arc.c revision 241773
1145519Sdarrenr/* 2145510Sdarrenr * CDDL HEADER START 3145510Sdarrenr * 4255332Scy * The contents of this file are subject to the terms of the 5145510Sdarrenr * Common Development and Distribution License (the "License"). 6145510Sdarrenr * You may not use this file except in compliance with the License. 7145510Sdarrenr * 8145510Sdarrenr * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9145510Sdarrenr * or http://www.opensolaris.org/os/licensing. 10145510Sdarrenr * See the License for the specific language governing permissions 11145510Sdarrenr * and limitations under the License. 12145510Sdarrenr * 13145510Sdarrenr * When distributing Covered Code, include this CDDL HEADER in each 14145510Sdarrenr * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15145510Sdarrenr * If applicable, add the following below this CDDL HEADER, with the 16145510Sdarrenr * fields enclosed by brackets "[]" replaced with your own identifying 17145510Sdarrenr * information: Portions Copyright [yyyy] [name of copyright owner] 18145510Sdarrenr * 19145510Sdarrenr * CDDL HEADER END 20145510Sdarrenr */ 21145510Sdarrenr/* 22145510Sdarrenr * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23145510Sdarrenr * Copyright 2011 Nexenta Systems, Inc. All rights reserved. 24145510Sdarrenr * Copyright (c) 2011 by Delphix. All rights reserved. 25145510Sdarrenr */ 26145510Sdarrenr 27170268Sdarrenr/* 28145510Sdarrenr * DVA-based Adjustable Replacement Cache 29145510Sdarrenr * 30170268Sdarrenr * While much of the theory of operation used here is 31145510Sdarrenr * based on the self-tuning, low overhead replacement cache 32145510Sdarrenr * presented by Megiddo and Modha at FAST 2003, there are some 33145510Sdarrenr * significant differences: 34145510Sdarrenr * 35145510Sdarrenr * 1. The Megiddo and Modha model assumes any page is evictable. 36145510Sdarrenr * Pages in its cache cannot be "locked" into memory. This makes 37145510Sdarrenr * the eviction algorithm simple: evict the last page in the list. 38145510Sdarrenr * This also make the performance characteristics easy to reason 39145510Sdarrenr * about. Our cache is not so simple. At any given moment, some 40145510Sdarrenr * subset of the blocks in the cache are un-evictable because we 41145510Sdarrenr * have handed out a reference to them. Blocks are only evictable 42145510Sdarrenr * when there are no external references active. This makes 43145510Sdarrenr * eviction far more problematic: we choose to evict the evictable 44145510Sdarrenr * blocks that are the "lowest" in the list. 45145510Sdarrenr * 46145510Sdarrenr * There are times when it is not possible to evict the requested 47145510Sdarrenr * space. In these circumstances we are unable to adjust the cache 48145510Sdarrenr * size. To prevent the cache growing unbounded at these times we 49170268Sdarrenr * implement a "cache throttle" that slows the flow of new data 50170268Sdarrenr * into the cache until we can make space available. 51145510Sdarrenr * 52145510Sdarrenr * 2. The Megiddo and Modha model assumes a fixed cache size. 53145510Sdarrenr * Pages are evicted when the cache is full and there is a cache 54145510Sdarrenr * miss. Our model has a variable sized cache. It grows with 55255332Scy * high use, but also tries to react to memory pressure from the 56255332Scy * operating system: decreasing its size when system memory is 57170268Sdarrenr * tight. 58255332Scy * 59145510Sdarrenr * 3. The Megiddo and Modha model assumes a fixed page size. All 60145510Sdarrenr * elements of the cache are therefor exactly the same size. So 61145510Sdarrenr * when adjusting the cache size following a cache miss, its simply 62145510Sdarrenr * a matter of choosing a single page to evict. In our model, we 63255332Scy * have variable sized cache blocks (rangeing from 512 bytes to 64255332Scy * 128K bytes). We therefor choose a set of blocks to evict to make 65145510Sdarrenr * space for a cache miss that approximates as closely as possible 66145510Sdarrenr * the space used by the new block. 67255332Scy * 68255332Scy * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 69255332Scy * by N. Megiddo & D. Modha, FAST 2003 70145510Sdarrenr */ 71145510Sdarrenr 72255332Scy/* 73255332Scy * The locking model: 74255332Scy * 75255332Scy * A new reference to a cache buffer can be obtained in two 76353076Scy * ways: 1) via a hash table lookup using the DVA as a key, 77255332Scy * or 2) via one of the ARC lists. The arc_read() interface 78255332Scy * uses method 1, while the internal arc algorithms for 79255332Scy * adjusting the cache use method 2. We therefor provide two 80145510Sdarrenr * types of locks: 1) the hash table lock array, and 2) the 81145510Sdarrenr * arc list locks. 82145510Sdarrenr * 83145510Sdarrenr * Buffers do not have their own mutexs, rather they rely on the 84255332Scy * hash table mutexs for the bulk of their protection (i.e. most 85255332Scy * fields in the arc_buf_hdr_t are protected by these mutexs). 86255332Scy * 87255332Scy * buf_hash_find() returns the appropriate mutex (held) when it 88145510Sdarrenr * locates the requested buffer in the hash table. It returns 89255332Scy * NULL for the mutex if the buffer was not in the table. 90145510Sdarrenr * 91145510Sdarrenr * buf_hash_remove() expects the appropriate hash mutex to be 92145510Sdarrenr * already held before it is invoked. 93145510Sdarrenr * 94255332Scy * Each arc state also has a mutex which is used to protect the 95255332Scy * buffer list associated with the state. When attempting to 96255332Scy * obtain a hash table lock while holding an arc list lock you 97145510Sdarrenr * must use: mutex_tryenter() to avoid deadlock. Also note that 98145510Sdarrenr * the active state mutex must be held before the ghost state mutex. 99145510Sdarrenr * 100145510Sdarrenr * Arc buffers may have an associated eviction callback function. 101145510Sdarrenr * This function will be invoked prior to removing the buffer (e.g. 102145510Sdarrenr * in arc_do_user_evicts()). Note however that the data associated 103145510Sdarrenr * with the buffer may be evicted prior to the callback. The callback 104145510Sdarrenr * must be made with *no locks held* (to prevent deadlock). Additionally, 105145510Sdarrenr * the users of callbacks must ensure that their private data is 106145510Sdarrenr * protected from simultaneous callbacks from arc_buf_evict() 107145510Sdarrenr * and arc_do_user_evicts(). 108145510Sdarrenr * 109145510Sdarrenr * Note that the majority of the performance stats are manipulated 110145510Sdarrenr * with atomic operations. 111145510Sdarrenr * 112145510Sdarrenr * The L2ARC uses the l2arc_buflist_mtx global mutex for the following: 113255332Scy * 114255332Scy * - L2ARC buflist creation 115255332Scy * - L2ARC buflist eviction 116145510Sdarrenr * - L2ARC write completion, which walks L2ARC buflists 117145510Sdarrenr * - ARC header destruction, as it removes from L2ARC buflists 118145510Sdarrenr * - ARC header release, as it removes from L2ARC buflists 119145510Sdarrenr */ 120145510Sdarrenr 121145510Sdarrenr#include <sys/spa.h> 122145510Sdarrenr#include <sys/zio.h> 123145510Sdarrenr#include <sys/zfs_context.h> 124145510Sdarrenr#include <sys/arc.h> 125255332Scy#include <sys/refcount.h> 126255332Scy#include <sys/vdev.h> 127255332Scy#include <sys/vdev_impl.h> 128145510Sdarrenr#ifdef _KERNEL 129145510Sdarrenr#include <sys/dnlc.h> 130145510Sdarrenr#endif 131145510Sdarrenr#include <sys/callb.h> 132170268Sdarrenr#include <sys/kstat.h> 133170268Sdarrenr#include <zfs_fletcher.h> 134170268Sdarrenr#include <sys/sdt.h> 135145510Sdarrenr 136145510Sdarrenr#include <vm/vm_pageout.h> 137145510Sdarrenr 138255332Scy#ifdef illumos 139255332Scy#ifndef _KERNEL 140255332Scy/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 141255332Scyboolean_t arc_watch = B_FALSE; 142145510Sdarrenrint arc_procfd; 143255332Scy#endif 144170268Sdarrenr#endif /* illumos */ 145255332Scy 146255332Scystatic kmutex_t arc_reclaim_thr_lock; 147255332Scystatic kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 148145510Sdarrenrstatic uint8_t arc_thread_exit; 149145510Sdarrenr 150145510Sdarrenrextern int zfs_write_limit_shift; 151255332Scyextern uint64_t zfs_write_limit_max; 152255332Scyextern kmutex_t zfs_write_limit_lock; 153145510Sdarrenr 154255332Scy#define ARC_REDUCE_DNLC_PERCENT 3 155145510Sdarrenruint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 156145510Sdarrenr 157145510Sdarrenrtypedef enum arc_reclaim_strategy { 158145510Sdarrenr ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 159145510Sdarrenr ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 160145510Sdarrenr} arc_reclaim_strategy_t; 161145510Sdarrenr 162255332Scy/* number of seconds before growing cache again */ 163170268Sdarrenrstatic int arc_grow_retry = 60; 164145510Sdarrenr 165145510Sdarrenr/* shift of arc_c for calculating both min and max arc_p */ 166145510Sdarrenrstatic int arc_p_min_shift = 4; 167145510Sdarrenr 168145510Sdarrenr/* log2(fraction of arc to reclaim) */ 169255332Scystatic int arc_shrink_shift = 5; 170145510Sdarrenr 171145510Sdarrenr/* 172255332Scy * minimum lifespan of a prefetch block in clock ticks 173255332Scy * (initialized in arc_init()) 174255332Scy */ 175255332Scystatic int arc_min_prefetch_lifespan; 176255332Scy 177145510Sdarrenrstatic int arc_dead; 178145510Sdarrenrextern int zfs_prefetch_disable; 179145510Sdarrenr 180145510Sdarrenr/* 181145510Sdarrenr * The arc has filled available memory and has now warmed up. 182145510Sdarrenr */ 183145510Sdarrenrstatic boolean_t arc_warm; 184255332Scy 185255332Scy/* 186255332Scy * These tunables are for performance analysis. 187255332Scy */ 188255332Scyuint64_t zfs_arc_max; 189255332Scyuint64_t zfs_arc_min; 190255332Scyuint64_t zfs_arc_meta_limit = 0; 191255332Scyint zfs_arc_grow_retry = 0; 192255332Scyint zfs_arc_shrink_shift = 0; 193255332Scyint zfs_arc_p_min_shift = 0; 194255332Scy 195255332ScyTUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max); 196255332ScyTUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 197255332ScyTUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 198255332ScySYSCTL_DECL(_vfs_zfs); 199255332ScySYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 200255332Scy "Maximum ARC size"); 201255332ScySYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 202255332Scy "Minimum ARC size"); 203255332Scy 204255332Scy/* 205255332Scy * Note that buffers can be in one of 6 states: 206255332Scy * ARC_anon - anonymous (discussed below) 207255332Scy * ARC_mru - recently used, currently cached 208255332Scy * ARC_mru_ghost - recentely used, no longer in cache 209255332Scy * ARC_mfu - frequently used, currently cached 210145510Sdarrenr * ARC_mfu_ghost - frequently used, no longer in cache 211145510Sdarrenr * ARC_l2c_only - exists in L2ARC but not other states 212145510Sdarrenr * When there are no active references to the buffer, they are 213145510Sdarrenr * are linked onto a list in one of these arc states. These are 214145510Sdarrenr * the only buffers that can be evicted or deleted. Within each 215170268Sdarrenr * state there are multiple lists, one for meta-data and one for 216255332Scy * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 217170268Sdarrenr * etc.) is tracked separately so that it can be managed more 218170268Sdarrenr * explicitly: favored over data, limited explicitly. 219170268Sdarrenr * 220145510Sdarrenr * Anonymous buffers are buffers that are not associated with 221145510Sdarrenr * a DVA. These are buffers that hold dirty block copies 222145510Sdarrenr * before they are written to stable storage. By definition, 223170268Sdarrenr * they are "ref'd" and are considered part of arc_mru 224170268Sdarrenr * that cannot be freed. Generally, they will aquire a DVA 225145510Sdarrenr * as they are written and migrate onto the arc_mru list. 226170268Sdarrenr * 227170268Sdarrenr * The ARC_l2c_only state is for buffers that are in the second 228145510Sdarrenr * level ARC but no longer in any of the ARC_m* lists. The second 229145510Sdarrenr * level ARC itself may also contain buffers that are in any of 230145510Sdarrenr * the ARC_m* states - meaning that a buffer can exist in two 231145510Sdarrenr * places. The reason for the ARC_l2c_only state is to keep the 232145510Sdarrenr * buffer header in the hash table, so that reads that hit the 233255332Scy * second level ARC benefit from these fast lookups. 234255332Scy */ 235255332Scy 236255332Scy#define ARCS_LOCK_PAD CACHE_LINE_SIZE 237255332Scystruct arcs_lock { 238255332Scy kmutex_t arcs_lock; 239255332Scy#ifdef _KERNEL 240255332Scy unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 241255332Scy#endif 242255332Scy}; 243255332Scy 244255332Scy/* 245255332Scy * must be power of two for mask use to work 246255332Scy * 247255332Scy */ 248255332Scy#define ARC_BUFC_NUMDATALISTS 16 249145510Sdarrenr#define ARC_BUFC_NUMMETADATALISTS 16 250145510Sdarrenr#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 251145510Sdarrenr 252145510Sdarrenrtypedef struct arc_state { 253255332Scy uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 254255332Scy uint64_t arcs_size; /* total amount of data in this state */ 255255332Scy list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 256255332Scy struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 257145510Sdarrenr} arc_state_t; 258145510Sdarrenr 259145510Sdarrenr#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 260145510Sdarrenr 261145510Sdarrenr/* The 6 states: */ 262145510Sdarrenrstatic arc_state_t ARC_anon; 263145510Sdarrenrstatic arc_state_t ARC_mru; 264145510Sdarrenrstatic arc_state_t ARC_mru_ghost; 265145510Sdarrenrstatic arc_state_t ARC_mfu; 266145510Sdarrenrstatic arc_state_t ARC_mfu_ghost; 267145510Sdarrenrstatic arc_state_t ARC_l2c_only; 268145510Sdarrenr 269145510Sdarrenrtypedef struct arc_stats { 270145510Sdarrenr kstat_named_t arcstat_hits; 271255332Scy kstat_named_t arcstat_misses; 272145510Sdarrenr kstat_named_t arcstat_demand_data_hits; 273145510Sdarrenr kstat_named_t arcstat_demand_data_misses; 274145510Sdarrenr kstat_named_t arcstat_demand_metadata_hits; 275145510Sdarrenr kstat_named_t arcstat_demand_metadata_misses; 276145510Sdarrenr kstat_named_t arcstat_prefetch_data_hits; 277145510Sdarrenr kstat_named_t arcstat_prefetch_data_misses; 278145510Sdarrenr kstat_named_t arcstat_prefetch_metadata_hits; 279145510Sdarrenr kstat_named_t arcstat_prefetch_metadata_misses; 280145510Sdarrenr kstat_named_t arcstat_mru_hits; 281145510Sdarrenr kstat_named_t arcstat_mru_ghost_hits; 282255332Scy kstat_named_t arcstat_mfu_hits; 283145510Sdarrenr kstat_named_t arcstat_mfu_ghost_hits; 284145510Sdarrenr kstat_named_t arcstat_allocated; 285145510Sdarrenr kstat_named_t arcstat_deleted; 286145510Sdarrenr kstat_named_t arcstat_stolen; 287145510Sdarrenr kstat_named_t arcstat_recycle_miss; 288145510Sdarrenr kstat_named_t arcstat_mutex_miss; 289145510Sdarrenr kstat_named_t arcstat_evict_skip; 290145510Sdarrenr kstat_named_t arcstat_evict_l2_cached; 291145510Sdarrenr kstat_named_t arcstat_evict_l2_eligible; 292145510Sdarrenr kstat_named_t arcstat_evict_l2_ineligible; 293145510Sdarrenr kstat_named_t arcstat_hash_elements; 294145510Sdarrenr kstat_named_t arcstat_hash_elements_max; 295145510Sdarrenr kstat_named_t arcstat_hash_collisions; 296145510Sdarrenr kstat_named_t arcstat_hash_chains; 297145510Sdarrenr kstat_named_t arcstat_hash_chain_max; 298145510Sdarrenr kstat_named_t arcstat_p; 299145510Sdarrenr kstat_named_t arcstat_c; 300145510Sdarrenr kstat_named_t arcstat_c_min; 301145510Sdarrenr kstat_named_t arcstat_c_max; 302145510Sdarrenr kstat_named_t arcstat_size; 303145510Sdarrenr kstat_named_t arcstat_hdr_size; 304145510Sdarrenr kstat_named_t arcstat_data_size; 305145510Sdarrenr kstat_named_t arcstat_other_size; 306145510Sdarrenr kstat_named_t arcstat_l2_hits; 307145510Sdarrenr kstat_named_t arcstat_l2_misses; 308145510Sdarrenr kstat_named_t arcstat_l2_feeds; 309145510Sdarrenr kstat_named_t arcstat_l2_rw_clash; 310255332Scy kstat_named_t arcstat_l2_read_bytes; 311255332Scy kstat_named_t arcstat_l2_write_bytes; 312255332Scy kstat_named_t arcstat_l2_writes_sent; 313255332Scy kstat_named_t arcstat_l2_writes_done; 314255332Scy kstat_named_t arcstat_l2_writes_error; 315255332Scy kstat_named_t arcstat_l2_writes_hdr_miss; 316145510Sdarrenr kstat_named_t arcstat_l2_evict_lock_retry; 317145510Sdarrenr kstat_named_t arcstat_l2_evict_reading; 318145510Sdarrenr kstat_named_t arcstat_l2_free_on_write; 319145510Sdarrenr kstat_named_t arcstat_l2_abort_lowmem; 320145510Sdarrenr kstat_named_t arcstat_l2_cksum_bad; 321145510Sdarrenr kstat_named_t arcstat_l2_io_error; 322145510Sdarrenr kstat_named_t arcstat_l2_size; 323145510Sdarrenr kstat_named_t arcstat_l2_hdr_size; 324145510Sdarrenr kstat_named_t arcstat_memory_throttle_count; 325145510Sdarrenr kstat_named_t arcstat_l2_write_trylock_fail; 326145510Sdarrenr kstat_named_t arcstat_l2_write_passed_headroom; 327145510Sdarrenr kstat_named_t arcstat_l2_write_spa_mismatch; 328145510Sdarrenr kstat_named_t arcstat_l2_write_in_l2; 329145510Sdarrenr kstat_named_t arcstat_l2_write_hdr_io_in_progress; 330145510Sdarrenr kstat_named_t arcstat_l2_write_not_cacheable; 331145510Sdarrenr kstat_named_t arcstat_l2_write_full; 332145510Sdarrenr kstat_named_t arcstat_l2_write_buffer_iter; 333145510Sdarrenr kstat_named_t arcstat_l2_write_pios; 334145510Sdarrenr kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 335145510Sdarrenr kstat_named_t arcstat_l2_write_buffer_list_iter; 336145510Sdarrenr kstat_named_t arcstat_l2_write_buffer_list_null_iter; 337145510Sdarrenr} arc_stats_t; 338145510Sdarrenr 339145510Sdarrenrstatic arc_stats_t arc_stats = { 340145510Sdarrenr { "hits", KSTAT_DATA_UINT64 }, 341145510Sdarrenr { "misses", KSTAT_DATA_UINT64 }, 342145510Sdarrenr { "demand_data_hits", KSTAT_DATA_UINT64 }, 343145510Sdarrenr { "demand_data_misses", KSTAT_DATA_UINT64 }, 344145510Sdarrenr { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 345145510Sdarrenr { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 346145510Sdarrenr { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 347145510Sdarrenr { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 348145510Sdarrenr { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 349145510Sdarrenr { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 350145510Sdarrenr { "mru_hits", KSTAT_DATA_UINT64 }, 351255332Scy { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 352255332Scy { "mfu_hits", KSTAT_DATA_UINT64 }, 353255332Scy { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 354255332Scy { "allocated", KSTAT_DATA_UINT64 }, 355145510Sdarrenr { "deleted", KSTAT_DATA_UINT64 }, 356145510Sdarrenr { "stolen", KSTAT_DATA_UINT64 }, 357145510Sdarrenr { "recycle_miss", KSTAT_DATA_UINT64 }, 358145510Sdarrenr { "mutex_miss", KSTAT_DATA_UINT64 }, 359145510Sdarrenr { "evict_skip", KSTAT_DATA_UINT64 }, 360145510Sdarrenr { "evict_l2_cached", KSTAT_DATA_UINT64 }, 361145510Sdarrenr { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 362145510Sdarrenr { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 363145510Sdarrenr { "hash_elements", KSTAT_DATA_UINT64 }, 364145510Sdarrenr { "hash_elements_max", KSTAT_DATA_UINT64 }, 365145510Sdarrenr { "hash_collisions", KSTAT_DATA_UINT64 }, 366255332Scy { "hash_chains", KSTAT_DATA_UINT64 }, 367145510Sdarrenr { "hash_chain_max", KSTAT_DATA_UINT64 }, 368145510Sdarrenr { "p", KSTAT_DATA_UINT64 }, 369145510Sdarrenr { "c", KSTAT_DATA_UINT64 }, 370145510Sdarrenr { "c_min", KSTAT_DATA_UINT64 }, 371145510Sdarrenr { "c_max", KSTAT_DATA_UINT64 }, 372145510Sdarrenr { "size", KSTAT_DATA_UINT64 }, 373145510Sdarrenr { "hdr_size", KSTAT_DATA_UINT64 }, 374145510Sdarrenr { "data_size", KSTAT_DATA_UINT64 }, 375145510Sdarrenr { "other_size", KSTAT_DATA_UINT64 }, 376145510Sdarrenr { "l2_hits", KSTAT_DATA_UINT64 }, 377353080Scy { "l2_misses", KSTAT_DATA_UINT64 }, 378353080Scy { "l2_feeds", KSTAT_DATA_UINT64 }, 379353080Scy { "l2_rw_clash", KSTAT_DATA_UINT64 }, 380145510Sdarrenr { "l2_read_bytes", KSTAT_DATA_UINT64 }, 381145510Sdarrenr { "l2_write_bytes", KSTAT_DATA_UINT64 }, 382353080Scy { "l2_writes_sent", KSTAT_DATA_UINT64 }, 383353080Scy { "l2_writes_done", KSTAT_DATA_UINT64 }, 384353080Scy { "l2_writes_error", KSTAT_DATA_UINT64 }, 385145510Sdarrenr { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 386145510Sdarrenr { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 387145510Sdarrenr { "l2_evict_reading", KSTAT_DATA_UINT64 }, 388255332Scy { "l2_free_on_write", KSTAT_DATA_UINT64 }, 389145510Sdarrenr { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 390145510Sdarrenr { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 391145510Sdarrenr { "l2_io_error", KSTAT_DATA_UINT64 }, 392145510Sdarrenr { "l2_size", KSTAT_DATA_UINT64 }, 393145510Sdarrenr { "l2_hdr_size", KSTAT_DATA_UINT64 }, 394145510Sdarrenr { "memory_throttle_count", KSTAT_DATA_UINT64 }, 395145510Sdarrenr { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 396145510Sdarrenr { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 397145510Sdarrenr { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 398145510Sdarrenr { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 399145510Sdarrenr { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 400145510Sdarrenr { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 401145510Sdarrenr { "l2_write_full", KSTAT_DATA_UINT64 }, 402255332Scy { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 403255332Scy { "l2_write_pios", KSTAT_DATA_UINT64 }, 404255332Scy { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 405255332Scy { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 406145510Sdarrenr { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 } 407145510Sdarrenr}; 408255332Scy 409255332Scy#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 410145510Sdarrenr 411145510Sdarrenr#define ARCSTAT_INCR(stat, val) \ 412145510Sdarrenr atomic_add_64(&arc_stats.stat.value.ui64, (val)); 413145510Sdarrenr 414145510Sdarrenr#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 415145510Sdarrenr#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 416145510Sdarrenr 417145510Sdarrenr#define ARCSTAT_MAX(stat, val) { \ 418145510Sdarrenr uint64_t m; \ 419145510Sdarrenr while ((val) > (m = arc_stats.stat.value.ui64) && \ 420145510Sdarrenr (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 421145510Sdarrenr continue; \ 422145510Sdarrenr} 423145510Sdarrenr 424145510Sdarrenr#define ARCSTAT_MAXSTAT(stat) \ 425145510Sdarrenr ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 426145510Sdarrenr 427145510Sdarrenr/* 428145510Sdarrenr * We define a macro to allow ARC hits/misses to be easily broken down by 429145510Sdarrenr * two separate conditions, giving a total of four different subtypes for 430145510Sdarrenr * each of hits and misses (so eight statistics total). 