arc.c revision 269230
1168404Spjd/* 2168404Spjd * CDDL HEADER START 3168404Spjd * 4168404Spjd * The contents of this file are subject to the terms of the 5168404Spjd * Common Development and Distribution License (the "License"). 6168404Spjd * You may not use this file except in compliance with the License. 7168404Spjd * 8168404Spjd * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9168404Spjd * or http://www.opensolaris.org/os/licensing. 10168404Spjd * See the License for the specific language governing permissions 11168404Spjd * and limitations under the License. 12168404Spjd * 13168404Spjd * When distributing Covered Code, include this CDDL HEADER in each 14168404Spjd * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15168404Spjd * If applicable, add the following below this CDDL HEADER, with the 16168404Spjd * fields enclosed by brackets "[]" replaced with your own identifying 17168404Spjd * information: Portions Copyright [yyyy] [name of copyright owner] 18168404Spjd * 19168404Spjd * CDDL HEADER END 20168404Spjd */ 21168404Spjd/* 22219089Spjd * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23268123Sdelphij * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 24260835Sdelphij * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 25268085Sdelphij * Copyright 2014 Nexenta Systems, Inc. All rights reserved. 26168404Spjd */ 27168404Spjd 28168404Spjd/* 29168404Spjd * DVA-based Adjustable Replacement Cache 30168404Spjd * 31168404Spjd * While much of the theory of operation used here is 32168404Spjd * based on the self-tuning, low overhead replacement cache 33168404Spjd * presented by Megiddo and Modha at FAST 2003, there are some 34168404Spjd * significant differences: 35168404Spjd * 36168404Spjd * 1. The Megiddo and Modha model assumes any page is evictable. 37168404Spjd * Pages in its cache cannot be "locked" into memory. This makes 38168404Spjd * the eviction algorithm simple: evict the last page in the list. 39168404Spjd * This also make the performance characteristics easy to reason 40168404Spjd * about. Our cache is not so simple. At any given moment, some 41168404Spjd * subset of the blocks in the cache are un-evictable because we 42168404Spjd * have handed out a reference to them. Blocks are only evictable 43168404Spjd * when there are no external references active. This makes 44168404Spjd * eviction far more problematic: we choose to evict the evictable 45168404Spjd * blocks that are the "lowest" in the list. 46168404Spjd * 47168404Spjd * There are times when it is not possible to evict the requested 48168404Spjd * space. In these circumstances we are unable to adjust the cache 49168404Spjd * size. To prevent the cache growing unbounded at these times we 50185029Spjd * implement a "cache throttle" that slows the flow of new data 51185029Spjd * into the cache until we can make space available. 52168404Spjd * 53168404Spjd * 2. The Megiddo and Modha model assumes a fixed cache size. 54168404Spjd * Pages are evicted when the cache is full and there is a cache 55168404Spjd * miss. Our model has a variable sized cache. It grows with 56185029Spjd * high use, but also tries to react to memory pressure from the 57168404Spjd * operating system: decreasing its size when system memory is 58168404Spjd * tight. 59168404Spjd * 60168404Spjd * 3. The Megiddo and Modha model assumes a fixed page size. All 61251631Sdelphij * elements of the cache are therefore exactly the same size. So 62168404Spjd * when adjusting the cache size following a cache miss, its simply 63168404Spjd * a matter of choosing a single page to evict. In our model, we 64168404Spjd * have variable sized cache blocks (rangeing from 512 bytes to 65251631Sdelphij * 128K bytes). We therefore choose a set of blocks to evict to make 66168404Spjd * space for a cache miss that approximates as closely as possible 67168404Spjd * the space used by the new block. 68168404Spjd * 69168404Spjd * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 70168404Spjd * by N. Megiddo & D. Modha, FAST 2003 71168404Spjd */ 72168404Spjd 73168404Spjd/* 74168404Spjd * The locking model: 75168404Spjd * 76168404Spjd * A new reference to a cache buffer can be obtained in two 77168404Spjd * ways: 1) via a hash table lookup using the DVA as a key, 78185029Spjd * or 2) via one of the ARC lists. The arc_read() interface 79168404Spjd * uses method 1, while the internal arc algorithms for 80251631Sdelphij * adjusting the cache use method 2. We therefore provide two 81168404Spjd * types of locks: 1) the hash table lock array, and 2) the 82168404Spjd * arc list locks. 83168404Spjd * 84168404Spjd * Buffers do not have their own mutexs, rather they rely on the 85168404Spjd * hash table mutexs for the bulk of their protection (i.e. most 86168404Spjd * fields in the arc_buf_hdr_t are protected by these mutexs). 87168404Spjd * 88168404Spjd * buf_hash_find() returns the appropriate mutex (held) when it 89168404Spjd * locates the requested buffer in the hash table. It returns 90168404Spjd * NULL for the mutex if the buffer was not in the table. 91168404Spjd * 92168404Spjd * buf_hash_remove() expects the appropriate hash mutex to be 93168404Spjd * already held before it is invoked. 94168404Spjd * 95168404Spjd * Each arc state also has a mutex which is used to protect the 96168404Spjd * buffer list associated with the state. When attempting to 97168404Spjd * obtain a hash table lock while holding an arc list lock you 98168404Spjd * must use: mutex_tryenter() to avoid deadlock. Also note that 99168404Spjd * the active state mutex must be held before the ghost state mutex. 100168404Spjd * 101168404Spjd * Arc buffers may have an associated eviction callback function. 102168404Spjd * This function will be invoked prior to removing the buffer (e.g. 103168404Spjd * in arc_do_user_evicts()). Note however that the data associated 104168404Spjd * with the buffer may be evicted prior to the callback. The callback 105168404Spjd * must be made with *no locks held* (to prevent deadlock). Additionally, 106168404Spjd * the users of callbacks must ensure that their private data is 107268858Sdelphij * protected from simultaneous callbacks from arc_clear_callback() 108168404Spjd * and arc_do_user_evicts(). 109168404Spjd * 110168404Spjd * Note that the majority of the performance stats are manipulated 111168404Spjd * with atomic operations. 112185029Spjd * 113185029Spjd * The L2ARC uses the l2arc_buflist_mtx global mutex for the following: 114185029Spjd * 115185029Spjd * - L2ARC buflist creation 116185029Spjd * - L2ARC buflist eviction 117185029Spjd * - L2ARC write completion, which walks L2ARC buflists 118185029Spjd * - ARC header destruction, as it removes from L2ARC buflists 119185029Spjd * - ARC header release, as it removes from L2ARC buflists 120168404Spjd */ 121168404Spjd 122168404Spjd#include <sys/spa.h> 123168404Spjd#include <sys/zio.h> 124251478Sdelphij#include <sys/zio_compress.h> 125168404Spjd#include <sys/zfs_context.h> 126168404Spjd#include <sys/arc.h> 127168404Spjd#include <sys/refcount.h> 128185029Spjd#include <sys/vdev.h> 129219089Spjd#include <sys/vdev_impl.h> 130258632Savg#include <sys/dsl_pool.h> 131168404Spjd#ifdef _KERNEL 132168404Spjd#include <sys/dnlc.h> 133168404Spjd#endif 134168404Spjd#include <sys/callb.h> 135168404Spjd#include <sys/kstat.h> 136248572Ssmh#include <sys/trim_map.h> 137219089Spjd#include <zfs_fletcher.h> 138168404Spjd#include <sys/sdt.h> 139168404Spjd 140191902Skmacy#include <vm/vm_pageout.h> 141191902Skmacy 142240133Smm#ifdef illumos 143240133Smm#ifndef _KERNEL 144240133Smm/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 145240133Smmboolean_t arc_watch = B_FALSE; 146240133Smmint arc_procfd; 147240133Smm#endif 148240133Smm#endif /* illumos */ 149240133Smm 150168404Spjdstatic kmutex_t arc_reclaim_thr_lock; 151168404Spjdstatic kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 152168404Spjdstatic uint8_t arc_thread_exit; 153168404Spjd 154168404Spjd#define ARC_REDUCE_DNLC_PERCENT 3 155168404Spjduint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 156168404Spjd 157168404Spjdtypedef enum arc_reclaim_strategy { 158168404Spjd ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 159168404Spjd ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 160168404Spjd} arc_reclaim_strategy_t; 161168404Spjd 162258632Savg/* 163258632Savg * The number of iterations through arc_evict_*() before we 164258632Savg * drop & reacquire the lock. 165258632Savg */ 166258632Savgint arc_evict_iterations = 100; 167258632Savg 168168404Spjd/* number of seconds before growing cache again */ 169168404Spjdstatic int arc_grow_retry = 60; 170168404Spjd 171208373Smm/* shift of arc_c for calculating both min and max arc_p */ 172208373Smmstatic int arc_p_min_shift = 4; 173208373Smm 174208373Smm/* log2(fraction of arc to reclaim) */ 175208373Smmstatic int arc_shrink_shift = 5; 176208373Smm 177168404Spjd/* 178168404Spjd * minimum lifespan of a prefetch block in clock ticks 179168404Spjd * (initialized in arc_init()) 180168404Spjd */ 181168404Spjdstatic int arc_min_prefetch_lifespan; 182168404Spjd 183258632Savg/* 184258632Savg * If this percent of memory is free, don't throttle. 185258632Savg */ 186258632Savgint arc_lotsfree_percent = 10; 187258632Savg 188208373Smmstatic int arc_dead; 189194043Skmacyextern int zfs_prefetch_disable; 190168404Spjd 191168404Spjd/* 192185029Spjd * The arc has filled available memory and has now warmed up. 193185029Spjd */ 194185029Spjdstatic boolean_t arc_warm; 195185029Spjd 196185029Spjd/* 197168404Spjd * These tunables are for performance analysis. 198168404Spjd */ 199185029Spjduint64_t zfs_arc_max; 200185029Spjduint64_t zfs_arc_min; 201185029Spjduint64_t zfs_arc_meta_limit = 0; 202208373Smmint zfs_arc_grow_retry = 0; 203208373Smmint zfs_arc_shrink_shift = 0; 204208373Smmint zfs_arc_p_min_shift = 0; 205242845Sdelphijint zfs_disable_dup_eviction = 0; 206269230Sdelphijuint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 207185029Spjd 208185029SpjdTUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 209168473SpjdSYSCTL_DECL(_vfs_zfs); 210217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 211168473Spjd "Maximum ARC size"); 212217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 213168473Spjd "Minimum ARC size"); 214269230SdelphijSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 215269230Sdelphij &zfs_arc_average_blocksize, 0, 216269230Sdelphij "ARC average blocksize"); 217168404Spjd 218168404Spjd/* 219185029Spjd * Note that buffers can be in one of 6 states: 220168404Spjd * ARC_anon - anonymous (discussed below) 221168404Spjd * ARC_mru - recently used, currently cached 222168404Spjd * ARC_mru_ghost - recentely used, no longer in cache 223168404Spjd * ARC_mfu - frequently used, currently cached 224168404Spjd * ARC_mfu_ghost - frequently used, no longer in cache 225185029Spjd * ARC_l2c_only - exists in L2ARC but not other states 226185029Spjd * When there are no active references to the buffer, they are 227185029Spjd * are linked onto a list in one of these arc states. These are 228185029Spjd * the only buffers that can be evicted or deleted. Within each 229185029Spjd * state there are multiple lists, one for meta-data and one for 230185029Spjd * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 231185029Spjd * etc.) is tracked separately so that it can be managed more 232185029Spjd * explicitly: favored over data, limited explicitly. 233168404Spjd * 234168404Spjd * Anonymous buffers are buffers that are not associated with 235168404Spjd * a DVA. These are buffers that hold dirty block copies 236168404Spjd * before they are written to stable storage. By definition, 237168404Spjd * they are "ref'd" and are considered part of arc_mru 238168404Spjd * that cannot be freed. Generally, they will aquire a DVA 239168404Spjd * as they are written and migrate onto the arc_mru list. 240185029Spjd * 241185029Spjd * The ARC_l2c_only state is for buffers that are in the second 242185029Spjd * level ARC but no longer in any of the ARC_m* lists. The second 243185029Spjd * level ARC itself may also contain buffers that are in any of 244185029Spjd * the ARC_m* states - meaning that a buffer can exist in two 245185029Spjd * places. The reason for the ARC_l2c_only state is to keep the 246185029Spjd * buffer header in the hash table, so that reads that hit the 247185029Spjd * second level ARC benefit from these fast lookups. 248168404Spjd */ 249168404Spjd 250205264Skmacy#define ARCS_LOCK_PAD CACHE_LINE_SIZE 251205231Skmacystruct arcs_lock { 252205231Skmacy kmutex_t arcs_lock; 253205231Skmacy#ifdef _KERNEL 254205231Skmacy unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 255205231Skmacy#endif 256205231Skmacy}; 257205231Skmacy 258205231Skmacy/* 259205231Skmacy * must be power of two for mask use to work 260205231Skmacy * 261205231Skmacy */ 262205231Skmacy#define ARC_BUFC_NUMDATALISTS 16 263205231Skmacy#define ARC_BUFC_NUMMETADATALISTS 16 264206796Spjd#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 265205231Skmacy 266168404Spjdtypedef struct arc_state { 267185029Spjd uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 268185029Spjd uint64_t arcs_size; /* total amount of data in this state */ 269205231Skmacy list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 270205264Skmacy struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 271168404Spjd} arc_state_t; 272168404Spjd 273206796Spjd#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 274205231Skmacy 275185029Spjd/* The 6 states: */ 276168404Spjdstatic arc_state_t ARC_anon; 277168404Spjdstatic arc_state_t ARC_mru; 278168404Spjdstatic arc_state_t ARC_mru_ghost; 279168404Spjdstatic arc_state_t ARC_mfu; 280168404Spjdstatic arc_state_t ARC_mfu_ghost; 281185029Spjdstatic arc_state_t ARC_l2c_only; 282168404Spjd 283168404Spjdtypedef struct arc_stats { 284168404Spjd kstat_named_t arcstat_hits; 285168404Spjd kstat_named_t arcstat_misses; 286168404Spjd kstat_named_t arcstat_demand_data_hits; 287168404Spjd kstat_named_t arcstat_demand_data_misses; 288168404Spjd kstat_named_t arcstat_demand_metadata_hits; 289168404Spjd kstat_named_t arcstat_demand_metadata_misses; 290168404Spjd kstat_named_t arcstat_prefetch_data_hits; 291168404Spjd kstat_named_t arcstat_prefetch_data_misses; 292168404Spjd kstat_named_t arcstat_prefetch_metadata_hits; 293168404Spjd kstat_named_t arcstat_prefetch_metadata_misses; 294168404Spjd kstat_named_t arcstat_mru_hits; 295168404Spjd kstat_named_t arcstat_mru_ghost_hits; 296168404Spjd kstat_named_t arcstat_mfu_hits; 297168404Spjd kstat_named_t arcstat_mfu_ghost_hits; 298205231Skmacy kstat_named_t arcstat_allocated; 299168404Spjd kstat_named_t arcstat_deleted; 300205231Skmacy kstat_named_t arcstat_stolen; 301168404Spjd kstat_named_t arcstat_recycle_miss; 302251629Sdelphij /* 303251629Sdelphij * Number of buffers that could not be evicted because the hash lock 304251629Sdelphij * was held by another thread. The lock may not necessarily be held 305251629Sdelphij * by something using the same buffer, since hash locks are shared 306251629Sdelphij * by multiple buffers. 307251629Sdelphij */ 308168404Spjd kstat_named_t arcstat_mutex_miss; 309251629Sdelphij /* 310251629Sdelphij * Number of buffers skipped because they have I/O in progress, are 311251629Sdelphij * indrect prefetch buffers that have not lived long enough, or are 312251629Sdelphij * not from the spa we're trying to evict from. 313251629Sdelphij */ 314168404Spjd kstat_named_t arcstat_evict_skip; 315208373Smm kstat_named_t arcstat_evict_l2_cached; 316208373Smm kstat_named_t arcstat_evict_l2_eligible; 317208373Smm kstat_named_t arcstat_evict_l2_ineligible; 318168404Spjd kstat_named_t arcstat_hash_elements; 319168404Spjd kstat_named_t arcstat_hash_elements_max; 320168404Spjd kstat_named_t arcstat_hash_collisions; 321168404Spjd kstat_named_t arcstat_hash_chains; 322168404Spjd kstat_named_t arcstat_hash_chain_max; 323168404Spjd kstat_named_t arcstat_p; 324168404Spjd kstat_named_t arcstat_c; 325168404Spjd kstat_named_t arcstat_c_min; 326168404Spjd kstat_named_t arcstat_c_max; 327168404Spjd kstat_named_t arcstat_size; 328185029Spjd kstat_named_t arcstat_hdr_size; 329208373Smm kstat_named_t arcstat_data_size; 330208373Smm kstat_named_t arcstat_other_size; 331185029Spjd kstat_named_t arcstat_l2_hits; 332185029Spjd kstat_named_t arcstat_l2_misses; 333185029Spjd kstat_named_t arcstat_l2_feeds; 334185029Spjd kstat_named_t arcstat_l2_rw_clash; 335208373Smm kstat_named_t arcstat_l2_read_bytes; 336208373Smm kstat_named_t arcstat_l2_write_bytes; 337185029Spjd kstat_named_t arcstat_l2_writes_sent; 338185029Spjd kstat_named_t arcstat_l2_writes_done; 339185029Spjd kstat_named_t arcstat_l2_writes_error; 340185029Spjd kstat_named_t arcstat_l2_writes_hdr_miss; 341185029Spjd kstat_named_t arcstat_l2_evict_lock_retry; 342185029Spjd kstat_named_t arcstat_l2_evict_reading; 343185029Spjd kstat_named_t arcstat_l2_free_on_write; 344185029Spjd kstat_named_t arcstat_l2_abort_lowmem; 345185029Spjd kstat_named_t arcstat_l2_cksum_bad; 346185029Spjd kstat_named_t arcstat_l2_io_error; 347185029Spjd kstat_named_t arcstat_l2_size; 348251478Sdelphij kstat_named_t arcstat_l2_asize; 349185029Spjd kstat_named_t arcstat_l2_hdr_size; 350251478Sdelphij kstat_named_t arcstat_l2_compress_successes; 351251478Sdelphij kstat_named_t arcstat_l2_compress_zeros; 352251478Sdelphij kstat_named_t arcstat_l2_compress_failures; 353205231Skmacy kstat_named_t arcstat_l2_write_trylock_fail; 354205231Skmacy kstat_named_t arcstat_l2_write_passed_headroom; 355205231Skmacy kstat_named_t arcstat_l2_write_spa_mismatch; 356206796Spjd kstat_named_t arcstat_l2_write_in_l2; 357205231Skmacy kstat_named_t arcstat_l2_write_hdr_io_in_progress; 358205231Skmacy kstat_named_t arcstat_l2_write_not_cacheable; 359205231Skmacy kstat_named_t arcstat_l2_write_full; 360205231Skmacy kstat_named_t arcstat_l2_write_buffer_iter; 361205231Skmacy kstat_named_t arcstat_l2_write_pios; 362205231Skmacy kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 363205231Skmacy kstat_named_t arcstat_l2_write_buffer_list_iter; 364205231Skmacy kstat_named_t arcstat_l2_write_buffer_list_null_iter; 365242845Sdelphij kstat_named_t arcstat_memory_throttle_count; 366242845Sdelphij kstat_named_t arcstat_duplicate_buffers; 367242845Sdelphij kstat_named_t arcstat_duplicate_buffers_size; 368242845Sdelphij kstat_named_t arcstat_duplicate_reads; 369168404Spjd} arc_stats_t; 370168404Spjd 371168404Spjdstatic arc_stats_t arc_stats = { 372168404Spjd { "hits", KSTAT_DATA_UINT64 }, 373168404Spjd { "misses", KSTAT_DATA_UINT64 }, 374168404Spjd { "demand_data_hits", KSTAT_DATA_UINT64 }, 375168404Spjd { "demand_data_misses", KSTAT_DATA_UINT64 }, 376168404Spjd { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 377168404Spjd { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 378168404Spjd { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 379168404Spjd { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 380168404Spjd { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 381168404Spjd { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 382168404Spjd { "mru_hits", KSTAT_DATA_UINT64 }, 383168404Spjd { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 384168404Spjd { "mfu_hits", KSTAT_DATA_UINT64 }, 385168404Spjd { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 386205231Skmacy { "allocated", KSTAT_DATA_UINT64 }, 387168404Spjd { "deleted", KSTAT_DATA_UINT64 }, 388205231Skmacy { "stolen", KSTAT_DATA_UINT64 }, 389168404Spjd { "recycle_miss", KSTAT_DATA_UINT64 }, 390168404Spjd { "mutex_miss", KSTAT_DATA_UINT64 }, 391168404Spjd { "evict_skip", KSTAT_DATA_UINT64 }, 392208373Smm { "evict_l2_cached", KSTAT_DATA_UINT64 }, 393208373Smm { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 394208373Smm { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 395168404Spjd { "hash_elements", KSTAT_DATA_UINT64 }, 396168404Spjd { "hash_elements_max", KSTAT_DATA_UINT64 }, 397168404Spjd { "hash_collisions", KSTAT_DATA_UINT64 }, 398168404Spjd { "hash_chains", KSTAT_DATA_UINT64 }, 399168404Spjd { "hash_chain_max", KSTAT_DATA_UINT64 }, 400168404Spjd { "p", KSTAT_DATA_UINT64 }, 401168404Spjd { "c", KSTAT_DATA_UINT64 }, 402168404Spjd { "c_min", KSTAT_DATA_UINT64 }, 403168404Spjd { "c_max", KSTAT_DATA_UINT64 }, 404185029Spjd { "size", KSTAT_DATA_UINT64 }, 405185029Spjd { "hdr_size", KSTAT_DATA_UINT64 }, 406208373Smm { "data_size", KSTAT_DATA_UINT64 }, 407208373Smm { "other_size", KSTAT_DATA_UINT64 }, 408185029Spjd { "l2_hits", KSTAT_DATA_UINT64 }, 409185029Spjd { "l2_misses", KSTAT_DATA_UINT64 }, 410185029Spjd { "l2_feeds", KSTAT_DATA_UINT64 }, 411185029Spjd { "l2_rw_clash", KSTAT_DATA_UINT64 }, 412208373Smm { "l2_read_bytes", KSTAT_DATA_UINT64 }, 413208373Smm { "l2_write_bytes", KSTAT_DATA_UINT64 }, 414185029Spjd { "l2_writes_sent", KSTAT_DATA_UINT64 }, 415185029Spjd { "l2_writes_done", KSTAT_DATA_UINT64 }, 416185029Spjd { "l2_writes_error", KSTAT_DATA_UINT64 }, 417185029Spjd { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 418185029Spjd { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 419185029Spjd { "l2_evict_reading", KSTAT_DATA_UINT64 }, 420185029Spjd { "l2_free_on_write", KSTAT_DATA_UINT64 }, 421185029Spjd { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 422185029Spjd { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 423185029Spjd { "l2_io_error", KSTAT_DATA_UINT64 }, 424185029Spjd { "l2_size", KSTAT_DATA_UINT64 }, 425251478Sdelphij { "l2_asize", KSTAT_DATA_UINT64 }, 426185029Spjd { "l2_hdr_size", KSTAT_DATA_UINT64 }, 427251478Sdelphij { "l2_compress_successes", KSTAT_DATA_UINT64 }, 428251478Sdelphij { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 429251478Sdelphij { "l2_compress_failures", KSTAT_DATA_UINT64 }, 430206796Spjd { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 431206796Spjd { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 432206796Spjd { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 433206796Spjd { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 434206796Spjd { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 435206796Spjd { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 436206796Spjd { "l2_write_full", KSTAT_DATA_UINT64 }, 437206796Spjd { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 438206796Spjd { "l2_write_pios", KSTAT_DATA_UINT64 }, 439206796Spjd { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 440206796Spjd { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 441242845Sdelphij { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 442242845Sdelphij { "memory_throttle_count", KSTAT_DATA_UINT64 }, 443242845Sdelphij { "duplicate_buffers", KSTAT_DATA_UINT64 }, 444242845Sdelphij { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 445242845Sdelphij { "duplicate_reads", KSTAT_DATA_UINT64 } 446168404Spjd}; 447168404Spjd 448168404Spjd#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 449168404Spjd 450168404Spjd#define ARCSTAT_INCR(stat, val) \ 451251631Sdelphij atomic_add_64(&arc_stats.stat.value.ui64, (val)) 452168404Spjd 453206796Spjd#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 454168404Spjd#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 455168404Spjd 456168404Spjd#define ARCSTAT_MAX(stat, val) { \ 457168404Spjd uint64_t m; \ 458168404Spjd while ((val) > (m = arc_stats.stat.value.ui64) && \ 459168404Spjd (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 460168404Spjd continue; \ 461168404Spjd} 462168404Spjd 463168404Spjd#define ARCSTAT_MAXSTAT(stat) \ 464168404Spjd ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 465168404Spjd 466168404Spjd/* 467168404Spjd * We define a macro to allow ARC hits/misses to be easily broken down by 468168404Spjd * two separate conditions, giving a total of four different subtypes for 469168404Spjd * each of hits and misses (so eight statistics total). 470168404Spjd */ 471168404Spjd#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 472168404Spjd if (cond1) { \ 473168404Spjd if (cond2) { \ 474168404Spjd ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 475168404Spjd } else { \ 476168404Spjd ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 477168404Spjd } \ 478168404Spjd } else { \ 479168404Spjd if (cond2) { \ 480168404Spjd ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 481168404Spjd } else { \ 482168404Spjd ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 483168404Spjd } \ 484168404Spjd } 485168404Spjd 486168404Spjdkstat_t *arc_ksp; 487206796Spjdstatic arc_state_t *arc_anon; 488168404Spjdstatic arc_state_t *arc_mru; 489168404Spjdstatic arc_state_t *arc_mru_ghost; 490168404Spjdstatic arc_state_t *arc_mfu; 491168404Spjdstatic arc_state_t *arc_mfu_ghost; 492185029Spjdstatic arc_state_t *arc_l2c_only; 493168404Spjd 494168404Spjd/* 495168404Spjd * There are several ARC variables that are critical to export as kstats -- 496168404Spjd * but we don't want to have to grovel around in the kstat whenever we wish to 497168404Spjd * manipulate them. For these variables, we therefore define them to be in 498168404Spjd * terms of the statistic variable. This assures that we are not introducing 499168404Spjd * the possibility of inconsistency by having shadow copies of the variables, 500168404Spjd * while still allowing the code to be readable. 501168404Spjd */ 502168404Spjd#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 503168404Spjd#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 504168404Spjd#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 505168404Spjd#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 506168404Spjd#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 507168404Spjd 508251478Sdelphij#define L2ARC_IS_VALID_COMPRESS(_c_) \ 509251478Sdelphij ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 510251478Sdelphij 511168404Spjdstatic int arc_no_grow; /* Don't try to grow cache size */ 512168404Spjdstatic uint64_t arc_tempreserve; 513209962Smmstatic uint64_t arc_loaned_bytes; 514185029Spjdstatic uint64_t arc_meta_used; 515185029Spjdstatic uint64_t arc_meta_limit; 516185029Spjdstatic uint64_t arc_meta_max = 0; 517229663SpjdSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RD, &arc_meta_used, 0, 518229663Spjd "ARC metadata used"); 519229663SpjdSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RW, &arc_meta_limit, 0, 520229663Spjd "ARC metadata limit"); 521168404Spjd 522185029Spjdtypedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 523185029Spjd 524168404Spjdtypedef struct arc_callback arc_callback_t; 525168404Spjd 526168404Spjdstruct arc_callback { 527168404Spjd void *acb_private; 528168404Spjd arc_done_func_t *acb_done; 529168404Spjd arc_buf_t *acb_buf; 530168404Spjd zio_t *acb_zio_dummy; 531168404Spjd arc_callback_t *acb_next; 532168404Spjd}; 533168404Spjd 534168404Spjdtypedef struct arc_write_callback arc_write_callback_t; 535168404Spjd 536168404Spjdstruct arc_write_callback { 537168404Spjd void *awcb_private; 538168404Spjd arc_done_func_t *awcb_ready; 539258632Savg arc_done_func_t *awcb_physdone; 540168404Spjd arc_done_func_t *awcb_done; 541168404Spjd arc_buf_t *awcb_buf; 542168404Spjd}; 543168404Spjd 544168404Spjdstruct arc_buf_hdr { 545168404Spjd /* protected by hash lock */ 546168404Spjd dva_t b_dva; 547168404Spjd uint64_t b_birth; 548168404Spjd uint64_t b_cksum0; 549168404Spjd 550168404Spjd kmutex_t b_freeze_lock; 551168404Spjd zio_cksum_t *b_freeze_cksum; 552219089Spjd void *b_thawed; 553168404Spjd 554168404Spjd arc_buf_hdr_t *b_hash_next; 555168404Spjd arc_buf_t *b_buf; 556168404Spjd uint32_t b_flags; 557168404Spjd uint32_t b_datacnt; 558168404Spjd 559168404Spjd arc_callback_t *b_acb; 560168404Spjd kcondvar_t b_cv; 561168404Spjd 562168404Spjd /* immutable */ 563168404Spjd arc_buf_contents_t b_type; 564168404Spjd uint64_t b_size; 565209962Smm uint64_t b_spa; 566168404Spjd 567168404Spjd /* protected by arc state mutex */ 568168404Spjd arc_state_t *b_state; 569168404Spjd list_node_t b_arc_node; 570168404Spjd 571168404Spjd /* updated atomically */ 572168404Spjd clock_t b_arc_access; 573168404Spjd 574168404Spjd /* self protecting */ 575168404Spjd refcount_t b_refcnt; 576185029Spjd 577185029Spjd l2arc_buf_hdr_t *b_l2hdr; 578185029Spjd list_node_t b_l2node; 579168404Spjd}; 580168404Spjd 581168404Spjdstatic arc_buf_t *arc_eviction_list; 582168404Spjdstatic kmutex_t arc_eviction_mtx; 583168404Spjdstatic arc_buf_hdr_t arc_eviction_hdr; 584168404Spjdstatic void arc_get_data_buf(arc_buf_t *buf); 585168404Spjdstatic void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock); 586185029Spjdstatic int arc_evict_needed(arc_buf_contents_t type); 587209962Smmstatic void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes); 588240133Smm#ifdef illumos 589240133Smmstatic void arc_buf_watch(arc_buf_t *buf); 590240133Smm#endif /* illumos */ 591168404Spjd 592209962Smmstatic boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab); 593208373Smm 594168404Spjd#define GHOST_STATE(state) \ 595185029Spjd ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 596185029Spjd (state) == arc_l2c_only) 597168404Spjd 598168404Spjd/* 599168404Spjd * Private ARC flags. These flags are private ARC only flags that will show up 600168404Spjd * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can 601168404Spjd * be passed in as arc_flags in things like arc_read. However, these flags 602168404Spjd * should never be passed and should only be set by ARC code. When adding new 603168404Spjd * public flags, make sure not to smash the private ones. 604168404Spjd */ 605168404Spjd 606168404Spjd#define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */ 607168404Spjd#define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */ 608168404Spjd#define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */ 609168404Spjd#define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */ 610168404Spjd#define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */ 611168404Spjd#define ARC_INDIRECT (1 << 14) /* this is an indirect block */ 612185029Spjd#define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */ 613185029Spjd#define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */ 614185029Spjd#define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */ 615185029Spjd#define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */ 616168404Spjd 617168404Spjd#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE) 618168404Spjd#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS) 619168404Spjd#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR) 620208373Smm#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH) 621168404Spjd#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ) 622168404Spjd#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE) 623185029Spjd#define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS) 624185029Spjd#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE) 625185029Spjd#define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \ 626185029Spjd (hdr)->b_l2hdr != NULL) 627185029Spjd#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING) 628185029Spjd#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED) 629185029Spjd#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD) 630168404Spjd 631168404Spjd/* 632185029Spjd * Other sizes 633185029Spjd */ 634185029Spjd 635185029Spjd#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 636185029Spjd#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 637185029Spjd 638185029Spjd/* 639168404Spjd * Hash table routines 640168404Spjd */ 641168404Spjd 642205253Skmacy#define HT_LOCK_PAD CACHE_LINE_SIZE 643168404Spjd 644168404Spjdstruct ht_lock { 645168404Spjd kmutex_t ht_lock; 646168404Spjd#ifdef _KERNEL 647168404Spjd unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 648168404Spjd#endif 649168404Spjd}; 650168404Spjd 651168404Spjd#define BUF_LOCKS 256 652168404Spjdtypedef struct buf_hash_table { 653168404Spjd uint64_t ht_mask; 654168404Spjd arc_buf_hdr_t **ht_table; 655205264Skmacy struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 656168404Spjd} buf_hash_table_t; 657168404Spjd 658168404Spjdstatic buf_hash_table_t buf_hash_table; 659168404Spjd 660168404Spjd#define BUF_HASH_INDEX(spa, dva, birth) \ 661168404Spjd (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 662168404Spjd#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 663168404Spjd#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 664219089Spjd#define HDR_LOCK(hdr) \ 665219089Spjd (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 666168404Spjd 667168404Spjduint64_t zfs_crc64_table[256]; 668168404Spjd 669185029Spjd/* 670185029Spjd * Level 2 ARC 671185029Spjd */ 672185029Spjd 673208373Smm#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 674251478Sdelphij#define L2ARC_HEADROOM 2 /* num of writes */ 675251478Sdelphij/* 676251478Sdelphij * If we discover during ARC scan any buffers to be compressed, we boost 677251478Sdelphij * our headroom for the next scanning cycle by this percentage multiple. 