arc.c revision 277452
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> 141272483Ssmh#include <machine/vmparam.h> 142191902Skmacy 143240133Smm#ifdef illumos 144240133Smm#ifndef _KERNEL 145240133Smm/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 146240133Smmboolean_t arc_watch = B_FALSE; 147240133Smmint arc_procfd; 148240133Smm#endif 149240133Smm#endif /* illumos */ 150240133Smm 151168404Spjdstatic kmutex_t arc_reclaim_thr_lock; 152168404Spjdstatic kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 153168404Spjdstatic uint8_t arc_thread_exit; 154168404Spjd 155168404Spjd#define ARC_REDUCE_DNLC_PERCENT 3 156168404Spjduint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 157168404Spjd 158168404Spjdtypedef enum arc_reclaim_strategy { 159168404Spjd ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 160168404Spjd ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 161168404Spjd} arc_reclaim_strategy_t; 162168404Spjd 163258632Savg/* 164258632Savg * The number of iterations through arc_evict_*() before we 165258632Savg * drop & reacquire the lock. 166258632Savg */ 167258632Savgint arc_evict_iterations = 100; 168258632Savg 169168404Spjd/* number of seconds before growing cache again */ 170168404Spjdstatic int arc_grow_retry = 60; 171168404Spjd 172208373Smm/* shift of arc_c for calculating both min and max arc_p */ 173208373Smmstatic int arc_p_min_shift = 4; 174208373Smm 175208373Smm/* log2(fraction of arc to reclaim) */ 176208373Smmstatic int arc_shrink_shift = 5; 177208373Smm 178168404Spjd/* 179168404Spjd * minimum lifespan of a prefetch block in clock ticks 180168404Spjd * (initialized in arc_init()) 181168404Spjd */ 182168404Spjdstatic int arc_min_prefetch_lifespan; 183168404Spjd 184258632Savg/* 185258632Savg * If this percent of memory is free, don't throttle. 186258632Savg */ 187258632Savgint arc_lotsfree_percent = 10; 188258632Savg 189208373Smmstatic int arc_dead; 190194043Skmacyextern int zfs_prefetch_disable; 191168404Spjd 192168404Spjd/* 193185029Spjd * The arc has filled available memory and has now warmed up. 194185029Spjd */ 195185029Spjdstatic boolean_t arc_warm; 196185029Spjd 197185029Spjduint64_t zfs_arc_max; 198185029Spjduint64_t zfs_arc_min; 199185029Spjduint64_t zfs_arc_meta_limit = 0; 200275780Sdelphijuint64_t zfs_arc_meta_min = 0; 201208373Smmint zfs_arc_grow_retry = 0; 202208373Smmint zfs_arc_shrink_shift = 0; 203208373Smmint zfs_arc_p_min_shift = 0; 204242845Sdelphijint zfs_disable_dup_eviction = 0; 205269230Sdelphijuint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 206272483Ssmhu_int zfs_arc_free_target = 0; 207185029Spjd 208270759Ssmhstatic int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 209275748Sdelphijstatic int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 210270759Ssmh 211270759Ssmh#ifdef _KERNEL 212270759Ssmhstatic void 213270759Ssmharc_free_target_init(void *unused __unused) 214270759Ssmh{ 215270759Ssmh 216272483Ssmh zfs_arc_free_target = vm_pageout_wakeup_thresh; 217270759Ssmh} 218270759SsmhSYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 219270759Ssmh arc_free_target_init, NULL); 220270759Ssmh 221185029SpjdTUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 222275780SdelphijTUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 223273026SdelphijTUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 224168473SpjdSYSCTL_DECL(_vfs_zfs); 225217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 226168473Spjd "Maximum ARC size"); 227217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 228168473Spjd "Minimum ARC size"); 229269230SdelphijSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 230269230Sdelphij &zfs_arc_average_blocksize, 0, 231269230Sdelphij "ARC average blocksize"); 232273026SdelphijSYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 233273026Sdelphij &arc_shrink_shift, 0, 234273026Sdelphij "log2(fraction of arc to reclaim)"); 235273026Sdelphij 236270759Ssmh/* 237270759Ssmh * We don't have a tunable for arc_free_target due to the dependency on 238270759Ssmh * pagedaemon initialisation. 239270759Ssmh */ 240270759SsmhSYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 241270759Ssmh CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 242270759Ssmh sysctl_vfs_zfs_arc_free_target, "IU", 243270759Ssmh "Desired number of free pages below which ARC triggers reclaim"); 244168404Spjd 245270759Ssmhstatic int 246270759Ssmhsysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 247270759Ssmh{ 248270759Ssmh u_int val; 249270759Ssmh int err; 250270759Ssmh 251270759Ssmh val = zfs_arc_free_target; 252270759Ssmh err = sysctl_handle_int(oidp, &val, 0, req); 253270759Ssmh if (err != 0 || req->newptr == NULL) 254270759Ssmh return (err); 255270759Ssmh 256272483Ssmh if (val < minfree) 257270759Ssmh return (EINVAL); 258272483Ssmh if (val > vm_cnt.v_page_count) 259270759Ssmh return (EINVAL); 260270759Ssmh 261270759Ssmh zfs_arc_free_target = val; 262270759Ssmh 263270759Ssmh return (0); 264270759Ssmh} 265275748Sdelphij 266275748Sdelphij/* 267275748Sdelphij * Must be declared here, before the definition of corresponding kstat 268275748Sdelphij * macro which uses the same names will confuse the compiler. 269275748Sdelphij */ 270275748SdelphijSYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 271275748Sdelphij CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 272275748Sdelphij sysctl_vfs_zfs_arc_meta_limit, "QU", 273275748Sdelphij "ARC metadata limit"); 274272483Ssmh#endif 275270759Ssmh 276168404Spjd/* 277185029Spjd * Note that buffers can be in one of 6 states: 278168404Spjd * ARC_anon - anonymous (discussed below) 279168404Spjd * ARC_mru - recently used, currently cached 280168404Spjd * ARC_mru_ghost - recentely used, no longer in cache 281168404Spjd * ARC_mfu - frequently used, currently cached 282168404Spjd * ARC_mfu_ghost - frequently used, no longer in cache 283185029Spjd * ARC_l2c_only - exists in L2ARC but not other states 284185029Spjd * When there are no active references to the buffer, they are 285185029Spjd * are linked onto a list in one of these arc states. These are 286185029Spjd * the only buffers that can be evicted or deleted. Within each 287185029Spjd * state there are multiple lists, one for meta-data and one for 288185029Spjd * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 289185029Spjd * etc.) is tracked separately so that it can be managed more 290185029Spjd * explicitly: favored over data, limited explicitly. 291168404Spjd * 292168404Spjd * Anonymous buffers are buffers that are not associated with 293168404Spjd * a DVA. These are buffers that hold dirty block copies 294168404Spjd * before they are written to stable storage. By definition, 295168404Spjd * they are "ref'd" and are considered part of arc_mru 296168404Spjd * that cannot be freed. Generally, they will aquire a DVA 297168404Spjd * as they are written and migrate onto the arc_mru list. 298185029Spjd * 299185029Spjd * The ARC_l2c_only state is for buffers that are in the second 300185029Spjd * level ARC but no longer in any of the ARC_m* lists. The second 301185029Spjd * level ARC itself may also contain buffers that are in any of 302185029Spjd * the ARC_m* states - meaning that a buffer can exist in two 303185029Spjd * places. The reason for the ARC_l2c_only state is to keep the 304185029Spjd * buffer header in the hash table, so that reads that hit the 305185029Spjd * second level ARC benefit from these fast lookups. 306168404Spjd */ 307168404Spjd 308205264Skmacy#define ARCS_LOCK_PAD CACHE_LINE_SIZE 309205231Skmacystruct arcs_lock { 310205231Skmacy kmutex_t arcs_lock; 311205231Skmacy#ifdef _KERNEL 312205231Skmacy unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 313205231Skmacy#endif 314205231Skmacy}; 315205231Skmacy 316205231Skmacy/* 317205231Skmacy * must be power of two for mask use to work 318205231Skmacy * 319205231Skmacy */ 320205231Skmacy#define ARC_BUFC_NUMDATALISTS 16 321205231Skmacy#define ARC_BUFC_NUMMETADATALISTS 16 322206796Spjd#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 323205231Skmacy 324168404Spjdtypedef struct arc_state { 325185029Spjd uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 326185029Spjd uint64_t arcs_size; /* total amount of data in this state */ 327205231Skmacy list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 328205264Skmacy struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 329168404Spjd} arc_state_t; 330168404Spjd 331206796Spjd#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 332205231Skmacy 333185029Spjd/* The 6 states: */ 334168404Spjdstatic arc_state_t ARC_anon; 335168404Spjdstatic arc_state_t ARC_mru; 336168404Spjdstatic arc_state_t ARC_mru_ghost; 337168404Spjdstatic arc_state_t ARC_mfu; 338168404Spjdstatic arc_state_t ARC_mfu_ghost; 339185029Spjdstatic arc_state_t ARC_l2c_only; 340168404Spjd 341168404Spjdtypedef struct arc_stats { 342168404Spjd kstat_named_t arcstat_hits; 343168404Spjd kstat_named_t arcstat_misses; 344168404Spjd kstat_named_t arcstat_demand_data_hits; 345168404Spjd kstat_named_t arcstat_demand_data_misses; 346168404Spjd kstat_named_t arcstat_demand_metadata_hits; 347168404Spjd kstat_named_t arcstat_demand_metadata_misses; 348168404Spjd kstat_named_t arcstat_prefetch_data_hits; 349168404Spjd kstat_named_t arcstat_prefetch_data_misses; 350168404Spjd kstat_named_t arcstat_prefetch_metadata_hits; 351168404Spjd kstat_named_t arcstat_prefetch_metadata_misses; 352168404Spjd kstat_named_t arcstat_mru_hits; 353168404Spjd kstat_named_t arcstat_mru_ghost_hits; 354168404Spjd kstat_named_t arcstat_mfu_hits; 355168404Spjd kstat_named_t arcstat_mfu_ghost_hits; 356205231Skmacy kstat_named_t arcstat_allocated; 357168404Spjd kstat_named_t arcstat_deleted; 358205231Skmacy kstat_named_t arcstat_stolen; 359168404Spjd kstat_named_t arcstat_recycle_miss; 360251629Sdelphij /* 361251629Sdelphij * Number of buffers that could not be evicted because the hash lock 362251629Sdelphij * was held by another thread. The lock may not necessarily be held 363251629Sdelphij * by something using the same buffer, since hash locks are shared 364251629Sdelphij * by multiple buffers. 365251629Sdelphij */ 366168404Spjd kstat_named_t arcstat_mutex_miss; 367251629Sdelphij /* 368251629Sdelphij * Number of buffers skipped because they have I/O in progress, are 369251629Sdelphij * indrect prefetch buffers that have not lived long enough, or are 370251629Sdelphij * not from the spa we're trying to evict from. 371251629Sdelphij */ 372168404Spjd kstat_named_t arcstat_evict_skip; 373208373Smm kstat_named_t arcstat_evict_l2_cached; 374208373Smm kstat_named_t arcstat_evict_l2_eligible; 375208373Smm kstat_named_t arcstat_evict_l2_ineligible; 376168404Spjd kstat_named_t arcstat_hash_elements; 377168404Spjd kstat_named_t arcstat_hash_elements_max; 378168404Spjd kstat_named_t arcstat_hash_collisions; 379168404Spjd kstat_named_t arcstat_hash_chains; 380168404Spjd kstat_named_t arcstat_hash_chain_max; 381168404Spjd kstat_named_t arcstat_p; 382168404Spjd kstat_named_t arcstat_c; 383168404Spjd kstat_named_t arcstat_c_min; 384168404Spjd kstat_named_t arcstat_c_max; 385168404Spjd kstat_named_t arcstat_size; 386185029Spjd kstat_named_t arcstat_hdr_size; 387208373Smm kstat_named_t arcstat_data_size; 388208373Smm kstat_named_t arcstat_other_size; 389185029Spjd kstat_named_t arcstat_l2_hits; 390185029Spjd kstat_named_t arcstat_l2_misses; 391185029Spjd kstat_named_t arcstat_l2_feeds; 392185029Spjd kstat_named_t arcstat_l2_rw_clash; 393208373Smm kstat_named_t arcstat_l2_read_bytes; 394208373Smm kstat_named_t arcstat_l2_write_bytes; 395185029Spjd kstat_named_t arcstat_l2_writes_sent; 396185029Spjd kstat_named_t arcstat_l2_writes_done; 397185029Spjd kstat_named_t arcstat_l2_writes_error; 398185029Spjd kstat_named_t arcstat_l2_writes_hdr_miss; 399185029Spjd kstat_named_t arcstat_l2_evict_lock_retry; 400185029Spjd kstat_named_t arcstat_l2_evict_reading; 401185029Spjd kstat_named_t arcstat_l2_free_on_write; 402274172Savg kstat_named_t arcstat_l2_cdata_free_on_write; 403185029Spjd kstat_named_t arcstat_l2_abort_lowmem; 404185029Spjd kstat_named_t arcstat_l2_cksum_bad; 405185029Spjd kstat_named_t arcstat_l2_io_error; 406185029Spjd kstat_named_t arcstat_l2_size; 407251478Sdelphij kstat_named_t arcstat_l2_asize; 408185029Spjd kstat_named_t arcstat_l2_hdr_size; 409251478Sdelphij kstat_named_t arcstat_l2_compress_successes; 410251478Sdelphij kstat_named_t arcstat_l2_compress_zeros; 411251478Sdelphij kstat_named_t arcstat_l2_compress_failures; 412205231Skmacy kstat_named_t arcstat_l2_write_trylock_fail; 413205231Skmacy kstat_named_t arcstat_l2_write_passed_headroom; 414205231Skmacy kstat_named_t arcstat_l2_write_spa_mismatch; 415206796Spjd kstat_named_t arcstat_l2_write_in_l2; 416205231Skmacy kstat_named_t arcstat_l2_write_hdr_io_in_progress; 417205231Skmacy kstat_named_t arcstat_l2_write_not_cacheable; 418205231Skmacy kstat_named_t arcstat_l2_write_full; 419205231Skmacy kstat_named_t arcstat_l2_write_buffer_iter; 420205231Skmacy kstat_named_t arcstat_l2_write_pios; 421205231Skmacy kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 422205231Skmacy kstat_named_t arcstat_l2_write_buffer_list_iter; 423205231Skmacy kstat_named_t arcstat_l2_write_buffer_list_null_iter; 424242845Sdelphij kstat_named_t arcstat_memory_throttle_count; 425242845Sdelphij kstat_named_t arcstat_duplicate_buffers; 426242845Sdelphij kstat_named_t arcstat_duplicate_buffers_size; 427242845Sdelphij kstat_named_t arcstat_duplicate_reads; 428275748Sdelphij kstat_named_t arcstat_meta_used; 429275748Sdelphij kstat_named_t arcstat_meta_limit; 430275748Sdelphij kstat_named_t arcstat_meta_max; 431275780Sdelphij kstat_named_t arcstat_meta_min; 432168404Spjd} arc_stats_t; 433168404Spjd 434168404Spjdstatic arc_stats_t arc_stats = { 435168404Spjd { "hits", KSTAT_DATA_UINT64 }, 436168404Spjd { "misses", KSTAT_DATA_UINT64 }, 437168404Spjd { "demand_data_hits", KSTAT_DATA_UINT64 }, 438168404Spjd { "demand_data_misses", KSTAT_DATA_UINT64 }, 439168404Spjd { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 440168404Spjd { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 441168404Spjd { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 442168404Spjd { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 443168404Spjd { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 444168404Spjd { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 445168404Spjd { "mru_hits", KSTAT_DATA_UINT64 }, 446168404Spjd { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 447168404Spjd { "mfu_hits", KSTAT_DATA_UINT64 }, 448168404Spjd { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 449205231Skmacy { "allocated", KSTAT_DATA_UINT64 }, 450168404Spjd { "deleted", KSTAT_DATA_UINT64 }, 451205231Skmacy { "stolen", KSTAT_DATA_UINT64 }, 452168404Spjd { "recycle_miss", KSTAT_DATA_UINT64 }, 453168404Spjd { "mutex_miss", KSTAT_DATA_UINT64 }, 454168404Spjd { "evict_skip", KSTAT_DATA_UINT64 }, 455208373Smm { "evict_l2_cached", KSTAT_DATA_UINT64 }, 456208373Smm { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 457208373Smm { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 458168404Spjd { "hash_elements", KSTAT_DATA_UINT64 }, 459168404Spjd { "hash_elements_max", KSTAT_DATA_UINT64 }, 460168404Spjd { "hash_collisions", KSTAT_DATA_UINT64 }, 461168404Spjd { "hash_chains", KSTAT_DATA_UINT64 }, 462168404Spjd { "hash_chain_max", KSTAT_DATA_UINT64 }, 463168404Spjd { "p", KSTAT_DATA_UINT64 }, 464168404Spjd { "c", KSTAT_DATA_UINT64 }, 465168404Spjd { "c_min", KSTAT_DATA_UINT64 }, 466168404Spjd { "c_max", KSTAT_DATA_UINT64 }, 467185029Spjd { "size", KSTAT_DATA_UINT64 }, 468185029Spjd { "hdr_size", KSTAT_DATA_UINT64 }, 469208373Smm { "data_size", KSTAT_DATA_UINT64 }, 470208373Smm { "other_size", KSTAT_DATA_UINT64 }, 471185029Spjd { "l2_hits", KSTAT_DATA_UINT64 }, 472185029Spjd { "l2_misses", KSTAT_DATA_UINT64 }, 473185029Spjd { "l2_feeds", KSTAT_DATA_UINT64 }, 474185029Spjd { "l2_rw_clash", KSTAT_DATA_UINT64 }, 475208373Smm { "l2_read_bytes", KSTAT_DATA_UINT64 }, 476208373Smm { "l2_write_bytes", KSTAT_DATA_UINT64 }, 477185029Spjd { "l2_writes_sent", KSTAT_DATA_UINT64 }, 478185029Spjd { "l2_writes_done", KSTAT_DATA_UINT64 }, 479185029Spjd { "l2_writes_error", KSTAT_DATA_UINT64 }, 480185029Spjd { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 481185029Spjd { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 482185029Spjd { "l2_evict_reading", KSTAT_DATA_UINT64 }, 483185029Spjd { "l2_free_on_write", KSTAT_DATA_UINT64 }, 484274172Savg { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, 485185029Spjd { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 486185029Spjd { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 487185029Spjd { "l2_io_error", KSTAT_DATA_UINT64 }, 488185029Spjd { "l2_size", KSTAT_DATA_UINT64 }, 489251478Sdelphij { "l2_asize", KSTAT_DATA_UINT64 }, 490185029Spjd { "l2_hdr_size", KSTAT_DATA_UINT64 }, 491251478Sdelphij { "l2_compress_successes", KSTAT_DATA_UINT64 }, 492251478Sdelphij { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 493251478Sdelphij { "l2_compress_failures", KSTAT_DATA_UINT64 }, 494206796Spjd { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 495206796Spjd { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 496206796Spjd { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 497206796Spjd { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 498206796Spjd { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 499206796Spjd { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 500206796Spjd { "l2_write_full", KSTAT_DATA_UINT64 }, 501206796Spjd { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 502206796Spjd { "l2_write_pios", KSTAT_DATA_UINT64 }, 503206796Spjd { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 504206796Spjd { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 505242845Sdelphij { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 506242845Sdelphij { "memory_throttle_count", KSTAT_DATA_UINT64 }, 507242845Sdelphij { "duplicate_buffers", KSTAT_DATA_UINT64 }, 508242845Sdelphij { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 509275748Sdelphij { "duplicate_reads", KSTAT_DATA_UINT64 }, 510275748Sdelphij { "arc_meta_used", KSTAT_DATA_UINT64 }, 511275748Sdelphij { "arc_meta_limit", KSTAT_DATA_UINT64 }, 512275780Sdelphij { "arc_meta_max", KSTAT_DATA_UINT64 }, 513275780Sdelphij { "arc_meta_min", KSTAT_DATA_UINT64 } 514168404Spjd}; 515168404Spjd 516168404Spjd#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 517168404Spjd 518168404Spjd#define ARCSTAT_INCR(stat, val) \ 519251631Sdelphij atomic_add_64(&arc_stats.stat.value.ui64, (val)) 520168404Spjd 521206796Spjd#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 522168404Spjd#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 523168404Spjd 524168404Spjd#define ARCSTAT_MAX(stat, val) { \ 525168404Spjd uint64_t m; \ 526168404Spjd while ((val) > (m = arc_stats.stat.value.ui64) && \ 527168404Spjd (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 528168404Spjd continue; \ 529168404Spjd} 530168404Spjd 531168404Spjd#define ARCSTAT_MAXSTAT(stat) \ 532168404Spjd ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 533168404Spjd 534168404Spjd/* 535168404Spjd * We define a macro to allow ARC hits/misses to be easily broken down by 536168404Spjd * two separate conditions, giving a total of four different subtypes for 537168404Spjd * each of hits and misses (so eight statistics total). 538168404Spjd */ 539168404Spjd#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 540168404Spjd if (cond1) { \ 541168404Spjd if (cond2) { \ 542168404Spjd ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 543168404Spjd } else { \ 544168404Spjd ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 545168404Spjd } \ 546168404Spjd } else { \ 547168404Spjd if (cond2) { \ 548168404Spjd ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 549168404Spjd } else { \ 550168404Spjd ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 551168404Spjd } \ 552168404Spjd } 553168404Spjd 554168404Spjdkstat_t *arc_ksp; 555206796Spjdstatic arc_state_t *arc_anon; 556168404Spjdstatic arc_state_t *arc_mru; 557168404Spjdstatic arc_state_t *arc_mru_ghost; 558168404Spjdstatic arc_state_t *arc_mfu; 559168404Spjdstatic arc_state_t *arc_mfu_ghost; 560185029Spjdstatic arc_state_t *arc_l2c_only; 561168404Spjd 562168404Spjd/* 563168404Spjd * There are several ARC variables that are critical to export as kstats -- 564168404Spjd * but we don't want to have to grovel around in the kstat whenever we wish to 565168404Spjd * manipulate them. For these variables, we therefore define them to be in 566168404Spjd * terms of the statistic variable. This assures that we are not introducing 567168404Spjd * the possibility of inconsistency by having shadow copies of the variables, 568168404Spjd * while still allowing the code to be readable. 569168404Spjd */ 570168404Spjd#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 571168404Spjd#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 572168404Spjd#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 573168404Spjd#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 574168404Spjd#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 575275748Sdelphij#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 576275780Sdelphij#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 577275748Sdelphij#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 578275748Sdelphij#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 579168404Spjd 580251478Sdelphij#define L2ARC_IS_VALID_COMPRESS(_c_) \ 581251478Sdelphij ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 582251478Sdelphij 583168404Spjdstatic int arc_no_grow; /* Don't try to grow cache size */ 584168404Spjdstatic uint64_t arc_tempreserve; 585209962Smmstatic uint64_t arc_loaned_bytes; 586168404Spjd 587185029Spjdtypedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 588185029Spjd 589168404Spjdtypedef struct arc_callback arc_callback_t; 590168404Spjd 591168404Spjdstruct arc_callback { 592168404Spjd void *acb_private; 593168404Spjd arc_done_func_t *acb_done; 594168404Spjd arc_buf_t *acb_buf; 595168404Spjd zio_t *acb_zio_dummy; 596168404Spjd arc_callback_t *acb_next; 597168404Spjd}; 598168404Spjd 599168404Spjdtypedef struct arc_write_callback arc_write_callback_t; 600168404Spjd 601168404Spjdstruct arc_write_callback { 602168404Spjd void *awcb_private; 603168404Spjd arc_done_func_t *awcb_ready; 604258632Savg arc_done_func_t *awcb_physdone; 605168404Spjd arc_done_func_t *awcb_done; 606168404Spjd arc_buf_t *awcb_buf; 607168404Spjd}; 608168404Spjd 609168404Spjdstruct arc_buf_hdr { 610168404Spjd /* protected by hash lock */ 611168404Spjd dva_t b_dva; 612168404Spjd uint64_t b_birth; 613168404Spjd uint64_t b_cksum0; 614168404Spjd 615168404Spjd kmutex_t b_freeze_lock; 616168404Spjd zio_cksum_t *b_freeze_cksum; 617219089Spjd void *b_thawed; 618168404Spjd 619168404Spjd arc_buf_hdr_t *b_hash_next; 620168404Spjd arc_buf_t *b_buf; 621275811Sdelphij arc_flags_t b_flags; 622168404Spjd uint32_t b_datacnt; 623168404Spjd 624168404Spjd arc_callback_t *b_acb; 625168404Spjd kcondvar_t b_cv; 626168404Spjd 627168404Spjd /* immutable */ 628168404Spjd arc_buf_contents_t b_type; 629168404Spjd uint64_t b_size; 630209962Smm uint64_t b_spa; 631168404Spjd 632168404Spjd /* protected by arc state mutex */ 633168404Spjd arc_state_t *b_state; 634168404Spjd list_node_t b_arc_node; 635168404Spjd 636168404Spjd /* updated atomically */ 637168404Spjd clock_t b_arc_access; 638168404Spjd 639168404Spjd /* self protecting */ 640168404Spjd refcount_t b_refcnt; 641185029Spjd 642185029Spjd l2arc_buf_hdr_t *b_l2hdr; 643185029Spjd list_node_t b_l2node; 644168404Spjd}; 645168404Spjd 646275748Sdelphij#ifdef _KERNEL 647275748Sdelphijstatic int 648275748Sdelphijsysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 649275748Sdelphij{ 650275748Sdelphij uint64_t val; 651275748Sdelphij int err; 652275748Sdelphij 653275748Sdelphij val = arc_meta_limit; 654275748Sdelphij err = sysctl_handle_64(oidp, &val, 0, req); 655275748Sdelphij if (err != 0 || req->newptr == NULL) 656275748Sdelphij return (err); 657275748Sdelphij 658275748Sdelphij if (val <= 0 || val > arc_c_max) 659275748Sdelphij return (EINVAL); 660275748Sdelphij 661275748Sdelphij arc_meta_limit = val; 662275748Sdelphij return (0); 663275748Sdelphij} 664275748Sdelphij#endif 665275748Sdelphij 666168404Spjdstatic arc_buf_t *arc_eviction_list; 667168404Spjdstatic kmutex_t arc_eviction_mtx; 668168404Spjdstatic arc_buf_hdr_t arc_eviction_hdr; 669168404Spjd 670168404Spjd#define GHOST_STATE(state) \ 671185029Spjd ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 672185029Spjd (state) == arc_l2c_only) 673168404Spjd 674275811Sdelphij#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 675275811Sdelphij#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 676275811Sdelphij#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 677275811Sdelphij#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 678275811Sdelphij#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ) 679275811Sdelphij#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE) 680275811Sdelphij#define HDR_FREE_IN_PROGRESS(hdr) \ 681275811Sdelphij ((hdr)->b_flags & ARC_FLAG_FREE_IN_PROGRESS) 682275811Sdelphij#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 683275811Sdelphij#define HDR_L2_READING(hdr) \ 684275811Sdelphij ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS && \ 685275811Sdelphij (hdr)->b_l2hdr != NULL) 686275811Sdelphij#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 687275811Sdelphij#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 688275811Sdelphij#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 689168404Spjd 690168404Spjd/* 691185029Spjd * Other sizes 692185029Spjd */ 693185029Spjd 694185029Spjd#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 695185029Spjd#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 696185029Spjd 697185029Spjd/* 698168404Spjd * Hash table routines 699168404Spjd */ 700168404Spjd 701205253Skmacy#define HT_LOCK_PAD CACHE_LINE_SIZE 702168404Spjd 703168404Spjdstruct ht_lock { 704168404Spjd kmutex_t ht_lock; 705168404Spjd#ifdef _KERNEL 706168404Spjd unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 707168404Spjd#endif 708168404Spjd}; 709168404Spjd 710168404Spjd#define BUF_LOCKS 256 711168404Spjdtypedef struct buf_hash_table { 712168404Spjd uint64_t ht_mask; 713168404Spjd arc_buf_hdr_t **ht_table; 714205264Skmacy struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 715168404Spjd} buf_hash_table_t; 716168404Spjd 717168404Spjdstatic buf_hash_table_t buf_hash_table; 718168404Spjd 719168404Spjd#define BUF_HASH_INDEX(spa, dva, birth) \ 720168404Spjd (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 721168404Spjd#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 722168404Spjd#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 723219089Spjd#define HDR_LOCK(hdr) \ 724219089Spjd (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 725168404Spjd 726168404Spjduint64_t zfs_crc64_table[256]; 727168404Spjd 728185029Spjd/* 729185029Spjd * Level 2 ARC 730185029Spjd */ 731185029Spjd 732272707Savg#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 733251478Sdelphij#define L2ARC_HEADROOM 2 /* num of writes */ 734251478Sdelphij/* 735251478Sdelphij * If we discover during ARC scan any buffers to be compressed, we boost 736251478Sdelphij * our headroom for the next scanning cycle by this percentage multiple. 