arc.c revision 206792
1168404Spjd/* 2168404Spjd * CDDL HEADER START 3168404Spjd * 4168404Spjd * The contents of this file are subject to the terms of the 5168404Spjd * Common Development and Distribution License (the "License"). 6168404Spjd * You may not use this file except in compliance with the License. 7168404Spjd * 8168404Spjd * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9168404Spjd * or http://www.opensolaris.org/os/licensing. 10168404Spjd * See the License for the specific language governing permissions 11168404Spjd * and limitations under the License. 12168404Spjd * 13168404Spjd * When distributing Covered Code, include this CDDL HEADER in each 14168404Spjd * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15168404Spjd * If applicable, add the following below this CDDL HEADER, with the 16168404Spjd * fields enclosed by brackets "[]" replaced with your own identifying 17168404Spjd * information: Portions Copyright [yyyy] [name of copyright owner] 18168404Spjd * 19168404Spjd * CDDL HEADER END 20168404Spjd */ 21168404Spjd/* 22185029Spjd * Copyright 2008 Sun Microsystems, Inc. All rights reserved. 23168404Spjd * Use is subject to license terms. 24168404Spjd */ 25168404Spjd 26168404Spjd/* 27168404Spjd * DVA-based Adjustable Replacement Cache 28168404Spjd * 29168404Spjd * While much of the theory of operation used here is 30168404Spjd * based on the self-tuning, low overhead replacement cache 31168404Spjd * presented by Megiddo and Modha at FAST 2003, there are some 32168404Spjd * significant differences: 33168404Spjd * 34168404Spjd * 1. The Megiddo and Modha model assumes any page is evictable. 35168404Spjd * Pages in its cache cannot be "locked" into memory. This makes 36168404Spjd * the eviction algorithm simple: evict the last page in the list. 37168404Spjd * This also make the performance characteristics easy to reason 38168404Spjd * about. Our cache is not so simple. At any given moment, some 39168404Spjd * subset of the blocks in the cache are un-evictable because we 40168404Spjd * have handed out a reference to them. Blocks are only evictable 41168404Spjd * when there are no external references active. This makes 42168404Spjd * eviction far more problematic: we choose to evict the evictable 43168404Spjd * blocks that are the "lowest" in the list. 44168404Spjd * 45168404Spjd * There are times when it is not possible to evict the requested 46168404Spjd * space. In these circumstances we are unable to adjust the cache 47168404Spjd * size. To prevent the cache growing unbounded at these times we 48185029Spjd * implement a "cache throttle" that slows the flow of new data 49185029Spjd * into the cache until we can make space available. 50168404Spjd * 51168404Spjd * 2. The Megiddo and Modha model assumes a fixed cache size. 52168404Spjd * Pages are evicted when the cache is full and there is a cache 53168404Spjd * miss. Our model has a variable sized cache. It grows with 54185029Spjd * high use, but also tries to react to memory pressure from the 55168404Spjd * operating system: decreasing its size when system memory is 56168404Spjd * tight. 57168404Spjd * 58168404Spjd * 3. The Megiddo and Modha model assumes a fixed page size. All 59168404Spjd * elements of the cache are therefor exactly the same size. So 60168404Spjd * when adjusting the cache size following a cache miss, its simply 61168404Spjd * a matter of choosing a single page to evict. In our model, we 62168404Spjd * have variable sized cache blocks (rangeing from 512 bytes to 63168404Spjd * 128K bytes). We therefor choose a set of blocks to evict to make 64168404Spjd * space for a cache miss that approximates as closely as possible 65168404Spjd * the space used by the new block. 66168404Spjd * 67168404Spjd * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 68168404Spjd * by N. Megiddo & D. Modha, FAST 2003 69168404Spjd */ 70168404Spjd 71168404Spjd/* 72168404Spjd * The locking model: 73168404Spjd * 74168404Spjd * A new reference to a cache buffer can be obtained in two 75168404Spjd * ways: 1) via a hash table lookup using the DVA as a key, 76185029Spjd * or 2) via one of the ARC lists. The arc_read() interface 77168404Spjd * uses method 1, while the internal arc algorithms for 78168404Spjd * adjusting the cache use method 2. We therefor provide two 79168404Spjd * types of locks: 1) the hash table lock array, and 2) the 80168404Spjd * arc list locks. 81168404Spjd * 82168404Spjd * Buffers do not have their own mutexs, rather they rely on the 83168404Spjd * hash table mutexs for the bulk of their protection (i.e. most 84168404Spjd * fields in the arc_buf_hdr_t are protected by these mutexs). 85168404Spjd * 86168404Spjd * buf_hash_find() returns the appropriate mutex (held) when it 87168404Spjd * locates the requested buffer in the hash table. It returns 88168404Spjd * NULL for the mutex if the buffer was not in the table. 89168404Spjd * 90168404Spjd * buf_hash_remove() expects the appropriate hash mutex to be 91168404Spjd * already held before it is invoked. 92168404Spjd * 93168404Spjd * Each arc state also has a mutex which is used to protect the 94168404Spjd * buffer list associated with the state. When attempting to 95168404Spjd * obtain a hash table lock while holding an arc list lock you 96168404Spjd * must use: mutex_tryenter() to avoid deadlock. Also note that 97168404Spjd * the active state mutex must be held before the ghost state mutex. 98168404Spjd * 99168404Spjd * Arc buffers may have an associated eviction callback function. 100168404Spjd * This function will be invoked prior to removing the buffer (e.g. 101168404Spjd * in arc_do_user_evicts()). Note however that the data associated 102168404Spjd * with the buffer may be evicted prior to the callback. The callback 103168404Spjd * must be made with *no locks held* (to prevent deadlock). Additionally, 104168404Spjd * the users of callbacks must ensure that their private data is 105168404Spjd * protected from simultaneous callbacks from arc_buf_evict() 106168404Spjd * and arc_do_user_evicts(). 107168404Spjd * 108168404Spjd * Note that the majority of the performance stats are manipulated 109168404Spjd * with atomic operations. 110185029Spjd * 111185029Spjd * The L2ARC uses the l2arc_buflist_mtx global mutex for the following: 112185029Spjd * 113185029Spjd * - L2ARC buflist creation 114185029Spjd * - L2ARC buflist eviction 115185029Spjd * - L2ARC write completion, which walks L2ARC buflists 116185029Spjd * - ARC header destruction, as it removes from L2ARC buflists 117185029Spjd * - ARC header release, as it removes from L2ARC buflists 118168404Spjd */ 119168404Spjd 120168404Spjd#include <sys/spa.h> 121168404Spjd#include <sys/zio.h> 122168404Spjd#include <sys/zio_checksum.h> 123168404Spjd#include <sys/zfs_context.h> 124168404Spjd#include <sys/arc.h> 125168404Spjd#include <sys/refcount.h> 126185029Spjd#include <sys/vdev.h> 127168404Spjd#ifdef _KERNEL 128168404Spjd#include <sys/dnlc.h> 129168404Spjd#endif 130168404Spjd#include <sys/callb.h> 131168404Spjd#include <sys/kstat.h> 132168404Spjd#include <sys/sdt.h> 133168404Spjd 134205231Skmacy#include <sys/ktr.h> 135191902Skmacy#include <vm/vm_pageout.h> 136191902Skmacy 137168404Spjdstatic kmutex_t arc_reclaim_thr_lock; 138168404Spjdstatic kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 139168404Spjdstatic uint8_t arc_thread_exit; 140168404Spjd 141185029Spjdextern int zfs_write_limit_shift; 142185029Spjdextern uint64_t zfs_write_limit_max; 143185029Spjdextern kmutex_t zfs_write_limit_lock; 144185029Spjd 145168404Spjd#define ARC_REDUCE_DNLC_PERCENT 3 146168404Spjduint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 147168404Spjd 148168404Spjdtypedef enum arc_reclaim_strategy { 149168404Spjd ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 150168404Spjd ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 151168404Spjd} arc_reclaim_strategy_t; 152168404Spjd 153168404Spjd/* number of seconds before growing cache again */ 154168404Spjdstatic int arc_grow_retry = 60; 155168404Spjd 156168404Spjd/* 157168404Spjd * minimum lifespan of a prefetch block in clock ticks 158168404Spjd * (initialized in arc_init()) 159168404Spjd */ 160168404Spjdstatic int arc_min_prefetch_lifespan; 161168404Spjd 162194043Skmacyextern int zfs_prefetch_disable; 163168404Spjdstatic int arc_dead; 164168404Spjd 165168404Spjd/* 166185029Spjd * The arc has filled available memory and has now warmed up. 167185029Spjd */ 168185029Spjdstatic boolean_t arc_warm; 169185029Spjd 170185029Spjd/* 171168404Spjd * These tunables are for performance analysis. 172168404Spjd */ 173185029Spjduint64_t zfs_arc_max; 174185029Spjduint64_t zfs_arc_min; 175185029Spjduint64_t zfs_arc_meta_limit = 0; 176185029Spjdint zfs_mdcomp_disable = 0; 177185029Spjd 178185029SpjdTUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max); 179185029SpjdTUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 180185029SpjdTUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 181185029SpjdTUNABLE_INT("vfs.zfs.mdcomp_disable", &zfs_mdcomp_disable); 182168473SpjdSYSCTL_DECL(_vfs_zfs); 183185029SpjdSYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 184168473Spjd "Maximum ARC size"); 185185029SpjdSYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 186168473Spjd "Minimum ARC size"); 187185029SpjdSYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RDTUN, 188185029Spjd &zfs_mdcomp_disable, 0, "Disable metadata compression"); 189168404Spjd 190205231Skmacy#ifdef ZIO_USE_UMA 191205231Skmacyextern kmem_cache_t *zio_buf_cache[]; 192205231Skmacyextern kmem_cache_t *zio_data_buf_cache[]; 193205231Skmacy#endif 194205231Skmacy 195168404Spjd/* 196185029Spjd * Note that buffers can be in one of 6 states: 197168404Spjd * ARC_anon - anonymous (discussed below) 198168404Spjd * ARC_mru - recently used, currently cached 199168404Spjd * ARC_mru_ghost - recentely used, no longer in cache 200168404Spjd * ARC_mfu - frequently used, currently cached 201168404Spjd * ARC_mfu_ghost - frequently used, no longer in cache 202185029Spjd * ARC_l2c_only - exists in L2ARC but not other states 203185029Spjd * When there are no active references to the buffer, they are 204185029Spjd * are linked onto a list in one of these arc states. These are 205185029Spjd * the only buffers that can be evicted or deleted. Within each 206185029Spjd * state there are multiple lists, one for meta-data and one for 207185029Spjd * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 208185029Spjd * etc.) is tracked separately so that it can be managed more 209185029Spjd * explicitly: favored over data, limited explicitly. 210168404Spjd * 211168404Spjd * Anonymous buffers are buffers that are not associated with 212168404Spjd * a DVA. These are buffers that hold dirty block copies 213168404Spjd * before they are written to stable storage. By definition, 214168404Spjd * they are "ref'd" and are considered part of arc_mru 215168404Spjd * that cannot be freed. Generally, they will aquire a DVA 216168404Spjd * as they are written and migrate onto the arc_mru list. 217185029Spjd * 218185029Spjd * The ARC_l2c_only state is for buffers that are in the second 219185029Spjd * level ARC but no longer in any of the ARC_m* lists. The second 220185029Spjd * level ARC itself may also contain buffers that are in any of 221185029Spjd * the ARC_m* states - meaning that a buffer can exist in two 222185029Spjd * places. The reason for the ARC_l2c_only state is to keep the 223185029Spjd * buffer header in the hash table, so that reads that hit the 224185029Spjd * second level ARC benefit from these fast lookups. 225168404Spjd */ 226168404Spjd 227205264Skmacy#define ARCS_LOCK_PAD CACHE_LINE_SIZE 228205231Skmacystruct arcs_lock { 229205231Skmacy kmutex_t arcs_lock; 230205231Skmacy#ifdef _KERNEL 231205231Skmacy unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 232205231Skmacy#endif 233205231Skmacy}; 234205231Skmacy 235205231Skmacy/* 236205231Skmacy * must be power of two for mask use to work 237205231Skmacy * 238205231Skmacy */ 239205231Skmacy#define ARC_BUFC_NUMDATALISTS 16 240205231Skmacy#define ARC_BUFC_NUMMETADATALISTS 16 241205231Skmacy#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS+ARC_BUFC_NUMDATALISTS) 242205231Skmacy 243168404Spjdtypedef struct arc_state { 244185029Spjd uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 245185029Spjd uint64_t arcs_size; /* total amount of data in this state */ 246205231Skmacy list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 247205264Skmacy struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 248168404Spjd} arc_state_t; 249168404Spjd 250205231Skmacy#define ARCS_LOCK(s, i) &((s)->arcs_locks[(i)].arcs_lock) 251205231Skmacy 252185029Spjd/* The 6 states: */ 253168404Spjdstatic arc_state_t ARC_anon; 254168404Spjdstatic arc_state_t ARC_mru; 255168404Spjdstatic arc_state_t ARC_mru_ghost; 256168404Spjdstatic arc_state_t ARC_mfu; 257168404Spjdstatic arc_state_t ARC_mfu_ghost; 258185029Spjdstatic arc_state_t ARC_l2c_only; 259168404Spjd 260168404Spjdtypedef struct arc_stats { 261168404Spjd kstat_named_t arcstat_hits; 262168404Spjd kstat_named_t arcstat_misses; 263168404Spjd kstat_named_t arcstat_demand_data_hits; 264168404Spjd kstat_named_t arcstat_demand_data_misses; 265168404Spjd kstat_named_t arcstat_demand_metadata_hits; 266168404Spjd kstat_named_t arcstat_demand_metadata_misses; 267168404Spjd kstat_named_t arcstat_prefetch_data_hits; 268168404Spjd kstat_named_t arcstat_prefetch_data_misses; 269168404Spjd kstat_named_t arcstat_prefetch_metadata_hits; 270168404Spjd kstat_named_t arcstat_prefetch_metadata_misses; 271168404Spjd kstat_named_t arcstat_mru_hits; 272168404Spjd kstat_named_t arcstat_mru_ghost_hits; 273168404Spjd kstat_named_t arcstat_mfu_hits; 274168404Spjd kstat_named_t arcstat_mfu_ghost_hits; 275205231Skmacy kstat_named_t arcstat_allocated; 276168404Spjd kstat_named_t arcstat_deleted; 277205231Skmacy kstat_named_t arcstat_stolen; 278168404Spjd kstat_named_t arcstat_recycle_miss; 279168404Spjd kstat_named_t arcstat_mutex_miss; 280168404Spjd kstat_named_t arcstat_evict_skip; 281168404Spjd kstat_named_t arcstat_hash_elements; 282168404Spjd kstat_named_t arcstat_hash_elements_max; 283168404Spjd kstat_named_t arcstat_hash_collisions; 284168404Spjd kstat_named_t arcstat_hash_chains; 285168404Spjd kstat_named_t arcstat_hash_chain_max; 286168404Spjd kstat_named_t arcstat_p; 287168404Spjd kstat_named_t arcstat_c; 288168404Spjd kstat_named_t arcstat_c_min; 289168404Spjd kstat_named_t arcstat_c_max; 290168404Spjd kstat_named_t arcstat_size; 291185029Spjd kstat_named_t arcstat_hdr_size; 292185029Spjd kstat_named_t arcstat_l2_hits; 293185029Spjd kstat_named_t arcstat_l2_misses; 294185029Spjd kstat_named_t arcstat_l2_feeds; 295185029Spjd kstat_named_t arcstat_l2_rw_clash; 296185029Spjd kstat_named_t arcstat_l2_writes_sent; 297185029Spjd kstat_named_t arcstat_l2_writes_done; 298185029Spjd kstat_named_t arcstat_l2_writes_error; 299185029Spjd kstat_named_t arcstat_l2_writes_hdr_miss; 300185029Spjd kstat_named_t arcstat_l2_evict_lock_retry; 301185029Spjd kstat_named_t arcstat_l2_evict_reading; 302185029Spjd kstat_named_t arcstat_l2_free_on_write; 303185029Spjd kstat_named_t arcstat_l2_abort_lowmem; 304185029Spjd kstat_named_t arcstat_l2_cksum_bad; 305185029Spjd kstat_named_t arcstat_l2_io_error; 306185029Spjd kstat_named_t arcstat_l2_size; 307185029Spjd kstat_named_t arcstat_l2_hdr_size; 308185029Spjd kstat_named_t arcstat_memory_throttle_count; 309205231Skmacy kstat_named_t arcstat_l2_write_trylock_fail; 310205231Skmacy kstat_named_t arcstat_l2_write_in_l2; 311205231Skmacy kstat_named_t arcstat_l2_write_passed_headroom; 312205231Skmacy kstat_named_t arcstat_l2_write_spa_mismatch; 313205231Skmacy kstat_named_t arcstat_l2_write_hdr_io_in_progress; 314205231Skmacy kstat_named_t arcstat_l2_write_not_cacheable; 315205231Skmacy kstat_named_t arcstat_l2_write_full; 316205231Skmacy kstat_named_t arcstat_l2_write_buffer_iter; 317205231Skmacy kstat_named_t arcstat_l2_write_pios; 318205231Skmacy kstat_named_t arcstat_l2_write_bytes_written; 319205231Skmacy kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 320205231Skmacy kstat_named_t arcstat_l2_write_buffer_list_iter; 321205231Skmacy kstat_named_t arcstat_l2_write_buffer_list_null_iter; 322168404Spjd} arc_stats_t; 323168404Spjd 324168404Spjdstatic arc_stats_t arc_stats = { 325168404Spjd { "hits", KSTAT_DATA_UINT64 }, 326168404Spjd { "misses", KSTAT_DATA_UINT64 }, 327168404Spjd { "demand_data_hits", KSTAT_DATA_UINT64 }, 328168404Spjd { "demand_data_misses", KSTAT_DATA_UINT64 }, 329168404Spjd { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 330168404Spjd { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 331168404Spjd { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 332168404Spjd { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 333168404Spjd { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 334168404Spjd { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 335168404Spjd { "mru_hits", KSTAT_DATA_UINT64 }, 336168404Spjd { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 337168404Spjd { "mfu_hits", KSTAT_DATA_UINT64 }, 338168404Spjd { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 339205231Skmacy { "allocated", KSTAT_DATA_UINT64 }, 340168404Spjd { "deleted", KSTAT_DATA_UINT64 }, 341205231Skmacy { "stolen", KSTAT_DATA_UINT64 }, 342168404Spjd { "recycle_miss", KSTAT_DATA_UINT64 }, 343168404Spjd { "mutex_miss", KSTAT_DATA_UINT64 }, 344168404Spjd { "evict_skip", KSTAT_DATA_UINT64 }, 345168404Spjd { "hash_elements", KSTAT_DATA_UINT64 }, 346168404Spjd { "hash_elements_max", KSTAT_DATA_UINT64 }, 347168404Spjd { "hash_collisions", KSTAT_DATA_UINT64 }, 348168404Spjd { "hash_chains", KSTAT_DATA_UINT64 }, 349168404Spjd { "hash_chain_max", KSTAT_DATA_UINT64 }, 350168404Spjd { "p", KSTAT_DATA_UINT64 }, 351168404Spjd { "c", KSTAT_DATA_UINT64 }, 352168404Spjd { "c_min", KSTAT_DATA_UINT64 }, 353168404Spjd { "c_max", KSTAT_DATA_UINT64 }, 354185029Spjd { "size", KSTAT_DATA_UINT64 }, 355185029Spjd { "hdr_size", KSTAT_DATA_UINT64 }, 356185029Spjd { "l2_hits", KSTAT_DATA_UINT64 }, 357185029Spjd { "l2_misses", KSTAT_DATA_UINT64 }, 358185029Spjd { "l2_feeds", KSTAT_DATA_UINT64 }, 359185029Spjd { "l2_rw_clash", KSTAT_DATA_UINT64 }, 360185029Spjd { "l2_writes_sent", KSTAT_DATA_UINT64 }, 361185029Spjd { "l2_writes_done", KSTAT_DATA_UINT64 }, 362185029Spjd { "l2_writes_error", KSTAT_DATA_UINT64 }, 363185029Spjd { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 364185029Spjd { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 365185029Spjd { "l2_evict_reading", KSTAT_DATA_UINT64 }, 366185029Spjd { "l2_free_on_write", KSTAT_DATA_UINT64 }, 367185029Spjd { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 368185029Spjd { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 369185029Spjd { "l2_io_error", KSTAT_DATA_UINT64 }, 370185029Spjd { "l2_size", KSTAT_DATA_UINT64 }, 371185029Spjd { "l2_hdr_size", KSTAT_DATA_UINT64 }, 372205231Skmacy { "memory_throttle_count", KSTAT_DATA_UINT64 }, 373205231Skmacy { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 374205231Skmacy { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 375205231Skmacy { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 376205231Skmacy { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 377205231Skmacy { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 378205231Skmacy { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 379205231Skmacy { "l2_write_full", KSTAT_DATA_UINT64 }, 380205231Skmacy { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 381205231Skmacy { "l2_write_pios", KSTAT_DATA_UINT64 }, 382205231Skmacy { "l2_write_bytes_written", KSTAT_DATA_UINT64 }, 383205231Skmacy { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 384205231Skmacy { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 385205231Skmacy { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 } 386168404Spjd}; 387168404Spjd 388168404Spjd#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 389168404Spjd 390168404Spjd#define ARCSTAT_INCR(stat, val) \ 391168404Spjd atomic_add_64(&arc_stats.stat.value.ui64, (val)); 392168404Spjd 393168404Spjd#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 394168404Spjd#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 395168404Spjd 396168404Spjd#define ARCSTAT_MAX(stat, val) { \ 397168404Spjd uint64_t m; \ 398168404Spjd while ((val) > (m = arc_stats.stat.value.ui64) && \ 399168404Spjd (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 400168404Spjd continue; \ 401168404Spjd} 402168404Spjd 403168404Spjd#define ARCSTAT_MAXSTAT(stat) \ 404168404Spjd ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 405168404Spjd 406168404Spjd/* 407168404Spjd * We define a macro to allow ARC hits/misses to be easily broken down by 408168404Spjd * two separate conditions, giving a total of four different subtypes for 409168404Spjd * each of hits and misses (so eight statistics total). 410168404Spjd */ 411168404Spjd#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 412168404Spjd if (cond1) { \ 413168404Spjd if (cond2) { \ 414168404Spjd ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 415168404Spjd } else { \ 416168404Spjd ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 417168404Spjd } \ 418168404Spjd } else { \ 419168404Spjd if (cond2) { \ 420168404Spjd ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 421168404Spjd } else { \ 422168404Spjd ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 423168404Spjd } \ 424168404Spjd } 425168404Spjd 426168404Spjdkstat_t *arc_ksp; 427168404Spjdstatic arc_state_t *arc_anon; 428168404Spjdstatic arc_state_t *arc_mru; 429168404Spjdstatic arc_state_t *arc_mru_ghost; 430168404Spjdstatic arc_state_t *arc_mfu; 431168404Spjdstatic arc_state_t *arc_mfu_ghost; 432185029Spjdstatic arc_state_t *arc_l2c_only; 433168404Spjd 434168404Spjd/* 435168404Spjd * There are several ARC variables that are critical to export as kstats -- 436168404Spjd * but we don't want to have to grovel around in the kstat whenever we wish to 437168404Spjd * manipulate them. For these variables, we therefore define them to be in 438168404Spjd * terms of the statistic variable. This assures that we are not introducing 439168404Spjd * the possibility of inconsistency by having shadow copies of the variables, 440168404Spjd * while still allowing the code to be readable. 