431145510Sdarrenr */ 432145510Sdarrenr#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 433145510Sdarrenr if (cond1) { \ 434145510Sdarrenr if (cond2) { \ 435145510Sdarrenr ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 436145510Sdarrenr } else { \ 437145510Sdarrenr ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 438145510Sdarrenr } \ 439145510Sdarrenr } else { \ 440145510Sdarrenr if (cond2) { \ 441145510Sdarrenr ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 442145510Sdarrenr } else { \ 443145510Sdarrenr ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 444145510Sdarrenr } \ 445145510Sdarrenr } 446145510Sdarrenr 447145510Sdarrenrkstat_t *arc_ksp; 448145510Sdarrenrstatic arc_state_t *arc_anon; 449145510Sdarrenrstatic arc_state_t *arc_mru; 450145510Sdarrenrstatic arc_state_t *arc_mru_ghost; 451145510Sdarrenrstatic arc_state_t *arc_mfu; 452145510Sdarrenrstatic arc_state_t *arc_mfu_ghost; 453145510Sdarrenrstatic arc_state_t *arc_l2c_only; 454353079Scy 455353079Scy/* 456353079Scy * There are several ARC variables that are critical to export as kstats -- 457145510Sdarrenr * but we don't want to have to grovel around in the kstat whenever we wish to 458145510Sdarrenr * manipulate them. For these variables, we therefore define them to be in 459353079Scy * terms of the statistic variable. This assures that we are not introducing 460353079Scy * the possibility of inconsistency by having shadow copies of the variables, 461353079Scy * while still allowing the code to be readable. 462145510Sdarrenr */ 463145510Sdarrenr#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 464145510Sdarrenr#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 465255332Scy#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 466145510Sdarrenr#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 467145510Sdarrenr#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 468145510Sdarrenr 469145510Sdarrenrstatic int arc_no_grow; /* Don't try to grow cache size */ 470145510Sdarrenrstatic uint64_t arc_tempreserve; 471145510Sdarrenrstatic uint64_t arc_loaned_bytes; 472145510Sdarrenrstatic uint64_t arc_meta_used; 473145510Sdarrenrstatic uint64_t arc_meta_limit; 474145510Sdarrenrstatic uint64_t arc_meta_max = 0; 475145510SdarrenrSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RD, &arc_meta_used, 0, 476145510Sdarrenr "ARC metadata used"); 477255332ScySYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RW, &arc_meta_limit, 0, 478145510Sdarrenr "ARC metadata limit"); 479145510Sdarrenr 480255332Scytypedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 481145510Sdarrenr 482145510Sdarrenrtypedef struct arc_callback arc_callback_t; 483255332Scy 484255332Scystruct arc_callback { 485145510Sdarrenr void *acb_private; 486145510Sdarrenr arc_done_func_t *acb_done; 487145510Sdarrenr arc_buf_t *acb_buf; 488145510Sdarrenr zio_t *acb_zio_dummy; 489145510Sdarrenr arc_callback_t *acb_next; 490145510Sdarrenr}; 491145510Sdarrenr 492255332Scytypedef struct arc_write_callback arc_write_callback_t; 493145510Sdarrenr 494145510Sdarrenrstruct arc_write_callback { 495255332Scy void *awcb_private; 496145510Sdarrenr arc_done_func_t *awcb_ready; 497145510Sdarrenr arc_done_func_t *awcb_done; 498255332Scy arc_buf_t *awcb_buf; 499255332Scy}; 500255332Scy 501255332Scystruct arc_buf_hdr { 502145510Sdarrenr /* protected by hash lock */ 503145510Sdarrenr dva_t b_dva; 504255332Scy uint64_t b_birth; 505255332Scy uint64_t b_cksum0; 506255332Scy 507255332Scy kmutex_t b_freeze_lock; 508255332Scy zio_cksum_t *b_freeze_cksum; 509255332Scy void *b_thawed; 510255332Scy 511255332Scy arc_buf_hdr_t *b_hash_next; 512255332Scy arc_buf_t *b_buf; 513255332Scy uint32_t b_flags; 514255332Scy uint32_t b_datacnt; 515255332Scy 516255332Scy arc_callback_t *b_acb; 517255332Scy kcondvar_t b_cv; 518255332Scy 519255332Scy /* immutable */ 520255332Scy arc_buf_contents_t b_type; 521255332Scy uint64_t b_size; 522255332Scy uint64_t b_spa; 523255332Scy 524255332Scy /* protected by arc state mutex */ 525255332Scy arc_state_t *b_state; 526255332Scy list_node_t b_arc_node; 527145510Sdarrenr 528145510Sdarrenr /* updated atomically */ 529145510Sdarrenr clock_t b_arc_access; 530145510Sdarrenr 531255332Scy /* self protecting */ 532255332Scy refcount_t b_refcnt; 533255332Scy 534255332Scy l2arc_buf_hdr_t *b_l2hdr; 535145510Sdarrenr list_node_t b_l2node; 536145510Sdarrenr}; 537145510Sdarrenr 538145510Sdarrenrstatic arc_buf_t *arc_eviction_list; 539145510Sdarrenrstatic kmutex_t arc_eviction_mtx; 540145510Sdarrenrstatic arc_buf_hdr_t arc_eviction_hdr; 541145510Sdarrenrstatic void arc_get_data_buf(arc_buf_t *buf); 542145510Sdarrenrstatic void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock); 543145510Sdarrenrstatic int arc_evict_needed(arc_buf_contents_t type); 544145510Sdarrenrstatic void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes); 545145510Sdarrenr#ifdef illumos 546145510Sdarrenrstatic void arc_buf_watch(arc_buf_t *buf); 547145510Sdarrenr#endif /* illumos */ 548145510Sdarrenr 549145510Sdarrenrstatic boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab); 550145510Sdarrenr 551145510Sdarrenr#define GHOST_STATE(state) \ 552145510Sdarrenr ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 553145510Sdarrenr (state) == arc_l2c_only) 554145510Sdarrenr 555145510Sdarrenr/* 556145510Sdarrenr * Private ARC flags. These flags are private ARC only flags that will show up 557145510Sdarrenr * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can 558145510Sdarrenr * be passed in as arc_flags in things like arc_read. However, these flags 559145510Sdarrenr * should never be passed and should only be set by ARC code. When adding new 560145510Sdarrenr * public flags, make sure not to smash the private ones. 561145510Sdarrenr */ 562145510Sdarrenr 563145510Sdarrenr#define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */ 564145510Sdarrenr#define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */ 565145510Sdarrenr#define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */ 566353078Scy#define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */ 567353078Scy#define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */ 568353078Scy#define ARC_INDIRECT (1 << 14) /* this is an indirect block */ 569145510Sdarrenr#define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */ 570145510Sdarrenr#define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */ 571353078Scy#define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */ 572353078Scy#define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */ 573353078Scy 574145510Sdarrenr#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE) 575145510Sdarrenr#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS) 576145510Sdarrenr#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR) 577255332Scy#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH) 578145510Sdarrenr#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ) 579145510Sdarrenr#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE) 580145510Sdarrenr#define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS) 581145510Sdarrenr#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE) 582145510Sdarrenr#define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \ 583145510Sdarrenr (hdr)->b_l2hdr != NULL) 584145510Sdarrenr#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING) 585145510Sdarrenr#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED) 586145510Sdarrenr#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD) 587145510Sdarrenr 588145510Sdarrenr/* 589145510Sdarrenr * Other sizes 590255332Scy */ 591145510Sdarrenr 592255332Scy#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 593145510Sdarrenr#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 594145510Sdarrenr 595145510Sdarrenr/* 596145510Sdarrenr * Hash table routines 597255332Scy */ 598145510Sdarrenr 599145510Sdarrenr#define HT_LOCK_PAD CACHE_LINE_SIZE 600145510Sdarrenr 601145510Sdarrenrstruct ht_lock { 602145510Sdarrenr kmutex_t ht_lock; 603145510Sdarrenr#ifdef _KERNEL 604255332Scy unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 605255332Scy#endif 606255332Scy}; 607145510Sdarrenr 608145510Sdarrenr#define BUF_LOCKS 256 609145510Sdarrenrtypedef struct buf_hash_table { 610145510Sdarrenr uint64_t ht_mask; 611145510Sdarrenr arc_buf_hdr_t **ht_table; 612145510Sdarrenr struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 613145510Sdarrenr} buf_hash_table_t; 614145510Sdarrenr 615145510Sdarrenrstatic buf_hash_table_t buf_hash_table; 616145510Sdarrenr 617145510Sdarrenr#define BUF_HASH_INDEX(spa, dva, birth) \ 618145510Sdarrenr (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 619145510Sdarrenr#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 620145510Sdarrenr#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 621145510Sdarrenr#define HDR_LOCK(hdr) \ 622145510Sdarrenr (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 623145510Sdarrenr 624145510Sdarrenruint64_t zfs_crc64_table[256]; 625145510Sdarrenr 626145510Sdarrenr/* 627145510Sdarrenr * Level 2 ARC 628145510Sdarrenr */ 629145510Sdarrenr 630145510Sdarrenr#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 631145510Sdarrenr#define L2ARC_HEADROOM 2 /* num of writes */ 632145510Sdarrenr#define L2ARC_FEED_SECS 1 /* caching interval secs */ 633145510Sdarrenr#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 634145510Sdarrenr 635145510Sdarrenr#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 636255332Scy#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 637255332Scy 638255332Scy/* 639255332Scy * L2ARC Performance Tunables 640145510Sdarrenr */ 641145510Sdarrenruint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 642145510Sdarrenruint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 643255332Scyuint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 644145510Sdarrenruint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 645255332Scyuint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 646145510Sdarrenrboolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 647145510Sdarrenrboolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 648145510Sdarrenrboolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 649255332Scy 650145510SdarrenrSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 651145510Sdarrenr &l2arc_write_max, 0, "max write size"); 652145510SdarrenrSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 653145510Sdarrenr &l2arc_write_boost, 0, "extra write during warmup"); 654145510SdarrenrSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 655145510Sdarrenr &l2arc_headroom, 0, "number of dev writes"); 656145510SdarrenrSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 657145510Sdarrenr &l2arc_feed_secs, 0, "interval seconds"); 658170268SdarrenrSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 659170268Sdarrenr &l2arc_feed_min_ms, 0, "min interval milliseconds"); 660255332Scy 661255332ScySYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 662255332Scy &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 663255332ScySYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 664170268Sdarrenr &l2arc_feed_again, 0, "turbo warmup"); 665170268SdarrenrSYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 666170268Sdarrenr &l2arc_norw, 0, "no reads during writes"); 667170268Sdarrenr 668170268SdarrenrSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 669170268Sdarrenr &ARC_anon.arcs_size, 0, "size of anonymous state"); 670170268SdarrenrSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 671170268Sdarrenr &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 672170268SdarrenrSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 673170268Sdarrenr &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 674170268Sdarrenr 675170268SdarrenrSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 676353074Scy &ARC_mru.arcs_size, 0, "size of mru state"); 677170268SdarrenrSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 678170268Sdarrenr &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 679170268SdarrenrSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 680170268Sdarrenr &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 681170268Sdarrenr 682170268SdarrenrSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 683170268Sdarrenr &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 684170268SdarrenrSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 685170268Sdarrenr &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 686170268Sdarrenr "size of metadata in mru ghost state"); 687170268SdarrenrSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 688170268Sdarrenr &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 689170268Sdarrenr "size of data in mru ghost state"); 690170268Sdarrenr 691170268SdarrenrSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 692170268Sdarrenr &ARC_mfu.arcs_size, 0, "size of mfu state"); 693170268SdarrenrSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 694170268Sdarrenr &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 695170268SdarrenrSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 696170268Sdarrenr &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 697170268Sdarrenr 698170268SdarrenrSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 699170268Sdarrenr &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 700353076ScySYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 701255332Scy &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 702353076Scy "size of metadata in mfu ghost state"); 703353076ScySYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 704353076Scy &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 705353076Scy "size of data in mfu ghost state"); 706353076Scy 707255332ScySYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 708255332Scy &ARC_l2c_only.arcs_size, 0, "size of mru state"); 709353076Scy 710170268Sdarrenr/* 711170268Sdarrenr * L2ARC Internals 712170268Sdarrenr */ 713170268Sdarrenrtypedef struct l2arc_dev { 714170268Sdarrenr vdev_t *l2ad_vdev; /* vdev */ 715170268Sdarrenr spa_t *l2ad_spa; /* spa */ 716170268Sdarrenr uint64_t l2ad_hand; /* next write location */ 717170268Sdarrenr uint64_t l2ad_write; /* desired write size, bytes */ 718170268Sdarrenr uint64_t l2ad_boost; /* warmup write boost, bytes */ 719170268Sdarrenr uint64_t l2ad_start; /* first addr on device */ 720353075Scy uint64_t l2ad_end; /* last addr on device */ 721353075Scy uint64_t l2ad_evict; /* last addr eviction reached */ 722353075Scy boolean_t l2ad_first; /* first sweep through */ 723170268Sdarrenr boolean_t l2ad_writing; /* currently writing */ 724170268Sdarrenr list_t *l2ad_buflist; /* buffer list */ 725353075Scy list_node_t l2ad_node; /* device list node */ 726353075Scy} l2arc_dev_t; 727353075Scy 728170268Sdarrenrstatic list_t L2ARC_dev_list; /* device list */ 729170268Sdarrenrstatic list_t *l2arc_dev_list; /* device list pointer */ 730170268Sdarrenrstatic kmutex_t l2arc_dev_mtx; /* device list mutex */ 731255332Scystatic l2arc_dev_t *l2arc_dev_last; /* last device used */ 732170268Sdarrenrstatic kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 733170268Sdarrenrstatic list_t L2ARC_free_on_write; /* free after write buf list */ 734170268Sdarrenrstatic list_t *l2arc_free_on_write; /* free after write list ptr */ 735170268Sdarrenrstatic kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 736170268Sdarrenrstatic uint64_t l2arc_ndev; /* number of devices */ 737170268Sdarrenr 738170268Sdarrenrtypedef struct l2arc_read_callback { 739170268Sdarrenr arc_buf_t *l2rcb_buf; /* read buffer */ 740170268Sdarrenr spa_t *l2rcb_spa; /* spa */ 741170268Sdarrenr blkptr_t l2rcb_bp; /* original blkptr */ 742170268Sdarrenr zbookmark_t l2rcb_zb; /* original bookmark */ 743170268Sdarrenr int l2rcb_flags; /* original flags */ 744170268Sdarrenr} l2arc_read_callback_t; 745170268Sdarrenr 746170268Sdarrenrtypedef struct l2arc_write_callback { 747170268Sdarrenr l2arc_dev_t *l2wcb_dev; /* device info */ 748170268Sdarrenr arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 749170268Sdarrenr} l2arc_write_callback_t; 750170268Sdarrenr 751170268Sdarrenrstruct l2arc_buf_hdr { 752170268Sdarrenr /* protected by arc_buf_hdr mutex */ 753170268Sdarrenr l2arc_dev_t *b_dev; /* L2ARC device */ 754255332Scy uint64_t b_daddr; /* disk address, offset byte */ 755255332Scy}; 756255332Scy 757255332Scytypedef struct l2arc_data_free { 758170268Sdarrenr /* protected by l2arc_free_on_write_mtx */ 759170268Sdarrenr void *l2df_data; 760170268Sdarrenr size_t l2df_size; 761170268Sdarrenr void (*l2df_func)(void *, size_t); 762170268Sdarrenr list_node_t l2df_list_node; 763170268Sdarrenr} l2arc_data_free_t; 764170268Sdarrenr 765170268Sdarrenrstatic kmutex_t l2arc_feed_thr_lock; 766170268Sdarrenrstatic kcondvar_t l2arc_feed_thr_cv; 767170268Sdarrenrstatic uint8_t l2arc_thread_exit; 768170268Sdarrenr 769170268Sdarrenrstatic void l2arc_read_done(zio_t *zio); 770170268Sdarrenrstatic void l2arc_hdr_stat_add(void); 771170268Sdarrenrstatic void l2arc_hdr_stat_remove(void); 772170268Sdarrenr 773170268Sdarrenrstatic uint64_t 774170268Sdarrenrbuf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 775170268Sdarrenr{ 776170268Sdarrenr uint8_t *vdva = (uint8_t *)dva; 777170268Sdarrenr uint64_t crc = -1ULL; 778170268Sdarrenr int i; 779170268Sdarrenr 780255332Scy ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 781170268Sdarrenr 782170268Sdarrenr for (i = 0; i < sizeof (dva_t); i++) 783170268Sdarrenr crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 784170268Sdarrenr 785170268Sdarrenr crc ^= (spa>>8) ^ birth; 786170268Sdarrenr 787255332Scy return (crc); 788255332Scy} 789170268Sdarrenr 790170268Sdarrenr#define BUF_EMPTY(buf) \ 791170268Sdarrenr ((buf)->b_dva.dva_word[0] == 0 && \ 792170268Sdarrenr (buf)->b_dva.dva_word[1] == 0 && \ 793170268Sdarrenr (buf)->b_birth == 0) 794170268Sdarrenr 795170268Sdarrenr#define BUF_EQUAL(spa, dva, birth, buf) \ 796170268Sdarrenr ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 797170268Sdarrenr ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 798170268Sdarrenr ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 799170268Sdarrenr 800170268Sdarrenrstatic void 801170268Sdarrenrbuf_discard_identity(arc_buf_hdr_t *hdr) 802170268Sdarrenr{ 803170268Sdarrenr hdr->b_dva.dva_word[0] = 0; 804170268Sdarrenr hdr->b_dva.dva_word[1] = 0; 805170268Sdarrenr hdr->b_birth = 0; 806170268Sdarrenr hdr->b_cksum0 = 0; 807170268Sdarrenr} 808170268Sdarrenr 809255332Scystatic arc_buf_hdr_t * 810170268Sdarrenrbuf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp) 811170268Sdarrenr{ 812170268Sdarrenr uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 813170268Sdarrenr kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 814170268Sdarrenr arc_buf_hdr_t *buf; 815170268Sdarrenr 816255332Scy mutex_enter(hash_lock); 817255332Scy for (buf = buf_hash_table.ht_table[idx]; buf != NULL; 818170268Sdarrenr buf = buf->b_hash_next) { 819170268Sdarrenr if (BUF_EQUAL(spa, dva, birth, buf)) { 820170268Sdarrenr *lockp = hash_lock; 821170268Sdarrenr return (buf); 822170268Sdarrenr } 823170268Sdarrenr } 824170268Sdarrenr mutex_exit(hash_lock); 825255332Scy *lockp = NULL; 826255332Scy return (NULL); 827255332Scy} 828255332Scy 829170268Sdarrenr/* 830255332Scy * Insert an entry into the hash table. If there is already an element 831170268Sdarrenr * equal to elem in the hash table, then the already existing element 832170268Sdarrenr * will be returned and the new element will not be inserted. 