678251478Sdelphij */ 679251478Sdelphij#define L2ARC_HEADROOM_BOOST 200 680208373Smm#define L2ARC_FEED_SECS 1 /* caching interval secs */ 681208373Smm#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 682185029Spjd 683185029Spjd#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 684185029Spjd#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 685185029Spjd 686251631Sdelphij/* L2ARC Performance Tunables */ 687185029Spjduint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 688185029Spjduint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 689185029Spjduint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 690251478Sdelphijuint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 691185029Spjduint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 692208373Smmuint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 693219089Spjdboolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 694208373Smmboolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 695208373Smmboolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 696185029Spjd 697217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 698205231Skmacy &l2arc_write_max, 0, "max write size"); 699217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 700205231Skmacy &l2arc_write_boost, 0, "extra write during warmup"); 701217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 702205231Skmacy &l2arc_headroom, 0, "number of dev writes"); 703217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 704205231Skmacy &l2arc_feed_secs, 0, "interval seconds"); 705217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 706208373Smm &l2arc_feed_min_ms, 0, "min interval milliseconds"); 707205231Skmacy 708205231SkmacySYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 709205231Skmacy &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 710208373SmmSYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 711208373Smm &l2arc_feed_again, 0, "turbo warmup"); 712208373SmmSYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 713208373Smm &l2arc_norw, 0, "no reads during writes"); 714205231Skmacy 715217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 716205231Skmacy &ARC_anon.arcs_size, 0, "size of anonymous state"); 717217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 718205231Skmacy &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 719217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 720205231Skmacy &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 721205231Skmacy 722217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 723205231Skmacy &ARC_mru.arcs_size, 0, "size of mru state"); 724217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 725205231Skmacy &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 726217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 727205231Skmacy &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 728205231Skmacy 729217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 730205231Skmacy &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 731217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 732205231Skmacy &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 733205231Skmacy "size of metadata in mru ghost state"); 734217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 735205231Skmacy &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 736205231Skmacy "size of data in mru ghost state"); 737205231Skmacy 738217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 739205231Skmacy &ARC_mfu.arcs_size, 0, "size of mfu state"); 740217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 741205231Skmacy &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 742217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 743205231Skmacy &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 744205231Skmacy 745217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 746205231Skmacy &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 747217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 748205231Skmacy &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 749205231Skmacy "size of metadata in mfu ghost state"); 750217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 751205231Skmacy &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 752205231Skmacy "size of data in mfu ghost state"); 753205231Skmacy 754217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 755205231Skmacy &ARC_l2c_only.arcs_size, 0, "size of mru state"); 756205231Skmacy 757185029Spjd/* 758185029Spjd * L2ARC Internals 759185029Spjd */ 760185029Spjdtypedef struct l2arc_dev { 761185029Spjd vdev_t *l2ad_vdev; /* vdev */ 762185029Spjd spa_t *l2ad_spa; /* spa */ 763185029Spjd uint64_t l2ad_hand; /* next write location */ 764185029Spjd uint64_t l2ad_start; /* first addr on device */ 765185029Spjd uint64_t l2ad_end; /* last addr on device */ 766185029Spjd uint64_t l2ad_evict; /* last addr eviction reached */ 767185029Spjd boolean_t l2ad_first; /* first sweep through */ 768208373Smm boolean_t l2ad_writing; /* currently writing */ 769185029Spjd list_t *l2ad_buflist; /* buffer list */ 770185029Spjd list_node_t l2ad_node; /* device list node */ 771185029Spjd} l2arc_dev_t; 772185029Spjd 773185029Spjdstatic list_t L2ARC_dev_list; /* device list */ 774185029Spjdstatic list_t *l2arc_dev_list; /* device list pointer */ 775185029Spjdstatic kmutex_t l2arc_dev_mtx; /* device list mutex */ 776185029Spjdstatic l2arc_dev_t *l2arc_dev_last; /* last device used */ 777185029Spjdstatic kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 778185029Spjdstatic list_t L2ARC_free_on_write; /* free after write buf list */ 779185029Spjdstatic list_t *l2arc_free_on_write; /* free after write list ptr */ 780185029Spjdstatic kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 781185029Spjdstatic uint64_t l2arc_ndev; /* number of devices */ 782185029Spjd 783185029Spjdtypedef struct l2arc_read_callback { 784251478Sdelphij arc_buf_t *l2rcb_buf; /* read buffer */ 785251478Sdelphij spa_t *l2rcb_spa; /* spa */ 786251478Sdelphij blkptr_t l2rcb_bp; /* original blkptr */ 787268123Sdelphij zbookmark_phys_t l2rcb_zb; /* original bookmark */ 788251478Sdelphij int l2rcb_flags; /* original flags */ 789251478Sdelphij enum zio_compress l2rcb_compress; /* applied compress */ 790185029Spjd} l2arc_read_callback_t; 791185029Spjd 792185029Spjdtypedef struct l2arc_write_callback { 793185029Spjd l2arc_dev_t *l2wcb_dev; /* device info */ 794185029Spjd arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 795185029Spjd} l2arc_write_callback_t; 796185029Spjd 797185029Spjdstruct l2arc_buf_hdr { 798185029Spjd /* protected by arc_buf_hdr mutex */ 799251478Sdelphij l2arc_dev_t *b_dev; /* L2ARC device */ 800251478Sdelphij uint64_t b_daddr; /* disk address, offset byte */ 801251478Sdelphij /* compression applied to buffer data */ 802251478Sdelphij enum zio_compress b_compress; 803251478Sdelphij /* real alloc'd buffer size depending on b_compress applied */ 804251478Sdelphij int b_asize; 805251478Sdelphij /* temporary buffer holder for in-flight compressed data */ 806251478Sdelphij void *b_tmp_cdata; 807185029Spjd}; 808185029Spjd 809185029Spjdtypedef struct l2arc_data_free { 810185029Spjd /* protected by l2arc_free_on_write_mtx */ 811185029Spjd void *l2df_data; 812185029Spjd size_t l2df_size; 813185029Spjd void (*l2df_func)(void *, size_t); 814185029Spjd list_node_t l2df_list_node; 815185029Spjd} l2arc_data_free_t; 816185029Spjd 817185029Spjdstatic kmutex_t l2arc_feed_thr_lock; 818185029Spjdstatic kcondvar_t l2arc_feed_thr_cv; 819185029Spjdstatic uint8_t l2arc_thread_exit; 820185029Spjd 821185029Spjdstatic void l2arc_read_done(zio_t *zio); 822185029Spjdstatic void l2arc_hdr_stat_add(void); 823185029Spjdstatic void l2arc_hdr_stat_remove(void); 824185029Spjd 825251478Sdelphijstatic boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr); 826251478Sdelphijstatic void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, 827251478Sdelphij enum zio_compress c); 828251478Sdelphijstatic void l2arc_release_cdata_buf(arc_buf_hdr_t *ab); 829251478Sdelphij 830168404Spjdstatic uint64_t 831209962Smmbuf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 832168404Spjd{ 833168404Spjd uint8_t *vdva = (uint8_t *)dva; 834168404Spjd uint64_t crc = -1ULL; 835168404Spjd int i; 836168404Spjd 837168404Spjd ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 838168404Spjd 839168404Spjd for (i = 0; i < sizeof (dva_t); i++) 840168404Spjd crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 841168404Spjd 842209962Smm crc ^= (spa>>8) ^ birth; 843168404Spjd 844168404Spjd return (crc); 845168404Spjd} 846168404Spjd 847168404Spjd#define BUF_EMPTY(buf) \ 848168404Spjd ((buf)->b_dva.dva_word[0] == 0 && \ 849168404Spjd (buf)->b_dva.dva_word[1] == 0 && \ 850260150Sdelphij (buf)->b_cksum0 == 0) 851168404Spjd 852168404Spjd#define BUF_EQUAL(spa, dva, birth, buf) \ 853168404Spjd ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 854168404Spjd ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 855168404Spjd ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 856168404Spjd 857219089Spjdstatic void 858219089Spjdbuf_discard_identity(arc_buf_hdr_t *hdr) 859219089Spjd{ 860219089Spjd hdr->b_dva.dva_word[0] = 0; 861219089Spjd hdr->b_dva.dva_word[1] = 0; 862219089Spjd hdr->b_birth = 0; 863219089Spjd hdr->b_cksum0 = 0; 864219089Spjd} 865219089Spjd 866168404Spjdstatic arc_buf_hdr_t * 867268075Sdelphijbuf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 868168404Spjd{ 869268075Sdelphij const dva_t *dva = BP_IDENTITY(bp); 870268075Sdelphij uint64_t birth = BP_PHYSICAL_BIRTH(bp); 871168404Spjd uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 872168404Spjd kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 873168404Spjd arc_buf_hdr_t *buf; 874168404Spjd 875168404Spjd mutex_enter(hash_lock); 876168404Spjd for (buf = buf_hash_table.ht_table[idx]; buf != NULL; 877168404Spjd buf = buf->b_hash_next) { 878168404Spjd if (BUF_EQUAL(spa, dva, birth, buf)) { 879168404Spjd *lockp = hash_lock; 880168404Spjd return (buf); 881168404Spjd } 882168404Spjd } 883168404Spjd mutex_exit(hash_lock); 884168404Spjd *lockp = NULL; 885168404Spjd return (NULL); 886168404Spjd} 887168404Spjd 888168404Spjd/* 889168404Spjd * Insert an entry into the hash table. If there is already an element 890168404Spjd * equal to elem in the hash table, then the already existing element 891168404Spjd * will be returned and the new element will not be inserted. 892168404Spjd * Otherwise returns NULL. 893168404Spjd */ 894168404Spjdstatic arc_buf_hdr_t * 895168404Spjdbuf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp) 896168404Spjd{ 897168404Spjd uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 898168404Spjd kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 899168404Spjd arc_buf_hdr_t *fbuf; 900168404Spjd uint32_t i; 901168404Spjd 902268075Sdelphij ASSERT(!DVA_IS_EMPTY(&buf->b_dva)); 903268075Sdelphij ASSERT(buf->b_birth != 0); 904168404Spjd ASSERT(!HDR_IN_HASH_TABLE(buf)); 905168404Spjd *lockp = hash_lock; 906168404Spjd mutex_enter(hash_lock); 907168404Spjd for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL; 908168404Spjd fbuf = fbuf->b_hash_next, i++) { 909168404Spjd if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf)) 910168404Spjd return (fbuf); 911168404Spjd } 912168404Spjd 913168404Spjd buf->b_hash_next = buf_hash_table.ht_table[idx]; 914168404Spjd buf_hash_table.ht_table[idx] = buf; 915168404Spjd buf->b_flags |= ARC_IN_HASH_TABLE; 916168404Spjd 917168404Spjd /* collect some hash table performance data */ 918168404Spjd if (i > 0) { 919168404Spjd ARCSTAT_BUMP(arcstat_hash_collisions); 920168404Spjd if (i == 1) 921168404Spjd ARCSTAT_BUMP(arcstat_hash_chains); 922168404Spjd 923168404Spjd ARCSTAT_MAX(arcstat_hash_chain_max, i); 924168404Spjd } 925168404Spjd 926168404Spjd ARCSTAT_BUMP(arcstat_hash_elements); 927168404Spjd ARCSTAT_MAXSTAT(arcstat_hash_elements); 928168404Spjd 929168404Spjd return (NULL); 930168404Spjd} 931168404Spjd 932168404Spjdstatic void 933168404Spjdbuf_hash_remove(arc_buf_hdr_t *buf) 934168404Spjd{ 935168404Spjd arc_buf_hdr_t *fbuf, **bufp; 936168404Spjd uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 937168404Spjd 938168404Spjd ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 939168404Spjd ASSERT(HDR_IN_HASH_TABLE(buf)); 940168404Spjd 941168404Spjd bufp = &buf_hash_table.ht_table[idx]; 942168404Spjd while ((fbuf = *bufp) != buf) { 943168404Spjd ASSERT(fbuf != NULL); 944168404Spjd bufp = &fbuf->b_hash_next; 945168404Spjd } 946168404Spjd *bufp = buf->b_hash_next; 947168404Spjd buf->b_hash_next = NULL; 948168404Spjd buf->b_flags &= ~ARC_IN_HASH_TABLE; 949168404Spjd 950168404Spjd /* collect some hash table performance data */ 951168404Spjd ARCSTAT_BUMPDOWN(arcstat_hash_elements); 952168404Spjd 953168404Spjd if (buf_hash_table.ht_table[idx] && 954168404Spjd buf_hash_table.ht_table[idx]->b_hash_next == NULL) 955168404Spjd ARCSTAT_BUMPDOWN(arcstat_hash_chains); 956168404Spjd} 957168404Spjd 958168404Spjd/* 959168404Spjd * Global data structures and functions for the buf kmem cache. 960168404Spjd */ 961168404Spjdstatic kmem_cache_t *hdr_cache; 962168404Spjdstatic kmem_cache_t *buf_cache; 963168404Spjd 964168404Spjdstatic void 965168404Spjdbuf_fini(void) 966168404Spjd{ 967168404Spjd int i; 968168404Spjd 969168404Spjd kmem_free(buf_hash_table.ht_table, 970168404Spjd (buf_hash_table.ht_mask + 1) * sizeof (void *)); 971168404Spjd for (i = 0; i < BUF_LOCKS; i++) 972168404Spjd mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 973168404Spjd kmem_cache_destroy(hdr_cache); 974168404Spjd kmem_cache_destroy(buf_cache); 975168404Spjd} 976168404Spjd 977168404Spjd/* 978168404Spjd * Constructor callback - called when the cache is empty 979168404Spjd * and a new buf is requested. 980168404Spjd */ 981168404Spjd/* ARGSUSED */ 982168404Spjdstatic int 983168404Spjdhdr_cons(void *vbuf, void *unused, int kmflag) 984168404Spjd{ 985168404Spjd arc_buf_hdr_t *buf = vbuf; 986168404Spjd 987168404Spjd bzero(buf, sizeof (arc_buf_hdr_t)); 988168404Spjd refcount_create(&buf->b_refcnt); 989168404Spjd cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL); 990185029Spjd mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 991208373Smm arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 992185029Spjd 993168404Spjd return (0); 994168404Spjd} 995168404Spjd 996185029Spjd/* ARGSUSED */ 997185029Spjdstatic int 998185029Spjdbuf_cons(void *vbuf, void *unused, int kmflag) 999185029Spjd{ 1000185029Spjd arc_buf_t *buf = vbuf; 1001185029Spjd 1002185029Spjd bzero(buf, sizeof (arc_buf_t)); 1003219089Spjd mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1004208373Smm arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1005208373Smm 1006185029Spjd return (0); 1007185029Spjd} 1008185029Spjd 1009168404Spjd/* 1010168404Spjd * Destructor callback - called when a cached buf is 1011168404Spjd * no longer required. 1012168404Spjd */ 1013168404Spjd/* ARGSUSED */ 1014168404Spjdstatic void 1015168404Spjdhdr_dest(void *vbuf, void *unused) 1016168404Spjd{ 1017168404Spjd arc_buf_hdr_t *buf = vbuf; 1018168404Spjd 1019219089Spjd ASSERT(BUF_EMPTY(buf)); 1020168404Spjd refcount_destroy(&buf->b_refcnt); 1021168404Spjd cv_destroy(&buf->b_cv); 1022185029Spjd mutex_destroy(&buf->b_freeze_lock); 1023208373Smm arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1024168404Spjd} 1025168404Spjd 1026185029Spjd/* ARGSUSED */ 1027185029Spjdstatic void 1028185029Spjdbuf_dest(void *vbuf, void *unused) 1029185029Spjd{ 1030185029Spjd arc_buf_t *buf = vbuf; 1031185029Spjd 1032219089Spjd mutex_destroy(&buf->b_evict_lock); 1033208373Smm arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1034185029Spjd} 1035185029Spjd 1036168404Spjd/* 1037168404Spjd * Reclaim callback -- invoked when memory is low. 1038168404Spjd */ 1039168404Spjd/* ARGSUSED */ 1040168404Spjdstatic void 1041168404Spjdhdr_recl(void *unused) 1042168404Spjd{ 1043168404Spjd dprintf("hdr_recl called\n"); 1044168404Spjd /* 1045168404Spjd * umem calls the reclaim func when we destroy the buf cache, 1046168404Spjd * which is after we do arc_fini(). 1047168404Spjd */ 1048168404Spjd if (!arc_dead) 1049168404Spjd cv_signal(&arc_reclaim_thr_cv); 1050168404Spjd} 1051168404Spjd 1052168404Spjdstatic void 1053168404Spjdbuf_init(void) 1054168404Spjd{ 1055168404Spjd uint64_t *ct; 1056168404Spjd uint64_t hsize = 1ULL << 12; 1057168404Spjd int i, j; 1058168404Spjd 1059168404Spjd /* 1060168404Spjd * The hash table is big enough to fill all of physical memory 1061269230Sdelphij * with an average block size of zfs_arc_average_blocksize (default 8K). 1062269230Sdelphij * By default, the table will take up 1063269230Sdelphij * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1064168404Spjd */ 1065269230Sdelphij while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1066168404Spjd hsize <<= 1; 1067168404Spjdretry: 1068168404Spjd buf_hash_table.ht_mask = hsize - 1; 1069168404Spjd buf_hash_table.ht_table = 1070168404Spjd kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1071168404Spjd if (buf_hash_table.ht_table == NULL) { 1072168404Spjd ASSERT(hsize > (1ULL << 8)); 1073168404Spjd hsize >>= 1; 1074168404Spjd goto retry; 1075168404Spjd } 1076168404Spjd 1077168404Spjd hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1078168404Spjd 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1079168404Spjd buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1080185029Spjd 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1081168404Spjd 1082168404Spjd for (i = 0; i < 256; i++) 1083168404Spjd for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1084168404Spjd *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1085168404Spjd 1086168404Spjd for (i = 0; i < BUF_LOCKS; i++) { 1087168404Spjd mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1088168404Spjd NULL, MUTEX_DEFAULT, NULL); 1089168404Spjd } 1090168404Spjd} 1091168404Spjd 1092168404Spjd#define ARC_MINTIME (hz>>4) /* 62 ms */ 1093168404Spjd 1094168404Spjdstatic void 1095168404Spjdarc_cksum_verify(arc_buf_t *buf) 1096168404Spjd{ 1097168404Spjd zio_cksum_t zc; 1098168404Spjd 1099168404Spjd if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1100168404Spjd return; 1101168404Spjd 1102168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1103168404Spjd if (buf->b_hdr->b_freeze_cksum == NULL || 1104168404Spjd (buf->b_hdr->b_flags & ARC_IO_ERROR)) { 1105168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1106168404Spjd return; 1107168404Spjd } 1108168404Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1109168404Spjd if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1110168404Spjd panic("buffer modified while frozen!"); 1111168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1112168404Spjd} 1113168404Spjd 1114185029Spjdstatic int 1115185029Spjdarc_cksum_equal(arc_buf_t *buf) 1116185029Spjd{ 1117185029Spjd zio_cksum_t zc; 1118185029Spjd int equal; 1119185029Spjd 1120185029Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1121185029Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1122185029Spjd equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1123185029Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1124185029Spjd 1125185029Spjd return (equal); 1126185029Spjd} 1127185029Spjd 1128168404Spjdstatic void 1129185029Spjdarc_cksum_compute(arc_buf_t *buf, boolean_t force) 1130168404Spjd{ 1131185029Spjd if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1132168404Spjd return; 1133168404Spjd 1134168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1135168404Spjd if (buf->b_hdr->b_freeze_cksum != NULL) { 1136168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1137168404Spjd return; 1138168404Spjd } 1139168404Spjd buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1140168404Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1141168404Spjd buf->b_hdr->b_freeze_cksum); 1142168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1143240133Smm#ifdef illumos 1144240133Smm arc_buf_watch(buf); 1145240133Smm#endif /* illumos */ 1146168404Spjd} 1147168404Spjd 1148240133Smm#ifdef illumos 1149240133Smm#ifndef _KERNEL 1150240133Smmtypedef struct procctl { 1151240133Smm long cmd; 1152240133Smm prwatch_t prwatch; 1153240133Smm} procctl_t; 1154240133Smm#endif 1155240133Smm 1156240133Smm/* ARGSUSED */ 1157240133Smmstatic void 1158240133Smmarc_buf_unwatch(arc_buf_t *buf) 1159240133Smm{ 1160240133Smm#ifndef _KERNEL 1161240133Smm if (arc_watch) { 1162240133Smm int result; 1163240133Smm procctl_t ctl; 1164240133Smm ctl.cmd = PCWATCH; 1165240133Smm ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1166240133Smm ctl.prwatch.pr_size = 0; 1167240133Smm ctl.prwatch.pr_wflags = 0; 1168240133Smm result = write(arc_procfd, &ctl, sizeof (ctl)); 1169240133Smm ASSERT3U(result, ==, sizeof (ctl)); 1170240133Smm } 1171240133Smm#endif 1172240133Smm} 1173240133Smm 1174240133Smm/* ARGSUSED */ 1175240133Smmstatic void 1176240133Smmarc_buf_watch(arc_buf_t *buf) 1177240133Smm{ 1178240133Smm#ifndef _KERNEL 1179240133Smm if (arc_watch) { 1180240133Smm int result; 1181240133Smm procctl_t ctl; 1182240133Smm ctl.cmd = PCWATCH; 1183240133Smm ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1184240133Smm ctl.prwatch.pr_size = buf->b_hdr->b_size; 1185240133Smm ctl.prwatch.pr_wflags = WA_WRITE; 1186240133Smm result = write(arc_procfd, &ctl, sizeof (ctl)); 1187240133Smm ASSERT3U(result, ==, sizeof (ctl)); 1188240133Smm } 1189240133Smm#endif 1190240133Smm} 1191240133Smm#endif /* illumos */ 1192240133Smm 1193168404Spjdvoid 1194168404Spjdarc_buf_thaw(arc_buf_t *buf) 1195168404Spjd{ 1196185029Spjd if (zfs_flags & ZFS_DEBUG_MODIFY) { 1197185029Spjd if (buf->b_hdr->b_state != arc_anon) 1198185029Spjd panic("modifying non-anon buffer!"); 1199185029Spjd if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS) 1200185029Spjd panic("modifying buffer while i/o in progress!"); 1201185029Spjd arc_cksum_verify(buf); 1202185029Spjd } 1203168404Spjd 1204168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1205168404Spjd if (buf->b_hdr->b_freeze_cksum != NULL) { 1206168404Spjd kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1207168404Spjd buf->b_hdr->b_freeze_cksum = NULL; 1208168404Spjd } 1209219089Spjd 1210219089Spjd if (zfs_flags & ZFS_DEBUG_MODIFY) { 1211219089Spjd if (buf->b_hdr->b_thawed) 1212219089Spjd kmem_free(buf->b_hdr->b_thawed, 1); 1213219089Spjd buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1214219089Spjd } 1215219089Spjd 1216168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1217240133Smm 1218240133Smm#ifdef illumos 1219240133Smm arc_buf_unwatch(buf); 1220240133Smm#endif /* illumos */ 1221168404Spjd} 1222168404Spjd 1223168404Spjdvoid 1224168404Spjdarc_buf_freeze(arc_buf_t *buf) 1225168404Spjd{ 1226219089Spjd kmutex_t *hash_lock; 1227219089Spjd 1228168404Spjd if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1229168404Spjd return; 1230168404Spjd 1231219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 1232219089Spjd mutex_enter(hash_lock); 1233219089Spjd 1234168404Spjd ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1235168404Spjd buf->b_hdr->b_state == arc_anon); 1236185029Spjd arc_cksum_compute(buf, B_FALSE); 1237219089Spjd mutex_exit(hash_lock); 1238240133Smm 1239168404Spjd} 1240168404Spjd 1241168404Spjdstatic void 1242205231Skmacyget_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock) 1243205231Skmacy{ 1244205231Skmacy uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth); 1245205231Skmacy 1246206796Spjd if (ab->b_type == ARC_BUFC_METADATA) 1247206796Spjd buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1248205231Skmacy else { 1249206796Spjd buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1250205231Skmacy buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1251205231Skmacy } 1252205231Skmacy 1253205231Skmacy *list = &state->arcs_lists[buf_hashid]; 1254205231Skmacy *lock = ARCS_LOCK(state, buf_hashid); 1255205231Skmacy} 1256205231Skmacy 1257205231Skmacy 1258205231Skmacystatic void 1259168404Spjdadd_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1260168404Spjd{ 1261168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 1262168404Spjd 1263168404Spjd if ((refcount_add(&ab->b_refcnt, tag) == 1) && 1264168404Spjd (ab->b_state != arc_anon)) { 1265206796Spjd uint64_t delta = ab->b_size * ab->b_datacnt; 1266206796Spjd uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type]; 1267205231Skmacy list_t *list; 1268205231Skmacy kmutex_t *lock; 1269168404Spjd 1270205231Skmacy get_buf_info(ab, ab->b_state, &list, &lock); 1271205231Skmacy ASSERT(!MUTEX_HELD(lock)); 1272205231Skmacy mutex_enter(lock); 1273168404Spjd ASSERT(list_link_active(&ab->b_arc_node)); 1274185029Spjd list_remove(list, ab); 1275168404Spjd if (GHOST_STATE(ab->b_state)) { 1276240415Smm ASSERT0(ab->b_datacnt); 1277168404Spjd ASSERT3P(ab->b_buf, ==, NULL); 1278168404Spjd delta = ab->b_size; 1279168404Spjd } 1280168404Spjd ASSERT(delta > 0); 1281185029Spjd ASSERT3U(*size, >=, delta); 1282185029Spjd atomic_add_64(size, -delta); 1283206794Spjd mutex_exit(lock); 1284185029Spjd /* remove the prefetch flag if we get a reference */ 1285168404Spjd if (ab->b_flags & ARC_PREFETCH) 1286168404Spjd ab->b_flags &= ~ARC_PREFETCH; 1287168404Spjd } 1288168404Spjd} 1289168404Spjd 1290168404Spjdstatic int 1291168404Spjdremove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1292168404Spjd{ 1293168404Spjd int cnt; 1294168404Spjd arc_state_t *state = ab->b_state; 1295168404Spjd 1296168404Spjd ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1297168404Spjd ASSERT(!GHOST_STATE(state)); 1298168404Spjd 1299168404Spjd if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) && 1300168404Spjd (state != arc_anon)) { 1301185029Spjd uint64_t *size = &state->arcs_lsize[ab->b_type]; 1302205231Skmacy list_t *list; 1303205231Skmacy kmutex_t *lock; 1304185029Spjd 1305205231Skmacy get_buf_info(ab, state, &list, &lock); 1306205231Skmacy ASSERT(!MUTEX_HELD(lock)); 1307205231Skmacy mutex_enter(lock); 1308168404Spjd ASSERT(!list_link_active(&ab->b_arc_node)); 1309205231Skmacy list_insert_head(list, ab); 1310168404Spjd ASSERT(ab->b_datacnt > 0); 1311185029Spjd atomic_add_64(size, ab->b_size * ab->b_datacnt); 1312206794Spjd mutex_exit(lock); 1313168404Spjd } 1314168404Spjd return (cnt); 1315168404Spjd} 1316168404Spjd 1317168404Spjd/* 1318168404Spjd * Move the supplied buffer to the indicated state. The mutex 1319168404Spjd * for the buffer must be held by the caller. 