520185029Spjd 4521185029Spjd /* if we were unable to find any usable vdevs, return NULL */ 4522185029Spjd if (vdev_is_dead(next->l2ad_vdev)) 4523185029Spjd next = NULL; 4524185029Spjd 4525185029Spjd l2arc_dev_last = next; 4526185029Spjd 4527185029Spjdout: 4528185029Spjd mutex_exit(&l2arc_dev_mtx); 4529185029Spjd 4530185029Spjd /* 4531185029Spjd * Grab the config lock to prevent the 'next' device from being 4532185029Spjd * removed while we are writing to it. 4533185029Spjd */ 4534185029Spjd if (next != NULL) 4535185029Spjd spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4536185029Spjd mutex_exit(&spa_namespace_lock); 4537185029Spjd 4538185029Spjd return (next); 4539185029Spjd} 4540185029Spjd 4541185029Spjd/* 4542185029Spjd * Free buffers that were tagged for destruction. 4543185029Spjd */ 4544185029Spjdstatic void 4545185029Spjdl2arc_do_free_on_write() 4546185029Spjd{ 4547185029Spjd list_t *buflist; 4548185029Spjd l2arc_data_free_t *df, *df_prev; 4549185029Spjd 4550185029Spjd mutex_enter(&l2arc_free_on_write_mtx); 4551185029Spjd buflist = l2arc_free_on_write; 4552185029Spjd 4553185029Spjd for (df = list_tail(buflist); df; df = df_prev) { 4554185029Spjd df_prev = list_prev(buflist, df); 4555185029Spjd ASSERT(df->l2df_data != NULL); 4556185029Spjd ASSERT(df->l2df_func != NULL); 4557185029Spjd df->l2df_func(df->l2df_data, df->l2df_size); 4558185029Spjd list_remove(buflist, df); 4559185029Spjd kmem_free(df, sizeof (l2arc_data_free_t)); 4560185029Spjd } 4561185029Spjd 4562185029Spjd mutex_exit(&l2arc_free_on_write_mtx); 4563185029Spjd} 4564185029Spjd 4565185029Spjd/* 4566185029Spjd * A write to a cache device has completed. Update all headers to allow 4567185029Spjd * reads from these buffers to begin. 4568185029Spjd */ 4569185029Spjdstatic void 4570185029Spjdl2arc_write_done(zio_t *zio) 4571185029Spjd{ 4572185029Spjd l2arc_write_callback_t *cb; 4573185029Spjd l2arc_dev_t *dev; 4574185029Spjd list_t *buflist; 4575185029Spjd arc_buf_hdr_t *head, *ab, *ab_prev; 4576185029Spjd l2arc_buf_hdr_t *abl2; 4577185029Spjd kmutex_t *hash_lock; 4578185029Spjd 4579185029Spjd cb = zio->io_private; 4580185029Spjd ASSERT(cb != NULL); 4581185029Spjd dev = cb->l2wcb_dev; 4582185029Spjd ASSERT(dev != NULL); 4583185029Spjd head = cb->l2wcb_head; 4584185029Spjd ASSERT(head != NULL); 4585185029Spjd buflist = dev->l2ad_buflist; 4586185029Spjd ASSERT(buflist != NULL); 4587185029Spjd DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4588185029Spjd l2arc_write_callback_t *, cb); 4589185029Spjd 4590185029Spjd if (zio->io_error != 0) 4591185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_error); 4592185029Spjd 4593185029Spjd mutex_enter(&l2arc_buflist_mtx); 4594185029Spjd 4595185029Spjd /* 4596185029Spjd * All writes completed, or an error was hit. 4597185029Spjd */ 4598185029Spjd for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 4599185029Spjd ab_prev = list_prev(buflist, ab); 4600262116Savg abl2 = ab->b_l2hdr; 4601185029Spjd 4602262116Savg /* 4603262116Savg * Release the temporary compressed buffer as soon as possible. 4604262116Savg */ 4605262116Savg if (abl2->b_compress != ZIO_COMPRESS_OFF) 4606262116Savg l2arc_release_cdata_buf(ab); 4607262116Savg 4608185029Spjd hash_lock = HDR_LOCK(ab); 4609185029Spjd if (!mutex_tryenter(hash_lock)) { 4610185029Spjd /* 4611185029Spjd * This buffer misses out. It may be in a stage 4612185029Spjd * of eviction. Its ARC_L2_WRITING flag will be 4613185029Spjd * left set, denying reads to this buffer. 4614185029Spjd */ 4615185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4616185029Spjd continue; 4617185029Spjd } 4618185029Spjd 4619185029Spjd if (zio->io_error != 0) { 4620185029Spjd /* 4621185029Spjd * Error - drop L2ARC entry. 4622185029Spjd */ 4623185029Spjd list_remove(buflist, ab); 4624252140Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4625185029Spjd ab->b_l2hdr = NULL; 4626251419Ssmh trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4627251419Ssmh ab->b_size, 0); 4628185029Spjd kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4629185029Spjd ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4630185029Spjd } 4631185029Spjd 4632185029Spjd /* 4633185029Spjd * Allow ARC to begin reads to this L2ARC entry. 4634185029Spjd */ 4635185029Spjd ab->b_flags &= ~ARC_L2_WRITING; 4636185029Spjd 4637185029Spjd mutex_exit(hash_lock); 4638185029Spjd } 4639185029Spjd 4640185029Spjd atomic_inc_64(&l2arc_writes_done); 4641185029Spjd list_remove(buflist, head); 4642185029Spjd kmem_cache_free(hdr_cache, head); 4643185029Spjd mutex_exit(&l2arc_buflist_mtx); 4644185029Spjd 4645185029Spjd l2arc_do_free_on_write(); 4646185029Spjd 4647185029Spjd kmem_free(cb, sizeof (l2arc_write_callback_t)); 4648185029Spjd} 4649185029Spjd 4650185029Spjd/* 4651185029Spjd * A read to a cache device completed. Validate buffer contents before 4652185029Spjd * handing over to the regular ARC routines. 