441168404Spjd */ 442168404Spjd#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 443168404Spjd#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 444168404Spjd#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 445168404Spjd#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 446168404Spjd#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 447168404Spjd 448168404Spjdstatic int arc_no_grow; /* Don't try to grow cache size */ 449168404Spjdstatic uint64_t arc_tempreserve; 450185029Spjdstatic uint64_t arc_meta_used; 451185029Spjdstatic uint64_t arc_meta_limit; 452185029Spjdstatic uint64_t arc_meta_max = 0; 453185029SpjdSYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RDTUN, 454185029Spjd &arc_meta_used, 0, "ARC metadata used"); 455185029SpjdSYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RDTUN, 456185029Spjd &arc_meta_limit, 0, "ARC metadata limit"); 457168404Spjd 458185029Spjdtypedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 459185029Spjd 460168404Spjdtypedef struct arc_callback arc_callback_t; 461168404Spjd 462168404Spjdstruct arc_callback { 463168404Spjd void *acb_private; 464168404Spjd arc_done_func_t *acb_done; 465168404Spjd arc_buf_t *acb_buf; 466168404Spjd zio_t *acb_zio_dummy; 467168404Spjd arc_callback_t *acb_next; 468168404Spjd}; 469168404Spjd 470168404Spjdtypedef struct arc_write_callback arc_write_callback_t; 471168404Spjd 472168404Spjdstruct arc_write_callback { 473168404Spjd void *awcb_private; 474168404Spjd arc_done_func_t *awcb_ready; 475168404Spjd arc_done_func_t *awcb_done; 476168404Spjd arc_buf_t *awcb_buf; 477168404Spjd}; 478168404Spjd 479168404Spjdstruct arc_buf_hdr { 480168404Spjd /* protected by hash lock */ 481168404Spjd dva_t b_dva; 482168404Spjd uint64_t b_birth; 483168404Spjd uint64_t b_cksum0; 484168404Spjd 485168404Spjd kmutex_t b_freeze_lock; 486168404Spjd zio_cksum_t *b_freeze_cksum; 487168404Spjd 488168404Spjd arc_buf_hdr_t *b_hash_next; 489168404Spjd arc_buf_t *b_buf; 490168404Spjd uint32_t b_flags; 491168404Spjd uint32_t b_datacnt; 492168404Spjd 493168404Spjd arc_callback_t *b_acb; 494168404Spjd kcondvar_t b_cv; 495168404Spjd 496168404Spjd /* immutable */ 497168404Spjd arc_buf_contents_t b_type; 498168404Spjd uint64_t b_size; 499168404Spjd spa_t *b_spa; 500168404Spjd 501168404Spjd /* protected by arc state mutex */ 502168404Spjd arc_state_t *b_state; 503168404Spjd list_node_t b_arc_node; 504168404Spjd 505168404Spjd /* updated atomically */ 506168404Spjd clock_t b_arc_access; 507168404Spjd 508168404Spjd /* self protecting */ 509168404Spjd refcount_t b_refcnt; 510185029Spjd 511185029Spjd l2arc_buf_hdr_t *b_l2hdr; 512185029Spjd list_node_t b_l2node; 513168404Spjd}; 514168404Spjd 515168404Spjdstatic arc_buf_t *arc_eviction_list; 516168404Spjdstatic kmutex_t arc_eviction_mtx; 517168404Spjdstatic arc_buf_hdr_t arc_eviction_hdr; 518168404Spjdstatic void arc_get_data_buf(arc_buf_t *buf); 519168404Spjdstatic void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock); 520185029Spjdstatic int arc_evict_needed(arc_buf_contents_t type); 521185029Spjdstatic void arc_evict_ghost(arc_state_t *state, spa_t *spa, int64_t bytes); 522168404Spjd 523168404Spjd#define GHOST_STATE(state) \ 524185029Spjd ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 525185029Spjd (state) == arc_l2c_only) 526168404Spjd 527168404Spjd/* 528168404Spjd * Private ARC flags. These flags are private ARC only flags that will show up 529168404Spjd * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can 530168404Spjd * be passed in as arc_flags in things like arc_read. However, these flags 531168404Spjd * should never be passed and should only be set by ARC code. When adding new 532168404Spjd * public flags, make sure not to smash the private ones. 533168404Spjd */ 534168404Spjd 535168404Spjd#define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */ 536168404Spjd#define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */ 537168404Spjd#define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */ 538168404Spjd#define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */ 539168404Spjd#define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */ 540168404Spjd#define ARC_INDIRECT (1 << 14) /* this is an indirect block */ 541185029Spjd#define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */ 542185029Spjd#define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */ 543185029Spjd#define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */ 544185029Spjd#define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */ 545185029Spjd#define ARC_STORED (1 << 19) /* has been store()d to */ 546168404Spjd 547168404Spjd#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE) 548168404Spjd#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS) 549168404Spjd#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR) 550168404Spjd#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ) 551168404Spjd#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE) 552185029Spjd#define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS) 553185029Spjd#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE) 554185029Spjd#define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \ 555185029Spjd (hdr)->b_l2hdr != NULL) 556185029Spjd#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING) 557185029Spjd#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED) 558185029Spjd#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD) 559168404Spjd 560168404Spjd/* 561185029Spjd * Other sizes 562185029Spjd */ 563185029Spjd 564185029Spjd#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 565185029Spjd#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 566185029Spjd 567185029Spjd/* 568168404Spjd * Hash table routines 569168404Spjd */ 570168404Spjd 571205253Skmacy#define HT_LOCK_PAD CACHE_LINE_SIZE 572168404Spjd 573168404Spjdstruct ht_lock { 574168404Spjd kmutex_t ht_lock; 575168404Spjd#ifdef _KERNEL 576168404Spjd unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 577168404Spjd#endif 578168404Spjd}; 579168404Spjd 580168404Spjd#define BUF_LOCKS 256 581168404Spjdtypedef struct buf_hash_table { 582168404Spjd uint64_t ht_mask; 583168404Spjd arc_buf_hdr_t **ht_table; 584205264Skmacy struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 585168404Spjd} buf_hash_table_t; 586168404Spjd 587168404Spjdstatic buf_hash_table_t buf_hash_table; 588168404Spjd 589168404Spjd#define BUF_HASH_INDEX(spa, dva, birth) \ 590168404Spjd (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 591168404Spjd#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 592168404Spjd#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 593168404Spjd#define HDR_LOCK(buf) \ 594168404Spjd (BUF_HASH_LOCK(BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth))) 595168404Spjd 596168404Spjduint64_t zfs_crc64_table[256]; 597168404Spjd 598205133Skmacy#ifdef ZIO_USE_UMA 599205133Skmacyextern kmem_cache_t *zio_buf_cache[]; 600205133Skmacyextern kmem_cache_t *zio_data_buf_cache[]; 601205133Skmacy#endif 602205133Skmacy 603185029Spjd/* 604185029Spjd * Level 2 ARC 605185029Spjd */ 606185029Spjd 607205231Skmacy#define L2ARC_WRITE_SIZE (64 * 1024 * 1024) /* initial write max */ 608205231Skmacy#define L2ARC_HEADROOM 128 /* num of writes */ 609185029Spjd#define L2ARC_FEED_SECS 1 /* caching interval */ 610205231Skmacy#define L2ARC_FEED_SECS_SHIFT 1 /* caching interval shift */ 611185029Spjd 612185029Spjd#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 613185029Spjd#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 614185029Spjd 615185029Spjd/* 616185029Spjd * L2ARC Performance Tunables 617185029Spjd */ 618185029Spjduint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 619185029Spjduint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 620185029Spjduint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 621185029Spjduint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 622205231Skmacyuint64_t l2arc_feed_secs_shift = L2ARC_FEED_SECS_SHIFT; /* interval seconds shift */ 623205231Skmacyboolean_t l2arc_noprefetch = B_FALSE; /* don't cache prefetch bufs */ 624185029Spjd 625205231Skmacy 626205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 627205231Skmacy &l2arc_write_max, 0, "max write size"); 628205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 629205231Skmacy &l2arc_write_boost, 0, "extra write during warmup"); 630205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 631205231Skmacy &l2arc_headroom, 0, "number of dev writes"); 632205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 633205231Skmacy &l2arc_feed_secs, 0, "interval seconds"); 634205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs_shift, CTLFLAG_RW, 635205231Skmacy &l2arc_feed_secs_shift, 0, "power of 2 division of feed seconds"); 636205231Skmacy 637205231SkmacySYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 638205231Skmacy &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 639205231Skmacy 640205231Skmacy 641205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 642205231Skmacy &ARC_anon.arcs_size, 0, "size of anonymous state"); 643205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 644205231Skmacy &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 645205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 646205231Skmacy &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 647205231Skmacy 648205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 649205231Skmacy &ARC_mru.arcs_size, 0, "size of mru state"); 650205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 651205231Skmacy &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 652205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 653205231Skmacy &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 654205231Skmacy 655205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 656205231Skmacy &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 657205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 658205231Skmacy &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 659205231Skmacy "size of metadata in mru ghost state"); 660205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 661205231Skmacy &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 662205231Skmacy "size of data in mru ghost state"); 663205231Skmacy 664205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 665205231Skmacy &ARC_mfu.arcs_size, 0, "size of mfu state"); 666205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 667205231Skmacy &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 668205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 669205231Skmacy &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 670205231Skmacy 671205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 672205231Skmacy &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 673205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 674205231Skmacy &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 675205231Skmacy "size of metadata in mfu ghost state"); 676205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 677205231Skmacy &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 678205231Skmacy "size of data in mfu ghost state"); 679205231Skmacy 680205231SkmacySYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 681205231Skmacy &ARC_l2c_only.arcs_size, 0, "size of mru state"); 682205231Skmacy 683185029Spjd/* 684185029Spjd * L2ARC Internals 685185029Spjd */ 686185029Spjdtypedef struct l2arc_dev { 687185029Spjd vdev_t *l2ad_vdev; /* vdev */ 688185029Spjd spa_t *l2ad_spa; /* spa */ 689185029Spjd uint64_t l2ad_hand; /* next write location */ 690185029Spjd uint64_t l2ad_write; /* desired write size, bytes */ 691185029Spjd uint64_t l2ad_boost; /* warmup write boost, bytes */ 692185029Spjd uint64_t l2ad_start; /* first addr on device */ 693185029Spjd uint64_t l2ad_end; /* last addr on device */ 694185029Spjd uint64_t l2ad_evict; /* last addr eviction reached */ 695185029Spjd boolean_t l2ad_first; /* first sweep through */ 696185029Spjd list_t *l2ad_buflist; /* buffer list */ 697185029Spjd list_node_t l2ad_node; /* device list node */ 698185029Spjd} l2arc_dev_t; 699185029Spjd 700185029Spjdstatic list_t L2ARC_dev_list; /* device list */ 701185029Spjdstatic list_t *l2arc_dev_list; /* device list pointer */ 702185029Spjdstatic kmutex_t l2arc_dev_mtx; /* device list mutex */ 703185029Spjdstatic l2arc_dev_t *l2arc_dev_last; /* last device used */ 704185029Spjdstatic kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 705185029Spjdstatic list_t L2ARC_free_on_write; /* free after write buf list */ 706185029Spjdstatic list_t *l2arc_free_on_write; /* free after write list ptr */ 707185029Spjdstatic kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 708185029Spjdstatic uint64_t l2arc_ndev; /* number of devices */ 709185029Spjd 710185029Spjdtypedef struct l2arc_read_callback { 711185029Spjd arc_buf_t *l2rcb_buf; /* read buffer */ 712185029Spjd spa_t *l2rcb_spa; /* spa */ 713185029Spjd blkptr_t l2rcb_bp; /* original blkptr */ 714185029Spjd zbookmark_t l2rcb_zb; /* original bookmark */ 715185029Spjd int l2rcb_flags; /* original flags */ 716185029Spjd} l2arc_read_callback_t; 717185029Spjd 718185029Spjdtypedef struct l2arc_write_callback { 719185029Spjd l2arc_dev_t *l2wcb_dev; /* device info */ 720185029Spjd arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 721185029Spjd} l2arc_write_callback_t; 722185029Spjd 723185029Spjdstruct l2arc_buf_hdr { 724185029Spjd /* protected by arc_buf_hdr mutex */ 725185029Spjd l2arc_dev_t *b_dev; /* L2ARC device */ 726185029Spjd daddr_t b_daddr; /* disk address, offset byte */ 727185029Spjd}; 728185029Spjd 729185029Spjdtypedef struct l2arc_data_free { 730185029Spjd /* protected by l2arc_free_on_write_mtx */ 731185029Spjd void *l2df_data; 732185029Spjd size_t l2df_size; 733185029Spjd void (*l2df_func)(void *, size_t); 734185029Spjd list_node_t l2df_list_node; 735185029Spjd} l2arc_data_free_t; 736185029Spjd 737185029Spjdstatic kmutex_t l2arc_feed_thr_lock; 738185029Spjdstatic kcondvar_t l2arc_feed_thr_cv; 739185029Spjdstatic uint8_t l2arc_thread_exit; 740185029Spjd 741185029Spjdstatic void l2arc_read_done(zio_t *zio); 742185029Spjdstatic void l2arc_hdr_stat_add(void); 743185029Spjdstatic void l2arc_hdr_stat_remove(void); 744185029Spjd 745168404Spjdstatic uint64_t 746185029Spjdbuf_hash(spa_t *spa, const dva_t *dva, uint64_t birth) 747168404Spjd{ 748168404Spjd uintptr_t spav = (uintptr_t)spa; 749168404Spjd uint8_t *vdva = (uint8_t *)dva; 750168404Spjd uint64_t crc = -1ULL; 751168404Spjd int i; 752168404Spjd 753168404Spjd ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 754168404Spjd 755168404Spjd for (i = 0; i < sizeof (dva_t); i++) 756168404Spjd crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 757168404Spjd 758168404Spjd crc ^= (spav>>8) ^ birth; 759168404Spjd 760168404Spjd return (crc); 761168404Spjd} 762168404Spjd 763168404Spjd#define BUF_EMPTY(buf) \ 764168404Spjd ((buf)->b_dva.dva_word[0] == 0 && \ 765168404Spjd (buf)->b_dva.dva_word[1] == 0 && \ 766168404Spjd (buf)->b_birth == 0) 767168404Spjd 768168404Spjd#define BUF_EQUAL(spa, dva, birth, buf) \ 769168404Spjd ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 770168404Spjd ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 771168404Spjd ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 772168404Spjd 773168404Spjdstatic arc_buf_hdr_t * 774185029Spjdbuf_hash_find(spa_t *spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp) 775168404Spjd{ 776168404Spjd uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 777168404Spjd kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 778168404Spjd arc_buf_hdr_t *buf; 779168404Spjd 780168404Spjd mutex_enter(hash_lock); 781168404Spjd for (buf = buf_hash_table.ht_table[idx]; buf != NULL; 782168404Spjd buf = buf->b_hash_next) { 783168404Spjd if (BUF_EQUAL(spa, dva, birth, buf)) { 784168404Spjd *lockp = hash_lock; 785168404Spjd return (buf); 786168404Spjd } 787168404Spjd } 788168404Spjd mutex_exit(hash_lock); 789168404Spjd *lockp = NULL; 790168404Spjd return (NULL); 791168404Spjd} 792168404Spjd 793168404Spjd/* 794168404Spjd * Insert an entry into the hash table. If there is already an element 795168404Spjd * equal to elem in the hash table, then the already existing element 796168404Spjd * will be returned and the new element will not be inserted. 797168404Spjd * Otherwise returns NULL. 798168404Spjd */ 799168404Spjdstatic arc_buf_hdr_t * 800168404Spjdbuf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp) 801168404Spjd{ 802168404Spjd uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 803168404Spjd kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 804168404Spjd arc_buf_hdr_t *fbuf; 805168404Spjd uint32_t i; 806168404Spjd 807168404Spjd ASSERT(!HDR_IN_HASH_TABLE(buf)); 808168404Spjd *lockp = hash_lock; 809168404Spjd mutex_enter(hash_lock); 810168404Spjd for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL; 811168404Spjd fbuf = fbuf->b_hash_next, i++) { 812168404Spjd if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf)) 813168404Spjd return (fbuf); 814168404Spjd } 815168404Spjd 816168404Spjd buf->b_hash_next = buf_hash_table.ht_table[idx]; 817168404Spjd buf_hash_table.ht_table[idx] = buf; 818168404Spjd buf->b_flags |= ARC_IN_HASH_TABLE; 819168404Spjd 820168404Spjd /* collect some hash table performance data */ 821168404Spjd if (i > 0) { 822168404Spjd ARCSTAT_BUMP(arcstat_hash_collisions); 823168404Spjd if (i == 1) 824168404Spjd ARCSTAT_BUMP(arcstat_hash_chains); 825168404Spjd 826168404Spjd ARCSTAT_MAX(arcstat_hash_chain_max, i); 827168404Spjd } 828168404Spjd 829168404Spjd ARCSTAT_BUMP(arcstat_hash_elements); 830168404Spjd ARCSTAT_MAXSTAT(arcstat_hash_elements); 831168404Spjd 832168404Spjd return (NULL); 833168404Spjd} 834168404Spjd 835168404Spjdstatic void 836168404Spjdbuf_hash_remove(arc_buf_hdr_t *buf) 837168404Spjd{ 838168404Spjd arc_buf_hdr_t *fbuf, **bufp; 839168404Spjd uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 840168404Spjd 841168404Spjd ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 842168404Spjd ASSERT(HDR_IN_HASH_TABLE(buf)); 843168404Spjd 844168404Spjd bufp = &buf_hash_table.ht_table[idx]; 845168404Spjd while ((fbuf = *bufp) != buf) { 846168404Spjd ASSERT(fbuf != NULL); 847168404Spjd bufp = &fbuf->b_hash_next; 848168404Spjd } 849168404Spjd *bufp = buf->b_hash_next; 850168404Spjd buf->b_hash_next = NULL; 851168404Spjd buf->b_flags &= ~ARC_IN_HASH_TABLE; 852168404Spjd 853168404Spjd /* collect some hash table performance data */ 854168404Spjd ARCSTAT_BUMPDOWN(arcstat_hash_elements); 855168404Spjd 856168404Spjd if (buf_hash_table.ht_table[idx] && 857168404Spjd buf_hash_table.ht_table[idx]->b_hash_next == NULL) 858168404Spjd ARCSTAT_BUMPDOWN(arcstat_hash_chains); 859168404Spjd} 860168404Spjd 861168404Spjd/* 862168404Spjd * Global data structures and functions for the buf kmem cache. 863168404Spjd */ 864168404Spjdstatic kmem_cache_t *hdr_cache; 865168404Spjdstatic kmem_cache_t *buf_cache; 866168404Spjd 867168404Spjdstatic void 868168404Spjdbuf_fini(void) 869168404Spjd{ 870168404Spjd int i; 871168404Spjd 872168404Spjd kmem_free(buf_hash_table.ht_table, 873168404Spjd (buf_hash_table.ht_mask + 1) * sizeof (void *)); 874168404Spjd for (i = 0; i < BUF_LOCKS; i++) 875168404Spjd mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 876168404Spjd kmem_cache_destroy(hdr_cache); 877168404Spjd kmem_cache_destroy(buf_cache); 878168404Spjd} 879168404Spjd 880168404Spjd/* 881168404Spjd * Constructor callback - called when the cache is empty 882168404Spjd * and a new buf is requested. 883168404Spjd */ 884168404Spjd/* ARGSUSED */ 885168404Spjdstatic int 886168404Spjdhdr_cons(void *vbuf, void *unused, int kmflag) 887168404Spjd{ 888168404Spjd arc_buf_hdr_t *buf = vbuf; 889168404Spjd 890168404Spjd bzero(buf, sizeof (arc_buf_hdr_t)); 891168404Spjd refcount_create(&buf->b_refcnt); 892168404Spjd cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL); 893185029Spjd mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 894185029Spjd 895185029Spjd ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 896168404Spjd return (0); 897168404Spjd} 898168404Spjd 899185029Spjd/* ARGSUSED */ 900185029Spjdstatic int 901185029Spjdbuf_cons(void *vbuf, void *unused, int kmflag) 902185029Spjd{ 903185029Spjd arc_buf_t *buf = vbuf; 904185029Spjd 905185029Spjd bzero(buf, sizeof (arc_buf_t)); 906185029Spjd rw_init(&buf->b_lock, NULL, RW_DEFAULT, NULL); 907185029Spjd return (0); 908185029Spjd} 909185029Spjd 910168404Spjd/* 911168404Spjd * Destructor callback - called when a cached buf is 912168404Spjd * no longer required. 913168404Spjd */ 914168404Spjd/* ARGSUSED */ 915168404Spjdstatic void 916168404Spjdhdr_dest(void *vbuf, void *unused) 917168404Spjd{ 918168404Spjd arc_buf_hdr_t *buf = vbuf; 919168404Spjd 920168404Spjd refcount_destroy(&buf->b_refcnt); 921168404Spjd cv_destroy(&buf->b_cv); 922185029Spjd mutex_destroy(&buf->b_freeze_lock); 923185029Spjd 924185029Spjd ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 925168404Spjd} 926168404Spjd 927185029Spjd/* ARGSUSED */ 928185029Spjdstatic void 929185029Spjdbuf_dest(void *vbuf, void *unused) 930185029Spjd{ 931185029Spjd arc_buf_t *buf = vbuf; 932185029Spjd 933185029Spjd rw_destroy(&buf->b_lock); 934185029Spjd} 935185029Spjd 936168404Spjd/* 937168404Spjd * Reclaim callback -- invoked when memory is low. 938168404Spjd */ 939168404Spjd/* ARGSUSED */ 940168404Spjdstatic void 941168404Spjdhdr_recl(void *unused) 942168404Spjd{ 943168404Spjd dprintf("hdr_recl called\n"); 944168404Spjd /* 945168404Spjd * umem calls the reclaim func when we destroy the buf cache, 946168404Spjd * which is after we do arc_fini(). 947168404Spjd */ 948168404Spjd if (!arc_dead) 949168404Spjd cv_signal(&arc_reclaim_thr_cv); 950168404Spjd} 951168404Spjd 952168404Spjdstatic void 953168404Spjdbuf_init(void) 954168404Spjd{ 955168404Spjd uint64_t *ct; 956168404Spjd uint64_t hsize = 1ULL << 12; 957168404Spjd int i, j; 958168404Spjd 959168404Spjd /* 960168404Spjd * The hash table is big enough to fill all of physical memory 961168404Spjd * with an average 64K block size. The table will take up 962168404Spjd * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers). 963168404Spjd */ 964168696Spjd while (hsize * 65536 < (uint64_t)physmem * PAGESIZE) 965168404Spjd hsize <<= 1; 966168404Spjdretry: 967168404Spjd buf_hash_table.ht_mask = hsize - 1; 968168404Spjd buf_hash_table.ht_table = 969168404Spjd kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 970168404Spjd if (buf_hash_table.ht_table == NULL) { 971168404Spjd ASSERT(hsize > (1ULL << 8)); 972168404Spjd hsize >>= 1; 973168404Spjd goto retry; 974168404Spjd } 975168404Spjd 976168404Spjd hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 977168404Spjd 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 978168404Spjd buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 979185029Spjd 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 980168404Spjd 981168404Spjd for (i = 0; i < 256; i++) 982168404Spjd for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 983168404Spjd *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 984168404Spjd 985168404Spjd for (i = 0; i < BUF_LOCKS; i++) { 986168404Spjd mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 987168404Spjd NULL, MUTEX_DEFAULT, NULL); 988168404Spjd } 989168404Spjd} 990168404Spjd 991168404Spjd#define ARC_MINTIME (hz>>4) /* 62 ms */ 992168404Spjd 993168404Spjdstatic void 994168404Spjdarc_cksum_verify(arc_buf_t *buf) 995168404Spjd{ 996168404Spjd zio_cksum_t zc; 997168404Spjd 998168404Spjd if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 999168404Spjd return; 1000168404Spjd 1001168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1002168404Spjd if (buf->b_hdr->b_freeze_cksum == NULL || 1003168404Spjd (buf->b_hdr->b_flags & ARC_IO_ERROR)) { 1004168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1005168404Spjd return; 1006168404Spjd } 1007168404Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1008168404Spjd if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1009168404Spjd panic("buffer modified while frozen!"); 1010168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1011168404Spjd} 1012168404Spjd 1013185029Spjdstatic int 1014185029Spjdarc_cksum_equal(arc_buf_t *buf) 1015185029Spjd{ 1016185029Spjd zio_cksum_t zc; 1017185029Spjd int equal; 1018185029Spjd 1019185029Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1020185029Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1021185029Spjd equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1022185029Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1023185029Spjd 1024185029Spjd return (equal); 1025185029Spjd} 1026185029Spjd 1027168404Spjdstatic void 1028185029Spjdarc_cksum_compute(arc_buf_t *buf, boolean_t force) 1029168404Spjd{ 1030185029Spjd if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1031168404Spjd return; 1032168404Spjd 1033168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1034168404Spjd if (buf->b_hdr->b_freeze_cksum != NULL) { 1035168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1036168404Spjd return; 1037168404Spjd } 1038168404Spjd buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1039168404Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1040168404Spjd buf->b_hdr->b_freeze_cksum); 1041168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1042168404Spjd} 1043168404Spjd 1044168404Spjdvoid 1045168404Spjdarc_buf_thaw(arc_buf_t *buf) 1046168404Spjd{ 1047185029Spjd if (zfs_flags & ZFS_DEBUG_MODIFY) { 1048185029Spjd if (buf->b_hdr->b_state != arc_anon) 1049185029Spjd panic("modifying non-anon buffer!"); 1050185029Spjd if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS) 1051185029Spjd panic("modifying buffer while i/o in progress!"); 1052185029Spjd arc_cksum_verify(buf); 1053185029Spjd } 1054168404Spjd 1055168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1056168404Spjd if (buf->b_hdr->b_freeze_cksum != NULL) { 1057168404Spjd kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1058168404Spjd buf->b_hdr->b_freeze_cksum = NULL; 1059168404Spjd } 1060168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1061168404Spjd} 1062168404Spjd 1063168404Spjdvoid 1064168404Spjdarc_buf_freeze(arc_buf_t *buf) 1065168404Spjd{ 1066168404Spjd if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1067168404Spjd return; 1068168404Spjd 1069168404Spjd ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1070168404Spjd buf->b_hdr->b_state == arc_anon); 1071185029Spjd arc_cksum_compute(buf, B_FALSE); 1072168404Spjd} 1073168404Spjd 1074168404Spjdstatic void 1075205231Skmacyget_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock) 1076205231Skmacy{ 1077205231Skmacy uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth); 1078205231Skmacy 1079205231Skmacy if (ab->b_type == ARC_BUFC_METADATA) 1080205231Skmacy buf_hashid &= (ARC_BUFC_NUMMETADATALISTS-1); 1081205231Skmacy else { 1082205231Skmacy buf_hashid &= (ARC_BUFC_NUMDATALISTS-1); 1083205231Skmacy buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1084205231Skmacy } 1085205231Skmacy 1086205231Skmacy *list = &state->arcs_lists[buf_hashid]; 1087205231Skmacy *lock = ARCS_LOCK(state, buf_hashid); 1088205231Skmacy} 1089205231Skmacy 1090205231Skmacy 1091205231Skmacystatic void 1092168404Spjdadd_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1093168404Spjd{ 1094205231Skmacy 1095168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 1096168404Spjd 1097168404Spjd if ((refcount_add(&ab->b_refcnt, tag) == 1) && 1098168404Spjd (ab->b_state != arc_anon)) { 1099205231Skmacy list_t *list; 1100205231Skmacy kmutex_t *lock; 1101168404Spjd uint64_t delta = ab->b_size * ab->b_datacnt; 1102185029Spjd uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type]; 1103168404Spjd 1104205231Skmacy get_buf_info(ab, ab->b_state, &list, &lock); 1105205231Skmacy ASSERT(!MUTEX_HELD(lock)); 1106205231Skmacy mutex_enter(lock); 1107168404Spjd ASSERT(list_link_active(&ab->b_arc_node)); 1108185029Spjd list_remove(list, ab); 1109205231Skmacy mutex_exit(lock); 1110205231Skmacy 1111168404Spjd if (GHOST_STATE(ab->b_state)) { 1112168404Spjd ASSERT3U(ab->b_datacnt, ==, 0); 1113168404Spjd ASSERT3P(ab->b_buf, ==, NULL); 1114168404Spjd delta = ab->b_size; 1115168404Spjd } 1116168404Spjd ASSERT(delta > 0); 1117185029Spjd ASSERT3U(*size, >=, delta); 1118185029Spjd atomic_add_64(size, -delta); 1119185029Spjd /* remove the prefetch flag if we get a reference */ 1120168404Spjd if (ab->b_flags & ARC_PREFETCH) 1121168404Spjd ab->b_flags &= ~ARC_PREFETCH; 1122168404Spjd } 1123168404Spjd} 1124168404Spjd 1125168404Spjdstatic int 1126168404Spjdremove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1127168404Spjd{ 1128168404Spjd int cnt; 1129168404Spjd arc_state_t *state = ab->b_state; 1130168404Spjd 1131168404Spjd ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1132168404Spjd ASSERT(!GHOST_STATE(state)); 1133168404Spjd 1134168404Spjd if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) && 1135168404Spjd (state != arc_anon)) { 1136185029Spjd uint64_t *size = &state->arcs_lsize[ab->b_type]; 1137205231Skmacy list_t *list; 1138205231Skmacy kmutex_t *lock; 1139185029Spjd 1140205231Skmacy get_buf_info(ab, state, &list, &lock); 1141205231Skmacy 1142205231Skmacy ASSERT(!MUTEX_HELD(lock)); 1143205231Skmacy mutex_enter(lock); 1144168404Spjd ASSERT(!list_link_active(&ab->b_arc_node)); 1145205231Skmacy list_insert_head(list, ab); 1146205231Skmacy mutex_exit(lock); 1147205231Skmacy 1148168404Spjd ASSERT(ab->b_datacnt > 0); 1149185029Spjd atomic_add_64(size, ab->b_size * ab->b_datacnt); 1150168404Spjd } 1151168404Spjd return (cnt); 1152168404Spjd} 1153168404Spjd 1154168404Spjd/* 1155168404Spjd * Move the supplied buffer to the indicated state. The mutex 1156168404Spjd * for the buffer must be held by the caller. 