arc.c revision 288517
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 24 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 25 * Copyright 2014 Nexenta Systems, Inc. All rights reserved. 26 */ 27 28/* 29 * DVA-based Adjustable Replacement Cache 30 * 31 * While much of the theory of operation used here is 32 * based on the self-tuning, low overhead replacement cache 33 * presented by Megiddo and Modha at FAST 2003, there are some 34 * significant differences: 35 * 36 * 1. The Megiddo and Modha model assumes any page is evictable. 37 * Pages in its cache cannot be "locked" into memory. This makes 38 * the eviction algorithm simple: evict the last page in the list. 39 * This also make the performance characteristics easy to reason 40 * about. Our cache is not so simple. At any given moment, some 41 * subset of the blocks in the cache are un-evictable because we 42 * have handed out a reference to them. Blocks are only evictable 43 * when there are no external references active. This makes 44 * eviction far more problematic: we choose to evict the evictable 45 * blocks that are the "lowest" in the list. 46 * 47 * There are times when it is not possible to evict the requested 48 * space. In these circumstances we are unable to adjust the cache 49 * size. To prevent the cache growing unbounded at these times we 50 * implement a "cache throttle" that slows the flow of new data 51 * into the cache until we can make space available. 52 * 53 * 2. The Megiddo and Modha model assumes a fixed cache size. 54 * Pages are evicted when the cache is full and there is a cache 55 * miss. Our model has a variable sized cache. It grows with 56 * high use, but also tries to react to memory pressure from the 57 * operating system: decreasing its size when system memory is 58 * tight. 59 * 60 * 3. The Megiddo and Modha model assumes a fixed page size. All 61 * elements of the cache are therefore exactly the same size. So 62 * when adjusting the cache size following a cache miss, its simply 63 * a matter of choosing a single page to evict. In our model, we 64 * have variable sized cache blocks (rangeing from 512 bytes to 65 * 128K bytes). We therefore choose a set of blocks to evict to make 66 * space for a cache miss that approximates as closely as possible 67 * the space used by the new block. 68 * 69 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 70 * by N. Megiddo & D. Modha, FAST 2003 71 */ 72 73/* 74 * The locking model: 75 * 76 * A new reference to a cache buffer can be obtained in two 77 * ways: 1) via a hash table lookup using the DVA as a key, 78 * or 2) via one of the ARC lists. The arc_read() interface 79 * uses method 1, while the internal arc algorithms for 80 * adjusting the cache use method 2. We therefore provide two 81 * types of locks: 1) the hash table lock array, and 2) the 82 * arc list locks. 83 * 84 * Buffers do not have their own mutexs, rather they rely on the 85 * hash table mutexs for the bulk of their protection (i.e. most 86 * fields in the arc_buf_hdr_t are protected by these mutexs). 87 * 88 * buf_hash_find() returns the appropriate mutex (held) when it 89 * locates the requested buffer in the hash table. It returns 90 * NULL for the mutex if the buffer was not in the table. 91 * 92 * buf_hash_remove() expects the appropriate hash mutex to be 93 * already held before it is invoked. 94 * 95 * Each arc state also has a mutex which is used to protect the 96 * buffer list associated with the state. When attempting to 97 * obtain a hash table lock while holding an arc list lock you 98 * must use: mutex_tryenter() to avoid deadlock. Also note that 99 * the active state mutex must be held before the ghost state mutex. 100 * 101 * Arc buffers may have an associated eviction callback function. 102 * This function will be invoked prior to removing the buffer (e.g. 103 * in arc_do_user_evicts()). Note however that the data associated 104 * with the buffer may be evicted prior to the callback. The callback 105 * must be made with *no locks held* (to prevent deadlock). Additionally, 106 * the users of callbacks must ensure that their private data is 107 * protected from simultaneous callbacks from arc_clear_callback() 108 * and arc_do_user_evicts(). 109 * 110 * Note that the majority of the performance stats are manipulated 111 * with atomic operations. 112 * 113 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following: 114 * 115 * - L2ARC buflist creation 116 * - L2ARC buflist eviction 117 * - L2ARC write completion, which walks L2ARC buflists 118 * - ARC header destruction, as it removes from L2ARC buflists 119 * - ARC header release, as it removes from L2ARC buflists 120 */ 121 122#include <sys/spa.h> 123#include <sys/zio.h> 124#include <sys/zio_compress.h> 125#include <sys/zfs_context.h> 126#include <sys/arc.h> 127#include <sys/refcount.h> 128#include <sys/vdev.h> 129#include <sys/vdev_impl.h> 130#include <sys/dsl_pool.h> 131#ifdef _KERNEL 132#include <sys/dnlc.h> 133#endif 134#include <sys/callb.h> 135#include <sys/kstat.h> 136#include <sys/trim_map.h> 137#include <zfs_fletcher.h> 138#include <sys/sdt.h> 139 140#include <vm/vm_pageout.h> 141#include <machine/vmparam.h> 142 143#ifdef illumos 144#ifndef _KERNEL 145/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 146boolean_t arc_watch = B_FALSE; 147int arc_procfd; 148#endif 149#endif /* illumos */ 150 151static kmutex_t arc_reclaim_thr_lock; 152static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 153static uint8_t arc_thread_exit; 154 155#define ARC_REDUCE_DNLC_PERCENT 3 156uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 157 158typedef enum arc_reclaim_strategy { 159 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 160 ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 161} arc_reclaim_strategy_t; 162 163/* 164 * The number of iterations through arc_evict_*() before we 165 * drop & reacquire the lock. 166 */ 167int arc_evict_iterations = 100; 168 169/* number of seconds before growing cache again */ 170static int arc_grow_retry = 60; 171 172/* shift of arc_c for calculating both min and max arc_p */ 173static int arc_p_min_shift = 4; 174 175/* log2(fraction of arc to reclaim) */ 176static int arc_shrink_shift = 5; 177 178/* 179 * minimum lifespan of a prefetch block in clock ticks 180 * (initialized in arc_init()) 181 */ 182static int arc_min_prefetch_lifespan; 183 184/* 185 * If this percent of memory is free, don't throttle. 186 */ 187int arc_lotsfree_percent = 10; 188 189static int arc_dead; 190extern int zfs_prefetch_disable; 191 192/* 193 * The arc has filled available memory and has now warmed up. 194 */ 195static boolean_t arc_warm; 196 197uint64_t zfs_arc_max; 198uint64_t zfs_arc_min; 199uint64_t zfs_arc_meta_limit = 0; 200uint64_t zfs_arc_meta_min = 0; 201int zfs_arc_grow_retry = 0; 202int zfs_arc_shrink_shift = 0; 203int zfs_arc_p_min_shift = 0; 204int zfs_disable_dup_eviction = 0; 205uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 206u_int zfs_arc_free_target = 0; 207 208static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 209static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 210 211#ifdef _KERNEL 212static void 213arc_free_target_init(void *unused __unused) 214{ 215 216 zfs_arc_free_target = vm_pageout_wakeup_thresh; 217} 218SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 219 arc_free_target_init, NULL); 220 221TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max); 222TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 223TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 224TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 225TUNABLE_QUAD("vfs.zfs.arc_average_blocksize", &zfs_arc_average_blocksize); 226TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 227SYSCTL_DECL(_vfs_zfs); 228SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 229 "Maximum ARC size"); 230SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 231 "Minimum ARC size"); 232SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 233 &zfs_arc_average_blocksize, 0, 234 "ARC average blocksize"); 235SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 236 &arc_shrink_shift, 0, 237 "log2(fraction of arc to reclaim)"); 238 239/* 240 * We don't have a tunable for arc_free_target due to the dependency on 241 * pagedaemon initialisation. 242 */ 243SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 244 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 245 sysctl_vfs_zfs_arc_free_target, "IU", 246 "Desired number of free pages below which ARC triggers reclaim"); 247 248static int 249sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 250{ 251 u_int val; 252 int err; 253 254 val = zfs_arc_free_target; 255 err = sysctl_handle_int(oidp, &val, 0, req); 256 if (err != 0 || req->newptr == NULL) 257 return (err); 258 259 if (val < minfree) 260 return (EINVAL); 261 if (val > cnt.v_page_count) 262 return (EINVAL); 263 264 zfs_arc_free_target = val; 265 266 return (0); 267} 268 269/* 270 * Must be declared here, before the definition of corresponding kstat 271 * macro which uses the same names will confuse the compiler. 272 */ 273SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 274 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 275 sysctl_vfs_zfs_arc_meta_limit, "QU", 276 "ARC metadata limit"); 277#endif 278 279/* 280 * Note that buffers can be in one of 6 states: 281 * ARC_anon - anonymous (discussed below) 282 * ARC_mru - recently used, currently cached 283 * ARC_mru_ghost - recentely used, no longer in cache 284 * ARC_mfu - frequently used, currently cached 285 * ARC_mfu_ghost - frequently used, no longer in cache 286 * ARC_l2c_only - exists in L2ARC but not other states 287 * When there are no active references to the buffer, they are 288 * are linked onto a list in one of these arc states. These are 289 * the only buffers that can be evicted or deleted. Within each 290 * state there are multiple lists, one for meta-data and one for 291 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 292 * etc.) is tracked separately so that it can be managed more 293 * explicitly: favored over data, limited explicitly. 294 * 295 * Anonymous buffers are buffers that are not associated with 296 * a DVA. These are buffers that hold dirty block copies 297 * before they are written to stable storage. By definition, 298 * they are "ref'd" and are considered part of arc_mru 299 * that cannot be freed. Generally, they will aquire a DVA 300 * as they are written and migrate onto the arc_mru list. 301 * 302 * The ARC_l2c_only state is for buffers that are in the second 303 * level ARC but no longer in any of the ARC_m* lists. The second 304 * level ARC itself may also contain buffers that are in any of 305 * the ARC_m* states - meaning that a buffer can exist in two 306 * places. The reason for the ARC_l2c_only state is to keep the 307 * buffer header in the hash table, so that reads that hit the 308 * second level ARC benefit from these fast lookups. 309 */ 310 311#define ARCS_LOCK_PAD CACHE_LINE_SIZE 312struct arcs_lock { 313 kmutex_t arcs_lock; 314#ifdef _KERNEL 315 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 316#endif 317}; 318 319/* 320 * must be power of two for mask use to work 321 * 322 */ 323#define ARC_BUFC_NUMDATALISTS 16 324#define ARC_BUFC_NUMMETADATALISTS 16 325#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 326 327typedef struct arc_state { 328 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 329 uint64_t arcs_size; /* total amount of data in this state */ 330 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 331 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 332} arc_state_t; 333 334#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 335 336/* The 6 states: */ 337static arc_state_t ARC_anon; 338static arc_state_t ARC_mru; 339static arc_state_t ARC_mru_ghost; 340static arc_state_t ARC_mfu; 341static arc_state_t ARC_mfu_ghost; 342static arc_state_t ARC_l2c_only; 343 344typedef struct arc_stats { 345 kstat_named_t arcstat_hits; 346 kstat_named_t arcstat_misses; 347 kstat_named_t arcstat_demand_data_hits; 348 kstat_named_t arcstat_demand_data_misses; 349 kstat_named_t arcstat_demand_metadata_hits; 350 kstat_named_t arcstat_demand_metadata_misses; 351 kstat_named_t arcstat_prefetch_data_hits; 352 kstat_named_t arcstat_prefetch_data_misses; 353 kstat_named_t arcstat_prefetch_metadata_hits; 354 kstat_named_t arcstat_prefetch_metadata_misses; 355 kstat_named_t arcstat_mru_hits; 356 kstat_named_t arcstat_mru_ghost_hits; 357 kstat_named_t arcstat_mfu_hits; 358 kstat_named_t arcstat_mfu_ghost_hits; 359 kstat_named_t arcstat_allocated; 360 kstat_named_t arcstat_deleted; 361 kstat_named_t arcstat_stolen; 362 kstat_named_t arcstat_recycle_miss; 363 /* 364 * Number of buffers that could not be evicted because the hash lock 365 * was held by another thread. The lock may not necessarily be held 366 * by something using the same buffer, since hash locks are shared 367 * by multiple buffers. 368 */ 369 kstat_named_t arcstat_mutex_miss; 370 /* 371 * Number of buffers skipped because they have I/O in progress, are 372 * indrect prefetch buffers that have not lived long enough, or are 373 * not from the spa we're trying to evict from. 374 */ 375 kstat_named_t arcstat_evict_skip; 376 kstat_named_t arcstat_evict_l2_cached; 377 kstat_named_t arcstat_evict_l2_eligible; 378 kstat_named_t arcstat_evict_l2_ineligible; 379 kstat_named_t arcstat_hash_elements; 380 kstat_named_t arcstat_hash_elements_max; 381 kstat_named_t arcstat_hash_collisions; 382 kstat_named_t arcstat_hash_chains; 383 kstat_named_t arcstat_hash_chain_max; 384 kstat_named_t arcstat_p; 385 kstat_named_t arcstat_c; 386 kstat_named_t arcstat_c_min; 387 kstat_named_t arcstat_c_max; 388 kstat_named_t arcstat_size; 389 kstat_named_t arcstat_hdr_size; 390 kstat_named_t arcstat_data_size; 391 kstat_named_t arcstat_other_size; 392 kstat_named_t arcstat_l2_hits; 393 kstat_named_t arcstat_l2_misses; 394 kstat_named_t arcstat_l2_feeds; 395 kstat_named_t arcstat_l2_rw_clash; 396 kstat_named_t arcstat_l2_read_bytes; 397 kstat_named_t arcstat_l2_write_bytes; 398 kstat_named_t arcstat_l2_writes_sent; 399 kstat_named_t arcstat_l2_writes_done; 400 kstat_named_t arcstat_l2_writes_error; 401 kstat_named_t arcstat_l2_writes_hdr_miss; 402 kstat_named_t arcstat_l2_evict_lock_retry; 403 kstat_named_t arcstat_l2_evict_reading; 404 kstat_named_t arcstat_l2_free_on_write; 405 kstat_named_t arcstat_l2_cdata_free_on_write; 406 kstat_named_t arcstat_l2_abort_lowmem; 407 kstat_named_t arcstat_l2_cksum_bad; 408 kstat_named_t arcstat_l2_io_error; 409 kstat_named_t arcstat_l2_size; 410 kstat_named_t arcstat_l2_asize; 411 kstat_named_t arcstat_l2_hdr_size; 412 kstat_named_t arcstat_l2_compress_successes; 413 kstat_named_t arcstat_l2_compress_zeros; 414 kstat_named_t arcstat_l2_compress_failures; 415 kstat_named_t arcstat_l2_write_trylock_fail; 416 kstat_named_t arcstat_l2_write_passed_headroom; 417 kstat_named_t arcstat_l2_write_spa_mismatch; 418 kstat_named_t arcstat_l2_write_in_l2; 419 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 420 kstat_named_t arcstat_l2_write_not_cacheable; 421 kstat_named_t arcstat_l2_write_full; 422 kstat_named_t arcstat_l2_write_buffer_iter; 423 kstat_named_t arcstat_l2_write_pios; 424 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 425 kstat_named_t arcstat_l2_write_buffer_list_iter; 426 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 427 kstat_named_t arcstat_memory_throttle_count; 428 kstat_named_t arcstat_duplicate_buffers; 429 kstat_named_t arcstat_duplicate_buffers_size; 430 kstat_named_t arcstat_duplicate_reads; 431 kstat_named_t arcstat_meta_used; 432 kstat_named_t arcstat_meta_limit; 433 kstat_named_t arcstat_meta_max; 434 kstat_named_t arcstat_meta_min; 435} arc_stats_t; 436 437static arc_stats_t arc_stats = { 438 { "hits", KSTAT_DATA_UINT64 }, 439 { "misses", KSTAT_DATA_UINT64 }, 440 { "demand_data_hits", KSTAT_DATA_UINT64 }, 441 { "demand_data_misses", KSTAT_DATA_UINT64 }, 442 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 443 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 444 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 445 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 446 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 447 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 448 { "mru_hits", KSTAT_DATA_UINT64 }, 449 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 450 { "mfu_hits", KSTAT_DATA_UINT64 }, 451 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 452 { "allocated", KSTAT_DATA_UINT64 }, 453 { "deleted", KSTAT_DATA_UINT64 }, 454 { "stolen", KSTAT_DATA_UINT64 }, 455 { "recycle_miss", KSTAT_DATA_UINT64 }, 456 { "mutex_miss", KSTAT_DATA_UINT64 }, 457 { "evict_skip", KSTAT_DATA_UINT64 }, 458 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 459 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 460 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 461 { "hash_elements", KSTAT_DATA_UINT64 }, 462 { "hash_elements_max", KSTAT_DATA_UINT64 }, 463 { "hash_collisions", KSTAT_DATA_UINT64 }, 464 { "hash_chains", KSTAT_DATA_UINT64 }, 465 { "hash_chain_max", KSTAT_DATA_UINT64 }, 466 { "p", KSTAT_DATA_UINT64 }, 467 { "c", KSTAT_DATA_UINT64 }, 468 { "c_min", KSTAT_DATA_UINT64 }, 469 { "c_max", KSTAT_DATA_UINT64 }, 470 { "size", KSTAT_DATA_UINT64 }, 471 { "hdr_size", KSTAT_DATA_UINT64 }, 472 { "data_size", KSTAT_DATA_UINT64 }, 473 { "other_size", KSTAT_DATA_UINT64 }, 474 { "l2_hits", KSTAT_DATA_UINT64 }, 475 { "l2_misses", KSTAT_DATA_UINT64 }, 476 { "l2_feeds", KSTAT_DATA_UINT64 }, 477 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 478 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 479 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 480 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 481 { "l2_writes_done", KSTAT_DATA_UINT64 }, 482 { "l2_writes_error", KSTAT_DATA_UINT64 }, 483 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 484 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 485 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 486 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 487 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, 488 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 489 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 490 { "l2_io_error", KSTAT_DATA_UINT64 }, 491 { "l2_size", KSTAT_DATA_UINT64 }, 492 { "l2_asize", KSTAT_DATA_UINT64 }, 493 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 494 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 495 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 496 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 497 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 498 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 499 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 500 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 501 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 502 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 503 { "l2_write_full", KSTAT_DATA_UINT64 }, 504 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 505 { "l2_write_pios", KSTAT_DATA_UINT64 }, 506 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 507 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 508 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 509 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 510 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 511 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 512 { "duplicate_reads", KSTAT_DATA_UINT64 }, 513 { "arc_meta_used", KSTAT_DATA_UINT64 }, 514 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 515 { "arc_meta_max", KSTAT_DATA_UINT64 }, 516 { "arc_meta_min", KSTAT_DATA_UINT64 } 517}; 518 519#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 520 521#define ARCSTAT_INCR(stat, val) \ 522 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 523 524#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 525#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 526 527#define ARCSTAT_MAX(stat, val) { \ 528 uint64_t m; \ 529 while ((val) > (m = arc_stats.stat.value.ui64) && \ 530 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 531 continue; \ 532} 533 534#define ARCSTAT_MAXSTAT(stat) \ 535 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 536 537/* 538 * We define a macro to allow ARC hits/misses to be easily broken down by 539 * two separate conditions, giving a total of four different subtypes for 540 * each of hits and misses (so eight statistics total). 541 */ 542#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 543 if (cond1) { \ 544 if (cond2) { \ 545 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 546 } else { \ 547 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 548 } \ 549 } else { \ 550 if (cond2) { \ 551 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 552 } else { \ 553 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 554 } \ 555 } 556 557kstat_t *arc_ksp; 558static arc_state_t *arc_anon; 559static arc_state_t *arc_mru; 560static arc_state_t *arc_mru_ghost; 561static arc_state_t *arc_mfu; 562static arc_state_t *arc_mfu_ghost; 563static arc_state_t *arc_l2c_only; 564 565/* 566 * There are several ARC variables that are critical to export as kstats -- 567 * but we don't want to have to grovel around in the kstat whenever we wish to 568 * manipulate them. For these variables, we therefore define them to be in 569 * terms of the statistic variable. This assures that we are not introducing 570 * the possibility of inconsistency by having shadow copies of the variables, 571 * while still allowing the code to be readable. 