4653185029Spjd */ 4654185029Spjdstatic void 4655185029Spjdl2arc_read_done(zio_t *zio) 4656185029Spjd{ 4657185029Spjd l2arc_read_callback_t *cb; 4658185029Spjd arc_buf_hdr_t *hdr; 4659185029Spjd arc_buf_t *buf; 4660185029Spjd kmutex_t *hash_lock; 4661185029Spjd int equal; 4662185029Spjd 4663185029Spjd ASSERT(zio->io_vd != NULL); 4664185029Spjd ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4665185029Spjd 4666185029Spjd spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4667185029Spjd 4668185029Spjd cb = zio->io_private; 4669185029Spjd ASSERT(cb != NULL); 4670185029Spjd buf = cb->l2rcb_buf; 4671185029Spjd ASSERT(buf != NULL); 4672185029Spjd 4673219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 4674185029Spjd mutex_enter(hash_lock); 4675219089Spjd hdr = buf->b_hdr; 4676219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4677185029Spjd 4678185029Spjd /* 4679252140Sdelphij * If the buffer was compressed, decompress it first. 4680252140Sdelphij */ 4681252140Sdelphij if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4682252140Sdelphij l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4683252140Sdelphij ASSERT(zio->io_data != NULL); 4684252140Sdelphij 4685252140Sdelphij /* 4686185029Spjd * Check this survived the L2ARC journey. 4687185029Spjd */ 4688185029Spjd equal = arc_cksum_equal(buf); 4689185029Spjd if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4690185029Spjd mutex_exit(hash_lock); 4691185029Spjd zio->io_private = buf; 4692185029Spjd zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4693185029Spjd zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4694185029Spjd arc_read_done(zio); 4695185029Spjd } else { 4696185029Spjd mutex_exit(hash_lock); 4697185029Spjd /* 4698185029Spjd * Buffer didn't survive caching. Increment stats and 4699185029Spjd * reissue to the original storage device. 4700185029Spjd */ 4701185029Spjd if (zio->io_error != 0) { 4702185029Spjd ARCSTAT_BUMP(arcstat_l2_io_error); 4703185029Spjd } else { 4704249643Smm zio->io_error = SET_ERROR(EIO); 4705185029Spjd } 4706185029Spjd if (!equal) 4707185029Spjd ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4708185029Spjd 4709185029Spjd /* 4710185029Spjd * If there's no waiter, issue an async i/o to the primary 4711185029Spjd * storage now. If there *is* a waiter, the caller must 4712185029Spjd * issue the i/o in a context where it's OK to block. 4713185029Spjd */ 4714209962Smm if (zio->io_waiter == NULL) { 4715209962Smm zio_t *pio = zio_unique_parent(zio); 4716209962Smm 4717209962Smm ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4718209962Smm 4719209962Smm zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4720185029Spjd buf->b_data, zio->io_size, arc_read_done, buf, 4721185029Spjd zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4722209962Smm } 4723185029Spjd } 4724185029Spjd 4725185029Spjd kmem_free(cb, sizeof (l2arc_read_callback_t)); 4726185029Spjd} 4727185029Spjd 4728185029Spjd/* 4729185029Spjd * This is the list priority from which the L2ARC will search for pages to 4730185029Spjd * cache. This is used within loops (0..3) to cycle through lists in the 4731185029Spjd * desired order. This order can have a significant effect on cache 4732185029Spjd * performance. 4733185029Spjd * 4734185029Spjd * Currently the metadata lists are hit first, MFU then MRU, followed by 4735185029Spjd * the data lists. This function returns a locked list, and also returns 4736185029Spjd * the lock pointer. 4737185029Spjd */ 4738185029Spjdstatic list_t * 4739185029Spjdl2arc_list_locked(int list_num, kmutex_t **lock) 4740185029Spjd{ 4741248369Smm list_t *list = NULL; 4742205231Skmacy int idx; 4743185029Spjd 4744206796Spjd ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4745206796Spjd 4746205231Skmacy if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4747205231Skmacy idx = list_num; 4748205231Skmacy list = &arc_mfu->arcs_lists[idx]; 4749205231Skmacy *lock = ARCS_LOCK(arc_mfu, idx); 4750206796Spjd } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4751205231Skmacy idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4752205231Skmacy list = &arc_mru->arcs_lists[idx]; 4753205231Skmacy *lock = ARCS_LOCK(arc_mru, idx); 4754206796Spjd } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4755205231Skmacy ARC_BUFC_NUMDATALISTS)) { 4756205231Skmacy idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4757205231Skmacy list = &arc_mfu->arcs_lists[idx]; 4758205231Skmacy *lock = ARCS_LOCK(arc_mfu, idx); 4759205231Skmacy } else { 4760205231Skmacy idx = list_num - ARC_BUFC_NUMLISTS; 4761205231Skmacy list = &arc_mru->arcs_lists[idx]; 4762205231Skmacy *lock = ARCS_LOCK(arc_mru, idx); 4763185029Spjd } 4764185029Spjd 4765185029Spjd ASSERT(!(MUTEX_HELD(*lock))); 4766185029Spjd mutex_enter(*lock); 4767185029Spjd return (list); 4768185029Spjd} 4769185029Spjd 4770185029Spjd/* 4771185029Spjd * Evict buffers from the device write hand to the distance specified in 4772185029Spjd * bytes. This distance may span populated buffers, it may span nothing. 4773185029Spjd * This is clearing a region on the L2ARC device ready for writing. 4774185029Spjd * If the 'all' boolean is set, every buffer is evicted. 4775185029Spjd */ 4776185029Spjdstatic void 4777185029Spjdl2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4778185029Spjd{ 4779185029Spjd list_t *buflist; 4780185029Spjd l2arc_buf_hdr_t *abl2; 4781185029Spjd arc_buf_hdr_t *ab, *ab_prev; 4782185029Spjd kmutex_t *hash_lock; 4783185029Spjd uint64_t taddr; 4784185029Spjd 4785185029Spjd buflist = dev->l2ad_buflist; 4786185029Spjd 4787185029Spjd if (buflist == NULL) 4788185029Spjd return; 4789185029Spjd 4790185029Spjd if (!all && dev->l2ad_first) { 4791185029Spjd /* 4792185029Spjd * This is the first sweep through the device. There is 4793185029Spjd * nothing to evict. 