833170268Sdarrenr * Otherwise returns NULL. 834170268Sdarrenr */ 835170268Sdarrenrstatic arc_buf_hdr_t * 836170268Sdarrenrbuf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp) 837170268Sdarrenr{ 838170268Sdarrenr uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 839170268Sdarrenr kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 840170268Sdarrenr arc_buf_hdr_t *fbuf; 841170268Sdarrenr uint32_t i; 842170268Sdarrenr 843170268Sdarrenr ASSERT(!HDR_IN_HASH_TABLE(buf)); 844170268Sdarrenr *lockp = hash_lock; 845170268Sdarrenr mutex_enter(hash_lock); 846255332Scy for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL; 847170268Sdarrenr fbuf = fbuf->b_hash_next, i++) { 848170268Sdarrenr if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf)) 849170268Sdarrenr return (fbuf); 850170268Sdarrenr } 851170268Sdarrenr 852255332Scy buf->b_hash_next = buf_hash_table.ht_table[idx]; 853170268Sdarrenr buf_hash_table.ht_table[idx] = buf; 854170268Sdarrenr buf->b_flags |= ARC_IN_HASH_TABLE; 855170268Sdarrenr 856170268Sdarrenr /* collect some hash table performance data */ 857255332Scy if (i > 0) { 858170268Sdarrenr ARCSTAT_BUMP(arcstat_hash_collisions); 859170268Sdarrenr if (i == 1) 860170268Sdarrenr ARCSTAT_BUMP(arcstat_hash_chains); 861170268Sdarrenr 862170268Sdarrenr ARCSTAT_MAX(arcstat_hash_chain_max, i); 863170268Sdarrenr } 864170268Sdarrenr 865170268Sdarrenr ARCSTAT_BUMP(arcstat_hash_elements); 866170268Sdarrenr ARCSTAT_MAXSTAT(arcstat_hash_elements); 867170268Sdarrenr 868170268Sdarrenr return (NULL); 869255332Scy} 870255332Scy 871170268Sdarrenrstatic void 872170268Sdarrenrbuf_hash_remove(arc_buf_hdr_t *buf) 873170268Sdarrenr{ 874170268Sdarrenr arc_buf_hdr_t *fbuf, **bufp; 875170268Sdarrenr uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 876170268Sdarrenr 877170268Sdarrenr ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 878170268Sdarrenr ASSERT(HDR_IN_HASH_TABLE(buf)); 879170268Sdarrenr 880170268Sdarrenr bufp = &buf_hash_table.ht_table[idx]; 881170268Sdarrenr while ((fbuf = *bufp) != buf) { 882255332Scy ASSERT(fbuf != NULL); 883170268Sdarrenr bufp = &fbuf->b_hash_next; 884170268Sdarrenr } 885170268Sdarrenr *bufp = buf->b_hash_next; 886170268Sdarrenr buf->b_hash_next = NULL; 887170268Sdarrenr buf->b_flags &= ~ARC_IN_HASH_TABLE; 888170268Sdarrenr 889170268Sdarrenr /* collect some hash table performance data */ 890170268Sdarrenr ARCSTAT_BUMPDOWN(arcstat_hash_elements); 891170268Sdarrenr 892255332Scy if (buf_hash_table.ht_table[idx] && 893170268Sdarrenr buf_hash_table.ht_table[idx]->b_hash_next == NULL) 894170268Sdarrenr ARCSTAT_BUMPDOWN(arcstat_hash_chains); 895170268Sdarrenr} 896170268Sdarrenr 897170268Sdarrenr/* 898255332Scy * Global data structures and functions for the buf kmem cache. 899170268Sdarrenr */ 900170268Sdarrenrstatic kmem_cache_t *hdr_cache; 901255332Scystatic kmem_cache_t *buf_cache; 902255332Scy 903255332Scystatic void 904255332Scybuf_fini(void) 905255332Scy{ 906255332Scy int i; 907255332Scy 908255332Scy kmem_free(buf_hash_table.ht_table, 909255332Scy (buf_hash_table.ht_mask + 1) * sizeof (void *)); 910255332Scy for (i = 0; i < BUF_LOCKS; i++) 911255332Scy mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 912255332Scy kmem_cache_destroy(hdr_cache); 913255332Scy kmem_cache_destroy(buf_cache); 914255332Scy} 915255332Scy 916255332Scy/* 917255332Scy * Constructor callback - called when the cache is empty 918255332Scy * and a new buf is requested. 919255332Scy */ 920255332Scy/* ARGSUSED */ 921255332Scystatic int 922255332Scyhdr_cons(void *vbuf, void *unused, int kmflag) 923255332Scy{ 924255332Scy arc_buf_hdr_t *buf = vbuf; 925255332Scy 926255332Scy bzero(buf, sizeof (arc_buf_hdr_t)); 927255332Scy refcount_create(&buf->b_refcnt); 928255332Scy cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL); 929255332Scy mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 930255332Scy arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 931255332Scy 932255332Scy return (0); 933255332Scy} 934255332Scy 935170268Sdarrenr/* ARGSUSED */ 936170268Sdarrenrstatic int 937170268Sdarrenrbuf_cons(void *vbuf, void *unused, int kmflag) 938255332Scy{ 939255332Scy arc_buf_t *buf = vbuf; 940255332Scy 941255332Scy bzero(buf, sizeof (arc_buf_t)); 942255332Scy mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 943170268Sdarrenr rw_init(&buf->b_data_lock, NULL, RW_DEFAULT, NULL); 944170268Sdarrenr arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 945170268Sdarrenr 946170268Sdarrenr return (0); 947170268Sdarrenr} 948170268Sdarrenr 949170268Sdarrenr/* 950170268Sdarrenr * Destructor callback - called when a cached buf is 951170268Sdarrenr * no longer required. 952170268Sdarrenr */ 953170268Sdarrenr/* ARGSUSED */ 954170268Sdarrenrstatic void 955170268Sdarrenrhdr_dest(void *vbuf, void *unused) 956170268Sdarrenr{ 957170268Sdarrenr arc_buf_hdr_t *buf = vbuf; 958170268Sdarrenr 959170268Sdarrenr ASSERT(BUF_EMPTY(buf)); 960170268Sdarrenr refcount_destroy(&buf->b_refcnt); 961255332Scy cv_destroy(&buf->b_cv); 962255332Scy mutex_destroy(&buf->b_freeze_lock); 963255332Scy arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 964170268Sdarrenr} 965255332Scy 966170268Sdarrenr/* ARGSUSED */ 967170268Sdarrenrstatic void 968255332Scybuf_dest(void *vbuf, void *unused) 969255332Scy{ 970255332Scy arc_buf_t *buf = vbuf; 971170268Sdarrenr 972255332Scy mutex_destroy(&buf->b_evict_lock); 973170268Sdarrenr rw_destroy(&buf->b_data_lock); 974170268Sdarrenr arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 975170268Sdarrenr} 976170268Sdarrenr 977255332Scy/* 978255332Scy * Reclaim callback -- invoked when memory is low. 979255332Scy */ 980255332Scy/* ARGSUSED */ 981255332Scystatic void 982170268Sdarrenrhdr_recl(void *unused) 983170268Sdarrenr{ 984170268Sdarrenr dprintf("hdr_recl called\n"); 985170268Sdarrenr /* 986170268Sdarrenr * umem calls the reclaim func when we destroy the buf cache, 987170268Sdarrenr * which is after we do arc_fini(). 988170268Sdarrenr */ 989170268Sdarrenr if (!arc_dead) 990170268Sdarrenr cv_signal(&arc_reclaim_thr_cv); 991170268Sdarrenr} 992170268Sdarrenr 993170268Sdarrenrstatic void 994170268Sdarrenrbuf_init(void) 995170268Sdarrenr{ 996170268Sdarrenr uint64_t *ct; 997170268Sdarrenr uint64_t hsize = 1ULL << 12; 998170268Sdarrenr int i, j; 999170268Sdarrenr 1000170268Sdarrenr /* 1001170268Sdarrenr * The hash table is big enough to fill all of physical memory 1002255332Scy * with an average 64K block size. The table will take up 1003170268Sdarrenr * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers). 1004170268Sdarrenr */ 1005170268Sdarrenr while (hsize * 65536 < (uint64_t)physmem * PAGESIZE) 1006255332Scy hsize <<= 1; 1007170268Sdarrenrretry: 1008170268Sdarrenr buf_hash_table.ht_mask = hsize - 1; 1009170268Sdarrenr buf_hash_table.ht_table = 1010170268Sdarrenr kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1011170268Sdarrenr if (buf_hash_table.ht_table == NULL) { 1012170268Sdarrenr ASSERT(hsize > (1ULL << 8)); 1013255332Scy hsize >>= 1; 1014255332Scy goto retry; 1015255332Scy } 1016255332Scy 1017255332Scy hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1018170268Sdarrenr 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1019255332Scy buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1020255332Scy 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1021255332Scy 1022255332Scy for (i = 0; i < 256; i++) 1023255332Scy for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1024255332Scy *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1025255332Scy 1026255332Scy for (i = 0; i < BUF_LOCKS; i++) { 1027255332Scy mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1028255332Scy NULL, MUTEX_DEFAULT, NULL); 1029255332Scy } 1030255332Scy} 1031255332Scy 1032255332Scy#define ARC_MINTIME (hz>>4) /* 62 ms */ 1033255332Scy 1034255332Scystatic void 1035255332Scyarc_cksum_verify(arc_buf_t *buf) 1036255332Scy{ 1037255332Scy zio_cksum_t zc; 1038255332Scy 1039255332Scy if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1040255332Scy return; 1041255332Scy 1042255332Scy mutex_enter(&buf->b_hdr->b_freeze_lock); 1043255332Scy if (buf->b_hdr->b_freeze_cksum == NULL || 1044255332Scy (buf->b_hdr->b_flags & ARC_IO_ERROR)) { 1045255332Scy mutex_exit(&buf->b_hdr->b_freeze_lock); 1046255332Scy return; 1047255332Scy } 1048255332Scy fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1049255332Scy if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1050255332Scy panic("buffer modified while frozen!"); 1051255332Scy mutex_exit(&buf->b_hdr->b_freeze_lock); 1052170268Sdarrenr} 1053170268Sdarrenr 1054170268Sdarrenrstatic int 1055170268Sdarrenrarc_cksum_equal(arc_buf_t *buf) 1056170268Sdarrenr{ 1057170268Sdarrenr zio_cksum_t zc; 1058170268Sdarrenr int equal; 1059255332Scy 1060170268Sdarrenr mutex_enter(&buf->b_hdr->b_freeze_lock); 1061170268Sdarrenr fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1062170268Sdarrenr equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1063170268Sdarrenr mutex_exit(&buf->b_hdr->b_freeze_lock); 1064170268Sdarrenr 1065170268Sdarrenr return (equal); 1066170268Sdarrenr} 1067255332Scy 1068255332Scystatic void 1069255332Scyarc_cksum_compute(arc_buf_t *buf, boolean_t force) 1070318206Scy{ 1071255332Scy if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1072318206Scy return; 1073255332Scy 1074255332Scy mutex_enter(&buf->b_hdr->b_freeze_lock); 1075255332Scy if (buf->b_hdr->b_freeze_cksum != NULL) { 1076255332Scy mutex_exit(&buf->b_hdr->b_freeze_lock); 1077255332Scy return; 1078255332Scy } 1079255332Scy buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1080255332Scy fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1081255332Scy buf->b_hdr->b_freeze_cksum); 1082255332Scy mutex_exit(&buf->b_hdr->b_freeze_lock); 1083255332Scy#ifdef illumos 1084255332Scy arc_buf_watch(buf); 1085255332Scy#endif /* illumos */ 1086255332Scy} 1087255332Scy 1088255332Scy#ifdef illumos 1089255332Scy#ifndef _KERNEL 1090255332Scytypedef struct procctl { 1091255332Scy long cmd; 1092255332Scy prwatch_t prwatch; 1093255332Scy} procctl_t; 1094170268Sdarrenr#endif 1095170268Sdarrenr 1096/* ARGSUSED */ 1097static void 1098arc_buf_unwatch(arc_buf_t *buf) 1099{ 1100#ifndef _KERNEL 1101 if (arc_watch) { 1102 int result; 1103 procctl_t ctl; 1104 ctl.cmd = PCWATCH; 1105 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1106 ctl.prwatch.pr_size = 0; 1107 ctl.prwatch.pr_wflags = 0; 1108 result = write(arc_procfd, &ctl, sizeof (ctl)); 1109 ASSERT3U(result, ==, sizeof (ctl)); 1110 } 1111#endif 1112} 1113 1114/* ARGSUSED */ 1115static void 1116arc_buf_watch(arc_buf_t *buf) 1117{ 1118#ifndef _KERNEL 1119 if (arc_watch) { 1120 int result; 1121 procctl_t ctl; 1122 ctl.cmd = PCWATCH; 1123 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1124 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1125 ctl.prwatch.pr_wflags = WA_WRITE; 1126 result = write(arc_procfd, &ctl, sizeof (ctl)); 1127 ASSERT3U(result, ==, sizeof (ctl)); 1128 } 1129#endif 1130} 1131#endif /* illumos */ 1132 1133void 1134arc_buf_thaw(arc_buf_t *buf) 1135{ 1136 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1137 if (buf->b_hdr->b_state != arc_anon) 1138 panic("modifying non-anon buffer!"); 1139 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS) 1140 panic("modifying buffer while i/o in progress!"); 1141 arc_cksum_verify(buf); 1142 } 1143 1144 mutex_enter(&buf->b_hdr->b_freeze_lock); 1145 if (buf->b_hdr->b_freeze_cksum != NULL) { 1146 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1147 buf->b_hdr->b_freeze_cksum = NULL; 1148 } 1149 1150 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1151 if (buf->b_hdr->b_thawed) 1152 kmem_free(buf->b_hdr->b_thawed, 1); 1153 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1154 } 1155 1156 mutex_exit(&buf->b_hdr->b_freeze_lock); 1157 1158#ifdef illumos 1159 arc_buf_unwatch(buf); 1160#endif /* illumos */ 1161} 1162 1163void 1164arc_buf_freeze(arc_buf_t *buf) 1165{ 1166 kmutex_t *hash_lock; 1167 1168 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1169 return; 1170 1171 hash_lock = HDR_LOCK(buf->b_hdr); 1172 mutex_enter(hash_lock); 1173 1174 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1175 buf->b_hdr->b_state == arc_anon); 1176 arc_cksum_compute(buf, B_FALSE); 1177 mutex_exit(hash_lock); 1178 1179} 1180 1181static void 1182get_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock) 1183{ 1184 uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth); 1185 1186 if (ab->b_type == ARC_BUFC_METADATA) 1187 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1188 else { 1189 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1190 buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1191 } 1192 1193 *list = &state->arcs_lists[buf_hashid]; 1194 *lock = ARCS_LOCK(state, buf_hashid); 1195} 1196 1197 1198static void 1199add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1200{ 1201 ASSERT(MUTEX_HELD(hash_lock)); 1202 1203 if ((refcount_add(&ab->b_refcnt, tag) == 1) && 1204 (ab->b_state != arc_anon)) { 1205 uint64_t delta = ab->b_size * ab->b_datacnt; 1206 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type]; 1207 list_t *list; 1208 kmutex_t *lock; 1209 1210 get_buf_info(ab, ab->b_state, &list, &lock); 1211 ASSERT(!MUTEX_HELD(lock)); 1212 mutex_enter(lock); 1213 ASSERT(list_link_active(&ab->b_arc_node)); 1214 list_remove(list, ab); 1215 if (GHOST_STATE(ab->b_state)) { 1216 ASSERT0(ab->b_datacnt); 1217 ASSERT3P(ab->b_buf, ==, NULL); 1218 delta = ab->b_size; 1219 } 1220 ASSERT(delta > 0); 1221 ASSERT3U(*size, >=, delta); 1222 atomic_add_64(size, -delta); 1223 mutex_exit(lock); 1224 /* remove the prefetch flag if we get a reference */ 1225 if (ab->b_flags & ARC_PREFETCH) 1226 ab->b_flags &= ~ARC_PREFETCH; 1227 } 1228} 1229 1230static int 1231remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1232{ 1233 int cnt; 1234 arc_state_t *state = ab->b_state; 1235 1236 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1237 ASSERT(!GHOST_STATE(state)); 1238 1239 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) && 1240 (state != arc_anon)) { 1241 uint64_t *size = &state->arcs_lsize[ab->b_type]; 1242 list_t *list; 1243 kmutex_t *lock; 1244 1245 get_buf_info(ab, state, &list, &lock); 1246 ASSERT(!MUTEX_HELD(lock)); 1247 mutex_enter(lock); 1248 ASSERT(!list_link_active(&ab->b_arc_node)); 1249 list_insert_head(list, ab); 1250 ASSERT(ab->b_datacnt > 0); 1251 atomic_add_64(size, ab->b_size * ab->b_datacnt); 1252 mutex_exit(lock); 1253 } 1254 return (cnt); 1255} 1256 1257/* 1258 * Move the supplied buffer to the indicated state. The mutex 1259 * for the buffer must be held by the caller. 1260 */ 1261static void 1262arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock) 1263{ 1264 arc_state_t *old_state = ab->b_state; 1265 int64_t refcnt = refcount_count(&ab->b_refcnt); 1266 uint64_t from_delta, to_delta; 1267 list_t *list; 1268 kmutex_t *lock; 1269 1270 ASSERT(MUTEX_HELD(hash_lock)); 1271 ASSERT(new_state != old_state); 1272 ASSERT(refcnt == 0 || ab->b_datacnt > 0); 1273 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state)); 1274 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon); 1275 1276 from_delta = to_delta = ab->b_datacnt * ab->b_size; 1277 1278 /* 1279 * If this buffer is evictable, transfer it from the 1280 * old state list to the new state list. 1281 */ 1282 if (refcnt == 0) { 1283 if (old_state != arc_anon) { 1284 int use_mutex; 1285 uint64_t *size = &old_state->arcs_lsize[ab->b_type]; 1286 1287 get_buf_info(ab, old_state, &list, &lock); 1288 use_mutex = !MUTEX_HELD(lock); 1289 if (use_mutex) 1290 mutex_enter(lock); 1291 1292 ASSERT(list_link_active(&ab->b_arc_node)); 1293 list_remove(list, ab); 1294 1295 /* 1296 * If prefetching out of the ghost cache, 1297 * we will have a non-zero datacnt. 1298 */ 1299 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) { 1300 /* ghost elements have a ghost size */ 1301 ASSERT(ab->b_buf == NULL); 1302 from_delta = ab->b_size; 1303 } 1304 ASSERT3U(*size, >=, from_delta); 1305 atomic_add_64(size, -from_delta); 1306 1307 if (use_mutex) 1308 mutex_exit(lock); 1309 } 1310 if (new_state != arc_anon) { 1311 int use_mutex; 1312 uint64_t *size = &new_state->arcs_lsize[ab->b_type]; 1313 1314 get_buf_info(ab, new_state, &list, &lock); 1315 use_mutex = !MUTEX_HELD(lock); 1316 if (use_mutex) 1317 mutex_enter(lock); 1318 1319 list_insert_head(list, ab); 1320 1321 /* ghost elements have a ghost size */ 1322 if (GHOST_STATE(new_state)) { 1323 ASSERT(ab->b_datacnt == 0); 1324 ASSERT(ab->b_buf == NULL); 1325 to_delta = ab->b_size; 1326 } 1327 atomic_add_64(size, to_delta); 1328 1329 if (use_mutex) 1330 mutex_exit(lock); 1331 } 1332 } 1333 1334 ASSERT(!BUF_EMPTY(ab)); 1335 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab)) 1336 buf_hash_remove(ab); 1337 1338 /* adjust state sizes */ 1339 if (to_delta) 1340 atomic_add_64(&new_state->arcs_size, to_delta); 1341 if (from_delta) { 1342 ASSERT3U(old_state->arcs_size, >=, from_delta); 1343 atomic_add_64(&old_state->arcs_size, -from_delta); 1344 } 1345 ab->b_state = new_state; 1346 1347 /* adjust l2arc hdr stats */ 1348 if (new_state == arc_l2c_only) 1349 l2arc_hdr_stat_add(); 1350 else if (old_state == arc_l2c_only) 1351 l2arc_hdr_stat_remove(); 1352} 1353 1354void 1355arc_space_consume(uint64_t space, arc_space_type_t type) 1356{ 1357 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1358 1359 switch (type) { 1360 case ARC_SPACE_DATA: 1361 ARCSTAT_INCR(arcstat_data_size, space); 1362 break; 1363 case ARC_SPACE_OTHER: 1364 ARCSTAT_INCR(arcstat_other_size, space); 1365 break; 1366 case ARC_SPACE_HDRS: 1367 ARCSTAT_INCR(arcstat_hdr_size, space); 1368 break; 1369 case ARC_SPACE_L2HDRS: 1370 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1371 break; 1372 } 1373 1374 atomic_add_64(&arc_meta_used, space); 1375 atomic_add_64(&arc_size, space); 1376} 1377 1378void 1379arc_space_return(uint64_t space, arc_space_type_t type) 1380{ 1381 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1382 1383 switch (type) { 1384 case ARC_SPACE_DATA: 1385 ARCSTAT_INCR(arcstat_data_size, -space); 1386 break; 1387 case ARC_SPACE_OTHER: 1388 ARCSTAT_INCR(arcstat_other_size, -space); 1389 break; 1390 case ARC_SPACE_HDRS: 1391 ARCSTAT_INCR(arcstat_hdr_size, -space); 1392 break; 1393 case ARC_SPACE_L2HDRS: 1394 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1395 break; 1396 } 1397 1398 ASSERT(arc_meta_used >= space); 1399 if (arc_meta_max < arc_meta_used) 1400 arc_meta_max = arc_meta_used; 1401 atomic_add_64(&arc_meta_used, -space); 1402 ASSERT(arc_size >= space); 1403 atomic_add_64(&arc_size, -space); 1404} 1405 1406void * 1407arc_data_buf_alloc(uint64_t size) 1408{ 1409 if (arc_evict_needed(ARC_BUFC_DATA)) 1410 cv_signal(&arc_reclaim_thr_cv); 1411 atomic_add_64(&arc_size, size); 1412 return (zio_data_buf_alloc(size)); 1413} 1414 1415void 1416arc_data_buf_free(void *buf, uint64_t size) 1417{ 1418 zio_data_buf_free(buf, size); 1419 ASSERT(arc_size >= size); 1420 atomic_add_64(&arc_size, -size); 1421} 1422 1423arc_buf_t * 1424arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1425{ 1426 arc_buf_hdr_t *hdr; 1427 arc_buf_t *buf; 1428 1429 ASSERT3U(size, >, 0); 1430 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1431 ASSERT(BUF_EMPTY(hdr)); 1432 hdr->b_size = size; 1433 hdr->b_type = type; 1434 hdr->b_spa = spa_load_guid(spa); 1435 hdr->b_state = arc_anon; 1436 hdr->b_arc_access = 0; 1437 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1438 buf->b_hdr = hdr; 1439 buf->b_data = NULL; 1440 buf->b_efunc = NULL; 1441 buf->b_private = NULL; 1442 buf->b_next = NULL; 1443 hdr->b_buf = buf; 1444 arc_get_data_buf(buf); 1445 hdr->b_datacnt = 1; 1446 hdr->b_flags = 0; 1447 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1448 (void) refcount_add(&hdr->b_refcnt, tag); 1449 1450 return (buf); 1451} 1452 1453static char *arc_onloan_tag = "onloan"; 1454 1455/* 1456 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1457 * flight data by arc_tempreserve_space() until they are "returned". Loaned 1458 * buffers must be returned to the arc before they can be used by the DMU or 1459 * freed. 1460 */ 1461arc_buf_t * 1462arc_loan_buf(spa_t *spa, int size) 1463{ 1464 arc_buf_t *buf; 1465 1466 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1467 1468 atomic_add_64(&arc_loaned_bytes, size); 1469 return (buf); 1470} 1471 1472/* 1473 * Return a loaned arc buffer to the arc. 1474 */ 1475void 1476arc_return_buf(arc_buf_t *buf, void *tag) 1477{ 1478 arc_buf_hdr_t *hdr = buf->b_hdr; 1479 1480 ASSERT(buf->b_data != NULL); 1481 (void) refcount_add(&hdr->b_refcnt, tag); 1482 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1483 1484 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1485} 1486 1487/* Detach an arc_buf from a dbuf (tag) */ 1488void 1489arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1490{ 1491 arc_buf_hdr_t *hdr; 1492 1493 ASSERT(buf->b_data != NULL); 1494 hdr = buf->b_hdr; 1495 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1496 (void) refcount_remove(&hdr->b_refcnt, tag); 1497 buf->b_efunc = NULL; 1498 buf->b_private = NULL; 1499 1500 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1501} 1502 1503static arc_buf_t * 1504arc_buf_clone(arc_buf_t *from) 1505{ 1506 arc_buf_t *buf; 1507 arc_buf_hdr_t *hdr = from->b_hdr; 1508 uint64_t size = hdr->b_size; 1509 1510 ASSERT(hdr->b_state != arc_anon); 1511 1512 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1513 buf->b_hdr = hdr; 1514 buf->b_data = NULL; 1515 buf->b_efunc = NULL; 1516 buf->b_private = NULL; 1517 buf->b_next = hdr->b_buf; 1518 hdr->b_buf = buf; 1519 arc_get_data_buf(buf); 1520 bcopy(from->b_data, buf->b_data, size); 1521 hdr->b_datacnt += 1; 1522 return (buf); 1523} 1524 1525void 1526arc_buf_add_ref(arc_buf_t *buf, void* tag) 1527{ 1528 arc_buf_hdr_t *hdr; 1529 kmutex_t *hash_lock; 1530 1531 /* 1532 * Check to see if this buffer is evicted. Callers 1533 * must verify b_data != NULL to know if the add_ref 1534 * was successful. 