1157168404Spjd */ 1158168404Spjdstatic void 1159168404Spjdarc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock) 1160168404Spjd{ 1161168404Spjd arc_state_t *old_state = ab->b_state; 1162168404Spjd int64_t refcnt = refcount_count(&ab->b_refcnt); 1163168404Spjd uint64_t from_delta, to_delta; 1164205231Skmacy list_t *list; 1165205231Skmacy kmutex_t *lock; 1166168404Spjd 1167168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 1168168404Spjd ASSERT(new_state != old_state); 1169168404Spjd ASSERT(refcnt == 0 || ab->b_datacnt > 0); 1170168404Spjd ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state)); 1171168404Spjd 1172168404Spjd from_delta = to_delta = ab->b_datacnt * ab->b_size; 1173168404Spjd 1174168404Spjd /* 1175168404Spjd * If this buffer is evictable, transfer it from the 1176168404Spjd * old state list to the new state list. 1177168404Spjd */ 1178168404Spjd if (refcnt == 0) { 1179168404Spjd if (old_state != arc_anon) { 1180205231Skmacy int use_mutex; 1181185029Spjd uint64_t *size = &old_state->arcs_lsize[ab->b_type]; 1182168404Spjd 1183205231Skmacy get_buf_info(ab, old_state, &list, &lock); 1184205231Skmacy use_mutex = !MUTEX_HELD(lock); 1185205231Skmacy 1186168404Spjd if (use_mutex) 1187205231Skmacy mutex_enter(lock); 1188168404Spjd 1189168404Spjd ASSERT(list_link_active(&ab->b_arc_node)); 1190205231Skmacy list_remove(list, ab); 1191168404Spjd 1192168404Spjd /* 1193168404Spjd * If prefetching out of the ghost cache, 1194168404Spjd * we will have a non-null datacnt. 1195168404Spjd */ 1196168404Spjd if (GHOST_STATE(old_state) && ab->b_datacnt == 0) { 1197168404Spjd /* ghost elements have a ghost size */ 1198168404Spjd ASSERT(ab->b_buf == NULL); 1199168404Spjd from_delta = ab->b_size; 1200168404Spjd } 1201185029Spjd ASSERT3U(*size, >=, from_delta); 1202185029Spjd atomic_add_64(size, -from_delta); 1203168404Spjd 1204168404Spjd if (use_mutex) 1205205231Skmacy mutex_exit(lock); 1206168404Spjd } 1207168404Spjd if (new_state != arc_anon) { 1208205231Skmacy int use_mutex; 1209185029Spjd uint64_t *size = &new_state->arcs_lsize[ab->b_type]; 1210168404Spjd 1211205231Skmacy get_buf_info(ab, new_state, &list, &lock); 1212205231Skmacy use_mutex = !MUTEX_HELD(lock); 1213205231Skmacy 1214205231Skmacy 1215168404Spjd if (use_mutex) 1216205231Skmacy mutex_enter(lock); 1217168404Spjd 1218205231Skmacy list_insert_head(list, ab); 1219168404Spjd 1220168404Spjd /* ghost elements have a ghost size */ 1221168404Spjd if (GHOST_STATE(new_state)) { 1222168404Spjd ASSERT(ab->b_datacnt == 0); 1223168404Spjd ASSERT(ab->b_buf == NULL); 1224168404Spjd to_delta = ab->b_size; 1225168404Spjd } 1226185029Spjd atomic_add_64(size, to_delta); 1227168404Spjd 1228168404Spjd if (use_mutex) 1229205231Skmacy mutex_exit(lock); 1230168404Spjd } 1231168404Spjd } 1232168404Spjd 1233168404Spjd ASSERT(!BUF_EMPTY(ab)); 1234185029Spjd if (new_state == arc_anon) { 1235168404Spjd buf_hash_remove(ab); 1236168404Spjd } 1237168404Spjd 1238168404Spjd /* adjust state sizes */ 1239168404Spjd if (to_delta) 1240168404Spjd atomic_add_64(&new_state->arcs_size, to_delta); 1241168404Spjd if (from_delta) { 1242168404Spjd ASSERT3U(old_state->arcs_size, >=, from_delta); 1243168404Spjd atomic_add_64(&old_state->arcs_size, -from_delta); 1244168404Spjd } 1245168404Spjd ab->b_state = new_state; 1246185029Spjd 1247185029Spjd /* adjust l2arc hdr stats */ 1248185029Spjd if (new_state == arc_l2c_only) 1249185029Spjd l2arc_hdr_stat_add(); 1250185029Spjd else if (old_state == arc_l2c_only) 1251185029Spjd l2arc_hdr_stat_remove(); 1252168404Spjd} 1253168404Spjd 1254185029Spjdvoid 1255185029Spjdarc_space_consume(uint64_t space) 1256185029Spjd{ 1257185029Spjd atomic_add_64(&arc_meta_used, space); 1258185029Spjd atomic_add_64(&arc_size, space); 1259185029Spjd} 1260185029Spjd 1261185029Spjdvoid 1262185029Spjdarc_space_return(uint64_t space) 1263185029Spjd{ 1264185029Spjd ASSERT(arc_meta_used >= space); 1265185029Spjd if (arc_meta_max < arc_meta_used) 1266185029Spjd arc_meta_max = arc_meta_used; 1267185029Spjd atomic_add_64(&arc_meta_used, -space); 1268185029Spjd ASSERT(arc_size >= space); 1269185029Spjd atomic_add_64(&arc_size, -space); 1270185029Spjd} 1271185029Spjd 1272185029Spjdvoid * 1273185029Spjdarc_data_buf_alloc(uint64_t size) 1274185029Spjd{ 1275185029Spjd if (arc_evict_needed(ARC_BUFC_DATA)) 1276185029Spjd cv_signal(&arc_reclaim_thr_cv); 1277185029Spjd atomic_add_64(&arc_size, size); 1278185029Spjd return (zio_data_buf_alloc(size)); 1279185029Spjd} 1280185029Spjd 1281185029Spjdvoid 1282185029Spjdarc_data_buf_free(void *buf, uint64_t size) 1283185029Spjd{ 1284185029Spjd zio_data_buf_free(buf, size); 1285185029Spjd ASSERT(arc_size >= size); 1286185029Spjd atomic_add_64(&arc_size, -size); 1287185029Spjd} 1288185029Spjd 1289168404Spjdarc_buf_t * 1290168404Spjdarc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1291168404Spjd{ 1292168404Spjd arc_buf_hdr_t *hdr; 1293168404Spjd arc_buf_t *buf; 1294168404Spjd 1295168404Spjd ASSERT3U(size, >, 0); 1296185029Spjd hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1297168404Spjd ASSERT(BUF_EMPTY(hdr)); 1298168404Spjd hdr->b_size = size; 1299168404Spjd hdr->b_type = type; 1300168404Spjd hdr->b_spa = spa; 1301168404Spjd hdr->b_state = arc_anon; 1302168404Spjd hdr->b_arc_access = 0; 1303185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1304168404Spjd buf->b_hdr = hdr; 1305168404Spjd buf->b_data = NULL; 1306168404Spjd buf->b_efunc = NULL; 1307168404Spjd buf->b_private = NULL; 1308168404Spjd buf->b_next = NULL; 1309168404Spjd hdr->b_buf = buf; 1310168404Spjd arc_get_data_buf(buf); 1311168404Spjd hdr->b_datacnt = 1; 1312168404Spjd hdr->b_flags = 0; 1313168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1314168404Spjd (void) refcount_add(&hdr->b_refcnt, tag); 1315168404Spjd 1316168404Spjd return (buf); 1317168404Spjd} 1318168404Spjd 1319168404Spjdstatic arc_buf_t * 1320168404Spjdarc_buf_clone(arc_buf_t *from) 1321168404Spjd{ 1322168404Spjd arc_buf_t *buf; 1323168404Spjd arc_buf_hdr_t *hdr = from->b_hdr; 1324168404Spjd uint64_t size = hdr->b_size; 1325168404Spjd 1326185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1327168404Spjd buf->b_hdr = hdr; 1328168404Spjd buf->b_data = NULL; 1329168404Spjd buf->b_efunc = NULL; 1330168404Spjd buf->b_private = NULL; 1331168404Spjd buf->b_next = hdr->b_buf; 1332168404Spjd hdr->b_buf = buf; 1333168404Spjd arc_get_data_buf(buf); 1334168404Spjd bcopy(from->b_data, buf->b_data, size); 1335168404Spjd hdr->b_datacnt += 1; 1336168404Spjd return (buf); 1337168404Spjd} 1338168404Spjd 1339168404Spjdvoid 1340168404Spjdarc_buf_add_ref(arc_buf_t *buf, void* tag) 1341168404Spjd{ 1342168404Spjd arc_buf_hdr_t *hdr; 1343168404Spjd kmutex_t *hash_lock; 1344168404Spjd 1345168404Spjd /* 1346185029Spjd * Check to see if this buffer is evicted. Callers 1347185029Spjd * must verify b_data != NULL to know if the add_ref 1348185029Spjd * was successful. 1349168404Spjd */ 1350185029Spjd rw_enter(&buf->b_lock, RW_READER); 1351185029Spjd if (buf->b_data == NULL) { 1352185029Spjd rw_exit(&buf->b_lock); 1353168404Spjd return; 1354168404Spjd } 1355185029Spjd hdr = buf->b_hdr; 1356185029Spjd ASSERT(hdr != NULL); 1357168404Spjd hash_lock = HDR_LOCK(hdr); 1358168404Spjd mutex_enter(hash_lock); 1359185029Spjd rw_exit(&buf->b_lock); 1360168404Spjd 1361168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1362168404Spjd add_reference(hdr, hash_lock, tag); 1363168404Spjd arc_access(hdr, hash_lock); 1364168404Spjd mutex_exit(hash_lock); 1365168404Spjd ARCSTAT_BUMP(arcstat_hits); 1366168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 1367168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1368168404Spjd data, metadata, hits); 1369168404Spjd} 1370168404Spjd 1371185029Spjd/* 1372185029Spjd * Free the arc data buffer. If it is an l2arc write in progress, 1373185029Spjd * the buffer is placed on l2arc_free_on_write to be freed later. 1374185029Spjd */ 1375168404Spjdstatic void 1376185029Spjdarc_buf_data_free(arc_buf_hdr_t *hdr, void (*free_func)(void *, size_t), 1377185029Spjd void *data, size_t size) 1378185029Spjd{ 1379185029Spjd if (HDR_L2_WRITING(hdr)) { 1380185029Spjd l2arc_data_free_t *df; 1381185029Spjd df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1382185029Spjd df->l2df_data = data; 1383185029Spjd df->l2df_size = size; 1384185029Spjd df->l2df_func = free_func; 1385185029Spjd mutex_enter(&l2arc_free_on_write_mtx); 1386185029Spjd list_insert_head(l2arc_free_on_write, df); 1387185029Spjd mutex_exit(&l2arc_free_on_write_mtx); 1388185029Spjd ARCSTAT_BUMP(arcstat_l2_free_on_write); 1389185029Spjd } else { 1390185029Spjd free_func(data, size); 1391185029Spjd } 1392185029Spjd} 1393185029Spjd 1394185029Spjdstatic void 1395168404Spjdarc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all) 1396168404Spjd{ 1397168404Spjd arc_buf_t **bufp; 1398168404Spjd 1399168404Spjd /* free up data associated with the buf */ 1400168404Spjd if (buf->b_data) { 1401168404Spjd arc_state_t *state = buf->b_hdr->b_state; 1402168404Spjd uint64_t size = buf->b_hdr->b_size; 1403168404Spjd arc_buf_contents_t type = buf->b_hdr->b_type; 1404168404Spjd 1405168404Spjd arc_cksum_verify(buf); 1406168404Spjd if (!recycle) { 1407168404Spjd if (type == ARC_BUFC_METADATA) { 1408185029Spjd arc_buf_data_free(buf->b_hdr, zio_buf_free, 1409185029Spjd buf->b_data, size); 1410185029Spjd arc_space_return(size); 1411168404Spjd } else { 1412168404Spjd ASSERT(type == ARC_BUFC_DATA); 1413185029Spjd arc_buf_data_free(buf->b_hdr, 1414185029Spjd zio_data_buf_free, buf->b_data, size); 1415185029Spjd atomic_add_64(&arc_size, -size); 1416168404Spjd } 1417168404Spjd } 1418168404Spjd if (list_link_active(&buf->b_hdr->b_arc_node)) { 1419185029Spjd uint64_t *cnt = &state->arcs_lsize[type]; 1420185029Spjd 1421168404Spjd ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1422168404Spjd ASSERT(state != arc_anon); 1423185029Spjd 1424185029Spjd ASSERT3U(*cnt, >=, size); 1425185029Spjd atomic_add_64(cnt, -size); 1426168404Spjd } 1427168404Spjd ASSERT3U(state->arcs_size, >=, size); 1428168404Spjd atomic_add_64(&state->arcs_size, -size); 1429168404Spjd buf->b_data = NULL; 1430168404Spjd ASSERT(buf->b_hdr->b_datacnt > 0); 1431168404Spjd buf->b_hdr->b_datacnt -= 1; 1432168404Spjd } 1433168404Spjd 1434168404Spjd /* only remove the buf if requested */ 1435168404Spjd if (!all) 1436168404Spjd return; 1437168404Spjd 1438168404Spjd /* remove the buf from the hdr list */ 1439168404Spjd for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1440168404Spjd continue; 1441168404Spjd *bufp = buf->b_next; 1442168404Spjd 1443168404Spjd ASSERT(buf->b_efunc == NULL); 1444168404Spjd 1445168404Spjd /* clean up the buf */ 1446168404Spjd buf->b_hdr = NULL; 1447168404Spjd kmem_cache_free(buf_cache, buf); 1448168404Spjd} 1449168404Spjd 1450168404Spjdstatic void 1451168404Spjdarc_hdr_destroy(arc_buf_hdr_t *hdr) 1452168404Spjd{ 1453168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1454168404Spjd ASSERT3P(hdr->b_state, ==, arc_anon); 1455168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1456185029Spjd ASSERT(!(hdr->b_flags & ARC_STORED)); 1457168404Spjd 1458185029Spjd if (hdr->b_l2hdr != NULL) { 1459185029Spjd if (!MUTEX_HELD(&l2arc_buflist_mtx)) { 1460185029Spjd /* 1461185029Spjd * To prevent arc_free() and l2arc_evict() from 1462185029Spjd * attempting to free the same buffer at the same time, 1463185029Spjd * a FREE_IN_PROGRESS flag is given to arc_free() to 1464185029Spjd * give it priority. l2arc_evict() can't destroy this 1465185029Spjd * header while we are waiting on l2arc_buflist_mtx. 1466185029Spjd * 1467185029Spjd * The hdr may be removed from l2ad_buflist before we 1468185029Spjd * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1469185029Spjd */ 1470185029Spjd mutex_enter(&l2arc_buflist_mtx); 1471185029Spjd if (hdr->b_l2hdr != NULL) { 1472185029Spjd list_remove(hdr->b_l2hdr->b_dev->l2ad_buflist, 1473185029Spjd hdr); 1474185029Spjd } 1475185029Spjd mutex_exit(&l2arc_buflist_mtx); 1476185029Spjd } else { 1477185029Spjd list_remove(hdr->b_l2hdr->b_dev->l2ad_buflist, hdr); 1478185029Spjd } 1479185029Spjd ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1480185029Spjd kmem_free(hdr->b_l2hdr, sizeof (l2arc_buf_hdr_t)); 1481185029Spjd if (hdr->b_state == arc_l2c_only) 1482185029Spjd l2arc_hdr_stat_remove(); 1483185029Spjd hdr->b_l2hdr = NULL; 1484185029Spjd } 1485185029Spjd 1486168404Spjd if (!BUF_EMPTY(hdr)) { 1487168404Spjd ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1488168404Spjd bzero(&hdr->b_dva, sizeof (dva_t)); 1489168404Spjd hdr->b_birth = 0; 1490168404Spjd hdr->b_cksum0 = 0; 1491168404Spjd } 1492168404Spjd while (hdr->b_buf) { 1493168404Spjd arc_buf_t *buf = hdr->b_buf; 1494168404Spjd 1495168404Spjd if (buf->b_efunc) { 1496168404Spjd mutex_enter(&arc_eviction_mtx); 1497185029Spjd rw_enter(&buf->b_lock, RW_WRITER); 1498168404Spjd ASSERT(buf->b_hdr != NULL); 1499168404Spjd arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1500168404Spjd hdr->b_buf = buf->b_next; 1501168404Spjd buf->b_hdr = &arc_eviction_hdr; 1502168404Spjd buf->b_next = arc_eviction_list; 1503168404Spjd arc_eviction_list = buf; 1504185029Spjd rw_exit(&buf->b_lock); 1505168404Spjd mutex_exit(&arc_eviction_mtx); 1506168404Spjd } else { 1507168404Spjd arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1508168404Spjd } 1509168404Spjd } 1510168404Spjd if (hdr->b_freeze_cksum != NULL) { 1511168404Spjd kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1512168404Spjd hdr->b_freeze_cksum = NULL; 1513168404Spjd } 1514168404Spjd 1515168404Spjd ASSERT(!list_link_active(&hdr->b_arc_node)); 1516168404Spjd ASSERT3P(hdr->b_hash_next, ==, NULL); 1517168404Spjd ASSERT3P(hdr->b_acb, ==, NULL); 1518168404Spjd kmem_cache_free(hdr_cache, hdr); 1519168404Spjd} 1520168404Spjd 1521168404Spjdvoid 1522168404Spjdarc_buf_free(arc_buf_t *buf, void *tag) 1523168404Spjd{ 1524168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 1525168404Spjd int hashed = hdr->b_state != arc_anon; 1526168404Spjd 1527168404Spjd ASSERT(buf->b_efunc == NULL); 1528168404Spjd ASSERT(buf->b_data != NULL); 1529168404Spjd 1530168404Spjd if (hashed) { 1531168404Spjd kmutex_t *hash_lock = HDR_LOCK(hdr); 1532168404Spjd 1533168404Spjd mutex_enter(hash_lock); 1534168404Spjd (void) remove_reference(hdr, hash_lock, tag); 1535168404Spjd if (hdr->b_datacnt > 1) 1536168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1537168404Spjd else 1538168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 1539168404Spjd mutex_exit(hash_lock); 1540168404Spjd } else if (HDR_IO_IN_PROGRESS(hdr)) { 1541168404Spjd int destroy_hdr; 1542168404Spjd /* 1543168404Spjd * We are in the middle of an async write. Don't destroy 1544168404Spjd * this buffer unless the write completes before we finish 1545168404Spjd * decrementing the reference count. 1546168404Spjd */ 1547168404Spjd mutex_enter(&arc_eviction_mtx); 1548168404Spjd (void) remove_reference(hdr, NULL, tag); 1549168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1550168404Spjd destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1551168404Spjd mutex_exit(&arc_eviction_mtx); 1552168404Spjd if (destroy_hdr) 1553168404Spjd arc_hdr_destroy(hdr); 1554168404Spjd } else { 1555168404Spjd if (remove_reference(hdr, NULL, tag) > 0) { 1556168404Spjd ASSERT(HDR_IO_ERROR(hdr)); 1557168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1558168404Spjd } else { 1559168404Spjd arc_hdr_destroy(hdr); 1560168404Spjd } 1561168404Spjd } 1562168404Spjd} 1563168404Spjd 1564168404Spjdint 1565168404Spjdarc_buf_remove_ref(arc_buf_t *buf, void* tag) 1566168404Spjd{ 1567168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 1568168404Spjd kmutex_t *hash_lock = HDR_LOCK(hdr); 1569168404Spjd int no_callback = (buf->b_efunc == NULL); 1570168404Spjd 1571168404Spjd if (hdr->b_state == arc_anon) { 1572168404Spjd arc_buf_free(buf, tag); 1573168404Spjd return (no_callback); 1574168404Spjd } 1575168404Spjd 1576168404Spjd mutex_enter(hash_lock); 1577168404Spjd ASSERT(hdr->b_state != arc_anon); 1578168404Spjd ASSERT(buf->b_data != NULL); 1579168404Spjd 1580168404Spjd (void) remove_reference(hdr, hash_lock, tag); 1581168404Spjd if (hdr->b_datacnt > 1) { 1582168404Spjd if (no_callback) 1583168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1584168404Spjd } else if (no_callback) { 1585168404Spjd ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1586168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 1587168404Spjd } 1588168404Spjd ASSERT(no_callback || hdr->b_datacnt > 1 || 1589168404Spjd refcount_is_zero(&hdr->b_refcnt)); 1590168404Spjd mutex_exit(hash_lock); 1591168404Spjd return (no_callback); 1592168404Spjd} 1593168404Spjd 1594168404Spjdint 1595168404Spjdarc_buf_size(arc_buf_t *buf) 1596168404Spjd{ 1597168404Spjd return (buf->b_hdr->b_size); 1598168404Spjd} 1599168404Spjd 1600168404Spjd/* 1601168404Spjd * Evict buffers from list until we've removed the specified number of 1602168404Spjd * bytes. Move the removed buffers to the appropriate evict state. 1603168404Spjd * If the recycle flag is set, then attempt to "recycle" a buffer: 1604168404Spjd * - look for a buffer to evict that is `bytes' long. 1605168404Spjd * - return the data block from this buffer rather than freeing it. 1606168404Spjd * This flag is used by callers that are trying to make space for a 1607168404Spjd * new buffer in a full arc cache. 1608185029Spjd * 1609185029Spjd * This function makes a "best effort". It skips over any buffers 1610185029Spjd * it can't get a hash_lock on, and so may not catch all candidates. 1611185029Spjd * It may also return without evicting as much space as requested. 1612168404Spjd */ 1613168404Spjdstatic void * 1614185029Spjdarc_evict(arc_state_t *state, spa_t *spa, int64_t bytes, boolean_t recycle, 1615168404Spjd arc_buf_contents_t type) 1616168404Spjd{ 1617168404Spjd arc_state_t *evicted_state; 1618168404Spjd uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 1619205231Skmacy int64_t bytes_remaining; 1620168404Spjd arc_buf_hdr_t *ab, *ab_prev = NULL; 1621205231Skmacy list_t *evicted_list, *list, *evicted_list_start, *list_start; 1622205231Skmacy kmutex_t *lock, *evicted_lock; 1623168404Spjd kmutex_t *hash_lock; 1624168404Spjd boolean_t have_lock; 1625168404Spjd void *stolen = NULL; 1626205231Skmacy static int evict_metadata_offset, evict_data_offset; 1627205231Skmacy int i, idx, offset, list_count, count; 1628168404Spjd 1629168404Spjd ASSERT(state == arc_mru || state == arc_mfu); 1630168404Spjd 1631168404Spjd evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 1632205231Skmacy 1633205231Skmacy if (type == ARC_BUFC_METADATA) { 1634205231Skmacy offset = 0; 1635205231Skmacy list_count = ARC_BUFC_NUMMETADATALISTS; 1636205231Skmacy list_start = &state->arcs_lists[0]; 1637205231Skmacy evicted_list_start = &evicted_state->arcs_lists[0]; 1638205231Skmacy idx = evict_metadata_offset; 1639205231Skmacy } else { 1640205231Skmacy offset = ARC_BUFC_NUMMETADATALISTS; 1641168404Spjd 1642205231Skmacy list_start = &state->arcs_lists[offset]; 1643205231Skmacy evicted_list_start = &evicted_state->arcs_lists[offset]; 1644205231Skmacy list_count = ARC_BUFC_NUMDATALISTS; 1645205231Skmacy idx = evict_data_offset; 1646205231Skmacy } 1647205231Skmacy bytes_remaining = evicted_state->arcs_lsize[type]; 1648205231Skmacy count = 0; 1649205231Skmacy 1650205231Skmacyevict_start: 1651205231Skmacy list = &list_start[idx]; 1652205231Skmacy evicted_list = &evicted_list_start[idx]; 1653205231Skmacy lock = ARCS_LOCK(state, (offset + idx)); 1654205231Skmacy evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 1655168404Spjd 1656205231Skmacy mutex_enter(lock); 1657205231Skmacy mutex_enter(evicted_lock); 1658205231Skmacy 1659185029Spjd for (ab = list_tail(list); ab; ab = ab_prev) { 1660185029Spjd ab_prev = list_prev(list, ab); 1661205231Skmacy bytes_remaining -= (ab->b_size * ab->b_datacnt); 1662168404Spjd /* prefetch buffers have a minimum lifespan */ 1663168404Spjd if (HDR_IO_IN_PROGRESS(ab) || 1664185029Spjd (spa && ab->b_spa != spa) || 1665168404Spjd (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 1666174049Sjb LBOLT - ab->b_arc_access < arc_min_prefetch_lifespan)) { 1667168404Spjd skipped++; 1668168404Spjd continue; 1669168404Spjd } 1670168404Spjd /* "lookahead" for better eviction candidate */ 1671168404Spjd if (recycle && ab->b_size != bytes && 1672168404Spjd ab_prev && ab_prev->b_size == bytes) 1673168404Spjd continue; 1674168404Spjd hash_lock = HDR_LOCK(ab); 1675168404Spjd have_lock = MUTEX_HELD(hash_lock); 1676168404Spjd if (have_lock || mutex_tryenter(hash_lock)) { 1677168404Spjd ASSERT3U(refcount_count(&ab->b_refcnt), ==, 0); 1678168404Spjd ASSERT(ab->b_datacnt > 0); 1679168404Spjd while (ab->b_buf) { 1680168404Spjd arc_buf_t *buf = ab->b_buf; 1681185029Spjd if (!rw_tryenter(&buf->b_lock, RW_WRITER)) { 1682185029Spjd missed += 1; 1683185029Spjd break; 1684185029Spjd } 1685168404Spjd if (buf->b_data) { 1686168404Spjd bytes_evicted += ab->b_size; 1687168404Spjd if (recycle && ab->b_type == type && 1688185029Spjd ab->b_size == bytes && 1689185029Spjd !HDR_L2_WRITING(ab)) { 1690168404Spjd stolen = buf->b_data; 1691168404Spjd recycle = FALSE; 1692168404Spjd } 1693168404Spjd } 1694168404Spjd if (buf->b_efunc) { 1695168404Spjd mutex_enter(&arc_eviction_mtx); 1696168404Spjd arc_buf_destroy(buf, 1697168404Spjd buf->b_data == stolen, FALSE); 1698168404Spjd ab->b_buf = buf->b_next; 1699168404Spjd buf->b_hdr = &arc_eviction_hdr; 1700168404Spjd buf->b_next = arc_eviction_list; 1701168404Spjd arc_eviction_list = buf; 1702168404Spjd mutex_exit(&arc_eviction_mtx); 1703185029Spjd rw_exit(&buf->b_lock); 1704168404Spjd } else { 1705185029Spjd rw_exit(&buf->b_lock); 1706168404Spjd arc_buf_destroy(buf, 1707168404Spjd buf->b_data == stolen, TRUE); 1708168404Spjd } 1709168404Spjd } 1710185029Spjd if (ab->b_datacnt == 0) { 1711185029Spjd arc_change_state(evicted_state, ab, hash_lock); 1712185029Spjd ASSERT(HDR_IN_HASH_TABLE(ab)); 1713185029Spjd ab->b_flags |= ARC_IN_HASH_TABLE; 1714185029Spjd ab->b_flags &= ~ARC_BUF_AVAILABLE; 1715185029Spjd DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 1716185029Spjd } 1717168404Spjd if (!have_lock) 1718168404Spjd mutex_exit(hash_lock); 1719168404Spjd if (bytes >= 0 && bytes_evicted >= bytes) 1720168404Spjd break; 1721205231Skmacy if (bytes_remaining > 0) { 1722205231Skmacy mutex_exit(evicted_lock); 1723205231Skmacy mutex_exit(lock); 1724205231Skmacy idx = ((idx + 1)&(list_count-1)); 1725205231Skmacy count++; 1726205231Skmacy goto evict_start; 1727205231Skmacy } 1728168404Spjd } else { 1729168404Spjd missed += 1; 1730168404Spjd } 1731168404Spjd } 1732168404Spjd 1733205231Skmacy mutex_exit(evicted_lock); 1734205231Skmacy mutex_exit(lock); 1735205231Skmacy 1736205231Skmacy idx = ((idx + 1)&(list_count-1)); 1737205231Skmacy count++; 1738168404Spjd 1739205231Skmacy if (bytes_evicted < bytes) { 1740205231Skmacy if (count < list_count) 1741205231Skmacy goto evict_start; 1742205231Skmacy else 1743205231Skmacy dprintf("only evicted %lld bytes from %x", 1744205231Skmacy (longlong_t)bytes_evicted, state); 1745205231Skmacy } 1746205231Skmacy if (type == ARC_BUFC_METADATA) 1747205231Skmacy evict_metadata_offset = idx; 1748205231Skmacy else 1749205231Skmacy evict_data_offset = idx; 1750205231Skmacy 1751168404Spjd if (skipped) 1752168404Spjd ARCSTAT_INCR(arcstat_evict_skip, skipped); 1753168404Spjd 1754168404Spjd if (missed) 1755168404Spjd ARCSTAT_INCR(arcstat_mutex_miss, missed); 1756168404Spjd 1757185029Spjd /* 1758185029Spjd * We have just evicted some date into the ghost state, make 1759185029Spjd * sure we also adjust the ghost state size if necessary. 