arc.c revision 288519
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) 2012, Joyent, Inc. All rights reserved. 24 * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 26 * Copyright 2014 Nexenta Systems, Inc. All rights reserved. 27 */ 28 29/* 30 * DVA-based Adjustable Replacement Cache 31 * 32 * While much of the theory of operation used here is 33 * based on the self-tuning, low overhead replacement cache 34 * presented by Megiddo and Modha at FAST 2003, there are some 35 * significant differences: 36 * 37 * 1. The Megiddo and Modha model assumes any page is evictable. 38 * Pages in its cache cannot be "locked" into memory. This makes 39 * the eviction algorithm simple: evict the last page in the list. 40 * This also make the performance characteristics easy to reason 41 * about. Our cache is not so simple. At any given moment, some 42 * subset of the blocks in the cache are un-evictable because we 43 * have handed out a reference to them. Blocks are only evictable 44 * when there are no external references active. This makes 45 * eviction far more problematic: we choose to evict the evictable 46 * blocks that are the "lowest" in the list. 47 * 48 * There are times when it is not possible to evict the requested 49 * space. In these circumstances we are unable to adjust the cache 50 * size. To prevent the cache growing unbounded at these times we 51 * implement a "cache throttle" that slows the flow of new data 52 * into the cache until we can make space available. 53 * 54 * 2. The Megiddo and Modha model assumes a fixed cache size. 55 * Pages are evicted when the cache is full and there is a cache 56 * miss. Our model has a variable sized cache. It grows with 57 * high use, but also tries to react to memory pressure from the 58 * operating system: decreasing its size when system memory is 59 * tight. 60 * 61 * 3. The Megiddo and Modha model assumes a fixed page size. All 62 * elements of the cache are therefore exactly the same size. So 63 * when adjusting the cache size following a cache miss, its simply 64 * a matter of choosing a single page to evict. In our model, we 65 * have variable sized cache blocks (rangeing from 512 bytes to 66 * 128K bytes). We therefore choose a set of blocks to evict to make 67 * space for a cache miss that approximates as closely as possible 68 * the space used by the new block. 69 * 70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 71 * by N. Megiddo & D. Modha, FAST 2003 72 */ 73 74/* 75 * The locking model: 76 * 77 * A new reference to a cache buffer can be obtained in two 78 * ways: 1) via a hash table lookup using the DVA as a key, 79 * or 2) via one of the ARC lists. The arc_read() interface 80 * uses method 1, while the internal arc algorithms for 81 * adjusting the cache use method 2. We therefore provide two 82 * types of locks: 1) the hash table lock array, and 2) the 83 * arc list locks. 84 * 85 * Buffers do not have their own mutexs, rather they rely on the 86 * hash table mutexs for the bulk of their protection (i.e. most 87 * fields in the arc_buf_hdr_t are protected by these mutexs). 88 * 89 * buf_hash_find() returns the appropriate mutex (held) when it 90 * locates the requested buffer in the hash table. It returns 91 * NULL for the mutex if the buffer was not in the table. 92 * 93 * buf_hash_remove() expects the appropriate hash mutex to be 94 * already held before it is invoked. 95 * 96 * Each arc state also has a mutex which is used to protect the 97 * buffer list associated with the state. When attempting to 98 * obtain a hash table lock while holding an arc list lock you 99 * must use: mutex_tryenter() to avoid deadlock. Also note that 100 * the active state mutex must be held before the ghost state mutex. 101 * 102 * Arc buffers may have an associated eviction callback function. 103 * This function will be invoked prior to removing the buffer (e.g. 104 * in arc_do_user_evicts()). Note however that the data associated 105 * with the buffer may be evicted prior to the callback. The callback 106 * must be made with *no locks held* (to prevent deadlock). Additionally, 107 * the users of callbacks must ensure that their private data is 108 * protected from simultaneous callbacks from arc_clear_callback() 109 * and arc_do_user_evicts(). 110 * 111 * Note that the majority of the performance stats are manipulated 112 * with atomic operations. 113 * 114 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following: 115 * 116 * - L2ARC buflist creation 117 * - L2ARC buflist eviction 118 * - L2ARC write completion, which walks L2ARC buflists 119 * - ARC header destruction, as it removes from L2ARC buflists 120 * - ARC header release, as it removes from L2ARC buflists 121 */ 122 123#include <sys/spa.h> 124#include <sys/zio.h> 125#include <sys/zio_compress.h> 126#include <sys/zfs_context.h> 127#include <sys/arc.h> 128#include <sys/refcount.h> 129#include <sys/vdev.h> 130#include <sys/vdev_impl.h> 131#include <sys/dsl_pool.h> 132#ifdef _KERNEL 133#include <sys/dnlc.h> 134#endif 135#include <sys/callb.h> 136#include <sys/kstat.h> 137#include <sys/trim_map.h> 138#include <zfs_fletcher.h> 139#include <sys/sdt.h> 140 141#include <vm/vm_pageout.h> 142#include <machine/vmparam.h> 143 144#ifdef illumos 145#ifndef _KERNEL 146/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 147boolean_t arc_watch = B_FALSE; 148int arc_procfd; 149#endif 150#endif /* illumos */ 151 152static kmutex_t arc_reclaim_thr_lock; 