1535 */ 1536 mutex_enter(&buf->b_evict_lock); 1537 if (buf->b_data == NULL) { 1538 mutex_exit(&buf->b_evict_lock); 1539 return; 1540 } 1541 hash_lock = HDR_LOCK(buf->b_hdr); 1542 mutex_enter(hash_lock); 1543 hdr = buf->b_hdr; 1544 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1545 mutex_exit(&buf->b_evict_lock); 1546 1547 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1548 add_reference(hdr, hash_lock, tag); 1549 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1550 arc_access(hdr, hash_lock); 1551 mutex_exit(hash_lock); 1552 ARCSTAT_BUMP(arcstat_hits); 1553 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 1554 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1555 data, metadata, hits); 1556} 1557 1558/* 1559 * Free the arc data buffer. If it is an l2arc write in progress, 1560 * the buffer is placed on l2arc_free_on_write to be freed later. 1561 */ 1562static void 1563arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1564{ 1565 arc_buf_hdr_t *hdr = buf->b_hdr; 1566 1567 if (HDR_L2_WRITING(hdr)) { 1568 l2arc_data_free_t *df; 1569 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1570 df->l2df_data = buf->b_data; 1571 df->l2df_size = hdr->b_size; 1572 df->l2df_func = free_func; 1573 mutex_enter(&l2arc_free_on_write_mtx); 1574 list_insert_head(l2arc_free_on_write, df); 1575 mutex_exit(&l2arc_free_on_write_mtx); 1576 ARCSTAT_BUMP(arcstat_l2_free_on_write); 1577 } else { 1578 free_func(buf->b_data, hdr->b_size); 1579 } 1580} 1581 1582static void 1583arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all) 1584{ 1585 arc_buf_t **bufp; 1586 1587 /* free up data associated with the buf */ 1588 if (buf->b_data) { 1589 arc_state_t *state = buf->b_hdr->b_state; 1590 uint64_t size = buf->b_hdr->b_size; 1591 arc_buf_contents_t type = buf->b_hdr->b_type; 1592 1593 arc_cksum_verify(buf); 1594#ifdef illumos 1595 arc_buf_unwatch(buf); 1596#endif /* illumos */ 1597 1598 if (!recycle) { 1599 if (type == ARC_BUFC_METADATA) { 1600 arc_buf_data_free(buf, zio_buf_free); 1601 arc_space_return(size, ARC_SPACE_DATA); 1602 } else { 1603 ASSERT(type == ARC_BUFC_DATA); 1604 arc_buf_data_free(buf, zio_data_buf_free); 1605 ARCSTAT_INCR(arcstat_data_size, -size); 1606 atomic_add_64(&arc_size, -size); 1607 } 1608 } 1609 if (list_link_active(&buf->b_hdr->b_arc_node)) { 1610 uint64_t *cnt = &state->arcs_lsize[type]; 1611 1612 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1613 ASSERT(state != arc_anon); 1614 1615 ASSERT3U(*cnt, >=, size); 1616 atomic_add_64(cnt, -size); 1617 } 1618 ASSERT3U(state->arcs_size, >=, size); 1619 atomic_add_64(&state->arcs_size, -size); 1620 buf->b_data = NULL; 1621 ASSERT(buf->b_hdr->b_datacnt > 0); 1622 buf->b_hdr->b_datacnt -= 1; 1623 } 1624 1625 /* only remove the buf if requested */ 1626 if (!all) 1627 return; 1628 1629 /* remove the buf from the hdr list */ 1630 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1631 continue; 1632 *bufp = buf->b_next; 1633 buf->b_next = NULL; 1634 1635 ASSERT(buf->b_efunc == NULL); 1636 1637 /* clean up the buf */ 1638 buf->b_hdr = NULL; 1639 kmem_cache_free(buf_cache, buf); 1640} 1641 1642static void 1643arc_hdr_destroy(arc_buf_hdr_t *hdr) 1644{ 1645 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1646 ASSERT3P(hdr->b_state, ==, arc_anon); 1647 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1648 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1649 1650 if (l2hdr != NULL) { 1651 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1652 /* 1653 * To prevent arc_free() and l2arc_evict() from 1654 * attempting to free the same buffer at the same time, 1655 * a FREE_IN_PROGRESS flag is given to arc_free() to 1656 * give it priority. l2arc_evict() can't destroy this 1657 * header while we are waiting on l2arc_buflist_mtx. 1658 * 1659 * The hdr may be removed from l2ad_buflist before we 1660 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1661 */ 1662 if (!buflist_held) { 1663 mutex_enter(&l2arc_buflist_mtx); 1664 l2hdr = hdr->b_l2hdr; 1665 } 1666 1667 if (l2hdr != NULL) { 1668 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1669 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1670 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1671 if (hdr->b_state == arc_l2c_only) 1672 l2arc_hdr_stat_remove(); 1673 hdr->b_l2hdr = NULL; 1674 } 1675 1676 if (!buflist_held) 1677 mutex_exit(&l2arc_buflist_mtx); 1678 } 1679 1680 if (!BUF_EMPTY(hdr)) { 1681 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1682 buf_discard_identity(hdr); 1683 } 1684 while (hdr->b_buf) { 1685 arc_buf_t *buf = hdr->b_buf; 1686 1687 if (buf->b_efunc) { 1688 mutex_enter(&arc_eviction_mtx); 1689 mutex_enter(&buf->b_evict_lock); 1690 ASSERT(buf->b_hdr != NULL); 1691 arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1692 hdr->b_buf = buf->b_next; 1693 buf->b_hdr = &arc_eviction_hdr; 1694 buf->b_next = arc_eviction_list; 1695 arc_eviction_list = buf; 1696 mutex_exit(&buf->b_evict_lock); 1697 mutex_exit(&arc_eviction_mtx); 1698 } else { 1699 arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1700 } 1701 } 1702 if (hdr->b_freeze_cksum != NULL) { 1703 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1704 hdr->b_freeze_cksum = NULL; 1705 } 1706 if (hdr->b_thawed) { 1707 kmem_free(hdr->b_thawed, 1); 1708 hdr->b_thawed = NULL; 1709 } 1710 1711 ASSERT(!list_link_active(&hdr->b_arc_node)); 1712 ASSERT3P(hdr->b_hash_next, ==, NULL); 1713 ASSERT3P(hdr->b_acb, ==, NULL); 1714 kmem_cache_free(hdr_cache, hdr); 1715} 1716 1717void 1718arc_buf_free(arc_buf_t *buf, void *tag) 1719{ 1720 arc_buf_hdr_t *hdr = buf->b_hdr; 1721 int hashed = hdr->b_state != arc_anon; 1722 1723 ASSERT(buf->b_efunc == NULL); 1724 ASSERT(buf->b_data != NULL); 1725 1726 if (hashed) { 1727 kmutex_t *hash_lock = HDR_LOCK(hdr); 1728 1729 mutex_enter(hash_lock); 1730 hdr = buf->b_hdr; 1731 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1732 1733 (void) remove_reference(hdr, hash_lock, tag); 1734 if (hdr->b_datacnt > 1) { 1735 arc_buf_destroy(buf, FALSE, TRUE); 1736 } else { 1737 ASSERT(buf == hdr->b_buf); 1738 ASSERT(buf->b_efunc == NULL); 1739 hdr->b_flags |= ARC_BUF_AVAILABLE; 1740 } 1741 mutex_exit(hash_lock); 1742 } else if (HDR_IO_IN_PROGRESS(hdr)) { 1743 int destroy_hdr; 1744 /* 1745 * We are in the middle of an async write. Don't destroy 1746 * this buffer unless the write completes before we finish 1747 * decrementing the reference count. 1748 */ 1749 mutex_enter(&arc_eviction_mtx); 1750 (void) remove_reference(hdr, NULL, tag); 1751 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1752 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1753 mutex_exit(&arc_eviction_mtx); 1754 if (destroy_hdr) 1755 arc_hdr_destroy(hdr); 1756 } else { 1757 if (remove_reference(hdr, NULL, tag) > 0) 1758 arc_buf_destroy(buf, FALSE, TRUE); 1759 else 1760 arc_hdr_destroy(hdr); 1761 } 1762} 1763 1764int 1765arc_buf_remove_ref(arc_buf_t *buf, void* tag) 1766{ 1767 arc_buf_hdr_t *hdr = buf->b_hdr; 1768 kmutex_t *hash_lock = HDR_LOCK(hdr); 1769 int no_callback = (buf->b_efunc == NULL); 1770 1771 if (hdr->b_state == arc_anon) { 1772 ASSERT(hdr->b_datacnt == 1); 1773 arc_buf_free(buf, tag); 1774 return (no_callback); 1775 } 1776 1777 mutex_enter(hash_lock); 1778 hdr = buf->b_hdr; 1779 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1780 ASSERT(hdr->b_state != arc_anon); 1781 ASSERT(buf->b_data != NULL); 1782 1783 (void) remove_reference(hdr, hash_lock, tag); 1784 if (hdr->b_datacnt > 1) { 1785 if (no_callback) 1786 arc_buf_destroy(buf, FALSE, TRUE); 1787 } else if (no_callback) { 1788 ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1789 ASSERT(buf->b_efunc == NULL); 1790 hdr->b_flags |= ARC_BUF_AVAILABLE; 1791 } 1792 ASSERT(no_callback || hdr->b_datacnt > 1 || 1793 refcount_is_zero(&hdr->b_refcnt)); 1794 mutex_exit(hash_lock); 1795 return (no_callback); 1796} 1797 1798int 1799arc_buf_size(arc_buf_t *buf) 1800{ 1801 return (buf->b_hdr->b_size); 1802} 1803 1804/* 1805 * Evict buffers from list until we've removed the specified number of 1806 * bytes. Move the removed buffers to the appropriate evict state. 1807 * If the recycle flag is set, then attempt to "recycle" a buffer: 1808 * - look for a buffer to evict that is `bytes' long. 1809 * - return the data block from this buffer rather than freeing it. 1810 * This flag is used by callers that are trying to make space for a 1811 * new buffer in a full arc cache. 1812 * 1813 * This function makes a "best effort". It skips over any buffers 1814 * it can't get a hash_lock on, and so may not catch all candidates. 1815 * It may also return without evicting as much space as requested. 1816 */ 1817static void * 1818arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 1819 arc_buf_contents_t type) 1820{ 1821 arc_state_t *evicted_state; 1822 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 1823 int64_t bytes_remaining; 1824 arc_buf_hdr_t *ab, *ab_prev = NULL; 1825 list_t *evicted_list, *list, *evicted_list_start, *list_start; 1826 kmutex_t *lock, *evicted_lock; 1827 kmutex_t *hash_lock; 1828 boolean_t have_lock; 1829 void *stolen = NULL; 1830 static int evict_metadata_offset, evict_data_offset; 1831 int i, idx, offset, list_count, count; 1832 1833 ASSERT(state == arc_mru || state == arc_mfu); 1834 1835 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 1836 1837 if (type == ARC_BUFC_METADATA) { 1838 offset = 0; 1839 list_count = ARC_BUFC_NUMMETADATALISTS; 1840 list_start = &state->arcs_lists[0]; 1841 evicted_list_start = &evicted_state->arcs_lists[0]; 1842 idx = evict_metadata_offset; 1843 } else { 1844 offset = ARC_BUFC_NUMMETADATALISTS; 1845 list_start = &state->arcs_lists[offset]; 1846 evicted_list_start = &evicted_state->arcs_lists[offset]; 1847 list_count = ARC_BUFC_NUMDATALISTS; 1848 idx = evict_data_offset; 1849 } 1850 bytes_remaining = evicted_state->arcs_lsize[type]; 1851 count = 0; 1852 1853evict_start: 1854 list = &list_start[idx]; 1855 evicted_list = &evicted_list_start[idx]; 1856 lock = ARCS_LOCK(state, (offset + idx)); 1857 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 1858 1859 mutex_enter(lock); 1860 mutex_enter(evicted_lock); 1861 1862 for (ab = list_tail(list); ab; ab = ab_prev) { 1863 ab_prev = list_prev(list, ab); 1864 bytes_remaining -= (ab->b_size * ab->b_datacnt); 1865 /* prefetch buffers have a minimum lifespan */ 1866 if (HDR_IO_IN_PROGRESS(ab) || 1867 (spa && ab->b_spa != spa) || 1868 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 1869 ddi_get_lbolt() - ab->b_arc_access < 1870 arc_min_prefetch_lifespan)) { 1871 skipped++; 1872 continue; 1873 } 1874 /* "lookahead" for better eviction candidate */ 1875 if (recycle && ab->b_size != bytes && 1876 ab_prev && ab_prev->b_size == bytes) 1877 continue; 1878 hash_lock = HDR_LOCK(ab); 1879 have_lock = MUTEX_HELD(hash_lock); 1880 if (have_lock || mutex_tryenter(hash_lock)) { 1881 ASSERT0(refcount_count(&ab->b_refcnt)); 1882 ASSERT(ab->b_datacnt > 0); 1883 while (ab->b_buf) { 1884 arc_buf_t *buf = ab->b_buf; 1885 if (!mutex_tryenter(&buf->b_evict_lock)) { 1886 missed += 1; 1887 break; 1888 } 1889 if (buf->b_data) { 1890 bytes_evicted += ab->b_size; 1891 if (recycle && ab->b_type == type && 1892 ab->b_size == bytes && 1893 !HDR_L2_WRITING(ab)) { 1894 stolen = buf->b_data; 1895 recycle = FALSE; 1896 } 1897 } 1898 if (buf->b_efunc) { 1899 mutex_enter(&arc_eviction_mtx); 1900 arc_buf_destroy(buf, 1901 buf->b_data == stolen, FALSE); 1902 ab->b_buf = buf->b_next; 1903 buf->b_hdr = &arc_eviction_hdr; 1904 buf->b_next = arc_eviction_list; 1905 arc_eviction_list = buf; 1906 mutex_exit(&arc_eviction_mtx); 1907 mutex_exit(&buf->b_evict_lock); 1908 } else { 1909 mutex_exit(&buf->b_evict_lock); 1910 arc_buf_destroy(buf, 1911 buf->b_data == stolen, TRUE); 1912 } 1913 } 1914 1915 if (ab->b_l2hdr) { 1916 ARCSTAT_INCR(arcstat_evict_l2_cached, 1917 ab->b_size); 1918 } else { 1919 if (l2arc_write_eligible(ab->b_spa, ab)) { 1920 ARCSTAT_INCR(arcstat_evict_l2_eligible, 1921 ab->b_size); 1922 } else { 1923 ARCSTAT_INCR( 1924 arcstat_evict_l2_ineligible, 1925 ab->b_size); 1926 } 1927 } 1928 1929 if (ab->b_datacnt == 0) { 1930 arc_change_state(evicted_state, ab, hash_lock); 1931 ASSERT(HDR_IN_HASH_TABLE(ab)); 1932 ab->b_flags |= ARC_IN_HASH_TABLE; 1933 ab->b_flags &= ~ARC_BUF_AVAILABLE; 1934 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 1935 } 1936 if (!have_lock) 1937 mutex_exit(hash_lock); 1938 if (bytes >= 0 && bytes_evicted >= bytes) 1939 break; 1940 if (bytes_remaining > 0) { 1941 mutex_exit(evicted_lock); 1942 mutex_exit(lock); 1943 idx = ((idx + 1) & (list_count - 1)); 1944 count++; 1945 goto evict_start; 1946 } 1947 } else { 1948 missed += 1; 1949 } 1950 } 1951 1952 mutex_exit(evicted_lock); 1953 mutex_exit(lock); 1954 1955 idx = ((idx + 1) & (list_count - 1)); 1956 count++; 1957 1958 if (bytes_evicted < bytes) { 1959 if (count < list_count) 1960 goto evict_start; 1961 else 1962 dprintf("only evicted %lld bytes from %x", 1963 (longlong_t)bytes_evicted, state); 1964 } 1965 if (type == ARC_BUFC_METADATA) 1966 evict_metadata_offset = idx; 1967 else 1968 evict_data_offset = idx; 1969 1970 if (skipped) 1971 ARCSTAT_INCR(arcstat_evict_skip, skipped); 1972 1973 if (missed) 1974 ARCSTAT_INCR(arcstat_mutex_miss, missed); 1975 1976 /* 1977 * We have just evicted some date into the ghost state, make 1978 * sure we also adjust the ghost state size if necessary. 1979 */ 1980 if (arc_no_grow && 1981 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) { 1982 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size + 1983 arc_mru_ghost->arcs_size - arc_c; 1984 1985 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) { 1986 int64_t todelete = 1987 MIN(arc_mru_ghost->arcs_lsize[type], mru_over); 1988 arc_evict_ghost(arc_mru_ghost, 0, todelete); 1989 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) { 1990 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type], 1991 arc_mru_ghost->arcs_size + 1992 arc_mfu_ghost->arcs_size - arc_c); 1993 arc_evict_ghost(arc_mfu_ghost, 0, todelete); 1994 } 1995 } 1996 if (stolen) 1997 ARCSTAT_BUMP(arcstat_stolen); 1998 1999 return (stolen); 2000} 2001 2002/* 2003 * Remove buffers from list until we've removed the specified number of 2004 * bytes. Destroy the buffers that are removed. 2005 */ 2006static void 2007arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2008{ 2009 arc_buf_hdr_t *ab, *ab_prev; 2010 arc_buf_hdr_t marker = { 0 }; 2011 list_t *list, *list_start; 2012 kmutex_t *hash_lock, *lock; 2013 uint64_t bytes_deleted = 0; 2014 uint64_t bufs_skipped = 0; 2015 static int evict_offset; 2016 int list_count, idx = evict_offset; 2017 int offset, count = 0; 2018 2019 ASSERT(GHOST_STATE(state)); 2020 2021 /* 2022 * data lists come after metadata lists 2023 */ 2024 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2025 list_count = ARC_BUFC_NUMDATALISTS; 2026 offset = ARC_BUFC_NUMMETADATALISTS; 2027 2028evict_start: 2029 list = &list_start[idx]; 2030 lock = ARCS_LOCK(state, idx + offset); 2031 2032 mutex_enter(lock); 2033 for (ab = list_tail(list); ab; ab = ab_prev) { 2034 ab_prev = list_prev(list, ab); 2035 if (spa && ab->b_spa != spa) 2036 continue; 2037 2038 /* ignore markers */ 2039 if (ab->b_spa == 0) 2040 continue; 2041 2042 hash_lock = HDR_LOCK(ab); 2043 /* caller may be trying to modify this buffer, skip it */ 2044 if (MUTEX_HELD(hash_lock)) 2045 continue; 2046 if (mutex_tryenter(hash_lock)) { 2047 ASSERT(!HDR_IO_IN_PROGRESS(ab)); 2048 ASSERT(ab->b_buf == NULL); 2049 ARCSTAT_BUMP(arcstat_deleted); 2050 bytes_deleted += ab->b_size; 2051 2052 if (ab->b_l2hdr != NULL) { 2053 /* 2054 * This buffer is cached on the 2nd Level ARC; 2055 * don't destroy the header. 2056 */ 2057 arc_change_state(arc_l2c_only, ab, hash_lock); 2058 mutex_exit(hash_lock); 2059 } else { 2060 arc_change_state(arc_anon, ab, hash_lock); 2061 mutex_exit(hash_lock); 2062 arc_hdr_destroy(ab); 2063 } 2064 2065 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 2066 if (bytes >= 0 && bytes_deleted >= bytes) 2067 break; 2068 } else if (bytes < 0) { 2069 /* 2070 * Insert a list marker and then wait for the 2071 * hash lock to become available. Once its 2072 * available, restart from where we left off. 2073 */ 2074 list_insert_after(list, ab, &marker); 2075 mutex_exit(lock); 2076 mutex_enter(hash_lock); 2077 mutex_exit(hash_lock); 2078 mutex_enter(lock); 2079 ab_prev = list_prev(list, &marker); 2080 list_remove(list, &marker); 2081 } else 2082 bufs_skipped += 1; 2083 } 2084 mutex_exit(lock); 2085 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2086 count++; 2087 2088 if (count < list_count) 2089 goto evict_start; 2090 2091 evict_offset = idx; 2092 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2093 (bytes < 0 || bytes_deleted < bytes)) { 2094 list_start = &state->arcs_lists[0]; 2095 list_count = ARC_BUFC_NUMMETADATALISTS; 2096 offset = count = 0; 2097 goto evict_start; 2098 } 2099 2100 if (bufs_skipped) { 2101 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2102 ASSERT(bytes >= 0); 2103 } 2104 2105 if (bytes_deleted < bytes) 2106 dprintf("only deleted %lld bytes from %p", 2107 (longlong_t)bytes_deleted, state); 2108} 2109 2110static void 2111arc_adjust(void) 2112{ 2113 int64_t adjustment, delta; 2114 2115 /* 2116 * Adjust MRU size 2117 */ 2118 2119 adjustment = MIN((int64_t)(arc_size - arc_c), 2120 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2121 arc_p)); 2122 2123 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2124 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2125 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2126 adjustment -= delta; 2127 } 2128 2129 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2130 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2131 (void) arc_evict(arc_mru, 0, delta, FALSE, 2132 ARC_BUFC_METADATA); 2133 } 2134 2135 /* 2136 * Adjust MFU size 2137 */ 2138 2139 adjustment = arc_size - arc_c; 2140 2141 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2142 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2143 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2144 adjustment -= delta; 2145 } 2146 2147 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2148 int64_t delta = MIN(adjustment, 2149 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2150 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2151 ARC_BUFC_METADATA); 2152 } 2153 2154 /* 2155 * Adjust ghost lists 2156 */ 2157 2158 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2159 2160 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2161 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2162 arc_evict_ghost(arc_mru_ghost, 0, delta); 2163 } 2164 2165 adjustment = 2166 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2167 2168 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2169 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2170 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2171 } 2172} 2173 2174static void 2175arc_do_user_evicts(void) 2176{ 2177 static arc_buf_t *tmp_arc_eviction_list; 2178 2179 /* 2180 * Move list over to avoid LOR 2181 */ 2182restart: 2183 mutex_enter(&arc_eviction_mtx); 2184 tmp_arc_eviction_list = arc_eviction_list; 2185 arc_eviction_list = NULL; 2186 mutex_exit(&arc_eviction_mtx); 2187 2188 while (tmp_arc_eviction_list != NULL) { 2189 arc_buf_t *buf = tmp_arc_eviction_list; 2190 tmp_arc_eviction_list = buf->b_next; 2191 mutex_enter(&buf->b_evict_lock); 2192 buf->b_hdr = NULL; 2193 mutex_exit(&buf->b_evict_lock); 2194 2195 if (buf->b_efunc != NULL) 2196 VERIFY(buf->b_efunc(buf) == 0); 2197 2198 buf->b_efunc = NULL; 2199 buf->b_private = NULL; 2200 kmem_cache_free(buf_cache, buf); 2201 } 2202 2203 if (arc_eviction_list != NULL) 2204 goto restart; 2205} 2206 2207/* 2208 * Flush all *evictable* data from the cache for the given spa. 2209 * NOTE: this will not touch "active" (i.e. referenced) data. 2210 */ 2211void 2212arc_flush(spa_t *spa) 2213{ 2214 uint64_t guid = 0; 2215 2216 if (spa) 2217 guid = spa_load_guid(spa); 2218 2219 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2220 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2221 if (spa) 2222 break; 2223 } 2224 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2225 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2226 if (spa) 2227 break; 2228 } 2229 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2230 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2231 if (spa) 2232 break; 2233 } 2234 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2235 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2236 if (spa) 2237 break; 2238 } 2239 2240 arc_evict_ghost(arc_mru_ghost, guid, -1); 2241 arc_evict_ghost(arc_mfu_ghost, guid, -1); 2242 2243 mutex_enter(&arc_reclaim_thr_lock); 2244 arc_do_user_evicts(); 2245 mutex_exit(&arc_reclaim_thr_lock); 2246 ASSERT(spa || arc_eviction_list == NULL); 2247} 2248 2249void 2250arc_shrink(void) 2251{ 2252 if (arc_c > arc_c_min) { 2253 uint64_t to_free; 2254 2255#ifdef _KERNEL 2256 to_free = arc_c >> arc_shrink_shift; 2257#else 2258 to_free = arc_c >> arc_shrink_shift; 2259#endif 2260 if (arc_c > arc_c_min + to_free) 2261 atomic_add_64(&arc_c, -to_free); 2262 else 2263 arc_c = arc_c_min; 2264 2265 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2266 if (arc_c > arc_size) 2267 arc_c = MAX(arc_size, arc_c_min); 2268 if (arc_p > arc_c) 2269 arc_p = (arc_c >> 1); 2270 ASSERT(arc_c >= arc_c_min); 2271 ASSERT((int64_t)arc_p >= 0); 2272 } 2273 2274 if (arc_size > arc_c) 2275 arc_adjust(); 2276} 2277 2278static int needfree = 0; 2279 2280static int 2281arc_reclaim_needed(void) 2282{ 2283 2284#ifdef _KERNEL 2285 2286 if (needfree) 2287 return (1); 2288 2289 /* 2290 * Cooperate with pagedaemon when it's time for it to scan 2291 * and reclaim some pages. 