1320168404Spjd */ 1321168404Spjdstatic void 1322168404Spjdarc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock) 1323168404Spjd{ 1324168404Spjd arc_state_t *old_state = ab->b_state; 1325168404Spjd int64_t refcnt = refcount_count(&ab->b_refcnt); 1326168404Spjd uint64_t from_delta, to_delta; 1327205231Skmacy list_t *list; 1328205231Skmacy kmutex_t *lock; 1329168404Spjd 1330168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 1331258632Savg ASSERT3P(new_state, !=, old_state); 1332168404Spjd ASSERT(refcnt == 0 || ab->b_datacnt > 0); 1333168404Spjd ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state)); 1334219089Spjd ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon); 1335168404Spjd 1336168404Spjd from_delta = to_delta = ab->b_datacnt * ab->b_size; 1337168404Spjd 1338168404Spjd /* 1339168404Spjd * If this buffer is evictable, transfer it from the 1340168404Spjd * old state list to the new state list. 1341168404Spjd */ 1342168404Spjd if (refcnt == 0) { 1343168404Spjd if (old_state != arc_anon) { 1344205231Skmacy int use_mutex; 1345185029Spjd uint64_t *size = &old_state->arcs_lsize[ab->b_type]; 1346168404Spjd 1347205231Skmacy get_buf_info(ab, old_state, &list, &lock); 1348205231Skmacy use_mutex = !MUTEX_HELD(lock); 1349168404Spjd if (use_mutex) 1350205231Skmacy mutex_enter(lock); 1351168404Spjd 1352168404Spjd ASSERT(list_link_active(&ab->b_arc_node)); 1353205231Skmacy list_remove(list, ab); 1354168404Spjd 1355168404Spjd /* 1356168404Spjd * If prefetching out of the ghost cache, 1357219089Spjd * we will have a non-zero datacnt. 1358168404Spjd */ 1359168404Spjd if (GHOST_STATE(old_state) && ab->b_datacnt == 0) { 1360168404Spjd /* ghost elements have a ghost size */ 1361168404Spjd ASSERT(ab->b_buf == NULL); 1362168404Spjd from_delta = ab->b_size; 1363168404Spjd } 1364185029Spjd ASSERT3U(*size, >=, from_delta); 1365185029Spjd atomic_add_64(size, -from_delta); 1366168404Spjd 1367168404Spjd if (use_mutex) 1368205231Skmacy mutex_exit(lock); 1369168404Spjd } 1370168404Spjd if (new_state != arc_anon) { 1371206796Spjd int use_mutex; 1372185029Spjd uint64_t *size = &new_state->arcs_lsize[ab->b_type]; 1373168404Spjd 1374205231Skmacy get_buf_info(ab, new_state, &list, &lock); 1375205231Skmacy use_mutex = !MUTEX_HELD(lock); 1376168404Spjd if (use_mutex) 1377205231Skmacy mutex_enter(lock); 1378168404Spjd 1379205231Skmacy list_insert_head(list, ab); 1380168404Spjd 1381168404Spjd /* ghost elements have a ghost size */ 1382168404Spjd if (GHOST_STATE(new_state)) { 1383168404Spjd ASSERT(ab->b_datacnt == 0); 1384168404Spjd ASSERT(ab->b_buf == NULL); 1385168404Spjd to_delta = ab->b_size; 1386168404Spjd } 1387185029Spjd atomic_add_64(size, to_delta); 1388168404Spjd 1389168404Spjd if (use_mutex) 1390205231Skmacy mutex_exit(lock); 1391168404Spjd } 1392168404Spjd } 1393168404Spjd 1394168404Spjd ASSERT(!BUF_EMPTY(ab)); 1395219089Spjd if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab)) 1396168404Spjd buf_hash_remove(ab); 1397168404Spjd 1398168404Spjd /* adjust state sizes */ 1399168404Spjd if (to_delta) 1400168404Spjd atomic_add_64(&new_state->arcs_size, to_delta); 1401168404Spjd if (from_delta) { 1402168404Spjd ASSERT3U(old_state->arcs_size, >=, from_delta); 1403168404Spjd atomic_add_64(&old_state->arcs_size, -from_delta); 1404168404Spjd } 1405168404Spjd ab->b_state = new_state; 1406185029Spjd 1407185029Spjd /* adjust l2arc hdr stats */ 1408185029Spjd if (new_state == arc_l2c_only) 1409185029Spjd l2arc_hdr_stat_add(); 1410185029Spjd else if (old_state == arc_l2c_only) 1411185029Spjd l2arc_hdr_stat_remove(); 1412168404Spjd} 1413168404Spjd 1414185029Spjdvoid 1415208373Smmarc_space_consume(uint64_t space, arc_space_type_t type) 1416185029Spjd{ 1417208373Smm ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1418208373Smm 1419208373Smm switch (type) { 1420208373Smm case ARC_SPACE_DATA: 1421208373Smm ARCSTAT_INCR(arcstat_data_size, space); 1422208373Smm break; 1423208373Smm case ARC_SPACE_OTHER: 1424208373Smm ARCSTAT_INCR(arcstat_other_size, space); 1425208373Smm break; 1426208373Smm case ARC_SPACE_HDRS: 1427208373Smm ARCSTAT_INCR(arcstat_hdr_size, space); 1428208373Smm break; 1429208373Smm case ARC_SPACE_L2HDRS: 1430208373Smm ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1431208373Smm break; 1432208373Smm } 1433208373Smm 1434185029Spjd atomic_add_64(&arc_meta_used, space); 1435185029Spjd atomic_add_64(&arc_size, space); 1436185029Spjd} 1437185029Spjd 1438185029Spjdvoid 1439208373Smmarc_space_return(uint64_t space, arc_space_type_t type) 1440185029Spjd{ 1441208373Smm ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1442208373Smm 1443208373Smm switch (type) { 1444208373Smm case ARC_SPACE_DATA: 1445208373Smm ARCSTAT_INCR(arcstat_data_size, -space); 1446208373Smm break; 1447208373Smm case ARC_SPACE_OTHER: 1448208373Smm ARCSTAT_INCR(arcstat_other_size, -space); 1449208373Smm break; 1450208373Smm case ARC_SPACE_HDRS: 1451208373Smm ARCSTAT_INCR(arcstat_hdr_size, -space); 1452208373Smm break; 1453208373Smm case ARC_SPACE_L2HDRS: 1454208373Smm ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1455208373Smm break; 1456208373Smm } 1457208373Smm 1458185029Spjd ASSERT(arc_meta_used >= space); 1459185029Spjd if (arc_meta_max < arc_meta_used) 1460185029Spjd arc_meta_max = arc_meta_used; 1461185029Spjd atomic_add_64(&arc_meta_used, -space); 1462185029Spjd ASSERT(arc_size >= space); 1463185029Spjd atomic_add_64(&arc_size, -space); 1464185029Spjd} 1465185029Spjd 1466185029Spjdvoid * 1467185029Spjdarc_data_buf_alloc(uint64_t size) 1468185029Spjd{ 1469185029Spjd if (arc_evict_needed(ARC_BUFC_DATA)) 1470185029Spjd cv_signal(&arc_reclaim_thr_cv); 1471185029Spjd atomic_add_64(&arc_size, size); 1472185029Spjd return (zio_data_buf_alloc(size)); 1473185029Spjd} 1474185029Spjd 1475185029Spjdvoid 1476185029Spjdarc_data_buf_free(void *buf, uint64_t size) 1477185029Spjd{ 1478185029Spjd zio_data_buf_free(buf, size); 1479185029Spjd ASSERT(arc_size >= size); 1480185029Spjd atomic_add_64(&arc_size, -size); 1481185029Spjd} 1482185029Spjd 1483168404Spjdarc_buf_t * 1484168404Spjdarc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1485168404Spjd{ 1486168404Spjd arc_buf_hdr_t *hdr; 1487168404Spjd arc_buf_t *buf; 1488168404Spjd 1489168404Spjd ASSERT3U(size, >, 0); 1490185029Spjd hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1491168404Spjd ASSERT(BUF_EMPTY(hdr)); 1492168404Spjd hdr->b_size = size; 1493168404Spjd hdr->b_type = type; 1494228103Smm hdr->b_spa = spa_load_guid(spa); 1495168404Spjd hdr->b_state = arc_anon; 1496168404Spjd hdr->b_arc_access = 0; 1497185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1498168404Spjd buf->b_hdr = hdr; 1499168404Spjd buf->b_data = NULL; 1500168404Spjd buf->b_efunc = NULL; 1501168404Spjd buf->b_private = NULL; 1502168404Spjd buf->b_next = NULL; 1503168404Spjd hdr->b_buf = buf; 1504168404Spjd arc_get_data_buf(buf); 1505168404Spjd hdr->b_datacnt = 1; 1506168404Spjd hdr->b_flags = 0; 1507168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1508168404Spjd (void) refcount_add(&hdr->b_refcnt, tag); 1509168404Spjd 1510168404Spjd return (buf); 1511168404Spjd} 1512168404Spjd 1513209962Smmstatic char *arc_onloan_tag = "onloan"; 1514209962Smm 1515209962Smm/* 1516209962Smm * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1517209962Smm * flight data by arc_tempreserve_space() until they are "returned". Loaned 1518209962Smm * buffers must be returned to the arc before they can be used by the DMU or 1519209962Smm * freed. 1520209962Smm */ 1521209962Smmarc_buf_t * 1522209962Smmarc_loan_buf(spa_t *spa, int size) 1523209962Smm{ 1524209962Smm arc_buf_t *buf; 1525209962Smm 1526209962Smm buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1527209962Smm 1528209962Smm atomic_add_64(&arc_loaned_bytes, size); 1529209962Smm return (buf); 1530209962Smm} 1531209962Smm 1532209962Smm/* 1533209962Smm * Return a loaned arc buffer to the arc. 1534209962Smm */ 1535209962Smmvoid 1536209962Smmarc_return_buf(arc_buf_t *buf, void *tag) 1537209962Smm{ 1538209962Smm arc_buf_hdr_t *hdr = buf->b_hdr; 1539209962Smm 1540209962Smm ASSERT(buf->b_data != NULL); 1541219089Spjd (void) refcount_add(&hdr->b_refcnt, tag); 1542219089Spjd (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1543209962Smm 1544209962Smm atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1545209962Smm} 1546209962Smm 1547219089Spjd/* Detach an arc_buf from a dbuf (tag) */ 1548219089Spjdvoid 1549219089Spjdarc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1550219089Spjd{ 1551219089Spjd arc_buf_hdr_t *hdr; 1552219089Spjd 1553219089Spjd ASSERT(buf->b_data != NULL); 1554219089Spjd hdr = buf->b_hdr; 1555219089Spjd (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1556219089Spjd (void) refcount_remove(&hdr->b_refcnt, tag); 1557219089Spjd buf->b_efunc = NULL; 1558219089Spjd buf->b_private = NULL; 1559219089Spjd 1560219089Spjd atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1561219089Spjd} 1562219089Spjd 1563168404Spjdstatic arc_buf_t * 1564168404Spjdarc_buf_clone(arc_buf_t *from) 1565168404Spjd{ 1566168404Spjd arc_buf_t *buf; 1567168404Spjd arc_buf_hdr_t *hdr = from->b_hdr; 1568168404Spjd uint64_t size = hdr->b_size; 1569168404Spjd 1570219089Spjd ASSERT(hdr->b_state != arc_anon); 1571219089Spjd 1572185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1573168404Spjd buf->b_hdr = hdr; 1574168404Spjd buf->b_data = NULL; 1575168404Spjd buf->b_efunc = NULL; 1576168404Spjd buf->b_private = NULL; 1577168404Spjd buf->b_next = hdr->b_buf; 1578168404Spjd hdr->b_buf = buf; 1579168404Spjd arc_get_data_buf(buf); 1580168404Spjd bcopy(from->b_data, buf->b_data, size); 1581242845Sdelphij 1582242845Sdelphij /* 1583242845Sdelphij * This buffer already exists in the arc so create a duplicate 1584242845Sdelphij * copy for the caller. If the buffer is associated with user data 1585242845Sdelphij * then track the size and number of duplicates. These stats will be 1586242845Sdelphij * updated as duplicate buffers are created and destroyed. 1587242845Sdelphij */ 1588242845Sdelphij if (hdr->b_type == ARC_BUFC_DATA) { 1589242845Sdelphij ARCSTAT_BUMP(arcstat_duplicate_buffers); 1590242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1591242845Sdelphij } 1592168404Spjd hdr->b_datacnt += 1; 1593168404Spjd return (buf); 1594168404Spjd} 1595168404Spjd 1596168404Spjdvoid 1597168404Spjdarc_buf_add_ref(arc_buf_t *buf, void* tag) 1598168404Spjd{ 1599168404Spjd arc_buf_hdr_t *hdr; 1600168404Spjd kmutex_t *hash_lock; 1601168404Spjd 1602168404Spjd /* 1603185029Spjd * Check to see if this buffer is evicted. Callers 1604185029Spjd * must verify b_data != NULL to know if the add_ref 1605185029Spjd * was successful. 1606168404Spjd */ 1607219089Spjd mutex_enter(&buf->b_evict_lock); 1608185029Spjd if (buf->b_data == NULL) { 1609219089Spjd mutex_exit(&buf->b_evict_lock); 1610168404Spjd return; 1611168404Spjd } 1612219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 1613219089Spjd mutex_enter(hash_lock); 1614185029Spjd hdr = buf->b_hdr; 1615219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1616219089Spjd mutex_exit(&buf->b_evict_lock); 1617168404Spjd 1618168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1619168404Spjd add_reference(hdr, hash_lock, tag); 1620208373Smm DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1621168404Spjd arc_access(hdr, hash_lock); 1622168404Spjd mutex_exit(hash_lock); 1623168404Spjd ARCSTAT_BUMP(arcstat_hits); 1624168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 1625168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1626168404Spjd data, metadata, hits); 1627168404Spjd} 1628168404Spjd 1629185029Spjd/* 1630185029Spjd * Free the arc data buffer. If it is an l2arc write in progress, 1631185029Spjd * the buffer is placed on l2arc_free_on_write to be freed later. 1632185029Spjd */ 1633168404Spjdstatic void 1634240133Smmarc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1635185029Spjd{ 1636240133Smm arc_buf_hdr_t *hdr = buf->b_hdr; 1637240133Smm 1638185029Spjd if (HDR_L2_WRITING(hdr)) { 1639185029Spjd l2arc_data_free_t *df; 1640185029Spjd df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1641240133Smm df->l2df_data = buf->b_data; 1642240133Smm df->l2df_size = hdr->b_size; 1643185029Spjd df->l2df_func = free_func; 1644185029Spjd mutex_enter(&l2arc_free_on_write_mtx); 1645185029Spjd list_insert_head(l2arc_free_on_write, df); 1646185029Spjd mutex_exit(&l2arc_free_on_write_mtx); 1647185029Spjd ARCSTAT_BUMP(arcstat_l2_free_on_write); 1648185029Spjd } else { 1649240133Smm free_func(buf->b_data, hdr->b_size); 1650185029Spjd } 1651185029Spjd} 1652185029Spjd 1653268858Sdelphij/* 1654268858Sdelphij * Free up buf->b_data and if 'remove' is set, then pull the 1655268858Sdelphij * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 1656268858Sdelphij */ 1657185029Spjdstatic void 1658268858Sdelphijarc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 1659168404Spjd{ 1660168404Spjd arc_buf_t **bufp; 1661168404Spjd 1662168404Spjd /* free up data associated with the buf */ 1663168404Spjd if (buf->b_data) { 1664168404Spjd arc_state_t *state = buf->b_hdr->b_state; 1665168404Spjd uint64_t size = buf->b_hdr->b_size; 1666168404Spjd arc_buf_contents_t type = buf->b_hdr->b_type; 1667168404Spjd 1668168404Spjd arc_cksum_verify(buf); 1669240133Smm#ifdef illumos 1670240133Smm arc_buf_unwatch(buf); 1671240133Smm#endif /* illumos */ 1672219089Spjd 1673168404Spjd if (!recycle) { 1674168404Spjd if (type == ARC_BUFC_METADATA) { 1675240133Smm arc_buf_data_free(buf, zio_buf_free); 1676208373Smm arc_space_return(size, ARC_SPACE_DATA); 1677168404Spjd } else { 1678168404Spjd ASSERT(type == ARC_BUFC_DATA); 1679240133Smm arc_buf_data_free(buf, zio_data_buf_free); 1680208373Smm ARCSTAT_INCR(arcstat_data_size, -size); 1681185029Spjd atomic_add_64(&arc_size, -size); 1682168404Spjd } 1683168404Spjd } 1684168404Spjd if (list_link_active(&buf->b_hdr->b_arc_node)) { 1685185029Spjd uint64_t *cnt = &state->arcs_lsize[type]; 1686185029Spjd 1687168404Spjd ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1688168404Spjd ASSERT(state != arc_anon); 1689185029Spjd 1690185029Spjd ASSERT3U(*cnt, >=, size); 1691185029Spjd atomic_add_64(cnt, -size); 1692168404Spjd } 1693168404Spjd ASSERT3U(state->arcs_size, >=, size); 1694168404Spjd atomic_add_64(&state->arcs_size, -size); 1695168404Spjd buf->b_data = NULL; 1696242845Sdelphij 1697242845Sdelphij /* 1698242845Sdelphij * If we're destroying a duplicate buffer make sure 1699242845Sdelphij * that the appropriate statistics are updated. 1700242845Sdelphij */ 1701242845Sdelphij if (buf->b_hdr->b_datacnt > 1 && 1702242845Sdelphij buf->b_hdr->b_type == ARC_BUFC_DATA) { 1703242845Sdelphij ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1704242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1705242845Sdelphij } 1706168404Spjd ASSERT(buf->b_hdr->b_datacnt > 0); 1707168404Spjd buf->b_hdr->b_datacnt -= 1; 1708168404Spjd } 1709168404Spjd 1710168404Spjd /* only remove the buf if requested */ 1711268858Sdelphij if (!remove) 1712168404Spjd return; 1713168404Spjd 1714168404Spjd /* remove the buf from the hdr list */ 1715168404Spjd for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1716168404Spjd continue; 1717168404Spjd *bufp = buf->b_next; 1718219089Spjd buf->b_next = NULL; 1719168404Spjd 1720168404Spjd ASSERT(buf->b_efunc == NULL); 1721168404Spjd 1722168404Spjd /* clean up the buf */ 1723168404Spjd buf->b_hdr = NULL; 1724168404Spjd kmem_cache_free(buf_cache, buf); 1725168404Spjd} 1726168404Spjd 1727168404Spjdstatic void 1728168404Spjdarc_hdr_destroy(arc_buf_hdr_t *hdr) 1729168404Spjd{ 1730168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1731168404Spjd ASSERT3P(hdr->b_state, ==, arc_anon); 1732168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1733219089Spjd l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1734168404Spjd 1735219089Spjd if (l2hdr != NULL) { 1736219089Spjd boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1737219089Spjd /* 1738219089Spjd * To prevent arc_free() and l2arc_evict() from 1739219089Spjd * attempting to free the same buffer at the same time, 1740219089Spjd * a FREE_IN_PROGRESS flag is given to arc_free() to 1741219089Spjd * give it priority. l2arc_evict() can't destroy this 1742219089Spjd * header while we are waiting on l2arc_buflist_mtx. 1743219089Spjd * 1744219089Spjd * The hdr may be removed from l2ad_buflist before we 1745219089Spjd * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1746219089Spjd */ 1747219089Spjd if (!buflist_held) { 1748185029Spjd mutex_enter(&l2arc_buflist_mtx); 1749219089Spjd l2hdr = hdr->b_l2hdr; 1750219089Spjd } 1751219089Spjd 1752219089Spjd if (l2hdr != NULL) { 1753248572Ssmh trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1754248574Ssmh hdr->b_size, 0); 1755219089Spjd list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1756219089Spjd ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1757251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1758268085Sdelphij vdev_space_update(l2hdr->b_dev->l2ad_vdev, 1759268085Sdelphij -l2hdr->b_asize, 0, 0); 1760219089Spjd kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1761219089Spjd if (hdr->b_state == arc_l2c_only) 1762219089Spjd l2arc_hdr_stat_remove(); 1763219089Spjd hdr->b_l2hdr = NULL; 1764219089Spjd } 1765219089Spjd 1766219089Spjd if (!buflist_held) 1767185029Spjd mutex_exit(&l2arc_buflist_mtx); 1768185029Spjd } 1769185029Spjd 1770168404Spjd if (!BUF_EMPTY(hdr)) { 1771168404Spjd ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1772219089Spjd buf_discard_identity(hdr); 1773168404Spjd } 1774168404Spjd while (hdr->b_buf) { 1775168404Spjd arc_buf_t *buf = hdr->b_buf; 1776168404Spjd 1777168404Spjd if (buf->b_efunc) { 1778168404Spjd mutex_enter(&arc_eviction_mtx); 1779219089Spjd mutex_enter(&buf->b_evict_lock); 1780168404Spjd ASSERT(buf->b_hdr != NULL); 1781168404Spjd arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1782168404Spjd hdr->b_buf = buf->b_next; 1783168404Spjd buf->b_hdr = &arc_eviction_hdr; 1784168404Spjd buf->b_next = arc_eviction_list; 1785168404Spjd arc_eviction_list = buf; 1786219089Spjd mutex_exit(&buf->b_evict_lock); 1787168404Spjd mutex_exit(&arc_eviction_mtx); 1788168404Spjd } else { 1789168404Spjd arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1790168404Spjd } 1791168404Spjd } 1792168404Spjd if (hdr->b_freeze_cksum != NULL) { 1793168404Spjd kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1794168404Spjd hdr->b_freeze_cksum = NULL; 1795168404Spjd } 1796219089Spjd if (hdr->b_thawed) { 1797219089Spjd kmem_free(hdr->b_thawed, 1); 1798219089Spjd hdr->b_thawed = NULL; 1799219089Spjd } 1800168404Spjd 1801168404Spjd ASSERT(!list_link_active(&hdr->b_arc_node)); 1802168404Spjd ASSERT3P(hdr->b_hash_next, ==, NULL); 1803168404Spjd ASSERT3P(hdr->b_acb, ==, NULL); 1804168404Spjd kmem_cache_free(hdr_cache, hdr); 1805168404Spjd} 1806168404Spjd 1807168404Spjdvoid 1808168404Spjdarc_buf_free(arc_buf_t *buf, void *tag) 1809168404Spjd{ 1810168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 1811168404Spjd int hashed = hdr->b_state != arc_anon; 1812168404Spjd 1813168404Spjd ASSERT(buf->b_efunc == NULL); 1814168404Spjd ASSERT(buf->b_data != NULL); 1815168404Spjd 1816168404Spjd if (hashed) { 1817168404Spjd kmutex_t *hash_lock = HDR_LOCK(hdr); 1818168404Spjd 1819168404Spjd mutex_enter(hash_lock); 1820219089Spjd hdr = buf->b_hdr; 1821219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1822219089Spjd 1823168404Spjd (void) remove_reference(hdr, hash_lock, tag); 1824219089Spjd if (hdr->b_datacnt > 1) { 1825168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1826219089Spjd } else { 1827219089Spjd ASSERT(buf == hdr->b_buf); 1828219089Spjd ASSERT(buf->b_efunc == NULL); 1829168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 1830219089Spjd } 1831168404Spjd mutex_exit(hash_lock); 1832168404Spjd } else if (HDR_IO_IN_PROGRESS(hdr)) { 1833168404Spjd int destroy_hdr; 1834168404Spjd /* 1835168404Spjd * We are in the middle of an async write. Don't destroy 1836168404Spjd * this buffer unless the write completes before we finish 1837168404Spjd * decrementing the reference count. 1838168404Spjd */ 1839168404Spjd mutex_enter(&arc_eviction_mtx); 1840168404Spjd (void) remove_reference(hdr, NULL, tag); 1841168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1842168404Spjd destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1843168404Spjd mutex_exit(&arc_eviction_mtx); 1844168404Spjd if (destroy_hdr) 1845168404Spjd arc_hdr_destroy(hdr); 1846168404Spjd } else { 1847219089Spjd if (remove_reference(hdr, NULL, tag) > 0) 1848168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1849219089Spjd else 1850168404Spjd arc_hdr_destroy(hdr); 1851168404Spjd } 1852168404Spjd} 1853168404Spjd 1854248571Smmboolean_t 1855168404Spjdarc_buf_remove_ref(arc_buf_t *buf, void* tag) 1856168404Spjd{ 1857168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 1858168404Spjd kmutex_t *hash_lock = HDR_LOCK(hdr); 1859248571Smm boolean_t no_callback = (buf->b_efunc == NULL); 1860168404Spjd 1861168404Spjd if (hdr->b_state == arc_anon) { 1862219089Spjd ASSERT(hdr->b_datacnt == 1); 1863168404Spjd arc_buf_free(buf, tag); 1864168404Spjd return (no_callback); 1865168404Spjd } 1866168404Spjd 1867168404Spjd mutex_enter(hash_lock); 1868219089Spjd hdr = buf->b_hdr; 1869219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1870168404Spjd ASSERT(hdr->b_state != arc_anon); 1871168404Spjd ASSERT(buf->b_data != NULL); 1872168404Spjd 1873168404Spjd (void) remove_reference(hdr, hash_lock, tag); 1874168404Spjd if (hdr->b_datacnt > 1) { 1875168404Spjd if (no_callback) 1876168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1877168404Spjd } else if (no_callback) { 1878168404Spjd ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1879219089Spjd ASSERT(buf->b_efunc == NULL); 1880168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 1881168404Spjd } 1882168404Spjd ASSERT(no_callback || hdr->b_datacnt > 1 || 1883168404Spjd refcount_is_zero(&hdr->b_refcnt)); 1884168404Spjd mutex_exit(hash_lock); 1885168404Spjd return (no_callback); 1886168404Spjd} 1887168404Spjd 1888168404Spjdint 1889168404Spjdarc_buf_size(arc_buf_t *buf) 1890168404Spjd{ 1891168404Spjd return (buf->b_hdr->b_size); 1892168404Spjd} 1893168404Spjd 1894168404Spjd/* 1895242845Sdelphij * Called from the DMU to determine if the current buffer should be 1896242845Sdelphij * evicted. In order to ensure proper locking, the eviction must be initiated 1897242845Sdelphij * from the DMU. Return true if the buffer is associated with user data and 1898242845Sdelphij * duplicate buffers still exist. 1899242845Sdelphij */ 1900242845Sdelphijboolean_t 1901242845Sdelphijarc_buf_eviction_needed(arc_buf_t *buf) 1902242845Sdelphij{ 1903242845Sdelphij arc_buf_hdr_t *hdr; 1904242845Sdelphij boolean_t evict_needed = B_FALSE; 1905242845Sdelphij 1906242845Sdelphij if (zfs_disable_dup_eviction) 1907242845Sdelphij return (B_FALSE); 1908242845Sdelphij 1909242845Sdelphij mutex_enter(&buf->b_evict_lock); 1910242845Sdelphij hdr = buf->b_hdr; 1911242845Sdelphij if (hdr == NULL) { 1912242845Sdelphij /* 1913242845Sdelphij * We are in arc_do_user_evicts(); let that function 1914242845Sdelphij * perform the eviction. 1915242845Sdelphij */ 1916242845Sdelphij ASSERT(buf->b_data == NULL); 1917242845Sdelphij mutex_exit(&buf->b_evict_lock); 1918242845Sdelphij return (B_FALSE); 1919242845Sdelphij } else if (buf->b_data == NULL) { 1920242845Sdelphij /* 1921242845Sdelphij * We have already been added to the arc eviction list; 1922242845Sdelphij * recommend eviction. 1923242845Sdelphij */ 1924242845Sdelphij ASSERT3P(hdr, ==, &arc_eviction_hdr); 1925242845Sdelphij mutex_exit(&buf->b_evict_lock); 1926242845Sdelphij return (B_TRUE); 1927242845Sdelphij } 1928242845Sdelphij 1929242845Sdelphij if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 1930242845Sdelphij evict_needed = B_TRUE; 1931242845Sdelphij 1932242845Sdelphij mutex_exit(&buf->b_evict_lock); 1933242845Sdelphij return (evict_needed); 1934242845Sdelphij} 1935242845Sdelphij 1936242845Sdelphij/* 1937168404Spjd * Evict buffers from list until we've removed the specified number of 1938168404Spjd * bytes. Move the removed buffers to the appropriate evict state. 1939168404Spjd * If the recycle flag is set, then attempt to "recycle" a buffer: 1940168404Spjd * - look for a buffer to evict that is `bytes' long. 1941168404Spjd * - return the data block from this buffer rather than freeing it. 1942168404Spjd * This flag is used by callers that are trying to make space for a 1943168404Spjd * new buffer in a full arc cache. 1944185029Spjd * 1945185029Spjd * This function makes a "best effort". It skips over any buffers 1946185029Spjd * it can't get a hash_lock on, and so may not catch all candidates. 1947185029Spjd * It may also return without evicting as much space as requested. 1948168404Spjd */ 1949168404Spjdstatic void * 1950209962Smmarc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 1951168404Spjd arc_buf_contents_t type) 1952168404Spjd{ 1953168404Spjd arc_state_t *evicted_state; 1954168404Spjd uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 1955205231Skmacy int64_t bytes_remaining; 1956168404Spjd arc_buf_hdr_t *ab, *ab_prev = NULL; 1957205231Skmacy list_t *evicted_list, *list, *evicted_list_start, *list_start; 1958205231Skmacy kmutex_t *lock, *evicted_lock; 1959168404Spjd kmutex_t *hash_lock; 1960168404Spjd boolean_t have_lock; 1961168404Spjd void *stolen = NULL; 1962258632Savg arc_buf_hdr_t marker = { 0 }; 1963258632Savg int count = 0; 1964205231Skmacy static int evict_metadata_offset, evict_data_offset; 1965258632Savg int i, idx, offset, list_count, lists; 1966168404Spjd 1967168404Spjd ASSERT(state == arc_mru || state == arc_mfu); 1968168404Spjd 1969168404Spjd evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 1970206796Spjd 1971205231Skmacy if (type == ARC_BUFC_METADATA) { 1972205231Skmacy offset = 0; 1973205231Skmacy list_count = ARC_BUFC_NUMMETADATALISTS; 1974205231Skmacy list_start = &state->arcs_lists[0]; 1975205231Skmacy evicted_list_start = &evicted_state->arcs_lists[0]; 1976205231Skmacy idx = evict_metadata_offset; 1977205231Skmacy } else { 1978205231Skmacy offset = ARC_BUFC_NUMMETADATALISTS; 1979205231Skmacy list_start = &state->arcs_lists[offset]; 1980205231Skmacy evicted_list_start = &evicted_state->arcs_lists[offset]; 1981205231Skmacy list_count = ARC_BUFC_NUMDATALISTS; 1982205231Skmacy idx = evict_data_offset; 1983205231Skmacy } 1984205231Skmacy bytes_remaining = evicted_state->arcs_lsize[type]; 1985258632Savg lists = 0; 1986206796Spjd 1987205231Skmacyevict_start: 1988205231Skmacy list = &list_start[idx]; 1989205231Skmacy evicted_list = &evicted_list_start[idx]; 1990205231Skmacy lock = ARCS_LOCK(state, (offset + idx)); 1991206796Spjd evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 1992168404Spjd 1993205231Skmacy mutex_enter(lock); 1994205231Skmacy mutex_enter(evicted_lock); 1995205231Skmacy 1996185029Spjd for (ab = list_tail(list); ab; ab = ab_prev) { 1997185029Spjd ab_prev = list_prev(list, ab); 1998205231Skmacy bytes_remaining -= (ab->b_size * ab->b_datacnt); 1999168404Spjd /* prefetch buffers have a minimum lifespan */ 2000168404Spjd if (HDR_IO_IN_PROGRESS(ab) || 2001185029Spjd (spa && ab->b_spa != spa) || 2002168404Spjd (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 2003219089Spjd ddi_get_lbolt() - ab->b_arc_access < 2004219089Spjd arc_min_prefetch_lifespan)) { 2005168404Spjd skipped++; 2006168404Spjd continue; 2007168404Spjd } 2008168404Spjd /* "lookahead" for better eviction candidate */ 2009168404Spjd if (recycle && ab->b_size != bytes && 2010168404Spjd ab_prev && ab_prev->b_size == bytes) 2011168404Spjd continue; 2012258632Savg 2013258632Savg /* ignore markers */ 2014258632Savg if (ab->b_spa == 0) 2015258632Savg continue; 2016258632Savg 2017258632Savg /* 2018258632Savg * It may take a long time to evict all the bufs requested. 2019258632Savg * To avoid blocking all arc activity, periodically drop 2020258632Savg * the arcs_mtx and give other threads a chance to run 2021258632Savg * before reacquiring the lock. 2022258632Savg * 2023258632Savg * If we are looking for a buffer to recycle, we are in 2024258632Savg * the hot code path, so don't sleep. 