arc.c revision 286570
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. 23277826Sdelphij * Copyright (c) 2012, Joyent, Inc. All rights reserved. 24268123Sdelphij * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 25260835Sdelphij * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 26268085Sdelphij * Copyright 2014 Nexenta Systems, Inc. All rights reserved. 27168404Spjd */ 28168404Spjd 29168404Spjd/* 30168404Spjd * DVA-based Adjustable Replacement Cache 31168404Spjd * 32168404Spjd * While much of the theory of operation used here is 33168404Spjd * based on the self-tuning, low overhead replacement cache 34168404Spjd * presented by Megiddo and Modha at FAST 2003, there are some 35168404Spjd * significant differences: 36168404Spjd * 37168404Spjd * 1. The Megiddo and Modha model assumes any page is evictable. 38168404Spjd * Pages in its cache cannot be "locked" into memory. This makes 39168404Spjd * the eviction algorithm simple: evict the last page in the list. 40168404Spjd * This also make the performance characteristics easy to reason 41168404Spjd * about. Our cache is not so simple. At any given moment, some 42168404Spjd * subset of the blocks in the cache are un-evictable because we 43168404Spjd * have handed out a reference to them. Blocks are only evictable 44168404Spjd * when there are no external references active. This makes 45168404Spjd * eviction far more problematic: we choose to evict the evictable 46168404Spjd * blocks that are the "lowest" in the list. 47168404Spjd * 48168404Spjd * There are times when it is not possible to evict the requested 49168404Spjd * space. In these circumstances we are unable to adjust the cache 50168404Spjd * size. To prevent the cache growing unbounded at these times we 51185029Spjd * implement a "cache throttle" that slows the flow of new data 52185029Spjd * into the cache until we can make space available. 53168404Spjd * 54168404Spjd * 2. The Megiddo and Modha model assumes a fixed cache size. 55168404Spjd * Pages are evicted when the cache is full and there is a cache 56168404Spjd * miss. Our model has a variable sized cache. It grows with 57185029Spjd * high use, but also tries to react to memory pressure from the 58168404Spjd * operating system: decreasing its size when system memory is 59168404Spjd * tight. 60168404Spjd * 61168404Spjd * 3. The Megiddo and Modha model assumes a fixed page size. All 62251631Sdelphij * elements of the cache are therefore exactly the same size. So 63168404Spjd * when adjusting the cache size following a cache miss, its simply 64168404Spjd * a matter of choosing a single page to evict. In our model, we 65168404Spjd * have variable sized cache blocks (rangeing from 512 bytes to 66251631Sdelphij * 128K bytes). We therefore choose a set of blocks to evict to make 67168404Spjd * space for a cache miss that approximates as closely as possible 68168404Spjd * the space used by the new block. 69168404Spjd * 70168404Spjd * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 71168404Spjd * by N. Megiddo & D. Modha, FAST 2003 72168404Spjd */ 73168404Spjd 74168404Spjd/* 75168404Spjd * The locking model: 76168404Spjd * 77168404Spjd * A new reference to a cache buffer can be obtained in two 78168404Spjd * ways: 1) via a hash table lookup using the DVA as a key, 79185029Spjd * or 2) via one of the ARC lists. The arc_read() interface 80168404Spjd * uses method 1, while the internal arc algorithms for 81251631Sdelphij * adjusting the cache use method 2. We therefore provide two 82168404Spjd * types of locks: 1) the hash table lock array, and 2) the 83168404Spjd * arc list locks. 84168404Spjd * 85168404Spjd * Buffers do not have their own mutexs, rather they rely on the 86168404Spjd * hash table mutexs for the bulk of their protection (i.e. most 87168404Spjd * fields in the arc_buf_hdr_t are protected by these mutexs). 88168404Spjd * 89168404Spjd * buf_hash_find() returns the appropriate mutex (held) when it 90168404Spjd * locates the requested buffer in the hash table. It returns 91168404Spjd * NULL for the mutex if the buffer was not in the table. 92168404Spjd * 93168404Spjd * buf_hash_remove() expects the appropriate hash mutex to be 94168404Spjd * already held before it is invoked. 95168404Spjd * 96168404Spjd * Each arc state also has a mutex which is used to protect the 97168404Spjd * buffer list associated with the state. When attempting to 98168404Spjd * obtain a hash table lock while holding an arc list lock you 99168404Spjd * must use: mutex_tryenter() to avoid deadlock. Also note that 100168404Spjd * the active state mutex must be held before the ghost state mutex. 101168404Spjd * 102168404Spjd * Arc buffers may have an associated eviction callback function. 103168404Spjd * This function will be invoked prior to removing the buffer (e.g. 104168404Spjd * in arc_do_user_evicts()). Note however that the data associated 105168404Spjd * with the buffer may be evicted prior to the callback. The callback 106168404Spjd * must be made with *no locks held* (to prevent deadlock). Additionally, 107168404Spjd * the users of callbacks must ensure that their private data is 108268858Sdelphij * protected from simultaneous callbacks from arc_clear_callback() 109168404Spjd * and arc_do_user_evicts(). 110168404Spjd * 111168404Spjd * Note that the majority of the performance stats are manipulated 112168404Spjd * with atomic operations. 113185029Spjd * 114286570Smav * The L2ARC uses the l2ad_mtx on each vdev for the following: 115185029Spjd * 116185029Spjd * - L2ARC buflist creation 117185029Spjd * - L2ARC buflist eviction 118185029Spjd * - L2ARC write completion, which walks L2ARC buflists 119185029Spjd * - ARC header destruction, as it removes from L2ARC buflists 120185029Spjd * - ARC header release, as it removes from L2ARC buflists 121168404Spjd */ 122168404Spjd 123168404Spjd#include <sys/spa.h> 124168404Spjd#include <sys/zio.h> 125251478Sdelphij#include <sys/zio_compress.h> 126168404Spjd#include <sys/zfs_context.h> 127168404Spjd#include <sys/arc.h> 128168404Spjd#include <sys/refcount.h> 129185029Spjd#include <sys/vdev.h> 130219089Spjd#include <sys/vdev_impl.h> 131258632Savg#include <sys/dsl_pool.h> 132168404Spjd#ifdef _KERNEL 133168404Spjd#include <sys/dnlc.h> 134168404Spjd#endif 135168404Spjd#include <sys/callb.h> 136168404Spjd#include <sys/kstat.h> 137248572Ssmh#include <sys/trim_map.h> 138219089Spjd#include <zfs_fletcher.h> 139168404Spjd#include <sys/sdt.h> 140168404Spjd 141191902Skmacy#include <vm/vm_pageout.h> 142272483Ssmh#include <machine/vmparam.h> 143191902Skmacy 144240133Smm#ifdef illumos 145240133Smm#ifndef _KERNEL 146240133Smm/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 147240133Smmboolean_t arc_watch = B_FALSE; 148240133Smmint arc_procfd; 149240133Smm#endif 150240133Smm#endif /* illumos */ 151240133Smm 152168404Spjdstatic kmutex_t arc_reclaim_thr_lock; 153168404Spjdstatic kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 154168404Spjdstatic uint8_t arc_thread_exit; 155168404Spjd 156168404Spjd#define ARC_REDUCE_DNLC_PERCENT 3 157168404Spjduint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 158168404Spjd 159168404Spjdtypedef enum arc_reclaim_strategy { 160168404Spjd ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 161168404Spjd ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 162168404Spjd} arc_reclaim_strategy_t; 163168404Spjd 164258632Savg/* 165258632Savg * The number of iterations through arc_evict_*() before we 166258632Savg * drop & reacquire the lock. 167258632Savg */ 168258632Savgint arc_evict_iterations = 100; 169258632Savg 170168404Spjd/* number of seconds before growing cache again */ 171168404Spjdstatic int arc_grow_retry = 60; 172168404Spjd 173208373Smm/* shift of arc_c for calculating both min and max arc_p */ 174208373Smmstatic int arc_p_min_shift = 4; 175208373Smm 176208373Smm/* log2(fraction of arc to reclaim) */ 177208373Smmstatic int arc_shrink_shift = 5; 178208373Smm 179168404Spjd/* 180168404Spjd * minimum lifespan of a prefetch block in clock ticks 181168404Spjd * (initialized in arc_init()) 182168404Spjd */ 183168404Spjdstatic int arc_min_prefetch_lifespan; 184168404Spjd 185258632Savg/* 186258632Savg * If this percent of memory is free, don't throttle. 187258632Savg */ 188258632Savgint arc_lotsfree_percent = 10; 189258632Savg 190208373Smmstatic int arc_dead; 191194043Skmacyextern int zfs_prefetch_disable; 192168404Spjd 193168404Spjd/* 194185029Spjd * The arc has filled available memory and has now warmed up. 195185029Spjd */ 196185029Spjdstatic boolean_t arc_warm; 197185029Spjd 198185029Spjduint64_t zfs_arc_max; 199185029Spjduint64_t zfs_arc_min; 200185029Spjduint64_t zfs_arc_meta_limit = 0; 201275780Sdelphijuint64_t zfs_arc_meta_min = 0; 202208373Smmint zfs_arc_grow_retry = 0; 203208373Smmint zfs_arc_shrink_shift = 0; 204208373Smmint zfs_arc_p_min_shift = 0; 205242845Sdelphijint zfs_disable_dup_eviction = 0; 206269230Sdelphijuint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 207272483Ssmhu_int zfs_arc_free_target = 0; 208185029Spjd 209270759Ssmhstatic int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 210275748Sdelphijstatic int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 211270759Ssmh 212270759Ssmh#ifdef _KERNEL 213270759Ssmhstatic void 214270759Ssmharc_free_target_init(void *unused __unused) 215270759Ssmh{ 216270759Ssmh 217272483Ssmh zfs_arc_free_target = vm_pageout_wakeup_thresh; 218270759Ssmh} 219270759SsmhSYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 220270759Ssmh arc_free_target_init, NULL); 221270759Ssmh 222185029SpjdTUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 223275780SdelphijTUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 224273026SdelphijTUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 225168473SpjdSYSCTL_DECL(_vfs_zfs); 226217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 227168473Spjd "Maximum ARC size"); 228217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 229168473Spjd "Minimum ARC size"); 230269230SdelphijSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 231269230Sdelphij &zfs_arc_average_blocksize, 0, 232269230Sdelphij "ARC average blocksize"); 233273026SdelphijSYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 234273026Sdelphij &arc_shrink_shift, 0, 235273026Sdelphij "log2(fraction of arc to reclaim)"); 236273026Sdelphij 237270759Ssmh/* 238270759Ssmh * We don't have a tunable for arc_free_target due to the dependency on 239270759Ssmh * pagedaemon initialisation. 240270759Ssmh */ 241270759SsmhSYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 242270759Ssmh CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 243270759Ssmh sysctl_vfs_zfs_arc_free_target, "IU", 244270759Ssmh "Desired number of free pages below which ARC triggers reclaim"); 245168404Spjd 246270759Ssmhstatic int 247270759Ssmhsysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 248270759Ssmh{ 249270759Ssmh u_int val; 250270759Ssmh int err; 251270759Ssmh 252270759Ssmh val = zfs_arc_free_target; 253270759Ssmh err = sysctl_handle_int(oidp, &val, 0, req); 254270759Ssmh if (err != 0 || req->newptr == NULL) 255270759Ssmh return (err); 256270759Ssmh 257272483Ssmh if (val < minfree) 258270759Ssmh return (EINVAL); 259272483Ssmh if (val > vm_cnt.v_page_count) 260270759Ssmh return (EINVAL); 261270759Ssmh 262270759Ssmh zfs_arc_free_target = val; 263270759Ssmh 264270759Ssmh return (0); 265270759Ssmh} 266275748Sdelphij 267275748Sdelphij/* 268275748Sdelphij * Must be declared here, before the definition of corresponding kstat 269275748Sdelphij * macro which uses the same names will confuse the compiler. 270275748Sdelphij */ 271275748SdelphijSYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 272275748Sdelphij CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 273275748Sdelphij sysctl_vfs_zfs_arc_meta_limit, "QU", 274275748Sdelphij "ARC metadata limit"); 275272483Ssmh#endif 276270759Ssmh 277168404Spjd/* 278185029Spjd * Note that buffers can be in one of 6 states: 279168404Spjd * ARC_anon - anonymous (discussed below) 280168404Spjd * ARC_mru - recently used, currently cached 281168404Spjd * ARC_mru_ghost - recentely used, no longer in cache 282168404Spjd * ARC_mfu - frequently used, currently cached 283168404Spjd * ARC_mfu_ghost - frequently used, no longer in cache 284185029Spjd * ARC_l2c_only - exists in L2ARC but not other states 285185029Spjd * When there are no active references to the buffer, they are 286185029Spjd * are linked onto a list in one of these arc states. These are 287185029Spjd * the only buffers that can be evicted or deleted. Within each 288185029Spjd * state there are multiple lists, one for meta-data and one for 289185029Spjd * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 290185029Spjd * etc.) is tracked separately so that it can be managed more 291185029Spjd * explicitly: favored over data, limited explicitly. 292168404Spjd * 293168404Spjd * Anonymous buffers are buffers that are not associated with 294168404Spjd * a DVA. These are buffers that hold dirty block copies 295168404Spjd * before they are written to stable storage. By definition, 296168404Spjd * they are "ref'd" and are considered part of arc_mru 297168404Spjd * that cannot be freed. Generally, they will aquire a DVA 298168404Spjd * as they are written and migrate onto the arc_mru list. 299185029Spjd * 300185029Spjd * The ARC_l2c_only state is for buffers that are in the second 301185029Spjd * level ARC but no longer in any of the ARC_m* lists. The second 302185029Spjd * level ARC itself may also contain buffers that are in any of 303185029Spjd * the ARC_m* states - meaning that a buffer can exist in two 304185029Spjd * places. The reason for the ARC_l2c_only state is to keep the 305185029Spjd * buffer header in the hash table, so that reads that hit the 306185029Spjd * second level ARC benefit from these fast lookups. 307168404Spjd */ 308168404Spjd 309205264Skmacy#define ARCS_LOCK_PAD CACHE_LINE_SIZE 310205231Skmacystruct arcs_lock { 311205231Skmacy kmutex_t arcs_lock; 312205231Skmacy#ifdef _KERNEL 313205231Skmacy unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 314205231Skmacy#endif 315205231Skmacy}; 316205231Skmacy 317205231Skmacy/* 318205231Skmacy * must be power of two for mask use to work 319205231Skmacy * 320205231Skmacy */ 321205231Skmacy#define ARC_BUFC_NUMDATALISTS 16 322205231Skmacy#define ARC_BUFC_NUMMETADATALISTS 16 323206796Spjd#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 324205231Skmacy 325168404Spjdtypedef struct arc_state { 326185029Spjd uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 327185029Spjd uint64_t arcs_size; /* total amount of data in this state */ 328205231Skmacy list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 329205264Skmacy struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 330168404Spjd} arc_state_t; 331168404Spjd 332206796Spjd#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 333205231Skmacy 334185029Spjd/* The 6 states: */ 335168404Spjdstatic arc_state_t ARC_anon; 336168404Spjdstatic arc_state_t ARC_mru; 337168404Spjdstatic arc_state_t ARC_mru_ghost; 338168404Spjdstatic arc_state_t ARC_mfu; 339168404Spjdstatic arc_state_t ARC_mfu_ghost; 340185029Spjdstatic arc_state_t ARC_l2c_only; 341168404Spjd 342168404Spjdtypedef struct arc_stats { 343168404Spjd kstat_named_t arcstat_hits; 344168404Spjd kstat_named_t arcstat_misses; 345168404Spjd kstat_named_t arcstat_demand_data_hits; 346168404Spjd kstat_named_t arcstat_demand_data_misses; 347168404Spjd kstat_named_t arcstat_demand_metadata_hits; 348168404Spjd kstat_named_t arcstat_demand_metadata_misses; 349168404Spjd kstat_named_t arcstat_prefetch_data_hits; 350168404Spjd kstat_named_t arcstat_prefetch_data_misses; 351168404Spjd kstat_named_t arcstat_prefetch_metadata_hits; 352168404Spjd kstat_named_t arcstat_prefetch_metadata_misses; 353168404Spjd kstat_named_t arcstat_mru_hits; 354168404Spjd kstat_named_t arcstat_mru_ghost_hits; 355168404Spjd kstat_named_t arcstat_mfu_hits; 356168404Spjd kstat_named_t arcstat_mfu_ghost_hits; 357205231Skmacy kstat_named_t arcstat_allocated; 358168404Spjd kstat_named_t arcstat_deleted; 359205231Skmacy kstat_named_t arcstat_stolen; 360168404Spjd kstat_named_t arcstat_recycle_miss; 361251629Sdelphij /* 362251629Sdelphij * Number of buffers that could not be evicted because the hash lock 363251629Sdelphij * was held by another thread. The lock may not necessarily be held 364251629Sdelphij * by something using the same buffer, since hash locks are shared 365251629Sdelphij * by multiple buffers. 366251629Sdelphij */ 367168404Spjd kstat_named_t arcstat_mutex_miss; 368251629Sdelphij /* 369251629Sdelphij * Number of buffers skipped because they have I/O in progress, are 370251629Sdelphij * indrect prefetch buffers that have not lived long enough, or are 371251629Sdelphij * not from the spa we're trying to evict from. 372251629Sdelphij */ 373168404Spjd kstat_named_t arcstat_evict_skip; 374208373Smm kstat_named_t arcstat_evict_l2_cached; 375208373Smm kstat_named_t arcstat_evict_l2_eligible; 376208373Smm kstat_named_t arcstat_evict_l2_ineligible; 377168404Spjd kstat_named_t arcstat_hash_elements; 378168404Spjd kstat_named_t arcstat_hash_elements_max; 379168404Spjd kstat_named_t arcstat_hash_collisions; 380168404Spjd kstat_named_t arcstat_hash_chains; 381168404Spjd kstat_named_t arcstat_hash_chain_max; 382168404Spjd kstat_named_t arcstat_p; 383168404Spjd kstat_named_t arcstat_c; 384168404Spjd kstat_named_t arcstat_c_min; 385168404Spjd kstat_named_t arcstat_c_max; 386168404Spjd kstat_named_t arcstat_size; 387185029Spjd kstat_named_t arcstat_hdr_size; 388208373Smm kstat_named_t arcstat_data_size; 389208373Smm kstat_named_t arcstat_other_size; 390185029Spjd kstat_named_t arcstat_l2_hits; 391185029Spjd kstat_named_t arcstat_l2_misses; 392185029Spjd kstat_named_t arcstat_l2_feeds; 393185029Spjd kstat_named_t arcstat_l2_rw_clash; 394208373Smm kstat_named_t arcstat_l2_read_bytes; 395208373Smm kstat_named_t arcstat_l2_write_bytes; 396185029Spjd kstat_named_t arcstat_l2_writes_sent; 397185029Spjd kstat_named_t arcstat_l2_writes_done; 398185029Spjd kstat_named_t arcstat_l2_writes_error; 399185029Spjd kstat_named_t arcstat_l2_writes_hdr_miss; 400185029Spjd kstat_named_t arcstat_l2_evict_lock_retry; 401185029Spjd kstat_named_t arcstat_l2_evict_reading; 402286570Smav kstat_named_t arcstat_l2_evict_l1cached; 403185029Spjd kstat_named_t arcstat_l2_free_on_write; 404274172Savg kstat_named_t arcstat_l2_cdata_free_on_write; 405185029Spjd kstat_named_t arcstat_l2_abort_lowmem; 406185029Spjd kstat_named_t arcstat_l2_cksum_bad; 407185029Spjd kstat_named_t arcstat_l2_io_error; 408185029Spjd kstat_named_t arcstat_l2_size; 409251478Sdelphij kstat_named_t arcstat_l2_asize; 410185029Spjd kstat_named_t arcstat_l2_hdr_size; 411251478Sdelphij kstat_named_t arcstat_l2_compress_successes; 412251478Sdelphij kstat_named_t arcstat_l2_compress_zeros; 413251478Sdelphij kstat_named_t arcstat_l2_compress_failures; 414205231Skmacy kstat_named_t arcstat_l2_write_trylock_fail; 415205231Skmacy kstat_named_t arcstat_l2_write_passed_headroom; 416205231Skmacy kstat_named_t arcstat_l2_write_spa_mismatch; 417206796Spjd kstat_named_t arcstat_l2_write_in_l2; 418205231Skmacy kstat_named_t arcstat_l2_write_hdr_io_in_progress; 419205231Skmacy kstat_named_t arcstat_l2_write_not_cacheable; 420205231Skmacy kstat_named_t arcstat_l2_write_full; 421205231Skmacy kstat_named_t arcstat_l2_write_buffer_iter; 422205231Skmacy kstat_named_t arcstat_l2_write_pios; 423205231Skmacy kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 424205231Skmacy kstat_named_t arcstat_l2_write_buffer_list_iter; 425205231Skmacy kstat_named_t arcstat_l2_write_buffer_list_null_iter; 426242845Sdelphij kstat_named_t arcstat_memory_throttle_count; 427242845Sdelphij kstat_named_t arcstat_duplicate_buffers; 428242845Sdelphij kstat_named_t arcstat_duplicate_buffers_size; 429242845Sdelphij kstat_named_t arcstat_duplicate_reads; 430275748Sdelphij kstat_named_t arcstat_meta_used; 431275748Sdelphij kstat_named_t arcstat_meta_limit; 432275748Sdelphij kstat_named_t arcstat_meta_max; 433275780Sdelphij kstat_named_t arcstat_meta_min; 434168404Spjd} arc_stats_t; 435168404Spjd 436168404Spjdstatic arc_stats_t arc_stats = { 437168404Spjd { "hits", KSTAT_DATA_UINT64 }, 438168404Spjd { "misses", KSTAT_DATA_UINT64 }, 439168404Spjd { "demand_data_hits", KSTAT_DATA_UINT64 }, 440168404Spjd { "demand_data_misses", KSTAT_DATA_UINT64 }, 441168404Spjd { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 442168404Spjd { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 443168404Spjd { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 444168404Spjd { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 445168404Spjd { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 446168404Spjd { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 447168404Spjd { "mru_hits", KSTAT_DATA_UINT64 }, 448168404Spjd { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 449168404Spjd { "mfu_hits", KSTAT_DATA_UINT64 }, 450168404Spjd { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 451205231Skmacy { "allocated", KSTAT_DATA_UINT64 }, 452168404Spjd { "deleted", KSTAT_DATA_UINT64 }, 453205231Skmacy { "stolen", KSTAT_DATA_UINT64 }, 454168404Spjd { "recycle_miss", KSTAT_DATA_UINT64 }, 455168404Spjd { "mutex_miss", KSTAT_DATA_UINT64 }, 456168404Spjd { "evict_skip", KSTAT_DATA_UINT64 }, 457208373Smm { "evict_l2_cached", KSTAT_DATA_UINT64 }, 458208373Smm { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 459208373Smm { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 460168404Spjd { "hash_elements", KSTAT_DATA_UINT64 }, 461168404Spjd { "hash_elements_max", KSTAT_DATA_UINT64 }, 462168404Spjd { "hash_collisions", KSTAT_DATA_UINT64 }, 463168404Spjd { "hash_chains", KSTAT_DATA_UINT64 }, 464168404Spjd { "hash_chain_max", KSTAT_DATA_UINT64 }, 465168404Spjd { "p", KSTAT_DATA_UINT64 }, 466168404Spjd { "c", KSTAT_DATA_UINT64 }, 467168404Spjd { "c_min", KSTAT_DATA_UINT64 }, 468168404Spjd { "c_max", KSTAT_DATA_UINT64 }, 469185029Spjd { "size", KSTAT_DATA_UINT64 }, 470185029Spjd { "hdr_size", KSTAT_DATA_UINT64 }, 471208373Smm { "data_size", KSTAT_DATA_UINT64 }, 472208373Smm { "other_size", KSTAT_DATA_UINT64 }, 473185029Spjd { "l2_hits", KSTAT_DATA_UINT64 }, 474185029Spjd { "l2_misses", KSTAT_DATA_UINT64 }, 475185029Spjd { "l2_feeds", KSTAT_DATA_UINT64 }, 476185029Spjd { "l2_rw_clash", KSTAT_DATA_UINT64 }, 477208373Smm { "l2_read_bytes", KSTAT_DATA_UINT64 }, 478208373Smm { "l2_write_bytes", KSTAT_DATA_UINT64 }, 479185029Spjd { "l2_writes_sent", KSTAT_DATA_UINT64 }, 480185029Spjd { "l2_writes_done", KSTAT_DATA_UINT64 }, 481185029Spjd { "l2_writes_error", KSTAT_DATA_UINT64 }, 482185029Spjd { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 483185029Spjd { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 484185029Spjd { "l2_evict_reading", KSTAT_DATA_UINT64 }, 485286570Smav { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 486185029Spjd { "l2_free_on_write", KSTAT_DATA_UINT64 }, 487274172Savg { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, 488185029Spjd { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 489185029Spjd { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 490185029Spjd { "l2_io_error", KSTAT_DATA_UINT64 }, 491185029Spjd { "l2_size", KSTAT_DATA_UINT64 }, 492251478Sdelphij { "l2_asize", KSTAT_DATA_UINT64 }, 493185029Spjd { "l2_hdr_size", KSTAT_DATA_UINT64 }, 494251478Sdelphij { "l2_compress_successes", KSTAT_DATA_UINT64 }, 495251478Sdelphij { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 496251478Sdelphij { "l2_compress_failures", KSTAT_DATA_UINT64 }, 497206796Spjd { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 498206796Spjd { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 499206796Spjd { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 500206796Spjd { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 501206796Spjd { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 502206796Spjd { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 503206796Spjd { "l2_write_full", KSTAT_DATA_UINT64 }, 504206796Spjd { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 505206796Spjd { "l2_write_pios", KSTAT_DATA_UINT64 }, 506206796Spjd { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 507206796Spjd { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 508242845Sdelphij { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 509242845Sdelphij { "memory_throttle_count", KSTAT_DATA_UINT64 }, 510242845Sdelphij { "duplicate_buffers", KSTAT_DATA_UINT64 }, 511242845Sdelphij { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 512275748Sdelphij { "duplicate_reads", KSTAT_DATA_UINT64 }, 513275748Sdelphij { "arc_meta_used", KSTAT_DATA_UINT64 }, 514275748Sdelphij { "arc_meta_limit", KSTAT_DATA_UINT64 }, 515275780Sdelphij { "arc_meta_max", KSTAT_DATA_UINT64 }, 516275780Sdelphij { "arc_meta_min", KSTAT_DATA_UINT64 } 517168404Spjd}; 518168404Spjd 519168404Spjd#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 520168404Spjd 521168404Spjd#define ARCSTAT_INCR(stat, val) \ 522251631Sdelphij atomic_add_64(&arc_stats.stat.value.ui64, (val)) 523168404Spjd 524206796Spjd#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 525168404Spjd#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 526168404Spjd 527168404Spjd#define ARCSTAT_MAX(stat, val) { \ 528168404Spjd uint64_t m; \ 529168404Spjd while ((val) > (m = arc_stats.stat.value.ui64) && \ 530168404Spjd (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 531168404Spjd continue; \ 532168404Spjd} 533168404Spjd 534168404Spjd#define ARCSTAT_MAXSTAT(stat) \ 535168404Spjd ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 536168404Spjd 537168404Spjd/* 538168404Spjd * We define a macro to allow ARC hits/misses to be easily broken down by 539168404Spjd * two separate conditions, giving a total of four different subtypes for 540168404Spjd * each of hits and misses (so eight statistics total). 541168404Spjd */ 542168404Spjd#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 543168404Spjd if (cond1) { \ 544168404Spjd if (cond2) { \ 545168404Spjd ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 546168404Spjd } else { \ 547168404Spjd ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 548168404Spjd } \ 549168404Spjd } else { \ 550168404Spjd if (cond2) { \ 551168404Spjd ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 552168404Spjd } else { \ 553168404Spjd ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 554168404Spjd } \ 555168404Spjd } 556168404Spjd 557168404Spjdkstat_t *arc_ksp; 558206796Spjdstatic arc_state_t *arc_anon; 559168404Spjdstatic arc_state_t *arc_mru; 560168404Spjdstatic arc_state_t *arc_mru_ghost; 561168404Spjdstatic arc_state_t *arc_mfu; 562168404Spjdstatic arc_state_t *arc_mfu_ghost; 563185029Spjdstatic arc_state_t *arc_l2c_only; 564168404Spjd 565168404Spjd/* 566168404Spjd * There are several ARC variables that are critical to export as kstats -- 567168404Spjd * but we don't want to have to grovel around in the kstat whenever we wish to 568168404Spjd * manipulate them. For these variables, we therefore define them to be in 569168404Spjd * terms of the statistic variable. This assures that we are not introducing 570168404Spjd * the possibility of inconsistency by having shadow copies of the variables, 571168404Spjd * while still allowing the code to be readable. 572168404Spjd */ 573168404Spjd#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 574168404Spjd#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 575168404Spjd#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 576168404Spjd#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 577168404Spjd#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 578275748Sdelphij#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 579275780Sdelphij#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 580275748Sdelphij#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 581275748Sdelphij#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 582168404Spjd 583251478Sdelphij#define L2ARC_IS_VALID_COMPRESS(_c_) \ 584251478Sdelphij ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 585251478Sdelphij 586168404Spjdstatic int arc_no_grow; /* Don't try to grow cache size */ 587168404Spjdstatic uint64_t arc_tempreserve; 588209962Smmstatic uint64_t arc_loaned_bytes; 589168404Spjd 590168404Spjdtypedef struct arc_callback arc_callback_t; 591168404Spjd 592168404Spjdstruct arc_callback { 593168404Spjd void *acb_private; 594168404Spjd arc_done_func_t *acb_done; 595168404Spjd arc_buf_t *acb_buf; 596168404Spjd zio_t *acb_zio_dummy; 597168404Spjd arc_callback_t *acb_next; 598168404Spjd}; 599168404Spjd 600168404Spjdtypedef struct arc_write_callback arc_write_callback_t; 601168404Spjd 602168404Spjdstruct arc_write_callback { 603168404Spjd void *awcb_private; 604168404Spjd arc_done_func_t *awcb_ready; 605258632Savg arc_done_func_t *awcb_physdone; 606168404Spjd arc_done_func_t *awcb_done; 607168404Spjd arc_buf_t *awcb_buf; 608168404Spjd}; 609168404Spjd 610286570Smav/* 611286570Smav * ARC buffers are separated into multiple structs as a memory saving measure: 612286570Smav * - Common fields struct, always defined, and embedded within it: 613286570Smav * - L2-only fields, always allocated but undefined when not in L2ARC 614286570Smav * - L1-only fields, only allocated when in L1ARC 615286570Smav * 616286570Smav * Buffer in L1 Buffer only in L2 617286570Smav * +------------------------+ +------------------------+ 618286570Smav * | arc_buf_hdr_t | | arc_buf_hdr_t | 619286570Smav * | | | | 620286570Smav * | | | | 621286570Smav * | | | | 622286570Smav * +------------------------+ +------------------------+ 623286570Smav * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 624286570Smav * | (undefined if L1-only) | | | 625286570Smav * +------------------------+ +------------------------+ 626286570Smav * | l1arc_buf_hdr_t | 627286570Smav * | | 628286570Smav * | | 629286570Smav * | | 630286570Smav * | | 631286570Smav * +------------------------+ 632286570Smav * 633286570Smav * Because it's possible for the L2ARC to become extremely large, we can wind 634286570Smav * up eating a lot of memory in L2ARC buffer headers, so the size of a header 635286570Smav * is minimized by only allocating the fields necessary for an L1-cached buffer 636286570Smav * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 637286570Smav * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 638286570Smav * words in pointers. arc_hdr_realloc() is used to switch a header between 639286570Smav * these two allocation states. 