2292 */ 2293 if (vm_paging_needed()) 2294 return (1); 2295 2296#ifdef sun 2297 /* 2298 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2299 */ 2300 extra = desfree; 2301 2302 /* 2303 * check that we're out of range of the pageout scanner. It starts to 2304 * schedule paging if freemem is less than lotsfree and needfree. 2305 * lotsfree is the high-water mark for pageout, and needfree is the 2306 * number of needed free pages. We add extra pages here to make sure 2307 * the scanner doesn't start up while we're freeing memory. 2308 */ 2309 if (freemem < lotsfree + needfree + extra) 2310 return (1); 2311 2312 /* 2313 * check to make sure that swapfs has enough space so that anon 2314 * reservations can still succeed. anon_resvmem() checks that the 2315 * availrmem is greater than swapfs_minfree, and the number of reserved 2316 * swap pages. We also add a bit of extra here just to prevent 2317 * circumstances from getting really dire. 2318 */ 2319 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2320 return (1); 2321 2322#if defined(__i386) 2323 /* 2324 * If we're on an i386 platform, it's possible that we'll exhaust the 2325 * kernel heap space before we ever run out of available physical 2326 * memory. Most checks of the size of the heap_area compare against 2327 * tune.t_minarmem, which is the minimum available real memory that we 2328 * can have in the system. However, this is generally fixed at 25 pages 2329 * which is so low that it's useless. In this comparison, we seek to 2330 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2331 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2332 * free) 2333 */ 2334 if (btop(vmem_size(heap_arena, VMEM_FREE)) < 2335 (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2)) 2336 return (1); 2337#endif 2338#else /* !sun */ 2339 if (kmem_used() > (kmem_size() * 3) / 4) 2340 return (1); 2341#endif /* sun */ 2342 2343#else 2344 if (spa_get_random(100) == 0) 2345 return (1); 2346#endif 2347 return (0); 2348} 2349 2350extern kmem_cache_t *zio_buf_cache[]; 2351extern kmem_cache_t *zio_data_buf_cache[]; 2352 2353static void 2354arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2355{ 2356 size_t i; 2357 kmem_cache_t *prev_cache = NULL; 2358 kmem_cache_t *prev_data_cache = NULL; 2359 2360#ifdef _KERNEL 2361 if (arc_meta_used >= arc_meta_limit) { 2362 /* 2363 * We are exceeding our meta-data cache limit. 2364 * Purge some DNLC entries to release holds on meta-data. 2365 */ 2366 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2367 } 2368#if defined(__i386) 2369 /* 2370 * Reclaim unused memory from all kmem caches. 2371 */ 2372 kmem_reap(); 2373#endif 2374#endif 2375 2376 /* 2377 * An aggressive reclamation will shrink the cache size as well as 2378 * reap free buffers from the arc kmem caches. 2379 */ 2380 if (strat == ARC_RECLAIM_AGGR) 2381 arc_shrink(); 2382 2383 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2384 if (zio_buf_cache[i] != prev_cache) { 2385 prev_cache = zio_buf_cache[i]; 2386 kmem_cache_reap_now(zio_buf_cache[i]); 2387 } 2388 if (zio_data_buf_cache[i] != prev_data_cache) { 2389 prev_data_cache = zio_data_buf_cache[i]; 2390 kmem_cache_reap_now(zio_data_buf_cache[i]); 2391 } 2392 } 2393 kmem_cache_reap_now(buf_cache); 2394 kmem_cache_reap_now(hdr_cache); 2395} 2396 2397static void 2398arc_reclaim_thread(void *dummy __unused) 2399{ 2400 clock_t growtime = 0; 2401 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2402 callb_cpr_t cpr; 2403 2404 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2405 2406 mutex_enter(&arc_reclaim_thr_lock); 2407 while (arc_thread_exit == 0) { 2408 if (arc_reclaim_needed()) { 2409 2410 if (arc_no_grow) { 2411 if (last_reclaim == ARC_RECLAIM_CONS) { 2412 last_reclaim = ARC_RECLAIM_AGGR; 2413 } else { 2414 last_reclaim = ARC_RECLAIM_CONS; 2415 } 2416 } else { 2417 arc_no_grow = TRUE; 2418 last_reclaim = ARC_RECLAIM_AGGR; 2419 membar_producer(); 2420 } 2421 2422 /* reset the growth delay for every reclaim */ 2423 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2424 2425 if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2426 /* 2427 * If needfree is TRUE our vm_lowmem hook 2428 * was called and in that case we must free some 2429 * memory, so switch to aggressive mode. 2430 */ 2431 arc_no_grow = TRUE; 2432 last_reclaim = ARC_RECLAIM_AGGR; 2433 } 2434 arc_kmem_reap_now(last_reclaim); 2435 arc_warm = B_TRUE; 2436 2437 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2438 arc_no_grow = FALSE; 2439 } 2440 2441 arc_adjust(); 2442 2443 if (arc_eviction_list != NULL) 2444 arc_do_user_evicts(); 2445 2446#ifdef _KERNEL 2447 if (needfree) { 2448 needfree = 0; 2449 wakeup(&needfree); 2450 } 2451#endif 2452 2453 /* block until needed, or one second, whichever is shorter */ 2454 CALLB_CPR_SAFE_BEGIN(&cpr); 2455 (void) cv_timedwait(&arc_reclaim_thr_cv, 2456 &arc_reclaim_thr_lock, hz); 2457 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2458 } 2459 2460 arc_thread_exit = 0; 2461 cv_broadcast(&arc_reclaim_thr_cv); 2462 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2463 thread_exit(); 2464} 2465 2466/* 2467 * Adapt arc info given the number of bytes we are trying to add and 2468 * the state that we are comming from. This function is only called 2469 * when we are adding new content to the cache. 2470 */ 2471static void 2472arc_adapt(int bytes, arc_state_t *state) 2473{ 2474 int mult; 2475 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2476 2477 if (state == arc_l2c_only) 2478 return; 2479 2480 ASSERT(bytes > 0); 2481 /* 2482 * Adapt the target size of the MRU list: 2483 * - if we just hit in the MRU ghost list, then increase 2484 * the target size of the MRU list. 2485 * - if we just hit in the MFU ghost list, then increase 2486 * the target size of the MFU list by decreasing the 2487 * target size of the MRU list. 2488 */ 2489 if (state == arc_mru_ghost) { 2490 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2491 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2492 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2493 2494 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2495 } else if (state == arc_mfu_ghost) { 2496 uint64_t delta; 2497 2498 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2499 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2500 mult = MIN(mult, 10); 2501 2502 delta = MIN(bytes * mult, arc_p); 2503 arc_p = MAX(arc_p_min, arc_p - delta); 2504 } 2505 ASSERT((int64_t)arc_p >= 0); 2506 2507 if (arc_reclaim_needed()) { 2508 cv_signal(&arc_reclaim_thr_cv); 2509 return; 2510 } 2511 2512 if (arc_no_grow) 2513 return; 2514 2515 if (arc_c >= arc_c_max) 2516 return; 2517 2518 /* 2519 * If we're within (2 * maxblocksize) bytes of the target 2520 * cache size, increment the target cache size 2521 */ 2522 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2523 atomic_add_64(&arc_c, (int64_t)bytes); 2524 if (arc_c > arc_c_max) 2525 arc_c = arc_c_max; 2526 else if (state == arc_anon) 2527 atomic_add_64(&arc_p, (int64_t)bytes); 2528 if (arc_p > arc_c) 2529 arc_p = arc_c; 2530 } 2531 ASSERT((int64_t)arc_p >= 0); 2532} 2533 2534/* 2535 * Check if the cache has reached its limits and eviction is required 2536 * prior to insert. 2537 */ 2538static int 2539arc_evict_needed(arc_buf_contents_t type) 2540{ 2541 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2542 return (1); 2543 2544#ifdef sun 2545#ifdef _KERNEL 2546 /* 2547 * If zio data pages are being allocated out of a separate heap segment, 2548 * then enforce that the size of available vmem for this area remains 2549 * above about 1/32nd free. 2550 */ 2551 if (type == ARC_BUFC_DATA && zio_arena != NULL && 2552 vmem_size(zio_arena, VMEM_FREE) < 2553 (vmem_size(zio_arena, VMEM_ALLOC) >> 5)) 2554 return (1); 2555#endif 2556#endif /* sun */ 2557 2558 if (arc_reclaim_needed()) 2559 return (1); 2560 2561 return (arc_size > arc_c); 2562} 2563 2564/* 2565 * The buffer, supplied as the first argument, needs a data block. 2566 * So, if we are at cache max, determine which cache should be victimized. 2567 * We have the following cases: 2568 * 2569 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2570 * In this situation if we're out of space, but the resident size of the MFU is 2571 * under the limit, victimize the MFU cache to satisfy this insertion request. 2572 * 2573 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2574 * Here, we've used up all of the available space for the MRU, so we need to 2575 * evict from our own cache instead. Evict from the set of resident MRU 2576 * entries. 2577 * 2578 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2579 * c minus p represents the MFU space in the cache, since p is the size of the 2580 * cache that is dedicated to the MRU. In this situation there's still space on 2581 * the MFU side, so the MRU side needs to be victimized. 2582 * 2583 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2584 * MFU's resident set is consuming more space than it has been allotted. In 2585 * this situation, we must victimize our own cache, the MFU, for this insertion. 2586 */ 2587static void 2588arc_get_data_buf(arc_buf_t *buf) 2589{ 2590 arc_state_t *state = buf->b_hdr->b_state; 2591 uint64_t size = buf->b_hdr->b_size; 2592 arc_buf_contents_t type = buf->b_hdr->b_type; 2593 2594 arc_adapt(size, state); 2595 2596 /* 2597 * We have not yet reached cache maximum size, 2598 * just allocate a new buffer. 2599 */ 2600 if (!arc_evict_needed(type)) { 2601 if (type == ARC_BUFC_METADATA) { 2602 buf->b_data = zio_buf_alloc(size); 2603 arc_space_consume(size, ARC_SPACE_DATA); 2604 } else { 2605 ASSERT(type == ARC_BUFC_DATA); 2606 buf->b_data = zio_data_buf_alloc(size); 2607 ARCSTAT_INCR(arcstat_data_size, size); 2608 atomic_add_64(&arc_size, size); 2609 } 2610 goto out; 2611 } 2612 2613 /* 2614 * If we are prefetching from the mfu ghost list, this buffer 2615 * will end up on the mru list; so steal space from there. 2616 */ 2617 if (state == arc_mfu_ghost) 2618 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2619 else if (state == arc_mru_ghost) 2620 state = arc_mru; 2621 2622 if (state == arc_mru || state == arc_anon) { 2623 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2624 state = (arc_mfu->arcs_lsize[type] >= size && 2625 arc_p > mru_used) ? arc_mfu : arc_mru; 2626 } else { 2627 /* MFU cases */ 2628 uint64_t mfu_space = arc_c - arc_p; 2629 state = (arc_mru->arcs_lsize[type] >= size && 2630 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2631 } 2632 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2633 if (type == ARC_BUFC_METADATA) { 2634 buf->b_data = zio_buf_alloc(size); 2635 arc_space_consume(size, ARC_SPACE_DATA); 2636 } else { 2637 ASSERT(type == ARC_BUFC_DATA); 2638 buf->b_data = zio_data_buf_alloc(size); 2639 ARCSTAT_INCR(arcstat_data_size, size); 2640 atomic_add_64(&arc_size, size); 2641 } 2642 ARCSTAT_BUMP(arcstat_recycle_miss); 2643 } 2644 ASSERT(buf->b_data != NULL); 2645out: 2646 /* 2647 * Update the state size. Note that ghost states have a 2648 * "ghost size" and so don't need to be updated. 2649 */ 2650 if (!GHOST_STATE(buf->b_hdr->b_state)) { 2651 arc_buf_hdr_t *hdr = buf->b_hdr; 2652 2653 atomic_add_64(&hdr->b_state->arcs_size, size); 2654 if (list_link_active(&hdr->b_arc_node)) { 2655 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2656 atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2657 } 2658 /* 2659 * If we are growing the cache, and we are adding anonymous 2660 * data, and we have outgrown arc_p, update arc_p 2661 */ 2662 if (arc_size < arc_c && hdr->b_state == arc_anon && 2663 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2664 arc_p = MIN(arc_c, arc_p + size); 2665 } 2666 ARCSTAT_BUMP(arcstat_allocated); 2667} 2668 2669/* 2670 * This routine is called whenever a buffer is accessed. 2671 * NOTE: the hash lock is dropped in this function. 2672 */ 2673static void 2674arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2675{ 2676 clock_t now; 2677 2678 ASSERT(MUTEX_HELD(hash_lock)); 2679 2680 if (buf->b_state == arc_anon) { 2681 /* 2682 * This buffer is not in the cache, and does not 2683 * appear in our "ghost" list. Add the new buffer 2684 * to the MRU state. 2685 */ 2686 2687 ASSERT(buf->b_arc_access == 0); 2688 buf->b_arc_access = ddi_get_lbolt(); 2689 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2690 arc_change_state(arc_mru, buf, hash_lock); 2691 2692 } else if (buf->b_state == arc_mru) { 2693 now = ddi_get_lbolt(); 2694 2695 /* 2696 * If this buffer is here because of a prefetch, then either: 2697 * - clear the flag if this is a "referencing" read 2698 * (any subsequent access will bump this into the MFU state). 2699 * or 2700 * - move the buffer to the head of the list if this is 2701 * another prefetch (to make it less likely to be evicted). 2702 */ 2703 if ((buf->b_flags & ARC_PREFETCH) != 0) { 2704 if (refcount_count(&buf->b_refcnt) == 0) { 2705 ASSERT(list_link_active(&buf->b_arc_node)); 2706 } else { 2707 buf->b_flags &= ~ARC_PREFETCH; 2708 ARCSTAT_BUMP(arcstat_mru_hits); 2709 } 2710 buf->b_arc_access = now; 2711 return; 2712 } 2713 2714 /* 2715 * This buffer has been "accessed" only once so far, 2716 * but it is still in the cache. Move it to the MFU 2717 * state. 2718 */ 2719 if (now > buf->b_arc_access + ARC_MINTIME) { 2720 /* 2721 * More than 125ms have passed since we 2722 * instantiated this buffer. Move it to the 2723 * most frequently used state. 2724 */ 2725 buf->b_arc_access = now; 2726 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2727 arc_change_state(arc_mfu, buf, hash_lock); 2728 } 2729 ARCSTAT_BUMP(arcstat_mru_hits); 2730 } else if (buf->b_state == arc_mru_ghost) { 2731 arc_state_t *new_state; 2732 /* 2733 * This buffer has been "accessed" recently, but 2734 * was evicted from the cache. Move it to the 2735 * MFU state. 2736 */ 2737 2738 if (buf->b_flags & ARC_PREFETCH) { 2739 new_state = arc_mru; 2740 if (refcount_count(&buf->b_refcnt) > 0) 2741 buf->b_flags &= ~ARC_PREFETCH; 2742 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2743 } else { 2744 new_state = arc_mfu; 2745 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2746 } 2747 2748 buf->b_arc_access = ddi_get_lbolt(); 2749 arc_change_state(new_state, buf, hash_lock); 2750 2751 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 2752 } else if (buf->b_state == arc_mfu) { 2753 /* 2754 * This buffer has been accessed more than once and is 2755 * still in the cache. Keep it in the MFU state. 2756 * 2757 * NOTE: an add_reference() that occurred when we did 2758 * the arc_read() will have kicked this off the list. 2759 * If it was a prefetch, we will explicitly move it to 2760 * the head of the list now. 2761 */ 2762 if ((buf->b_flags & ARC_PREFETCH) != 0) { 2763 ASSERT(refcount_count(&buf->b_refcnt) == 0); 2764 ASSERT(list_link_active(&buf->b_arc_node)); 2765 } 2766 ARCSTAT_BUMP(arcstat_mfu_hits); 2767 buf->b_arc_access = ddi_get_lbolt(); 2768 } else if (buf->b_state == arc_mfu_ghost) { 2769 arc_state_t *new_state = arc_mfu; 2770 /* 2771 * This buffer has been accessed more than once but has 2772 * been evicted from the cache. Move it back to the 2773 * MFU state. 2774 */ 2775 2776 if (buf->b_flags & ARC_PREFETCH) { 2777 /* 2778 * This is a prefetch access... 2779 * move this block back to the MRU state. 2780 */ 2781 ASSERT0(refcount_count(&buf->b_refcnt)); 2782 new_state = arc_mru; 2783 } 2784 2785 buf->b_arc_access = ddi_get_lbolt(); 2786 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2787 arc_change_state(new_state, buf, hash_lock); 2788 2789 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 2790 } else if (buf->b_state == arc_l2c_only) { 2791 /* 2792 * This buffer is on the 2nd Level ARC. 2793 */ 2794 2795 buf->b_arc_access = ddi_get_lbolt(); 2796 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2797 arc_change_state(arc_mfu, buf, hash_lock); 2798 } else { 2799 ASSERT(!"invalid arc state"); 2800 } 2801} 2802 2803/* a generic arc_done_func_t which you can use */ 2804/* ARGSUSED */ 2805void 2806arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 2807{ 2808 if (zio == NULL || zio->io_error == 0) 2809 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 2810 VERIFY(arc_buf_remove_ref(buf, arg) == 1); 2811} 2812 2813/* a generic arc_done_func_t */ 2814void 2815arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 2816{ 2817 arc_buf_t **bufp = arg; 2818 if (zio && zio->io_error) { 2819 VERIFY(arc_buf_remove_ref(buf, arg) == 1); 2820 *bufp = NULL; 2821 } else { 2822 *bufp = buf; 2823 ASSERT(buf->b_data); 2824 } 2825} 2826 2827static void 2828arc_read_done(zio_t *zio) 2829{ 2830 arc_buf_hdr_t *hdr, *found; 2831 arc_buf_t *buf; 2832 arc_buf_t *abuf; /* buffer we're assigning to callback */ 2833 kmutex_t *hash_lock; 2834 arc_callback_t *callback_list, *acb; 2835 int freeable = FALSE; 2836 2837 buf = zio->io_private; 2838 hdr = buf->b_hdr; 2839 2840 /* 2841 * The hdr was inserted into hash-table and removed from lists 2842 * prior to starting I/O. We should find this header, since 2843 * it's in the hash table, and it should be legit since it's 2844 * not possible to evict it during the I/O. The only possible 2845 * reason for it not to be found is if we were freed during the 2846 * read. 2847 */ 2848 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth, 2849 &hash_lock); 2850 2851 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) || 2852 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 2853 (found == hdr && HDR_L2_READING(hdr))); 2854 2855 hdr->b_flags &= ~ARC_L2_EVICTED; 2856 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 2857 hdr->b_flags &= ~ARC_L2CACHE; 2858 2859 /* byteswap if necessary */ 2860 callback_list = hdr->b_acb; 2861 ASSERT(callback_list != NULL); 2862 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 2863 dmu_object_byteswap_t bswap = 2864 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 2865 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 2866 byteswap_uint64_array : 2867 dmu_ot_byteswap[bswap].ob_func; 2868 func(buf->b_data, hdr->b_size); 2869 } 2870 2871 arc_cksum_compute(buf, B_FALSE); 2872#ifdef illumos 2873 arc_buf_watch(buf); 2874#endif /* illumos */ 2875 2876 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 2877 /* 2878 * Only call arc_access on anonymous buffers. This is because 2879 * if we've issued an I/O for an evicted buffer, we've already 2880 * called arc_access (to prevent any simultaneous readers from 2881 * getting confused). 2882 */ 2883 arc_access(hdr, hash_lock); 2884 } 2885 2886 /* create copies of the data buffer for the callers */ 2887 abuf = buf; 2888 for (acb = callback_list; acb; acb = acb->acb_next) { 2889 if (acb->acb_done) { 2890 if (abuf == NULL) 2891 abuf = arc_buf_clone(buf); 2892 acb->acb_buf = abuf; 2893 abuf = NULL; 2894 } 2895 } 2896 hdr->b_acb = NULL; 2897 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 2898 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 2899 if (abuf == buf) { 2900 ASSERT(buf->b_efunc == NULL); 2901 ASSERT(hdr->b_datacnt == 1); 2902 hdr->b_flags |= ARC_BUF_AVAILABLE; 2903 } 2904 2905 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 2906 2907 if (zio->io_error != 0) { 2908 hdr->b_flags |= ARC_IO_ERROR; 2909 if (hdr->b_state != arc_anon) 2910 arc_change_state(arc_anon, hdr, hash_lock); 2911 if (HDR_IN_HASH_TABLE(hdr)) 2912 buf_hash_remove(hdr); 2913 freeable = refcount_is_zero(&hdr->b_refcnt); 2914 } 2915 2916 /* 2917 * Broadcast before we drop the hash_lock to avoid the possibility 2918 * that the hdr (and hence the cv) might be freed before we get to 2919 * the cv_broadcast(). 2920 */ 2921 cv_broadcast(&hdr->b_cv); 2922 2923 if (hash_lock) { 2924 mutex_exit(hash_lock); 2925 } else { 2926 /* 2927 * This block was freed while we waited for the read to 2928 * complete. It has been removed from the hash table and 2929 * moved to the anonymous state (so that it won't show up 2930 * in the cache). 2931 */ 2932 ASSERT3P(hdr->b_state, ==, arc_anon); 2933 freeable = refcount_is_zero(&hdr->b_refcnt); 2934 } 2935 2936 /* execute each callback and free its structure */ 2937 while ((acb = callback_list) != NULL) { 2938 if (acb->acb_done) 2939 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 2940 2941 if (acb->acb_zio_dummy != NULL) { 2942 acb->acb_zio_dummy->io_error = zio->io_error; 2943 zio_nowait(acb->acb_zio_dummy); 2944 } 2945 2946 callback_list = acb->acb_next; 2947 kmem_free(acb, sizeof (arc_callback_t)); 2948 } 2949 2950 if (freeable) 2951 arc_hdr_destroy(hdr); 2952} 2953 2954/* 2955 * "Read" the block block at the specified DVA (in bp) via the 2956 * cache. If the block is found in the cache, invoke the provided 2957 * callback immediately and return. Note that the `zio' parameter 2958 * in the callback will be NULL in this case, since no IO was 2959 * required. If the block is not in the cache pass the read request 2960 * on to the spa with a substitute callback function, so that the 2961 * requested block will be added to the cache. 