1760185029Spjd */ 1761185029Spjd if (arc_no_grow && 1762185029Spjd arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) { 1763185029Spjd int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size + 1764185029Spjd arc_mru_ghost->arcs_size - arc_c; 1765185029Spjd 1766185029Spjd if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) { 1767185029Spjd int64_t todelete = 1768185029Spjd MIN(arc_mru_ghost->arcs_lsize[type], mru_over); 1769185029Spjd arc_evict_ghost(arc_mru_ghost, NULL, todelete); 1770185029Spjd } else if (arc_mfu_ghost->arcs_lsize[type] > 0) { 1771185029Spjd int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type], 1772185029Spjd arc_mru_ghost->arcs_size + 1773185029Spjd arc_mfu_ghost->arcs_size - arc_c); 1774185029Spjd arc_evict_ghost(arc_mfu_ghost, NULL, todelete); 1775185029Spjd } 1776185029Spjd } 1777205231Skmacy if (stolen) 1778205231Skmacy ARCSTAT_BUMP(arcstat_stolen); 1779185029Spjd 1780168404Spjd return (stolen); 1781168404Spjd} 1782168404Spjd 1783168404Spjd/* 1784168404Spjd * Remove buffers from list until we've removed the specified number of 1785168404Spjd * bytes. Destroy the buffers that are removed. 1786168404Spjd */ 1787168404Spjdstatic void 1788185029Spjdarc_evict_ghost(arc_state_t *state, spa_t *spa, int64_t bytes) 1789168404Spjd{ 1790168404Spjd arc_buf_hdr_t *ab, *ab_prev; 1791205231Skmacy list_t *list, *list_start; 1792205231Skmacy kmutex_t *hash_lock, *lock; 1793168404Spjd uint64_t bytes_deleted = 0; 1794168404Spjd uint64_t bufs_skipped = 0; 1795205231Skmacy static int evict_offset; 1796205231Skmacy int list_count, idx = evict_offset; 1797205231Skmacy int offset, count = 0; 1798168404Spjd 1799168404Spjd ASSERT(GHOST_STATE(state)); 1800205231Skmacy 1801205231Skmacy /* 1802205231Skmacy * data lists come after metadata lists 1803205231Skmacy */ 1804205231Skmacy list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 1805205231Skmacy list_count = ARC_BUFC_NUMDATALISTS; 1806205231Skmacy offset = ARC_BUFC_NUMMETADATALISTS; 1807205231Skmacy 1808205231Skmacyevict_start: 1809205231Skmacy list = &list_start[idx]; 1810205231Skmacy lock = ARCS_LOCK(state, idx + offset); 1811205231Skmacy 1812205231Skmacy mutex_enter(lock); 1813185029Spjd for (ab = list_tail(list); ab; ab = ab_prev) { 1814185029Spjd ab_prev = list_prev(list, ab); 1815185029Spjd if (spa && ab->b_spa != spa) 1816185029Spjd continue; 1817168404Spjd hash_lock = HDR_LOCK(ab); 1818168404Spjd if (mutex_tryenter(hash_lock)) { 1819168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(ab)); 1820168404Spjd ASSERT(ab->b_buf == NULL); 1821168404Spjd ARCSTAT_BUMP(arcstat_deleted); 1822168404Spjd bytes_deleted += ab->b_size; 1823185029Spjd 1824185029Spjd if (ab->b_l2hdr != NULL) { 1825185029Spjd /* 1826185029Spjd * This buffer is cached on the 2nd Level ARC; 1827185029Spjd * don't destroy the header. 1828185029Spjd */ 1829185029Spjd arc_change_state(arc_l2c_only, ab, hash_lock); 1830185029Spjd mutex_exit(hash_lock); 1831185029Spjd } else { 1832185029Spjd arc_change_state(arc_anon, ab, hash_lock); 1833185029Spjd mutex_exit(hash_lock); 1834185029Spjd arc_hdr_destroy(ab); 1835185029Spjd } 1836185029Spjd 1837168404Spjd DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 1838168404Spjd if (bytes >= 0 && bytes_deleted >= bytes) 1839168404Spjd break; 1840168404Spjd } else { 1841168404Spjd if (bytes < 0) { 1842205231Skmacy /* 1843205231Skmacy * we're draining the ARC, retry 1844205231Skmacy */ 1845205231Skmacy mutex_exit(lock); 1846168404Spjd mutex_enter(hash_lock); 1847168404Spjd mutex_exit(hash_lock); 1848205231Skmacy goto evict_start; 1849168404Spjd } 1850168404Spjd bufs_skipped += 1; 1851168404Spjd } 1852168404Spjd } 1853205231Skmacy mutex_exit(lock); 1854205231Skmacy idx = ((idx + 1)&(ARC_BUFC_NUMDATALISTS-1)); 1855205231Skmacy count++; 1856205231Skmacy 1857205231Skmacy if (count < list_count) 1858205231Skmacy goto evict_start; 1859205231Skmacy 1860205231Skmacy evict_offset = idx; 1861205231Skmacy if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 1862185029Spjd (bytes < 0 || bytes_deleted < bytes)) { 1863205231Skmacy list_start = &state->arcs_lists[0]; 1864205231Skmacy list_count = ARC_BUFC_NUMMETADATALISTS; 1865205231Skmacy offset = count = 0; 1866205231Skmacy goto evict_start; 1867185029Spjd } 1868185029Spjd 1869168404Spjd if (bufs_skipped) { 1870168404Spjd ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 1871168404Spjd ASSERT(bytes >= 0); 1872168404Spjd } 1873168404Spjd 1874168404Spjd if (bytes_deleted < bytes) 1875168404Spjd dprintf("only deleted %lld bytes from %p", 1876168404Spjd (longlong_t)bytes_deleted, state); 1877168404Spjd} 1878168404Spjd 1879168404Spjdstatic void 1880168404Spjdarc_adjust(void) 1881168404Spjd{ 1882168404Spjd int64_t top_sz, mru_over, arc_over, todelete; 1883168404Spjd 1884185029Spjd top_sz = arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used; 1885168404Spjd 1886185029Spjd if (top_sz > arc_p && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 1887185029Spjd int64_t toevict = 1888185029Spjd MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], top_sz - arc_p); 1889185029Spjd (void) arc_evict(arc_mru, NULL, toevict, FALSE, ARC_BUFC_DATA); 1890168404Spjd top_sz = arc_anon->arcs_size + arc_mru->arcs_size; 1891168404Spjd } 1892168404Spjd 1893185029Spjd if (top_sz > arc_p && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 1894185029Spjd int64_t toevict = 1895185029Spjd MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], top_sz - arc_p); 1896185029Spjd (void) arc_evict(arc_mru, NULL, toevict, FALSE, 1897185029Spjd ARC_BUFC_METADATA); 1898185029Spjd top_sz = arc_anon->arcs_size + arc_mru->arcs_size; 1899185029Spjd } 1900185029Spjd 1901168404Spjd mru_over = top_sz + arc_mru_ghost->arcs_size - arc_c; 1902168404Spjd 1903168404Spjd if (mru_over > 0) { 1904185029Spjd if (arc_mru_ghost->arcs_size > 0) { 1905185029Spjd todelete = MIN(arc_mru_ghost->arcs_size, mru_over); 1906185029Spjd arc_evict_ghost(arc_mru_ghost, NULL, todelete); 1907168404Spjd } 1908168404Spjd } 1909168404Spjd 1910168404Spjd if ((arc_over = arc_size - arc_c) > 0) { 1911168404Spjd int64_t tbl_over; 1912168404Spjd 1913185029Spjd if (arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 1914185029Spjd int64_t toevict = 1915185029Spjd MIN(arc_mfu->arcs_lsize[ARC_BUFC_DATA], arc_over); 1916185029Spjd (void) arc_evict(arc_mfu, NULL, toevict, FALSE, 1917185029Spjd ARC_BUFC_DATA); 1918185029Spjd arc_over = arc_size - arc_c; 1919168404Spjd } 1920168404Spjd 1921185029Spjd if (arc_over > 0 && 1922185029Spjd arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 1923185029Spjd int64_t toevict = 1924185029Spjd MIN(arc_mfu->arcs_lsize[ARC_BUFC_METADATA], 1925185029Spjd arc_over); 1926185029Spjd (void) arc_evict(arc_mfu, NULL, toevict, FALSE, 1927185029Spjd ARC_BUFC_METADATA); 1928185029Spjd } 1929168404Spjd 1930185029Spjd tbl_over = arc_size + arc_mru_ghost->arcs_size + 1931185029Spjd arc_mfu_ghost->arcs_size - arc_c * 2; 1932185029Spjd 1933185029Spjd if (tbl_over > 0 && arc_mfu_ghost->arcs_size > 0) { 1934185029Spjd todelete = MIN(arc_mfu_ghost->arcs_size, tbl_over); 1935185029Spjd arc_evict_ghost(arc_mfu_ghost, NULL, todelete); 1936168404Spjd } 1937168404Spjd } 1938168404Spjd} 1939168404Spjd 1940168404Spjdstatic void 1941168404Spjdarc_do_user_evicts(void) 1942168404Spjd{ 1943191903Skmacy static arc_buf_t *tmp_arc_eviction_list; 1944191903Skmacy 1945191903Skmacy /* 1946191903Skmacy * Move list over to avoid LOR 1947191903Skmacy */ 1948191903Skmacyrestart: 1949168404Spjd mutex_enter(&arc_eviction_mtx); 1950191903Skmacy tmp_arc_eviction_list = arc_eviction_list; 1951191903Skmacy arc_eviction_list = NULL; 1952191903Skmacy mutex_exit(&arc_eviction_mtx); 1953191903Skmacy 1954191903Skmacy while (tmp_arc_eviction_list != NULL) { 1955191903Skmacy arc_buf_t *buf = tmp_arc_eviction_list; 1956191903Skmacy tmp_arc_eviction_list = buf->b_next; 1957185029Spjd rw_enter(&buf->b_lock, RW_WRITER); 1958168404Spjd buf->b_hdr = NULL; 1959185029Spjd rw_exit(&buf->b_lock); 1960168404Spjd 1961168404Spjd if (buf->b_efunc != NULL) 1962168404Spjd VERIFY(buf->b_efunc(buf) == 0); 1963168404Spjd 1964168404Spjd buf->b_efunc = NULL; 1965168404Spjd buf->b_private = NULL; 1966168404Spjd kmem_cache_free(buf_cache, buf); 1967168404Spjd } 1968191903Skmacy 1969191903Skmacy if (arc_eviction_list != NULL) 1970191903Skmacy goto restart; 1971168404Spjd} 1972168404Spjd 1973168404Spjd/* 1974185029Spjd * Flush all *evictable* data from the cache for the given spa. 1975168404Spjd * NOTE: this will not touch "active" (i.e. referenced) data. 1976168404Spjd */ 1977168404Spjdvoid 1978185029Spjdarc_flush(spa_t *spa) 1979168404Spjd{ 1980205231Skmacy while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 1981185029Spjd (void) arc_evict(arc_mru, spa, -1, FALSE, ARC_BUFC_DATA); 1982185029Spjd if (spa) 1983185029Spjd break; 1984185029Spjd } 1985205231Skmacy while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 1986185029Spjd (void) arc_evict(arc_mru, spa, -1, FALSE, ARC_BUFC_METADATA); 1987185029Spjd if (spa) 1988185029Spjd break; 1989185029Spjd } 1990205231Skmacy while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 1991185029Spjd (void) arc_evict(arc_mfu, spa, -1, FALSE, ARC_BUFC_DATA); 1992185029Spjd if (spa) 1993185029Spjd break; 1994185029Spjd } 1995205231Skmacy while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 1996185029Spjd (void) arc_evict(arc_mfu, spa, -1, FALSE, ARC_BUFC_METADATA); 1997185029Spjd if (spa) 1998185029Spjd break; 1999185029Spjd } 2000168404Spjd 2001185029Spjd arc_evict_ghost(arc_mru_ghost, spa, -1); 2002185029Spjd arc_evict_ghost(arc_mfu_ghost, spa, -1); 2003168404Spjd 2004168404Spjd mutex_enter(&arc_reclaim_thr_lock); 2005168404Spjd arc_do_user_evicts(); 2006168404Spjd mutex_exit(&arc_reclaim_thr_lock); 2007185029Spjd ASSERT(spa || arc_eviction_list == NULL); 2008168404Spjd} 2009168404Spjd 2010168404Spjdint arc_shrink_shift = 5; /* log2(fraction of arc to reclaim) */ 2011168404Spjd 2012168404Spjdvoid 2013168404Spjdarc_shrink(void) 2014168404Spjd{ 2015168404Spjd if (arc_c > arc_c_min) { 2016168404Spjd uint64_t to_free; 2017168404Spjd 2018168404Spjd#ifdef _KERNEL 2019168404Spjd to_free = arc_c >> arc_shrink_shift; 2020168404Spjd#else 2021168404Spjd to_free = arc_c >> arc_shrink_shift; 2022168404Spjd#endif 2023168404Spjd if (arc_c > arc_c_min + to_free) 2024168404Spjd atomic_add_64(&arc_c, -to_free); 2025168404Spjd else 2026168404Spjd arc_c = arc_c_min; 2027168404Spjd 2028168404Spjd atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2029168404Spjd if (arc_c > arc_size) 2030168404Spjd arc_c = MAX(arc_size, arc_c_min); 2031168404Spjd if (arc_p > arc_c) 2032168404Spjd arc_p = (arc_c >> 1); 2033168404Spjd ASSERT(arc_c >= arc_c_min); 2034168404Spjd ASSERT((int64_t)arc_p >= 0); 2035168404Spjd } 2036168404Spjd 2037168404Spjd if (arc_size > arc_c) 2038168404Spjd arc_adjust(); 2039168404Spjd} 2040168404Spjd 2041185029Spjdstatic int needfree = 0; 2042168404Spjd 2043168404Spjdstatic int 2044168404Spjdarc_reclaim_needed(void) 2045168404Spjd{ 2046168404Spjd#if 0 2047168404Spjd uint64_t extra; 2048168404Spjd#endif 2049168404Spjd 2050168404Spjd#ifdef _KERNEL 2051197816Skmacy if (needfree) 2052197816Skmacy return (1); 2053197816Skmacy if (arc_size > arc_c_max) 2054197816Skmacy return (1); 2055197816Skmacy if (arc_size <= arc_c_min) 2056197816Skmacy return (0); 2057168404Spjd 2058191902Skmacy /* 2059191902Skmacy * If pages are needed or we're within 2048 pages 2060191902Skmacy * of needing to page need to reclaim 2061191902Skmacy */ 2062191902Skmacy if (vm_pages_needed || (vm_paging_target() > -2048)) 2063191902Skmacy return (1); 2064191902Skmacy 2065168404Spjd#if 0 2066168404Spjd /* 2067185029Spjd * take 'desfree' extra pages, so we reclaim sooner, rather than later 2068185029Spjd */ 2069185029Spjd extra = desfree; 2070185029Spjd 2071185029Spjd /* 2072185029Spjd * check that we're out of range of the pageout scanner. It starts to 2073185029Spjd * schedule paging if freemem is less than lotsfree and needfree. 2074185029Spjd * lotsfree is the high-water mark for pageout, and needfree is the 2075185029Spjd * number of needed free pages. We add extra pages here to make sure 2076185029Spjd * the scanner doesn't start up while we're freeing memory. 2077185029Spjd */ 2078185029Spjd if (freemem < lotsfree + needfree + extra) 2079185029Spjd return (1); 2080185029Spjd 2081185029Spjd /* 2082168404Spjd * check to make sure that swapfs has enough space so that anon 2083185029Spjd * reservations can still succeed. anon_resvmem() checks that the 2084168404Spjd * availrmem is greater than swapfs_minfree, and the number of reserved 2085168404Spjd * swap pages. We also add a bit of extra here just to prevent 2086168404Spjd * circumstances from getting really dire. 2087168404Spjd */ 2088168404Spjd if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2089168404Spjd return (1); 2090168404Spjd 2091168404Spjd#if defined(__i386) 2092168404Spjd /* 2093168404Spjd * If we're on an i386 platform, it's possible that we'll exhaust the 2094168404Spjd * kernel heap space before we ever run out of available physical 2095168404Spjd * memory. Most checks of the size of the heap_area compare against 2096168404Spjd * tune.t_minarmem, which is the minimum available real memory that we 2097168404Spjd * can have in the system. However, this is generally fixed at 25 pages 2098168404Spjd * which is so low that it's useless. In this comparison, we seek to 2099168404Spjd * calculate the total heap-size, and reclaim if more than 3/4ths of the 2100185029Spjd * heap is allocated. (Or, in the calculation, if less than 1/4th is 2101168404Spjd * free) 2102168404Spjd */ 2103168404Spjd if (btop(vmem_size(heap_arena, VMEM_FREE)) < 2104168404Spjd (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2)) 2105168404Spjd return (1); 2106168404Spjd#endif 2107168404Spjd#else 2108175633Spjd if (kmem_used() > (kmem_size() * 3) / 4) 2109168404Spjd return (1); 2110168404Spjd#endif 2111168404Spjd 2112168404Spjd#else 2113168404Spjd if (spa_get_random(100) == 0) 2114168404Spjd return (1); 2115168404Spjd#endif 2116168404Spjd return (0); 2117168404Spjd} 2118168404Spjd 2119168404Spjdstatic void 2120168404Spjdarc_kmem_reap_now(arc_reclaim_strategy_t strat) 2121168404Spjd{ 2122168404Spjd#ifdef ZIO_USE_UMA 2123168404Spjd size_t i; 2124168404Spjd kmem_cache_t *prev_cache = NULL; 2125168404Spjd kmem_cache_t *prev_data_cache = NULL; 2126168404Spjd#endif 2127168404Spjd 2128168404Spjd#ifdef _KERNEL 2129185029Spjd if (arc_meta_used >= arc_meta_limit) { 2130185029Spjd /* 2131185029Spjd * We are exceeding our meta-data cache limit. 2132185029Spjd * Purge some DNLC entries to release holds on meta-data. 2133185029Spjd */ 2134185029Spjd dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2135185029Spjd } 2136168404Spjd#if defined(__i386) 2137168404Spjd /* 2138168404Spjd * Reclaim unused memory from all kmem caches. 2139168404Spjd */ 2140168404Spjd kmem_reap(); 2141168404Spjd#endif 2142168404Spjd#endif 2143168404Spjd 2144168404Spjd /* 2145185029Spjd * An aggressive reclamation will shrink the cache size as well as 2146168404Spjd * reap free buffers from the arc kmem caches. 2147168404Spjd */ 2148168404Spjd if (strat == ARC_RECLAIM_AGGR) 2149168404Spjd arc_shrink(); 2150168404Spjd 2151168404Spjd#ifdef ZIO_USE_UMA 2152168404Spjd for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2153168404Spjd if (zio_buf_cache[i] != prev_cache) { 2154168404Spjd prev_cache = zio_buf_cache[i]; 2155168404Spjd kmem_cache_reap_now(zio_buf_cache[i]); 2156168404Spjd } 2157168404Spjd if (zio_data_buf_cache[i] != prev_data_cache) { 2158168404Spjd prev_data_cache = zio_data_buf_cache[i]; 2159168404Spjd kmem_cache_reap_now(zio_data_buf_cache[i]); 2160168404Spjd } 2161168404Spjd } 2162168404Spjd#endif 2163168404Spjd kmem_cache_reap_now(buf_cache); 2164168404Spjd kmem_cache_reap_now(hdr_cache); 2165168404Spjd} 2166168404Spjd 2167168404Spjdstatic void 2168168404Spjdarc_reclaim_thread(void *dummy __unused) 2169168404Spjd{ 2170168404Spjd clock_t growtime = 0; 2171168404Spjd arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2172168404Spjd callb_cpr_t cpr; 2173168404Spjd 2174168404Spjd CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2175168404Spjd 2176168404Spjd mutex_enter(&arc_reclaim_thr_lock); 2177168404Spjd while (arc_thread_exit == 0) { 2178168404Spjd if (arc_reclaim_needed()) { 2179168404Spjd 2180168404Spjd if (arc_no_grow) { 2181168404Spjd if (last_reclaim == ARC_RECLAIM_CONS) { 2182168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2183168404Spjd } else { 2184168404Spjd last_reclaim = ARC_RECLAIM_CONS; 2185168404Spjd } 2186168404Spjd } else { 2187168404Spjd arc_no_grow = TRUE; 2188168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2189168404Spjd membar_producer(); 2190168404Spjd } 2191168404Spjd 2192168404Spjd /* reset the growth delay for every reclaim */ 2193174049Sjb growtime = LBOLT + (arc_grow_retry * hz); 2194168404Spjd 2195185029Spjd if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2196168404Spjd /* 2197185029Spjd * If needfree is TRUE our vm_lowmem hook 2198168404Spjd * was called and in that case we must free some 2199168404Spjd * memory, so switch to aggressive mode. 2200168404Spjd */ 2201168404Spjd arc_no_grow = TRUE; 2202168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2203168404Spjd } 2204168404Spjd arc_kmem_reap_now(last_reclaim); 2205185029Spjd arc_warm = B_TRUE; 2206185029Spjd 2207185029Spjd } else if (arc_no_grow && LBOLT >= growtime) { 2208168404Spjd arc_no_grow = FALSE; 2209168404Spjd } 2210168404Spjd 2211185029Spjd if (needfree || 2212168404Spjd (2 * arc_c < arc_size + 2213168404Spjd arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size)) 2214168404Spjd arc_adjust(); 2215168404Spjd 2216168404Spjd if (arc_eviction_list != NULL) 2217168404Spjd arc_do_user_evicts(); 2218168404Spjd 2219168404Spjd if (arc_reclaim_needed()) { 2220185029Spjd needfree = 0; 2221168404Spjd#ifdef _KERNEL 2222185029Spjd wakeup(&needfree); 2223168404Spjd#endif 2224168404Spjd } 2225168404Spjd 2226168404Spjd /* block until needed, or one second, whichever is shorter */ 2227168404Spjd CALLB_CPR_SAFE_BEGIN(&cpr); 2228168404Spjd (void) cv_timedwait(&arc_reclaim_thr_cv, 2229168404Spjd &arc_reclaim_thr_lock, hz); 2230168404Spjd CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2231168404Spjd } 2232168404Spjd 2233168404Spjd arc_thread_exit = 0; 2234168404Spjd cv_broadcast(&arc_reclaim_thr_cv); 2235168404Spjd CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2236168404Spjd thread_exit(); 2237168404Spjd} 2238168404Spjd 2239168404Spjd/* 2240168404Spjd * Adapt arc info given the number of bytes we are trying to add and 2241168404Spjd * the state that we are comming from. This function is only called 2242168404Spjd * when we are adding new content to the cache. 2243168404Spjd */ 2244168404Spjdstatic void 2245168404Spjdarc_adapt(int bytes, arc_state_t *state) 2246168404Spjd{ 2247168404Spjd int mult; 2248168404Spjd 2249185029Spjd if (state == arc_l2c_only) 2250185029Spjd return; 2251185029Spjd 2252168404Spjd ASSERT(bytes > 0); 2253168404Spjd /* 2254168404Spjd * Adapt the target size of the MRU list: 2255168404Spjd * - if we just hit in the MRU ghost list, then increase 2256168404Spjd * the target size of the MRU list. 2257168404Spjd * - if we just hit in the MFU ghost list, then increase 2258168404Spjd * the target size of the MFU list by decreasing the 2259168404Spjd * target size of the MRU list. 2260168404Spjd */ 2261168404Spjd if (state == arc_mru_ghost) { 2262168404Spjd mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2263168404Spjd 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2264168404Spjd 2265168404Spjd arc_p = MIN(arc_c, arc_p + bytes * mult); 2266168404Spjd } else if (state == arc_mfu_ghost) { 2267168404Spjd mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2268168404Spjd 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2269168404Spjd 2270168404Spjd arc_p = MAX(0, (int64_t)arc_p - bytes * mult); 2271168404Spjd } 2272168404Spjd ASSERT((int64_t)arc_p >= 0); 2273168404Spjd 2274168404Spjd if (arc_reclaim_needed()) { 2275168404Spjd cv_signal(&arc_reclaim_thr_cv); 2276168404Spjd return; 2277168404Spjd } 2278168404Spjd 2279168404Spjd if (arc_no_grow) 2280168404Spjd return; 2281168404Spjd 2282168404Spjd if (arc_c >= arc_c_max) 2283168404Spjd return; 2284168404Spjd 2285168404Spjd /* 2286168404Spjd * If we're within (2 * maxblocksize) bytes of the target 2287168404Spjd * cache size, increment the target cache size 2288168404Spjd */ 2289168404Spjd if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2290168404Spjd atomic_add_64(&arc_c, (int64_t)bytes); 2291168404Spjd if (arc_c > arc_c_max) 2292168404Spjd arc_c = arc_c_max; 2293168404Spjd else if (state == arc_anon) 2294168404Spjd atomic_add_64(&arc_p, (int64_t)bytes); 2295168404Spjd if (arc_p > arc_c) 2296168404Spjd arc_p = arc_c; 2297168404Spjd } 2298168404Spjd ASSERT((int64_t)arc_p >= 0); 2299168404Spjd} 2300168404Spjd 2301168404Spjd/* 2302168404Spjd * Check if the cache has reached its limits and eviction is required 2303168404Spjd * prior to insert. 2304168404Spjd */ 2305168404Spjdstatic int 2306185029Spjdarc_evict_needed(arc_buf_contents_t type) 2307168404Spjd{ 2308185029Spjd if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2309185029Spjd return (1); 2310185029Spjd 2311185029Spjd#if 0 2312185029Spjd#ifdef _KERNEL 2313185029Spjd /* 2314185029Spjd * If zio data pages are being allocated out of a separate heap segment, 2315185029Spjd * then enforce that the size of available vmem for this area remains 2316185029Spjd * above about 1/32nd free. 2317185029Spjd */ 2318185029Spjd if (type == ARC_BUFC_DATA && zio_arena != NULL && 2319185029Spjd vmem_size(zio_arena, VMEM_FREE) < 2320185029Spjd (vmem_size(zio_arena, VMEM_ALLOC) >> 5)) 2321185029Spjd return (1); 2322185029Spjd#endif 2323185029Spjd#endif 2324185029Spjd 2325168404Spjd if (arc_reclaim_needed()) 2326168404Spjd return (1); 2327168404Spjd 2328168404Spjd return (arc_size > arc_c); 2329168404Spjd} 2330168404Spjd 2331168404Spjd/* 2332168404Spjd * The buffer, supplied as the first argument, needs a data block. 2333168404Spjd * So, if we are at cache max, determine which cache should be victimized. 2334168404Spjd * We have the following cases: 2335168404Spjd * 2336168404Spjd * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2337168404Spjd * In this situation if we're out of space, but the resident size of the MFU is 2338168404Spjd * under the limit, victimize the MFU cache to satisfy this insertion request. 2339168404Spjd * 2340168404Spjd * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2341168404Spjd * Here, we've used up all of the available space for the MRU, so we need to 2342168404Spjd * evict from our own cache instead. Evict from the set of resident MRU 2343168404Spjd * entries. 2344168404Spjd * 2345168404Spjd * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2346168404Spjd * c minus p represents the MFU space in the cache, since p is the size of the 2347168404Spjd * cache that is dedicated to the MRU. In this situation there's still space on 2348168404Spjd * the MFU side, so the MRU side needs to be victimized. 2349168404Spjd * 2350168404Spjd * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2351168404Spjd * MFU's resident set is consuming more space than it has been allotted. In 2352168404Spjd * this situation, we must victimize our own cache, the MFU, for this insertion. 2353168404Spjd */ 2354168404Spjdstatic void 2355168404Spjdarc_get_data_buf(arc_buf_t *buf) 2356168404Spjd{ 2357168404Spjd arc_state_t *state = buf->b_hdr->b_state; 2358168404Spjd uint64_t size = buf->b_hdr->b_size; 2359168404Spjd arc_buf_contents_t type = buf->b_hdr->b_type; 2360168404Spjd 2361168404Spjd arc_adapt(size, state); 2362168404Spjd 2363168404Spjd /* 2364168404Spjd * We have not yet reached cache maximum size, 2365168404Spjd * just allocate a new buffer. 2366168404Spjd */ 2367185029Spjd if (!arc_evict_needed(type)) { 2368168404Spjd if (type == ARC_BUFC_METADATA) { 2369168404Spjd buf->b_data = zio_buf_alloc(size); 2370185029Spjd arc_space_consume(size); 2371168404Spjd } else { 2372168404Spjd ASSERT(type == ARC_BUFC_DATA); 2373168404Spjd buf->b_data = zio_data_buf_alloc(size); 2374185029Spjd atomic_add_64(&arc_size, size); 2375168404Spjd } 2376168404Spjd goto out; 2377168404Spjd } 2378168404Spjd 2379168404Spjd /* 2380168404Spjd * If we are prefetching from the mfu ghost list, this buffer 2381168404Spjd * will end up on the mru list; so steal space from there. 