640286570Smav */ 641286570Smavtypedef struct l1arc_buf_hdr { 642168404Spjd kmutex_t b_freeze_lock; 643286570Smav#ifdef ZFS_DEBUG 644286570Smav /* 645286570Smav * used for debugging wtih kmem_flags - by allocating and freeing 646286570Smav * b_thawed when the buffer is thawed, we get a record of the stack 647286570Smav * trace that thawed it. 648286570Smav */ 649219089Spjd void *b_thawed; 650286570Smav#endif 651168404Spjd 652168404Spjd arc_buf_t *b_buf; 653168404Spjd uint32_t b_datacnt; 654286570Smav /* for waiting on writes to complete */ 655168404Spjd kcondvar_t b_cv; 656168404Spjd 657168404Spjd /* protected by arc state mutex */ 658168404Spjd arc_state_t *b_state; 659168404Spjd list_node_t b_arc_node; 660168404Spjd 661168404Spjd /* updated atomically */ 662168404Spjd clock_t b_arc_access; 663168404Spjd 664168404Spjd /* self protecting */ 665168404Spjd refcount_t b_refcnt; 666185029Spjd 667286570Smav arc_callback_t *b_acb; 668286570Smav /* temporary buffer holder for in-flight compressed data */ 669286570Smav void *b_tmp_cdata; 670286570Smav} l1arc_buf_hdr_t; 671286570Smav 672286570Smavtypedef struct l2arc_dev l2arc_dev_t; 673286570Smav 674286570Smavtypedef struct l2arc_buf_hdr { 675286570Smav /* protected by arc_buf_hdr mutex */ 676286570Smav l2arc_dev_t *b_dev; /* L2ARC device */ 677286570Smav uint64_t b_daddr; /* disk address, offset byte */ 678286570Smav /* real alloc'd buffer size depending on b_compress applied */ 679286570Smav int32_t b_asize; 680286570Smav 681185029Spjd list_node_t b_l2node; 682286570Smav} l2arc_buf_hdr_t; 683286570Smav 684286570Smavstruct arc_buf_hdr { 685286570Smav /* protected by hash lock */ 686286570Smav dva_t b_dva; 687286570Smav uint64_t b_birth; 688286570Smav /* 689286570Smav * Even though this checksum is only set/verified when a buffer is in 690286570Smav * the L1 cache, it needs to be in the set of common fields because it 691286570Smav * must be preserved from the time before a buffer is written out to 692286570Smav * L2ARC until after it is read back in. 693286570Smav */ 694286570Smav zio_cksum_t *b_freeze_cksum; 695286570Smav 696286570Smav arc_buf_hdr_t *b_hash_next; 697286570Smav arc_flags_t b_flags; 698286570Smav 699286570Smav /* immutable */ 700286570Smav int32_t b_size; 701286570Smav uint64_t b_spa; 702286570Smav 703286570Smav /* L2ARC fields. Undefined when not in L2ARC. */ 704286570Smav l2arc_buf_hdr_t b_l2hdr; 705286570Smav /* L1ARC fields. Undefined when in l2arc_only state */ 706286570Smav l1arc_buf_hdr_t b_l1hdr; 707168404Spjd}; 708168404Spjd 709275748Sdelphij#ifdef _KERNEL 710275748Sdelphijstatic int 711275748Sdelphijsysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 712275748Sdelphij{ 713275748Sdelphij uint64_t val; 714275748Sdelphij int err; 715275748Sdelphij 716275748Sdelphij val = arc_meta_limit; 717275748Sdelphij err = sysctl_handle_64(oidp, &val, 0, req); 718275748Sdelphij if (err != 0 || req->newptr == NULL) 719275748Sdelphij return (err); 720275748Sdelphij 721275748Sdelphij if (val <= 0 || val > arc_c_max) 722275748Sdelphij return (EINVAL); 723275748Sdelphij 724275748Sdelphij arc_meta_limit = val; 725275748Sdelphij return (0); 726275748Sdelphij} 727275748Sdelphij#endif 728275748Sdelphij 729168404Spjdstatic arc_buf_t *arc_eviction_list; 730168404Spjdstatic kmutex_t arc_eviction_mtx; 731168404Spjdstatic arc_buf_hdr_t arc_eviction_hdr; 732168404Spjd 733168404Spjd#define GHOST_STATE(state) \ 734185029Spjd ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 735185029Spjd (state) == arc_l2c_only) 736168404Spjd 737275811Sdelphij#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 738275811Sdelphij#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 739275811Sdelphij#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 740275811Sdelphij#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 741275811Sdelphij#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ) 742275811Sdelphij#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE) 743286570Smav 744275811Sdelphij#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 745286570Smav#define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS) 746275811Sdelphij#define HDR_L2_READING(hdr) \ 747286570Smav (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 748286570Smav ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 749275811Sdelphij#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 750275811Sdelphij#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 751275811Sdelphij#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 752168404Spjd 753286570Smav#define HDR_ISTYPE_METADATA(hdr) \ 754286570Smav ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 755286570Smav#define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 756286570Smav 757286570Smav#define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 758286570Smav#define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 759286570Smav 760286570Smav/* For storing compression mode in b_flags */ 761286570Smav#define HDR_COMPRESS_OFFSET 24 762286570Smav#define HDR_COMPRESS_NBITS 7 763286570Smav 764286570Smav#define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET(hdr->b_flags, \ 765286570Smav HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS)) 766286570Smav#define HDR_SET_COMPRESS(hdr, cmp) BF32_SET(hdr->b_flags, \ 767286570Smav HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS, (cmp)) 768286570Smav 769168404Spjd/* 770185029Spjd * Other sizes 771185029Spjd */ 772185029Spjd 773286570Smav#define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 774286570Smav#define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 775185029Spjd 776185029Spjd/* 777168404Spjd * Hash table routines 778168404Spjd */ 779168404Spjd 780205253Skmacy#define HT_LOCK_PAD CACHE_LINE_SIZE 781168404Spjd 782168404Spjdstruct ht_lock { 783168404Spjd kmutex_t ht_lock; 784168404Spjd#ifdef _KERNEL 785168404Spjd unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 786168404Spjd#endif 787168404Spjd}; 788168404Spjd 789168404Spjd#define BUF_LOCKS 256 790168404Spjdtypedef struct buf_hash_table { 791168404Spjd uint64_t ht_mask; 792168404Spjd arc_buf_hdr_t **ht_table; 793205264Skmacy struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 794168404Spjd} buf_hash_table_t; 795168404Spjd 796168404Spjdstatic buf_hash_table_t buf_hash_table; 797168404Spjd 798168404Spjd#define BUF_HASH_INDEX(spa, dva, birth) \ 799168404Spjd (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 800168404Spjd#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 801168404Spjd#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 802219089Spjd#define HDR_LOCK(hdr) \ 803219089Spjd (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 804168404Spjd 805168404Spjduint64_t zfs_crc64_table[256]; 806168404Spjd 807185029Spjd/* 808185029Spjd * Level 2 ARC 809185029Spjd */ 810185029Spjd 811272707Savg#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 812251478Sdelphij#define L2ARC_HEADROOM 2 /* num of writes */ 813251478Sdelphij/* 814251478Sdelphij * If we discover during ARC scan any buffers to be compressed, we boost 815251478Sdelphij * our headroom for the next scanning cycle by this percentage multiple. 816251478Sdelphij */ 817251478Sdelphij#define L2ARC_HEADROOM_BOOST 200 818208373Smm#define L2ARC_FEED_SECS 1 /* caching interval secs */ 819208373Smm#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 820185029Spjd 821185029Spjd#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 822185029Spjd#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 823185029Spjd 824251631Sdelphij/* L2ARC Performance Tunables */ 825185029Spjduint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 826185029Spjduint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 827185029Spjduint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 828251478Sdelphijuint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 829185029Spjduint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 830208373Smmuint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 831219089Spjdboolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 832208373Smmboolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 833208373Smmboolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 834185029Spjd 835217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 836205231Skmacy &l2arc_write_max, 0, "max write size"); 837217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 838205231Skmacy &l2arc_write_boost, 0, "extra write during warmup"); 839217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 840205231Skmacy &l2arc_headroom, 0, "number of dev writes"); 841217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 842205231Skmacy &l2arc_feed_secs, 0, "interval seconds"); 843217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 844208373Smm &l2arc_feed_min_ms, 0, "min interval milliseconds"); 845205231Skmacy 846205231SkmacySYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 847205231Skmacy &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 848208373SmmSYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 849208373Smm &l2arc_feed_again, 0, "turbo warmup"); 850208373SmmSYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 851208373Smm &l2arc_norw, 0, "no reads during writes"); 852205231Skmacy 853217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 854205231Skmacy &ARC_anon.arcs_size, 0, "size of anonymous state"); 855217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 856205231Skmacy &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 857217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 858205231Skmacy &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 859205231Skmacy 860217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 861205231Skmacy &ARC_mru.arcs_size, 0, "size of mru state"); 862217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 863205231Skmacy &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 864217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 865205231Skmacy &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 866205231Skmacy 867217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 868205231Skmacy &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 869217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 870205231Skmacy &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 871205231Skmacy "size of metadata in mru ghost state"); 872217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 873205231Skmacy &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 874205231Skmacy "size of data in mru ghost state"); 875205231Skmacy 876217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 877205231Skmacy &ARC_mfu.arcs_size, 0, "size of mfu state"); 878217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 879205231Skmacy &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 880217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 881205231Skmacy &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 882205231Skmacy 883217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 884205231Skmacy &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 885217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 886205231Skmacy &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 887205231Skmacy "size of metadata in mfu ghost state"); 888217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 889205231Skmacy &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 890205231Skmacy "size of data in mfu ghost state"); 891205231Skmacy 892217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 893205231Skmacy &ARC_l2c_only.arcs_size, 0, "size of mru state"); 894205231Skmacy 895185029Spjd/* 896185029Spjd * L2ARC Internals 897185029Spjd */ 898286570Smavstruct l2arc_dev { 899185029Spjd vdev_t *l2ad_vdev; /* vdev */ 900185029Spjd spa_t *l2ad_spa; /* spa */ 901185029Spjd uint64_t l2ad_hand; /* next write location */ 902185029Spjd uint64_t l2ad_start; /* first addr on device */ 903185029Spjd uint64_t l2ad_end; /* last addr on device */ 904185029Spjd uint64_t l2ad_evict; /* last addr eviction reached */ 905185029Spjd boolean_t l2ad_first; /* first sweep through */ 906208373Smm boolean_t l2ad_writing; /* currently writing */ 907286570Smav kmutex_t l2ad_mtx; /* lock for buffer list */ 908286570Smav list_t l2ad_buflist; /* buffer list */ 909185029Spjd list_node_t l2ad_node; /* device list node */ 910286570Smav}; 911185029Spjd 912185029Spjdstatic list_t L2ARC_dev_list; /* device list */ 913185029Spjdstatic list_t *l2arc_dev_list; /* device list pointer */ 914185029Spjdstatic kmutex_t l2arc_dev_mtx; /* device list mutex */ 915185029Spjdstatic l2arc_dev_t *l2arc_dev_last; /* last device used */ 916185029Spjdstatic list_t L2ARC_free_on_write; /* free after write buf list */ 917185029Spjdstatic list_t *l2arc_free_on_write; /* free after write list ptr */ 918185029Spjdstatic kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 919185029Spjdstatic uint64_t l2arc_ndev; /* number of devices */ 920185029Spjd 921185029Spjdtypedef struct l2arc_read_callback { 922251478Sdelphij arc_buf_t *l2rcb_buf; /* read buffer */ 923251478Sdelphij spa_t *l2rcb_spa; /* spa */ 924251478Sdelphij blkptr_t l2rcb_bp; /* original blkptr */ 925268123Sdelphij zbookmark_phys_t l2rcb_zb; /* original bookmark */ 926251478Sdelphij int l2rcb_flags; /* original flags */ 927251478Sdelphij enum zio_compress l2rcb_compress; /* applied compress */ 928185029Spjd} l2arc_read_callback_t; 929185029Spjd 930185029Spjdtypedef struct l2arc_write_callback { 931185029Spjd l2arc_dev_t *l2wcb_dev; /* device info */ 932185029Spjd arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 933185029Spjd} l2arc_write_callback_t; 934185029Spjd 935185029Spjdtypedef struct l2arc_data_free { 936185029Spjd /* protected by l2arc_free_on_write_mtx */ 937185029Spjd void *l2df_data; 938185029Spjd size_t l2df_size; 939185029Spjd void (*l2df_func)(void *, size_t); 940185029Spjd list_node_t l2df_list_node; 941185029Spjd} l2arc_data_free_t; 942185029Spjd 943185029Spjdstatic kmutex_t l2arc_feed_thr_lock; 944185029Spjdstatic kcondvar_t l2arc_feed_thr_cv; 945185029Spjdstatic uint8_t l2arc_thread_exit; 946185029Spjd 947275811Sdelphijstatic void arc_get_data_buf(arc_buf_t *); 948275811Sdelphijstatic void arc_access(arc_buf_hdr_t *, kmutex_t *); 949275811Sdelphijstatic int arc_evict_needed(arc_buf_contents_t); 950275811Sdelphijstatic void arc_evict_ghost(arc_state_t *, uint64_t, int64_t); 951275811Sdelphijstatic void arc_buf_watch(arc_buf_t *); 952275811Sdelphij 953286570Smavstatic arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 954286570Smavstatic uint32_t arc_bufc_to_flags(arc_buf_contents_t); 955286570Smav 956275811Sdelphijstatic boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 957275811Sdelphijstatic void l2arc_read_done(zio_t *); 958185029Spjd 959286570Smavstatic boolean_t l2arc_compress_buf(arc_buf_hdr_t *); 960275811Sdelphijstatic void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress); 961275811Sdelphijstatic void l2arc_release_cdata_buf(arc_buf_hdr_t *); 962251478Sdelphij 963168404Spjdstatic uint64_t 964209962Smmbuf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 965168404Spjd{ 966168404Spjd uint8_t *vdva = (uint8_t *)dva; 967168404Spjd uint64_t crc = -1ULL; 968168404Spjd int i; 969168404Spjd 970168404Spjd ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 971168404Spjd 972168404Spjd for (i = 0; i < sizeof (dva_t); i++) 973168404Spjd crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 974168404Spjd 975209962Smm crc ^= (spa>>8) ^ birth; 976168404Spjd 977168404Spjd return (crc); 978168404Spjd} 979168404Spjd 980168404Spjd#define BUF_EMPTY(buf) \ 981168404Spjd ((buf)->b_dva.dva_word[0] == 0 && \ 982286570Smav (buf)->b_dva.dva_word[1] == 0) 983168404Spjd 984168404Spjd#define BUF_EQUAL(spa, dva, birth, buf) \ 985168404Spjd ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 986168404Spjd ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 987168404Spjd ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 988168404Spjd 989219089Spjdstatic void 990219089Spjdbuf_discard_identity(arc_buf_hdr_t *hdr) 991219089Spjd{ 992219089Spjd hdr->b_dva.dva_word[0] = 0; 993219089Spjd hdr->b_dva.dva_word[1] = 0; 994219089Spjd hdr->b_birth = 0; 995219089Spjd} 996219089Spjd 997168404Spjdstatic arc_buf_hdr_t * 998268075Sdelphijbuf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 999168404Spjd{ 1000268075Sdelphij const dva_t *dva = BP_IDENTITY(bp); 1001268075Sdelphij uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1002168404Spjd uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1003168404Spjd kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1004275811Sdelphij arc_buf_hdr_t *hdr; 1005168404Spjd 1006168404Spjd mutex_enter(hash_lock); 1007275811Sdelphij for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1008275811Sdelphij hdr = hdr->b_hash_next) { 1009275811Sdelphij if (BUF_EQUAL(spa, dva, birth, hdr)) { 1010168404Spjd *lockp = hash_lock; 1011275811Sdelphij return (hdr); 1012168404Spjd } 1013168404Spjd } 1014168404Spjd mutex_exit(hash_lock); 1015168404Spjd *lockp = NULL; 1016168404Spjd return (NULL); 1017168404Spjd} 1018168404Spjd 1019168404Spjd/* 1020168404Spjd * Insert an entry into the hash table. If there is already an element 1021168404Spjd * equal to elem in the hash table, then the already existing element 1022168404Spjd * will be returned and the new element will not be inserted. 1023168404Spjd * Otherwise returns NULL. 1024286570Smav * If lockp == NULL, the caller is assumed to already hold the hash lock. 1025168404Spjd */ 1026168404Spjdstatic arc_buf_hdr_t * 1027275811Sdelphijbuf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1028168404Spjd{ 1029275811Sdelphij uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1030168404Spjd kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1031275811Sdelphij arc_buf_hdr_t *fhdr; 1032168404Spjd uint32_t i; 1033168404Spjd 1034275811Sdelphij ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1035275811Sdelphij ASSERT(hdr->b_birth != 0); 1036275811Sdelphij ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1037286570Smav 1038286570Smav if (lockp != NULL) { 1039286570Smav *lockp = hash_lock; 1040286570Smav mutex_enter(hash_lock); 1041286570Smav } else { 1042286570Smav ASSERT(MUTEX_HELD(hash_lock)); 1043286570Smav } 1044286570Smav 1045275811Sdelphij for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1046275811Sdelphij fhdr = fhdr->b_hash_next, i++) { 1047275811Sdelphij if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1048275811Sdelphij return (fhdr); 1049168404Spjd } 1050168404Spjd 1051275811Sdelphij hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1052275811Sdelphij buf_hash_table.ht_table[idx] = hdr; 1053275811Sdelphij hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 1054168404Spjd 1055168404Spjd /* collect some hash table performance data */ 1056168404Spjd if (i > 0) { 1057168404Spjd ARCSTAT_BUMP(arcstat_hash_collisions); 1058168404Spjd if (i == 1) 1059168404Spjd ARCSTAT_BUMP(arcstat_hash_chains); 1060168404Spjd 1061168404Spjd ARCSTAT_MAX(arcstat_hash_chain_max, i); 1062168404Spjd } 1063168404Spjd 1064168404Spjd ARCSTAT_BUMP(arcstat_hash_elements); 1065168404Spjd ARCSTAT_MAXSTAT(arcstat_hash_elements); 1066168404Spjd 1067168404Spjd return (NULL); 1068168404Spjd} 1069168404Spjd 1070168404Spjdstatic void 1071275811Sdelphijbuf_hash_remove(arc_buf_hdr_t *hdr) 1072168404Spjd{ 1073275811Sdelphij arc_buf_hdr_t *fhdr, **hdrp; 1074275811Sdelphij uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1075168404Spjd 1076168404Spjd ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1077275811Sdelphij ASSERT(HDR_IN_HASH_TABLE(hdr)); 1078168404Spjd 1079275811Sdelphij hdrp = &buf_hash_table.ht_table[idx]; 1080275811Sdelphij while ((fhdr = *hdrp) != hdr) { 1081275811Sdelphij ASSERT(fhdr != NULL); 1082275811Sdelphij hdrp = &fhdr->b_hash_next; 1083168404Spjd } 1084275811Sdelphij *hdrp = hdr->b_hash_next; 1085275811Sdelphij hdr->b_hash_next = NULL; 1086275811Sdelphij hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; 1087168404Spjd 1088168404Spjd /* collect some hash table performance data */ 1089168404Spjd ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1090168404Spjd 1091168404Spjd if (buf_hash_table.ht_table[idx] && 1092168404Spjd buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1093168404Spjd ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1094168404Spjd} 1095168404Spjd 1096168404Spjd/* 1097168404Spjd * Global data structures and functions for the buf kmem cache. 1098168404Spjd */ 1099286570Smavstatic kmem_cache_t *hdr_full_cache; 1100286570Smavstatic kmem_cache_t *hdr_l2only_cache; 1101168404Spjdstatic kmem_cache_t *buf_cache; 1102168404Spjd 1103168404Spjdstatic void 1104168404Spjdbuf_fini(void) 1105168404Spjd{ 1106168404Spjd int i; 1107168404Spjd 1108168404Spjd kmem_free(buf_hash_table.ht_table, 1109168404Spjd (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1110168404Spjd for (i = 0; i < BUF_LOCKS; i++) 1111168404Spjd mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1112286570Smav kmem_cache_destroy(hdr_full_cache); 1113286570Smav kmem_cache_destroy(hdr_l2only_cache); 1114168404Spjd kmem_cache_destroy(buf_cache); 1115168404Spjd} 1116168404Spjd 1117168404Spjd/* 1118168404Spjd * Constructor callback - called when the cache is empty 1119168404Spjd * and a new buf is requested. 1120168404Spjd */ 1121168404Spjd/* ARGSUSED */ 1122168404Spjdstatic int 1123286570Smavhdr_full_cons(void *vbuf, void *unused, int kmflag) 1124168404Spjd{ 1125275811Sdelphij arc_buf_hdr_t *hdr = vbuf; 1126168404Spjd 1127286570Smav bzero(hdr, HDR_FULL_SIZE); 1128286570Smav cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1129286570Smav refcount_create(&hdr->b_l1hdr.b_refcnt); 1130286570Smav mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1131286570Smav arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1132185029Spjd 1133168404Spjd return (0); 1134168404Spjd} 1135168404Spjd 1136185029Spjd/* ARGSUSED */ 1137185029Spjdstatic int 1138286570Smavhdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1139286570Smav{ 1140286570Smav arc_buf_hdr_t *hdr = vbuf; 1141286570Smav 1142286570Smav bzero(hdr, HDR_L2ONLY_SIZE); 1143286570Smav arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1144286570Smav 1145286570Smav return (0); 1146286570Smav} 1147286570Smav 1148286570Smav/* ARGSUSED */ 1149286570Smavstatic int 1150185029Spjdbuf_cons(void *vbuf, void *unused, int kmflag) 1151185029Spjd{ 1152185029Spjd arc_buf_t *buf = vbuf; 1153185029Spjd 1154185029Spjd bzero(buf, sizeof (arc_buf_t)); 1155219089Spjd mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1156208373Smm arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1157208373Smm 1158185029Spjd return (0); 1159185029Spjd} 1160185029Spjd 1161168404Spjd/* 1162168404Spjd * Destructor callback - called when a cached buf is 1163168404Spjd * no longer required. 1164168404Spjd */ 1165168404Spjd/* ARGSUSED */ 1166168404Spjdstatic void 1167286570Smavhdr_full_dest(void *vbuf, void *unused) 1168168404Spjd{ 1169275811Sdelphij arc_buf_hdr_t *hdr = vbuf; 1170168404Spjd 1171275811Sdelphij ASSERT(BUF_EMPTY(hdr)); 1172286570Smav cv_destroy(&hdr->b_l1hdr.b_cv); 1173286570Smav refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1174286570Smav mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1175286570Smav arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1176168404Spjd} 1177168404Spjd 1178185029Spjd/* ARGSUSED */ 1179185029Spjdstatic void 1180286570Smavhdr_l2only_dest(void *vbuf, void *unused) 1181286570Smav{ 1182286570Smav arc_buf_hdr_t *hdr = vbuf; 1183286570Smav 1184286570Smav ASSERT(BUF_EMPTY(hdr)); 1185286570Smav arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1186286570Smav} 1187286570Smav 1188286570Smav/* ARGSUSED */ 1189286570Smavstatic void 1190185029Spjdbuf_dest(void *vbuf, void *unused) 1191185029Spjd{ 1192185029Spjd arc_buf_t *buf = vbuf; 1193185029Spjd 1194219089Spjd mutex_destroy(&buf->b_evict_lock); 1195208373Smm arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1196185029Spjd} 1197185029Spjd 1198168404Spjd/* 1199168404Spjd * Reclaim callback -- invoked when memory is low. 1200168404Spjd */ 1201168404Spjd/* ARGSUSED */ 1202168404Spjdstatic void 1203168404Spjdhdr_recl(void *unused) 1204168404Spjd{ 1205168404Spjd dprintf("hdr_recl called\n"); 1206168404Spjd /* 1207168404Spjd * umem calls the reclaim func when we destroy the buf cache, 1208168404Spjd * which is after we do arc_fini(). 1209168404Spjd */ 1210168404Spjd if (!arc_dead) 1211168404Spjd cv_signal(&arc_reclaim_thr_cv); 1212168404Spjd} 1213168404Spjd 1214168404Spjdstatic void 1215168404Spjdbuf_init(void) 1216168404Spjd{ 1217168404Spjd uint64_t *ct; 1218168404Spjd uint64_t hsize = 1ULL << 12; 1219168404Spjd int i, j; 1220168404Spjd 1221168404Spjd /* 1222168404Spjd * The hash table is big enough to fill all of physical memory 1223269230Sdelphij * with an average block size of zfs_arc_average_blocksize (default 8K). 1224269230Sdelphij * By default, the table will take up 1225269230Sdelphij * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1226168404Spjd */ 1227269230Sdelphij while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1228168404Spjd hsize <<= 1; 1229168404Spjdretry: 1230168404Spjd buf_hash_table.ht_mask = hsize - 1; 1231168404Spjd buf_hash_table.ht_table = 1232168404Spjd kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1233168404Spjd if (buf_hash_table.ht_table == NULL) { 1234168404Spjd ASSERT(hsize > (1ULL << 8)); 1235168404Spjd hsize >>= 1; 1236168404Spjd goto retry; 1237168404Spjd } 1238168404Spjd 1239286570Smav hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1240286570Smav 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1241286570Smav hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1242286570Smav HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1243286570Smav NULL, NULL, 0); 1244168404Spjd buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1245185029Spjd 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1246168404Spjd 1247168404Spjd for (i = 0; i < 256; i++) 1248168404Spjd for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1249168404Spjd *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1250168404Spjd 1251168404Spjd for (i = 0; i < BUF_LOCKS; i++) { 1252168404Spjd mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1253168404Spjd NULL, MUTEX_DEFAULT, NULL); 1254168404Spjd } 1255168404Spjd} 1256168404Spjd 1257286570Smav/* 1258286570Smav * Transition between the two allocation states for the arc_buf_hdr struct. 1259286570Smav * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 1260286570Smav * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 1261286570Smav * version is used when a cache buffer is only in the L2ARC in order to reduce 1262286570Smav * memory usage. 1263286570Smav */ 1264286570Smavstatic arc_buf_hdr_t * 1265286570Smavarc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 1266286570Smav{ 1267286570Smav ASSERT(HDR_HAS_L2HDR(hdr)); 1268286570Smav 1269286570Smav arc_buf_hdr_t *nhdr; 1270286570Smav l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1271286570Smav 1272286570Smav ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 1273286570Smav (old == hdr_l2only_cache && new == hdr_full_cache)); 1274286570Smav 1275286570Smav nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 1276286570Smav 1277286570Smav ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 1278286570Smav buf_hash_remove(hdr); 1279286570Smav 1280286570Smav bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 1281286570Smav if (new == hdr_full_cache) { 1282286570Smav nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 1283286570Smav /* 1284286570Smav * arc_access and arc_change_state need to be aware that a 1285286570Smav * header has just come out of L2ARC, so we set its state to 1286286570Smav * l2c_only even though it's about to change. 1287286570Smav */ 1288286570Smav nhdr->b_l1hdr.b_state = arc_l2c_only; 1289286570Smav } else { 1290286570Smav ASSERT(hdr->b_l1hdr.b_buf == NULL); 1291286570Smav ASSERT0(hdr->b_l1hdr.b_datacnt); 1292286570Smav ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 1293286570Smav /* 1294286570Smav * We might be removing the L1hdr of a buffer which was just 1295286570Smav * written out to L2ARC. If such a buffer is compressed then we 1296286570Smav * need to free its b_tmp_cdata before destroying the header. 1297286570Smav */ 1298286570Smav if (hdr->b_l1hdr.b_tmp_cdata != NULL && 1299286570Smav HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) 1300286570Smav l2arc_release_cdata_buf(hdr); 1301286570Smav nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR; 1302286570Smav } 1303286570Smav /* 1304286570Smav * The header has been reallocated so we need to re-insert it into any 1305286570Smav * lists it was on. 1306286570Smav */ 1307286570Smav (void) buf_hash_insert(nhdr, NULL); 1308286570Smav 1309286570Smav ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 1310286570Smav 1311286570Smav mutex_enter(&dev->l2ad_mtx); 1312286570Smav 1313286570Smav /* 1314286570Smav * We must place the realloc'ed header back into the list at 1315286570Smav * the same spot. Otherwise, if it's placed earlier in the list, 1316286570Smav * l2arc_write_buffers() could find it during the function's 1317286570Smav * write phase, and try to write it out to the l2arc. 1318286570Smav */ 1319286570Smav list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 1320286570Smav list_remove(&dev->l2ad_buflist, hdr); 1321286570Smav 1322286570Smav mutex_exit(&dev->l2ad_mtx); 1323286570Smav 1324286570Smav buf_discard_identity(hdr); 1325286570Smav hdr->b_freeze_cksum = NULL; 1326286570Smav kmem_cache_free(old, hdr); 1327286570Smav 1328286570Smav return (nhdr); 1329286570Smav} 1330286570Smav 1331286570Smav 1332168404Spjd#define ARC_MINTIME (hz>>4) /* 62 ms */ 1333168404Spjd 1334168404Spjdstatic void 1335168404Spjdarc_cksum_verify(arc_buf_t *buf) 1336168404Spjd{ 1337168404Spjd zio_cksum_t zc; 1338168404Spjd 1339168404Spjd if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1340168404Spjd return; 1341168404Spjd 1342286570Smav mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1343286570Smav if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) { 1344286570Smav mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1345168404Spjd return; 1346168404Spjd } 1347168404Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1348168404Spjd if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1349168404Spjd panic("buffer modified while frozen!"); 1350286570Smav mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1351168404Spjd} 1352168404Spjd 1353185029Spjdstatic int 1354185029Spjdarc_cksum_equal(arc_buf_t *buf) 1355185029Spjd{ 1356185029Spjd zio_cksum_t zc; 1357185029Spjd int equal; 1358185029Spjd 1359286570Smav mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1360185029Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1361185029Spjd equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1362286570Smav mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1363185029Spjd 1364185029Spjd return (equal); 1365185029Spjd} 1366185029Spjd 1367168404Spjdstatic void 1368185029Spjdarc_cksum_compute(arc_buf_t *buf, boolean_t force) 1369168404Spjd{ 1370185029Spjd if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1371168404Spjd return; 1372168404Spjd 1373286570Smav mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1374168404Spjd if (buf->b_hdr->b_freeze_cksum != NULL) { 1375286570Smav mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1376168404Spjd return; 1377168404Spjd } 1378168404Spjd buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1379168404Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1380168404Spjd buf->b_hdr->b_freeze_cksum); 1381286570Smav mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1382240133Smm#ifdef illumos 1383240133Smm arc_buf_watch(buf); 1384277300Ssmh#endif 1385168404Spjd} 1386168404Spjd 1387240133Smm#ifdef illumos 1388240133Smm#ifndef _KERNEL 1389240133Smmtypedef struct procctl { 1390240133Smm long cmd; 1391240133Smm prwatch_t prwatch; 1392240133Smm} procctl_t; 1393240133Smm#endif 1394240133Smm 1395240133Smm/* ARGSUSED */ 1396240133Smmstatic void 1397240133Smmarc_buf_unwatch(arc_buf_t *buf) 1398240133Smm{ 1399240133Smm#ifndef _KERNEL 1400240133Smm if (arc_watch) { 1401240133Smm int result; 1402240133Smm procctl_t ctl; 1403240133Smm ctl.cmd = PCWATCH; 1404240133Smm ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1405240133Smm ctl.prwatch.pr_size = 0; 1406240133Smm ctl.prwatch.pr_wflags = 0; 1407240133Smm result = write(arc_procfd, &ctl, sizeof (ctl)); 1408240133Smm ASSERT3U(result, ==, sizeof (ctl)); 1409240133Smm } 1410240133Smm#endif 1411240133Smm} 1412240133Smm 1413240133Smm/* ARGSUSED */ 1414240133Smmstatic void 1415240133Smmarc_buf_watch(arc_buf_t *buf) 1416240133Smm{ 1417240133Smm#ifndef _KERNEL 1418240133Smm if (arc_watch) { 1419240133Smm int result; 1420240133Smm procctl_t ctl; 1421240133Smm ctl.cmd = PCWATCH; 1422240133Smm ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1423240133Smm ctl.prwatch.pr_size = buf->b_hdr->b_size; 1424240133Smm ctl.prwatch.pr_wflags = WA_WRITE; 1425240133Smm result = write(arc_procfd, &ctl, sizeof (ctl)); 1426240133Smm ASSERT3U(result, ==, sizeof (ctl)); 1427240133Smm } 1428240133Smm#endif 1429240133Smm} 1430240133Smm#endif /* illumos */ 1431240133Smm 1432286570Smavstatic arc_buf_contents_t 1433286570Smavarc_buf_type(arc_buf_hdr_t *hdr) 1434286570Smav{ 1435286570Smav if (HDR_ISTYPE_METADATA(hdr)) { 1436286570Smav return (ARC_BUFC_METADATA); 1437286570Smav } else { 1438286570Smav return (ARC_BUFC_DATA); 1439286570Smav } 1440286570Smav} 1441286570Smav 1442286570Smavstatic uint32_t 1443286570Smavarc_bufc_to_flags(arc_buf_contents_t type) 1444286570Smav{ 1445286570Smav switch (type) { 1446286570Smav case ARC_BUFC_DATA: 1447286570Smav /* metadata field is 0 if buffer contains normal