2962 * 2963 * If a read request arrives for a block that has a read in-progress, 2964 * either wait for the in-progress read to complete (and return the 2965 * results); or, if this is a read with a "done" func, add a record 2966 * to the read to invoke the "done" func when the read completes, 2967 * and return; or just return. 2968 * 2969 * arc_read_done() will invoke all the requested "done" functions 2970 * for readers of this block. 2971 * 2972 * Normal callers should use arc_read and pass the arc buffer and offset 2973 * for the bp. But if you know you don't need locking, you can use 2974 * arc_read_nolock. 2975 */ 2976int 2977arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_buf_t *pbuf, 2978 arc_done_func_t *done, void *private, int priority, int zio_flags, 2979 uint32_t *arc_flags, const zbookmark_t *zb) 2980{ 2981 int err; 2982 2983 if (pbuf == NULL) { 2984 /* 2985 * XXX This happens from traverse callback funcs, for 2986 * the objset_phys_t block. 2987 */ 2988 return (arc_read_nolock(pio, spa, bp, done, private, priority, 2989 zio_flags, arc_flags, zb)); 2990 } 2991 2992 ASSERT(!refcount_is_zero(&pbuf->b_hdr->b_refcnt)); 2993 ASSERT3U((char *)bp - (char *)pbuf->b_data, <, pbuf->b_hdr->b_size); 2994 rw_enter(&pbuf->b_data_lock, RW_READER); 2995 2996 err = arc_read_nolock(pio, spa, bp, done, private, priority, 2997 zio_flags, arc_flags, zb); 2998 rw_exit(&pbuf->b_data_lock); 2999 3000 return (err); 3001} 3002 3003int 3004arc_read_nolock(zio_t *pio, spa_t *spa, const blkptr_t *bp, 3005 arc_done_func_t *done, void *private, int priority, int zio_flags, 3006 uint32_t *arc_flags, const zbookmark_t *zb) 3007{ 3008 arc_buf_hdr_t *hdr; 3009 arc_buf_t *buf; 3010 kmutex_t *hash_lock; 3011 zio_t *rzio; 3012 uint64_t guid = spa_load_guid(spa); 3013 3014top: 3015 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp), 3016 &hash_lock); 3017 if (hdr && hdr->b_datacnt > 0) { 3018 3019 *arc_flags |= ARC_CACHED; 3020 3021 if (HDR_IO_IN_PROGRESS(hdr)) { 3022 3023 if (*arc_flags & ARC_WAIT) { 3024 cv_wait(&hdr->b_cv, hash_lock); 3025 mutex_exit(hash_lock); 3026 goto top; 3027 } 3028 ASSERT(*arc_flags & ARC_NOWAIT); 3029 3030 if (done) { 3031 arc_callback_t *acb = NULL; 3032 3033 acb = kmem_zalloc(sizeof (arc_callback_t), 3034 KM_SLEEP); 3035 acb->acb_done = done; 3036 acb->acb_private = private; 3037 if (pio != NULL) 3038 acb->acb_zio_dummy = zio_null(pio, 3039 spa, NULL, NULL, NULL, zio_flags); 3040 3041 ASSERT(acb->acb_done != NULL); 3042 acb->acb_next = hdr->b_acb; 3043 hdr->b_acb = acb; 3044 add_reference(hdr, hash_lock, private); 3045 mutex_exit(hash_lock); 3046 return (0); 3047 } 3048 mutex_exit(hash_lock); 3049 return (0); 3050 } 3051 3052 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3053 3054 if (done) { 3055 add_reference(hdr, hash_lock, private); 3056 /* 3057 * If this block is already in use, create a new 3058 * copy of the data so that we will be guaranteed 3059 * that arc_release() will always succeed. 3060 */ 3061 buf = hdr->b_buf; 3062 ASSERT(buf); 3063 ASSERT(buf->b_data); 3064 if (HDR_BUF_AVAILABLE(hdr)) { 3065 ASSERT(buf->b_efunc == NULL); 3066 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3067 } else { 3068 buf = arc_buf_clone(buf); 3069 } 3070 3071 } else if (*arc_flags & ARC_PREFETCH && 3072 refcount_count(&hdr->b_refcnt) == 0) { 3073 hdr->b_flags |= ARC_PREFETCH; 3074 } 3075 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3076 arc_access(hdr, hash_lock); 3077 if (*arc_flags & ARC_L2CACHE) 3078 hdr->b_flags |= ARC_L2CACHE; 3079 mutex_exit(hash_lock); 3080 ARCSTAT_BUMP(arcstat_hits); 3081 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3082 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3083 data, metadata, hits); 3084 3085 if (done) 3086 done(NULL, buf, private); 3087 } else { 3088 uint64_t size = BP_GET_LSIZE(bp); 3089 arc_callback_t *acb; 3090 vdev_t *vd = NULL; 3091 uint64_t addr; 3092 boolean_t devw = B_FALSE; 3093 3094 if (hdr == NULL) { 3095 /* this block is not in the cache */ 3096 arc_buf_hdr_t *exists; 3097 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3098 buf = arc_buf_alloc(spa, size, private, type); 3099 hdr = buf->b_hdr; 3100 hdr->b_dva = *BP_IDENTITY(bp); 3101 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3102 hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3103 exists = buf_hash_insert(hdr, &hash_lock); 3104 if (exists) { 3105 /* somebody beat us to the hash insert */ 3106 mutex_exit(hash_lock); 3107 buf_discard_identity(hdr); 3108 (void) arc_buf_remove_ref(buf, private); 3109 goto top; /* restart the IO request */ 3110 } 3111 /* if this is a prefetch, we don't have a reference */ 3112 if (*arc_flags & ARC_PREFETCH) { 3113 (void) remove_reference(hdr, hash_lock, 3114 private); 3115 hdr->b_flags |= ARC_PREFETCH; 3116 } 3117 if (*arc_flags & ARC_L2CACHE) 3118 hdr->b_flags |= ARC_L2CACHE; 3119 if (BP_GET_LEVEL(bp) > 0) 3120 hdr->b_flags |= ARC_INDIRECT; 3121 } else { 3122 /* this block is in the ghost cache */ 3123 ASSERT(GHOST_STATE(hdr->b_state)); 3124 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3125 ASSERT0(refcount_count(&hdr->b_refcnt)); 3126 ASSERT(hdr->b_buf == NULL); 3127 3128 /* if this is a prefetch, we don't have a reference */ 3129 if (*arc_flags & ARC_PREFETCH) 3130 hdr->b_flags |= ARC_PREFETCH; 3131 else 3132 add_reference(hdr, hash_lock, private); 3133 if (*arc_flags & ARC_L2CACHE) 3134 hdr->b_flags |= ARC_L2CACHE; 3135 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3136 buf->b_hdr = hdr; 3137 buf->b_data = NULL; 3138 buf->b_efunc = NULL; 3139 buf->b_private = NULL; 3140 buf->b_next = NULL; 3141 hdr->b_buf = buf; 3142 ASSERT(hdr->b_datacnt == 0); 3143 hdr->b_datacnt = 1; 3144 arc_get_data_buf(buf); 3145 arc_access(hdr, hash_lock); 3146 } 3147 3148 ASSERT(!GHOST_STATE(hdr->b_state)); 3149 3150 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3151 acb->acb_done = done; 3152 acb->acb_private = private; 3153 3154 ASSERT(hdr->b_acb == NULL); 3155 hdr->b_acb = acb; 3156 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3157 3158 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL && 3159 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3160 devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3161 addr = hdr->b_l2hdr->b_daddr; 3162 /* 3163 * Lock out device removal. 3164 */ 3165 if (vdev_is_dead(vd) || 3166 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3167 vd = NULL; 3168 } 3169 3170 mutex_exit(hash_lock); 3171 3172 ASSERT3U(hdr->b_size, ==, size); 3173 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3174 uint64_t, size, zbookmark_t *, zb); 3175 ARCSTAT_BUMP(arcstat_misses); 3176 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3177 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3178 data, metadata, misses); 3179#ifdef _KERNEL 3180 curthread->td_ru.ru_inblock++; 3181#endif 3182 3183 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3184 /* 3185 * Read from the L2ARC if the following are true: 3186 * 1. The L2ARC vdev was previously cached. 3187 * 2. This buffer still has L2ARC metadata. 3188 * 3. This buffer isn't currently writing to the L2ARC. 3189 * 4. The L2ARC entry wasn't evicted, which may 3190 * also have invalidated the vdev. 3191 * 5. This isn't prefetch and l2arc_noprefetch is set. 3192 */ 3193 if (hdr->b_l2hdr != NULL && 3194 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3195 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3196 l2arc_read_callback_t *cb; 3197 3198 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3199 ARCSTAT_BUMP(arcstat_l2_hits); 3200 3201 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3202 KM_SLEEP); 3203 cb->l2rcb_buf = buf; 3204 cb->l2rcb_spa = spa; 3205 cb->l2rcb_bp = *bp; 3206 cb->l2rcb_zb = *zb; 3207 cb->l2rcb_flags = zio_flags; 3208 3209 /* 3210 * l2arc read. The SCL_L2ARC lock will be 3211 * released by l2arc_read_done(). 3212 */ 3213 rzio = zio_read_phys(pio, vd, addr, size, 3214 buf->b_data, ZIO_CHECKSUM_OFF, 3215 l2arc_read_done, cb, priority, zio_flags | 3216 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL | 3217 ZIO_FLAG_DONT_PROPAGATE | 3218 ZIO_FLAG_DONT_RETRY, B_FALSE); 3219 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3220 zio_t *, rzio); 3221 ARCSTAT_INCR(arcstat_l2_read_bytes, size); 3222 3223 if (*arc_flags & ARC_NOWAIT) { 3224 zio_nowait(rzio); 3225 return (0); 3226 } 3227 3228 ASSERT(*arc_flags & ARC_WAIT); 3229 if (zio_wait(rzio) == 0) 3230 return (0); 3231 3232 /* l2arc read error; goto zio_read() */ 3233 } else { 3234 DTRACE_PROBE1(l2arc__miss, 3235 arc_buf_hdr_t *, hdr); 3236 ARCSTAT_BUMP(arcstat_l2_misses); 3237 if (HDR_L2_WRITING(hdr)) 3238 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3239 spa_config_exit(spa, SCL_L2ARC, vd); 3240 } 3241 } else { 3242 if (vd != NULL) 3243 spa_config_exit(spa, SCL_L2ARC, vd); 3244 if (l2arc_ndev != 0) { 3245 DTRACE_PROBE1(l2arc__miss, 3246 arc_buf_hdr_t *, hdr); 3247 ARCSTAT_BUMP(arcstat_l2_misses); 3248 } 3249 } 3250 3251 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3252 arc_read_done, buf, priority, zio_flags, zb); 3253 3254 if (*arc_flags & ARC_WAIT) 3255 return (zio_wait(rzio)); 3256 3257 ASSERT(*arc_flags & ARC_NOWAIT); 3258 zio_nowait(rzio); 3259 } 3260 return (0); 3261} 3262 3263void 3264arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3265{ 3266 ASSERT(buf->b_hdr != NULL); 3267 ASSERT(buf->b_hdr->b_state != arc_anon); 3268 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3269 ASSERT(buf->b_efunc == NULL); 3270 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3271 3272 buf->b_efunc = func; 3273 buf->b_private = private; 3274} 3275 3276/* 3277 * This is used by the DMU to let the ARC know that a buffer is 3278 * being evicted, so the ARC should clean up. If this arc buf 3279 * is not yet in the evicted state, it will be put there. 3280 */ 3281int 3282arc_buf_evict(arc_buf_t *buf) 3283{ 3284 arc_buf_hdr_t *hdr; 3285 kmutex_t *hash_lock; 3286 arc_buf_t **bufp; 3287 list_t *list, *evicted_list; 3288 kmutex_t *lock, *evicted_lock; 3289 3290 mutex_enter(&buf->b_evict_lock); 3291 hdr = buf->b_hdr; 3292 if (hdr == NULL) { 3293 /* 3294 * We are in arc_do_user_evicts(). 3295 */ 3296 ASSERT(buf->b_data == NULL); 3297 mutex_exit(&buf->b_evict_lock); 3298 return (0); 3299 } else if (buf->b_data == NULL) { 3300 arc_buf_t copy = *buf; /* structure assignment */ 3301 /* 3302 * We are on the eviction list; process this buffer now 3303 * but let arc_do_user_evicts() do the reaping. 3304 */ 3305 buf->b_efunc = NULL; 3306 mutex_exit(&buf->b_evict_lock); 3307 VERIFY(copy.b_efunc(©) == 0); 3308 return (1); 3309 } 3310 hash_lock = HDR_LOCK(hdr); 3311 mutex_enter(hash_lock); 3312 hdr = buf->b_hdr; 3313 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3314 3315 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3316 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3317 3318 /* 3319 * Pull this buffer off of the hdr 3320 */ 3321 bufp = &hdr->b_buf; 3322 while (*bufp != buf) 3323 bufp = &(*bufp)->b_next; 3324 *bufp = buf->b_next; 3325 3326 ASSERT(buf->b_data != NULL); 3327 arc_buf_destroy(buf, FALSE, FALSE); 3328 3329 if (hdr->b_datacnt == 0) { 3330 arc_state_t *old_state = hdr->b_state; 3331 arc_state_t *evicted_state; 3332 3333 ASSERT(hdr->b_buf == NULL); 3334 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 3335 3336 evicted_state = 3337 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3338 3339 get_buf_info(hdr, old_state, &list, &lock); 3340 get_buf_info(hdr, evicted_state, &evicted_list, &evicted_lock); 3341 mutex_enter(lock); 3342 mutex_enter(evicted_lock); 3343 3344 arc_change_state(evicted_state, hdr, hash_lock); 3345 ASSERT(HDR_IN_HASH_TABLE(hdr)); 3346 hdr->b_flags |= ARC_IN_HASH_TABLE; 3347 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3348 3349 mutex_exit(evicted_lock); 3350 mutex_exit(lock); 3351 } 3352 mutex_exit(hash_lock); 3353 mutex_exit(&buf->b_evict_lock); 3354 3355 VERIFY(buf->b_efunc(buf) == 0); 3356 buf->b_efunc = NULL; 3357 buf->b_private = NULL; 3358 buf->b_hdr = NULL; 3359 buf->b_next = NULL; 3360 kmem_cache_free(buf_cache, buf); 3361 return (1); 3362} 3363 3364/* 3365 * Release this buffer from the cache. This must be done 3366 * after a read and prior to modifying the buffer contents. 3367 * If the buffer has more than one reference, we must make 3368 * a new hdr for the buffer. 3369 */ 3370void 3371arc_release(arc_buf_t *buf, void *tag) 3372{ 3373 arc_buf_hdr_t *hdr; 3374 kmutex_t *hash_lock = NULL; 3375 l2arc_buf_hdr_t *l2hdr; 3376 uint64_t buf_size; 3377 3378 /* 3379 * It would be nice to assert that if it's DMU metadata (level > 3380 * 0 || it's the dnode file), then it must be syncing context. 3381 * But we don't know that information at this level. 3382 */ 3383 3384 mutex_enter(&buf->b_evict_lock); 3385 hdr = buf->b_hdr; 3386 3387 /* this buffer is not on any list */ 3388 ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3389 3390 if (hdr->b_state == arc_anon) { 3391 /* this buffer is already released */ 3392 ASSERT(buf->b_efunc == NULL); 3393 } else { 3394 hash_lock = HDR_LOCK(hdr); 3395 mutex_enter(hash_lock); 3396 hdr = buf->b_hdr; 3397 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3398 } 3399 3400 l2hdr = hdr->b_l2hdr; 3401 if (l2hdr) { 3402 mutex_enter(&l2arc_buflist_mtx); 3403 hdr->b_l2hdr = NULL; 3404 buf_size = hdr->b_size; 3405 } 3406 3407 /* 3408 * Do we have more than one buf? 3409 */ 3410 if (hdr->b_datacnt > 1) { 3411 arc_buf_hdr_t *nhdr; 3412 arc_buf_t **bufp; 3413 uint64_t blksz = hdr->b_size; 3414 uint64_t spa = hdr->b_spa; 3415 arc_buf_contents_t type = hdr->b_type; 3416 uint32_t flags = hdr->b_flags; 3417 3418 ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3419 /* 3420 * Pull the data off of this hdr and attach it to 3421 * a new anonymous hdr. 3422 */ 3423 (void) remove_reference(hdr, hash_lock, tag); 3424 bufp = &hdr->b_buf; 3425 while (*bufp != buf) 3426 bufp = &(*bufp)->b_next; 3427 *bufp = buf->b_next; 3428 buf->b_next = NULL; 3429 3430 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3431 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3432 if (refcount_is_zero(&hdr->b_refcnt)) { 3433 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3434 ASSERT3U(*size, >=, hdr->b_size); 3435 atomic_add_64(size, -hdr->b_size); 3436 } 3437 hdr->b_datacnt -= 1; 3438 arc_cksum_verify(buf); 3439#ifdef illumos 3440 arc_buf_unwatch(buf); 3441#endif /* illumos */ 3442 3443 mutex_exit(hash_lock); 3444 3445 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3446 nhdr->b_size = blksz; 3447 nhdr->b_spa = spa; 3448 nhdr->b_type = type; 3449 nhdr->b_buf = buf; 3450 nhdr->b_state = arc_anon; 3451 nhdr->b_arc_access = 0; 3452 nhdr->b_flags = flags & ARC_L2_WRITING; 3453 nhdr->b_l2hdr = NULL; 3454 nhdr->b_datacnt = 1; 3455 nhdr->b_freeze_cksum = NULL; 3456 (void) refcount_add(&nhdr->b_refcnt, tag); 3457 buf->b_hdr = nhdr; 3458 mutex_exit(&buf->b_evict_lock); 3459 atomic_add_64(&arc_anon->arcs_size, blksz); 3460 } else { 3461 mutex_exit(&buf->b_evict_lock); 3462 ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3463 ASSERT(!list_link_active(&hdr->b_arc_node)); 3464 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3465 if (hdr->b_state != arc_anon) 3466 arc_change_state(arc_anon, hdr, hash_lock); 3467 hdr->b_arc_access = 0; 3468 if (hash_lock) 3469 mutex_exit(hash_lock); 3470 3471 buf_discard_identity(hdr); 3472 arc_buf_thaw(buf); 3473 } 3474 buf->b_efunc = NULL; 3475 buf->b_private = NULL; 3476 3477 if (l2hdr) { 3478 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3479 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3480 ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3481 mutex_exit(&l2arc_buflist_mtx); 3482 } 3483} 3484 3485/* 3486 * Release this buffer. If it does not match the provided BP, fill it 3487 * with that block's contents. 3488 */ 3489/* ARGSUSED */ 3490int 3491arc_release_bp(arc_buf_t *buf, void *tag, blkptr_t *bp, spa_t *spa, 3492 zbookmark_t *zb) 3493{ 3494 arc_release(buf, tag); 3495 return (0); 3496} 3497 3498int 3499arc_released(arc_buf_t *buf) 3500{ 3501 int released; 3502 3503 mutex_enter(&buf->b_evict_lock); 3504 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3505 mutex_exit(&buf->b_evict_lock); 3506 return (released); 3507} 3508 3509int 3510arc_has_callback(arc_buf_t *buf) 3511{ 3512 int callback; 3513 3514 mutex_enter(&buf->b_evict_lock); 3515 callback = (buf->b_efunc != NULL); 3516 mutex_exit(&buf->b_evict_lock); 3517 return (callback); 3518} 3519 3520#ifdef ZFS_DEBUG 3521int 3522arc_referenced(arc_buf_t *buf) 3523{ 3524 int referenced; 3525 3526 mutex_enter(&buf->b_evict_lock); 3527 referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3528 mutex_exit(&buf->b_evict_lock); 3529 return (referenced); 3530} 3531#endif 3532 3533static void 3534arc_write_ready(zio_t *zio) 3535{ 3536 arc_write_callback_t *callback = zio->io_private; 3537 arc_buf_t *buf = callback->awcb_buf; 3538 arc_buf_hdr_t *hdr = buf->b_hdr; 3539 3540 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3541 callback->awcb_ready(zio, buf, callback->awcb_private); 3542 3543 /* 3544 * If the IO is already in progress, then this is a re-write 3545 * attempt, so we need to thaw and re-compute the cksum. 3546 * It is the responsibility of the callback to handle the 3547 * accounting for any re-write attempt. 3548 */ 3549 if (HDR_IO_IN_PROGRESS(hdr)) { 3550 mutex_enter(&hdr->b_freeze_lock); 3551 if (hdr->b_freeze_cksum != NULL) { 3552 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3553 hdr->b_freeze_cksum = NULL; 3554 } 3555 mutex_exit(&hdr->b_freeze_lock); 3556 } 3557 arc_cksum_compute(buf, B_FALSE); 3558 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3559} 3560 3561static void 3562arc_write_done(zio_t *zio) 3563{ 3564 arc_write_callback_t *callback = zio->io_private; 3565 arc_buf_t *buf = callback->awcb_buf; 3566 arc_buf_hdr_t *hdr = buf->b_hdr; 3567 3568 ASSERT(hdr->b_acb == NULL); 3569 3570 if (zio->io_error == 0) { 3571 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3572 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3573 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3574 } else { 3575 ASSERT(BUF_EMPTY(hdr)); 3576 } 3577 3578 /* 3579 * If the block to be written was all-zero, we may have 3580 * compressed it away. In this case no write was performed 3581 * so there will be no dva/birth/checksum. The buffer must 3582 * therefore remain anonymous (and uncached). 3583 */ 3584 if (!BUF_EMPTY(hdr)) { 3585 arc_buf_hdr_t *exists; 3586 kmutex_t *hash_lock; 3587 3588 ASSERT(zio->io_error == 0); 3589 3590 arc_cksum_verify(buf); 3591 3592 exists = buf_hash_insert(hdr, &hash_lock); 3593 if (exists) { 3594 /* 3595 * This can only happen if we overwrite for 3596 * sync-to-convergence, because we remove 3597 * buffers from the hash table when we arc_free(). 3598 */ 3599 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3600 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3601 panic("bad overwrite, hdr=%p exists=%p", 3602 (void *)hdr, (void *)exists); 3603 ASSERT(refcount_is_zero(&exists->b_refcnt)); 3604 arc_change_state(arc_anon, exists, hash_lock); 3605 mutex_exit(hash_lock); 3606 arc_hdr_destroy(exists); 3607 exists = buf_hash_insert(hdr, &hash_lock); 3608 ASSERT3P(exists, ==, NULL); 3609 } else { 3610 /* Dedup */ 3611 ASSERT(hdr->b_datacnt == 1); 3612 ASSERT(hdr->b_state == arc_anon); 3613 ASSERT(BP_GET_DEDUP(zio->io_bp)); 3614 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3615 } 3616 } 3617 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3618 /* if it's not anon, we are doing a scrub */ 3619 if (!exists && hdr->b_state == arc_anon) 3620 arc_access(hdr, hash_lock); 3621 mutex_exit(hash_lock); 3622 } else { 3623 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3624 } 3625 3626 ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3627 callback->awcb_done(zio, buf, callback->awcb_private); 3628 3629 kmem_free(callback, sizeof (arc_write_callback_t)); 3630} 3631 3632zio_t * 3633arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 3634 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, const zio_prop_t *zp, 3635 arc_done_func_t *ready, arc_done_func_t *done, void *private, 3636 int priority, int zio_flags, const zbookmark_t *zb) 3637{ 3638 arc_buf_hdr_t *hdr = buf->b_hdr; 3639 arc_write_callback_t *callback; 3640 zio_t *zio; 3641 3642 ASSERT(ready != NULL); 3643 ASSERT(done != NULL); 3644 ASSERT(!HDR_IO_ERROR(hdr)); 3645 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3646 ASSERT(hdr->b_acb == NULL); 3647 if (l2arc) 3648 hdr->b_flags |= ARC_L2CACHE; 3649 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 3650 callback->awcb_ready = ready; 3651 callback->awcb_done = done; 3652 callback->awcb_private = private; 3653 callback->awcb_buf = buf; 3654 3655 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 3656 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb); 3657 3658 return (zio); 3659} 3660 3661static int 3662arc_memory_throttle(uint64_t reserve, uint64_t inflight_data, uint64_t txg) 3663{ 3664#ifdef _KERNEL 3665 uint64_t available_memory = 3666 ptoa((uintmax_t)cnt.v_free_count + cnt.v_cache_count); 3667 static uint64_t page_load = 0; 3668 static uint64_t last_txg = 0; 3669 3670#ifdef sun 3671#if defined(__i386) 3672 available_memory = 3673 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 3674#endif 3675#endif /* sun */ 3676 if (available_memory >= zfs_write_limit_max) 3677 return (0); 3678 3679 if (txg > last_txg) { 3680 last_txg = txg; 3681 page_load = 0; 3682 } 3683 /* 3684 * If we are in pageout, we know that memory is already tight, 3685 * the arc is already going to be evicting, so we just want to 3686 * continue to let page writes occur as quickly as possible. 