2382168404Spjd */ 2383168404Spjd if (state == arc_mfu_ghost) 2384168404Spjd state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2385168404Spjd else if (state == arc_mru_ghost) 2386168404Spjd state = arc_mru; 2387168404Spjd 2388168404Spjd if (state == arc_mru || state == arc_anon) { 2389168404Spjd uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2390185029Spjd state = (arc_mfu->arcs_lsize[type] > 0 && 2391185029Spjd arc_p > mru_used) ? arc_mfu : arc_mru; 2392168404Spjd } else { 2393168404Spjd /* MFU cases */ 2394168404Spjd uint64_t mfu_space = arc_c - arc_p; 2395185029Spjd state = (arc_mru->arcs_lsize[type] > 0 && 2396185029Spjd mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2397168404Spjd } 2398185029Spjd if ((buf->b_data = arc_evict(state, NULL, size, TRUE, type)) == NULL) { 2399168404Spjd if (type == ARC_BUFC_METADATA) { 2400168404Spjd buf->b_data = zio_buf_alloc(size); 2401185029Spjd arc_space_consume(size); 2402168404Spjd } else { 2403168404Spjd ASSERT(type == ARC_BUFC_DATA); 2404168404Spjd buf->b_data = zio_data_buf_alloc(size); 2405185029Spjd atomic_add_64(&arc_size, size); 2406168404Spjd } 2407168404Spjd ARCSTAT_BUMP(arcstat_recycle_miss); 2408168404Spjd } 2409168404Spjd ASSERT(buf->b_data != NULL); 2410168404Spjdout: 2411168404Spjd /* 2412168404Spjd * Update the state size. Note that ghost states have a 2413168404Spjd * "ghost size" and so don't need to be updated. 2414168404Spjd */ 2415168404Spjd if (!GHOST_STATE(buf->b_hdr->b_state)) { 2416168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 2417168404Spjd 2418168404Spjd atomic_add_64(&hdr->b_state->arcs_size, size); 2419168404Spjd if (list_link_active(&hdr->b_arc_node)) { 2420168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2421185029Spjd atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2422168404Spjd } 2423168404Spjd /* 2424168404Spjd * If we are growing the cache, and we are adding anonymous 2425168404Spjd * data, and we have outgrown arc_p, update arc_p 2426168404Spjd */ 2427168404Spjd if (arc_size < arc_c && hdr->b_state == arc_anon && 2428168404Spjd arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2429168404Spjd arc_p = MIN(arc_c, arc_p + size); 2430168404Spjd } 2431205231Skmacy ARCSTAT_BUMP(arcstat_allocated); 2432168404Spjd} 2433168404Spjd 2434168404Spjd/* 2435168404Spjd * This routine is called whenever a buffer is accessed. 2436168404Spjd * NOTE: the hash lock is dropped in this function. 2437168404Spjd */ 2438168404Spjdstatic void 2439168404Spjdarc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2440168404Spjd{ 2441168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 2442168404Spjd 2443168404Spjd if (buf->b_state == arc_anon) { 2444168404Spjd /* 2445168404Spjd * This buffer is not in the cache, and does not 2446168404Spjd * appear in our "ghost" list. Add the new buffer 2447168404Spjd * to the MRU state. 2448168404Spjd */ 2449168404Spjd 2450168404Spjd ASSERT(buf->b_arc_access == 0); 2451174049Sjb buf->b_arc_access = LBOLT; 2452168404Spjd DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2453168404Spjd arc_change_state(arc_mru, buf, hash_lock); 2454168404Spjd 2455168404Spjd } else if (buf->b_state == arc_mru) { 2456168404Spjd /* 2457168404Spjd * If this buffer is here because of a prefetch, then either: 2458168404Spjd * - clear the flag if this is a "referencing" read 2459168404Spjd * (any subsequent access will bump this into the MFU state). 2460168404Spjd * or 2461168404Spjd * - move the buffer to the head of the list if this is 2462168404Spjd * another prefetch (to make it less likely to be evicted). 2463168404Spjd */ 2464168404Spjd if ((buf->b_flags & ARC_PREFETCH) != 0) { 2465168404Spjd if (refcount_count(&buf->b_refcnt) == 0) { 2466168404Spjd ASSERT(list_link_active(&buf->b_arc_node)); 2467168404Spjd } else { 2468168404Spjd buf->b_flags &= ~ARC_PREFETCH; 2469168404Spjd ARCSTAT_BUMP(arcstat_mru_hits); 2470168404Spjd } 2471174049Sjb buf->b_arc_access = LBOLT; 2472168404Spjd return; 2473168404Spjd } 2474168404Spjd 2475168404Spjd /* 2476168404Spjd * This buffer has been "accessed" only once so far, 2477168404Spjd * but it is still in the cache. Move it to the MFU 2478168404Spjd * state. 2479168404Spjd */ 2480174049Sjb if (LBOLT > buf->b_arc_access + ARC_MINTIME) { 2481168404Spjd /* 2482168404Spjd * More than 125ms have passed since we 2483168404Spjd * instantiated this buffer. Move it to the 2484168404Spjd * most frequently used state. 2485168404Spjd */ 2486174049Sjb buf->b_arc_access = LBOLT; 2487168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2488168404Spjd arc_change_state(arc_mfu, buf, hash_lock); 2489168404Spjd } 2490168404Spjd ARCSTAT_BUMP(arcstat_mru_hits); 2491168404Spjd } else if (buf->b_state == arc_mru_ghost) { 2492168404Spjd arc_state_t *new_state; 2493168404Spjd /* 2494168404Spjd * This buffer has been "accessed" recently, but 2495168404Spjd * was evicted from the cache. Move it to the 2496168404Spjd * MFU state. 2497168404Spjd */ 2498168404Spjd 2499168404Spjd if (buf->b_flags & ARC_PREFETCH) { 2500168404Spjd new_state = arc_mru; 2501168404Spjd if (refcount_count(&buf->b_refcnt) > 0) 2502168404Spjd buf->b_flags &= ~ARC_PREFETCH; 2503168404Spjd DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2504168404Spjd } else { 2505168404Spjd new_state = arc_mfu; 2506168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2507168404Spjd } 2508168404Spjd 2509174049Sjb buf->b_arc_access = LBOLT; 2510168404Spjd arc_change_state(new_state, buf, hash_lock); 2511168404Spjd 2512168404Spjd ARCSTAT_BUMP(arcstat_mru_ghost_hits); 2513168404Spjd } else if (buf->b_state == arc_mfu) { 2514168404Spjd /* 2515168404Spjd * This buffer has been accessed more than once and is 2516168404Spjd * still in the cache. Keep it in the MFU state. 2517168404Spjd * 2518168404Spjd * NOTE: an add_reference() that occurred when we did 2519168404Spjd * the arc_read() will have kicked this off the list. 2520168404Spjd * If it was a prefetch, we will explicitly move it to 2521168404Spjd * the head of the list now. 2522168404Spjd */ 2523168404Spjd if ((buf->b_flags & ARC_PREFETCH) != 0) { 2524168404Spjd ASSERT(refcount_count(&buf->b_refcnt) == 0); 2525168404Spjd ASSERT(list_link_active(&buf->b_arc_node)); 2526168404Spjd } 2527168404Spjd ARCSTAT_BUMP(arcstat_mfu_hits); 2528174049Sjb buf->b_arc_access = LBOLT; 2529168404Spjd } else if (buf->b_state == arc_mfu_ghost) { 2530168404Spjd arc_state_t *new_state = arc_mfu; 2531168404Spjd /* 2532168404Spjd * This buffer has been accessed more than once but has 2533168404Spjd * been evicted from the cache. Move it back to the 2534168404Spjd * MFU state. 2535168404Spjd */ 2536168404Spjd 2537168404Spjd if (buf->b_flags & ARC_PREFETCH) { 2538168404Spjd /* 2539168404Spjd * This is a prefetch access... 2540168404Spjd * move this block back to the MRU state. 2541168404Spjd */ 2542168404Spjd ASSERT3U(refcount_count(&buf->b_refcnt), ==, 0); 2543168404Spjd new_state = arc_mru; 2544168404Spjd } 2545168404Spjd 2546174049Sjb buf->b_arc_access = LBOLT; 2547168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2548168404Spjd arc_change_state(new_state, buf, hash_lock); 2549168404Spjd 2550168404Spjd ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 2551185029Spjd } else if (buf->b_state == arc_l2c_only) { 2552185029Spjd /* 2553185029Spjd * This buffer is on the 2nd Level ARC. 2554185029Spjd */ 2555185029Spjd 2556185029Spjd buf->b_arc_access = LBOLT; 2557185029Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2558185029Spjd arc_change_state(arc_mfu, buf, hash_lock); 2559168404Spjd } else { 2560168404Spjd ASSERT(!"invalid arc state"); 2561168404Spjd } 2562168404Spjd} 2563168404Spjd 2564168404Spjd/* a generic arc_done_func_t which you can use */ 2565168404Spjd/* ARGSUSED */ 2566168404Spjdvoid 2567168404Spjdarc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 2568168404Spjd{ 2569168404Spjd bcopy(buf->b_data, arg, buf->b_hdr->b_size); 2570168404Spjd VERIFY(arc_buf_remove_ref(buf, arg) == 1); 2571168404Spjd} 2572168404Spjd 2573185029Spjd/* a generic arc_done_func_t */ 2574168404Spjdvoid 2575168404Spjdarc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 2576168404Spjd{ 2577168404Spjd arc_buf_t **bufp = arg; 2578168404Spjd if (zio && zio->io_error) { 2579168404Spjd VERIFY(arc_buf_remove_ref(buf, arg) == 1); 2580168404Spjd *bufp = NULL; 2581168404Spjd } else { 2582168404Spjd *bufp = buf; 2583168404Spjd } 2584168404Spjd} 2585168404Spjd 2586168404Spjdstatic void 2587168404Spjdarc_read_done(zio_t *zio) 2588168404Spjd{ 2589168404Spjd arc_buf_hdr_t *hdr, *found; 2590168404Spjd arc_buf_t *buf; 2591168404Spjd arc_buf_t *abuf; /* buffer we're assigning to callback */ 2592168404Spjd kmutex_t *hash_lock; 2593168404Spjd arc_callback_t *callback_list, *acb; 2594168404Spjd int freeable = FALSE; 2595168404Spjd 2596168404Spjd buf = zio->io_private; 2597168404Spjd hdr = buf->b_hdr; 2598168404Spjd 2599168404Spjd /* 2600168404Spjd * The hdr was inserted into hash-table and removed from lists 2601168404Spjd * prior to starting I/O. We should find this header, since 2602168404Spjd * it's in the hash table, and it should be legit since it's 2603168404Spjd * not possible to evict it during the I/O. The only possible 2604168404Spjd * reason for it not to be found is if we were freed during the 2605168404Spjd * read. 2606168404Spjd */ 2607168404Spjd found = buf_hash_find(zio->io_spa, &hdr->b_dva, hdr->b_birth, 2608168404Spjd &hash_lock); 2609168404Spjd 2610168404Spjd ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) || 2611185029Spjd (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 2612185029Spjd (found == hdr && HDR_L2_READING(hdr))); 2613168404Spjd 2614185029Spjd hdr->b_flags &= ~ARC_L2_EVICTED; 2615185029Spjd if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 2616185029Spjd hdr->b_flags &= ~ARC_L2CACHE; 2617205231Skmacy#if 0 2618205231Skmacy else if ((hdr->b_flags & ARC_PREFETCH) == 0) 2619205231Skmacy hdr->b_flags |= ARC_L2CACHE; 2620205231Skmacy#endif 2621168404Spjd /* byteswap if necessary */ 2622168404Spjd callback_list = hdr->b_acb; 2623168404Spjd ASSERT(callback_list != NULL); 2624185029Spjd if (BP_SHOULD_BYTESWAP(zio->io_bp)) { 2625185029Spjd arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 2626185029Spjd byteswap_uint64_array : 2627185029Spjd dmu_ot[BP_GET_TYPE(zio->io_bp)].ot_byteswap; 2628185029Spjd func(buf->b_data, hdr->b_size); 2629185029Spjd } 2630168404Spjd 2631185029Spjd arc_cksum_compute(buf, B_FALSE); 2632168404Spjd 2633168404Spjd /* create copies of the data buffer for the callers */ 2634168404Spjd abuf = buf; 2635168404Spjd for (acb = callback_list; acb; acb = acb->acb_next) { 2636168404Spjd if (acb->acb_done) { 2637168404Spjd if (abuf == NULL) 2638168404Spjd abuf = arc_buf_clone(buf); 2639168404Spjd acb->acb_buf = abuf; 2640168404Spjd abuf = NULL; 2641168404Spjd } 2642168404Spjd } 2643168404Spjd hdr->b_acb = NULL; 2644168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 2645168404Spjd ASSERT(!HDR_BUF_AVAILABLE(hdr)); 2646168404Spjd if (abuf == buf) 2647168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 2648168404Spjd 2649168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 2650168404Spjd 2651168404Spjd if (zio->io_error != 0) { 2652168404Spjd hdr->b_flags |= ARC_IO_ERROR; 2653168404Spjd if (hdr->b_state != arc_anon) 2654168404Spjd arc_change_state(arc_anon, hdr, hash_lock); 2655168404Spjd if (HDR_IN_HASH_TABLE(hdr)) 2656168404Spjd buf_hash_remove(hdr); 2657168404Spjd freeable = refcount_is_zero(&hdr->b_refcnt); 2658168404Spjd } 2659168404Spjd 2660168404Spjd /* 2661168404Spjd * Broadcast before we drop the hash_lock to avoid the possibility 2662168404Spjd * that the hdr (and hence the cv) might be freed before we get to 2663168404Spjd * the cv_broadcast(). 2664168404Spjd */ 2665168404Spjd cv_broadcast(&hdr->b_cv); 2666168404Spjd 2667168404Spjd if (hash_lock) { 2668168404Spjd /* 2669168404Spjd * Only call arc_access on anonymous buffers. This is because 2670168404Spjd * if we've issued an I/O for an evicted buffer, we've already 2671168404Spjd * called arc_access (to prevent any simultaneous readers from 2672168404Spjd * getting confused). 2673168404Spjd */ 2674168404Spjd if (zio->io_error == 0 && hdr->b_state == arc_anon) 2675168404Spjd arc_access(hdr, hash_lock); 2676168404Spjd mutex_exit(hash_lock); 2677168404Spjd } else { 2678168404Spjd /* 2679168404Spjd * This block was freed while we waited for the read to 2680168404Spjd * complete. It has been removed from the hash table and 2681168404Spjd * moved to the anonymous state (so that it won't show up 2682168404Spjd * in the cache). 2683168404Spjd */ 2684168404Spjd ASSERT3P(hdr->b_state, ==, arc_anon); 2685168404Spjd freeable = refcount_is_zero(&hdr->b_refcnt); 2686168404Spjd } 2687168404Spjd 2688168404Spjd /* execute each callback and free its structure */ 2689168404Spjd while ((acb = callback_list) != NULL) { 2690168404Spjd if (acb->acb_done) 2691168404Spjd acb->acb_done(zio, acb->acb_buf, acb->acb_private); 2692168404Spjd 2693168404Spjd if (acb->acb_zio_dummy != NULL) { 2694168404Spjd acb->acb_zio_dummy->io_error = zio->io_error; 2695168404Spjd zio_nowait(acb->acb_zio_dummy); 2696168404Spjd } 2697168404Spjd 2698168404Spjd callback_list = acb->acb_next; 2699168404Spjd kmem_free(acb, sizeof (arc_callback_t)); 2700168404Spjd } 2701168404Spjd 2702168404Spjd if (freeable) 2703168404Spjd arc_hdr_destroy(hdr); 2704168404Spjd} 2705168404Spjd 2706168404Spjd/* 2707168404Spjd * "Read" the block block at the specified DVA (in bp) via the 2708168404Spjd * cache. If the block is found in the cache, invoke the provided 2709168404Spjd * callback immediately and return. Note that the `zio' parameter 2710168404Spjd * in the callback will be NULL in this case, since no IO was 2711168404Spjd * required. If the block is not in the cache pass the read request 2712168404Spjd * on to the spa with a substitute callback function, so that the 2713168404Spjd * requested block will be added to the cache. 2714168404Spjd * 2715168404Spjd * If a read request arrives for a block that has a read in-progress, 2716168404Spjd * either wait for the in-progress read to complete (and return the 2717168404Spjd * results); or, if this is a read with a "done" func, add a record 2718168404Spjd * to the read to invoke the "done" func when the read completes, 2719168404Spjd * and return; or just return. 2720168404Spjd * 2721168404Spjd * arc_read_done() will invoke all the requested "done" functions 2722168404Spjd * for readers of this block. 2723185029Spjd * 2724185029Spjd * Normal callers should use arc_read and pass the arc buffer and offset 2725185029Spjd * for the bp. But if you know you don't need locking, you can use 2726185029Spjd * arc_read_bp. 2727168404Spjd */ 2728168404Spjdint 2729185029Spjdarc_read(zio_t *pio, spa_t *spa, blkptr_t *bp, arc_buf_t *pbuf, 2730185029Spjd arc_done_func_t *done, void *private, int priority, int zio_flags, 2731185029Spjd uint32_t *arc_flags, const zbookmark_t *zb) 2732168404Spjd{ 2733185029Spjd int err; 2734205231Skmacy arc_buf_hdr_t *hdr = pbuf->b_hdr; 2735185029Spjd 2736185029Spjd ASSERT(!refcount_is_zero(&pbuf->b_hdr->b_refcnt)); 2737185029Spjd ASSERT3U((char *)bp - (char *)pbuf->b_data, <, pbuf->b_hdr->b_size); 2738185029Spjd rw_enter(&pbuf->b_lock, RW_READER); 2739185029Spjd 2740185029Spjd err = arc_read_nolock(pio, spa, bp, done, private, priority, 2741185029Spjd zio_flags, arc_flags, zb); 2742185029Spjd 2743205231Skmacy ASSERT3P(hdr, ==, pbuf->b_hdr); 2744185029Spjd rw_exit(&pbuf->b_lock); 2745185029Spjd return (err); 2746185029Spjd} 2747185029Spjd 2748185029Spjdint 2749185029Spjdarc_read_nolock(zio_t *pio, spa_t *spa, blkptr_t *bp, 2750185029Spjd arc_done_func_t *done, void *private, int priority, int zio_flags, 2751185029Spjd uint32_t *arc_flags, const zbookmark_t *zb) 2752185029Spjd{ 2753168404Spjd arc_buf_hdr_t *hdr; 2754168404Spjd arc_buf_t *buf; 2755168404Spjd kmutex_t *hash_lock; 2756185029Spjd zio_t *rzio; 2757168404Spjd 2758168404Spjdtop: 2759168404Spjd hdr = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_lock); 2760168404Spjd if (hdr && hdr->b_datacnt > 0) { 2761168404Spjd 2762168404Spjd *arc_flags |= ARC_CACHED; 2763168404Spjd 2764168404Spjd if (HDR_IO_IN_PROGRESS(hdr)) { 2765168404Spjd 2766168404Spjd if (*arc_flags & ARC_WAIT) { 2767168404Spjd cv_wait(&hdr->b_cv, hash_lock); 2768168404Spjd mutex_exit(hash_lock); 2769168404Spjd goto top; 2770168404Spjd } 2771168404Spjd ASSERT(*arc_flags & ARC_NOWAIT); 2772168404Spjd 2773168404Spjd if (done) { 2774168404Spjd arc_callback_t *acb = NULL; 2775168404Spjd 2776168404Spjd acb = kmem_zalloc(sizeof (arc_callback_t), 2777168404Spjd KM_SLEEP); 2778168404Spjd acb->acb_done = done; 2779168404Spjd acb->acb_private = private; 2780168404Spjd if (pio != NULL) 2781168404Spjd acb->acb_zio_dummy = zio_null(pio, 2782185029Spjd spa, NULL, NULL, zio_flags); 2783168404Spjd 2784168404Spjd ASSERT(acb->acb_done != NULL); 2785168404Spjd acb->acb_next = hdr->b_acb; 2786168404Spjd hdr->b_acb = acb; 2787168404Spjd add_reference(hdr, hash_lock, private); 2788168404Spjd mutex_exit(hash_lock); 2789168404Spjd return (0); 2790168404Spjd } 2791168404Spjd mutex_exit(hash_lock); 2792168404Spjd return (0); 2793168404Spjd } 2794168404Spjd 2795168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 2796168404Spjd 2797168404Spjd if (done) { 2798168404Spjd add_reference(hdr, hash_lock, private); 2799168404Spjd /* 2800168404Spjd * If this block is already in use, create a new 2801168404Spjd * copy of the data so that we will be guaranteed 2802168404Spjd * that arc_release() will always succeed. 2803168404Spjd */ 2804168404Spjd buf = hdr->b_buf; 2805168404Spjd ASSERT(buf); 2806168404Spjd ASSERT(buf->b_data); 2807168404Spjd if (HDR_BUF_AVAILABLE(hdr)) { 2808168404Spjd ASSERT(buf->b_efunc == NULL); 2809168404Spjd hdr->b_flags &= ~ARC_BUF_AVAILABLE; 2810168404Spjd } else { 2811168404Spjd buf = arc_buf_clone(buf); 2812168404Spjd } 2813168404Spjd } else if (*arc_flags & ARC_PREFETCH && 2814168404Spjd refcount_count(&hdr->b_refcnt) == 0) { 2815168404Spjd hdr->b_flags |= ARC_PREFETCH; 2816168404Spjd } 2817168404Spjd DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 2818168404Spjd arc_access(hdr, hash_lock); 2819185029Spjd if (*arc_flags & ARC_L2CACHE) 2820185029Spjd hdr->b_flags |= ARC_L2CACHE; 2821168404Spjd mutex_exit(hash_lock); 2822168404Spjd ARCSTAT_BUMP(arcstat_hits); 2823168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 2824168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 2825168404Spjd data, metadata, hits); 2826168404Spjd 2827168404Spjd if (done) 2828168404Spjd done(NULL, buf, private); 2829168404Spjd } else { 2830168404Spjd uint64_t size = BP_GET_LSIZE(bp); 2831168404Spjd arc_callback_t *acb; 2832185029Spjd vdev_t *vd = NULL; 2833185029Spjd daddr_t addr; 2834168404Spjd 2835168404Spjd if (hdr == NULL) { 2836168404Spjd /* this block is not in the cache */ 2837168404Spjd arc_buf_hdr_t *exists; 2838168404Spjd arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 2839168404Spjd buf = arc_buf_alloc(spa, size, private, type); 2840168404Spjd hdr = buf->b_hdr; 2841168404Spjd hdr->b_dva = *BP_IDENTITY(bp); 2842168404Spjd hdr->b_birth = bp->blk_birth; 2843168404Spjd hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 2844168404Spjd exists = buf_hash_insert(hdr, &hash_lock); 2845168404Spjd if (exists) { 2846168404Spjd /* somebody beat us to the hash insert */ 2847168404Spjd mutex_exit(hash_lock); 2848168404Spjd bzero(&hdr->b_dva, sizeof (dva_t)); 2849168404Spjd hdr->b_birth = 0; 2850168404Spjd hdr->b_cksum0 = 0; 2851168404Spjd (void) arc_buf_remove_ref(buf, private); 2852168404Spjd goto top; /* restart the IO request */ 2853168404Spjd } 2854168404Spjd /* if this is a prefetch, we don't have a reference */ 2855168404Spjd if (*arc_flags & ARC_PREFETCH) { 2856168404Spjd (void) remove_reference(hdr, hash_lock, 2857168404Spjd private); 2858168404Spjd hdr->b_flags |= ARC_PREFETCH; 2859168404Spjd } 2860185029Spjd if (*arc_flags & ARC_L2CACHE) 2861185029Spjd hdr->b_flags |= ARC_L2CACHE; 2862168404Spjd if (BP_GET_LEVEL(bp) > 0) 2863168404Spjd hdr->b_flags |= ARC_INDIRECT; 2864168404Spjd } else { 2865168404Spjd /* this block is in the ghost cache */ 2866168404Spjd ASSERT(GHOST_STATE(hdr->b_state)); 2867168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2868168404Spjd ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 0); 2869168404Spjd ASSERT(hdr->b_buf == NULL); 2870168404Spjd 2871168404Spjd /* if this is a prefetch, we don't have a reference */ 2872168404Spjd if (*arc_flags & ARC_PREFETCH) 2873168404Spjd hdr->b_flags |= ARC_PREFETCH; 2874168404Spjd else 2875168404Spjd add_reference(hdr, hash_lock, private); 2876185029Spjd if (*arc_flags & ARC_L2CACHE) 2877185029Spjd hdr->b_flags |= ARC_L2CACHE; 2878185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2879168404Spjd buf->b_hdr = hdr; 2880168404Spjd buf->b_data = NULL; 2881168404Spjd buf->b_efunc = NULL; 2882168404Spjd buf->b_private = NULL; 2883168404Spjd buf->b_next = NULL; 2884168404Spjd hdr->b_buf = buf; 2885168404Spjd arc_get_data_buf(buf); 2886168404Spjd ASSERT(hdr->b_datacnt == 0); 2887168404Spjd hdr->b_datacnt = 1; 2888168404Spjd 2889168404Spjd } 2890168404Spjd 2891168404Spjd acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 2892168404Spjd acb->acb_done = done; 2893168404Spjd acb->acb_private = private; 2894168404Spjd 2895168404Spjd ASSERT(hdr->b_acb == NULL); 2896168404Spjd hdr->b_acb = acb; 2897168404Spjd hdr->b_flags |= ARC_IO_IN_PROGRESS; 2898168404Spjd 2899168404Spjd /* 2900168404Spjd * If the buffer has been evicted, migrate it to a present state 2901168404Spjd * before issuing the I/O. Once we drop the hash-table lock, 2902168404Spjd * the header will be marked as I/O in progress and have an 2903168404Spjd * attached buffer. At this point, anybody who finds this 2904168404Spjd * buffer ought to notice that it's legit but has a pending I/O. 2905168404Spjd */ 2906168404Spjd 2907168404Spjd if (GHOST_STATE(hdr->b_state)) 2908168404Spjd arc_access(hdr, hash_lock); 2909185029Spjd 2910185029Spjd if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL && 2911185029Spjd (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 2912185029Spjd addr = hdr->b_l2hdr->b_daddr; 2913185029Spjd /* 2914185029Spjd * Lock out device removal. 2915185029Spjd */ 2916185029Spjd if (vdev_is_dead(vd) || 2917185029Spjd !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 2918185029Spjd vd = NULL; 2919185029Spjd } 2920185029Spjd 2921168404Spjd mutex_exit(hash_lock); 2922168404Spjd 2923168404Spjd ASSERT3U(hdr->b_size, ==, size); 2924168404Spjd DTRACE_PROBE3(arc__miss, blkptr_t *, bp, uint64_t, size, 2925168404Spjd zbookmark_t *, zb); 2926168404Spjd ARCSTAT_BUMP(arcstat_misses); 2927168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 2928168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 2929168404Spjd data, metadata, misses); 2930168404Spjd 2931185029Spjd if (vd != NULL) { 2932185029Spjd /* 2933185029Spjd * Read from the L2ARC if the following are true: 2934185029Spjd * 1. The L2ARC vdev was previously cached. 2935185029Spjd * 2. This buffer still has L2ARC metadata. 2936185029Spjd * 3. This buffer isn't currently writing to the L2ARC. 2937185029Spjd * 4. The L2ARC entry wasn't evicted, which may 2938185029Spjd * also have invalidated the vdev. 2939185029Spjd */ 2940185029Spjd if (hdr->b_l2hdr != NULL && 2941185029Spjd !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr)) { 2942185029Spjd l2arc_read_callback_t *cb; 2943185029Spjd 2944185029Spjd DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 2945185029Spjd ARCSTAT_BUMP(arcstat_l2_hits); 2946185029Spjd 2947185029Spjd cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 2948185029Spjd KM_SLEEP); 2949185029Spjd cb->l2rcb_buf = buf; 2950185029Spjd cb->l2rcb_spa = spa; 2951185029Spjd cb->l2rcb_bp = *bp; 2952185029Spjd cb->l2rcb_zb = *zb; 2953185029Spjd cb->l2rcb_flags = zio_flags; 2954185029Spjd 2955185029Spjd /* 2956185029Spjd * l2arc read. The SCL_L2ARC lock will be 2957185029Spjd * released by l2arc_read_done(). 