data */ 1448286570Smav return (0); 1449286570Smav case ARC_BUFC_METADATA: 1450286570Smav return (ARC_FLAG_BUFC_METADATA); 1451286570Smav default: 1452286570Smav break; 1453286570Smav } 1454286570Smav panic("undefined ARC buffer type!"); 1455286570Smav return ((uint32_t)-1); 1456286570Smav} 1457286570Smav 1458168404Spjdvoid 1459168404Spjdarc_buf_thaw(arc_buf_t *buf) 1460168404Spjd{ 1461185029Spjd if (zfs_flags & ZFS_DEBUG_MODIFY) { 1462286570Smav if (buf->b_hdr->b_l1hdr.b_state != arc_anon) 1463185029Spjd panic("modifying non-anon buffer!"); 1464286570Smav if (HDR_IO_IN_PROGRESS(buf->b_hdr)) 1465185029Spjd panic("modifying buffer while i/o in progress!"); 1466185029Spjd arc_cksum_verify(buf); 1467185029Spjd } 1468168404Spjd 1469286570Smav mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1470168404Spjd if (buf->b_hdr->b_freeze_cksum != NULL) { 1471168404Spjd kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1472168404Spjd buf->b_hdr->b_freeze_cksum = NULL; 1473168404Spjd } 1474219089Spjd 1475286570Smav#ifdef ZFS_DEBUG 1476219089Spjd if (zfs_flags & ZFS_DEBUG_MODIFY) { 1477286570Smav if (buf->b_hdr->b_l1hdr.b_thawed != NULL) 1478286570Smav kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1); 1479286570Smav buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 1480219089Spjd } 1481286570Smav#endif 1482219089Spjd 1483286570Smav mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1484240133Smm 1485240133Smm#ifdef illumos 1486240133Smm arc_buf_unwatch(buf); 1487277300Ssmh#endif 1488168404Spjd} 1489168404Spjd 1490168404Spjdvoid 1491168404Spjdarc_buf_freeze(arc_buf_t *buf) 1492168404Spjd{ 1493219089Spjd kmutex_t *hash_lock; 1494219089Spjd 1495168404Spjd if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1496168404Spjd return; 1497168404Spjd 1498219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 1499219089Spjd mutex_enter(hash_lock); 1500219089Spjd 1501168404Spjd ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1502286570Smav buf->b_hdr->b_l1hdr.b_state == arc_anon); 1503185029Spjd arc_cksum_compute(buf, B_FALSE); 1504219089Spjd mutex_exit(hash_lock); 1505240133Smm 1506168404Spjd} 1507168404Spjd 1508168404Spjdstatic void 1509275811Sdelphijget_buf_info(arc_buf_hdr_t *hdr, arc_state_t *state, list_t **list, kmutex_t **lock) 1510205231Skmacy{ 1511275811Sdelphij uint64_t buf_hashid = buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1512205231Skmacy 1513286570Smav if (arc_buf_type(hdr) == ARC_BUFC_METADATA) 1514206796Spjd buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1515205231Skmacy else { 1516206796Spjd buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1517205231Skmacy buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1518205231Skmacy } 1519205231Skmacy 1520205231Skmacy *list = &state->arcs_lists[buf_hashid]; 1521205231Skmacy *lock = ARCS_LOCK(state, buf_hashid); 1522205231Skmacy} 1523205231Skmacy 1524205231Skmacy 1525205231Skmacystatic void 1526275811Sdelphijadd_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1527168404Spjd{ 1528286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 1529168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 1530286570Smav arc_state_t *state = hdr->b_l1hdr.b_state; 1531168404Spjd 1532286570Smav if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 1533286570Smav (state != arc_anon)) { 1534286570Smav /* We don't use the L2-only state list. */ 1535286570Smav if (state != arc_l2c_only) { 1536286570Smav uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt; 1537286570Smav uint64_t *size = &state->arcs_lsize[arc_buf_type(hdr)]; 1538286570Smav list_t *list; 1539286570Smav kmutex_t *lock; 1540168404Spjd 1541286570Smav get_buf_info(hdr, state, &list, &lock); 1542286570Smav ASSERT(!MUTEX_HELD(lock)); 1543286570Smav mutex_enter(lock); 1544286570Smav ASSERT(list_link_active(&hdr->b_l1hdr.b_arc_node)); 1545286570Smav list_remove(list, hdr); 1546286570Smav if (GHOST_STATE(state)) { 1547286570Smav ASSERT0(hdr->b_l1hdr.b_datacnt); 1548286570Smav ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 1549286570Smav delta = hdr->b_size; 1550286570Smav } 1551286570Smav ASSERT(delta > 0); 1552286570Smav ASSERT3U(*size, >=, delta); 1553286570Smav atomic_add_64(size, -delta); 1554286570Smav mutex_exit(lock); 1555168404Spjd } 1556185029Spjd /* remove the prefetch flag if we get a reference */ 1557286570Smav hdr->b_flags &= ~ARC_FLAG_PREFETCH; 1558168404Spjd } 1559168404Spjd} 1560168404Spjd 1561168404Spjdstatic int 1562275811Sdelphijremove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1563168404Spjd{ 1564168404Spjd int cnt; 1565286570Smav arc_state_t *state = hdr->b_l1hdr.b_state; 1566168404Spjd 1567286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 1568168404Spjd ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1569168404Spjd ASSERT(!GHOST_STATE(state)); 1570168404Spjd 1571286570Smav /* 1572286570Smav * arc_l2c_only counts as a ghost state so we don't need to explicitly 1573286570Smav * check to prevent usage of the arc_l2c_only list. 1574286570Smav */ 1575286570Smav if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 1576168404Spjd (state != arc_anon)) { 1577286570Smav uint64_t *size = &state->arcs_lsize[arc_buf_type(hdr)]; 1578205231Skmacy list_t *list; 1579205231Skmacy kmutex_t *lock; 1580185029Spjd 1581275811Sdelphij get_buf_info(hdr, state, &list, &lock); 1582205231Skmacy ASSERT(!MUTEX_HELD(lock)); 1583205231Skmacy mutex_enter(lock); 1584286570Smav ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 1585275811Sdelphij list_insert_head(list, hdr); 1586286570Smav ASSERT(hdr->b_l1hdr.b_datacnt > 0); 1587286570Smav atomic_add_64(size, hdr->b_size * 1588286570Smav hdr->b_l1hdr.b_datacnt); 1589206794Spjd mutex_exit(lock); 1590168404Spjd } 1591168404Spjd return (cnt); 1592168404Spjd} 1593168404Spjd 1594168404Spjd/* 1595168404Spjd * Move the supplied buffer to the indicated state. The mutex 1596168404Spjd * for the buffer must be held by the caller. 1597168404Spjd */ 1598168404Spjdstatic void 1599275811Sdelphijarc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 1600275811Sdelphij kmutex_t *hash_lock) 1601168404Spjd{ 1602286570Smav arc_state_t *old_state; 1603286570Smav int64_t refcnt; 1604286570Smav uint32_t datacnt; 1605168404Spjd uint64_t from_delta, to_delta; 1606286570Smav arc_buf_contents_t buftype = arc_buf_type(hdr); 1607205231Skmacy list_t *list; 1608205231Skmacy kmutex_t *lock; 1609168404Spjd 1610286570Smav /* 1611286570Smav * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 1612286570Smav * in arc_read() when bringing a buffer out of the L2ARC. However, the 1613286570Smav * L1 hdr doesn't always exist when we change state to arc_anon before 1614286570Smav * destroying a header, in which case reallocating to add the L1 hdr is 1615286570Smav * pointless. 1616286570Smav */ 1617286570Smav if (HDR_HAS_L1HDR(hdr)) { 1618286570Smav old_state = hdr->b_l1hdr.b_state; 1619286570Smav refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 1620286570Smav datacnt = hdr->b_l1hdr.b_datacnt; 1621286570Smav } else { 1622286570Smav old_state = arc_l2c_only; 1623286570Smav refcnt = 0; 1624286570Smav datacnt = 0; 1625286570Smav } 1626286570Smav 1627168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 1628258632Savg ASSERT3P(new_state, !=, old_state); 1629286570Smav ASSERT(refcnt == 0 || datacnt > 0); 1630286570Smav ASSERT(!GHOST_STATE(new_state) || datacnt == 0); 1631286570Smav ASSERT(old_state != arc_anon || datacnt <= 1); 1632168404Spjd 1633286570Smav from_delta = to_delta = datacnt * hdr->b_size; 1634168404Spjd 1635168404Spjd /* 1636168404Spjd * If this buffer is evictable, transfer it from the 1637168404Spjd * old state list to the new state list. 1638168404Spjd */ 1639168404Spjd if (refcnt == 0) { 1640286570Smav if (old_state != arc_anon && old_state != arc_l2c_only) { 1641205231Skmacy int use_mutex; 1642286570Smav uint64_t *size = &old_state->arcs_lsize[buftype]; 1643168404Spjd 1644275811Sdelphij get_buf_info(hdr, old_state, &list, &lock); 1645205231Skmacy use_mutex = !MUTEX_HELD(lock); 1646168404Spjd if (use_mutex) 1647205231Skmacy mutex_enter(lock); 1648168404Spjd 1649286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 1650286570Smav ASSERT(list_link_active(&hdr->b_l1hdr.b_arc_node)); 1651275811Sdelphij list_remove(list, hdr); 1652168404Spjd 1653168404Spjd /* 1654168404Spjd * If prefetching out of the ghost cache, 1655219089Spjd * we will have a non-zero datacnt. 1656168404Spjd */ 1657286570Smav if (GHOST_STATE(old_state) && datacnt == 0) { 1658168404Spjd /* ghost elements have a ghost size */ 1659286570Smav ASSERT(hdr->b_l1hdr.b_buf == NULL); 1660275811Sdelphij from_delta = hdr->b_size; 1661168404Spjd } 1662185029Spjd ASSERT3U(*size, >=, from_delta); 1663185029Spjd atomic_add_64(size, -from_delta); 1664168404Spjd 1665168404Spjd if (use_mutex) 1666205231Skmacy mutex_exit(lock); 1667168404Spjd } 1668286570Smav if (new_state != arc_anon && new_state != arc_l2c_only) { 1669206796Spjd int use_mutex; 1670286570Smav uint64_t *size = &new_state->arcs_lsize[buftype]; 1671168404Spjd 1672286570Smav /* 1673286570Smav * An L1 header always exists here, since if we're 1674286570Smav * moving to some L1-cached state (i.e. not l2c_only or 1675286570Smav * anonymous), we realloc the header to add an L1hdr 1676286570Smav * beforehand. 1677286570Smav */ 1678286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 1679275811Sdelphij get_buf_info(hdr, new_state, &list, &lock); 1680205231Skmacy use_mutex = !MUTEX_HELD(lock); 1681168404Spjd if (use_mutex) 1682205231Skmacy mutex_enter(lock); 1683168404Spjd 1684275811Sdelphij list_insert_head(list, hdr); 1685168404Spjd 1686168404Spjd /* ghost elements have a ghost size */ 1687168404Spjd if (GHOST_STATE(new_state)) { 1688286570Smav ASSERT(datacnt == 0); 1689286570Smav ASSERT(hdr->b_l1hdr.b_buf == NULL); 1690275811Sdelphij to_delta = hdr->b_size; 1691168404Spjd } 1692185029Spjd atomic_add_64(size, to_delta); 1693168404Spjd 1694168404Spjd if (use_mutex) 1695205231Skmacy mutex_exit(lock); 1696168404Spjd } 1697168404Spjd } 1698168404Spjd 1699275811Sdelphij ASSERT(!BUF_EMPTY(hdr)); 1700275811Sdelphij if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 1701275811Sdelphij buf_hash_remove(hdr); 1702168404Spjd 1703286570Smav /* adjust state sizes (ignore arc_l2c_only) */ 1704286570Smav if (to_delta && new_state != arc_l2c_only) 1705168404Spjd atomic_add_64(&new_state->arcs_size, to_delta); 1706286570Smav if (from_delta && old_state != arc_l2c_only) { 1707168404Spjd ASSERT3U(old_state->arcs_size, >=, from_delta); 1708168404Spjd atomic_add_64(&old_state->arcs_size, -from_delta); 1709168404Spjd } 1710286570Smav if (HDR_HAS_L1HDR(hdr)) 1711286570Smav hdr->b_l1hdr.b_state = new_state; 1712185029Spjd 1713286570Smav /* 1714286570Smav * L2 headers should never be on the L2 state list since they don't 1715286570Smav * have L1 headers allocated. 1716286570Smav */ 1717286570Smav ASSERT(list_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 1718286570Smav list_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 1719168404Spjd} 1720168404Spjd 1721185029Spjdvoid 1722208373Smmarc_space_consume(uint64_t space, arc_space_type_t type) 1723185029Spjd{ 1724208373Smm ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1725208373Smm 1726208373Smm switch (type) { 1727208373Smm case ARC_SPACE_DATA: 1728208373Smm ARCSTAT_INCR(arcstat_data_size, space); 1729208373Smm break; 1730208373Smm case ARC_SPACE_OTHER: 1731208373Smm ARCSTAT_INCR(arcstat_other_size, space); 1732208373Smm break; 1733208373Smm case ARC_SPACE_HDRS: 1734208373Smm ARCSTAT_INCR(arcstat_hdr_size, space); 1735208373Smm break; 1736208373Smm case ARC_SPACE_L2HDRS: 1737208373Smm ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1738208373Smm break; 1739208373Smm } 1740208373Smm 1741275748Sdelphij ARCSTAT_INCR(arcstat_meta_used, space); 1742185029Spjd atomic_add_64(&arc_size, space); 1743185029Spjd} 1744185029Spjd 1745185029Spjdvoid 1746208373Smmarc_space_return(uint64_t space, arc_space_type_t type) 1747185029Spjd{ 1748208373Smm ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1749208373Smm 1750208373Smm switch (type) { 1751208373Smm case ARC_SPACE_DATA: 1752208373Smm ARCSTAT_INCR(arcstat_data_size, -space); 1753208373Smm break; 1754208373Smm case ARC_SPACE_OTHER: 1755208373Smm ARCSTAT_INCR(arcstat_other_size, -space); 1756208373Smm break; 1757208373Smm case ARC_SPACE_HDRS: 1758208373Smm ARCSTAT_INCR(arcstat_hdr_size, -space); 1759208373Smm break; 1760208373Smm case ARC_SPACE_L2HDRS: 1761208373Smm ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1762208373Smm break; 1763208373Smm } 1764208373Smm 1765185029Spjd ASSERT(arc_meta_used >= space); 1766185029Spjd if (arc_meta_max < arc_meta_used) 1767185029Spjd arc_meta_max = arc_meta_used; 1768275748Sdelphij ARCSTAT_INCR(arcstat_meta_used, -space); 1769185029Spjd ASSERT(arc_size >= space); 1770185029Spjd atomic_add_64(&arc_size, -space); 1771185029Spjd} 1772185029Spjd 1773168404Spjdarc_buf_t * 1774286570Smavarc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type) 1775168404Spjd{ 1776168404Spjd arc_buf_hdr_t *hdr; 1777168404Spjd arc_buf_t *buf; 1778168404Spjd 1779168404Spjd ASSERT3U(size, >, 0); 1780286570Smav hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 1781168404Spjd ASSERT(BUF_EMPTY(hdr)); 1782286570Smav ASSERT3P(hdr->b_freeze_cksum, ==, NULL); 1783168404Spjd hdr->b_size = size; 1784228103Smm hdr->b_spa = spa_load_guid(spa); 1785286570Smav 1786185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1787168404Spjd buf->b_hdr = hdr; 1788168404Spjd buf->b_data = NULL; 1789168404Spjd buf->b_efunc = NULL; 1790168404Spjd buf->b_private = NULL; 1791168404Spjd buf->b_next = NULL; 1792286570Smav 1793286570Smav hdr->b_flags = arc_bufc_to_flags(type); 1794286570Smav hdr->b_flags |= ARC_FLAG_HAS_L1HDR; 1795286570Smav 1796286570Smav hdr->b_l1hdr.b_buf = buf; 1797286570Smav hdr->b_l1hdr.b_state = arc_anon; 1798286570Smav hdr->b_l1hdr.b_arc_access = 0; 1799286570Smav hdr->b_l1hdr.b_datacnt = 1; 1800286570Smav 1801168404Spjd arc_get_data_buf(buf); 1802286570Smav ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 1803286570Smav (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 1804168404Spjd 1805168404Spjd return (buf); 1806168404Spjd} 1807168404Spjd 1808209962Smmstatic char *arc_onloan_tag = "onloan"; 1809209962Smm 1810209962Smm/* 1811209962Smm * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1812209962Smm * flight data by arc_tempreserve_space() until they are "returned". Loaned 1813209962Smm * buffers must be returned to the arc before they can be used by the DMU or 1814209962Smm * freed. 1815209962Smm */ 1816209962Smmarc_buf_t * 1817209962Smmarc_loan_buf(spa_t *spa, int size) 1818209962Smm{ 1819209962Smm arc_buf_t *buf; 1820209962Smm 1821209962Smm buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1822209962Smm 1823209962Smm atomic_add_64(&arc_loaned_bytes, size); 1824209962Smm return (buf); 1825209962Smm} 1826209962Smm 1827209962Smm/* 1828209962Smm * Return a loaned arc buffer to the arc. 1829209962Smm */ 1830209962Smmvoid 1831209962Smmarc_return_buf(arc_buf_t *buf, void *tag) 1832209962Smm{ 1833209962Smm arc_buf_hdr_t *hdr = buf->b_hdr; 1834209962Smm 1835209962Smm ASSERT(buf->b_data != NULL); 1836286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 1837286570Smav (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 1838286570Smav (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 1839209962Smm 1840209962Smm atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1841209962Smm} 1842209962Smm 1843219089Spjd/* Detach an arc_buf from a dbuf (tag) */ 1844219089Spjdvoid 1845219089Spjdarc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1846219089Spjd{ 1847286570Smav arc_buf_hdr_t *hdr = buf->b_hdr; 1848219089Spjd 1849219089Spjd ASSERT(buf->b_data != NULL); 1850286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 1851286570Smav (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 1852286570Smav (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 1853219089Spjd buf->b_efunc = NULL; 1854219089Spjd buf->b_private = NULL; 1855219089Spjd 1856219089Spjd atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1857219089Spjd} 1858219089Spjd 1859168404Spjdstatic arc_buf_t * 1860168404Spjdarc_buf_clone(arc_buf_t *from) 1861168404Spjd{ 1862168404Spjd arc_buf_t *buf; 1863168404Spjd arc_buf_hdr_t *hdr = from->b_hdr; 1864168404Spjd uint64_t size = hdr->b_size; 1865168404Spjd 1866286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 1867286570Smav ASSERT(hdr->b_l1hdr.b_state != arc_anon); 1868219089Spjd 1869185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1870168404Spjd buf->b_hdr = hdr; 1871168404Spjd buf->b_data = NULL; 1872168404Spjd buf->b_efunc = NULL; 1873168404Spjd buf->b_private = NULL; 1874286570Smav buf->b_next = hdr->b_l1hdr.b_buf; 1875286570Smav hdr->b_l1hdr.b_buf = buf; 1876168404Spjd arc_get_data_buf(buf); 1877168404Spjd bcopy(from->b_data, buf->b_data, size); 1878242845Sdelphij 1879242845Sdelphij /* 1880242845Sdelphij * This buffer already exists in the arc so create a duplicate 1881242845Sdelphij * copy for the caller. If the buffer is associated with user data 1882242845Sdelphij * then track the size and number of duplicates. These stats will be 1883242845Sdelphij * updated as duplicate buffers are created and destroyed. 1884242845Sdelphij */ 1885286570Smav if (HDR_ISTYPE_DATA(hdr)) { 1886242845Sdelphij ARCSTAT_BUMP(arcstat_duplicate_buffers); 1887242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1888242845Sdelphij } 1889286570Smav hdr->b_l1hdr.b_datacnt += 1; 1890168404Spjd return (buf); 1891168404Spjd} 1892168404Spjd 1893168404Spjdvoid 1894168404Spjdarc_buf_add_ref(arc_buf_t *buf, void* tag) 1895168404Spjd{ 1896168404Spjd arc_buf_hdr_t *hdr; 1897168404Spjd kmutex_t *hash_lock; 1898168404Spjd 1899168404Spjd /* 1900185029Spjd * Check to see if this buffer is evicted. Callers 1901185029Spjd * must verify b_data != NULL to know if the add_ref 1902185029Spjd * was successful. 1903168404Spjd */ 1904219089Spjd mutex_enter(&buf->b_evict_lock); 1905185029Spjd if (buf->b_data == NULL) { 1906219089Spjd mutex_exit(&buf->b_evict_lock); 1907168404Spjd return; 1908168404Spjd } 1909219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 1910219089Spjd mutex_enter(hash_lock); 1911185029Spjd hdr = buf->b_hdr; 1912286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 1913219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1914219089Spjd mutex_exit(&buf->b_evict_lock); 1915168404Spjd 1916286570Smav ASSERT(hdr->b_l1hdr.b_state == arc_mru || 1917286570Smav hdr->b_l1hdr.b_state == arc_mfu); 1918286570Smav 1919168404Spjd add_reference(hdr, hash_lock, tag); 1920208373Smm DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1921168404Spjd arc_access(hdr, hash_lock); 1922168404Spjd mutex_exit(hash_lock); 1923168404Spjd ARCSTAT_BUMP(arcstat_hits); 1924286570Smav ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 1925286570Smav demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 1926168404Spjd data, metadata, hits); 1927168404Spjd} 1928168404Spjd 1929274172Savgstatic void 1930274172Savgarc_buf_free_on_write(void *data, size_t size, 1931274172Savg void (*free_func)(void *, size_t)) 1932274172Savg{ 1933274172Savg l2arc_data_free_t *df; 1934274172Savg 1935274172Savg df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1936274172Savg df->l2df_data = data; 1937274172Savg df->l2df_size = size; 1938274172Savg df->l2df_func = free_func; 1939274172Savg mutex_enter(&l2arc_free_on_write_mtx); 1940274172Savg list_insert_head(l2arc_free_on_write, df); 1941274172Savg mutex_exit(&l2arc_free_on_write_mtx); 1942274172Savg} 1943274172Savg 1944185029Spjd/* 1945185029Spjd * Free the arc data buffer. If it is an l2arc write in progress, 1946185029Spjd * the buffer is placed on l2arc_free_on_write to be freed later. 1947185029Spjd */ 1948168404Spjdstatic void 1949240133Smmarc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1950185029Spjd{ 1951240133Smm arc_buf_hdr_t *hdr = buf->b_hdr; 1952240133Smm 1953185029Spjd if (HDR_L2_WRITING(hdr)) { 1954274172Savg arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 1955185029Spjd ARCSTAT_BUMP(arcstat_l2_free_on_write); 1956185029Spjd } else { 1957240133Smm free_func(buf->b_data, hdr->b_size); 1958185029Spjd } 1959185029Spjd} 1960185029Spjd 1961268858Sdelphij/* 1962268858Sdelphij * Free up buf->b_data and if 'remove' is set, then pull the 1963268858Sdelphij * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 1964268858Sdelphij */ 1965185029Spjdstatic void 1966274172Savgarc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 1967274172Savg{ 1968286570Smav ASSERT(HDR_HAS_L2HDR(hdr)); 1969286570Smav ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx)); 1970274172Savg 1971286570Smav /* 1972286570Smav * The b_tmp_cdata field is linked off of the b_l1hdr, so if 1973286570Smav * that doesn't exist, the header is in the arc_l2c_only state, 1974286570Smav * and there isn't anything to free (it's already been freed). 1975286570Smav */ 1976286570Smav if (!HDR_HAS_L1HDR(hdr)) 1977286570Smav return; 1978274172Savg 1979286570Smav if (hdr->b_l1hdr.b_tmp_cdata == NULL) 1980274172Savg return; 1981274172Savg 1982274172Savg ASSERT(HDR_L2_WRITING(hdr)); 1983286570Smav arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata, hdr->b_size, 1984274172Savg zio_data_buf_free); 1985286570Smav 1986274172Savg ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 1987286570Smav hdr->b_l1hdr.b_tmp_cdata = NULL; 1988274172Savg} 1989274172Savg 1990274172Savgstatic void 1991268858Sdelphijarc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 1992168404Spjd{ 1993168404Spjd arc_buf_t **bufp; 1994168404Spjd 1995168404Spjd /* free up data associated with the buf */ 1996286570Smav if (buf->b_data != NULL) { 1997286570Smav arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 1998168404Spjd uint64_t size = buf->b_hdr->b_size; 1999286570Smav arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 2000168404Spjd 2001168404Spjd arc_cksum_verify(buf); 2002240133Smm#ifdef illumos 2003240133Smm arc_buf_unwatch(buf); 2004277300Ssmh#endif 2005219089Spjd 2006168404Spjd if (!recycle) { 2007168404Spjd if (type == ARC_BUFC_METADATA) { 2008240133Smm arc_buf_data_free(buf, zio_buf_free); 2009208373Smm arc_space_return(size, ARC_SPACE_DATA); 2010168404Spjd } else { 2011168404Spjd ASSERT(type == ARC_BUFC_DATA); 2012240133Smm arc_buf_data_free(buf, zio_data_buf_free); 2013208373Smm ARCSTAT_INCR(arcstat_data_size, -size); 2014185029Spjd atomic_add_64(&arc_size, -size); 2015168404Spjd } 2016168404Spjd } 2017286570Smav if (list_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) { 2018185029Spjd uint64_t *cnt = &state->arcs_lsize[type]; 2019185029Spjd 2020286570Smav ASSERT(refcount_is_zero( 2021286570Smav &buf->b_hdr->b_l1hdr.b_refcnt)); 2022286570Smav ASSERT(state != arc_anon && state != arc_l2c_only); 2023185029Spjd 2024185029Spjd ASSERT3U(*cnt, >=, size); 2025185029Spjd atomic_add_64(cnt, -size); 2026168404Spjd } 2027168404Spjd ASSERT3U(state->arcs_size, >=, size); 2028168404Spjd atomic_add_64(&state->arcs_size, -size); 2029168404Spjd buf->b_data = NULL; 2030242845Sdelphij 2031242845Sdelphij /* 2032242845Sdelphij * If we're destroying a duplicate buffer make sure 2033242845Sdelphij * that the appropriate statistics are updated. 2034242845Sdelphij */ 2035286570Smav if (buf->b_hdr->b_l1hdr.b_datacnt > 1 && 2036286570Smav HDR_ISTYPE_DATA(buf->b_hdr)) { 2037242845Sdelphij ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 2038242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 2039242845Sdelphij } 2040286570Smav ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0); 2041286570Smav buf->b_hdr->b_l1hdr.b_datacnt -= 1; 2042168404Spjd } 2043168404Spjd 2044168404Spjd /* only remove the buf if requested */ 2045268858Sdelphij if (!remove) 2046168404Spjd return; 2047168404Spjd 2048168404Spjd /* remove the buf from the hdr list */ 2049286570Smav for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf; 2050286570Smav bufp = &(*bufp)->b_next) 2051168404Spjd continue; 2052168404Spjd *bufp = buf->b_next; 2053219089Spjd buf->b_next = NULL; 2054168404Spjd 2055168404Spjd ASSERT(buf->b_efunc == NULL); 2056168404Spjd 2057168404Spjd /* clean up the buf */ 2058168404Spjd buf->b_hdr = NULL; 2059168404Spjd kmem_cache_free(buf_cache, buf); 2060168404Spjd} 2061168404Spjd 2062168404Spjdstatic void 2063168404Spjdarc_hdr_destroy(arc_buf_hdr_t *hdr) 2064168404Spjd{ 2065286570Smav if (HDR_HAS_L1HDR(hdr)) { 2066286570Smav ASSERT(hdr->b_l1hdr.b_buf == NULL || 2067286570Smav hdr->b_l1hdr.b_datacnt > 0); 2068286570Smav ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2069286570Smav ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 2070286570Smav } 2071168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2072286570Smav ASSERT(!HDR_IN_HASH_TABLE(hdr)); 2073168404Spjd 2074286570Smav if (HDR_HAS_L2HDR(hdr)) { 2075286570Smav l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 2076286570Smav boolean_t buflist_held = MUTEX_HELD(&l2hdr->b_dev->l2ad_mtx); 2077286570Smav 2078219089Spjd if (!buflist_held) { 2079286570Smav mutex_enter(&l2hdr->b_dev->l2ad_mtx); 2080286570Smav l2hdr = &hdr->b_l2hdr; 2081219089Spjd } 2082219089Spjd 2083286570Smav trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 2084286570Smav l2hdr->b_asize, 0); 2085286570Smav list_remove(&l2hdr->b_dev->l2ad_buflist, hdr); 2086219089Spjd 2087286570Smav /* 2088286570Smav * We don't want to leak the b_tmp_cdata buffer that was 2089286570Smav * allocated in l2arc_write_buffers() 2090286570Smav */ 2091286570Smav arc_buf_l2_cdata_free(hdr); 2092286570Smav 2093286570Smav ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 2094286570Smav ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 2095286570Smav 2096219089Spjd if (!buflist_held) 2097286570Smav mutex_exit(&l2hdr->b_dev->l2ad_mtx); 2098286570Smav 2099286570Smav hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 2100185029Spjd } 2101185029Spjd 2102286570Smav if (!BUF_EMPTY(hdr)) 2103219089Spjd buf_discard_identity(hdr); 2104168404Spjd if (hdr->b_freeze_cksum != NULL) { 2105168404Spjd kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 2106168404Spjd hdr->b_freeze_cksum = NULL; 2107168404Spjd } 2108286570Smav 2109286570Smav if (HDR_HAS_L1HDR(hdr)) { 2110286570Smav while (hdr->b_l1hdr.b_buf) { 2111286570Smav arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2112286570Smav 2113286570Smav if (buf->b_efunc != NULL) { 2114286570Smav mutex_enter(&arc_eviction_mtx); 2115286570Smav mutex_enter(&buf->b_evict_lock); 2116286570Smav ASSERT(buf->b_hdr != NULL); 2117286570Smav arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE, 2118286570Smav FALSE); 2119286570Smav hdr->b_l1hdr.b_buf = buf->b_next; 2120286570Smav buf->b_hdr = &arc_eviction_hdr; 2121286570Smav buf->b_next = arc_eviction_list; 2122286570Smav arc_eviction_list = buf; 2123286570Smav mutex_exit(&buf->b_evict_lock); 2124286570Smav mutex_exit(&arc_eviction_mtx); 2125286570Smav } else { 2126286570Smav arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE, 2127286570Smav TRUE); 2128286570Smav } 2129286570Smav } 2130286570Smav#ifdef ZFS_DEBUG 2131286570Smav if (hdr->b_l1hdr.b_thawed != NULL) { 2132286570Smav kmem_free(hdr->b_l1hdr.b_thawed, 1); 2133286570Smav hdr->b_l1hdr.b_thawed = NULL; 2134286570Smav } 2135286570Smav#endif 2136219089Spjd } 2137168404Spjd 2138168404Spjd ASSERT3P(hdr->b_hash_next, ==, NULL); 2139286570Smav if (HDR_HAS_L1HDR(hdr)) { 2140286570Smav ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 2141286570Smav ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 2142286570Smav kmem_cache_free(hdr_full_cache, hdr); 2143286570Smav } else { 2144286570Smav kmem_cache_free(hdr_l2only_cache, hdr); 2145286570Smav } 2146168404Spjd} 2147168404Spjd 2148168404Spjdvoid 2149168404Spjdarc_buf_free(arc_buf_t *buf, void *tag) 2150168404Spjd{ 2151168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 2152286570Smav int hashed = hdr->b_l1hdr.b_state != arc_anon; 2153168404Spjd 2154168404Spjd ASSERT(buf->b_efunc == NULL); 2155168404Spjd ASSERT(buf->b_data != NULL); 2156168404Spjd 2157168404Spjd if (hashed) { 2158168404Spjd kmutex_t *hash_lock = HDR_LOCK(hdr); 2159168404Spjd 2160168404Spjd mutex_enter(hash_lock); 2161219089Spjd hdr = buf->b_hdr; 2162219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2163219089Spjd 2164168404Spjd (void) remove_reference(hdr, hash_lock, tag); 2165286570Smav if (hdr->b_l1hdr.b_datacnt > 1) { 2166168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 2167219089Spjd } else { 2168286570Smav ASSERT(buf == hdr->b_l1hdr.b_buf); 2169219089Spjd ASSERT(buf->b_efunc == NULL); 2170275811Sdelphij hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2171219089Spjd } 2172168404Spjd mutex_exit(hash_lock); 2173168404Spjd } else if (HDR_IO_IN_PROGRESS(hdr)) { 2174168404Spjd int destroy_hdr; 2175168404Spjd /* 2176168404Spjd * We are in the middle of an async write. Don't destroy 2177168404Spjd * this buffer unless the write completes before we finish 2178168404Spjd * decrementing the reference count. 2179168404Spjd */ 2180168404Spjd mutex_enter(&arc_eviction_mtx); 2181168404Spjd (void) remove_reference(hdr, NULL, tag); 2182286570Smav ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2183168404Spjd destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 2184168404Spjd mutex_exit(&arc_eviction_mtx); 2185168404Spjd if (destroy_hdr) 2186168404Spjd arc_hdr_destroy(hdr); 2187168404Spjd } else { 2188219089Spjd if (remove_reference(hdr, NULL, tag) > 0) 2189168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 2190219089Spjd else 2191168404Spjd arc_hdr_destroy(hdr); 2192168404Spjd } 2193168404Spjd} 2194168404Spjd 2195248571Smmboolean_t 2196168404Spjdarc_buf_remove_ref(arc_buf_t *buf, void* tag) 2197168404Spjd{ 2198168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 2199168404Spjd kmutex_t *hash_lock = HDR_LOCK(hdr); 2200248571Smm boolean_t no_callback = (buf->b_efunc == NULL); 2201168404Spjd 2202286570Smav if (hdr->b_l1hdr.b_state == arc_anon) { 2203286570Smav ASSERT(hdr->b_l1hdr.b_datacnt == 1); 2204168404Spjd arc_buf_free(buf, tag); 2205168404Spjd return (no_callback); 2206168404Spjd } 2207168404Spjd 2208168404Spjd mutex_enter(hash_lock); 2209219089Spjd hdr = buf->b_hdr; 2210286570Smav ASSERT(hdr->b_l1hdr.b_datacnt > 0); 2211219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2212286570Smav ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2213168404Spjd ASSERT(buf->b_data != NULL); 2214168404Spjd 2215168404Spjd (void) remove_reference(hdr, hash_lock, tag); 2216286570Smav if (hdr->b_l1hdr.b_datacnt > 1) { 2217168404Spjd if (no_callback) 2218168404Spjd arc_buf_destroy(buf, FALSE, TRUE); 2219168404Spjd } else if (no_callback) { 2220286570Smav ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL); 2221219089Spjd ASSERT(buf->b_efunc == NULL); 2222275811Sdelphij hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2223168404Spjd } 2224286570Smav ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 || 2225286570Smav refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2226168404Spjd mutex_exit(hash_lock); 2227168404Spjd return (no_callback); 2228168404Spjd} 2229168404Spjd 2230286570Smavint32_t 2231168404Spjdarc_buf_size(arc_buf_t *buf) 2232168404Spjd{ 2233168404Spjd return (buf->b_hdr->b_size); 2234168404Spjd} 2235168404Spjd 2236168404Spjd/* 2237242845Sdelphij * Called from the DMU to determine if the current buffer should be 2238242845Sdelphij * evicted. In order to ensure proper locking, the eviction must be initiated 2239242845Sdelphij * from the DMU. Return true if the buffer is associated with user data and 2240242845Sdelphij * duplicate buffers still exist. 2241242845Sdelphij */ 2242242845Sdelphijboolean_t 2243242845Sdelphijarc_buf_eviction_needed(arc_buf_t *buf) 2244242845Sdelphij{ 2245242845Sdelphij arc_buf_hdr_t *hdr; 2246242845Sdelphij boolean_t evict_needed = B_FALSE; 2247242845Sdelphij 2248242845Sdelphij if (zfs_disable_dup_eviction) 2249242845Sdelphij return (B_FALSE); 2250242845Sdelphij 2251242845Sdelphij mutex_enter(&buf->b_evict_lock); 2252242845Sdelphij hdr = buf->b_hdr; 2253242845Sdelphij if (hdr == NULL) { 2254242845Sdelphij /* 2255242845Sdelphij * We are in arc_do_user_evicts(); let that function 2256242845Sdelphij * perform the eviction. 