572 */ 573#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 574#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 575#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 576#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 577#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 578#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 579#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 580#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 581#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 582 583#define L2ARC_IS_VALID_COMPRESS(_c_) \ 584 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 585 586static int arc_no_grow; /* Don't try to grow cache size */ 587static uint64_t arc_tempreserve; 588static uint64_t arc_loaned_bytes; 589 590typedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 591 592typedef struct arc_callback arc_callback_t; 593 594struct arc_callback { 595 void *acb_private; 596 arc_done_func_t *acb_done; 597 arc_buf_t *acb_buf; 598 zio_t *acb_zio_dummy; 599 arc_callback_t *acb_next; 600}; 601 602typedef struct arc_write_callback arc_write_callback_t; 603 604struct arc_write_callback { 605 void *awcb_private; 606 arc_done_func_t *awcb_ready; 607 arc_done_func_t *awcb_physdone; 608 arc_done_func_t *awcb_done; 609 arc_buf_t *awcb_buf; 610}; 611 612struct arc_buf_hdr { 613 /* protected by hash lock */ 614 dva_t b_dva; 615 uint64_t b_birth; 616 uint64_t b_cksum0; 617 618 kmutex_t b_freeze_lock; 619 zio_cksum_t *b_freeze_cksum; 620 void *b_thawed; 621 622 arc_buf_hdr_t *b_hash_next; 623 arc_buf_t *b_buf; 624 arc_flags_t b_flags; 625 uint32_t b_datacnt; 626 627 arc_callback_t *b_acb; 628 kcondvar_t b_cv; 629 630 /* immutable */ 631 arc_buf_contents_t b_type; 632 uint64_t b_size; 633 uint64_t b_spa; 634 635 /* protected by arc state mutex */ 636 arc_state_t *b_state; 637 list_node_t b_arc_node; 638 639 /* updated atomically */ 640 clock_t b_arc_access; 641 642 /* self protecting */ 643 refcount_t b_refcnt; 644 645 l2arc_buf_hdr_t *b_l2hdr; 646 list_node_t b_l2node; 647}; 648 649#ifdef _KERNEL 650static int 651sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 652{ 653 uint64_t val; 654 int err; 655 656 val = arc_meta_limit; 657 err = sysctl_handle_64(oidp, &val, 0, req); 658 if (err != 0 || req->newptr == NULL) 659 return (err); 660 661 if (val <= 0 || val > arc_c_max) 662 return (EINVAL); 663 664 arc_meta_limit = val; 665 return (0); 666} 667#endif 668 669static arc_buf_t *arc_eviction_list; 670static kmutex_t arc_eviction_mtx; 671static arc_buf_hdr_t arc_eviction_hdr; 672 673#define GHOST_STATE(state) \ 674 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 675 (state) == arc_l2c_only) 676 677#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 678#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 679#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 680#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 681#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ) 682#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE) 683#define HDR_FREE_IN_PROGRESS(hdr) \ 684 ((hdr)->b_flags & ARC_FLAG_FREE_IN_PROGRESS) 685#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 686#define HDR_L2_READING(hdr) \ 687 ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS && \ 688 (hdr)->b_l2hdr != NULL) 689#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 690#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 691#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 692 693/* 694 * Other sizes 695 */ 696 697#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 698#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 699 700/* 701 * Hash table routines 702 */ 703 704#define HT_LOCK_PAD CACHE_LINE_SIZE 705 706struct ht_lock { 707 kmutex_t ht_lock; 708#ifdef _KERNEL 709 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 710#endif 711}; 712 713#define BUF_LOCKS 256 714typedef struct buf_hash_table { 715 uint64_t ht_mask; 716 arc_buf_hdr_t **ht_table; 717 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 718} buf_hash_table_t; 719 720static buf_hash_table_t buf_hash_table; 721 722#define BUF_HASH_INDEX(spa, dva, birth) \ 723 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 724#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 725#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 726#define HDR_LOCK(hdr) \ 727 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 728 729uint64_t zfs_crc64_table[256]; 730 731/* 732 * Level 2 ARC 733 */ 734 735#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 736#define L2ARC_HEADROOM 2 /* num of writes */ 737/* 738 * If we discover during ARC scan any buffers to be compressed, we boost 739 * our headroom for the next scanning cycle by this percentage multiple. 740 */ 741#define L2ARC_HEADROOM_BOOST 200 742#define L2ARC_FEED_SECS 1 /* caching interval secs */ 743#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 744 745#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 746#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 747 748/* L2ARC Performance Tunables */ 749uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 750uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 751uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 752uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 753uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 754uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 755boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 756boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 757boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 758 759SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 760 &l2arc_write_max, 0, "max write size"); 761SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 762 &l2arc_write_boost, 0, "extra write during warmup"); 763SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 764 &l2arc_headroom, 0, "number of dev writes"); 765SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 766 &l2arc_feed_secs, 0, "interval seconds"); 767SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 768 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 769 770SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 771 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 772SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 773 &l2arc_feed_again, 0, "turbo warmup"); 774SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 775 &l2arc_norw, 0, "no reads during writes"); 776 777SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 778 &ARC_anon.arcs_size, 0, "size of anonymous state"); 779SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 780 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 781SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 782 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 783 784SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 785 &ARC_mru.arcs_size, 0, "size of mru state"); 786SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 787 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 788SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 789 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 790 791SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 792 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 793SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 794 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 795 "size of metadata in mru ghost state"); 796SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 797 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 798 "size of data in mru ghost state"); 799 800SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 801 &ARC_mfu.arcs_size, 0, "size of mfu state"); 802SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 803 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 804SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 805 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 806 807SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 808 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 809SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 810 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 811 "size of metadata in mfu ghost state"); 812SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 813 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 814 "size of data in mfu ghost state"); 815 816SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 817 &ARC_l2c_only.arcs_size, 0, "size of mru state"); 818 819/* 820 * L2ARC Internals 821 */ 822typedef struct l2arc_dev { 823 vdev_t *l2ad_vdev; /* vdev */ 824 spa_t *l2ad_spa; /* spa */ 825 uint64_t l2ad_hand; /* next write location */ 826 uint64_t l2ad_start; /* first addr on device */ 827 uint64_t l2ad_end; /* last addr on device */ 828 uint64_t l2ad_evict; /* last addr eviction reached */ 829 boolean_t l2ad_first; /* first sweep through */ 830 boolean_t l2ad_writing; /* currently writing */ 831 list_t *l2ad_buflist; /* buffer list */ 832 list_node_t l2ad_node; /* device list node */ 833} l2arc_dev_t; 834 835static list_t L2ARC_dev_list; /* device list */ 836static list_t *l2arc_dev_list; /* device list pointer */ 837static kmutex_t l2arc_dev_mtx; /* device list mutex */ 838static l2arc_dev_t *l2arc_dev_last; /* last device used */ 839static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 840static list_t L2ARC_free_on_write; /* free after write buf list */ 841static list_t *l2arc_free_on_write; /* free after write list ptr */ 842static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 843static uint64_t l2arc_ndev; /* number of devices */ 844 845typedef struct l2arc_read_callback { 846 arc_buf_t *l2rcb_buf; /* read buffer */ 847 spa_t *l2rcb_spa; /* spa */ 848 blkptr_t l2rcb_bp; /* original blkptr */ 849 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 850 int l2rcb_flags; /* original flags */ 851 enum zio_compress l2rcb_compress; /* applied compress */ 852} l2arc_read_callback_t; 853 854typedef struct l2arc_write_callback { 855 l2arc_dev_t *l2wcb_dev; /* device info */ 856 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 857} l2arc_write_callback_t; 858 859struct l2arc_buf_hdr { 860 /* protected by arc_buf_hdr mutex */ 861 l2arc_dev_t *b_dev; /* L2ARC device */ 862 uint64_t b_daddr; /* disk address, offset byte */ 863 /* compression applied to buffer data */ 864 enum zio_compress b_compress; 865 /* real alloc'd buffer size depending on b_compress applied */ 866 int b_asize; 867 /* temporary buffer holder for in-flight compressed data */ 868 void *b_tmp_cdata; 869}; 870 871typedef struct l2arc_data_free { 872 /* protected by l2arc_free_on_write_mtx */ 873 void *l2df_data; 874 size_t l2df_size; 875 void (*l2df_func)(void *, size_t); 876 list_node_t l2df_list_node; 877} l2arc_data_free_t; 878 879static kmutex_t l2arc_feed_thr_lock; 880static kcondvar_t l2arc_feed_thr_cv; 881static uint8_t l2arc_thread_exit; 882 883static void arc_get_data_buf(arc_buf_t *); 884static void arc_access(arc_buf_hdr_t *, kmutex_t *); 885static int arc_evict_needed(arc_buf_contents_t); 886static void arc_evict_ghost(arc_state_t *, uint64_t, int64_t); 887static void arc_buf_watch(arc_buf_t *); 888 889static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 890static void l2arc_read_done(zio_t *); 891static void l2arc_hdr_stat_add(void); 892static void l2arc_hdr_stat_remove(void); 893 894static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *); 895static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress); 896static void l2arc_release_cdata_buf(arc_buf_hdr_t *); 897 898static uint64_t 899buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 900{ 901 uint8_t *vdva = (uint8_t *)dva; 902 uint64_t crc = -1ULL; 903 int i; 904 905 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 906 907 for (i = 0; i < sizeof (dva_t); i++) 908 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 909 910 crc ^= (spa>>8) ^ birth; 911 912 return (crc); 913} 914 915#define BUF_EMPTY(buf) \ 916 ((buf)->b_dva.dva_word[0] == 0 && \ 917 (buf)->b_dva.dva_word[1] == 0 && \ 918 (buf)->b_cksum0 == 0) 919 920#define BUF_EQUAL(spa, dva, birth, buf) \ 921 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 922 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 923 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 924 925static void 926buf_discard_identity(arc_buf_hdr_t *hdr) 927{ 928 hdr->b_dva.dva_word[0] = 0; 929 hdr->b_dva.dva_word[1] = 0; 930 hdr->b_birth = 0; 931 hdr->b_cksum0 = 0; 932} 933 934static arc_buf_hdr_t * 935buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 936{ 937 const dva_t *dva = BP_IDENTITY(bp); 938 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 939 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 940 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 941 arc_buf_hdr_t *hdr; 942 943 mutex_enter(hash_lock); 944 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 945 hdr = hdr->b_hash_next) { 946 if (BUF_EQUAL(spa, dva, birth, hdr)) { 947 *lockp = hash_lock; 948 return (hdr); 949 } 950 } 951 mutex_exit(hash_lock); 952 *lockp = NULL; 953 return (NULL); 954} 955 956/* 957 * Insert an entry into the hash table. If there is already an element 958 * equal to elem in the hash table, then the already existing element 959 * will be returned and the new element will not be inserted. 960 * Otherwise returns NULL. 961 */ 962static arc_buf_hdr_t * 963buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 964{ 965 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 966 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 967 arc_buf_hdr_t *fhdr; 968 uint32_t i; 969 970 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 971 ASSERT(hdr->b_birth != 0); 972 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 973 *lockp = hash_lock; 974 mutex_enter(hash_lock); 975 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 976 fhdr = fhdr->b_hash_next, i++) { 977 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 978 return (fhdr); 979 } 980 981 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 982 buf_hash_table.ht_table[idx] = hdr; 983 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 984 985 /* collect some hash table performance data */ 986 if (i > 0) { 987 ARCSTAT_BUMP(arcstat_hash_collisions); 988 if (i == 1) 989 ARCSTAT_BUMP(arcstat_hash_chains); 990 991 ARCSTAT_MAX(arcstat_hash_chain_max, i); 992 } 993 994 ARCSTAT_BUMP(arcstat_hash_elements); 995 ARCSTAT_MAXSTAT(arcstat_hash_elements); 996 997 return (NULL); 998} 999 1000static void 1001buf_hash_remove(arc_buf_hdr_t *hdr) 1002{ 1003 arc_buf_hdr_t *fhdr, **hdrp; 1004 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1005 1006 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1007 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1008 1009 hdrp = &buf_hash_table.ht_table[idx]; 1010 while ((fhdr = *hdrp) != hdr) { 1011 ASSERT(fhdr != NULL); 1012 hdrp = &fhdr->b_hash_next; 1013 } 1014 *hdrp = hdr->b_hash_next; 1015 hdr->b_hash_next = NULL; 1016 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; 1017 1018 /* collect some hash table performance data */ 1019 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1020 1021 if (buf_hash_table.ht_table[idx] && 1022 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1023 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1024} 1025 1026/* 1027 * Global data structures and functions for the buf kmem cache. 1028 */ 1029static kmem_cache_t *hdr_cache; 1030static kmem_cache_t *buf_cache; 1031 1032static void 1033buf_fini(void) 1034{ 1035 int i; 1036 1037 kmem_free(buf_hash_table.ht_table, 1038 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1039 for (i = 0; i < BUF_LOCKS; i++) 1040 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1041 kmem_cache_destroy(hdr_cache); 1042 kmem_cache_destroy(buf_cache); 1043} 1044 1045/* 1046 * Constructor callback - called when the cache is empty 1047 * and a new buf is requested. 1048 */ 1049/* ARGSUSED */ 1050static int 1051hdr_cons(void *vbuf, void *unused, int kmflag) 1052{ 1053 arc_buf_hdr_t *hdr = vbuf; 1054 1055 bzero(hdr, sizeof (arc_buf_hdr_t)); 1056 refcount_create(&hdr->b_refcnt); 1057 cv_init(&hdr->b_cv, NULL, CV_DEFAULT, NULL); 1058 mutex_init(&hdr->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1059 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1060 1061 return (0); 1062} 1063 1064/* ARGSUSED */ 1065static int 1066buf_cons(void *vbuf, void *unused, int kmflag) 1067{ 1068 arc_buf_t *buf = vbuf; 1069 1070 bzero(buf, sizeof (arc_buf_t)); 1071 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1072 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1073 1074 return (0); 1075} 1076 1077/* 1078 * Destructor callback - called when a cached buf is 1079 * no longer required. 1080 */ 1081/* ARGSUSED */ 1082static void 1083hdr_dest(void *vbuf, void *unused) 1084{ 1085 arc_buf_hdr_t *hdr = vbuf; 1086 1087 ASSERT(BUF_EMPTY(hdr)); 1088 refcount_destroy(&hdr->b_refcnt); 1089 cv_destroy(&hdr->b_cv); 1090 mutex_destroy(&hdr->b_freeze_lock); 1091 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1092} 1093 1094/* ARGSUSED */ 1095static void 1096buf_dest(void *vbuf, void *unused) 1097{ 1098 arc_buf_t *buf = vbuf; 1099 1100 mutex_destroy(&buf->b_evict_lock); 1101 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1102} 1103 1104/* 1105 * Reclaim callback -- invoked when memory is low. 1106 */ 1107/* ARGSUSED */ 1108static void 1109hdr_recl(void *unused) 1110{ 1111 dprintf("hdr_recl called\n"); 1112 /* 1113 * umem calls the reclaim func when we destroy the buf cache, 1114 * which is after we do arc_fini(). 1115 */ 1116 if (!arc_dead) 1117 cv_signal(&arc_reclaim_thr_cv); 1118} 1119 1120static void 1121buf_init(void) 1122{ 1123 uint64_t *ct; 1124 uint64_t hsize = 1ULL << 12; 1125 int i, j; 1126 1127 /* 1128 * The hash table is big enough to fill all of physical memory 1129 * with an average block size of zfs_arc_average_blocksize (default 8K). 1130 * By default, the table will take up 1131 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1132 */ 1133 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1134 hsize <<= 1; 1135retry: 1136 buf_hash_table.ht_mask = hsize - 1; 1137 buf_hash_table.ht_table = 1138 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1139 if (buf_hash_table.ht_table == NULL) { 1140 ASSERT(hsize > (1ULL << 8)); 1141 hsize >>= 1; 1142 goto retry; 1143 } 1144 1145 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1146 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1147 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1148 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1149 1150 for (i = 0; i < 256; i++) 1151 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1152 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1153 1154 for (i = 0; i < BUF_LOCKS; i++) { 1155 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1156 NULL, MUTEX_DEFAULT, NULL); 1157 } 1158} 1159 1160#define ARC_MINTIME (hz>>4) /* 62 ms */ 1161 1162static void 1163arc_cksum_verify(arc_buf_t *buf) 1164{ 1165 zio_cksum_t zc; 1166 1167 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1168 return; 1169 1170 mutex_enter(&buf->b_hdr->b_freeze_lock); 1171 if (buf->b_hdr->b_freeze_cksum == NULL || 1172 (buf->b_hdr->b_flags & ARC_FLAG_IO_ERROR)) { 1173 mutex_exit(&buf->b_hdr->b_freeze_lock); 1174 return; 1175 } 1176 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1177 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1178 panic("buffer modified while frozen!"); 1179 mutex_exit(&buf->b_hdr->b_freeze_lock); 1180} 1181 1182static int 1183arc_cksum_equal(arc_buf_t *buf) 1184{ 1185 zio_cksum_t zc; 1186 int equal; 1187 1188 mutex_enter(&buf->b_hdr->b_freeze_lock); 1189 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1190 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1191 mutex_exit(&buf->b_hdr->b_freeze_lock); 1192 1193 return (equal); 1194} 1195 1196static void 1197arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1198{ 1199 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1200 return; 1201 1202 mutex_enter(&buf->b_hdr->b_freeze_lock); 1203 if (buf->b_hdr->b_freeze_cksum != NULL) { 1204 mutex_exit(&buf->b_hdr->b_freeze_lock); 1205 return; 1206 } 1207 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1208 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1209 buf->b_hdr->b_freeze_cksum); 1210 mutex_exit(&buf->b_hdr->b_freeze_lock); 1211#ifdef illumos 1212 arc_buf_watch(buf); 1213#endif /* illumos */ 1214} 1215 1216#ifdef illumos 1217#ifndef _KERNEL 1218typedef struct procctl { 1219 long cmd; 1220 prwatch_t prwatch; 1221} procctl_t; 1222#endif 1223 1224/* ARGSUSED */ 1225static void 1226arc_buf_unwatch(arc_buf_t *buf) 1227{ 1228#ifndef _KERNEL 1229 if (arc_watch) { 1230 int result; 1231 procctl_t ctl; 1232 ctl.cmd = PCWATCH; 1233 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1234 ctl.prwatch.pr_size = 0; 1235 ctl.prwatch.pr_wflags = 0; 1236 result = write(arc_procfd, &ctl, sizeof (ctl)); 1237 ASSERT3U(result, ==, sizeof (ctl)); 1238 } 1239#endif 1240} 1241 1242/* ARGSUSED */ 1243static void 1244arc_buf_watch(arc_buf_t *buf) 1245{ 1246#ifndef _KERNEL 1247 if (arc_watch) { 1248 int result; 1249 procctl_t ctl; 1250 ctl.cmd = PCWATCH; 1251 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1252 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1253 ctl.prwatch.pr_wflags = WA_WRITE; 1254 result = write(arc_procfd, &ctl, sizeof (ctl)); 1255 ASSERT3U(result, ==, sizeof (ctl)); 1256 } 1257#endif 1258} 1259#endif /* illumos */ 1260 1261void 1262arc_buf_thaw(arc_buf_t *buf) 1263{ 1264 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1265 if (buf->b_hdr->b_state != arc_anon) 1266 panic("modifying non-anon buffer!"); 1267 if (buf->b_hdr->b_flags & ARC_FLAG_IO_IN_PROGRESS) 1268 panic("modifying buffer while i/o in progress!"); 1269 arc_cksum_verify(buf); 1270 } 1271 1272 mutex_enter(&buf->b_hdr->b_freeze_lock); 1273 if (buf->b_hdr->b_freeze_cksum != NULL) { 1274 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1275 buf->b_hdr->b_freeze_cksum = NULL; 1276 } 1277 1278 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1279 if (buf->b_hdr->b_thawed) 1280 kmem_free(buf->b_hdr->b_thawed, 1); 1281 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1282 } 1283 1284 mutex_exit(&buf->b_hdr->b_freeze_lock); 1285 1286#ifdef illumos 1287 arc_buf_unwatch(buf); 1288#endif /* illumos */ 1289} 1290 1291void 1292arc_buf_freeze(arc_buf_t *buf) 1293{ 1294 kmutex_t *hash_lock; 1295 1296 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1297 return; 1298 1299 hash_lock = HDR_LOCK(buf->b_hdr); 1300 mutex_enter(hash_lock); 1301 1302 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1303 buf->b_hdr->b_state == arc_anon); 1304 arc_cksum_compute(buf, B_FALSE); 1305 mutex_exit(hash_lock); 1306 1307} 1308 1309static void 1310get_buf_info(arc_buf_hdr_t *hdr, arc_state_t *state, list_t **list, kmutex_t **lock) 1311{ 1312 uint64_t buf_hashid = buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1313 1314 if (hdr->b_type == ARC_BUFC_METADATA) 1315 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1316 else { 1317 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1318 buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1319 } 1320 1321 *list = &state->arcs_lists[buf_hashid]; 1322 *lock = ARCS_LOCK(state, buf_hashid); 1323} 1324 1325 1326static void 1327add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1328{ 1329 ASSERT(MUTEX_HELD(hash_lock)); 1330 1331 if ((refcount_add(&hdr->b_refcnt, tag) == 1) && 1332 (hdr->b_state != arc_anon)) { 1333 uint64_t delta = hdr->b_size * hdr->b_datacnt; 1334 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 1335 list_t *list; 1336 kmutex_t *lock; 1337 1338 get_buf_info(hdr, hdr->b_state, &list, &lock); 1339 ASSERT(!MUTEX_HELD(lock)); 1340 mutex_enter(lock); 1341 ASSERT(list_link_active(&hdr->b_arc_node)); 1342 list_remove(list, hdr); 1343 if (GHOST_STATE(hdr->b_state)) { 1344 ASSERT0(hdr->b_datacnt); 1345 ASSERT3P(hdr->b_buf, ==, NULL); 1346 delta = hdr->b_size; 1347 } 1348 ASSERT(delta > 0); 1349 ASSERT3U(*size, >=, delta); 1350 atomic_add_64(size, -delta); 1351 mutex_exit(lock); 1352 /* remove the prefetch flag if we get a reference */ 1353 if (hdr->b_flags & ARC_FLAG_PREFETCH) 1354 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 1355 } 1356} 1357 1358static int 1359remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1360{ 1361 int cnt; 1362 arc_state_t *state = hdr->b_state; 1363 1364 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1365 ASSERT(!GHOST_STATE(state)); 1366 1367 if (((cnt = refcount_remove(&hdr->b_refcnt, tag)) == 0) && 1368 (state != arc_anon)) { 1369 uint64_t *size = &state->arcs_lsize[hdr->b_type]; 1370 list_t *list; 1371 kmutex_t *lock; 1372 1373 get_buf_info(hdr, state, &list, &lock); 1374 ASSERT(!MUTEX_HELD(lock)); 1375 mutex_enter(lock); 1376 ASSERT(!list_link_active(&hdr->b_arc_node)); 1377 list_insert_head(list, hdr); 1378 ASSERT(hdr->b_datacnt > 0); 1379 atomic_add_64(size, hdr->b_size * hdr->b_datacnt); 1380 mutex_exit(lock); 1381 } 1382 return (cnt); 1383} 1384 1385/* 1386 * Move the supplied buffer to the indicated state. The mutex 1387 * for the buffer must be held by the caller. 