737251478Sdelphij */ 738251478Sdelphij#define L2ARC_HEADROOM_BOOST 200 739208373Smm#define L2ARC_FEED_SECS 1 /* caching interval secs */ 740208373Smm#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 741185029Spjd 742185029Spjd#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 743185029Spjd#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 744185029Spjd 745251631Sdelphij/* L2ARC Performance Tunables */ 746185029Spjduint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 747185029Spjduint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 748185029Spjduint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 749251478Sdelphijuint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 750185029Spjduint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 751208373Smmuint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 752219089Spjdboolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 753208373Smmboolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 754208373Smmboolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 755185029Spjd 756217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 757205231Skmacy &l2arc_write_max, 0, "max write size"); 758217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 759205231Skmacy &l2arc_write_boost, 0, "extra write during warmup"); 760217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 761205231Skmacy &l2arc_headroom, 0, "number of dev writes"); 762217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 763205231Skmacy &l2arc_feed_secs, 0, "interval seconds"); 764217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 765208373Smm &l2arc_feed_min_ms, 0, "min interval milliseconds"); 766205231Skmacy 767205231SkmacySYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 768205231Skmacy &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 769208373SmmSYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 770208373Smm &l2arc_feed_again, 0, "turbo warmup"); 771208373SmmSYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 772208373Smm &l2arc_norw, 0, "no reads during writes"); 773205231Skmacy 774217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 775205231Skmacy &ARC_anon.arcs_size, 0, "size of anonymous state"); 776217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 777205231Skmacy &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 778217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 779205231Skmacy &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 780205231Skmacy 781217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 782205231Skmacy &ARC_mru.arcs_size, 0, "size of mru state"); 783217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 784205231Skmacy &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 785217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 786205231Skmacy &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 787205231Skmacy 788217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 789205231Skmacy &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 790217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 791205231Skmacy &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 792205231Skmacy "size of metadata in mru ghost state"); 793217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 794205231Skmacy &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 795205231Skmacy "size of data in mru ghost state"); 796205231Skmacy 797217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 798205231Skmacy &ARC_mfu.arcs_size, 0, "size of mfu state"); 799217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 800205231Skmacy &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 801217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 802205231Skmacy &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 803205231Skmacy 804217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 805205231Skmacy &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 806217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 807205231Skmacy &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 808205231Skmacy "size of metadata in mfu ghost state"); 809217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 810205231Skmacy &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 811205231Skmacy "size of data in mfu ghost state"); 812205231Skmacy 813217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 814205231Skmacy &ARC_l2c_only.arcs_size, 0, "size of mru state"); 815205231Skmacy 816185029Spjd/* 817185029Spjd * L2ARC Internals 818185029Spjd */ 819185029Spjdtypedef struct l2arc_dev { 820185029Spjd vdev_t *l2ad_vdev; /* vdev */ 821185029Spjd spa_t *l2ad_spa; /* spa */ 822185029Spjd uint64_t l2ad_hand; /* next write location */ 823185029Spjd uint64_t l2ad_start; /* first addr on device */ 824185029Spjd uint64_t l2ad_end; /* last addr on device */ 825185029Spjd uint64_t l2ad_evict; /* last addr eviction reached */ 826185029Spjd boolean_t l2ad_first; /* first sweep through */ 827208373Smm boolean_t l2ad_writing; /* currently writing */ 828185029Spjd list_t *l2ad_buflist; /* buffer list */ 829185029Spjd list_node_t l2ad_node; /* device list node */ 830185029Spjd} l2arc_dev_t; 831185029Spjd 832185029Spjdstatic list_t L2ARC_dev_list; /* device list */ 833185029Spjdstatic list_t *l2arc_dev_list; /* device list pointer */ 834185029Spjdstatic kmutex_t l2arc_dev_mtx; /* device list mutex */ 835185029Spjdstatic l2arc_dev_t *l2arc_dev_last; /* last device used */ 836185029Spjdstatic kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 837185029Spjdstatic list_t L2ARC_free_on_write; /* free after write buf list */ 838185029Spjdstatic list_t *l2arc_free_on_write; /* free after write list ptr */ 839185029Spjdstatic kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 840185029Spjdstatic uint64_t l2arc_ndev; /* number of devices */ 841185029Spjd 842185029Spjdtypedef struct l2arc_read_callback { 843251478Sdelphij arc_buf_t *l2rcb_buf; /* read buffer */ 844251478Sdelphij spa_t *l2rcb_spa; /* spa */ 845251478Sdelphij blkptr_t l2rcb_bp; /* original blkptr */ 846268123Sdelphij zbookmark_phys_t l2rcb_zb; /* original bookmark */ 847251478Sdelphij int l2rcb_flags; /* original flags */ 848251478Sdelphij enum zio_compress l2rcb_compress; /* applied compress */ 849185029Spjd} l2arc_read_callback_t; 850185029Spjd 851185029Spjdtypedef struct l2arc_write_callback { 852185029Spjd l2arc_dev_t *l2wcb_dev; /* device info */ 853185029Spjd arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 854185029Spjd} l2arc_write_callback_t; 855185029Spjd 856185029Spjdstruct l2arc_buf_hdr { 857185029Spjd /* protected by arc_buf_hdr mutex */ 858251478Sdelphij l2arc_dev_t *b_dev; /* L2ARC device */ 859251478Sdelphij uint64_t b_daddr; /* disk address, offset byte */ 860251478Sdelphij /* compression applied to buffer data */ 861251478Sdelphij enum zio_compress b_compress; 862251478Sdelphij /* real alloc'd buffer size depending on b_compress applied */ 863251478Sdelphij int b_asize; 864251478Sdelphij /* temporary buffer holder for in-flight compressed data */ 865251478Sdelphij void *b_tmp_cdata; 866185029Spjd}; 867185029Spjd 868185029Spjdtypedef struct l2arc_data_free { 869185029Spjd /* protected by l2arc_free_on_write_mtx */ 870185029Spjd void *l2df_data; 871185029Spjd size_t l2df_size; 872185029Spjd void (*l2df_func)(void *, size_t); 873185029Spjd list_node_t l2df_list_node; 874185029Spjd} l2arc_data_free_t; 875185029Spjd 876185029Spjdstatic kmutex_t l2arc_feed_thr_lock; 877185029Spjdstatic kcondvar_t l2arc_feed_thr_cv; 878185029Spjdstatic uint8_t l2arc_thread_exit; 879185029Spjd 880275811Sdelphijstatic void arc_get_data_buf(arc_buf_t *); 881275811Sdelphijstatic void arc_access(arc_buf_hdr_t *, kmutex_t *); 882275811Sdelphijstatic int arc_evict_needed(arc_buf_contents_t); 883275811Sdelphijstatic void arc_evict_ghost(arc_state_t *, uint64_t, int64_t); 884275811Sdelphijstatic void arc_buf_watch(arc_buf_t *); 885275811Sdelphij 886275811Sdelphijstatic boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 887275811Sdelphijstatic void l2arc_read_done(zio_t *); 888185029Spjdstatic void l2arc_hdr_stat_add(void); 889185029Spjdstatic void l2arc_hdr_stat_remove(void); 890185029Spjd 891275811Sdelphijstatic boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *); 892275811Sdelphijstatic void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress); 893275811Sdelphijstatic void l2arc_release_cdata_buf(arc_buf_hdr_t *); 894251478Sdelphij 895168404Spjdstatic uint64_t 896209962Smmbuf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 897168404Spjd{ 898168404Spjd uint8_t *vdva = (uint8_t *)dva; 899168404Spjd uint64_t crc = -1ULL; 900168404Spjd int i; 901168404Spjd 902168404Spjd ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 903168404Spjd 904168404Spjd for (i = 0; i < sizeof (dva_t); i++) 905168404Spjd crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 906168404Spjd 907209962Smm crc ^= (spa>>8) ^ birth; 908168404Spjd 909168404Spjd return (crc); 910168404Spjd} 911168404Spjd 912168404Spjd#define BUF_EMPTY(buf) \ 913168404Spjd ((buf)->b_dva.dva_word[0] == 0 && \ 914168404Spjd (buf)->b_dva.dva_word[1] == 0 && \ 915260150Sdelphij (buf)->b_cksum0 == 0) 916168404Spjd 917168404Spjd#define BUF_EQUAL(spa, dva, birth, buf) \ 918168404Spjd ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 919168404Spjd ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 920168404Spjd ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 921168404Spjd 922219089Spjdstatic void 923219089Spjdbuf_discard_identity(arc_buf_hdr_t *hdr) 924219089Spjd{ 925219089Spjd hdr->b_dva.dva_word[0] = 0; 926219089Spjd hdr->b_dva.dva_word[1] = 0; 927219089Spjd hdr->b_birth = 0; 928219089Spjd hdr->b_cksum0 = 0; 929219089Spjd} 930219089Spjd 931168404Spjdstatic arc_buf_hdr_t * 932268075Sdelphijbuf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 933168404Spjd{ 934268075Sdelphij const dva_t *dva = BP_IDENTITY(bp); 935268075Sdelphij uint64_t birth = BP_PHYSICAL_BIRTH(bp); 936168404Spjd uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 937168404Spjd kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 938275811Sdelphij arc_buf_hdr_t *hdr; 939168404Spjd 940168404Spjd mutex_enter(hash_lock); 941275811Sdelphij for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 942275811Sdelphij hdr = hdr->b_hash_next) { 943275811Sdelphij if (BUF_EQUAL(spa, dva, birth, hdr)) { 944168404Spjd *lockp = hash_lock; 945275811Sdelphij return (hdr); 946168404Spjd } 947168404Spjd } 948168404Spjd mutex_exit(hash_lock); 949168404Spjd *lockp = NULL; 950168404Spjd return (NULL); 951168404Spjd} 952168404Spjd 953168404Spjd/* 954168404Spjd * Insert an entry into the hash table. If there is already an element 955168404Spjd * equal to elem in the hash table, then the already existing element 956168404Spjd * will be returned and the new element will not be inserted. 957168404Spjd * Otherwise returns NULL. 958168404Spjd */ 959168404Spjdstatic arc_buf_hdr_t * 960275811Sdelphijbuf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 961168404Spjd{ 962275811Sdelphij uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 963168404Spjd kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 964275811Sdelphij arc_buf_hdr_t *fhdr; 965168404Spjd uint32_t i; 966168404Spjd 967275811Sdelphij ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 968275811Sdelphij ASSERT(hdr->b_birth != 0); 969275811Sdelphij ASSERT(!HDR_IN_HASH_TABLE(hdr)); 970168404Spjd *lockp = hash_lock; 971168404Spjd mutex_enter(hash_lock); 972275811Sdelphij for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 973275811Sdelphij fhdr = fhdr->b_hash_next, i++) { 974275811Sdelphij if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 975275811Sdelphij return (fhdr); 976168404Spjd } 977168404Spjd 978275811Sdelphij hdr->b_hash_next = buf_hash_table.ht_table[idx]; 979275811Sdelphij buf_hash_table.ht_table[idx] = hdr; 980275811Sdelphij hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 981168404Spjd 982168404Spjd /* collect some hash table performance data */ 983168404Spjd if (i > 0) { 984168404Spjd ARCSTAT_BUMP(arcstat_hash_collisions); 985168404Spjd if (i == 1) 986168404Spjd ARCSTAT_BUMP(arcstat_hash_chains); 987168404Spjd 988168404Spjd ARCSTAT_MAX(arcstat_hash_chain_max, i); 989168404Spjd } 990168404Spjd 991168404Spjd ARCSTAT_BUMP(arcstat_hash_elements); 992168404Spjd ARCSTAT_MAXSTAT(arcstat_hash_elements); 993168404Spjd 994168404Spjd return (NULL); 995168404Spjd} 996168404Spjd 997168404Spjdstatic void 998275811Sdelphijbuf_hash_remove(arc_buf_hdr_t *hdr) 999168404Spjd{ 1000275811Sdelphij arc_buf_hdr_t *fhdr, **hdrp; 1001275811Sdelphij uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1002168404Spjd 1003168404Spjd ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1004275811Sdelphij ASSERT(HDR_IN_HASH_TABLE(hdr)); 1005168404Spjd 1006275811Sdelphij hdrp = &buf_hash_table.ht_table[idx]; 1007275811Sdelphij while ((fhdr = *hdrp) != hdr) { 1008275811Sdelphij ASSERT(fhdr != NULL); 1009275811Sdelphij hdrp = &fhdr->b_hash_next; 1010168404Spjd } 1011275811Sdelphij *hdrp = hdr->b_hash_next; 1012275811Sdelphij hdr->b_hash_next = NULL; 1013275811Sdelphij hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; 1014168404Spjd 1015168404Spjd /* collect some hash table performance data */ 1016168404Spjd ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1017168404Spjd 1018168404Spjd if (buf_hash_table.ht_table[idx] && 1019168404Spjd buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1020168404Spjd ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1021168404Spjd} 1022168404Spjd 1023168404Spjd/* 1024168404Spjd * Global data structures and functions for the buf kmem cache. 1025168404Spjd */ 1026168404Spjdstatic kmem_cache_t *hdr_cache; 1027168404Spjdstatic kmem_cache_t *buf_cache; 1028168404Spjd 1029168404Spjdstatic void 1030168404Spjdbuf_fini(void) 1031168404Spjd{ 1032168404Spjd int i; 1033168404Spjd 1034168404Spjd kmem_free(buf_hash_table.ht_table, 1035168404Spjd (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1036168404Spjd for (i = 0; i < BUF_LOCKS; i++) 1037168404Spjd mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1038168404Spjd kmem_cache_destroy(hdr_cache); 1039168404Spjd kmem_cache_destroy(buf_cache); 1040168404Spjd} 1041168404Spjd 1042168404Spjd/* 1043168404Spjd * Constructor callback - called when the cache is empty 1044168404Spjd * and a new buf is requested. 1045168404Spjd */ 1046168404Spjd/* ARGSUSED */ 1047168404Spjdstatic int 1048168404Spjdhdr_cons(void *vbuf, void *unused, int kmflag) 1049168404Spjd{ 1050275811Sdelphij arc_buf_hdr_t *hdr = vbuf; 1051168404Spjd 1052275811Sdelphij bzero(hdr, sizeof (arc_buf_hdr_t)); 1053275811Sdelphij refcount_create(&hdr->b_refcnt); 1054275811Sdelphij cv_init(&hdr->b_cv, NULL, CV_DEFAULT, NULL); 1055275811Sdelphij mutex_init(&hdr->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1056208373Smm arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1057185029Spjd 1058168404Spjd return (0); 1059168404Spjd} 1060168404Spjd 1061185029Spjd/* ARGSUSED */ 1062185029Spjdstatic int 1063185029Spjdbuf_cons(void *vbuf, void *unused, int kmflag) 1064185029Spjd{ 1065185029Spjd arc_buf_t *buf = vbuf; 1066185029Spjd 1067185029Spjd bzero(buf, sizeof (arc_buf_t)); 1068219089Spjd mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1069208373Smm arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1070208373Smm 1071185029Spjd return (0); 1072185029Spjd} 1073185029Spjd 1074168404Spjd/* 1075168404Spjd * Destructor callback - called when a cached buf is 1076168404Spjd * no longer required. 1077168404Spjd */ 1078168404Spjd/* ARGSUSED */ 1079168404Spjdstatic void 1080168404Spjdhdr_dest(void *vbuf, void *unused) 1081168404Spjd{ 1082275811Sdelphij arc_buf_hdr_t *hdr = vbuf; 1083168404Spjd 1084275811Sdelphij ASSERT(BUF_EMPTY(hdr)); 1085275811Sdelphij refcount_destroy(&hdr->b_refcnt); 1086275811Sdelphij cv_destroy(&hdr->b_cv); 1087275811Sdelphij mutex_destroy(&hdr->b_freeze_lock); 1088208373Smm arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1089168404Spjd} 1090168404Spjd 1091185029Spjd/* ARGSUSED */ 1092185029Spjdstatic void 1093185029Spjdbuf_dest(void *vbuf, void *unused) 1094185029Spjd{ 1095185029Spjd arc_buf_t *buf = vbuf; 1096185029Spjd 1097219089Spjd mutex_destroy(&buf->b_evict_lock); 1098208373Smm arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1099185029Spjd} 1100185029Spjd 1101168404Spjd/* 1102168404Spjd * Reclaim callback -- invoked when memory is low. 1103168404Spjd */ 1104168404Spjd/* ARGSUSED */ 1105168404Spjdstatic void 1106168404Spjdhdr_recl(void *unused) 1107168404Spjd{ 1108168404Spjd dprintf("hdr_recl called\n"); 1109168404Spjd /* 1110168404Spjd * umem calls the reclaim func when we destroy the buf cache, 1111168404Spjd * which is after we do arc_fini(). 1112168404Spjd */ 1113168404Spjd if (!arc_dead) 1114168404Spjd cv_signal(&arc_reclaim_thr_cv); 1115168404Spjd} 1116168404Spjd 1117168404Spjdstatic void 1118168404Spjdbuf_init(void) 1119168404Spjd{ 1120168404Spjd uint64_t *ct; 1121168404Spjd uint64_t hsize = 1ULL << 12; 1122168404Spjd int i, j; 1123168404Spjd 1124168404Spjd /* 1125168404Spjd * The hash table is big enough to fill all of physical memory 1126269230Sdelphij * with an average block size of zfs_arc_average_blocksize (default 8K). 1127269230Sdelphij * By default, the table will take up 1128269230Sdelphij * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1129168404Spjd */ 1130269230Sdelphij while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1131168404Spjd hsize <<= 1; 1132168404Spjdretry: 1133168404Spjd buf_hash_table.ht_mask = hsize - 1; 1134168404Spjd buf_hash_table.ht_table = 1135168404Spjd kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1136168404Spjd if (buf_hash_table.ht_table == NULL) { 1137168404Spjd ASSERT(hsize > (1ULL << 8)); 1138168404Spjd hsize >>= 1; 1139168404Spjd goto retry; 1140168404Spjd } 1141168404Spjd 1142168404Spjd hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1143168404Spjd 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1144168404Spjd buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1145185029Spjd 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1146168404Spjd 1147168404Spjd for (i = 0; i < 256; i++) 1148168404Spjd for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1149168404Spjd *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1150168404Spjd 1151168404Spjd for (i = 0; i < BUF_LOCKS; i++) { 1152168404Spjd mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1153168404Spjd NULL, MUTEX_DEFAULT, NULL); 1154168404Spjd } 1155168404Spjd} 1156168404Spjd 1157168404Spjd#define ARC_MINTIME (hz>>4) /* 62 ms */ 1158168404Spjd 1159168404Spjdstatic void 1160168404Spjdarc_cksum_verify(arc_buf_t *buf) 1161168404Spjd{ 1162168404Spjd zio_cksum_t zc; 1163168404Spjd 1164168404Spjd if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1165168404Spjd return; 1166168404Spjd 1167168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1168168404Spjd if (buf->b_hdr->b_freeze_cksum == NULL || 1169275811Sdelphij (buf->b_hdr->b_flags & ARC_FLAG_IO_ERROR)) { 1170168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1171168404Spjd return; 1172168404Spjd } 1173168404Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1174168404Spjd if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1175168404Spjd panic("buffer modified while frozen!"); 1176168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1177168404Spjd} 1178168404Spjd 1179185029Spjdstatic int 1180185029Spjdarc_cksum_equal(arc_buf_t *buf) 1181185029Spjd{ 1182185029Spjd zio_cksum_t zc; 1183185029Spjd int equal; 1184185029Spjd 1185185029Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1186185029Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1187185029Spjd equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1188185029Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1189185029Spjd 1190185029Spjd return (equal); 1191185029Spjd} 1192185029Spjd 1193168404Spjdstatic void 1194185029Spjdarc_cksum_compute(arc_buf_t *buf, boolean_t force) 1195168404Spjd{ 1196185029Spjd if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1197168404Spjd return; 1198168404Spjd 1199168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1200168404Spjd if (buf->b_hdr->b_freeze_cksum != NULL) { 1201168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1202168404Spjd return; 1203168404Spjd } 1204168404Spjd buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1205168404Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1206168404Spjd buf->b_hdr->b_freeze_cksum); 1207168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1208240133Smm#ifdef illumos 1209240133Smm arc_buf_watch(buf); 1210277300Ssmh#endif 1211168404Spjd} 1212168404Spjd 1213240133Smm#ifdef illumos 1214240133Smm#ifndef _KERNEL 1215240133Smmtypedef struct procctl { 1216240133Smm long cmd; 1217240133Smm prwatch_t prwatch; 1218240133Smm} procctl_t; 1219240133Smm#endif 1220240133Smm 1221240133Smm/* ARGSUSED */ 1222240133Smmstatic void 1223240133Smmarc_buf_unwatch(arc_buf_t *buf) 1224240133Smm{ 1225240133Smm#ifndef _KERNEL 1226240133Smm if (arc_watch) { 1227240133Smm int result; 1228240133Smm procctl_t ctl; 1229240133Smm ctl.cmd = PCWATCH; 1230240133Smm ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1231240133Smm ctl.prwatch.pr_size = 0; 1232240133Smm ctl.prwatch.pr_wflags = 0; 1233240133Smm result = write(arc_procfd, &ctl, sizeof (ctl)); 1234240133Smm ASSERT3U(result, ==, sizeof (ctl)); 1235240133Smm } 1236240133Smm#endif 1237240133Smm} 1238240133Smm 1239240133Smm/* ARGSUSED */ 1240240133Smmstatic void 1241240133Smmarc_buf_watch(arc_buf_t *buf) 1242240133Smm{ 1243240133Smm#ifndef _KERNEL 1244240133Smm if (arc_watch) { 1245240133Smm int result; 1246240133Smm procctl_t ctl; 1247240133Smm ctl.cmd = PCWATCH; 1248240133Smm ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1249240133Smm ctl.prwatch.pr_size = buf->b_hdr->b_size; 1250240133Smm ctl.prwatch.pr_wflags = WA_WRITE; 1251240133Smm result = write(arc_procfd, &ctl, sizeof (ctl)); 1252240133Smm ASSERT3U(result, ==, sizeof (ctl)); 1253240133Smm } 1254240133Smm#endif 1255240133Smm} 1256240133Smm#endif /* illumos */ 1257240133Smm 1258168404Spjdvoid 1259168404Spjdarc_buf_thaw(arc_buf_t *buf) 1260168404Spjd{ 1261185029Spjd if (zfs_flags & ZFS_DEBUG_MODIFY) { 1262185029Spjd if (buf->b_hdr->b_state != arc_anon) 1263185029Spjd panic("modifying non-anon buffer!"); 1264275811Sdelphij if (buf->b_hdr->b_flags & ARC_FLAG_IO_IN_PROGRESS) 1265185029Spjd panic("modifying buffer while i/o in progress!"); 1266185029Spjd arc_cksum_verify(buf); 1267185029Spjd } 1268168404Spjd 1269168404Spjd mutex_enter(&buf->b_hdr->b_freeze_lock); 1270168404Spjd if (buf->b_hdr->b_freeze_cksum != NULL) { 1271168404Spjd kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1272168404Spjd buf->b_hdr->b_freeze_cksum = NULL; 1273168404Spjd } 1274219089Spjd 1275219089Spjd if (zfs_flags & ZFS_DEBUG_MODIFY) { 1276219089Spjd if (buf->b_hdr->b_thawed) 1277219089Spjd kmem_free(buf->b_hdr->b_thawed, 1); 1278219089Spjd buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1279219089Spjd } 1280219089Spjd 1281168404Spjd mutex_exit(&buf->b_hdr->b_freeze_lock); 1282240133Smm 1283240133Smm#ifdef illumos 1284240133Smm arc_buf_unwatch(buf); 1285277300Ssmh#endif 1286168404Spjd} 1287168404Spjd 1288168404Spjdvoid 1289168404Spjdarc_buf_freeze(arc_buf_t *buf) 1290168404Spjd{ 1291219089Spjd kmutex_t *hash_lock; 1292219089Spjd 1293168404Spjd if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1294168404Spjd return; 1295168404Spjd 1296219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 1297219089Spjd mutex_enter(hash_lock); 1298219089Spjd 1299168404Spjd ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1300168404Spjd buf->b_hdr->b_state == arc_anon); 1301185029Spjd arc_cksum_compute(buf, B_FALSE); 1302219089Spjd mutex_exit(hash_lock); 1303240133Smm 1304168404Spjd} 1305168404Spjd 1306168404Spjdstatic void 1307275811Sdelphijget_buf_info(arc_buf_hdr_t *hdr, arc_state_t *state, list_t **list, kmutex_t **lock) 1308205231Skmacy{ 1309275811Sdelphij uint64_t buf_hashid = buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1310205231Skmacy 1311275811Sdelphij if (hdr->b_type == ARC_BUFC_METADATA) 1312206796Spjd buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1313205231Skmacy else { 1314206796Spjd buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1315205231Skmacy buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1316205231Skmacy } 1317205231Skmacy 1318205231Skmacy *list = &state->arcs_lists[buf_hashid]; 1319205231Skmacy *lock = ARCS_LOCK(state, buf_hashid); 1320205231Skmacy} 1321205231Skmacy 1322205231Skmacy 1323205231Skmacystatic void 1324275811Sdelphijadd_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1325168404Spjd{ 1326168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 1327168404Spjd 1328275811Sdelphij if ((refcount_add(&hdr->b_refcnt, tag) == 1) && 1329275811Sdelphij (hdr->b_state != arc_anon)) { 1330275811Sdelphij uint64_t delta = hdr->b_size * hdr->b_datacnt; 1331275811Sdelphij uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 1332205231Skmacy list_t *list; 1333205231Skmacy kmutex_t *lock; 1334168404Spjd 1335275811Sdelphij get_buf_info(hdr, hdr->b_state, &list, &lock); 1336205231Skmacy ASSERT(!MUTEX_HELD(lock)); 1337205231Skmacy mutex_enter(lock); 1338275811Sdelphij ASSERT(list_link_active(&hdr->b_arc_node)); 1339275811Sdelphij list_remove(list, hdr); 1340275811Sdelphij if (GHOST_STATE(hdr->b_state)) { 1341275811Sdelphij ASSERT0(hdr->b_datacnt); 1342275811Sdelphij ASSERT3P(hdr->b_buf, ==, NULL); 1343275811Sdelphij delta = hdr->b_size; 1344168404Spjd } 1345168404Spjd ASSERT(delta > 0); 1346185029Spjd ASSERT3U(*size, >=, delta); 1347185029Spjd atomic_add_64(size, -delta); 1348206794Spjd mutex_exit(lock); 1349185029Spjd /* remove the prefetch flag if we get a reference */ 1350275811Sdelphij if (hdr->b_flags & ARC_FLAG_PREFETCH) 1351275811Sdelphij hdr->b_flags &= ~ARC_FLAG_PREFETCH; 1352168404Spjd } 1353168404Spjd} 1354168404Spjd 1355168404Spjdstatic int 1356275811Sdelphijremove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1357168404Spjd{ 1358168404Spjd int cnt; 1359275811Sdelphij arc_state_t *state = hdr->b_state; 1360168404Spjd 1361168404Spjd ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1362168404Spjd ASSERT(!GHOST_STATE(state)); 1363168404Spjd 1364275811Sdelphij if (((cnt = refcount_remove(&hdr->b_refcnt, tag)) == 0) && 1365168404Spjd (state != arc_anon)) { 1366275811Sdelphij uint64_t *size = &state->arcs_lsize[hdr->b_type]; 1367205231Skmacy list_t *list; 1368205231Skmacy kmutex_t *lock; 1369185029Spjd 1370275811Sdelphij get_buf_info(hdr, state, &list, &lock); 1371205231Skmacy ASSERT(!MUTEX_HELD(lock)); 1372205231Skmacy mutex_enter(lock); 1373275811Sdelphij ASSERT(!list_link_active(&hdr->b_arc_node)); 1374275811Sdelphij list_insert_head(list, hdr); 1375275811Sdelphij ASSERT(hdr->b_datacnt > 0); 1376275811Sdelphij atomic_add_64(size, hdr->b_size * hdr->b_datacnt); 1377206794Spjd mutex_exit(lock); 1378168404Spjd } 1379168404Spjd return (cnt); 1380168404Spjd} 1381168404Spjd 1382168404Spjd/* 1383168404Spjd * Move the supplied buffer to the indicated state. The mutex 1384168404Spjd * for the buffer must be held by the caller. 1385168404Spjd */ 1386168404Spjdstatic void 1387275811Sdelphijarc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 1388275811Sdelphij kmutex_t *hash_lock) 1389168404Spjd{ 1390275811Sdelphij arc_state_t *old_state = hdr->b_state; 1391275811Sdelphij int64_t refcnt = refcount_count(&hdr->b_refcnt); 1392168404Spjd uint64_t from_delta, to_delta; 1393205231Skmacy list_t *list; 1394205231Skmacy kmutex_t *lock; 1395168404Spjd 1396168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 1397258632Savg ASSERT3P(new_state, !=, old_state); 1398275811Sdelphij ASSERT(refcnt == 0 || hdr->b_datacnt > 0); 1399275811Sdelphij ASSERT(hdr->b_datacnt == 0 || !GHOST_STATE(new_state)); 1400275811Sdelphij ASSERT(hdr->b_datacnt <= 1 || old_state != arc_anon); 1401168404Spjd 1402275811Sdelphij from_delta = to_delta = hdr->b_datacnt * hdr->b_size; 1403168404Spjd 1404168404Spjd /* 1405168404Spjd * If this buffer is evictable, transfer it from the 1406168404Spjd * old state list to the new state list. 1407168404Spjd */ 1408168404Spjd if (refcnt == 0) { 1409168404Spjd if (old_state != arc_anon) { 1410205231Skmacy int use_mutex; 1411275811Sdelphij uint64_t *size = &old_state->arcs_lsize[hdr->b_type]; 1412168404Spjd 1413275811Sdelphij get_buf_info(hdr, old_state, &list, &lock); 1414205231Skmacy use_mutex = !MUTEX_HELD(lock); 1415168404Spjd if (use_mutex) 1416205231Skmacy mutex_enter(lock); 1417168404Spjd 1418275811Sdelphij ASSERT(list_link_active(&hdr->b_arc_node)); 1419275811Sdelphij list_remove(list, hdr); 1420168404Spjd 1421168404Spjd /* 1422168404Spjd * If prefetching out of the ghost cache, 1423219089Spjd * we will have a non-zero datacnt. 