2257242845Sdelphij */ 2258242845Sdelphij ASSERT(buf->b_data == NULL); 2259242845Sdelphij mutex_exit(&buf->b_evict_lock); 2260242845Sdelphij return (B_FALSE); 2261242845Sdelphij } else if (buf->b_data == NULL) { 2262242845Sdelphij /* 2263242845Sdelphij * We have already been added to the arc eviction list; 2264242845Sdelphij * recommend eviction. 2265242845Sdelphij */ 2266242845Sdelphij ASSERT3P(hdr, ==, &arc_eviction_hdr); 2267242845Sdelphij mutex_exit(&buf->b_evict_lock); 2268242845Sdelphij return (B_TRUE); 2269242845Sdelphij } 2270242845Sdelphij 2271286570Smav if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr)) 2272242845Sdelphij evict_needed = B_TRUE; 2273242845Sdelphij 2274242845Sdelphij mutex_exit(&buf->b_evict_lock); 2275242845Sdelphij return (evict_needed); 2276242845Sdelphij} 2277242845Sdelphij 2278242845Sdelphij/* 2279168404Spjd * Evict buffers from list until we've removed the specified number of 2280168404Spjd * bytes. Move the removed buffers to the appropriate evict state. 2281168404Spjd * If the recycle flag is set, then attempt to "recycle" a buffer: 2282168404Spjd * - look for a buffer to evict that is `bytes' long. 2283168404Spjd * - return the data block from this buffer rather than freeing it. 2284168404Spjd * This flag is used by callers that are trying to make space for a 2285168404Spjd * new buffer in a full arc cache. 2286185029Spjd * 2287185029Spjd * This function makes a "best effort". It skips over any buffers 2288185029Spjd * it can't get a hash_lock on, and so may not catch all candidates. 2289185029Spjd * It may also return without evicting as much space as requested. 2290168404Spjd */ 2291168404Spjdstatic void * 2292209962Smmarc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 2293168404Spjd arc_buf_contents_t type) 2294168404Spjd{ 2295168404Spjd arc_state_t *evicted_state; 2296168404Spjd uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 2297205231Skmacy int64_t bytes_remaining; 2298275811Sdelphij arc_buf_hdr_t *hdr, *hdr_prev = NULL; 2299205231Skmacy list_t *evicted_list, *list, *evicted_list_start, *list_start; 2300205231Skmacy kmutex_t *lock, *evicted_lock; 2301168404Spjd kmutex_t *hash_lock; 2302168404Spjd boolean_t have_lock; 2303168404Spjd void *stolen = NULL; 2304258632Savg arc_buf_hdr_t marker = { 0 }; 2305258632Savg int count = 0; 2306205231Skmacy static int evict_metadata_offset, evict_data_offset; 2307258632Savg int i, idx, offset, list_count, lists; 2308168404Spjd 2309168404Spjd ASSERT(state == arc_mru || state == arc_mfu); 2310168404Spjd 2311168404Spjd evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2312206796Spjd 2313275780Sdelphij /* 2314275780Sdelphij * Decide which "type" (data vs metadata) to recycle from. 2315275780Sdelphij * 2316275780Sdelphij * If we are over the metadata limit, recycle from metadata. 2317275780Sdelphij * If we are under the metadata minimum, recycle from data. 2318275780Sdelphij * Otherwise, recycle from whichever type has the oldest (least 2319275780Sdelphij * recently accessed) header. This is not yet implemented. 2320275780Sdelphij */ 2321275780Sdelphij if (recycle) { 2322275780Sdelphij arc_buf_contents_t realtype; 2323275780Sdelphij if (state->arcs_lsize[ARC_BUFC_DATA] == 0) { 2324275780Sdelphij realtype = ARC_BUFC_METADATA; 2325275780Sdelphij } else if (state->arcs_lsize[ARC_BUFC_METADATA] == 0) { 2326275780Sdelphij realtype = ARC_BUFC_DATA; 2327275780Sdelphij } else if (arc_meta_used >= arc_meta_limit) { 2328275780Sdelphij realtype = ARC_BUFC_METADATA; 2329275780Sdelphij } else if (arc_meta_used <= arc_meta_min) { 2330275780Sdelphij realtype = ARC_BUFC_DATA; 2331286570Smav#ifdef illumos 2332286570Smav } else if (HDR_HAS_L1HDR(data_hdr) && 2333286570Smav HDR_HAS_L1HDR(metadata_hdr) && 2334286570Smav data_hdr->b_l1hdr.b_arc_access < 2335286570Smav metadata_hdr->b_l1hdr.b_arc_access) { 2336286570Smav realtype = ARC_BUFC_DATA; 2337275780Sdelphij } else { 2338286570Smav realtype = ARC_BUFC_METADATA; 2339275780Sdelphij#else 2340286570Smav } else { 2341275780Sdelphij /* TODO */ 2342275780Sdelphij realtype = type; 2343275780Sdelphij#endif 2344275780Sdelphij } 2345275780Sdelphij if (realtype != type) { 2346275780Sdelphij /* 2347275780Sdelphij * If we want to evict from a different list, 2348275780Sdelphij * we can not recycle, because DATA vs METADATA 2349275780Sdelphij * buffers are segregated into different kmem 2350275780Sdelphij * caches (and vmem arenas). 2351275780Sdelphij */ 2352275780Sdelphij type = realtype; 2353275780Sdelphij recycle = B_FALSE; 2354275780Sdelphij } 2355275780Sdelphij } 2356275780Sdelphij 2357205231Skmacy if (type == ARC_BUFC_METADATA) { 2358205231Skmacy offset = 0; 2359205231Skmacy list_count = ARC_BUFC_NUMMETADATALISTS; 2360205231Skmacy list_start = &state->arcs_lists[0]; 2361205231Skmacy evicted_list_start = &evicted_state->arcs_lists[0]; 2362205231Skmacy idx = evict_metadata_offset; 2363205231Skmacy } else { 2364205231Skmacy offset = ARC_BUFC_NUMMETADATALISTS; 2365205231Skmacy list_start = &state->arcs_lists[offset]; 2366205231Skmacy evicted_list_start = &evicted_state->arcs_lists[offset]; 2367205231Skmacy list_count = ARC_BUFC_NUMDATALISTS; 2368205231Skmacy idx = evict_data_offset; 2369205231Skmacy } 2370205231Skmacy bytes_remaining = evicted_state->arcs_lsize[type]; 2371258632Savg lists = 0; 2372206796Spjd 2373205231Skmacyevict_start: 2374205231Skmacy list = &list_start[idx]; 2375205231Skmacy evicted_list = &evicted_list_start[idx]; 2376205231Skmacy lock = ARCS_LOCK(state, (offset + idx)); 2377206796Spjd evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 2378168404Spjd 2379286570Smav /* 2380286570Smav * The ghost list lock must be acquired first in order to prevent 2381286570Smav * a 3 party deadlock: 2382286570Smav * 2383286570Smav * - arc_evict_ghost acquires arc_*_ghost->arcs_mtx, followed by 2384286570Smav * l2ad_mtx in arc_hdr_realloc 2385286570Smav * - l2arc_write_buffers acquires l2ad_mtx, followed by arc_*->arcs_mtx 2386286570Smav * - arc_evict acquires arc_*_ghost->arcs_mtx, followed by 2387286570Smav * arc_*_ghost->arcs_mtx and forms a deadlock cycle. 2388286570Smav * 2389286570Smav * This situation is avoided by acquiring the ghost list lock first. 2390286570Smav */ 2391286570Smav mutex_enter(evicted_lock); 2392205231Skmacy mutex_enter(lock); 2393205231Skmacy 2394275811Sdelphij for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2395275811Sdelphij hdr_prev = list_prev(list, hdr); 2396286570Smav if (HDR_HAS_L1HDR(hdr)) { 2397286570Smav bytes_remaining -= 2398286570Smav (hdr->b_size * hdr->b_l1hdr.b_datacnt); 2399286570Smav } 2400168404Spjd /* prefetch buffers have a minimum lifespan */ 2401275811Sdelphij if (HDR_IO_IN_PROGRESS(hdr) || 2402275811Sdelphij (spa && hdr->b_spa != spa) || 2403286570Smav ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 2404286570Smav ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 2405219089Spjd arc_min_prefetch_lifespan)) { 2406168404Spjd skipped++; 2407168404Spjd continue; 2408168404Spjd } 2409168404Spjd /* "lookahead" for better eviction candidate */ 2410275811Sdelphij if (recycle && hdr->b_size != bytes && 2411275811Sdelphij hdr_prev && hdr_prev->b_size == bytes) 2412168404Spjd continue; 2413258632Savg 2414258632Savg /* ignore markers */ 2415275811Sdelphij if (hdr->b_spa == 0) 2416258632Savg continue; 2417258632Savg 2418258632Savg /* 2419258632Savg * It may take a long time to evict all the bufs requested. 2420258632Savg * To avoid blocking all arc activity, periodically drop 2421258632Savg * the arcs_mtx and give other threads a chance to run 2422258632Savg * before reacquiring the lock. 2423258632Savg * 2424258632Savg * If we are looking for a buffer to recycle, we are in 2425258632Savg * the hot code path, so don't sleep. 2426258632Savg */ 2427258632Savg if (!recycle && count++ > arc_evict_iterations) { 2428275811Sdelphij list_insert_after(list, hdr, &marker); 2429286570Smav mutex_exit(lock); 2430258632Savg mutex_exit(evicted_lock); 2431258632Savg kpreempt(KPREEMPT_SYNC); 2432286570Smav mutex_enter(evicted_lock); 2433258632Savg mutex_enter(lock); 2434275811Sdelphij hdr_prev = list_prev(list, &marker); 2435258632Savg list_remove(list, &marker); 2436258632Savg count = 0; 2437258632Savg continue; 2438258632Savg } 2439258632Savg 2440275811Sdelphij hash_lock = HDR_LOCK(hdr); 2441168404Spjd have_lock = MUTEX_HELD(hash_lock); 2442168404Spjd if (have_lock || mutex_tryenter(hash_lock)) { 2443286570Smav ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 2444286570Smav ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0); 2445286570Smav while (hdr->b_l1hdr.b_buf) { 2446286570Smav arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2447219089Spjd if (!mutex_tryenter(&buf->b_evict_lock)) { 2448185029Spjd missed += 1; 2449185029Spjd break; 2450185029Spjd } 2451286570Smav if (buf->b_data != NULL) { 2452275811Sdelphij bytes_evicted += hdr->b_size; 2453286570Smav if (recycle && 2454286570Smav arc_buf_type(hdr) == type && 2455275811Sdelphij hdr->b_size == bytes && 2456275811Sdelphij !HDR_L2_WRITING(hdr)) { 2457168404Spjd stolen = buf->b_data; 2458168404Spjd recycle = FALSE; 2459168404Spjd } 2460168404Spjd } 2461286570Smav if (buf->b_efunc != NULL) { 2462168404Spjd mutex_enter(&arc_eviction_mtx); 2463168404Spjd arc_buf_destroy(buf, 2464168404Spjd buf->b_data == stolen, FALSE); 2465286570Smav hdr->b_l1hdr.b_buf = buf->b_next; 2466168404Spjd buf->b_hdr = &arc_eviction_hdr; 2467168404Spjd buf->b_next = arc_eviction_list; 2468168404Spjd arc_eviction_list = buf; 2469168404Spjd mutex_exit(&arc_eviction_mtx); 2470219089Spjd mutex_exit(&buf->b_evict_lock); 2471168404Spjd } else { 2472219089Spjd mutex_exit(&buf->b_evict_lock); 2473168404Spjd arc_buf_destroy(buf, 2474168404Spjd buf->b_data == stolen, TRUE); 2475168404Spjd } 2476168404Spjd } 2477208373Smm 2478286570Smav if (HDR_HAS_L2HDR(hdr)) { 2479208373Smm ARCSTAT_INCR(arcstat_evict_l2_cached, 2480275811Sdelphij hdr->b_size); 2481208373Smm } else { 2482275811Sdelphij if (l2arc_write_eligible(hdr->b_spa, hdr)) { 2483208373Smm ARCSTAT_INCR(arcstat_evict_l2_eligible, 2484275811Sdelphij hdr->b_size); 2485208373Smm } else { 2486208373Smm ARCSTAT_INCR( 2487208373Smm arcstat_evict_l2_ineligible, 2488275811Sdelphij hdr->b_size); 2489208373Smm } 2490208373Smm } 2491208373Smm 2492286570Smav if (hdr->b_l1hdr.b_datacnt == 0) { 2493275811Sdelphij arc_change_state(evicted_state, hdr, hash_lock); 2494275811Sdelphij ASSERT(HDR_IN_HASH_TABLE(hdr)); 2495275811Sdelphij hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 2496275811Sdelphij hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 2497275811Sdelphij DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 2498185029Spjd } 2499168404Spjd if (!have_lock) 2500168404Spjd mutex_exit(hash_lock); 2501168404Spjd if (bytes >= 0 && bytes_evicted >= bytes) 2502168404Spjd break; 2503205231Skmacy if (bytes_remaining > 0) { 2504205231Skmacy mutex_exit(evicted_lock); 2505205231Skmacy mutex_exit(lock); 2506206796Spjd idx = ((idx + 1) & (list_count - 1)); 2507258632Savg lists++; 2508205231Skmacy goto evict_start; 2509205231Skmacy } 2510168404Spjd } else { 2511168404Spjd missed += 1; 2512168404Spjd } 2513168404Spjd } 2514168404Spjd 2515286570Smav mutex_exit(lock); 2516205231Skmacy mutex_exit(evicted_lock); 2517206796Spjd 2518206796Spjd idx = ((idx + 1) & (list_count - 1)); 2519258632Savg lists++; 2520168404Spjd 2521205231Skmacy if (bytes_evicted < bytes) { 2522258632Savg if (lists < list_count) 2523205231Skmacy goto evict_start; 2524205231Skmacy else 2525205231Skmacy dprintf("only evicted %lld bytes from %x", 2526205231Skmacy (longlong_t)bytes_evicted, state); 2527205231Skmacy } 2528206796Spjd if (type == ARC_BUFC_METADATA) 2529205231Skmacy evict_metadata_offset = idx; 2530205231Skmacy else 2531205231Skmacy evict_data_offset = idx; 2532206796Spjd 2533168404Spjd if (skipped) 2534168404Spjd ARCSTAT_INCR(arcstat_evict_skip, skipped); 2535168404Spjd 2536168404Spjd if (missed) 2537168404Spjd ARCSTAT_INCR(arcstat_mutex_miss, missed); 2538168404Spjd 2539185029Spjd /* 2540258632Savg * Note: we have just evicted some data into the ghost state, 2541258632Savg * potentially putting the ghost size over the desired size. Rather 2542258632Savg * that evicting from the ghost list in this hot code path, leave 2543258632Savg * this chore to the arc_reclaim_thread(). 2544185029Spjd */ 2545185029Spjd 2546205231Skmacy if (stolen) 2547205231Skmacy ARCSTAT_BUMP(arcstat_stolen); 2548168404Spjd return (stolen); 2549168404Spjd} 2550168404Spjd 2551168404Spjd/* 2552168404Spjd * Remove buffers from list until we've removed the specified number of 2553168404Spjd * bytes. Destroy the buffers that are removed. 2554168404Spjd */ 2555168404Spjdstatic void 2556209962Smmarc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2557168404Spjd{ 2558275811Sdelphij arc_buf_hdr_t *hdr, *hdr_prev; 2559219089Spjd arc_buf_hdr_t marker = { 0 }; 2560205231Skmacy list_t *list, *list_start; 2561205231Skmacy kmutex_t *hash_lock, *lock; 2562168404Spjd uint64_t bytes_deleted = 0; 2563168404Spjd uint64_t bufs_skipped = 0; 2564258632Savg int count = 0; 2565205231Skmacy static int evict_offset; 2566205231Skmacy int list_count, idx = evict_offset; 2567258632Savg int offset, lists = 0; 2568168404Spjd 2569168404Spjd ASSERT(GHOST_STATE(state)); 2570205231Skmacy 2571205231Skmacy /* 2572205231Skmacy * data lists come after metadata lists 2573205231Skmacy */ 2574205231Skmacy list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2575205231Skmacy list_count = ARC_BUFC_NUMDATALISTS; 2576205231Skmacy offset = ARC_BUFC_NUMMETADATALISTS; 2577206796Spjd 2578205231Skmacyevict_start: 2579205231Skmacy list = &list_start[idx]; 2580205231Skmacy lock = ARCS_LOCK(state, idx + offset); 2581205231Skmacy 2582205231Skmacy mutex_enter(lock); 2583275811Sdelphij for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2584275811Sdelphij hdr_prev = list_prev(list, hdr); 2585286570Smav if (arc_buf_type(hdr) >= ARC_BUFC_NUMTYPES) 2586275811Sdelphij panic("invalid hdr=%p", (void *)hdr); 2587275811Sdelphij if (spa && hdr->b_spa != spa) 2588185029Spjd continue; 2589219089Spjd 2590219089Spjd /* ignore markers */ 2591275811Sdelphij if (hdr->b_spa == 0) 2592219089Spjd continue; 2593219089Spjd 2594275811Sdelphij hash_lock = HDR_LOCK(hdr); 2595219089Spjd /* caller may be trying to modify this buffer, skip it */ 2596219089Spjd if (MUTEX_HELD(hash_lock)) 2597219089Spjd continue; 2598258632Savg 2599258632Savg /* 2600258632Savg * It may take a long time to evict all the bufs requested. 2601258632Savg * To avoid blocking all arc activity, periodically drop 2602258632Savg * the arcs_mtx and give other threads a chance to run 2603258632Savg * before reacquiring the lock. 2604258632Savg */ 2605258632Savg if (count++ > arc_evict_iterations) { 2606275811Sdelphij list_insert_after(list, hdr, &marker); 2607258632Savg mutex_exit(lock); 2608258632Savg kpreempt(KPREEMPT_SYNC); 2609258632Savg mutex_enter(lock); 2610275811Sdelphij hdr_prev = list_prev(list, &marker); 2611258632Savg list_remove(list, &marker); 2612258632Savg count = 0; 2613258632Savg continue; 2614258632Savg } 2615168404Spjd if (mutex_tryenter(hash_lock)) { 2616275811Sdelphij ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2617286570Smav ASSERT(!HDR_HAS_L1HDR(hdr) || 2618286570Smav hdr->b_l1hdr.b_buf == NULL); 2619168404Spjd ARCSTAT_BUMP(arcstat_deleted); 2620275811Sdelphij bytes_deleted += hdr->b_size; 2621185029Spjd 2622286570Smav if (HDR_HAS_L2HDR(hdr)) { 2623185029Spjd /* 2624185029Spjd * This buffer is cached on the 2nd Level ARC; 2625185029Spjd * don't destroy the header. 2626185029Spjd */ 2627275811Sdelphij arc_change_state(arc_l2c_only, hdr, hash_lock); 2628286570Smav /* 2629286570Smav * dropping from L1+L2 cached to L2-only, 2630286570Smav * realloc to remove the L1 header. 2631286570Smav */ 2632286570Smav hdr = arc_hdr_realloc(hdr, hdr_full_cache, 2633286570Smav hdr_l2only_cache); 2634185029Spjd mutex_exit(hash_lock); 2635185029Spjd } else { 2636275811Sdelphij arc_change_state(arc_anon, hdr, hash_lock); 2637185029Spjd mutex_exit(hash_lock); 2638275811Sdelphij arc_hdr_destroy(hdr); 2639185029Spjd } 2640185029Spjd 2641275811Sdelphij DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 2642168404Spjd if (bytes >= 0 && bytes_deleted >= bytes) 2643168404Spjd break; 2644219089Spjd } else if (bytes < 0) { 2645219089Spjd /* 2646219089Spjd * Insert a list marker and then wait for the 2647219089Spjd * hash lock to become available. Once its 2648219089Spjd * available, restart from where we left off. 2649219089Spjd */ 2650275811Sdelphij list_insert_after(list, hdr, &marker); 2651219089Spjd mutex_exit(lock); 2652219089Spjd mutex_enter(hash_lock); 2653219089Spjd mutex_exit(hash_lock); 2654219089Spjd mutex_enter(lock); 2655275811Sdelphij hdr_prev = list_prev(list, &marker); 2656219089Spjd list_remove(list, &marker); 2657258632Savg } else { 2658168404Spjd bufs_skipped += 1; 2659258632Savg } 2660258632Savg 2661168404Spjd } 2662205231Skmacy mutex_exit(lock); 2663206796Spjd idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2664258632Savg lists++; 2665206796Spjd 2666258632Savg if (lists < list_count) 2667205231Skmacy goto evict_start; 2668206796Spjd 2669205231Skmacy evict_offset = idx; 2670205231Skmacy if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2671185029Spjd (bytes < 0 || bytes_deleted < bytes)) { 2672205231Skmacy list_start = &state->arcs_lists[0]; 2673205231Skmacy list_count = ARC_BUFC_NUMMETADATALISTS; 2674258632Savg offset = lists = 0; 2675205231Skmacy goto evict_start; 2676185029Spjd } 2677185029Spjd 2678168404Spjd if (bufs_skipped) { 2679168404Spjd ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2680168404Spjd ASSERT(bytes >= 0); 2681168404Spjd } 2682168404Spjd 2683168404Spjd if (bytes_deleted < bytes) 2684168404Spjd dprintf("only deleted %lld bytes from %p", 2685168404Spjd (longlong_t)bytes_deleted, state); 2686168404Spjd} 2687168404Spjd 2688168404Spjdstatic void 2689168404Spjdarc_adjust(void) 2690168404Spjd{ 2691208373Smm int64_t adjustment, delta; 2692168404Spjd 2693208373Smm /* 2694208373Smm * Adjust MRU size 2695208373Smm */ 2696168404Spjd 2697209275Smm adjustment = MIN((int64_t)(arc_size - arc_c), 2698209275Smm (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2699209275Smm arc_p)); 2700208373Smm 2701208373Smm if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2702208373Smm delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2703209962Smm (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2704208373Smm adjustment -= delta; 2705168404Spjd } 2706168404Spjd 2707208373Smm if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2708208373Smm delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2709209962Smm (void) arc_evict(arc_mru, 0, delta, FALSE, 2710185029Spjd ARC_BUFC_METADATA); 2711185029Spjd } 2712185029Spjd 2713208373Smm /* 2714208373Smm * Adjust MFU size 2715208373Smm */ 2716168404Spjd 2717208373Smm adjustment = arc_size - arc_c; 2718208373Smm 2719208373Smm if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2720208373Smm delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2721209962Smm (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2722208373Smm adjustment -= delta; 2723168404Spjd } 2724168404Spjd 2725208373Smm if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2726208373Smm int64_t delta = MIN(adjustment, 2727208373Smm arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2728209962Smm (void) arc_evict(arc_mfu, 0, delta, FALSE, 2729208373Smm ARC_BUFC_METADATA); 2730208373Smm } 2731168404Spjd 2732208373Smm /* 2733208373Smm * Adjust ghost lists 2734208373Smm */ 2735168404Spjd 2736208373Smm adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2737168404Spjd 2738208373Smm if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2739208373Smm delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2740209962Smm arc_evict_ghost(arc_mru_ghost, 0, delta); 2741208373Smm } 2742185029Spjd 2743208373Smm adjustment = 2744208373Smm arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2745208373Smm 2746208373Smm if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2747208373Smm delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2748209962Smm arc_evict_ghost(arc_mfu_ghost, 0, delta); 2749168404Spjd } 2750168404Spjd} 2751168404Spjd 2752168404Spjdstatic void 2753168404Spjdarc_do_user_evicts(void) 2754168404Spjd{ 2755191903Skmacy static arc_buf_t *tmp_arc_eviction_list; 2756191903Skmacy 2757191903Skmacy /* 2758191903Skmacy * Move list over to avoid LOR 2759191903Skmacy */ 2760206796Spjdrestart: 2761168404Spjd mutex_enter(&arc_eviction_mtx); 2762191903Skmacy tmp_arc_eviction_list = arc_eviction_list; 2763191903Skmacy arc_eviction_list = NULL; 2764191903Skmacy mutex_exit(&arc_eviction_mtx); 2765191903Skmacy 2766191903Skmacy while (tmp_arc_eviction_list != NULL) { 2767191903Skmacy arc_buf_t *buf = tmp_arc_eviction_list; 2768191903Skmacy tmp_arc_eviction_list = buf->b_next; 2769219089Spjd mutex_enter(&buf->b_evict_lock); 2770168404Spjd buf->b_hdr = NULL; 2771219089Spjd mutex_exit(&buf->b_evict_lock); 2772168404Spjd 2773168404Spjd if (buf->b_efunc != NULL) 2774268858Sdelphij VERIFY0(buf->b_efunc(buf->b_private)); 2775168404Spjd 2776168404Spjd buf->b_efunc = NULL; 2777168404Spjd buf->b_private = NULL; 2778168404Spjd kmem_cache_free(buf_cache, buf); 2779168404Spjd } 2780191903Skmacy 2781191903Skmacy if (arc_eviction_list != NULL) 2782191903Skmacy goto restart; 2783168404Spjd} 2784168404Spjd 2785168404Spjd/* 2786185029Spjd * Flush all *evictable* data from the cache for the given spa. 2787168404Spjd * NOTE: this will not touch "active" (i.e. referenced) data. 2788168404Spjd */ 2789168404Spjdvoid 2790185029Spjdarc_flush(spa_t *spa) 2791168404Spjd{ 2792209962Smm uint64_t guid = 0; 2793209962Smm 2794286570Smav if (spa != NULL) 2795228103Smm guid = spa_load_guid(spa); 2796209962Smm 2797205231Skmacy while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2798209962Smm (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2799286570Smav if (spa != NULL) 2800185029Spjd break; 2801185029Spjd } 2802205231Skmacy while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2803209962Smm (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2804286570Smav if (spa != NULL) 2805185029Spjd break; 2806185029Spjd } 2807205231Skmacy while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2808209962Smm (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2809286570Smav if (spa != NULL) 2810185029Spjd break; 2811185029Spjd } 2812205231Skmacy while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2813209962Smm (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2814286570Smav if (spa != NULL) 2815185029Spjd break; 2816185029Spjd } 2817168404Spjd 2818209962Smm arc_evict_ghost(arc_mru_ghost, guid, -1); 2819209962Smm arc_evict_ghost(arc_mfu_ghost, guid, -1); 2820168404Spjd 2821168404Spjd mutex_enter(&arc_reclaim_thr_lock); 2822168404Spjd arc_do_user_evicts(); 2823168404Spjd mutex_exit(&arc_reclaim_thr_lock); 2824185029Spjd ASSERT(spa || arc_eviction_list == NULL); 2825168404Spjd} 2826168404Spjd 2827168404Spjdvoid 2828168404Spjdarc_shrink(void) 2829168404Spjd{ 2830270759Ssmh 2831168404Spjd if (arc_c > arc_c_min) { 2832168404Spjd uint64_t to_free; 2833168404Spjd 2834277452Swill to_free = arc_c >> arc_shrink_shift; 2835272483Ssmh DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 2836272483Ssmh arc_c_min, uint64_t, arc_p, uint64_t, to_free); 2837168404Spjd if (arc_c > arc_c_min + to_free) 2838168404Spjd atomic_add_64(&arc_c, -to_free); 2839168404Spjd else 2840168404Spjd arc_c = arc_c_min; 2841168404Spjd 2842168404Spjd atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2843168404Spjd if (arc_c > arc_size) 2844168404Spjd arc_c = MAX(arc_size, arc_c_min); 2845168404Spjd if (arc_p > arc_c) 2846168404Spjd arc_p = (arc_c >> 1); 2847272483Ssmh 2848272483Ssmh DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 2849272483Ssmh arc_p); 2850272483Ssmh 2851168404Spjd ASSERT(arc_c >= arc_c_min); 2852168404Spjd ASSERT((int64_t)arc_p >= 0); 2853168404Spjd } 2854168404Spjd 2855270759Ssmh if (arc_size > arc_c) { 2856270759Ssmh DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 2857270759Ssmh uint64_t, arc_c); 2858168404Spjd arc_adjust(); 2859270759Ssmh } 2860168404Spjd} 2861168404Spjd 2862185029Spjdstatic int needfree = 0; 2863168404Spjd 2864168404Spjdstatic int 2865168404Spjdarc_reclaim_needed(void) 2866168404Spjd{ 2867168404Spjd 2868168404Spjd#ifdef _KERNEL 2869219089Spjd 2870270759Ssmh if (needfree) { 2871270759Ssmh DTRACE_PROBE(arc__reclaim_needfree); 2872197816Skmacy return (1); 2873270759Ssmh } 2874168404Spjd 2875191902Skmacy /* 2876212780Savg * Cooperate with pagedaemon when it's time for it to scan 2877212780Savg * and reclaim some pages. 2878191902Skmacy */ 2879272483Ssmh if (freemem < zfs_arc_free_target) { 2880272483Ssmh DTRACE_PROBE2(arc__reclaim_freemem, uint64_t, 2881272483Ssmh freemem, uint64_t, zfs_arc_free_target); 2882191902Skmacy return (1); 2883270759Ssmh } 2884191902Skmacy 2885277300Ssmh#ifdef illumos 2886168404Spjd /* 2887185029Spjd * take 'desfree' extra pages, so we reclaim sooner, rather than later 2888185029Spjd */ 2889185029Spjd extra = desfree; 2890185029Spjd 2891185029Spjd /* 2892185029Spjd * check that we're out of range of the pageout scanner. It starts to 2893185029Spjd * schedule paging if freemem is less than lotsfree and needfree. 2894185029Spjd * lotsfree is the high-water mark for pageout, and needfree is the 2895185029Spjd * number of needed free pages. We add extra pages here to make sure 2896185029Spjd * the scanner doesn't start up while we're freeing memory. 2897185029Spjd */ 2898185029Spjd if (freemem < lotsfree + needfree + extra) 2899185029Spjd return (1); 2900185029Spjd 2901185029Spjd /* 2902168404Spjd * check to make sure that swapfs has enough space so that anon 2903185029Spjd * reservations can still succeed. anon_resvmem() checks that the 2904168404Spjd * availrmem is greater than swapfs_minfree, and the number of reserved 2905168404Spjd * swap pages. We also add a bit of extra here just to prevent 2906168404Spjd * circumstances from getting really dire. 2907168404Spjd */ 2908168404Spjd if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2909168404Spjd return (1); 2910168404Spjd 2911168404Spjd /* 2912272483Ssmh * Check that we have enough availrmem that memory locking (e.g., via 2913272483Ssmh * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 2914272483Ssmh * stores the number of pages that cannot be locked; when availrmem 2915272483Ssmh * drops below pages_pp_maximum, page locking mechanisms such as 2916272483Ssmh * page_pp_lock() will fail.) 2917272483Ssmh */ 2918272483Ssmh if (availrmem <= pages_pp_maximum) 2919272483Ssmh return (1); 2920272483Ssmh 2921277300Ssmh#endif /* illumos */ 2922272483Ssmh#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 2923272483Ssmh /* 2924168404Spjd * If we're on an i386 platform, it's possible that we'll exhaust the 2925168404Spjd * kernel heap space before we ever run out of available physical 2926168404Spjd * memory. Most checks of the size of the heap_area compare against 2927168404Spjd * tune.t_minarmem, which is the minimum available real memory that we 2928168404Spjd * can have in the system. However, this is generally fixed at 25 pages 2929168404Spjd * which is so low that it's useless. In this comparison, we seek to 2930168404Spjd * calculate the total heap-size, and reclaim if more than 3/4ths of the 2931185029Spjd * heap is allocated. (Or, in the calculation, if less than 1/4th is 2932168404Spjd * free) 2933168404Spjd */ 2934272483Ssmh if (vmem_size(heap_arena, VMEM_FREE) < 2935272483Ssmh (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2)) { 2936270861Ssmh DTRACE_PROBE2(arc__reclaim_used, uint64_t, 2937272483Ssmh vmem_size(heap_arena, VMEM_FREE), uint64_t, 2938272483Ssmh (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2); 2939270861Ssmh return (1); 2940270861Ssmh } 2941281026Smav#define zio_arena NULL 2942281026Smav#else 2943281026Smav#define zio_arena heap_arena 2944270861Ssmh#endif 2945281026Smav 2946272483Ssmh /* 2947272483Ssmh * If zio data pages are being allocated out of a separate heap segment, 2948272483Ssmh * then enforce that the size of available vmem for this arena remains 2949272483Ssmh * above about 1/16th free. 2950272483Ssmh * 2951272483Ssmh * Note: The 1/16th arena free requirement was put in place 2952272483Ssmh * to aggressively evict memory from the arc in order to avoid 2953272483Ssmh * memory fragmentation issues. 2954272483Ssmh */ 2955272483Ssmh if (zio_arena != NULL && 2956272483Ssmh vmem_size(zio_arena, VMEM_FREE) < 2957272483Ssmh (vmem_size(zio_arena, VMEM_ALLOC) >> 4)) 2958272483Ssmh return (1); 2959281026Smav 2960281026Smav /* 2961281026Smav * Above limits know nothing about real level of KVA fragmentation. 2962281026Smav * Start aggressive reclamation if too little sequential KVA left. 2963281026Smav */ 2964281109Smav if (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) { 2965281109Smav DTRACE_PROBE2(arc__reclaim_maxfree, uint64_t, 2966281109Smav vmem_size(heap_arena, VMEM_MAXFREE), 2967281109Smav uint64_t, zfs_max_recordsize); 2968281026Smav return (1); 2969281109Smav } 2970281026Smav 2971272483Ssmh#else /* _KERNEL */ 2972168404Spjd if (spa_get_random(100) == 0) 2973168404Spjd return (1); 2974272483Ssmh#endif /* _KERNEL */ 2975270759Ssmh DTRACE_PROBE(arc__reclaim_no); 2976270759Ssmh 2977168404Spjd return (0); 2978168404Spjd} 2979168404Spjd 2980208454Spjdextern kmem_cache_t *zio_buf_cache[]; 2981208454Spjdextern kmem_cache_t *zio_data_buf_cache[]; 2982272527Sdelphijextern kmem_cache_t *range_seg_cache; 2983208454Spjd 2984278040Ssmhstatic __noinline void 2985168404Spjdarc_kmem_reap_now(arc_reclaim_strategy_t strat) 2986168404Spjd{ 2987168404Spjd size_t i; 2988168404Spjd kmem_cache_t *prev_cache = NULL; 2989168404Spjd kmem_cache_t *prev_data_cache = NULL; 2990168404Spjd 2991272483Ssmh DTRACE_PROBE(arc__kmem_reap_start); 2992168404Spjd#ifdef _KERNEL 2993185029Spjd if (arc_meta_used >= arc_meta_limit) { 2994185029Spjd /* 2995185029Spjd * We are exceeding our meta-data cache limit. 2996185029Spjd * Purge some DNLC entries to release holds on meta-data. 2997185029Spjd */ 2998185029Spjd dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2999185029Spjd } 3000168404Spjd#if defined(__i386) 3001168404Spjd /* 3002168404Spjd * Reclaim unused memory from all kmem caches. 3003168404Spjd */ 3004168404Spjd kmem_reap(); 3005168404Spjd#endif 3006168404Spjd#endif 3007168404Spjd 3008168404Spjd /* 3009185029Spjd * An aggressive reclamation will shrink the cache size as well as 3010168404Spjd * reap free buffers from the arc kmem caches. 3011168404Spjd */ 3012168404Spjd if (strat == ARC_RECLAIM_AGGR) 3013168404Spjd arc_shrink(); 3014168404Spjd 3015168404Spjd for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 3016168404Spjd if (zio_buf_cache[i] != prev_cache) { 3017168404Spjd prev_cache = zio_buf_cache[i]; 3018168404Spjd kmem_cache_reap_now(zio_buf_cache[i]); 3019168404Spjd } 3020168404Spjd if (zio_data_buf_cache[i] != prev_data_cache) { 3021168404Spjd prev_data_cache = zio_data_buf_cache[i]; 3022168404Spjd kmem_cache_reap_now(zio_data_buf_cache[i]); 3023168404Spjd } 3024168404Spjd } 3025168404Spjd kmem_cache_reap_now(buf_cache); 3026286570Smav kmem_cache_reap_now(hdr_full_cache); 3027286570Smav kmem_cache_reap_now(hdr_l2only_cache); 3028272506Sdelphij kmem_cache_reap_now(range_seg_cache); 3029272483Ssmh 3030277300Ssmh#ifdef illumos 3031272483Ssmh /* 3032272483Ssmh * Ask the vmem arena to reclaim unused memory from its 3033272483Ssmh * quantum caches. 