2958185029Spjd */ 2959185029Spjd rzio = zio_read_phys(pio, vd, addr, size, 2960185029Spjd buf->b_data, ZIO_CHECKSUM_OFF, 2961185029Spjd l2arc_read_done, cb, priority, zio_flags | 2962185029Spjd ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL | 2963185029Spjd ZIO_FLAG_DONT_PROPAGATE | 2964185029Spjd ZIO_FLAG_DONT_RETRY, B_FALSE); 2965185029Spjd DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 2966185029Spjd zio_t *, rzio); 2967185029Spjd 2968185029Spjd if (*arc_flags & ARC_NOWAIT) { 2969185029Spjd zio_nowait(rzio); 2970185029Spjd return (0); 2971185029Spjd } 2972185029Spjd 2973185029Spjd ASSERT(*arc_flags & ARC_WAIT); 2974185029Spjd if (zio_wait(rzio) == 0) 2975185029Spjd return (0); 2976185029Spjd 2977185029Spjd /* l2arc read error; goto zio_read() */ 2978185029Spjd } else { 2979185029Spjd DTRACE_PROBE1(l2arc__miss, 2980185029Spjd arc_buf_hdr_t *, hdr); 2981185029Spjd ARCSTAT_BUMP(arcstat_l2_misses); 2982185029Spjd if (HDR_L2_WRITING(hdr)) 2983185029Spjd ARCSTAT_BUMP(arcstat_l2_rw_clash); 2984185029Spjd spa_config_exit(spa, SCL_L2ARC, vd); 2985185029Spjd } 2986185029Spjd } 2987185029Spjd 2988168404Spjd rzio = zio_read(pio, spa, bp, buf->b_data, size, 2989185029Spjd arc_read_done, buf, priority, zio_flags, zb); 2990168404Spjd 2991168404Spjd if (*arc_flags & ARC_WAIT) 2992168404Spjd return (zio_wait(rzio)); 2993168404Spjd 2994168404Spjd ASSERT(*arc_flags & ARC_NOWAIT); 2995168404Spjd zio_nowait(rzio); 2996168404Spjd } 2997168404Spjd return (0); 2998168404Spjd} 2999168404Spjd 3000168404Spjd/* 3001168404Spjd * arc_read() variant to support pool traversal. If the block is already 3002168404Spjd * in the ARC, make a copy of it; otherwise, the caller will do the I/O. 3003168404Spjd * The idea is that we don't want pool traversal filling up memory, but 3004168404Spjd * if the ARC already has the data anyway, we shouldn't pay for the I/O. 3005168404Spjd */ 3006168404Spjdint 3007168404Spjdarc_tryread(spa_t *spa, blkptr_t *bp, void *data) 3008168404Spjd{ 3009168404Spjd arc_buf_hdr_t *hdr; 3010168404Spjd kmutex_t *hash_mtx; 3011168404Spjd int rc = 0; 3012168404Spjd 3013168404Spjd hdr = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_mtx); 3014168404Spjd 3015168404Spjd if (hdr && hdr->b_datacnt > 0 && !HDR_IO_IN_PROGRESS(hdr)) { 3016168404Spjd arc_buf_t *buf = hdr->b_buf; 3017168404Spjd 3018168404Spjd ASSERT(buf); 3019168404Spjd while (buf->b_data == NULL) { 3020168404Spjd buf = buf->b_next; 3021168404Spjd ASSERT(buf); 3022168404Spjd } 3023168404Spjd bcopy(buf->b_data, data, hdr->b_size); 3024168404Spjd } else { 3025168404Spjd rc = ENOENT; 3026168404Spjd } 3027168404Spjd 3028168404Spjd if (hash_mtx) 3029168404Spjd mutex_exit(hash_mtx); 3030168404Spjd 3031168404Spjd return (rc); 3032168404Spjd} 3033168404Spjd 3034168404Spjdvoid 3035168404Spjdarc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3036168404Spjd{ 3037168404Spjd ASSERT(buf->b_hdr != NULL); 3038168404Spjd ASSERT(buf->b_hdr->b_state != arc_anon); 3039168404Spjd ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3040168404Spjd buf->b_efunc = func; 3041168404Spjd buf->b_private = private; 3042168404Spjd} 3043168404Spjd 3044168404Spjd/* 3045168404Spjd * This is used by the DMU to let the ARC know that a buffer is 3046168404Spjd * being evicted, so the ARC should clean up. If this arc buf 3047168404Spjd * is not yet in the evicted state, it will be put there. 3048168404Spjd */ 3049168404Spjdint 3050168404Spjdarc_buf_evict(arc_buf_t *buf) 3051168404Spjd{ 3052168404Spjd arc_buf_hdr_t *hdr; 3053168404Spjd kmutex_t *hash_lock; 3054168404Spjd arc_buf_t **bufp; 3055205231Skmacy list_t *list, *evicted_list; 3056205231Skmacy kmutex_t *lock, *evicted_lock; 3057205231Skmacy 3058185029Spjd rw_enter(&buf->b_lock, RW_WRITER); 3059168404Spjd hdr = buf->b_hdr; 3060168404Spjd if (hdr == NULL) { 3061168404Spjd /* 3062168404Spjd * We are in arc_do_user_evicts(). 3063168404Spjd */ 3064168404Spjd ASSERT(buf->b_data == NULL); 3065185029Spjd rw_exit(&buf->b_lock); 3066168404Spjd return (0); 3067185029Spjd } else if (buf->b_data == NULL) { 3068185029Spjd arc_buf_t copy = *buf; /* structure assignment */ 3069185029Spjd /* 3070185029Spjd * We are on the eviction list; process this buffer now 3071185029Spjd * but let arc_do_user_evicts() do the reaping. 3072185029Spjd */ 3073185029Spjd buf->b_efunc = NULL; 3074185029Spjd rw_exit(&buf->b_lock); 3075185029Spjd VERIFY(copy.b_efunc(©) == 0); 3076185029Spjd return (1); 3077168404Spjd } 3078168404Spjd hash_lock = HDR_LOCK(hdr); 3079168404Spjd mutex_enter(hash_lock); 3080168404Spjd 3081168404Spjd ASSERT(buf->b_hdr == hdr); 3082168404Spjd ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3083168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3084168404Spjd 3085168404Spjd /* 3086168404Spjd * Pull this buffer off of the hdr 3087168404Spjd */ 3088168404Spjd bufp = &hdr->b_buf; 3089168404Spjd while (*bufp != buf) 3090168404Spjd bufp = &(*bufp)->b_next; 3091168404Spjd *bufp = buf->b_next; 3092168404Spjd 3093168404Spjd ASSERT(buf->b_data != NULL); 3094168404Spjd arc_buf_destroy(buf, FALSE, FALSE); 3095168404Spjd 3096168404Spjd if (hdr->b_datacnt == 0) { 3097168404Spjd arc_state_t *old_state = hdr->b_state; 3098168404Spjd arc_state_t *evicted_state; 3099168404Spjd 3100168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 3101168404Spjd 3102168404Spjd evicted_state = 3103168404Spjd (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3104168404Spjd 3105205231Skmacy get_buf_info(hdr, old_state, &list, &lock); 3106205231Skmacy get_buf_info(hdr, evicted_state, &evicted_list, &evicted_lock); 3107205231Skmacy mutex_enter(lock); 3108205231Skmacy mutex_enter(evicted_lock); 3109168404Spjd 3110168404Spjd arc_change_state(evicted_state, hdr, hash_lock); 3111168404Spjd ASSERT(HDR_IN_HASH_TABLE(hdr)); 3112185029Spjd hdr->b_flags |= ARC_IN_HASH_TABLE; 3113185029Spjd hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3114168404Spjd 3115205231Skmacy mutex_exit(evicted_lock); 3116205231Skmacy mutex_exit(lock); 3117168404Spjd } 3118168404Spjd mutex_exit(hash_lock); 3119185029Spjd rw_exit(&buf->b_lock); 3120168404Spjd 3121168404Spjd VERIFY(buf->b_efunc(buf) == 0); 3122168404Spjd buf->b_efunc = NULL; 3123168404Spjd buf->b_private = NULL; 3124168404Spjd buf->b_hdr = NULL; 3125168404Spjd kmem_cache_free(buf_cache, buf); 3126168404Spjd return (1); 3127168404Spjd} 3128168404Spjd 3129168404Spjd/* 3130168404Spjd * Release this buffer from the cache. This must be done 3131168404Spjd * after a read and prior to modifying the buffer contents. 3132168404Spjd * If the buffer has more than one reference, we must make 3133185029Spjd * a new hdr for the buffer. 3134168404Spjd */ 3135168404Spjdvoid 3136168404Spjdarc_release(arc_buf_t *buf, void *tag) 3137168404Spjd{ 3138185029Spjd arc_buf_hdr_t *hdr; 3139185029Spjd kmutex_t *hash_lock; 3140185029Spjd l2arc_buf_hdr_t *l2hdr; 3141185029Spjd uint64_t buf_size; 3142168404Spjd 3143185029Spjd rw_enter(&buf->b_lock, RW_WRITER); 3144185029Spjd hdr = buf->b_hdr; 3145185029Spjd 3146168404Spjd /* this buffer is not on any list */ 3147168404Spjd ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3148185029Spjd ASSERT(!(hdr->b_flags & ARC_STORED)); 3149168404Spjd 3150168404Spjd if (hdr->b_state == arc_anon) { 3151168404Spjd /* this buffer is already released */ 3152168404Spjd ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 1); 3153168404Spjd ASSERT(BUF_EMPTY(hdr)); 3154168404Spjd ASSERT(buf->b_efunc == NULL); 3155168404Spjd arc_buf_thaw(buf); 3156185029Spjd rw_exit(&buf->b_lock); 3157168404Spjd return; 3158168404Spjd } 3159168404Spjd 3160185029Spjd hash_lock = HDR_LOCK(hdr); 3161168404Spjd mutex_enter(hash_lock); 3162168404Spjd 3163185029Spjd l2hdr = hdr->b_l2hdr; 3164185029Spjd if (l2hdr) { 3165185029Spjd mutex_enter(&l2arc_buflist_mtx); 3166185029Spjd hdr->b_l2hdr = NULL; 3167185029Spjd buf_size = hdr->b_size; 3168185029Spjd } 3169185029Spjd 3170168404Spjd /* 3171168404Spjd * Do we have more than one buf? 3172168404Spjd */ 3173185029Spjd if (hdr->b_datacnt > 1) { 3174168404Spjd arc_buf_hdr_t *nhdr; 3175168404Spjd arc_buf_t **bufp; 3176168404Spjd uint64_t blksz = hdr->b_size; 3177168404Spjd spa_t *spa = hdr->b_spa; 3178168404Spjd arc_buf_contents_t type = hdr->b_type; 3179185029Spjd uint32_t flags = hdr->b_flags; 3180168404Spjd 3181185029Spjd ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3182168404Spjd /* 3183168404Spjd * Pull the data off of this buf and attach it to 3184168404Spjd * a new anonymous buf. 3185168404Spjd */ 3186168404Spjd (void) remove_reference(hdr, hash_lock, tag); 3187168404Spjd bufp = &hdr->b_buf; 3188168404Spjd while (*bufp != buf) 3189168404Spjd bufp = &(*bufp)->b_next; 3190168404Spjd *bufp = (*bufp)->b_next; 3191168404Spjd buf->b_next = NULL; 3192168404Spjd 3193168404Spjd ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3194168404Spjd atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3195168404Spjd if (refcount_is_zero(&hdr->b_refcnt)) { 3196185029Spjd uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3197185029Spjd ASSERT3U(*size, >=, hdr->b_size); 3198185029Spjd atomic_add_64(size, -hdr->b_size); 3199168404Spjd } 3200168404Spjd hdr->b_datacnt -= 1; 3201168404Spjd arc_cksum_verify(buf); 3202168404Spjd 3203168404Spjd mutex_exit(hash_lock); 3204168404Spjd 3205185029Spjd nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3206168404Spjd nhdr->b_size = blksz; 3207168404Spjd nhdr->b_spa = spa; 3208168404Spjd nhdr->b_type = type; 3209168404Spjd nhdr->b_buf = buf; 3210168404Spjd nhdr->b_state = arc_anon; 3211168404Spjd nhdr->b_arc_access = 0; 3212185029Spjd nhdr->b_flags = flags & ARC_L2_WRITING; 3213185029Spjd nhdr->b_l2hdr = NULL; 3214168404Spjd nhdr->b_datacnt = 1; 3215168404Spjd nhdr->b_freeze_cksum = NULL; 3216168404Spjd (void) refcount_add(&nhdr->b_refcnt, tag); 3217168404Spjd buf->b_hdr = nhdr; 3218185029Spjd rw_exit(&buf->b_lock); 3219168404Spjd atomic_add_64(&arc_anon->arcs_size, blksz); 3220168404Spjd } else { 3221185029Spjd rw_exit(&buf->b_lock); 3222168404Spjd ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3223168404Spjd ASSERT(!list_link_active(&hdr->b_arc_node)); 3224168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3225168404Spjd arc_change_state(arc_anon, hdr, hash_lock); 3226168404Spjd hdr->b_arc_access = 0; 3227168404Spjd mutex_exit(hash_lock); 3228185029Spjd 3229168404Spjd bzero(&hdr->b_dva, sizeof (dva_t)); 3230168404Spjd hdr->b_birth = 0; 3231168404Spjd hdr->b_cksum0 = 0; 3232168404Spjd arc_buf_thaw(buf); 3233168404Spjd } 3234168404Spjd buf->b_efunc = NULL; 3235168404Spjd buf->b_private = NULL; 3236185029Spjd 3237185029Spjd if (l2hdr) { 3238185029Spjd list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3239185029Spjd kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3240185029Spjd ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3241185029Spjd mutex_exit(&l2arc_buflist_mtx); 3242185029Spjd } 3243168404Spjd} 3244168404Spjd 3245168404Spjdint 3246168404Spjdarc_released(arc_buf_t *buf) 3247168404Spjd{ 3248185029Spjd int released; 3249185029Spjd 3250185029Spjd rw_enter(&buf->b_lock, RW_READER); 3251185029Spjd released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3252185029Spjd rw_exit(&buf->b_lock); 3253185029Spjd return (released); 3254168404Spjd} 3255168404Spjd 3256168404Spjdint 3257168404Spjdarc_has_callback(arc_buf_t *buf) 3258168404Spjd{ 3259185029Spjd int callback; 3260185029Spjd 3261185029Spjd rw_enter(&buf->b_lock, RW_READER); 3262185029Spjd callback = (buf->b_efunc != NULL); 3263185029Spjd rw_exit(&buf->b_lock); 3264185029Spjd return (callback); 3265168404Spjd} 3266168404Spjd 3267168404Spjd#ifdef ZFS_DEBUG 3268168404Spjdint 3269168404Spjdarc_referenced(arc_buf_t *buf) 3270168404Spjd{ 3271185029Spjd int referenced; 3272185029Spjd 3273185029Spjd rw_enter(&buf->b_lock, RW_READER); 3274185029Spjd referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3275185029Spjd rw_exit(&buf->b_lock); 3276185029Spjd return (referenced); 3277168404Spjd} 3278168404Spjd#endif 3279168404Spjd 3280168404Spjdstatic void 3281168404Spjdarc_write_ready(zio_t *zio) 3282168404Spjd{ 3283168404Spjd arc_write_callback_t *callback = zio->io_private; 3284168404Spjd arc_buf_t *buf = callback->awcb_buf; 3285185029Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3286168404Spjd 3287185029Spjd ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3288185029Spjd callback->awcb_ready(zio, buf, callback->awcb_private); 3289185029Spjd 3290185029Spjd /* 3291185029Spjd * If the IO is already in progress, then this is a re-write 3292185029Spjd * attempt, so we need to thaw and re-compute the cksum. 3293185029Spjd * It is the responsibility of the callback to handle the 3294185029Spjd * accounting for any re-write attempt. 3295185029Spjd */ 3296185029Spjd if (HDR_IO_IN_PROGRESS(hdr)) { 3297185029Spjd mutex_enter(&hdr->b_freeze_lock); 3298185029Spjd if (hdr->b_freeze_cksum != NULL) { 3299185029Spjd kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3300185029Spjd hdr->b_freeze_cksum = NULL; 3301185029Spjd } 3302185029Spjd mutex_exit(&hdr->b_freeze_lock); 3303168404Spjd } 3304185029Spjd arc_cksum_compute(buf, B_FALSE); 3305185029Spjd hdr->b_flags |= ARC_IO_IN_PROGRESS; 3306168404Spjd} 3307168404Spjd 3308168404Spjdstatic void 3309168404Spjdarc_write_done(zio_t *zio) 3310168404Spjd{ 3311168404Spjd arc_write_callback_t *callback = zio->io_private; 3312168404Spjd arc_buf_t *buf = callback->awcb_buf; 3313168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3314168404Spjd 3315168404Spjd hdr->b_acb = NULL; 3316168404Spjd 3317168404Spjd hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3318168404Spjd hdr->b_birth = zio->io_bp->blk_birth; 3319168404Spjd hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3320168404Spjd /* 3321168404Spjd * If the block to be written was all-zero, we may have 3322168404Spjd * compressed it away. In this case no write was performed 3323168404Spjd * so there will be no dva/birth-date/checksum. The buffer 3324168404Spjd * must therefor remain anonymous (and uncached). 3325168404Spjd */ 3326168404Spjd if (!BUF_EMPTY(hdr)) { 3327168404Spjd arc_buf_hdr_t *exists; 3328168404Spjd kmutex_t *hash_lock; 3329168404Spjd 3330168404Spjd arc_cksum_verify(buf); 3331168404Spjd 3332168404Spjd exists = buf_hash_insert(hdr, &hash_lock); 3333168404Spjd if (exists) { 3334168404Spjd /* 3335168404Spjd * This can only happen if we overwrite for 3336168404Spjd * sync-to-convergence, because we remove 3337168404Spjd * buffers from the hash table when we arc_free(). 3338168404Spjd */ 3339185029Spjd ASSERT(zio->io_flags & ZIO_FLAG_IO_REWRITE); 3340168404Spjd ASSERT(DVA_EQUAL(BP_IDENTITY(&zio->io_bp_orig), 3341168404Spjd BP_IDENTITY(zio->io_bp))); 3342168404Spjd ASSERT3U(zio->io_bp_orig.blk_birth, ==, 3343168404Spjd zio->io_bp->blk_birth); 3344168404Spjd 3345168404Spjd ASSERT(refcount_is_zero(&exists->b_refcnt)); 3346168404Spjd arc_change_state(arc_anon, exists, hash_lock); 3347168404Spjd mutex_exit(hash_lock); 3348168404Spjd arc_hdr_destroy(exists); 3349168404Spjd exists = buf_hash_insert(hdr, &hash_lock); 3350168404Spjd ASSERT3P(exists, ==, NULL); 3351168404Spjd } 3352168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3353185029Spjd /* if it's not anon, we are doing a scrub */ 3354185029Spjd if (hdr->b_state == arc_anon) 3355185029Spjd arc_access(hdr, hash_lock); 3356168404Spjd mutex_exit(hash_lock); 3357168404Spjd } else if (callback->awcb_done == NULL) { 3358168404Spjd int destroy_hdr; 3359168404Spjd /* 3360168404Spjd * This is an anonymous buffer with no user callback, 3361168404Spjd * destroy it if there are no active references. 3362168404Spjd */ 3363168404Spjd mutex_enter(&arc_eviction_mtx); 3364168404Spjd destroy_hdr = refcount_is_zero(&hdr->b_refcnt); 3365168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3366168404Spjd mutex_exit(&arc_eviction_mtx); 3367168404Spjd if (destroy_hdr) 3368168404Spjd arc_hdr_destroy(hdr); 3369168404Spjd } else { 3370168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3371168404Spjd } 3372185029Spjd hdr->b_flags &= ~ARC_STORED; 3373168404Spjd 3374168404Spjd if (callback->awcb_done) { 3375168404Spjd ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3376168404Spjd callback->awcb_done(zio, buf, callback->awcb_private); 3377168404Spjd } 3378168404Spjd 3379168404Spjd kmem_free(callback, sizeof (arc_write_callback_t)); 3380168404Spjd} 3381168404Spjd 3382185029Spjdstatic void 3383185029Spjdwrite_policy(spa_t *spa, const writeprops_t *wp, zio_prop_t *zp) 3384185029Spjd{ 3385185029Spjd boolean_t ismd = (wp->wp_level > 0 || dmu_ot[wp->wp_type].ot_metadata); 3386185029Spjd 3387185029Spjd /* Determine checksum setting */ 3388185029Spjd if (ismd) { 3389185029Spjd /* 3390185029Spjd * Metadata always gets checksummed. If the data 3391185029Spjd * checksum is multi-bit correctable, and it's not a 3392185029Spjd * ZBT-style checksum, then it's suitable for metadata 3393185029Spjd * as well. Otherwise, the metadata checksum defaults 3394185029Spjd * to fletcher4. 3395185029Spjd */ 3396185029Spjd if (zio_checksum_table[wp->wp_oschecksum].ci_correctable && 3397185029Spjd !zio_checksum_table[wp->wp_oschecksum].ci_zbt) 3398185029Spjd zp->zp_checksum = wp->wp_oschecksum; 3399185029Spjd else 3400185029Spjd zp->zp_checksum = ZIO_CHECKSUM_FLETCHER_4; 3401185029Spjd } else { 3402185029Spjd zp->zp_checksum = zio_checksum_select(wp->wp_dnchecksum, 3403185029Spjd wp->wp_oschecksum); 3404185029Spjd } 3405185029Spjd 3406185029Spjd /* Determine compression setting */ 3407185029Spjd if (ismd) { 3408185029Spjd /* 3409185029Spjd * XXX -- we should design a compression algorithm 3410185029Spjd * that specializes in arrays of bps. 3411185029Spjd */ 3412185029Spjd zp->zp_compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY : 3413185029Spjd ZIO_COMPRESS_LZJB; 3414185029Spjd } else { 3415185029Spjd zp->zp_compress = zio_compress_select(wp->wp_dncompress, 3416185029Spjd wp->wp_oscompress); 3417185029Spjd } 3418185029Spjd 3419185029Spjd zp->zp_type = wp->wp_type; 3420185029Spjd zp->zp_level = wp->wp_level; 3421185029Spjd zp->zp_ndvas = MIN(wp->wp_copies + ismd, spa_max_replication(spa)); 3422185029Spjd} 3423185029Spjd 3424168404Spjdzio_t * 3425185029Spjdarc_write(zio_t *pio, spa_t *spa, const writeprops_t *wp, 3426185029Spjd boolean_t l2arc, uint64_t txg, blkptr_t *bp, arc_buf_t *buf, 3427168404Spjd arc_done_func_t *ready, arc_done_func_t *done, void *private, int priority, 3428185029Spjd int zio_flags, const zbookmark_t *zb) 3429168404Spjd{ 3430168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3431168404Spjd arc_write_callback_t *callback; 3432185029Spjd zio_t *zio; 3433185029Spjd zio_prop_t zp; 3434168404Spjd 3435185029Spjd ASSERT(ready != NULL); 3436168404Spjd ASSERT(!HDR_IO_ERROR(hdr)); 3437168404Spjd ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3438168404Spjd ASSERT(hdr->b_acb == 0); 3439185029Spjd if (l2arc) 3440185029Spjd hdr->b_flags |= ARC_L2CACHE; 3441168404Spjd callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 3442168404Spjd callback->awcb_ready = ready; 3443168404Spjd callback->awcb_done = done; 3444168404Spjd callback->awcb_private = private; 3445168404Spjd callback->awcb_buf = buf; 3446168404Spjd 3447185029Spjd write_policy(spa, wp, &zp); 3448185029Spjd zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, &zp, 3449185029Spjd arc_write_ready, arc_write_done, callback, priority, zio_flags, zb); 3450185029Spjd 3451168404Spjd return (zio); 3452168404Spjd} 3453168404Spjd 3454168404Spjdint 3455168404Spjdarc_free(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 3456168404Spjd zio_done_func_t *done, void *private, uint32_t arc_flags) 3457168404Spjd{ 3458168404Spjd arc_buf_hdr_t *ab; 3459168404Spjd kmutex_t *hash_lock; 3460168404Spjd zio_t *zio; 3461168404Spjd 3462168404Spjd /* 3463168404Spjd * If this buffer is in the cache, release it, so it 3464168404Spjd * can be re-used. 3465168404Spjd */ 3466168404Spjd ab = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_lock); 3467168404Spjd if (ab != NULL) { 3468168404Spjd /* 3469168404Spjd * The checksum of blocks to free is not always 3470168404Spjd * preserved (eg. on the deadlist). However, if it is 3471168404Spjd * nonzero, it should match what we have in the cache. 3472168404Spjd */ 3473168404Spjd ASSERT(bp->blk_cksum.zc_word[0] == 0 || 3474185029Spjd bp->blk_cksum.zc_word[0] == ab->b_cksum0 || 3475185029Spjd bp->blk_fill == BLK_FILL_ALREADY_FREED); 3476185029Spjd 3477168404Spjd if (ab->b_state != arc_anon) 3478168404Spjd arc_change_state(arc_anon, ab, hash_lock); 3479168404Spjd if (HDR_IO_IN_PROGRESS(ab)) { 3480168404Spjd /* 3481168404Spjd * This should only happen when we prefetch. 3482168404Spjd */ 3483168404Spjd ASSERT(ab->b_flags & ARC_PREFETCH); 3484168404Spjd ASSERT3U(ab->b_datacnt, ==, 1); 3485168404Spjd ab->b_flags |= ARC_FREED_IN_READ; 3486168404Spjd if (HDR_IN_HASH_TABLE(ab)) 3487168404Spjd buf_hash_remove(ab); 3488168404Spjd ab->b_arc_access = 0; 3489168404Spjd bzero(&ab->b_dva, sizeof (dva_t)); 3490168404Spjd ab->b_birth = 0; 3491168404Spjd ab->b_cksum0 = 0; 3492168404Spjd ab->b_buf->b_efunc = NULL; 3493168404Spjd ab->b_buf->b_private = NULL; 3494168404Spjd mutex_exit(hash_lock); 3495168404Spjd } else if (refcount_is_zero(&ab->b_refcnt)) { 3496185029Spjd ab->b_flags |= ARC_FREE_IN_PROGRESS; 3497168404Spjd mutex_exit(hash_lock); 3498168404Spjd arc_hdr_destroy(ab); 3499168404Spjd ARCSTAT_BUMP(arcstat_deleted); 3500168404Spjd } else { 3501168404Spjd /* 3502168404Spjd * We still have an active reference on this 3503168404Spjd * buffer. This can happen, e.g., from 3504168404Spjd * dbuf_unoverride(). 3505168404Spjd */ 3506168404Spjd ASSERT(!HDR_IN_HASH_TABLE(ab)); 3507168404Spjd ab->b_arc_access = 0; 3508168404Spjd bzero(&ab->b_dva, sizeof (dva_t)); 3509168404Spjd ab->b_birth = 0; 3510168404Spjd ab->b_cksum0 = 0; 3511168404Spjd ab->b_buf->b_efunc = NULL; 3512168404Spjd ab->b_buf->b_private = NULL; 3513168404Spjd mutex_exit(hash_lock); 3514168404Spjd } 3515168404Spjd } 3516168404Spjd 3517185029Spjd zio = zio_free(pio, spa, txg, bp, done, private, ZIO_FLAG_MUSTSUCCEED); 3518168404Spjd 3519168404Spjd if (arc_flags & ARC_WAIT) 3520168404Spjd return (zio_wait(zio)); 3521168404Spjd 3522168404Spjd ASSERT(arc_flags & ARC_NOWAIT); 3523168404Spjd zio_nowait(zio); 3524168404Spjd 3525168404Spjd return (0); 3526168404Spjd} 3527168404Spjd 3528185029Spjdstatic int 3529185029Spjdarc_memory_throttle(uint64_t reserve, uint64_t txg) 3530185029Spjd{ 3531185029Spjd#ifdef _KERNEL 3532185029Spjd uint64_t inflight_data = arc_anon->arcs_size; 3533185029Spjd uint64_t available_memory = ptoa((uintmax_t)cnt.v_free_count); 3534185029Spjd static uint64_t page_load = 0; 3535185029Spjd static uint64_t last_txg = 0; 3536185029Spjd 3537185029Spjd#if 0 3538185029Spjd#if defined(__i386) 3539185029Spjd available_memory = 3540185029Spjd MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 3541185029Spjd#endif 3542185029Spjd#endif 3543185029Spjd if (available_memory >= zfs_write_limit_max) 3544185029Spjd return (0); 3545185029Spjd 3546185029Spjd if (txg > last_txg) { 3547185029Spjd last_txg = txg; 3548185029Spjd page_load = 0; 3549185029Spjd } 3550185029Spjd /* 3551185029Spjd * If we are in pageout, we know that memory is already tight, 3552185029Spjd * the arc is already going to be evicting, so we just want to 3553185029Spjd * continue to let page writes occur as quickly as possible. 3554185029Spjd */ 3555185029Spjd if (curproc == pageproc) { 3556185029Spjd if (page_load > available_memory / 4) 3557185029Spjd return (ERESTART); 3558185029Spjd /* Note: reserve is inflated, so we deflate */ 3559185029Spjd page_load += reserve / 8; 3560185029Spjd return (0); 3561185029Spjd } else if (page_load > 0 && arc_reclaim_needed()) { 3562185029Spjd /* memory is low, delay before restarting */ 3563185029Spjd ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3564185029Spjd return (EAGAIN); 3565185029Spjd } 3566185029Spjd page_load = 0; 3567185029Spjd 3568185029Spjd if (arc_size > arc_c_min) { 3569185029Spjd uint64_t evictable_memory = 3570185029Spjd arc_mru->arcs_lsize[ARC_BUFC_DATA] + 3571185029Spjd arc_mru->arcs_lsize[ARC_BUFC_METADATA] + 3572185029Spjd arc_mfu->arcs_lsize[ARC_BUFC_DATA] + 3573185029Spjd arc_mfu->arcs_lsize[ARC_BUFC_METADATA]; 3574185029Spjd available_memory += MIN(evictable_memory, arc_size - arc_c_min); 3575185029Spjd } 3576185029Spjd 3577185029Spjd if (inflight_data > available_memory / 4) { 3578185029Spjd ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3579185029Spjd return (ERESTART); 3580185029Spjd } 3581185029Spjd#endif 3582185029Spjd return (0); 3583185029Spjd} 3584185029Spjd 3585168404Spjdvoid 3586185029Spjdarc_tempreserve_clear(uint64_t reserve) 3587168404Spjd{ 3588185029Spjd atomic_add_64(&arc_tempreserve, -reserve); 3589168404Spjd ASSERT((int64_t)arc_tempreserve >= 0); 3590168404Spjd} 3591168404Spjd 3592168404Spjdint 3593185029Spjdarc_tempreserve_space(uint64_t reserve, uint64_t txg) 3594168404Spjd{ 3595185029Spjd int error; 3596185029Spjd 3597168404Spjd#ifdef ZFS_DEBUG 3598168404Spjd /* 3599168404Spjd * Once in a while, fail for no reason. Everything should cope. 3600168404Spjd */ 3601168404Spjd if (spa_get_random(10000) == 0) { 3602168404Spjd dprintf("forcing random failure\n"); 3603168404Spjd return (ERESTART); 3604168404Spjd } 3605168404Spjd#endif 3606185029Spjd if (reserve > arc_c/4 && !arc_no_grow) 3607185029Spjd arc_c = MIN(arc_c_max, reserve * 4); 3608185029Spjd if (reserve > arc_c) 3609168404Spjd return (ENOMEM); 3610168404Spjd 3611168404Spjd /* 3612185029Spjd * Writes will, almost always, require additional memory allocations 3613185029Spjd * in order to compress/encrypt/etc the data. We therefor need to 3614185029Spjd * make sure that there is sufficient available memory for this. 3615185029Spjd */ 3616185029Spjd if (error = arc_memory_throttle(reserve, txg)) 3617185029Spjd return (error); 3618185029Spjd 3619185029Spjd /* 3620168404Spjd * Throttle writes when the amount of dirty data in the cache 3621168404Spjd * gets too large. We try to keep the cache less than half full 3622168404Spjd * of dirty blocks so that our sync times don't grow too large. 