1424168404Spjd */ 1425275811Sdelphij if (GHOST_STATE(old_state) && hdr->b_datacnt == 0) { 1426168404Spjd /* ghost elements have a ghost size */ 1427275811Sdelphij ASSERT(hdr->b_buf == NULL); 1428275811Sdelphij from_delta = hdr->b_size; 1429168404Spjd } 1430185029Spjd ASSERT3U(*size, >=, from_delta); 1431185029Spjd atomic_add_64(size, -from_delta); 1432168404Spjd 1433168404Spjd if (use_mutex) 1434205231Skmacy mutex_exit(lock); 1435168404Spjd } 1436168404Spjd if (new_state != arc_anon) { 1437206796Spjd int use_mutex; 1438275811Sdelphij uint64_t *size = &new_state->arcs_lsize[hdr->b_type]; 1439168404Spjd 1440275811Sdelphij get_buf_info(hdr, new_state, &list, &lock); 1441205231Skmacy use_mutex = !MUTEX_HELD(lock); 1442168404Spjd if (use_mutex) 1443205231Skmacy mutex_enter(lock); 1444168404Spjd 1445275811Sdelphij list_insert_head(list, hdr); 1446168404Spjd 1447168404Spjd /* ghost elements have a ghost size */ 1448168404Spjd if (GHOST_STATE(new_state)) { 1449275811Sdelphij ASSERT(hdr->b_datacnt == 0); 1450275811Sdelphij ASSERT(hdr->b_buf == NULL); 1451275811Sdelphij to_delta = hdr->b_size; 1452168404Spjd } 1453185029Spjd atomic_add_64(size, to_delta); 1454168404Spjd 1455168404Spjd if (use_mutex) 1456205231Skmacy mutex_exit(lock); 1457168404Spjd } 1458168404Spjd } 1459168404Spjd 1460275811Sdelphij ASSERT(!BUF_EMPTY(hdr)); 1461275811Sdelphij if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 1462275811Sdelphij buf_hash_remove(hdr); 1463168404Spjd 1464168404Spjd /* adjust state sizes */ 1465168404Spjd if (to_delta) 1466168404Spjd atomic_add_64(&new_state->arcs_size, to_delta); 1467168404Spjd if (from_delta) { 1468168404Spjd ASSERT3U(old_state->arcs_size, >=, from_delta); 1469168404Spjd atomic_add_64(&old_state->arcs_size, -from_delta); 1470168404Spjd } 1471275811Sdelphij hdr->b_state = new_state; 1472185029Spjd 1473185029Spjd /* adjust l2arc hdr stats */ 1474185029Spjd if (new_state == arc_l2c_only) 1475185029Spjd l2arc_hdr_stat_add(); 1476185029Spjd else if (old_state == arc_l2c_only) 1477185029Spjd l2arc_hdr_stat_remove(); 1478168404Spjd} 1479168404Spjd 1480185029Spjdvoid 1481208373Smmarc_space_consume(uint64_t space, arc_space_type_t type) 1482185029Spjd{ 1483208373Smm ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1484208373Smm 1485208373Smm switch (type) { 1486208373Smm case ARC_SPACE_DATA: 1487208373Smm ARCSTAT_INCR(arcstat_data_size, space); 1488208373Smm break; 1489208373Smm case ARC_SPACE_OTHER: 1490208373Smm ARCSTAT_INCR(arcstat_other_size, space); 1491208373Smm break; 1492208373Smm case ARC_SPACE_HDRS: 1493208373Smm ARCSTAT_INCR(arcstat_hdr_size, space); 1494208373Smm break; 1495208373Smm case ARC_SPACE_L2HDRS: 1496208373Smm ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1497208373Smm break; 1498208373Smm } 1499208373Smm 1500275748Sdelphij ARCSTAT_INCR(arcstat_meta_used, space); 1501185029Spjd atomic_add_64(&arc_size, space); 1502185029Spjd} 1503185029Spjd 1504185029Spjdvoid 1505208373Smmarc_space_return(uint64_t space, arc_space_type_t type) 1506185029Spjd{ 1507208373Smm ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1508208373Smm 1509208373Smm switch (type) { 1510208373Smm case ARC_SPACE_DATA: 1511208373Smm ARCSTAT_INCR(arcstat_data_size, -space); 1512208373Smm break; 1513208373Smm case ARC_SPACE_OTHER: 1514208373Smm ARCSTAT_INCR(arcstat_other_size, -space); 1515208373Smm break; 1516208373Smm case ARC_SPACE_HDRS: 1517208373Smm ARCSTAT_INCR(arcstat_hdr_size, -space); 1518208373Smm break; 1519208373Smm case ARC_SPACE_L2HDRS: 1520208373Smm ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1521208373Smm break; 1522208373Smm } 1523208373Smm 1524185029Spjd ASSERT(arc_meta_used >= space); 1525185029Spjd if (arc_meta_max < arc_meta_used) 1526185029Spjd arc_meta_max = arc_meta_used; 1527275748Sdelphij ARCSTAT_INCR(arcstat_meta_used, -space); 1528185029Spjd ASSERT(arc_size >= space); 1529185029Spjd atomic_add_64(&arc_size, -space); 1530185029Spjd} 1531185029Spjd 1532168404Spjdarc_buf_t * 1533168404Spjdarc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1534168404Spjd{ 1535168404Spjd arc_buf_hdr_t *hdr; 1536168404Spjd arc_buf_t *buf; 1537168404Spjd 1538168404Spjd ASSERT3U(size, >, 0); 1539185029Spjd hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1540168404Spjd ASSERT(BUF_EMPTY(hdr)); 1541168404Spjd hdr->b_size = size; 1542168404Spjd hdr->b_type = type; 1543228103Smm hdr->b_spa = spa_load_guid(spa); 1544168404Spjd hdr->b_state = arc_anon; 1545168404Spjd hdr->b_arc_access = 0; 1546185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1547168404Spjd buf->b_hdr = hdr; 1548168404Spjd buf->b_data = NULL; 1549168404Spjd buf->b_efunc = NULL; 1550168404Spjd buf->b_private = NULL; 1551168404Spjd buf->b_next = NULL; 1552168404Spjd hdr->b_buf = buf; 1553168404Spjd arc_get_data_buf(buf); 1554168404Spjd hdr->b_datacnt = 1; 1555168404Spjd hdr->b_flags = 0; 1556168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1557168404Spjd (void) refcount_add(&hdr->b_refcnt, tag); 1558168404Spjd 1559168404Spjd return (buf); 1560168404Spjd} 1561168404Spjd 1562209962Smmstatic char *arc_onloan_tag = "onloan"; 1563209962Smm 1564209962Smm/* 1565209962Smm * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1566209962Smm * flight data by arc_tempreserve_space() until they are "returned". Loaned 1567209962Smm * buffers must be returned to the arc before they can be used by the DMU or 1568209962Smm * freed. 1569209962Smm */ 1570209962Smmarc_buf_t * 1571209962Smmarc_loan_buf(spa_t *spa, int size) 1572209962Smm{ 1573209962Smm arc_buf_t *buf; 1574209962Smm 1575209962Smm buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1576209962Smm 1577209962Smm atomic_add_64(&arc_loaned_bytes, size); 1578209962Smm return (buf); 1579209962Smm} 1580209962Smm 1581209962Smm/* 1582209962Smm * Return a loaned arc buffer to the arc. 1583209962Smm */ 1584209962Smmvoid 1585209962Smmarc_return_buf(arc_buf_t *buf, void *tag) 1586209962Smm{ 1587209962Smm arc_buf_hdr_t *hdr = buf->b_hdr; 1588209962Smm 1589209962Smm ASSERT(buf->b_data != NULL); 1590219089Spjd (void) refcount_add(&hdr->b_refcnt, tag); 1591219089Spjd (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1592209962Smm 1593209962Smm atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1594209962Smm} 1595209962Smm 1596219089Spjd/* Detach an arc_buf from a dbuf (tag) */ 1597219089Spjdvoid 1598219089Spjdarc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1599219089Spjd{ 1600219089Spjd arc_buf_hdr_t *hdr; 1601219089Spjd 1602219089Spjd ASSERT(buf->b_data != NULL); 1603219089Spjd hdr = buf->b_hdr; 1604219089Spjd (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1605219089Spjd (void) refcount_remove(&hdr->b_refcnt, tag); 1606219089Spjd buf->b_efunc = NULL; 1607219089Spjd buf->b_private = NULL; 1608219089Spjd 1609219089Spjd atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1610219089Spjd} 1611219089Spjd 1612168404Spjdstatic arc_buf_t * 1613168404Spjdarc_buf_clone(arc_buf_t *from) 1614168404Spjd{ 1615168404Spjd arc_buf_t *buf; 1616168404Spjd arc_buf_hdr_t *hdr = from->b_hdr; 1617168404Spjd uint64_t size = hdr->b_size; 1618168404Spjd 1619219089Spjd ASSERT(hdr->b_state != arc_anon); 1620219089Spjd 1621185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1622168404Spjd buf->b_hdr = hdr; 1623168404Spjd buf->b_data = NULL; 1624168404Spjd buf->b_efunc = NULL; 1625168404Spjd buf->b_private = NULL; 1626168404Spjd buf->b_next = hdr->b_buf; 1627168404Spjd hdr->b_buf = buf; 1628168404Spjd arc_get_data_buf(buf); 1629168404Spjd bcopy(from->b_data, buf->b_data, size); 1630242845Sdelphij 1631242845Sdelphij /* 1632242845Sdelphij * This buffer already exists in the arc so create a duplicate 1633242845Sdelphij * copy for the caller. If the buffer is associated with user data 1634242845Sdelphij * then track the size and number of duplicates. These stats will be 1635242845Sdelphij * updated as duplicate buffers are created and destroyed. 1636242845Sdelphij */ 1637242845Sdelphij if (hdr->b_type == ARC_BUFC_DATA) { 1638242845Sdelphij ARCSTAT_BUMP(arcstat_duplicate_buffers); 1639242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1640242845Sdelphij } 1641168404Spjd hdr->b_datacnt += 1; 1642168404Spjd return (buf); 1643168404Spjd} 1644168404Spjd 1645168404Spjdvoid 1646168404Spjdarc_buf_add_ref(arc_buf_t *buf, void* tag) 1647168404Spjd{ 1648168404Spjd arc_buf_hdr_t *hdr; 1649168404Spjd kmutex_t *hash_lock; 1650168404Spjd 1651168404Spjd /* 1652185029Spjd * Check to see if this buffer is evicted. Callers 1653185029Spjd * must verify b_data != NULL to know if the add_ref 1654185029Spjd * was successful. 1655168404Spjd */ 1656219089Spjd mutex_enter(&buf->b_evict_lock); 1657185029Spjd if (buf->b_data == NULL) { 1658219089Spjd mutex_exit(&buf->b_evict_lock); 1659168404Spjd return; 1660168404Spjd } 1661219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 1662219089Spjd mutex_enter(hash_lock); 1663185029Spjd hdr = buf->b_hdr; 1664219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1665219089Spjd mutex_exit(&buf->b_evict_lock); 1666168404Spjd 1667168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1668168404Spjd add_reference(hdr, hash_lock, tag); 1669208373Smm DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1670168404Spjd arc_access(hdr, hash_lock); 1671168404Spjd mutex_exit(hash_lock); 1672168404Spjd ARCSTAT_BUMP(arcstat_hits); 1673275811Sdelphij ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 1674168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1675168404Spjd data, metadata, hits); 1676168404Spjd} 1677168404Spjd 1678274172Savgstatic void 1679274172Savgarc_buf_free_on_write(void *data, size_t size, 1680274172Savg void (*free_func)(void *, size_t)) 1681274172Savg{ 1682274172Savg l2arc_data_free_t *df; 1683274172Savg 1684274172Savg df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1685274172Savg df->l2df_data = data; 1686274172Savg df->l2df_size = size; 1687274172Savg df->l2df_func = free_func; 1688274172Savg mutex_enter(&l2arc_free_on_write_mtx); 1689274172Savg list_insert_head(l2arc_free_on_write, df); 1690274172Savg mutex_exit(&l2arc_free_on_write_mtx); 1691274172Savg} 1692274172Savg 1693185029Spjd/* 1694185029Spjd * Free the arc data buffer. If it is an l2arc write in progress, 1695185029Spjd * the buffer is placed on l2arc_free_on_write to be freed later. 1696185029Spjd */ 1697168404Spjdstatic void 1698240133Smmarc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1699185029Spjd{ 1700240133Smm arc_buf_hdr_t *hdr = buf->b_hdr; 1701240133Smm 1702185029Spjd if (HDR_L2_WRITING(hdr)) { 1703274172Savg arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 1704185029Spjd ARCSTAT_BUMP(arcstat_l2_free_on_write); 1705185029Spjd } else { 1706240133Smm free_func(buf->b_data, hdr->b_size); 1707185029Spjd } 1708185029Spjd} 1709185029Spjd 1710268858Sdelphij/* 1711268858Sdelphij * Free up buf->b_data and if 'remove' is set, then pull the 1712268858Sdelphij * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 1713268858Sdelphij */ 1714185029Spjdstatic void 1715274172Savgarc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 1716274172Savg{ 1717274172Savg l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1718274172Savg 1719274172Savg ASSERT(MUTEX_HELD(&l2arc_buflist_mtx)); 1720274172Savg 1721274172Savg if (l2hdr->b_tmp_cdata == NULL) 1722274172Savg return; 1723274172Savg 1724274172Savg ASSERT(HDR_L2_WRITING(hdr)); 1725274172Savg arc_buf_free_on_write(l2hdr->b_tmp_cdata, hdr->b_size, 1726274172Savg zio_data_buf_free); 1727274172Savg ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 1728274172Savg l2hdr->b_tmp_cdata = NULL; 1729274172Savg} 1730274172Savg 1731274172Savgstatic void 1732268858Sdelphijarc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 1733168404Spjd{ 1734168404Spjd arc_buf_t **bufp; 1735168404Spjd 1736168404Spjd /* free up data associated with the buf */ 1737168404Spjd if (buf->b_data) { 1738168404Spjd arc_state_t *state = buf->b_hdr->b_state; 1739168404Spjd uint64_t size = buf->b_hdr->b_size; 1740168404Spjd arc_buf_contents_t type = buf->b_hdr->b_type; 1741168404Spjd 1742168404Spjd arc_cksum_verify(buf); 1743240133Smm#ifdef illumos 1744240133Smm arc_buf_unwatch(buf); 1745277300Ssmh#endif 1746219089Spjd 1747168404Spjd if (!recycle) { 1748168404Spjd if (type == ARC_BUFC_METADATA) { 1749240133Smm arc_buf_data_free(buf, zio_buf_free); 1750208373Smm arc_space_return(size, ARC_SPACE_DATA); 1751168404Spjd } else { 1752168404Spjd ASSERT(type == ARC_BUFC_DATA); 1753240133Smm arc_buf_data_free(buf, zio_data_buf_free); 1754208373Smm ARCSTAT_INCR(arcstat_data_size, -size); 1755185029Spjd atomic_add_64(&arc_size, -size); 1756168404Spjd } 1757168404Spjd } 1758168404Spjd if (list_link_active(&buf->b_hdr->b_arc_node)) { 1759185029Spjd uint64_t *cnt = &state->arcs_lsize[type]; 1760185029Spjd 1761168404Spjd ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1762168404Spjd ASSERT(state != arc_anon); 1763185029Spjd 1764185029Spjd ASSERT3U(*cnt, >=, size); 1765185029Spjd atomic_add_64(cnt, -size); 1766168404Spjd } 1767168404Spjd ASSERT3U(state->arcs_size, >=, size); 1768168404Spjd atomic_add_64(&state->arcs_size, -size); 1769168404Spjd buf->b_data = NULL; 1770242845Sdelphij 1771242845Sdelphij /* 1772242845Sdelphij * If we're destroying a duplicate buffer make sure 1773242845Sdelphij * that the appropriate statistics are updated. 1774242845Sdelphij */ 1775242845Sdelphij if (buf->b_hdr->b_datacnt > 1 && 1776242845Sdelphij buf->b_hdr->b_type == ARC_BUFC_DATA) { 1777242845Sdelphij ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1778242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1779242845Sdelphij } 1780168404Spjd ASSERT(buf->b_hdr->b_datacnt > 0); 1781168404Spjd buf->b_hdr->b_datacnt -= 1; 1782168404Spjd } 1783168404Spjd 1784168404Spjd /* only remove the buf if requested */ 1785268858Sdelphij if (!remove) 1786168404Spjd return; 1787168404Spjd 1788168404Spjd /* remove the buf from the hdr list */ 1789168404Spjd for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1790168404Spjd continue; 1791168404Spjd *bufp = buf->b_next; 1792219089Spjd buf->b_next = NULL; 1793168404Spjd 1794168404Spjd ASSERT(buf->b_efunc == NULL); 1795168404Spjd 1796168404Spjd /* clean up the buf */ 1797168404Spjd buf->b_hdr = NULL; 1798168404Spjd kmem_cache_free(buf_cache, buf); 1799168404Spjd} 1800168404Spjd 1801168404Spjdstatic void 1802168404Spjdarc_hdr_destroy(arc_buf_hdr_t *hdr) 1803168404Spjd{ 1804168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1805168404Spjd ASSERT3P(hdr->b_state, ==, arc_anon); 1806168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1807219089Spjd l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1808168404Spjd 1809219089Spjd if (l2hdr != NULL) { 1810219089Spjd boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1811219089Spjd /* 1812219089Spjd * To prevent arc_free() and l2arc_evict() from 1813219089Spjd * attempting to free the same buffer at the same time, 1814219089Spjd * a FREE_IN_PROGRESS flag is given to arc_free() to 1815219089Spjd * give it priority. l2arc_evict() can't destroy this 1816219089Spjd * header while we are waiting on l2arc_buflist_mtx. 1817219089Spjd * 1818219089Spjd * The hdr may be removed from l2ad_buflist before we 1819219089Spjd * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1820219089Spjd */ 1821219089Spjd if (!buflist_held) { 1822185029Spjd mutex_enter(&l2arc_buflist_mtx); 1823219089Spjd l2hdr = hdr->b_l2hdr; 1824219089Spjd } 1825219089Spjd 1826219089Spjd if (l2hdr != NULL) { 1827248572Ssmh trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1828248574Ssmh hdr->b_size, 0); 1829219089Spjd list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1830274172Savg arc_buf_l2_cdata_free(hdr); 1831219089Spjd ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1832251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1833268085Sdelphij vdev_space_update(l2hdr->b_dev->l2ad_vdev, 1834268085Sdelphij -l2hdr->b_asize, 0, 0); 1835219089Spjd kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1836219089Spjd if (hdr->b_state == arc_l2c_only) 1837219089Spjd l2arc_hdr_stat_remove(); 1838219089Spjd hdr->b_l2hdr = NULL; 1839219089Spjd } 1840219089Spjd 1841219089Spjd if (!buflist_held) 1842185029Spjd mutex_exit(&l2arc_buflist_mtx); 1843185029Spjd } 1844185029Spjd 1845168404Spjd if (!BUF_EMPTY(hdr)) { 1846168404Spjd ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1847219089Spjd buf_discard_identity(hdr); 1848168404Spjd } 1849168404Spjd while (hdr->b_buf) { 1850168404Spjd arc_buf_t *buf = hdr->b_buf; 1851168404Spjd 1852168404Spjd if (buf->b_efunc) { 1853168404Spjd mutex_enter(&arc_eviction_mtx); 1854219089Spjd mutex_enter(&buf->b_evict_lock); 1855168404Spjd ASSERT(buf->b_hdr != NULL); 1856168404Spjd arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1857168404Spjd hdr->b_buf = buf->b_next; 1858168404Spjd buf->b_hdr = &arc_eviction_hdr; 1859168404Spjd buf->b_next = arc_eviction_list; 1860168404Spjd arc_eviction_list = buf; 1861219089Spjd mutex_exit(&buf->b_evict_lock); 1862168404Spjd mutex_exit(&arc_eviction_mtx); 1863168404Spjd } else { 1864168404Spjd arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1865168404Spjd } 1866168404Spjd } 1867168404Spjd if (hdr->b_freeze_cksum != NULL) { 1868168404Spjd kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1869168404Spjd hdr->b_freeze_cksum = NULL; 1870168404Spjd } 1871219089Spjd if (hdr->b_thawed) { 1872219089Spjd kmem_free(hdr->b_thawed, 1); 1873219089Spjd hdr->b_thawed = NULL; 1874219089Spjd } 1875168404Spjd 1876168404Spjd ASSERT(!list_link_active(&hdr->b_arc_node)); 1877168404Spjd ASSERT3P(hdr->b_hash_next, ==, NULL); 1878168404Spjd ASSERT3P(hdr->b_acb, ==, NULL); 1879168404Spjd kmem_cache_free(hdr_cache, hdr); 1880168404Spjd} 1881168404Spjd 1882168404Spjdvoid 1883168404Spjdarc_buf_free(arc_buf_t *buf, void *tag) 1884168404Spjd{ 1885168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 1886168404Spjd int hashed = hdr->b_state != arc_anon; 1887168404Spjd 1888168404Spjd ASSERT(buf->b_efunc == NULL); 1889168404Spjd ASSERT(buf->b_data != NULL); 1890168404Spjd 1891168404Spjd if (hashed) { 1892168404Spjd kmutex_t *hash_lock = HDR_LOCK(hdr); 1893168404Spjd 1894168404Spjd mutex_enter(hash_lock); 1895219089Spjd hdr = buf->b_hdr; 1896219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1897219089Spjd 1898168404Spjd (void) remove_reference(hdr, hash_lock, tag); 1899219089Spjd if (hdr->b_datacnt > 1) { 1900168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1901219089Spjd } else { 1902219089Spjd ASSERT(buf == hdr->b_buf); 1903219089Spjd ASSERT(buf->b_efunc == NULL); 1904275811Sdelphij hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 1905219089Spjd } 1906168404Spjd mutex_exit(hash_lock); 1907168404Spjd } else if (HDR_IO_IN_PROGRESS(hdr)) { 1908168404Spjd int destroy_hdr; 1909168404Spjd /* 1910168404Spjd * We are in the middle of an async write. Don't destroy 1911168404Spjd * this buffer unless the write completes before we finish 1912168404Spjd * decrementing the reference count. 1913168404Spjd */ 1914168404Spjd mutex_enter(&arc_eviction_mtx); 1915168404Spjd (void) remove_reference(hdr, NULL, tag); 1916168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1917168404Spjd destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1918168404Spjd mutex_exit(&arc_eviction_mtx); 1919168404Spjd if (destroy_hdr) 1920168404Spjd arc_hdr_destroy(hdr); 1921168404Spjd } else { 1922219089Spjd if (remove_reference(hdr, NULL, tag) > 0) 1923168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1924219089Spjd else 1925168404Spjd arc_hdr_destroy(hdr); 1926168404Spjd } 1927168404Spjd} 1928168404Spjd 1929248571Smmboolean_t 1930168404Spjdarc_buf_remove_ref(arc_buf_t *buf, void* tag) 1931168404Spjd{ 1932168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 1933168404Spjd kmutex_t *hash_lock = HDR_LOCK(hdr); 1934248571Smm boolean_t no_callback = (buf->b_efunc == NULL); 1935168404Spjd 1936168404Spjd if (hdr->b_state == arc_anon) { 1937219089Spjd ASSERT(hdr->b_datacnt == 1); 1938168404Spjd arc_buf_free(buf, tag); 1939168404Spjd return (no_callback); 1940168404Spjd } 1941168404Spjd 1942168404Spjd mutex_enter(hash_lock); 1943219089Spjd hdr = buf->b_hdr; 1944219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1945168404Spjd ASSERT(hdr->b_state != arc_anon); 1946168404Spjd ASSERT(buf->b_data != NULL); 1947168404Spjd 1948168404Spjd (void) remove_reference(hdr, hash_lock, tag); 1949168404Spjd if (hdr->b_datacnt > 1) { 1950168404Spjd if (no_callback) 1951168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 1952168404Spjd } else if (no_callback) { 1953168404Spjd ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1954219089Spjd ASSERT(buf->b_efunc == NULL); 1955275811Sdelphij hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 1956168404Spjd } 1957168404Spjd ASSERT(no_callback || hdr->b_datacnt > 1 || 1958168404Spjd refcount_is_zero(&hdr->b_refcnt)); 1959168404Spjd mutex_exit(hash_lock); 1960168404Spjd return (no_callback); 1961168404Spjd} 1962168404Spjd 1963168404Spjdint 1964168404Spjdarc_buf_size(arc_buf_t *buf) 1965168404Spjd{ 1966168404Spjd return (buf->b_hdr->b_size); 1967168404Spjd} 1968168404Spjd 1969168404Spjd/* 1970242845Sdelphij * Called from the DMU to determine if the current buffer should be 1971242845Sdelphij * evicted. In order to ensure proper locking, the eviction must be initiated 1972242845Sdelphij * from the DMU. Return true if the buffer is associated with user data and 1973242845Sdelphij * duplicate buffers still exist. 1974242845Sdelphij */ 1975242845Sdelphijboolean_t 1976242845Sdelphijarc_buf_eviction_needed(arc_buf_t *buf) 1977242845Sdelphij{ 1978242845Sdelphij arc_buf_hdr_t *hdr; 1979242845Sdelphij boolean_t evict_needed = B_FALSE; 1980242845Sdelphij 1981242845Sdelphij if (zfs_disable_dup_eviction) 1982242845Sdelphij return (B_FALSE); 1983242845Sdelphij 1984242845Sdelphij mutex_enter(&buf->b_evict_lock); 1985242845Sdelphij hdr = buf->b_hdr; 1986242845Sdelphij if (hdr == NULL) { 1987242845Sdelphij /* 1988242845Sdelphij * We are in arc_do_user_evicts(); let that function 1989242845Sdelphij * perform the eviction. 1990242845Sdelphij */ 1991242845Sdelphij ASSERT(buf->b_data == NULL); 1992242845Sdelphij mutex_exit(&buf->b_evict_lock); 1993242845Sdelphij return (B_FALSE); 1994242845Sdelphij } else if (buf->b_data == NULL) { 1995242845Sdelphij /* 1996242845Sdelphij * We have already been added to the arc eviction list; 1997242845Sdelphij * recommend eviction. 1998242845Sdelphij */ 1999242845Sdelphij ASSERT3P(hdr, ==, &arc_eviction_hdr); 2000242845Sdelphij mutex_exit(&buf->b_evict_lock); 2001242845Sdelphij return (B_TRUE); 2002242845Sdelphij } 2003242845Sdelphij 2004242845Sdelphij if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 2005242845Sdelphij evict_needed = B_TRUE; 2006242845Sdelphij 2007242845Sdelphij mutex_exit(&buf->b_evict_lock); 2008242845Sdelphij return (evict_needed); 2009242845Sdelphij} 2010242845Sdelphij 2011242845Sdelphij/* 2012168404Spjd * Evict buffers from list until we've removed the specified number of 2013168404Spjd * bytes. Move the removed buffers to the appropriate evict state. 2014168404Spjd * If the recycle flag is set, then attempt to "recycle" a buffer: 2015168404Spjd * - look for a buffer to evict that is `bytes' long. 2016168404Spjd * - return the data block from this buffer rather than freeing it. 2017168404Spjd * This flag is used by callers that are trying to make space for a 2018168404Spjd * new buffer in a full arc cache. 2019185029Spjd * 2020185029Spjd * This function makes a "best effort". It skips over any buffers 2021185029Spjd * it can't get a hash_lock on, and so may not catch all candidates. 2022185029Spjd * It may also return without evicting as much space as requested. 2023168404Spjd */ 2024168404Spjdstatic void * 2025209962Smmarc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 2026168404Spjd arc_buf_contents_t type) 2027168404Spjd{ 2028168404Spjd arc_state_t *evicted_state; 2029168404Spjd uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 2030205231Skmacy int64_t bytes_remaining; 2031275811Sdelphij arc_buf_hdr_t *hdr, *hdr_prev = NULL; 2032205231Skmacy list_t *evicted_list, *list, *evicted_list_start, *list_start; 2033205231Skmacy kmutex_t *lock, *evicted_lock; 2034168404Spjd kmutex_t *hash_lock; 2035168404Spjd boolean_t have_lock; 2036168404Spjd void *stolen = NULL; 2037258632Savg arc_buf_hdr_t marker = { 0 }; 2038258632Savg int count = 0; 2039205231Skmacy static int evict_metadata_offset, evict_data_offset; 2040258632Savg int i, idx, offset, list_count, lists; 2041168404Spjd 2042168404Spjd ASSERT(state == arc_mru || state == arc_mfu); 2043168404Spjd 2044168404Spjd evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2045206796Spjd 2046275780Sdelphij /* 2047275780Sdelphij * Decide which "type" (data vs metadata) to recycle from. 2048275780Sdelphij * 2049275780Sdelphij * If we are over the metadata limit, recycle from metadata. 2050275780Sdelphij * If we are under the metadata minimum, recycle from data. 2051275780Sdelphij * Otherwise, recycle from whichever type has the oldest (least 2052275780Sdelphij * recently accessed) header. This is not yet implemented. 2053275780Sdelphij */ 2054275780Sdelphij if (recycle) { 2055275780Sdelphij arc_buf_contents_t realtype; 2056275780Sdelphij if (state->arcs_lsize[ARC_BUFC_DATA] == 0) { 2057275780Sdelphij realtype = ARC_BUFC_METADATA; 2058275780Sdelphij } else if (state->arcs_lsize[ARC_BUFC_METADATA] == 0) { 2059275780Sdelphij realtype = ARC_BUFC_DATA; 2060275780Sdelphij } else if (arc_meta_used >= arc_meta_limit) { 2061275780Sdelphij realtype = ARC_BUFC_METADATA; 2062275780Sdelphij } else if (arc_meta_used <= arc_meta_min) { 2063275780Sdelphij realtype = ARC_BUFC_DATA; 2064275780Sdelphij } else { 2065275780Sdelphij#ifdef illumos 2066275780Sdelphij if (data_hdr->b_arc_access < 2067275780Sdelphij metadata_hdr->b_arc_access) { 2068275780Sdelphij realtype = ARC_BUFC_DATA; 2069275780Sdelphij } else { 2070275780Sdelphij realtype = ARC_BUFC_METADATA; 2071275780Sdelphij } 2072275780Sdelphij#else 2073275780Sdelphij /* TODO */ 2074275780Sdelphij realtype = type; 2075275780Sdelphij#endif 2076275780Sdelphij } 2077275780Sdelphij if (realtype != type) { 2078275780Sdelphij /* 2079275780Sdelphij * If we want to evict from a different list, 2080275780Sdelphij * we can not recycle, because DATA vs METADATA 2081275780Sdelphij * buffers are segregated into different kmem 2082275780Sdelphij * caches (and vmem arenas). 2083275780Sdelphij */ 2084275780Sdelphij type = realtype; 2085275780Sdelphij recycle = B_FALSE; 2086275780Sdelphij } 2087275780Sdelphij } 2088275780Sdelphij 2089205231Skmacy if (type == ARC_BUFC_METADATA) { 2090205231Skmacy offset = 0; 2091205231Skmacy list_count = ARC_BUFC_NUMMETADATALISTS; 2092205231Skmacy list_start = &state->arcs_lists[0]; 2093205231Skmacy evicted_list_start = &evicted_state->arcs_lists[0]; 2094205231Skmacy idx = evict_metadata_offset; 2095205231Skmacy } else { 2096205231Skmacy offset = ARC_BUFC_NUMMETADATALISTS; 2097205231Skmacy list_start = &state->arcs_lists[offset]; 2098205231Skmacy evicted_list_start = &evicted_state->arcs_lists[offset]; 2099205231Skmacy list_count = ARC_BUFC_NUMDATALISTS; 2100205231Skmacy idx = evict_data_offset; 2101205231Skmacy } 2102205231Skmacy bytes_remaining = evicted_state->arcs_lsize[type]; 2103258632Savg lists = 0; 2104206796Spjd 2105205231Skmacyevict_start: 2106205231Skmacy list = &list_start[idx]; 2107205231Skmacy evicted_list = &evicted_list_start[idx]; 2108205231Skmacy lock = ARCS_LOCK(state, (offset + idx)); 2109206796Spjd evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 2110168404Spjd 2111205231Skmacy mutex_enter(lock); 2112205231Skmacy mutex_enter(evicted_lock); 2113205231Skmacy 2114275811Sdelphij for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2115275811Sdelphij hdr_prev = list_prev(list, hdr); 2116275811Sdelphij bytes_remaining -= (hdr->b_size * hdr->b_datacnt); 2117168404Spjd /* prefetch buffers have a minimum lifespan */ 2118275811Sdelphij if (HDR_IO_IN_PROGRESS(hdr) || 2119275811Sdelphij (spa && hdr->b_spa != spa) || 2120275811Sdelphij (hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT) && 2121275811Sdelphij ddi_get_lbolt() - hdr->b_arc_access < 2122219089Spjd arc_min_prefetch_lifespan)) { 2123168404Spjd skipped++; 2124168404Spjd continue; 2125168404Spjd } 2126168404Spjd /* "lookahead" for better eviction candidate */ 2127275811Sdelphij if (recycle && hdr->b_size != bytes && 2128275811Sdelphij hdr_prev && hdr_prev->b_size == bytes) 2129168404Spjd continue; 2130258632Savg 2131258632Savg /* ignore markers */ 2132275811Sdelphij if (hdr->b_spa == 0) 2133258632Savg continue; 2134258632Savg 2135258632Savg /* 2136258632Savg * It may take a long time to evict all the bufs requested. 2137258632Savg * To avoid blocking all arc activity, periodically drop 2138258632Savg * the arcs_mtx and give other threads a chance to run 2139258632Savg * before reacquiring the lock. 2140258632Savg * 2141258632Savg * If we are looking for a buffer to recycle, we are in 2142258632Savg * the hot code path, so don't sleep. 2143258632Savg */ 2144258632Savg if (!recycle && count++ > arc_evict_iterations) { 2145275811Sdelphij list_insert_after(list, hdr, &marker); 2146258632Savg mutex_exit(evicted_lock); 2147258632Savg mutex_exit(lock); 2148258632Savg kpreempt(KPREEMPT_SYNC); 2149258632Savg mutex_enter(lock); 2150258632Savg mutex_enter(evicted_lock); 2151275811Sdelphij hdr_prev = list_prev(list, &marker); 2152258632Savg list_remove(list, &marker); 2153258632Savg count = 0; 2154258632Savg continue; 2155258632Savg } 2156258632Savg 2157275811Sdelphij hash_lock = HDR_LOCK(hdr); 2158168404Spjd have_lock = MUTEX_HELD(hash_lock); 2159168404Spjd if (have_lock || mutex_tryenter(hash_lock)) { 2160275811Sdelphij ASSERT0(refcount_count(&hdr->b_refcnt)); 2161275811Sdelphij ASSERT(hdr->b_datacnt > 0); 2162275811Sdelphij while (hdr->b_buf) { 2163275811Sdelphij arc_buf_t *buf = hdr->b_buf; 2164219089Spjd if (!mutex_tryenter(&buf->b_evict_lock)) { 2165185029Spjd missed += 1; 2166185029Spjd break; 2167185029Spjd } 2168168404Spjd if (buf->b_data) { 2169275811Sdelphij bytes_evicted += hdr->b_size; 2170275811Sdelphij if (recycle && hdr->b_type == type && 2171275811Sdelphij hdr->b_size == bytes && 2172275811Sdelphij !HDR_L2_WRITING(hdr)) { 2173168404Spjd stolen = buf->b_data; 2174168404Spjd recycle = FALSE; 2175168404Spjd } 2176168404Spjd } 2177168404Spjd if (buf->b_efunc) { 2178168404Spjd mutex_enter(&arc_eviction_mtx); 2179168404Spjd arc_buf_destroy(buf, 2180168404Spjd buf->b_data == stolen, FALSE); 2181275811Sdelphij hdr->b_buf = buf->b_next; 2182168404Spjd buf->b_hdr = &arc_eviction_hdr; 2183168404Spjd buf->b_next = arc_eviction_list; 2184168404Spjd arc_eviction_list = buf; 2185168404Spjd mutex_exit(&arc_eviction_mtx); 2186219089Spjd mutex_exit(&buf->b_evict_lock); 2187168404Spjd } else { 2188219089Spjd mutex_exit(&buf->b_evict_lock); 2189168404Spjd arc_buf_destroy(buf, 2190168404Spjd buf->b_data == stolen, TRUE); 2191168404Spjd } 2192168404Spjd } 2193208373Smm 2194275811Sdelphij if (hdr->b_l2hdr) { 2195208373Smm ARCSTAT_INCR(arcstat_evict_l2_cached, 2196275811Sdelphij hdr->b_size); 2197208373Smm } else { 2198275811Sdelphij if (l2arc_write_eligible(hdr->b_spa, hdr)) { 2199208373Smm ARCSTAT_INCR(arcstat_evict_l2_eligible, 2200275811Sdelphij hdr->b_size); 2201208373Smm } else { 2202208373Smm ARCSTAT_INCR( 2203208373Smm arcstat_evict_l2_ineligible, 2204275811Sdelphij hdr->b_size); 2205208373Smm } 2206208373Smm } 2207208373Smm 2208275811Sdelphij if (hdr->b_datacnt == 0) { 2209275811Sdelphij arc_change_state(evicted_state, hdr, hash_lock); 2210275811Sdelphij ASSERT(HDR_IN_HASH_TABLE(hdr)); 2211275811Sdelphij hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 2212275811Sdelphij hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 2213275811Sdelphij DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 2214185029Spjd } 2215168404Spjd if (!have_lock) 2216168404Spjd mutex_exit(hash_lock); 2217168404Spjd if (bytes >= 0 && bytes_evicted >= bytes) 2218168404Spjd break; 2219205231Skmacy if (bytes_remaining > 0) { 2220205231Skmacy mutex_exit(evicted_lock); 2221205231Skmacy mutex_exit(lock); 2222206796Spjd idx = ((idx + 1) & (list_count - 1)); 2223258632Savg lists++; 2224205231Skmacy goto evict_start; 2225205231Skmacy } 2226168404Spjd } else { 2227168404Spjd missed += 1; 2228168404Spjd } 2229168404Spjd } 2230168404Spjd 2231205231Skmacy mutex_exit(evicted_lock); 2232205231Skmacy mutex_exit(lock); 2233206796Spjd 2234206796Spjd idx = ((idx + 1) & (list_count - 1)); 2235258632Savg lists++; 2236168404Spjd 2237205231Skmacy if (bytes_evicted < bytes) { 2238258632Savg if (lists < list_count) 2239205231Skmacy goto evict_start; 2240205231Skmacy else 2241205231Skmacy dprintf("only evicted %lld bytes from %x", 2242205231Skmacy (longlong_t)bytes_evicted, state); 2243205231Skmacy } 2244206796Spjd if (type == ARC_BUFC_METADATA) 2245205231Skmacy evict_metadata_offset = idx; 2246205231Skmacy else 2247205231Skmacy evict_data_offset = idx; 2248206796Spjd 2249168404Spjd if (skipped) 2250168404Spjd ARCSTAT_INCR(arcstat_evict_skip, skipped); 2251168404Spjd 2252168404Spjd if (missed) 2253168404Spjd ARCSTAT_INCR(arcstat_mutex_miss, missed); 2254168404Spjd 2255185029Spjd /* 2256258632Savg * Note: we have just evicted some data into the ghost state, 2257258632Savg * potentially putting the ghost size over the desired size. Rather 2258258632Savg * that evicting from the ghost list in this hot code path, leave 2259258632Savg * this chore to the arc_reclaim_thread(). 