1388 */ 1389static void 1390arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 1391 kmutex_t *hash_lock) 1392{ 1393 arc_state_t *old_state = hdr->b_state; 1394 int64_t refcnt = refcount_count(&hdr->b_refcnt); 1395 uint64_t from_delta, to_delta; 1396 list_t *list; 1397 kmutex_t *lock; 1398 1399 ASSERT(MUTEX_HELD(hash_lock)); 1400 ASSERT3P(new_state, !=, old_state); 1401 ASSERT(refcnt == 0 || hdr->b_datacnt > 0); 1402 ASSERT(hdr->b_datacnt == 0 || !GHOST_STATE(new_state)); 1403 ASSERT(hdr->b_datacnt <= 1 || old_state != arc_anon); 1404 1405 from_delta = to_delta = hdr->b_datacnt * hdr->b_size; 1406 1407 /* 1408 * If this buffer is evictable, transfer it from the 1409 * old state list to the new state list. 1410 */ 1411 if (refcnt == 0) { 1412 if (old_state != arc_anon) { 1413 int use_mutex; 1414 uint64_t *size = &old_state->arcs_lsize[hdr->b_type]; 1415 1416 get_buf_info(hdr, old_state, &list, &lock); 1417 use_mutex = !MUTEX_HELD(lock); 1418 if (use_mutex) 1419 mutex_enter(lock); 1420 1421 ASSERT(list_link_active(&hdr->b_arc_node)); 1422 list_remove(list, hdr); 1423 1424 /* 1425 * If prefetching out of the ghost cache, 1426 * we will have a non-zero datacnt. 1427 */ 1428 if (GHOST_STATE(old_state) && hdr->b_datacnt == 0) { 1429 /* ghost elements have a ghost size */ 1430 ASSERT(hdr->b_buf == NULL); 1431 from_delta = hdr->b_size; 1432 } 1433 ASSERT3U(*size, >=, from_delta); 1434 atomic_add_64(size, -from_delta); 1435 1436 if (use_mutex) 1437 mutex_exit(lock); 1438 } 1439 if (new_state != arc_anon) { 1440 int use_mutex; 1441 uint64_t *size = &new_state->arcs_lsize[hdr->b_type]; 1442 1443 get_buf_info(hdr, new_state, &list, &lock); 1444 use_mutex = !MUTEX_HELD(lock); 1445 if (use_mutex) 1446 mutex_enter(lock); 1447 1448 list_insert_head(list, hdr); 1449 1450 /* ghost elements have a ghost size */ 1451 if (GHOST_STATE(new_state)) { 1452 ASSERT(hdr->b_datacnt == 0); 1453 ASSERT(hdr->b_buf == NULL); 1454 to_delta = hdr->b_size; 1455 } 1456 atomic_add_64(size, to_delta); 1457 1458 if (use_mutex) 1459 mutex_exit(lock); 1460 } 1461 } 1462 1463 ASSERT(!BUF_EMPTY(hdr)); 1464 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 1465 buf_hash_remove(hdr); 1466 1467 /* adjust state sizes */ 1468 if (to_delta) 1469 atomic_add_64(&new_state->arcs_size, to_delta); 1470 if (from_delta) { 1471 ASSERT3U(old_state->arcs_size, >=, from_delta); 1472 atomic_add_64(&old_state->arcs_size, -from_delta); 1473 } 1474 hdr->b_state = new_state; 1475 1476 /* adjust l2arc hdr stats */ 1477 if (new_state == arc_l2c_only) 1478 l2arc_hdr_stat_add(); 1479 else if (old_state == arc_l2c_only) 1480 l2arc_hdr_stat_remove(); 1481} 1482 1483void 1484arc_space_consume(uint64_t space, arc_space_type_t type) 1485{ 1486 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1487 1488 switch (type) { 1489 case ARC_SPACE_DATA: 1490 ARCSTAT_INCR(arcstat_data_size, space); 1491 break; 1492 case ARC_SPACE_OTHER: 1493 ARCSTAT_INCR(arcstat_other_size, space); 1494 break; 1495 case ARC_SPACE_HDRS: 1496 ARCSTAT_INCR(arcstat_hdr_size, space); 1497 break; 1498 case ARC_SPACE_L2HDRS: 1499 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1500 break; 1501 } 1502 1503 ARCSTAT_INCR(arcstat_meta_used, space); 1504 atomic_add_64(&arc_size, space); 1505} 1506 1507void 1508arc_space_return(uint64_t space, arc_space_type_t type) 1509{ 1510 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1511 1512 switch (type) { 1513 case ARC_SPACE_DATA: 1514 ARCSTAT_INCR(arcstat_data_size, -space); 1515 break; 1516 case ARC_SPACE_OTHER: 1517 ARCSTAT_INCR(arcstat_other_size, -space); 1518 break; 1519 case ARC_SPACE_HDRS: 1520 ARCSTAT_INCR(arcstat_hdr_size, -space); 1521 break; 1522 case ARC_SPACE_L2HDRS: 1523 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1524 break; 1525 } 1526 1527 ASSERT(arc_meta_used >= space); 1528 if (arc_meta_max < arc_meta_used) 1529 arc_meta_max = arc_meta_used; 1530 ARCSTAT_INCR(arcstat_meta_used, -space); 1531 ASSERT(arc_size >= space); 1532 atomic_add_64(&arc_size, -space); 1533} 1534 1535arc_buf_t * 1536arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1537{ 1538 arc_buf_hdr_t *hdr; 1539 arc_buf_t *buf; 1540 1541 ASSERT3U(size, >, 0); 1542 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1543 ASSERT(BUF_EMPTY(hdr)); 1544 hdr->b_size = size; 1545 hdr->b_type = type; 1546 hdr->b_spa = spa_load_guid(spa); 1547 hdr->b_state = arc_anon; 1548 hdr->b_arc_access = 0; 1549 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1550 buf->b_hdr = hdr; 1551 buf->b_data = NULL; 1552 buf->b_efunc = NULL; 1553 buf->b_private = NULL; 1554 buf->b_next = NULL; 1555 hdr->b_buf = buf; 1556 arc_get_data_buf(buf); 1557 hdr->b_datacnt = 1; 1558 hdr->b_flags = 0; 1559 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1560 (void) refcount_add(&hdr->b_refcnt, tag); 1561 1562 return (buf); 1563} 1564 1565static char *arc_onloan_tag = "onloan"; 1566 1567/* 1568 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1569 * flight data by arc_tempreserve_space() until they are "returned". Loaned 1570 * buffers must be returned to the arc before they can be used by the DMU or 1571 * freed. 1572 */ 1573arc_buf_t * 1574arc_loan_buf(spa_t *spa, int size) 1575{ 1576 arc_buf_t *buf; 1577 1578 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1579 1580 atomic_add_64(&arc_loaned_bytes, size); 1581 return (buf); 1582} 1583 1584/* 1585 * Return a loaned arc buffer to the arc. 1586 */ 1587void 1588arc_return_buf(arc_buf_t *buf, void *tag) 1589{ 1590 arc_buf_hdr_t *hdr = buf->b_hdr; 1591 1592 ASSERT(buf->b_data != NULL); 1593 (void) refcount_add(&hdr->b_refcnt, tag); 1594 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1595 1596 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1597} 1598 1599/* Detach an arc_buf from a dbuf (tag) */ 1600void 1601arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1602{ 1603 arc_buf_hdr_t *hdr; 1604 1605 ASSERT(buf->b_data != NULL); 1606 hdr = buf->b_hdr; 1607 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1608 (void) refcount_remove(&hdr->b_refcnt, tag); 1609 buf->b_efunc = NULL; 1610 buf->b_private = NULL; 1611 1612 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1613} 1614 1615static arc_buf_t * 1616arc_buf_clone(arc_buf_t *from) 1617{ 1618 arc_buf_t *buf; 1619 arc_buf_hdr_t *hdr = from->b_hdr; 1620 uint64_t size = hdr->b_size; 1621 1622 ASSERT(hdr->b_state != arc_anon); 1623 1624 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1625 buf->b_hdr = hdr; 1626 buf->b_data = NULL; 1627 buf->b_efunc = NULL; 1628 buf->b_private = NULL; 1629 buf->b_next = hdr->b_buf; 1630 hdr->b_buf = buf; 1631 arc_get_data_buf(buf); 1632 bcopy(from->b_data, buf->b_data, size); 1633 1634 /* 1635 * This buffer already exists in the arc so create a duplicate 1636 * copy for the caller. If the buffer is associated with user data 1637 * then track the size and number of duplicates. These stats will be 1638 * updated as duplicate buffers are created and destroyed. 1639 */ 1640 if (hdr->b_type == ARC_BUFC_DATA) { 1641 ARCSTAT_BUMP(arcstat_duplicate_buffers); 1642 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1643 } 1644 hdr->b_datacnt += 1; 1645 return (buf); 1646} 1647 1648void 1649arc_buf_add_ref(arc_buf_t *buf, void* tag) 1650{ 1651 arc_buf_hdr_t *hdr; 1652 kmutex_t *hash_lock; 1653 1654 /* 1655 * Check to see if this buffer is evicted. Callers 1656 * must verify b_data != NULL to know if the add_ref 1657 * was successful. 1658 */ 1659 mutex_enter(&buf->b_evict_lock); 1660 if (buf->b_data == NULL) { 1661 mutex_exit(&buf->b_evict_lock); 1662 return; 1663 } 1664 hash_lock = HDR_LOCK(buf->b_hdr); 1665 mutex_enter(hash_lock); 1666 hdr = buf->b_hdr; 1667 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1668 mutex_exit(&buf->b_evict_lock); 1669 1670 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1671 add_reference(hdr, hash_lock, tag); 1672 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1673 arc_access(hdr, hash_lock); 1674 mutex_exit(hash_lock); 1675 ARCSTAT_BUMP(arcstat_hits); 1676 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 1677 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1678 data, metadata, hits); 1679} 1680 1681static void 1682arc_buf_free_on_write(void *data, size_t size, 1683 void (*free_func)(void *, size_t)) 1684{ 1685 l2arc_data_free_t *df; 1686 1687 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1688 df->l2df_data = data; 1689 df->l2df_size = size; 1690 df->l2df_func = free_func; 1691 mutex_enter(&l2arc_free_on_write_mtx); 1692 list_insert_head(l2arc_free_on_write, df); 1693 mutex_exit(&l2arc_free_on_write_mtx); 1694} 1695 1696/* 1697 * Free the arc data buffer. If it is an l2arc write in progress, 1698 * the buffer is placed on l2arc_free_on_write to be freed later. 1699 */ 1700static void 1701arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1702{ 1703 arc_buf_hdr_t *hdr = buf->b_hdr; 1704 1705 if (HDR_L2_WRITING(hdr)) { 1706 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 1707 ARCSTAT_BUMP(arcstat_l2_free_on_write); 1708 } else { 1709 free_func(buf->b_data, hdr->b_size); 1710 } 1711} 1712 1713/* 1714 * Free up buf->b_data and if 'remove' is set, then pull the 1715 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 1716 */ 1717static void 1718arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 1719{ 1720 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1721 1722 ASSERT(MUTEX_HELD(&l2arc_buflist_mtx)); 1723 1724 if (l2hdr->b_tmp_cdata == NULL) 1725 return; 1726 1727 ASSERT(HDR_L2_WRITING(hdr)); 1728 arc_buf_free_on_write(l2hdr->b_tmp_cdata, hdr->b_size, 1729 zio_data_buf_free); 1730 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 1731 l2hdr->b_tmp_cdata = NULL; 1732} 1733 1734static void 1735arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 1736{ 1737 arc_buf_t **bufp; 1738 1739 /* free up data associated with the buf */ 1740 if (buf->b_data) { 1741 arc_state_t *state = buf->b_hdr->b_state; 1742 uint64_t size = buf->b_hdr->b_size; 1743 arc_buf_contents_t type = buf->b_hdr->b_type; 1744 1745 arc_cksum_verify(buf); 1746#ifdef illumos 1747 arc_buf_unwatch(buf); 1748#endif /* illumos */ 1749 1750 if (!recycle) { 1751 if (type == ARC_BUFC_METADATA) { 1752 arc_buf_data_free(buf, zio_buf_free); 1753 arc_space_return(size, ARC_SPACE_DATA); 1754 } else { 1755 ASSERT(type == ARC_BUFC_DATA); 1756 arc_buf_data_free(buf, zio_data_buf_free); 1757 ARCSTAT_INCR(arcstat_data_size, -size); 1758 atomic_add_64(&arc_size, -size); 1759 } 1760 } 1761 if (list_link_active(&buf->b_hdr->b_arc_node)) { 1762 uint64_t *cnt = &state->arcs_lsize[type]; 1763 1764 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1765 ASSERT(state != arc_anon); 1766 1767 ASSERT3U(*cnt, >=, size); 1768 atomic_add_64(cnt, -size); 1769 } 1770 ASSERT3U(state->arcs_size, >=, size); 1771 atomic_add_64(&state->arcs_size, -size); 1772 buf->b_data = NULL; 1773 1774 /* 1775 * If we're destroying a duplicate buffer make sure 1776 * that the appropriate statistics are updated. 1777 */ 1778 if (buf->b_hdr->b_datacnt > 1 && 1779 buf->b_hdr->b_type == ARC_BUFC_DATA) { 1780 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1781 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1782 } 1783 ASSERT(buf->b_hdr->b_datacnt > 0); 1784 buf->b_hdr->b_datacnt -= 1; 1785 } 1786 1787 /* only remove the buf if requested */ 1788 if (!remove) 1789 return; 1790 1791 /* remove the buf from the hdr list */ 1792 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1793 continue; 1794 *bufp = buf->b_next; 1795 buf->b_next = NULL; 1796 1797 ASSERT(buf->b_efunc == NULL); 1798 1799 /* clean up the buf */ 1800 buf->b_hdr = NULL; 1801 kmem_cache_free(buf_cache, buf); 1802} 1803 1804static void 1805arc_hdr_destroy(arc_buf_hdr_t *hdr) 1806{ 1807 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1808 ASSERT3P(hdr->b_state, ==, arc_anon); 1809 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1810 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1811 1812 if (l2hdr != NULL) { 1813 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1814 /* 1815 * To prevent arc_free() and l2arc_evict() from 1816 * attempting to free the same buffer at the same time, 1817 * a FREE_IN_PROGRESS flag is given to arc_free() to 1818 * give it priority. l2arc_evict() can't destroy this 1819 * header while we are waiting on l2arc_buflist_mtx. 1820 * 1821 * The hdr may be removed from l2ad_buflist before we 1822 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1823 */ 1824 if (!buflist_held) { 1825 mutex_enter(&l2arc_buflist_mtx); 1826 l2hdr = hdr->b_l2hdr; 1827 } 1828 1829 if (l2hdr != NULL) { 1830 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1831 l2hdr->b_asize, 0); 1832 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1833 arc_buf_l2_cdata_free(hdr); 1834 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1835 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1836 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 1837 -l2hdr->b_asize, 0, 0); 1838 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1839 if (hdr->b_state == arc_l2c_only) 1840 l2arc_hdr_stat_remove(); 1841 hdr->b_l2hdr = NULL; 1842 } 1843 1844 if (!buflist_held) 1845 mutex_exit(&l2arc_buflist_mtx); 1846 } 1847 1848 if (!BUF_EMPTY(hdr)) { 1849 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1850 buf_discard_identity(hdr); 1851 } 1852 while (hdr->b_buf) { 1853 arc_buf_t *buf = hdr->b_buf; 1854 1855 if (buf->b_efunc) { 1856 mutex_enter(&arc_eviction_mtx); 1857 mutex_enter(&buf->b_evict_lock); 1858 ASSERT(buf->b_hdr != NULL); 1859 arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1860 hdr->b_buf = buf->b_next; 1861 buf->b_hdr = &arc_eviction_hdr; 1862 buf->b_next = arc_eviction_list; 1863 arc_eviction_list = buf; 1864 mutex_exit(&buf->b_evict_lock); 1865 mutex_exit(&arc_eviction_mtx); 1866 } else { 1867 arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1868 } 1869 } 1870 if (hdr->b_freeze_cksum != NULL) { 1871 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1872 hdr->b_freeze_cksum = NULL; 1873 } 1874 if (hdr->b_thawed) { 1875 kmem_free(hdr->b_thawed, 1); 1876 hdr->b_thawed = NULL; 1877 } 1878 1879 ASSERT(!list_link_active(&hdr->b_arc_node)); 1880 ASSERT3P(hdr->b_hash_next, ==, NULL); 1881 ASSERT3P(hdr->b_acb, ==, NULL); 1882 kmem_cache_free(hdr_cache, hdr); 1883} 1884 1885void 1886arc_buf_free(arc_buf_t *buf, void *tag) 1887{ 1888 arc_buf_hdr_t *hdr = buf->b_hdr; 1889 int hashed = hdr->b_state != arc_anon; 1890 1891 ASSERT(buf->b_efunc == NULL); 1892 ASSERT(buf->b_data != NULL); 1893 1894 if (hashed) { 1895 kmutex_t *hash_lock = HDR_LOCK(hdr); 1896 1897 mutex_enter(hash_lock); 1898 hdr = buf->b_hdr; 1899 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1900 1901 (void) remove_reference(hdr, hash_lock, tag); 1902 if (hdr->b_datacnt > 1) { 1903 arc_buf_destroy(buf, FALSE, TRUE); 1904 } else { 1905 ASSERT(buf == hdr->b_buf); 1906 ASSERT(buf->b_efunc == NULL); 1907 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 1908 } 1909 mutex_exit(hash_lock); 1910 } else if (HDR_IO_IN_PROGRESS(hdr)) { 1911 int destroy_hdr; 1912 /* 1913 * We are in the middle of an async write. Don't destroy 1914 * this buffer unless the write completes before we finish 1915 * decrementing the reference count. 1916 */ 1917 mutex_enter(&arc_eviction_mtx); 1918 (void) remove_reference(hdr, NULL, tag); 1919 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1920 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1921 mutex_exit(&arc_eviction_mtx); 1922 if (destroy_hdr) 1923 arc_hdr_destroy(hdr); 1924 } else { 1925 if (remove_reference(hdr, NULL, tag) > 0) 1926 arc_buf_destroy(buf, FALSE, TRUE); 1927 else 1928 arc_hdr_destroy(hdr); 1929 } 1930} 1931 1932boolean_t 1933arc_buf_remove_ref(arc_buf_t *buf, void* tag) 1934{ 1935 arc_buf_hdr_t *hdr = buf->b_hdr; 1936 kmutex_t *hash_lock = HDR_LOCK(hdr); 1937 boolean_t no_callback = (buf->b_efunc == NULL); 1938 1939 if (hdr->b_state == arc_anon) { 1940 ASSERT(hdr->b_datacnt == 1); 1941 arc_buf_free(buf, tag); 1942 return (no_callback); 1943 } 1944 1945 mutex_enter(hash_lock); 1946 hdr = buf->b_hdr; 1947 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1948 ASSERT(hdr->b_state != arc_anon); 1949 ASSERT(buf->b_data != NULL); 1950 1951 (void) remove_reference(hdr, hash_lock, tag); 1952 if (hdr->b_datacnt > 1) { 1953 if (no_callback) 1954 arc_buf_destroy(buf, FALSE, TRUE); 1955 } else if (no_callback) { 1956 ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1957 ASSERT(buf->b_efunc == NULL); 1958 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 1959 } 1960 ASSERT(no_callback || hdr->b_datacnt > 1 || 1961 refcount_is_zero(&hdr->b_refcnt)); 1962 mutex_exit(hash_lock); 1963 return (no_callback); 1964} 1965 1966int 1967arc_buf_size(arc_buf_t *buf) 1968{ 1969 return (buf->b_hdr->b_size); 1970} 1971 1972/* 1973 * Called from the DMU to determine if the current buffer should be 1974 * evicted. In order to ensure proper locking, the eviction must be initiated 1975 * from the DMU. Return true if the buffer is associated with user data and 1976 * duplicate buffers still exist. 1977 */ 1978boolean_t 1979arc_buf_eviction_needed(arc_buf_t *buf) 1980{ 1981 arc_buf_hdr_t *hdr; 1982 boolean_t evict_needed = B_FALSE; 1983 1984 if (zfs_disable_dup_eviction) 1985 return (B_FALSE); 1986 1987 mutex_enter(&buf->b_evict_lock); 1988 hdr = buf->b_hdr; 1989 if (hdr == NULL) { 1990 /* 1991 * We are in arc_do_user_evicts(); let that function 1992 * perform the eviction. 1993 */ 1994 ASSERT(buf->b_data == NULL); 1995 mutex_exit(&buf->b_evict_lock); 1996 return (B_FALSE); 1997 } else if (buf->b_data == NULL) { 1998 /* 1999 * We have already been added to the arc eviction list; 2000 * recommend eviction. 2001 */ 2002 ASSERT3P(hdr, ==, &arc_eviction_hdr); 2003 mutex_exit(&buf->b_evict_lock); 2004 return (B_TRUE); 2005 } 2006 2007 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 2008 evict_needed = B_TRUE; 2009 2010 mutex_exit(&buf->b_evict_lock); 2011 return (evict_needed); 2012} 2013 2014/* 2015 * Evict buffers from list until we've removed the specified number of 2016 * bytes. Move the removed buffers to the appropriate evict state. 2017 * If the recycle flag is set, then attempt to "recycle" a buffer: 2018 * - look for a buffer to evict that is `bytes' long. 2019 * - return the data block from this buffer rather than freeing it. 2020 * This flag is used by callers that are trying to make space for a 2021 * new buffer in a full arc cache. 2022 * 2023 * This function makes a "best effort". It skips over any buffers 2024 * it can't get a hash_lock on, and so may not catch all candidates. 2025 * It may also return without evicting as much space as requested. 2026 */ 2027static void * 2028arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 2029 arc_buf_contents_t type) 2030{ 2031 arc_state_t *evicted_state; 2032 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 2033 int64_t bytes_remaining; 2034 arc_buf_hdr_t *hdr, *hdr_prev = NULL; 2035 list_t *evicted_list, *list, *evicted_list_start, *list_start; 2036 kmutex_t *lock, *evicted_lock; 2037 kmutex_t *hash_lock; 2038 boolean_t have_lock; 2039 void *stolen = NULL; 2040 arc_buf_hdr_t marker = { 0 }; 2041 int count = 0; 2042 static int evict_metadata_offset, evict_data_offset; 2043 int i, idx, offset, list_count, lists; 2044 2045 ASSERT(state == arc_mru || state == arc_mfu); 2046 2047 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2048 2049 /* 2050 * Decide which "type" (data vs metadata) to recycle from. 2051 * 2052 * If we are over the metadata limit, recycle from metadata. 2053 * If we are under the metadata minimum, recycle from data. 2054 * Otherwise, recycle from whichever type has the oldest (least 2055 * recently accessed) header. This is not yet implemented. 2056 */ 2057 if (recycle) { 2058 arc_buf_contents_t realtype; 2059 if (state->arcs_lsize[ARC_BUFC_DATA] == 0) { 2060 realtype = ARC_BUFC_METADATA; 2061 } else if (state->arcs_lsize[ARC_BUFC_METADATA] == 0) { 2062 realtype = ARC_BUFC_DATA; 2063 } else if (arc_meta_used >= arc_meta_limit) { 2064 realtype = ARC_BUFC_METADATA; 2065 } else if (arc_meta_used <= arc_meta_min) { 2066 realtype = ARC_BUFC_DATA; 2067 } else { 2068#ifdef illumos 2069 if (data_hdr->b_arc_access < 2070 metadata_hdr->b_arc_access) { 2071 realtype = ARC_BUFC_DATA; 2072 } else { 2073 realtype = ARC_BUFC_METADATA; 2074 } 2075#else 2076 /* TODO */ 2077 realtype = type; 2078#endif 2079 } 2080 if (realtype != type) { 2081 /* 2082 * If we want to evict from a different list, 2083 * we can not recycle, because DATA vs METADATA 2084 * buffers are segregated into different kmem 2085 * caches (and vmem arenas). 