4794185029Spjd */ 4795185029Spjd return; 4796185029Spjd } 4797185029Spjd 4798185029Spjd if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4799185029Spjd /* 4800185029Spjd * When nearing the end of the device, evict to the end 4801185029Spjd * before the device write hand jumps to the start. 4802185029Spjd */ 4803185029Spjd taddr = dev->l2ad_end; 4804185029Spjd } else { 4805185029Spjd taddr = dev->l2ad_hand + distance; 4806185029Spjd } 4807185029Spjd DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4808185029Spjd uint64_t, taddr, boolean_t, all); 4809185029Spjd 4810185029Spjdtop: 4811185029Spjd mutex_enter(&l2arc_buflist_mtx); 4812185029Spjd for (ab = list_tail(buflist); ab; ab = ab_prev) { 4813185029Spjd ab_prev = list_prev(buflist, ab); 4814185029Spjd 4815185029Spjd hash_lock = HDR_LOCK(ab); 4816185029Spjd if (!mutex_tryenter(hash_lock)) { 4817185029Spjd /* 4818185029Spjd * Missed the hash lock. Retry. 4819185029Spjd */ 4820185029Spjd ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4821185029Spjd mutex_exit(&l2arc_buflist_mtx); 4822185029Spjd mutex_enter(hash_lock); 4823185029Spjd mutex_exit(hash_lock); 4824185029Spjd goto top; 4825185029Spjd } 4826185029Spjd 4827185029Spjd if (HDR_L2_WRITE_HEAD(ab)) { 4828185029Spjd /* 4829185029Spjd * We hit a write head node. Leave it for 4830185029Spjd * l2arc_write_done(). 4831185029Spjd */ 4832185029Spjd list_remove(buflist, ab); 4833185029Spjd mutex_exit(hash_lock); 4834185029Spjd continue; 4835185029Spjd } 4836185029Spjd 4837185029Spjd if (!all && ab->b_l2hdr != NULL && 4838185029Spjd (ab->b_l2hdr->b_daddr > taddr || 4839185029Spjd ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4840185029Spjd /* 4841185029Spjd * We've evicted to the target address, 4842185029Spjd * or the end of the device. 4843185029Spjd */ 4844185029Spjd mutex_exit(hash_lock); 4845185029Spjd break; 4846185029Spjd } 4847185029Spjd 4848185029Spjd if (HDR_FREE_IN_PROGRESS(ab)) { 4849185029Spjd /* 4850185029Spjd * Already on the path to destruction. 4851185029Spjd */ 4852185029Spjd mutex_exit(hash_lock); 4853185029Spjd continue; 4854185029Spjd } 4855185029Spjd 4856185029Spjd if (ab->b_state == arc_l2c_only) { 4857185029Spjd ASSERT(!HDR_L2_READING(ab)); 4858185029Spjd /* 4859185029Spjd * This doesn't exist in the ARC. Destroy. 4860185029Spjd * arc_hdr_destroy() will call list_remove() 4861185029Spjd * and decrement arcstat_l2_size. 4862185029Spjd */ 4863185029Spjd arc_change_state(arc_anon, ab, hash_lock); 4864185029Spjd arc_hdr_destroy(ab); 4865185029Spjd } else { 4866185029Spjd /* 4867185029Spjd * Invalidate issued or about to be issued 4868185029Spjd * reads, since we may be about to write 4869185029Spjd * over this location. 4870185029Spjd */ 4871185029Spjd if (HDR_L2_READING(ab)) { 4872185029Spjd ARCSTAT_BUMP(arcstat_l2_evict_reading); 4873185029Spjd ab->b_flags |= ARC_L2_EVICTED; 4874185029Spjd } 4875185029Spjd 4876185029Spjd /* 4877185029Spjd * Tell ARC this no longer exists in L2ARC. 4878185029Spjd */ 4879185029Spjd if (ab->b_l2hdr != NULL) { 4880185029Spjd abl2 = ab->b_l2hdr; 4881252140Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4882185029Spjd ab->b_l2hdr = NULL; 4883185029Spjd kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4884185029Spjd ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4885185029Spjd } 4886185029Spjd list_remove(buflist, ab); 4887185029Spjd 4888185029Spjd /* 4889185029Spjd * This may have been leftover after a 4890185029Spjd * failed write. 4891185029Spjd */ 4892185029Spjd ab->b_flags &= ~ARC_L2_WRITING; 4893185029Spjd } 4894185029Spjd mutex_exit(hash_lock); 4895185029Spjd } 4896185029Spjd mutex_exit(&l2arc_buflist_mtx); 4897185029Spjd 4898219089Spjd vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0); 4899185029Spjd dev->l2ad_evict = taddr; 4900185029Spjd} 4901185029Spjd 4902185029Spjd/* 4903185029Spjd * Find and write ARC buffers to the L2ARC device. 4904185029Spjd * 4905185029Spjd * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 4906185029Spjd * for reading until they have completed writing. 4907252140Sdelphij * The headroom_boost is an in-out parameter used to maintain headroom boost 4908252140Sdelphij * state between calls to this function. 4909252140Sdelphij * 4910252140Sdelphij * Returns the number of bytes actually written (which may be smaller than 4911252140Sdelphij * the delta by which the device hand has changed due to alignment). 4912185029Spjd */ 4913208373Smmstatic uint64_t 4914252140Sdelphijl2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 4915252140Sdelphij boolean_t *headroom_boost) 4916185029Spjd{ 4917185029Spjd arc_buf_hdr_t *ab, *ab_prev, *head; 4918185029Spjd list_t *list; 4919252140Sdelphij uint64_t write_asize, write_psize, write_sz, headroom, 4920252140Sdelphij buf_compress_minsz; 4921185029Spjd void *buf_data; 4922252140Sdelphij kmutex_t *list_lock; 4923252140Sdelphij boolean_t full; 4924185029Spjd l2arc_write_callback_t *cb; 4925185029Spjd zio_t *pio, *wzio; 4926229578Smm uint64_t guid = spa_load_guid(spa); 4927252140Sdelphij const boolean_t do_headroom_boost = *headroom_boost; 4928185029Spjd int try; 4929185029Spjd 4930185029Spjd ASSERT(dev->l2ad_vdev != NULL); 4931185029Spjd 4932252140Sdelphij /* Lower the flag now, we might want to raise it again later. */ 4933252140Sdelphij *headroom_boost = B_FALSE; 4934252140Sdelphij 4935185029Spjd pio = NULL; 4936252140Sdelphij write_sz = write_asize = write_psize = 0; 4937185029Spjd full = B_FALSE; 4938185029Spjd head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 4939185029Spjd head->b_flags |= ARC_L2_WRITE_HEAD; 4940185029Spjd 4941205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 4942185029Spjd /* 4943252140Sdelphij * We will want to try to compress buffers that are at least 2x the 4944252140Sdelphij * device sector size. 4945252140Sdelphij */ 4946252140Sdelphij buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 4947252140Sdelphij 4948252140Sdelphij /* 4949185029Spjd * Copy buffers for L2ARC writing. 