2260185029Spjd */ 2261185029Spjd 2262205231Skmacy if (stolen) 2263205231Skmacy ARCSTAT_BUMP(arcstat_stolen); 2264168404Spjd return (stolen); 2265168404Spjd} 2266168404Spjd 2267168404Spjd/* 2268168404Spjd * Remove buffers from list until we've removed the specified number of 2269168404Spjd * bytes. Destroy the buffers that are removed. 2270168404Spjd */ 2271168404Spjdstatic void 2272209962Smmarc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2273168404Spjd{ 2274275811Sdelphij arc_buf_hdr_t *hdr, *hdr_prev; 2275219089Spjd arc_buf_hdr_t marker = { 0 }; 2276205231Skmacy list_t *list, *list_start; 2277205231Skmacy kmutex_t *hash_lock, *lock; 2278168404Spjd uint64_t bytes_deleted = 0; 2279168404Spjd uint64_t bufs_skipped = 0; 2280258632Savg int count = 0; 2281205231Skmacy static int evict_offset; 2282205231Skmacy int list_count, idx = evict_offset; 2283258632Savg int offset, lists = 0; 2284168404Spjd 2285168404Spjd ASSERT(GHOST_STATE(state)); 2286205231Skmacy 2287205231Skmacy /* 2288205231Skmacy * data lists come after metadata lists 2289205231Skmacy */ 2290205231Skmacy list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2291205231Skmacy list_count = ARC_BUFC_NUMDATALISTS; 2292205231Skmacy offset = ARC_BUFC_NUMMETADATALISTS; 2293206796Spjd 2294205231Skmacyevict_start: 2295205231Skmacy list = &list_start[idx]; 2296205231Skmacy lock = ARCS_LOCK(state, idx + offset); 2297205231Skmacy 2298205231Skmacy mutex_enter(lock); 2299275811Sdelphij for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2300275811Sdelphij hdr_prev = list_prev(list, hdr); 2301275811Sdelphij if (hdr->b_type > ARC_BUFC_NUMTYPES) 2302275811Sdelphij panic("invalid hdr=%p", (void *)hdr); 2303275811Sdelphij if (spa && hdr->b_spa != spa) 2304185029Spjd continue; 2305219089Spjd 2306219089Spjd /* ignore markers */ 2307275811Sdelphij if (hdr->b_spa == 0) 2308219089Spjd continue; 2309219089Spjd 2310275811Sdelphij hash_lock = HDR_LOCK(hdr); 2311219089Spjd /* caller may be trying to modify this buffer, skip it */ 2312219089Spjd if (MUTEX_HELD(hash_lock)) 2313219089Spjd continue; 2314258632Savg 2315258632Savg /* 2316258632Savg * It may take a long time to evict all the bufs requested. 2317258632Savg * To avoid blocking all arc activity, periodically drop 2318258632Savg * the arcs_mtx and give other threads a chance to run 2319258632Savg * before reacquiring the lock. 2320258632Savg */ 2321258632Savg if (count++ > arc_evict_iterations) { 2322275811Sdelphij list_insert_after(list, hdr, &marker); 2323258632Savg mutex_exit(lock); 2324258632Savg kpreempt(KPREEMPT_SYNC); 2325258632Savg mutex_enter(lock); 2326275811Sdelphij hdr_prev = list_prev(list, &marker); 2327258632Savg list_remove(list, &marker); 2328258632Savg count = 0; 2329258632Savg continue; 2330258632Savg } 2331168404Spjd if (mutex_tryenter(hash_lock)) { 2332275811Sdelphij ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2333275811Sdelphij ASSERT(hdr->b_buf == NULL); 2334168404Spjd ARCSTAT_BUMP(arcstat_deleted); 2335275811Sdelphij bytes_deleted += hdr->b_size; 2336185029Spjd 2337275811Sdelphij if (hdr->b_l2hdr != NULL) { 2338185029Spjd /* 2339185029Spjd * This buffer is cached on the 2nd Level ARC; 2340185029Spjd * don't destroy the header. 2341185029Spjd */ 2342275811Sdelphij arc_change_state(arc_l2c_only, hdr, hash_lock); 2343185029Spjd mutex_exit(hash_lock); 2344185029Spjd } else { 2345275811Sdelphij arc_change_state(arc_anon, hdr, hash_lock); 2346185029Spjd mutex_exit(hash_lock); 2347275811Sdelphij arc_hdr_destroy(hdr); 2348185029Spjd } 2349185029Spjd 2350275811Sdelphij DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 2351168404Spjd if (bytes >= 0 && bytes_deleted >= bytes) 2352168404Spjd break; 2353219089Spjd } else if (bytes < 0) { 2354219089Spjd /* 2355219089Spjd * Insert a list marker and then wait for the 2356219089Spjd * hash lock to become available. Once its 2357219089Spjd * available, restart from where we left off. 2358219089Spjd */ 2359275811Sdelphij list_insert_after(list, hdr, &marker); 2360219089Spjd mutex_exit(lock); 2361219089Spjd mutex_enter(hash_lock); 2362219089Spjd mutex_exit(hash_lock); 2363219089Spjd mutex_enter(lock); 2364275811Sdelphij hdr_prev = list_prev(list, &marker); 2365219089Spjd list_remove(list, &marker); 2366258632Savg } else { 2367168404Spjd bufs_skipped += 1; 2368258632Savg } 2369258632Savg 2370168404Spjd } 2371205231Skmacy mutex_exit(lock); 2372206796Spjd idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2373258632Savg lists++; 2374206796Spjd 2375258632Savg if (lists < list_count) 2376205231Skmacy goto evict_start; 2377206796Spjd 2378205231Skmacy evict_offset = idx; 2379205231Skmacy if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2380185029Spjd (bytes < 0 || bytes_deleted < bytes)) { 2381205231Skmacy list_start = &state->arcs_lists[0]; 2382205231Skmacy list_count = ARC_BUFC_NUMMETADATALISTS; 2383258632Savg offset = lists = 0; 2384205231Skmacy goto evict_start; 2385185029Spjd } 2386185029Spjd 2387168404Spjd if (bufs_skipped) { 2388168404Spjd ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2389168404Spjd ASSERT(bytes >= 0); 2390168404Spjd } 2391168404Spjd 2392168404Spjd if (bytes_deleted < bytes) 2393168404Spjd dprintf("only deleted %lld bytes from %p", 2394168404Spjd (longlong_t)bytes_deleted, state); 2395168404Spjd} 2396168404Spjd 2397168404Spjdstatic void 2398168404Spjdarc_adjust(void) 2399168404Spjd{ 2400208373Smm int64_t adjustment, delta; 2401168404Spjd 2402208373Smm /* 2403208373Smm * Adjust MRU size 2404208373Smm */ 2405168404Spjd 2406209275Smm adjustment = MIN((int64_t)(arc_size - arc_c), 2407209275Smm (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2408209275Smm arc_p)); 2409208373Smm 2410208373Smm if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2411208373Smm delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2412209962Smm (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2413208373Smm adjustment -= delta; 2414168404Spjd } 2415168404Spjd 2416208373Smm if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2417208373Smm delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2418209962Smm (void) arc_evict(arc_mru, 0, delta, FALSE, 2419185029Spjd ARC_BUFC_METADATA); 2420185029Spjd } 2421185029Spjd 2422208373Smm /* 2423208373Smm * Adjust MFU size 2424208373Smm */ 2425168404Spjd 2426208373Smm adjustment = arc_size - arc_c; 2427208373Smm 2428208373Smm if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2429208373Smm delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2430209962Smm (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2431208373Smm adjustment -= delta; 2432168404Spjd } 2433168404Spjd 2434208373Smm if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2435208373Smm int64_t delta = MIN(adjustment, 2436208373Smm arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2437209962Smm (void) arc_evict(arc_mfu, 0, delta, FALSE, 2438208373Smm ARC_BUFC_METADATA); 2439208373Smm } 2440168404Spjd 2441208373Smm /* 2442208373Smm * Adjust ghost lists 2443208373Smm */ 2444168404Spjd 2445208373Smm adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2446168404Spjd 2447208373Smm if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2448208373Smm delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2449209962Smm arc_evict_ghost(arc_mru_ghost, 0, delta); 2450208373Smm } 2451185029Spjd 2452208373Smm adjustment = 2453208373Smm arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2454208373Smm 2455208373Smm if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2456208373Smm delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2457209962Smm arc_evict_ghost(arc_mfu_ghost, 0, delta); 2458168404Spjd } 2459168404Spjd} 2460168404Spjd 2461168404Spjdstatic void 2462168404Spjdarc_do_user_evicts(void) 2463168404Spjd{ 2464191903Skmacy static arc_buf_t *tmp_arc_eviction_list; 2465191903Skmacy 2466191903Skmacy /* 2467191903Skmacy * Move list over to avoid LOR 2468191903Skmacy */ 2469206796Spjdrestart: 2470168404Spjd mutex_enter(&arc_eviction_mtx); 2471191903Skmacy tmp_arc_eviction_list = arc_eviction_list; 2472191903Skmacy arc_eviction_list = NULL; 2473191903Skmacy mutex_exit(&arc_eviction_mtx); 2474191903Skmacy 2475191903Skmacy while (tmp_arc_eviction_list != NULL) { 2476191903Skmacy arc_buf_t *buf = tmp_arc_eviction_list; 2477191903Skmacy tmp_arc_eviction_list = buf->b_next; 2478219089Spjd mutex_enter(&buf->b_evict_lock); 2479168404Spjd buf->b_hdr = NULL; 2480219089Spjd mutex_exit(&buf->b_evict_lock); 2481168404Spjd 2482168404Spjd if (buf->b_efunc != NULL) 2483268858Sdelphij VERIFY0(buf->b_efunc(buf->b_private)); 2484168404Spjd 2485168404Spjd buf->b_efunc = NULL; 2486168404Spjd buf->b_private = NULL; 2487168404Spjd kmem_cache_free(buf_cache, buf); 2488168404Spjd } 2489191903Skmacy 2490191903Skmacy if (arc_eviction_list != NULL) 2491191903Skmacy goto restart; 2492168404Spjd} 2493168404Spjd 2494168404Spjd/* 2495185029Spjd * Flush all *evictable* data from the cache for the given spa. 2496168404Spjd * NOTE: this will not touch "active" (i.e. referenced) data. 2497168404Spjd */ 2498168404Spjdvoid 2499185029Spjdarc_flush(spa_t *spa) 2500168404Spjd{ 2501209962Smm uint64_t guid = 0; 2502209962Smm 2503209962Smm if (spa) 2504228103Smm guid = spa_load_guid(spa); 2505209962Smm 2506205231Skmacy while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2507209962Smm (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2508185029Spjd if (spa) 2509185029Spjd break; 2510185029Spjd } 2511205231Skmacy while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2512209962Smm (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2513185029Spjd if (spa) 2514185029Spjd break; 2515185029Spjd } 2516205231Skmacy while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2517209962Smm (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2518185029Spjd if (spa) 2519185029Spjd break; 2520185029Spjd } 2521205231Skmacy while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2522209962Smm (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2523185029Spjd if (spa) 2524185029Spjd break; 2525185029Spjd } 2526168404Spjd 2527209962Smm arc_evict_ghost(arc_mru_ghost, guid, -1); 2528209962Smm arc_evict_ghost(arc_mfu_ghost, guid, -1); 2529168404Spjd 2530168404Spjd mutex_enter(&arc_reclaim_thr_lock); 2531168404Spjd arc_do_user_evicts(); 2532168404Spjd mutex_exit(&arc_reclaim_thr_lock); 2533185029Spjd ASSERT(spa || arc_eviction_list == NULL); 2534168404Spjd} 2535168404Spjd 2536168404Spjdvoid 2537168404Spjdarc_shrink(void) 2538168404Spjd{ 2539270759Ssmh 2540168404Spjd if (arc_c > arc_c_min) { 2541168404Spjd uint64_t to_free; 2542168404Spjd 2543277452Swill to_free = arc_c >> arc_shrink_shift; 2544272483Ssmh DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 2545272483Ssmh arc_c_min, uint64_t, arc_p, uint64_t, to_free); 2546168404Spjd if (arc_c > arc_c_min + to_free) 2547168404Spjd atomic_add_64(&arc_c, -to_free); 2548168404Spjd else 2549168404Spjd arc_c = arc_c_min; 2550168404Spjd 2551168404Spjd atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2552168404Spjd if (arc_c > arc_size) 2553168404Spjd arc_c = MAX(arc_size, arc_c_min); 2554168404Spjd if (arc_p > arc_c) 2555168404Spjd arc_p = (arc_c >> 1); 2556272483Ssmh 2557272483Ssmh DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 2558272483Ssmh arc_p); 2559272483Ssmh 2560168404Spjd ASSERT(arc_c >= arc_c_min); 2561168404Spjd ASSERT((int64_t)arc_p >= 0); 2562168404Spjd } 2563168404Spjd 2564270759Ssmh if (arc_size > arc_c) { 2565270759Ssmh DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 2566270759Ssmh uint64_t, arc_c); 2567168404Spjd arc_adjust(); 2568270759Ssmh } 2569168404Spjd} 2570168404Spjd 2571185029Spjdstatic int needfree = 0; 2572168404Spjd 2573168404Spjdstatic int 2574168404Spjdarc_reclaim_needed(void) 2575168404Spjd{ 2576168404Spjd 2577168404Spjd#ifdef _KERNEL 2578219089Spjd 2579270759Ssmh if (needfree) { 2580270759Ssmh DTRACE_PROBE(arc__reclaim_needfree); 2581197816Skmacy return (1); 2582270759Ssmh } 2583168404Spjd 2584191902Skmacy /* 2585212780Savg * Cooperate with pagedaemon when it's time for it to scan 2586212780Savg * and reclaim some pages. 2587191902Skmacy */ 2588272483Ssmh if (freemem < zfs_arc_free_target) { 2589272483Ssmh DTRACE_PROBE2(arc__reclaim_freemem, uint64_t, 2590272483Ssmh freemem, uint64_t, zfs_arc_free_target); 2591191902Skmacy return (1); 2592270759Ssmh } 2593191902Skmacy 2594277300Ssmh#ifdef illumos 2595168404Spjd /* 2596185029Spjd * take 'desfree' extra pages, so we reclaim sooner, rather than later 2597185029Spjd */ 2598185029Spjd extra = desfree; 2599185029Spjd 2600185029Spjd /* 2601185029Spjd * check that we're out of range of the pageout scanner. It starts to 2602185029Spjd * schedule paging if freemem is less than lotsfree and needfree. 2603185029Spjd * lotsfree is the high-water mark for pageout, and needfree is the 2604185029Spjd * number of needed free pages. We add extra pages here to make sure 2605185029Spjd * the scanner doesn't start up while we're freeing memory. 2606185029Spjd */ 2607185029Spjd if (freemem < lotsfree + needfree + extra) 2608185029Spjd return (1); 2609185029Spjd 2610185029Spjd /* 2611168404Spjd * check to make sure that swapfs has enough space so that anon 2612185029Spjd * reservations can still succeed. anon_resvmem() checks that the 2613168404Spjd * availrmem is greater than swapfs_minfree, and the number of reserved 2614168404Spjd * swap pages. We also add a bit of extra here just to prevent 2615168404Spjd * circumstances from getting really dire. 2616168404Spjd */ 2617168404Spjd if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2618168404Spjd return (1); 2619168404Spjd 2620168404Spjd /* 2621272483Ssmh * Check that we have enough availrmem that memory locking (e.g., via 2622272483Ssmh * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 2623272483Ssmh * stores the number of pages that cannot be locked; when availrmem 2624272483Ssmh * drops below pages_pp_maximum, page locking mechanisms such as 2625272483Ssmh * page_pp_lock() will fail.) 2626272483Ssmh */ 2627272483Ssmh if (availrmem <= pages_pp_maximum) 2628272483Ssmh return (1); 2629272483Ssmh 2630277300Ssmh#endif /* illumos */ 2631272483Ssmh#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 2632272483Ssmh /* 2633168404Spjd * If we're on an i386 platform, it's possible that we'll exhaust the 2634168404Spjd * kernel heap space before we ever run out of available physical 2635168404Spjd * memory. Most checks of the size of the heap_area compare against 2636168404Spjd * tune.t_minarmem, which is the minimum available real memory that we 2637168404Spjd * can have in the system. However, this is generally fixed at 25 pages 2638168404Spjd * which is so low that it's useless. In this comparison, we seek to 2639168404Spjd * calculate the total heap-size, and reclaim if more than 3/4ths of the 2640185029Spjd * heap is allocated. (Or, in the calculation, if less than 1/4th is 2641168404Spjd * free) 2642168404Spjd */ 2643272483Ssmh if (vmem_size(heap_arena, VMEM_FREE) < 2644272483Ssmh (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2)) { 2645270861Ssmh DTRACE_PROBE2(arc__reclaim_used, uint64_t, 2646272483Ssmh vmem_size(heap_arena, VMEM_FREE), uint64_t, 2647272483Ssmh (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2); 2648270861Ssmh return (1); 2649270861Ssmh } 2650270861Ssmh#endif 2651277300Ssmh#ifdef illumos 2652272483Ssmh /* 2653272483Ssmh * If zio data pages are being allocated out of a separate heap segment, 2654272483Ssmh * then enforce that the size of available vmem for this arena remains 2655272483Ssmh * above about 1/16th free. 2656272483Ssmh * 2657272483Ssmh * Note: The 1/16th arena free requirement was put in place 2658272483Ssmh * to aggressively evict memory from the arc in order to avoid 2659272483Ssmh * memory fragmentation issues. 2660272483Ssmh */ 2661272483Ssmh if (zio_arena != NULL && 2662272483Ssmh vmem_size(zio_arena, VMEM_FREE) < 2663272483Ssmh (vmem_size(zio_arena, VMEM_ALLOC) >> 4)) 2664272483Ssmh return (1); 2665277300Ssmh#endif /* illumos */ 2666272483Ssmh#else /* _KERNEL */ 2667168404Spjd if (spa_get_random(100) == 0) 2668168404Spjd return (1); 2669272483Ssmh#endif /* _KERNEL */ 2670270759Ssmh DTRACE_PROBE(arc__reclaim_no); 2671270759Ssmh 2672168404Spjd return (0); 2673168404Spjd} 2674168404Spjd 2675208454Spjdextern kmem_cache_t *zio_buf_cache[]; 2676208454Spjdextern kmem_cache_t *zio_data_buf_cache[]; 2677272527Sdelphijextern kmem_cache_t *range_seg_cache; 2678208454Spjd 2679272483Ssmhstatic void __noinline 2680168404Spjdarc_kmem_reap_now(arc_reclaim_strategy_t strat) 2681168404Spjd{ 2682168404Spjd size_t i; 2683168404Spjd kmem_cache_t *prev_cache = NULL; 2684168404Spjd kmem_cache_t *prev_data_cache = NULL; 2685168404Spjd 2686272483Ssmh DTRACE_PROBE(arc__kmem_reap_start); 2687168404Spjd#ifdef _KERNEL 2688185029Spjd if (arc_meta_used >= arc_meta_limit) { 2689185029Spjd /* 2690185029Spjd * We are exceeding our meta-data cache limit. 2691185029Spjd * Purge some DNLC entries to release holds on meta-data. 2692185029Spjd */ 2693185029Spjd dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2694185029Spjd } 2695168404Spjd#if defined(__i386) 2696168404Spjd /* 2697168404Spjd * Reclaim unused memory from all kmem caches. 2698168404Spjd */ 2699168404Spjd kmem_reap(); 2700168404Spjd#endif 2701168404Spjd#endif 2702168404Spjd 2703168404Spjd /* 2704185029Spjd * An aggressive reclamation will shrink the cache size as well as 2705168404Spjd * reap free buffers from the arc kmem caches. 2706168404Spjd */ 2707168404Spjd if (strat == ARC_RECLAIM_AGGR) 2708168404Spjd arc_shrink(); 2709168404Spjd 2710168404Spjd for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2711168404Spjd if (zio_buf_cache[i] != prev_cache) { 2712168404Spjd prev_cache = zio_buf_cache[i]; 2713168404Spjd kmem_cache_reap_now(zio_buf_cache[i]); 2714168404Spjd } 2715168404Spjd if (zio_data_buf_cache[i] != prev_data_cache) { 2716168404Spjd prev_data_cache = zio_data_buf_cache[i]; 2717168404Spjd kmem_cache_reap_now(zio_data_buf_cache[i]); 2718168404Spjd } 2719168404Spjd } 2720168404Spjd kmem_cache_reap_now(buf_cache); 2721168404Spjd kmem_cache_reap_now(hdr_cache); 2722272506Sdelphij kmem_cache_reap_now(range_seg_cache); 2723272483Ssmh 2724277300Ssmh#ifdef illumos 2725272483Ssmh /* 2726272483Ssmh * Ask the vmem arena to reclaim unused memory from its 2727272483Ssmh * quantum caches. 2728272483Ssmh */ 2729272483Ssmh if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR) 2730272483Ssmh vmem_qcache_reap(zio_arena); 2731272483Ssmh#endif 2732272483Ssmh DTRACE_PROBE(arc__kmem_reap_end); 2733168404Spjd} 2734168404Spjd 2735168404Spjdstatic void 2736168404Spjdarc_reclaim_thread(void *dummy __unused) 2737168404Spjd{ 2738168404Spjd clock_t growtime = 0; 2739168404Spjd arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2740168404Spjd callb_cpr_t cpr; 2741168404Spjd 2742168404Spjd CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2743168404Spjd 2744168404Spjd mutex_enter(&arc_reclaim_thr_lock); 2745168404Spjd while (arc_thread_exit == 0) { 2746168404Spjd if (arc_reclaim_needed()) { 2747168404Spjd 2748168404Spjd if (arc_no_grow) { 2749168404Spjd if (last_reclaim == ARC_RECLAIM_CONS) { 2750272483Ssmh DTRACE_PROBE(arc__reclaim_aggr_no_grow); 2751168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2752168404Spjd } else { 2753168404Spjd last_reclaim = ARC_RECLAIM_CONS; 2754168404Spjd } 2755168404Spjd } else { 2756168404Spjd arc_no_grow = TRUE; 2757168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2758272483Ssmh DTRACE_PROBE(arc__reclaim_aggr); 2759168404Spjd membar_producer(); 2760168404Spjd } 2761168404Spjd 2762168404Spjd /* reset the growth delay for every reclaim */ 2763219089Spjd growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2764168404Spjd 2765185029Spjd if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2766168404Spjd /* 2767185029Spjd * If needfree is TRUE our vm_lowmem hook 2768168404Spjd * was called and in that case we must free some 2769168404Spjd * memory, so switch to aggressive mode. 2770168404Spjd */ 2771168404Spjd arc_no_grow = TRUE; 2772168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 2773168404Spjd } 2774168404Spjd arc_kmem_reap_now(last_reclaim); 2775185029Spjd arc_warm = B_TRUE; 2776185029Spjd 2777219089Spjd } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2778168404Spjd arc_no_grow = FALSE; 2779168404Spjd } 2780168404Spjd 2781209275Smm arc_adjust(); 2782168404Spjd 2783168404Spjd if (arc_eviction_list != NULL) 2784168404Spjd arc_do_user_evicts(); 2785168404Spjd 2786211762Savg#ifdef _KERNEL 2787211762Savg if (needfree) { 2788185029Spjd needfree = 0; 2789185029Spjd wakeup(&needfree); 2790211762Savg } 2791168404Spjd#endif 2792168404Spjd 2793168404Spjd /* block until needed, or one second, whichever is shorter */ 2794168404Spjd CALLB_CPR_SAFE_BEGIN(&cpr); 2795168404Spjd (void) cv_timedwait(&arc_reclaim_thr_cv, 2796168404Spjd &arc_reclaim_thr_lock, hz); 2797168404Spjd CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2798168404Spjd } 2799168404Spjd 2800168404Spjd arc_thread_exit = 0; 2801168404Spjd cv_broadcast(&arc_reclaim_thr_cv); 2802168404Spjd CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2803168404Spjd thread_exit(); 2804168404Spjd} 2805168404Spjd 2806168404Spjd/* 2807168404Spjd * Adapt arc info given the number of bytes we are trying to add and 2808168404Spjd * the state that we are comming from. This function is only called 2809168404Spjd * when we are adding new content to the cache. 2810168404Spjd */ 2811168404Spjdstatic void 2812168404Spjdarc_adapt(int bytes, arc_state_t *state) 2813168404Spjd{ 2814168404Spjd int mult; 2815208373Smm uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2816168404Spjd 2817185029Spjd if (state == arc_l2c_only) 2818185029Spjd return; 2819185029Spjd 2820168404Spjd ASSERT(bytes > 0); 2821168404Spjd /* 2822168404Spjd * Adapt the target size of the MRU list: 2823168404Spjd * - if we just hit in the MRU ghost list, then increase 2824168404Spjd * the target size of the MRU list. 2825168404Spjd * - if we just hit in the MFU ghost list, then increase 2826168404Spjd * the target size of the MFU list by decreasing the 2827168404Spjd * target size of the MRU list. 2828168404Spjd */ 2829168404Spjd if (state == arc_mru_ghost) { 2830168404Spjd mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2831168404Spjd 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2832209275Smm mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2833168404Spjd 2834208373Smm arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2835168404Spjd } else if (state == arc_mfu_ghost) { 2836208373Smm uint64_t delta; 2837208373Smm 2838168404Spjd mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2839168404Spjd 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2840209275Smm mult = MIN(mult, 10); 2841168404Spjd 2842208373Smm delta = MIN(bytes * mult, arc_p); 2843208373Smm arc_p = MAX(arc_p_min, arc_p - delta); 2844168404Spjd } 2845168404Spjd ASSERT((int64_t)arc_p >= 0); 2846168404Spjd 2847168404Spjd if (arc_reclaim_needed()) { 2848168404Spjd cv_signal(&arc_reclaim_thr_cv); 2849168404Spjd return; 2850168404Spjd } 2851168404Spjd 2852168404Spjd if (arc_no_grow) 2853168404Spjd return; 2854168404Spjd 2855168404Spjd if (arc_c >= arc_c_max) 2856168404Spjd return; 2857168404Spjd 2858168404Spjd /* 2859168404Spjd * If we're within (2 * maxblocksize) bytes of the target 2860168404Spjd * cache size, increment the target cache size 2861168404Spjd */ 2862168404Spjd if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2863272483Ssmh DTRACE_PROBE1(arc__inc_adapt, int, bytes); 2864168404Spjd atomic_add_64(&arc_c, (int64_t)bytes); 2865168404Spjd if (arc_c > arc_c_max) 2866168404Spjd arc_c = arc_c_max; 2867168404Spjd else if (state == arc_anon) 2868168404Spjd atomic_add_64(&arc_p, (int64_t)bytes); 2869168404Spjd if (arc_p > arc_c) 2870168404Spjd arc_p = arc_c; 2871168404Spjd } 2872168404Spjd ASSERT((int64_t)arc_p >= 0); 2873168404Spjd} 2874168404Spjd 2875168404Spjd/* 2876168404Spjd * Check if the cache has reached its limits and eviction is required 2877168404Spjd * prior to insert. 2878168404Spjd */ 2879168404Spjdstatic int 2880185029Spjdarc_evict_needed(arc_buf_contents_t type) 2881168404Spjd{ 2882185029Spjd if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2883185029Spjd return (1); 2884185029Spjd 2885168404Spjd if (arc_reclaim_needed()) 2886168404Spjd return (1); 2887168404Spjd 2888168404Spjd return (arc_size > arc_c); 2889168404Spjd} 2890168404Spjd 2891168404Spjd/* 2892168404Spjd * The buffer, supplied as the first argument, needs a data block. 2893168404Spjd * So, if we are at cache max, determine which cache should be victimized. 2894168404Spjd * We have the following cases: 2895168404Spjd * 2896168404Spjd * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2897168404Spjd * In this situation if we're out of space, but the resident size of the MFU is 2898168404Spjd * under the limit, victimize the MFU cache to satisfy this insertion request. 2899168404Spjd * 2900168404Spjd * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2901168404Spjd * Here, we've used up all of the available space for the MRU, so we need to 2902168404Spjd * evict from our own cache instead. Evict from the set of resident MRU 2903168404Spjd * entries. 2904168404Spjd * 2905168404Spjd * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2906168404Spjd * c minus p represents the MFU space in the cache, since p is the size of the 2907168404Spjd * cache that is dedicated to the MRU. In this situation there's still space on 2908168404Spjd * the MFU side, so the MRU side needs to be victimized. 2909168404Spjd * 2910168404Spjd * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2911168404Spjd * MFU's resident set is consuming more space than it has been allotted. In 2912168404Spjd * this situation, we must victimize our own cache, the MFU, for this insertion. 2913168404Spjd */ 2914168404Spjdstatic void 2915168404Spjdarc_get_data_buf(arc_buf_t *buf) 2916168404Spjd{ 2917168404Spjd arc_state_t *state = buf->b_hdr->b_state; 2918168404Spjd uint64_t size = buf->b_hdr->b_size; 2919168404Spjd arc_buf_contents_t type = buf->b_hdr->b_type; 2920168404Spjd 2921168404Spjd arc_adapt(size, state); 2922168404Spjd 2923168404Spjd /* 2924168404Spjd * We have not yet reached cache maximum size, 2925168404Spjd * just allocate a new buffer. 2926168404Spjd */ 2927185029Spjd if (!arc_evict_needed(type)) { 2928168404Spjd if (type == ARC_BUFC_METADATA) { 2929168404Spjd buf->b_data = zio_buf_alloc(size); 2930208373Smm arc_space_consume(size, ARC_SPACE_DATA); 2931168404Spjd } else { 2932168404Spjd ASSERT(type == ARC_BUFC_DATA); 2933168404Spjd buf->b_data = zio_data_buf_alloc(size); 2934208373Smm ARCSTAT_INCR(arcstat_data_size, size); 2935185029Spjd atomic_add_64(&arc_size, size); 2936168404Spjd } 2937168404Spjd goto out; 2938168404Spjd } 2939168404Spjd 2940168404Spjd /* 2941168404Spjd * If we are prefetching from the mfu ghost list, this buffer 2942168404Spjd * will end up on the mru list; so steal space from there. 2943168404Spjd */ 2944168404Spjd if (state == arc_mfu_ghost) 2945275811Sdelphij state = buf->b_hdr->b_flags & ARC_FLAG_PREFETCH ? 2946275811Sdelphij arc_mru : arc_mfu; 2947168404Spjd else if (state == arc_mru_ghost) 2948168404Spjd state = arc_mru; 2949168404Spjd 2950168404Spjd if (state == arc_mru || state == arc_anon) { 2951168404Spjd uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2952208373Smm state = (arc_mfu->arcs_lsize[type] >= size && 2953185029Spjd arc_p > mru_used) ? arc_mfu : arc_mru; 2954168404Spjd } else { 2955168404Spjd /* MFU cases */ 2956168404Spjd uint64_t mfu_space = arc_c - arc_p; 2957208373Smm state = (arc_mru->arcs_lsize[type] >= size && 2958185029Spjd mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2959168404Spjd } 2960209962Smm if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2961168404Spjd if (type == ARC_BUFC_METADATA) { 2962168404Spjd buf->b_data = zio_buf_alloc(size); 2963208373Smm arc_space_consume(size, ARC_SPACE_DATA); 2964168404Spjd } else { 2965168404Spjd ASSERT(type == ARC_BUFC_DATA); 2966168404Spjd buf->b_data = zio_data_buf_alloc(size); 2967208373Smm ARCSTAT_INCR(arcstat_data_size, size); 2968185029Spjd atomic_add_64(&arc_size, size); 2969168404Spjd } 2970168404Spjd ARCSTAT_BUMP(arcstat_recycle_miss); 2971168404Spjd } 2972168404Spjd ASSERT(buf->b_data != NULL); 2973168404Spjdout: 2974168404Spjd /* 2975168404Spjd * Update the state size. Note that ghost states have a 2976168404Spjd * "ghost size" and so don't need to be updated. 