3623168404Spjd * Note: if two requests come in concurrently, we might let them 3624168404Spjd * both succeed, when one of them should fail. Not a huge deal. 3625168404Spjd */ 3626185029Spjd if (reserve + arc_tempreserve + arc_anon->arcs_size > arc_c / 2 && 3627185029Spjd arc_anon->arcs_size > arc_c / 4) { 3628185029Spjd dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 3629185029Spjd "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 3630185029Spjd arc_tempreserve>>10, 3631185029Spjd arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 3632185029Spjd arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 3633185029Spjd reserve>>10, arc_c>>10); 3634168404Spjd return (ERESTART); 3635168404Spjd } 3636185029Spjd atomic_add_64(&arc_tempreserve, reserve); 3637168404Spjd return (0); 3638168404Spjd} 3639168404Spjd 3640168582Spjdstatic kmutex_t arc_lowmem_lock; 3641168404Spjd#ifdef _KERNEL 3642168566Spjdstatic eventhandler_tag arc_event_lowmem = NULL; 3643168404Spjd 3644168404Spjdstatic void 3645168566Spjdarc_lowmem(void *arg __unused, int howto __unused) 3646168404Spjd{ 3647168404Spjd 3648168566Spjd /* Serialize access via arc_lowmem_lock. */ 3649168566Spjd mutex_enter(&arc_lowmem_lock); 3650185029Spjd needfree = 1; 3651168404Spjd cv_signal(&arc_reclaim_thr_cv); 3652185029Spjd while (needfree) 3653185029Spjd tsleep(&needfree, 0, "zfs:lowmem", hz / 5); 3654168566Spjd mutex_exit(&arc_lowmem_lock); 3655168404Spjd} 3656168404Spjd#endif 3657168404Spjd 3658168404Spjdvoid 3659168404Spjdarc_init(void) 3660168404Spjd{ 3661193953Skmacy int prefetch_tunable_set = 0; 3662205231Skmacy int i; 3663205231Skmacy 3664168404Spjd mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 3665168404Spjd cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 3666168566Spjd mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 3667168404Spjd 3668168404Spjd /* Convert seconds to clock ticks */ 3669168404Spjd arc_min_prefetch_lifespan = 1 * hz; 3670168404Spjd 3671168404Spjd /* Start out with 1/8 of all memory */ 3672168566Spjd arc_c = kmem_size() / 8; 3673192360Skmacy#if 0 3674192360Skmacy#ifdef _KERNEL 3675192360Skmacy /* 3676192360Skmacy * On architectures where the physical memory can be larger 3677192360Skmacy * than the addressable space (intel in 32-bit mode), we may 3678192360Skmacy * need to limit the cache to 1/8 of VM size. 3679192360Skmacy */ 3680192360Skmacy arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 3681192360Skmacy#endif 3682192360Skmacy#endif 3683168566Spjd /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 3684168566Spjd arc_c_min = MAX(arc_c / 4, 64<<18); 3685168566Spjd /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 3686168404Spjd if (arc_c * 8 >= 1<<30) 3687168404Spjd arc_c_max = (arc_c * 8) - (1<<30); 3688168404Spjd else 3689168404Spjd arc_c_max = arc_c_min; 3690175633Spjd arc_c_max = MAX(arc_c * 5, arc_c_max); 3691168481Spjd#ifdef _KERNEL 3692168404Spjd /* 3693168404Spjd * Allow the tunables to override our calculations if they are 3694168566Spjd * reasonable (ie. over 16MB) 3695168404Spjd */ 3696168566Spjd if (zfs_arc_max >= 64<<18 && zfs_arc_max < kmem_size()) 3697168404Spjd arc_c_max = zfs_arc_max; 3698168566Spjd if (zfs_arc_min >= 64<<18 && zfs_arc_min <= arc_c_max) 3699168404Spjd arc_c_min = zfs_arc_min; 3700168481Spjd#endif 3701168404Spjd arc_c = arc_c_max; 3702168404Spjd arc_p = (arc_c >> 1); 3703168404Spjd 3704185029Spjd /* limit meta-data to 1/4 of the arc capacity */ 3705185029Spjd arc_meta_limit = arc_c_max / 4; 3706185029Spjd 3707185029Spjd /* Allow the tunable to override if it is reasonable */ 3708185029Spjd if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 3709185029Spjd arc_meta_limit = zfs_arc_meta_limit; 3710185029Spjd 3711185029Spjd if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 3712185029Spjd arc_c_min = arc_meta_limit / 2; 3713185029Spjd 3714168404Spjd /* if kmem_flags are set, lets try to use less memory */ 3715168404Spjd if (kmem_debugging()) 3716168404Spjd arc_c = arc_c / 2; 3717168404Spjd if (arc_c < arc_c_min) 3718168404Spjd arc_c = arc_c_min; 3719168404Spjd 3720168473Spjd zfs_arc_min = arc_c_min; 3721168473Spjd zfs_arc_max = arc_c_max; 3722168473Spjd 3723168404Spjd arc_anon = &ARC_anon; 3724168404Spjd arc_mru = &ARC_mru; 3725168404Spjd arc_mru_ghost = &ARC_mru_ghost; 3726168404Spjd arc_mfu = &ARC_mfu; 3727168404Spjd arc_mfu_ghost = &ARC_mfu_ghost; 3728185029Spjd arc_l2c_only = &ARC_l2c_only; 3729168404Spjd arc_size = 0; 3730168404Spjd 3731205231Skmacy for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 3732205231Skmacy 3733205231Skmacy mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 3734205231Skmacy NULL, MUTEX_DEFAULT, NULL); 3735205231Skmacy mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 3736205231Skmacy NULL, MUTEX_DEFAULT, NULL); 3737205231Skmacy mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 3738205231Skmacy NULL, MUTEX_DEFAULT, NULL); 3739205231Skmacy mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 3740205231Skmacy NULL, MUTEX_DEFAULT, NULL); 3741205231Skmacy mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 3742205231Skmacy NULL, MUTEX_DEFAULT, NULL); 3743205231Skmacy mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 3744205231Skmacy NULL, MUTEX_DEFAULT, NULL); 3745205231Skmacy 3746205231Skmacy list_create(&arc_mru->arcs_lists[i], 3747205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3748205231Skmacy list_create(&arc_mru_ghost->arcs_lists[i], 3749205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3750205231Skmacy list_create(&arc_mfu->arcs_lists[i], 3751205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3752205231Skmacy list_create(&arc_mfu_ghost->arcs_lists[i], 3753205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3754205231Skmacy list_create(&arc_mfu_ghost->arcs_lists[i], 3755205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3756205231Skmacy list_create(&arc_l2c_only->arcs_lists[i], 3757205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3758205231Skmacy } 3759168404Spjd 3760168404Spjd buf_init(); 3761168404Spjd 3762168404Spjd arc_thread_exit = 0; 3763168404Spjd arc_eviction_list = NULL; 3764168404Spjd mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 3765168404Spjd bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 3766168404Spjd 3767168404Spjd arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 3768168404Spjd sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 3769168404Spjd 3770168404Spjd if (arc_ksp != NULL) { 3771168404Spjd arc_ksp->ks_data = &arc_stats; 3772168404Spjd kstat_install(arc_ksp); 3773168404Spjd } 3774168404Spjd 3775168404Spjd (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 3776168404Spjd TS_RUN, minclsyspri); 3777168404Spjd 3778168404Spjd#ifdef _KERNEL 3779168566Spjd arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 3780168404Spjd EVENTHANDLER_PRI_FIRST); 3781168404Spjd#endif 3782168404Spjd 3783168404Spjd arc_dead = FALSE; 3784185029Spjd arc_warm = B_FALSE; 3785168566Spjd 3786185029Spjd if (zfs_write_limit_max == 0) 3787185029Spjd zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift; 3788185029Spjd else 3789185029Spjd zfs_write_limit_shift = 0; 3790185029Spjd mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL); 3791185029Spjd 3792168566Spjd#ifdef _KERNEL 3793194043Skmacy if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 3794193953Skmacy prefetch_tunable_set = 1; 3795193953Skmacy 3796193878Skmacy#ifdef __i386__ 3797193953Skmacy if (prefetch_tunable_set == 0) { 3798196863Strasz printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 3799196863Strasz "-- to enable,\n"); 3800196863Strasz printf(" add \"vfs.zfs.prefetch_disable=0\" " 3801196863Strasz "to /boot/loader.conf.\n"); 3802194043Skmacy zfs_prefetch_disable=1; 3803193878Skmacy } 3804193953Skmacy#else 3805193878Skmacy if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 3806193953Skmacy prefetch_tunable_set == 0) { 3807196863Strasz printf("ZFS NOTICE: Prefetch is disabled by default if less " 3808196941Strasz "than 4GB of RAM is present;\n" 3809196863Strasz " to enable, add \"vfs.zfs.prefetch_disable=0\" " 3810196863Strasz "to /boot/loader.conf.\n"); 3811194043Skmacy zfs_prefetch_disable=1; 3812193878Skmacy } 3813193953Skmacy#endif 3814175633Spjd /* Warn about ZFS memory and address space requirements. */ 3815168696Spjd if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 3816168987Sbmah printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 3817168987Sbmah "expect unstable behavior.\n"); 3818175633Spjd } 3819175633Spjd if (kmem_size() < 512 * (1 << 20)) { 3820173419Spjd printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 3821168987Sbmah "expect unstable behavior.\n"); 3822185029Spjd printf(" Consider tuning vm.kmem_size and " 3823173419Spjd "vm.kmem_size_max\n"); 3824185029Spjd printf(" in /boot/loader.conf.\n"); 3825168566Spjd } 3826168566Spjd#endif 3827168404Spjd} 3828168404Spjd 3829168404Spjdvoid 3830168404Spjdarc_fini(void) 3831168404Spjd{ 3832205231Skmacy int i; 3833205231Skmacy 3834168404Spjd mutex_enter(&arc_reclaim_thr_lock); 3835168404Spjd arc_thread_exit = 1; 3836168404Spjd cv_signal(&arc_reclaim_thr_cv); 3837168404Spjd while (arc_thread_exit != 0) 3838168404Spjd cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 3839168404Spjd mutex_exit(&arc_reclaim_thr_lock); 3840168404Spjd 3841185029Spjd arc_flush(NULL); 3842168404Spjd 3843168404Spjd arc_dead = TRUE; 3844168404Spjd 3845168404Spjd if (arc_ksp != NULL) { 3846168404Spjd kstat_delete(arc_ksp); 3847168404Spjd arc_ksp = NULL; 3848168404Spjd } 3849168404Spjd 3850168404Spjd mutex_destroy(&arc_eviction_mtx); 3851168404Spjd mutex_destroy(&arc_reclaim_thr_lock); 3852168404Spjd cv_destroy(&arc_reclaim_thr_cv); 3853168404Spjd 3854205231Skmacy for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 3855205231Skmacy list_destroy(&arc_mru->arcs_lists[i]); 3856205231Skmacy list_destroy(&arc_mru_ghost->arcs_lists[i]); 3857205231Skmacy list_destroy(&arc_mfu->arcs_lists[i]); 3858205231Skmacy list_destroy(&arc_mfu_ghost->arcs_lists[i]); 3859168404Spjd 3860205231Skmacy mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 3861205231Skmacy mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 3862205231Skmacy mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 3863205231Skmacy mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 3864205231Skmacy mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 3865205231Skmacy } 3866205231Skmacy 3867185029Spjd mutex_destroy(&zfs_write_limit_lock); 3868185029Spjd 3869168404Spjd buf_fini(); 3870168404Spjd 3871168582Spjd mutex_destroy(&arc_lowmem_lock); 3872168404Spjd#ifdef _KERNEL 3873168566Spjd if (arc_event_lowmem != NULL) 3874168566Spjd EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 3875168404Spjd#endif 3876168404Spjd} 3877185029Spjd 3878185029Spjd/* 3879185029Spjd * Level 2 ARC 3880185029Spjd * 3881185029Spjd * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 3882185029Spjd * It uses dedicated storage devices to hold cached data, which are populated 3883185029Spjd * using large infrequent writes. The main role of this cache is to boost 3884185029Spjd * the performance of random read workloads. The intended L2ARC devices 3885185029Spjd * include short-stroked disks, solid state disks, and other media with 3886185029Spjd * substantially faster read latency than disk. 3887185029Spjd * 3888185029Spjd * +-----------------------+ 3889185029Spjd * | ARC | 3890185029Spjd * +-----------------------+ 3891185029Spjd * | ^ ^ 3892185029Spjd * | | | 3893185029Spjd * l2arc_feed_thread() arc_read() 3894185029Spjd * | | | 3895185029Spjd * | l2arc read | 3896185029Spjd * V | | 3897185029Spjd * +---------------+ | 3898185029Spjd * | L2ARC | | 3899185029Spjd * +---------------+ | 3900185029Spjd * | ^ | 3901185029Spjd * l2arc_write() | | 3902185029Spjd * | | | 3903185029Spjd * V | | 3904185029Spjd * +-------+ +-------+ 3905185029Spjd * | vdev | | vdev | 3906185029Spjd * | cache | | cache | 3907185029Spjd * +-------+ +-------+ 3908185029Spjd * +=========+ .-----. 3909185029Spjd * : L2ARC : |-_____-| 3910185029Spjd * : devices : | Disks | 3911185029Spjd * +=========+ `-_____-' 3912185029Spjd * 3913185029Spjd * Read requests are satisfied from the following sources, in order: 3914185029Spjd * 3915185029Spjd * 1) ARC 3916185029Spjd * 2) vdev cache of L2ARC devices 3917185029Spjd * 3) L2ARC devices 3918185029Spjd * 4) vdev cache of disks 3919185029Spjd * 5) disks 3920185029Spjd * 3921185029Spjd * Some L2ARC device types exhibit extremely slow write performance. 3922185029Spjd * To accommodate for this there are some significant differences between 3923185029Spjd * the L2ARC and traditional cache design: 3924185029Spjd * 3925185029Spjd * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 3926185029Spjd * the ARC behave as usual, freeing buffers and placing headers on ghost 3927185029Spjd * lists. The ARC does not send buffers to the L2ARC during eviction as 3928185029Spjd * this would add inflated write latencies for all ARC memory pressure. 3929185029Spjd * 3930185029Spjd * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 3931185029Spjd * It does this by periodically scanning buffers from the eviction-end of 3932185029Spjd * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 3933185029Spjd * not already there. It scans until a headroom of buffers is satisfied, 3934185029Spjd * which itself is a buffer for ARC eviction. The thread that does this is 3935185029Spjd * l2arc_feed_thread(), illustrated below; example sizes are included to 3936185029Spjd * provide a better sense of ratio than this diagram: 3937185029Spjd * 3938185029Spjd * head --> tail 3939185029Spjd * +---------------------+----------+ 3940185029Spjd * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 3941185029Spjd * +---------------------+----------+ | o L2ARC eligible 3942185029Spjd * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 3943185029Spjd * +---------------------+----------+ | 3944185029Spjd * 15.9 Gbytes ^ 32 Mbytes | 3945185029Spjd * headroom | 3946185029Spjd * l2arc_feed_thread() 3947185029Spjd * | 3948185029Spjd * l2arc write hand <--[oooo]--' 3949185029Spjd * | 8 Mbyte 3950185029Spjd * | write max 3951185029Spjd * V 3952185029Spjd * +==============================+ 3953185029Spjd * L2ARC dev |####|#|###|###| |####| ... | 3954185029Spjd * +==============================+ 3955185029Spjd * 32 Gbytes 3956185029Spjd * 3957185029Spjd * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 3958185029Spjd * evicted, then the L2ARC has cached a buffer much sooner than it probably 3959185029Spjd * needed to, potentially wasting L2ARC device bandwidth and storage. It is 3960185029Spjd * safe to say that this is an uncommon case, since buffers at the end of 3961185029Spjd * the ARC lists have moved there due to inactivity. 3962185029Spjd * 3963185029Spjd * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 3964185029Spjd * then the L2ARC simply misses copying some buffers. This serves as a 3965185029Spjd * pressure valve to prevent heavy read workloads from both stalling the ARC 3966185029Spjd * with waits and clogging the L2ARC with writes. This also helps prevent 3967185029Spjd * the potential for the L2ARC to churn if it attempts to cache content too 3968185029Spjd * quickly, such as during backups of the entire pool. 3969185029Spjd * 3970185029Spjd * 5. After system boot and before the ARC has filled main memory, there are 3971185029Spjd * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 3972185029Spjd * lists can remain mostly static. Instead of searching from tail of these 3973185029Spjd * lists as pictured, the l2arc_feed_thread() will search from the list heads 3974185029Spjd * for eligible buffers, greatly increasing its chance of finding them. 3975185029Spjd * 3976185029Spjd * The L2ARC device write speed is also boosted during this time so that 3977185029Spjd * the L2ARC warms up faster. Since there have been no ARC evictions yet, 3978185029Spjd * there are no L2ARC reads, and no fear of degrading read performance 3979185029Spjd * through increased writes. 3980185029Spjd * 3981185029Spjd * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 3982185029Spjd * the vdev queue can aggregate them into larger and fewer writes. Each 3983185029Spjd * device is written to in a rotor fashion, sweeping writes through 3984185029Spjd * available space then repeating. 3985185029Spjd * 3986185029Spjd * 7. The L2ARC does not store dirty content. It never needs to flush 3987185029Spjd * write buffers back to disk based storage. 3988185029Spjd * 3989185029Spjd * 8. If an ARC buffer is written (and dirtied) which also exists in the 3990185029Spjd * L2ARC, the now stale L2ARC buffer is immediately dropped. 3991185029Spjd * 3992185029Spjd * The performance of the L2ARC can be tweaked by a number of tunables, which 3993185029Spjd * may be necessary for different workloads: 3994185029Spjd * 3995185029Spjd * l2arc_write_max max write bytes per interval 3996185029Spjd * l2arc_write_boost extra write bytes during device warmup 3997185029Spjd * l2arc_noprefetch skip caching prefetched buffers 3998185029Spjd * l2arc_headroom number of max device writes to precache 3999185029Spjd * l2arc_feed_secs seconds between L2ARC writing 4000185029Spjd * 4001185029Spjd * Tunables may be removed or added as future performance improvements are 4002185029Spjd * integrated, and also may become zpool properties. 4003185029Spjd */ 4004185029Spjd 4005185029Spjdstatic void 4006185029Spjdl2arc_hdr_stat_add(void) 4007185029Spjd{ 4008185029Spjd ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4009185029Spjd ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4010185029Spjd} 4011185029Spjd 4012185029Spjdstatic void 4013185029Spjdl2arc_hdr_stat_remove(void) 4014185029Spjd{ 4015185029Spjd ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4016185029Spjd ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4017185029Spjd} 4018185029Spjd 4019185029Spjd/* 4020185029Spjd * Cycle through L2ARC devices. This is how L2ARC load balances. 4021185029Spjd * If a device is returned, this also returns holding the spa config lock. 4022185029Spjd */ 4023185029Spjdstatic l2arc_dev_t * 4024185029Spjdl2arc_dev_get_next(void) 4025185029Spjd{ 4026185029Spjd l2arc_dev_t *first, *next = NULL; 4027185029Spjd 4028185029Spjd /* 4029185029Spjd * Lock out the removal of spas (spa_namespace_lock), then removal 4030185029Spjd * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4031185029Spjd * both locks will be dropped and a spa config lock held instead. 4032185029Spjd */ 4033185029Spjd mutex_enter(&spa_namespace_lock); 4034185029Spjd mutex_enter(&l2arc_dev_mtx); 4035185029Spjd 4036185029Spjd /* if there are no vdevs, there is nothing to do */ 4037185029Spjd if (l2arc_ndev == 0) 4038185029Spjd goto out; 4039185029Spjd 4040185029Spjd first = NULL; 4041185029Spjd next = l2arc_dev_last; 4042185029Spjd do { 4043185029Spjd /* loop around the list looking for a non-faulted vdev */ 4044185029Spjd if (next == NULL) { 4045185029Spjd next = list_head(l2arc_dev_list); 4046185029Spjd } else { 4047185029Spjd next = list_next(l2arc_dev_list, next); 4048185029Spjd if (next == NULL) 4049185029Spjd next = list_head(l2arc_dev_list); 4050185029Spjd } 4051185029Spjd 4052185029Spjd /* if we have come back to the start, bail out */ 4053185029Spjd if (first == NULL) 4054185029Spjd first = next; 4055185029Spjd else if (next == first) 4056185029Spjd break; 4057185029Spjd 4058185029Spjd } while (vdev_is_dead(next->l2ad_vdev)); 4059185029Spjd 4060185029Spjd /* if we were unable to find any usable vdevs, return NULL */ 4061185029Spjd if (vdev_is_dead(next->l2ad_vdev)) 4062185029Spjd next = NULL; 4063185029Spjd 4064185029Spjd l2arc_dev_last = next; 4065185029Spjd 4066185029Spjdout: 4067185029Spjd mutex_exit(&l2arc_dev_mtx); 4068185029Spjd 4069185029Spjd /* 4070185029Spjd * Grab the config lock to prevent the 'next' device from being 4071185029Spjd * removed while we are writing to it. 4072185029Spjd */ 4073185029Spjd if (next != NULL) 4074185029Spjd spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4075185029Spjd mutex_exit(&spa_namespace_lock); 4076185029Spjd 4077185029Spjd return (next); 4078185029Spjd} 4079185029Spjd 4080185029Spjd/* 4081185029Spjd * Free buffers that were tagged for destruction. 4082185029Spjd */ 4083185029Spjdstatic void 4084185029Spjdl2arc_do_free_on_write() 4085185029Spjd{ 4086185029Spjd list_t *buflist; 4087185029Spjd l2arc_data_free_t *df, *df_prev; 4088185029Spjd 4089185029Spjd mutex_enter(&l2arc_free_on_write_mtx); 4090185029Spjd buflist = l2arc_free_on_write; 4091185029Spjd 4092185029Spjd for (df = list_tail(buflist); df; df = df_prev) { 4093185029Spjd df_prev = list_prev(buflist, df); 4094185029Spjd ASSERT(df->l2df_data != NULL); 4095185029Spjd ASSERT(df->l2df_func != NULL); 4096185029Spjd df->l2df_func(df->l2df_data, df->l2df_size); 4097185029Spjd list_remove(buflist, df); 4098185029Spjd kmem_free(df, sizeof (l2arc_data_free_t)); 4099185029Spjd } 4100185029Spjd 4101185029Spjd mutex_exit(&l2arc_free_on_write_mtx); 4102185029Spjd} 4103185029Spjd 4104185029Spjd/* 4105185029Spjd * A write to a cache device has completed. Update all headers to allow 4106185029Spjd * reads from these buffers to begin. 4107185029Spjd */ 4108185029Spjdstatic void 4109185029Spjdl2arc_write_done(zio_t *zio) 4110185029Spjd{ 4111185029Spjd l2arc_write_callback_t *cb; 4112185029Spjd l2arc_dev_t *dev; 4113185029Spjd list_t *buflist; 4114185029Spjd arc_buf_hdr_t *head, *ab, *ab_prev; 4115185029Spjd l2arc_buf_hdr_t *abl2; 4116185029Spjd kmutex_t *hash_lock; 4117185029Spjd 4118185029Spjd cb = zio->io_private; 4119185029Spjd ASSERT(cb != NULL); 4120185029Spjd dev = cb->l2wcb_dev; 4121185029Spjd ASSERT(dev != NULL); 4122185029Spjd head = cb->l2wcb_head; 4123185029Spjd ASSERT(head != NULL); 4124185029Spjd buflist = dev->l2ad_buflist; 4125185029Spjd ASSERT(buflist != NULL); 4126185029Spjd DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4127185029Spjd l2arc_write_callback_t *, cb); 4128185029Spjd 4129185029Spjd if (zio->io_error != 0) 4130185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_error); 4131185029Spjd 4132185029Spjd mutex_enter(&l2arc_buflist_mtx); 4133185029Spjd 4134185029Spjd /* 4135185029Spjd * All writes completed, or an error was hit. 4136185029Spjd */ 4137185029Spjd for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 4138185029Spjd ab_prev = list_prev(buflist, ab); 4139185029Spjd 4140185029Spjd hash_lock = HDR_LOCK(ab); 4141185029Spjd if (!mutex_tryenter(hash_lock)) { 4142185029Spjd /* 4143185029Spjd * This buffer misses out. It may be in a stage 4144185029Spjd * of eviction. Its ARC_L2_WRITING flag will be 4145185029Spjd * left set, denying reads to this buffer. 4146185029Spjd */ 4147185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4148185029Spjd continue; 4149185029Spjd } 4150185029Spjd 4151185029Spjd if (zio->io_error != 0) { 4152185029Spjd /* 4153185029Spjd * Error - drop L2ARC entry. 4154185029Spjd */ 4155185029Spjd list_remove(buflist, ab); 4156185029Spjd abl2 = ab->b_l2hdr; 4157185029Spjd ab->b_l2hdr = NULL; 4158185029Spjd kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4159185029Spjd ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4160185029Spjd } 4161185029Spjd 4162185029Spjd /* 4163185029Spjd * Allow ARC to begin reads to this L2ARC entry. 4164185029Spjd */ 4165185029Spjd ab->b_flags &= ~ARC_L2_WRITING; 4166185029Spjd 4167185029Spjd mutex_exit(hash_lock); 4168185029Spjd } 4169185029Spjd 4170185029Spjd atomic_inc_64(&l2arc_writes_done); 4171185029Spjd list_remove(buflist, head); 4172185029Spjd kmem_cache_free(hdr_cache, head); 4173185029Spjd mutex_exit(&l2arc_buflist_mtx); 4174185029Spjd 4175185029Spjd l2arc_do_free_on_write(); 4176185029Spjd 4177185029Spjd kmem_free(cb, sizeof (l2arc_write_callback_t)); 4178185029Spjd} 4179185029Spjd 4180185029Spjd/* 4181185029Spjd * A read to a cache device completed. Validate buffer contents before 4182185029Spjd * handing over to the regular ARC routines. 4183185029Spjd */ 4184185029Spjdstatic void 4185185029Spjdl2arc_read_done(zio_t *zio) 4186185029Spjd{ 4187185029Spjd l2arc_read_callback_t *cb; 4188185029Spjd arc_buf_hdr_t *hdr; 4189185029Spjd arc_buf_t *buf; 4190185029Spjd kmutex_t *hash_lock; 4191185029Spjd int equal; 4192185029Spjd 4193185029Spjd ASSERT(zio->io_vd != NULL); 4194185029Spjd ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4195185029Spjd 4196185029Spjd spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4197185029Spjd 4198185029Spjd cb = zio->io_private; 4199185029Spjd ASSERT(cb != NULL); 4200185029Spjd buf = cb->l2rcb_buf; 4201185029Spjd ASSERT(buf != NULL); 4202185029Spjd hdr = buf->b_hdr; 4203185029Spjd ASSERT(hdr != NULL); 4204185029Spjd 4205185029Spjd hash_lock = HDR_LOCK(hdr); 4206185029Spjd mutex_enter(hash_lock); 4207185029Spjd 4208185029Spjd /* 4209185029Spjd * Check this survived the L2ARC journey. 