2025258632Savg */ 2026258632Savg if (!recycle && count++ > arc_evict_iterations) { 2027258632Savg list_insert_after(list, ab, &marker); 2028258632Savg mutex_exit(evicted_lock); 2029258632Savg mutex_exit(lock); 2030258632Savg kpreempt(KPREEMPT_SYNC); 2031258632Savg mutex_enter(lock); 2032258632Savg mutex_enter(evicted_lock); 2033258632Savg ab_prev = list_prev(list, &marker); 2034258632Savg list_remove(list, &marker); 2035258632Savg count = 0; 2036258632Savg continue; 2037258632Savg } 2038258632Savg 2039168404Spjd hash_lock = HDR_LOCK(ab); 2040168404Spjd have_lock = MUTEX_HELD(hash_lock); 2041168404Spjd if (have_lock || mutex_tryenter(hash_lock)) { 2042240415Smm ASSERT0(refcount_count(&ab->b_refcnt)); 2043168404Spjd ASSERT(ab->b_datacnt > 0); 2044168404Spjd while (ab->b_buf) { 2045168404Spjd arc_buf_t *buf = ab->b_buf; 2046219089Spjd if (!mutex_tryenter(&buf->b_evict_lock)) { 2047185029Spjd missed += 1; 2048185029Spjd break; 2049185029Spjd } 2050168404Spjd if (buf->b_data) { 2051168404Spjd bytes_evicted += ab->b_size; 2052168404Spjd if (recycle && ab->b_type == type && 2053185029Spjd ab->b_size == bytes && 2054185029Spjd !HDR_L2_WRITING(ab)) { 2055168404Spjd stolen = buf->b_data; 2056168404Spjd recycle = FALSE; 2057168404Spjd } 2058168404Spjd } 2059168404Spjd if (buf->b_efunc) { 2060168404Spjd mutex_enter(&arc_eviction_mtx); 2061168404Spjd arc_buf_destroy(buf, 2062168404Spjd buf->b_data == stolen, FALSE); 2063168404Spjd ab->b_buf = buf->b_next; 2064168404Spjd buf->b_hdr = &arc_eviction_hdr; 2065168404Spjd buf->b_next = arc_eviction_list; 2066168404Spjd arc_eviction_list = buf; 2067168404Spjd mutex_exit(&arc_eviction_mtx); 2068219089Spjd mutex_exit(&buf->b_evict_lock); 2069168404Spjd } else { 2070219089Spjd mutex_exit(&buf->b_evict_lock); 2071168404Spjd arc_buf_destroy(buf, 2072168404Spjd buf->b_data == stolen, TRUE); 2073168404Spjd } 2074168404Spjd } 2075208373Smm 2076208373Smm if (ab->b_l2hdr) { 2077208373Smm ARCSTAT_INCR(arcstat_evict_l2_cached, 2078208373Smm ab->b_size); 2079208373Smm } else { 2080208373Smm if (l2arc_write_eligible(ab->b_spa, ab)) { 2081208373Smm ARCSTAT_INCR(arcstat_evict_l2_eligible, 2082208373Smm ab->b_size); 2083208373Smm } else { 2084208373Smm ARCSTAT_INCR( 2085208373Smm arcstat_evict_l2_ineligible, 2086208373Smm ab->b_size); 2087208373Smm } 2088208373Smm } 2089208373Smm 2090185029Spjd if (ab->b_datacnt == 0) { 2091185029Spjd arc_change_state(evicted_state, ab, hash_lock); 2092185029Spjd ASSERT(HDR_IN_HASH_TABLE(ab)); 2093185029Spjd ab->b_flags |= ARC_IN_HASH_TABLE; 2094185029Spjd ab->b_flags &= ~ARC_BUF_AVAILABLE; 2095185029Spjd DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 2096185029Spjd } 2097168404Spjd if (!have_lock) 2098168404Spjd mutex_exit(hash_lock); 2099168404Spjd if (bytes >= 0 && bytes_evicted >= bytes) 2100168404Spjd break; 2101205231Skmacy if (bytes_remaining > 0) { 2102205231Skmacy mutex_exit(evicted_lock); 2103205231Skmacy mutex_exit(lock); 2104206796Spjd idx = ((idx + 1) & (list_count - 1)); 2105258632Savg lists++; 2106205231Skmacy goto evict_start; 2107205231Skmacy } 2108168404Spjd } else { 2109168404Spjd missed += 1; 2110168404Spjd } 2111168404Spjd } 2112168404Spjd 2113205231Skmacy mutex_exit(evicted_lock); 2114205231Skmacy mutex_exit(lock); 2115206796Spjd 2116206796Spjd idx = ((idx + 1) & (list_count - 1)); 2117258632Savg lists++; 2118168404Spjd 2119205231Skmacy if (bytes_evicted < bytes) { 2120258632Savg if (lists < list_count) 2121205231Skmacy goto evict_start; 2122205231Skmacy else 2123205231Skmacy dprintf("only evicted %lld bytes from %x", 2124205231Skmacy (longlong_t)bytes_evicted, state); 2125205231Skmacy } 2126206796Spjd if (type == ARC_BUFC_METADATA) 2127205231Skmacy evict_metadata_offset = idx; 2128205231Skmacy else 2129205231Skmacy evict_data_offset = idx; 2130206796Spjd 2131168404Spjd if (skipped) 2132168404Spjd ARCSTAT_INCR(arcstat_evict_skip, skipped); 2133168404Spjd 2134168404Spjd if (missed) 2135168404Spjd ARCSTAT_INCR(arcstat_mutex_miss, missed); 2136168404Spjd 2137185029Spjd /* 2138258632Savg * Note: we have just evicted some data into the ghost state, 2139258632Savg * potentially putting the ghost size over the desired size. Rather 2140258632Savg * that evicting from the ghost list in this hot code path, leave 2141258632Savg * this chore to the arc_reclaim_thread(). 2142185029Spjd */ 2143185029Spjd 2144205231Skmacy if (stolen) 2145205231Skmacy ARCSTAT_BUMP(arcstat_stolen); 2146168404Spjd return (stolen); 2147168404Spjd} 2148168404Spjd 2149168404Spjd/* 2150168404Spjd * Remove buffers from list until we've removed the specified number of 2151168404Spjd * bytes. Destroy the buffers that are removed. 2152168404Spjd */ 2153168404Spjdstatic void 2154209962Smmarc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2155168404Spjd{ 2156168404Spjd arc_buf_hdr_t *ab, *ab_prev; 2157219089Spjd arc_buf_hdr_t marker = { 0 }; 2158205231Skmacy list_t *list, *list_start; 2159205231Skmacy kmutex_t *hash_lock, *lock; 2160168404Spjd uint64_t bytes_deleted = 0; 2161168404Spjd uint64_t bufs_skipped = 0; 2162258632Savg int count = 0; 2163205231Skmacy static int evict_offset; 2164205231Skmacy int list_count, idx = evict_offset; 2165258632Savg int offset, lists = 0; 2166168404Spjd 2167168404Spjd ASSERT(GHOST_STATE(state)); 2168205231Skmacy 2169205231Skmacy /* 2170205231Skmacy * data lists come after metadata lists 2171205231Skmacy */ 2172205231Skmacy list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2173205231Skmacy list_count = ARC_BUFC_NUMDATALISTS; 2174205231Skmacy offset = ARC_BUFC_NUMMETADATALISTS; 2175206796Spjd 2176205231Skmacyevict_start: 2177205231Skmacy list = &list_start[idx]; 2178205231Skmacy lock = ARCS_LOCK(state, idx + offset); 2179205231Skmacy 2180205231Skmacy mutex_enter(lock); 2181185029Spjd for (ab = list_tail(list); ab; ab = ab_prev) { 2182185029Spjd ab_prev = list_prev(list, ab); 2183258632Savg if (ab->b_type > ARC_BUFC_NUMTYPES) 2184258632Savg panic("invalid ab=%p", (void *)ab); 2185185029Spjd if (spa && ab->b_spa != spa) 2186185029Spjd continue; 2187219089Spjd 2188219089Spjd /* ignore markers */ 2189219089Spjd if (ab->b_spa == 0) 2190219089Spjd continue; 2191219089Spjd 2192168404Spjd hash_lock = HDR_LOCK(ab); 2193219089Spjd /* caller may be trying to modify this buffer, skip it */ 2194219089Spjd if (MUTEX_HELD(hash_lock)) 2195219089Spjd continue; 2196258632Savg 2197258632Savg /* 2198258632Savg * It may take a long time to evict all the bufs requested. 2199258632Savg * To avoid blocking all arc activity, periodically drop 2200258632Savg * the arcs_mtx and give other threads a chance to run 2201258632Savg * before reacquiring the lock. 2202258632Savg */ 2203258632Savg if (count++ > arc_evict_iterations) { 2204258632Savg list_insert_after(list, ab, &marker); 2205258632Savg mutex_exit(lock); 2206258632Savg kpreempt(KPREEMPT_SYNC); 2207258632Savg mutex_enter(lock); 2208258632Savg ab_prev = list_prev(list, &marker); 2209258632Savg list_remove(list, &marker); 2210258632Savg count = 0; 2211258632Savg continue; 2212258632Savg } 2213168404Spjd if (mutex_tryenter(hash_lock)) { 2214168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(ab)); 2215168404Spjd ASSERT(ab->b_buf == NULL); 2216168404Spjd ARCSTAT_BUMP(arcstat_deleted); 2217168404Spjd bytes_deleted += ab->b_size; 2218185029Spjd 2219185029Spjd if (ab->b_l2hdr != NULL) { 2220185029Spjd /* 2221185029Spjd * This buffer is cached on the 2nd Level ARC; 2222185029Spjd * don't destroy the header. 2223185029Spjd */ 2224185029Spjd arc_change_state(arc_l2c_only, ab, hash_lock); 2225185029Spjd mutex_exit(hash_lock); 2226185029Spjd } else { 2227185029Spjd arc_change_state(arc_anon, ab, hash_lock); 2228185029Spjd mutex_exit(hash_lock); 2229185029Spjd arc_hdr_destroy(ab); 2230185029Spjd } 2231185029Spjd 2232168404Spjd DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 2233168404Spjd if (bytes >= 0 && bytes_deleted >= bytes) 2234168404Spjd break; 2235219089Spjd } else if (bytes < 0) { 2236219089Spjd /* 2237219089Spjd * Insert a list marker and then wait for the 2238219089Spjd * hash lock to become available. Once its 2239219089Spjd * available, restart from where we left off. 2240219089Spjd */ 2241219089Spjd list_insert_after(list, ab, &marker); 2242219089Spjd mutex_exit(lock); 2243219089Spjd mutex_enter(hash_lock); 2244219089Spjd mutex_exit(hash_lock); 2245219089Spjd mutex_enter(lock); 2246219089Spjd ab_prev = list_prev(list, &marker); 2247219089Spjd list_remove(list, &marker); 2248258632Savg } else { 2249168404Spjd bufs_skipped += 1; 2250258632Savg } 2251258632Savg 2252168404Spjd } 2253205231Skmacy mutex_exit(lock); 2254206796Spjd idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2255258632Savg lists++; 2256206796Spjd 2257258632Savg if (lists < list_count) 2258205231Skmacy goto evict_start; 2259206796Spjd 2260205231Skmacy evict_offset = idx; 2261205231Skmacy if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2262185029Spjd (bytes < 0 || bytes_deleted < bytes)) { 2263205231Skmacy list_start = &state->arcs_lists[0]; 2264205231Skmacy list_count = ARC_BUFC_NUMMETADATALISTS; 2265258632Savg offset = lists = 0; 2266205231Skmacy goto evict_start; 2267185029Spjd } 2268185029Spjd 2269168404Spjd if (bufs_skipped) { 2270168404Spjd ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2271168404Spjd ASSERT(bytes >= 0); 2272168404Spjd } 2273168404Spjd 2274168404Spjd if (bytes_deleted < bytes) 2275168404Spjd dprintf("only deleted %lld bytes from %p", 2276168404Spjd (longlong_t)bytes_deleted, state); 2277168404Spjd} 2278168404Spjd 2279168404Spjdstatic void 2280168404Spjdarc_adjust(void) 2281168404Spjd{ 2282208373Smm int64_t adjustment, delta; 2283168404Spjd 2284208373Smm /* 2285208373Smm * Adjust MRU size 2286208373Smm */ 2287168404Spjd 2288209275Smm adjustment = MIN((int64_t)(arc_size - arc_c), 2289209275Smm (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2290209275Smm arc_p)); 2291208373Smm 2292208373Smm if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2293208373Smm delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2294209962Smm (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2295208373Smm adjustment -= delta; 2296168404Spjd } 2297168404Spjd 2298208373Smm if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2299208373Smm delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2300209962Smm (void) arc_evict(arc_mru, 0, delta, FALSE, 2301185029Spjd ARC_BUFC_METADATA); 2302185029Spjd } 2303185029Spjd 2304208373Smm /* 2305208373Smm * Adjust MFU size 2306208373Smm */ 2307168404Spjd 2308208373Smm adjustment = arc_size - arc_c; 2309208373Smm 2310208373Smm if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2311208373Smm delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2312209962Smm (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2313208373Smm adjustment -= delta; 2314168404Spjd } 2315168404Spjd 2316208373Smm if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2317208373Smm int64_t delta = MIN(adjustment, 2318208373Smm arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2319209962Smm (void) arc_evict(arc_mfu, 0, delta, FALSE, 2320208373Smm ARC_BUFC_METADATA); 2321208373Smm } 2322168404Spjd 2323208373Smm /* 2324208373Smm * Adjust ghost lists 2325208373Smm */ 2326168404Spjd 2327208373Smm adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2328168404Spjd 2329208373Smm if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2330208373Smm delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2331209962Smm arc_evict_ghost(arc_mru_ghost, 0, delta); 2332208373Smm } 2333185029Spjd 2334208373Smm adjustment = 2335208373Smm arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2336208373Smm 2337208373Smm if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2338208373Smm delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2339209962Smm arc_evict_ghost(arc_mfu_ghost, 0, delta); 2340168404Spjd } 2341168404Spjd} 2342168404Spjd 2343168404Spjdstatic void 2344168404Spjdarc_do_user_evicts(void) 2345168404Spjd{ 2346191903Skmacy static arc_buf_t *tmp_arc_eviction_list; 2347191903Skmacy 2348191903Skmacy /* 2349191903Skmacy * Move list over to avoid LOR 2350191903Skmacy */ 2351206796Spjdrestart: 2352168404Spjd mutex_enter(&arc_eviction_mtx); 2353191903Skmacy tmp_arc_eviction_list = arc_eviction_list; 2354191903Skmacy arc_eviction_list = NULL; 2355191903Skmacy mutex_exit(&arc_eviction_mtx); 2356191903Skmacy 2357191903Skmacy while (tmp_arc_eviction_list != NULL) { 2358191903Skmacy arc_buf_t *buf = tmp_arc_eviction_list; 2359191903Skmacy tmp_arc_eviction_list = buf->b_next; 2360219089Spjd mutex_enter(&buf->b_evict_lock); 2361168404Spjd buf->b_hdr = NULL; 2362219089Spjd mutex_exit(&buf->b_evict_lock); 2363168404Spjd 2364168404Spjd if (buf->b_efunc != NULL) 2365268858Sdelphij VERIFY0(buf->b_efunc(buf->b_private)); 2366168404Spjd 2367168404Spjd buf->b_efunc = NULL; 2368168404Spjd buf->b_private = NULL; 2369168404Spjd kmem_cache_free(buf_cache, buf); 2370168404Spjd } 2371191903Skmacy 2372191903Skmacy if (arc_eviction_list != NULL) 2373191903Skmacy goto restart; 2374168404Spjd} 2375168404Spjd 2376168404Spjd/* 2377185029Spjd * Flush all *evictable* data from the cache for the given spa. 2378168404Spjd * NOTE: this will not touch "active" (i.e. referenced) data. 2379168404Spjd */ 2380168404Spjdvoid 2381185029Spjdarc_flush(spa_t *spa) 2382168404Spjd{ 2383209962Smm uint64_t guid = 0; 2384209962Smm 2385209962Smm if (spa) 2386228103Smm guid = spa_load_guid(spa); 2387209962Smm 2388205231Skmacy while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2389209962Smm (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2390185029Spjd if (spa) 2391185029Spjd break; 2392185029Spjd } 2393205231Skmacy while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2394209962Smm (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2395185029Spjd if (spa) 2396185029Spjd break; 2397185029Spjd } 2398205231Skmacy while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2399209962Smm (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2400185029Spjd if (spa) 2401185029Spjd break; 2402185029Spjd } 2403205231Skmacy while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2404209962Smm (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2405185029Spjd if (spa) 2406185029Spjd break; 2407185029Spjd } 2408168404Spjd 2409209962Smm arc_evict_ghost(arc_mru_ghost, guid, -1); 2410209962Smm arc_evict_ghost(arc_mfu_ghost, guid, -1); 2411168404Spjd 2412168404Spjd mutex_enter(&arc_reclaim_thr_lock); 2413168404Spjd arc_do_user_evicts(); 2414168404Spjd mutex_exit(&arc_reclaim_thr_lock); 2415185029Spjd ASSERT(spa || arc_eviction_list == NULL); 2416168404Spjd} 2417168404Spjd 2418168404Spjdvoid 2419168404Spjdarc_shrink(void) 2420168404Spjd{ 2421168404Spjd if (arc_c > arc_c_min) { 2422168404Spjd uint64_t to_free; 2423168404Spjd 2424168404Spjd#ifdef _KERNEL 2425168404Spjd to_free = arc_c >> arc_shrink_shift; 2426168404Spjd#else 2427168404Spjd to_free = arc_c >> arc_shrink_shift; 2428168404Spjd#endif 2429168404Spjd if (arc_c > arc_c_min + to_free) 2430168404Spjd atomic_add_64(&arc_c, -to_free); 2431168404Spjd else 2432168404Spjd arc_c = arc_c_min; 2433168404Spjd 2434168404Spjd atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2435168404Spjd if (arc_c > arc_size) 2436168404Spjd arc_c = MAX(arc_size, arc_c_min); 2437168404Spjd if (arc_p > arc_c) 2438168404Spjd arc_p = (arc_c >> 1); 2439168404Spjd ASSERT(arc_c >= arc_c_min); 2440168404Spjd ASSERT((int64_t)arc_p >= 0); 2441168404Spjd } 2442168404Spjd 2443168404Spjd if (arc_size > arc_c) 2444168404Spjd arc_adjust(); 2445168404Spjd} 2446168404Spjd 2447185029Spjdstatic int needfree = 0; 2448168404Spjd 2449168404Spjdstatic int 2450168404Spjdarc_reclaim_needed(void) 2451168404Spjd{ 2452168404Spjd 2453168404Spjd#ifdef _KERNEL 2454219089Spjd 2455197816Skmacy if (needfree) 2456197816Skmacy return (1); 2457168404Spjd 2458191902Skmacy /* 2459212780Savg * Cooperate with pagedaemon when it's time for it to scan 2460212780Savg * and reclaim some pages. 2461191902Skmacy */ 2462212783Savg if (vm_paging_needed()) 2463191902Skmacy return (1); 2464191902Skmacy 2465219089Spjd#ifdef sun 2466168404Spjd /* 2467185029Spjd * take 'desfree' extra pages, so we reclaim sooner, rather than later 2468185029Spjd */ 2469185029Spjd extra = desfree; 2470185029Spjd 2471185029Spjd /* 2472185029Spjd * check that we're out of range of the pageout scanner. It starts to 2473185029Spjd * schedule paging if freemem is less than lotsfree and needfree. 2474185029Spjd * lotsfree is the high-water mark for pageout, and needfree is the 2475185029Spjd * number of needed free pages. We add extra pages here to make sure 2476185029Spjd * the scanner doesn't start up while we're freeing memory. 2477185029Spjd */ 2478185029Spjd if (freemem < lotsfree + needfree + extra) 2479185029Spjd return (1); 2480185029Spjd 2481185029Spjd /* 2482168404Spjd * check to make sure that swapfs has enough space so that anon 2483185029Spjd * reservations can still succeed. anon_resvmem() checks that the 2484168404Spjd * availrmem is greater than swapfs_minfree, and the number of reserved 2485168404Spjd * swap pages. We also add a bit of extra here just to prevent 2486168404Spjd * circumstances from getting really dire. 2487168404Spjd */ 2488168404Spjd if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2489168404Spjd return (1); 2490168404Spjd 2491168404Spjd#if defined(__i386) 2492168404Spjd /* 2493168404Spjd * If we're on an i386 platform, it's possible that we'll exhaust the 2494168404Spjd * kernel heap space before we ever run out of available physical 2495168404Spjd * memory. Most checks of the size of the heap_area compare against 2496168404Spjd * tune.t_minarmem, which is the minimum available real memory that we 2497168404Spjd * can have in the system. However, this is generally fixed at 25 pages 2498168404Spjd * which is so low that it's useless. In this comparison, we seek to 2499168404Spjd * calculate the total heap-size, and reclaim if more than 3/4ths of the 2500185029Spjd * heap is allocated. (Or, in the calculation, if less than 1/4th is 2501168404Spjd * free) 2502168404Spjd */ 2503168404Spjd if (btop(vmem_size(heap_arena, VMEM_FREE)) < 2504168404Spjd (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2)) 2505168404Spjd return (1); 2506168404Spjd#endif 2507219089Spjd#else /* !sun */ 2508175633Spjd if (kmem_used() > (kmem_size() * 3) / 4) 2509168404Spjd return (1); 2510219089Spjd#endif /* sun */ 2511168404Spjd 2512168404Spjd#else 2513168404Spjd if (spa_get_random(100) == 0) 2514168404Spjd return (1); 2515168404Spjd#endif 2516168404Spjd return (0); 2517168404Spjd} 2518168404Spjd 2519208454Spjdextern kmem_cache_t *zio_buf_cache[]; 2520208454Spjdextern kmem_cache_t *zio_data_buf_cache[]; 2521208454Spjd 2522168404Spjdstatic void 2523168404Spjdarc_kmem_reap_now(arc_reclaim_strategy_t strat) 2524168404Spjd{ 2525168404Spjd size_t i; 2526168404Spjd kmem_cache_t *prev_cache = NULL; 2527168404Spjd kmem_cache_t *prev_data_cache = NULL; 2528168404Spjd 2529168404Spjd#ifdef _KERNEL 2530185029Spjd if (arc_meta_used >= arc_meta_limit) { 2531185029Spjd /* 2532185029Spjd * We are exceeding our meta-data cache limit. 2533185029Spjd * Purge some DNLC entries to release holds on meta-data. 2534185029Spjd */ 2535185029Spjd dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2536185029Spjd } 2537168404Spjd#if defined(__i386) 2538168404Spjd /* 2539168404Spjd * Reclaim unused memory from all kmem caches. 2540168404Spjd */ 2541168404Spjd kmem_reap(); 2542168404Spjd#endif 2543168404Spjd#endif 2544168404Spjd 2545168404Spjd /* 2546185029Spjd * An aggressive reclamation will shrink the cache size as well as 2547168404Spjd * reap free buffers from the arc kmem caches. 2548168404Spjd */ 2549168404Spjd if (strat == ARC_RECLAIM_AGGR) 2550168404Spjd arc_shrink(); 2551168404Spjd 2552168404Spjd for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2553168404Spjd if (zio_buf_cache[i] != prev_cache) { 2554168404Spjd prev_cache = zio_buf_cache[i]; 2555168404Spjd kmem_cache_reap_now(zio_buf_cache[i]); 2556168404Spjd } 2557168404Spjd if (zio_data_buf_cache[i] != prev_data_cache) { 2558168404Spjd prev_data_cache = zio_data_buf_cache[i]; 2559168404Spjd kmem_cache_reap_now(zio_data_buf_cache[i]); 2560168404Spjd } 2561168404Spjd } 2562168404Spjd kmem_cache_reap_now(buf_cache); 2563168404Spjd kmem_cache_reap_now(hdr_cache); 2564168404Spjd} 2565168404Spjd 2566168404Spjdstatic void 2567168404Spjdarc_reclaim_thread(void *dummy __unused) 2568168404Spjd{ 2569168404Spjd clock_t growtime = 0; 2570168404Spjd arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2571168404Spjd callb_cpr_t cpr; 2572168404Spjd 2573168404Spjd CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2574168404Spjd 2575168404Spjd mutex_enter(&arc_reclaim_thr_lock); 2576168404Spjd while (arc_thread_exit == 0) { 2577168404Spjd if (arc_reclaim_needed()) { 2578168404Spjd 2579168404Spjd if (arc_no_grow) { 2580168404Spjd if (last_reclaim == ARC_RECLAIM_CONS) { 2581168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2582168404Spjd } else { 2583168404Spjd last_reclaim = ARC_RECLAIM_CONS; 2584168404Spjd } 2585168404Spjd } else { 2586168404Spjd arc_no_grow = TRUE; 2587168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2588168404Spjd membar_producer(); 2589168404Spjd } 2590168404Spjd 2591168404Spjd /* reset the growth delay for every reclaim */ 2592219089Spjd growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2593168404Spjd 2594185029Spjd if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2595168404Spjd /* 2596185029Spjd * If needfree is TRUE our vm_lowmem hook 2597168404Spjd * was called and in that case we must free some 2598168404Spjd * memory, so switch to aggressive mode. 2599168404Spjd */ 2600168404Spjd arc_no_grow = TRUE; 2601168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2602168404Spjd } 2603168404Spjd arc_kmem_reap_now(last_reclaim); 2604185029Spjd arc_warm = B_TRUE; 2605185029Spjd 2606219089Spjd } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2607168404Spjd arc_no_grow = FALSE; 2608168404Spjd } 2609168404Spjd 2610209275Smm arc_adjust(); 2611168404Spjd 2612168404Spjd if (arc_eviction_list != NULL) 2613168404Spjd arc_do_user_evicts(); 2614168404Spjd 2615211762Savg#ifdef _KERNEL 2616211762Savg if (needfree) { 2617185029Spjd needfree = 0; 2618185029Spjd wakeup(&needfree); 2619211762Savg } 2620168404Spjd#endif 2621168404Spjd 2622168404Spjd /* block until needed, or one second, whichever is shorter */ 2623168404Spjd CALLB_CPR_SAFE_BEGIN(&cpr); 2624168404Spjd (void) cv_timedwait(&arc_reclaim_thr_cv, 2625168404Spjd &arc_reclaim_thr_lock, hz); 2626168404Spjd CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2627168404Spjd } 2628168404Spjd 2629168404Spjd arc_thread_exit = 0; 2630168404Spjd cv_broadcast(&arc_reclaim_thr_cv); 2631168404Spjd CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2632168404Spjd thread_exit(); 2633168404Spjd} 2634168404Spjd 2635168404Spjd/* 2636168404Spjd * Adapt arc info given the number of bytes we are trying to add and 2637168404Spjd * the state that we are comming from. This function is only called 2638168404Spjd * when we are adding new content to the cache. 2639168404Spjd */ 2640168404Spjdstatic void 2641168404Spjdarc_adapt(int bytes, arc_state_t *state) 2642168404Spjd{ 2643168404Spjd int mult; 2644208373Smm uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2645168404Spjd 2646185029Spjd if (state == arc_l2c_only) 2647185029Spjd return; 2648185029Spjd 2649168404Spjd ASSERT(bytes > 0); 2650168404Spjd /* 2651168404Spjd * Adapt the target size of the MRU list: 2652168404Spjd * - if we just hit in the MRU ghost list, then increase 2653168404Spjd * the target size of the MRU list. 2654168404Spjd * - if we just hit in the MFU ghost list, then increase 2655168404Spjd * the target size of the MFU list by decreasing the 2656168404Spjd * target size of the MRU list. 2657168404Spjd */ 2658168404Spjd if (state == arc_mru_ghost) { 2659168404Spjd mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2660168404Spjd 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2661209275Smm mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2662168404Spjd 2663208373Smm arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2664168404Spjd } else if (state == arc_mfu_ghost) { 2665208373Smm uint64_t delta; 2666208373Smm 2667168404Spjd mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2668168404Spjd 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2669209275Smm mult = MIN(mult, 10); 2670168404Spjd 2671208373Smm delta = MIN(bytes * mult, arc_p); 2672208373Smm arc_p = MAX(arc_p_min, arc_p - delta); 2673168404Spjd } 2674168404Spjd ASSERT((int64_t)arc_p >= 0); 2675168404Spjd 2676168404Spjd if (arc_reclaim_needed()) { 2677168404Spjd cv_signal(&arc_reclaim_thr_cv); 2678168404Spjd return; 2679168404Spjd } 2680168404Spjd 2681168404Spjd if (arc_no_grow) 2682168404Spjd return; 2683168404Spjd 2684168404Spjd if (arc_c >= arc_c_max) 2685168404Spjd return; 2686168404Spjd 2687168404Spjd /* 2688168404Spjd * If we're within (2 * maxblocksize) bytes of the target 2689168404Spjd * cache size, increment the target cache size 2690168404Spjd */ 2691168404Spjd if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2692168404Spjd atomic_add_64(&arc_c, (int64_t)bytes); 2693168404Spjd if (arc_c > arc_c_max) 2694168404Spjd arc_c = arc_c_max; 2695168404Spjd else if (state == arc_anon) 2696168404Spjd atomic_add_64(&arc_p, (int64_t)bytes); 2697168404Spjd if (arc_p > arc_c) 2698168404Spjd arc_p = arc_c; 2699168404Spjd } 2700168404Spjd ASSERT((int64_t)arc_p >= 0); 2701168404Spjd} 2702168404Spjd 2703168404Spjd/* 2704168404Spjd * Check if the cache has reached its limits and eviction is required 2705168404Spjd * prior to insert. 2706168404Spjd */ 2707168404Spjdstatic int 2708185029Spjdarc_evict_needed(arc_buf_contents_t type) 2709168404Spjd{ 2710185029Spjd if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2711185029Spjd return (1); 2712185029Spjd 2713219089Spjd#ifdef sun 2714185029Spjd#ifdef _KERNEL 2715185029Spjd /* 2716185029Spjd * If zio data pages are being allocated out of a separate heap segment, 2717185029Spjd * then enforce that the size of available vmem for this area remains 2718185029Spjd * above about 1/32nd free. 2719185029Spjd */ 2720185029Spjd if (type == ARC_BUFC_DATA && zio_arena != NULL && 2721185029Spjd vmem_size(zio_arena, VMEM_FREE) < 2722185029Spjd (vmem_size(zio_arena, VMEM_ALLOC) >> 5)) 2723185029Spjd return (1); 2724185029Spjd#endif 2725219089Spjd#endif /* sun */ 2726185029Spjd 2727168404Spjd if (arc_reclaim_needed()) 2728168404Spjd return (1); 2729168404Spjd 2730168404Spjd return (arc_size > arc_c); 2731168404Spjd} 2732168404Spjd 2733168404Spjd/* 2734168404Spjd * The buffer, supplied as the first argument, needs a data block. 2735168404Spjd * So, if we are at cache max, determine which cache should be victimized. 2736168404Spjd * We have the following cases: 2737168404Spjd * 2738168404Spjd * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2739168404Spjd * In this situation if we're out of space, but the resident size of the MFU is 2740168404Spjd * under the limit, victimize the MFU cache to satisfy this insertion request. 2741168404Spjd * 2742168404Spjd * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2743168404Spjd * Here, we've used up all of the available space for the MRU, so we need to 2744168404Spjd * evict from our own cache instead. Evict from the set of resident MRU 2745168404Spjd * entries. 2746168404Spjd * 2747168404Spjd * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2748168404Spjd * c minus p represents the MFU space in the cache, since p is the size of the 2749168404Spjd * cache that is dedicated to the MRU. In this situation there's still space on 2750168404Spjd * the MFU side, so the MRU side needs to be victimized. 2751168404Spjd * 2752168404Spjd * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2753168404Spjd * MFU's resident set is consuming more space than it has been allotted. In 2754168404Spjd * this situation, we must victimize our own cache, the MFU, for this insertion. 2755168404Spjd */ 2756168404Spjdstatic void 2757168404Spjdarc_get_data_buf(arc_buf_t *buf) 2758168404Spjd{ 2759168404Spjd arc_state_t *state = buf->b_hdr->b_state; 2760168404Spjd uint64_t size = buf->b_hdr->b_size; 2761168404Spjd arc_buf_contents_t type = buf->b_hdr->b_type; 2762168404Spjd 2763168404Spjd arc_adapt(size, state); 2764168404Spjd 2765168404Spjd /* 2766168404Spjd * We have not yet reached cache maximum size, 2767168404Spjd * just allocate a new buffer. 2768168404Spjd */ 2769185029Spjd if (!arc_evict_needed(type)) { 2770168404Spjd if (type == ARC_BUFC_METADATA) { 2771168404Spjd buf->b_data = zio_buf_alloc(size); 2772208373Smm arc_space_consume(size, ARC_SPACE_DATA); 2773168404Spjd } else { 2774168404Spjd ASSERT(type == ARC_BUFC_DATA); 2775168404Spjd buf->b_data = zio_data_buf_alloc(size); 2776208373Smm ARCSTAT_INCR(arcstat_data_size, size); 2777185029Spjd atomic_add_64(&arc_size, size); 2778168404Spjd } 2779168404Spjd goto out; 2780168404Spjd } 2781168404Spjd 2782168404Spjd /* 2783168404Spjd * If we are prefetching from the mfu ghost list, this buffer 2784168404Spjd * will end up on the mru list; so steal space from there. 