2977168404Spjd */ 2978168404Spjd if (!GHOST_STATE(buf->b_hdr->b_state)) { 2979168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 2980168404Spjd 2981168404Spjd atomic_add_64(&hdr->b_state->arcs_size, size); 2982168404Spjd if (list_link_active(&hdr->b_arc_node)) { 2983168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2984185029Spjd atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2985168404Spjd } 2986168404Spjd /* 2987168404Spjd * If we are growing the cache, and we are adding anonymous 2988168404Spjd * data, and we have outgrown arc_p, update arc_p 2989168404Spjd */ 2990168404Spjd if (arc_size < arc_c && hdr->b_state == arc_anon && 2991168404Spjd arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2992168404Spjd arc_p = MIN(arc_c, arc_p + size); 2993168404Spjd } 2994205231Skmacy ARCSTAT_BUMP(arcstat_allocated); 2995168404Spjd} 2996168404Spjd 2997168404Spjd/* 2998168404Spjd * This routine is called whenever a buffer is accessed. 2999168404Spjd * NOTE: the hash lock is dropped in this function. 3000168404Spjd */ 3001168404Spjdstatic void 3002275811Sdelphijarc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3003168404Spjd{ 3004219089Spjd clock_t now; 3005219089Spjd 3006168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 3007168404Spjd 3008275811Sdelphij if (hdr->b_state == arc_anon) { 3009168404Spjd /* 3010168404Spjd * This buffer is not in the cache, and does not 3011168404Spjd * appear in our "ghost" list. Add the new buffer 3012168404Spjd * to the MRU state. 3013168404Spjd */ 3014168404Spjd 3015275811Sdelphij ASSERT(hdr->b_arc_access == 0); 3016275811Sdelphij hdr->b_arc_access = ddi_get_lbolt(); 3017275811Sdelphij DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3018275811Sdelphij arc_change_state(arc_mru, hdr, hash_lock); 3019168404Spjd 3020275811Sdelphij } else if (hdr->b_state == arc_mru) { 3021219089Spjd now = ddi_get_lbolt(); 3022219089Spjd 3023168404Spjd /* 3024168404Spjd * If this buffer is here because of a prefetch, then either: 3025168404Spjd * - clear the flag if this is a "referencing" read 3026168404Spjd * (any subsequent access will bump this into the MFU state). 3027168404Spjd * or 3028168404Spjd * - move the buffer to the head of the list if this is 3029168404Spjd * another prefetch (to make it less likely to be evicted). 3030168404Spjd */ 3031275811Sdelphij if ((hdr->b_flags & ARC_FLAG_PREFETCH) != 0) { 3032275811Sdelphij if (refcount_count(&hdr->b_refcnt) == 0) { 3033275811Sdelphij ASSERT(list_link_active(&hdr->b_arc_node)); 3034168404Spjd } else { 3035275811Sdelphij hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3036168404Spjd ARCSTAT_BUMP(arcstat_mru_hits); 3037168404Spjd } 3038275811Sdelphij hdr->b_arc_access = now; 3039168404Spjd return; 3040168404Spjd } 3041168404Spjd 3042168404Spjd /* 3043168404Spjd * This buffer has been "accessed" only once so far, 3044168404Spjd * but it is still in the cache. Move it to the MFU 3045168404Spjd * state. 3046168404Spjd */ 3047275811Sdelphij if (now > hdr->b_arc_access + ARC_MINTIME) { 3048168404Spjd /* 3049168404Spjd * More than 125ms have passed since we 3050168404Spjd * instantiated this buffer. Move it to the 3051168404Spjd * most frequently used state. 3052168404Spjd */ 3053275811Sdelphij hdr->b_arc_access = now; 3054275811Sdelphij DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3055275811Sdelphij arc_change_state(arc_mfu, hdr, hash_lock); 3056168404Spjd } 3057168404Spjd ARCSTAT_BUMP(arcstat_mru_hits); 3058275811Sdelphij } else if (hdr->b_state == arc_mru_ghost) { 3059168404Spjd arc_state_t *new_state; 3060168404Spjd /* 3061168404Spjd * This buffer has been "accessed" recently, but 3062168404Spjd * was evicted from the cache. Move it to the 3063168404Spjd * MFU state. 3064168404Spjd */ 3065168404Spjd 3066275811Sdelphij if (hdr->b_flags & ARC_FLAG_PREFETCH) { 3067168404Spjd new_state = arc_mru; 3068275811Sdelphij if (refcount_count(&hdr->b_refcnt) > 0) 3069275811Sdelphij hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3070275811Sdelphij DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3071168404Spjd } else { 3072168404Spjd new_state = arc_mfu; 3073275811Sdelphij DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3074168404Spjd } 3075168404Spjd 3076275811Sdelphij hdr->b_arc_access = ddi_get_lbolt(); 3077275811Sdelphij arc_change_state(new_state, hdr, hash_lock); 3078168404Spjd 3079168404Spjd ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3080275811Sdelphij } else if (hdr->b_state == arc_mfu) { 3081168404Spjd /* 3082168404Spjd * This buffer has been accessed more than once and is 3083168404Spjd * still in the cache. Keep it in the MFU state. 3084168404Spjd * 3085168404Spjd * NOTE: an add_reference() that occurred when we did 3086168404Spjd * the arc_read() will have kicked this off the list. 3087168404Spjd * If it was a prefetch, we will explicitly move it to 3088168404Spjd * the head of the list now. 3089168404Spjd */ 3090275811Sdelphij if ((hdr->b_flags & ARC_FLAG_PREFETCH) != 0) { 3091275811Sdelphij ASSERT(refcount_count(&hdr->b_refcnt) == 0); 3092275811Sdelphij ASSERT(list_link_active(&hdr->b_arc_node)); 3093168404Spjd } 3094168404Spjd ARCSTAT_BUMP(arcstat_mfu_hits); 3095275811Sdelphij hdr->b_arc_access = ddi_get_lbolt(); 3096275811Sdelphij } else if (hdr->b_state == arc_mfu_ghost) { 3097168404Spjd arc_state_t *new_state = arc_mfu; 3098168404Spjd /* 3099168404Spjd * This buffer has been accessed more than once but has 3100168404Spjd * been evicted from the cache. Move it back to the 3101168404Spjd * MFU state. 3102168404Spjd */ 3103168404Spjd 3104275811Sdelphij if (hdr->b_flags & ARC_FLAG_PREFETCH) { 3105168404Spjd /* 3106168404Spjd * This is a prefetch access... 3107168404Spjd * move this block back to the MRU state. 3108168404Spjd */ 3109275811Sdelphij ASSERT0(refcount_count(&hdr->b_refcnt)); 3110168404Spjd new_state = arc_mru; 3111168404Spjd } 3112168404Spjd 3113275811Sdelphij hdr->b_arc_access = ddi_get_lbolt(); 3114275811Sdelphij DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3115275811Sdelphij arc_change_state(new_state, hdr, hash_lock); 3116168404Spjd 3117168404Spjd ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 3118275811Sdelphij } else if (hdr->b_state == arc_l2c_only) { 3119185029Spjd /* 3120185029Spjd * This buffer is on the 2nd Level ARC. 3121185029Spjd */ 3122185029Spjd 3123275811Sdelphij hdr->b_arc_access = ddi_get_lbolt(); 3124275811Sdelphij DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3125275811Sdelphij arc_change_state(arc_mfu, hdr, hash_lock); 3126168404Spjd } else { 3127168404Spjd ASSERT(!"invalid arc state"); 3128168404Spjd } 3129168404Spjd} 3130168404Spjd 3131168404Spjd/* a generic arc_done_func_t which you can use */ 3132168404Spjd/* ARGSUSED */ 3133168404Spjdvoid 3134168404Spjdarc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 3135168404Spjd{ 3136219089Spjd if (zio == NULL || zio->io_error == 0) 3137219089Spjd bcopy(buf->b_data, arg, buf->b_hdr->b_size); 3138248571Smm VERIFY(arc_buf_remove_ref(buf, arg)); 3139168404Spjd} 3140168404Spjd 3141185029Spjd/* a generic arc_done_func_t */ 3142168404Spjdvoid 3143168404Spjdarc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 3144168404Spjd{ 3145168404Spjd arc_buf_t **bufp = arg; 3146168404Spjd if (zio && zio->io_error) { 3147248571Smm VERIFY(arc_buf_remove_ref(buf, arg)); 3148168404Spjd *bufp = NULL; 3149168404Spjd } else { 3150168404Spjd *bufp = buf; 3151219089Spjd ASSERT(buf->b_data); 3152168404Spjd } 3153168404Spjd} 3154168404Spjd 3155168404Spjdstatic void 3156168404Spjdarc_read_done(zio_t *zio) 3157168404Spjd{ 3158268075Sdelphij arc_buf_hdr_t *hdr; 3159168404Spjd arc_buf_t *buf; 3160168404Spjd arc_buf_t *abuf; /* buffer we're assigning to callback */ 3161268075Sdelphij kmutex_t *hash_lock = NULL; 3162168404Spjd arc_callback_t *callback_list, *acb; 3163168404Spjd int freeable = FALSE; 3164168404Spjd 3165168404Spjd buf = zio->io_private; 3166168404Spjd hdr = buf->b_hdr; 3167168404Spjd 3168168404Spjd /* 3169168404Spjd * The hdr was inserted into hash-table and removed from lists 3170168404Spjd * prior to starting I/O. We should find this header, since 3171168404Spjd * it's in the hash table, and it should be legit since it's 3172168404Spjd * not possible to evict it during the I/O. The only possible 3173168404Spjd * reason for it not to be found is if we were freed during the 3174168404Spjd * read. 3175168404Spjd */ 3176268075Sdelphij if (HDR_IN_HASH_TABLE(hdr)) { 3177268075Sdelphij ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3178268075Sdelphij ASSERT3U(hdr->b_dva.dva_word[0], ==, 3179268075Sdelphij BP_IDENTITY(zio->io_bp)->dva_word[0]); 3180268075Sdelphij ASSERT3U(hdr->b_dva.dva_word[1], ==, 3181268075Sdelphij BP_IDENTITY(zio->io_bp)->dva_word[1]); 3182168404Spjd 3183268075Sdelphij arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3184268075Sdelphij &hash_lock); 3185168404Spjd 3186268075Sdelphij ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3187268075Sdelphij hash_lock == NULL) || 3188268075Sdelphij (found == hdr && 3189268075Sdelphij DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3190268075Sdelphij (found == hdr && HDR_L2_READING(hdr))); 3191268075Sdelphij } 3192268075Sdelphij 3193275811Sdelphij hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; 3194275811Sdelphij if (l2arc_noprefetch && (hdr->b_flags & ARC_FLAG_PREFETCH)) 3195275811Sdelphij hdr->b_flags &= ~ARC_FLAG_L2CACHE; 3196206796Spjd 3197168404Spjd /* byteswap if necessary */ 3198168404Spjd callback_list = hdr->b_acb; 3199168404Spjd ASSERT(callback_list != NULL); 3200209101Smm if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3201236884Smm dmu_object_byteswap_t bswap = 3202236884Smm DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3203185029Spjd arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3204185029Spjd byteswap_uint64_array : 3205236884Smm dmu_ot_byteswap[bswap].ob_func; 3206185029Spjd func(buf->b_data, hdr->b_size); 3207185029Spjd } 3208168404Spjd 3209185029Spjd arc_cksum_compute(buf, B_FALSE); 3210240133Smm#ifdef illumos 3211240133Smm arc_buf_watch(buf); 3212277300Ssmh#endif 3213168404Spjd 3214219089Spjd if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3215219089Spjd /* 3216219089Spjd * Only call arc_access on anonymous buffers. This is because 3217219089Spjd * if we've issued an I/O for an evicted buffer, we've already 3218219089Spjd * called arc_access (to prevent any simultaneous readers from 3219219089Spjd * getting confused). 3220219089Spjd */ 3221219089Spjd arc_access(hdr, hash_lock); 3222219089Spjd } 3223219089Spjd 3224168404Spjd /* create copies of the data buffer for the callers */ 3225168404Spjd abuf = buf; 3226168404Spjd for (acb = callback_list; acb; acb = acb->acb_next) { 3227168404Spjd if (acb->acb_done) { 3228242845Sdelphij if (abuf == NULL) { 3229242845Sdelphij ARCSTAT_BUMP(arcstat_duplicate_reads); 3230168404Spjd abuf = arc_buf_clone(buf); 3231242845Sdelphij } 3232168404Spjd acb->acb_buf = abuf; 3233168404Spjd abuf = NULL; 3234168404Spjd } 3235168404Spjd } 3236168404Spjd hdr->b_acb = NULL; 3237275811Sdelphij hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 3238168404Spjd ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3239219089Spjd if (abuf == buf) { 3240219089Spjd ASSERT(buf->b_efunc == NULL); 3241219089Spjd ASSERT(hdr->b_datacnt == 1); 3242275811Sdelphij hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 3243219089Spjd } 3244168404Spjd 3245168404Spjd ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3246168404Spjd 3247168404Spjd if (zio->io_error != 0) { 3248275811Sdelphij hdr->b_flags |= ARC_FLAG_IO_ERROR; 3249168404Spjd if (hdr->b_state != arc_anon) 3250168404Spjd arc_change_state(arc_anon, hdr, hash_lock); 3251168404Spjd if (HDR_IN_HASH_TABLE(hdr)) 3252168404Spjd buf_hash_remove(hdr); 3253168404Spjd freeable = refcount_is_zero(&hdr->b_refcnt); 3254168404Spjd } 3255168404Spjd 3256168404Spjd /* 3257168404Spjd * Broadcast before we drop the hash_lock to avoid the possibility 3258168404Spjd * that the hdr (and hence the cv) might be freed before we get to 3259168404Spjd * the cv_broadcast(). 3260168404Spjd */ 3261168404Spjd cv_broadcast(&hdr->b_cv); 3262168404Spjd 3263168404Spjd if (hash_lock) { 3264168404Spjd mutex_exit(hash_lock); 3265168404Spjd } else { 3266168404Spjd /* 3267168404Spjd * This block was freed while we waited for the read to 3268168404Spjd * complete. It has been removed from the hash table and 3269168404Spjd * moved to the anonymous state (so that it won't show up 3270168404Spjd * in the cache). 3271168404Spjd */ 3272168404Spjd ASSERT3P(hdr->b_state, ==, arc_anon); 3273168404Spjd freeable = refcount_is_zero(&hdr->b_refcnt); 3274168404Spjd } 3275168404Spjd 3276168404Spjd /* execute each callback and free its structure */ 3277168404Spjd while ((acb = callback_list) != NULL) { 3278168404Spjd if (acb->acb_done) 3279168404Spjd acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3280168404Spjd 3281168404Spjd if (acb->acb_zio_dummy != NULL) { 3282168404Spjd acb->acb_zio_dummy->io_error = zio->io_error; 3283168404Spjd zio_nowait(acb->acb_zio_dummy); 3284168404Spjd } 3285168404Spjd 3286168404Spjd callback_list = acb->acb_next; 3287168404Spjd kmem_free(acb, sizeof (arc_callback_t)); 3288168404Spjd } 3289168404Spjd 3290168404Spjd if (freeable) 3291168404Spjd arc_hdr_destroy(hdr); 3292168404Spjd} 3293168404Spjd 3294168404Spjd/* 3295168404Spjd * "Read" the block block at the specified DVA (in bp) via the 3296168404Spjd * cache. If the block is found in the cache, invoke the provided 3297168404Spjd * callback immediately and return. Note that the `zio' parameter 3298168404Spjd * in the callback will be NULL in this case, since no IO was 3299168404Spjd * required. If the block is not in the cache pass the read request 3300168404Spjd * on to the spa with a substitute callback function, so that the 3301168404Spjd * requested block will be added to the cache. 3302168404Spjd * 3303168404Spjd * If a read request arrives for a block that has a read in-progress, 3304168404Spjd * either wait for the in-progress read to complete (and return the 3305168404Spjd * results); or, if this is a read with a "done" func, add a record 3306168404Spjd * to the read to invoke the "done" func when the read completes, 3307168404Spjd * and return; or just return. 3308168404Spjd * 3309168404Spjd * arc_read_done() will invoke all the requested "done" functions 3310168404Spjd * for readers of this block. 3311168404Spjd */ 3312168404Spjdint 3313246666Smmarc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3314275811Sdelphij void *private, zio_priority_t priority, int zio_flags, 3315275811Sdelphij arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 3316168404Spjd{ 3317268075Sdelphij arc_buf_hdr_t *hdr = NULL; 3318247187Smm arc_buf_t *buf = NULL; 3319268075Sdelphij kmutex_t *hash_lock = NULL; 3320185029Spjd zio_t *rzio; 3321228103Smm uint64_t guid = spa_load_guid(spa); 3322168404Spjd 3323268075Sdelphij ASSERT(!BP_IS_EMBEDDED(bp) || 3324268075Sdelphij BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3325268075Sdelphij 3326168404Spjdtop: 3327268075Sdelphij if (!BP_IS_EMBEDDED(bp)) { 3328268075Sdelphij /* 3329268075Sdelphij * Embedded BP's have no DVA and require no I/O to "read". 3330268075Sdelphij * Create an anonymous arc buf to back it. 3331268075Sdelphij */ 3332268075Sdelphij hdr = buf_hash_find(guid, bp, &hash_lock); 3333268075Sdelphij } 3334168404Spjd 3335268075Sdelphij if (hdr != NULL && hdr->b_datacnt > 0) { 3336268075Sdelphij 3337275811Sdelphij *arc_flags |= ARC_FLAG_CACHED; 3338168404Spjd 3339168404Spjd if (HDR_IO_IN_PROGRESS(hdr)) { 3340168404Spjd 3341275811Sdelphij if (*arc_flags & ARC_FLAG_WAIT) { 3342168404Spjd cv_wait(&hdr->b_cv, hash_lock); 3343168404Spjd mutex_exit(hash_lock); 3344168404Spjd goto top; 3345168404Spjd } 3346275811Sdelphij ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 3347168404Spjd 3348168404Spjd if (done) { 3349168404Spjd arc_callback_t *acb = NULL; 3350168404Spjd 3351168404Spjd acb = kmem_zalloc(sizeof (arc_callback_t), 3352168404Spjd KM_SLEEP); 3353168404Spjd acb->acb_done = done; 3354168404Spjd acb->acb_private = private; 3355168404Spjd if (pio != NULL) 3356168404Spjd acb->acb_zio_dummy = zio_null(pio, 3357209962Smm spa, NULL, NULL, NULL, zio_flags); 3358168404Spjd 3359168404Spjd ASSERT(acb->acb_done != NULL); 3360168404Spjd acb->acb_next = hdr->b_acb; 3361168404Spjd hdr->b_acb = acb; 3362168404Spjd add_reference(hdr, hash_lock, private); 3363168404Spjd mutex_exit(hash_lock); 3364168404Spjd return (0); 3365168404Spjd } 3366168404Spjd mutex_exit(hash_lock); 3367168404Spjd return (0); 3368168404Spjd } 3369168404Spjd 3370168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3371168404Spjd 3372168404Spjd if (done) { 3373168404Spjd add_reference(hdr, hash_lock, private); 3374168404Spjd /* 3375168404Spjd * If this block is already in use, create a new 3376168404Spjd * copy of the data so that we will be guaranteed 3377168404Spjd * that arc_release() will always succeed. 3378168404Spjd */ 3379168404Spjd buf = hdr->b_buf; 3380168404Spjd ASSERT(buf); 3381168404Spjd ASSERT(buf->b_data); 3382168404Spjd if (HDR_BUF_AVAILABLE(hdr)) { 3383168404Spjd ASSERT(buf->b_efunc == NULL); 3384275811Sdelphij hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 3385168404Spjd } else { 3386168404Spjd buf = arc_buf_clone(buf); 3387168404Spjd } 3388219089Spjd 3389275811Sdelphij } else if (*arc_flags & ARC_FLAG_PREFETCH && 3390168404Spjd refcount_count(&hdr->b_refcnt) == 0) { 3391275811Sdelphij hdr->b_flags |= ARC_FLAG_PREFETCH; 3392168404Spjd } 3393168404Spjd DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3394168404Spjd arc_access(hdr, hash_lock); 3395275811Sdelphij if (*arc_flags & ARC_FLAG_L2CACHE) 3396275811Sdelphij hdr->b_flags |= ARC_FLAG_L2CACHE; 3397275811Sdelphij if (*arc_flags & ARC_FLAG_L2COMPRESS) 3398275811Sdelphij hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3399168404Spjd mutex_exit(hash_lock); 3400168404Spjd ARCSTAT_BUMP(arcstat_hits); 3401275811Sdelphij ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 3402168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3403168404Spjd data, metadata, hits); 3404168404Spjd 3405168404Spjd if (done) 3406168404Spjd done(NULL, buf, private); 3407168404Spjd } else { 3408168404Spjd uint64_t size = BP_GET_LSIZE(bp); 3409268075Sdelphij arc_callback_t *acb; 3410185029Spjd vdev_t *vd = NULL; 3411247187Smm uint64_t addr = 0; 3412208373Smm boolean_t devw = B_FALSE; 3413258389Savg enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3414258389Savg uint64_t b_asize = 0; 3415168404Spjd 3416168404Spjd if (hdr == NULL) { 3417168404Spjd /* this block is not in the cache */ 3418268075Sdelphij arc_buf_hdr_t *exists = NULL; 3419168404Spjd arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3420168404Spjd buf = arc_buf_alloc(spa, size, private, type); 3421168404Spjd hdr = buf->b_hdr; 3422268075Sdelphij if (!BP_IS_EMBEDDED(bp)) { 3423268075Sdelphij hdr->b_dva = *BP_IDENTITY(bp); 3424268075Sdelphij hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3425268075Sdelphij hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3426268075Sdelphij exists = buf_hash_insert(hdr, &hash_lock); 3427268075Sdelphij } 3428268075Sdelphij if (exists != NULL) { 3429168404Spjd /* somebody beat us to the hash insert */ 3430168404Spjd mutex_exit(hash_lock); 3431219089Spjd buf_discard_identity(hdr); 3432168404Spjd (void) arc_buf_remove_ref(buf, private); 3433168404Spjd goto top; /* restart the IO request */ 3434168404Spjd } 3435275811Sdelphij 3436168404Spjd /* if this is a prefetch, we don't have a reference */ 3437275811Sdelphij if (*arc_flags & ARC_FLAG_PREFETCH) { 3438168404Spjd (void) remove_reference(hdr, hash_lock, 3439168404Spjd private); 3440275811Sdelphij hdr->b_flags |= ARC_FLAG_PREFETCH; 3441168404Spjd } 3442275811Sdelphij if (*arc_flags & ARC_FLAG_L2CACHE) 3443275811Sdelphij hdr->b_flags |= ARC_FLAG_L2CACHE; 3444275811Sdelphij if (*arc_flags & ARC_FLAG_L2COMPRESS) 3445275811Sdelphij hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3446168404Spjd if (BP_GET_LEVEL(bp) > 0) 3447275811Sdelphij hdr->b_flags |= ARC_FLAG_INDIRECT; 3448168404Spjd } else { 3449168404Spjd /* this block is in the ghost cache */ 3450168404Spjd ASSERT(GHOST_STATE(hdr->b_state)); 3451168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3452240415Smm ASSERT0(refcount_count(&hdr->b_refcnt)); 3453168404Spjd ASSERT(hdr->b_buf == NULL); 3454168404Spjd 3455168404Spjd /* if this is a prefetch, we don't have a reference */ 3456275811Sdelphij if (*arc_flags & ARC_FLAG_PREFETCH) 3457275811Sdelphij hdr->b_flags |= ARC_FLAG_PREFETCH; 3458168404Spjd else 3459168404Spjd add_reference(hdr, hash_lock, private); 3460275811Sdelphij if (*arc_flags & ARC_FLAG_L2CACHE) 3461275811Sdelphij hdr->b_flags |= ARC_FLAG_L2CACHE; 3462275811Sdelphij if (*arc_flags & ARC_FLAG_L2COMPRESS) 3463275811Sdelphij hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3464185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3465168404Spjd buf->b_hdr = hdr; 3466168404Spjd buf->b_data = NULL; 3467168404Spjd buf->b_efunc = NULL; 3468168404Spjd buf->b_private = NULL; 3469168404Spjd buf->b_next = NULL; 3470168404Spjd hdr->b_buf = buf; 3471168404Spjd ASSERT(hdr->b_datacnt == 0); 3472168404Spjd hdr->b_datacnt = 1; 3473219089Spjd arc_get_data_buf(buf); 3474219089Spjd arc_access(hdr, hash_lock); 3475168404Spjd } 3476168404Spjd 3477219089Spjd ASSERT(!GHOST_STATE(hdr->b_state)); 3478219089Spjd 3479168404Spjd acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3480168404Spjd acb->acb_done = done; 3481168404Spjd acb->acb_private = private; 3482168404Spjd 3483168404Spjd ASSERT(hdr->b_acb == NULL); 3484168404Spjd hdr->b_acb = acb; 3485275811Sdelphij hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 3486168404Spjd 3487258389Savg if (hdr->b_l2hdr != NULL && 3488185029Spjd (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3489208373Smm devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3490185029Spjd addr = hdr->b_l2hdr->b_daddr; 3491258389Savg b_compress = hdr->b_l2hdr->b_compress; 3492258389Savg b_asize = hdr->b_l2hdr->b_asize; 3493185029Spjd /* 3494185029Spjd * Lock out device removal. 3495185029Spjd */ 3496185029Spjd if (vdev_is_dead(vd) || 3497185029Spjd !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3498185029Spjd vd = NULL; 3499185029Spjd } 3500185029Spjd 3501268075Sdelphij if (hash_lock != NULL) 3502268075Sdelphij mutex_exit(hash_lock); 3503168404Spjd 3504251629Sdelphij /* 3505251629Sdelphij * At this point, we have a level 1 cache miss. Try again in 3506251629Sdelphij * L2ARC if possible. 3507251629Sdelphij */ 3508168404Spjd ASSERT3U(hdr->b_size, ==, size); 3509219089Spjd DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3510268123Sdelphij uint64_t, size, zbookmark_phys_t *, zb); 3511168404Spjd ARCSTAT_BUMP(arcstat_misses); 3512275811Sdelphij ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 3513168404Spjd demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3514168404Spjd data, metadata, misses); 3515228392Spjd#ifdef _KERNEL 3516228392Spjd curthread->td_ru.ru_inblock++; 3517228392Spjd#endif 3518168404Spjd 3519208373Smm if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3520185029Spjd /* 3521185029Spjd * Read from the L2ARC if the following are true: 3522185029Spjd * 1. The L2ARC vdev was previously cached. 3523185029Spjd * 2. This buffer still has L2ARC metadata. 3524185029Spjd * 3. This buffer isn't currently writing to the L2ARC. 3525185029Spjd * 4. The L2ARC entry wasn't evicted, which may 3526185029Spjd * also have invalidated the vdev. 3527208373Smm * 5. This isn't prefetch and l2arc_noprefetch is set. 3528185029Spjd */ 3529185029Spjd if (hdr->b_l2hdr != NULL && 3530208373Smm !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3531208373Smm !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3532185029Spjd l2arc_read_callback_t *cb; 3533185029Spjd 3534185029Spjd DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3535185029Spjd ARCSTAT_BUMP(arcstat_l2_hits); 3536185029Spjd 3537185029Spjd cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3538185029Spjd KM_SLEEP); 3539185029Spjd cb->l2rcb_buf = buf; 3540185029Spjd cb->l2rcb_spa = spa; 3541185029Spjd cb->l2rcb_bp = *bp; 3542185029Spjd cb->l2rcb_zb = *zb; 3543185029Spjd cb->l2rcb_flags = zio_flags; 3544258389Savg cb->l2rcb_compress = b_compress; 3545185029Spjd 3546247187Smm ASSERT(addr >= VDEV_LABEL_START_SIZE && 3547247187Smm addr + size < vd->vdev_psize - 3548247187Smm VDEV_LABEL_END_SIZE); 3549247187Smm 3550185029Spjd /* 3551185029Spjd * l2arc read. The SCL_L2ARC lock will be 3552185029Spjd * released by l2arc_read_done(). 3553251478Sdelphij * Issue a null zio if the underlying buffer 3554251478Sdelphij * was squashed to zero size by compression. 3555185029Spjd */ 3556258389Savg if (b_compress == ZIO_COMPRESS_EMPTY) { 3557251478Sdelphij rzio = zio_null(pio, spa, vd, 3558251478Sdelphij l2arc_read_done, cb, 3559251478Sdelphij zio_flags | ZIO_FLAG_DONT_CACHE | 3560251478Sdelphij ZIO_FLAG_CANFAIL | 3561251478Sdelphij ZIO_FLAG_DONT_PROPAGATE | 3562251478Sdelphij ZIO_FLAG_DONT_RETRY); 3563251478Sdelphij } else { 3564251478Sdelphij rzio = zio_read_phys(pio, vd, addr, 3565258389Savg b_asize, buf->b_data, 3566258389Savg ZIO_CHECKSUM_OFF, 3567251478Sdelphij l2arc_read_done, cb, priority, 3568251478Sdelphij zio_flags | ZIO_FLAG_DONT_CACHE | 3569251478Sdelphij ZIO_FLAG_CANFAIL | 3570251478Sdelphij ZIO_FLAG_DONT_PROPAGATE | 3571251478Sdelphij ZIO_FLAG_DONT_RETRY, B_FALSE); 3572251478Sdelphij } 3573185029Spjd DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3574185029Spjd zio_t *, rzio); 3575258389Savg ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3576185029Spjd 3577275811Sdelphij if (*arc_flags & ARC_FLAG_NOWAIT) { 3578185029Spjd zio_nowait(rzio); 3579185029Spjd return (0); 3580185029Spjd } 3581185029Spjd 3582275811Sdelphij ASSERT(*arc_flags & ARC_FLAG_WAIT); 3583185029Spjd if (zio_wait(rzio) == 0) 3584185029Spjd return (0); 3585185029Spjd 3586185029Spjd /* l2arc read error; goto zio_read() */ 3587185029Spjd } else { 3588185029Spjd DTRACE_PROBE1(l2arc__miss, 3589185029Spjd arc_buf_hdr_t *, hdr); 3590185029Spjd ARCSTAT_BUMP(arcstat_l2_misses); 3591185029Spjd if (HDR_L2_WRITING(hdr)) 3592185029Spjd ARCSTAT_BUMP(arcstat_l2_rw_clash); 3593185029Spjd spa_config_exit(spa, SCL_L2ARC, vd); 3594185029Spjd } 3595208373Smm } else { 3596208373Smm if (vd != NULL) 3597208373Smm spa_config_exit(spa, SCL_L2ARC, vd); 3598208373Smm if (l2arc_ndev != 0) { 3599208373Smm DTRACE_PROBE1(l2arc__miss, 3600208373Smm arc_buf_hdr_t *, hdr); 3601208373Smm ARCSTAT_BUMP(arcstat_l2_misses); 3602208373Smm } 3603185029Spjd } 3604185029Spjd 3605168404Spjd rzio = zio_read(pio, spa, bp, buf->b_data, size, 3606185029Spjd arc_read_done, buf, priority, zio_flags, zb); 3607168404Spjd 3608275811Sdelphij if (*arc_flags & ARC_FLAG_WAIT) 3609168404Spjd return (zio_wait(rzio)); 3610168404Spjd 3611275811Sdelphij ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 3612168404Spjd zio_nowait(rzio); 3613168404Spjd } 3614168404Spjd return (0); 3615168404Spjd} 3616168404Spjd 3617168404Spjdvoid 3618168404Spjdarc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3619168404Spjd{ 3620168404Spjd ASSERT(buf->b_hdr != NULL); 3621168404Spjd ASSERT(buf->b_hdr->b_state != arc_anon); 3622168404Spjd ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3623219089Spjd ASSERT(buf->b_efunc == NULL); 3624219089Spjd ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3625219089Spjd 3626168404Spjd buf->b_efunc = func; 3627168404Spjd buf->b_private = private; 3628168404Spjd} 3629168404Spjd 3630168404Spjd/* 3631251520Sdelphij * Notify the arc that a block was freed, and thus will never be used again. 3632251520Sdelphij */ 3633251520Sdelphijvoid 3634251520Sdelphijarc_freed(spa_t *spa, const blkptr_t *bp) 3635251520Sdelphij{ 3636251520Sdelphij arc_buf_hdr_t *hdr; 3637251520Sdelphij kmutex_t *hash_lock; 3638251520Sdelphij uint64_t guid = spa_load_guid(spa); 3639251520Sdelphij 3640268075Sdelphij ASSERT(!BP_IS_EMBEDDED(bp)); 3641268075Sdelphij 3642268075Sdelphij hdr = buf_hash_find(guid, bp, &hash_lock); 3643251520Sdelphij if (hdr == NULL) 3644251520Sdelphij return; 3645251520Sdelphij if (HDR_BUF_AVAILABLE(hdr)) { 3646251520Sdelphij arc_buf_t *buf = hdr->b_buf; 3647251520Sdelphij add_reference(hdr, hash_lock, FTAG); 3648275811Sdelphij hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 3649251520Sdelphij mutex_exit(hash_lock); 3650251520Sdelphij 3651251520Sdelphij arc_release(buf, FTAG); 3652251520Sdelphij (void) arc_buf_remove_ref(buf, FTAG); 3653251520Sdelphij } else { 3654251520Sdelphij mutex_exit(hash_lock); 3655251520Sdelphij } 3656251520Sdelphij 3657251520Sdelphij} 3658251520Sdelphij 3659251520Sdelphij/* 3660268858Sdelphij * Clear the user eviction callback set by arc_set_callback(), first calling 3661268858Sdelphij * it if it exists. Because the presence of a callback keeps an arc_buf cached 3662268858Sdelphij * clearing the callback may result in the arc_buf being destroyed. However, 3663268858Sdelphij * it will not result in the *last* arc_buf being destroyed, hence the data 3664268858Sdelphij * will remain cached in the ARC. We make a copy of the arc buffer here so 3665268858Sdelphij * that we can process the callback without holding any locks. 