2086 */ 2087 type = realtype; 2088 recycle = B_FALSE; 2089 } 2090 } 2091 2092 if (type == ARC_BUFC_METADATA) { 2093 offset = 0; 2094 list_count = ARC_BUFC_NUMMETADATALISTS; 2095 list_start = &state->arcs_lists[0]; 2096 evicted_list_start = &evicted_state->arcs_lists[0]; 2097 idx = evict_metadata_offset; 2098 } else { 2099 offset = ARC_BUFC_NUMMETADATALISTS; 2100 list_start = &state->arcs_lists[offset]; 2101 evicted_list_start = &evicted_state->arcs_lists[offset]; 2102 list_count = ARC_BUFC_NUMDATALISTS; 2103 idx = evict_data_offset; 2104 } 2105 bytes_remaining = evicted_state->arcs_lsize[type]; 2106 lists = 0; 2107 2108evict_start: 2109 list = &list_start[idx]; 2110 evicted_list = &evicted_list_start[idx]; 2111 lock = ARCS_LOCK(state, (offset + idx)); 2112 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 2113 2114 mutex_enter(lock); 2115 mutex_enter(evicted_lock); 2116 2117 for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2118 hdr_prev = list_prev(list, hdr); 2119 bytes_remaining -= (hdr->b_size * hdr->b_datacnt); 2120 /* prefetch buffers have a minimum lifespan */ 2121 if (HDR_IO_IN_PROGRESS(hdr) || 2122 (spa && hdr->b_spa != spa) || 2123 (hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT) && 2124 ddi_get_lbolt() - hdr->b_arc_access < 2125 arc_min_prefetch_lifespan)) { 2126 skipped++; 2127 continue; 2128 } 2129 /* "lookahead" for better eviction candidate */ 2130 if (recycle && hdr->b_size != bytes && 2131 hdr_prev && hdr_prev->b_size == bytes) 2132 continue; 2133 2134 /* ignore markers */ 2135 if (hdr->b_spa == 0) 2136 continue; 2137 2138 /* 2139 * It may take a long time to evict all the bufs requested. 2140 * To avoid blocking all arc activity, periodically drop 2141 * the arcs_mtx and give other threads a chance to run 2142 * before reacquiring the lock. 2143 * 2144 * If we are looking for a buffer to recycle, we are in 2145 * the hot code path, so don't sleep. 2146 */ 2147 if (!recycle && count++ > arc_evict_iterations) { 2148 list_insert_after(list, hdr, &marker); 2149 mutex_exit(evicted_lock); 2150 mutex_exit(lock); 2151 kpreempt(KPREEMPT_SYNC); 2152 mutex_enter(lock); 2153 mutex_enter(evicted_lock); 2154 hdr_prev = list_prev(list, &marker); 2155 list_remove(list, &marker); 2156 count = 0; 2157 continue; 2158 } 2159 2160 hash_lock = HDR_LOCK(hdr); 2161 have_lock = MUTEX_HELD(hash_lock); 2162 if (have_lock || mutex_tryenter(hash_lock)) { 2163 ASSERT0(refcount_count(&hdr->b_refcnt)); 2164 ASSERT(hdr->b_datacnt > 0); 2165 while (hdr->b_buf) { 2166 arc_buf_t *buf = hdr->b_buf; 2167 if (!mutex_tryenter(&buf->b_evict_lock)) { 2168 missed += 1; 2169 break; 2170 } 2171 if (buf->b_data) { 2172 bytes_evicted += hdr->b_size; 2173 if (recycle && hdr->b_type == type && 2174 hdr->b_size == bytes && 2175 !HDR_L2_WRITING(hdr)) { 2176 stolen = buf->b_data; 2177 recycle = FALSE; 2178 } 2179 } 2180 if (buf->b_efunc) { 2181 mutex_enter(&arc_eviction_mtx); 2182 arc_buf_destroy(buf, 2183 buf->b_data == stolen, FALSE); 2184 hdr->b_buf = buf->b_next; 2185 buf->b_hdr = &arc_eviction_hdr; 2186 buf->b_next = arc_eviction_list; 2187 arc_eviction_list = buf; 2188 mutex_exit(&arc_eviction_mtx); 2189 mutex_exit(&buf->b_evict_lock); 2190 } else { 2191 mutex_exit(&buf->b_evict_lock); 2192 arc_buf_destroy(buf, 2193 buf->b_data == stolen, TRUE); 2194 } 2195 } 2196 2197 if (hdr->b_l2hdr) { 2198 ARCSTAT_INCR(arcstat_evict_l2_cached, 2199 hdr->b_size); 2200 } else { 2201 if (l2arc_write_eligible(hdr->b_spa, hdr)) { 2202 ARCSTAT_INCR(arcstat_evict_l2_eligible, 2203 hdr->b_size); 2204 } else { 2205 ARCSTAT_INCR( 2206 arcstat_evict_l2_ineligible, 2207 hdr->b_size); 2208 } 2209 } 2210 2211 if (hdr->b_datacnt == 0) { 2212 arc_change_state(evicted_state, hdr, hash_lock); 2213 ASSERT(HDR_IN_HASH_TABLE(hdr)); 2214 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 2215 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 2216 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 2217 } 2218 if (!have_lock) 2219 mutex_exit(hash_lock); 2220 if (bytes >= 0 && bytes_evicted >= bytes) 2221 break; 2222 if (bytes_remaining > 0) { 2223 mutex_exit(evicted_lock); 2224 mutex_exit(lock); 2225 idx = ((idx + 1) & (list_count - 1)); 2226 lists++; 2227 goto evict_start; 2228 } 2229 } else { 2230 missed += 1; 2231 } 2232 } 2233 2234 mutex_exit(evicted_lock); 2235 mutex_exit(lock); 2236 2237 idx = ((idx + 1) & (list_count - 1)); 2238 lists++; 2239 2240 if (bytes_evicted < bytes) { 2241 if (lists < list_count) 2242 goto evict_start; 2243 else 2244 dprintf("only evicted %lld bytes from %x", 2245 (longlong_t)bytes_evicted, state); 2246 } 2247 if (type == ARC_BUFC_METADATA) 2248 evict_metadata_offset = idx; 2249 else 2250 evict_data_offset = idx; 2251 2252 if (skipped) 2253 ARCSTAT_INCR(arcstat_evict_skip, skipped); 2254 2255 if (missed) 2256 ARCSTAT_INCR(arcstat_mutex_miss, missed); 2257 2258 /* 2259 * Note: we have just evicted some data into the ghost state, 2260 * potentially putting the ghost size over the desired size. Rather 2261 * that evicting from the ghost list in this hot code path, leave 2262 * this chore to the arc_reclaim_thread(). 2263 */ 2264 2265 if (stolen) 2266 ARCSTAT_BUMP(arcstat_stolen); 2267 return (stolen); 2268} 2269 2270/* 2271 * Remove buffers from list until we've removed the specified number of 2272 * bytes. Destroy the buffers that are removed. 2273 */ 2274static void 2275arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2276{ 2277 arc_buf_hdr_t *hdr, *hdr_prev; 2278 arc_buf_hdr_t marker = { 0 }; 2279 list_t *list, *list_start; 2280 kmutex_t *hash_lock, *lock; 2281 uint64_t bytes_deleted = 0; 2282 uint64_t bufs_skipped = 0; 2283 int count = 0; 2284 static int evict_offset; 2285 int list_count, idx = evict_offset; 2286 int offset, lists = 0; 2287 2288 ASSERT(GHOST_STATE(state)); 2289 2290 /* 2291 * data lists come after metadata lists 2292 */ 2293 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2294 list_count = ARC_BUFC_NUMDATALISTS; 2295 offset = ARC_BUFC_NUMMETADATALISTS; 2296 2297evict_start: 2298 list = &list_start[idx]; 2299 lock = ARCS_LOCK(state, idx + offset); 2300 2301 mutex_enter(lock); 2302 for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2303 hdr_prev = list_prev(list, hdr); 2304 if (hdr->b_type > ARC_BUFC_NUMTYPES) 2305 panic("invalid hdr=%p", (void *)hdr); 2306 if (spa && hdr->b_spa != spa) 2307 continue; 2308 2309 /* ignore markers */ 2310 if (hdr->b_spa == 0) 2311 continue; 2312 2313 hash_lock = HDR_LOCK(hdr); 2314 /* caller may be trying to modify this buffer, skip it */ 2315 if (MUTEX_HELD(hash_lock)) 2316 continue; 2317 2318 /* 2319 * It may take a long time to evict all the bufs requested. 2320 * To avoid blocking all arc activity, periodically drop 2321 * the arcs_mtx and give other threads a chance to run 2322 * before reacquiring the lock. 2323 */ 2324 if (count++ > arc_evict_iterations) { 2325 list_insert_after(list, hdr, &marker); 2326 mutex_exit(lock); 2327 kpreempt(KPREEMPT_SYNC); 2328 mutex_enter(lock); 2329 hdr_prev = list_prev(list, &marker); 2330 list_remove(list, &marker); 2331 count = 0; 2332 continue; 2333 } 2334 if (mutex_tryenter(hash_lock)) { 2335 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2336 ASSERT(hdr->b_buf == NULL); 2337 ARCSTAT_BUMP(arcstat_deleted); 2338 bytes_deleted += hdr->b_size; 2339 2340 if (hdr->b_l2hdr != NULL) { 2341 /* 2342 * This buffer is cached on the 2nd Level ARC; 2343 * don't destroy the header. 2344 */ 2345 arc_change_state(arc_l2c_only, hdr, hash_lock); 2346 mutex_exit(hash_lock); 2347 } else { 2348 arc_change_state(arc_anon, hdr, hash_lock); 2349 mutex_exit(hash_lock); 2350 arc_hdr_destroy(hdr); 2351 } 2352 2353 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 2354 if (bytes >= 0 && bytes_deleted >= bytes) 2355 break; 2356 } else if (bytes < 0) { 2357 /* 2358 * Insert a list marker and then wait for the 2359 * hash lock to become available. Once its 2360 * available, restart from where we left off. 2361 */ 2362 list_insert_after(list, hdr, &marker); 2363 mutex_exit(lock); 2364 mutex_enter(hash_lock); 2365 mutex_exit(hash_lock); 2366 mutex_enter(lock); 2367 hdr_prev = list_prev(list, &marker); 2368 list_remove(list, &marker); 2369 } else { 2370 bufs_skipped += 1; 2371 } 2372 2373 } 2374 mutex_exit(lock); 2375 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2376 lists++; 2377 2378 if (lists < list_count) 2379 goto evict_start; 2380 2381 evict_offset = idx; 2382 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2383 (bytes < 0 || bytes_deleted < bytes)) { 2384 list_start = &state->arcs_lists[0]; 2385 list_count = ARC_BUFC_NUMMETADATALISTS; 2386 offset = lists = 0; 2387 goto evict_start; 2388 } 2389 2390 if (bufs_skipped) { 2391 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2392 ASSERT(bytes >= 0); 2393 } 2394 2395 if (bytes_deleted < bytes) 2396 dprintf("only deleted %lld bytes from %p", 2397 (longlong_t)bytes_deleted, state); 2398} 2399 2400static void 2401arc_adjust(void) 2402{ 2403 int64_t adjustment, delta; 2404 2405 /* 2406 * Adjust MRU size 2407 */ 2408 2409 adjustment = MIN((int64_t)(arc_size - arc_c), 2410 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2411 arc_p)); 2412 2413 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2414 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2415 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2416 adjustment -= delta; 2417 } 2418 2419 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2420 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2421 (void) arc_evict(arc_mru, 0, delta, FALSE, 2422 ARC_BUFC_METADATA); 2423 } 2424 2425 /* 2426 * Adjust MFU size 2427 */ 2428 2429 adjustment = arc_size - arc_c; 2430 2431 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2432 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2433 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2434 adjustment -= delta; 2435 } 2436 2437 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2438 int64_t delta = MIN(adjustment, 2439 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2440 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2441 ARC_BUFC_METADATA); 2442 } 2443 2444 /* 2445 * Adjust ghost lists 2446 */ 2447 2448 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2449 2450 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2451 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2452 arc_evict_ghost(arc_mru_ghost, 0, delta); 2453 } 2454 2455 adjustment = 2456 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2457 2458 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2459 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2460 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2461 } 2462} 2463 2464static void 2465arc_do_user_evicts(void) 2466{ 2467 static arc_buf_t *tmp_arc_eviction_list; 2468 2469 /* 2470 * Move list over to avoid LOR 2471 */ 2472restart: 2473 mutex_enter(&arc_eviction_mtx); 2474 tmp_arc_eviction_list = arc_eviction_list; 2475 arc_eviction_list = NULL; 2476 mutex_exit(&arc_eviction_mtx); 2477 2478 while (tmp_arc_eviction_list != NULL) { 2479 arc_buf_t *buf = tmp_arc_eviction_list; 2480 tmp_arc_eviction_list = buf->b_next; 2481 mutex_enter(&buf->b_evict_lock); 2482 buf->b_hdr = NULL; 2483 mutex_exit(&buf->b_evict_lock); 2484 2485 if (buf->b_efunc != NULL) 2486 VERIFY0(buf->b_efunc(buf->b_private)); 2487 2488 buf->b_efunc = NULL; 2489 buf->b_private = NULL; 2490 kmem_cache_free(buf_cache, buf); 2491 } 2492 2493 if (arc_eviction_list != NULL) 2494 goto restart; 2495} 2496 2497/* 2498 * Flush all *evictable* data from the cache for the given spa. 2499 * NOTE: this will not touch "active" (i.e. referenced) data. 2500 */ 2501void 2502arc_flush(spa_t *spa) 2503{ 2504 uint64_t guid = 0; 2505 2506 if (spa) 2507 guid = spa_load_guid(spa); 2508 2509 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2510 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2511 if (spa) 2512 break; 2513 } 2514 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2515 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2516 if (spa) 2517 break; 2518 } 2519 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2520 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2521 if (spa) 2522 break; 2523 } 2524 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2525 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2526 if (spa) 2527 break; 2528 } 2529 2530 arc_evict_ghost(arc_mru_ghost, guid, -1); 2531 arc_evict_ghost(arc_mfu_ghost, guid, -1); 2532 2533 mutex_enter(&arc_reclaim_thr_lock); 2534 arc_do_user_evicts(); 2535 mutex_exit(&arc_reclaim_thr_lock); 2536 ASSERT(spa || arc_eviction_list == NULL); 2537} 2538 2539void 2540arc_shrink(void) 2541{ 2542 2543 if (arc_c > arc_c_min) { 2544 uint64_t to_free; 2545 2546 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 2547 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 2548#ifdef _KERNEL 2549 to_free = arc_c >> arc_shrink_shift; 2550#else 2551 to_free = arc_c >> arc_shrink_shift; 2552#endif 2553 if (arc_c > arc_c_min + to_free) 2554 atomic_add_64(&arc_c, -to_free); 2555 else 2556 arc_c = arc_c_min; 2557 2558 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2559 if (arc_c > arc_size) 2560 arc_c = MAX(arc_size, arc_c_min); 2561 if (arc_p > arc_c) 2562 arc_p = (arc_c >> 1); 2563 2564 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 2565 arc_p); 2566 2567 ASSERT(arc_c >= arc_c_min); 2568 ASSERT((int64_t)arc_p >= 0); 2569 } 2570 2571 if (arc_size > arc_c) { 2572 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 2573 uint64_t, arc_c); 2574 arc_adjust(); 2575 } 2576} 2577 2578static int needfree = 0; 2579 2580static int 2581arc_reclaim_needed(void) 2582{ 2583 2584#ifdef _KERNEL 2585 2586 if (needfree) { 2587 DTRACE_PROBE(arc__reclaim_needfree); 2588 return (1); 2589 } 2590 2591 /* 2592 * Cooperate with pagedaemon when it's time for it to scan 2593 * and reclaim some pages. 2594 */ 2595 if (freemem < zfs_arc_free_target) { 2596 DTRACE_PROBE2(arc__reclaim_freemem, uint64_t, 2597 freemem, uint64_t, zfs_arc_free_target); 2598 return (1); 2599 } 2600 2601#ifdef sun 2602 /* 2603 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2604 */ 2605 extra = desfree; 2606 2607 /* 2608 * check that we're out of range of the pageout scanner. It starts to 2609 * schedule paging if freemem is less than lotsfree and needfree. 2610 * lotsfree is the high-water mark for pageout, and needfree is the 2611 * number of needed free pages. We add extra pages here to make sure 2612 * the scanner doesn't start up while we're freeing memory. 2613 */ 2614 if (freemem < lotsfree + needfree + extra) 2615 return (1); 2616 2617 /* 2618 * check to make sure that swapfs has enough space so that anon 2619 * reservations can still succeed. anon_resvmem() checks that the 2620 * availrmem is greater than swapfs_minfree, and the number of reserved 2621 * swap pages. We also add a bit of extra here just to prevent 2622 * circumstances from getting really dire. 2623 */ 2624 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2625 return (1); 2626 2627 /* 2628 * Check that we have enough availrmem that memory locking (e.g., via 2629 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 2630 * stores the number of pages that cannot be locked; when availrmem 2631 * drops below pages_pp_maximum, page locking mechanisms such as 2632 * page_pp_lock() will fail.) 2633 */ 2634 if (availrmem <= pages_pp_maximum) 2635 return (1); 2636 2637#endif /* sun */ 2638#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 2639 /* 2640 * If we're on an i386 platform, it's possible that we'll exhaust the 2641 * kernel heap space before we ever run out of available physical 2642 * memory. Most checks of the size of the heap_area compare against 2643 * tune.t_minarmem, which is the minimum available real memory that we 2644 * can have in the system. However, this is generally fixed at 25 pages 2645 * which is so low that it's useless. In this comparison, we seek to 2646 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2647 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2648 * free) 2649 */ 2650 if (vmem_size(heap_arena, VMEM_FREE) < 2651 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2)) { 2652 DTRACE_PROBE2(arc__reclaim_used, uint64_t, 2653 vmem_size(heap_arena, VMEM_FREE), uint64_t, 2654 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2); 2655 return (1); 2656 } 2657#define zio_arena NULL 2658#else 2659#define zio_arena heap_arena 2660#endif 2661 2662 /* 2663 * If zio data pages are being allocated out of a separate heap segment, 2664 * then enforce that the size of available vmem for this arena remains 2665 * above about 1/16th free. 2666 * 2667 * Note: The 1/16th arena free requirement was put in place 2668 * to aggressively evict memory from the arc in order to avoid 2669 * memory fragmentation issues. 2670 */ 2671 if (zio_arena != NULL && 2672 vmem_size(zio_arena, VMEM_FREE) < 2673 (vmem_size(zio_arena, VMEM_ALLOC) >> 4)) 2674 return (1); 2675 2676 /* 2677 * Above limits know nothing about real level of KVA fragmentation. 2678 * Start aggressive reclamation if too little sequential KVA left. 2679 */ 2680 if (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) 2681 return (1); 2682 2683#else /* _KERNEL */ 2684 if (spa_get_random(100) == 0) 2685 return (1); 2686#endif /* _KERNEL */ 2687 DTRACE_PROBE(arc__reclaim_no); 2688 2689 return (0); 2690} 2691 2692extern kmem_cache_t *zio_buf_cache[]; 2693extern kmem_cache_t *zio_data_buf_cache[]; 2694extern kmem_cache_t *range_seg_cache; 2695 2696static __noinline void 2697arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2698{ 2699 size_t i; 2700 kmem_cache_t *prev_cache = NULL; 2701 kmem_cache_t *prev_data_cache = NULL; 2702 2703 DTRACE_PROBE(arc__kmem_reap_start); 2704#ifdef _KERNEL 2705 if (arc_meta_used >= arc_meta_limit) { 2706 /* 2707 * We are exceeding our meta-data cache limit. 2708 * Purge some DNLC entries to release holds on meta-data. 2709 */ 2710 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2711 } 2712#if defined(__i386) 2713 /* 2714 * Reclaim unused memory from all kmem caches. 2715 */ 2716 kmem_reap(); 2717#endif 2718#endif 2719 2720 /* 2721 * An aggressive reclamation will shrink the cache size as well as 2722 * reap free buffers from the arc kmem caches. 2723 */ 2724 if (strat == ARC_RECLAIM_AGGR) 2725 arc_shrink(); 2726 2727 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2728 if (zio_buf_cache[i] != prev_cache) { 2729 prev_cache = zio_buf_cache[i]; 2730 kmem_cache_reap_now(zio_buf_cache[i]); 2731 } 2732 if (zio_data_buf_cache[i] != prev_data_cache) { 2733 prev_data_cache = zio_data_buf_cache[i]; 2734 kmem_cache_reap_now(zio_data_buf_cache[i]); 2735 } 2736 } 2737 kmem_cache_reap_now(buf_cache); 2738 kmem_cache_reap_now(hdr_cache); 2739 kmem_cache_reap_now(range_seg_cache); 2740 2741#ifdef sun 2742 /* 2743 * Ask the vmem arena to reclaim unused memory from its 2744 * quantum caches. 2745 */ 2746 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR) 2747 vmem_qcache_reap(zio_arena); 2748#endif 2749 DTRACE_PROBE(arc__kmem_reap_end); 2750} 2751 2752static void 2753arc_reclaim_thread(void *dummy __unused) 2754{ 2755 clock_t growtime = 0; 2756 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2757 callb_cpr_t cpr; 2758 2759 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2760 2761 mutex_enter(&arc_reclaim_thr_lock); 2762 while (arc_thread_exit == 0) { 2763 if (arc_reclaim_needed()) { 2764 2765 if (arc_no_grow) { 2766 if (last_reclaim == ARC_RECLAIM_CONS) { 2767 DTRACE_PROBE(arc__reclaim_aggr_no_grow); 2768 last_reclaim = ARC_RECLAIM_AGGR; 2769 } else { 2770 last_reclaim = ARC_RECLAIM_CONS; 2771 } 2772 } else { 2773 arc_no_grow = TRUE; 2774 last_reclaim = ARC_RECLAIM_AGGR; 2775 DTRACE_PROBE(arc__reclaim_aggr); 2776 membar_producer(); 2777 } 2778 2779 /* reset the growth delay for every reclaim */ 2780 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2781 2782 if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2783 /* 2784 * If needfree is TRUE our vm_lowmem hook 2785 * was called and in that case we must free some 2786 * memory, so switch to aggressive mode. 2787 */ 2788 arc_no_grow = TRUE; 2789 last_reclaim = ARC_RECLAIM_AGGR; 2790 } 2791 arc_kmem_reap_now(last_reclaim); 2792 arc_warm = B_TRUE; 2793 2794 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2795 arc_no_grow = FALSE; 2796 } 2797 2798 arc_adjust(); 2799 2800 if (arc_eviction_list != NULL) 2801 arc_do_user_evicts(); 2802 2803#ifdef _KERNEL 2804 if (needfree) { 2805 needfree = 0; 2806 wakeup(&needfree); 2807 } 2808#endif 2809 2810 /* block until needed, or one second, whichever is shorter */ 2811 CALLB_CPR_SAFE_BEGIN(&cpr); 2812 (void) cv_timedwait(&arc_reclaim_thr_cv, 2813 &arc_reclaim_thr_lock, hz); 2814 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2815 } 2816 2817 arc_thread_exit = 0; 2818 cv_broadcast(&arc_reclaim_thr_cv); 2819 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2820 thread_exit(); 2821} 2822 2823/* 2824 * Adapt arc info given the number of bytes we are trying to add and 2825 * the state that we are comming from. This function is only called 2826 * when we are adding new content to the cache. 2827 */ 2828static void 2829arc_adapt(int bytes, arc_state_t *state) 2830{ 2831 int mult; 2832 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2833 2834 if (state == arc_l2c_only) 2835 return; 2836 2837 ASSERT(bytes > 0); 2838 /* 2839 * Adapt the target size of the MRU list: 2840 * - if we just hit in the MRU ghost list, then increase 2841 * the target size of the MRU list. 2842 * - if we just hit in the MFU ghost list, then increase 2843 * the target size of the MFU list by decreasing the 2844 * target size of the MRU list. 2845 */ 2846 if (state == arc_mru_ghost) { 2847 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2848 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2849 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2850 2851 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2852 } else if (state == arc_mfu_ghost) { 2853 uint64_t delta; 2854 2855 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2856 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2857 mult = MIN(mult, 10); 2858 2859 delta = MIN(bytes * mult, arc_p); 2860 arc_p = MAX(arc_p_min, arc_p - delta); 2861 } 2862 ASSERT((int64_t)arc_p >= 0); 2863 2864 if (arc_reclaim_needed()) { 2865 cv_signal(&arc_reclaim_thr_cv); 2866 return; 2867 } 2868 2869 if (arc_no_grow) 2870 return; 2871 2872 if (arc_c >= arc_c_max) 2873 return; 2874 2875 /* 2876 * If we're within (2 * maxblocksize) bytes of the target 2877 * cache size, increment the target cache size 2878 */ 2879 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2880 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 2881 atomic_add_64(&arc_c, (int64_t)bytes); 2882 if (arc_c > arc_c_max) 2883 arc_c = arc_c_max; 2884 else if (state == arc_anon) 2885 atomic_add_64(&arc_p, (int64_t)bytes); 2886 if (arc_p > arc_c) 2887 arc_p = arc_c; 2888 } 2889 ASSERT((int64_t)arc_p >= 0); 2890} 2891 2892/* 2893 * Check if the cache has reached its limits and eviction is required 2894 * prior to insert. 2895 */ 2896static int 2897arc_evict_needed(arc_buf_contents_t type) 2898{ 2899 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2900 return (1); 2901 2902 if (arc_reclaim_needed()) 2903 return (1); 2904 2905 return (arc_size > arc_c); 2906} 2907 2908/* 2909 * The buffer, supplied as the first argument, needs a data block. 