4950185029Spjd */ 4951185029Spjd mutex_enter(&l2arc_buflist_mtx); 4952206796Spjd for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 4953252140Sdelphij uint64_t passed_sz = 0; 4954252140Sdelphij 4955185029Spjd list = l2arc_list_locked(try, &list_lock); 4956205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 4957185029Spjd 4958185029Spjd /* 4959185029Spjd * L2ARC fast warmup. 4960185029Spjd * 4961185029Spjd * Until the ARC is warm and starts to evict, read from the 4962185029Spjd * head of the ARC lists rather than the tail. 4963185029Spjd */ 4964185029Spjd if (arc_warm == B_FALSE) 4965185029Spjd ab = list_head(list); 4966185029Spjd else 4967185029Spjd ab = list_tail(list); 4968206796Spjd if (ab == NULL) 4969205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 4970185029Spjd 4971252140Sdelphij headroom = target_sz * l2arc_headroom; 4972252140Sdelphij if (do_headroom_boost) 4973252140Sdelphij headroom = (headroom * l2arc_headroom_boost) / 100; 4974252140Sdelphij 4975185029Spjd for (; ab; ab = ab_prev) { 4976252140Sdelphij l2arc_buf_hdr_t *l2hdr; 4977252140Sdelphij kmutex_t *hash_lock; 4978252140Sdelphij uint64_t buf_sz; 4979252140Sdelphij 4980185029Spjd if (arc_warm == B_FALSE) 4981185029Spjd ab_prev = list_next(list, ab); 4982185029Spjd else 4983185029Spjd ab_prev = list_prev(list, ab); 4984205231Skmacy ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size); 4985206796Spjd 4986185029Spjd hash_lock = HDR_LOCK(ab); 4987252140Sdelphij if (!mutex_tryenter(hash_lock)) { 4988205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 4989185029Spjd /* 4990185029Spjd * Skip this buffer rather than waiting. 4991185029Spjd */ 4992185029Spjd continue; 4993185029Spjd } 4994185029Spjd 4995185029Spjd passed_sz += ab->b_size; 4996185029Spjd if (passed_sz > headroom) { 4997185029Spjd /* 4998185029Spjd * Searched too far. 4999185029Spjd */ 5000185029Spjd mutex_exit(hash_lock); 5001205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5002185029Spjd break; 5003185029Spjd } 5004185029Spjd 5005209962Smm if (!l2arc_write_eligible(guid, ab)) { 5006185029Spjd mutex_exit(hash_lock); 5007185029Spjd continue; 5008185029Spjd } 5009185029Spjd 5010185029Spjd if ((write_sz + ab->b_size) > target_sz) { 5011185029Spjd full = B_TRUE; 5012185029Spjd mutex_exit(hash_lock); 5013205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_full); 5014185029Spjd break; 5015185029Spjd } 5016185029Spjd 5017185029Spjd if (pio == NULL) { 5018185029Spjd /* 5019185029Spjd * Insert a dummy header on the buflist so 5020185029Spjd * l2arc_write_done() can find where the 5021185029Spjd * write buffers begin without searching. 5022185029Spjd */ 5023185029Spjd list_insert_head(dev->l2ad_buflist, head); 5024185029Spjd 5025185029Spjd cb = kmem_alloc( 5026185029Spjd sizeof (l2arc_write_callback_t), KM_SLEEP); 5027185029Spjd cb->l2wcb_dev = dev; 5028185029Spjd cb->l2wcb_head = head; 5029185029Spjd pio = zio_root(spa, l2arc_write_done, cb, 5030185029Spjd ZIO_FLAG_CANFAIL); 5031205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_pios); 5032185029Spjd } 5033185029Spjd 5034185029Spjd /* 5035185029Spjd * Create and add a new L2ARC header. 5036185029Spjd */ 5037252140Sdelphij l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5038252140Sdelphij l2hdr->b_dev = dev; 5039252140Sdelphij ab->b_flags |= ARC_L2_WRITING; 5040185029Spjd 5041252140Sdelphij /* 5042252140Sdelphij * Temporarily stash the data buffer in b_tmp_cdata. 5043252140Sdelphij * The subsequent write step will pick it up from 5044252140Sdelphij * there. This is because can't access ab->b_buf 5045252140Sdelphij * without holding the hash_lock, which we in turn 5046252140Sdelphij * can't access without holding the ARC list locks 5047252140Sdelphij * (which we want to avoid during compression/writing). 5048252140Sdelphij */ 5049252140Sdelphij l2hdr->b_compress = ZIO_COMPRESS_OFF; 5050252140Sdelphij l2hdr->b_asize = ab->b_size; 5051252140Sdelphij l2hdr->b_tmp_cdata = ab->b_buf->b_data; 5052252140Sdelphij 5053185029Spjd buf_sz = ab->b_size; 5054252140Sdelphij ab->b_l2hdr = l2hdr; 5055185029Spjd 5056252140Sdelphij list_insert_head(dev->l2ad_buflist, ab); 5057252140Sdelphij 5058185029Spjd /* 5059185029Spjd * Compute and store the buffer cksum before 5060185029Spjd * writing. On debug the cksum is verified first. 