2785168404Spjd */ 2786168404Spjd if (state == arc_mfu_ghost) 2787168404Spjd state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2788168404Spjd else if (state == arc_mru_ghost) 2789168404Spjd state = arc_mru; 2790168404Spjd 2791168404Spjd if (state == arc_mru || state == arc_anon) { 2792168404Spjd uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2793208373Smm state = (arc_mfu->arcs_lsize[type] >= size && 2794185029Spjd arc_p > mru_used) ? arc_mfu : arc_mru; 2795168404Spjd } else { 2796168404Spjd /* MFU cases */ 2797168404Spjd uint64_t mfu_space = arc_c - arc_p; 2798208373Smm state = (arc_mru->arcs_lsize[type] >= size && 2799185029Spjd mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2800168404Spjd } 2801209962Smm if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2802168404Spjd if (type == ARC_BUFC_METADATA) { 2803168404Spjd buf->b_data = zio_buf_alloc(size); 2804208373Smm arc_space_consume(size, ARC_SPACE_DATA); 2805168404Spjd } else { 2806168404Spjd ASSERT(type == ARC_BUFC_DATA); 2807168404Spjd buf->b_data = zio_data_buf_alloc(size); 2808208373Smm ARCSTAT_INCR(arcstat_data_size, size); 2809185029Spjd atomic_add_64(&arc_size, size); 2810168404Spjd } 2811168404Spjd ARCSTAT_BUMP(arcstat_recycle_miss); 2812168404Spjd } 2813168404Spjd ASSERT(buf->b_data != NULL); 2814168404Spjdout: 2815168404Spjd /* 2816168404Spjd * Update the state size. Note that ghost states have a 2817168404Spjd * "ghost size" and so don't need to be updated. 2818168404Spjd */ 2819168404Spjd if (!GHOST_STATE(buf->b_hdr->b_state)) { 2820168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 2821168404Spjd 2822168404Spjd atomic_add_64(&hdr->b_state->arcs_size, size); 2823168404Spjd if (list_link_active(&hdr->b_arc_node)) { 2824168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2825185029Spjd atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2826168404Spjd } 2827168404Spjd /* 2828168404Spjd * If we are growing the cache, and we are adding anonymous 2829168404Spjd * data, and we have outgrown arc_p, update arc_p 2830168404Spjd */ 2831168404Spjd if (arc_size < arc_c && hdr->b_state == arc_anon && 2832168404Spjd arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2833168404Spjd arc_p = MIN(arc_c, arc_p + size); 2834168404Spjd } 2835205231Skmacy ARCSTAT_BUMP(arcstat_allocated); 2836168404Spjd} 2837168404Spjd 2838168404Spjd/* 2839168404Spjd * This routine is called whenever a buffer is accessed. 2840168404Spjd * NOTE: the hash lock is dropped in this function. 2841168404Spjd */ 2842168404Spjdstatic void 2843168404Spjdarc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2844168404Spjd{ 2845219089Spjd clock_t now; 2846219089Spjd 2847168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 2848168404Spjd 2849168404Spjd if (buf->b_state == arc_anon) { 2850168404Spjd /* 2851168404Spjd * This buffer is not in the cache, and does not 2852168404Spjd * appear in our "ghost" list. Add the new buffer 2853168404Spjd * to the MRU state. 2854168404Spjd */ 2855168404Spjd 2856168404Spjd ASSERT(buf->b_arc_access == 0); 2857219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2858168404Spjd DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2859168404Spjd arc_change_state(arc_mru, buf, hash_lock); 2860168404Spjd 2861168404Spjd } else if (buf->b_state == arc_mru) { 2862219089Spjd now = ddi_get_lbolt(); 2863219089Spjd 2864168404Spjd /* 2865168404Spjd * If this buffer is here because of a prefetch, then either: 2866168404Spjd * - clear the flag if this is a "referencing" read 2867168404Spjd * (any subsequent access will bump this into the MFU state). 2868168404Spjd * or 2869168404Spjd * - move the buffer to the head of the list if this is 2870168404Spjd * another prefetch (to make it less likely to be evicted). 2871168404Spjd */ 2872168404Spjd if ((buf->b_flags & ARC_PREFETCH) != 0) { 2873168404Spjd if (refcount_count(&buf->b_refcnt) == 0) { 2874168404Spjd ASSERT(list_link_active(&buf->b_arc_node)); 2875168404Spjd } else { 2876168404Spjd buf->b_flags &= ~ARC_PREFETCH; 2877168404Spjd ARCSTAT_BUMP(arcstat_mru_hits); 2878168404Spjd } 2879219089Spjd buf->b_arc_access = now; 2880168404Spjd return; 2881168404Spjd } 2882168404Spjd 2883168404Spjd /* 2884168404Spjd * This buffer has been "accessed" only once so far, 2885168404Spjd * but it is still in the cache. Move it to the MFU 2886168404Spjd * state. 2887168404Spjd */ 2888219089Spjd if (now > buf->b_arc_access + ARC_MINTIME) { 2889168404Spjd /* 2890168404Spjd * More than 125ms have passed since we 2891168404Spjd * instantiated this buffer. Move it to the 2892168404Spjd * most frequently used state. 2893168404Spjd */ 2894219089Spjd buf->b_arc_access = now; 2895168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2896168404Spjd arc_change_state(arc_mfu, buf, hash_lock); 2897168404Spjd } 2898168404Spjd ARCSTAT_BUMP(arcstat_mru_hits); 2899168404Spjd } else if (buf->b_state == arc_mru_ghost) { 2900168404Spjd arc_state_t *new_state; 2901168404Spjd /* 2902168404Spjd * This buffer has been "accessed" recently, but 2903168404Spjd * was evicted from the cache. Move it to the 2904168404Spjd * MFU state. 2905168404Spjd */ 2906168404Spjd 2907168404Spjd if (buf->b_flags & ARC_PREFETCH) { 2908168404Spjd new_state = arc_mru; 2909168404Spjd if (refcount_count(&buf->b_refcnt) > 0) 2910168404Spjd buf->b_flags &= ~ARC_PREFETCH; 2911168404Spjd DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2912168404Spjd } else { 2913168404Spjd new_state = arc_mfu; 2914168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2915168404Spjd } 2916168404Spjd 2917219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2918168404Spjd arc_change_state(new_state, buf, hash_lock); 2919168404Spjd 2920168404Spjd ARCSTAT_BUMP(arcstat_mru_ghost_hits); 2921168404Spjd } else if (buf->b_state == arc_mfu) { 2922168404Spjd /* 2923168404Spjd * This buffer has been accessed more than once and is 2924168404Spjd * still in the cache. Keep it in the MFU state. 2925168404Spjd * 2926168404Spjd * NOTE: an add_reference() that occurred when we did 2927168404Spjd * the arc_read() will have kicked this off the list. 2928168404Spjd * If it was a prefetch, we will explicitly move it to 2929168404Spjd * the head of the list now. 2930168404Spjd */ 2931168404Spjd if ((buf->b_flags & ARC_PREFETCH) != 0) { 2932168404Spjd ASSERT(refcount_count(&buf->b_refcnt) == 0); 2933168404Spjd ASSERT(list_link_active(&buf->b_arc_node)); 2934168404Spjd } 2935168404Spjd ARCSTAT_BUMP(arcstat_mfu_hits); 2936219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2937168404Spjd } else if (buf->b_state == arc_mfu_ghost) { 2938168404Spjd arc_state_t *new_state = arc_mfu; 2939168404Spjd /* 2940168404Spjd * This buffer has been accessed more than once but has 2941168404Spjd * been evicted from the cache. Move it back to the 2942168404Spjd * MFU state. 2943168404Spjd */ 2944168404Spjd 2945168404Spjd if (buf->b_flags & ARC_PREFETCH) { 2946168404Spjd /* 2947168404Spjd * This is a prefetch access... 2948168404Spjd * move this block back to the MRU state. 2949168404Spjd */ 2950240415Smm ASSERT0(refcount_count(&buf->b_refcnt)); 2951168404Spjd new_state = arc_mru; 2952168404Spjd } 2953168404Spjd 2954219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2955168404Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2956168404Spjd arc_change_state(new_state, buf, hash_lock); 2957168404Spjd 2958168404Spjd ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 2959185029Spjd } else if (buf->b_state == arc_l2c_only) { 2960185029Spjd /* 2961185029Spjd * This buffer is on the 2nd Level ARC. 2962185029Spjd */ 2963185029Spjd 2964219089Spjd buf->b_arc_access = ddi_get_lbolt(); 2965185029Spjd DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2966185029Spjd arc_change_state(arc_mfu, buf, hash_lock); 2967168404Spjd } else { 2968168404Spjd ASSERT(!"invalid arc state"); 2969168404Spjd } 2970168404Spjd} 2971168404Spjd 2972168404Spjd/* a generic arc_done_func_t which you can use */ 2973168404Spjd/* ARGSUSED */ 2974168404Spjdvoid 2975168404Spjdarc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 2976168404Spjd{ 2977219089Spjd if (zio == NULL || zio->io_error == 0) 2978219089Spjd bcopy(buf->b_data, arg, buf->b_hdr->b_size); 2979248571Smm VERIFY(arc_buf_remove_ref(buf, arg)); 2980168404Spjd} 2981168404Spjd 2982185029Spjd/* a generic arc_done_func_t */ 2983168404Spjdvoid 2984168404Spjdarc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 2985168404Spjd{ 2986168404Spjd arc_buf_t **bufp = arg; 2987168404Spjd if (zio && zio->io_error) { 2988248571Smm VERIFY(arc_buf_remove_ref(buf, arg)); 2989168404Spjd *bufp = NULL; 2990168404Spjd } else { 2991168404Spjd *bufp = buf; 2992219089Spjd ASSERT(buf->b_data); 2993168404Spjd } 2994168404Spjd} 2995168404Spjd 2996168404Spjdstatic void 2997168404Spjdarc_read_done(zio_t *zio) 2998168404Spjd{ 2999268075Sdelphij arc_buf_hdr_t *hdr; 3000168404Spjd arc_buf_t *buf; 3001168404Spjd arc_buf_t *abuf; /* buffer we're assigning to callback */ 3002268075Sdelphij kmutex_t *hash_lock = NULL; 3003168404Spjd arc_callback_t *callback_list, *acb; 3004168404Spjd int freeable = FALSE; 3005168404Spjd 3006168404Spjd buf = zio->io_private; 3007168404Spjd hdr = buf->b_hdr; 3008168404Spjd 3009168404Spjd /* 3010168404Spjd * The hdr was inserted into hash-table and removed from lists 3011168404Spjd * prior to starting I/O. We should find this header, since 3012168404Spjd * it's in the hash table, and it should be legit since it's 3013168404Spjd * not possible to evict it during the I/O. The only possible 3014168404Spjd * reason for it not to be found is if we were freed during the 3015168404Spjd * read. 3016168404Spjd */ 3017268075Sdelphij if (HDR_IN_HASH_TABLE(hdr)) { 3018268075Sdelphij ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3019268075Sdelphij ASSERT3U(hdr->b_dva.dva_word[0], ==, 3020268075Sdelphij BP_IDENTITY(zio->io_bp)->dva_word[0]); 3021268075Sdelphij ASSERT3U(hdr->b_dva.dva_word[1], ==, 3022268075Sdelphij BP_IDENTITY(zio->io_bp)->dva_word[1]); 3023168404Spjd 3024268075Sdelphij arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3025268075Sdelphij &hash_lock); 3026168404Spjd 3027268075Sdelphij ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3028268075Sdelphij hash_lock == NULL) || 3029268075Sdelphij (found == hdr && 3030268075Sdelphij DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3031268075Sdelphij (found == hdr && HDR_L2_READING(hdr))); 3032268075Sdelphij } 3033268075Sdelphij 3034185029Spjd hdr->b_flags &= ~ARC_L2_EVICTED; 3035185029Spjd if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 3036185029Spjd hdr->b_flags &= ~ARC_L2CACHE; 3037206796Spjd 3038168404Spjd /* byteswap if necessary */ 3039168404Spjd callback_list = hdr->b_acb; 3040168404Spjd ASSERT(callback_list != NULL); 3041209101Smm if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3042236884Smm dmu_object_byteswap_t bswap = 3043236884Smm DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3044185029Spjd arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3045185029Spjd byteswap_uint64_array : 3046236884Smm dmu_ot_byteswap[bswap].ob_func; 3047185029Spjd func(buf->b_data, hdr->b_size); 3048185029Spjd } 3049168404Spjd 3050185029Spjd arc_cksum_compute(buf, B_FALSE); 3051240133Smm#ifdef illumos 3052240133Smm arc_buf_watch(buf); 3053240133Smm#endif /* illumos */ 3054168404Spjd 3055219089Spjd if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3056219089Spjd /* 3057219089Spjd * Only call arc_access on anonymous buffers. This is because 3058219089Spjd * if we've issued an I/O for an evicted buffer, we've already 3059219089Spjd * called arc_access (to prevent any simultaneous readers from 3060219089Spjd * getting confused). 3061219089Spjd */ 3062219089Spjd arc_access(hdr, hash_lock); 3063219089Spjd } 3064219089Spjd 3065168404Spjd /* create copies of the data buffer for the callers */ 3066168404Spjd abuf = buf; 3067168404Spjd for (acb = callback_list; acb; acb = acb->acb_next) { 3068168404Spjd if (acb->acb_done) { 3069242845Sdelphij if (abuf == NULL) { 3070242845Sdelphij ARCSTAT_BUMP(arcstat_duplicate_reads); 3071168404Spjd abuf = arc_buf_clone(buf); 3072242845Sdelphij } 3073168404Spjd acb->acb_buf = abuf; 3074168404Spjd abuf = NULL; 3075168404Spjd } 3076168404Spjd } 3077168404Spjd hdr->b_acb = NULL; 3078168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3079168404Spjd ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3080219089Spjd if (abuf == buf) { 3081219089Spjd ASSERT(buf->b_efunc == NULL); 3082219089Spjd ASSERT(hdr->b_datacnt == 1); 3083168404Spjd hdr->b_flags |= ARC_BUF_AVAILABLE; 3084219089Spjd } 3085168404Spjd 3086168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3087168404Spjd 3088168404Spjd if (zio->io_error != 0) { 3089168404Spjd hdr->b_flags |= ARC_IO_ERROR; 3090168404Spjd if (hdr->b_state != arc_anon) 3091168404Spjd arc_change_state(arc_anon, hdr, hash_lock); 3092168404Spjd if (HDR_IN_HASH_TABLE(hdr)) 3093168404Spjd buf_hash_remove(hdr); 3094168404Spjd freeable = refcount_is_zero(&hdr->b_refcnt); 3095168404Spjd } 3096168404Spjd 3097168404Spjd /* 3098168404Spjd * Broadcast before we drop the hash_lock to avoid the possibility 3099168404Spjd * that the hdr (and hence the cv) might be freed before we get to 3100168404Spjd * the cv_broadcast(). 3101168404Spjd */ 3102168404Spjd cv_broadcast(&hdr->b_cv); 3103168404Spjd 3104168404Spjd if (hash_lock) { 3105168404Spjd mutex_exit(hash_lock); 3106168404Spjd } else { 3107168404Spjd /* 3108168404Spjd * This block was freed while we waited for the read to 3109168404Spjd * complete. It has been removed from the hash table and 3110168404Spjd * moved to the anonymous state (so that it won't show up 3111168404Spjd * in the cache). 3112168404Spjd */ 3113168404Spjd ASSERT3P(hdr->b_state, ==, arc_anon); 3114168404Spjd freeable = refcount_is_zero(&hdr->b_refcnt); 3115168404Spjd } 3116168404Spjd 3117168404Spjd /* execute each callback and free its structure */ 3118168404Spjd while ((acb = callback_list) != NULL) { 3119168404Spjd if (acb->acb_done) 3120168404Spjd acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3121168404Spjd 3122168404Spjd if (acb->acb_zio_dummy != NULL) { 3123168404Spjd acb->acb_zio_dummy->io_error = zio->io_error; 3124168404Spjd zio_nowait(acb->acb_zio_dummy); 3125168404Spjd } 3126168404Spjd 3127168404Spjd callback_list = acb->acb_next; 3128168404Spjd kmem_free(acb, sizeof (arc_callback_t)); 3129168404Spjd } 3130168404Spjd 3131168404Spjd if (freeable) 3132168404Spjd arc_hdr_destroy(hdr); 3133168404Spjd} 3134168404Spjd 3135168404Spjd/* 3136168404Spjd * "Read" the block block at the specified DVA (in bp) via the 3137168404Spjd * cache. If the block is found in the cache, invoke the provided 3138168404Spjd * callback immediately and return. Note that the `zio' parameter 3139168404Spjd * in the callback will be NULL in this case, since no IO was 3140168404Spjd * required. If the block is not in the cache pass the read request 3141168404Spjd * on to the spa with a substitute callback function, so that the 3142168404Spjd * requested block will be added to the cache. 3143168404Spjd * 3144168404Spjd * If a read request arrives for a block that has a read in-progress, 3145168404Spjd * either wait for the in-progress read to complete (and return the 3146168404Spjd * results); or, if this is a read with a "done" func, add a record 3147168404Spjd * to the read to invoke the "done" func when the read completes, 3148168404Spjd * and return; or just return. 3149168404Spjd * 3150168404Spjd * arc_read_done() will invoke all the requested "done" functions 3151168404Spjd * for readers of this block. 3152168404Spjd */ 3153168404Spjdint 3154246666Smmarc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3155258632Savg void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags, 3156268123Sdelphij const zbookmark_phys_t *zb) 3157168404Spjd{ 3158268075Sdelphij arc_buf_hdr_t *hdr = NULL; 3159247187Smm arc_buf_t *buf = NULL; 3160268075Sdelphij kmutex_t *hash_lock = NULL; 3161185029Spjd zio_t *rzio; 3162228103Smm uint64_t guid = spa_load_guid(spa); 3163168404Spjd 3164268075Sdelphij ASSERT(!BP_IS_EMBEDDED(bp) || 3165268075Sdelphij BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3166268075Sdelphij 3167168404Spjdtop: 3168268075Sdelphij if (!BP_IS_EMBEDDED(bp)) { 3169268075Sdelphij /* 3170268075Sdelphij * Embedded BP's have no DVA and require no I/O to "read". 3171268075Sdelphij * Create an anonymous arc buf to back it. 3172268075Sdelphij */ 3173268075Sdelphij hdr = buf_hash_find(guid, bp, &hash_lock); 3174268075Sdelphij } 3175168404Spjd 3176268075Sdelphij if (hdr != NULL && hdr->b_datacnt > 0) { 3177268075Sdelphij 3178168404Spjd *arc_flags |= ARC_CACHED; 3179168404Spjd 3180168404Spjd if (HDR_IO_IN_PROGRESS(hdr)) { 3181168404Spjd 3182168404Spjd if (*arc_flags & ARC_WAIT) { 3183168404Spjd cv_wait(&hdr->b_cv, hash_lock); 3184168404Spjd mutex_exit(hash_lock); 3185168404Spjd goto top; 3186168404Spjd } 3187168404Spjd ASSERT(*arc_flags & ARC_NOWAIT); 3188168404Spjd 3189168404Spjd if (done) { 3190168404Spjd arc_callback_t *acb = NULL; 3191168404Spjd 3192168404Spjd acb = kmem_zalloc(sizeof (arc_callback_t), 3193168404Spjd KM_SLEEP); 3194168404Spjd acb->acb_done = done; 3195168404Spjd acb->acb_private = private; 3196168404Spjd if (pio != NULL) 3197168404Spjd acb->acb_zio_dummy = zio_null(pio, 3198209962Smm spa, NULL, NULL, NULL, zio_flags); 3199168404Spjd 3200168404Spjd ASSERT(acb->acb_done != NULL); 3201168404Spjd acb->acb_next = hdr->b_acb; 3202168404Spjd hdr->b_acb = acb; 3203168404Spjd add_reference(hdr, hash_lock, private); 3204168404Spjd mutex_exit(hash_lock); 3205168404Spjd return (0); 3206168404Spjd } 3207168404Spjd mutex_exit(hash_lock); 3208168404Spjd return (0); 3209168404Spjd } 3210168404Spjd 3211168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3212168404Spjd 3213168404Spjd if (done) { 3214168404Spjd add_reference(hdr, hash_lock, private); 3215168404Spjd /* 3216168404Spjd * If this block is already in use, create a new 3217168404Spjd * copy of the data so that we will be guaranteed 3218168404Spjd * that arc_release() will always succeed. 3219168404Spjd */ 3220168404Spjd buf = hdr->b_buf; 3221168404Spjd ASSERT(buf); 3222168404Spjd ASSERT(buf->b_data); 3223168404Spjd if (HDR_BUF_AVAILABLE(hdr)) { 3224168404Spjd ASSERT(buf->b_efunc == NULL); 3225168404Spjd hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3226168404Spjd } else { 3227168404Spjd buf = arc_buf_clone(buf); 3228168404Spjd } 3229219089Spjd 3230168404Spjd } else if (*arc_flags & ARC_PREFETCH && 3231168404Spjd refcount_count(&hdr->b_refcnt) == 0) { 3232168404Spjd hdr->b_flags |= ARC_PREFETCH; 3233168404Spjd } 3234168404Spjd DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3235168404Spjd arc_access(hdr, hash_lock); 3236185029Spjd if (*arc_flags & ARC_L2CACHE) 3237185029Spjd hdr->b_flags |= ARC_L2CACHE; 3238251478Sdelphij if (*arc_flags & ARC_L2COMPRESS) 3239251478Sdelphij hdr->b_flags |= ARC_L2COMPRESS; 3240168404Spjd mutex_exit(hash_lock); 3241168404Spjd ARCSTAT_BUMP(arcstat_hits); 3242168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3243168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3244168404Spjd data, metadata, hits); 3245168404Spjd 3246168404Spjd if (done) 3247168404Spjd done(NULL, buf, private); 3248168404Spjd } else { 3249168404Spjd uint64_t size = BP_GET_LSIZE(bp); 3250268075Sdelphij arc_callback_t *acb; 3251185029Spjd vdev_t *vd = NULL; 3252247187Smm uint64_t addr = 0; 3253208373Smm boolean_t devw = B_FALSE; 3254258389Savg enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3255258389Savg uint64_t b_asize = 0; 3256168404Spjd 3257168404Spjd if (hdr == NULL) { 3258168404Spjd /* this block is not in the cache */ 3259268075Sdelphij arc_buf_hdr_t *exists = NULL; 3260168404Spjd arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3261168404Spjd buf = arc_buf_alloc(spa, size, private, type); 3262168404Spjd hdr = buf->b_hdr; 3263268075Sdelphij if (!BP_IS_EMBEDDED(bp)) { 3264268075Sdelphij hdr->b_dva = *BP_IDENTITY(bp); 3265268075Sdelphij hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3266268075Sdelphij hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3267268075Sdelphij exists = buf_hash_insert(hdr, &hash_lock); 3268268075Sdelphij } 3269268075Sdelphij if (exists != NULL) { 3270168404Spjd /* somebody beat us to the hash insert */ 3271168404Spjd mutex_exit(hash_lock); 3272219089Spjd buf_discard_identity(hdr); 3273168404Spjd (void) arc_buf_remove_ref(buf, private); 3274168404Spjd goto top; /* restart the IO request */ 3275168404Spjd } 3276168404Spjd /* if this is a prefetch, we don't have a reference */ 3277168404Spjd if (*arc_flags & ARC_PREFETCH) { 3278168404Spjd (void) remove_reference(hdr, hash_lock, 3279168404Spjd private); 3280168404Spjd hdr->b_flags |= ARC_PREFETCH; 3281168404Spjd } 3282185029Spjd if (*arc_flags & ARC_L2CACHE) 3283185029Spjd hdr->b_flags |= ARC_L2CACHE; 3284251478Sdelphij if (*arc_flags & ARC_L2COMPRESS) 3285251478Sdelphij hdr->b_flags |= ARC_L2COMPRESS; 3286168404Spjd if (BP_GET_LEVEL(bp) > 0) 3287168404Spjd hdr->b_flags |= ARC_INDIRECT; 3288168404Spjd } else { 3289168404Spjd /* this block is in the ghost cache */ 3290168404Spjd ASSERT(GHOST_STATE(hdr->b_state)); 3291168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3292240415Smm ASSERT0(refcount_count(&hdr->b_refcnt)); 3293168404Spjd ASSERT(hdr->b_buf == NULL); 3294168404Spjd 3295168404Spjd /* if this is a prefetch, we don't have a reference */ 3296168404Spjd if (*arc_flags & ARC_PREFETCH) 3297168404Spjd hdr->b_flags |= ARC_PREFETCH; 3298168404Spjd else 3299168404Spjd add_reference(hdr, hash_lock, private); 3300185029Spjd if (*arc_flags & ARC_L2CACHE) 3301185029Spjd hdr->b_flags |= ARC_L2CACHE; 3302251478Sdelphij if (*arc_flags & ARC_L2COMPRESS) 3303251478Sdelphij hdr->b_flags |= ARC_L2COMPRESS; 3304185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3305168404Spjd buf->b_hdr = hdr; 3306168404Spjd buf->b_data = NULL; 3307168404Spjd buf->b_efunc = NULL; 3308168404Spjd buf->b_private = NULL; 3309168404Spjd buf->b_next = NULL; 3310168404Spjd hdr->b_buf = buf; 3311168404Spjd ASSERT(hdr->b_datacnt == 0); 3312168404Spjd hdr->b_datacnt = 1; 3313219089Spjd arc_get_data_buf(buf); 3314219089Spjd arc_access(hdr, hash_lock); 3315168404Spjd } 3316168404Spjd 3317219089Spjd ASSERT(!GHOST_STATE(hdr->b_state)); 3318219089Spjd 3319168404Spjd acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3320168404Spjd acb->acb_done = done; 3321168404Spjd acb->acb_private = private; 3322168404Spjd 3323168404Spjd ASSERT(hdr->b_acb == NULL); 3324168404Spjd hdr->b_acb = acb; 3325168404Spjd hdr->b_flags |= ARC_IO_IN_PROGRESS; 3326168404Spjd 3327258389Savg if (hdr->b_l2hdr != NULL && 3328185029Spjd (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3329208373Smm devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3330185029Spjd addr = hdr->b_l2hdr->b_daddr; 3331258389Savg b_compress = hdr->b_l2hdr->b_compress; 3332258389Savg b_asize = hdr->b_l2hdr->b_asize; 3333185029Spjd /* 3334185029Spjd * Lock out device removal. 3335185029Spjd */ 3336185029Spjd if (vdev_is_dead(vd) || 3337185029Spjd !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3338185029Spjd vd = NULL; 3339185029Spjd } 3340185029Spjd 3341268075Sdelphij if (hash_lock != NULL) 3342268075Sdelphij mutex_exit(hash_lock); 3343168404Spjd 3344251629Sdelphij /* 3345251629Sdelphij * At this point, we have a level 1 cache miss. Try again in 3346251629Sdelphij * L2ARC if possible. 3347251629Sdelphij */ 3348168404Spjd ASSERT3U(hdr->b_size, ==, size); 3349219089Spjd DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3350268123Sdelphij uint64_t, size, zbookmark_phys_t *, zb); 3351168404Spjd ARCSTAT_BUMP(arcstat_misses); 3352168404Spjd ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3353168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3354168404Spjd data, metadata, misses); 3355228392Spjd#ifdef _KERNEL 3356228392Spjd curthread->td_ru.ru_inblock++; 3357228392Spjd#endif 3358168404Spjd 3359208373Smm if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3360185029Spjd /* 3361185029Spjd * Read from the L2ARC if the following are true: 3362185029Spjd * 1. The L2ARC vdev was previously cached. 3363185029Spjd * 2. This buffer still has L2ARC metadata. 3364185029Spjd * 3. This buffer isn't currently writing to the L2ARC. 3365185029Spjd * 4. The L2ARC entry wasn't evicted, which may 3366185029Spjd * also have invalidated the vdev. 3367208373Smm * 5. This isn't prefetch and l2arc_noprefetch is set. 3368185029Spjd */ 3369185029Spjd if (hdr->b_l2hdr != NULL && 3370208373Smm !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3371208373Smm !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3372185029Spjd l2arc_read_callback_t *cb; 3373185029Spjd 3374185029Spjd DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3375185029Spjd ARCSTAT_BUMP(arcstat_l2_hits); 3376185029Spjd 3377185029Spjd cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3378185029Spjd KM_SLEEP); 3379185029Spjd cb->l2rcb_buf = buf; 3380185029Spjd cb->l2rcb_spa = spa; 3381185029Spjd cb->l2rcb_bp = *bp; 3382185029Spjd cb->l2rcb_zb = *zb; 3383185029Spjd cb->l2rcb_flags = zio_flags; 3384258389Savg cb->l2rcb_compress = b_compress; 3385185029Spjd 3386247187Smm ASSERT(addr >= VDEV_LABEL_START_SIZE && 3387247187Smm addr + size < vd->vdev_psize - 3388247187Smm VDEV_LABEL_END_SIZE); 3389247187Smm 3390185029Spjd /* 3391185029Spjd * l2arc read. The SCL_L2ARC lock will be 3392185029Spjd * released by l2arc_read_done(). 3393251478Sdelphij * Issue a null zio if the underlying buffer 3394251478Sdelphij * was squashed to zero size by compression. 3395185029Spjd */ 3396258389Savg if (b_compress == ZIO_COMPRESS_EMPTY) { 3397251478Sdelphij rzio = zio_null(pio, spa, vd, 3398251478Sdelphij l2arc_read_done, cb, 3399251478Sdelphij zio_flags | ZIO_FLAG_DONT_CACHE | 3400251478Sdelphij ZIO_FLAG_CANFAIL | 3401251478Sdelphij ZIO_FLAG_DONT_PROPAGATE | 3402251478Sdelphij ZIO_FLAG_DONT_RETRY); 3403251478Sdelphij } else { 3404251478Sdelphij rzio = zio_read_phys(pio, vd, addr, 3405258389Savg b_asize, buf->b_data, 3406258389Savg ZIO_CHECKSUM_OFF, 3407251478Sdelphij l2arc_read_done, cb, priority, 3408251478Sdelphij zio_flags | ZIO_FLAG_DONT_CACHE | 3409251478Sdelphij ZIO_FLAG_CANFAIL | 3410251478Sdelphij ZIO_FLAG_DONT_PROPAGATE | 3411251478Sdelphij ZIO_FLAG_DONT_RETRY, B_FALSE); 3412251478Sdelphij } 3413185029Spjd DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3414185029Spjd zio_t *, rzio); 3415258389Savg ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3416185029Spjd 3417185029Spjd if (*arc_flags & ARC_NOWAIT) { 3418185029Spjd zio_nowait(rzio); 3419185029Spjd return (0); 3420185029Spjd } 3421185029Spjd 3422185029Spjd ASSERT(*arc_flags & ARC_WAIT); 3423185029Spjd if (zio_wait(rzio) == 0) 3424185029Spjd return (0); 3425185029Spjd 3426185029Spjd /* l2arc read error; goto zio_read() */ 3427185029Spjd } else { 3428185029Spjd DTRACE_PROBE1(l2arc__miss, 3429185029Spjd arc_buf_hdr_t *, hdr); 3430185029Spjd ARCSTAT_BUMP(arcstat_l2_misses); 3431185029Spjd if (HDR_L2_WRITING(hdr)) 3432185029Spjd ARCSTAT_BUMP(arcstat_l2_rw_clash); 3433185029Spjd spa_config_exit(spa, SCL_L2ARC, vd); 3434185029Spjd } 3435208373Smm } else { 3436208373Smm if (vd != NULL) 3437208373Smm spa_config_exit(spa, SCL_L2ARC, vd); 3438208373Smm if (l2arc_ndev != 0) { 3439208373Smm DTRACE_PROBE1(l2arc__miss, 3440208373Smm arc_buf_hdr_t *, hdr); 3441208373Smm ARCSTAT_BUMP(arcstat_l2_misses); 3442208373Smm } 3443185029Spjd } 3444185029Spjd 3445168404Spjd rzio = zio_read(pio, spa, bp, buf->b_data, size, 3446185029Spjd arc_read_done, buf, priority, zio_flags, zb); 3447168404Spjd 3448168404Spjd if (*arc_flags & ARC_WAIT) 3449168404Spjd return (zio_wait(rzio)); 3450168404Spjd 3451168404Spjd ASSERT(*arc_flags & ARC_NOWAIT); 3452168404Spjd zio_nowait(rzio); 3453168404Spjd } 3454168404Spjd return (0); 3455168404Spjd} 3456168404Spjd 3457168404Spjdvoid 3458168404Spjdarc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3459168404Spjd{ 3460168404Spjd ASSERT(buf->b_hdr != NULL); 3461168404Spjd ASSERT(buf->b_hdr->b_state != arc_anon); 3462168404Spjd ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3463219089Spjd ASSERT(buf->b_efunc == NULL); 3464219089Spjd ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3465219089Spjd 3466168404Spjd buf->b_efunc = func; 3467168404Spjd buf->b_private = private; 3468168404Spjd} 3469168404Spjd 3470168404Spjd/* 3471251520Sdelphij * Notify the arc that a block was freed, and thus will never be used again. 3472251520Sdelphij */ 3473251520Sdelphijvoid 3474251520Sdelphijarc_freed(spa_t *spa, const blkptr_t *bp) 3475251520Sdelphij{ 3476251520Sdelphij arc_buf_hdr_t *hdr; 3477251520Sdelphij kmutex_t *hash_lock; 3478251520Sdelphij uint64_t guid = spa_load_guid(spa); 3479251520Sdelphij 3480268075Sdelphij ASSERT(!BP_IS_EMBEDDED(bp)); 3481268075Sdelphij 3482268075Sdelphij hdr = buf_hash_find(guid, bp, &hash_lock); 3483251520Sdelphij if (hdr == NULL) 3484251520Sdelphij return; 3485251520Sdelphij if (HDR_BUF_AVAILABLE(hdr)) { 3486251520Sdelphij arc_buf_t *buf = hdr->b_buf; 3487251520Sdelphij add_reference(hdr, hash_lock, FTAG); 3488251520Sdelphij hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3489251520Sdelphij mutex_exit(hash_lock); 3490251520Sdelphij 3491251520Sdelphij arc_release(buf, FTAG); 3492251520Sdelphij (void) arc_buf_remove_ref(buf, FTAG); 3493251520Sdelphij } else { 3494251520Sdelphij mutex_exit(hash_lock); 3495251520Sdelphij } 3496251520Sdelphij 3497251520Sdelphij} 3498251520Sdelphij 3499251520Sdelphij/* 3500268858Sdelphij * Clear the user eviction callback set by arc_set_callback(), first calling 3501268858Sdelphij * it if it exists. Because the presence of a callback keeps an arc_buf cached 3502268858Sdelphij * clearing the callback may result in the arc_buf being destroyed. However, 3503268858Sdelphij * it will not result in the *last* arc_buf being destroyed, hence the data 3504268858Sdelphij * will remain cached in the ARC. We make a copy of the arc buffer here so 3505268858Sdelphij * that we can process the callback without holding any locks. 