2910 * So, if we are at cache max, determine which cache should be victimized. 2911 * We have the following cases: 2912 * 2913 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2914 * In this situation if we're out of space, but the resident size of the MFU is 2915 * under the limit, victimize the MFU cache to satisfy this insertion request. 2916 * 2917 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2918 * Here, we've used up all of the available space for the MRU, so we need to 2919 * evict from our own cache instead. Evict from the set of resident MRU 2920 * entries. 2921 * 2922 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2923 * c minus p represents the MFU space in the cache, since p is the size of the 2924 * cache that is dedicated to the MRU. In this situation there's still space on 2925 * the MFU side, so the MRU side needs to be victimized. 2926 * 2927 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2928 * MFU's resident set is consuming more space than it has been allotted. In 2929 * this situation, we must victimize our own cache, the MFU, for this insertion. 2930 */ 2931static void 2932arc_get_data_buf(arc_buf_t *buf) 2933{ 2934 arc_state_t *state = buf->b_hdr->b_state; 2935 uint64_t size = buf->b_hdr->b_size; 2936 arc_buf_contents_t type = buf->b_hdr->b_type; 2937 2938 arc_adapt(size, state); 2939 2940 /* 2941 * We have not yet reached cache maximum size, 2942 * just allocate a new buffer. 2943 */ 2944 if (!arc_evict_needed(type)) { 2945 if (type == ARC_BUFC_METADATA) { 2946 buf->b_data = zio_buf_alloc(size); 2947 arc_space_consume(size, ARC_SPACE_DATA); 2948 } else { 2949 ASSERT(type == ARC_BUFC_DATA); 2950 buf->b_data = zio_data_buf_alloc(size); 2951 ARCSTAT_INCR(arcstat_data_size, size); 2952 atomic_add_64(&arc_size, size); 2953 } 2954 goto out; 2955 } 2956 2957 /* 2958 * If we are prefetching from the mfu ghost list, this buffer 2959 * will end up on the mru list; so steal space from there. 2960 */ 2961 if (state == arc_mfu_ghost) 2962 state = buf->b_hdr->b_flags & ARC_FLAG_PREFETCH ? 2963 arc_mru : arc_mfu; 2964 else if (state == arc_mru_ghost) 2965 state = arc_mru; 2966 2967 if (state == arc_mru || state == arc_anon) { 2968 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2969 state = (arc_mfu->arcs_lsize[type] >= size && 2970 arc_p > mru_used) ? arc_mfu : arc_mru; 2971 } else { 2972 /* MFU cases */ 2973 uint64_t mfu_space = arc_c - arc_p; 2974 state = (arc_mru->arcs_lsize[type] >= size && 2975 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2976 } 2977 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2978 if (type == ARC_BUFC_METADATA) { 2979 buf->b_data = zio_buf_alloc(size); 2980 arc_space_consume(size, ARC_SPACE_DATA); 2981 } else { 2982 ASSERT(type == ARC_BUFC_DATA); 2983 buf->b_data = zio_data_buf_alloc(size); 2984 ARCSTAT_INCR(arcstat_data_size, size); 2985 atomic_add_64(&arc_size, size); 2986 } 2987 ARCSTAT_BUMP(arcstat_recycle_miss); 2988 } 2989 ASSERT(buf->b_data != NULL); 2990out: 2991 /* 2992 * Update the state size. Note that ghost states have a 2993 * "ghost size" and so don't need to be updated. 2994 */ 2995 if (!GHOST_STATE(buf->b_hdr->b_state)) { 2996 arc_buf_hdr_t *hdr = buf->b_hdr; 2997 2998 atomic_add_64(&hdr->b_state->arcs_size, size); 2999 if (list_link_active(&hdr->b_arc_node)) { 3000 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 3001 atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 3002 } 3003 /* 3004 * If we are growing the cache, and we are adding anonymous 3005 * data, and we have outgrown arc_p, update arc_p 3006 */ 3007 if (arc_size < arc_c && hdr->b_state == arc_anon && 3008 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 3009 arc_p = MIN(arc_c, arc_p + size); 3010 } 3011 ARCSTAT_BUMP(arcstat_allocated); 3012} 3013 3014/* 3015 * This routine is called whenever a buffer is accessed. 3016 * NOTE: the hash lock is dropped in this function. 3017 */ 3018static void 3019arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3020{ 3021 clock_t now; 3022 3023 ASSERT(MUTEX_HELD(hash_lock)); 3024 3025 if (hdr->b_state == arc_anon) { 3026 /* 3027 * This buffer is not in the cache, and does not 3028 * appear in our "ghost" list. Add the new buffer 3029 * to the MRU state. 3030 */ 3031 3032 ASSERT(hdr->b_arc_access == 0); 3033 hdr->b_arc_access = ddi_get_lbolt(); 3034 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3035 arc_change_state(arc_mru, hdr, hash_lock); 3036 3037 } else if (hdr->b_state == arc_mru) { 3038 now = ddi_get_lbolt(); 3039 3040 /* 3041 * If this buffer is here because of a prefetch, then either: 3042 * - clear the flag if this is a "referencing" read 3043 * (any subsequent access will bump this into the MFU state). 3044 * or 3045 * - move the buffer to the head of the list if this is 3046 * another prefetch (to make it less likely to be evicted). 3047 */ 3048 if ((hdr->b_flags & ARC_FLAG_PREFETCH) != 0) { 3049 if (refcount_count(&hdr->b_refcnt) == 0) { 3050 ASSERT(list_link_active(&hdr->b_arc_node)); 3051 } else { 3052 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3053 ARCSTAT_BUMP(arcstat_mru_hits); 3054 } 3055 hdr->b_arc_access = now; 3056 return; 3057 } 3058 3059 /* 3060 * This buffer has been "accessed" only once so far, 3061 * but it is still in the cache. Move it to the MFU 3062 * state. 3063 */ 3064 if (now > hdr->b_arc_access + ARC_MINTIME) { 3065 /* 3066 * More than 125ms have passed since we 3067 * instantiated this buffer. Move it to the 3068 * most frequently used state. 3069 */ 3070 hdr->b_arc_access = now; 3071 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3072 arc_change_state(arc_mfu, hdr, hash_lock); 3073 } 3074 ARCSTAT_BUMP(arcstat_mru_hits); 3075 } else if (hdr->b_state == arc_mru_ghost) { 3076 arc_state_t *new_state; 3077 /* 3078 * This buffer has been "accessed" recently, but 3079 * was evicted from the cache. Move it to the 3080 * MFU state. 3081 */ 3082 3083 if (hdr->b_flags & ARC_FLAG_PREFETCH) { 3084 new_state = arc_mru; 3085 if (refcount_count(&hdr->b_refcnt) > 0) 3086 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3087 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3088 } else { 3089 new_state = arc_mfu; 3090 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3091 } 3092 3093 hdr->b_arc_access = ddi_get_lbolt(); 3094 arc_change_state(new_state, hdr, hash_lock); 3095 3096 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3097 } else if (hdr->b_state == arc_mfu) { 3098 /* 3099 * This buffer has been accessed more than once and is 3100 * still in the cache. Keep it in the MFU state. 3101 * 3102 * NOTE: an add_reference() that occurred when we did 3103 * the arc_read() will have kicked this off the list. 3104 * If it was a prefetch, we will explicitly move it to 3105 * the head of the list now. 3106 */ 3107 if ((hdr->b_flags & ARC_FLAG_PREFETCH) != 0) { 3108 ASSERT(refcount_count(&hdr->b_refcnt) == 0); 3109 ASSERT(list_link_active(&hdr->b_arc_node)); 3110 } 3111 ARCSTAT_BUMP(arcstat_mfu_hits); 3112 hdr->b_arc_access = ddi_get_lbolt(); 3113 } else if (hdr->b_state == arc_mfu_ghost) { 3114 arc_state_t *new_state = arc_mfu; 3115 /* 3116 * This buffer has been accessed more than once but has 3117 * been evicted from the cache. Move it back to the 3118 * MFU state. 3119 */ 3120 3121 if (hdr->b_flags & ARC_FLAG_PREFETCH) { 3122 /* 3123 * This is a prefetch access... 3124 * move this block back to the MRU state. 3125 */ 3126 ASSERT0(refcount_count(&hdr->b_refcnt)); 3127 new_state = arc_mru; 3128 } 3129 3130 hdr->b_arc_access = ddi_get_lbolt(); 3131 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3132 arc_change_state(new_state, hdr, hash_lock); 3133 3134 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 3135 } else if (hdr->b_state == arc_l2c_only) { 3136 /* 3137 * This buffer is on the 2nd Level ARC. 3138 */ 3139 3140 hdr->b_arc_access = ddi_get_lbolt(); 3141 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3142 arc_change_state(arc_mfu, hdr, hash_lock); 3143 } else { 3144 ASSERT(!"invalid arc state"); 3145 } 3146} 3147 3148/* a generic arc_done_func_t which you can use */ 3149/* ARGSUSED */ 3150void 3151arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 3152{ 3153 if (zio == NULL || zio->io_error == 0) 3154 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 3155 VERIFY(arc_buf_remove_ref(buf, arg)); 3156} 3157 3158/* a generic arc_done_func_t */ 3159void 3160arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 3161{ 3162 arc_buf_t **bufp = arg; 3163 if (zio && zio->io_error) { 3164 VERIFY(arc_buf_remove_ref(buf, arg)); 3165 *bufp = NULL; 3166 } else { 3167 *bufp = buf; 3168 ASSERT(buf->b_data); 3169 } 3170} 3171 3172static void 3173arc_read_done(zio_t *zio) 3174{ 3175 arc_buf_hdr_t *hdr; 3176 arc_buf_t *buf; 3177 arc_buf_t *abuf; /* buffer we're assigning to callback */ 3178 kmutex_t *hash_lock = NULL; 3179 arc_callback_t *callback_list, *acb; 3180 int freeable = FALSE; 3181 3182 buf = zio->io_private; 3183 hdr = buf->b_hdr; 3184 3185 /* 3186 * The hdr was inserted into hash-table and removed from lists 3187 * prior to starting I/O. We should find this header, since 3188 * it's in the hash table, and it should be legit since it's 3189 * not possible to evict it during the I/O. The only possible 3190 * reason for it not to be found is if we were freed during the 3191 * read. 3192 */ 3193 if (HDR_IN_HASH_TABLE(hdr)) { 3194 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3195 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3196 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3197 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3198 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3199 3200 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3201 &hash_lock); 3202 3203 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3204 hash_lock == NULL) || 3205 (found == hdr && 3206 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3207 (found == hdr && HDR_L2_READING(hdr))); 3208 } 3209 3210 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; 3211 if (l2arc_noprefetch && (hdr->b_flags & ARC_FLAG_PREFETCH)) 3212 hdr->b_flags &= ~ARC_FLAG_L2CACHE; 3213 3214 /* byteswap if necessary */ 3215 callback_list = hdr->b_acb; 3216 ASSERT(callback_list != NULL); 3217 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3218 dmu_object_byteswap_t bswap = 3219 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3220 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3221 byteswap_uint64_array : 3222 dmu_ot_byteswap[bswap].ob_func; 3223 func(buf->b_data, hdr->b_size); 3224 } 3225 3226 arc_cksum_compute(buf, B_FALSE); 3227#ifdef illumos 3228 arc_buf_watch(buf); 3229#endif /* illumos */ 3230 3231 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3232 /* 3233 * Only call arc_access on anonymous buffers. This is because 3234 * if we've issued an I/O for an evicted buffer, we've already 3235 * called arc_access (to prevent any simultaneous readers from 3236 * getting confused). 3237 */ 3238 arc_access(hdr, hash_lock); 3239 } 3240 3241 /* create copies of the data buffer for the callers */ 3242 abuf = buf; 3243 for (acb = callback_list; acb; acb = acb->acb_next) { 3244 if (acb->acb_done) { 3245 if (abuf == NULL) { 3246 ARCSTAT_BUMP(arcstat_duplicate_reads); 3247 abuf = arc_buf_clone(buf); 3248 } 3249 acb->acb_buf = abuf; 3250 abuf = NULL; 3251 } 3252 } 3253 hdr->b_acb = NULL; 3254 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 3255 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3256 if (abuf == buf) { 3257 ASSERT(buf->b_efunc == NULL); 3258 ASSERT(hdr->b_datacnt == 1); 3259 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 3260 } 3261 3262 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3263 3264 if (zio->io_error != 0) { 3265 hdr->b_flags |= ARC_FLAG_IO_ERROR; 3266 if (hdr->b_state != arc_anon) 3267 arc_change_state(arc_anon, hdr, hash_lock); 3268 if (HDR_IN_HASH_TABLE(hdr)) 3269 buf_hash_remove(hdr); 3270 freeable = refcount_is_zero(&hdr->b_refcnt); 3271 } 3272 3273 /* 3274 * Broadcast before we drop the hash_lock to avoid the possibility 3275 * that the hdr (and hence the cv) might be freed before we get to 3276 * the cv_broadcast(). 3277 */ 3278 cv_broadcast(&hdr->b_cv); 3279 3280 if (hash_lock) { 3281 mutex_exit(hash_lock); 3282 } else { 3283 /* 3284 * This block was freed while we waited for the read to 3285 * complete. It has been removed from the hash table and 3286 * moved to the anonymous state (so that it won't show up 3287 * in the cache). 3288 */ 3289 ASSERT3P(hdr->b_state, ==, arc_anon); 3290 freeable = refcount_is_zero(&hdr->b_refcnt); 3291 } 3292 3293 /* execute each callback and free its structure */ 3294 while ((acb = callback_list) != NULL) { 3295 if (acb->acb_done) 3296 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3297 3298 if (acb->acb_zio_dummy != NULL) { 3299 acb->acb_zio_dummy->io_error = zio->io_error; 3300 zio_nowait(acb->acb_zio_dummy); 3301 } 3302 3303 callback_list = acb->acb_next; 3304 kmem_free(acb, sizeof (arc_callback_t)); 3305 } 3306 3307 if (freeable) 3308 arc_hdr_destroy(hdr); 3309} 3310 3311/* 3312 * "Read" the block block at the specified DVA (in bp) via the 3313 * cache. If the block is found in the cache, invoke the provided 3314 * callback immediately and return. Note that the `zio' parameter 3315 * in the callback will be NULL in this case, since no IO was 3316 * required. If the block is not in the cache pass the read request 3317 * on to the spa with a substitute callback function, so that the 3318 * requested block will be added to the cache. 3319 * 3320 * If a read request arrives for a block that has a read in-progress, 3321 * either wait for the in-progress read to complete (and return the 3322 * results); or, if this is a read with a "done" func, add a record 3323 * to the read to invoke the "done" func when the read completes, 3324 * and return; or just return. 3325 * 3326 * arc_read_done() will invoke all the requested "done" functions 3327 * for readers of this block. 3328 */ 3329int 3330arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3331 void *private, zio_priority_t priority, int zio_flags, 3332 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 3333{ 3334 arc_buf_hdr_t *hdr = NULL; 3335 arc_buf_t *buf = NULL; 3336 kmutex_t *hash_lock = NULL; 3337 zio_t *rzio; 3338 uint64_t guid = spa_load_guid(spa); 3339 3340 ASSERT(!BP_IS_EMBEDDED(bp) || 3341 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3342 3343top: 3344 if (!BP_IS_EMBEDDED(bp)) { 3345 /* 3346 * Embedded BP's have no DVA and require no I/O to "read". 3347 * Create an anonymous arc buf to back it. 3348 */ 3349 hdr = buf_hash_find(guid, bp, &hash_lock); 3350 } 3351 3352 if (hdr != NULL && hdr->b_datacnt > 0) { 3353 3354 *arc_flags |= ARC_FLAG_CACHED; 3355 3356 if (HDR_IO_IN_PROGRESS(hdr)) { 3357 3358 if (*arc_flags & ARC_FLAG_WAIT) { 3359 cv_wait(&hdr->b_cv, hash_lock); 3360 mutex_exit(hash_lock); 3361 goto top; 3362 } 3363 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 3364 3365 if (done) { 3366 arc_callback_t *acb = NULL; 3367 3368 acb = kmem_zalloc(sizeof (arc_callback_t), 3369 KM_SLEEP); 3370 acb->acb_done = done; 3371 acb->acb_private = private; 3372 if (pio != NULL) 3373 acb->acb_zio_dummy = zio_null(pio, 3374 spa, NULL, NULL, NULL, zio_flags); 3375 3376 ASSERT(acb->acb_done != NULL); 3377 acb->acb_next = hdr->b_acb; 3378 hdr->b_acb = acb; 3379 add_reference(hdr, hash_lock, private); 3380 mutex_exit(hash_lock); 3381 return (0); 3382 } 3383 mutex_exit(hash_lock); 3384 return (0); 3385 } 3386 3387 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3388 3389 if (done) { 3390 add_reference(hdr, hash_lock, private); 3391 /* 3392 * If this block is already in use, create a new 3393 * copy of the data so that we will be guaranteed 3394 * that arc_release() will always succeed. 3395 */ 3396 buf = hdr->b_buf; 3397 ASSERT(buf); 3398 ASSERT(buf->b_data); 3399 if (HDR_BUF_AVAILABLE(hdr)) { 3400 ASSERT(buf->b_efunc == NULL); 3401 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 3402 } else { 3403 buf = arc_buf_clone(buf); 3404 } 3405 3406 } else if (*arc_flags & ARC_FLAG_PREFETCH && 3407 refcount_count(&hdr->b_refcnt) == 0) { 3408 hdr->b_flags |= ARC_FLAG_PREFETCH; 3409 } 3410 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3411 arc_access(hdr, hash_lock); 3412 if (*arc_flags & ARC_FLAG_L2CACHE) 3413 hdr->b_flags |= ARC_FLAG_L2CACHE; 3414 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3415 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3416 mutex_exit(hash_lock); 3417 ARCSTAT_BUMP(arcstat_hits); 3418 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 3419 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3420 data, metadata, hits); 3421 3422 if (done) 3423 done(NULL, buf, private); 3424 } else { 3425 uint64_t size = BP_GET_LSIZE(bp); 3426 arc_callback_t *acb; 3427 vdev_t *vd = NULL; 3428 uint64_t addr = 0; 3429 boolean_t devw = B_FALSE; 3430 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3431 uint64_t b_asize = 0; 3432 3433 if (hdr == NULL) { 3434 /* this block is not in the cache */ 3435 arc_buf_hdr_t *exists = NULL; 3436 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3437 buf = arc_buf_alloc(spa, size, private, type); 3438 hdr = buf->b_hdr; 3439 if (!BP_IS_EMBEDDED(bp)) { 3440 hdr->b_dva = *BP_IDENTITY(bp); 3441 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3442 hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3443 exists = buf_hash_insert(hdr, &hash_lock); 3444 } 3445 if (exists != NULL) { 3446 /* somebody beat us to the hash insert */ 3447 mutex_exit(hash_lock); 3448 buf_discard_identity(hdr); 3449 (void) arc_buf_remove_ref(buf, private); 3450 goto top; /* restart the IO request */ 3451 } 3452 3453 /* if this is a prefetch, we don't have a reference */ 3454 if (*arc_flags & ARC_FLAG_PREFETCH) { 3455 (void) remove_reference(hdr, hash_lock, 3456 private); 3457 hdr->b_flags |= ARC_FLAG_PREFETCH; 3458 } 3459 if (*arc_flags & ARC_FLAG_L2CACHE) 3460 hdr->b_flags |= ARC_FLAG_L2CACHE; 3461 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3462 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3463 if (BP_GET_LEVEL(bp) > 0) 3464 hdr->b_flags |= ARC_FLAG_INDIRECT; 3465 } else { 3466 /* this block is in the ghost cache */ 3467 ASSERT(GHOST_STATE(hdr->b_state)); 3468 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3469 ASSERT0(refcount_count(&hdr->b_refcnt)); 3470 ASSERT(hdr->b_buf == NULL); 3471 3472 /* if this is a prefetch, we don't have a reference */ 3473 if (*arc_flags & ARC_FLAG_PREFETCH) 3474 hdr->b_flags |= ARC_FLAG_PREFETCH; 3475 else 3476 add_reference(hdr, hash_lock, private); 3477 if (*arc_flags & ARC_FLAG_L2CACHE) 3478 hdr->b_flags |= ARC_FLAG_L2CACHE; 3479 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3480 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3481 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3482 buf->b_hdr = hdr; 3483 buf->b_data = NULL; 3484 buf->b_efunc = NULL; 3485 buf->b_private = NULL; 3486 buf->b_next = NULL; 3487 hdr->b_buf = buf; 3488 ASSERT(hdr->b_datacnt == 0); 3489 hdr->b_datacnt = 1; 3490 arc_get_data_buf(buf); 3491 arc_access(hdr, hash_lock); 3492 } 3493 3494 ASSERT(!GHOST_STATE(hdr->b_state)); 3495 3496 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3497 acb->acb_done = done; 3498 acb->acb_private = private; 3499 3500 ASSERT(hdr->b_acb == NULL); 3501 hdr->b_acb = acb; 3502 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 3503 3504 if (hdr->b_l2hdr != NULL && 3505 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3506 devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3507 addr = hdr->b_l2hdr->b_daddr; 3508 b_compress = hdr->b_l2hdr->b_compress; 3509 b_asize = hdr->b_l2hdr->b_asize; 3510 /* 3511 * Lock out device removal. 3512 */ 3513 if (vdev_is_dead(vd) || 3514 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3515 vd = NULL; 3516 } 3517 3518 if (hash_lock != NULL) 3519 mutex_exit(hash_lock); 3520 3521 /* 3522 * At this point, we have a level 1 cache miss. Try again in 3523 * L2ARC if possible. 3524 */ 3525 ASSERT3U(hdr->b_size, ==, size); 3526 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3527 uint64_t, size, zbookmark_phys_t *, zb); 3528 ARCSTAT_BUMP(arcstat_misses); 3529 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 3530 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3531 data, metadata, misses); 3532#ifdef _KERNEL 3533 curthread->td_ru.ru_inblock++; 3534#endif 3535 3536 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3537 /* 3538 * Read from the L2ARC if the following are true: 3539 * 1. The L2ARC vdev was previously cached. 3540 * 2. This buffer still has L2ARC metadata. 3541 * 3. This buffer isn't currently writing to the L2ARC. 3542 * 4. The L2ARC entry wasn't evicted, which may 3543 * also have invalidated the vdev. 3544 * 5. This isn't prefetch and l2arc_noprefetch is set. 3545 */ 3546 if (hdr->b_l2hdr != NULL && 3547 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3548 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3549 l2arc_read_callback_t *cb; 3550 3551 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3552 ARCSTAT_BUMP(arcstat_l2_hits); 3553 3554 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3555 KM_SLEEP); 3556 cb->l2rcb_buf = buf; 3557 cb->l2rcb_spa = spa; 3558 cb->l2rcb_bp = *bp; 3559 cb->l2rcb_zb = *zb; 3560 cb->l2rcb_flags = zio_flags; 3561 cb->l2rcb_compress = b_compress; 3562 3563 ASSERT(addr >= VDEV_LABEL_START_SIZE && 3564 addr + size < vd->vdev_psize - 3565 VDEV_LABEL_END_SIZE); 3566 3567 /* 3568 * l2arc read. The SCL_L2ARC lock will be 3569 * released by l2arc_read_done(). 3570 * Issue a null zio if the underlying buffer 3571 * was squashed to zero size by compression. 