5061185029Spjd */ 5062185029Spjd arc_cksum_verify(ab->b_buf); 5063185029Spjd arc_cksum_compute(ab->b_buf, B_TRUE); 5064185029Spjd 5065185029Spjd mutex_exit(hash_lock); 5066185029Spjd 5067252140Sdelphij write_sz += buf_sz; 5068252140Sdelphij } 5069252140Sdelphij 5070252140Sdelphij mutex_exit(list_lock); 5071252140Sdelphij 5072252140Sdelphij if (full == B_TRUE) 5073252140Sdelphij break; 5074252140Sdelphij } 5075252140Sdelphij 5076252140Sdelphij /* No buffers selected for writing? */ 5077252140Sdelphij if (pio == NULL) { 5078252140Sdelphij ASSERT0(write_sz); 5079252140Sdelphij mutex_exit(&l2arc_buflist_mtx); 5080252140Sdelphij kmem_cache_free(hdr_cache, head); 5081252140Sdelphij return (0); 5082252140Sdelphij } 5083252140Sdelphij 5084252140Sdelphij /* 5085252140Sdelphij * Now start writing the buffers. We're starting at the write head 5086252140Sdelphij * and work backwards, retracing the course of the buffer selector 5087252140Sdelphij * loop above. 5088252140Sdelphij */ 5089252140Sdelphij for (ab = list_prev(dev->l2ad_buflist, head); ab; 5090252140Sdelphij ab = list_prev(dev->l2ad_buflist, ab)) { 5091252140Sdelphij l2arc_buf_hdr_t *l2hdr; 5092252140Sdelphij uint64_t buf_sz; 5093252140Sdelphij 5094252140Sdelphij /* 5095252140Sdelphij * We shouldn't need to lock the buffer here, since we flagged 5096252140Sdelphij * it as ARC_L2_WRITING in the previous step, but we must take 5097252140Sdelphij * care to only access its L2 cache parameters. In particular, 5098252140Sdelphij * ab->b_buf may be invalid by now due to ARC eviction. 5099252140Sdelphij */ 5100252140Sdelphij l2hdr = ab->b_l2hdr; 5101252140Sdelphij l2hdr->b_daddr = dev->l2ad_hand; 5102252140Sdelphij 5103252140Sdelphij if ((ab->b_flags & ARC_L2COMPRESS) && 5104252140Sdelphij l2hdr->b_asize >= buf_compress_minsz) { 5105252140Sdelphij if (l2arc_compress_buf(l2hdr)) { 5106252140Sdelphij /* 5107252140Sdelphij * If compression succeeded, enable headroom 5108252140Sdelphij * boost on the next scan cycle. 5109252140Sdelphij */ 5110252140Sdelphij *headroom_boost = B_TRUE; 5111252140Sdelphij } 5112252140Sdelphij } 5113252140Sdelphij 5114252140Sdelphij /* 5115252140Sdelphij * Pick up the buffer data we had previously stashed away 5116252140Sdelphij * (and now potentially also compressed). 5117252140Sdelphij */ 5118252140Sdelphij buf_data = l2hdr->b_tmp_cdata; 5119252140Sdelphij buf_sz = l2hdr->b_asize; 5120252140Sdelphij 5121252140Sdelphij /* Compression may have squashed the buffer to zero length. */ 5122252140Sdelphij if (buf_sz != 0) { 5123252140Sdelphij uint64_t buf_p_sz; 5124252140Sdelphij 5125185029Spjd wzio = zio_write_phys(pio, dev->l2ad_vdev, 5126185029Spjd dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5127185029Spjd NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5128185029Spjd ZIO_FLAG_CANFAIL, B_FALSE); 5129185029Spjd 5130185029Spjd DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5131185029Spjd zio_t *, wzio); 5132185029Spjd (void) zio_nowait(wzio); 5133185029Spjd 5134252140Sdelphij write_asize += buf_sz; 5135185029Spjd /* 5136185029Spjd * Keep the clock hand suitably device-aligned. 5137185029Spjd */ 5138252140Sdelphij buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5139252140Sdelphij write_psize += buf_p_sz; 5140252140Sdelphij dev->l2ad_hand += buf_p_sz; 5141185029Spjd } 5142252140Sdelphij } 5143185029Spjd 5144185029Spjd mutex_exit(&l2arc_buflist_mtx); 5145185029Spjd 5146252140Sdelphij ASSERT3U(write_asize, <=, target_sz); 5147185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_sent); 5148252140Sdelphij ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5149185029Spjd ARCSTAT_INCR(arcstat_l2_size, write_sz); 5150252140Sdelphij ARCSTAT_INCR(arcstat_l2_asize, write_asize); 5151252140Sdelphij vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); 5152185029Spjd 5153185029Spjd /* 5154185029Spjd * Bump device hand to the device start if it is approaching the end. 5155185029Spjd * l2arc_evict() will already have evicted ahead for this case. 5156185029Spjd */ 5157185029Spjd if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5158219089Spjd vdev_space_update(dev->l2ad_vdev, 5159219089Spjd dev->l2ad_end - dev->l2ad_hand, 0, 0); 5160185029Spjd dev->l2ad_hand = dev->l2ad_start; 5161185029Spjd dev->l2ad_evict = dev->l2ad_start; 5162185029Spjd dev->l2ad_first = B_FALSE; 5163185029Spjd } 5164185029Spjd 5165208373Smm dev->l2ad_writing = B_TRUE; 5166185029Spjd (void) zio_wait(pio); 5167208373Smm dev->l2ad_writing = B_FALSE; 5168208373Smm 5169252140Sdelphij return (write_asize); 5170185029Spjd} 5171185029Spjd 5172185029Spjd/* 5173252140Sdelphij * Compresses an L2ARC buffer. 5174252140Sdelphij * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5175252140Sdelphij * size in l2hdr->b_asize. This routine tries to compress the data and 5176252140Sdelphij * depending on the compression result there are three possible outcomes: 5177252140Sdelphij * *) The buffer was incompressible. The original l2hdr contents were left 5178252140Sdelphij * untouched and are ready for writing to an L2 device. 5179252140Sdelphij * *) The buffer was all-zeros, so there is no need to write it to an L2 5180252140Sdelphij * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5181252140Sdelphij * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5182252140Sdelphij * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5183252140Sdelphij * data buffer which holds the compressed data to be written, and b_asize 5184252140Sdelphij * tells us how much data there is. b_compress is set to the appropriate 5185252140Sdelphij * compression algorithm. Once writing is done, invoke 5186252140Sdelphij * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5187252140Sdelphij * 5188252140Sdelphij * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5189252140Sdelphij * buffer was incompressible). 