3506268858Sdelphij * 3507268858Sdelphij * It's possible that the callback is already in the process of being cleared 3508268858Sdelphij * by another thread. In this case we can not clear the callback. 3509268858Sdelphij * 3510268858Sdelphij * Returns B_TRUE if the callback was successfully called and cleared. 3511168404Spjd */ 3512268858Sdelphijboolean_t 3513268858Sdelphijarc_clear_callback(arc_buf_t *buf) 3514168404Spjd{ 3515168404Spjd arc_buf_hdr_t *hdr; 3516168404Spjd kmutex_t *hash_lock; 3517268858Sdelphij arc_evict_func_t *efunc = buf->b_efunc; 3518268858Sdelphij void *private = buf->b_private; 3519205231Skmacy list_t *list, *evicted_list; 3520205231Skmacy kmutex_t *lock, *evicted_lock; 3521206796Spjd 3522219089Spjd mutex_enter(&buf->b_evict_lock); 3523168404Spjd hdr = buf->b_hdr; 3524168404Spjd if (hdr == NULL) { 3525168404Spjd /* 3526168404Spjd * We are in arc_do_user_evicts(). 3527168404Spjd */ 3528168404Spjd ASSERT(buf->b_data == NULL); 3529219089Spjd mutex_exit(&buf->b_evict_lock); 3530268858Sdelphij return (B_FALSE); 3531185029Spjd } else if (buf->b_data == NULL) { 3532185029Spjd /* 3533185029Spjd * We are on the eviction list; process this buffer now 3534185029Spjd * but let arc_do_user_evicts() do the reaping. 3535185029Spjd */ 3536185029Spjd buf->b_efunc = NULL; 3537219089Spjd mutex_exit(&buf->b_evict_lock); 3538268858Sdelphij VERIFY0(efunc(private)); 3539268858Sdelphij return (B_TRUE); 3540168404Spjd } 3541168404Spjd hash_lock = HDR_LOCK(hdr); 3542168404Spjd mutex_enter(hash_lock); 3543219089Spjd hdr = buf->b_hdr; 3544219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3545168404Spjd 3546168404Spjd ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3547168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3548168404Spjd 3549268858Sdelphij buf->b_efunc = NULL; 3550268858Sdelphij buf->b_private = NULL; 3551168404Spjd 3552268858Sdelphij if (hdr->b_datacnt > 1) { 3553268858Sdelphij mutex_exit(&buf->b_evict_lock); 3554268858Sdelphij arc_buf_destroy(buf, FALSE, TRUE); 3555268858Sdelphij } else { 3556268858Sdelphij ASSERT(buf == hdr->b_buf); 3557268858Sdelphij hdr->b_flags |= ARC_BUF_AVAILABLE; 3558268858Sdelphij mutex_exit(&buf->b_evict_lock); 3559268858Sdelphij } 3560168404Spjd 3561168404Spjd mutex_exit(hash_lock); 3562268858Sdelphij VERIFY0(efunc(private)); 3563268858Sdelphij return (B_TRUE); 3564168404Spjd} 3565168404Spjd 3566168404Spjd/* 3567251629Sdelphij * Release this buffer from the cache, making it an anonymous buffer. This 3568251629Sdelphij * must be done after a read and prior to modifying the buffer contents. 3569168404Spjd * If the buffer has more than one reference, we must make 3570185029Spjd * a new hdr for the buffer. 3571168404Spjd */ 3572168404Spjdvoid 3573168404Spjdarc_release(arc_buf_t *buf, void *tag) 3574168404Spjd{ 3575185029Spjd arc_buf_hdr_t *hdr; 3576219089Spjd kmutex_t *hash_lock = NULL; 3577185029Spjd l2arc_buf_hdr_t *l2hdr; 3578185029Spjd uint64_t buf_size; 3579168404Spjd 3580219089Spjd /* 3581219089Spjd * It would be nice to assert that if it's DMU metadata (level > 3582219089Spjd * 0 || it's the dnode file), then it must be syncing context. 3583219089Spjd * But we don't know that information at this level. 3584219089Spjd */ 3585219089Spjd 3586219089Spjd mutex_enter(&buf->b_evict_lock); 3587185029Spjd hdr = buf->b_hdr; 3588185029Spjd 3589168404Spjd /* this buffer is not on any list */ 3590168404Spjd ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3591168404Spjd 3592168404Spjd if (hdr->b_state == arc_anon) { 3593168404Spjd /* this buffer is already released */ 3594168404Spjd ASSERT(buf->b_efunc == NULL); 3595208373Smm } else { 3596208373Smm hash_lock = HDR_LOCK(hdr); 3597208373Smm mutex_enter(hash_lock); 3598219089Spjd hdr = buf->b_hdr; 3599219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3600168404Spjd } 3601168404Spjd 3602185029Spjd l2hdr = hdr->b_l2hdr; 3603185029Spjd if (l2hdr) { 3604185029Spjd mutex_enter(&l2arc_buflist_mtx); 3605185029Spjd hdr->b_l2hdr = NULL; 3606258388Savg list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3607185029Spjd } 3608247187Smm buf_size = hdr->b_size; 3609185029Spjd 3610168404Spjd /* 3611168404Spjd * Do we have more than one buf? 3612168404Spjd */ 3613185029Spjd if (hdr->b_datacnt > 1) { 3614168404Spjd arc_buf_hdr_t *nhdr; 3615168404Spjd arc_buf_t **bufp; 3616168404Spjd uint64_t blksz = hdr->b_size; 3617209962Smm uint64_t spa = hdr->b_spa; 3618168404Spjd arc_buf_contents_t type = hdr->b_type; 3619185029Spjd uint32_t flags = hdr->b_flags; 3620168404Spjd 3621185029Spjd ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3622168404Spjd /* 3623219089Spjd * Pull the data off of this hdr and attach it to 3624219089Spjd * a new anonymous hdr. 3625168404Spjd */ 3626168404Spjd (void) remove_reference(hdr, hash_lock, tag); 3627168404Spjd bufp = &hdr->b_buf; 3628168404Spjd while (*bufp != buf) 3629168404Spjd bufp = &(*bufp)->b_next; 3630219089Spjd *bufp = buf->b_next; 3631168404Spjd buf->b_next = NULL; 3632168404Spjd 3633168404Spjd ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3634168404Spjd atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3635168404Spjd if (refcount_is_zero(&hdr->b_refcnt)) { 3636185029Spjd uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3637185029Spjd ASSERT3U(*size, >=, hdr->b_size); 3638185029Spjd atomic_add_64(size, -hdr->b_size); 3639168404Spjd } 3640242845Sdelphij 3641242845Sdelphij /* 3642242845Sdelphij * We're releasing a duplicate user data buffer, update 3643242845Sdelphij * our statistics accordingly. 3644242845Sdelphij */ 3645242845Sdelphij if (hdr->b_type == ARC_BUFC_DATA) { 3646242845Sdelphij ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3647242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3648242845Sdelphij -hdr->b_size); 3649242845Sdelphij } 3650168404Spjd hdr->b_datacnt -= 1; 3651168404Spjd arc_cksum_verify(buf); 3652240133Smm#ifdef illumos 3653240133Smm arc_buf_unwatch(buf); 3654240133Smm#endif /* illumos */ 3655168404Spjd 3656168404Spjd mutex_exit(hash_lock); 3657168404Spjd 3658185029Spjd nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3659168404Spjd nhdr->b_size = blksz; 3660168404Spjd nhdr->b_spa = spa; 3661168404Spjd nhdr->b_type = type; 3662168404Spjd nhdr->b_buf = buf; 3663168404Spjd nhdr->b_state = arc_anon; 3664168404Spjd nhdr->b_arc_access = 0; 3665185029Spjd nhdr->b_flags = flags & ARC_L2_WRITING; 3666185029Spjd nhdr->b_l2hdr = NULL; 3667168404Spjd nhdr->b_datacnt = 1; 3668168404Spjd nhdr->b_freeze_cksum = NULL; 3669168404Spjd (void) refcount_add(&nhdr->b_refcnt, tag); 3670168404Spjd buf->b_hdr = nhdr; 3671219089Spjd mutex_exit(&buf->b_evict_lock); 3672168404Spjd atomic_add_64(&arc_anon->arcs_size, blksz); 3673168404Spjd } else { 3674219089Spjd mutex_exit(&buf->b_evict_lock); 3675168404Spjd ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3676168404Spjd ASSERT(!list_link_active(&hdr->b_arc_node)); 3677168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3678219089Spjd if (hdr->b_state != arc_anon) 3679219089Spjd arc_change_state(arc_anon, hdr, hash_lock); 3680168404Spjd hdr->b_arc_access = 0; 3681219089Spjd if (hash_lock) 3682219089Spjd mutex_exit(hash_lock); 3683185029Spjd 3684219089Spjd buf_discard_identity(hdr); 3685168404Spjd arc_buf_thaw(buf); 3686168404Spjd } 3687168404Spjd buf->b_efunc = NULL; 3688168404Spjd buf->b_private = NULL; 3689185029Spjd 3690185029Spjd if (l2hdr) { 3691251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3692268085Sdelphij vdev_space_update(l2hdr->b_dev->l2ad_vdev, 3693268085Sdelphij -l2hdr->b_asize, 0, 0); 3694248572Ssmh trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3695248574Ssmh hdr->b_size, 0); 3696185029Spjd kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3697185029Spjd ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3698185029Spjd mutex_exit(&l2arc_buflist_mtx); 3699185029Spjd } 3700168404Spjd} 3701168404Spjd 3702168404Spjdint 3703168404Spjdarc_released(arc_buf_t *buf) 3704168404Spjd{ 3705185029Spjd int released; 3706185029Spjd 3707219089Spjd mutex_enter(&buf->b_evict_lock); 3708185029Spjd released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3709219089Spjd mutex_exit(&buf->b_evict_lock); 3710185029Spjd return (released); 3711168404Spjd} 3712168404Spjd 3713168404Spjd#ifdef ZFS_DEBUG 3714168404Spjdint 3715168404Spjdarc_referenced(arc_buf_t *buf) 3716168404Spjd{ 3717185029Spjd int referenced; 3718185029Spjd 3719219089Spjd mutex_enter(&buf->b_evict_lock); 3720185029Spjd referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3721219089Spjd mutex_exit(&buf->b_evict_lock); 3722185029Spjd return (referenced); 3723168404Spjd} 3724168404Spjd#endif 3725168404Spjd 3726168404Spjdstatic void 3727168404Spjdarc_write_ready(zio_t *zio) 3728168404Spjd{ 3729168404Spjd arc_write_callback_t *callback = zio->io_private; 3730168404Spjd arc_buf_t *buf = callback->awcb_buf; 3731185029Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3732168404Spjd 3733185029Spjd ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3734185029Spjd callback->awcb_ready(zio, buf, callback->awcb_private); 3735185029Spjd 3736185029Spjd /* 3737185029Spjd * If the IO is already in progress, then this is a re-write 3738185029Spjd * attempt, so we need to thaw and re-compute the cksum. 3739185029Spjd * It is the responsibility of the callback to handle the 3740185029Spjd * accounting for any re-write attempt. 3741185029Spjd */ 3742185029Spjd if (HDR_IO_IN_PROGRESS(hdr)) { 3743185029Spjd mutex_enter(&hdr->b_freeze_lock); 3744185029Spjd if (hdr->b_freeze_cksum != NULL) { 3745185029Spjd kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3746185029Spjd hdr->b_freeze_cksum = NULL; 3747185029Spjd } 3748185029Spjd mutex_exit(&hdr->b_freeze_lock); 3749168404Spjd } 3750185029Spjd arc_cksum_compute(buf, B_FALSE); 3751185029Spjd hdr->b_flags |= ARC_IO_IN_PROGRESS; 3752168404Spjd} 3753168404Spjd 3754258632Savg/* 3755258632Savg * The SPA calls this callback for each physical write that happens on behalf 3756258632Savg * of a logical write. See the comment in dbuf_write_physdone() for details. 3757258632Savg */ 3758168404Spjdstatic void 3759258632Savgarc_write_physdone(zio_t *zio) 3760258632Savg{ 3761258632Savg arc_write_callback_t *cb = zio->io_private; 3762258632Savg if (cb->awcb_physdone != NULL) 3763258632Savg cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3764258632Savg} 3765258632Savg 3766258632Savgstatic void 3767168404Spjdarc_write_done(zio_t *zio) 3768168404Spjd{ 3769168404Spjd arc_write_callback_t *callback = zio->io_private; 3770168404Spjd arc_buf_t *buf = callback->awcb_buf; 3771168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3772168404Spjd 3773219089Spjd ASSERT(hdr->b_acb == NULL); 3774168404Spjd 3775219089Spjd if (zio->io_error == 0) { 3776268075Sdelphij if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 3777260150Sdelphij buf_discard_identity(hdr); 3778260150Sdelphij } else { 3779260150Sdelphij hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3780260150Sdelphij hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3781260150Sdelphij hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3782260150Sdelphij } 3783219089Spjd } else { 3784219089Spjd ASSERT(BUF_EMPTY(hdr)); 3785219089Spjd } 3786219089Spjd 3787168404Spjd /* 3788268075Sdelphij * If the block to be written was all-zero or compressed enough to be 3789268075Sdelphij * embedded in the BP, no write was performed so there will be no 3790268075Sdelphij * dva/birth/checksum. The buffer must therefore remain anonymous 3791268075Sdelphij * (and uncached). 3792168404Spjd */ 3793168404Spjd if (!BUF_EMPTY(hdr)) { 3794168404Spjd arc_buf_hdr_t *exists; 3795168404Spjd kmutex_t *hash_lock; 3796168404Spjd 3797219089Spjd ASSERT(zio->io_error == 0); 3798219089Spjd 3799168404Spjd arc_cksum_verify(buf); 3800168404Spjd 3801168404Spjd exists = buf_hash_insert(hdr, &hash_lock); 3802168404Spjd if (exists) { 3803168404Spjd /* 3804168404Spjd * This can only happen if we overwrite for 3805168404Spjd * sync-to-convergence, because we remove 3806168404Spjd * buffers from the hash table when we arc_free(). 3807168404Spjd */ 3808219089Spjd if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3809219089Spjd if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3810219089Spjd panic("bad overwrite, hdr=%p exists=%p", 3811219089Spjd (void *)hdr, (void *)exists); 3812219089Spjd ASSERT(refcount_is_zero(&exists->b_refcnt)); 3813219089Spjd arc_change_state(arc_anon, exists, hash_lock); 3814219089Spjd mutex_exit(hash_lock); 3815219089Spjd arc_hdr_destroy(exists); 3816219089Spjd exists = buf_hash_insert(hdr, &hash_lock); 3817219089Spjd ASSERT3P(exists, ==, NULL); 3818243524Smm } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3819243524Smm /* nopwrite */ 3820243524Smm ASSERT(zio->io_prop.zp_nopwrite); 3821243524Smm if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3822243524Smm panic("bad nopwrite, hdr=%p exists=%p", 3823243524Smm (void *)hdr, (void *)exists); 3824219089Spjd } else { 3825219089Spjd /* Dedup */ 3826219089Spjd ASSERT(hdr->b_datacnt == 1); 3827219089Spjd ASSERT(hdr->b_state == arc_anon); 3828219089Spjd ASSERT(BP_GET_DEDUP(zio->io_bp)); 3829219089Spjd ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3830219089Spjd } 3831168404Spjd } 3832168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3833185029Spjd /* if it's not anon, we are doing a scrub */ 3834219089Spjd if (!exists && hdr->b_state == arc_anon) 3835185029Spjd arc_access(hdr, hash_lock); 3836168404Spjd mutex_exit(hash_lock); 3837168404Spjd } else { 3838168404Spjd hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3839168404Spjd } 3840168404Spjd 3841219089Spjd ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3842219089Spjd callback->awcb_done(zio, buf, callback->awcb_private); 3843168404Spjd 3844168404Spjd kmem_free(callback, sizeof (arc_write_callback_t)); 3845168404Spjd} 3846168404Spjd 3847168404Spjdzio_t * 3848219089Spjdarc_write(zio_t *pio, spa_t *spa, uint64_t txg, 3849251478Sdelphij blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 3850258632Savg const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 3851258632Savg arc_done_func_t *done, void *private, zio_priority_t priority, 3852268123Sdelphij int zio_flags, const zbookmark_phys_t *zb) 3853168404Spjd{ 3854168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3855168404Spjd arc_write_callback_t *callback; 3856185029Spjd zio_t *zio; 3857168404Spjd 3858185029Spjd ASSERT(ready != NULL); 3859219089Spjd ASSERT(done != NULL); 3860168404Spjd ASSERT(!HDR_IO_ERROR(hdr)); 3861168404Spjd ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3862219089Spjd ASSERT(hdr->b_acb == NULL); 3863185029Spjd if (l2arc) 3864185029Spjd hdr->b_flags |= ARC_L2CACHE; 3865251478Sdelphij if (l2arc_compress) 3866251478Sdelphij hdr->b_flags |= ARC_L2COMPRESS; 3867168404Spjd callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 3868168404Spjd callback->awcb_ready = ready; 3869258632Savg callback->awcb_physdone = physdone; 3870168404Spjd callback->awcb_done = done; 3871168404Spjd callback->awcb_private = private; 3872168404Spjd callback->awcb_buf = buf; 3873168404Spjd 3874219089Spjd zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 3875258632Savg arc_write_ready, arc_write_physdone, arc_write_done, callback, 3876258632Savg priority, zio_flags, zb); 3877185029Spjd 3878168404Spjd return (zio); 3879168404Spjd} 3880168404Spjd 3881185029Spjdstatic int 3882258632Savgarc_memory_throttle(uint64_t reserve, uint64_t txg) 3883185029Spjd{ 3884185029Spjd#ifdef _KERNEL 3885219089Spjd uint64_t available_memory = 3886263620Sbdrewery ptoa((uintmax_t)vm_cnt.v_free_count + vm_cnt.v_cache_count); 3887185029Spjd static uint64_t page_load = 0; 3888185029Spjd static uint64_t last_txg = 0; 3889185029Spjd 3890219089Spjd#ifdef sun 3891185029Spjd#if defined(__i386) 3892185029Spjd available_memory = 3893185029Spjd MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 3894185029Spjd#endif 3895219089Spjd#endif /* sun */ 3896258632Savg 3897263620Sbdrewery if (vm_cnt.v_free_count + vm_cnt.v_cache_count > 3898258632Savg (uint64_t)physmem * arc_lotsfree_percent / 100) 3899185029Spjd return (0); 3900185029Spjd 3901185029Spjd if (txg > last_txg) { 3902185029Spjd last_txg = txg; 3903185029Spjd page_load = 0; 3904185029Spjd } 3905185029Spjd /* 3906185029Spjd * If we are in pageout, we know that memory is already tight, 3907185029Spjd * the arc is already going to be evicting, so we just want to 3908185029Spjd * continue to let page writes occur as quickly as possible. 3909185029Spjd */ 3910185029Spjd if (curproc == pageproc) { 3911185029Spjd if (page_load > available_memory / 4) 3912249195Smm return (SET_ERROR(ERESTART)); 3913185029Spjd /* Note: reserve is inflated, so we deflate */ 3914185029Spjd page_load += reserve / 8; 3915185029Spjd return (0); 3916185029Spjd } else if (page_load > 0 && arc_reclaim_needed()) { 3917185029Spjd /* memory is low, delay before restarting */ 3918185029Spjd ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3919249195Smm return (SET_ERROR(EAGAIN)); 3920185029Spjd } 3921185029Spjd page_load = 0; 3922185029Spjd#endif 3923185029Spjd return (0); 3924185029Spjd} 3925185029Spjd 3926168404Spjdvoid 3927185029Spjdarc_tempreserve_clear(uint64_t reserve) 3928168404Spjd{ 3929185029Spjd atomic_add_64(&arc_tempreserve, -reserve); 3930168404Spjd ASSERT((int64_t)arc_tempreserve >= 0); 3931168404Spjd} 3932168404Spjd 3933168404Spjdint 3934185029Spjdarc_tempreserve_space(uint64_t reserve, uint64_t txg) 3935168404Spjd{ 3936185029Spjd int error; 3937209962Smm uint64_t anon_size; 3938185029Spjd 3939185029Spjd if (reserve > arc_c/4 && !arc_no_grow) 3940185029Spjd arc_c = MIN(arc_c_max, reserve * 4); 3941185029Spjd if (reserve > arc_c) 3942249195Smm return (SET_ERROR(ENOMEM)); 3943168404Spjd 3944168404Spjd /* 3945209962Smm * Don't count loaned bufs as in flight dirty data to prevent long 3946209962Smm * network delays from blocking transactions that are ready to be 3947209962Smm * assigned to a txg. 3948209962Smm */ 3949209962Smm anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 3950209962Smm 3951209962Smm /* 3952185029Spjd * Writes will, almost always, require additional memory allocations 3953251631Sdelphij * in order to compress/encrypt/etc the data. We therefore need to 3954185029Spjd * make sure that there is sufficient available memory for this. 3955185029Spjd */ 3956258632Savg error = arc_memory_throttle(reserve, txg); 3957258632Savg if (error != 0) 3958185029Spjd return (error); 3959185029Spjd 3960185029Spjd /* 3961168404Spjd * Throttle writes when the amount of dirty data in the cache 3962168404Spjd * gets too large. We try to keep the cache less than half full 3963168404Spjd * of dirty blocks so that our sync times don't grow too large. 3964168404Spjd * Note: if two requests come in concurrently, we might let them 3965168404Spjd * both succeed, when one of them should fail. Not a huge deal. 3966168404Spjd */ 3967209962Smm 3968209962Smm if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 3969209962Smm anon_size > arc_c / 4) { 3970185029Spjd dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 3971185029Spjd "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 3972185029Spjd arc_tempreserve>>10, 3973185029Spjd arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 3974185029Spjd arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 3975185029Spjd reserve>>10, arc_c>>10); 3976249195Smm return (SET_ERROR(ERESTART)); 3977168404Spjd } 3978185029Spjd atomic_add_64(&arc_tempreserve, reserve); 3979168404Spjd return (0); 3980168404Spjd} 3981168404Spjd 3982168582Spjdstatic kmutex_t arc_lowmem_lock; 3983168404Spjd#ifdef _KERNEL 3984168566Spjdstatic eventhandler_tag arc_event_lowmem = NULL; 3985168404Spjd 3986168404Spjdstatic void 3987168566Spjdarc_lowmem(void *arg __unused, int howto __unused) 3988168404Spjd{ 3989168404Spjd 3990168566Spjd /* Serialize access via arc_lowmem_lock. */ 3991168566Spjd mutex_enter(&arc_lowmem_lock); 3992219089Spjd mutex_enter(&arc_reclaim_thr_lock); 3993185029Spjd needfree = 1; 3994168404Spjd cv_signal(&arc_reclaim_thr_cv); 3995241773Savg 3996241773Savg /* 3997241773Savg * It is unsafe to block here in arbitrary threads, because we can come 3998241773Savg * here from ARC itself and may hold ARC locks and thus risk a deadlock 3999241773Savg * with ARC reclaim thread. 4000241773Savg */ 4001241773Savg if (curproc == pageproc) { 4002241773Savg while (needfree) 4003241773Savg msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4004241773Savg } 4005219089Spjd mutex_exit(&arc_reclaim_thr_lock); 4006168566Spjd mutex_exit(&arc_lowmem_lock); 4007168404Spjd} 4008168404Spjd#endif 4009168404Spjd 4010168404Spjdvoid 4011168404Spjdarc_init(void) 4012168404Spjd{ 4013219089Spjd int i, prefetch_tunable_set = 0; 4014205231Skmacy 4015168404Spjd mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4016168404Spjd cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4017168566Spjd mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4018168404Spjd 4019168404Spjd /* Convert seconds to clock ticks */ 4020168404Spjd arc_min_prefetch_lifespan = 1 * hz; 4021168404Spjd 4022168404Spjd /* Start out with 1/8 of all memory */ 4023168566Spjd arc_c = kmem_size() / 8; 4024219089Spjd 4025219089Spjd#ifdef sun 4026192360Skmacy#ifdef _KERNEL 4027192360Skmacy /* 4028192360Skmacy * On architectures where the physical memory can be larger 4029192360Skmacy * than the addressable space (intel in 32-bit mode), we may 4030192360Skmacy * need to limit the cache to 1/8 of VM size. 4031192360Skmacy */ 4032192360Skmacy arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4033192360Skmacy#endif 4034219089Spjd#endif /* sun */ 4035168566Spjd /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4036168566Spjd arc_c_min = MAX(arc_c / 4, 64<<18); 4037168566Spjd /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4038168404Spjd if (arc_c * 8 >= 1<<30) 4039168404Spjd arc_c_max = (arc_c * 8) - (1<<30); 4040168404Spjd else 4041168404Spjd arc_c_max = arc_c_min; 4042175633Spjd arc_c_max = MAX(arc_c * 5, arc_c_max); 4043219089Spjd 4044168481Spjd#ifdef _KERNEL 4045168404Spjd /* 4046168404Spjd * Allow the tunables to override our calculations if they are 4047168566Spjd * reasonable (ie. over 16MB) 4048168404Spjd */ 4049219089Spjd if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size()) 4050168404Spjd arc_c_max = zfs_arc_max; 4051219089Spjd if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max) 4052168404Spjd arc_c_min = zfs_arc_min; 4053168481Spjd#endif 4054219089Spjd 4055168404Spjd arc_c = arc_c_max; 4056168404Spjd arc_p = (arc_c >> 1); 4057168404Spjd 4058185029Spjd /* limit meta-data to 1/4 of the arc capacity */ 4059185029Spjd arc_meta_limit = arc_c_max / 4; 4060185029Spjd 4061185029Spjd /* Allow the tunable to override if it is reasonable */ 4062185029Spjd if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4063185029Spjd arc_meta_limit = zfs_arc_meta_limit; 4064185029Spjd 4065185029Spjd if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4066185029Spjd arc_c_min = arc_meta_limit / 2; 4067185029Spjd 4068208373Smm if (zfs_arc_grow_retry > 0) 4069208373Smm arc_grow_retry = zfs_arc_grow_retry; 4070208373Smm 4071208373Smm if (zfs_arc_shrink_shift > 0) 4072208373Smm arc_shrink_shift = zfs_arc_shrink_shift; 4073208373Smm 4074208373Smm if (zfs_arc_p_min_shift > 0) 4075208373Smm arc_p_min_shift = zfs_arc_p_min_shift; 4076208373Smm 4077168404Spjd /* if kmem_flags are set, lets try to use less memory */ 4078168404Spjd if (kmem_debugging()) 4079168404Spjd arc_c = arc_c / 2; 4080168404Spjd if (arc_c < arc_c_min) 4081168404Spjd arc_c = arc_c_min; 4082168404Spjd 4083168473Spjd zfs_arc_min = arc_c_min; 4084168473Spjd zfs_arc_max = arc_c_max; 4085168473Spjd 4086168404Spjd arc_anon = &ARC_anon; 4087168404Spjd arc_mru = &ARC_mru; 4088168404Spjd arc_mru_ghost = &ARC_mru_ghost; 4089168404Spjd arc_mfu = &ARC_mfu; 4090168404Spjd arc_mfu_ghost = &ARC_mfu_ghost; 4091185029Spjd arc_l2c_only = &ARC_l2c_only; 4092168404Spjd arc_size = 0; 4093168404Spjd 4094205231Skmacy for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4095205231Skmacy mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4096205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4097205231Skmacy mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4098205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4099205231Skmacy mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4100205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4101205231Skmacy mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4102205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4103205231Skmacy mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4104205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4105205231Skmacy mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4106205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4107206796Spjd 4108205231Skmacy list_create(&arc_mru->arcs_lists[i], 4109205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4110205231Skmacy list_create(&arc_mru_ghost->arcs_lists[i], 4111205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4112205231Skmacy list_create(&arc_mfu->arcs_lists[i], 4113205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4114205231Skmacy list_create(&arc_mfu_ghost->arcs_lists[i], 4115205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4116205231Skmacy list_create(&arc_mfu_ghost->arcs_lists[i], 4117205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4118205231Skmacy list_create(&arc_l2c_only->arcs_lists[i], 4119205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4120205231Skmacy } 4121168404Spjd 4122168404Spjd buf_init(); 4123168404Spjd 4124168404Spjd arc_thread_exit = 0; 4125168404Spjd arc_eviction_list = NULL; 4126168404Spjd mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4127168404Spjd bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4128168404Spjd 4129168404Spjd arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4130168404Spjd sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4131168404Spjd 4132168404Spjd if (arc_ksp != NULL) { 4133168404Spjd arc_ksp->ks_data = &arc_stats; 4134168404Spjd kstat_install(arc_ksp); 4135168404Spjd } 4136168404Spjd 4137168404Spjd (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4138168404Spjd TS_RUN, minclsyspri); 4139168404Spjd 4140168404Spjd#ifdef _KERNEL 4141168566Spjd arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4142168404Spjd EVENTHANDLER_PRI_FIRST); 4143168404Spjd#endif 4144168404Spjd 4145168404Spjd arc_dead = FALSE; 4146185029Spjd arc_warm = B_FALSE; 4147168566Spjd 4148258632Savg /* 4149258632Savg * Calculate maximum amount of dirty data per pool. 4150258632Savg * 4151258632Savg * If it has been set by /etc/system, take that. 4152258632Savg * Otherwise, use a percentage of physical memory defined by 4153258632Savg * zfs_dirty_data_max_percent (default 10%) with a cap at 4154258632Savg * zfs_dirty_data_max_max (default 4GB). 4155258632Savg */ 4156258632Savg if (zfs_dirty_data_max == 0) { 4157258632Savg zfs_dirty_data_max = ptob(physmem) * 4158258632Savg zfs_dirty_data_max_percent / 100; 4159258632Savg zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4160258632Savg zfs_dirty_data_max_max); 4161258632Savg } 4162185029Spjd 4163168566Spjd#ifdef _KERNEL 4164194043Skmacy if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4165193953Skmacy prefetch_tunable_set = 1; 4166206796Spjd 4167193878Skmacy#ifdef __i386__ 4168193953Skmacy if (prefetch_tunable_set == 0) { 4169196863Strasz printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4170196863Strasz "-- to enable,\n"); 4171196863Strasz printf(" add \"vfs.zfs.prefetch_disable=0\" " 4172196863Strasz "to /boot/loader.conf.\n"); 4173219089Spjd zfs_prefetch_disable = 1; 4174193878Skmacy } 4175206796Spjd#else 4176193878Skmacy if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4177193953Skmacy prefetch_tunable_set == 0) { 4178196863Strasz printf("ZFS NOTICE: Prefetch is disabled by default if less " 4179196941Strasz "than 4GB of RAM is present;\n" 4180196863Strasz " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4181196863Strasz "to /boot/loader.conf.\n"); 4182219089Spjd zfs_prefetch_disable = 1; 4183193878Skmacy } 4184206796Spjd#endif 4185175633Spjd /* Warn about ZFS memory and address space requirements. */ 4186168696Spjd if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4187168987Sbmah printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4188168987Sbmah "expect unstable behavior.