4210185029Spjd */ 4211185029Spjd equal = arc_cksum_equal(buf); 4212185029Spjd if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4213185029Spjd mutex_exit(hash_lock); 4214185029Spjd zio->io_private = buf; 4215185029Spjd zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4216185029Spjd zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4217185029Spjd arc_read_done(zio); 4218185029Spjd } else { 4219185029Spjd mutex_exit(hash_lock); 4220185029Spjd /* 4221185029Spjd * Buffer didn't survive caching. Increment stats and 4222185029Spjd * reissue to the original storage device. 4223185029Spjd */ 4224185029Spjd if (zio->io_error != 0) { 4225185029Spjd ARCSTAT_BUMP(arcstat_l2_io_error); 4226185029Spjd } else { 4227185029Spjd zio->io_error = EIO; 4228185029Spjd } 4229185029Spjd if (!equal) 4230185029Spjd ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4231185029Spjd 4232185029Spjd /* 4233185029Spjd * If there's no waiter, issue an async i/o to the primary 4234185029Spjd * storage now. If there *is* a waiter, the caller must 4235185029Spjd * issue the i/o in a context where it's OK to block. 4236185029Spjd */ 4237185029Spjd if (zio->io_waiter == NULL) 4238185029Spjd zio_nowait(zio_read(zio->io_parent, 4239185029Spjd cb->l2rcb_spa, &cb->l2rcb_bp, 4240185029Spjd buf->b_data, zio->io_size, arc_read_done, buf, 4241185029Spjd zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4242185029Spjd } 4243185029Spjd 4244185029Spjd kmem_free(cb, sizeof (l2arc_read_callback_t)); 4245185029Spjd} 4246185029Spjd 4247185029Spjd/* 4248185029Spjd * This is the list priority from which the L2ARC will search for pages to 4249185029Spjd * cache. This is used within loops (0..3) to cycle through lists in the 4250185029Spjd * desired order. This order can have a significant effect on cache 4251185029Spjd * performance. 4252185029Spjd * 4253185029Spjd * Currently the metadata lists are hit first, MFU then MRU, followed by 4254185029Spjd * the data lists. This function returns a locked list, and also returns 4255185029Spjd * the lock pointer. 4256185029Spjd */ 4257185029Spjdstatic list_t * 4258185029Spjdl2arc_list_locked(int list_num, kmutex_t **lock) 4259185029Spjd{ 4260185029Spjd list_t *list; 4261205231Skmacy int idx; 4262205231Skmacy 4263205231Skmacy ASSERT(list_num >= 0 && list_num < 2*ARC_BUFC_NUMLISTS); 4264185029Spjd 4265205231Skmacy if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4266205231Skmacy idx = list_num; 4267205231Skmacy list = &arc_mfu->arcs_lists[idx]; 4268205231Skmacy *lock = ARCS_LOCK(arc_mfu, idx); 4269205231Skmacy } else if (list_num < ARC_BUFC_NUMMETADATALISTS*2) { 4270205231Skmacy idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4271205231Skmacy list = &arc_mru->arcs_lists[idx]; 4272205231Skmacy *lock = ARCS_LOCK(arc_mru, idx); 4273205231Skmacy } else if (list_num < (ARC_BUFC_NUMMETADATALISTS*2 + 4274205231Skmacy ARC_BUFC_NUMDATALISTS)) { 4275205231Skmacy idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4276205231Skmacy list = &arc_mfu->arcs_lists[idx]; 4277205231Skmacy *lock = ARCS_LOCK(arc_mfu, idx); 4278205231Skmacy } else { 4279205231Skmacy idx = list_num - ARC_BUFC_NUMLISTS; 4280205231Skmacy list = &arc_mru->arcs_lists[idx]; 4281205231Skmacy *lock = ARCS_LOCK(arc_mru, idx); 4282185029Spjd } 4283185029Spjd 4284205231Skmacy CTR3(KTR_SPARE2, "list=%p list_num=%d idx=%d", 4285205231Skmacy list, list_num, idx); 4286185029Spjd ASSERT(!(MUTEX_HELD(*lock))); 4287185029Spjd mutex_enter(*lock); 4288185029Spjd return (list); 4289185029Spjd} 4290185029Spjd 4291185029Spjd/* 4292185029Spjd * Evict buffers from the device write hand to the distance specified in 4293185029Spjd * bytes. This distance may span populated buffers, it may span nothing. 4294185029Spjd * This is clearing a region on the L2ARC device ready for writing. 4295185029Spjd * If the 'all' boolean is set, every buffer is evicted. 4296185029Spjd */ 4297185029Spjdstatic void 4298185029Spjdl2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4299185029Spjd{ 4300185029Spjd list_t *buflist; 4301185029Spjd l2arc_buf_hdr_t *abl2; 4302185029Spjd arc_buf_hdr_t *ab, *ab_prev; 4303185029Spjd kmutex_t *hash_lock; 4304185029Spjd uint64_t taddr; 4305185029Spjd 4306185029Spjd buflist = dev->l2ad_buflist; 4307185029Spjd 4308185029Spjd if (buflist == NULL) 4309185029Spjd return; 4310185029Spjd 4311185029Spjd if (!all && dev->l2ad_first) { 4312185029Spjd /* 4313185029Spjd * This is the first sweep through the device. There is 4314185029Spjd * nothing to evict. 4315185029Spjd */ 4316185029Spjd return; 4317185029Spjd } 4318185029Spjd 4319185029Spjd if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4320185029Spjd /* 4321185029Spjd * When nearing the end of the device, evict to the end 4322185029Spjd * before the device write hand jumps to the start. 4323185029Spjd */ 4324185029Spjd taddr = dev->l2ad_end; 4325185029Spjd } else { 4326185029Spjd taddr = dev->l2ad_hand + distance; 4327185029Spjd } 4328185029Spjd DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4329185029Spjd uint64_t, taddr, boolean_t, all); 4330185029Spjd 4331185029Spjdtop: 4332185029Spjd mutex_enter(&l2arc_buflist_mtx); 4333185029Spjd for (ab = list_tail(buflist); ab; ab = ab_prev) { 4334185029Spjd ab_prev = list_prev(buflist, ab); 4335185029Spjd 4336185029Spjd hash_lock = HDR_LOCK(ab); 4337185029Spjd if (!mutex_tryenter(hash_lock)) { 4338185029Spjd /* 4339185029Spjd * Missed the hash lock. Retry. 4340185029Spjd */ 4341185029Spjd ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4342185029Spjd mutex_exit(&l2arc_buflist_mtx); 4343185029Spjd mutex_enter(hash_lock); 4344185029Spjd mutex_exit(hash_lock); 4345185029Spjd goto top; 4346185029Spjd } 4347185029Spjd 4348185029Spjd if (HDR_L2_WRITE_HEAD(ab)) { 4349185029Spjd /* 4350185029Spjd * We hit a write head node. Leave it for 4351185029Spjd * l2arc_write_done(). 4352185029Spjd */ 4353185029Spjd list_remove(buflist, ab); 4354185029Spjd mutex_exit(hash_lock); 4355185029Spjd continue; 4356185029Spjd } 4357185029Spjd 4358185029Spjd if (!all && ab->b_l2hdr != NULL && 4359185029Spjd (ab->b_l2hdr->b_daddr > taddr || 4360185029Spjd ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4361185029Spjd /* 4362185029Spjd * We've evicted to the target address, 4363185029Spjd * or the end of the device. 4364185029Spjd */ 4365185029Spjd mutex_exit(hash_lock); 4366185029Spjd break; 4367185029Spjd } 4368185029Spjd 4369185029Spjd if (HDR_FREE_IN_PROGRESS(ab)) { 4370185029Spjd /* 4371185029Spjd * Already on the path to destruction. 4372185029Spjd */ 4373185029Spjd mutex_exit(hash_lock); 4374185029Spjd continue; 4375185029Spjd } 4376185029Spjd 4377185029Spjd if (ab->b_state == arc_l2c_only) { 4378185029Spjd ASSERT(!HDR_L2_READING(ab)); 4379185029Spjd /* 4380185029Spjd * This doesn't exist in the ARC. Destroy. 4381185029Spjd * arc_hdr_destroy() will call list_remove() 4382185029Spjd * and decrement arcstat_l2_size. 4383185029Spjd */ 4384185029Spjd arc_change_state(arc_anon, ab, hash_lock); 4385185029Spjd arc_hdr_destroy(ab); 4386185029Spjd } else { 4387185029Spjd /* 4388185029Spjd * Invalidate issued or about to be issued 4389185029Spjd * reads, since we may be about to write 4390185029Spjd * over this location. 4391185029Spjd */ 4392185029Spjd if (HDR_L2_READING(ab)) { 4393185029Spjd ARCSTAT_BUMP(arcstat_l2_evict_reading); 4394185029Spjd ab->b_flags |= ARC_L2_EVICTED; 4395185029Spjd } 4396185029Spjd 4397185029Spjd /* 4398185029Spjd * Tell ARC this no longer exists in L2ARC. 4399185029Spjd */ 4400185029Spjd if (ab->b_l2hdr != NULL) { 4401185029Spjd abl2 = ab->b_l2hdr; 4402185029Spjd ab->b_l2hdr = NULL; 4403185029Spjd kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4404185029Spjd ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4405185029Spjd } 4406185029Spjd list_remove(buflist, ab); 4407185029Spjd 4408185029Spjd /* 4409185029Spjd * This may have been leftover after a 4410185029Spjd * failed write. 4411185029Spjd */ 4412185029Spjd ab->b_flags &= ~ARC_L2_WRITING; 4413185029Spjd } 4414185029Spjd mutex_exit(hash_lock); 4415185029Spjd } 4416185029Spjd mutex_exit(&l2arc_buflist_mtx); 4417185029Spjd 4418185029Spjd spa_l2cache_space_update(dev->l2ad_vdev, 0, -(taddr - dev->l2ad_evict)); 4419185029Spjd dev->l2ad_evict = taddr; 4420185029Spjd} 4421185029Spjd 4422185029Spjd/* 4423185029Spjd * Find and write ARC buffers to the L2ARC device. 4424185029Spjd * 4425185029Spjd * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 4426185029Spjd * for reading until they have completed writing. 4427185029Spjd */ 4428185029Spjdstatic void 4429185029Spjdl2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) 4430185029Spjd{ 4431185029Spjd arc_buf_hdr_t *ab, *ab_prev, *head; 4432185029Spjd l2arc_buf_hdr_t *hdrl2; 4433185029Spjd list_t *list; 4434185029Spjd uint64_t passed_sz, write_sz, buf_sz, headroom; 4435185029Spjd void *buf_data; 4436185029Spjd kmutex_t *hash_lock, *list_lock; 4437185029Spjd boolean_t have_lock, full; 4438185029Spjd l2arc_write_callback_t *cb; 4439185029Spjd zio_t *pio, *wzio; 4440185029Spjd int try; 4441185029Spjd 4442185029Spjd ASSERT(dev->l2ad_vdev != NULL); 4443185029Spjd 4444185029Spjd pio = NULL; 4445185029Spjd write_sz = 0; 4446185029Spjd full = B_FALSE; 4447185029Spjd head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 4448185029Spjd head->b_flags |= ARC_L2_WRITE_HEAD; 4449185029Spjd 4450205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 4451185029Spjd /* 4452185029Spjd * Copy buffers for L2ARC writing. 4453185029Spjd */ 4454185029Spjd mutex_enter(&l2arc_buflist_mtx); 4455205231Skmacy for (try = 0; try < 2*ARC_BUFC_NUMLISTS; try++) { 4456185029Spjd list = l2arc_list_locked(try, &list_lock); 4457185029Spjd passed_sz = 0; 4458205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 4459185029Spjd 4460185029Spjd /* 4461185029Spjd * L2ARC fast warmup. 4462185029Spjd * 4463185029Spjd * Until the ARC is warm and starts to evict, read from the 4464185029Spjd * head of the ARC lists rather than the tail. 4465185029Spjd */ 4466185029Spjd headroom = target_sz * l2arc_headroom; 4467185029Spjd if (arc_warm == B_FALSE) 4468185029Spjd ab = list_head(list); 4469185029Spjd else 4470185029Spjd ab = list_tail(list); 4471205231Skmacy if (ab == NULL) { 4472205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 4473205231Skmacy } 4474185029Spjd 4475185029Spjd for (; ab; ab = ab_prev) { 4476185029Spjd if (arc_warm == B_FALSE) 4477185029Spjd ab_prev = list_next(list, ab); 4478185029Spjd else 4479185029Spjd ab_prev = list_prev(list, ab); 4480205231Skmacy ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size); 4481205231Skmacy 4482185029Spjd hash_lock = HDR_LOCK(ab); 4483185029Spjd have_lock = MUTEX_HELD(hash_lock); 4484185029Spjd if (!have_lock && !mutex_tryenter(hash_lock)) { 4485205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 4486185029Spjd /* 4487185029Spjd * Skip this buffer rather than waiting. 4488185029Spjd */ 4489185029Spjd continue; 4490185029Spjd } 4491185029Spjd 4492205231Skmacy if (ab->b_l2hdr != NULL) { 4493205231Skmacy /* 4494205231Skmacy * Already in L2ARC. 4495205231Skmacy */ 4496205231Skmacy mutex_exit(hash_lock); 4497205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4498205231Skmacy continue; 4499205231Skmacy } 4500205231Skmacy 4501185029Spjd passed_sz += ab->b_size; 4502185029Spjd if (passed_sz > headroom) { 4503185029Spjd /* 4504185029Spjd * Searched too far. 4505185029Spjd */ 4506185029Spjd mutex_exit(hash_lock); 4507205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 4508185029Spjd break; 4509185029Spjd } 4510185029Spjd 4511185029Spjd if (ab->b_spa != spa) { 4512185029Spjd mutex_exit(hash_lock); 4513205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4514185029Spjd continue; 4515185029Spjd } 4516185029Spjd 4517205231Skmacy if (HDR_IO_IN_PROGRESS(ab)) { 4518185029Spjd mutex_exit(hash_lock); 4519205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4520185029Spjd continue; 4521185029Spjd } 4522205231Skmacy if (!HDR_L2CACHE(ab)) { 4523185029Spjd mutex_exit(hash_lock); 4524205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4525185029Spjd continue; 4526185029Spjd } 4527185029Spjd if ((write_sz + ab->b_size) > target_sz) { 4528185029Spjd full = B_TRUE; 4529185029Spjd mutex_exit(hash_lock); 4530205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_full); 4531185029Spjd break; 4532185029Spjd } 4533185029Spjd 4534185029Spjd if (ab->b_buf == NULL) { 4535185029Spjd DTRACE_PROBE1(l2arc__buf__null, void *, ab); 4536185029Spjd mutex_exit(hash_lock); 4537185029Spjd continue; 4538185029Spjd } 4539185029Spjd 4540185029Spjd if (pio == NULL) { 4541185029Spjd /* 4542185029Spjd * Insert a dummy header on the buflist so 4543185029Spjd * l2arc_write_done() can find where the 4544185029Spjd * write buffers begin without searching. 4545185029Spjd */ 4546185029Spjd list_insert_head(dev->l2ad_buflist, head); 4547185029Spjd 4548185029Spjd cb = kmem_alloc( 4549185029Spjd sizeof (l2arc_write_callback_t), KM_SLEEP); 4550185029Spjd cb->l2wcb_dev = dev; 4551185029Spjd cb->l2wcb_head = head; 4552185029Spjd pio = zio_root(spa, l2arc_write_done, cb, 4553185029Spjd ZIO_FLAG_CANFAIL); 4554205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_pios); 4555185029Spjd } 4556185029Spjd 4557205231Skmacy ARCSTAT_INCR(arcstat_l2_write_bytes_written, ab->b_size); 4558185029Spjd /* 4559185029Spjd * Create and add a new L2ARC header. 4560185029Spjd */ 4561185029Spjd hdrl2 = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 4562185029Spjd hdrl2->b_dev = dev; 4563185029Spjd hdrl2->b_daddr = dev->l2ad_hand; 4564185029Spjd 4565206792Spjd ab->b_flags |= ARC_L2_WRITING; 4566185029Spjd ab->b_l2hdr = hdrl2; 4567185029Spjd list_insert_head(dev->l2ad_buflist, ab); 4568185029Spjd buf_data = ab->b_buf->b_data; 4569185029Spjd buf_sz = ab->b_size; 4570185029Spjd 4571185029Spjd /* 4572185029Spjd * Compute and store the buffer cksum before 4573185029Spjd * writing. On debug the cksum is verified first. 4574185029Spjd */ 4575185029Spjd arc_cksum_verify(ab->b_buf); 4576185029Spjd arc_cksum_compute(ab->b_buf, B_TRUE); 4577185029Spjd 4578185029Spjd mutex_exit(hash_lock); 4579185029Spjd 4580185029Spjd wzio = zio_write_phys(pio, dev->l2ad_vdev, 4581185029Spjd dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 4582185029Spjd NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 4583185029Spjd ZIO_FLAG_CANFAIL, B_FALSE); 4584185029Spjd 4585185029Spjd DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 4586185029Spjd zio_t *, wzio); 4587185029Spjd (void) zio_nowait(wzio); 4588185029Spjd 4589185029Spjd /* 4590185029Spjd * Keep the clock hand suitably device-aligned. 4591185029Spjd */ 4592185029Spjd buf_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 4593185029Spjd 4594185029Spjd write_sz += buf_sz; 4595185029Spjd dev->l2ad_hand += buf_sz; 4596185029Spjd } 4597185029Spjd 4598185029Spjd mutex_exit(list_lock); 4599185029Spjd 4600185029Spjd if (full == B_TRUE) 4601185029Spjd break; 4602185029Spjd } 4603185029Spjd mutex_exit(&l2arc_buflist_mtx); 4604185029Spjd 4605185029Spjd if (pio == NULL) { 4606185029Spjd ASSERT3U(write_sz, ==, 0); 4607185029Spjd kmem_cache_free(hdr_cache, head); 4608185029Spjd return; 4609185029Spjd } 4610185029Spjd 4611185029Spjd ASSERT3U(write_sz, <=, target_sz); 4612185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_sent); 4613185029Spjd ARCSTAT_INCR(arcstat_l2_size, write_sz); 4614185029Spjd spa_l2cache_space_update(dev->l2ad_vdev, 0, write_sz); 4615185029Spjd 4616185029Spjd /* 4617185029Spjd * Bump device hand to the device start if it is approaching the end. 4618185029Spjd * l2arc_evict() will already have evicted ahead for this case. 4619185029Spjd */ 4620185029Spjd if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 4621185029Spjd spa_l2cache_space_update(dev->l2ad_vdev, 0, 4622185029Spjd dev->l2ad_end - dev->l2ad_hand); 4623185029Spjd dev->l2ad_hand = dev->l2ad_start; 4624185029Spjd dev->l2ad_evict = dev->l2ad_start; 4625185029Spjd dev->l2ad_first = B_FALSE; 4626185029Spjd } 4627185029Spjd 4628185029Spjd (void) zio_wait(pio); 4629185029Spjd} 4630185029Spjd 4631185029Spjd/* 4632185029Spjd * This thread feeds the L2ARC at regular intervals. This is the beating 4633185029Spjd * heart of the L2ARC. 4634185029Spjd */ 4635185029Spjdstatic void 4636185029Spjdl2arc_feed_thread(void *dummy __unused) 4637185029Spjd{ 4638185029Spjd callb_cpr_t cpr; 4639185029Spjd l2arc_dev_t *dev; 4640185029Spjd spa_t *spa; 4641185029Spjd uint64_t size; 4642185029Spjd 4643185029Spjd CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 4644185029Spjd 4645185029Spjd mutex_enter(&l2arc_feed_thr_lock); 4646185029Spjd 4647185029Spjd while (l2arc_thread_exit == 0) { 4648185029Spjd /* 4649185029Spjd * Pause for l2arc_feed_secs seconds between writes. 4650185029Spjd */ 4651185029Spjd CALLB_CPR_SAFE_BEGIN(&cpr); 4652185029Spjd (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 4653205231Skmacy hz * l2arc_feed_secs >> l2arc_feed_secs_shift); 4654185029Spjd CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 4655185029Spjd 4656185029Spjd /* 4657185029Spjd * Quick check for L2ARC devices. 4658185029Spjd */ 4659185029Spjd mutex_enter(&l2arc_dev_mtx); 4660185029Spjd if (l2arc_ndev == 0) { 4661185029Spjd mutex_exit(&l2arc_dev_mtx); 4662185029Spjd continue; 4663185029Spjd } 4664185029Spjd mutex_exit(&l2arc_dev_mtx); 4665185029Spjd 4666185029Spjd /* 4667185029Spjd * This selects the next l2arc device to write to, and in 4668185029Spjd * doing so the next spa to feed from: dev->l2ad_spa. This 4669185029Spjd * will return NULL if there are now no l2arc devices or if 4670185029Spjd * they are all faulted. 4671185029Spjd * 4672185029Spjd * If a device is returned, its spa's config lock is also 4673185029Spjd * held to prevent device removal. l2arc_dev_get_next() 4674185029Spjd * will grab and release l2arc_dev_mtx. 4675185029Spjd */ 4676185029Spjd if ((dev = l2arc_dev_get_next()) == NULL) 4677185029Spjd continue; 4678185029Spjd 4679185029Spjd spa = dev->l2ad_spa; 4680185029Spjd ASSERT(spa != NULL); 4681185029Spjd 4682185029Spjd /* 4683185029Spjd * Avoid contributing to memory pressure. 4684185029Spjd */ 4685185029Spjd if (arc_reclaim_needed()) { 4686185029Spjd ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 4687185029Spjd spa_config_exit(spa, SCL_L2ARC, dev); 4688185029Spjd continue; 4689185029Spjd } 4690185029Spjd 4691185029Spjd ARCSTAT_BUMP(arcstat_l2_feeds); 4692185029Spjd 4693185029Spjd size = dev->l2ad_write; 4694185029Spjd if (arc_warm == B_FALSE) 4695185029Spjd size += dev->l2ad_boost; 4696185029Spjd 4697185029Spjd /* 4698185029Spjd * Evict L2ARC buffers that will be overwritten. 4699185029Spjd */ 4700185029Spjd l2arc_evict(dev, size, B_FALSE); 4701185029Spjd 4702185029Spjd /* 4703185029Spjd * Write ARC buffers. 4704185029Spjd */ 4705185029Spjd l2arc_write_buffers(spa, dev, size); 4706185029Spjd spa_config_exit(spa, SCL_L2ARC, dev); 4707185029Spjd } 4708185029Spjd 4709185029Spjd l2arc_thread_exit = 0; 4710185029Spjd cv_broadcast(&l2arc_feed_thr_cv); 4711185029Spjd CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 4712185029Spjd thread_exit(); 4713185029Spjd} 4714185029Spjd 4715185029Spjdboolean_t 4716185029Spjdl2arc_vdev_present(vdev_t *vd) 4717185029Spjd{ 4718185029Spjd l2arc_dev_t *dev; 4719185029Spjd 4720185029Spjd mutex_enter(&l2arc_dev_mtx); 4721185029Spjd for (dev = list_head(l2arc_dev_list); dev != NULL; 4722185029Spjd dev = list_next(l2arc_dev_list, dev)) { 4723185029Spjd if (dev->l2ad_vdev == vd) 4724185029Spjd break; 4725185029Spjd } 4726185029Spjd mutex_exit(&l2arc_dev_mtx); 4727185029Spjd 4728185029Spjd return (dev != NULL); 4729185029Spjd} 4730185029Spjd 4731185029Spjd/* 4732185029Spjd * Add a vdev for use by the L2ARC. By this point the spa has already 4733185029Spjd * validated the vdev and opened it. 4734185029Spjd */ 4735185029Spjdvoid 4736185029Spjdl2arc_add_vdev(spa_t *spa, vdev_t *vd, uint64_t start, uint64_t end) 4737185029Spjd{ 4738185029Spjd l2arc_dev_t *adddev; 4739185029Spjd 4740185029Spjd ASSERT(!l2arc_vdev_present(vd)); 4741185029Spjd 4742185029Spjd /* 4743185029Spjd * Create a new l2arc device entry. 4744185029Spjd */ 4745185029Spjd adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 4746185029Spjd adddev->l2ad_spa = spa; 4747185029Spjd adddev->l2ad_vdev = vd; 4748185029Spjd adddev->l2ad_write = l2arc_write_max; 4749185029Spjd adddev->l2ad_boost = l2arc_write_boost; 4750185029Spjd adddev->l2ad_start = start; 4751185029Spjd adddev->l2ad_end = end; 4752185029Spjd adddev->l2ad_hand = adddev->l2ad_start; 4753185029Spjd adddev->l2ad_evict = adddev->l2ad_start; 4754185029Spjd adddev->l2ad_first = B_TRUE; 4755185029Spjd ASSERT3U(adddev->l2ad_write, >, 0); 4756185029Spjd 4757185029Spjd /* 4758185029Spjd * This is a list of all ARC buffers that are still valid on the 4759185029Spjd * device. 4760185029Spjd */ 4761185029Spjd adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 4762185029Spjd list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 4763185029Spjd offsetof(arc_buf_hdr_t, b_l2node)); 4764185029Spjd 4765185029Spjd spa_l2cache_space_update(vd, adddev->l2ad_end - adddev->l2ad_hand, 0); 4766185029Spjd 4767185029Spjd /* 4768185029Spjd * Add device to global list 4769185029Spjd */ 4770185029Spjd mutex_enter(&l2arc_dev_mtx); 4771185029Spjd list_insert_head(l2arc_dev_list, adddev); 4772185029Spjd atomic_inc_64(&l2arc_ndev); 4773185029Spjd mutex_exit(&l2arc_dev_mtx); 4774185029Spjd} 4775185029Spjd 4776185029Spjd/* 4777185029Spjd * Remove a vdev from the L2ARC. 4778185029Spjd */ 4779185029Spjdvoid 4780185029Spjdl2arc_remove_vdev(vdev_t *vd) 4781185029Spjd{ 4782185029Spjd l2arc_dev_t *dev, *nextdev, *remdev = NULL; 4783185029Spjd 4784185029Spjd /* 4785185029Spjd * Find the device by vdev 4786185029Spjd */ 4787185029Spjd mutex_enter(&l2arc_dev_mtx); 4788185029Spjd for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 4789185029Spjd nextdev = list_next(l2arc_dev_list, dev); 4790185029Spjd if (vd == dev->l2ad_vdev) { 4791185029Spjd remdev = dev; 4792185029Spjd break; 4793185029Spjd } 4794185029Spjd } 4795185029Spjd ASSERT(remdev != NULL); 4796185029Spjd 4797185029Spjd /* 4798185029Spjd * Remove device from global list 4799185029Spjd */ 4800185029Spjd list_remove(l2arc_dev_list, remdev); 4801185029Spjd l2arc_dev_last = NULL; /* may have been invalidated */ 4802185029Spjd atomic_dec_64(&l2arc_ndev); 4803185029Spjd mutex_exit(&l2arc_dev_mtx); 4804185029Spjd 4805185029Spjd /* 4806185029Spjd * Clear all buflists and ARC references. L2ARC device flush. 4807185029Spjd */ 4808185029Spjd l2arc_evict(remdev, 0, B_TRUE); 4809185029Spjd list_destroy(remdev->l2ad_buflist); 4810185029Spjd kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 4811185029Spjd kmem_free(remdev, sizeof (l2arc_dev_t)); 4812185029Spjd} 4813185029Spjd 4814185029Spjdvoid 4815185029Spjdl2arc_init(void) 4816185029Spjd{ 4817185029Spjd l2arc_thread_exit = 0; 4818185029Spjd l2arc_ndev = 0; 4819185029Spjd l2arc_writes_sent = 0; 4820185029Spjd l2arc_writes_done = 0; 4821185029Spjd 4822185029Spjd mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4823185029Spjd cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 4824185029Spjd mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 4825185029Spjd mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 4826185029Spjd mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 4827185029Spjd 4828185029Spjd l2arc_dev_list = &L2ARC_dev_list; 4829185029Spjd l2arc_free_on_write = &L2ARC_free_on_write; 4830185029Spjd list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 4831185029Spjd offsetof(l2arc_dev_t, l2ad_node)); 4832185029Spjd list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 4833185029Spjd offsetof(l2arc_data_free_t, l2df_list_node)); 4834185029Spjd} 4835185029Spjd 4836185029Spjdvoid 4837185029Spjdl2arc_fini(void) 4838185029Spjd{ 4839185029Spjd /* 4840185029Spjd * This is called from dmu_fini(), which is called from spa_fini(); 4841185029Spjd * Because of this, we can assume that all l2arc devices have 4842185029Spjd * already been removed when the pools themselves were removed. 4843185029Spjd */ 4844185029Spjd 4845185029Spjd l2arc_do_free_on_write(); 4846185029Spjd 4847185029Spjd mutex_destroy(&l2arc_feed_thr_lock); 4848185029Spjd cv_destroy(&l2arc_feed_thr_cv); 4849185029Spjd mutex_destroy(&l2arc_dev_mtx); 4850185029Spjd mutex_destroy(&l2arc_buflist_mtx); 4851185029Spjd mutex_destroy(&l2arc_free_on_write_mtx); 4852185029Spjd 4853185029Spjd list_destroy(l2arc_dev_list); 4854185029Spjd list_destroy(l2arc_free_on_write); 4855185029Spjd} 4856185029Spjd 4857185029Spjdvoid 4858185029Spjdl2arc_start(void) 4859185029Spjd{ 4860185029Spjd if (!(spa_mode & FWRITE)) 4861185029Spjd return; 4862185029Spjd 4863185029Spjd (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 4864185029Spjd TS_RUN, minclsyspri); 4865185029Spjd} 4866185029Spjd 4867185029Spjdvoid 4868185029Spjdl2arc_stop(void) 4869185029Spjd{ 4870185029Spjd if (!(spa_mode & FWRITE)) 4871185029Spjd return; 4872185029Spjd 4873185029Spjd mutex_enter(&l2arc_feed_thr_lock); 4874185029Spjd cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 4875185029Spjd l2arc_thread_exit = 1; 4876185029Spjd while (l2arc_thread_exit != 0) 4877185029Spjd cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 4878185029Spjd mutex_exit(&l2arc_feed_thr_lock); 4879185029Spjd} 4880