3034272483Ssmh */ 3035272483Ssmh if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR) 3036272483Ssmh vmem_qcache_reap(zio_arena); 3037272483Ssmh#endif 3038272483Ssmh DTRACE_PROBE(arc__kmem_reap_end); 3039168404Spjd} 3040168404Spjd 3041168404Spjdstatic void 3042168404Spjdarc_reclaim_thread(void *dummy __unused) 3043168404Spjd{ 3044168404Spjd clock_t growtime = 0; 3045168404Spjd arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 3046168404Spjd callb_cpr_t cpr; 3047168404Spjd 3048168404Spjd CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 3049168404Spjd 3050168404Spjd mutex_enter(&arc_reclaim_thr_lock); 3051168404Spjd while (arc_thread_exit == 0) { 3052168404Spjd if (arc_reclaim_needed()) { 3053168404Spjd 3054168404Spjd if (arc_no_grow) { 3055168404Spjd if (last_reclaim == ARC_RECLAIM_CONS) { 3056272483Ssmh DTRACE_PROBE(arc__reclaim_aggr_no_grow); 3057168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 3058168404Spjd } else { 3059168404Spjd last_reclaim = ARC_RECLAIM_CONS; 3060168404Spjd } 3061168404Spjd } else { 3062168404Spjd arc_no_grow = TRUE; 3063168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 3064272483Ssmh DTRACE_PROBE(arc__reclaim_aggr); 3065168404Spjd membar_producer(); 3066168404Spjd } 3067168404Spjd 3068168404Spjd /* reset the growth delay for every reclaim */ 3069219089Spjd growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 3070168404Spjd 3071185029Spjd if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 3072168404Spjd /* 3073185029Spjd * If needfree is TRUE our vm_lowmem hook 3074168404Spjd * was called and in that case we must free some 3075168404Spjd * memory, so switch to aggressive mode. 3076168404Spjd */ 3077168404Spjd arc_no_grow = TRUE; 3078168404Spjd last_reclaim = ARC_RECLAIM_AGGR; 3079168404Spjd } 3080168404Spjd arc_kmem_reap_now(last_reclaim); 3081185029Spjd arc_warm = B_TRUE; 3082185029Spjd 3083219089Spjd } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 3084168404Spjd arc_no_grow = FALSE; 3085168404Spjd } 3086168404Spjd 3087209275Smm arc_adjust(); 3088168404Spjd 3089168404Spjd if (arc_eviction_list != NULL) 3090168404Spjd arc_do_user_evicts(); 3091168404Spjd 3092211762Savg#ifdef _KERNEL 3093211762Savg if (needfree) { 3094185029Spjd needfree = 0; 3095185029Spjd wakeup(&needfree); 3096211762Savg } 3097168404Spjd#endif 3098168404Spjd 3099168404Spjd /* block until needed, or one second, whichever is shorter */ 3100168404Spjd CALLB_CPR_SAFE_BEGIN(&cpr); 3101168404Spjd (void) cv_timedwait(&arc_reclaim_thr_cv, 3102168404Spjd &arc_reclaim_thr_lock, hz); 3103168404Spjd CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 3104168404Spjd } 3105168404Spjd 3106168404Spjd arc_thread_exit = 0; 3107168404Spjd cv_broadcast(&arc_reclaim_thr_cv); 3108168404Spjd CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 3109168404Spjd thread_exit(); 3110168404Spjd} 3111168404Spjd 3112168404Spjd/* 3113168404Spjd * Adapt arc info given the number of bytes we are trying to add and 3114168404Spjd * the state that we are comming from. This function is only called 3115168404Spjd * when we are adding new content to the cache. 3116168404Spjd */ 3117168404Spjdstatic void 3118168404Spjdarc_adapt(int bytes, arc_state_t *state) 3119168404Spjd{ 3120168404Spjd int mult; 3121208373Smm uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 3122168404Spjd 3123185029Spjd if (state == arc_l2c_only) 3124185029Spjd return; 3125185029Spjd 3126168404Spjd ASSERT(bytes > 0); 3127168404Spjd /* 3128168404Spjd * Adapt the target size of the MRU list: 3129168404Spjd * - if we just hit in the MRU ghost list, then increase 3130168404Spjd * the target size of the MRU list. 3131168404Spjd * - if we just hit in the MFU ghost list, then increase 3132168404Spjd * the target size of the MFU list by decreasing the 3133168404Spjd * target size of the MRU list. 3134168404Spjd */ 3135168404Spjd if (state == arc_mru_ghost) { 3136168404Spjd mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 3137168404Spjd 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 3138209275Smm mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 3139168404Spjd 3140208373Smm arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 3141168404Spjd } else if (state == arc_mfu_ghost) { 3142208373Smm uint64_t delta; 3143208373Smm 3144168404Spjd mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 3145168404Spjd 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 3146209275Smm mult = MIN(mult, 10); 3147168404Spjd 3148208373Smm delta = MIN(bytes * mult, arc_p); 3149208373Smm arc_p = MAX(arc_p_min, arc_p - delta); 3150168404Spjd } 3151168404Spjd ASSERT((int64_t)arc_p >= 0); 3152168404Spjd 3153168404Spjd if (arc_reclaim_needed()) { 3154168404Spjd cv_signal(&arc_reclaim_thr_cv); 3155168404Spjd return; 3156168404Spjd } 3157168404Spjd 3158168404Spjd if (arc_no_grow) 3159168404Spjd return; 3160168404Spjd 3161168404Spjd if (arc_c >= arc_c_max) 3162168404Spjd return; 3163168404Spjd 3164168404Spjd /* 3165168404Spjd * If we're within (2 * maxblocksize) bytes of the target 3166168404Spjd * cache size, increment the target cache size 3167168404Spjd */ 3168168404Spjd if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 3169272483Ssmh DTRACE_PROBE1(arc__inc_adapt, int, bytes); 3170168404Spjd atomic_add_64(&arc_c, (int64_t)bytes); 3171168404Spjd if (arc_c > arc_c_max) 3172168404Spjd arc_c = arc_c_max; 3173168404Spjd else if (state == arc_anon) 3174168404Spjd atomic_add_64(&arc_p, (int64_t)bytes); 3175168404Spjd if (arc_p > arc_c) 3176168404Spjd arc_p = arc_c; 3177168404Spjd } 3178168404Spjd ASSERT((int64_t)arc_p >= 0); 3179168404Spjd} 3180168404Spjd 3181168404Spjd/* 3182168404Spjd * Check if the cache has reached its limits and eviction is required 3183168404Spjd * prior to insert. 3184168404Spjd */ 3185168404Spjdstatic int 3186185029Spjdarc_evict_needed(arc_buf_contents_t type) 3187168404Spjd{ 3188185029Spjd if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 3189185029Spjd return (1); 3190185029Spjd 3191168404Spjd if (arc_reclaim_needed()) 3192168404Spjd return (1); 3193168404Spjd 3194168404Spjd return (arc_size > arc_c); 3195168404Spjd} 3196168404Spjd 3197168404Spjd/* 3198168404Spjd * The buffer, supplied as the first argument, needs a data block. 3199168404Spjd * So, if we are at cache max, determine which cache should be victimized. 3200168404Spjd * We have the following cases: 3201168404Spjd * 3202168404Spjd * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 3203168404Spjd * In this situation if we're out of space, but the resident size of the MFU is 3204168404Spjd * under the limit, victimize the MFU cache to satisfy this insertion request. 3205168404Spjd * 3206168404Spjd * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 3207168404Spjd * Here, we've used up all of the available space for the MRU, so we need to 3208168404Spjd * evict from our own cache instead. Evict from the set of resident MRU 3209168404Spjd * entries. 3210168404Spjd * 3211168404Spjd * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 3212168404Spjd * c minus p represents the MFU space in the cache, since p is the size of the 3213168404Spjd * cache that is dedicated to the MRU. In this situation there's still space on 3214168404Spjd * the MFU side, so the MRU side needs to be victimized. 3215168404Spjd * 3216168404Spjd * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 3217168404Spjd * MFU's resident set is consuming more space than it has been allotted. In 3218168404Spjd * this situation, we must victimize our own cache, the MFU, for this insertion. 3219168404Spjd */ 3220168404Spjdstatic void 3221168404Spjdarc_get_data_buf(arc_buf_t *buf) 3222168404Spjd{ 3223286570Smav arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 3224168404Spjd uint64_t size = buf->b_hdr->b_size; 3225286570Smav arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 3226168404Spjd 3227168404Spjd arc_adapt(size, state); 3228168404Spjd 3229168404Spjd /* 3230168404Spjd * We have not yet reached cache maximum size, 3231168404Spjd * just allocate a new buffer. 3232168404Spjd */ 3233185029Spjd if (!arc_evict_needed(type)) { 3234168404Spjd if (type == ARC_BUFC_METADATA) { 3235168404Spjd buf->b_data = zio_buf_alloc(size); 3236208373Smm arc_space_consume(size, ARC_SPACE_DATA); 3237168404Spjd } else { 3238168404Spjd ASSERT(type == ARC_BUFC_DATA); 3239168404Spjd buf->b_data = zio_data_buf_alloc(size); 3240208373Smm ARCSTAT_INCR(arcstat_data_size, size); 3241185029Spjd atomic_add_64(&arc_size, size); 3242168404Spjd } 3243168404Spjd goto out; 3244168404Spjd } 3245168404Spjd 3246168404Spjd /* 3247168404Spjd * If we are prefetching from the mfu ghost list, this buffer 3248168404Spjd * will end up on the mru list; so steal space from there. 3249168404Spjd */ 3250168404Spjd if (state == arc_mfu_ghost) 3251286570Smav state = HDR_PREFETCH(buf->b_hdr) ? arc_mru : arc_mfu; 3252168404Spjd else if (state == arc_mru_ghost) 3253168404Spjd state = arc_mru; 3254168404Spjd 3255168404Spjd if (state == arc_mru || state == arc_anon) { 3256168404Spjd uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 3257208373Smm state = (arc_mfu->arcs_lsize[type] >= size && 3258185029Spjd arc_p > mru_used) ? arc_mfu : arc_mru; 3259168404Spjd } else { 3260168404Spjd /* MFU cases */ 3261168404Spjd uint64_t mfu_space = arc_c - arc_p; 3262208373Smm state = (arc_mru->arcs_lsize[type] >= size && 3263185029Spjd mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 3264168404Spjd } 3265209962Smm if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 3266168404Spjd if (type == ARC_BUFC_METADATA) { 3267168404Spjd buf->b_data = zio_buf_alloc(size); 3268208373Smm arc_space_consume(size, ARC_SPACE_DATA); 3269168404Spjd } else { 3270168404Spjd ASSERT(type == ARC_BUFC_DATA); 3271168404Spjd buf->b_data = zio_data_buf_alloc(size); 3272208373Smm ARCSTAT_INCR(arcstat_data_size, size); 3273185029Spjd atomic_add_64(&arc_size, size); 3274168404Spjd } 3275168404Spjd ARCSTAT_BUMP(arcstat_recycle_miss); 3276168404Spjd } 3277168404Spjd ASSERT(buf->b_data != NULL); 3278168404Spjdout: 3279168404Spjd /* 3280168404Spjd * Update the state size. Note that ghost states have a 3281168404Spjd * "ghost size" and so don't need to be updated. 3282168404Spjd */ 3283286570Smav if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) { 3284168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3285168404Spjd 3286286570Smav atomic_add_64(&hdr->b_l1hdr.b_state->arcs_size, size); 3287286570Smav if (list_link_active(&hdr->b_l1hdr.b_arc_node)) { 3288286570Smav ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3289286570Smav atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type], 3290286570Smav size); 3291168404Spjd } 3292168404Spjd /* 3293168404Spjd * If we are growing the cache, and we are adding anonymous 3294168404Spjd * data, and we have outgrown arc_p, update arc_p 3295168404Spjd */ 3296286570Smav if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 3297168404Spjd arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 3298168404Spjd arc_p = MIN(arc_c, arc_p + size); 3299168404Spjd } 3300205231Skmacy ARCSTAT_BUMP(arcstat_allocated); 3301168404Spjd} 3302168404Spjd 3303168404Spjd/* 3304168404Spjd * This routine is called whenever a buffer is accessed. 3305168404Spjd * NOTE: the hash lock is dropped in this function. 3306168404Spjd */ 3307168404Spjdstatic void 3308275811Sdelphijarc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3309168404Spjd{ 3310219089Spjd clock_t now; 3311219089Spjd 3312168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 3313286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 3314168404Spjd 3315286570Smav if (hdr->b_l1hdr.b_state == arc_anon) { 3316168404Spjd /* 3317168404Spjd * This buffer is not in the cache, and does not 3318168404Spjd * appear in our "ghost" list. Add the new buffer 3319168404Spjd * to the MRU state. 3320168404Spjd */ 3321168404Spjd 3322286570Smav ASSERT0(hdr->b_l1hdr.b_arc_access); 3323286570Smav hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3324275811Sdelphij DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3325275811Sdelphij arc_change_state(arc_mru, hdr, hash_lock); 3326168404Spjd 3327286570Smav } else if (hdr->b_l1hdr.b_state == arc_mru) { 3328219089Spjd now = ddi_get_lbolt(); 3329219089Spjd 3330168404Spjd /* 3331168404Spjd * If this buffer is here because of a prefetch, then either: 3332168404Spjd * - clear the flag if this is a "referencing" read 3333168404Spjd * (any subsequent access will bump this into the MFU state). 3334168404Spjd * or 3335168404Spjd * - move the buffer to the head of the list if this is 3336168404Spjd * another prefetch (to make it less likely to be evicted). 3337168404Spjd */ 3338286570Smav if (HDR_PREFETCH(hdr)) { 3339286570Smav if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 3340286570Smav ASSERT(list_link_active( 3341286570Smav &hdr->b_l1hdr.b_arc_node)); 3342168404Spjd } else { 3343275811Sdelphij hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3344168404Spjd ARCSTAT_BUMP(arcstat_mru_hits); 3345168404Spjd } 3346286570Smav hdr->b_l1hdr.b_arc_access = now; 3347168404Spjd return; 3348168404Spjd } 3349168404Spjd 3350168404Spjd /* 3351168404Spjd * This buffer has been "accessed" only once so far, 3352168404Spjd * but it is still in the cache. Move it to the MFU 3353168404Spjd * state. 3354168404Spjd */ 3355286570Smav if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 3356168404Spjd /* 3357168404Spjd * More than 125ms have passed since we 3358168404Spjd * instantiated this buffer. Move it to the 3359168404Spjd * most frequently used state. 3360168404Spjd */ 3361286570Smav hdr->b_l1hdr.b_arc_access = now; 3362275811Sdelphij DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3363275811Sdelphij arc_change_state(arc_mfu, hdr, hash_lock); 3364168404Spjd } 3365168404Spjd ARCSTAT_BUMP(arcstat_mru_hits); 3366286570Smav } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 3367168404Spjd arc_state_t *new_state; 3368168404Spjd /* 3369168404Spjd * This buffer has been "accessed" recently, but 3370168404Spjd * was evicted from the cache. Move it to the 3371168404Spjd * MFU state. 3372168404Spjd */ 3373168404Spjd 3374286570Smav if (HDR_PREFETCH(hdr)) { 3375168404Spjd new_state = arc_mru; 3376286570Smav if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 3377275811Sdelphij hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3378275811Sdelphij DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3379168404Spjd } else { 3380168404Spjd new_state = arc_mfu; 3381275811Sdelphij DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3382168404Spjd } 3383168404Spjd 3384286570Smav hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3385275811Sdelphij arc_change_state(new_state, hdr, hash_lock); 3386168404Spjd 3387168404Spjd ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3388286570Smav } else if (hdr->b_l1hdr.b_state == arc_mfu) { 3389168404Spjd /* 3390168404Spjd * This buffer has been accessed more than once and is 3391168404Spjd * still in the cache. Keep it in the MFU state. 3392168404Spjd * 3393168404Spjd * NOTE: an add_reference() that occurred when we did 3394168404Spjd * the arc_read() will have kicked this off the list. 3395168404Spjd * If it was a prefetch, we will explicitly move it to 3396168404Spjd * the head of the list now. 3397168404Spjd */ 3398286570Smav if ((HDR_PREFETCH(hdr)) != 0) { 3399286570Smav ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3400286570Smav ASSERT(list_link_active(&hdr->b_l1hdr.b_arc_node)); 3401168404Spjd } 3402168404Spjd ARCSTAT_BUMP(arcstat_mfu_hits); 3403286570Smav hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3404286570Smav } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 3405168404Spjd arc_state_t *new_state = arc_mfu; 3406168404Spjd /* 3407168404Spjd * This buffer has been accessed more than once but has 3408168404Spjd * been evicted from the cache. Move it back to the 3409168404Spjd * MFU state. 3410168404Spjd */ 3411168404Spjd 3412286570Smav if (HDR_PREFETCH(hdr)) { 3413168404Spjd /* 3414168404Spjd * This is a prefetch access... 3415168404Spjd * move this block back to the MRU state. 3416168404Spjd */ 3417286570Smav ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 3418168404Spjd new_state = arc_mru; 3419168404Spjd } 3420168404Spjd 3421286570Smav hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3422275811Sdelphij DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3423275811Sdelphij arc_change_state(new_state, hdr, hash_lock); 3424168404Spjd 3425168404Spjd ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 3426286570Smav } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 3427185029Spjd /* 3428185029Spjd * This buffer is on the 2nd Level ARC. 3429185029Spjd */ 3430185029Spjd 3431286570Smav hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3432275811Sdelphij DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3433275811Sdelphij arc_change_state(arc_mfu, hdr, hash_lock); 3434168404Spjd } else { 3435168404Spjd ASSERT(!"invalid arc state"); 3436168404Spjd } 3437168404Spjd} 3438168404Spjd 3439168404Spjd/* a generic arc_done_func_t which you can use */ 3440168404Spjd/* ARGSUSED */ 3441168404Spjdvoid 3442168404Spjdarc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 3443168404Spjd{ 3444219089Spjd if (zio == NULL || zio->io_error == 0) 3445219089Spjd bcopy(buf->b_data, arg, buf->b_hdr->b_size); 3446248571Smm VERIFY(arc_buf_remove_ref(buf, arg)); 3447168404Spjd} 3448168404Spjd 3449185029Spjd/* a generic arc_done_func_t */ 3450168404Spjdvoid 3451168404Spjdarc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 3452168404Spjd{ 3453168404Spjd arc_buf_t **bufp = arg; 3454168404Spjd if (zio && zio->io_error) { 3455248571Smm VERIFY(arc_buf_remove_ref(buf, arg)); 3456168404Spjd *bufp = NULL; 3457168404Spjd } else { 3458168404Spjd *bufp = buf; 3459219089Spjd ASSERT(buf->b_data); 3460168404Spjd } 3461168404Spjd} 3462168404Spjd 3463168404Spjdstatic void 3464168404Spjdarc_read_done(zio_t *zio) 3465168404Spjd{ 3466268075Sdelphij arc_buf_hdr_t *hdr; 3467168404Spjd arc_buf_t *buf; 3468168404Spjd arc_buf_t *abuf; /* buffer we're assigning to callback */ 3469268075Sdelphij kmutex_t *hash_lock = NULL; 3470168404Spjd arc_callback_t *callback_list, *acb; 3471168404Spjd int freeable = FALSE; 3472168404Spjd 3473168404Spjd buf = zio->io_private; 3474168404Spjd hdr = buf->b_hdr; 3475168404Spjd 3476168404Spjd /* 3477168404Spjd * The hdr was inserted into hash-table and removed from lists 3478168404Spjd * prior to starting I/O. We should find this header, since 3479168404Spjd * it's in the hash table, and it should be legit since it's 3480168404Spjd * not possible to evict it during the I/O. The only possible 3481168404Spjd * reason for it not to be found is if we were freed during the 3482168404Spjd * read. 3483168404Spjd */ 3484268075Sdelphij if (HDR_IN_HASH_TABLE(hdr)) { 3485268075Sdelphij ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3486268075Sdelphij ASSERT3U(hdr->b_dva.dva_word[0], ==, 3487268075Sdelphij BP_IDENTITY(zio->io_bp)->dva_word[0]); 3488268075Sdelphij ASSERT3U(hdr->b_dva.dva_word[1], ==, 3489268075Sdelphij BP_IDENTITY(zio->io_bp)->dva_word[1]); 3490168404Spjd 3491268075Sdelphij arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3492268075Sdelphij &hash_lock); 3493168404Spjd 3494268075Sdelphij ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3495268075Sdelphij hash_lock == NULL) || 3496268075Sdelphij (found == hdr && 3497268075Sdelphij DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3498268075Sdelphij (found == hdr && HDR_L2_READING(hdr))); 3499268075Sdelphij } 3500268075Sdelphij 3501275811Sdelphij hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; 3502286570Smav if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 3503275811Sdelphij hdr->b_flags &= ~ARC_FLAG_L2CACHE; 3504206796Spjd 3505168404Spjd /* byteswap if necessary */ 3506286570Smav callback_list = hdr->b_l1hdr.b_acb; 3507168404Spjd ASSERT(callback_list != NULL); 3508209101Smm if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3509236884Smm dmu_object_byteswap_t bswap = 3510236884Smm DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3511185029Spjd arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3512185029Spjd byteswap_uint64_array : 3513236884Smm dmu_ot_byteswap[bswap].ob_func; 3514185029Spjd func(buf->b_data, hdr->b_size); 3515185029Spjd } 3516168404Spjd 3517185029Spjd arc_cksum_compute(buf, B_FALSE); 3518240133Smm#ifdef illumos 3519240133Smm arc_buf_watch(buf); 3520277300Ssmh#endif 3521168404Spjd 3522286570Smav if (hash_lock && zio->io_error == 0 && 3523286570Smav hdr->b_l1hdr.b_state == arc_anon) { 3524219089Spjd /* 3525219089Spjd * Only call arc_access on anonymous buffers. This is because 3526219089Spjd * if we've issued an I/O for an evicted buffer, we've already 3527219089Spjd * called arc_access (to prevent any simultaneous readers from 3528219089Spjd * getting confused). 3529219089Spjd */ 3530219089Spjd arc_access(hdr, hash_lock); 3531219089Spjd } 3532219089Spjd 3533168404Spjd /* create copies of the data buffer for the callers */ 3534168404Spjd abuf = buf; 3535168404Spjd for (acb = callback_list; acb; acb = acb->acb_next) { 3536168404Spjd if (acb->acb_done) { 3537242845Sdelphij if (abuf == NULL) { 3538242845Sdelphij ARCSTAT_BUMP(arcstat_duplicate_reads); 3539168404Spjd abuf = arc_buf_clone(buf); 3540242845Sdelphij } 3541168404Spjd acb->acb_buf = abuf; 3542168404Spjd abuf = NULL; 3543168404Spjd } 3544168404Spjd } 3545286570Smav hdr->b_l1hdr.b_acb = NULL; 3546275811Sdelphij hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 3547168404Spjd ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3548219089Spjd if (abuf == buf) { 3549219089Spjd ASSERT(buf->b_efunc == NULL); 3550286570Smav ASSERT(hdr->b_l1hdr.b_datacnt == 1); 3551275811Sdelphij hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 3552219089Spjd } 3553168404Spjd 3554286570Smav ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 3555286570Smav callback_list != NULL); 3556168404Spjd 3557168404Spjd if (zio->io_error != 0) { 3558275811Sdelphij hdr->b_flags |= ARC_FLAG_IO_ERROR; 3559286570Smav if (hdr->b_l1hdr.b_state != arc_anon) 3560168404Spjd arc_change_state(arc_anon, hdr, hash_lock); 3561168404Spjd if (HDR_IN_HASH_TABLE(hdr)) 3562168404Spjd buf_hash_remove(hdr); 3563286570Smav freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 3564168404Spjd } 3565168404Spjd 3566168404Spjd /* 3567168404Spjd * Broadcast before we drop the hash_lock to avoid the possibility 3568168404Spjd * that the hdr (and hence the cv) might be freed before we get to 3569168404Spjd * the cv_broadcast(). 3570168404Spjd */ 3571286570Smav cv_broadcast(&hdr->b_l1hdr.b_cv); 3572168404Spjd 3573286570Smav if (hash_lock != NULL) { 3574168404Spjd mutex_exit(hash_lock); 3575168404Spjd } else { 3576168404Spjd /* 3577168404Spjd * This block was freed while we waited for the read to 3578168404Spjd * complete. It has been removed from the hash table and 3579168404Spjd * moved to the anonymous state (so that it won't show up 3580168404Spjd * in the cache). 3581168404Spjd */ 3582286570Smav ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 3583286570Smav freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 3584168404Spjd } 3585168404Spjd 3586168404Spjd /* execute each callback and free its structure */ 3587168404Spjd while ((acb = callback_list) != NULL) { 3588168404Spjd if (acb->acb_done) 3589168404Spjd acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3590168404Spjd 3591168404Spjd if (acb->acb_zio_dummy != NULL) { 3592168404Spjd acb->acb_zio_dummy->io_error = zio->io_error; 3593168404Spjd zio_nowait(acb->acb_zio_dummy); 3594168404Spjd } 3595168404Spjd 3596168404Spjd callback_list = acb->acb_next; 3597168404Spjd kmem_free(acb, sizeof (arc_callback_t)); 3598168404Spjd } 3599168404Spjd 3600168404Spjd if (freeable) 3601168404Spjd arc_hdr_destroy(hdr); 3602168404Spjd} 3603168404Spjd 3604168404Spjd/* 3605168404Spjd * "Read" the block block at the specified DVA (in bp) via the 3606168404Spjd * cache. If the block is found in the cache, invoke the provided 3607168404Spjd * callback immediately and return. Note that the `zio' parameter 3608168404Spjd * in the callback will be NULL in this case, since no IO was 3609168404Spjd * required. If the block is not in the cache pass the read request 3610168404Spjd * on to the spa with a substitute callback function, so that the 3611168404Spjd * requested block will be added to the cache. 3612168404Spjd * 3613168404Spjd * If a read request arrives for a block that has a read in-progress, 3614168404Spjd * either wait for the in-progress read to complete (and return the 3615168404Spjd * results); or, if this is a read with a "done" func, add a record 3616168404Spjd * to the read to invoke the "done" func when the read completes, 3617168404Spjd * and return; or just return. 3618168404Spjd * 3619168404Spjd * arc_read_done() will invoke all the requested "done" functions 3620168404Spjd * for readers of this block. 3621168404Spjd */ 3622168404Spjdint 3623246666Smmarc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3624275811Sdelphij void *private, zio_priority_t priority, int zio_flags, 3625275811Sdelphij arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 3626168404Spjd{ 3627268075Sdelphij arc_buf_hdr_t *hdr = NULL; 3628247187Smm arc_buf_t *buf = NULL; 3629268075Sdelphij kmutex_t *hash_lock = NULL; 3630185029Spjd zio_t *rzio; 3631228103Smm uint64_t guid = spa_load_guid(spa); 3632168404Spjd 3633268075Sdelphij ASSERT(!BP_IS_EMBEDDED(bp) || 3634268075Sdelphij BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3635268075Sdelphij 3636168404Spjdtop: 3637268075Sdelphij if (!BP_IS_EMBEDDED(bp)) { 3638268075Sdelphij /* 3639268075Sdelphij * Embedded BP's have no DVA and require no I/O to "read". 3640268075Sdelphij * Create an anonymous arc buf to back it. 3641268075Sdelphij */ 3642268075Sdelphij hdr = buf_hash_find(guid, bp, &hash_lock); 3643268075Sdelphij } 3644168404Spjd 3645286570Smav if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) { 3646268075Sdelphij 3647275811Sdelphij *arc_flags |= ARC_FLAG_CACHED; 3648168404Spjd 3649168404Spjd if (HDR_IO_IN_PROGRESS(hdr)) { 3650168404Spjd 3651275811Sdelphij if (*arc_flags & ARC_FLAG_WAIT) { 3652286570Smav cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 3653168404Spjd mutex_exit(hash_lock); 3654168404Spjd goto top; 3655168404Spjd } 3656275811Sdelphij ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 3657168404Spjd 3658168404Spjd if (done) { 3659168404Spjd arc_callback_t *acb = NULL; 3660168404Spjd 3661168404Spjd acb = kmem_zalloc(sizeof (arc_callback_t), 3662168404Spjd KM_SLEEP); 3663168404Spjd acb->acb_done = done; 3664168404Spjd acb->acb_private = private; 3665168404Spjd if (pio != NULL) 3666168404Spjd acb->acb_zio_dummy = zio_null(pio, 3667209962Smm spa, NULL, NULL, NULL, zio_flags); 3668168404Spjd 3669168404Spjd ASSERT(acb->acb_done != NULL); 3670286570Smav acb->acb_next = hdr->b_l1hdr.b_acb; 3671286570Smav hdr->b_l1hdr.b_acb = acb; 3672168404Spjd add_reference(hdr, hash_lock, private); 3673168404Spjd mutex_exit(hash_lock); 3674168404Spjd return (0); 3675168404Spjd } 3676168404Spjd mutex_exit(hash_lock); 3677168404Spjd return (0); 3678168404Spjd } 3679168404Spjd 3680286570Smav ASSERT(hdr->b_l1hdr.b_state == arc_mru || 3681286570Smav hdr->b_l1hdr.b_state == arc_mfu); 3682168404Spjd 3683168404Spjd if (done) { 3684168404Spjd add_reference(hdr, hash_lock, private); 3685168404Spjd /* 3686168404Spjd * If this block is already in use, create a new 3687168404Spjd * copy of the data so that we will be guaranteed 3688168404Spjd * that arc_release() will always succeed. 3689168404Spjd */ 3690286570Smav buf = hdr->b_l1hdr.b_buf; 3691168404Spjd ASSERT(buf); 3692168404Spjd ASSERT(buf->b_data); 3693168404Spjd if (HDR_BUF_AVAILABLE(hdr)) { 3694168404Spjd ASSERT(buf->b_efunc == NULL); 3695275811Sdelphij hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 3696168404Spjd } else { 3697168404Spjd buf = arc_buf_clone(buf); 3698168404Spjd } 3699219089Spjd 3700275811Sdelphij } else if (*arc_flags & ARC_FLAG_PREFETCH && 3701286570Smav refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 3702275811Sdelphij hdr->b_flags |= ARC_FLAG_PREFETCH; 3703168404Spjd } 3704168404Spjd DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3705168404Spjd arc_access(hdr, hash_lock); 3706275811Sdelphij if (*arc_flags & ARC_FLAG_L2CACHE) 3707275811Sdelphij hdr->b_flags |= ARC_FLAG_L2CACHE; 3708275811Sdelphij if (*arc_flags & ARC_FLAG_L2COMPRESS) 3709275811Sdelphij hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3710168404Spjd mutex_exit(hash_lock); 3711168404Spjd ARCSTAT_BUMP(arcstat_hits); 3712286570Smav ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 3713286570Smav demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 3714168404Spjd data, metadata, hits); 3715168404Spjd 3716168404Spjd if (done) 3717168404Spjd done(NULL, buf, private); 3718168404Spjd } else { 3719168404Spjd uint64_t size = BP_GET_LSIZE(bp); 3720268075Sdelphij arc_callback_t *acb; 3721185029Spjd vdev_t *vd = NULL; 3722247187Smm uint64_t addr = 0; 3723208373Smm boolean_t devw = B_FALSE; 3724258389Savg enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3725286570Smav int32_t b_asize = 0; 3726168404Spjd 3727168404Spjd if (hdr == NULL) { 3728168404Spjd /* this block is not in the cache */ 3729268075Sdelphij arc_buf_hdr_t *exists = NULL; 3730168404Spjd arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3731168404Spjd buf = arc_buf_alloc(spa, size, private, type); 3732168404Spjd hdr = buf->b_hdr; 3733268075Sdelphij if (!BP_IS_EMBEDDED(bp)) { 3734268075Sdelphij hdr->b_dva = *BP_IDENTITY(bp); 3735268075Sdelphij hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3736268075Sdelphij exists = buf_hash_insert(hdr, &hash_lock); 3737268075Sdelphij } 3738268075Sdelphij if (exists != NULL) { 3739168404Spjd /* somebody beat us to the hash insert */ 3740168404Spjd mutex_exit(hash_lock); 3741219089Spjd buf_discard_identity(hdr); 3742168404Spjd (void) arc_buf_remove_ref(buf, private); 3743168404Spjd goto top; /* restart the IO request */ 3744168404Spjd } 3745275811Sdelphij 3746168404Spjd /* if this is a prefetch, we don't have a reference */ 3747275811Sdelphij if (*arc_flags & ARC_FLAG_PREFETCH) { 3748168404Spjd (void) remove_reference(hdr, hash_lock, 3749168404Spjd private); 3750275811Sdelphij hdr->b_flags |= ARC_FLAG_PREFETCH; 3751168404Spjd } 3752275811Sdelphij if (*arc_flags & ARC_FLAG_L2CACHE) 3753275811Sdelphij hdr->b_flags |= ARC_FLAG_L2CACHE; 3754275811Sdelphij if (*arc_flags & ARC_FLAG_L2COMPRESS) 3755275811Sdelphij hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3756168404Spjd if (BP_GET_LEVEL(bp) > 0) 3757275811Sdelphij hdr->b_flags |= ARC_FLAG_INDIRECT; 3758168404Spjd } else { 3759286570Smav /* 3760286570Smav * This block is in the ghost cache. If it was L2-only 3761286570Smav * (and thus didn't have an L1 hdr), we realloc the 3762286570Smav * header to add an L1 hdr. 3763286570Smav */ 3764286570Smav if (!HDR_HAS_L1HDR(hdr)) { 3765286570Smav hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 3766286570Smav hdr_full_cache); 3767286570Smav } 3768286570Smav 3769286570Smav ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 3770168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3771286570Smav ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3772286570Smav ASSERT(hdr->b_l1hdr.b_buf == NULL); 3773168404Spjd 3774168404Spjd /* if this is a prefetch, we don't have a reference */ 3775275811Sdelphij if (*arc_flags & ARC_FLAG_PREFETCH) 3776275811Sdelphij hdr->b_flags |= ARC_FLAG_PREFETCH; 3777168404Spjd else 3778168404Spjd add_reference(hdr, hash_lock, private); 3779275811Sdelphij if (*arc_flags & ARC_FLAG_L2CACHE) 3780275811Sdelphij hdr->b_flags |= ARC_FLAG_L2CACHE; 3781275811Sdelphij if (*arc_flags & ARC_FLAG_L2COMPRESS) 3782275811Sdelphij hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3783185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3784168404Spjd buf->b_hdr = hdr; 3785168404Spjd buf->b_data = NULL; 3786168404Spjd buf->b_efunc = NULL; 3787168404Spjd buf->b_private = NULL; 3788168404Spjd buf->b_next = NULL; 3789286570Smav hdr->b_l1hdr.b_buf = buf; 3790286570Smav ASSERT0(hdr->b_l1hdr.b_datacnt); 3791286570Smav hdr->b_l1hdr.b_datacnt = 1; 3792219089Spjd arc_get_data_buf(buf); 3793219089Spjd arc_access(hdr, hash_lock); 3794168404Spjd } 3795168404Spjd 3796286570Smav ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 3797219089Spjd 3798168404Spjd acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3799168404Spjd acb->acb_done = done; 3800168404Spjd acb->acb_private = private; 3801168404Spjd 3802286570Smav ASSERT(hdr->b_l1hdr.b_acb == NULL); 3803286570Smav hdr->b_l1hdr.b_acb = acb; 3804275811Sdelphij hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 3805168404Spjd 3806286570Smav if (HDR_HAS_L2HDR(hdr) && 3807286570Smav (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 3808286570Smav devw = hdr->b_l2hdr.b_dev->l2ad_writing; 3809286570Smav addr = hdr->b_l2hdr.b_daddr; 3810286570Smav b_compress = HDR_GET_COMPRESS(hdr); 3811286570Smav b_asize = hdr->b_l2hdr.b_asize; 3812185029Spjd /* 3813185029Spjd * Lock out device removal. 