5190252140Sdelphij */ 5191252140Sdelphijstatic boolean_t 5192252140Sdelphijl2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5193252140Sdelphij{ 5194252140Sdelphij void *cdata; 5195252140Sdelphij size_t csize, len; 5196252140Sdelphij 5197252140Sdelphij ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5198252140Sdelphij ASSERT(l2hdr->b_tmp_cdata != NULL); 5199252140Sdelphij 5200252140Sdelphij len = l2hdr->b_asize; 5201252140Sdelphij cdata = zio_data_buf_alloc(len); 5202252140Sdelphij csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5203262173Savg cdata, l2hdr->b_asize, (size_t)(1ULL << l2hdr->b_dev->l2ad_vdev->vdev_ashift)); 5204252140Sdelphij 5205252140Sdelphij if (csize == 0) { 5206252140Sdelphij /* zero block, indicate that there's nothing to write */ 5207252140Sdelphij zio_data_buf_free(cdata, len); 5208252140Sdelphij l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5209252140Sdelphij l2hdr->b_asize = 0; 5210252140Sdelphij l2hdr->b_tmp_cdata = NULL; 5211252140Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5212252140Sdelphij return (B_TRUE); 5213252140Sdelphij } else if (csize > 0 && csize < len) { 5214252140Sdelphij /* 5215252140Sdelphij * Compression succeeded, we'll keep the cdata around for 5216252140Sdelphij * writing and release it afterwards. 5217252140Sdelphij */ 5218252140Sdelphij l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5219252140Sdelphij l2hdr->b_asize = csize; 5220252140Sdelphij l2hdr->b_tmp_cdata = cdata; 5221252140Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_successes); 5222252140Sdelphij return (B_TRUE); 5223252140Sdelphij } else { 5224252140Sdelphij /* 5225252140Sdelphij * Compression failed, release the compressed buffer. 5226252140Sdelphij * l2hdr will be left unmodified. 5227252140Sdelphij */ 5228252140Sdelphij zio_data_buf_free(cdata, len); 5229252140Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_failures); 5230252140Sdelphij return (B_FALSE); 5231252140Sdelphij } 5232252140Sdelphij} 5233252140Sdelphij 5234252140Sdelphij/* 5235252140Sdelphij * Decompresses a zio read back from an l2arc device. On success, the 5236252140Sdelphij * underlying zio's io_data buffer is overwritten by the uncompressed 5237252140Sdelphij * version. On decompression error (corrupt compressed stream), the 5238252140Sdelphij * zio->io_error value is set to signal an I/O error. 5239252140Sdelphij * 5240252140Sdelphij * Please note that the compressed data stream is not checksummed, so 5241252140Sdelphij * if the underlying device is experiencing data corruption, we may feed 5242252140Sdelphij * corrupt data to the decompressor, so the decompressor needs to be 5243252140Sdelphij * able to handle this situation (LZ4 does). 5244252140Sdelphij */ 5245252140Sdelphijstatic void 5246252140Sdelphijl2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5247252140Sdelphij{ 5248252140Sdelphij ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5249252140Sdelphij 5250252140Sdelphij if (zio->io_error != 0) { 5251252140Sdelphij /* 5252252140Sdelphij * An io error has occured, just restore the original io 5253252140Sdelphij * size in preparation for a main pool read. 5254252140Sdelphij */ 5255252140Sdelphij zio->io_orig_size = zio->io_size = hdr->b_size; 5256252140Sdelphij return; 5257252140Sdelphij } 5258252140Sdelphij 5259252140Sdelphij if (c == ZIO_COMPRESS_EMPTY) { 5260252140Sdelphij /* 5261252140Sdelphij * An empty buffer results in a null zio, which means we 5262252140Sdelphij * need to fill its io_data after we're done restoring the 5263252140Sdelphij * buffer's contents. 5264252140Sdelphij */ 5265252140Sdelphij ASSERT(hdr->b_buf != NULL); 5266252140Sdelphij bzero(hdr->b_buf->b_data, hdr->b_size); 5267252140Sdelphij zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5268252140Sdelphij } else { 5269252140Sdelphij ASSERT(zio->io_data != NULL); 5270252140Sdelphij /* 5271252140Sdelphij * We copy the compressed data from the start of the arc buffer 5272252140Sdelphij * (the zio_read will have pulled in only what we need, the 5273252140Sdelphij * rest is garbage which we will overwrite at decompression) 5274252140Sdelphij * and then decompress back to the ARC data buffer. This way we 5275252140Sdelphij * can minimize copying by simply decompressing back over the 5276252140Sdelphij * original compressed data (rather than decompressing to an 5277252140Sdelphij * aux buffer and then copying back the uncompressed buffer, 5278252140Sdelphij * which is likely to be much larger). 5279252140Sdelphij */ 5280252140Sdelphij uint64_t csize; 5281252140Sdelphij void *cdata; 5282252140Sdelphij 5283252140Sdelphij csize = zio->io_size; 5284252140Sdelphij cdata = zio_data_buf_alloc(csize); 5285252140Sdelphij bcopy(zio->io_data, cdata, csize); 5286252140Sdelphij if (zio_decompress_data(c, cdata, zio->io_data, csize, 5287252140Sdelphij hdr->b_size) != 0) 5288252140Sdelphij zio->io_error = EIO; 5289252140Sdelphij zio_data_buf_free(cdata, csize); 5290252140Sdelphij } 5291252140Sdelphij 5292252140Sdelphij /* Restore the expected uncompressed IO size. */ 5293252140Sdelphij zio->io_orig_size = zio->io_size = hdr->b_size; 5294252140Sdelphij} 5295252140Sdelphij 5296252140Sdelphij/* 5297252140Sdelphij * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5298252140Sdelphij * This buffer serves as