3666268858Sdelphij * 3667268858Sdelphij * It's possible that the callback is already in the process of being cleared 3668268858Sdelphij * by another thread. In this case we can not clear the callback. 3669268858Sdelphij * 3670268858Sdelphij * Returns B_TRUE if the callback was successfully called and cleared. 3671168404Spjd */ 3672268858Sdelphijboolean_t 3673268858Sdelphijarc_clear_callback(arc_buf_t *buf) 3674168404Spjd{ 3675168404Spjd arc_buf_hdr_t *hdr; 3676168404Spjd kmutex_t *hash_lock; 3677268858Sdelphij arc_evict_func_t *efunc = buf->b_efunc; 3678268858Sdelphij void *private = buf->b_private; 3679205231Skmacy list_t *list, *evicted_list; 3680205231Skmacy kmutex_t *lock, *evicted_lock; 3681206796Spjd 3682219089Spjd mutex_enter(&buf->b_evict_lock); 3683168404Spjd hdr = buf->b_hdr; 3684168404Spjd if (hdr == NULL) { 3685168404Spjd /* 3686168404Spjd * We are in arc_do_user_evicts(). 3687168404Spjd */ 3688168404Spjd ASSERT(buf->b_data == NULL); 3689219089Spjd mutex_exit(&buf->b_evict_lock); 3690268858Sdelphij return (B_FALSE); 3691185029Spjd } else if (buf->b_data == NULL) { 3692185029Spjd /* 3693185029Spjd * We are on the eviction list; process this buffer now 3694185029Spjd * but let arc_do_user_evicts() do the reaping. 3695185029Spjd */ 3696185029Spjd buf->b_efunc = NULL; 3697219089Spjd mutex_exit(&buf->b_evict_lock); 3698268858Sdelphij VERIFY0(efunc(private)); 3699268858Sdelphij return (B_TRUE); 3700168404Spjd } 3701168404Spjd hash_lock = HDR_LOCK(hdr); 3702168404Spjd mutex_enter(hash_lock); 3703219089Spjd hdr = buf->b_hdr; 3704219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3705168404Spjd 3706168404Spjd ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3707168404Spjd ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3708168404Spjd 3709268858Sdelphij buf->b_efunc = NULL; 3710268858Sdelphij buf->b_private = NULL; 3711168404Spjd 3712268858Sdelphij if (hdr->b_datacnt > 1) { 3713268858Sdelphij mutex_exit(&buf->b_evict_lock); 3714268858Sdelphij arc_buf_destroy(buf, FALSE, TRUE); 3715268858Sdelphij } else { 3716268858Sdelphij ASSERT(buf == hdr->b_buf); 3717275811Sdelphij hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 3718268858Sdelphij mutex_exit(&buf->b_evict_lock); 3719268858Sdelphij } 3720168404Spjd 3721168404Spjd mutex_exit(hash_lock); 3722268858Sdelphij VERIFY0(efunc(private)); 3723268858Sdelphij return (B_TRUE); 3724168404Spjd} 3725168404Spjd 3726168404Spjd/* 3727251629Sdelphij * Release this buffer from the cache, making it an anonymous buffer. This 3728251629Sdelphij * must be done after a read and prior to modifying the buffer contents. 3729168404Spjd * If the buffer has more than one reference, we must make 3730185029Spjd * a new hdr for the buffer. 3731168404Spjd */ 3732168404Spjdvoid 3733168404Spjdarc_release(arc_buf_t *buf, void *tag) 3734168404Spjd{ 3735185029Spjd arc_buf_hdr_t *hdr; 3736219089Spjd kmutex_t *hash_lock = NULL; 3737185029Spjd l2arc_buf_hdr_t *l2hdr; 3738185029Spjd uint64_t buf_size; 3739168404Spjd 3740219089Spjd /* 3741219089Spjd * It would be nice to assert that if it's DMU metadata (level > 3742219089Spjd * 0 || it's the dnode file), then it must be syncing context. 3743219089Spjd * But we don't know that information at this level. 3744219089Spjd */ 3745219089Spjd 3746219089Spjd mutex_enter(&buf->b_evict_lock); 3747185029Spjd hdr = buf->b_hdr; 3748185029Spjd 3749168404Spjd /* this buffer is not on any list */ 3750168404Spjd ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3751168404Spjd 3752168404Spjd if (hdr->b_state == arc_anon) { 3753168404Spjd /* this buffer is already released */ 3754168404Spjd ASSERT(buf->b_efunc == NULL); 3755208373Smm } else { 3756208373Smm hash_lock = HDR_LOCK(hdr); 3757208373Smm mutex_enter(hash_lock); 3758219089Spjd hdr = buf->b_hdr; 3759219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3760168404Spjd } 3761168404Spjd 3762185029Spjd l2hdr = hdr->b_l2hdr; 3763185029Spjd if (l2hdr) { 3764185029Spjd mutex_enter(&l2arc_buflist_mtx); 3765274172Savg arc_buf_l2_cdata_free(hdr); 3766185029Spjd hdr->b_l2hdr = NULL; 3767258388Savg list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3768185029Spjd } 3769247187Smm buf_size = hdr->b_size; 3770185029Spjd 3771168404Spjd /* 3772168404Spjd * Do we have more than one buf? 3773168404Spjd */ 3774185029Spjd if (hdr->b_datacnt > 1) { 3775168404Spjd arc_buf_hdr_t *nhdr; 3776168404Spjd arc_buf_t **bufp; 3777168404Spjd uint64_t blksz = hdr->b_size; 3778209962Smm uint64_t spa = hdr->b_spa; 3779168404Spjd arc_buf_contents_t type = hdr->b_type; 3780185029Spjd uint32_t flags = hdr->b_flags; 3781168404Spjd 3782185029Spjd ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3783168404Spjd /* 3784219089Spjd * Pull the data off of this hdr and attach it to 3785219089Spjd * a new anonymous hdr. 3786168404Spjd */ 3787168404Spjd (void) remove_reference(hdr, hash_lock, tag); 3788168404Spjd bufp = &hdr->b_buf; 3789168404Spjd while (*bufp != buf) 3790168404Spjd bufp = &(*bufp)->b_next; 3791219089Spjd *bufp = buf->b_next; 3792168404Spjd buf->b_next = NULL; 3793168404Spjd 3794168404Spjd ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3795168404Spjd atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3796168404Spjd if (refcount_is_zero(&hdr->b_refcnt)) { 3797185029Spjd uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3798185029Spjd ASSERT3U(*size, >=, hdr->b_size); 3799185029Spjd atomic_add_64(size, -hdr->b_size); 3800168404Spjd } 3801242845Sdelphij 3802242845Sdelphij /* 3803242845Sdelphij * We're releasing a duplicate user data buffer, update 3804242845Sdelphij * our statistics accordingly. 3805242845Sdelphij */ 3806242845Sdelphij if (hdr->b_type == ARC_BUFC_DATA) { 3807242845Sdelphij ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3808242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3809242845Sdelphij -hdr->b_size); 3810242845Sdelphij } 3811168404Spjd hdr->b_datacnt -= 1; 3812168404Spjd arc_cksum_verify(buf); 3813240133Smm#ifdef illumos 3814240133Smm arc_buf_unwatch(buf); 3815277300Ssmh#endif 3816168404Spjd 3817168404Spjd mutex_exit(hash_lock); 3818168404Spjd 3819185029Spjd nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3820168404Spjd nhdr->b_size = blksz; 3821168404Spjd nhdr->b_spa = spa; 3822168404Spjd nhdr->b_type = type; 3823168404Spjd nhdr->b_buf = buf; 3824168404Spjd nhdr->b_state = arc_anon; 3825168404Spjd nhdr->b_arc_access = 0; 3826275811Sdelphij nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; 3827185029Spjd nhdr->b_l2hdr = NULL; 3828168404Spjd nhdr->b_datacnt = 1; 3829168404Spjd nhdr->b_freeze_cksum = NULL; 3830168404Spjd (void) refcount_add(&nhdr->b_refcnt, tag); 3831168404Spjd buf->b_hdr = nhdr; 3832219089Spjd mutex_exit(&buf->b_evict_lock); 3833168404Spjd atomic_add_64(&arc_anon->arcs_size, blksz); 3834168404Spjd } else { 3835219089Spjd mutex_exit(&buf->b_evict_lock); 3836168404Spjd ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3837168404Spjd ASSERT(!list_link_active(&hdr->b_arc_node)); 3838168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3839219089Spjd if (hdr->b_state != arc_anon) 3840219089Spjd arc_change_state(arc_anon, hdr, hash_lock); 3841168404Spjd hdr->b_arc_access = 0; 3842219089Spjd if (hash_lock) 3843219089Spjd mutex_exit(hash_lock); 3844185029Spjd 3845219089Spjd buf_discard_identity(hdr); 3846168404Spjd arc_buf_thaw(buf); 3847168404Spjd } 3848168404Spjd buf->b_efunc = NULL; 3849168404Spjd buf->b_private = NULL; 3850185029Spjd 3851185029Spjd if (l2hdr) { 3852251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3853268085Sdelphij vdev_space_update(l2hdr->b_dev->l2ad_vdev, 3854268085Sdelphij -l2hdr->b_asize, 0, 0); 3855248572Ssmh trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3856248574Ssmh hdr->b_size, 0); 3857185029Spjd kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3858185029Spjd ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3859185029Spjd mutex_exit(&l2arc_buflist_mtx); 3860185029Spjd } 3861168404Spjd} 3862168404Spjd 3863168404Spjdint 3864168404Spjdarc_released(arc_buf_t *buf) 3865168404Spjd{ 3866185029Spjd int released; 3867185029Spjd 3868219089Spjd mutex_enter(&buf->b_evict_lock); 3869185029Spjd released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3870219089Spjd mutex_exit(&buf->b_evict_lock); 3871185029Spjd return (released); 3872168404Spjd} 3873168404Spjd 3874168404Spjd#ifdef ZFS_DEBUG 3875168404Spjdint 3876168404Spjdarc_referenced(arc_buf_t *buf) 3877168404Spjd{ 3878185029Spjd int referenced; 3879185029Spjd 3880219089Spjd mutex_enter(&buf->b_evict_lock); 3881185029Spjd referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3882219089Spjd mutex_exit(&buf->b_evict_lock); 3883185029Spjd return (referenced); 3884168404Spjd} 3885168404Spjd#endif 3886168404Spjd 3887168404Spjdstatic void 3888168404Spjdarc_write_ready(zio_t *zio) 3889168404Spjd{ 3890168404Spjd arc_write_callback_t *callback = zio->io_private; 3891168404Spjd arc_buf_t *buf = callback->awcb_buf; 3892185029Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3893168404Spjd 3894185029Spjd ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3895185029Spjd callback->awcb_ready(zio, buf, callback->awcb_private); 3896185029Spjd 3897185029Spjd /* 3898185029Spjd * If the IO is already in progress, then this is a re-write 3899185029Spjd * attempt, so we need to thaw and re-compute the cksum. 3900185029Spjd * It is the responsibility of the callback to handle the 3901185029Spjd * accounting for any re-write attempt. 3902185029Spjd */ 3903185029Spjd if (HDR_IO_IN_PROGRESS(hdr)) { 3904185029Spjd mutex_enter(&hdr->b_freeze_lock); 3905185029Spjd if (hdr->b_freeze_cksum != NULL) { 3906185029Spjd kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3907185029Spjd hdr->b_freeze_cksum = NULL; 3908185029Spjd } 3909185029Spjd mutex_exit(&hdr->b_freeze_lock); 3910168404Spjd } 3911185029Spjd arc_cksum_compute(buf, B_FALSE); 3912275811Sdelphij hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 3913168404Spjd} 3914168404Spjd 3915258632Savg/* 3916258632Savg * The SPA calls this callback for each physical write that happens on behalf 3917258632Savg * of a logical write. See the comment in dbuf_write_physdone() for details. 3918258632Savg */ 3919168404Spjdstatic void 3920258632Savgarc_write_physdone(zio_t *zio) 3921258632Savg{ 3922258632Savg arc_write_callback_t *cb = zio->io_private; 3923258632Savg if (cb->awcb_physdone != NULL) 3924258632Savg cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3925258632Savg} 3926258632Savg 3927258632Savgstatic void 3928168404Spjdarc_write_done(zio_t *zio) 3929168404Spjd{ 3930168404Spjd arc_write_callback_t *callback = zio->io_private; 3931168404Spjd arc_buf_t *buf = callback->awcb_buf; 3932168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3933168404Spjd 3934219089Spjd ASSERT(hdr->b_acb == NULL); 3935168404Spjd 3936219089Spjd if (zio->io_error == 0) { 3937268075Sdelphij if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 3938260150Sdelphij buf_discard_identity(hdr); 3939260150Sdelphij } else { 3940260150Sdelphij hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3941260150Sdelphij hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3942260150Sdelphij hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3943260150Sdelphij } 3944219089Spjd } else { 3945219089Spjd ASSERT(BUF_EMPTY(hdr)); 3946219089Spjd } 3947219089Spjd 3948168404Spjd /* 3949268075Sdelphij * If the block to be written was all-zero or compressed enough to be 3950268075Sdelphij * embedded in the BP, no write was performed so there will be no 3951268075Sdelphij * dva/birth/checksum. The buffer must therefore remain anonymous 3952268075Sdelphij * (and uncached). 3953168404Spjd */ 3954168404Spjd if (!BUF_EMPTY(hdr)) { 3955168404Spjd arc_buf_hdr_t *exists; 3956168404Spjd kmutex_t *hash_lock; 3957168404Spjd 3958219089Spjd ASSERT(zio->io_error == 0); 3959219089Spjd 3960168404Spjd arc_cksum_verify(buf); 3961168404Spjd 3962168404Spjd exists = buf_hash_insert(hdr, &hash_lock); 3963168404Spjd if (exists) { 3964168404Spjd /* 3965168404Spjd * This can only happen if we overwrite for 3966168404Spjd * sync-to-convergence, because we remove 3967168404Spjd * buffers from the hash table when we arc_free(). 3968168404Spjd */ 3969219089Spjd if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3970219089Spjd if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3971219089Spjd panic("bad overwrite, hdr=%p exists=%p", 3972219089Spjd (void *)hdr, (void *)exists); 3973219089Spjd ASSERT(refcount_is_zero(&exists->b_refcnt)); 3974219089Spjd arc_change_state(arc_anon, exists, hash_lock); 3975219089Spjd mutex_exit(hash_lock); 3976219089Spjd arc_hdr_destroy(exists); 3977219089Spjd exists = buf_hash_insert(hdr, &hash_lock); 3978219089Spjd ASSERT3P(exists, ==, NULL); 3979243524Smm } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3980243524Smm /* nopwrite */ 3981243524Smm ASSERT(zio->io_prop.zp_nopwrite); 3982243524Smm if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3983243524Smm panic("bad nopwrite, hdr=%p exists=%p", 3984243524Smm (void *)hdr, (void *)exists); 3985219089Spjd } else { 3986219089Spjd /* Dedup */ 3987219089Spjd ASSERT(hdr->b_datacnt == 1); 3988219089Spjd ASSERT(hdr->b_state == arc_anon); 3989219089Spjd ASSERT(BP_GET_DEDUP(zio->io_bp)); 3990219089Spjd ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3991219089Spjd } 3992168404Spjd } 3993275811Sdelphij hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 3994185029Spjd /* if it's not anon, we are doing a scrub */ 3995219089Spjd if (!exists && hdr->b_state == arc_anon) 3996185029Spjd arc_access(hdr, hash_lock); 3997168404Spjd mutex_exit(hash_lock); 3998168404Spjd } else { 3999275811Sdelphij hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4000168404Spjd } 4001168404Spjd 4002219089Spjd ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 4003219089Spjd callback->awcb_done(zio, buf, callback->awcb_private); 4004168404Spjd 4005168404Spjd kmem_free(callback, sizeof (arc_write_callback_t)); 4006168404Spjd} 4007168404Spjd 4008168404Spjdzio_t * 4009219089Spjdarc_write(zio_t *pio, spa_t *spa, uint64_t txg, 4010251478Sdelphij blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 4011258632Savg const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 4012258632Savg arc_done_func_t *done, void *private, zio_priority_t priority, 4013268123Sdelphij int zio_flags, const zbookmark_phys_t *zb) 4014168404Spjd{ 4015168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 4016168404Spjd arc_write_callback_t *callback; 4017185029Spjd zio_t *zio; 4018168404Spjd 4019185029Spjd ASSERT(ready != NULL); 4020219089Spjd ASSERT(done != NULL); 4021168404Spjd ASSERT(!HDR_IO_ERROR(hdr)); 4022275811Sdelphij ASSERT((hdr->b_flags & ARC_FLAG_IO_IN_PROGRESS) == 0); 4023219089Spjd ASSERT(hdr->b_acb == NULL); 4024185029Spjd if (l2arc) 4025275811Sdelphij hdr->b_flags |= ARC_FLAG_L2CACHE; 4026251478Sdelphij if (l2arc_compress) 4027275811Sdelphij hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4028168404Spjd callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 4029168404Spjd callback->awcb_ready = ready; 4030258632Savg callback->awcb_physdone = physdone; 4031168404Spjd callback->awcb_done = done; 4032168404Spjd callback->awcb_private = private; 4033168404Spjd callback->awcb_buf = buf; 4034168404Spjd 4035219089Spjd zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 4036258632Savg arc_write_ready, arc_write_physdone, arc_write_done, callback, 4037258632Savg priority, zio_flags, zb); 4038185029Spjd 4039168404Spjd return (zio); 4040168404Spjd} 4041168404Spjd 4042185029Spjdstatic int 4043258632Savgarc_memory_throttle(uint64_t reserve, uint64_t txg) 4044185029Spjd{ 4045185029Spjd#ifdef _KERNEL 4046272483Ssmh uint64_t available_memory = ptob(freemem); 4047185029Spjd static uint64_t page_load = 0; 4048185029Spjd static uint64_t last_txg = 0; 4049185029Spjd 4050272483Ssmh#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 4051185029Spjd available_memory = 4052272483Ssmh MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 4053185029Spjd#endif 4054258632Savg 4055272483Ssmh if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 4056185029Spjd return (0); 4057185029Spjd 4058185029Spjd if (txg > last_txg) { 4059185029Spjd last_txg = txg; 4060185029Spjd page_load = 0; 4061185029Spjd } 4062185029Spjd /* 4063185029Spjd * If we are in pageout, we know that memory is already tight, 4064185029Spjd * the arc is already going to be evicting, so we just want to 4065185029Spjd * continue to let page writes occur as quickly as possible. 4066185029Spjd */ 4067185029Spjd if (curproc == pageproc) { 4068272483Ssmh if (page_load > MAX(ptob(minfree), available_memory) / 4) 4069249195Smm return (SET_ERROR(ERESTART)); 4070185029Spjd /* Note: reserve is inflated, so we deflate */ 4071185029Spjd page_load += reserve / 8; 4072185029Spjd return (0); 4073185029Spjd } else if (page_load > 0 && arc_reclaim_needed()) { 4074185029Spjd /* memory is low, delay before restarting */ 4075185029Spjd ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 4076249195Smm return (SET_ERROR(EAGAIN)); 4077185029Spjd } 4078185029Spjd page_load = 0; 4079185029Spjd#endif 4080185029Spjd return (0); 4081185029Spjd} 4082185029Spjd 4083168404Spjdvoid 4084185029Spjdarc_tempreserve_clear(uint64_t reserve) 4085168404Spjd{ 4086185029Spjd atomic_add_64(&arc_tempreserve, -reserve); 4087168404Spjd ASSERT((int64_t)arc_tempreserve >= 0); 4088168404Spjd} 4089168404Spjd 4090168404Spjdint 4091185029Spjdarc_tempreserve_space(uint64_t reserve, uint64_t txg) 4092168404Spjd{ 4093185029Spjd int error; 4094209962Smm uint64_t anon_size; 4095185029Spjd 4096272483Ssmh if (reserve > arc_c/4 && !arc_no_grow) { 4097185029Spjd arc_c = MIN(arc_c_max, reserve * 4); 4098272483Ssmh DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 4099272483Ssmh } 4100185029Spjd if (reserve > arc_c) 4101249195Smm return (SET_ERROR(ENOMEM)); 4102168404Spjd 4103168404Spjd /* 4104209962Smm * Don't count loaned bufs as in flight dirty data to prevent long 4105209962Smm * network delays from blocking transactions that are ready to be 4106209962Smm * assigned to a txg. 4107209962Smm */ 4108209962Smm anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 4109209962Smm 4110209962Smm /* 4111185029Spjd * Writes will, almost always, require additional memory allocations 4112251631Sdelphij * in order to compress/encrypt/etc the data. We therefore need to 4113185029Spjd * make sure that there is sufficient available memory for this. 4114185029Spjd */ 4115258632Savg error = arc_memory_throttle(reserve, txg); 4116258632Savg if (error != 0) 4117185029Spjd return (error); 4118185029Spjd 4119185029Spjd /* 4120168404Spjd * Throttle writes when the amount of dirty data in the cache 4121168404Spjd * gets too large. We try to keep the cache less than half full 4122168404Spjd * of dirty blocks so that our sync times don't grow too large. 4123168404Spjd * Note: if two requests come in concurrently, we might let them 4124168404Spjd * both succeed, when one of them should fail. Not a huge deal. 4125168404Spjd */ 4126209962Smm 4127209962Smm if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 4128209962Smm anon_size > arc_c / 4) { 4129185029Spjd dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 4130185029Spjd "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 4131185029Spjd arc_tempreserve>>10, 4132185029Spjd arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 4133185029Spjd arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 4134185029Spjd reserve>>10, arc_c>>10); 4135249195Smm return (SET_ERROR(ERESTART)); 4136168404Spjd } 4137185029Spjd atomic_add_64(&arc_tempreserve, reserve); 4138168404Spjd return (0); 4139168404Spjd} 4140168404Spjd 4141168582Spjdstatic kmutex_t arc_lowmem_lock; 4142168404Spjd#ifdef _KERNEL 4143168566Spjdstatic eventhandler_tag arc_event_lowmem = NULL; 4144168404Spjd 4145168404Spjdstatic void 4146168566Spjdarc_lowmem(void *arg __unused, int howto __unused) 4147168404Spjd{ 4148168404Spjd 4149168566Spjd /* Serialize access via arc_lowmem_lock. */ 4150168566Spjd mutex_enter(&arc_lowmem_lock); 4151219089Spjd mutex_enter(&arc_reclaim_thr_lock); 4152185029Spjd needfree = 1; 4153272483Ssmh DTRACE_PROBE(arc__needfree); 4154168404Spjd cv_signal(&arc_reclaim_thr_cv); 4155241773Savg 4156241773Savg /* 4157241773Savg * It is unsafe to block here in arbitrary threads, because we can come 4158241773Savg * here from ARC itself and may hold ARC locks and thus risk a deadlock 4159241773Savg * with ARC reclaim thread. 4160241773Savg */ 4161241773Savg if (curproc == pageproc) { 4162241773Savg while (needfree) 4163241773Savg msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4164241773Savg } 4165219089Spjd mutex_exit(&arc_reclaim_thr_lock); 4166168566Spjd mutex_exit(&arc_lowmem_lock); 4167168404Spjd} 4168168404Spjd#endif 4169168404Spjd 4170168404Spjdvoid 4171168404Spjdarc_init(void) 4172168404Spjd{ 4173219089Spjd int i, prefetch_tunable_set = 0; 4174205231Skmacy 4175168404Spjd mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4176168404Spjd cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4177168566Spjd mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4178168404Spjd 4179168404Spjd /* Convert seconds to clock ticks */ 4180168404Spjd arc_min_prefetch_lifespan = 1 * hz; 4181168404Spjd 4182168404Spjd /* Start out with 1/8 of all memory */ 4183168566Spjd arc_c = kmem_size() / 8; 4184219089Spjd 4185277300Ssmh#ifdef illumos 4186192360Skmacy#ifdef _KERNEL 4187192360Skmacy /* 4188192360Skmacy * On architectures where the physical memory can be larger 4189192360Skmacy * than the addressable space (intel in 32-bit mode), we may 4190192360Skmacy * need to limit the cache to 1/8 of VM size. 4191192360Skmacy */ 4192192360Skmacy arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4193192360Skmacy#endif 4194277300Ssmh#endif /* illumos */ 4195168566Spjd /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4196168566Spjd arc_c_min = MAX(arc_c / 4, 64<<18); 4197168566Spjd /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4198168404Spjd if (arc_c * 8 >= 1<<30) 4199168404Spjd arc_c_max = (arc_c * 8) - (1<<30); 4200168404Spjd else 4201168404Spjd arc_c_max = arc_c_min; 4202175633Spjd arc_c_max = MAX(arc_c * 5, arc_c_max); 4203219089Spjd 4204168481Spjd#ifdef _KERNEL 4205168404Spjd /* 4206168404Spjd * Allow the tunables to override our calculations if they are 4207168566Spjd * reasonable (ie. over 16MB) 4208168404Spjd */ 4209219089Spjd if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size()) 4210168404Spjd arc_c_max = zfs_arc_max; 4211219089Spjd if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max) 4212168404Spjd arc_c_min = zfs_arc_min; 4213168481Spjd#endif 4214219089Spjd 4215168404Spjd arc_c = arc_c_max; 4216168404Spjd arc_p = (arc_c >> 1); 4217168404Spjd 4218185029Spjd /* limit meta-data to 1/4 of the arc capacity */ 4219185029Spjd arc_meta_limit = arc_c_max / 4; 4220185029Spjd 4221185029Spjd /* Allow the tunable to override if it is reasonable */ 4222185029Spjd if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4223185029Spjd arc_meta_limit = zfs_arc_meta_limit; 4224185029Spjd 4225185029Spjd if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4226185029Spjd arc_c_min = arc_meta_limit / 2; 4227185029Spjd 4228275780Sdelphij if (zfs_arc_meta_min > 0) { 4229275780Sdelphij arc_meta_min = zfs_arc_meta_min; 4230275780Sdelphij } else { 4231275780Sdelphij arc_meta_min = arc_c_min / 2; 4232275780Sdelphij } 4233275780Sdelphij 4234208373Smm if (zfs_arc_grow_retry > 0) 4235208373Smm arc_grow_retry = zfs_arc_grow_retry; 4236208373Smm 4237208373Smm if (zfs_arc_shrink_shift > 0) 4238208373Smm arc_shrink_shift = zfs_arc_shrink_shift; 4239208373Smm 4240208373Smm if (zfs_arc_p_min_shift > 0) 4241208373Smm arc_p_min_shift = zfs_arc_p_min_shift; 4242208373Smm 4243168404Spjd /* if kmem_flags are set, lets try to use less memory */ 4244168404Spjd if (kmem_debugging()) 4245168404Spjd arc_c = arc_c / 2; 4246168404Spjd if (arc_c < arc_c_min) 4247168404Spjd arc_c = arc_c_min; 4248168404Spjd 4249168473Spjd zfs_arc_min = arc_c_min; 4250168473Spjd zfs_arc_max = arc_c_max; 4251168473Spjd 4252168404Spjd arc_anon = &ARC_anon; 4253168404Spjd arc_mru = &ARC_mru; 4254168404Spjd arc_mru_ghost = &ARC_mru_ghost; 4255168404Spjd arc_mfu = &ARC_mfu; 4256168404Spjd arc_mfu_ghost = &ARC_mfu_ghost; 4257185029Spjd arc_l2c_only = &ARC_l2c_only; 4258168404Spjd arc_size = 0; 4259168404Spjd 4260205231Skmacy for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4261205231Skmacy mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4262205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4263205231Skmacy mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4264205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4265205231Skmacy mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4266205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4267205231Skmacy mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4268205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4269205231Skmacy mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4270205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4271205231Skmacy mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4272205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4273206796Spjd 4274205231Skmacy list_create(&arc_mru->arcs_lists[i], 4275205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4276205231Skmacy list_create(&arc_mru_ghost->arcs_lists[i], 4277205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4278205231Skmacy list_create(&arc_mfu->arcs_lists[i], 4279205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4280205231Skmacy list_create(&arc_mfu_ghost->arcs_lists[i], 4281205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4282205231Skmacy list_create(&arc_mfu_ghost->arcs_lists[i], 4283205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4284205231Skmacy list_create(&arc_l2c_only->arcs_lists[i], 4285205231Skmacy sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4286205231Skmacy } 4287168404Spjd 4288168404Spjd buf_init(); 4289168404Spjd 4290168404Spjd arc_thread_exit = 0; 4291168404Spjd arc_eviction_list = NULL; 4292168404Spjd mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4293168404Spjd bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4294168404Spjd 4295168404Spjd arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4296168404Spjd sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4297168404Spjd 4298168404Spjd if (arc_ksp != NULL) { 4299168404Spjd arc_ksp->ks_data = &arc_stats; 4300168404Spjd kstat_install(arc_ksp); 4301168404Spjd } 4302168404Spjd 4303168404Spjd (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4304168404Spjd TS_RUN, minclsyspri); 4305168404Spjd 4306168404Spjd#ifdef _KERNEL 4307168566Spjd arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4308168404Spjd EVENTHANDLER_PRI_FIRST); 4309168404Spjd#endif 4310168404Spjd 4311168404Spjd arc_dead = FALSE; 4312185029Spjd arc_warm = B_FALSE; 4313168566Spjd 4314258632Savg /* 4315258632Savg * Calculate maximum amount of dirty data per pool. 4316258632Savg * 4317258632Savg * If it has been set by /etc/system, take that. 4318258632Savg * Otherwise, use a percentage of physical memory defined by 4319258632Savg * zfs_dirty_data_max_percent (default 10%) with a cap at 4320258632Savg * zfs_dirty_data_max_max (default 4GB). 4321258632Savg */ 4322258632Savg if (zfs_dirty_data_max == 0) { 4323258632Savg zfs_dirty_data_max = ptob(physmem) * 4324258632Savg zfs_dirty_data_max_percent / 100; 4325258632Savg zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4326258632Savg zfs_dirty_data_max_max); 4327258632Savg } 4328185029Spjd 4329168566Spjd#ifdef _KERNEL 4330194043Skmacy if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4331193953Skmacy prefetch_tunable_set = 1; 4332206796Spjd 4333193878Skmacy#ifdef __i386__ 4334193953Skmacy if (prefetch_tunable_set == 0) { 4335196863Strasz printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4336196863Strasz "-- to enable,\n"); 4337196863Strasz printf(" add \"vfs.zfs.prefetch_disable=0\" " 4338196863Strasz "to /boot/loader.conf.\n"); 4339219089Spjd zfs_prefetch_disable = 1; 4340193878Skmacy } 4341206796Spjd#else 4342193878Skmacy if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4343193953Skmacy prefetch_tunable_set == 0) { 4344196863Strasz printf("ZFS NOTICE: Prefetch is disabled by default if less " 4345196941Strasz "than 4GB of RAM is present;\n" 4346196863Strasz " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4347196863Strasz "to /boot/loader.conf.\n"); 4348219089Spjd zfs_prefetch_disable = 1; 4349193878Skmacy } 4350206796Spjd#endif 4351175633Spjd /* Warn about ZFS memory and address space requirements. */ 4352168696Spjd if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4353168987Sbmah printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4354168987Sbmah "expect unstable behavior.\n"); 4355175633Spjd } 4356175633Spjd if (kmem_size() < 512 * (1 << 20)) { 4357173419Spjd printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4358168987Sbmah "expect unstable behavior.\n"); 4359185029Spjd printf(" Consider tuning vm.kmem_size and " 4360173419Spjd "vm.kmem_size_max\n"); 4361185029Spjd printf(" in /boot/loader.conf.