3814185029Spjd */ 3815185029Spjd if (vdev_is_dead(vd) || 3816185029Spjd !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3817185029Spjd vd = NULL; 3818185029Spjd } 3819185029Spjd 3820268075Sdelphij if (hash_lock != NULL) 3821268075Sdelphij mutex_exit(hash_lock); 3822168404Spjd 3823251629Sdelphij /* 3824251629Sdelphij * At this point, we have a level 1 cache miss. Try again in 3825251629Sdelphij * L2ARC if possible. 3826251629Sdelphij */ 3827168404Spjd ASSERT3U(hdr->b_size, ==, size); 3828219089Spjd DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3829268123Sdelphij uint64_t, size, zbookmark_phys_t *, zb); 3830168404Spjd ARCSTAT_BUMP(arcstat_misses); 3831286570Smav ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 3832286570Smav demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 3833168404Spjd data, metadata, misses); 3834228392Spjd#ifdef _KERNEL 3835228392Spjd curthread->td_ru.ru_inblock++; 3836228392Spjd#endif 3837168404Spjd 3838208373Smm if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3839185029Spjd /* 3840185029Spjd * Read from the L2ARC if the following are true: 3841185029Spjd * 1. The L2ARC vdev was previously cached. 3842185029Spjd * 2. This buffer still has L2ARC metadata. 3843185029Spjd * 3. This buffer isn't currently writing to the L2ARC. 3844185029Spjd * 4. The L2ARC entry wasn't evicted, which may 3845185029Spjd * also have invalidated the vdev. 3846208373Smm * 5. This isn't prefetch and l2arc_noprefetch is set. 3847185029Spjd */ 3848286570Smav if (HDR_HAS_L2HDR(hdr) && 3849208373Smm !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3850208373Smm !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3851185029Spjd l2arc_read_callback_t *cb; 3852185029Spjd 3853185029Spjd DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3854185029Spjd ARCSTAT_BUMP(arcstat_l2_hits); 3855185029Spjd 3856185029Spjd cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3857185029Spjd KM_SLEEP); 3858185029Spjd cb->l2rcb_buf = buf; 3859185029Spjd cb->l2rcb_spa = spa; 3860185029Spjd cb->l2rcb_bp = *bp; 3861185029Spjd cb->l2rcb_zb = *zb; 3862185029Spjd cb->l2rcb_flags = zio_flags; 3863258389Savg cb->l2rcb_compress = b_compress; 3864185029Spjd 3865247187Smm ASSERT(addr >= VDEV_LABEL_START_SIZE && 3866247187Smm addr + size < vd->vdev_psize - 3867247187Smm VDEV_LABEL_END_SIZE); 3868247187Smm 3869185029Spjd /* 3870185029Spjd * l2arc read. The SCL_L2ARC lock will be 3871185029Spjd * released by l2arc_read_done(). 3872251478Sdelphij * Issue a null zio if the underlying buffer 3873251478Sdelphij * was squashed to zero size by compression. 3874185029Spjd */ 3875258389Savg if (b_compress == ZIO_COMPRESS_EMPTY) { 3876251478Sdelphij rzio = zio_null(pio, spa, vd, 3877251478Sdelphij l2arc_read_done, cb, 3878251478Sdelphij zio_flags | ZIO_FLAG_DONT_CACHE | 3879251478Sdelphij ZIO_FLAG_CANFAIL | 3880251478Sdelphij ZIO_FLAG_DONT_PROPAGATE | 3881251478Sdelphij ZIO_FLAG_DONT_RETRY); 3882251478Sdelphij } else { 3883251478Sdelphij rzio = zio_read_phys(pio, vd, addr, 3884258389Savg b_asize, buf->b_data, 3885258389Savg ZIO_CHECKSUM_OFF, 3886251478Sdelphij l2arc_read_done, cb, priority, 3887251478Sdelphij zio_flags | ZIO_FLAG_DONT_CACHE | 3888251478Sdelphij ZIO_FLAG_CANFAIL | 3889251478Sdelphij ZIO_FLAG_DONT_PROPAGATE | 3890251478Sdelphij ZIO_FLAG_DONT_RETRY, B_FALSE); 3891251478Sdelphij } 3892185029Spjd DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3893185029Spjd zio_t *, rzio); 3894258389Savg ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3895185029Spjd 3896275811Sdelphij if (*arc_flags & ARC_FLAG_NOWAIT) { 3897185029Spjd zio_nowait(rzio); 3898185029Spjd return (0); 3899185029Spjd } 3900185029Spjd 3901275811Sdelphij ASSERT(*arc_flags & ARC_FLAG_WAIT); 3902185029Spjd if (zio_wait(rzio) == 0) 3903185029Spjd return (0); 3904185029Spjd 3905185029Spjd /* l2arc read error; goto zio_read() */ 3906185029Spjd } else { 3907185029Spjd DTRACE_PROBE1(l2arc__miss, 3908185029Spjd arc_buf_hdr_t *, hdr); 3909185029Spjd ARCSTAT_BUMP(arcstat_l2_misses); 3910185029Spjd if (HDR_L2_WRITING(hdr)) 3911185029Spjd ARCSTAT_BUMP(arcstat_l2_rw_clash); 3912185029Spjd spa_config_exit(spa, SCL_L2ARC, vd); 3913185029Spjd } 3914208373Smm } else { 3915208373Smm if (vd != NULL) 3916208373Smm spa_config_exit(spa, SCL_L2ARC, vd); 3917208373Smm if (l2arc_ndev != 0) { 3918208373Smm DTRACE_PROBE1(l2arc__miss, 3919208373Smm arc_buf_hdr_t *, hdr); 3920208373Smm ARCSTAT_BUMP(arcstat_l2_misses); 3921208373Smm } 3922185029Spjd } 3923185029Spjd 3924168404Spjd rzio = zio_read(pio, spa, bp, buf->b_data, size, 3925185029Spjd arc_read_done, buf, priority, zio_flags, zb); 3926168404Spjd 3927275811Sdelphij if (*arc_flags & ARC_FLAG_WAIT) 3928168404Spjd return (zio_wait(rzio)); 3929168404Spjd 3930275811Sdelphij ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 3931168404Spjd zio_nowait(rzio); 3932168404Spjd } 3933168404Spjd return (0); 3934168404Spjd} 3935168404Spjd 3936168404Spjdvoid 3937168404Spjdarc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3938168404Spjd{ 3939168404Spjd ASSERT(buf->b_hdr != NULL); 3940286570Smav ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon); 3941286570Smav ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) || 3942286570Smav func == NULL); 3943219089Spjd ASSERT(buf->b_efunc == NULL); 3944219089Spjd ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3945219089Spjd 3946168404Spjd buf->b_efunc = func; 3947168404Spjd buf->b_private = private; 3948168404Spjd} 3949168404Spjd 3950168404Spjd/* 3951251520Sdelphij * Notify the arc that a block was freed, and thus will never be used again. 3952251520Sdelphij */ 3953251520Sdelphijvoid 3954251520Sdelphijarc_freed(spa_t *spa, const blkptr_t *bp) 3955251520Sdelphij{ 3956251520Sdelphij arc_buf_hdr_t *hdr; 3957251520Sdelphij kmutex_t *hash_lock; 3958251520Sdelphij uint64_t guid = spa_load_guid(spa); 3959251520Sdelphij 3960268075Sdelphij ASSERT(!BP_IS_EMBEDDED(bp)); 3961268075Sdelphij 3962268075Sdelphij hdr = buf_hash_find(guid, bp, &hash_lock); 3963251520Sdelphij if (hdr == NULL) 3964251520Sdelphij return; 3965251520Sdelphij if (HDR_BUF_AVAILABLE(hdr)) { 3966286570Smav arc_buf_t *buf = hdr->b_l1hdr.b_buf; 3967251520Sdelphij add_reference(hdr, hash_lock, FTAG); 3968275811Sdelphij hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 3969251520Sdelphij mutex_exit(hash_lock); 3970251520Sdelphij 3971251520Sdelphij arc_release(buf, FTAG); 3972251520Sdelphij (void) arc_buf_remove_ref(buf, FTAG); 3973251520Sdelphij } else { 3974251520Sdelphij mutex_exit(hash_lock); 3975251520Sdelphij } 3976251520Sdelphij 3977251520Sdelphij} 3978251520Sdelphij 3979251520Sdelphij/* 3980268858Sdelphij * Clear the user eviction callback set by arc_set_callback(), first calling 3981268858Sdelphij * it if it exists. Because the presence of a callback keeps an arc_buf cached 3982268858Sdelphij * clearing the callback may result in the arc_buf being destroyed. However, 3983268858Sdelphij * it will not result in the *last* arc_buf being destroyed, hence the data 3984268858Sdelphij * will remain cached in the ARC. We make a copy of the arc buffer here so 3985268858Sdelphij * that we can process the callback without holding any locks. 3986268858Sdelphij * 3987268858Sdelphij * It's possible that the callback is already in the process of being cleared 3988268858Sdelphij * by another thread. In this case we can not clear the callback. 3989268858Sdelphij * 3990268858Sdelphij * Returns B_TRUE if the callback was successfully called and cleared. 3991168404Spjd */ 3992268858Sdelphijboolean_t 3993268858Sdelphijarc_clear_callback(arc_buf_t *buf) 3994168404Spjd{ 3995168404Spjd arc_buf_hdr_t *hdr; 3996168404Spjd kmutex_t *hash_lock; 3997268858Sdelphij arc_evict_func_t *efunc = buf->b_efunc; 3998268858Sdelphij void *private = buf->b_private; 3999205231Skmacy list_t *list, *evicted_list; 4000205231Skmacy kmutex_t *lock, *evicted_lock; 4001206796Spjd 4002219089Spjd mutex_enter(&buf->b_evict_lock); 4003168404Spjd hdr = buf->b_hdr; 4004168404Spjd if (hdr == NULL) { 4005168404Spjd /* 4006168404Spjd * We are in arc_do_user_evicts(). 4007168404Spjd */ 4008168404Spjd ASSERT(buf->b_data == NULL); 4009219089Spjd mutex_exit(&buf->b_evict_lock); 4010268858Sdelphij return (B_FALSE); 4011185029Spjd } else if (buf->b_data == NULL) { 4012185029Spjd /* 4013185029Spjd * We are on the eviction list; process this buffer now 4014185029Spjd * but let arc_do_user_evicts() do the reaping. 4015185029Spjd */ 4016185029Spjd buf->b_efunc = NULL; 4017219089Spjd mutex_exit(&buf->b_evict_lock); 4018268858Sdelphij VERIFY0(efunc(private)); 4019268858Sdelphij return (B_TRUE); 4020168404Spjd } 4021168404Spjd hash_lock = HDR_LOCK(hdr); 4022168404Spjd mutex_enter(hash_lock); 4023219089Spjd hdr = buf->b_hdr; 4024219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4025168404Spjd 4026286570Smav ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <, 4027286570Smav hdr->b_l1hdr.b_datacnt); 4028286570Smav ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4029286570Smav hdr->b_l1hdr.b_state == arc_mfu); 4030168404Spjd 4031268858Sdelphij buf->b_efunc = NULL; 4032268858Sdelphij buf->b_private = NULL; 4033168404Spjd 4034286570Smav if (hdr->b_l1hdr.b_datacnt > 1) { 4035268858Sdelphij mutex_exit(&buf->b_evict_lock); 4036268858Sdelphij arc_buf_destroy(buf, FALSE, TRUE); 4037268858Sdelphij } else { 4038286570Smav ASSERT(buf == hdr->b_l1hdr.b_buf); 4039275811Sdelphij hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4040268858Sdelphij mutex_exit(&buf->b_evict_lock); 4041268858Sdelphij } 4042168404Spjd 4043168404Spjd mutex_exit(hash_lock); 4044268858Sdelphij VERIFY0(efunc(private)); 4045268858Sdelphij return (B_TRUE); 4046168404Spjd} 4047168404Spjd 4048168404Spjd/* 4049251629Sdelphij * Release this buffer from the cache, making it an anonymous buffer. This 4050251629Sdelphij * must be done after a read and prior to modifying the buffer contents. 4051168404Spjd * If the buffer has more than one reference, we must make 4052185029Spjd * a new hdr for the buffer. 4053168404Spjd */ 4054168404Spjdvoid 4055168404Spjdarc_release(arc_buf_t *buf, void *tag) 4056168404Spjd{ 4057286570Smav arc_buf_hdr_t *hdr = buf->b_hdr; 4058168404Spjd 4059219089Spjd /* 4060219089Spjd * It would be nice to assert that if it's DMU metadata (level > 4061219089Spjd * 0 || it's the dnode file), then it must be syncing context. 4062219089Spjd * But we don't know that information at this level. 4063219089Spjd */ 4064219089Spjd 4065219089Spjd mutex_enter(&buf->b_evict_lock); 4066286570Smav /* 4067286570Smav * We don't grab the hash lock prior to this check, because if 4068286570Smav * the buffer's header is in the arc_anon state, it won't be 4069286570Smav * linked into the hash table. 4070286570Smav */ 4071286570Smav if (hdr->b_l1hdr.b_state == arc_anon) { 4072286570Smav mutex_exit(&buf->b_evict_lock); 4073286570Smav ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4074286570Smav ASSERT(!HDR_IN_HASH_TABLE(hdr)); 4075286570Smav ASSERT(!HDR_HAS_L2HDR(hdr)); 4076286570Smav ASSERT(BUF_EMPTY(hdr)); 4077286570Smav ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1); 4078286570Smav ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 4079286570Smav ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 4080185029Spjd 4081286570Smav ASSERT3P(buf->b_efunc, ==, NULL); 4082286570Smav ASSERT3P(buf->b_private, ==, NULL); 4083168404Spjd 4084286570Smav hdr->b_l1hdr.b_arc_access = 0; 4085286570Smav arc_buf_thaw(buf); 4086286570Smav 4087286570Smav return; 4088168404Spjd } 4089168404Spjd 4090286570Smav kmutex_t *hash_lock = HDR_LOCK(hdr); 4091286570Smav mutex_enter(hash_lock); 4092286570Smav 4093286570Smav /* 4094286570Smav * This assignment is only valid as long as the hash_lock is 4095286570Smav * held, we must be careful not to reference state or the 4096286570Smav * b_state field after dropping the lock. 4097286570Smav */ 4098286570Smav arc_state_t *state = hdr->b_l1hdr.b_state; 4099286570Smav ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4100286570Smav ASSERT3P(state, !=, arc_anon); 4101286570Smav 4102286570Smav /* this buffer is not on any list */ 4103286570Smav ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0); 4104286570Smav 4105286570Smav if (HDR_HAS_L2HDR(hdr)) { 4106286570Smav ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize); 4107286570Smav ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 4108286570Smav 4109286570Smav mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4110286570Smav trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev, 4111286570Smav hdr->b_l2hdr.b_daddr, hdr->b_l2hdr.b_asize, 0); 4112286570Smav list_remove(&hdr->b_l2hdr.b_dev->l2ad_buflist, hdr); 4113286570Smav 4114286570Smav /* 4115286570Smav * We don't want to leak the b_tmp_cdata buffer that was 4116286570Smav * allocated in l2arc_write_buffers() 4117286570Smav */ 4118274172Savg arc_buf_l2_cdata_free(hdr); 4119286570Smav 4120286570Smav mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4121286570Smav 4122286570Smav hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 4123185029Spjd } 4124185029Spjd 4125168404Spjd /* 4126168404Spjd * Do we have more than one buf? 4127168404Spjd */ 4128286570Smav if (hdr->b_l1hdr.b_datacnt > 1) { 4129168404Spjd arc_buf_hdr_t *nhdr; 4130168404Spjd arc_buf_t **bufp; 4131168404Spjd uint64_t blksz = hdr->b_size; 4132209962Smm uint64_t spa = hdr->b_spa; 4133286570Smav arc_buf_contents_t type = arc_buf_type(hdr); 4134185029Spjd uint32_t flags = hdr->b_flags; 4135168404Spjd 4136286570Smav ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 4137168404Spjd /* 4138219089Spjd * Pull the data off of this hdr and attach it to 4139219089Spjd * a new anonymous hdr. 4140168404Spjd */ 4141168404Spjd (void) remove_reference(hdr, hash_lock, tag); 4142286570Smav bufp = &hdr->b_l1hdr.b_buf; 4143168404Spjd while (*bufp != buf) 4144168404Spjd bufp = &(*bufp)->b_next; 4145219089Spjd *bufp = buf->b_next; 4146168404Spjd buf->b_next = NULL; 4147168404Spjd 4148286570Smav ASSERT3P(state, !=, arc_l2c_only); 4149286570Smav ASSERT3U(state->arcs_size, >=, hdr->b_size); 4150286570Smav atomic_add_64(&state->arcs_size, -hdr->b_size); 4151286570Smav if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 4152286570Smav ASSERT3P(state, !=, arc_l2c_only); 4153286570Smav uint64_t *size = &state->arcs_lsize[type]; 4154185029Spjd ASSERT3U(*size, >=, hdr->b_size); 4155185029Spjd atomic_add_64(size, -hdr->b_size); 4156168404Spjd } 4157242845Sdelphij 4158242845Sdelphij /* 4159242845Sdelphij * We're releasing a duplicate user data buffer, update 4160242845Sdelphij * our statistics accordingly. 4161242845Sdelphij */ 4162286570Smav if (HDR_ISTYPE_DATA(hdr)) { 4163242845Sdelphij ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 4164242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, 4165242845Sdelphij -hdr->b_size); 4166242845Sdelphij } 4167286570Smav hdr->b_l1hdr.b_datacnt -= 1; 4168168404Spjd arc_cksum_verify(buf); 4169240133Smm#ifdef illumos 4170240133Smm arc_buf_unwatch(buf); 4171277300Ssmh#endif 4172168404Spjd 4173168404Spjd mutex_exit(hash_lock); 4174168404Spjd 4175286570Smav nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 4176168404Spjd nhdr->b_size = blksz; 4177168404Spjd nhdr->b_spa = spa; 4178286570Smav 4179275811Sdelphij nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; 4180286570Smav nhdr->b_flags |= arc_bufc_to_flags(type); 4181286570Smav nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 4182286570Smav 4183286570Smav nhdr->b_l1hdr.b_buf = buf; 4184286570Smav nhdr->b_l1hdr.b_datacnt = 1; 4185286570Smav nhdr->b_l1hdr.b_state = arc_anon; 4186286570Smav nhdr->b_l1hdr.b_arc_access = 0; 4187168404Spjd nhdr->b_freeze_cksum = NULL; 4188286570Smav 4189286570Smav (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 4190168404Spjd buf->b_hdr = nhdr; 4191219089Spjd mutex_exit(&buf->b_evict_lock); 4192168404Spjd atomic_add_64(&arc_anon->arcs_size, blksz); 4193168404Spjd } else { 4194219089Spjd mutex_exit(&buf->b_evict_lock); 4195286570Smav ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 4196286570Smav /* protected by hash lock */ 4197286570Smav ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 4198168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4199286570Smav arc_change_state(arc_anon, hdr, hash_lock); 4200286570Smav hdr->b_l1hdr.b_arc_access = 0; 4201286570Smav mutex_exit(hash_lock); 4202185029Spjd 4203219089Spjd buf_discard_identity(hdr); 4204168404Spjd arc_buf_thaw(buf); 4205168404Spjd } 4206168404Spjd buf->b_efunc = NULL; 4207168404Spjd buf->b_private = NULL; 4208168404Spjd} 4209168404Spjd 4210168404Spjdint 4211168404Spjdarc_released(arc_buf_t *buf) 4212168404Spjd{ 4213185029Spjd int released; 4214185029Spjd 4215219089Spjd mutex_enter(&buf->b_evict_lock); 4216286570Smav released = (buf->b_data != NULL && 4217286570Smav buf->b_hdr->b_l1hdr.b_state == arc_anon); 4218219089Spjd mutex_exit(&buf->b_evict_lock); 4219185029Spjd return (released); 4220168404Spjd} 4221168404Spjd 4222168404Spjd#ifdef ZFS_DEBUG 4223168404Spjdint 4224168404Spjdarc_referenced(arc_buf_t *buf) 4225168404Spjd{ 4226185029Spjd int referenced; 4227185029Spjd 4228219089Spjd mutex_enter(&buf->b_evict_lock); 4229286570Smav referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 4230219089Spjd mutex_exit(&buf->b_evict_lock); 4231185029Spjd return (referenced); 4232168404Spjd} 4233168404Spjd#endif 4234168404Spjd 4235168404Spjdstatic void 4236168404Spjdarc_write_ready(zio_t *zio) 4237168404Spjd{ 4238168404Spjd arc_write_callback_t *callback = zio->io_private; 4239168404Spjd arc_buf_t *buf = callback->awcb_buf; 4240185029Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 4241168404Spjd 4242286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 4243286570Smav ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 4244286570Smav ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4245185029Spjd callback->awcb_ready(zio, buf, callback->awcb_private); 4246185029Spjd 4247185029Spjd /* 4248185029Spjd * If the IO is already in progress, then this is a re-write 4249185029Spjd * attempt, so we need to thaw and re-compute the cksum. 4250185029Spjd * It is the responsibility of the callback to handle the 4251185029Spjd * accounting for any re-write attempt. 4252185029Spjd */ 4253185029Spjd if (HDR_IO_IN_PROGRESS(hdr)) { 4254286570Smav mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 4255185029Spjd if (hdr->b_freeze_cksum != NULL) { 4256185029Spjd kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 4257185029Spjd hdr->b_freeze_cksum = NULL; 4258185029Spjd } 4259286570Smav mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 4260168404Spjd } 4261185029Spjd arc_cksum_compute(buf, B_FALSE); 4262275811Sdelphij hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4263168404Spjd} 4264168404Spjd 4265258632Savg/* 4266258632Savg * The SPA calls this callback for each physical write that happens on behalf 4267258632Savg * of a logical write. See the comment in dbuf_write_physdone() for details. 4268258632Savg */ 4269168404Spjdstatic void 4270258632Savgarc_write_physdone(zio_t *zio) 4271258632Savg{ 4272258632Savg arc_write_callback_t *cb = zio->io_private; 4273258632Savg if (cb->awcb_physdone != NULL) 4274258632Savg cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 4275258632Savg} 4276258632Savg 4277258632Savgstatic void 4278168404Spjdarc_write_done(zio_t *zio) 4279168404Spjd{ 4280168404Spjd arc_write_callback_t *callback = zio->io_private; 4281168404Spjd arc_buf_t *buf = callback->awcb_buf; 4282168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 4283168404Spjd 4284286570Smav ASSERT(hdr->b_l1hdr.b_acb == NULL); 4285168404Spjd 4286219089Spjd if (zio->io_error == 0) { 4287268075Sdelphij if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 4288260150Sdelphij buf_discard_identity(hdr); 4289260150Sdelphij } else { 4290260150Sdelphij hdr->b_dva = *BP_IDENTITY(zio->io_bp); 4291260150Sdelphij hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 4292260150Sdelphij } 4293219089Spjd } else { 4294219089Spjd ASSERT(BUF_EMPTY(hdr)); 4295219089Spjd } 4296219089Spjd 4297168404Spjd /* 4298268075Sdelphij * If the block to be written was all-zero or compressed enough to be 4299268075Sdelphij * embedded in the BP, no write was performed so there will be no 4300268075Sdelphij * dva/birth/checksum. The buffer must therefore remain anonymous 4301268075Sdelphij * (and uncached). 4302168404Spjd */ 4303168404Spjd if (!BUF_EMPTY(hdr)) { 4304168404Spjd arc_buf_hdr_t *exists; 4305168404Spjd kmutex_t *hash_lock; 4306168404Spjd 4307219089Spjd ASSERT(zio->io_error == 0); 4308219089Spjd 4309168404Spjd arc_cksum_verify(buf); 4310168404Spjd 4311168404Spjd exists = buf_hash_insert(hdr, &hash_lock); 4312286570Smav if (exists != NULL) { 4313168404Spjd /* 4314168404Spjd * This can only happen if we overwrite for 4315168404Spjd * sync-to-convergence, because we remove 4316168404Spjd * buffers from the hash table when we arc_free(). 4317168404Spjd */ 4318219089Spjd if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 4319219089Spjd if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 4320219089Spjd panic("bad overwrite, hdr=%p exists=%p", 4321219089Spjd (void *)hdr, (void *)exists); 4322286570Smav ASSERT(refcount_is_zero( 4323286570Smav &exists->b_l1hdr.b_refcnt)); 4324219089Spjd arc_change_state(arc_anon, exists, hash_lock); 4325219089Spjd mutex_exit(hash_lock); 4326219089Spjd arc_hdr_destroy(exists); 4327219089Spjd exists = buf_hash_insert(hdr, &hash_lock); 4328219089Spjd ASSERT3P(exists, ==, NULL); 4329243524Smm } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 4330243524Smm /* nopwrite */ 4331243524Smm ASSERT(zio->io_prop.zp_nopwrite); 4332243524Smm if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 4333243524Smm panic("bad nopwrite, hdr=%p exists=%p", 4334243524Smm (void *)hdr, (void *)exists); 4335219089Spjd } else { 4336219089Spjd /* Dedup */ 4337286570Smav ASSERT(hdr->b_l1hdr.b_datacnt == 1); 4338286570Smav ASSERT(hdr->b_l1hdr.b_state == arc_anon); 4339219089Spjd ASSERT(BP_GET_DEDUP(zio->io_bp)); 4340219089Spjd ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 4341219089Spjd } 4342168404Spjd } 4343275811Sdelphij hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4344185029Spjd /* if it's not anon, we are doing a scrub */ 4345286570Smav if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 4346185029Spjd arc_access(hdr, hash_lock); 4347168404Spjd mutex_exit(hash_lock); 4348168404Spjd } else { 4349275811Sdelphij hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4350168404Spjd } 4351168404Spjd 4352286570Smav ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4353219089Spjd callback->awcb_done(zio, buf, callback->awcb_private); 4354168404Spjd 4355168404Spjd kmem_free(callback, sizeof (arc_write_callback_t)); 4356168404Spjd} 4357168404Spjd 4358168404Spjdzio_t * 4359219089Spjdarc_write(zio_t *pio, spa_t *spa, uint64_t txg, 4360251478Sdelphij blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 4361258632Savg const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 4362258632Savg arc_done_func_t *done, void *private, zio_priority_t priority, 4363268123Sdelphij int zio_flags, const zbookmark_phys_t *zb) 4364168404Spjd{ 4365168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 4366168404Spjd arc_write_callback_t *callback; 4367185029Spjd zio_t *zio; 4368168404Spjd 4369185029Spjd ASSERT(ready != NULL); 4370219089Spjd ASSERT(done != NULL); 4371168404Spjd ASSERT(!HDR_IO_ERROR(hdr)); 4372286570Smav ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4373286570Smav ASSERT(hdr->b_l1hdr.b_acb == NULL); 4374286570Smav ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4375185029Spjd if (l2arc) 4376275811Sdelphij hdr->b_flags |= ARC_FLAG_L2CACHE; 4377251478Sdelphij if (l2arc_compress) 4378275811Sdelphij hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4379168404Spjd callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 4380168404Spjd callback->awcb_ready = ready; 4381258632Savg callback->awcb_physdone = physdone; 4382168404Spjd callback->awcb_done = done; 4383168404Spjd callback->awcb_private = private; 4384168404Spjd callback->awcb_buf = buf; 4385168404Spjd 4386219089Spjd zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 4387258632Savg arc_write_ready, arc_write_physdone, arc_write_done, callback, 4388258632Savg priority, zio_flags, zb); 4389185029Spjd 4390168404Spjd return (zio); 4391168404Spjd} 4392168404Spjd 4393185029Spjdstatic int 4394258632Savgarc_memory_throttle(uint64_t reserve, uint64_t txg) 4395185029Spjd{ 4396185029Spjd#ifdef _KERNEL 4397272483Ssmh uint64_t available_memory = ptob(freemem); 4398185029Spjd static uint64_t page_load = 0; 4399185029Spjd static uint64_t last_txg = 0; 4400185029Spjd 4401272483Ssmh#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 4402185029Spjd available_memory = 4403272483Ssmh MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 4404185029Spjd#endif 4405258632Savg 4406272483Ssmh if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 4407185029Spjd return (0); 4408185029Spjd 4409185029Spjd if (txg > last_txg) { 4410185029Spjd last_txg = txg; 4411185029Spjd page_load = 0; 4412185029Spjd } 4413185029Spjd /* 4414185029Spjd * If we are in pageout, we know that memory is already tight, 4415185029Spjd * the arc is already going to be evicting, so we just want to 4416185029Spjd * continue to let page writes occur as quickly as possible. 4417185029Spjd */ 4418185029Spjd if (curproc == pageproc) { 4419272483Ssmh if (page_load > MAX(ptob(minfree), available_memory) / 4) 4420249195Smm return (SET_ERROR(ERESTART)); 4421185029Spjd /* Note: reserve is inflated, so we deflate */ 4422185029Spjd page_load += reserve / 8; 4423185029Spjd return (0); 4424185029Spjd } else if (page_load > 0 && arc_reclaim_needed()) { 4425185029Spjd /* memory is low, delay before restarting */ 4426185029Spjd ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 4427249195Smm return (SET_ERROR(EAGAIN)); 4428185029Spjd } 4429185029Spjd page_load = 0; 4430185029Spjd#endif 4431185029Spjd return (0); 4432185029Spjd} 4433185029Spjd 4434168404Spjdvoid 4435185029Spjdarc_tempreserve_clear(uint64_t reserve) 4436168404Spjd{ 4437185029Spjd atomic_add_64(&arc_tempreserve, -reserve); 4438168404Spjd ASSERT((int64_t)arc_tempreserve >= 0); 4439168404Spjd} 4440168404Spjd 4441168404Spjdint 4442185029Spjdarc_tempreserve_space(uint64_t reserve, uint64_t txg) 4443168404Spjd{ 4444185029Spjd int error; 4445209962Smm uint64_t anon_size; 4446185029Spjd 4447272483Ssmh if (reserve > arc_c/4 && !arc_no_grow) { 4448185029Spjd arc_c = MIN(arc_c_max, reserve * 4); 4449272483Ssmh DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 4450272483Ssmh } 4451185029Spjd if (reserve > arc_c) 4452249195Smm return (SET_ERROR(ENOMEM)); 4453168404Spjd 4454168404Spjd /* 4455209962Smm * Don't count loaned bufs as in flight dirty data to prevent long 4456209962Smm * network delays from blocking transactions that are ready to be 4457209962Smm * assigned to a txg. 4458209962Smm */ 4459209962Smm anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 4460209962Smm 4461209962Smm /* 4462185029Spjd * Writes will, almost always, require additional memory allocations 4463251631Sdelphij * in order to compress/encrypt/etc the data. We therefore need to 4464185029Spjd * make sure that there is sufficient available memory for this. 4465185029Spjd */ 4466258632Savg error = arc_memory_throttle(reserve, txg); 4467258632Savg if (error != 0) 4468185029Spjd return (error); 4469185029Spjd 4470185029Spjd /* 4471168404Spjd * Throttle writes when the amount of dirty data in the cache 4472168404Spjd * gets too large. We try to keep the cache less than half full 4473168404Spjd * of dirty blocks so that our sync times don't grow too large. 4474168404Spjd * Note: if two requests come in concurrently, we might let them 4475168404Spjd * both succeed, when one of them should fail. Not a huge deal. 4476168404Spjd */ 4477209962Smm 4478209962Smm if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 4479209962Smm anon_size > arc_c / 4) { 4480185029Spjd dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 4481185029Spjd "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 4482185029Spjd arc_tempreserve>>10, 4483185029Spjd arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 4484185029Spjd arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 4485185029Spjd reserve>>10, arc_c>>10); 4486249195Smm return (SET_ERROR(ERESTART)); 4487168404Spjd } 4488185029Spjd atomic_add_64(&arc_tempreserve, reserve); 4489168404Spjd return (0); 4490168404Spjd} 4491168404Spjd 4492168582Spjdstatic kmutex_t arc_lowmem_lock; 4493168404Spjd#ifdef _KERNEL 4494168566Spjdstatic eventhandler_tag arc_event_lowmem = NULL; 4495168404Spjd 4496168404Spjdstatic void 4497168566Spjdarc_lowmem(void *arg __unused, int howto __unused) 4498168404Spjd{ 4499168404Spjd 4500168566Spjd /* Serialize access via arc_lowmem_lock. */ 4501168566Spjd mutex_enter(&arc_lowmem_lock); 4502219089Spjd mutex_enter(&arc_reclaim_thr_lock); 4503185029Spjd needfree = 1; 4504272483Ssmh DTRACE_PROBE(arc__needfree); 4505168404Spjd cv_signal(&arc_reclaim_thr_cv); 4506241773Savg 4507241773Savg /* 4508241773Savg * It is unsafe to block here in arbitrary threads, because we can come 4509241773Savg * here from ARC itself and may hold ARC locks and thus risk a deadlock 4510241773Savg * with ARC reclaim thread. 