\n"); 4189175633Spjd } 4190175633Spjd if (kmem_size() < 512 * (1 << 20)) { 4191173419Spjd printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4192168987Sbmah "expect unstable behavior.\n"); 4193185029Spjd printf(" Consider tuning vm.kmem_size and " 4194173419Spjd "vm.kmem_size_max\n"); 4195185029Spjd printf(" in /boot/loader.conf.\n"); 4196168566Spjd } 4197168566Spjd#endif 4198168404Spjd} 4199168404Spjd 4200168404Spjdvoid 4201168404Spjdarc_fini(void) 4202168404Spjd{ 4203205231Skmacy int i; 4204206796Spjd 4205168404Spjd mutex_enter(&arc_reclaim_thr_lock); 4206168404Spjd arc_thread_exit = 1; 4207168404Spjd cv_signal(&arc_reclaim_thr_cv); 4208168404Spjd while (arc_thread_exit != 0) 4209168404Spjd cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4210168404Spjd mutex_exit(&arc_reclaim_thr_lock); 4211168404Spjd 4212185029Spjd arc_flush(NULL); 4213168404Spjd 4214168404Spjd arc_dead = TRUE; 4215168404Spjd 4216168404Spjd if (arc_ksp != NULL) { 4217168404Spjd kstat_delete(arc_ksp); 4218168404Spjd arc_ksp = NULL; 4219168404Spjd } 4220168404Spjd 4221168404Spjd mutex_destroy(&arc_eviction_mtx); 4222168404Spjd mutex_destroy(&arc_reclaim_thr_lock); 4223168404Spjd cv_destroy(&arc_reclaim_thr_cv); 4224168404Spjd 4225205231Skmacy for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4226205231Skmacy list_destroy(&arc_mru->arcs_lists[i]); 4227205231Skmacy list_destroy(&arc_mru_ghost->arcs_lists[i]); 4228205231Skmacy list_destroy(&arc_mfu->arcs_lists[i]); 4229205231Skmacy list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4230206795Spjd list_destroy(&arc_l2c_only->arcs_lists[i]); 4231168404Spjd 4232205231Skmacy mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4233205231Skmacy mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4234205231Skmacy mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4235205231Skmacy mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4236205231Skmacy mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4237206795Spjd mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4238205231Skmacy } 4239206796Spjd 4240168404Spjd buf_fini(); 4241168404Spjd 4242209962Smm ASSERT(arc_loaned_bytes == 0); 4243209962Smm 4244168582Spjd mutex_destroy(&arc_lowmem_lock); 4245168404Spjd#ifdef _KERNEL 4246168566Spjd if (arc_event_lowmem != NULL) 4247168566Spjd EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4248168404Spjd#endif 4249168404Spjd} 4250185029Spjd 4251185029Spjd/* 4252185029Spjd * Level 2 ARC 4253185029Spjd * 4254185029Spjd * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4255185029Spjd * It uses dedicated storage devices to hold cached data, which are populated 4256185029Spjd * using large infrequent writes. The main role of this cache is to boost 4257185029Spjd * the performance of random read workloads. The intended L2ARC devices 4258185029Spjd * include short-stroked disks, solid state disks, and other media with 4259185029Spjd * substantially faster read latency than disk. 4260185029Spjd * 4261185029Spjd * +-----------------------+ 4262185029Spjd * | ARC | 4263185029Spjd * +-----------------------+ 4264185029Spjd * | ^ ^ 4265185029Spjd * | | | 4266185029Spjd * l2arc_feed_thread() arc_read() 4267185029Spjd * | | | 4268185029Spjd * | l2arc read | 4269185029Spjd * V | | 4270185029Spjd * +---------------+ | 4271185029Spjd * | L2ARC | | 4272185029Spjd * +---------------+ | 4273185029Spjd * | ^ | 4274185029Spjd * l2arc_write() | | 4275185029Spjd * | | | 4276185029Spjd * V | | 4277185029Spjd * +-------+ +-------+ 4278185029Spjd * | vdev | | vdev | 4279185029Spjd * | cache | | cache | 4280185029Spjd * +-------+ +-------+ 4281185029Spjd * +=========+ .-----. 4282185029Spjd * : L2ARC : |-_____-| 4283185029Spjd * : devices : | Disks | 4284185029Spjd * +=========+ `-_____-' 4285185029Spjd * 4286185029Spjd * Read requests are satisfied from the following sources, in order: 4287185029Spjd * 4288185029Spjd * 1) ARC 4289185029Spjd * 2) vdev cache of L2ARC devices 4290185029Spjd * 3) L2ARC devices 4291185029Spjd * 4) vdev cache of disks 4292185029Spjd * 5) disks 4293185029Spjd * 4294185029Spjd * Some L2ARC device types exhibit extremely slow write performance. 4295185029Spjd * To accommodate for this there are some significant differences between 4296185029Spjd * the L2ARC and traditional cache design: 4297185029Spjd * 4298185029Spjd * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4299185029Spjd * the ARC behave as usual, freeing buffers and placing headers on ghost 4300185029Spjd * lists. The ARC does not send buffers to the L2ARC during eviction as 4301185029Spjd * this would add inflated write latencies for all ARC memory pressure. 4302185029Spjd * 4303185029Spjd * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4304185029Spjd * It does this by periodically scanning buffers from the eviction-end of 4305185029Spjd * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4306251478Sdelphij * not already there. It scans until a headroom of buffers is satisfied, 4307251478Sdelphij * which itself is a buffer for ARC eviction. If a compressible buffer is 4308251478Sdelphij * found during scanning and selected for writing to an L2ARC device, we 4309251478Sdelphij * temporarily boost scanning headroom during the next scan cycle to make 4310251478Sdelphij * sure we adapt to compression effects (which might significantly reduce 4311251478Sdelphij * the data volume we write to L2ARC). The thread that does this is 4312185029Spjd * l2arc_feed_thread(), illustrated below; example sizes are included to 4313185029Spjd * provide a better sense of ratio than this diagram: 4314185029Spjd * 4315185029Spjd * head --> tail 4316185029Spjd * +---------------------+----------+ 4317185029Spjd * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4318185029Spjd * +---------------------+----------+ | o L2ARC eligible 4319185029Spjd * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4320185029Spjd * +---------------------+----------+ | 4321185029Spjd * 15.9 Gbytes ^ 32 Mbytes | 4322185029Spjd * headroom | 4323185029Spjd * l2arc_feed_thread() 4324185029Spjd * | 4325185029Spjd * l2arc write hand <--[oooo]--' 4326185029Spjd * | 8 Mbyte 4327185029Spjd * | write max 4328185029Spjd * V 4329185029Spjd * +==============================+ 4330185029Spjd * L2ARC dev |####|#|###|###| |####| ... | 4331185029Spjd * +==============================+ 4332185029Spjd * 32 Gbytes 4333185029Spjd * 4334185029Spjd * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4335185029Spjd * evicted, then the L2ARC has cached a buffer much sooner than it probably 4336185029Spjd * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4337185029Spjd * safe to say that this is an uncommon case, since buffers at the end of 4338185029Spjd * the ARC lists have moved there due to inactivity. 4339185029Spjd * 4340185029Spjd * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4341185029Spjd * then the L2ARC simply misses copying some buffers. This serves as a 4342185029Spjd * pressure valve to prevent heavy read workloads from both stalling the ARC 4343185029Spjd * with waits and clogging the L2ARC with writes. This also helps prevent 4344185029Spjd * the potential for the L2ARC to churn if it attempts to cache content too 4345185029Spjd * quickly, such as during backups of the entire pool. 4346185029Spjd * 4347185029Spjd * 5. After system boot and before the ARC has filled main memory, there are 4348185029Spjd * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4349185029Spjd * lists can remain mostly static. Instead of searching from tail of these 4350185029Spjd * lists as pictured, the l2arc_feed_thread() will search from the list heads 4351185029Spjd * for eligible buffers, greatly increasing its chance of finding them. 4352185029Spjd * 4353185029Spjd * The L2ARC device write speed is also boosted during this time so that 4354185029Spjd * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4355185029Spjd * there are no L2ARC reads, and no fear of degrading read performance 4356185029Spjd * through increased writes. 4357185029Spjd * 4358185029Spjd * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4359185029Spjd * the vdev queue can aggregate them into larger and fewer writes. Each 4360185029Spjd * device is written to in a rotor fashion, sweeping writes through 4361185029Spjd * available space then repeating. 4362185029Spjd * 4363185029Spjd * 7. The L2ARC does not store dirty content. It never needs to flush 4364185029Spjd * write buffers back to disk based storage. 4365185029Spjd * 4366185029Spjd * 8. If an ARC buffer is written (and dirtied) which also exists in the 4367185029Spjd * L2ARC, the now stale L2ARC buffer is immediately dropped. 4368185029Spjd * 4369185029Spjd * The performance of the L2ARC can be tweaked by a number of tunables, which 4370185029Spjd * may be necessary for different workloads: 4371185029Spjd * 4372185029Spjd * l2arc_write_max max write bytes per interval 4373185029Spjd * l2arc_write_boost extra write bytes during device warmup 4374185029Spjd * l2arc_noprefetch skip caching prefetched buffers 4375185029Spjd * l2arc_headroom number of max device writes to precache 4376251478Sdelphij * l2arc_headroom_boost when we find compressed buffers during ARC 4377251478Sdelphij * scanning, we multiply headroom by this 4378251478Sdelphij * percentage factor for the next scan cycle, 4379251478Sdelphij * since more compressed buffers are likely to 4380251478Sdelphij * be present 4381185029Spjd * l2arc_feed_secs seconds between L2ARC writing 4382185029Spjd * 4383185029Spjd * Tunables may be removed or added as future performance improvements are 4384185029Spjd * integrated, and also may become zpool properties. 4385208373Smm * 4386208373Smm * There are three key functions that control how the L2ARC warms up: 4387208373Smm * 4388208373Smm * l2arc_write_eligible() check if a buffer is eligible to cache 4389208373Smm * l2arc_write_size() calculate how much to write 4390208373Smm * l2arc_write_interval() calculate sleep delay between writes 4391208373Smm * 4392208373Smm * These three functions determine what to write, how much, and how quickly 4393208373Smm * to send writes. 4394185029Spjd */ 4395185029Spjd 4396208373Smmstatic boolean_t 4397209962Smml2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab) 4398208373Smm{ 4399208373Smm /* 4400208373Smm * A buffer is *not* eligible for the L2ARC if it: 4401208373Smm * 1. belongs to a different spa. 4402208373Smm * 2. is already cached on the L2ARC. 4403208373Smm * 3. has an I/O in progress (it may be an incomplete read). 4404208373Smm * 4. is flagged not eligible (zfs property). 4405208373Smm */ 4406209962Smm if (ab->b_spa != spa_guid) { 4407208373Smm ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4408208373Smm return (B_FALSE); 4409208373Smm } 4410208373Smm if (ab->b_l2hdr != NULL) { 4411208373Smm ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4412208373Smm return (B_FALSE); 4413208373Smm } 4414208373Smm if (HDR_IO_IN_PROGRESS(ab)) { 4415208373Smm ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4416208373Smm return (B_FALSE); 4417208373Smm } 4418208373Smm if (!HDR_L2CACHE(ab)) { 4419208373Smm ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4420208373Smm return (B_FALSE); 4421208373Smm } 4422208373Smm 4423208373Smm return (B_TRUE); 4424208373Smm} 4425208373Smm 4426208373Smmstatic uint64_t 4427251478Sdelphijl2arc_write_size(void) 4428208373Smm{ 4429208373Smm uint64_t size; 4430208373Smm 4431251478Sdelphij /* 4432251478Sdelphij * Make sure our globals have meaningful values in case the user 4433251478Sdelphij * altered them. 4434251478Sdelphij */ 4435251478Sdelphij size = l2arc_write_max; 4436251478Sdelphij if (size == 0) { 4437251478Sdelphij cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4438251478Sdelphij "be greater than zero, resetting it to the default (%d)", 4439251478Sdelphij L2ARC_WRITE_SIZE); 4440251478Sdelphij size = l2arc_write_max = L2ARC_WRITE_SIZE; 4441251478Sdelphij } 4442208373Smm 4443208373Smm if (arc_warm == B_FALSE) 4444251478Sdelphij size += l2arc_write_boost; 4445208373Smm 4446208373Smm return (size); 4447208373Smm 4448208373Smm} 4449208373Smm 4450208373Smmstatic clock_t 4451208373Smml2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4452208373Smm{ 4453219089Spjd clock_t interval, next, now; 4454208373Smm 4455208373Smm /* 4456208373Smm * If the ARC lists are busy, increase our write rate; if the 4457208373Smm * lists are stale, idle back. This is achieved by checking 4458208373Smm * how much we previously wrote - if it was more than half of 4459208373Smm * what we wanted, schedule the next write much sooner. 4460208373Smm */ 4461208373Smm if (l2arc_feed_again && wrote > (wanted / 2)) 4462208373Smm interval = (hz * l2arc_feed_min_ms) / 1000; 4463208373Smm else 4464208373Smm interval = hz * l2arc_feed_secs; 4465208373Smm 4466219089Spjd now = ddi_get_lbolt(); 4467219089Spjd next = MAX(now, MIN(now + interval, began + interval)); 4468208373Smm 4469208373Smm return (next); 4470208373Smm} 4471208373Smm 4472185029Spjdstatic void 4473185029Spjdl2arc_hdr_stat_add(void) 4474185029Spjd{ 4475185029Spjd ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4476185029Spjd ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4477185029Spjd} 4478185029Spjd 4479185029Spjdstatic void 4480185029Spjdl2arc_hdr_stat_remove(void) 4481185029Spjd{ 4482185029Spjd ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4483185029Spjd ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4484185029Spjd} 4485185029Spjd 4486185029Spjd/* 4487185029Spjd * Cycle through L2ARC devices. This is how L2ARC load balances. 4488185029Spjd * If a device is returned, this also returns holding the spa config lock. 4489185029Spjd */ 4490185029Spjdstatic l2arc_dev_t * 4491185029Spjdl2arc_dev_get_next(void) 4492185029Spjd{ 4493185029Spjd l2arc_dev_t *first, *next = NULL; 4494185029Spjd 4495185029Spjd /* 4496185029Spjd * Lock out the removal of spas (spa_namespace_lock), then removal 4497185029Spjd * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4498185029Spjd * both locks will be dropped and a spa config lock held instead. 4499185029Spjd */ 4500185029Spjd mutex_enter(&spa_namespace_lock); 4501185029Spjd mutex_enter(&l2arc_dev_mtx); 4502185029Spjd 4503185029Spjd /* if there are no vdevs, there is nothing to do */ 4504185029Spjd if (l2arc_ndev == 0) 4505185029Spjd goto out; 4506185029Spjd 4507185029Spjd first = NULL; 4508185029Spjd next = l2arc_dev_last; 4509185029Spjd do { 4510185029Spjd /* loop around the list looking for a non-faulted vdev */ 4511185029Spjd if (next == NULL) { 4512185029Spjd next = list_head(l2arc_dev_list); 4513185029Spjd } else { 4514185029Spjd next = list_next(l2arc_dev_list, next); 4515185029Spjd if (next == NULL) 4516185029Spjd next = list_head(l2arc_dev_list); 4517185029Spjd } 4518185029Spjd 4519185029Spjd /* if we have come back to the start, bail out */ 4520185029Spjd if (first == NULL) 4521185029Spjd first = next; 4522185029Spjd else if (next == first) 4523185029Spjd break; 4524185029Spjd 4525185029Spjd } while (vdev_is_dead(next->l2ad_vdev)); 4526185029Spjd 4527185029Spjd /* if we were unable to find any usable vdevs, return NULL */ 4528185029Spjd if (vdev_is_dead(next->l2ad_vdev)) 4529185029Spjd next = NULL; 4530185029Spjd 4531185029Spjd l2arc_dev_last = next; 4532185029Spjd 4533185029Spjdout: 4534185029Spjd mutex_exit(&l2arc_dev_mtx); 4535185029Spjd 4536185029Spjd /* 4537185029Spjd * Grab the config lock to prevent the 'next' device from being 4538185029Spjd * removed while we are writing to it. 4539185029Spjd */ 4540185029Spjd if (next != NULL) 4541185029Spjd spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4542185029Spjd mutex_exit(&spa_namespace_lock); 4543185029Spjd 4544185029Spjd return (next); 4545185029Spjd} 4546185029Spjd 4547185029Spjd/* 4548185029Spjd * Free buffers that were tagged for destruction. 4549185029Spjd */ 4550185029Spjdstatic void 4551185029Spjdl2arc_do_free_on_write() 4552185029Spjd{ 4553185029Spjd list_t *buflist; 4554185029Spjd l2arc_data_free_t *df, *df_prev; 4555185029Spjd 4556185029Spjd mutex_enter(&l2arc_free_on_write_mtx); 4557185029Spjd buflist = l2arc_free_on_write; 4558185029Spjd 4559185029Spjd for (df = list_tail(buflist); df; df = df_prev) { 4560185029Spjd df_prev = list_prev(buflist, df); 4561185029Spjd ASSERT(df->l2df_data != NULL); 4562185029Spjd ASSERT(df->l2df_func != NULL); 4563185029Spjd df->l2df_func(df->l2df_data, df->l2df_size); 4564185029Spjd list_remove(buflist, df); 4565185029Spjd kmem_free(df, sizeof (l2arc_data_free_t)); 4566185029Spjd } 4567185029Spjd 4568185029Spjd mutex_exit(&l2arc_free_on_write_mtx); 4569185029Spjd} 4570185029Spjd 4571185029Spjd/* 4572185029Spjd * A write to a cache device has completed. Update all headers to allow 4573185029Spjd * reads from these buffers to begin. 4574185029Spjd */ 4575185029Spjdstatic void 4576185029Spjdl2arc_write_done(zio_t *zio) 4577185029Spjd{ 4578185029Spjd l2arc_write_callback_t *cb; 4579185029Spjd l2arc_dev_t *dev; 4580185029Spjd list_t *buflist; 4581185029Spjd arc_buf_hdr_t *head, *ab, *ab_prev; 4582185029Spjd l2arc_buf_hdr_t *abl2; 4583185029Spjd kmutex_t *hash_lock; 4584268085Sdelphij int64_t bytes_dropped = 0; 4585185029Spjd 4586185029Spjd cb = zio->io_private; 4587185029Spjd ASSERT(cb != NULL); 4588185029Spjd dev = cb->l2wcb_dev; 4589185029Spjd ASSERT(dev != NULL); 4590185029Spjd head = cb->l2wcb_head; 4591185029Spjd ASSERT(head != NULL); 4592185029Spjd buflist = dev->l2ad_buflist; 4593185029Spjd ASSERT(buflist != NULL); 4594185029Spjd DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4595185029Spjd l2arc_write_callback_t *, cb); 4596185029Spjd 4597185029Spjd if (zio->io_error != 0) 4598185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_error); 4599185029Spjd 4600185029Spjd mutex_enter(&l2arc_buflist_mtx); 4601185029Spjd 4602185029Spjd /* 4603185029Spjd * All writes completed, or an error was hit. 4604185029Spjd */ 4605185029Spjd for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 4606185029Spjd ab_prev = list_prev(buflist, ab); 4607260835Sdelphij abl2 = ab->b_l2hdr; 4608185029Spjd 4609260835Sdelphij /* 4610260835Sdelphij * Release the temporary compressed buffer as soon as possible. 4611260835Sdelphij */ 4612260835Sdelphij if (abl2->b_compress != ZIO_COMPRESS_OFF) 4613260835Sdelphij l2arc_release_cdata_buf(ab); 4614260835Sdelphij 4615185029Spjd hash_lock = HDR_LOCK(ab); 4616185029Spjd if (!mutex_tryenter(hash_lock)) { 4617185029Spjd /* 4618185029Spjd * This buffer misses out. It may be in a stage 4619185029Spjd * of eviction. Its ARC_L2_WRITING flag will be 4620185029Spjd * left set, denying reads to this buffer. 4621185029Spjd */ 4622185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4623185029Spjd continue; 4624185029Spjd } 4625185029Spjd 4626185029Spjd if (zio->io_error != 0) { 4627185029Spjd /* 4628185029Spjd * Error - drop L2ARC entry. 4629185029Spjd */ 4630185029Spjd list_remove(buflist, ab); 4631251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4632268085Sdelphij bytes_dropped += abl2->b_asize; 4633185029Spjd ab->b_l2hdr = NULL; 4634248572Ssmh trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4635248574Ssmh ab->b_size, 0); 4636185029Spjd kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4637185029Spjd ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4638185029Spjd } 4639185029Spjd 4640185029Spjd /* 4641185029Spjd * Allow ARC to begin reads to this L2ARC entry. 4642185029Spjd */ 4643185029Spjd ab->b_flags &= ~ARC_L2_WRITING; 4644185029Spjd 4645185029Spjd mutex_exit(hash_lock); 4646185029Spjd } 4647185029Spjd 4648185029Spjd atomic_inc_64(&l2arc_writes_done); 4649185029Spjd list_remove(buflist, head); 4650185029Spjd kmem_cache_free(hdr_cache, head); 4651185029Spjd mutex_exit(&l2arc_buflist_mtx); 4652185029Spjd 4653268085Sdelphij vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 4654268085Sdelphij 4655185029Spjd l2arc_do_free_on_write(); 4656185029Spjd 4657185029Spjd kmem_free(cb, sizeof (l2arc_write_callback_t)); 4658185029Spjd} 4659185029Spjd 4660185029Spjd/* 4661185029Spjd * A read to a cache device completed. Validate buffer contents before 4662185029Spjd * handing over to the regular ARC routines. 4663185029Spjd */ 4664185029Spjdstatic void 4665185029Spjdl2arc_read_done(zio_t *zio) 4666185029Spjd{ 4667185029Spjd l2arc_read_callback_t *cb; 4668185029Spjd arc_buf_hdr_t *hdr; 4669185029Spjd arc_buf_t *buf; 4670185029Spjd kmutex_t *hash_lock; 4671185029Spjd int equal; 4672185029Spjd 4673185029Spjd ASSERT(zio->io_vd != NULL); 4674185029Spjd ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4675185029Spjd 4676185029Spjd spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4677185029Spjd 4678185029Spjd cb = zio->io_private; 4679185029Spjd ASSERT(cb != NULL); 4680185029Spjd buf = cb->l2rcb_buf; 4681185029Spjd ASSERT(buf != NULL); 4682185029Spjd 4683219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 4684185029Spjd mutex_enter(hash_lock); 4685219089Spjd hdr = buf->b_hdr; 4686219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4687185029Spjd 4688185029Spjd /* 4689251478Sdelphij * If the buffer was compressed, decompress it first. 4690251478Sdelphij */ 4691251478Sdelphij if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4692251478Sdelphij l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4693251478Sdelphij ASSERT(zio->io_data != NULL); 4694251478Sdelphij 4695251478Sdelphij /* 4696185029Spjd * Check this survived the L2ARC journey. 4697185029Spjd */ 4698185029Spjd equal = arc_cksum_equal(buf); 4699185029Spjd if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4700185029Spjd mutex_exit(hash_lock); 4701185029Spjd zio->io_private = buf; 4702185029Spjd zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4703185029Spjd zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4704185029Spjd arc_read_done(zio); 4705185029Spjd } else { 4706185029Spjd mutex_exit(hash_lock); 4707185029Spjd /* 4708185029Spjd * Buffer didn't survive caching. Increment stats and 4709185029Spjd * reissue to the original storage device. 4710185029Spjd */ 4711185029Spjd if (zio->io_error != 0) { 4712185029Spjd ARCSTAT_BUMP(arcstat_l2_io_error); 4713185029Spjd } else { 4714249195Smm zio->io_error = SET_ERROR(EIO); 4715185029Spjd } 4716185029Spjd if (!equal) 4717185029Spjd ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4718185029Spjd 4719185029Spjd /* 4720185029Spjd * If there's no waiter, issue an async i/o to the primary 4721185029Spjd * storage now. If there *is* a waiter, the caller must 4722185029Spjd * issue the i/o in a context where it's OK to block. 4723185029Spjd */ 4724209962Smm if (zio->io_waiter == NULL) { 4725209962Smm zio_t *pio = zio_unique_parent(zio); 4726209962Smm 4727209962Smm ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4728209962Smm 4729209962Smm zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4730185029Spjd buf->b_data, zio->io_size, arc_read_done, buf, 4731185029Spjd zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4732209962Smm } 4733185029Spjd } 4734185029Spjd 4735185029Spjd kmem_free(cb, sizeof (l2arc_read_callback_t)); 4736185029Spjd} 4737185029Spjd 4738185029Spjd/* 4739185029Spjd * This is the list priority from which the L2ARC will search for pages to 4740185029Spjd * cache. This is used within loops (0..3) to cycle through lists in the 4741185029Spjd * desired order. This order can have a significant effect on cache 4742185029Spjd * performance. 4743185029Spjd * 4744185029Spjd * Currently the metadata lists are hit first, MFU then MRU, followed by 4745185029Spjd * the data lists. This function returns a locked list, and also returns 4746185029Spjd * the lock pointer. 4747185029Spjd */ 4748185029Spjdstatic list_t * 4749185029Spjdl2arc_list_locked(int list_num, kmutex_t **lock) 4750185029Spjd{ 4751247187Smm list_t *list = NULL; 4752205231Skmacy int idx; 4753185029Spjd 4754206796Spjd ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4755206796Spjd 4756205231Skmacy if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4757205231Skmacy idx = list_num; 4758205231Skmacy list = &arc_mfu->arcs_lists[idx]; 4759205231Skmacy *lock = ARCS_LOCK(arc_mfu, idx); 4760206796Spjd } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4761205231Skmacy idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4762205231Skmacy list = &arc_mru->arcs_lists[idx]; 4763205231Skmacy *lock = ARCS_LOCK(arc_mru, idx); 4764206796Spjd } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4765205231Skmacy ARC_BUFC_NUMDATALISTS)) { 4766205231Skmacy idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4767205231Skmacy list = &arc_mfu->arcs_lists[idx]; 4768205231Skmacy *lock = ARCS_LOCK(arc_mfu, idx); 4769205231Skmacy } else { 4770205231Skmacy idx = list_num - ARC_BUFC_NUMLISTS; 4771205231Skmacy list = &arc_mru->arcs_lists[idx]; 4772205231Skmacy *lock = ARCS_LOCK(arc_mru, idx); 4773185029Spjd } 4774185029Spjd 4775185029Spjd ASSERT(!(MUTEX_HELD(*lock))); 4776185029Spjd mutex_enter(*lock); 4777185029Spjd return (list); 4778185029Spjd} 4779185029Spjd 4780185029Spjd/* 4781185029Spjd * Evict buffers from the device write hand to the distance specified in 4782185029Spjd * bytes. This distance may span populated buffers, it may span nothing. 4783185029Spjd * This is clearing a region on the L2ARC device ready for writing. 4784185029Spjd * If the 'all' boolean is set, every buffer is evicted. 4785185029Spjd */ 4786185029Spjdstatic void 4787185029Spjdl2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4788185029Spjd{ 4789185029Spjd list_t *buflist; 4790185029Spjd l2arc_buf_hdr_t *abl2; 4791185029Spjd arc_buf_hdr_t *ab, *ab_prev; 4792185029Spjd kmutex_t *hash_lock; 4793185029Spjd uint64_t taddr; 4794268085Sdelphij int64_t bytes_evicted = 0; 4795185029Spjd 4796185029Spjd buflist = dev->l2ad_buflist; 4797185029Spjd 4798185029Spjd if (buflist == NULL) 4799185029Spjd return; 4800185029Spjd 4801185029Spjd if (!all && dev->l2ad_first) { 4802185029Spjd /* 4803185029Spjd * This is the first sweep through the device. There is 4804185029Spjd * nothing to evict. 4805185029Spjd */ 4806185029Spjd return; 4807185029Spjd } 4808185029Spjd 4809185029Spjd if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4810185029Spjd /* 4811185029Spjd * When nearing the end of the device, evict to the end 4812185029Spjd * before the device write hand jumps to the start. 4813185029Spjd */ 4814185029Spjd taddr = dev->l2ad_end; 4815185029Spjd } else { 4816185029Spjd taddr = dev->l2ad_hand + distance; 4817185029Spjd } 4818185029Spjd DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4819185029Spjd uint64_t, taddr, boolean_t, all); 4820185029Spjd 4821185029Spjdtop: 4822185029Spjd mutex_enter(&l2arc_buflist_mtx); 4823185029Spjd for (ab = list_tail(buflist); ab; ab = ab_prev) { 4824185029Spjd ab_prev = list_prev(buflist, ab); 4825185029Spjd 4826185029Spjd hash_lock = HDR_LOCK(ab); 4827185029Spjd if (!mutex_tryenter(hash_lock)) { 4828185029Spjd /* 4829185029Spjd * Missed the hash lock. Retry. 4830185029Spjd */ 4831185029Spjd ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4832185029Spjd mutex_exit(&l2arc_buflist_mtx); 4833185029Spjd mutex_enter(hash_lock); 4834185029Spjd mutex_exit(hash_lock); 4835185029Spjd goto top; 4836185029Spjd } 4837185029Spjd 4838185029Spjd if (HDR_L2_WRITE_HEAD(ab)) { 4839185029Spjd /* 4840185029Spjd * We hit a write head node. Leave it for 4841185029Spjd * l2arc_write_done(). 4842185029Spjd */ 4843185029Spjd list_remove(buflist, ab); 4844185029Spjd mutex_exit(hash_lock); 4845185029Spjd continue; 4846185029Spjd } 4847185029Spjd 4848185029Spjd if (!all && ab->b_l2hdr != NULL && 4849185029Spjd (ab->b_l2hdr->b_daddr > taddr || 4850185029Spjd ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4851185029Spjd /* 4852185029Spjd * We've evicted to the target address, 4853185029Spjd * or the end of the device. 4854185029Spjd */ 4855185029Spjd mutex_exit(hash_lock); 4856185029Spjd break; 4857185029Spjd } 4858185029Spjd 4859185029Spjd if (HDR_FREE_IN_PROGRESS(ab)) { 4860185029Spjd /* 4861185029Spjd * Already on the path to destruction. 4862185029Spjd */ 4863185029Spjd mutex_exit(hash_lock); 4864185029Spjd continue; 4865185029Spjd } 4866185029Spjd 4867185029Spjd if (ab->b_state == arc_l2c_only) { 4868185029Spjd ASSERT(!HDR_L2_READING(ab)); 4869185029Spjd /* 4870185029Spjd * This doesn't exist in the ARC. Destroy. 4871185029Spjd * arc_hdr_destroy() will call list_remove() 4872185029Spjd * and decrement arcstat_l2_size. 4873185029Spjd */ 4874185029Spjd arc_change_state(arc_anon, ab, hash_lock); 4875185029Spjd arc_hdr_destroy(ab); 4876185029Spjd } else { 4877185029Spjd /* 4878185029Spjd * Invalidate issued or about to be issued 4879185029Spjd * reads, since we may be about to write 4880185029Spjd * over this location. 