\n"); 4362168566Spjd } 4363168566Spjd#endif 4364168404Spjd} 4365168404Spjd 4366168404Spjdvoid 4367168404Spjdarc_fini(void) 4368168404Spjd{ 4369205231Skmacy int i; 4370206796Spjd 4371168404Spjd mutex_enter(&arc_reclaim_thr_lock); 4372168404Spjd arc_thread_exit = 1; 4373168404Spjd cv_signal(&arc_reclaim_thr_cv); 4374168404Spjd while (arc_thread_exit != 0) 4375168404Spjd cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4376168404Spjd mutex_exit(&arc_reclaim_thr_lock); 4377168404Spjd 4378185029Spjd arc_flush(NULL); 4379168404Spjd 4380168404Spjd arc_dead = TRUE; 4381168404Spjd 4382168404Spjd if (arc_ksp != NULL) { 4383168404Spjd kstat_delete(arc_ksp); 4384168404Spjd arc_ksp = NULL; 4385168404Spjd } 4386168404Spjd 4387168404Spjd mutex_destroy(&arc_eviction_mtx); 4388168404Spjd mutex_destroy(&arc_reclaim_thr_lock); 4389168404Spjd cv_destroy(&arc_reclaim_thr_cv); 4390168404Spjd 4391205231Skmacy for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4392205231Skmacy list_destroy(&arc_mru->arcs_lists[i]); 4393205231Skmacy list_destroy(&arc_mru_ghost->arcs_lists[i]); 4394205231Skmacy list_destroy(&arc_mfu->arcs_lists[i]); 4395205231Skmacy list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4396206795Spjd list_destroy(&arc_l2c_only->arcs_lists[i]); 4397168404Spjd 4398205231Skmacy mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4399205231Skmacy mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4400205231Skmacy mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4401205231Skmacy mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4402205231Skmacy mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4403206795Spjd mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4404205231Skmacy } 4405206796Spjd 4406168404Spjd buf_fini(); 4407168404Spjd 4408209962Smm ASSERT(arc_loaned_bytes == 0); 4409209962Smm 4410168582Spjd mutex_destroy(&arc_lowmem_lock); 4411168404Spjd#ifdef _KERNEL 4412168566Spjd if (arc_event_lowmem != NULL) 4413168566Spjd EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4414168404Spjd#endif 4415168404Spjd} 4416185029Spjd 4417185029Spjd/* 4418185029Spjd * Level 2 ARC 4419185029Spjd * 4420185029Spjd * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4421185029Spjd * It uses dedicated storage devices to hold cached data, which are populated 4422185029Spjd * using large infrequent writes. The main role of this cache is to boost 4423185029Spjd * the performance of random read workloads. The intended L2ARC devices 4424185029Spjd * include short-stroked disks, solid state disks, and other media with 4425185029Spjd * substantially faster read latency than disk. 4426185029Spjd * 4427185029Spjd * +-----------------------+ 4428185029Spjd * | ARC | 4429185029Spjd * +-----------------------+ 4430185029Spjd * | ^ ^ 4431185029Spjd * | | | 4432185029Spjd * l2arc_feed_thread() arc_read() 4433185029Spjd * | | | 4434185029Spjd * | l2arc read | 4435185029Spjd * V | | 4436185029Spjd * +---------------+ | 4437185029Spjd * | L2ARC | | 4438185029Spjd * +---------------+ | 4439185029Spjd * | ^ | 4440185029Spjd * l2arc_write() | | 4441185029Spjd * | | | 4442185029Spjd * V | | 4443185029Spjd * +-------+ +-------+ 4444185029Spjd * | vdev | | vdev | 4445185029Spjd * | cache | | cache | 4446185029Spjd * +-------+ +-------+ 4447185029Spjd * +=========+ .-----. 4448185029Spjd * : L2ARC : |-_____-| 4449185029Spjd * : devices : | Disks | 4450185029Spjd * +=========+ `-_____-' 4451185029Spjd * 4452185029Spjd * Read requests are satisfied from the following sources, in order: 4453185029Spjd * 4454185029Spjd * 1) ARC 4455185029Spjd * 2) vdev cache of L2ARC devices 4456185029Spjd * 3) L2ARC devices 4457185029Spjd * 4) vdev cache of disks 4458185029Spjd * 5) disks 4459185029Spjd * 4460185029Spjd * Some L2ARC device types exhibit extremely slow write performance. 4461185029Spjd * To accommodate for this there are some significant differences between 4462185029Spjd * the L2ARC and traditional cache design: 4463185029Spjd * 4464185029Spjd * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4465185029Spjd * the ARC behave as usual, freeing buffers and placing headers on ghost 4466185029Spjd * lists. The ARC does not send buffers to the L2ARC during eviction as 4467185029Spjd * this would add inflated write latencies for all ARC memory pressure. 4468185029Spjd * 4469185029Spjd * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4470185029Spjd * It does this by periodically scanning buffers from the eviction-end of 4471185029Spjd * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4472251478Sdelphij * not already there. It scans until a headroom of buffers is satisfied, 4473251478Sdelphij * which itself is a buffer for ARC eviction. If a compressible buffer is 4474251478Sdelphij * found during scanning and selected for writing to an L2ARC device, we 4475251478Sdelphij * temporarily boost scanning headroom during the next scan cycle to make 4476251478Sdelphij * sure we adapt to compression effects (which might significantly reduce 4477251478Sdelphij * the data volume we write to L2ARC). The thread that does this is 4478185029Spjd * l2arc_feed_thread(), illustrated below; example sizes are included to 4479185029Spjd * provide a better sense of ratio than this diagram: 4480185029Spjd * 4481185029Spjd * head --> tail 4482185029Spjd * +---------------------+----------+ 4483185029Spjd * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4484185029Spjd * +---------------------+----------+ | o L2ARC eligible 4485185029Spjd * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4486185029Spjd * +---------------------+----------+ | 4487185029Spjd * 15.9 Gbytes ^ 32 Mbytes | 4488185029Spjd * headroom | 4489185029Spjd * l2arc_feed_thread() 4490185029Spjd * | 4491185029Spjd * l2arc write hand <--[oooo]--' 4492185029Spjd * | 8 Mbyte 4493185029Spjd * | write max 4494185029Spjd * V 4495185029Spjd * +==============================+ 4496185029Spjd * L2ARC dev |####|#|###|###| |####| ... | 4497185029Spjd * +==============================+ 4498185029Spjd * 32 Gbytes 4499185029Spjd * 4500185029Spjd * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4501185029Spjd * evicted, then the L2ARC has cached a buffer much sooner than it probably 4502185029Spjd * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4503185029Spjd * safe to say that this is an uncommon case, since buffers at the end of 4504185029Spjd * the ARC lists have moved there due to inactivity. 4505185029Spjd * 4506185029Spjd * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4507185029Spjd * then the L2ARC simply misses copying some buffers. This serves as a 4508185029Spjd * pressure valve to prevent heavy read workloads from both stalling the ARC 4509185029Spjd * with waits and clogging the L2ARC with writes. This also helps prevent 4510185029Spjd * the potential for the L2ARC to churn if it attempts to cache content too 4511185029Spjd * quickly, such as during backups of the entire pool. 4512185029Spjd * 4513185029Spjd * 5. After system boot and before the ARC has filled main memory, there are 4514185029Spjd * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4515185029Spjd * lists can remain mostly static. Instead of searching from tail of these 4516185029Spjd * lists as pictured, the l2arc_feed_thread() will search from the list heads 4517185029Spjd * for eligible buffers, greatly increasing its chance of finding them. 4518185029Spjd * 4519185029Spjd * The L2ARC device write speed is also boosted during this time so that 4520185029Spjd * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4521185029Spjd * there are no L2ARC reads, and no fear of degrading read performance 4522185029Spjd * through increased writes. 4523185029Spjd * 4524185029Spjd * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4525185029Spjd * the vdev queue can aggregate them into larger and fewer writes. Each 4526185029Spjd * device is written to in a rotor fashion, sweeping writes through 4527185029Spjd * available space then repeating. 4528185029Spjd * 4529185029Spjd * 7. The L2ARC does not store dirty content. It never needs to flush 4530185029Spjd * write buffers back to disk based storage. 4531185029Spjd * 4532185029Spjd * 8. If an ARC buffer is written (and dirtied) which also exists in the 4533185029Spjd * L2ARC, the now stale L2ARC buffer is immediately dropped. 4534185029Spjd * 4535185029Spjd * The performance of the L2ARC can be tweaked by a number of tunables, which 4536185029Spjd * may be necessary for different workloads: 4537185029Spjd * 4538185029Spjd * l2arc_write_max max write bytes per interval 4539185029Spjd * l2arc_write_boost extra write bytes during device warmup 4540185029Spjd * l2arc_noprefetch skip caching prefetched buffers 4541185029Spjd * l2arc_headroom number of max device writes to precache 4542251478Sdelphij * l2arc_headroom_boost when we find compressed buffers during ARC 4543251478Sdelphij * scanning, we multiply headroom by this 4544251478Sdelphij * percentage factor for the next scan cycle, 4545251478Sdelphij * since more compressed buffers are likely to 4546251478Sdelphij * be present 4547185029Spjd * l2arc_feed_secs seconds between L2ARC writing 4548185029Spjd * 4549185029Spjd * Tunables may be removed or added as future performance improvements are 4550185029Spjd * integrated, and also may become zpool properties. 4551208373Smm * 4552208373Smm * There are three key functions that control how the L2ARC warms up: 4553208373Smm * 4554208373Smm * l2arc_write_eligible() check if a buffer is eligible to cache 4555208373Smm * l2arc_write_size() calculate how much to write 4556208373Smm * l2arc_write_interval() calculate sleep delay between writes 4557208373Smm * 4558208373Smm * These three functions determine what to write, how much, and how quickly 4559208373Smm * to send writes. 4560185029Spjd */ 4561185029Spjd 4562208373Smmstatic boolean_t 4563275811Sdelphijl2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 4564208373Smm{ 4565208373Smm /* 4566208373Smm * A buffer is *not* eligible for the L2ARC if it: 4567208373Smm * 1. belongs to a different spa. 4568208373Smm * 2. is already cached on the L2ARC. 4569208373Smm * 3. has an I/O in progress (it may be an incomplete read). 4570208373Smm * 4. is flagged not eligible (zfs property). 4571208373Smm */ 4572275811Sdelphij if (hdr->b_spa != spa_guid) { 4573208373Smm ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4574208373Smm return (B_FALSE); 4575208373Smm } 4576275811Sdelphij if (hdr->b_l2hdr != NULL) { 4577208373Smm ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4578208373Smm return (B_FALSE); 4579208373Smm } 4580275811Sdelphij if (HDR_IO_IN_PROGRESS(hdr)) { 4581208373Smm ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4582208373Smm return (B_FALSE); 4583208373Smm } 4584275811Sdelphij if (!HDR_L2CACHE(hdr)) { 4585208373Smm ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4586208373Smm return (B_FALSE); 4587208373Smm } 4588208373Smm 4589208373Smm return (B_TRUE); 4590208373Smm} 4591208373Smm 4592208373Smmstatic uint64_t 4593251478Sdelphijl2arc_write_size(void) 4594208373Smm{ 4595208373Smm uint64_t size; 4596208373Smm 4597251478Sdelphij /* 4598251478Sdelphij * Make sure our globals have meaningful values in case the user 4599251478Sdelphij * altered them. 4600251478Sdelphij */ 4601251478Sdelphij size = l2arc_write_max; 4602251478Sdelphij if (size == 0) { 4603251478Sdelphij cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4604251478Sdelphij "be greater than zero, resetting it to the default (%d)", 4605251478Sdelphij L2ARC_WRITE_SIZE); 4606251478Sdelphij size = l2arc_write_max = L2ARC_WRITE_SIZE; 4607251478Sdelphij } 4608208373Smm 4609208373Smm if (arc_warm == B_FALSE) 4610251478Sdelphij size += l2arc_write_boost; 4611208373Smm 4612208373Smm return (size); 4613208373Smm 4614208373Smm} 4615208373Smm 4616208373Smmstatic clock_t 4617208373Smml2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4618208373Smm{ 4619219089Spjd clock_t interval, next, now; 4620208373Smm 4621208373Smm /* 4622208373Smm * If the ARC lists are busy, increase our write rate; if the 4623208373Smm * lists are stale, idle back. This is achieved by checking 4624208373Smm * how much we previously wrote - if it was more than half of 4625208373Smm * what we wanted, schedule the next write much sooner. 4626208373Smm */ 4627208373Smm if (l2arc_feed_again && wrote > (wanted / 2)) 4628208373Smm interval = (hz * l2arc_feed_min_ms) / 1000; 4629208373Smm else 4630208373Smm interval = hz * l2arc_feed_secs; 4631208373Smm 4632219089Spjd now = ddi_get_lbolt(); 4633219089Spjd next = MAX(now, MIN(now + interval, began + interval)); 4634208373Smm 4635208373Smm return (next); 4636208373Smm} 4637208373Smm 4638185029Spjdstatic void 4639185029Spjdl2arc_hdr_stat_add(void) 4640185029Spjd{ 4641185029Spjd ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4642185029Spjd ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4643185029Spjd} 4644185029Spjd 4645185029Spjdstatic void 4646185029Spjdl2arc_hdr_stat_remove(void) 4647185029Spjd{ 4648185029Spjd ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4649185029Spjd ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4650185029Spjd} 4651185029Spjd 4652185029Spjd/* 4653185029Spjd * Cycle through L2ARC devices. This is how L2ARC load balances. 4654185029Spjd * If a device is returned, this also returns holding the spa config lock. 4655185029Spjd */ 4656185029Spjdstatic l2arc_dev_t * 4657185029Spjdl2arc_dev_get_next(void) 4658185029Spjd{ 4659185029Spjd l2arc_dev_t *first, *next = NULL; 4660185029Spjd 4661185029Spjd /* 4662185029Spjd * Lock out the removal of spas (spa_namespace_lock), then removal 4663185029Spjd * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4664185029Spjd * both locks will be dropped and a spa config lock held instead. 4665185029Spjd */ 4666185029Spjd mutex_enter(&spa_namespace_lock); 4667185029Spjd mutex_enter(&l2arc_dev_mtx); 4668185029Spjd 4669185029Spjd /* if there are no vdevs, there is nothing to do */ 4670185029Spjd if (l2arc_ndev == 0) 4671185029Spjd goto out; 4672185029Spjd 4673185029Spjd first = NULL; 4674185029Spjd next = l2arc_dev_last; 4675185029Spjd do { 4676185029Spjd /* loop around the list looking for a non-faulted vdev */ 4677185029Spjd if (next == NULL) { 4678185029Spjd next = list_head(l2arc_dev_list); 4679185029Spjd } else { 4680185029Spjd next = list_next(l2arc_dev_list, next); 4681185029Spjd if (next == NULL) 4682185029Spjd next = list_head(l2arc_dev_list); 4683185029Spjd } 4684185029Spjd 4685185029Spjd /* if we have come back to the start, bail out */ 4686185029Spjd if (first == NULL) 4687185029Spjd first = next; 4688185029Spjd else if (next == first) 4689185029Spjd break; 4690185029Spjd 4691185029Spjd } while (vdev_is_dead(next->l2ad_vdev)); 4692185029Spjd 4693185029Spjd /* if we were unable to find any usable vdevs, return NULL */ 4694185029Spjd if (vdev_is_dead(next->l2ad_vdev)) 4695185029Spjd next = NULL; 4696185029Spjd 4697185029Spjd l2arc_dev_last = next; 4698185029Spjd 4699185029Spjdout: 4700185029Spjd mutex_exit(&l2arc_dev_mtx); 4701185029Spjd 4702185029Spjd /* 4703185029Spjd * Grab the config lock to prevent the 'next' device from being 4704185029Spjd * removed while we are writing to it. 4705185029Spjd */ 4706185029Spjd if (next != NULL) 4707185029Spjd spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4708185029Spjd mutex_exit(&spa_namespace_lock); 4709185029Spjd 4710185029Spjd return (next); 4711185029Spjd} 4712185029Spjd 4713185029Spjd/* 4714185029Spjd * Free buffers that were tagged for destruction. 4715185029Spjd */ 4716185029Spjdstatic void 4717185029Spjdl2arc_do_free_on_write() 4718185029Spjd{ 4719185029Spjd list_t *buflist; 4720185029Spjd l2arc_data_free_t *df, *df_prev; 4721185029Spjd 4722185029Spjd mutex_enter(&l2arc_free_on_write_mtx); 4723185029Spjd buflist = l2arc_free_on_write; 4724185029Spjd 4725185029Spjd for (df = list_tail(buflist); df; df = df_prev) { 4726185029Spjd df_prev = list_prev(buflist, df); 4727185029Spjd ASSERT(df->l2df_data != NULL); 4728185029Spjd ASSERT(df->l2df_func != NULL); 4729185029Spjd df->l2df_func(df->l2df_data, df->l2df_size); 4730185029Spjd list_remove(buflist, df); 4731185029Spjd kmem_free(df, sizeof (l2arc_data_free_t)); 4732185029Spjd } 4733185029Spjd 4734185029Spjd mutex_exit(&l2arc_free_on_write_mtx); 4735185029Spjd} 4736185029Spjd 4737185029Spjd/* 4738185029Spjd * A write to a cache device has completed. Update all headers to allow 4739185029Spjd * reads from these buffers to begin. 4740185029Spjd */ 4741185029Spjdstatic void 4742185029Spjdl2arc_write_done(zio_t *zio) 4743185029Spjd{ 4744185029Spjd l2arc_write_callback_t *cb; 4745185029Spjd l2arc_dev_t *dev; 4746185029Spjd list_t *buflist; 4747275811Sdelphij arc_buf_hdr_t *head, *hdr, *hdr_prev; 4748185029Spjd l2arc_buf_hdr_t *abl2; 4749185029Spjd kmutex_t *hash_lock; 4750268085Sdelphij int64_t bytes_dropped = 0; 4751185029Spjd 4752185029Spjd cb = zio->io_private; 4753185029Spjd ASSERT(cb != NULL); 4754185029Spjd dev = cb->l2wcb_dev; 4755185029Spjd ASSERT(dev != NULL); 4756185029Spjd head = cb->l2wcb_head; 4757185029Spjd ASSERT(head != NULL); 4758185029Spjd buflist = dev->l2ad_buflist; 4759185029Spjd ASSERT(buflist != NULL); 4760185029Spjd DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4761185029Spjd l2arc_write_callback_t *, cb); 4762185029Spjd 4763185029Spjd if (zio->io_error != 0) 4764185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_error); 4765185029Spjd 4766185029Spjd mutex_enter(&l2arc_buflist_mtx); 4767185029Spjd 4768185029Spjd /* 4769185029Spjd * All writes completed, or an error was hit. 4770185029Spjd */ 4771275811Sdelphij for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 4772275811Sdelphij hdr_prev = list_prev(buflist, hdr); 4773275811Sdelphij abl2 = hdr->b_l2hdr; 4774185029Spjd 4775260835Sdelphij /* 4776260835Sdelphij * Release the temporary compressed buffer as soon as possible. 4777260835Sdelphij */ 4778260835Sdelphij if (abl2->b_compress != ZIO_COMPRESS_OFF) 4779275811Sdelphij l2arc_release_cdata_buf(hdr); 4780260835Sdelphij 4781275811Sdelphij hash_lock = HDR_LOCK(hdr); 4782185029Spjd if (!mutex_tryenter(hash_lock)) { 4783185029Spjd /* 4784185029Spjd * This buffer misses out. It may be in a stage 4785185029Spjd * of eviction. Its ARC_L2_WRITING flag will be 4786185029Spjd * left set, denying reads to this buffer. 4787185029Spjd */ 4788185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4789185029Spjd continue; 4790185029Spjd } 4791185029Spjd 4792185029Spjd if (zio->io_error != 0) { 4793185029Spjd /* 4794185029Spjd * Error - drop L2ARC entry. 4795185029Spjd */ 4796275811Sdelphij list_remove(buflist, hdr); 4797251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4798268085Sdelphij bytes_dropped += abl2->b_asize; 4799275811Sdelphij hdr->b_l2hdr = NULL; 4800248572Ssmh trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4801275811Sdelphij hdr->b_size, 0); 4802185029Spjd kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4803275811Sdelphij ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 4804185029Spjd } 4805185029Spjd 4806185029Spjd /* 4807185029Spjd * Allow ARC to begin reads to this L2ARC entry. 4808185029Spjd */ 4809275811Sdelphij hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 4810185029Spjd 4811185029Spjd mutex_exit(hash_lock); 4812185029Spjd } 4813185029Spjd 4814185029Spjd atomic_inc_64(&l2arc_writes_done); 4815185029Spjd list_remove(buflist, head); 4816185029Spjd kmem_cache_free(hdr_cache, head); 4817185029Spjd mutex_exit(&l2arc_buflist_mtx); 4818185029Spjd 4819268085Sdelphij vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 4820268085Sdelphij 4821185029Spjd l2arc_do_free_on_write(); 4822185029Spjd 4823185029Spjd kmem_free(cb, sizeof (l2arc_write_callback_t)); 4824185029Spjd} 4825185029Spjd 4826185029Spjd/* 4827185029Spjd * A read to a cache device completed. Validate buffer contents before 4828185029Spjd * handing over to the regular ARC routines. 4829185029Spjd */ 4830185029Spjdstatic void 4831185029Spjdl2arc_read_done(zio_t *zio) 4832185029Spjd{ 4833185029Spjd l2arc_read_callback_t *cb; 4834185029Spjd arc_buf_hdr_t *hdr; 4835185029Spjd arc_buf_t *buf; 4836185029Spjd kmutex_t *hash_lock; 4837185029Spjd int equal; 4838185029Spjd 4839185029Spjd ASSERT(zio->io_vd != NULL); 4840185029Spjd ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4841185029Spjd 4842185029Spjd spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4843185029Spjd 4844185029Spjd cb = zio->io_private; 4845185029Spjd ASSERT(cb != NULL); 4846185029Spjd buf = cb->l2rcb_buf; 4847185029Spjd ASSERT(buf != NULL); 4848185029Spjd 4849219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 4850185029Spjd mutex_enter(hash_lock); 4851219089Spjd hdr = buf->b_hdr; 4852219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4853185029Spjd 4854185029Spjd /* 4855251478Sdelphij * If the buffer was compressed, decompress it first. 4856251478Sdelphij */ 4857251478Sdelphij if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4858251478Sdelphij l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4859251478Sdelphij ASSERT(zio->io_data != NULL); 4860251478Sdelphij 4861251478Sdelphij /* 4862185029Spjd * Check this survived the L2ARC journey. 4863185029Spjd */ 4864185029Spjd equal = arc_cksum_equal(buf); 4865185029Spjd if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4866185029Spjd mutex_exit(hash_lock); 4867185029Spjd zio->io_private = buf; 4868185029Spjd zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4869185029Spjd zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4870185029Spjd arc_read_done(zio); 4871185029Spjd } else { 4872185029Spjd mutex_exit(hash_lock); 4873185029Spjd /* 4874185029Spjd * Buffer didn't survive caching. Increment stats and 4875185029Spjd * reissue to the original storage device. 4876185029Spjd */ 4877185029Spjd if (zio->io_error != 0) { 4878185029Spjd ARCSTAT_BUMP(arcstat_l2_io_error); 4879185029Spjd } else { 4880249195Smm zio->io_error = SET_ERROR(EIO); 4881185029Spjd } 4882185029Spjd if (!equal) 4883185029Spjd ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4884185029Spjd 4885185029Spjd /* 4886185029Spjd * If there's no waiter, issue an async i/o to the primary 4887185029Spjd * storage now. If there *is* a waiter, the caller must 4888185029Spjd * issue the i/o in a context where it's OK to block. 4889185029Spjd */ 4890209962Smm if (zio->io_waiter == NULL) { 4891209962Smm zio_t *pio = zio_unique_parent(zio); 4892209962Smm 4893209962Smm ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4894209962Smm 4895209962Smm zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4896185029Spjd buf->b_data, zio->io_size, arc_read_done, buf, 4897185029Spjd zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4898209962Smm } 4899185029Spjd } 4900185029Spjd 4901185029Spjd kmem_free(cb, sizeof (l2arc_read_callback_t)); 4902185029Spjd} 4903185029Spjd 4904185029Spjd/* 4905185029Spjd * This is the list priority from which the L2ARC will search for pages to 4906185029Spjd * cache. This is used within loops (0..3) to cycle through lists in the 4907185029Spjd * desired order. This order can have a significant effect on cache 4908185029Spjd * performance. 4909185029Spjd * 4910185029Spjd * Currently the metadata lists are hit first, MFU then MRU, followed by 4911185029Spjd * the data lists. This function returns a locked list, and also returns 4912185029Spjd * the lock pointer. 4913185029Spjd */ 4914185029Spjdstatic list_t * 4915185029Spjdl2arc_list_locked(int list_num, kmutex_t **lock) 4916185029Spjd{ 4917247187Smm list_t *list = NULL; 4918205231Skmacy int idx; 4919185029Spjd 4920206796Spjd ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4921206796Spjd 4922205231Skmacy if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4923205231Skmacy idx = list_num; 4924205231Skmacy list = &arc_mfu->arcs_lists[idx]; 4925205231Skmacy *lock = ARCS_LOCK(arc_mfu, idx); 4926206796Spjd } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4927205231Skmacy idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4928205231Skmacy list = &arc_mru->arcs_lists[idx]; 4929205231Skmacy *lock = ARCS_LOCK(arc_mru, idx); 4930206796Spjd } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4931205231Skmacy ARC_BUFC_NUMDATALISTS)) { 4932205231Skmacy idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4933205231Skmacy list = &arc_mfu->arcs_lists[idx]; 4934205231Skmacy *lock = ARCS_LOCK(arc_mfu, idx); 4935205231Skmacy } else { 4936205231Skmacy idx = list_num - ARC_BUFC_NUMLISTS; 4937205231Skmacy list = &arc_mru->arcs_lists[idx]; 4938205231Skmacy *lock = ARCS_LOCK(arc_mru, idx); 4939185029Spjd } 4940185029Spjd 4941185029Spjd ASSERT(!(MUTEX_HELD(*lock))); 4942185029Spjd mutex_enter(*lock); 4943185029Spjd return (list); 4944185029Spjd} 4945185029Spjd 4946185029Spjd/* 4947185029Spjd * Evict buffers from the device write hand to the distance specified in 4948185029Spjd * bytes. This distance may span populated buffers, it may span nothing. 4949185029Spjd * This is clearing a region on the L2ARC device ready for writing. 4950185029Spjd * If the 'all' boolean is set, every buffer is evicted. 4951185029Spjd */ 4952185029Spjdstatic void 4953185029Spjdl2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4954185029Spjd{ 4955185029Spjd list_t *buflist; 4956185029Spjd l2arc_buf_hdr_t *abl2; 4957275811Sdelphij arc_buf_hdr_t *hdr, *hdr_prev; 4958185029Spjd kmutex_t *hash_lock; 4959185029Spjd uint64_t taddr; 4960268085Sdelphij int64_t bytes_evicted = 0; 4961185029Spjd 4962185029Spjd buflist = dev->l2ad_buflist; 4963185029Spjd 4964185029Spjd if (buflist == NULL) 4965185029Spjd return; 4966185029Spjd 4967185029Spjd if (!all && dev->l2ad_first) { 4968185029Spjd /* 4969185029Spjd * This is the first sweep through the device. There is 4970185029Spjd * nothing to evict. 4971185029Spjd */ 4972185029Spjd return; 4973185029Spjd } 4974185029Spjd 4975185029Spjd if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4976185029Spjd /* 4977185029Spjd * When nearing the end of the device, evict to the end 4978185029Spjd * before the device write hand jumps to the start. 4979185029Spjd */ 4980185029Spjd taddr = dev->l2ad_end; 4981185029Spjd } else { 4982185029Spjd taddr = dev->l2ad_hand + distance; 4983185029Spjd } 4984185029Spjd DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4985185029Spjd uint64_t, taddr, boolean_t, all); 4986185029Spjd 4987185029Spjdtop: 4988185029Spjd mutex_enter(&l2arc_buflist_mtx); 4989275811Sdelphij for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 4990275811Sdelphij hdr_prev = list_prev(buflist, hdr); 4991185029Spjd 4992275811Sdelphij hash_lock = HDR_LOCK(hdr); 4993185029Spjd if (!mutex_tryenter(hash_lock)) { 4994185029Spjd /* 4995185029Spjd * Missed the hash lock. Retry. 4996185029Spjd */ 4997185029Spjd ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4998185029Spjd mutex_exit(&l2arc_buflist_mtx); 4999185029Spjd mutex_enter(hash_lock); 5000185029Spjd mutex_exit(hash_lock); 5001185029Spjd goto top; 5002185029Spjd } 5003185029Spjd 5004275811Sdelphij if (HDR_L2_WRITE_HEAD(hdr)) { 5005185029Spjd /* 5006185029Spjd * We hit a write head node. Leave it for 5007185029Spjd * l2arc_write_done(). 5008185029Spjd */ 5009275811Sdelphij list_remove(buflist, hdr); 5010185029Spjd mutex_exit(hash_lock); 5011185029Spjd continue; 5012185029Spjd } 5013185029Spjd 5014275811Sdelphij if (!all && hdr->b_l2hdr != NULL && 5015275811Sdelphij (hdr->b_l2hdr->b_daddr > taddr || 5016275811Sdelphij hdr->b_l2hdr->b_daddr < dev->l2ad_hand)) { 5017185029Spjd /* 5018185029Spjd * We've evicted to the target address, 5019185029Spjd * or the end of the device. 5020185029Spjd */ 5021185029Spjd mutex_exit(hash_lock); 5022185029Spjd break; 5023185029Spjd } 5024185029Spjd 5025275811Sdelphij if (HDR_FREE_IN_PROGRESS(hdr)) { 5026185029Spjd /* 5027185029Spjd * Already on the path to destruction. 5028185029Spjd */ 5029185029Spjd mutex_exit(hash_lock); 5030185029Spjd continue; 5031185029Spjd } 5032185029Spjd 5033275811Sdelphij if (hdr->b_state == arc_l2c_only) { 5034275811Sdelphij ASSERT(!HDR_L2_READING(hdr)); 5035185029Spjd /* 5036185029Spjd * This doesn't exist in the ARC. Destroy. 5037185029Spjd * arc_hdr_destroy() will call list_remove() 5038185029Spjd * and decrement arcstat_l2_size. 5039185029Spjd */ 5040275811Sdelphij arc_change_state(arc_anon, hdr, hash_lock); 5041275811Sdelphij arc_hdr_destroy(hdr); 5042185029Spjd } else { 5043185029Spjd /* 5044185029Spjd * Invalidate issued or about to be issued 5045185029Spjd * reads, since we may be about to write 5046185029Spjd * over this location. 5047185029Spjd */ 5048275811Sdelphij if (HDR_L2_READING(hdr)) { 5049185029Spjd ARCSTAT_BUMP(arcstat_l2_evict_reading); 5050275811Sdelphij hdr->b_flags |= ARC_FLAG_L2_EVICTED; 5051185029Spjd } 5052185029Spjd 5053185029Spjd /* 5054185029Spjd * Tell ARC this no longer exists in L2ARC. 5055185029Spjd */ 5056275811Sdelphij if (hdr->b_l2hdr != NULL) { 5057275811Sdelphij abl2 = hdr->b_l2hdr; 5058251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 5059268085Sdelphij bytes_evicted += abl2->b_asize; 5060275811Sdelphij hdr->b_l2hdr = NULL; 5061274172Savg /* 5062274172Savg * We are destroying l2hdr, so ensure that 5063274172Savg * its compressed buffer, if any, is not leaked. 5064274172Savg */ 5065274172Savg ASSERT(abl2->b_tmp_cdata == NULL); 5066185029Spjd kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 5067275811Sdelphij ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 5068185029Spjd } 5069275811Sdelphij list_remove(buflist, hdr); 5070185029Spjd 5071185029Spjd /* 5072185029Spjd * This may have been leftover after a 5073185029Spjd * failed write. 