3572 */ 3573 if (b_compress == ZIO_COMPRESS_EMPTY) { 3574 rzio = zio_null(pio, spa, vd, 3575 l2arc_read_done, cb, 3576 zio_flags | ZIO_FLAG_DONT_CACHE | 3577 ZIO_FLAG_CANFAIL | 3578 ZIO_FLAG_DONT_PROPAGATE | 3579 ZIO_FLAG_DONT_RETRY); 3580 } else { 3581 rzio = zio_read_phys(pio, vd, addr, 3582 b_asize, buf->b_data, 3583 ZIO_CHECKSUM_OFF, 3584 l2arc_read_done, cb, priority, 3585 zio_flags | ZIO_FLAG_DONT_CACHE | 3586 ZIO_FLAG_CANFAIL | 3587 ZIO_FLAG_DONT_PROPAGATE | 3588 ZIO_FLAG_DONT_RETRY, B_FALSE); 3589 } 3590 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3591 zio_t *, rzio); 3592 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3593 3594 if (*arc_flags & ARC_FLAG_NOWAIT) { 3595 zio_nowait(rzio); 3596 return (0); 3597 } 3598 3599 ASSERT(*arc_flags & ARC_FLAG_WAIT); 3600 if (zio_wait(rzio) == 0) 3601 return (0); 3602 3603 /* l2arc read error; goto zio_read() */ 3604 } else { 3605 DTRACE_PROBE1(l2arc__miss, 3606 arc_buf_hdr_t *, hdr); 3607 ARCSTAT_BUMP(arcstat_l2_misses); 3608 if (HDR_L2_WRITING(hdr)) 3609 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3610 spa_config_exit(spa, SCL_L2ARC, vd); 3611 } 3612 } else { 3613 if (vd != NULL) 3614 spa_config_exit(spa, SCL_L2ARC, vd); 3615 if (l2arc_ndev != 0) { 3616 DTRACE_PROBE1(l2arc__miss, 3617 arc_buf_hdr_t *, hdr); 3618 ARCSTAT_BUMP(arcstat_l2_misses); 3619 } 3620 } 3621 3622 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3623 arc_read_done, buf, priority, zio_flags, zb); 3624 3625 if (*arc_flags & ARC_FLAG_WAIT) 3626 return (zio_wait(rzio)); 3627 3628 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 3629 zio_nowait(rzio); 3630 } 3631 return (0); 3632} 3633 3634void 3635arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3636{ 3637 ASSERT(buf->b_hdr != NULL); 3638 ASSERT(buf->b_hdr->b_state != arc_anon); 3639 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3640 ASSERT(buf->b_efunc == NULL); 3641 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3642 3643 buf->b_efunc = func; 3644 buf->b_private = private; 3645} 3646 3647/* 3648 * Notify the arc that a block was freed, and thus will never be used again. 3649 */ 3650void 3651arc_freed(spa_t *spa, const blkptr_t *bp) 3652{ 3653 arc_buf_hdr_t *hdr; 3654 kmutex_t *hash_lock; 3655 uint64_t guid = spa_load_guid(spa); 3656 3657 ASSERT(!BP_IS_EMBEDDED(bp)); 3658 3659 hdr = buf_hash_find(guid, bp, &hash_lock); 3660 if (hdr == NULL) 3661 return; 3662 if (HDR_BUF_AVAILABLE(hdr)) { 3663 arc_buf_t *buf = hdr->b_buf; 3664 add_reference(hdr, hash_lock, FTAG); 3665 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 3666 mutex_exit(hash_lock); 3667 3668 arc_release(buf, FTAG); 3669 (void) arc_buf_remove_ref(buf, FTAG); 3670 } else { 3671 mutex_exit(hash_lock); 3672 } 3673 3674} 3675 3676/* 3677 * Clear the user eviction callback set by arc_set_callback(), first calling 3678 * it if it exists. Because the presence of a callback keeps an arc_buf cached 3679 * clearing the callback may result in the arc_buf being destroyed. However, 3680 * it will not result in the *last* arc_buf being destroyed, hence the data 3681 * will remain cached in the ARC. We make a copy of the arc buffer here so 3682 * that we can process the callback without holding any locks. 3683 * 3684 * It's possible that the callback is already in the process of being cleared 3685 * by another thread. In this case we can not clear the callback. 3686 * 3687 * Returns B_TRUE if the callback was successfully called and cleared. 3688 */ 3689boolean_t 3690arc_clear_callback(arc_buf_t *buf) 3691{ 3692 arc_buf_hdr_t *hdr; 3693 kmutex_t *hash_lock; 3694 arc_evict_func_t *efunc = buf->b_efunc; 3695 void *private = buf->b_private; 3696 list_t *list, *evicted_list; 3697 kmutex_t *lock, *evicted_lock; 3698 3699 mutex_enter(&buf->b_evict_lock); 3700 hdr = buf->b_hdr; 3701 if (hdr == NULL) { 3702 /* 3703 * We are in arc_do_user_evicts(). 3704 */ 3705 ASSERT(buf->b_data == NULL); 3706 mutex_exit(&buf->b_evict_lock); 3707 return (B_FALSE); 3708 } else if (buf->b_data == NULL) { 3709 /* 3710 * We are on the eviction list; process this buffer now 3711 * but let arc_do_user_evicts() do the reaping. 3712 */ 3713 buf->b_efunc = NULL; 3714 mutex_exit(&buf->b_evict_lock); 3715 VERIFY0(efunc(private)); 3716 return (B_TRUE); 3717 } 3718 hash_lock = HDR_LOCK(hdr); 3719 mutex_enter(hash_lock); 3720 hdr = buf->b_hdr; 3721 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3722 3723 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3724 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3725 3726 buf->b_efunc = NULL; 3727 buf->b_private = NULL; 3728 3729 if (hdr->b_datacnt > 1) { 3730 mutex_exit(&buf->b_evict_lock); 3731 arc_buf_destroy(buf, FALSE, TRUE); 3732 } else { 3733 ASSERT(buf == hdr->b_buf); 3734 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 3735 mutex_exit(&buf->b_evict_lock); 3736 } 3737 3738 mutex_exit(hash_lock); 3739 VERIFY0(efunc(private)); 3740 return (B_TRUE); 3741} 3742 3743/* 3744 * Release this buffer from the cache, making it an anonymous buffer. This 3745 * must be done after a read and prior to modifying the buffer contents. 3746 * If the buffer has more than one reference, we must make 3747 * a new hdr for the buffer. 3748 */ 3749void 3750arc_release(arc_buf_t *buf, void *tag) 3751{ 3752 arc_buf_hdr_t *hdr; 3753 kmutex_t *hash_lock = NULL; 3754 l2arc_buf_hdr_t *l2hdr; 3755 uint64_t buf_size; 3756 3757 /* 3758 * It would be nice to assert that if it's DMU metadata (level > 3759 * 0 || it's the dnode file), then it must be syncing context. 3760 * But we don't know that information at this level. 3761 */ 3762 3763 mutex_enter(&buf->b_evict_lock); 3764 hdr = buf->b_hdr; 3765 3766 /* this buffer is not on any list */ 3767 ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3768 3769 if (hdr->b_state == arc_anon) { 3770 /* this buffer is already released */ 3771 ASSERT(buf->b_efunc == NULL); 3772 } else { 3773 hash_lock = HDR_LOCK(hdr); 3774 mutex_enter(hash_lock); 3775 hdr = buf->b_hdr; 3776 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3777 } 3778 3779 l2hdr = hdr->b_l2hdr; 3780 if (l2hdr) { 3781 mutex_enter(&l2arc_buflist_mtx); 3782 arc_buf_l2_cdata_free(hdr); 3783 hdr->b_l2hdr = NULL; 3784 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3785 } 3786 buf_size = hdr->b_size; 3787 3788 /* 3789 * Do we have more than one buf? 3790 */ 3791 if (hdr->b_datacnt > 1) { 3792 arc_buf_hdr_t *nhdr; 3793 arc_buf_t **bufp; 3794 uint64_t blksz = hdr->b_size; 3795 uint64_t spa = hdr->b_spa; 3796 arc_buf_contents_t type = hdr->b_type; 3797 uint32_t flags = hdr->b_flags; 3798 3799 ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3800 /* 3801 * Pull the data off of this hdr and attach it to 3802 * a new anonymous hdr. 3803 */ 3804 (void) remove_reference(hdr, hash_lock, tag); 3805 bufp = &hdr->b_buf; 3806 while (*bufp != buf) 3807 bufp = &(*bufp)->b_next; 3808 *bufp = buf->b_next; 3809 buf->b_next = NULL; 3810 3811 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3812 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3813 if (refcount_is_zero(&hdr->b_refcnt)) { 3814 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3815 ASSERT3U(*size, >=, hdr->b_size); 3816 atomic_add_64(size, -hdr->b_size); 3817 } 3818 3819 /* 3820 * We're releasing a duplicate user data buffer, update 3821 * our statistics accordingly. 3822 */ 3823 if (hdr->b_type == ARC_BUFC_DATA) { 3824 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3825 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3826 -hdr->b_size); 3827 } 3828 hdr->b_datacnt -= 1; 3829 arc_cksum_verify(buf); 3830#ifdef illumos 3831 arc_buf_unwatch(buf); 3832#endif /* illumos */ 3833 3834 mutex_exit(hash_lock); 3835 3836 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3837 nhdr->b_size = blksz; 3838 nhdr->b_spa = spa; 3839 nhdr->b_type = type; 3840 nhdr->b_buf = buf; 3841 nhdr->b_state = arc_anon; 3842 nhdr->b_arc_access = 0; 3843 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; 3844 nhdr->b_l2hdr = NULL; 3845 nhdr->b_datacnt = 1; 3846 nhdr->b_freeze_cksum = NULL; 3847 (void) refcount_add(&nhdr->b_refcnt, tag); 3848 buf->b_hdr = nhdr; 3849 mutex_exit(&buf->b_evict_lock); 3850 atomic_add_64(&arc_anon->arcs_size, blksz); 3851 } else { 3852 mutex_exit(&buf->b_evict_lock); 3853 ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3854 ASSERT(!list_link_active(&hdr->b_arc_node)); 3855 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3856 if (hdr->b_state != arc_anon) 3857 arc_change_state(arc_anon, hdr, hash_lock); 3858 hdr->b_arc_access = 0; 3859 if (hash_lock) 3860 mutex_exit(hash_lock); 3861 3862 buf_discard_identity(hdr); 3863 arc_buf_thaw(buf); 3864 } 3865 buf->b_efunc = NULL; 3866 buf->b_private = NULL; 3867 3868 if (l2hdr) { 3869 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3870 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 3871 -l2hdr->b_asize, 0, 0); 3872 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3873 l2hdr->b_asize, 0); 3874 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3875 ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3876 mutex_exit(&l2arc_buflist_mtx); 3877 } 3878} 3879 3880int 3881arc_released(arc_buf_t *buf) 3882{ 3883 int released; 3884 3885 mutex_enter(&buf->b_evict_lock); 3886 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3887 mutex_exit(&buf->b_evict_lock); 3888 return (released); 3889} 3890 3891#ifdef ZFS_DEBUG 3892int 3893arc_referenced(arc_buf_t *buf) 3894{ 3895 int referenced; 3896 3897 mutex_enter(&buf->b_evict_lock); 3898 referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3899 mutex_exit(&buf->b_evict_lock); 3900 return (referenced); 3901} 3902#endif 3903 3904static void 3905arc_write_ready(zio_t *zio) 3906{ 3907 arc_write_callback_t *callback = zio->io_private; 3908 arc_buf_t *buf = callback->awcb_buf; 3909 arc_buf_hdr_t *hdr = buf->b_hdr; 3910 3911 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3912 callback->awcb_ready(zio, buf, callback->awcb_private); 3913 3914 /* 3915 * If the IO is already in progress, then this is a re-write 3916 * attempt, so we need to thaw and re-compute the cksum. 3917 * It is the responsibility of the callback to handle the 3918 * accounting for any re-write attempt. 3919 */ 3920 if (HDR_IO_IN_PROGRESS(hdr)) { 3921 mutex_enter(&hdr->b_freeze_lock); 3922 if (hdr->b_freeze_cksum != NULL) { 3923 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3924 hdr->b_freeze_cksum = NULL; 3925 } 3926 mutex_exit(&hdr->b_freeze_lock); 3927 } 3928 arc_cksum_compute(buf, B_FALSE); 3929 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 3930} 3931 3932/* 3933 * The SPA calls this callback for each physical write that happens on behalf 3934 * of a logical write. See the comment in dbuf_write_physdone() for details. 3935 */ 3936static void 3937arc_write_physdone(zio_t *zio) 3938{ 3939 arc_write_callback_t *cb = zio->io_private; 3940 if (cb->awcb_physdone != NULL) 3941 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3942} 3943 3944static void 3945arc_write_done(zio_t *zio) 3946{ 3947 arc_write_callback_t *callback = zio->io_private; 3948 arc_buf_t *buf = callback->awcb_buf; 3949 arc_buf_hdr_t *hdr = buf->b_hdr; 3950 3951 ASSERT(hdr->b_acb == NULL); 3952 3953 if (zio->io_error == 0) { 3954 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 3955 buf_discard_identity(hdr); 3956 } else { 3957 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3958 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3959 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3960 } 3961 } else { 3962 ASSERT(BUF_EMPTY(hdr)); 3963 } 3964 3965 /* 3966 * If the block to be written was all-zero or compressed enough to be 3967 * embedded in the BP, no write was performed so there will be no 3968 * dva/birth/checksum. The buffer must therefore remain anonymous 3969 * (and uncached). 3970 */ 3971 if (!BUF_EMPTY(hdr)) { 3972 arc_buf_hdr_t *exists; 3973 kmutex_t *hash_lock; 3974 3975 ASSERT(zio->io_error == 0); 3976 3977 arc_cksum_verify(buf); 3978 3979 exists = buf_hash_insert(hdr, &hash_lock); 3980 if (exists) { 3981 /* 3982 * This can only happen if we overwrite for 3983 * sync-to-convergence, because we remove 3984 * buffers from the hash table when we arc_free(). 3985 */ 3986 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3987 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3988 panic("bad overwrite, hdr=%p exists=%p", 3989 (void *)hdr, (void *)exists); 3990 ASSERT(refcount_is_zero(&exists->b_refcnt)); 3991 arc_change_state(arc_anon, exists, hash_lock); 3992 mutex_exit(hash_lock); 3993 arc_hdr_destroy(exists); 3994 exists = buf_hash_insert(hdr, &hash_lock); 3995 ASSERT3P(exists, ==, NULL); 3996 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3997 /* nopwrite */ 3998 ASSERT(zio->io_prop.zp_nopwrite); 3999 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 4000 panic("bad nopwrite, hdr=%p exists=%p", 4001 (void *)hdr, (void *)exists); 4002 } else { 4003 /* Dedup */ 4004 ASSERT(hdr->b_datacnt == 1); 4005 ASSERT(hdr->b_state == arc_anon); 4006 ASSERT(BP_GET_DEDUP(zio->io_bp)); 4007 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 4008 } 4009 } 4010 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4011 /* if it's not anon, we are doing a scrub */ 4012 if (!exists && hdr->b_state == arc_anon) 4013 arc_access(hdr, hash_lock); 4014 mutex_exit(hash_lock); 4015 } else { 4016 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4017 } 4018 4019 ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 4020 callback->awcb_done(zio, buf, callback->awcb_private); 4021 4022 kmem_free(callback, sizeof (arc_write_callback_t)); 4023} 4024 4025zio_t * 4026arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 4027 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 4028 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 4029 arc_done_func_t *done, void *private, zio_priority_t priority, 4030 int zio_flags, const zbookmark_phys_t *zb) 4031{ 4032 arc_buf_hdr_t *hdr = buf->b_hdr; 4033 arc_write_callback_t *callback; 4034 zio_t *zio; 4035 4036 ASSERT(ready != NULL); 4037 ASSERT(done != NULL); 4038 ASSERT(!HDR_IO_ERROR(hdr)); 4039 ASSERT((hdr->b_flags & ARC_FLAG_IO_IN_PROGRESS) == 0); 4040 ASSERT(hdr->b_acb == NULL); 4041 if (l2arc) 4042 hdr->b_flags |= ARC_FLAG_L2CACHE; 4043 if (l2arc_compress) 4044 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4045 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 4046 callback->awcb_ready = ready; 4047 callback->awcb_physdone = physdone; 4048 callback->awcb_done = done; 4049 callback->awcb_private = private; 4050 callback->awcb_buf = buf; 4051 4052 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 4053 arc_write_ready, arc_write_physdone, arc_write_done, callback, 4054 priority, zio_flags, zb); 4055 4056 return (zio); 4057} 4058 4059static int 4060arc_memory_throttle(uint64_t reserve, uint64_t txg) 4061{ 4062#ifdef _KERNEL 4063 uint64_t available_memory = ptob(freemem); 4064 static uint64_t page_load = 0; 4065 static uint64_t last_txg = 0; 4066 4067#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 4068 available_memory = 4069 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 4070#endif 4071 4072 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 4073 return (0); 4074 4075 if (txg > last_txg) { 4076 last_txg = txg; 4077 page_load = 0; 4078 } 4079 /* 4080 * If we are in pageout, we know that memory is already tight, 4081 * the arc is already going to be evicting, so we just want to 4082 * continue to let page writes occur as quickly as possible. 4083 */ 4084 if (curproc == pageproc) { 4085 if (page_load > MAX(ptob(minfree), available_memory) / 4) 4086 return (SET_ERROR(ERESTART)); 4087 /* Note: reserve is inflated, so we deflate */ 4088 page_load += reserve / 8; 4089 return (0); 4090 } else if (page_load > 0 && arc_reclaim_needed()) { 4091 /* memory is low, delay before restarting */ 4092 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 4093 return (SET_ERROR(EAGAIN)); 4094 } 4095 page_load = 0; 4096#endif 4097 return (0); 4098} 4099 4100void 4101arc_tempreserve_clear(uint64_t reserve) 4102{ 4103 atomic_add_64(&arc_tempreserve, -reserve); 4104 ASSERT((int64_t)arc_tempreserve >= 0); 4105} 4106 4107int 4108arc_tempreserve_space(uint64_t reserve, uint64_t txg) 4109{ 4110 int error; 4111 uint64_t anon_size; 4112 4113 if (reserve > arc_c/4 && !arc_no_grow) { 4114 arc_c = MIN(arc_c_max, reserve * 4); 4115 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 4116 } 4117 if (reserve > arc_c) 4118 return (SET_ERROR(ENOMEM)); 4119 4120 /* 4121 * Don't count loaned bufs as in flight dirty data to prevent long 4122 * network delays from blocking transactions that are ready to be 4123 * assigned to a txg. 4124 */ 4125 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 4126 4127 /* 4128 * Writes will, almost always, require additional memory allocations 4129 * in order to compress/encrypt/etc the data. We therefore need to 4130 * make sure that there is sufficient available memory for this. 4131 */ 4132 error = arc_memory_throttle(reserve, txg); 4133 if (error != 0) 4134 return (error); 4135 4136 /* 4137 * Throttle writes when the amount of dirty data in the cache 4138 * gets too large. We try to keep the cache less than half full 4139 * of dirty blocks so that our sync times don't grow too large. 4140 * Note: if two requests come in concurrently, we might let them 4141 * both succeed, when one of them should fail. Not a huge deal. 4142 */ 4143 4144 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 4145 anon_size > arc_c / 4) { 4146 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 4147 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 4148 arc_tempreserve>>10, 4149 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 4150 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 4151 reserve>>10, arc_c>>10); 4152 return (SET_ERROR(ERESTART)); 4153 } 4154 atomic_add_64(&arc_tempreserve, reserve); 4155 return (0); 4156} 4157 4158static kmutex_t arc_lowmem_lock; 4159#ifdef _KERNEL 4160static eventhandler_tag arc_event_lowmem = NULL; 4161 4162static void 4163arc_lowmem(void *arg __unused, int howto __unused) 4164{ 4165 4166 /* Serialize access via arc_lowmem_lock. */ 4167 mutex_enter(&arc_lowmem_lock); 4168 mutex_enter(&arc_reclaim_thr_lock); 4169 needfree = 1; 4170 DTRACE_PROBE(arc__needfree); 4171 cv_signal(&arc_reclaim_thr_cv); 4172 4173 /* 4174 * It is unsafe to block here in arbitrary threads, because we can come 4175 * here from ARC itself and may hold ARC locks and thus risk a deadlock 4176 * with ARC reclaim thread. 4177 */ 4178 if (curproc == pageproc) { 4179 while (needfree) 4180 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4181 } 4182 mutex_exit(&arc_reclaim_thr_lock); 4183 mutex_exit(&arc_lowmem_lock); 4184} 4185#endif 4186 4187void 4188arc_init(void) 4189{ 4190 int i, prefetch_tunable_set = 0; 4191 4192 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4193 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4194 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4195 4196 /* Convert seconds to clock ticks */ 4197 arc_min_prefetch_lifespan = 1 * hz; 4198 4199 /* Start out with 1/8 of all memory */ 4200 arc_c = kmem_size() / 8; 4201 4202#ifdef sun 4203#ifdef _KERNEL 4204 /* 4205 * On architectures where the physical memory can be larger 4206 * than the addressable space (intel in 32-bit mode), we may 4207 * need to limit the cache to 1/8 of VM size. 4208 */ 4209 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4210#endif 4211#endif /* sun */ 4212 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4213 arc_c_min = MAX(arc_c / 4, 16 << 20); 4214 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4215 if (arc_c * 8 >= 1 << 30) 4216 arc_c_max = (arc_c * 8) - (1 << 30); 4217 else 4218 arc_c_max = arc_c_min; 4219 arc_c_max = MAX(arc_c * 5, arc_c_max); 4220 4221#ifdef _KERNEL 4222 /* 4223 * Allow the tunables to override our calculations if they are 4224 * reasonable (ie. over 16MB) 4225 */ 4226 if (zfs_arc_max > 16 << 20 && zfs_arc_max < kmem_size()) 4227 arc_c_max = zfs_arc_max; 4228 if (zfs_arc_min > 16 << 20 && zfs_arc_min <= arc_c_max) 4229 arc_c_min = zfs_arc_min; 4230#endif 4231 4232 arc_c = arc_c_max; 4233 arc_p = (arc_c >> 1); 4234 4235 /* limit meta-data to 1/4 of the arc capacity */ 4236 arc_meta_limit = arc_c_max / 4; 4237 4238 /* Allow the tunable to override if it is reasonable */ 4239 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4240 arc_meta_limit = zfs_arc_meta_limit; 4241 4242 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4243 arc_c_min = arc_meta_limit / 2; 4244 4245 if (zfs_arc_meta_min > 0) { 4246 arc_meta_min = zfs_arc_meta_min; 4247 } else { 4248 arc_meta_min = arc_c_min / 2; 4249 } 4250 4251 if (zfs_arc_grow_retry > 0) 4252 arc_grow_retry = zfs_arc_grow_retry; 4253 4254 if (zfs_arc_shrink_shift > 0) 4255 arc_shrink_shift = zfs_arc_shrink_shift; 4256 4257 if (zfs_arc_p_min_shift > 0) 4258 arc_p_min_shift = zfs_arc_p_min_shift; 4259 4260 /* if kmem_flags are set, lets try to use less memory */ 4261 if (kmem_debugging()) 4262 arc_c = arc_c / 2; 4263 if (arc_c < arc_c_min) 4264 arc_c = arc_c_min; 4265 4266 zfs_arc_min = arc_c_min; 4267 zfs_arc_max = arc_c_max; 4268 4269 arc_anon = &ARC_anon; 4270 arc_mru = &ARC_mru; 4271 arc_mru_ghost = &ARC_mru_ghost; 4272 arc_mfu = &ARC_mfu; 4273 arc_mfu_ghost = &ARC_mfu_ghost; 4274 arc_l2c_only = &ARC_l2c_only; 4275 arc_size = 0; 4276 4277 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4278 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4279 NULL, MUTEX_DEFAULT, NULL); 4280 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4281 NULL, MUTEX_DEFAULT, NULL); 4282 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4283 NULL, MUTEX_DEFAULT, NULL); 4284 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4285 NULL, MUTEX_DEFAULT, NULL); 4286 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4287 NULL, MUTEX_DEFAULT, NULL); 4288 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4289 NULL, MUTEX_DEFAULT, NULL); 4290 4291 list_create(&arc_mru->arcs_lists[i], 4292 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4293 list_create(&arc_mru_ghost->arcs_lists[i], 4294 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4295 list_create(&arc_mfu->arcs_lists[i], 4296 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4297 list_create(&arc_mfu_ghost->arcs_lists[i], 4298 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4299 list_create(&arc_mfu_ghost->arcs_lists[i], 4300 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4301 list_create(&arc_l2c_only->arcs_lists[i], 4302 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4303 } 4304 4305 buf_init(); 4306 4307 arc_thread_exit = 0; 4308 arc_eviction_list = NULL; 4309 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4310 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4311 4312 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4313 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4314 4315 if (arc_ksp != NULL) { 4316 arc_ksp->ks_data = &arc_stats; 4317 kstat_install(arc_ksp); 4318 } 4319 4320 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4321 TS_RUN, minclsyspri); 4322 4323#ifdef _KERNEL 4324 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4325 EVENTHANDLER_PRI_FIRST); 4326#endif 4327 4328 arc_dead = FALSE; 4329 arc_warm = B_FALSE; 4330 4331 /* 4332 * Calculate maximum amount of dirty data per pool. 4333 * 4334 * If it has been set by /etc/system, take that. 4335 * Otherwise, use a percentage of physical memory defined by 4336 * zfs_dirty_data_max_percent (default 10%) with a cap at 4337 * zfs_dirty_data_max_max (default 4GB). 