4881185029Spjd */ 4882185029Spjd if (HDR_L2_READING(ab)) { 4883185029Spjd ARCSTAT_BUMP(arcstat_l2_evict_reading); 4884185029Spjd ab->b_flags |= ARC_L2_EVICTED; 4885185029Spjd } 4886185029Spjd 4887185029Spjd /* 4888185029Spjd * Tell ARC this no longer exists in L2ARC. 4889185029Spjd */ 4890185029Spjd if (ab->b_l2hdr != NULL) { 4891185029Spjd abl2 = ab->b_l2hdr; 4892251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4893268085Sdelphij bytes_evicted += abl2->b_asize; 4894185029Spjd ab->b_l2hdr = NULL; 4895185029Spjd kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4896185029Spjd ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4897185029Spjd } 4898185029Spjd list_remove(buflist, ab); 4899185029Spjd 4900185029Spjd /* 4901185029Spjd * This may have been leftover after a 4902185029Spjd * failed write. 4903185029Spjd */ 4904185029Spjd ab->b_flags &= ~ARC_L2_WRITING; 4905185029Spjd } 4906185029Spjd mutex_exit(hash_lock); 4907185029Spjd } 4908185029Spjd mutex_exit(&l2arc_buflist_mtx); 4909185029Spjd 4910268085Sdelphij vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); 4911185029Spjd dev->l2ad_evict = taddr; 4912185029Spjd} 4913185029Spjd 4914185029Spjd/* 4915185029Spjd * Find and write ARC buffers to the L2ARC device. 4916185029Spjd * 4917185029Spjd * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 4918185029Spjd * for reading until they have completed writing. 4919251478Sdelphij * The headroom_boost is an in-out parameter used to maintain headroom boost 4920251478Sdelphij * state between calls to this function. 4921251478Sdelphij * 4922251478Sdelphij * Returns the number of bytes actually written (which may be smaller than 4923251478Sdelphij * the delta by which the device hand has changed due to alignment). 4924185029Spjd */ 4925208373Smmstatic uint64_t 4926251478Sdelphijl2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 4927251478Sdelphij boolean_t *headroom_boost) 4928185029Spjd{ 4929185029Spjd arc_buf_hdr_t *ab, *ab_prev, *head; 4930185029Spjd list_t *list; 4931251478Sdelphij uint64_t write_asize, write_psize, write_sz, headroom, 4932251478Sdelphij buf_compress_minsz; 4933185029Spjd void *buf_data; 4934251478Sdelphij kmutex_t *list_lock; 4935251478Sdelphij boolean_t full; 4936185029Spjd l2arc_write_callback_t *cb; 4937185029Spjd zio_t *pio, *wzio; 4938228103Smm uint64_t guid = spa_load_guid(spa); 4939251478Sdelphij const boolean_t do_headroom_boost = *headroom_boost; 4940185029Spjd int try; 4941185029Spjd 4942185029Spjd ASSERT(dev->l2ad_vdev != NULL); 4943185029Spjd 4944251478Sdelphij /* Lower the flag now, we might want to raise it again later. */ 4945251478Sdelphij *headroom_boost = B_FALSE; 4946251478Sdelphij 4947185029Spjd pio = NULL; 4948251478Sdelphij write_sz = write_asize = write_psize = 0; 4949185029Spjd full = B_FALSE; 4950185029Spjd head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 4951185029Spjd head->b_flags |= ARC_L2_WRITE_HEAD; 4952185029Spjd 4953205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 4954185029Spjd /* 4955251478Sdelphij * We will want to try to compress buffers that are at least 2x the 4956251478Sdelphij * device sector size. 4957251478Sdelphij */ 4958251478Sdelphij buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 4959251478Sdelphij 4960251478Sdelphij /* 4961185029Spjd * Copy buffers for L2ARC writing. 4962185029Spjd */ 4963185029Spjd mutex_enter(&l2arc_buflist_mtx); 4964206796Spjd for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 4965251478Sdelphij uint64_t passed_sz = 0; 4966251478Sdelphij 4967185029Spjd list = l2arc_list_locked(try, &list_lock); 4968205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 4969185029Spjd 4970185029Spjd /* 4971185029Spjd * L2ARC fast warmup. 4972185029Spjd * 4973185029Spjd * Until the ARC is warm and starts to evict, read from the 4974185029Spjd * head of the ARC lists rather than the tail. 4975185029Spjd */ 4976185029Spjd if (arc_warm == B_FALSE) 4977185029Spjd ab = list_head(list); 4978185029Spjd else 4979185029Spjd ab = list_tail(list); 4980206796Spjd if (ab == NULL) 4981205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 4982185029Spjd 4983251478Sdelphij headroom = target_sz * l2arc_headroom; 4984251478Sdelphij if (do_headroom_boost) 4985251478Sdelphij headroom = (headroom * l2arc_headroom_boost) / 100; 4986251478Sdelphij 4987185029Spjd for (; ab; ab = ab_prev) { 4988251478Sdelphij l2arc_buf_hdr_t *l2hdr; 4989251478Sdelphij kmutex_t *hash_lock; 4990251478Sdelphij uint64_t buf_sz; 4991251478Sdelphij 4992185029Spjd if (arc_warm == B_FALSE) 4993185029Spjd ab_prev = list_next(list, ab); 4994185029Spjd else 4995185029Spjd ab_prev = list_prev(list, ab); 4996205231Skmacy ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size); 4997206796Spjd 4998185029Spjd hash_lock = HDR_LOCK(ab); 4999251478Sdelphij if (!mutex_tryenter(hash_lock)) { 5000205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5001185029Spjd /* 5002185029Spjd * Skip this buffer rather than waiting. 5003185029Spjd */ 5004185029Spjd continue; 5005185029Spjd } 5006185029Spjd 5007185029Spjd passed_sz += ab->b_size; 5008185029Spjd if (passed_sz > headroom) { 5009185029Spjd /* 5010185029Spjd * Searched too far. 5011185029Spjd */ 5012185029Spjd mutex_exit(hash_lock); 5013205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5014185029Spjd break; 5015185029Spjd } 5016185029Spjd 5017209962Smm if (!l2arc_write_eligible(guid, ab)) { 5018185029Spjd mutex_exit(hash_lock); 5019185029Spjd continue; 5020185029Spjd } 5021185029Spjd 5022185029Spjd if ((write_sz + ab->b_size) > target_sz) { 5023185029Spjd full = B_TRUE; 5024185029Spjd mutex_exit(hash_lock); 5025205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_full); 5026185029Spjd break; 5027185029Spjd } 5028185029Spjd 5029185029Spjd if (pio == NULL) { 5030185029Spjd /* 5031185029Spjd * Insert a dummy header on the buflist so 5032185029Spjd * l2arc_write_done() can find where the 5033185029Spjd * write buffers begin without searching. 5034185029Spjd */ 5035185029Spjd list_insert_head(dev->l2ad_buflist, head); 5036185029Spjd 5037185029Spjd cb = kmem_alloc( 5038185029Spjd sizeof (l2arc_write_callback_t), KM_SLEEP); 5039185029Spjd cb->l2wcb_dev = dev; 5040185029Spjd cb->l2wcb_head = head; 5041185029Spjd pio = zio_root(spa, l2arc_write_done, cb, 5042185029Spjd ZIO_FLAG_CANFAIL); 5043205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_pios); 5044185029Spjd } 5045185029Spjd 5046185029Spjd /* 5047185029Spjd * Create and add a new L2ARC header. 5048185029Spjd */ 5049251478Sdelphij l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5050251478Sdelphij l2hdr->b_dev = dev; 5051251478Sdelphij ab->b_flags |= ARC_L2_WRITING; 5052185029Spjd 5053251478Sdelphij /* 5054251478Sdelphij * Temporarily stash the data buffer in b_tmp_cdata. 5055251478Sdelphij * The subsequent write step will pick it up from 5056251478Sdelphij * there. This is because can't access ab->b_buf 5057251478Sdelphij * without holding the hash_lock, which we in turn 5058251478Sdelphij * can't access without holding the ARC list locks 5059251478Sdelphij * (which we want to avoid during compression/writing). 5060251478Sdelphij */ 5061251478Sdelphij l2hdr->b_compress = ZIO_COMPRESS_OFF; 5062251478Sdelphij l2hdr->b_asize = ab->b_size; 5063251478Sdelphij l2hdr->b_tmp_cdata = ab->b_buf->b_data; 5064251478Sdelphij 5065185029Spjd buf_sz = ab->b_size; 5066251478Sdelphij ab->b_l2hdr = l2hdr; 5067185029Spjd 5068251478Sdelphij list_insert_head(dev->l2ad_buflist, ab); 5069251478Sdelphij 5070185029Spjd /* 5071185029Spjd * Compute and store the buffer cksum before 5072185029Spjd * writing. On debug the cksum is verified first. 5073185029Spjd */ 5074185029Spjd arc_cksum_verify(ab->b_buf); 5075185029Spjd arc_cksum_compute(ab->b_buf, B_TRUE); 5076185029Spjd 5077185029Spjd mutex_exit(hash_lock); 5078185029Spjd 5079251478Sdelphij write_sz += buf_sz; 5080251478Sdelphij } 5081251478Sdelphij 5082251478Sdelphij mutex_exit(list_lock); 5083251478Sdelphij 5084251478Sdelphij if (full == B_TRUE) 5085251478Sdelphij break; 5086251478Sdelphij } 5087251478Sdelphij 5088251478Sdelphij /* No buffers selected for writing? */ 5089251478Sdelphij if (pio == NULL) { 5090251478Sdelphij ASSERT0(write_sz); 5091251478Sdelphij mutex_exit(&l2arc_buflist_mtx); 5092251478Sdelphij kmem_cache_free(hdr_cache, head); 5093251478Sdelphij return (0); 5094251478Sdelphij } 5095251478Sdelphij 5096251478Sdelphij /* 5097251478Sdelphij * Now start writing the buffers. We're starting at the write head 5098251478Sdelphij * and work backwards, retracing the course of the buffer selector 5099251478Sdelphij * loop above. 5100251478Sdelphij */ 5101251478Sdelphij for (ab = list_prev(dev->l2ad_buflist, head); ab; 5102251478Sdelphij ab = list_prev(dev->l2ad_buflist, ab)) { 5103251478Sdelphij l2arc_buf_hdr_t *l2hdr; 5104251478Sdelphij uint64_t buf_sz; 5105251478Sdelphij 5106251478Sdelphij /* 5107251478Sdelphij * We shouldn't need to lock the buffer here, since we flagged 5108251478Sdelphij * it as ARC_L2_WRITING in the previous step, but we must take 5109251478Sdelphij * care to only access its L2 cache parameters. In particular, 5110251478Sdelphij * ab->b_buf may be invalid by now due to ARC eviction. 5111251478Sdelphij */ 5112251478Sdelphij l2hdr = ab->b_l2hdr; 5113251478Sdelphij l2hdr->b_daddr = dev->l2ad_hand; 5114251478Sdelphij 5115251478Sdelphij if ((ab->b_flags & ARC_L2COMPRESS) && 5116251478Sdelphij l2hdr->b_asize >= buf_compress_minsz) { 5117251478Sdelphij if (l2arc_compress_buf(l2hdr)) { 5118251478Sdelphij /* 5119251478Sdelphij * If compression succeeded, enable headroom 5120251478Sdelphij * boost on the next scan cycle. 5121251478Sdelphij */ 5122251478Sdelphij *headroom_boost = B_TRUE; 5123251478Sdelphij } 5124251478Sdelphij } 5125251478Sdelphij 5126251478Sdelphij /* 5127251478Sdelphij * Pick up the buffer data we had previously stashed away 5128251478Sdelphij * (and now potentially also compressed). 5129251478Sdelphij */ 5130251478Sdelphij buf_data = l2hdr->b_tmp_cdata; 5131251478Sdelphij buf_sz = l2hdr->b_asize; 5132251478Sdelphij 5133251478Sdelphij /* Compression may have squashed the buffer to zero length. */ 5134251478Sdelphij if (buf_sz != 0) { 5135251478Sdelphij uint64_t buf_p_sz; 5136251478Sdelphij 5137185029Spjd wzio = zio_write_phys(pio, dev->l2ad_vdev, 5138185029Spjd dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5139185029Spjd NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5140185029Spjd ZIO_FLAG_CANFAIL, B_FALSE); 5141185029Spjd 5142185029Spjd DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5143185029Spjd zio_t *, wzio); 5144185029Spjd (void) zio_nowait(wzio); 5145185029Spjd 5146251478Sdelphij write_asize += buf_sz; 5147185029Spjd /* 5148185029Spjd * Keep the clock hand suitably device-aligned. 5149185029Spjd */ 5150251478Sdelphij buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5151251478Sdelphij write_psize += buf_p_sz; 5152251478Sdelphij dev->l2ad_hand += buf_p_sz; 5153185029Spjd } 5154251478Sdelphij } 5155185029Spjd 5156185029Spjd mutex_exit(&l2arc_buflist_mtx); 5157185029Spjd 5158251478Sdelphij ASSERT3U(write_asize, <=, target_sz); 5159185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_sent); 5160251478Sdelphij ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5161185029Spjd ARCSTAT_INCR(arcstat_l2_size, write_sz); 5162251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, write_asize); 5163268085Sdelphij vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 5164185029Spjd 5165185029Spjd /* 5166185029Spjd * Bump device hand to the device start if it is approaching the end. 5167185029Spjd * l2arc_evict() will already have evicted ahead for this case. 5168185029Spjd */ 5169185029Spjd if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5170185029Spjd dev->l2ad_hand = dev->l2ad_start; 5171185029Spjd dev->l2ad_evict = dev->l2ad_start; 5172185029Spjd dev->l2ad_first = B_FALSE; 5173185029Spjd } 5174185029Spjd 5175208373Smm dev->l2ad_writing = B_TRUE; 5176185029Spjd (void) zio_wait(pio); 5177208373Smm dev->l2ad_writing = B_FALSE; 5178208373Smm 5179251478Sdelphij return (write_asize); 5180185029Spjd} 5181185029Spjd 5182185029Spjd/* 5183251478Sdelphij * Compresses an L2ARC buffer. 5184251478Sdelphij * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5185251478Sdelphij * size in l2hdr->b_asize. This routine tries to compress the data and 5186251478Sdelphij * depending on the compression result there are three possible outcomes: 5187251478Sdelphij * *) The buffer was incompressible. The original l2hdr contents were left 5188251478Sdelphij * untouched and are ready for writing to an L2 device. 5189251478Sdelphij * *) The buffer was all-zeros, so there is no need to write it to an L2 5190251478Sdelphij * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5191251478Sdelphij * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5192251478Sdelphij * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5193251478Sdelphij * data buffer which holds the compressed data to be written, and b_asize 5194251478Sdelphij * tells us how much data there is. b_compress is set to the appropriate 5195251478Sdelphij * compression algorithm. Once writing is done, invoke 5196251478Sdelphij * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5197251478Sdelphij * 5198251478Sdelphij * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5199251478Sdelphij * buffer was incompressible). 5200251478Sdelphij */ 5201251478Sdelphijstatic boolean_t 5202251478Sdelphijl2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5203251478Sdelphij{ 5204251478Sdelphij void *cdata; 5205268075Sdelphij size_t csize, len, rounded; 5206251478Sdelphij 5207251478Sdelphij ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5208251478Sdelphij ASSERT(l2hdr->b_tmp_cdata != NULL); 5209251478Sdelphij 5210251478Sdelphij len = l2hdr->b_asize; 5211251478Sdelphij cdata = zio_data_buf_alloc(len); 5212251478Sdelphij csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5213269086Sdelphij cdata, l2hdr->b_asize); 5214251478Sdelphij 5215268075Sdelphij rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE); 5216268075Sdelphij if (rounded > csize) { 5217268075Sdelphij bzero((char *)cdata + csize, rounded - csize); 5218268075Sdelphij csize = rounded; 5219268075Sdelphij } 5220268075Sdelphij 5221251478Sdelphij if (csize == 0) { 5222251478Sdelphij /* zero block, indicate that there's nothing to write */ 5223251478Sdelphij zio_data_buf_free(cdata, len); 5224251478Sdelphij l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5225251478Sdelphij l2hdr->b_asize = 0; 5226251478Sdelphij l2hdr->b_tmp_cdata = NULL; 5227251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5228251478Sdelphij return (B_TRUE); 5229251478Sdelphij } else if (csize > 0 && csize < len) { 5230251478Sdelphij /* 5231251478Sdelphij * Compression succeeded, we'll keep the cdata around for 5232251478Sdelphij * writing and release it afterwards. 5233251478Sdelphij */ 5234251478Sdelphij l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5235251478Sdelphij l2hdr->b_asize = csize; 5236251478Sdelphij l2hdr->b_tmp_cdata = cdata; 5237251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_successes); 5238251478Sdelphij return (B_TRUE); 5239251478Sdelphij } else { 5240251478Sdelphij /* 5241251478Sdelphij * Compression failed, release the compressed buffer. 5242251478Sdelphij * l2hdr will be left unmodified. 5243251478Sdelphij */ 5244251478Sdelphij zio_data_buf_free(cdata, len); 5245251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_failures); 5246251478Sdelphij return (B_FALSE); 5247251478Sdelphij } 5248251478Sdelphij} 5249251478Sdelphij 5250251478Sdelphij/* 5251251478Sdelphij * Decompresses a zio read back from an l2arc device. On success, the 5252251478Sdelphij * underlying zio's io_data buffer is overwritten by the uncompressed 5253251478Sdelphij * version. On decompression error (corrupt compressed stream), the 5254251478Sdelphij * zio->io_error value is set to signal an I/O error. 5255251478Sdelphij * 5256251478Sdelphij * Please note that the compressed data stream is not checksummed, so 5257251478Sdelphij * if the underlying device is experiencing data corruption, we may feed 5258251478Sdelphij * corrupt data to the decompressor, so the decompressor needs to be 5259251478Sdelphij * able to handle this situation (LZ4 does). 5260251478Sdelphij */ 5261251478Sdelphijstatic void 5262251478Sdelphijl2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5263251478Sdelphij{ 5264251478Sdelphij ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5265251478Sdelphij 5266251478Sdelphij if (zio->io_error != 0) { 5267251478Sdelphij /* 5268251478Sdelphij * An io error has occured, just restore the original io 5269251478Sdelphij * size in preparation for a main pool read. 5270251478Sdelphij */ 5271251478Sdelphij zio->io_orig_size = zio->io_size = hdr->b_size; 5272251478Sdelphij return; 5273251478Sdelphij } 5274251478Sdelphij 5275251478Sdelphij if (c == ZIO_COMPRESS_EMPTY) { 5276251478Sdelphij /* 5277251478Sdelphij * An empty buffer results in a null zio, which means we 5278251478Sdelphij * need to fill its io_data after we're done restoring the 5279251478Sdelphij * buffer's contents. 5280251478Sdelphij */ 5281251478Sdelphij ASSERT(hdr->b_buf != NULL); 5282251478Sdelphij bzero(hdr->b_buf->b_data, hdr->b_size); 5283251478Sdelphij zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5284251478Sdelphij } else { 5285251478Sdelphij ASSERT(zio->io_data != NULL); 5286251478Sdelphij /* 5287251478Sdelphij * We copy the compressed data from the start of the arc buffer 5288251478Sdelphij * (the zio_read will have pulled in only what we need, the 5289251478Sdelphij * rest is garbage which we will overwrite at decompression) 5290251478Sdelphij * and then decompress back to the ARC data buffer. This way we 5291251478Sdelphij * can minimize copying by simply decompressing back over the 5292251478Sdelphij * original compressed data (rather than decompressing to an 5293251478Sdelphij * aux buffer and then copying back the uncompressed buffer, 5294251478Sdelphij * which is likely to be much larger). 5295251478Sdelphij */ 5296251478Sdelphij uint64_t csize; 5297251478Sdelphij void *cdata; 5298251478Sdelphij 5299251478Sdelphij csize = zio->io_size; 5300251478Sdelphij cdata = zio_data_buf_alloc(csize); 5301251478Sdelphij bcopy(zio->io_data, cdata, csize); 5302251478Sdelphij if (zio_decompress_data(c, cdata, zio->io_data, csize, 5303251478Sdelphij hdr->b_size) != 0) 5304251478Sdelphij zio->io_error = EIO; 5305251478Sdelphij zio_data_buf_free(cdata, csize); 5306251478Sdelphij } 5307251478Sdelphij 5308251478Sdelphij /* Restore the expected uncompressed IO size. */ 5309251478Sdelphij zio->io_orig_size = zio->io_size = hdr->b_size; 5310251478Sdelphij} 5311251478Sdelphij 5312251478Sdelphij/* 5313251478Sdelphij * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5314251478Sdelphij * This buffer serves as a temporary holder of compressed data while 5315251478Sdelphij * the buffer entry is being written to an l2arc device. Once that is 5316251478Sdelphij * done, we can dispose of it. 5317251478Sdelphij */ 5318251478Sdelphijstatic void 5319251478Sdelphijl2arc_release_cdata_buf(arc_buf_hdr_t *ab) 5320251478Sdelphij{ 5321251478Sdelphij l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr; 5322251478Sdelphij 5323251478Sdelphij if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) { 5324251478Sdelphij /* 5325251478Sdelphij * If the data was compressed, then we've allocated a 5326251478Sdelphij * temporary buffer for it, so now we need to release it. 5327251478Sdelphij */ 5328251478Sdelphij ASSERT(l2hdr->b_tmp_cdata != NULL); 5329251478Sdelphij zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size); 5330251478Sdelphij } 5331251478Sdelphij l2hdr->b_tmp_cdata = NULL; 5332251478Sdelphij} 5333251478Sdelphij 5334251478Sdelphij/* 5335185029Spjd * This thread feeds the L2ARC at regular intervals. This is the beating 5336185029Spjd * heart of the L2ARC. 5337185029Spjd */ 5338185029Spjdstatic void 5339185029Spjdl2arc_feed_thread(void *dummy __unused) 5340185029Spjd{ 5341185029Spjd callb_cpr_t cpr; 5342185029Spjd l2arc_dev_t *dev; 5343185029Spjd spa_t *spa; 5344208373Smm uint64_t size, wrote; 5345219089Spjd clock_t begin, next = ddi_get_lbolt(); 5346251478Sdelphij boolean_t headroom_boost = B_FALSE; 5347185029Spjd 5348185029Spjd CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5349185029Spjd 5350185029Spjd mutex_enter(&l2arc_feed_thr_lock); 5351185029Spjd 5352185029Spjd while (l2arc_thread_exit == 0) { 5353185029Spjd CALLB_CPR_SAFE_BEGIN(&cpr); 5354185029Spjd (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5355219089Spjd next - ddi_get_lbolt()); 5356185029Spjd CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5357219089Spjd next = ddi_get_lbolt() + hz; 5358185029Spjd 5359185029Spjd /* 5360185029Spjd * Quick check for L2ARC devices. 5361185029Spjd */ 5362185029Spjd mutex_enter(&l2arc_dev_mtx); 5363185029Spjd if (l2arc_ndev == 0) { 5364185029Spjd mutex_exit(&l2arc_dev_mtx); 5365185029Spjd continue; 5366185029Spjd } 5367185029Spjd mutex_exit(&l2arc_dev_mtx); 5368219089Spjd begin = ddi_get_lbolt(); 5369185029Spjd 5370185029Spjd /* 5371185029Spjd * This selects the next l2arc device to write to, and in 5372185029Spjd * doing so the next spa to feed from: dev->l2ad_spa. This 5373185029Spjd * will return NULL if there are now no l2arc devices or if 5374185029Spjd * they are all faulted. 5375185029Spjd * 5376185029Spjd * If a device is returned, its spa's config lock is also 5377185029Spjd * held to prevent device removal. l2arc_dev_get_next() 5378185029Spjd * will grab and release l2arc_dev_mtx. 5379185029Spjd */ 5380185029Spjd if ((dev = l2arc_dev_get_next()) == NULL) 5381185029Spjd continue; 5382185029Spjd 5383185029Spjd spa = dev->l2ad_spa; 5384185029Spjd ASSERT(spa != NULL); 5385185029Spjd 5386185029Spjd /* 5387219089Spjd * If the pool is read-only then force the feed thread to 5388219089Spjd * sleep a little longer. 5389219089Spjd */ 5390219089Spjd if (!spa_writeable(spa)) { 5391219089Spjd next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5392219089Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5393219089Spjd continue; 5394219089Spjd } 5395219089Spjd 5396219089Spjd /* 5397185029Spjd * Avoid contributing to memory pressure. 5398185029Spjd */ 5399185029Spjd if (arc_reclaim_needed()) { 5400185029Spjd ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5401185029Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5402185029Spjd continue; 5403185029Spjd } 5404185029Spjd 5405185029Spjd ARCSTAT_BUMP(arcstat_l2_feeds); 5406185029Spjd 5407251478Sdelphij size = l2arc_write_size(); 5408185029Spjd 5409185029Spjd /* 5410185029Spjd * Evict L2ARC buffers that will be overwritten. 5411185029Spjd */ 5412185029Spjd l2arc_evict(dev, size, B_FALSE); 5413185029Spjd 5414185029Spjd /* 5415185029Spjd * Write ARC buffers. 5416185029Spjd */ 5417251478Sdelphij wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5418208373Smm 5419208373Smm /* 5420208373Smm * Calculate interval between writes. 5421208373Smm */ 5422208373Smm next = l2arc_write_interval(begin, size, wrote); 5423185029Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5424185029Spjd } 5425185029Spjd 5426185029Spjd l2arc_thread_exit = 0; 5427185029Spjd cv_broadcast(&l2arc_feed_thr_cv); 5428185029Spjd CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5429185029Spjd thread_exit(); 5430185029Spjd} 5431185029Spjd 5432185029Spjdboolean_t 5433185029Spjdl2arc_vdev_present(vdev_t *vd) 5434185029Spjd{ 5435185029Spjd l2arc_dev_t *dev; 5436185029Spjd 5437185029Spjd mutex_enter(&l2arc_dev_mtx); 5438185029Spjd for (dev = list_head(l2arc_dev_list); dev != NULL; 5439185029Spjd dev = list_next(l2arc_dev_list, dev)) { 5440185029Spjd if (dev->l2ad_vdev == vd) 5441185029Spjd break; 5442185029Spjd } 5443185029Spjd mutex_exit(&l2arc_dev_mtx); 5444185029Spjd 5445185029Spjd return (dev != NULL); 5446185029Spjd} 5447185029Spjd 5448185029Spjd/* 5449185029Spjd * Add a vdev for use by the L2ARC. By this point the spa has already 5450185029Spjd * validated the vdev and opened it. 5451185029Spjd */ 5452185029Spjdvoid 5453219089Spjdl2arc_add_vdev(spa_t *spa, vdev_t *vd) 5454185029Spjd{ 5455185029Spjd l2arc_dev_t *adddev; 5456185029Spjd 5457185029Spjd ASSERT(!l2arc_vdev_present(vd)); 5458185029Spjd 5459255753Sgibbs vdev_ashift_optimize(vd); 5460255753Sgibbs 5461185029Spjd /* 5462185029Spjd * Create a new l2arc device entry. 5463185029Spjd */ 5464185029Spjd adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5465185029Spjd adddev->l2ad_spa = spa; 5466185029Spjd adddev->l2ad_vdev = vd; 5467219089Spjd adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5468219089Spjd adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5469185029Spjd adddev->l2ad_hand = adddev->l2ad_start; 5470185029Spjd adddev->l2ad_evict = adddev->l2ad_start; 5471185029Spjd adddev->l2ad_first = B_TRUE; 5472208373Smm adddev->l2ad_writing = B_FALSE; 5473185029Spjd 5474185029Spjd /* 5475185029Spjd * This is a list of all ARC buffers that are still valid on the 5476185029Spjd * device. 5477185029Spjd */ 5478185029Spjd adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5479185029Spjd list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5480185029Spjd offsetof(arc_buf_hdr_t, b_l2node)); 5481185029Spjd 5482219089Spjd vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5483185029Spjd 5484185029Spjd /* 5485185029Spjd * Add device to global list 5486185029Spjd */ 5487185029Spjd mutex_enter(&l2arc_dev_mtx); 5488185029Spjd list_insert_head(l2arc_dev_list, adddev); 5489185029Spjd atomic_inc_64(&l2arc_ndev); 5490185029Spjd mutex_exit(&l2arc_dev_mtx); 5491185029Spjd} 5492185029Spjd 5493185029Spjd/* 5494185029Spjd * Remove a vdev from the L2ARC. 5495185029Spjd */ 5496185029Spjdvoid 5497185029Spjdl2arc_remove_vdev(vdev_t *vd) 5498185029Spjd{ 5499185029Spjd l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5500185029Spjd 5501185029Spjd /* 5502185029Spjd * Find the device by vdev 5503185029Spjd */ 5504185029Spjd mutex_enter(&l2arc_dev_mtx); 5505185029Spjd for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5506185029Spjd nextdev = list_next(l2arc_dev_list, dev); 5507185029Spjd if (vd == dev->l2ad_vdev) { 5508185029Spjd remdev = dev; 5509185029Spjd break; 5510185029Spjd } 5511185029Spjd } 5512185029Spjd ASSERT(remdev != NULL); 5513185029Spjd 5514185029Spjd /* 5515185029Spjd * Remove device from global list 5516185029Spjd */ 5517185029Spjd list_remove(l2arc_dev_list, remdev); 5518185029Spjd l2arc_dev_last = NULL; /* may have been invalidated */ 5519185029Spjd atomic_dec_64(&l2arc_ndev); 5520185029Spjd mutex_exit(&l2arc_dev_mtx); 5521185029Spjd 5522185029Spjd /* 5523185029Spjd * Clear all buflists and ARC references. L2ARC device flush. 5524185029Spjd */ 5525185029Spjd l2arc_evict(remdev, 0, B_TRUE); 5526185029Spjd list_destroy(remdev->l2ad_buflist); 5527185029Spjd kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5528185029Spjd kmem_free(remdev, sizeof (l2arc_dev_t)); 5529185029Spjd} 5530185029Spjd 5531185029Spjdvoid 5532185029Spjdl2arc_init(void) 5533185029Spjd{ 5534185029Spjd l2arc_thread_exit = 0; 5535185029Spjd l2arc_ndev = 0; 5536185029Spjd l2arc_writes_sent = 0; 5537185029Spjd l2arc_writes_done = 0; 5538185029Spjd 5539185029Spjd mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5540185029Spjd cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5541185029Spjd mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5542185029Spjd mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5543185029Spjd mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5544185029Spjd 5545185029Spjd l2arc_dev_list = &L2ARC_dev_list; 5546185029Spjd l2arc_free_on_write = &L2ARC_free_on_write; 5547185029Spjd list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5548185029Spjd offsetof(l2arc_dev_t, l2ad_node)); 5549185029Spjd list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5550185029Spjd offsetof(l2arc_data_free_t, l2df_list_node)); 5551185029Spjd} 5552185029Spjd 5553185029Spjdvoid 5554185029Spjdl2arc_fini(void) 5555185029Spjd{ 5556185029Spjd /* 5557185029Spjd * This is called from dmu_fini(), which is called from spa_fini(); 5558185029Spjd * Because of this, we can assume that all l2arc devices have 5559185029Spjd * already been removed when the pools themselves were removed. 5560185029Spjd */ 5561185029Spjd 5562185029Spjd l2arc_do_free_on_write(); 5563185029Spjd 5564185029Spjd mutex_destroy(&l2arc_feed_thr_lock); 5565185029Spjd cv_destroy(&l2arc_feed_thr_cv); 5566185029Spjd mutex_destroy(&l2arc_dev_mtx); 5567185029Spjd mutex_destroy(&l2arc_buflist_mtx); 5568185029Spjd mutex_destroy(&l2arc_free_on_write_mtx); 5569185029Spjd 5570185029Spjd list_destroy(l2arc_dev_list); 5571185029Spjd list_destroy(l2arc_free_on_write); 5572185029Spjd} 5573185029Spjd 5574185029Spjdvoid 5575185029Spjdl2arc_start(void) 5576185029Spjd{ 5577209962Smm if (!(spa_mode_global & FWRITE)) 5578185029Spjd return; 5579185029Spjd 5580185029Spjd (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5581185029Spjd TS_RUN, minclsyspri); 5582185029Spjd} 5583185029Spjd 5584185029Spjdvoid 5585185029Spjdl2arc_stop(void) 5586185029Spjd{ 5587209962Smm if (!(spa_mode_global & FWRITE)) 5588185029Spjd return; 5589185029Spjd 5590185029Spjd mutex_enter(&l2arc_feed_thr_lock); 5591185029Spjd cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5592185029Spjd l2arc_thread_exit = 1; 5593185029Spjd while (l2arc_thread_exit != 0) 5594185029Spjd cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5595185029Spjd mutex_exit(&l2arc_feed_thr_lock); 5596185029Spjd} 5597