3687 */ 3688 if (curproc == pageproc) { 3689 if (page_load > available_memory / 4) 3690 return (ERESTART); 3691 /* Note: reserve is inflated, so we deflate */ 3692 page_load += reserve / 8; 3693 return (0); 3694 } else if (page_load > 0 && arc_reclaim_needed()) { 3695 /* memory is low, delay before restarting */ 3696 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3697 return (EAGAIN); 3698 } 3699 page_load = 0; 3700 3701 if (arc_size > arc_c_min) { 3702 uint64_t evictable_memory = 3703 arc_mru->arcs_lsize[ARC_BUFC_DATA] + 3704 arc_mru->arcs_lsize[ARC_BUFC_METADATA] + 3705 arc_mfu->arcs_lsize[ARC_BUFC_DATA] + 3706 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]; 3707 available_memory += MIN(evictable_memory, arc_size - arc_c_min); 3708 } 3709 3710 if (inflight_data > available_memory / 4) { 3711 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3712 return (ERESTART); 3713 } 3714#endif 3715 return (0); 3716} 3717 3718void 3719arc_tempreserve_clear(uint64_t reserve) 3720{ 3721 atomic_add_64(&arc_tempreserve, -reserve); 3722 ASSERT((int64_t)arc_tempreserve >= 0); 3723} 3724 3725int 3726arc_tempreserve_space(uint64_t reserve, uint64_t txg) 3727{ 3728 int error; 3729 uint64_t anon_size; 3730 3731#ifdef ZFS_DEBUG 3732 /* 3733 * Once in a while, fail for no reason. Everything should cope. 3734 */ 3735 if (spa_get_random(10000) == 0) { 3736 dprintf("forcing random failure\n"); 3737 return (ERESTART); 3738 } 3739#endif 3740 if (reserve > arc_c/4 && !arc_no_grow) 3741 arc_c = MIN(arc_c_max, reserve * 4); 3742 if (reserve > arc_c) 3743 return (ENOMEM); 3744 3745 /* 3746 * Don't count loaned bufs as in flight dirty data to prevent long 3747 * network delays from blocking transactions that are ready to be 3748 * assigned to a txg. 3749 */ 3750 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 3751 3752 /* 3753 * Writes will, almost always, require additional memory allocations 3754 * in order to compress/encrypt/etc the data. We therefor need to 3755 * make sure that there is sufficient available memory for this. 3756 */ 3757 if (error = arc_memory_throttle(reserve, anon_size, txg)) 3758 return (error); 3759 3760 /* 3761 * Throttle writes when the amount of dirty data in the cache 3762 * gets too large. We try to keep the cache less than half full 3763 * of dirty blocks so that our sync times don't grow too large. 3764 * Note: if two requests come in concurrently, we might let them 3765 * both succeed, when one of them should fail. Not a huge deal. 3766 */ 3767 3768 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 3769 anon_size > arc_c / 4) { 3770 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 3771 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 3772 arc_tempreserve>>10, 3773 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 3774 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 3775 reserve>>10, arc_c>>10); 3776 return (ERESTART); 3777 } 3778 atomic_add_64(&arc_tempreserve, reserve); 3779 return (0); 3780} 3781 3782static kmutex_t arc_lowmem_lock; 3783#ifdef _KERNEL 3784static eventhandler_tag arc_event_lowmem = NULL; 3785 3786static void 3787arc_lowmem(void *arg __unused, int howto __unused) 3788{ 3789 3790 /* Serialize access via arc_lowmem_lock. */ 3791 mutex_enter(&arc_lowmem_lock); 3792 mutex_enter(&arc_reclaim_thr_lock); 3793 needfree = 1; 3794 cv_signal(&arc_reclaim_thr_cv); 3795 3796 /* 3797 * It is unsafe to block here in arbitrary threads, because we can come 3798 * here from ARC itself and may hold ARC locks and thus risk a deadlock 3799 * with ARC reclaim thread. 3800 */ 3801 if (curproc == pageproc) { 3802 while (needfree) 3803 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 3804 } 3805 mutex_exit(&arc_reclaim_thr_lock); 3806 mutex_exit(&arc_lowmem_lock); 3807} 3808#endif 3809 3810void 3811arc_init(void) 3812{ 3813 int i, prefetch_tunable_set = 0; 3814 3815 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 3816 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 3817 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 3818 3819 /* Convert seconds to clock ticks */ 3820 arc_min_prefetch_lifespan = 1 * hz; 3821 3822 /* Start out with 1/8 of all memory */ 3823 arc_c = kmem_size() / 8; 3824 3825#ifdef sun 3826#ifdef _KERNEL 3827 /* 3828 * On architectures where the physical memory can be larger 3829 * than the addressable space (intel in 32-bit mode), we may 3830 * need to limit the cache to 1/8 of VM size. 3831 */ 3832 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 3833#endif 3834#endif /* sun */ 3835 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 3836 arc_c_min = MAX(arc_c / 4, 64<<18); 3837 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 3838 if (arc_c * 8 >= 1<<30) 3839 arc_c_max = (arc_c * 8) - (1<<30); 3840 else 3841 arc_c_max = arc_c_min; 3842 arc_c_max = MAX(arc_c * 5, arc_c_max); 3843 3844#ifdef _KERNEL 3845 /* 3846 * Allow the tunables to override our calculations if they are 3847 * reasonable (ie. over 16MB) 3848 */ 3849 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size()) 3850 arc_c_max = zfs_arc_max; 3851 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max) 3852 arc_c_min = zfs_arc_min; 3853#endif 3854 3855 arc_c = arc_c_max; 3856 arc_p = (arc_c >> 1); 3857 3858 /* limit meta-data to 1/4 of the arc capacity */ 3859 arc_meta_limit = arc_c_max / 4; 3860 3861 /* Allow the tunable to override if it is reasonable */ 3862 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 3863 arc_meta_limit = zfs_arc_meta_limit; 3864 3865 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 3866 arc_c_min = arc_meta_limit / 2; 3867 3868 if (zfs_arc_grow_retry > 0) 3869 arc_grow_retry = zfs_arc_grow_retry; 3870 3871 if (zfs_arc_shrink_shift > 0) 3872 arc_shrink_shift = zfs_arc_shrink_shift; 3873 3874 if (zfs_arc_p_min_shift > 0) 3875 arc_p_min_shift = zfs_arc_p_min_shift; 3876 3877 /* if kmem_flags are set, lets try to use less memory */ 3878 if (kmem_debugging()) 3879 arc_c = arc_c / 2; 3880 if (arc_c < arc_c_min) 3881 arc_c = arc_c_min; 3882 3883 zfs_arc_min = arc_c_min; 3884 zfs_arc_max = arc_c_max; 3885 3886 arc_anon = &ARC_anon; 3887 arc_mru = &ARC_mru; 3888 arc_mru_ghost = &ARC_mru_ghost; 3889 arc_mfu = &ARC_mfu; 3890 arc_mfu_ghost = &ARC_mfu_ghost; 3891 arc_l2c_only = &ARC_l2c_only; 3892 arc_size = 0; 3893 3894 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 3895 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 3896 NULL, MUTEX_DEFAULT, NULL); 3897 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 3898 NULL, MUTEX_DEFAULT, NULL); 3899 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 3900 NULL, MUTEX_DEFAULT, NULL); 3901 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 3902 NULL, MUTEX_DEFAULT, NULL); 3903 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 3904 NULL, MUTEX_DEFAULT, NULL); 3905 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 3906 NULL, MUTEX_DEFAULT, NULL); 3907 3908 list_create(&arc_mru->arcs_lists[i], 3909 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3910 list_create(&arc_mru_ghost->arcs_lists[i], 3911 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3912 list_create(&arc_mfu->arcs_lists[i], 3913 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3914 list_create(&arc_mfu_ghost->arcs_lists[i], 3915 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3916 list_create(&arc_mfu_ghost->arcs_lists[i], 3917 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3918 list_create(&arc_l2c_only->arcs_lists[i], 3919 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3920 } 3921 3922 buf_init(); 3923 3924 arc_thread_exit = 0; 3925 arc_eviction_list = NULL; 3926 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 3927 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 3928 3929 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 3930 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 3931 3932 if (arc_ksp != NULL) { 3933 arc_ksp->ks_data = &arc_stats; 3934 kstat_install(arc_ksp); 3935 } 3936 3937 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 3938 TS_RUN, minclsyspri); 3939 3940#ifdef _KERNEL 3941 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 3942 EVENTHANDLER_PRI_FIRST); 3943#endif 3944 3945 arc_dead = FALSE; 3946 arc_warm = B_FALSE; 3947 3948 if (zfs_write_limit_max == 0) 3949 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift; 3950 else 3951 zfs_write_limit_shift = 0; 3952 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL); 3953 3954#ifdef _KERNEL 3955 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 3956 prefetch_tunable_set = 1; 3957 3958#ifdef __i386__ 3959 if (prefetch_tunable_set == 0) { 3960 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 3961 "-- to enable,\n"); 3962 printf(" add \"vfs.zfs.prefetch_disable=0\" " 3963 "to /boot/loader.conf.\n"); 3964 zfs_prefetch_disable = 1; 3965 } 3966#else 3967 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 3968 prefetch_tunable_set == 0) { 3969 printf("ZFS NOTICE: Prefetch is disabled by default if less " 3970 "than 4GB of RAM is present;\n" 3971 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 3972 "to /boot/loader.conf.\n"); 3973 zfs_prefetch_disable = 1; 3974 } 3975#endif 3976 /* Warn about ZFS memory and address space requirements. */ 3977 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 3978 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 3979 "expect unstable behavior.\n"); 3980 } 3981 if (kmem_size() < 512 * (1 << 20)) { 3982 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 3983 "expect unstable behavior.\n"); 3984 printf(" Consider tuning vm.kmem_size and " 3985 "vm.kmem_size_max\n"); 3986 printf(" in /boot/loader.conf.\n"); 3987 } 3988#endif 3989} 3990 3991void 3992arc_fini(void) 3993{ 3994 int i; 3995 3996 mutex_enter(&arc_reclaim_thr_lock); 3997 arc_thread_exit = 1; 3998 cv_signal(&arc_reclaim_thr_cv); 3999 while (arc_thread_exit != 0) 4000 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4001 mutex_exit(&arc_reclaim_thr_lock); 4002 4003 arc_flush(NULL); 4004 4005 arc_dead = TRUE; 4006 4007 if (arc_ksp != NULL) { 4008 kstat_delete(arc_ksp); 4009 arc_ksp = NULL; 4010 } 4011 4012 mutex_destroy(&arc_eviction_mtx); 4013 mutex_destroy(&arc_reclaim_thr_lock); 4014 cv_destroy(&arc_reclaim_thr_cv); 4015 4016 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4017 list_destroy(&arc_mru->arcs_lists[i]); 4018 list_destroy(&arc_mru_ghost->arcs_lists[i]); 4019 list_destroy(&arc_mfu->arcs_lists[i]); 4020 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4021 list_destroy(&arc_l2c_only->arcs_lists[i]); 4022 4023 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4024 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4025 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4026 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4027 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4028 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4029 } 4030 4031 mutex_destroy(&zfs_write_limit_lock); 4032 4033 buf_fini(); 4034 4035 ASSERT(arc_loaned_bytes == 0); 4036 4037 mutex_destroy(&arc_lowmem_lock); 4038#ifdef _KERNEL 4039 if (arc_event_lowmem != NULL) 4040 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4041#endif 4042} 4043 4044/* 4045 * Level 2 ARC 4046 * 4047 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4048 * It uses dedicated storage devices to hold cached data, which are populated 4049 * using large infrequent writes. The main role of this cache is to boost 4050 * the performance of random read workloads. The intended L2ARC devices 4051 * include short-stroked disks, solid state disks, and other media with 4052 * substantially faster read latency than disk. 4053 * 4054 * +-----------------------+ 4055 * | ARC | 4056 * +-----------------------+ 4057 * | ^ ^ 4058 * | | | 4059 * l2arc_feed_thread() arc_read() 4060 * | | | 4061 * | l2arc read | 4062 * V | | 4063 * +---------------+ | 4064 * | L2ARC | | 4065 * +---------------+ | 4066 * | ^ | 4067 * l2arc_write() | | 4068 * | | | 4069 * V | | 4070 * +-------+ +-------+ 4071 * | vdev | | vdev | 4072 * | cache | | cache | 4073 * +-------+ +-------+ 4074 * +=========+ .-----. 4075 * : L2ARC : |-_____-| 4076 * : devices : | Disks | 4077 * +=========+ `-_____-' 4078 * 4079 * Read requests are satisfied from the following sources, in order: 4080 * 4081 * 1) ARC 4082 * 2) vdev cache of L2ARC devices 4083 * 3) L2ARC devices 4084 * 4) vdev cache of disks 4085 * 5) disks 4086 * 4087 * Some L2ARC device types exhibit extremely slow write performance. 4088 * To accommodate for this there are some significant differences between 4089 * the L2ARC and traditional cache design: 4090 * 4091 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4092 * the ARC behave as usual, freeing buffers and placing headers on ghost 4093 * lists. The ARC does not send buffers to the L2ARC during eviction as 4094 * this would add inflated write latencies for all ARC memory pressure. 4095 * 4096 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4097 * It does this by periodically scanning buffers from the eviction-end of 4098 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4099 * not already there. It scans until a headroom of buffers is satisfied, 4100 * which itself is a buffer for ARC eviction. The thread that does this is 4101 * l2arc_feed_thread(), illustrated below; example sizes are included to 4102 * provide a better sense of ratio than this diagram: 4103 * 4104 * head --> tail 4105 * +---------------------+----------+ 4106 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4107 * +---------------------+----------+ | o L2ARC eligible 4108 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4109 * +---------------------+----------+ | 4110 * 15.9 Gbytes ^ 32 Mbytes | 4111 * headroom | 4112 * l2arc_feed_thread() 4113 * | 4114 * l2arc write hand <--[oooo]--' 4115 * | 8 Mbyte 4116 * | write max 4117 * V 4118 * +==============================+ 4119 * L2ARC dev |####|#|###|###| |####| ... | 4120 * +==============================+ 4121 * 32 Gbytes 4122 * 4123 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4124 * evicted, then the L2ARC has cached a buffer much sooner than it probably 4125 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4126 * safe to say that this is an uncommon case, since buffers at the end of 4127 * the ARC lists have moved there due to inactivity. 4128 * 4129 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4130 * then the L2ARC simply misses copying some buffers. This serves as a 4131 * pressure valve to prevent heavy read workloads from both stalling the ARC 4132 * with waits and clogging the L2ARC with writes. This also helps prevent 4133 * the potential for the L2ARC to churn if it attempts to cache content too 4134 * quickly, such as during backups of the entire pool. 4135 * 4136 * 5. After system boot and before the ARC has filled main memory, there are 4137 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4138 * lists can remain mostly static. Instead of searching from tail of these 4139 * lists as pictured, the l2arc_feed_thread() will search from the list heads 4140 * for eligible buffers, greatly increasing its chance of finding them. 4141 * 4142 * The L2ARC device write speed is also boosted during this time so that 4143 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4144 * there are no L2ARC reads, and no fear of degrading read performance 4145 * through increased writes. 4146 * 4147 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4148 * the vdev queue can aggregate them into larger and fewer writes. Each 4149 * device is written to in a rotor fashion, sweeping writes through 4150 * available space then repeating. 4151 * 4152 * 7. The L2ARC does not store dirty content. It never needs to flush 4153 * write buffers back to disk based storage. 4154 * 4155 * 8. If an ARC buffer is written (and dirtied) which also exists in the 4156 * L2ARC, the now stale L2ARC buffer is immediately dropped. 4157 * 4158 * The performance of the L2ARC can be tweaked by a number of tunables, which 4159 * may be necessary for different workloads: 4160 * 4161 * l2arc_write_max max write bytes per interval 4162 * l2arc_write_boost extra write bytes during device warmup 4163 * l2arc_noprefetch skip caching prefetched buffers 4164 * l2arc_headroom number of max device writes to precache 4165 * l2arc_feed_secs seconds between L2ARC writing 4166 * 4167 * Tunables may be removed or added as future performance improvements are 4168 * integrated, and also may become zpool properties. 4169 * 4170 * There are three key functions that control how the L2ARC warms up: 4171 * 4172 * l2arc_write_eligible() check if a buffer is eligible to cache 4173 * l2arc_write_size() calculate how much to write 4174 * l2arc_write_interval() calculate sleep delay between writes 4175 * 4176 * These three functions determine what to write, how much, and how quickly 4177 * to send writes. 4178 */ 4179 4180static boolean_t 4181l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab) 4182{ 4183 /* 4184 * A buffer is *not* eligible for the L2ARC if it: 4185 * 1. belongs to a different spa. 4186 * 2. is already cached on the L2ARC. 4187 * 3. has an I/O in progress (it may be an incomplete read). 4188 * 4. is flagged not eligible (zfs property). 4189 */ 4190 if (ab->b_spa != spa_guid) { 4191 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4192 return (B_FALSE); 4193 } 4194 if (ab->b_l2hdr != NULL) { 4195 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4196 return (B_FALSE); 4197 } 4198 if (HDR_IO_IN_PROGRESS(ab)) { 4199 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4200 return (B_FALSE); 4201 } 4202 if (!HDR_L2CACHE(ab)) { 4203 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4204 return (B_FALSE); 4205 } 4206 4207 return (B_TRUE); 4208} 4209 4210static uint64_t 4211l2arc_write_size(l2arc_dev_t *dev) 4212{ 4213 uint64_t size; 4214 4215 size = dev->l2ad_write; 4216 4217 if (arc_warm == B_FALSE) 4218 size += dev->l2ad_boost; 4219 4220 return (size); 4221 4222} 4223 4224static clock_t 4225l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4226{ 4227 clock_t interval, next, now; 4228 4229 /* 4230 * If the ARC lists are busy, increase our write rate; if the 4231 * lists are stale, idle back. This is achieved by checking 4232 * how much we previously wrote - if it was more than half of 4233 * what we wanted, schedule the next write much sooner. 4234 */ 4235 if (l2arc_feed_again && wrote > (wanted / 2)) 4236 interval = (hz * l2arc_feed_min_ms) / 1000; 4237 else 4238 interval = hz * l2arc_feed_secs; 4239 4240 now = ddi_get_lbolt(); 4241 next = MAX(now, MIN(now + interval, began + interval)); 4242 4243 return (next); 4244} 4245 4246static void 4247l2arc_hdr_stat_add(void) 4248{ 4249 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4250 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4251} 4252 4253static void 4254l2arc_hdr_stat_remove(void) 4255{ 4256 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4257 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4258} 4259 4260/* 4261 * Cycle through L2ARC devices. This is how L2ARC load balances. 4262 * If a device is returned, this also returns holding the spa config lock. 4263 */ 4264static l2arc_dev_t * 4265l2arc_dev_get_next(void) 4266{ 4267 l2arc_dev_t *first, *next = NULL; 4268 4269 /* 4270 * Lock out the removal of spas (spa_namespace_lock), then removal 4271 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4272 * both locks will be dropped and a spa config lock held instead. 4273 */ 4274 mutex_enter(&spa_namespace_lock); 4275 mutex_enter(&l2arc_dev_mtx); 4276 4277 /* if there are no vdevs, there is nothing to do */ 4278 if (l2arc_ndev == 0) 4279 goto out; 4280 4281 first = NULL; 4282 next = l2arc_dev_last; 4283 do { 4284 /* loop around the list looking for a non-faulted vdev */ 4285 if (next == NULL) { 4286 next = list_head(l2arc_dev_list); 4287 } else { 4288 next = list_next(l2arc_dev_list, next); 4289 if (next == NULL) 4290 next = list_head(l2arc_dev_list); 4291 } 4292 4293 /* if we have come back to the start, bail out */ 4294 if (first == NULL) 4295 first = next; 4296 else if (next == first) 4297 break; 4298 4299 } while (vdev_is_dead(next->l2ad_vdev)); 4300 4301 /* if we were unable to find any usable vdevs, return NULL */ 4302 if (vdev_is_dead(next->l2ad_vdev)) 4303 next = NULL; 4304 4305 l2arc_dev_last = next; 4306 4307out: 4308 mutex_exit(&l2arc_dev_mtx); 4309 4310 /* 4311 * Grab the config lock to prevent the 'next' device from being 4312 * removed while we are writing to it. 4313 */ 4314 if (next != NULL) 4315 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4316 mutex_exit(&spa_namespace_lock); 4317 4318 return (next); 4319} 4320 4321/* 4322 * Free buffers that were tagged for destruction. 