5074185029Spjd */ 5075275811Sdelphij hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 5076185029Spjd } 5077185029Spjd mutex_exit(hash_lock); 5078185029Spjd } 5079185029Spjd mutex_exit(&l2arc_buflist_mtx); 5080185029Spjd 5081268085Sdelphij vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); 5082185029Spjd dev->l2ad_evict = taddr; 5083185029Spjd} 5084185029Spjd 5085185029Spjd/* 5086185029Spjd * Find and write ARC buffers to the L2ARC device. 5087185029Spjd * 5088275811Sdelphij * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 5089185029Spjd * for reading until they have completed writing. 5090251478Sdelphij * The headroom_boost is an in-out parameter used to maintain headroom boost 5091251478Sdelphij * state between calls to this function. 5092251478Sdelphij * 5093251478Sdelphij * Returns the number of bytes actually written (which may be smaller than 5094251478Sdelphij * the delta by which the device hand has changed due to alignment). 5095185029Spjd */ 5096208373Smmstatic uint64_t 5097251478Sdelphijl2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 5098251478Sdelphij boolean_t *headroom_boost) 5099185029Spjd{ 5100275811Sdelphij arc_buf_hdr_t *hdr, *hdr_prev, *head; 5101185029Spjd list_t *list; 5102251478Sdelphij uint64_t write_asize, write_psize, write_sz, headroom, 5103251478Sdelphij buf_compress_minsz; 5104185029Spjd void *buf_data; 5105251478Sdelphij kmutex_t *list_lock; 5106251478Sdelphij boolean_t full; 5107185029Spjd l2arc_write_callback_t *cb; 5108185029Spjd zio_t *pio, *wzio; 5109228103Smm uint64_t guid = spa_load_guid(spa); 5110251478Sdelphij const boolean_t do_headroom_boost = *headroom_boost; 5111185029Spjd int try; 5112185029Spjd 5113185029Spjd ASSERT(dev->l2ad_vdev != NULL); 5114185029Spjd 5115251478Sdelphij /* Lower the flag now, we might want to raise it again later. */ 5116251478Sdelphij *headroom_boost = B_FALSE; 5117251478Sdelphij 5118185029Spjd pio = NULL; 5119251478Sdelphij write_sz = write_asize = write_psize = 0; 5120185029Spjd full = B_FALSE; 5121185029Spjd head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 5122275811Sdelphij head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; 5123185029Spjd 5124205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 5125185029Spjd /* 5126251478Sdelphij * We will want to try to compress buffers that are at least 2x the 5127251478Sdelphij * device sector size. 5128251478Sdelphij */ 5129251478Sdelphij buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 5130251478Sdelphij 5131251478Sdelphij /* 5132185029Spjd * Copy buffers for L2ARC writing. 5133185029Spjd */ 5134185029Spjd mutex_enter(&l2arc_buflist_mtx); 5135206796Spjd for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 5136251478Sdelphij uint64_t passed_sz = 0; 5137251478Sdelphij 5138185029Spjd list = l2arc_list_locked(try, &list_lock); 5139205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 5140185029Spjd 5141185029Spjd /* 5142185029Spjd * L2ARC fast warmup. 5143185029Spjd * 5144185029Spjd * Until the ARC is warm and starts to evict, read from the 5145185029Spjd * head of the ARC lists rather than the tail. 5146185029Spjd */ 5147185029Spjd if (arc_warm == B_FALSE) 5148275811Sdelphij hdr = list_head(list); 5149185029Spjd else 5150275811Sdelphij hdr = list_tail(list); 5151275811Sdelphij if (hdr == NULL) 5152205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 5153185029Spjd 5154272708Savg headroom = target_sz * l2arc_headroom * 2 / ARC_BUFC_NUMLISTS; 5155251478Sdelphij if (do_headroom_boost) 5156251478Sdelphij headroom = (headroom * l2arc_headroom_boost) / 100; 5157251478Sdelphij 5158275811Sdelphij for (; hdr; hdr = hdr_prev) { 5159251478Sdelphij l2arc_buf_hdr_t *l2hdr; 5160251478Sdelphij kmutex_t *hash_lock; 5161251478Sdelphij uint64_t buf_sz; 5162251478Sdelphij 5163185029Spjd if (arc_warm == B_FALSE) 5164275811Sdelphij hdr_prev = list_next(list, hdr); 5165185029Spjd else 5166275811Sdelphij hdr_prev = list_prev(list, hdr); 5167275811Sdelphij ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size); 5168206796Spjd 5169275811Sdelphij hash_lock = HDR_LOCK(hdr); 5170251478Sdelphij if (!mutex_tryenter(hash_lock)) { 5171205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5172185029Spjd /* 5173185029Spjd * Skip this buffer rather than waiting. 5174185029Spjd */ 5175185029Spjd continue; 5176185029Spjd } 5177185029Spjd 5178275811Sdelphij passed_sz += hdr->b_size; 5179185029Spjd if (passed_sz > headroom) { 5180185029Spjd /* 5181185029Spjd * Searched too far. 5182185029Spjd */ 5183185029Spjd mutex_exit(hash_lock); 5184205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5185185029Spjd break; 5186185029Spjd } 5187185029Spjd 5188275811Sdelphij if (!l2arc_write_eligible(guid, hdr)) { 5189185029Spjd mutex_exit(hash_lock); 5190185029Spjd continue; 5191185029Spjd } 5192185029Spjd 5193275811Sdelphij if ((write_sz + hdr->b_size) > target_sz) { 5194185029Spjd full = B_TRUE; 5195185029Spjd mutex_exit(hash_lock); 5196205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_full); 5197185029Spjd break; 5198185029Spjd } 5199185029Spjd 5200185029Spjd if (pio == NULL) { 5201185029Spjd /* 5202185029Spjd * Insert a dummy header on the buflist so 5203185029Spjd * l2arc_write_done() can find where the 5204185029Spjd * write buffers begin without searching. 5205185029Spjd */ 5206185029Spjd list_insert_head(dev->l2ad_buflist, head); 5207185029Spjd 5208185029Spjd cb = kmem_alloc( 5209185029Spjd sizeof (l2arc_write_callback_t), KM_SLEEP); 5210185029Spjd cb->l2wcb_dev = dev; 5211185029Spjd cb->l2wcb_head = head; 5212185029Spjd pio = zio_root(spa, l2arc_write_done, cb, 5213185029Spjd ZIO_FLAG_CANFAIL); 5214205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_pios); 5215185029Spjd } 5216185029Spjd 5217185029Spjd /* 5218185029Spjd * Create and add a new L2ARC header. 5219185029Spjd */ 5220251478Sdelphij l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5221251478Sdelphij l2hdr->b_dev = dev; 5222275811Sdelphij hdr->b_flags |= ARC_FLAG_L2_WRITING; 5223185029Spjd 5224251478Sdelphij /* 5225251478Sdelphij * Temporarily stash the data buffer in b_tmp_cdata. 5226251478Sdelphij * The subsequent write step will pick it up from 5227275811Sdelphij * there. This is because can't access hdr->b_buf 5228251478Sdelphij * without holding the hash_lock, which we in turn 5229251478Sdelphij * can't access without holding the ARC list locks 5230251478Sdelphij * (which we want to avoid during compression/writing). 5231251478Sdelphij */ 5232251478Sdelphij l2hdr->b_compress = ZIO_COMPRESS_OFF; 5233275811Sdelphij l2hdr->b_asize = hdr->b_size; 5234275811Sdelphij l2hdr->b_tmp_cdata = hdr->b_buf->b_data; 5235251478Sdelphij 5236275811Sdelphij buf_sz = hdr->b_size; 5237275811Sdelphij hdr->b_l2hdr = l2hdr; 5238185029Spjd 5239275811Sdelphij list_insert_head(dev->l2ad_buflist, hdr); 5240251478Sdelphij 5241185029Spjd /* 5242185029Spjd * Compute and store the buffer cksum before 5243185029Spjd * writing. On debug the cksum is verified first. 5244185029Spjd */ 5245275811Sdelphij arc_cksum_verify(hdr->b_buf); 5246275811Sdelphij arc_cksum_compute(hdr->b_buf, B_TRUE); 5247185029Spjd 5248185029Spjd mutex_exit(hash_lock); 5249185029Spjd 5250251478Sdelphij write_sz += buf_sz; 5251251478Sdelphij } 5252251478Sdelphij 5253251478Sdelphij mutex_exit(list_lock); 5254251478Sdelphij 5255251478Sdelphij if (full == B_TRUE) 5256251478Sdelphij break; 5257251478Sdelphij } 5258251478Sdelphij 5259251478Sdelphij /* No buffers selected for writing? */ 5260251478Sdelphij if (pio == NULL) { 5261251478Sdelphij ASSERT0(write_sz); 5262251478Sdelphij mutex_exit(&l2arc_buflist_mtx); 5263251478Sdelphij kmem_cache_free(hdr_cache, head); 5264251478Sdelphij return (0); 5265251478Sdelphij } 5266251478Sdelphij 5267251478Sdelphij /* 5268251478Sdelphij * Now start writing the buffers. We're starting at the write head 5269251478Sdelphij * and work backwards, retracing the course of the buffer selector 5270251478Sdelphij * loop above. 5271251478Sdelphij */ 5272275811Sdelphij for (hdr = list_prev(dev->l2ad_buflist, head); hdr; 5273275811Sdelphij hdr = list_prev(dev->l2ad_buflist, hdr)) { 5274251478Sdelphij l2arc_buf_hdr_t *l2hdr; 5275251478Sdelphij uint64_t buf_sz; 5276251478Sdelphij 5277251478Sdelphij /* 5278251478Sdelphij * We shouldn't need to lock the buffer here, since we flagged 5279275811Sdelphij * it as ARC_FLAG_L2_WRITING in the previous step, but we must 5280275811Sdelphij * take care to only access its L2 cache parameters. In 5281275811Sdelphij * particular, hdr->b_buf may be invalid by now due to 5282275811Sdelphij * ARC eviction. 5283251478Sdelphij */ 5284275811Sdelphij l2hdr = hdr->b_l2hdr; 5285251478Sdelphij l2hdr->b_daddr = dev->l2ad_hand; 5286251478Sdelphij 5287275811Sdelphij if ((hdr->b_flags & ARC_FLAG_L2COMPRESS) && 5288251478Sdelphij l2hdr->b_asize >= buf_compress_minsz) { 5289251478Sdelphij if (l2arc_compress_buf(l2hdr)) { 5290251478Sdelphij /* 5291251478Sdelphij * If compression succeeded, enable headroom 5292251478Sdelphij * boost on the next scan cycle. 5293251478Sdelphij */ 5294251478Sdelphij *headroom_boost = B_TRUE; 5295251478Sdelphij } 5296251478Sdelphij } 5297251478Sdelphij 5298251478Sdelphij /* 5299251478Sdelphij * Pick up the buffer data we had previously stashed away 5300251478Sdelphij * (and now potentially also compressed). 5301251478Sdelphij */ 5302251478Sdelphij buf_data = l2hdr->b_tmp_cdata; 5303251478Sdelphij buf_sz = l2hdr->b_asize; 5304251478Sdelphij 5305274172Savg /* 5306274172Savg * If the data has not been compressed, then clear b_tmp_cdata 5307274172Savg * to make sure that it points only to a temporary compression 5308274172Savg * buffer. 5309274172Savg */ 5310274172Savg if (!L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)) 5311274172Savg l2hdr->b_tmp_cdata = NULL; 5312274172Savg 5313251478Sdelphij /* Compression may have squashed the buffer to zero length. */ 5314251478Sdelphij if (buf_sz != 0) { 5315251478Sdelphij uint64_t buf_p_sz; 5316251478Sdelphij 5317185029Spjd wzio = zio_write_phys(pio, dev->l2ad_vdev, 5318185029Spjd dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5319185029Spjd NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5320185029Spjd ZIO_FLAG_CANFAIL, B_FALSE); 5321185029Spjd 5322185029Spjd DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5323185029Spjd zio_t *, wzio); 5324185029Spjd (void) zio_nowait(wzio); 5325185029Spjd 5326251478Sdelphij write_asize += buf_sz; 5327185029Spjd /* 5328185029Spjd * Keep the clock hand suitably device-aligned. 5329185029Spjd */ 5330251478Sdelphij buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5331251478Sdelphij write_psize += buf_p_sz; 5332251478Sdelphij dev->l2ad_hand += buf_p_sz; 5333185029Spjd } 5334251478Sdelphij } 5335185029Spjd 5336185029Spjd mutex_exit(&l2arc_buflist_mtx); 5337185029Spjd 5338251478Sdelphij ASSERT3U(write_asize, <=, target_sz); 5339185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_sent); 5340251478Sdelphij ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5341185029Spjd ARCSTAT_INCR(arcstat_l2_size, write_sz); 5342251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, write_asize); 5343275096Sdelphij vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 5344185029Spjd 5345185029Spjd /* 5346185029Spjd * Bump device hand to the device start if it is approaching the end. 5347185029Spjd * l2arc_evict() will already have evicted ahead for this case. 5348185029Spjd */ 5349185029Spjd if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5350185029Spjd dev->l2ad_hand = dev->l2ad_start; 5351185029Spjd dev->l2ad_evict = dev->l2ad_start; 5352185029Spjd dev->l2ad_first = B_FALSE; 5353185029Spjd } 5354185029Spjd 5355208373Smm dev->l2ad_writing = B_TRUE; 5356185029Spjd (void) zio_wait(pio); 5357208373Smm dev->l2ad_writing = B_FALSE; 5358208373Smm 5359251478Sdelphij return (write_asize); 5360185029Spjd} 5361185029Spjd 5362185029Spjd/* 5363251478Sdelphij * Compresses an L2ARC buffer. 5364251478Sdelphij * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5365251478Sdelphij * size in l2hdr->b_asize. This routine tries to compress the data and 5366251478Sdelphij * depending on the compression result there are three possible outcomes: 5367251478Sdelphij * *) The buffer was incompressible. The original l2hdr contents were left 5368251478Sdelphij * untouched and are ready for writing to an L2 device. 5369251478Sdelphij * *) The buffer was all-zeros, so there is no need to write it to an L2 5370251478Sdelphij * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5371251478Sdelphij * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5372251478Sdelphij * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5373251478Sdelphij * data buffer which holds the compressed data to be written, and b_asize 5374251478Sdelphij * tells us how much data there is. b_compress is set to the appropriate 5375251478Sdelphij * compression algorithm. Once writing is done, invoke 5376251478Sdelphij * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5377251478Sdelphij * 5378251478Sdelphij * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5379251478Sdelphij * buffer was incompressible). 5380251478Sdelphij */ 5381251478Sdelphijstatic boolean_t 5382251478Sdelphijl2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5383251478Sdelphij{ 5384251478Sdelphij void *cdata; 5385268075Sdelphij size_t csize, len, rounded; 5386251478Sdelphij 5387251478Sdelphij ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5388251478Sdelphij ASSERT(l2hdr->b_tmp_cdata != NULL); 5389251478Sdelphij 5390251478Sdelphij len = l2hdr->b_asize; 5391251478Sdelphij cdata = zio_data_buf_alloc(len); 5392251478Sdelphij csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5393269086Sdelphij cdata, l2hdr->b_asize); 5394251478Sdelphij 5395251478Sdelphij if (csize == 0) { 5396251478Sdelphij /* zero block, indicate that there's nothing to write */ 5397251478Sdelphij zio_data_buf_free(cdata, len); 5398251478Sdelphij l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5399251478Sdelphij l2hdr->b_asize = 0; 5400251478Sdelphij l2hdr->b_tmp_cdata = NULL; 5401251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5402251478Sdelphij return (B_TRUE); 5403274628Savg } 5404274628Savg 5405274628Savg rounded = P2ROUNDUP(csize, 5406274628Savg (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift); 5407274628Savg if (rounded < len) { 5408251478Sdelphij /* 5409251478Sdelphij * Compression succeeded, we'll keep the cdata around for 5410251478Sdelphij * writing and release it afterwards. 5411251478Sdelphij */ 5412274628Savg if (rounded > csize) { 5413274628Savg bzero((char *)cdata + csize, rounded - csize); 5414274628Savg csize = rounded; 5415274628Savg } 5416251478Sdelphij l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5417251478Sdelphij l2hdr->b_asize = csize; 5418251478Sdelphij l2hdr->b_tmp_cdata = cdata; 5419251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_successes); 5420251478Sdelphij return (B_TRUE); 5421251478Sdelphij } else { 5422251478Sdelphij /* 5423251478Sdelphij * Compression failed, release the compressed buffer. 5424251478Sdelphij * l2hdr will be left unmodified. 5425251478Sdelphij */ 5426251478Sdelphij zio_data_buf_free(cdata, len); 5427251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_failures); 5428251478Sdelphij return (B_FALSE); 5429251478Sdelphij } 5430251478Sdelphij} 5431251478Sdelphij 5432251478Sdelphij/* 5433251478Sdelphij * Decompresses a zio read back from an l2arc device. On success, the 5434251478Sdelphij * underlying zio's io_data buffer is overwritten by the uncompressed 5435251478Sdelphij * version. On decompression error (corrupt compressed stream), the 5436251478Sdelphij * zio->io_error value is set to signal an I/O error. 5437251478Sdelphij * 5438251478Sdelphij * Please note that the compressed data stream is not checksummed, so 5439251478Sdelphij * if the underlying device is experiencing data corruption, we may feed 5440251478Sdelphij * corrupt data to the decompressor, so the decompressor needs to be 5441251478Sdelphij * able to handle this situation (LZ4 does). 5442251478Sdelphij */ 5443251478Sdelphijstatic void 5444251478Sdelphijl2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5445251478Sdelphij{ 5446251478Sdelphij ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5447251478Sdelphij 5448251478Sdelphij if (zio->io_error != 0) { 5449251478Sdelphij /* 5450251478Sdelphij * An io error has occured, just restore the original io 5451251478Sdelphij * size in preparation for a main pool read. 5452251478Sdelphij */ 5453251478Sdelphij zio->io_orig_size = zio->io_size = hdr->b_size; 5454251478Sdelphij return; 5455251478Sdelphij } 5456251478Sdelphij 5457251478Sdelphij if (c == ZIO_COMPRESS_EMPTY) { 5458251478Sdelphij /* 5459251478Sdelphij * An empty buffer results in a null zio, which means we 5460251478Sdelphij * need to fill its io_data after we're done restoring the 5461251478Sdelphij * buffer's contents. 5462251478Sdelphij */ 5463251478Sdelphij ASSERT(hdr->b_buf != NULL); 5464251478Sdelphij bzero(hdr->b_buf->b_data, hdr->b_size); 5465251478Sdelphij zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5466251478Sdelphij } else { 5467251478Sdelphij ASSERT(zio->io_data != NULL); 5468251478Sdelphij /* 5469251478Sdelphij * We copy the compressed data from the start of the arc buffer 5470251478Sdelphij * (the zio_read will have pulled in only what we need, the 5471251478Sdelphij * rest is garbage which we will overwrite at decompression) 5472251478Sdelphij * and then decompress back to the ARC data buffer. This way we 5473251478Sdelphij * can minimize copying by simply decompressing back over the 5474251478Sdelphij * original compressed data (rather than decompressing to an 5475251478Sdelphij * aux buffer and then copying back the uncompressed buffer, 5476251478Sdelphij * which is likely to be much larger). 5477251478Sdelphij */ 5478251478Sdelphij uint64_t csize; 5479251478Sdelphij void *cdata; 5480251478Sdelphij 5481251478Sdelphij csize = zio->io_size; 5482251478Sdelphij cdata = zio_data_buf_alloc(csize); 5483251478Sdelphij bcopy(zio->io_data, cdata, csize); 5484251478Sdelphij if (zio_decompress_data(c, cdata, zio->io_data, csize, 5485251478Sdelphij hdr->b_size) != 0) 5486251478Sdelphij zio->io_error = EIO; 5487251478Sdelphij zio_data_buf_free(cdata, csize); 5488251478Sdelphij } 5489251478Sdelphij 5490251478Sdelphij /* Restore the expected uncompressed IO size. */ 5491251478Sdelphij zio->io_orig_size = zio->io_size = hdr->b_size; 5492251478Sdelphij} 5493251478Sdelphij 5494251478Sdelphij/* 5495251478Sdelphij * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5496251478Sdelphij * This buffer serves as a temporary holder of compressed data while 5497251478Sdelphij * the buffer entry is being written to an l2arc device. Once that is 5498251478Sdelphij * done, we can dispose of it. 5499251478Sdelphij */ 5500251478Sdelphijstatic void 5501275811Sdelphijl2arc_release_cdata_buf(arc_buf_hdr_t *hdr) 5502251478Sdelphij{ 5503275811Sdelphij l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 5504251478Sdelphij 5505274172Savg ASSERT(L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)); 5506274172Savg if (l2hdr->b_compress != ZIO_COMPRESS_EMPTY) { 5507251478Sdelphij /* 5508251478Sdelphij * If the data was compressed, then we've allocated a 5509251478Sdelphij * temporary buffer for it, so now we need to release it. 5510251478Sdelphij */ 5511251478Sdelphij ASSERT(l2hdr->b_tmp_cdata != NULL); 5512275811Sdelphij zio_data_buf_free(l2hdr->b_tmp_cdata, hdr->b_size); 5513274172Savg l2hdr->b_tmp_cdata = NULL; 5514274172Savg } else { 5515274172Savg ASSERT(l2hdr->b_tmp_cdata == NULL); 5516251478Sdelphij } 5517251478Sdelphij} 5518251478Sdelphij 5519251478Sdelphij/* 5520185029Spjd * This thread feeds the L2ARC at regular intervals. This is the beating 5521185029Spjd * heart of the L2ARC. 5522185029Spjd */ 5523185029Spjdstatic void 5524185029Spjdl2arc_feed_thread(void *dummy __unused) 5525185029Spjd{ 5526185029Spjd callb_cpr_t cpr; 5527185029Spjd l2arc_dev_t *dev; 5528185029Spjd spa_t *spa; 5529208373Smm uint64_t size, wrote; 5530219089Spjd clock_t begin, next = ddi_get_lbolt(); 5531251478Sdelphij boolean_t headroom_boost = B_FALSE; 5532185029Spjd 5533185029Spjd CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5534185029Spjd 5535185029Spjd mutex_enter(&l2arc_feed_thr_lock); 5536185029Spjd 5537185029Spjd while (l2arc_thread_exit == 0) { 5538185029Spjd CALLB_CPR_SAFE_BEGIN(&cpr); 5539185029Spjd (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5540219089Spjd next - ddi_get_lbolt()); 5541185029Spjd CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5542219089Spjd next = ddi_get_lbolt() + hz; 5543185029Spjd 5544185029Spjd /* 5545185029Spjd * Quick check for L2ARC devices. 5546185029Spjd */ 5547185029Spjd mutex_enter(&l2arc_dev_mtx); 5548185029Spjd if (l2arc_ndev == 0) { 5549185029Spjd mutex_exit(&l2arc_dev_mtx); 5550185029Spjd continue; 5551185029Spjd } 5552185029Spjd mutex_exit(&l2arc_dev_mtx); 5553219089Spjd begin = ddi_get_lbolt(); 5554185029Spjd 5555185029Spjd /* 5556185029Spjd * This selects the next l2arc device to write to, and in 5557185029Spjd * doing so the next spa to feed from: dev->l2ad_spa. This 5558185029Spjd * will return NULL if there are now no l2arc devices or if 5559185029Spjd * they are all faulted. 5560185029Spjd * 5561185029Spjd * If a device is returned, its spa's config lock is also 5562185029Spjd * held to prevent device removal. l2arc_dev_get_next() 5563185029Spjd * will grab and release l2arc_dev_mtx. 5564185029Spjd */ 5565185029Spjd if ((dev = l2arc_dev_get_next()) == NULL) 5566185029Spjd continue; 5567185029Spjd 5568185029Spjd spa = dev->l2ad_spa; 5569185029Spjd ASSERT(spa != NULL); 5570185029Spjd 5571185029Spjd /* 5572219089Spjd * If the pool is read-only then force the feed thread to 5573219089Spjd * sleep a little longer. 5574219089Spjd */ 5575219089Spjd if (!spa_writeable(spa)) { 5576219089Spjd next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5577219089Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5578219089Spjd continue; 5579219089Spjd } 5580219089Spjd 5581219089Spjd /* 5582185029Spjd * Avoid contributing to memory pressure. 5583185029Spjd */ 5584185029Spjd if (arc_reclaim_needed()) { 5585185029Spjd ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5586185029Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5587185029Spjd continue; 5588185029Spjd } 5589185029Spjd 5590185029Spjd ARCSTAT_BUMP(arcstat_l2_feeds); 5591185029Spjd 5592251478Sdelphij size = l2arc_write_size(); 5593185029Spjd 5594185029Spjd /* 5595185029Spjd * Evict L2ARC buffers that will be overwritten. 5596185029Spjd */ 5597185029Spjd l2arc_evict(dev, size, B_FALSE); 5598185029Spjd 5599185029Spjd /* 5600185029Spjd * Write ARC buffers. 5601185029Spjd */ 5602251478Sdelphij wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5603208373Smm 5604208373Smm /* 5605208373Smm * Calculate interval between writes. 5606208373Smm */ 5607208373Smm next = l2arc_write_interval(begin, size, wrote); 5608185029Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5609185029Spjd } 5610185029Spjd 5611185029Spjd l2arc_thread_exit = 0; 5612185029Spjd cv_broadcast(&l2arc_feed_thr_cv); 5613185029Spjd CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5614185029Spjd thread_exit(); 5615185029Spjd} 5616185029Spjd 5617185029Spjdboolean_t 5618185029Spjdl2arc_vdev_present(vdev_t *vd) 5619185029Spjd{ 5620185029Spjd l2arc_dev_t *dev; 5621185029Spjd 5622185029Spjd mutex_enter(&l2arc_dev_mtx); 5623185029Spjd for (dev = list_head(l2arc_dev_list); dev != NULL; 5624185029Spjd dev = list_next(l2arc_dev_list, dev)) { 5625185029Spjd if (dev->l2ad_vdev == vd) 5626185029Spjd break; 5627185029Spjd } 5628185029Spjd mutex_exit(&l2arc_dev_mtx); 5629185029Spjd 5630185029Spjd return (dev != NULL); 5631185029Spjd} 5632185029Spjd 5633185029Spjd/* 5634185029Spjd * Add a vdev for use by the L2ARC. By this point the spa has already 5635185029Spjd * validated the vdev and opened it. 5636185029Spjd */ 5637185029Spjdvoid 5638219089Spjdl2arc_add_vdev(spa_t *spa, vdev_t *vd) 5639185029Spjd{ 5640185029Spjd l2arc_dev_t *adddev; 5641185029Spjd 5642185029Spjd ASSERT(!l2arc_vdev_present(vd)); 5643185029Spjd 5644255753Sgibbs vdev_ashift_optimize(vd); 5645255753Sgibbs 5646185029Spjd /* 5647185029Spjd * Create a new l2arc device entry. 5648185029Spjd */ 5649185029Spjd adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5650185029Spjd adddev->l2ad_spa = spa; 5651185029Spjd adddev->l2ad_vdev = vd; 5652219089Spjd adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5653219089Spjd adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5654185029Spjd adddev->l2ad_hand = adddev->l2ad_start; 5655185029Spjd adddev->l2ad_evict = adddev->l2ad_start; 5656185029Spjd adddev->l2ad_first = B_TRUE; 5657208373Smm adddev->l2ad_writing = B_FALSE; 5658185029Spjd 5659185029Spjd /* 5660185029Spjd * This is a list of all ARC buffers that are still valid on the 5661185029Spjd * device. 5662185029Spjd */ 5663185029Spjd adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5664185029Spjd list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5665185029Spjd offsetof(arc_buf_hdr_t, b_l2node)); 5666185029Spjd 5667219089Spjd vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5668185029Spjd 5669185029Spjd /* 5670185029Spjd * Add device to global list 5671185029Spjd */ 5672185029Spjd mutex_enter(&l2arc_dev_mtx); 5673185029Spjd list_insert_head(l2arc_dev_list, adddev); 5674185029Spjd atomic_inc_64(&l2arc_ndev); 5675185029Spjd mutex_exit(&l2arc_dev_mtx); 5676185029Spjd} 5677185029Spjd 5678185029Spjd/* 5679185029Spjd * Remove a vdev from the L2ARC. 5680185029Spjd */ 5681185029Spjdvoid 5682185029Spjdl2arc_remove_vdev(vdev_t *vd) 5683185029Spjd{ 5684185029Spjd l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5685185029Spjd 5686185029Spjd /* 5687185029Spjd * Find the device by vdev 5688185029Spjd */ 5689185029Spjd mutex_enter(&l2arc_dev_mtx); 5690185029Spjd for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5691185029Spjd nextdev = list_next(l2arc_dev_list, dev); 5692185029Spjd if (vd == dev->l2ad_vdev) { 5693185029Spjd remdev = dev; 5694185029Spjd break; 5695185029Spjd } 5696185029Spjd } 5697185029Spjd ASSERT(remdev != NULL); 5698185029Spjd 5699185029Spjd /* 5700185029Spjd * Remove device from global list 5701185029Spjd */ 5702185029Spjd list_remove(l2arc_dev_list, remdev); 5703185029Spjd l2arc_dev_last = NULL; /* may have been invalidated */ 5704185029Spjd atomic_dec_64(&l2arc_ndev); 5705185029Spjd mutex_exit(&l2arc_dev_mtx); 5706185029Spjd 5707185029Spjd /* 5708185029Spjd * Clear all buflists and ARC references. L2ARC device flush. 5709185029Spjd */ 5710185029Spjd l2arc_evict(remdev, 0, B_TRUE); 5711185029Spjd list_destroy(remdev->l2ad_buflist); 5712185029Spjd kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5713185029Spjd kmem_free(remdev, sizeof (l2arc_dev_t)); 5714185029Spjd} 5715185029Spjd 5716185029Spjdvoid 5717185029Spjdl2arc_init(void) 5718185029Spjd{ 5719185029Spjd l2arc_thread_exit = 0; 5720185029Spjd l2arc_ndev = 0; 5721185029Spjd l2arc_writes_sent = 0; 5722185029Spjd l2arc_writes_done = 0; 5723185029Spjd 5724185029Spjd mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5725185029Spjd cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5726185029Spjd mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5727185029Spjd mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5728185029Spjd mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5729185029Spjd 5730185029Spjd l2arc_dev_list = &L2ARC_dev_list; 5731185029Spjd l2arc_free_on_write = &L2ARC_free_on_write; 5732185029Spjd list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5733185029Spjd offsetof(l2arc_dev_t, l2ad_node)); 5734185029Spjd list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5735185029Spjd offsetof(l2arc_data_free_t, l2df_list_node)); 5736185029Spjd} 5737185029Spjd 5738185029Spjdvoid 5739185029Spjdl2arc_fini(void) 5740185029Spjd{ 5741185029Spjd /* 5742185029Spjd * This is called from dmu_fini(), which is called from spa_fini(); 5743185029Spjd * Because of this, we can assume that all l2arc devices have 5744185029Spjd * already been removed when the pools themselves were removed. 5745185029Spjd */ 5746185029Spjd 5747185029Spjd l2arc_do_free_on_write(); 5748185029Spjd 5749185029Spjd mutex_destroy(&l2arc_feed_thr_lock); 5750185029Spjd cv_destroy(&l2arc_feed_thr_cv); 5751185029Spjd mutex_destroy(&l2arc_dev_mtx); 5752185029Spjd mutex_destroy(&l2arc_buflist_mtx); 5753185029Spjd mutex_destroy(&l2arc_free_on_write_mtx); 5754185029Spjd 5755185029Spjd list_destroy(l2arc_dev_list); 5756185029Spjd list_destroy(l2arc_free_on_write); 5757185029Spjd} 5758185029Spjd 5759185029Spjdvoid 5760185029Spjdl2arc_start(void) 5761185029Spjd{ 5762209962Smm if (!(spa_mode_global & FWRITE)) 5763185029Spjd return; 5764185029Spjd 5765185029Spjd (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5766185029Spjd TS_RUN, minclsyspri); 5767185029Spjd} 5768185029Spjd 5769185029Spjdvoid 5770185029Spjdl2arc_stop(void) 5771185029Spjd{ 5772209962Smm if (!(spa_mode_global & FWRITE)) 5773185029Spjd return; 5774185029Spjd 5775185029Spjd mutex_enter(&l2arc_feed_thr_lock); 5776185029Spjd cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5777185029Spjd l2arc_thread_exit = 1; 5778185029Spjd while (l2arc_thread_exit != 0) 5779185029Spjd cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5780185029Spjd mutex_exit(&l2arc_feed_thr_lock); 5781185029Spjd} 5782