a temporary holder of compressed data while 5299252140Sdelphij * the buffer entry is being written to an l2arc device. Once that is 5300252140Sdelphij * done, we can dispose of it. 5301252140Sdelphij */ 5302252140Sdelphijstatic void 5303252140Sdelphijl2arc_release_cdata_buf(arc_buf_hdr_t *ab) 5304252140Sdelphij{ 5305252140Sdelphij l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr; 5306252140Sdelphij 5307252140Sdelphij if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) { 5308252140Sdelphij /* 5309252140Sdelphij * If the data was compressed, then we've allocated a 5310252140Sdelphij * temporary buffer for it, so now we need to release it. 5311252140Sdelphij */ 5312252140Sdelphij ASSERT(l2hdr->b_tmp_cdata != NULL); 5313252140Sdelphij zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size); 5314252140Sdelphij } 5315252140Sdelphij l2hdr->b_tmp_cdata = NULL; 5316252140Sdelphij} 5317252140Sdelphij 5318252140Sdelphij/* 5319185029Spjd * This thread feeds the L2ARC at regular intervals. This is the beating 5320185029Spjd * heart of the L2ARC. 5321185029Spjd */ 5322185029Spjdstatic void 5323185029Spjdl2arc_feed_thread(void *dummy __unused) 5324185029Spjd{ 5325185029Spjd callb_cpr_t cpr; 5326185029Spjd l2arc_dev_t *dev; 5327185029Spjd spa_t *spa; 5328208373Smm uint64_t size, wrote; 5329219089Spjd clock_t begin, next = ddi_get_lbolt(); 5330252140Sdelphij boolean_t headroom_boost = B_FALSE; 5331185029Spjd 5332185029Spjd CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5333185029Spjd 5334185029Spjd mutex_enter(&l2arc_feed_thr_lock); 5335185029Spjd 5336185029Spjd while (l2arc_thread_exit == 0) { 5337185029Spjd CALLB_CPR_SAFE_BEGIN(&cpr); 5338185029Spjd (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5339219089Spjd next - ddi_get_lbolt()); 5340185029Spjd CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5341219089Spjd next = ddi_get_lbolt() + hz; 5342185029Spjd 5343185029Spjd /* 5344185029Spjd * Quick check for L2ARC devices. 5345185029Spjd */ 5346185029Spjd mutex_enter(&l2arc_dev_mtx); 5347185029Spjd if (l2arc_ndev == 0) { 5348185029Spjd mutex_exit(&l2arc_dev_mtx); 5349185029Spjd continue; 5350185029Spjd } 5351185029Spjd mutex_exit(&l2arc_dev_mtx); 5352219089Spjd begin = ddi_get_lbolt(); 5353185029Spjd 5354185029Spjd /* 5355185029Spjd * This selects the next l2arc device to write to, and in 5356185029Spjd * doing so the next spa to feed from: dev->l2ad_spa. This 5357185029Spjd * will return NULL if there are now no l2arc devices or if 5358185029Spjd * they are all faulted. 5359185029Spjd * 5360185029Spjd * If a device is returned, its spa's config lock is also 5361185029Spjd * held to prevent device removal. l2arc_dev_get_next() 5362185029Spjd * will grab and release l2arc_dev_mtx. 5363185029Spjd */ 5364185029Spjd if ((dev = l2arc_dev_get_next()) == NULL) 5365185029Spjd continue; 5366185029Spjd 5367185029Spjd spa = dev->l2ad_spa; 5368185029Spjd ASSERT(spa != NULL); 5369185029Spjd 5370185029Spjd /* 5371219089Spjd * If the pool is read-only then force the feed thread to 5372219089Spjd * sleep a little longer. 5373219089Spjd */ 5374219089Spjd if (!spa_writeable(spa)) { 5375219089Spjd next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5376219089Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5377219089Spjd continue; 5378219089Spjd } 5379219089Spjd 5380219089Spjd /* 5381185029Spjd * Avoid contributing to memory pressure. 5382185029Spjd */ 5383185029Spjd if (arc_reclaim_needed()) { 5384185029Spjd ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5385185029Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5386185029Spjd continue; 5387185029Spjd } 5388185029Spjd 5389185029Spjd ARCSTAT_BUMP(arcstat_l2_feeds); 5390185029Spjd 5391252140Sdelphij size = l2arc_write_size(); 5392185029Spjd 5393185029Spjd /* 5394185029Spjd * Evict L2ARC buffers that will be overwritten. 5395185029Spjd */ 5396185029Spjd l2arc_evict(dev, size, B_FALSE); 5397185029Spjd 5398185029Spjd /* 5399185029Spjd * Write ARC buffers. 5400185029Spjd */ 5401252140Sdelphij wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5402208373Smm 5403208373Smm /* 5404208373Smm * Calculate interval between writes. 5405208373Smm */ 5406208373Smm next = l2arc_write_interval(begin, size, wrote); 5407185029Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5408185029Spjd } 5409185029Spjd 5410185029Spjd l2arc_thread_exit = 0; 5411185029Spjd cv_broadcast(&l2arc_feed_thr_cv); 5412185029Spjd CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5413185029Spjd thread_exit(); 5414185029Spjd} 5415185029Spjd 5416185029Spjdboolean_t 5417185029Spjdl2arc_vdev_present(vdev_t *vd) 5418185029Spjd{ 5419185029Spjd l2arc_dev_t *dev; 5420185029Spjd 5421185029Spjd mutex_enter(&l2arc_dev_mtx); 5422185029Spjd for (dev = list_head(l2arc_dev_list); dev != NULL; 5423185029Spjd dev = list_next(l2arc_dev_list, dev)) { 5424185029Spjd if (dev->l2ad_vdev == vd) 5425185029Spjd break; 5426185029Spjd } 5427185029Spjd mutex_exit(&l2arc_dev_mtx); 5428185029Spjd 5429185029Spjd return (dev != NULL); 5430185029Spjd} 5431185029Spjd 5432185029Spjd/* 5433185029Spjd * Add a vdev for use by the L2ARC. By this point the spa has already 5434185029Spjd * validated the vdev and opened it. 5435185029Spjd */ 5436185029Spjdvoid 5437219089Spjdl2arc_add_vdev(spa_t *spa, vdev_t *vd) 5438185029Spjd{ 5439185029Spjd l2arc_dev_t *adddev; 5440185029Spjd 5441185029Spjd ASSERT(!l2arc_vdev_present(vd)); 5442185029Spjd 5443262081Savg vdev_ashift_optimize(vd); 5444262081Savg 5445185029Spjd /* 5446185029Spjd * Create a new l2arc device entry. 