4323 */ 4324static void 4325l2arc_do_free_on_write() 4326{ 4327 list_t *buflist; 4328 l2arc_data_free_t *df, *df_prev; 4329 4330 mutex_enter(&l2arc_free_on_write_mtx); 4331 buflist = l2arc_free_on_write; 4332 4333 for (df = list_tail(buflist); df; df = df_prev) { 4334 df_prev = list_prev(buflist, df); 4335 ASSERT(df->l2df_data != NULL); 4336 ASSERT(df->l2df_func != NULL); 4337 df->l2df_func(df->l2df_data, df->l2df_size); 4338 list_remove(buflist, df); 4339 kmem_free(df, sizeof (l2arc_data_free_t)); 4340 } 4341 4342 mutex_exit(&l2arc_free_on_write_mtx); 4343} 4344 4345/* 4346 * A write to a cache device has completed. Update all headers to allow 4347 * reads from these buffers to begin. 4348 */ 4349static void 4350l2arc_write_done(zio_t *zio) 4351{ 4352 l2arc_write_callback_t *cb; 4353 l2arc_dev_t *dev; 4354 list_t *buflist; 4355 arc_buf_hdr_t *head, *ab, *ab_prev; 4356 l2arc_buf_hdr_t *abl2; 4357 kmutex_t *hash_lock; 4358 4359 cb = zio->io_private; 4360 ASSERT(cb != NULL); 4361 dev = cb->l2wcb_dev; 4362 ASSERT(dev != NULL); 4363 head = cb->l2wcb_head; 4364 ASSERT(head != NULL); 4365 buflist = dev->l2ad_buflist; 4366 ASSERT(buflist != NULL); 4367 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4368 l2arc_write_callback_t *, cb); 4369 4370 if (zio->io_error != 0) 4371 ARCSTAT_BUMP(arcstat_l2_writes_error); 4372 4373 mutex_enter(&l2arc_buflist_mtx); 4374 4375 /* 4376 * All writes completed, or an error was hit. 4377 */ 4378 for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 4379 ab_prev = list_prev(buflist, ab); 4380 4381 hash_lock = HDR_LOCK(ab); 4382 if (!mutex_tryenter(hash_lock)) { 4383 /* 4384 * This buffer misses out. It may be in a stage 4385 * of eviction. Its ARC_L2_WRITING flag will be 4386 * left set, denying reads to this buffer. 4387 */ 4388 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4389 continue; 4390 } 4391 4392 if (zio->io_error != 0) { 4393 /* 4394 * Error - drop L2ARC entry. 4395 */ 4396 list_remove(buflist, ab); 4397 abl2 = ab->b_l2hdr; 4398 ab->b_l2hdr = NULL; 4399 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4400 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4401 } 4402 4403 /* 4404 * Allow ARC to begin reads to this L2ARC entry. 4405 */ 4406 ab->b_flags &= ~ARC_L2_WRITING; 4407 4408 mutex_exit(hash_lock); 4409 } 4410 4411 atomic_inc_64(&l2arc_writes_done); 4412 list_remove(buflist, head); 4413 kmem_cache_free(hdr_cache, head); 4414 mutex_exit(&l2arc_buflist_mtx); 4415 4416 l2arc_do_free_on_write(); 4417 4418 kmem_free(cb, sizeof (l2arc_write_callback_t)); 4419} 4420 4421/* 4422 * A read to a cache device completed. Validate buffer contents before 4423 * handing over to the regular ARC routines. 4424 */ 4425static void 4426l2arc_read_done(zio_t *zio) 4427{ 4428 l2arc_read_callback_t *cb; 4429 arc_buf_hdr_t *hdr; 4430 arc_buf_t *buf; 4431 kmutex_t *hash_lock; 4432 int equal; 4433 4434 ASSERT(zio->io_vd != NULL); 4435 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4436 4437 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4438 4439 cb = zio->io_private; 4440 ASSERT(cb != NULL); 4441 buf = cb->l2rcb_buf; 4442 ASSERT(buf != NULL); 4443 4444 hash_lock = HDR_LOCK(buf->b_hdr); 4445 mutex_enter(hash_lock); 4446 hdr = buf->b_hdr; 4447 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4448 4449 /* 4450 * Check this survived the L2ARC journey. 4451 */ 4452 equal = arc_cksum_equal(buf); 4453 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4454 mutex_exit(hash_lock); 4455 zio->io_private = buf; 4456 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4457 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4458 arc_read_done(zio); 4459 } else { 4460 mutex_exit(hash_lock); 4461 /* 4462 * Buffer didn't survive caching. Increment stats and 4463 * reissue to the original storage device. 4464 */ 4465 if (zio->io_error != 0) { 4466 ARCSTAT_BUMP(arcstat_l2_io_error); 4467 } else { 4468 zio->io_error = EIO; 4469 } 4470 if (!equal) 4471 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4472 4473 /* 4474 * If there's no waiter, issue an async i/o to the primary 4475 * storage now. If there *is* a waiter, the caller must 4476 * issue the i/o in a context where it's OK to block. 4477 */ 4478 if (zio->io_waiter == NULL) { 4479 zio_t *pio = zio_unique_parent(zio); 4480 4481 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4482 4483 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4484 buf->b_data, zio->io_size, arc_read_done, buf, 4485 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4486 } 4487 } 4488 4489 kmem_free(cb, sizeof (l2arc_read_callback_t)); 4490} 4491 4492/* 4493 * This is the list priority from which the L2ARC will search for pages to 4494 * cache. This is used within loops (0..3) to cycle through lists in the 4495 * desired order. This order can have a significant effect on cache 4496 * performance. 4497 * 4498 * Currently the metadata lists are hit first, MFU then MRU, followed by 4499 * the data lists. This function returns a locked list, and also returns 4500 * the lock pointer. 4501 */ 4502static list_t * 4503l2arc_list_locked(int list_num, kmutex_t **lock) 4504{ 4505 list_t *list; 4506 int idx; 4507 4508 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4509 4510 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4511 idx = list_num; 4512 list = &arc_mfu->arcs_lists[idx]; 4513 *lock = ARCS_LOCK(arc_mfu, idx); 4514 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4515 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4516 list = &arc_mru->arcs_lists[idx]; 4517 *lock = ARCS_LOCK(arc_mru, idx); 4518 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4519 ARC_BUFC_NUMDATALISTS)) { 4520 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4521 list = &arc_mfu->arcs_lists[idx]; 4522 *lock = ARCS_LOCK(arc_mfu, idx); 4523 } else { 4524 idx = list_num - ARC_BUFC_NUMLISTS; 4525 list = &arc_mru->arcs_lists[idx]; 4526 *lock = ARCS_LOCK(arc_mru, idx); 4527 } 4528 4529 ASSERT(!(MUTEX_HELD(*lock))); 4530 mutex_enter(*lock); 4531 return (list); 4532} 4533 4534/* 4535 * Evict buffers from the device write hand to the distance specified in 4536 * bytes. This distance may span populated buffers, it may span nothing. 4537 * This is clearing a region on the L2ARC device ready for writing. 4538 * If the 'all' boolean is set, every buffer is evicted. 4539 */ 4540static void 4541l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4542{ 4543 list_t *buflist; 4544 l2arc_buf_hdr_t *abl2; 4545 arc_buf_hdr_t *ab, *ab_prev; 4546 kmutex_t *hash_lock; 4547 uint64_t taddr; 4548 4549 buflist = dev->l2ad_buflist; 4550 4551 if (buflist == NULL) 4552 return; 4553 4554 if (!all && dev->l2ad_first) { 4555 /* 4556 * This is the first sweep through the device. There is 4557 * nothing to evict. 4558 */ 4559 return; 4560 } 4561 4562 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4563 /* 4564 * When nearing the end of the device, evict to the end 4565 * before the device write hand jumps to the start. 4566 */ 4567 taddr = dev->l2ad_end; 4568 } else { 4569 taddr = dev->l2ad_hand + distance; 4570 } 4571 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4572 uint64_t, taddr, boolean_t, all); 4573 4574top: 4575 mutex_enter(&l2arc_buflist_mtx); 4576 for (ab = list_tail(buflist); ab; ab = ab_prev) { 4577 ab_prev = list_prev(buflist, ab); 4578 4579 hash_lock = HDR_LOCK(ab); 4580 if (!mutex_tryenter(hash_lock)) { 4581 /* 4582 * Missed the hash lock. Retry. 4583 */ 4584 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4585 mutex_exit(&l2arc_buflist_mtx); 4586 mutex_enter(hash_lock); 4587 mutex_exit(hash_lock); 4588 goto top; 4589 } 4590 4591 if (HDR_L2_WRITE_HEAD(ab)) { 4592 /* 4593 * We hit a write head node. Leave it for 4594 * l2arc_write_done(). 4595 */ 4596 list_remove(buflist, ab); 4597 mutex_exit(hash_lock); 4598 continue; 4599 } 4600 4601 if (!all && ab->b_l2hdr != NULL && 4602 (ab->b_l2hdr->b_daddr > taddr || 4603 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4604 /* 4605 * We've evicted to the target address, 4606 * or the end of the device. 4607 */ 4608 mutex_exit(hash_lock); 4609 break; 4610 } 4611 4612 if (HDR_FREE_IN_PROGRESS(ab)) { 4613 /* 4614 * Already on the path to destruction. 4615 */ 4616 mutex_exit(hash_lock); 4617 continue; 4618 } 4619 4620 if (ab->b_state == arc_l2c_only) { 4621 ASSERT(!HDR_L2_READING(ab)); 4622 /* 4623 * This doesn't exist in the ARC. Destroy. 4624 * arc_hdr_destroy() will call list_remove() 4625 * and decrement arcstat_l2_size. 4626 */ 4627 arc_change_state(arc_anon, ab, hash_lock); 4628 arc_hdr_destroy(ab); 4629 } else { 4630 /* 4631 * Invalidate issued or about to be issued 4632 * reads, since we may be about to write 4633 * over this location. 4634 */ 4635 if (HDR_L2_READING(ab)) { 4636 ARCSTAT_BUMP(arcstat_l2_evict_reading); 4637 ab->b_flags |= ARC_L2_EVICTED; 4638 } 4639 4640 /* 4641 * Tell ARC this no longer exists in L2ARC. 4642 */ 4643 if (ab->b_l2hdr != NULL) { 4644 abl2 = ab->b_l2hdr; 4645 ab->b_l2hdr = NULL; 4646 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4647 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4648 } 4649 list_remove(buflist, ab); 4650 4651 /* 4652 * This may have been leftover after a 4653 * failed write. 4654 */ 4655 ab->b_flags &= ~ARC_L2_WRITING; 4656 } 4657 mutex_exit(hash_lock); 4658 } 4659 mutex_exit(&l2arc_buflist_mtx); 4660 4661 vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0); 4662 dev->l2ad_evict = taddr; 4663} 4664 4665/* 4666 * Find and write ARC buffers to the L2ARC device. 4667 * 4668 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 4669 * for reading until they have completed writing. 4670 */ 4671static uint64_t 4672l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) 4673{ 4674 arc_buf_hdr_t *ab, *ab_prev, *head; 4675 l2arc_buf_hdr_t *hdrl2; 4676 list_t *list; 4677 uint64_t passed_sz, write_sz, buf_sz, headroom; 4678 void *buf_data; 4679 kmutex_t *hash_lock, *list_lock; 4680 boolean_t have_lock, full; 4681 l2arc_write_callback_t *cb; 4682 zio_t *pio, *wzio; 4683 uint64_t guid = spa_load_guid(spa); 4684 int try; 4685 4686 ASSERT(dev->l2ad_vdev != NULL); 4687 4688 pio = NULL; 4689 write_sz = 0; 4690 full = B_FALSE; 4691 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 4692 head->b_flags |= ARC_L2_WRITE_HEAD; 4693 4694 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 4695 /* 4696 * Copy buffers for L2ARC writing. 4697 */ 4698 mutex_enter(&l2arc_buflist_mtx); 4699 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 4700 list = l2arc_list_locked(try, &list_lock); 4701 passed_sz = 0; 4702 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 4703 4704 /* 4705 * L2ARC fast warmup. 4706 * 4707 * Until the ARC is warm and starts to evict, read from the 4708 * head of the ARC lists rather than the tail. 4709 */ 4710 headroom = target_sz * l2arc_headroom; 4711 if (arc_warm == B_FALSE) 4712 ab = list_head(list); 4713 else 4714 ab = list_tail(list); 4715 if (ab == NULL) 4716 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 4717 4718 for (; ab; ab = ab_prev) { 4719 if (arc_warm == B_FALSE) 4720 ab_prev = list_next(list, ab); 4721 else 4722 ab_prev = list_prev(list, ab); 4723 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size); 4724 4725 hash_lock = HDR_LOCK(ab); 4726 have_lock = MUTEX_HELD(hash_lock); 4727 if (!have_lock && !mutex_tryenter(hash_lock)) { 4728 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 4729 /* 4730 * Skip this buffer rather than waiting. 4731 */ 4732 continue; 4733 } 4734 4735 passed_sz += ab->b_size; 4736 if (passed_sz > headroom) { 4737 /* 4738 * Searched too far. 4739 */ 4740 mutex_exit(hash_lock); 4741 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 4742 break; 4743 } 4744 4745 if (!l2arc_write_eligible(guid, ab)) { 4746 mutex_exit(hash_lock); 4747 continue; 4748 } 4749 4750 if ((write_sz + ab->b_size) > target_sz) { 4751 full = B_TRUE; 4752 mutex_exit(hash_lock); 4753 ARCSTAT_BUMP(arcstat_l2_write_full); 4754 break; 4755 } 4756 4757 if (pio == NULL) { 4758 /* 4759 * Insert a dummy header on the buflist so 4760 * l2arc_write_done() can find where the 4761 * write buffers begin without searching. 4762 */ 4763 list_insert_head(dev->l2ad_buflist, head); 4764 4765 cb = kmem_alloc( 4766 sizeof (l2arc_write_callback_t), KM_SLEEP); 4767 cb->l2wcb_dev = dev; 4768 cb->l2wcb_head = head; 4769 pio = zio_root(spa, l2arc_write_done, cb, 4770 ZIO_FLAG_CANFAIL); 4771 ARCSTAT_BUMP(arcstat_l2_write_pios); 4772 } 4773 4774 /* 4775 * Create and add a new L2ARC header. 4776 */ 4777 hdrl2 = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 4778 hdrl2->b_dev = dev; 4779 hdrl2->b_daddr = dev->l2ad_hand; 4780 4781 ab->b_flags |= ARC_L2_WRITING; 4782 ab->b_l2hdr = hdrl2; 4783 list_insert_head(dev->l2ad_buflist, ab); 4784 buf_data = ab->b_buf->b_data; 4785 buf_sz = ab->b_size; 4786 4787 /* 4788 * Compute and store the buffer cksum before 4789 * writing. On debug the cksum is verified first. 4790 */ 4791 arc_cksum_verify(ab->b_buf); 4792 arc_cksum_compute(ab->b_buf, B_TRUE); 4793 4794 mutex_exit(hash_lock); 4795 4796 wzio = zio_write_phys(pio, dev->l2ad_vdev, 4797 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 4798 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 4799 ZIO_FLAG_CANFAIL, B_FALSE); 4800 4801 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 4802 zio_t *, wzio); 4803 (void) zio_nowait(wzio); 4804 4805 /* 4806 * Keep the clock hand suitably device-aligned. 4807 */ 4808 buf_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 4809 4810 write_sz += buf_sz; 4811 dev->l2ad_hand += buf_sz; 4812 } 4813 4814 mutex_exit(list_lock); 4815 4816 if (full == B_TRUE) 4817 break; 4818 } 4819 mutex_exit(&l2arc_buflist_mtx); 4820 4821 if (pio == NULL) { 4822 ASSERT0(write_sz); 4823 kmem_cache_free(hdr_cache, head); 4824 return (0); 4825 } 4826 4827 ASSERT3U(write_sz, <=, target_sz); 4828 ARCSTAT_BUMP(arcstat_l2_writes_sent); 4829 ARCSTAT_INCR(arcstat_l2_write_bytes, write_sz); 4830 ARCSTAT_INCR(arcstat_l2_size, write_sz); 4831 vdev_space_update(dev->l2ad_vdev, write_sz, 0, 0); 4832 4833 /* 4834 * Bump device hand to the device start if it is approaching the end. 4835 * l2arc_evict() will already have evicted ahead for this case. 4836 */ 4837 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 4838 vdev_space_update(dev->l2ad_vdev, 4839 dev->l2ad_end - dev->l2ad_hand, 0, 0); 4840 dev->l2ad_hand = dev->l2ad_start; 4841 dev->l2ad_evict = dev->l2ad_start; 4842 dev->l2ad_first = B_FALSE; 4843 } 4844 4845 dev->l2ad_writing = B_TRUE; 4846 (void) zio_wait(pio); 4847 dev->l2ad_writing = B_FALSE; 4848 4849 return (write_sz); 4850} 4851 4852/* 4853 * This thread feeds the L2ARC at regular intervals. This is the beating 4854 * heart of the L2ARC. 4855 */ 4856static void 4857l2arc_feed_thread(void *dummy __unused) 4858{ 4859 callb_cpr_t cpr; 4860 l2arc_dev_t *dev; 4861 spa_t *spa; 4862 uint64_t size, wrote; 4863 clock_t begin, next = ddi_get_lbolt(); 4864 4865 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 4866 4867 mutex_enter(&l2arc_feed_thr_lock); 4868 4869 while (l2arc_thread_exit == 0) { 4870 CALLB_CPR_SAFE_BEGIN(&cpr); 4871 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 4872 next - ddi_get_lbolt()); 4873 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 4874 next = ddi_get_lbolt() + hz; 4875 4876 /* 4877 * Quick check for L2ARC devices. 4878 */ 4879 mutex_enter(&l2arc_dev_mtx); 4880 if (l2arc_ndev == 0) { 4881 mutex_exit(&l2arc_dev_mtx); 4882 continue; 4883 } 4884 mutex_exit(&l2arc_dev_mtx); 4885 begin = ddi_get_lbolt(); 4886 4887 /* 4888 * This selects the next l2arc device to write to, and in 4889 * doing so the next spa to feed from: dev->l2ad_spa. This 4890 * will return NULL if there are now no l2arc devices or if 4891 * they are all faulted. 4892 * 4893 * If a device is returned, its spa's config lock is also 4894 * held to prevent device removal. l2arc_dev_get_next() 4895 * will grab and release l2arc_dev_mtx. 4896 */ 4897 if ((dev = l2arc_dev_get_next()) == NULL) 4898 continue; 4899 4900 spa = dev->l2ad_spa; 4901 ASSERT(spa != NULL); 4902 4903 /* 4904 * If the pool is read-only then force the feed thread to 4905 * sleep a little longer. 4906 */ 4907 if (!spa_writeable(spa)) { 4908 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 4909 spa_config_exit(spa, SCL_L2ARC, dev); 4910 continue; 4911 } 4912 4913 /* 4914 * Avoid contributing to memory pressure. 4915 */ 4916 if (arc_reclaim_needed()) { 4917 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 4918 spa_config_exit(spa, SCL_L2ARC, dev); 4919 continue; 4920 } 4921 4922 ARCSTAT_BUMP(arcstat_l2_feeds); 4923 4924 size = l2arc_write_size(dev); 4925 4926 /* 4927 * Evict L2ARC buffers that will be overwritten. 4928 */ 4929 l2arc_evict(dev, size, B_FALSE); 4930 4931 /* 4932 * Write ARC buffers. 4933 */ 4934 wrote = l2arc_write_buffers(spa, dev, size); 4935 4936 /* 4937 * Calculate interval between writes. 4938 */ 4939 next = l2arc_write_interval(begin, size, wrote); 4940 spa_config_exit(spa, SCL_L2ARC, dev); 4941 } 4942 4943 l2arc_thread_exit = 0; 4944 cv_broadcast(&l2arc_feed_thr_cv); 4945 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 4946 thread_exit(); 4947} 4948 4949boolean_t 4950l2arc_vdev_present(vdev_t *vd) 4951{ 4952 l2arc_dev_t *dev; 4953 4954 mutex_enter(&l2arc_dev_mtx); 4955 for (dev = list_head(l2arc_dev_list); dev != NULL; 4956 dev = list_next(l2arc_dev_list, dev)) { 4957 if (dev->l2ad_vdev == vd) 4958 break; 4959 } 4960 mutex_exit(&l2arc_dev_mtx); 4961 4962 return (dev != NULL); 4963} 4964 4965/* 4966 * Add a vdev for use by the L2ARC. By this point the spa has already 4967 * validated the vdev and opened it. 4968 */ 4969void 4970l2arc_add_vdev(spa_t *spa, vdev_t *vd) 4971{ 4972 l2arc_dev_t *adddev; 4973 4974 ASSERT(!l2arc_vdev_present(vd)); 4975 4976 /* 4977 * Create a new l2arc device entry. 4978 */ 4979 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 4980 adddev->l2ad_spa = spa; 4981 adddev->l2ad_vdev = vd; 4982 adddev->l2ad_write = l2arc_write_max; 4983 adddev->l2ad_boost = l2arc_write_boost; 4984 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 4985 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 4986 adddev->l2ad_hand = adddev->l2ad_start; 4987 adddev->l2ad_evict = adddev->l2ad_start; 4988 adddev->l2ad_first = B_TRUE; 4989 adddev->l2ad_writing = B_FALSE; 4990 ASSERT3U(adddev->l2ad_write, >, 0); 4991 4992 /* 4993 * This is a list of all ARC buffers that are still valid on the 4994 * device. 4995 */ 4996 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 4997 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 4998 offsetof(arc_buf_hdr_t, b_l2node)); 4999 5000 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5001 5002 /* 5003 * Add device to global list 5004 */ 5005 mutex_enter(&l2arc_dev_mtx); 5006 list_insert_head(l2arc_dev_list, adddev); 5007 atomic_inc_64(&l2arc_ndev); 5008 mutex_exit(&l2arc_dev_mtx); 5009} 5010 5011/* 5012 * Remove a vdev from the L2ARC. 5013 */ 5014void 5015l2arc_remove_vdev(vdev_t *vd) 5016{ 5017 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5018 5019 /* 5020 * Find the device by vdev 5021 */ 5022 mutex_enter(&l2arc_dev_mtx); 5023 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5024 nextdev = list_next(l2arc_dev_list, dev); 5025 if (vd == dev->l2ad_vdev) { 5026 remdev = dev; 5027 break; 5028 } 5029 } 5030 ASSERT(remdev != NULL); 5031 5032 /* 5033 * Remove device from global list 5034 */ 5035 list_remove(l2arc_dev_list, remdev); 5036 l2arc_dev_last = NULL; /* may have been invalidated */ 5037 atomic_dec_64(&l2arc_ndev); 5038 mutex_exit(&l2arc_dev_mtx); 5039 5040 /* 5041 * Clear all buflists and ARC references. L2ARC device flush. 5042 */ 5043 l2arc_evict(remdev, 0, B_TRUE); 5044 list_destroy(remdev->l2ad_buflist); 5045 kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5046 kmem_free(remdev, sizeof (l2arc_dev_t)); 5047} 5048 5049void 5050l2arc_init(void) 5051{ 5052 l2arc_thread_exit = 0; 5053 l2arc_ndev = 0; 5054 l2arc_writes_sent = 0; 5055 l2arc_writes_done = 0; 5056 5057 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5058 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5059 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5060 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5061 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5062 5063 l2arc_dev_list = &L2ARC_dev_list; 5064 l2arc_free_on_write = &L2ARC_free_on_write; 5065 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5066 offsetof(l2arc_dev_t, l2ad_node)); 5067 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5068 offsetof(l2arc_data_free_t, l2df_list_node)); 5069} 5070 5071void 5072l2arc_fini(void) 5073{ 5074 /* 5075 * This is called from dmu_fini(), which is called from spa_fini(); 5076 * Because of this, we can assume that all l2arc devices have 5077 * already been removed when the pools themselves were removed. 5078 */ 5079 5080 l2arc_do_free_on_write(); 5081 5082 mutex_destroy(&l2arc_feed_thr_lock); 5083 cv_destroy(&l2arc_feed_thr_cv); 5084 mutex_destroy(&l2arc_dev_mtx); 5085 mutex_destroy(&l2arc_buflist_mtx); 5086 mutex_destroy(&l2arc_free_on_write_mtx); 5087 5088 list_destroy(l2arc_dev_list); 5089 list_destroy(l2arc_free_on_write); 5090} 5091 5092void 5093l2arc_start(void) 5094{ 5095 if (!(spa_mode_global & FWRITE)) 5096 return; 5097 5098 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5099 TS_RUN, minclsyspri); 5100} 5101 5102void 5103l2arc_stop(void) 5104{ 5105 if (!(spa_mode_global & FWRITE)) 5106 return; 5107 5108 mutex_enter(&l2arc_feed_thr_lock); 5109 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5110 l2arc_thread_exit = 1; 5111 while (l2arc_thread_exit != 0) 5112 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5113 mutex_exit(&l2arc_feed_thr_lock); 5114} 5115