4511241773Savg */ 4512241773Savg if (curproc == pageproc) { 4513241773Savg while (needfree) 4514241773Savg msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4515241773Savg } 4516219089Spjd mutex_exit(&arc_reclaim_thr_lock); 4517168566Spjd mutex_exit(&arc_lowmem_lock); 4518168404Spjd} 4519168404Spjd#endif 4520168404Spjd 4521168404Spjdvoid 4522168404Spjdarc_init(void) 4523168404Spjd{ 4524219089Spjd int i, prefetch_tunable_set = 0; 4525205231Skmacy 4526168404Spjd mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4527168404Spjd cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4528168566Spjd mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4529168404Spjd 4530168404Spjd /* Convert seconds to clock ticks */ 4531168404Spjd arc_min_prefetch_lifespan = 1 * hz; 4532168404Spjd 4533168404Spjd /* Start out with 1/8 of all memory */ 4534168566Spjd arc_c = kmem_size() / 8; 4535219089Spjd 4536277300Ssmh#ifdef illumos 4537192360Skmacy#ifdef _KERNEL 4538192360Skmacy /* 4539192360Skmacy * On architectures where the physical memory can be larger 4540192360Skmacy * than the addressable space (intel in 32-bit mode), we may 4541192360Skmacy * need to limit the cache to 1/8 of VM size. 4542192360Skmacy */ 4543192360Skmacy arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4544192360Skmacy#endif 4545277300Ssmh#endif /* illumos */ 4546168566Spjd /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4547280822Smav arc_c_min = MAX(arc_c / 4, 16 << 20); 4548168566Spjd /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4549280822Smav if (arc_c * 8 >= 1 << 30) 4550280822Smav arc_c_max = (arc_c * 8) - (1 << 30); 4551168404Spjd else 4552168404Spjd arc_c_max = arc_c_min; 4553175633Spjd arc_c_max = MAX(arc_c * 5, arc_c_max); 4554219089Spjd 4555168481Spjd#ifdef _KERNEL 4556168404Spjd /* 4557168404Spjd * Allow the tunables to override our calculations if they are 4558168566Spjd * reasonable (ie. over 16MB) 4559168404Spjd */ 4560280822Smav if (zfs_arc_max > 16 << 20 && zfs_arc_max < kmem_size()) 4561168404Spjd arc_c_max = zfs_arc_max; 4562280822Smav if (zfs_arc_min > 16 << 20 && zfs_arc_min <= arc_c_max) 4563168404Spjd arc_c_min = zfs_arc_min; 4564168481Spjd#endif 4565219089Spjd 4566168404Spjd arc_c = arc_c_max; 4567168404Spjd arc_p = (arc_c >> 1); 4568168404Spjd 4569185029Spjd /* limit meta-data to 1/4 of the arc capacity */ 4570185029Spjd arc_meta_limit = arc_c_max / 4; 4571185029Spjd 4572185029Spjd /* Allow the tunable to override if it is reasonable */ 4573185029Spjd if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4574185029Spjd arc_meta_limit = zfs_arc_meta_limit; 4575185029Spjd 4576185029Spjd if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4577185029Spjd arc_c_min = arc_meta_limit / 2; 4578185029Spjd 4579275780Sdelphij if (zfs_arc_meta_min > 0) { 4580275780Sdelphij arc_meta_min = zfs_arc_meta_min; 4581275780Sdelphij } else { 4582275780Sdelphij arc_meta_min = arc_c_min / 2; 4583275780Sdelphij } 4584275780Sdelphij 4585208373Smm if (zfs_arc_grow_retry > 0) 4586208373Smm arc_grow_retry = zfs_arc_grow_retry; 4587208373Smm 4588208373Smm if (zfs_arc_shrink_shift > 0) 4589208373Smm arc_shrink_shift = zfs_arc_shrink_shift; 4590208373Smm 4591208373Smm if (zfs_arc_p_min_shift > 0) 4592208373Smm arc_p_min_shift = zfs_arc_p_min_shift; 4593208373Smm 4594168404Spjd /* if kmem_flags are set, lets try to use less memory */ 4595168404Spjd if (kmem_debugging()) 4596168404Spjd arc_c = arc_c / 2; 4597168404Spjd if (arc_c < arc_c_min) 4598168404Spjd arc_c = arc_c_min; 4599168404Spjd 4600168473Spjd zfs_arc_min = arc_c_min; 4601168473Spjd zfs_arc_max = arc_c_max; 4602168473Spjd 4603168404Spjd arc_anon = &ARC_anon; 4604168404Spjd arc_mru = &ARC_mru; 4605168404Spjd arc_mru_ghost = &ARC_mru_ghost; 4606168404Spjd arc_mfu = &ARC_mfu; 4607168404Spjd arc_mfu_ghost = &ARC_mfu_ghost; 4608185029Spjd arc_l2c_only = &ARC_l2c_only; 4609168404Spjd arc_size = 0; 4610168404Spjd 4611205231Skmacy for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4612205231Skmacy mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4613205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4614205231Skmacy mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4615205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4616205231Skmacy mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4617205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4618205231Skmacy mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4619205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4620205231Skmacy mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4621205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4622205231Skmacy mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4623205231Skmacy NULL, MUTEX_DEFAULT, NULL); 4624206796Spjd 4625205231Skmacy list_create(&arc_mru->arcs_lists[i], 4626286570Smav sizeof (arc_buf_hdr_t), 4627286570Smav offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4628205231Skmacy list_create(&arc_mru_ghost->arcs_lists[i], 4629286570Smav sizeof (arc_buf_hdr_t), 4630286570Smav offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4631205231Skmacy list_create(&arc_mfu->arcs_lists[i], 4632286570Smav sizeof (arc_buf_hdr_t), 4633286570Smav offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4634205231Skmacy list_create(&arc_mfu_ghost->arcs_lists[i], 4635286570Smav sizeof (arc_buf_hdr_t), 4636286570Smav offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4637205231Skmacy list_create(&arc_mfu_ghost->arcs_lists[i], 4638286570Smav sizeof (arc_buf_hdr_t), 4639286570Smav offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4640205231Skmacy list_create(&arc_l2c_only->arcs_lists[i], 4641286570Smav sizeof (arc_buf_hdr_t), 4642286570Smav offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4643205231Skmacy } 4644168404Spjd 4645168404Spjd buf_init(); 4646168404Spjd 4647168404Spjd arc_thread_exit = 0; 4648168404Spjd arc_eviction_list = NULL; 4649168404Spjd mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4650168404Spjd bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4651168404Spjd 4652168404Spjd arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4653168404Spjd sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4654168404Spjd 4655168404Spjd if (arc_ksp != NULL) { 4656168404Spjd arc_ksp->ks_data = &arc_stats; 4657168404Spjd kstat_install(arc_ksp); 4658168404Spjd } 4659168404Spjd 4660168404Spjd (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4661168404Spjd TS_RUN, minclsyspri); 4662168404Spjd 4663168404Spjd#ifdef _KERNEL 4664168566Spjd arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4665168404Spjd EVENTHANDLER_PRI_FIRST); 4666168404Spjd#endif 4667168404Spjd 4668168404Spjd arc_dead = FALSE; 4669185029Spjd arc_warm = B_FALSE; 4670168566Spjd 4671258632Savg /* 4672258632Savg * Calculate maximum amount of dirty data per pool. 4673258632Savg * 4674258632Savg * If it has been set by /etc/system, take that. 4675258632Savg * Otherwise, use a percentage of physical memory defined by 4676258632Savg * zfs_dirty_data_max_percent (default 10%) with a cap at 4677258632Savg * zfs_dirty_data_max_max (default 4GB). 4678258632Savg */ 4679258632Savg if (zfs_dirty_data_max == 0) { 4680258632Savg zfs_dirty_data_max = ptob(physmem) * 4681258632Savg zfs_dirty_data_max_percent / 100; 4682258632Savg zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4683258632Savg zfs_dirty_data_max_max); 4684258632Savg } 4685185029Spjd 4686168566Spjd#ifdef _KERNEL 4687194043Skmacy if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4688193953Skmacy prefetch_tunable_set = 1; 4689206796Spjd 4690193878Skmacy#ifdef __i386__ 4691193953Skmacy if (prefetch_tunable_set == 0) { 4692196863Strasz printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4693196863Strasz "-- to enable,\n"); 4694196863Strasz printf(" add \"vfs.zfs.prefetch_disable=0\" " 4695196863Strasz "to /boot/loader.conf.\n"); 4696219089Spjd zfs_prefetch_disable = 1; 4697193878Skmacy } 4698206796Spjd#else 4699193878Skmacy if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4700193953Skmacy prefetch_tunable_set == 0) { 4701196863Strasz printf("ZFS NOTICE: Prefetch is disabled by default if less " 4702196941Strasz "than 4GB of RAM is present;\n" 4703196863Strasz " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4704196863Strasz "to /boot/loader.conf.\n"); 4705219089Spjd zfs_prefetch_disable = 1; 4706193878Skmacy } 4707206796Spjd#endif 4708175633Spjd /* Warn about ZFS memory and address space requirements. */ 4709168696Spjd if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4710168987Sbmah printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4711168987Sbmah "expect unstable behavior.\n"); 4712175633Spjd } 4713175633Spjd if (kmem_size() < 512 * (1 << 20)) { 4714173419Spjd printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4715168987Sbmah "expect unstable behavior.\n"); 4716185029Spjd printf(" Consider tuning vm.kmem_size and " 4717173419Spjd "vm.kmem_size_max\n"); 4718185029Spjd printf(" in /boot/loader.conf.\n"); 4719168566Spjd } 4720168566Spjd#endif 4721168404Spjd} 4722168404Spjd 4723168404Spjdvoid 4724168404Spjdarc_fini(void) 4725168404Spjd{ 4726205231Skmacy int i; 4727206796Spjd 4728168404Spjd mutex_enter(&arc_reclaim_thr_lock); 4729168404Spjd arc_thread_exit = 1; 4730168404Spjd cv_signal(&arc_reclaim_thr_cv); 4731168404Spjd while (arc_thread_exit != 0) 4732168404Spjd cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4733168404Spjd mutex_exit(&arc_reclaim_thr_lock); 4734168404Spjd 4735185029Spjd arc_flush(NULL); 4736168404Spjd 4737168404Spjd arc_dead = TRUE; 4738168404Spjd 4739168404Spjd if (arc_ksp != NULL) { 4740168404Spjd kstat_delete(arc_ksp); 4741168404Spjd arc_ksp = NULL; 4742168404Spjd } 4743168404Spjd 4744168404Spjd mutex_destroy(&arc_eviction_mtx); 4745168404Spjd mutex_destroy(&arc_reclaim_thr_lock); 4746168404Spjd cv_destroy(&arc_reclaim_thr_cv); 4747168404Spjd 4748205231Skmacy for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4749205231Skmacy list_destroy(&arc_mru->arcs_lists[i]); 4750205231Skmacy list_destroy(&arc_mru_ghost->arcs_lists[i]); 4751205231Skmacy list_destroy(&arc_mfu->arcs_lists[i]); 4752205231Skmacy list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4753206795Spjd list_destroy(&arc_l2c_only->arcs_lists[i]); 4754168404Spjd 4755205231Skmacy mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4756205231Skmacy mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4757205231Skmacy mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4758205231Skmacy mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4759205231Skmacy mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4760206795Spjd mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4761205231Skmacy } 4762206796Spjd 4763168404Spjd buf_fini(); 4764168404Spjd 4765286570Smav ASSERT0(arc_loaned_bytes); 4766209962Smm 4767168582Spjd mutex_destroy(&arc_lowmem_lock); 4768168404Spjd#ifdef _KERNEL 4769168566Spjd if (arc_event_lowmem != NULL) 4770168566Spjd EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4771168404Spjd#endif 4772168404Spjd} 4773185029Spjd 4774185029Spjd/* 4775185029Spjd * Level 2 ARC 4776185029Spjd * 4777185029Spjd * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4778185029Spjd * It uses dedicated storage devices to hold cached data, which are populated 4779185029Spjd * using large infrequent writes. The main role of this cache is to boost 4780185029Spjd * the performance of random read workloads. The intended L2ARC devices 4781185029Spjd * include short-stroked disks, solid state disks, and other media with 4782185029Spjd * substantially faster read latency than disk. 4783185029Spjd * 4784185029Spjd * +-----------------------+ 4785185029Spjd * | ARC | 4786185029Spjd * +-----------------------+ 4787185029Spjd * | ^ ^ 4788185029Spjd * | | | 4789185029Spjd * l2arc_feed_thread() arc_read() 4790185029Spjd * | | | 4791185029Spjd * | l2arc read | 4792185029Spjd * V | | 4793185029Spjd * +---------------+ | 4794185029Spjd * | L2ARC | | 4795185029Spjd * +---------------+ | 4796185029Spjd * | ^ | 4797185029Spjd * l2arc_write() | | 4798185029Spjd * | | | 4799185029Spjd * V | | 4800185029Spjd * +-------+ +-------+ 4801185029Spjd * | vdev | | vdev | 4802185029Spjd * | cache | | cache | 4803185029Spjd * +-------+ +-------+ 4804185029Spjd * +=========+ .-----. 4805185029Spjd * : L2ARC : |-_____-| 4806185029Spjd * : devices : | Disks | 4807185029Spjd * +=========+ `-_____-' 4808185029Spjd * 4809185029Spjd * Read requests are satisfied from the following sources, in order: 4810185029Spjd * 4811185029Spjd * 1) ARC 4812185029Spjd * 2) vdev cache of L2ARC devices 4813185029Spjd * 3) L2ARC devices 4814185029Spjd * 4) vdev cache of disks 4815185029Spjd * 5) disks 4816185029Spjd * 4817185029Spjd * Some L2ARC device types exhibit extremely slow write performance. 4818185029Spjd * To accommodate for this there are some significant differences between 4819185029Spjd * the L2ARC and traditional cache design: 4820185029Spjd * 4821185029Spjd * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4822185029Spjd * the ARC behave as usual, freeing buffers and placing headers on ghost 4823185029Spjd * lists. The ARC does not send buffers to the L2ARC during eviction as 4824185029Spjd * this would add inflated write latencies for all ARC memory pressure. 4825185029Spjd * 4826185029Spjd * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4827185029Spjd * It does this by periodically scanning buffers from the eviction-end of 4828185029Spjd * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4829251478Sdelphij * not already there. It scans until a headroom of buffers is satisfied, 4830251478Sdelphij * which itself is a buffer for ARC eviction. If a compressible buffer is 4831251478Sdelphij * found during scanning and selected for writing to an L2ARC device, we 4832251478Sdelphij * temporarily boost scanning headroom during the next scan cycle to make 4833251478Sdelphij * sure we adapt to compression effects (which might significantly reduce 4834251478Sdelphij * the data volume we write to L2ARC). The thread that does this is 4835185029Spjd * l2arc_feed_thread(), illustrated below; example sizes are included to 4836185029Spjd * provide a better sense of ratio than this diagram: 4837185029Spjd * 4838185029Spjd * head --> tail 4839185029Spjd * +---------------------+----------+ 4840185029Spjd * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4841185029Spjd * +---------------------+----------+ | o L2ARC eligible 4842185029Spjd * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4843185029Spjd * +---------------------+----------+ | 4844185029Spjd * 15.9 Gbytes ^ 32 Mbytes | 4845185029Spjd * headroom | 4846185029Spjd * l2arc_feed_thread() 4847185029Spjd * | 4848185029Spjd * l2arc write hand <--[oooo]--' 4849185029Spjd * | 8 Mbyte 4850185029Spjd * | write max 4851185029Spjd * V 4852185029Spjd * +==============================+ 4853185029Spjd * L2ARC dev |####|#|###|###| |####| ... | 4854185029Spjd * +==============================+ 4855185029Spjd * 32 Gbytes 4856185029Spjd * 4857185029Spjd * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4858185029Spjd * evicted, then the L2ARC has cached a buffer much sooner than it probably 4859185029Spjd * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4860185029Spjd * safe to say that this is an uncommon case, since buffers at the end of 4861185029Spjd * the ARC lists have moved there due to inactivity. 4862185029Spjd * 4863185029Spjd * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4864185029Spjd * then the L2ARC simply misses copying some buffers. This serves as a 4865185029Spjd * pressure valve to prevent heavy read workloads from both stalling the ARC 4866185029Spjd * with waits and clogging the L2ARC with writes. This also helps prevent 4867185029Spjd * the potential for the L2ARC to churn if it attempts to cache content too 4868185029Spjd * quickly, such as during backups of the entire pool. 4869185029Spjd * 4870185029Spjd * 5. After system boot and before the ARC has filled main memory, there are 4871185029Spjd * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4872185029Spjd * lists can remain mostly static. Instead of searching from tail of these 4873185029Spjd * lists as pictured, the l2arc_feed_thread() will search from the list heads 4874185029Spjd * for eligible buffers, greatly increasing its chance of finding them. 4875185029Spjd * 4876185029Spjd * The L2ARC device write speed is also boosted during this time so that 4877185029Spjd * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4878185029Spjd * there are no L2ARC reads, and no fear of degrading read performance 4879185029Spjd * through increased writes. 4880185029Spjd * 4881185029Spjd * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4882185029Spjd * the vdev queue can aggregate them into larger and fewer writes. Each 4883185029Spjd * device is written to in a rotor fashion, sweeping writes through 4884185029Spjd * available space then repeating. 4885185029Spjd * 4886185029Spjd * 7. The L2ARC does not store dirty content. It never needs to flush 4887185029Spjd * write buffers back to disk based storage. 4888185029Spjd * 4889185029Spjd * 8. If an ARC buffer is written (and dirtied) which also exists in the 4890185029Spjd * L2ARC, the now stale L2ARC buffer is immediately dropped. 4891185029Spjd * 4892185029Spjd * The performance of the L2ARC can be tweaked by a number of tunables, which 4893185029Spjd * may be necessary for different workloads: 4894185029Spjd * 4895185029Spjd * l2arc_write_max max write bytes per interval 4896185029Spjd * l2arc_write_boost extra write bytes during device warmup 4897185029Spjd * l2arc_noprefetch skip caching prefetched buffers 4898185029Spjd * l2arc_headroom number of max device writes to precache 4899251478Sdelphij * l2arc_headroom_boost when we find compressed buffers during ARC 4900251478Sdelphij * scanning, we multiply headroom by this 4901251478Sdelphij * percentage factor for the next scan cycle, 4902251478Sdelphij * since more compressed buffers are likely to 4903251478Sdelphij * be present 4904185029Spjd * l2arc_feed_secs seconds between L2ARC writing 4905185029Spjd * 4906185029Spjd * Tunables may be removed or added as future performance improvements are 4907185029Spjd * integrated, and also may become zpool properties. 4908208373Smm * 4909208373Smm * There are three key functions that control how the L2ARC warms up: 4910208373Smm * 4911208373Smm * l2arc_write_eligible() check if a buffer is eligible to cache 4912208373Smm * l2arc_write_size() calculate how much to write 4913208373Smm * l2arc_write_interval() calculate sleep delay between writes 4914208373Smm * 4915208373Smm * These three functions determine what to write, how much, and how quickly 4916208373Smm * to send writes. 4917185029Spjd */ 4918185029Spjd 4919208373Smmstatic boolean_t 4920275811Sdelphijl2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 4921208373Smm{ 4922208373Smm /* 4923208373Smm * A buffer is *not* eligible for the L2ARC if it: 4924208373Smm * 1. belongs to a different spa. 4925208373Smm * 2. is already cached on the L2ARC. 4926208373Smm * 3. has an I/O in progress (it may be an incomplete read). 4927208373Smm * 4. is flagged not eligible (zfs property). 4928208373Smm */ 4929275811Sdelphij if (hdr->b_spa != spa_guid) { 4930208373Smm ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4931208373Smm return (B_FALSE); 4932208373Smm } 4933286570Smav if (HDR_HAS_L2HDR(hdr)) { 4934208373Smm ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4935208373Smm return (B_FALSE); 4936208373Smm } 4937275811Sdelphij if (HDR_IO_IN_PROGRESS(hdr)) { 4938208373Smm ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4939208373Smm return (B_FALSE); 4940208373Smm } 4941275811Sdelphij if (!HDR_L2CACHE(hdr)) { 4942208373Smm ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4943208373Smm return (B_FALSE); 4944208373Smm } 4945208373Smm 4946208373Smm return (B_TRUE); 4947208373Smm} 4948208373Smm 4949208373Smmstatic uint64_t 4950251478Sdelphijl2arc_write_size(void) 4951208373Smm{ 4952208373Smm uint64_t size; 4953208373Smm 4954251478Sdelphij /* 4955251478Sdelphij * Make sure our globals have meaningful values in case the user 4956251478Sdelphij * altered them. 4957251478Sdelphij */ 4958251478Sdelphij size = l2arc_write_max; 4959251478Sdelphij if (size == 0) { 4960251478Sdelphij cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4961251478Sdelphij "be greater than zero, resetting it to the default (%d)", 4962251478Sdelphij L2ARC_WRITE_SIZE); 4963251478Sdelphij size = l2arc_write_max = L2ARC_WRITE_SIZE; 4964251478Sdelphij } 4965208373Smm 4966208373Smm if (arc_warm == B_FALSE) 4967251478Sdelphij size += l2arc_write_boost; 4968208373Smm 4969208373Smm return (size); 4970208373Smm 4971208373Smm} 4972208373Smm 4973208373Smmstatic clock_t 4974208373Smml2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4975208373Smm{ 4976219089Spjd clock_t interval, next, now; 4977208373Smm 4978208373Smm /* 4979208373Smm * If the ARC lists are busy, increase our write rate; if the 4980208373Smm * lists are stale, idle back. This is achieved by checking 4981208373Smm * how much we previously wrote - if it was more than half of 4982208373Smm * what we wanted, schedule the next write much sooner. 4983208373Smm */ 4984208373Smm if (l2arc_feed_again && wrote > (wanted / 2)) 4985208373Smm interval = (hz * l2arc_feed_min_ms) / 1000; 4986208373Smm else 4987208373Smm interval = hz * l2arc_feed_secs; 4988208373Smm 4989219089Spjd now = ddi_get_lbolt(); 4990219089Spjd next = MAX(now, MIN(now + interval, began + interval)); 4991208373Smm 4992208373Smm return (next); 4993208373Smm} 4994208373Smm 4995185029Spjd/* 4996185029Spjd * Cycle through L2ARC devices. This is how L2ARC load balances. 4997185029Spjd * If a device is returned, this also returns holding the spa config lock. 4998185029Spjd */ 4999185029Spjdstatic l2arc_dev_t * 5000185029Spjdl2arc_dev_get_next(void) 5001185029Spjd{ 5002185029Spjd l2arc_dev_t *first, *next = NULL; 5003185029Spjd 5004185029Spjd /* 5005185029Spjd * Lock out the removal of spas (spa_namespace_lock), then removal 5006185029Spjd * of cache devices (l2arc_dev_mtx). Once a device has been selected, 5007185029Spjd * both locks will be dropped and a spa config lock held instead. 5008185029Spjd */ 5009185029Spjd mutex_enter(&spa_namespace_lock); 5010185029Spjd mutex_enter(&l2arc_dev_mtx); 5011185029Spjd 5012185029Spjd /* if there are no vdevs, there is nothing to do */ 5013185029Spjd if (l2arc_ndev == 0) 5014185029Spjd goto out; 5015185029Spjd 5016185029Spjd first = NULL; 5017185029Spjd next = l2arc_dev_last; 5018185029Spjd do { 5019185029Spjd /* loop around the list looking for a non-faulted vdev */ 5020185029Spjd if (next == NULL) { 5021185029Spjd next = list_head(l2arc_dev_list); 5022185029Spjd } else { 5023185029Spjd next = list_next(l2arc_dev_list, next); 5024185029Spjd if (next == NULL) 5025185029Spjd next = list_head(l2arc_dev_list); 5026185029Spjd } 5027185029Spjd 5028185029Spjd /* if we have come back to the start, bail out */ 5029185029Spjd if (first == NULL) 5030185029Spjd first = next; 5031185029Spjd else if (next == first) 5032185029Spjd break; 5033185029Spjd 5034185029Spjd } while (vdev_is_dead(next->l2ad_vdev)); 5035185029Spjd 5036185029Spjd /* if we were unable to find any usable vdevs, return NULL */ 5037185029Spjd if (vdev_is_dead(next->l2ad_vdev)) 5038185029Spjd next = NULL; 5039185029Spjd 5040185029Spjd l2arc_dev_last = next; 5041185029Spjd 5042185029Spjdout: 5043185029Spjd mutex_exit(&l2arc_dev_mtx); 5044185029Spjd 5045185029Spjd /* 5046185029Spjd * Grab the config lock to prevent the 'next' device from being 5047185029Spjd * removed while we are writing to it. 5048185029Spjd */ 5049185029Spjd if (next != NULL) 5050185029Spjd spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 5051185029Spjd mutex_exit(&spa_namespace_lock); 5052185029Spjd 5053185029Spjd return (next); 5054185029Spjd} 5055185029Spjd 5056185029Spjd/* 5057185029Spjd * Free buffers that were tagged for destruction. 5058185029Spjd */ 5059185029Spjdstatic void 5060185029Spjdl2arc_do_free_on_write() 5061185029Spjd{ 5062185029Spjd list_t *buflist; 5063185029Spjd l2arc_data_free_t *df, *df_prev; 5064185029Spjd 5065185029Spjd mutex_enter(&l2arc_free_on_write_mtx); 5066185029Spjd buflist = l2arc_free_on_write; 5067185029Spjd 5068185029Spjd for (df = list_tail(buflist); df; df = df_prev) { 5069185029Spjd df_prev = list_prev(buflist, df); 5070185029Spjd ASSERT(df->l2df_data != NULL); 5071185029Spjd ASSERT(df->l2df_func != NULL); 5072185029Spjd df->l2df_func(df->l2df_data, df->l2df_size); 5073185029Spjd list_remove(buflist, df); 5074185029Spjd kmem_free(df, sizeof (l2arc_data_free_t)); 5075185029Spjd } 5076185029Spjd 5077185029Spjd mutex_exit(&l2arc_free_on_write_mtx); 5078185029Spjd} 5079185029Spjd 5080185029Spjd/* 5081185029Spjd * A write to a cache device has completed. Update all headers to allow 5082185029Spjd * reads from these buffers to begin. 5083185029Spjd */ 5084185029Spjdstatic void 5085185029Spjdl2arc_write_done(zio_t *zio) 5086185029Spjd{ 5087185029Spjd l2arc_write_callback_t *cb; 5088185029Spjd l2arc_dev_t *dev; 5089185029Spjd list_t *buflist; 5090275811Sdelphij arc_buf_hdr_t *head, *hdr, *hdr_prev; 5091185029Spjd kmutex_t *hash_lock; 5092268085Sdelphij int64_t bytes_dropped = 0; 5093185029Spjd 5094185029Spjd cb = zio->io_private; 5095185029Spjd ASSERT(cb != NULL); 5096185029Spjd dev = cb->l2wcb_dev; 5097185029Spjd ASSERT(dev != NULL); 5098185029Spjd head = cb->l2wcb_head; 5099185029Spjd ASSERT(head != NULL); 5100286570Smav buflist = &dev->l2ad_buflist; 5101185029Spjd ASSERT(buflist != NULL); 5102185029Spjd DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 5103185029Spjd l2arc_write_callback_t *, cb); 5104185029Spjd 5105185029Spjd if (zio->io_error != 0) 5106185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_error); 5107185029Spjd 5108286570Smav mutex_enter(&dev->l2ad_mtx); 5109185029Spjd 5110185029Spjd /* 5111185029Spjd * All writes completed, or an error was hit. 5112185029Spjd */ 5113275811Sdelphij for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 5114275811Sdelphij hdr_prev = list_prev(buflist, hdr); 5115185029Spjd 5116275811Sdelphij hash_lock = HDR_LOCK(hdr); 5117185029Spjd if (!mutex_tryenter(hash_lock)) { 5118185029Spjd /* 5119185029Spjd * This buffer misses out. It may be in a stage 5120286570Smav * of eviction. Its ARC_FLAG_L2_WRITING flag will be 5121185029Spjd * left set, denying reads to this buffer. 5122185029Spjd */ 5123185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 5124185029Spjd continue; 5125185029Spjd } 5126185029Spjd 5127286570Smav /* 5128286570Smav * It's possible that this buffer got evicted from the L1 cache 5129286570Smav * before we grabbed the vdev + hash locks, in which case 5130286570Smav * arc_hdr_realloc freed b_tmp_cdata for us if it was allocated. 5131286570Smav * Only free the buffer if we still have an L1 hdr. 5132286570Smav */ 5133286570Smav if (HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_tmp_cdata != NULL && 5134286570Smav HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) 5135286570Smav l2arc_release_cdata_buf(hdr); 5136286570Smav 5137185029Spjd if (zio->io_error != 0) { 5138185029Spjd /* 5139185029Spjd * Error - drop L2ARC entry. 5140185029Spjd */ 5141286570Smav trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev, 5142286570Smav hdr->b_l2hdr.b_daddr, hdr->b_l2hdr.b_asize, 0); 5143286570Smav hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 5144286570Smav 5145286570Smav ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize); 5146275811Sdelphij ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 5147185029Spjd } 5148185029Spjd 5149185029Spjd /* 5150185029Spjd * Allow ARC to begin reads to this L2ARC entry. 5151185029Spjd */ 5152275811Sdelphij hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 5153185029Spjd 5154185029Spjd mutex_exit(hash_lock); 5155185029Spjd } 5156185029Spjd 5157185029Spjd atomic_inc_64(&l2arc_writes_done); 5158185029Spjd list_remove(buflist, head); 5159286570Smav ASSERT(!HDR_HAS_L1HDR(head)); 5160286570Smav kmem_cache_free(hdr_l2only_cache, head); 5161286570Smav mutex_exit(&dev->l2ad_mtx); 5162185029Spjd 5163268085Sdelphij vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 5164268085Sdelphij 5165185029Spjd l2arc_do_free_on_write(); 5166185029Spjd 5167185029Spjd kmem_free(cb, sizeof (l2arc_write_callback_t)); 5168185029Spjd} 5169185029Spjd 5170185029Spjd/* 5171185029Spjd * A read to a cache device completed. Validate buffer contents before 5172185029Spjd * handing over to the regular ARC routines. 5173185029Spjd */ 5174185029Spjdstatic void 5175185029Spjdl2arc_read_done(zio_t *zio) 5176185029Spjd{ 5177185029Spjd l2arc_read_callback_t *cb; 5178185029Spjd arc_buf_hdr_t *hdr; 5179185029Spjd arc_buf_t *buf; 5180185029Spjd kmutex_t *hash_lock; 5181185029Spjd int equal; 5182185029Spjd 5183185029Spjd ASSERT(zio->io_vd != NULL); 5184185029Spjd ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 5185185029Spjd 5186185029Spjd spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 5187185029Spjd 5188185029Spjd cb = zio->io_private; 5189185029Spjd ASSERT(cb != NULL); 5190185029Spjd buf = cb->l2rcb_buf; 5191185029Spjd ASSERT(buf != NULL); 5192185029Spjd 5193219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 5194185029Spjd mutex_enter(hash_lock); 5195219089Spjd hdr = buf->b_hdr; 5196219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 5197185029Spjd 5198185029Spjd /* 5199251478Sdelphij * If the buffer was compressed, decompress it first. 5200251478Sdelphij */ 5201251478Sdelphij if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 5202251478Sdelphij l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 5203251478Sdelphij ASSERT(zio->io_data != NULL); 5204251478Sdelphij 5205251478Sdelphij /* 5206185029Spjd * Check this survived the L2ARC journey. 5207185029Spjd */ 5208185029Spjd equal = arc_cksum_equal(buf); 5209185029Spjd if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 5210185029Spjd mutex_exit(hash_lock); 5211185029Spjd zio->io_private = buf; 5212185029Spjd zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 5213185029Spjd zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 5214185029Spjd arc_read_done(zio); 5215185029Spjd } else { 5216185029Spjd mutex_exit(hash_lock); 5217185029Spjd /* 5218185029Spjd * Buffer didn't survive caching. Increment stats and 5219185029Spjd * reissue to the original storage device. 5220185029Spjd */ 5221185029Spjd if (zio->io_error != 0) { 5222185029Spjd ARCSTAT_BUMP(arcstat_l2_io_error); 5223185029Spjd } else { 5224249195Smm zio->io_error = SET_ERROR(EIO); 5225185029Spjd } 5226185029Spjd if (!equal) 5227185029Spjd ARCSTAT_BUMP(arcstat_l2_cksum_bad); 5228185029Spjd 5229185029Spjd /* 5230185029Spjd * If there's no waiter, issue an async i/o to the primary 5231185029Spjd * storage now. If there *is* a waiter, the caller must 5232185029Spjd * issue the i/o in a context where it's OK to block. 5233185029Spjd */ 5234209962Smm if (zio->io_waiter == NULL) { 5235209962Smm zio_t *pio = zio_unique_parent(zio); 5236209962Smm 5237209962Smm ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 5238209962Smm 5239209962Smm zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 5240185029Spjd buf->b_data, zio->io_size, arc_read_done, buf, 5241185029Spjd zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 5242209962Smm } 5243185029Spjd } 5244185029Spjd 5245185029Spjd kmem_free(cb, sizeof (l2arc_read_callback_t)); 5246185029Spjd} 5247185029Spjd 5248185029Spjd/* 5249185029Spjd * This is the list priority from which the L2ARC will search for pages to 5250185029Spjd * cache. This is used within loops (0..3) to cycle through lists in the 5251185029Spjd * desired order. This order can have a significant effect on cache 5252185029Spjd * performance. 5253185029Spjd * 5254185029Spjd * Currently the metadata lists are hit first, MFU then MRU, followed by 5255185029Spjd * the data lists. This function returns a locked list, and also returns 5256185029Spjd * the lock pointer. 5257185029Spjd */ 5258185029Spjdstatic list_t * 5259185029Spjdl2arc_list_locked(int list_num, kmutex_t **lock) 5260185029Spjd{ 5261247187Smm list_t *list = NULL; 5262205231Skmacy int idx; 5263185029Spjd 5264206796Spjd ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 5265206796Spjd 5266205231Skmacy if (list_num < ARC_BUFC_NUMMETADATALISTS) { 5267205231Skmacy idx = list_num; 5268205231Skmacy list = &arc_mfu->arcs_lists[idx]; 5269205231Skmacy *lock = ARCS_LOCK(arc_mfu, idx); 5270206796Spjd } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 5271205231Skmacy idx = list_num - ARC_BUFC_NUMMETADATALISTS; 5272205231Skmacy list = &arc_mru->arcs_lists[idx]; 5273205231Skmacy *lock = ARCS_LOCK(arc_mru, idx); 5274206796Spjd } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 5275205231Skmacy ARC_BUFC_NUMDATALISTS)) { 5276205231Skmacy idx = list_num - ARC_BUFC_NUMMETADATALISTS; 5277205231Skmacy list = &arc_mfu->arcs_lists[idx]; 5278205231Skmacy *lock = ARCS_LOCK(arc_mfu, idx); 5279205231Skmacy } else { 5280205231Skmacy idx = list_num - ARC_BUFC_NUMLISTS; 5281205231Skmacy list = &arc_mru->arcs_lists[idx]; 5282205231Skmacy *lock = ARCS_LOCK(arc_mru, idx); 5283185029Spjd } 5284185029Spjd 5285185029Spjd ASSERT(!(MUTEX_HELD(*lock))); 5286185029Spjd mutex_enter(*lock); 5287185029Spjd return (list); 5288185029Spjd} 5289185029Spjd 5290185029Spjd/* 5291185029Spjd * Evict buffers from the device write hand to the distance specified in 5292185029Spjd * bytes. This distance may span populated buffers, it may span nothing. 5293185029Spjd * This is clearing a region on the L2ARC device ready for writing. 5294185029Spjd * If the 'all' boolean is set, every buffer is evicted. 5295185029Spjd */ 5296185029Spjdstatic void 5297185029Spjdl2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 5298185029Spjd{ 5299185029Spjd list_t *buflist; 5300275811Sdelphij arc_buf_hdr_t *hdr, *hdr_prev; 5301185029Spjd kmutex_t *hash_lock; 5302185029Spjd uint64_t taddr; 5303268085Sdelphij int64_t bytes_evicted = 0; 5304185029Spjd 5305286570Smav buflist = &dev->l2ad_buflist; 5306185029Spjd 5307185029Spjd if (!all && dev->l2ad_first) { 5308185029Spjd /* 5309185029Spjd * This is the first sweep through the device. There is 5310185029Spjd * nothing to evict. 5311185029Spjd */ 5312185029Spjd return; 5313185029Spjd } 5314185029Spjd 5315185029Spjd if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 5316185029Spjd /* 5317185029Spjd * When nearing the end of the device, evict to the end 5318185029Spjd * before the device write hand jumps to the start. 