5447185029Spjd */ 5448185029Spjd adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5449185029Spjd adddev->l2ad_spa = spa; 5450185029Spjd adddev->l2ad_vdev = vd; 5451219089Spjd adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5452219089Spjd adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5453185029Spjd adddev->l2ad_hand = adddev->l2ad_start; 5454185029Spjd adddev->l2ad_evict = adddev->l2ad_start; 5455185029Spjd adddev->l2ad_first = B_TRUE; 5456208373Smm adddev->l2ad_writing = B_FALSE; 5457185029Spjd 5458185029Spjd /* 5459185029Spjd * This is a list of all ARC buffers that are still valid on the 5460185029Spjd * device. 5461185029Spjd */ 5462185029Spjd adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5463185029Spjd list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5464185029Spjd offsetof(arc_buf_hdr_t, b_l2node)); 5465185029Spjd 5466219089Spjd vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5467185029Spjd 5468185029Spjd /* 5469185029Spjd * Add device to global list 5470185029Spjd */ 5471185029Spjd mutex_enter(&l2arc_dev_mtx); 5472185029Spjd list_insert_head(l2arc_dev_list, adddev); 5473185029Spjd atomic_inc_64(&l2arc_ndev); 5474185029Spjd mutex_exit(&l2arc_dev_mtx); 5475185029Spjd} 5476185029Spjd 5477185029Spjd/* 5478185029Spjd * Remove a vdev from the L2ARC. 5479185029Spjd */ 5480185029Spjdvoid 5481185029Spjdl2arc_remove_vdev(vdev_t *vd) 5482185029Spjd{ 5483185029Spjd l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5484185029Spjd 5485185029Spjd /* 5486185029Spjd * Find the device by vdev 5487185029Spjd */ 5488185029Spjd mutex_enter(&l2arc_dev_mtx); 5489185029Spjd for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5490185029Spjd nextdev = list_next(l2arc_dev_list, dev); 5491185029Spjd if (vd == dev->l2ad_vdev) { 5492185029Spjd remdev = dev; 5493185029Spjd break; 5494185029Spjd } 5495185029Spjd } 5496185029Spjd ASSERT(remdev != NULL); 5497185029Spjd 5498185029Spjd /* 5499185029Spjd * Remove device from global list 5500185029Spjd */ 5501185029Spjd list_remove(l2arc_dev_list, remdev); 5502185029Spjd l2arc_dev_last = NULL; /* may have been invalidated */ 5503185029Spjd atomic_dec_64(&l2arc_ndev); 5504185029Spjd mutex_exit(&l2arc_dev_mtx); 5505185029Spjd 5506185029Spjd /* 5507185029Spjd * Clear all buflists and ARC references. L2ARC device flush. 5508185029Spjd */ 5509185029Spjd l2arc_evict(remdev, 0, B_TRUE); 5510185029Spjd list_destroy(remdev->l2ad_buflist); 5511185029Spjd kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5512185029Spjd kmem_free(remdev, sizeof (l2arc_dev_t)); 5513185029Spjd} 5514185029Spjd 5515185029Spjdvoid 5516185029Spjdl2arc_init(void) 5517185029Spjd{ 5518185029Spjd l2arc_thread_exit = 0; 5519185029Spjd l2arc_ndev = 0; 5520185029Spjd l2arc_writes_sent = 0; 5521185029Spjd l2arc_writes_done = 0; 5522185029Spjd 5523185029Spjd mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5524185029Spjd cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5525185029Spjd mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5526185029Spjd mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5527185029Spjd mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5528185029Spjd 5529185029Spjd l2arc_dev_list = &L2ARC_dev_list; 5530185029Spjd l2arc_free_on_write = &L2ARC_free_on_write; 5531185029Spjd list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5532185029Spjd offsetof(l2arc_dev_t, l2ad_node)); 5533185029Spjd list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5534185029Spjd offsetof(l2arc_data_free_t, l2df_list_node)); 5535185029Spjd} 5536185029Spjd 5537185029Spjdvoid 5538185029Spjdl2arc_fini(void) 5539185029Spjd{ 5540185029Spjd /* 5541185029Spjd * This is called from dmu_fini(), which is called from spa_fini(); 5542185029Spjd * Because of this, we can assume that all l2arc devices have 5543185029Spjd * already been removed when the pools themselves were removed. 5544185029Spjd */ 5545185029Spjd 5546185029Spjd l2arc_do_free_on_write(); 5547185029Spjd 5548185029Spjd mutex_destroy(&l2arc_feed_thr_lock); 5549185029Spjd cv_destroy(&l2arc_feed_thr_cv); 5550185029Spjd mutex_destroy(&l2arc_dev_mtx); 5551185029Spjd mutex_destroy(&l2arc_buflist_mtx); 5552185029Spjd mutex_destroy(&l2arc_free_on_write_mtx); 5553185029Spjd 5554185029Spjd list_destroy(l2arc_dev_list); 5555185029Spjd list_destroy(l2arc_free_on_write); 5556185029Spjd} 5557185029Spjd 5558185029Spjdvoid 5559185029Spjdl2arc_start(void) 5560185029Spjd{ 5561209962Smm if (!(spa_mode_global & FWRITE)) 5562185029Spjd return; 5563185029Spjd 5564185029Spjd (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5565185029Spjd TS_RUN, minclsyspri); 5566185029Spjd} 5567185029Spjd 5568185029Spjdvoid 5569185029Spjdl2arc_stop(void) 5570185029Spjd{ 5571209962Smm if (!(spa_mode_global & FWRITE)) 5572185029Spjd return; 5573185029Spjd 5574185029Spjd mutex_enter(&l2arc_feed_thr_lock); 5575185029Spjd cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5576185029Spjd l2arc_thread_exit = 1; 5577185029Spjd while (l2arc_thread_exit != 0) 5578185029Spjd cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5579185029Spjd mutex_exit(&l2arc_feed_thr_lock); 5580185029Spjd} 5581