4338 */ 4339 if (zfs_dirty_data_max == 0) { 4340 zfs_dirty_data_max = ptob(physmem) * 4341 zfs_dirty_data_max_percent / 100; 4342 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4343 zfs_dirty_data_max_max); 4344 } 4345 4346#ifdef _KERNEL 4347 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4348 prefetch_tunable_set = 1; 4349 4350#ifdef __i386__ 4351 if (prefetch_tunable_set == 0) { 4352 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4353 "-- to enable,\n"); 4354 printf(" add \"vfs.zfs.prefetch_disable=0\" " 4355 "to /boot/loader.conf.\n"); 4356 zfs_prefetch_disable = 1; 4357 } 4358#else 4359 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4360 prefetch_tunable_set == 0) { 4361 printf("ZFS NOTICE: Prefetch is disabled by default if less " 4362 "than 4GB of RAM is present;\n" 4363 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4364 "to /boot/loader.conf.\n"); 4365 zfs_prefetch_disable = 1; 4366 } 4367#endif 4368 /* Warn about ZFS memory and address space requirements. */ 4369 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4370 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4371 "expect unstable behavior.\n"); 4372 } 4373 if (kmem_size() < 512 * (1 << 20)) { 4374 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4375 "expect unstable behavior.\n"); 4376 printf(" Consider tuning vm.kmem_size and " 4377 "vm.kmem_size_max\n"); 4378 printf(" in /boot/loader.conf.\n"); 4379 } 4380#endif 4381} 4382 4383void 4384arc_fini(void) 4385{ 4386 int i; 4387 4388 mutex_enter(&arc_reclaim_thr_lock); 4389 arc_thread_exit = 1; 4390 cv_signal(&arc_reclaim_thr_cv); 4391 while (arc_thread_exit != 0) 4392 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4393 mutex_exit(&arc_reclaim_thr_lock); 4394 4395 arc_flush(NULL); 4396 4397 arc_dead = TRUE; 4398 4399 if (arc_ksp != NULL) { 4400 kstat_delete(arc_ksp); 4401 arc_ksp = NULL; 4402 } 4403 4404 mutex_destroy(&arc_eviction_mtx); 4405 mutex_destroy(&arc_reclaim_thr_lock); 4406 cv_destroy(&arc_reclaim_thr_cv); 4407 4408 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4409 list_destroy(&arc_mru->arcs_lists[i]); 4410 list_destroy(&arc_mru_ghost->arcs_lists[i]); 4411 list_destroy(&arc_mfu->arcs_lists[i]); 4412 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4413 list_destroy(&arc_l2c_only->arcs_lists[i]); 4414 4415 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4416 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4417 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4418 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4419 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4420 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4421 } 4422 4423 buf_fini(); 4424 4425 ASSERT(arc_loaned_bytes == 0); 4426 4427 mutex_destroy(&arc_lowmem_lock); 4428#ifdef _KERNEL 4429 if (arc_event_lowmem != NULL) 4430 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4431#endif 4432} 4433 4434/* 4435 * Level 2 ARC 4436 * 4437 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4438 * It uses dedicated storage devices to hold cached data, which are populated 4439 * using large infrequent writes. The main role of this cache is to boost 4440 * the performance of random read workloads. The intended L2ARC devices 4441 * include short-stroked disks, solid state disks, and other media with 4442 * substantially faster read latency than disk. 4443 * 4444 * +-----------------------+ 4445 * | ARC | 4446 * +-----------------------+ 4447 * | ^ ^ 4448 * | | | 4449 * l2arc_feed_thread() arc_read() 4450 * | | | 4451 * | l2arc read | 4452 * V | | 4453 * +---------------+ | 4454 * | L2ARC | | 4455 * +---------------+ | 4456 * | ^ | 4457 * l2arc_write() | | 4458 * | | | 4459 * V | | 4460 * +-------+ +-------+ 4461 * | vdev | | vdev | 4462 * | cache | | cache | 4463 * +-------+ +-------+ 4464 * +=========+ .-----. 4465 * : L2ARC : |-_____-| 4466 * : devices : | Disks | 4467 * +=========+ `-_____-' 4468 * 4469 * Read requests are satisfied from the following sources, in order: 4470 * 4471 * 1) ARC 4472 * 2) vdev cache of L2ARC devices 4473 * 3) L2ARC devices 4474 * 4) vdev cache of disks 4475 * 5) disks 4476 * 4477 * Some L2ARC device types exhibit extremely slow write performance. 4478 * To accommodate for this there are some significant differences between 4479 * the L2ARC and traditional cache design: 4480 * 4481 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4482 * the ARC behave as usual, freeing buffers and placing headers on ghost 4483 * lists. The ARC does not send buffers to the L2ARC during eviction as 4484 * this would add inflated write latencies for all ARC memory pressure. 4485 * 4486 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4487 * It does this by periodically scanning buffers from the eviction-end of 4488 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4489 * not already there. It scans until a headroom of buffers is satisfied, 4490 * which itself is a buffer for ARC eviction. If a compressible buffer is 4491 * found during scanning and selected for writing to an L2ARC device, we 4492 * temporarily boost scanning headroom during the next scan cycle to make 4493 * sure we adapt to compression effects (which might significantly reduce 4494 * the data volume we write to L2ARC). The thread that does this is 4495 * l2arc_feed_thread(), illustrated below; example sizes are included to 4496 * provide a better sense of ratio than this diagram: 4497 * 4498 * head --> tail 4499 * +---------------------+----------+ 4500 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4501 * +---------------------+----------+ | o L2ARC eligible 4502 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4503 * +---------------------+----------+ | 4504 * 15.9 Gbytes ^ 32 Mbytes | 4505 * headroom | 4506 * l2arc_feed_thread() 4507 * | 4508 * l2arc write hand <--[oooo]--' 4509 * | 8 Mbyte 4510 * | write max 4511 * V 4512 * +==============================+ 4513 * L2ARC dev |####|#|###|###| |####| ... | 4514 * +==============================+ 4515 * 32 Gbytes 4516 * 4517 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4518 * evicted, then the L2ARC has cached a buffer much sooner than it probably 4519 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4520 * safe to say that this is an uncommon case, since buffers at the end of 4521 * the ARC lists have moved there due to inactivity. 4522 * 4523 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4524 * then the L2ARC simply misses copying some buffers. This serves as a 4525 * pressure valve to prevent heavy read workloads from both stalling the ARC 4526 * with waits and clogging the L2ARC with writes. This also helps prevent 4527 * the potential for the L2ARC to churn if it attempts to cache content too 4528 * quickly, such as during backups of the entire pool. 4529 * 4530 * 5. After system boot and before the ARC has filled main memory, there are 4531 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4532 * lists can remain mostly static. Instead of searching from tail of these 4533 * lists as pictured, the l2arc_feed_thread() will search from the list heads 4534 * for eligible buffers, greatly increasing its chance of finding them. 4535 * 4536 * The L2ARC device write speed is also boosted during this time so that 4537 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4538 * there are no L2ARC reads, and no fear of degrading read performance 4539 * through increased writes. 4540 * 4541 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4542 * the vdev queue can aggregate them into larger and fewer writes. Each 4543 * device is written to in a rotor fashion, sweeping writes through 4544 * available space then repeating. 4545 * 4546 * 7. The L2ARC does not store dirty content. It never needs to flush 4547 * write buffers back to disk based storage. 4548 * 4549 * 8. If an ARC buffer is written (and dirtied) which also exists in the 4550 * L2ARC, the now stale L2ARC buffer is immediately dropped. 4551 * 4552 * The performance of the L2ARC can be tweaked by a number of tunables, which 4553 * may be necessary for different workloads: 4554 * 4555 * l2arc_write_max max write bytes per interval 4556 * l2arc_write_boost extra write bytes during device warmup 4557 * l2arc_noprefetch skip caching prefetched buffers 4558 * l2arc_headroom number of max device writes to precache 4559 * l2arc_headroom_boost when we find compressed buffers during ARC 4560 * scanning, we multiply headroom by this 4561 * percentage factor for the next scan cycle, 4562 * since more compressed buffers are likely to 4563 * be present 4564 * l2arc_feed_secs seconds between L2ARC writing 4565 * 4566 * Tunables may be removed or added as future performance improvements are 4567 * integrated, and also may become zpool properties. 4568 * 4569 * There are three key functions that control how the L2ARC warms up: 4570 * 4571 * l2arc_write_eligible() check if a buffer is eligible to cache 4572 * l2arc_write_size() calculate how much to write 4573 * l2arc_write_interval() calculate sleep delay between writes 4574 * 4575 * These three functions determine what to write, how much, and how quickly 4576 * to send writes. 4577 */ 4578 4579static boolean_t 4580l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 4581{ 4582 /* 4583 * A buffer is *not* eligible for the L2ARC if it: 4584 * 1. belongs to a different spa. 4585 * 2. is already cached on the L2ARC. 4586 * 3. has an I/O in progress (it may be an incomplete read). 4587 * 4. is flagged not eligible (zfs property). 4588 */ 4589 if (hdr->b_spa != spa_guid) { 4590 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4591 return (B_FALSE); 4592 } 4593 if (hdr->b_l2hdr != NULL) { 4594 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4595 return (B_FALSE); 4596 } 4597 if (HDR_IO_IN_PROGRESS(hdr)) { 4598 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4599 return (B_FALSE); 4600 } 4601 if (!HDR_L2CACHE(hdr)) { 4602 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4603 return (B_FALSE); 4604 } 4605 4606 return (B_TRUE); 4607} 4608 4609static uint64_t 4610l2arc_write_size(void) 4611{ 4612 uint64_t size; 4613 4614 /* 4615 * Make sure our globals have meaningful values in case the user 4616 * altered them. 4617 */ 4618 size = l2arc_write_max; 4619 if (size == 0) { 4620 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4621 "be greater than zero, resetting it to the default (%d)", 4622 L2ARC_WRITE_SIZE); 4623 size = l2arc_write_max = L2ARC_WRITE_SIZE; 4624 } 4625 4626 if (arc_warm == B_FALSE) 4627 size += l2arc_write_boost; 4628 4629 return (size); 4630 4631} 4632 4633static clock_t 4634l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4635{ 4636 clock_t interval, next, now; 4637 4638 /* 4639 * If the ARC lists are busy, increase our write rate; if the 4640 * lists are stale, idle back. This is achieved by checking 4641 * how much we previously wrote - if it was more than half of 4642 * what we wanted, schedule the next write much sooner. 4643 */ 4644 if (l2arc_feed_again && wrote > (wanted / 2)) 4645 interval = (hz * l2arc_feed_min_ms) / 1000; 4646 else 4647 interval = hz * l2arc_feed_secs; 4648 4649 now = ddi_get_lbolt(); 4650 next = MAX(now, MIN(now + interval, began + interval)); 4651 4652 return (next); 4653} 4654 4655static void 4656l2arc_hdr_stat_add(void) 4657{ 4658 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4659 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4660} 4661 4662static void 4663l2arc_hdr_stat_remove(void) 4664{ 4665 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4666 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4667} 4668 4669/* 4670 * Cycle through L2ARC devices. This is how L2ARC load balances. 4671 * If a device is returned, this also returns holding the spa config lock. 4672 */ 4673static l2arc_dev_t * 4674l2arc_dev_get_next(void) 4675{ 4676 l2arc_dev_t *first, *next = NULL; 4677 4678 /* 4679 * Lock out the removal of spas (spa_namespace_lock), then removal 4680 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4681 * both locks will be dropped and a spa config lock held instead. 4682 */ 4683 mutex_enter(&spa_namespace_lock); 4684 mutex_enter(&l2arc_dev_mtx); 4685 4686 /* if there are no vdevs, there is nothing to do */ 4687 if (l2arc_ndev == 0) 4688 goto out; 4689 4690 first = NULL; 4691 next = l2arc_dev_last; 4692 do { 4693 /* loop around the list looking for a non-faulted vdev */ 4694 if (next == NULL) { 4695 next = list_head(l2arc_dev_list); 4696 } else { 4697 next = list_next(l2arc_dev_list, next); 4698 if (next == NULL) 4699 next = list_head(l2arc_dev_list); 4700 } 4701 4702 /* if we have come back to the start, bail out */ 4703 if (first == NULL) 4704 first = next; 4705 else if (next == first) 4706 break; 4707 4708 } while (vdev_is_dead(next->l2ad_vdev)); 4709 4710 /* if we were unable to find any usable vdevs, return NULL */ 4711 if (vdev_is_dead(next->l2ad_vdev)) 4712 next = NULL; 4713 4714 l2arc_dev_last = next; 4715 4716out: 4717 mutex_exit(&l2arc_dev_mtx); 4718 4719 /* 4720 * Grab the config lock to prevent the 'next' device from being 4721 * removed while we are writing to it. 4722 */ 4723 if (next != NULL) 4724 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4725 mutex_exit(&spa_namespace_lock); 4726 4727 return (next); 4728} 4729 4730/* 4731 * Free buffers that were tagged for destruction. 4732 */ 4733static void 4734l2arc_do_free_on_write() 4735{ 4736 list_t *buflist; 4737 l2arc_data_free_t *df, *df_prev; 4738 4739 mutex_enter(&l2arc_free_on_write_mtx); 4740 buflist = l2arc_free_on_write; 4741 4742 for (df = list_tail(buflist); df; df = df_prev) { 4743 df_prev = list_prev(buflist, df); 4744 ASSERT(df->l2df_data != NULL); 4745 ASSERT(df->l2df_func != NULL); 4746 df->l2df_func(df->l2df_data, df->l2df_size); 4747 list_remove(buflist, df); 4748 kmem_free(df, sizeof (l2arc_data_free_t)); 4749 } 4750 4751 mutex_exit(&l2arc_free_on_write_mtx); 4752} 4753 4754/* 4755 * A write to a cache device has completed. Update all headers to allow 4756 * reads from these buffers to begin. 4757 */ 4758static void 4759l2arc_write_done(zio_t *zio) 4760{ 4761 l2arc_write_callback_t *cb; 4762 l2arc_dev_t *dev; 4763 list_t *buflist; 4764 arc_buf_hdr_t *head, *hdr, *hdr_prev; 4765 l2arc_buf_hdr_t *abl2; 4766 kmutex_t *hash_lock; 4767 int64_t bytes_dropped = 0; 4768 4769 cb = zio->io_private; 4770 ASSERT(cb != NULL); 4771 dev = cb->l2wcb_dev; 4772 ASSERT(dev != NULL); 4773 head = cb->l2wcb_head; 4774 ASSERT(head != NULL); 4775 buflist = dev->l2ad_buflist; 4776 ASSERT(buflist != NULL); 4777 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4778 l2arc_write_callback_t *, cb); 4779 4780 if (zio->io_error != 0) 4781 ARCSTAT_BUMP(arcstat_l2_writes_error); 4782 4783 mutex_enter(&l2arc_buflist_mtx); 4784 4785 /* 4786 * All writes completed, or an error was hit. 4787 */ 4788 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 4789 hdr_prev = list_prev(buflist, hdr); 4790 abl2 = hdr->b_l2hdr; 4791 4792 /* 4793 * Release the temporary compressed buffer as soon as possible. 4794 */ 4795 if (abl2->b_compress != ZIO_COMPRESS_OFF) 4796 l2arc_release_cdata_buf(hdr); 4797 4798 hash_lock = HDR_LOCK(hdr); 4799 if (!mutex_tryenter(hash_lock)) { 4800 /* 4801 * This buffer misses out. It may be in a stage 4802 * of eviction. Its ARC_L2_WRITING flag will be 4803 * left set, denying reads to this buffer. 4804 */ 4805 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4806 continue; 4807 } 4808 4809 if (zio->io_error != 0) { 4810 /* 4811 * Error - drop L2ARC entry. 4812 */ 4813 list_remove(buflist, hdr); 4814 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4815 bytes_dropped += abl2->b_asize; 4816 hdr->b_l2hdr = NULL; 4817 trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4818 abl2->b_asize, 0); 4819 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4820 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 4821 } 4822 4823 /* 4824 * Allow ARC to begin reads to this L2ARC entry. 4825 */ 4826 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 4827 4828 mutex_exit(hash_lock); 4829 } 4830 4831 atomic_inc_64(&l2arc_writes_done); 4832 list_remove(buflist, head); 4833 kmem_cache_free(hdr_cache, head); 4834 mutex_exit(&l2arc_buflist_mtx); 4835 4836 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 4837 4838 l2arc_do_free_on_write(); 4839 4840 kmem_free(cb, sizeof (l2arc_write_callback_t)); 4841} 4842 4843/* 4844 * A read to a cache device completed. Validate buffer contents before 4845 * handing over to the regular ARC routines. 4846 */ 4847static void 4848l2arc_read_done(zio_t *zio) 4849{ 4850 l2arc_read_callback_t *cb; 4851 arc_buf_hdr_t *hdr; 4852 arc_buf_t *buf; 4853 kmutex_t *hash_lock; 4854 int equal; 4855 4856 ASSERT(zio->io_vd != NULL); 4857 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4858 4859 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4860 4861 cb = zio->io_private; 4862 ASSERT(cb != NULL); 4863 buf = cb->l2rcb_buf; 4864 ASSERT(buf != NULL); 4865 4866 hash_lock = HDR_LOCK(buf->b_hdr); 4867 mutex_enter(hash_lock); 4868 hdr = buf->b_hdr; 4869 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4870 4871 /* 4872 * If the buffer was compressed, decompress it first. 4873 */ 4874 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4875 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4876 ASSERT(zio->io_data != NULL); 4877 4878 /* 4879 * Check this survived the L2ARC journey. 4880 */ 4881 equal = arc_cksum_equal(buf); 4882 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4883 mutex_exit(hash_lock); 4884 zio->io_private = buf; 4885 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4886 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4887 arc_read_done(zio); 4888 } else { 4889 mutex_exit(hash_lock); 4890 /* 4891 * Buffer didn't survive caching. Increment stats and 4892 * reissue to the original storage device. 4893 */ 4894 if (zio->io_error != 0) { 4895 ARCSTAT_BUMP(arcstat_l2_io_error); 4896 } else { 4897 zio->io_error = SET_ERROR(EIO); 4898 } 4899 if (!equal) 4900 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4901 4902 /* 4903 * If there's no waiter, issue an async i/o to the primary 4904 * storage now. If there *is* a waiter, the caller must 4905 * issue the i/o in a context where it's OK to block. 4906 */ 4907 if (zio->io_waiter == NULL) { 4908 zio_t *pio = zio_unique_parent(zio); 4909 4910 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4911 4912 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4913 buf->b_data, zio->io_size, arc_read_done, buf, 4914 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4915 } 4916 } 4917 4918 kmem_free(cb, sizeof (l2arc_read_callback_t)); 4919} 4920 4921/* 4922 * This is the list priority from which the L2ARC will search for pages to 4923 * cache. This is used within loops (0..3) to cycle through lists in the 4924 * desired order. This order can have a significant effect on cache 4925 * performance. 4926 * 4927 * Currently the metadata lists are hit first, MFU then MRU, followed by 4928 * the data lists. This function returns a locked list, and also returns 4929 * the lock pointer. 4930 */ 4931static list_t * 4932l2arc_list_locked(int list_num, kmutex_t **lock) 4933{ 4934 list_t *list = NULL; 4935 int idx; 4936 4937 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4938 4939 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4940 idx = list_num; 4941 list = &arc_mfu->arcs_lists[idx]; 4942 *lock = ARCS_LOCK(arc_mfu, idx); 4943 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4944 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4945 list = &arc_mru->arcs_lists[idx]; 4946 *lock = ARCS_LOCK(arc_mru, idx); 4947 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4948 ARC_BUFC_NUMDATALISTS)) { 4949 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4950 list = &arc_mfu->arcs_lists[idx]; 4951 *lock = ARCS_LOCK(arc_mfu, idx); 4952 } else { 4953 idx = list_num - ARC_BUFC_NUMLISTS; 4954 list = &arc_mru->arcs_lists[idx]; 4955 *lock = ARCS_LOCK(arc_mru, idx); 4956 } 4957 4958 ASSERT(!(MUTEX_HELD(*lock))); 4959 mutex_enter(*lock); 4960 return (list); 4961} 4962 4963/* 4964 * Evict buffers from the device write hand to the distance specified in 4965 * bytes. This distance may span populated buffers, it may span nothing. 4966 * This is clearing a region on the L2ARC device ready for writing. 4967 * If the 'all' boolean is set, every buffer is evicted. 4968 */ 4969static void 4970l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4971{ 4972 list_t *buflist; 4973 l2arc_buf_hdr_t *abl2; 4974 arc_buf_hdr_t *hdr, *hdr_prev; 4975 kmutex_t *hash_lock; 4976 uint64_t taddr; 4977 int64_t bytes_evicted = 0; 4978 4979 buflist = dev->l2ad_buflist; 4980 4981 if (buflist == NULL) 4982 return; 4983 4984 if (!all && dev->l2ad_first) { 4985 /* 4986 * This is the first sweep through the device. There is 4987 * nothing to evict. 4988 */ 4989 return; 4990 } 4991 4992 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4993 /* 4994 * When nearing the end of the device, evict to the end 4995 * before the device write hand jumps to the start. 4996 */ 4997 taddr = dev->l2ad_end; 4998 } else { 4999 taddr = dev->l2ad_hand + distance; 5000 } 5001 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 5002 uint64_t, taddr, boolean_t, all); 5003 5004top: 5005 mutex_enter(&l2arc_buflist_mtx); 5006 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 5007 hdr_prev = list_prev(buflist, hdr); 5008 5009 hash_lock = HDR_LOCK(hdr); 5010 if (!mutex_tryenter(hash_lock)) { 5011 /* 5012 * Missed the hash lock. Retry. 5013 */ 5014 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 5015 mutex_exit(&l2arc_buflist_mtx); 5016 mutex_enter(hash_lock); 5017 mutex_exit(hash_lock); 5018 goto top; 5019 } 5020 5021 if (HDR_L2_WRITE_HEAD(hdr)) { 5022 /* 5023 * We hit a write head node. Leave it for 5024 * l2arc_write_done(). 5025 */ 5026 list_remove(buflist, hdr); 5027 mutex_exit(hash_lock); 5028 continue; 5029 } 5030 5031 if (!all && hdr->b_l2hdr != NULL && 5032 (hdr->b_l2hdr->b_daddr > taddr || 5033 hdr->b_l2hdr->b_daddr < dev->l2ad_hand)) { 5034 /* 5035 * We've evicted to the target address, 5036 * or the end of the device. 5037 */ 5038 mutex_exit(hash_lock); 5039 break; 5040 } 5041 5042 if (HDR_FREE_IN_PROGRESS(hdr)) { 5043 /* 5044 * Already on the path to destruction. 5045 */ 5046 mutex_exit(hash_lock); 5047 continue; 5048 } 5049 5050 if (hdr->b_state == arc_l2c_only) { 5051 ASSERT(!HDR_L2_READING(hdr)); 5052 /* 5053 * This doesn't exist in the ARC. Destroy. 5054 * arc_hdr_destroy() will call list_remove() 5055 * and decrement arcstat_l2_size. 5056 */ 5057 arc_change_state(arc_anon, hdr, hash_lock); 5058 arc_hdr_destroy(hdr); 5059 } else { 5060 /* 5061 * Invalidate issued or about to be issued 5062 * reads, since we may be about to write 5063 * over this location. 