5319185029Spjd */ 5320185029Spjd taddr = dev->l2ad_end; 5321185029Spjd } else { 5322185029Spjd taddr = dev->l2ad_hand + distance; 5323185029Spjd } 5324185029Spjd DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 5325185029Spjd uint64_t, taddr, boolean_t, all); 5326185029Spjd 5327185029Spjdtop: 5328286570Smav mutex_enter(&dev->l2ad_mtx); 5329275811Sdelphij for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 5330275811Sdelphij hdr_prev = list_prev(buflist, hdr); 5331185029Spjd 5332275811Sdelphij hash_lock = HDR_LOCK(hdr); 5333185029Spjd if (!mutex_tryenter(hash_lock)) { 5334185029Spjd /* 5335185029Spjd * Missed the hash lock. Retry. 5336185029Spjd */ 5337185029Spjd ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 5338286570Smav mutex_exit(&dev->l2ad_mtx); 5339185029Spjd mutex_enter(hash_lock); 5340185029Spjd mutex_exit(hash_lock); 5341185029Spjd goto top; 5342185029Spjd } 5343185029Spjd 5344275811Sdelphij if (HDR_L2_WRITE_HEAD(hdr)) { 5345185029Spjd /* 5346185029Spjd * We hit a write head node. Leave it for 5347185029Spjd * l2arc_write_done(). 5348185029Spjd */ 5349275811Sdelphij list_remove(buflist, hdr); 5350185029Spjd mutex_exit(hash_lock); 5351185029Spjd continue; 5352185029Spjd } 5353185029Spjd 5354286570Smav if (!all && HDR_HAS_L2HDR(hdr) && 5355286570Smav (hdr->b_l2hdr.b_daddr > taddr || 5356286570Smav hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 5357185029Spjd /* 5358185029Spjd * We've evicted to the target address, 5359185029Spjd * or the end of the device. 5360185029Spjd */ 5361185029Spjd mutex_exit(hash_lock); 5362185029Spjd break; 5363185029Spjd } 5364185029Spjd 5365286570Smav ASSERT(HDR_HAS_L2HDR(hdr)); 5366286570Smav if (!HDR_HAS_L1HDR(hdr)) { 5367275811Sdelphij ASSERT(!HDR_L2_READING(hdr)); 5368185029Spjd /* 5369185029Spjd * This doesn't exist in the ARC. Destroy. 5370185029Spjd * arc_hdr_destroy() will call list_remove() 5371185029Spjd * and decrement arcstat_l2_size. 5372185029Spjd */ 5373275811Sdelphij arc_change_state(arc_anon, hdr, hash_lock); 5374275811Sdelphij arc_hdr_destroy(hdr); 5375185029Spjd } else { 5376286570Smav ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 5377286570Smav ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 5378185029Spjd /* 5379185029Spjd * Invalidate issued or about to be issued 5380185029Spjd * reads, since we may be about to write 5381185029Spjd * over this location. 5382185029Spjd */ 5383275811Sdelphij if (HDR_L2_READING(hdr)) { 5384185029Spjd ARCSTAT_BUMP(arcstat_l2_evict_reading); 5385275811Sdelphij hdr->b_flags |= ARC_FLAG_L2_EVICTED; 5386185029Spjd } 5387185029Spjd 5388286570Smav /* Tell ARC this no longer exists in L2ARC. */ 5389286570Smav ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize); 5390286570Smav ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 5391286570Smav hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 5392275811Sdelphij list_remove(buflist, hdr); 5393185029Spjd 5394286570Smav /* This may have been leftover after a failed write. */ 5395275811Sdelphij hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 5396185029Spjd } 5397185029Spjd mutex_exit(hash_lock); 5398185029Spjd } 5399286570Smav mutex_exit(&dev->l2ad_mtx); 5400185029Spjd 5401268085Sdelphij vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); 5402185029Spjd dev->l2ad_evict = taddr; 5403185029Spjd} 5404185029Spjd 5405185029Spjd/* 5406185029Spjd * Find and write ARC buffers to the L2ARC device. 5407185029Spjd * 5408275811Sdelphij * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 5409185029Spjd * for reading until they have completed writing. 5410251478Sdelphij * The headroom_boost is an in-out parameter used to maintain headroom boost 5411251478Sdelphij * state between calls to this function. 5412251478Sdelphij * 5413251478Sdelphij * Returns the number of bytes actually written (which may be smaller than 5414251478Sdelphij * the delta by which the device hand has changed due to alignment). 5415185029Spjd */ 5416208373Smmstatic uint64_t 5417251478Sdelphijl2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 5418251478Sdelphij boolean_t *headroom_boost) 5419185029Spjd{ 5420275811Sdelphij arc_buf_hdr_t *hdr, *hdr_prev, *head; 5421185029Spjd list_t *list; 5422251478Sdelphij uint64_t write_asize, write_psize, write_sz, headroom, 5423251478Sdelphij buf_compress_minsz; 5424185029Spjd void *buf_data; 5425251478Sdelphij kmutex_t *list_lock; 5426251478Sdelphij boolean_t full; 5427185029Spjd l2arc_write_callback_t *cb; 5428185029Spjd zio_t *pio, *wzio; 5429228103Smm uint64_t guid = spa_load_guid(spa); 5430251478Sdelphij const boolean_t do_headroom_boost = *headroom_boost; 5431185029Spjd int try; 5432185029Spjd 5433185029Spjd ASSERT(dev->l2ad_vdev != NULL); 5434185029Spjd 5435251478Sdelphij /* Lower the flag now, we might want to raise it again later. */ 5436251478Sdelphij *headroom_boost = B_FALSE; 5437251478Sdelphij 5438185029Spjd pio = NULL; 5439251478Sdelphij write_sz = write_asize = write_psize = 0; 5440185029Spjd full = B_FALSE; 5441286570Smav head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 5442275811Sdelphij head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; 5443286570Smav head->b_flags |= ARC_FLAG_HAS_L2HDR; 5444185029Spjd 5445205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 5446185029Spjd /* 5447251478Sdelphij * We will want to try to compress buffers that are at least 2x the 5448251478Sdelphij * device sector size. 5449251478Sdelphij */ 5450251478Sdelphij buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 5451251478Sdelphij 5452251478Sdelphij /* 5453185029Spjd * Copy buffers for L2ARC writing. 5454185029Spjd */ 5455286570Smav mutex_enter(&dev->l2ad_mtx); 5456206796Spjd for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 5457251478Sdelphij uint64_t passed_sz = 0; 5458251478Sdelphij 5459185029Spjd list = l2arc_list_locked(try, &list_lock); 5460205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 5461185029Spjd 5462185029Spjd /* 5463185029Spjd * L2ARC fast warmup. 5464185029Spjd * 5465185029Spjd * Until the ARC is warm and starts to evict, read from the 5466185029Spjd * head of the ARC lists rather than the tail. 5467185029Spjd */ 5468185029Spjd if (arc_warm == B_FALSE) 5469275811Sdelphij hdr = list_head(list); 5470185029Spjd else 5471275811Sdelphij hdr = list_tail(list); 5472275811Sdelphij if (hdr == NULL) 5473205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 5474185029Spjd 5475272708Savg headroom = target_sz * l2arc_headroom * 2 / ARC_BUFC_NUMLISTS; 5476251478Sdelphij if (do_headroom_boost) 5477251478Sdelphij headroom = (headroom * l2arc_headroom_boost) / 100; 5478251478Sdelphij 5479275811Sdelphij for (; hdr; hdr = hdr_prev) { 5480251478Sdelphij kmutex_t *hash_lock; 5481251478Sdelphij uint64_t buf_sz; 5482251478Sdelphij 5483185029Spjd if (arc_warm == B_FALSE) 5484275811Sdelphij hdr_prev = list_next(list, hdr); 5485185029Spjd else 5486275811Sdelphij hdr_prev = list_prev(list, hdr); 5487275811Sdelphij ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size); 5488206796Spjd 5489275811Sdelphij hash_lock = HDR_LOCK(hdr); 5490251478Sdelphij if (!mutex_tryenter(hash_lock)) { 5491205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5492185029Spjd /* 5493185029Spjd * Skip this buffer rather than waiting. 5494185029Spjd */ 5495185029Spjd continue; 5496185029Spjd } 5497185029Spjd 5498275811Sdelphij passed_sz += hdr->b_size; 5499185029Spjd if (passed_sz > headroom) { 5500185029Spjd /* 5501185029Spjd * Searched too far. 5502185029Spjd */ 5503185029Spjd mutex_exit(hash_lock); 5504205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5505185029Spjd break; 5506185029Spjd } 5507185029Spjd 5508275811Sdelphij if (!l2arc_write_eligible(guid, hdr)) { 5509185029Spjd mutex_exit(hash_lock); 5510185029Spjd continue; 5511185029Spjd } 5512185029Spjd 5513275811Sdelphij if ((write_sz + hdr->b_size) > target_sz) { 5514185029Spjd full = B_TRUE; 5515185029Spjd mutex_exit(hash_lock); 5516205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_full); 5517185029Spjd break; 5518185029Spjd } 5519185029Spjd 5520185029Spjd if (pio == NULL) { 5521185029Spjd /* 5522185029Spjd * Insert a dummy header on the buflist so 5523185029Spjd * l2arc_write_done() can find where the 5524185029Spjd * write buffers begin without searching. 5525185029Spjd */ 5526286570Smav list_insert_head(&dev->l2ad_buflist, head); 5527185029Spjd 5528185029Spjd cb = kmem_alloc( 5529185029Spjd sizeof (l2arc_write_callback_t), KM_SLEEP); 5530185029Spjd cb->l2wcb_dev = dev; 5531185029Spjd cb->l2wcb_head = head; 5532185029Spjd pio = zio_root(spa, l2arc_write_done, cb, 5533185029Spjd ZIO_FLAG_CANFAIL); 5534205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_pios); 5535185029Spjd } 5536185029Spjd 5537185029Spjd /* 5538185029Spjd * Create and add a new L2ARC header. 5539185029Spjd */ 5540286570Smav hdr->b_l2hdr.b_dev = dev; 5541275811Sdelphij hdr->b_flags |= ARC_FLAG_L2_WRITING; 5542251478Sdelphij /* 5543251478Sdelphij * Temporarily stash the data buffer in b_tmp_cdata. 5544251478Sdelphij * The subsequent write step will pick it up from 5545286570Smav * there. This is because can't access b_l1hdr.b_buf 5546251478Sdelphij * without holding the hash_lock, which we in turn 5547251478Sdelphij * can't access without holding the ARC list locks 5548251478Sdelphij * (which we want to avoid during compression/writing). 5549251478Sdelphij */ 5550286570Smav HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); 5551286570Smav hdr->b_l2hdr.b_asize = hdr->b_size; 5552286570Smav hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data; 5553251478Sdelphij 5554275811Sdelphij buf_sz = hdr->b_size; 5555286570Smav hdr->b_flags |= ARC_FLAG_HAS_L2HDR; 5556185029Spjd 5557286570Smav list_insert_head(&dev->l2ad_buflist, hdr); 5558251478Sdelphij 5559185029Spjd /* 5560185029Spjd * Compute and store the buffer cksum before 5561185029Spjd * writing. On debug the cksum is verified first. 5562185029Spjd */ 5563286570Smav arc_cksum_verify(hdr->b_l1hdr.b_buf); 5564286570Smav arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE); 5565185029Spjd 5566185029Spjd mutex_exit(hash_lock); 5567185029Spjd 5568251478Sdelphij write_sz += buf_sz; 5569251478Sdelphij } 5570251478Sdelphij 5571251478Sdelphij mutex_exit(list_lock); 5572251478Sdelphij 5573251478Sdelphij if (full == B_TRUE) 5574251478Sdelphij break; 5575251478Sdelphij } 5576251478Sdelphij 5577251478Sdelphij /* No buffers selected for writing? */ 5578251478Sdelphij if (pio == NULL) { 5579251478Sdelphij ASSERT0(write_sz); 5580286570Smav mutex_exit(&dev->l2ad_mtx); 5581286570Smav ASSERT(!HDR_HAS_L1HDR(head)); 5582286570Smav kmem_cache_free(hdr_l2only_cache, head); 5583251478Sdelphij return (0); 5584251478Sdelphij } 5585251478Sdelphij 5586251478Sdelphij /* 5587251478Sdelphij * Now start writing the buffers. We're starting at the write head 5588251478Sdelphij * and work backwards, retracing the course of the buffer selector 5589251478Sdelphij * loop above. 5590251478Sdelphij */ 5591286570Smav for (hdr = list_prev(&dev->l2ad_buflist, head); hdr; 5592286570Smav hdr = list_prev(&dev->l2ad_buflist, hdr)) { 5593251478Sdelphij uint64_t buf_sz; 5594251478Sdelphij 5595251478Sdelphij /* 5596251478Sdelphij * We shouldn't need to lock the buffer here, since we flagged 5597275811Sdelphij * it as ARC_FLAG_L2_WRITING in the previous step, but we must 5598275811Sdelphij * take care to only access its L2 cache parameters. In 5599286570Smav * particular, hdr->l1hdr.b_buf may be invalid by now due to 5600275811Sdelphij * ARC eviction. 5601251478Sdelphij */ 5602286570Smav hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 5603251478Sdelphij 5604286570Smav if ((HDR_L2COMPRESS(hdr)) && 5605286570Smav hdr->b_l2hdr.b_asize >= buf_compress_minsz) { 5606286570Smav if (l2arc_compress_buf(hdr)) { 5607251478Sdelphij /* 5608251478Sdelphij * If compression succeeded, enable headroom 5609251478Sdelphij * boost on the next scan cycle. 5610251478Sdelphij */ 5611251478Sdelphij *headroom_boost = B_TRUE; 5612251478Sdelphij } 5613251478Sdelphij } 5614251478Sdelphij 5615251478Sdelphij /* 5616251478Sdelphij * Pick up the buffer data we had previously stashed away 5617251478Sdelphij * (and now potentially also compressed). 5618251478Sdelphij */ 5619286570Smav buf_data = hdr->b_l1hdr.b_tmp_cdata; 5620286570Smav buf_sz = hdr->b_l2hdr.b_asize; 5621251478Sdelphij 5622274172Savg /* 5623274172Savg * If the data has not been compressed, then clear b_tmp_cdata 5624274172Savg * to make sure that it points only to a temporary compression 5625274172Savg * buffer. 5626274172Savg */ 5627286570Smav if (!L2ARC_IS_VALID_COMPRESS(HDR_GET_COMPRESS(hdr))) 5628286570Smav hdr->b_l1hdr.b_tmp_cdata = NULL; 5629274172Savg 5630251478Sdelphij /* Compression may have squashed the buffer to zero length. */ 5631251478Sdelphij if (buf_sz != 0) { 5632251478Sdelphij uint64_t buf_p_sz; 5633251478Sdelphij 5634185029Spjd wzio = zio_write_phys(pio, dev->l2ad_vdev, 5635185029Spjd dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5636185029Spjd NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5637185029Spjd ZIO_FLAG_CANFAIL, B_FALSE); 5638185029Spjd 5639185029Spjd DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5640185029Spjd zio_t *, wzio); 5641185029Spjd (void) zio_nowait(wzio); 5642185029Spjd 5643251478Sdelphij write_asize += buf_sz; 5644185029Spjd /* 5645185029Spjd * Keep the clock hand suitably device-aligned. 5646185029Spjd */ 5647251478Sdelphij buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5648251478Sdelphij write_psize += buf_p_sz; 5649251478Sdelphij dev->l2ad_hand += buf_p_sz; 5650185029Spjd } 5651251478Sdelphij } 5652185029Spjd 5653286570Smav mutex_exit(&dev->l2ad_mtx); 5654185029Spjd 5655251478Sdelphij ASSERT3U(write_asize, <=, target_sz); 5656185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_sent); 5657251478Sdelphij ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5658185029Spjd ARCSTAT_INCR(arcstat_l2_size, write_sz); 5659251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, write_asize); 5660275096Sdelphij vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 5661185029Spjd 5662185029Spjd /* 5663185029Spjd * Bump device hand to the device start if it is approaching the end. 5664185029Spjd * l2arc_evict() will already have evicted ahead for this case. 5665185029Spjd */ 5666185029Spjd if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5667185029Spjd dev->l2ad_hand = dev->l2ad_start; 5668185029Spjd dev->l2ad_evict = dev->l2ad_start; 5669185029Spjd dev->l2ad_first = B_FALSE; 5670185029Spjd } 5671185029Spjd 5672208373Smm dev->l2ad_writing = B_TRUE; 5673185029Spjd (void) zio_wait(pio); 5674208373Smm dev->l2ad_writing = B_FALSE; 5675208373Smm 5676251478Sdelphij return (write_asize); 5677185029Spjd} 5678185029Spjd 5679185029Spjd/* 5680251478Sdelphij * Compresses an L2ARC buffer. 5681286570Smav * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its 5682251478Sdelphij * size in l2hdr->b_asize. This routine tries to compress the data and 5683251478Sdelphij * depending on the compression result there are three possible outcomes: 5684251478Sdelphij * *) The buffer was incompressible. The original l2hdr contents were left 5685251478Sdelphij * untouched and are ready for writing to an L2 device. 5686251478Sdelphij * *) The buffer was all-zeros, so there is no need to write it to an L2 5687251478Sdelphij * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5688251478Sdelphij * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5689251478Sdelphij * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5690251478Sdelphij * data buffer which holds the compressed data to be written, and b_asize 5691251478Sdelphij * tells us how much data there is. b_compress is set to the appropriate 5692251478Sdelphij * compression algorithm. Once writing is done, invoke 5693251478Sdelphij * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5694251478Sdelphij * 5695251478Sdelphij * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5696251478Sdelphij * buffer was incompressible). 5697251478Sdelphij */ 5698251478Sdelphijstatic boolean_t 5699286570Smavl2arc_compress_buf(arc_buf_hdr_t *hdr) 5700251478Sdelphij{ 5701251478Sdelphij void *cdata; 5702268075Sdelphij size_t csize, len, rounded; 5703286570Smav ASSERT(HDR_HAS_L2HDR(hdr)); 5704286570Smav l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 5705251478Sdelphij 5706286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 5707286570Smav ASSERT(HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF); 5708286570Smav ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 5709251478Sdelphij 5710251478Sdelphij len = l2hdr->b_asize; 5711251478Sdelphij cdata = zio_data_buf_alloc(len); 5712286570Smav ASSERT3P(cdata, !=, NULL); 5713286570Smav csize = zio_compress_data(ZIO_COMPRESS_LZ4, hdr->b_l1hdr.b_tmp_cdata, 5714269086Sdelphij cdata, l2hdr->b_asize); 5715251478Sdelphij 5716251478Sdelphij if (csize == 0) { 5717251478Sdelphij /* zero block, indicate that there's nothing to write */ 5718251478Sdelphij zio_data_buf_free(cdata, len); 5719286570Smav HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_EMPTY); 5720251478Sdelphij l2hdr->b_asize = 0; 5721286570Smav hdr->b_l1hdr.b_tmp_cdata = NULL; 5722251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5723251478Sdelphij return (B_TRUE); 5724274628Savg } 5725274628Savg 5726274628Savg rounded = P2ROUNDUP(csize, 5727274628Savg (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift); 5728274628Savg if (rounded < len) { 5729251478Sdelphij /* 5730251478Sdelphij * Compression succeeded, we'll keep the cdata around for 5731251478Sdelphij * writing and release it afterwards. 5732251478Sdelphij */ 5733274628Savg if (rounded > csize) { 5734274628Savg bzero((char *)cdata + csize, rounded - csize); 5735274628Savg csize = rounded; 5736274628Savg } 5737286570Smav HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_LZ4); 5738251478Sdelphij l2hdr->b_asize = csize; 5739286570Smav hdr->b_l1hdr.b_tmp_cdata = cdata; 5740251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_successes); 5741251478Sdelphij return (B_TRUE); 5742251478Sdelphij } else { 5743251478Sdelphij /* 5744251478Sdelphij * Compression failed, release the compressed buffer. 5745251478Sdelphij * l2hdr will be left unmodified. 5746251478Sdelphij */ 5747251478Sdelphij zio_data_buf_free(cdata, len); 5748251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_failures); 5749251478Sdelphij return (B_FALSE); 5750251478Sdelphij } 5751251478Sdelphij} 5752251478Sdelphij 5753251478Sdelphij/* 5754251478Sdelphij * Decompresses a zio read back from an l2arc device. On success, the 5755251478Sdelphij * underlying zio's io_data buffer is overwritten by the uncompressed 5756251478Sdelphij * version. On decompression error (corrupt compressed stream), the 5757251478Sdelphij * zio->io_error value is set to signal an I/O error. 5758251478Sdelphij * 5759251478Sdelphij * Please note that the compressed data stream is not checksummed, so 5760251478Sdelphij * if the underlying device is experiencing data corruption, we may feed 5761251478Sdelphij * corrupt data to the decompressor, so the decompressor needs to be 5762251478Sdelphij * able to handle this situation (LZ4 does). 5763251478Sdelphij */ 5764251478Sdelphijstatic void 5765251478Sdelphijl2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5766251478Sdelphij{ 5767251478Sdelphij ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5768251478Sdelphij 5769251478Sdelphij if (zio->io_error != 0) { 5770251478Sdelphij /* 5771251478Sdelphij * An io error has occured, just restore the original io 5772251478Sdelphij * size in preparation for a main pool read. 5773251478Sdelphij */ 5774251478Sdelphij zio->io_orig_size = zio->io_size = hdr->b_size; 5775251478Sdelphij return; 5776251478Sdelphij } 5777251478Sdelphij 5778251478Sdelphij if (c == ZIO_COMPRESS_EMPTY) { 5779251478Sdelphij /* 5780251478Sdelphij * An empty buffer results in a null zio, which means we 5781251478Sdelphij * need to fill its io_data after we're done restoring the 5782251478Sdelphij * buffer's contents. 5783251478Sdelphij */ 5784286570Smav ASSERT(hdr->b_l1hdr.b_buf != NULL); 5785286570Smav bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size); 5786286570Smav zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data; 5787251478Sdelphij } else { 5788251478Sdelphij ASSERT(zio->io_data != NULL); 5789251478Sdelphij /* 5790251478Sdelphij * We copy the compressed data from the start of the arc buffer 5791251478Sdelphij * (the zio_read will have pulled in only what we need, the 5792251478Sdelphij * rest is garbage which we will overwrite at decompression) 5793251478Sdelphij * and then decompress back to the ARC data buffer. This way we 5794251478Sdelphij * can minimize copying by simply decompressing back over the 5795251478Sdelphij * original compressed data (rather than decompressing to an 5796251478Sdelphij * aux buffer and then copying back the uncompressed buffer, 5797251478Sdelphij * which is likely to be much larger). 5798251478Sdelphij */ 5799251478Sdelphij uint64_t csize; 5800251478Sdelphij void *cdata; 5801251478Sdelphij 5802251478Sdelphij csize = zio->io_size; 5803251478Sdelphij cdata = zio_data_buf_alloc(csize); 5804251478Sdelphij bcopy(zio->io_data, cdata, csize); 5805251478Sdelphij if (zio_decompress_data(c, cdata, zio->io_data, csize, 5806251478Sdelphij hdr->b_size) != 0) 5807251478Sdelphij zio->io_error = EIO; 5808251478Sdelphij zio_data_buf_free(cdata, csize); 5809251478Sdelphij } 5810251478Sdelphij 5811251478Sdelphij /* Restore the expected uncompressed IO size. */ 5812251478Sdelphij zio->io_orig_size = zio->io_size = hdr->b_size; 5813251478Sdelphij} 5814251478Sdelphij 5815251478Sdelphij/* 5816251478Sdelphij * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5817251478Sdelphij * This buffer serves as a temporary holder of compressed data while 5818251478Sdelphij * the buffer entry is being written to an l2arc device. Once that is 5819251478Sdelphij * done, we can dispose of it. 5820251478Sdelphij */ 5821251478Sdelphijstatic void 5822275811Sdelphijl2arc_release_cdata_buf(arc_buf_hdr_t *hdr) 5823251478Sdelphij{ 5824286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 5825286570Smav if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_EMPTY) { 5826251478Sdelphij /* 5827251478Sdelphij * If the data was compressed, then we've allocated a 5828251478Sdelphij * temporary buffer for it, so now we need to release it. 5829251478Sdelphij */ 5830286570Smav ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 5831286570Smav zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata, 5832286570Smav hdr->b_size); 5833286570Smav hdr->b_l1hdr.b_tmp_cdata = NULL; 5834274172Savg } else { 5835286570Smav ASSERT(hdr->b_l1hdr.b_tmp_cdata == NULL); 5836251478Sdelphij } 5837251478Sdelphij} 5838251478Sdelphij 5839251478Sdelphij/* 5840185029Spjd * This thread feeds the L2ARC at regular intervals. This is the beating 5841185029Spjd * heart of the L2ARC. 5842185029Spjd */ 5843185029Spjdstatic void 5844185029Spjdl2arc_feed_thread(void *dummy __unused) 5845185029Spjd{ 5846185029Spjd callb_cpr_t cpr; 5847185029Spjd l2arc_dev_t *dev; 5848185029Spjd spa_t *spa; 5849208373Smm uint64_t size, wrote; 5850219089Spjd clock_t begin, next = ddi_get_lbolt(); 5851251478Sdelphij boolean_t headroom_boost = B_FALSE; 5852185029Spjd 5853185029Spjd CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5854185029Spjd 5855185029Spjd mutex_enter(&l2arc_feed_thr_lock); 5856185029Spjd 5857185029Spjd while (l2arc_thread_exit == 0) { 5858185029Spjd CALLB_CPR_SAFE_BEGIN(&cpr); 5859185029Spjd (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5860219089Spjd next - ddi_get_lbolt()); 5861185029Spjd CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5862219089Spjd next = ddi_get_lbolt() + hz; 5863185029Spjd 5864185029Spjd /* 5865185029Spjd * Quick check for L2ARC devices. 5866185029Spjd */ 5867185029Spjd mutex_enter(&l2arc_dev_mtx); 5868185029Spjd if (l2arc_ndev == 0) { 5869185029Spjd mutex_exit(&l2arc_dev_mtx); 5870185029Spjd continue; 5871185029Spjd } 5872185029Spjd mutex_exit(&l2arc_dev_mtx); 5873219089Spjd begin = ddi_get_lbolt(); 5874185029Spjd 5875185029Spjd /* 5876185029Spjd * This selects the next l2arc device to write to, and in 5877185029Spjd * doing so the next spa to feed from: dev->l2ad_spa. This 5878185029Spjd * will return NULL if there are now no l2arc devices or if 5879185029Spjd * they are all faulted. 5880185029Spjd * 5881185029Spjd * If a device is returned, its spa's config lock is also 5882185029Spjd * held to prevent device removal. l2arc_dev_get_next() 5883185029Spjd * will grab and release l2arc_dev_mtx. 5884185029Spjd */ 5885185029Spjd if ((dev = l2arc_dev_get_next()) == NULL) 5886185029Spjd continue; 5887185029Spjd 5888185029Spjd spa = dev->l2ad_spa; 5889185029Spjd ASSERT(spa != NULL); 5890185029Spjd 5891185029Spjd /* 5892219089Spjd * If the pool is read-only then force the feed thread to 5893219089Spjd * sleep a little longer. 5894219089Spjd */ 5895219089Spjd if (!spa_writeable(spa)) { 5896219089Spjd next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5897219089Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5898219089Spjd continue; 5899219089Spjd } 5900219089Spjd 5901219089Spjd /* 5902185029Spjd * Avoid contributing to memory pressure. 5903185029Spjd */ 5904185029Spjd if (arc_reclaim_needed()) { 5905185029Spjd ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5906185029Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5907185029Spjd continue; 5908185029Spjd } 5909185029Spjd 5910185029Spjd ARCSTAT_BUMP(arcstat_l2_feeds); 5911185029Spjd 5912251478Sdelphij size = l2arc_write_size(); 5913185029Spjd 5914185029Spjd /* 5915185029Spjd * Evict L2ARC buffers that will be overwritten. 5916185029Spjd */ 5917185029Spjd l2arc_evict(dev, size, B_FALSE); 5918185029Spjd 5919185029Spjd /* 5920185029Spjd * Write ARC buffers. 5921185029Spjd */ 5922251478Sdelphij wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5923208373Smm 5924208373Smm /* 5925208373Smm * Calculate interval between writes. 5926208373Smm */ 5927208373Smm next = l2arc_write_interval(begin, size, wrote); 5928185029Spjd spa_config_exit(spa, SCL_L2ARC, dev); 5929185029Spjd } 5930185029Spjd 5931185029Spjd l2arc_thread_exit = 0; 5932185029Spjd cv_broadcast(&l2arc_feed_thr_cv); 5933185029Spjd CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5934185029Spjd thread_exit(); 5935185029Spjd} 5936185029Spjd 5937185029Spjdboolean_t 5938185029Spjdl2arc_vdev_present(vdev_t *vd) 5939185029Spjd{ 5940185029Spjd l2arc_dev_t *dev; 5941185029Spjd 5942185029Spjd mutex_enter(&l2arc_dev_mtx); 5943185029Spjd for (dev = list_head(l2arc_dev_list); dev != NULL; 5944185029Spjd dev = list_next(l2arc_dev_list, dev)) { 5945185029Spjd if (dev->l2ad_vdev == vd) 5946185029Spjd break; 5947185029Spjd } 5948185029Spjd mutex_exit(&l2arc_dev_mtx); 5949185029Spjd 5950185029Spjd return (dev != NULL); 5951185029Spjd} 5952185029Spjd 5953185029Spjd/* 5954185029Spjd * Add a vdev for use by the L2ARC. By this point the spa has already 5955185029Spjd * validated the vdev and opened it. 5956185029Spjd */ 5957185029Spjdvoid 5958219089Spjdl2arc_add_vdev(spa_t *spa, vdev_t *vd) 5959185029Spjd{ 5960185029Spjd l2arc_dev_t *adddev; 5961185029Spjd 5962185029Spjd ASSERT(!l2arc_vdev_present(vd)); 5963185029Spjd 5964255753Sgibbs vdev_ashift_optimize(vd); 5965255753Sgibbs 5966185029Spjd /* 5967185029Spjd * Create a new l2arc device entry. 5968185029Spjd */ 5969185029Spjd adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5970185029Spjd adddev->l2ad_spa = spa; 5971185029Spjd adddev->l2ad_vdev = vd; 5972219089Spjd adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5973219089Spjd adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5974185029Spjd adddev->l2ad_hand = adddev->l2ad_start; 5975185029Spjd adddev->l2ad_evict = adddev->l2ad_start; 5976185029Spjd adddev->l2ad_first = B_TRUE; 5977208373Smm adddev->l2ad_writing = B_FALSE; 5978185029Spjd 5979286570Smav mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 5980185029Spjd /* 5981185029Spjd * This is a list of all ARC buffers that are still valid on the 5982185029Spjd * device. 5983185029Spjd */ 5984286570Smav list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5985286570Smav offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 5986185029Spjd 5987219089Spjd vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5988185029Spjd 5989185029Spjd /* 5990185029Spjd * Add device to global list 5991185029Spjd */ 5992185029Spjd mutex_enter(&l2arc_dev_mtx); 5993185029Spjd list_insert_head(l2arc_dev_list, adddev); 5994185029Spjd atomic_inc_64(&l2arc_ndev); 5995185029Spjd mutex_exit(&l2arc_dev_mtx); 5996185029Spjd} 5997185029Spjd 5998185029Spjd/* 5999185029Spjd * Remove a vdev from the L2ARC. 6000185029Spjd */ 6001185029Spjdvoid 6002185029Spjdl2arc_remove_vdev(vdev_t *vd) 6003185029Spjd{ 6004185029Spjd l2arc_dev_t *dev, *nextdev, *remdev = NULL; 6005185029Spjd 6006185029Spjd /* 6007185029Spjd * Find the device by vdev 6008185029Spjd */ 6009185029Spjd mutex_enter(&l2arc_dev_mtx); 6010185029Spjd for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 6011185029Spjd nextdev = list_next(l2arc_dev_list, dev); 6012185029Spjd if (vd == dev->l2ad_vdev) { 6013185029Spjd remdev = dev; 6014185029Spjd break; 6015185029Spjd } 6016185029Spjd } 6017185029Spjd ASSERT(remdev != NULL); 6018185029Spjd 6019185029Spjd /* 6020185029Spjd * Remove device from global list 6021185029Spjd */ 6022185029Spjd list_remove(l2arc_dev_list, remdev); 6023185029Spjd l2arc_dev_last = NULL; /* may have been invalidated */ 6024185029Spjd atomic_dec_64(&l2arc_ndev); 6025185029Spjd mutex_exit(&l2arc_dev_mtx); 6026185029Spjd 6027185029Spjd /* 6028185029Spjd * Clear all buflists and ARC references. L2ARC device flush. 6029185029Spjd */ 6030185029Spjd l2arc_evict(remdev, 0, B_TRUE); 6031286570Smav list_destroy(&remdev->l2ad_buflist); 6032286570Smav mutex_destroy(&remdev->l2ad_mtx); 6033185029Spjd kmem_free(remdev, sizeof (l2arc_dev_t)); 6034185029Spjd} 6035185029Spjd 6036185029Spjdvoid 6037185029Spjdl2arc_init(void) 6038185029Spjd{ 6039185029Spjd l2arc_thread_exit = 0; 6040185029Spjd l2arc_ndev = 0; 6041185029Spjd l2arc_writes_sent = 0; 6042185029Spjd l2arc_writes_done = 0; 6043185029Spjd 6044185029Spjd mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 6045185029Spjd cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 6046185029Spjd mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 6047185029Spjd mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 6048185029Spjd 6049185029Spjd l2arc_dev_list = &L2ARC_dev_list; 6050185029Spjd l2arc_free_on_write = &L2ARC_free_on_write; 6051185029Spjd list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 6052185029Spjd offsetof(l2arc_dev_t, l2ad_node)); 6053185029Spjd list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 6054185029Spjd offsetof(l2arc_data_free_t, l2df_list_node)); 6055185029Spjd} 6056185029Spjd 6057185029Spjdvoid 6058185029Spjdl2arc_fini(void) 6059185029Spjd{ 6060185029Spjd /* 6061185029Spjd * This is called from dmu_fini(), which is called from spa_fini(); 6062185029Spjd * Because of this, we can assume that all l2arc devices have 6063185029Spjd * already been removed when the pools themselves were removed. 6064185029Spjd */ 6065185029Spjd 6066185029Spjd l2arc_do_free_on_write(); 6067185029Spjd 6068185029Spjd mutex_destroy(&l2arc_feed_thr_lock); 6069185029Spjd cv_destroy(&l2arc_feed_thr_cv); 6070185029Spjd mutex_destroy(&l2arc_dev_mtx); 6071185029Spjd mutex_destroy(&l2arc_free_on_write_mtx); 6072185029Spjd 6073185029Spjd list_destroy(l2arc_dev_list); 6074185029Spjd list_destroy(l2arc_free_on_write); 6075185029Spjd} 6076185029Spjd 6077185029Spjdvoid 6078185029Spjdl2arc_start(void) 6079185029Spjd{ 6080209962Smm if (!(spa_mode_global & FWRITE)) 6081185029Spjd return; 6082185029Spjd 6083185029Spjd (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 6084185029Spjd TS_RUN, minclsyspri); 6085185029Spjd} 6086185029Spjd 6087185029Spjdvoid 6088185029Spjdl2arc_stop(void) 6089185029Spjd{ 6090209962Smm if (!(spa_mode_global & FWRITE)) 6091185029Spjd return; 6092185029Spjd 6093185029Spjd mutex_enter(&l2arc_feed_thr_lock); 6094185029Spjd cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 6095185029Spjd l2arc_thread_exit = 1; 6096185029Spjd while (l2arc_thread_exit != 0) 6097185029Spjd cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 6098185029Spjd mutex_exit(&l2arc_feed_thr_lock); 6099185029Spjd} 6100