5064 */ 5065 if (HDR_L2_READING(hdr)) { 5066 ARCSTAT_BUMP(arcstat_l2_evict_reading); 5067 hdr->b_flags |= ARC_FLAG_L2_EVICTED; 5068 } 5069 5070 /* 5071 * Tell ARC this no longer exists in L2ARC. 5072 */ 5073 if (hdr->b_l2hdr != NULL) { 5074 abl2 = hdr->b_l2hdr; 5075 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 5076 bytes_evicted += abl2->b_asize; 5077 hdr->b_l2hdr = NULL; 5078 /* 5079 * We are destroying l2hdr, so ensure that 5080 * its compressed buffer, if any, is not leaked. 5081 */ 5082 ASSERT(abl2->b_tmp_cdata == NULL); 5083 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 5084 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 5085 } 5086 list_remove(buflist, hdr); 5087 5088 /* 5089 * This may have been leftover after a 5090 * failed write. 5091 */ 5092 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 5093 } 5094 mutex_exit(hash_lock); 5095 } 5096 mutex_exit(&l2arc_buflist_mtx); 5097 5098 vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); 5099 dev->l2ad_evict = taddr; 5100} 5101 5102/* 5103 * Find and write ARC buffers to the L2ARC device. 5104 * 5105 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 5106 * for reading until they have completed writing. 5107 * The headroom_boost is an in-out parameter used to maintain headroom boost 5108 * state between calls to this function. 5109 * 5110 * Returns the number of bytes actually written (which may be smaller than 5111 * the delta by which the device hand has changed due to alignment). 5112 */ 5113static uint64_t 5114l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 5115 boolean_t *headroom_boost) 5116{ 5117 arc_buf_hdr_t *hdr, *hdr_prev, *head; 5118 list_t *list; 5119 uint64_t write_asize, write_sz, headroom, buf_compress_minsz; 5120 void *buf_data; 5121 kmutex_t *list_lock; 5122 boolean_t full; 5123 l2arc_write_callback_t *cb; 5124 zio_t *pio, *wzio; 5125 uint64_t guid = spa_load_guid(spa); 5126 const boolean_t do_headroom_boost = *headroom_boost; 5127 int try; 5128 5129 ASSERT(dev->l2ad_vdev != NULL); 5130 5131 /* Lower the flag now, we might want to raise it again later. */ 5132 *headroom_boost = B_FALSE; 5133 5134 pio = NULL; 5135 write_sz = write_asize = 0; 5136 full = B_FALSE; 5137 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 5138 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; 5139 5140 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 5141 /* 5142 * We will want to try to compress buffers that are at least 2x the 5143 * device sector size. 5144 */ 5145 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 5146 5147 /* 5148 * Copy buffers for L2ARC writing. 5149 */ 5150 mutex_enter(&l2arc_buflist_mtx); 5151 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 5152 uint64_t passed_sz = 0; 5153 5154 list = l2arc_list_locked(try, &list_lock); 5155 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 5156 5157 /* 5158 * L2ARC fast warmup. 5159 * 5160 * Until the ARC is warm and starts to evict, read from the 5161 * head of the ARC lists rather than the tail. 5162 */ 5163 if (arc_warm == B_FALSE) 5164 hdr = list_head(list); 5165 else 5166 hdr = list_tail(list); 5167 if (hdr == NULL) 5168 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 5169 5170 headroom = target_sz * l2arc_headroom * 2 / ARC_BUFC_NUMLISTS; 5171 if (do_headroom_boost) 5172 headroom = (headroom * l2arc_headroom_boost) / 100; 5173 5174 for (; hdr; hdr = hdr_prev) { 5175 l2arc_buf_hdr_t *l2hdr; 5176 kmutex_t *hash_lock; 5177 uint64_t buf_sz; 5178 uint64_t buf_a_sz; 5179 5180 if (arc_warm == B_FALSE) 5181 hdr_prev = list_next(list, hdr); 5182 else 5183 hdr_prev = list_prev(list, hdr); 5184 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size); 5185 5186 hash_lock = HDR_LOCK(hdr); 5187 if (!mutex_tryenter(hash_lock)) { 5188 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5189 /* 5190 * Skip this buffer rather than waiting. 5191 */ 5192 continue; 5193 } 5194 5195 passed_sz += hdr->b_size; 5196 if (passed_sz > headroom) { 5197 /* 5198 * Searched too far. 5199 */ 5200 mutex_exit(hash_lock); 5201 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5202 break; 5203 } 5204 5205 if (!l2arc_write_eligible(guid, hdr)) { 5206 mutex_exit(hash_lock); 5207 continue; 5208 } 5209 5210 /* 5211 * Assume that the buffer is not going to be compressed 5212 * and could take more space on disk because of a larger 5213 * disk block size. 5214 */ 5215 buf_sz = hdr->b_size; 5216 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5217 5218 if ((write_asize + buf_a_sz) > target_sz) { 5219 full = B_TRUE; 5220 mutex_exit(hash_lock); 5221 ARCSTAT_BUMP(arcstat_l2_write_full); 5222 break; 5223 } 5224 5225 if (pio == NULL) { 5226 /* 5227 * Insert a dummy header on the buflist so 5228 * l2arc_write_done() can find where the 5229 * write buffers begin without searching. 5230 */ 5231 list_insert_head(dev->l2ad_buflist, head); 5232 5233 cb = kmem_alloc( 5234 sizeof (l2arc_write_callback_t), KM_SLEEP); 5235 cb->l2wcb_dev = dev; 5236 cb->l2wcb_head = head; 5237 pio = zio_root(spa, l2arc_write_done, cb, 5238 ZIO_FLAG_CANFAIL); 5239 ARCSTAT_BUMP(arcstat_l2_write_pios); 5240 } 5241 5242 /* 5243 * Create and add a new L2ARC header. 5244 */ 5245 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5246 l2hdr->b_dev = dev; 5247 hdr->b_flags |= ARC_FLAG_L2_WRITING; 5248 5249 /* 5250 * Temporarily stash the data buffer in b_tmp_cdata. 5251 * The subsequent write step will pick it up from 5252 * there. This is because can't access hdr->b_buf 5253 * without holding the hash_lock, which we in turn 5254 * can't access without holding the ARC list locks 5255 * (which we want to avoid during compression/writing). 5256 */ 5257 l2hdr->b_compress = ZIO_COMPRESS_OFF; 5258 l2hdr->b_asize = hdr->b_size; 5259 l2hdr->b_tmp_cdata = hdr->b_buf->b_data; 5260 5261 hdr->b_l2hdr = l2hdr; 5262 5263 list_insert_head(dev->l2ad_buflist, hdr); 5264 5265 /* 5266 * Compute and store the buffer cksum before 5267 * writing. On debug the cksum is verified first. 5268 */ 5269 arc_cksum_verify(hdr->b_buf); 5270 arc_cksum_compute(hdr->b_buf, B_TRUE); 5271 5272 mutex_exit(hash_lock); 5273 5274 write_sz += buf_sz; 5275 write_asize += buf_a_sz; 5276 } 5277 5278 mutex_exit(list_lock); 5279 5280 if (full == B_TRUE) 5281 break; 5282 } 5283 5284 /* No buffers selected for writing? */ 5285 if (pio == NULL) { 5286 ASSERT0(write_sz); 5287 mutex_exit(&l2arc_buflist_mtx); 5288 kmem_cache_free(hdr_cache, head); 5289 return (0); 5290 } 5291 5292 /* 5293 * Note that elsewhere in this file arcstat_l2_asize 5294 * and the used space on l2ad_vdev are updated using b_asize, 5295 * which is not necessarily rounded up to the device block size. 5296 * Too keep accounting consistent we do the same here as well: 5297 * stats_size accumulates the sum of b_asize of the written buffers, 5298 * while write_asize accumulates the sum of b_asize rounded up 5299 * to the device block size. 5300 * The latter sum is used only to validate the corectness of the code. 5301 */ 5302 uint64_t stats_size = 0; 5303 write_asize = 0; 5304 5305 /* 5306 * Now start writing the buffers. We're starting at the write head 5307 * and work backwards, retracing the course of the buffer selector 5308 * loop above. 5309 */ 5310 for (hdr = list_prev(dev->l2ad_buflist, head); hdr; 5311 hdr = list_prev(dev->l2ad_buflist, hdr)) { 5312 l2arc_buf_hdr_t *l2hdr; 5313 uint64_t buf_sz; 5314 5315 /* 5316 * We shouldn't need to lock the buffer here, since we flagged 5317 * it as ARC_FLAG_L2_WRITING in the previous step, but we must 5318 * take care to only access its L2 cache parameters. In 5319 * particular, hdr->b_buf may be invalid by now due to 5320 * ARC eviction. 5321 */ 5322 l2hdr = hdr->b_l2hdr; 5323 l2hdr->b_daddr = dev->l2ad_hand; 5324 5325 if ((hdr->b_flags & ARC_FLAG_L2COMPRESS) && 5326 l2hdr->b_asize >= buf_compress_minsz) { 5327 if (l2arc_compress_buf(l2hdr)) { 5328 /* 5329 * If compression succeeded, enable headroom 5330 * boost on the next scan cycle. 5331 */ 5332 *headroom_boost = B_TRUE; 5333 } 5334 } 5335 5336 /* 5337 * Pick up the buffer data we had previously stashed away 5338 * (and now potentially also compressed). 5339 */ 5340 buf_data = l2hdr->b_tmp_cdata; 5341 buf_sz = l2hdr->b_asize; 5342 5343 /* 5344 * If the data has not been compressed, then clear b_tmp_cdata 5345 * to make sure that it points only to a temporary compression 5346 * buffer. 5347 */ 5348 if (!L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)) 5349 l2hdr->b_tmp_cdata = NULL; 5350 5351 /* Compression may have squashed the buffer to zero length. */ 5352 if (buf_sz != 0) { 5353 uint64_t buf_a_sz; 5354 5355 wzio = zio_write_phys(pio, dev->l2ad_vdev, 5356 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5357 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5358 ZIO_FLAG_CANFAIL, B_FALSE); 5359 5360 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5361 zio_t *, wzio); 5362 (void) zio_nowait(wzio); 5363 5364 stats_size += buf_sz; 5365 /* 5366 * Keep the clock hand suitably device-aligned. 5367 */ 5368 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5369 write_asize += buf_a_sz; 5370 dev->l2ad_hand += buf_a_sz; 5371 } 5372 } 5373 5374 mutex_exit(&l2arc_buflist_mtx); 5375 5376 ASSERT3U(write_asize, <=, target_sz); 5377 ARCSTAT_BUMP(arcstat_l2_writes_sent); 5378 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5379 ARCSTAT_INCR(arcstat_l2_size, write_sz); 5380 ARCSTAT_INCR(arcstat_l2_asize, stats_size); 5381 vdev_space_update(dev->l2ad_vdev, stats_size, 0, 0); 5382 5383 /* 5384 * Bump device hand to the device start if it is approaching the end. 5385 * l2arc_evict() will already have evicted ahead for this case. 5386 */ 5387 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5388 dev->l2ad_hand = dev->l2ad_start; 5389 dev->l2ad_evict = dev->l2ad_start; 5390 dev->l2ad_first = B_FALSE; 5391 } 5392 5393 dev->l2ad_writing = B_TRUE; 5394 (void) zio_wait(pio); 5395 dev->l2ad_writing = B_FALSE; 5396 5397 return (write_asize); 5398} 5399 5400/* 5401 * Compresses an L2ARC buffer. 5402 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5403 * size in l2hdr->b_asize. This routine tries to compress the data and 5404 * depending on the compression result there are three possible outcomes: 5405 * *) The buffer was incompressible. The original l2hdr contents were left 5406 * untouched and are ready for writing to an L2 device. 5407 * *) The buffer was all-zeros, so there is no need to write it to an L2 5408 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5409 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5410 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5411 * data buffer which holds the compressed data to be written, and b_asize 5412 * tells us how much data there is. b_compress is set to the appropriate 5413 * compression algorithm. Once writing is done, invoke 5414 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5415 * 5416 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5417 * buffer was incompressible). 5418 */ 5419static boolean_t 5420l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5421{ 5422 void *cdata; 5423 size_t csize, len, rounded; 5424 5425 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5426 ASSERT(l2hdr->b_tmp_cdata != NULL); 5427 5428 len = l2hdr->b_asize; 5429 cdata = zio_data_buf_alloc(len); 5430 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5431 cdata, l2hdr->b_asize); 5432 5433 if (csize == 0) { 5434 /* zero block, indicate that there's nothing to write */ 5435 zio_data_buf_free(cdata, len); 5436 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5437 l2hdr->b_asize = 0; 5438 l2hdr->b_tmp_cdata = NULL; 5439 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5440 return (B_TRUE); 5441 } 5442 5443 rounded = P2ROUNDUP(csize, 5444 (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift); 5445 if (rounded < len) { 5446 /* 5447 * Compression succeeded, we'll keep the cdata around for 5448 * writing and release it afterwards. 5449 */ 5450 if (rounded > csize) { 5451 bzero((char *)cdata + csize, rounded - csize); 5452 csize = rounded; 5453 } 5454 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5455 l2hdr->b_asize = csize; 5456 l2hdr->b_tmp_cdata = cdata; 5457 ARCSTAT_BUMP(arcstat_l2_compress_successes); 5458 return (B_TRUE); 5459 } else { 5460 /* 5461 * Compression failed, release the compressed buffer. 5462 * l2hdr will be left unmodified. 5463 */ 5464 zio_data_buf_free(cdata, len); 5465 ARCSTAT_BUMP(arcstat_l2_compress_failures); 5466 return (B_FALSE); 5467 } 5468} 5469 5470/* 5471 * Decompresses a zio read back from an l2arc device. On success, the 5472 * underlying zio's io_data buffer is overwritten by the uncompressed 5473 * version. On decompression error (corrupt compressed stream), the 5474 * zio->io_error value is set to signal an I/O error. 5475 * 5476 * Please note that the compressed data stream is not checksummed, so 5477 * if the underlying device is experiencing data corruption, we may feed 5478 * corrupt data to the decompressor, so the decompressor needs to be 5479 * able to handle this situation (LZ4 does). 5480 */ 5481static void 5482l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5483{ 5484 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5485 5486 if (zio->io_error != 0) { 5487 /* 5488 * An io error has occured, just restore the original io 5489 * size in preparation for a main pool read. 5490 */ 5491 zio->io_orig_size = zio->io_size = hdr->b_size; 5492 return; 5493 } 5494 5495 if (c == ZIO_COMPRESS_EMPTY) { 5496 /* 5497 * An empty buffer results in a null zio, which means we 5498 * need to fill its io_data after we're done restoring the 5499 * buffer's contents. 5500 */ 5501 ASSERT(hdr->b_buf != NULL); 5502 bzero(hdr->b_buf->b_data, hdr->b_size); 5503 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5504 } else { 5505 ASSERT(zio->io_data != NULL); 5506 /* 5507 * We copy the compressed data from the start of the arc buffer 5508 * (the zio_read will have pulled in only what we need, the 5509 * rest is garbage which we will overwrite at decompression) 5510 * and then decompress back to the ARC data buffer. This way we 5511 * can minimize copying by simply decompressing back over the 5512 * original compressed data (rather than decompressing to an 5513 * aux buffer and then copying back the uncompressed buffer, 5514 * which is likely to be much larger). 5515 */ 5516 uint64_t csize; 5517 void *cdata; 5518 5519 csize = zio->io_size; 5520 cdata = zio_data_buf_alloc(csize); 5521 bcopy(zio->io_data, cdata, csize); 5522 if (zio_decompress_data(c, cdata, zio->io_data, csize, 5523 hdr->b_size) != 0) 5524 zio->io_error = EIO; 5525 zio_data_buf_free(cdata, csize); 5526 } 5527 5528 /* Restore the expected uncompressed IO size. */ 5529 zio->io_orig_size = zio->io_size = hdr->b_size; 5530} 5531 5532/* 5533 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5534 * This buffer serves as a temporary holder of compressed data while 5535 * the buffer entry is being written to an l2arc device. Once that is 5536 * done, we can dispose of it. 5537 */ 5538static void 5539l2arc_release_cdata_buf(arc_buf_hdr_t *hdr) 5540{ 5541 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 5542 5543 ASSERT(L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)); 5544 if (l2hdr->b_compress != ZIO_COMPRESS_EMPTY) { 5545 /* 5546 * If the data was compressed, then we've allocated a 5547 * temporary buffer for it, so now we need to release it. 5548 */ 5549 ASSERT(l2hdr->b_tmp_cdata != NULL); 5550 zio_data_buf_free(l2hdr->b_tmp_cdata, hdr->b_size); 5551 l2hdr->b_tmp_cdata = NULL; 5552 } else { 5553 ASSERT(l2hdr->b_tmp_cdata == NULL); 5554 } 5555} 5556 5557/* 5558 * This thread feeds the L2ARC at regular intervals. This is the beating 5559 * heart of the L2ARC. 5560 */ 5561static void 5562l2arc_feed_thread(void *dummy __unused) 5563{ 5564 callb_cpr_t cpr; 5565 l2arc_dev_t *dev; 5566 spa_t *spa; 5567 uint64_t size, wrote; 5568 clock_t begin, next = ddi_get_lbolt(); 5569 boolean_t headroom_boost = B_FALSE; 5570 5571 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5572 5573 mutex_enter(&l2arc_feed_thr_lock); 5574 5575 while (l2arc_thread_exit == 0) { 5576 CALLB_CPR_SAFE_BEGIN(&cpr); 5577 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5578 next - ddi_get_lbolt()); 5579 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5580 next = ddi_get_lbolt() + hz; 5581 5582 /* 5583 * Quick check for L2ARC devices. 5584 */ 5585 mutex_enter(&l2arc_dev_mtx); 5586 if (l2arc_ndev == 0) { 5587 mutex_exit(&l2arc_dev_mtx); 5588 continue; 5589 } 5590 mutex_exit(&l2arc_dev_mtx); 5591 begin = ddi_get_lbolt(); 5592 5593 /* 5594 * This selects the next l2arc device to write to, and in 5595 * doing so the next spa to feed from: dev->l2ad_spa. This 5596 * will return NULL if there are now no l2arc devices or if 5597 * they are all faulted. 5598 * 5599 * If a device is returned, its spa's config lock is also 5600 * held to prevent device removal. l2arc_dev_get_next() 5601 * will grab and release l2arc_dev_mtx. 5602 */ 5603 if ((dev = l2arc_dev_get_next()) == NULL) 5604 continue; 5605 5606 spa = dev->l2ad_spa; 5607 ASSERT(spa != NULL); 5608 5609 /* 5610 * If the pool is read-only then force the feed thread to 5611 * sleep a little longer. 5612 */ 5613 if (!spa_writeable(spa)) { 5614 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5615 spa_config_exit(spa, SCL_L2ARC, dev); 5616 continue; 5617 } 5618 5619 /* 5620 * Avoid contributing to memory pressure. 5621 */ 5622 if (arc_reclaim_needed()) { 5623 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5624 spa_config_exit(spa, SCL_L2ARC, dev); 5625 continue; 5626 } 5627 5628 ARCSTAT_BUMP(arcstat_l2_feeds); 5629 5630 size = l2arc_write_size(); 5631 5632 /* 5633 * Evict L2ARC buffers that will be overwritten. 5634 */ 5635 l2arc_evict(dev, size, B_FALSE); 5636 5637 /* 5638 * Write ARC buffers. 5639 */ 5640 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5641 5642 /* 5643 * Calculate interval between writes. 5644 */ 5645 next = l2arc_write_interval(begin, size, wrote); 5646 spa_config_exit(spa, SCL_L2ARC, dev); 5647 } 5648 5649 l2arc_thread_exit = 0; 5650 cv_broadcast(&l2arc_feed_thr_cv); 5651 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5652 thread_exit(); 5653} 5654 5655boolean_t 5656l2arc_vdev_present(vdev_t *vd) 5657{ 5658 l2arc_dev_t *dev; 5659 5660 mutex_enter(&l2arc_dev_mtx); 5661 for (dev = list_head(l2arc_dev_list); dev != NULL; 5662 dev = list_next(l2arc_dev_list, dev)) { 5663 if (dev->l2ad_vdev == vd) 5664 break; 5665 } 5666 mutex_exit(&l2arc_dev_mtx); 5667 5668 return (dev != NULL); 5669} 5670 5671/* 5672 * Add a vdev for use by the L2ARC. By this point the spa has already 5673 * validated the vdev and opened it. 5674 */ 5675void 5676l2arc_add_vdev(spa_t *spa, vdev_t *vd) 5677{ 5678 l2arc_dev_t *adddev; 5679 5680 ASSERT(!l2arc_vdev_present(vd)); 5681 5682 vdev_ashift_optimize(vd); 5683 5684 /* 5685 * Create a new l2arc device entry. 5686 */ 5687 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5688 adddev->l2ad_spa = spa; 5689 adddev->l2ad_vdev = vd; 5690 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5691 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5692 adddev->l2ad_hand = adddev->l2ad_start; 5693 adddev->l2ad_evict = adddev->l2ad_start; 5694 adddev->l2ad_first = B_TRUE; 5695 adddev->l2ad_writing = B_FALSE; 5696 5697 /* 5698 * This is a list of all ARC buffers that are still valid on the 5699 * device. 5700 */ 5701 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5702 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5703 offsetof(arc_buf_hdr_t, b_l2node)); 5704 5705 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5706 5707 /* 5708 * Add device to global list 5709 */ 5710 mutex_enter(&l2arc_dev_mtx); 5711 list_insert_head(l2arc_dev_list, adddev); 5712 atomic_inc_64(&l2arc_ndev); 5713 mutex_exit(&l2arc_dev_mtx); 5714} 5715 5716/* 5717 * Remove a vdev from the L2ARC. 5718 */ 5719void 5720l2arc_remove_vdev(vdev_t *vd) 5721{ 5722 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5723 5724 /* 5725 * Find the device by vdev 5726 */ 5727 mutex_enter(&l2arc_dev_mtx); 5728 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5729 nextdev = list_next(l2arc_dev_list, dev); 5730 if (vd == dev->l2ad_vdev) { 5731 remdev = dev; 5732 break; 5733 } 5734 } 5735 ASSERT(remdev != NULL); 5736 5737 /* 5738 * Remove device from global list 5739 */ 5740 list_remove(l2arc_dev_list, remdev); 5741 l2arc_dev_last = NULL; /* may have been invalidated */ 5742 atomic_dec_64(&l2arc_ndev); 5743 mutex_exit(&l2arc_dev_mtx); 5744 5745 /* 5746 * Clear all buflists and ARC references. L2ARC device flush. 5747 */ 5748 l2arc_evict(remdev, 0, B_TRUE); 5749 list_destroy(remdev->l2ad_buflist); 5750 kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5751 kmem_free(remdev, sizeof (l2arc_dev_t)); 5752} 5753 5754void 5755l2arc_init(void) 5756{ 5757 l2arc_thread_exit = 0; 5758 l2arc_ndev = 0; 5759 l2arc_writes_sent = 0; 5760 l2arc_writes_done = 0; 5761 5762 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5763 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5764 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5765 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5766 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5767 5768 l2arc_dev_list = &L2ARC_dev_list; 5769 l2arc_free_on_write = &L2ARC_free_on_write; 5770 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5771 offsetof(l2arc_dev_t, l2ad_node)); 5772 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5773 offsetof(l2arc_data_free_t, l2df_list_node)); 5774} 5775 5776void 5777l2arc_fini(void) 5778{ 5779 /* 5780 * This is called from dmu_fini(), which is called from spa_fini(); 5781 * Because of this, we can assume that all l2arc devices have 5782 * already been removed when the pools themselves were removed. 5783 */ 5784 5785 l2arc_do_free_on_write(); 5786 5787 mutex_destroy(&l2arc_feed_thr_lock); 5788 cv_destroy(&l2arc_feed_thr_cv); 5789 mutex_destroy(&l2arc_dev_mtx); 5790 mutex_destroy(&l2arc_buflist_mtx); 5791 mutex_destroy(&l2arc_free_on_write_mtx); 5792 5793 list_destroy(l2arc_dev_list); 5794 list_destroy(l2arc_free_on_write); 5795} 5796 5797void 5798l2arc_start(void) 5799{ 5800 if (!(spa_mode_global & FWRITE)) 5801 return; 5802 5803 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5804 TS_RUN, minclsyspri); 5805} 5806 5807void 5808l2arc_stop(void) 5809{ 5810 if (!(spa_mode_global & FWRITE)) 5811 return; 5812 5813 mutex_enter(&l2arc_feed_thr_lock); 5814 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5815 l2arc_thread_exit = 1; 5816 while (l2arc_thread_exit != 0) 5817 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5818 mutex_exit(&l2arc_feed_thr_lock); 5819} 5820