153static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 154static uint8_t arc_thread_exit; 155 156#define ARC_REDUCE_DNLC_PERCENT 3 157uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 158 159typedef enum arc_reclaim_strategy { 160 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 161 ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 162} arc_reclaim_strategy_t; 163 164/* 165 * The number of iterations through arc_evict_*() before we 166 * drop & reacquire the lock. 167 */ 168int arc_evict_iterations = 100; 169 170/* number of seconds before growing cache again */ 171static int arc_grow_retry = 60; 172 173/* shift of arc_c for calculating both min and max arc_p */ 174static int arc_p_min_shift = 4; 175 176/* log2(fraction of arc to reclaim) */ 177static int arc_shrink_shift = 5; 178 179/* 180 * minimum lifespan of a prefetch block in clock ticks 181 * (initialized in arc_init()) 182 */ 183static int arc_min_prefetch_lifespan; 184 185/* 186 * If this percent of memory is free, don't throttle. 187 */ 188int arc_lotsfree_percent = 10; 189 190static int arc_dead; 191extern int zfs_prefetch_disable; 192 193/* 194 * The arc has filled available memory and has now warmed up. 195 */ 196static boolean_t arc_warm; 197 198uint64_t zfs_arc_max; 199uint64_t zfs_arc_min; 200uint64_t zfs_arc_meta_limit = 0; 201uint64_t zfs_arc_meta_min = 0; 202int zfs_arc_grow_retry = 0; 203int zfs_arc_shrink_shift = 0; 204int zfs_arc_p_min_shift = 0; 205int zfs_disable_dup_eviction = 0; 206uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 207u_int zfs_arc_free_target = 0; 208 209static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 210static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 211 212#ifdef _KERNEL 213static void 214arc_free_target_init(void *unused __unused) 215{ 216 217 zfs_arc_free_target = vm_pageout_wakeup_thresh; 218} 219SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 220 arc_free_target_init, NULL); 221 222TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max); 223TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 224TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 225TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 226TUNABLE_QUAD("vfs.zfs.arc_average_blocksize", &zfs_arc_average_blocksize); 227TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 228SYSCTL_DECL(_vfs_zfs); 229SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 230 "Maximum ARC size"); 231SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 232 "Minimum ARC size"); 233SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 234 &zfs_arc_average_blocksize, 0, 235 "ARC average blocksize"); 236SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 237 &arc_shrink_shift, 0, 238 "log2(fraction of arc to reclaim)"); 239 240/* 241 * We don't have a tunable for arc_free_target due to the dependency on 242 * pagedaemon initialisation. 243 */ 244SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 245 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 246 sysctl_vfs_zfs_arc_free_target, "IU", 247 "Desired number of free pages below which ARC triggers reclaim"); 248 249static int 250sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 251{ 252 u_int val; 253 int err; 254 255 val = zfs_arc_free_target; 256 err = sysctl_handle_int(oidp, &val, 0, req); 257 if (err != 0 || req->newptr == NULL) 258 return (err); 259 260 if (val < minfree) 261 return (EINVAL); 262 if (val > cnt.v_page_count) 263 return (EINVAL); 264 265 zfs_arc_free_target = val; 266 267 return (0); 268} 269 270/* 271 * Must be declared here, before the definition of corresponding kstat 272 * macro which uses the same names will confuse the compiler. 273 */ 274SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 275 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 276 sysctl_vfs_zfs_arc_meta_limit, "QU", 277 "ARC metadata limit"); 278#endif 279 280/* 281 * Note that buffers can be in one of 6 states: 282 * ARC_anon - anonymous (discussed below) 283 * ARC_mru - recently used, currently cached 284 * ARC_mru_ghost - recentely used, no longer in cache 285 * ARC_mfu - frequently used, currently cached 286 * ARC_mfu_ghost - frequently used, no longer in cache 287 * ARC_l2c_only - exists in L2ARC but not other states 288 * When there are no active references to the buffer, they are 289 * are linked onto a list in one of these arc states. These are 290 * the only buffers that can be evicted or deleted. Within each 291 * state there are multiple lists, one for meta-data and one for 292 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 293 * etc.) is tracked separately so that it can be managed more 294 * explicitly: favored over data, limited explicitly. 295 * 296 * Anonymous buffers are buffers that are not associated with 297 * a DVA. These are buffers that hold dirty block copies 298 * before they are written to stable storage. By definition, 299 * they are "ref'd" and are considered part of arc_mru 300 * that cannot be freed. Generally, they will aquire a DVA 301 * as they are written and migrate onto the arc_mru list. 302 * 303 * The ARC_l2c_only state is for buffers that are in the second 304 * level ARC but no longer in any of the ARC_m* lists. The second 305 * level ARC itself may also contain buffers that are in any of 306 * the ARC_m* states - meaning that a buffer can exist in two 307 * places. The reason for the ARC_l2c_only state is to keep the 308 * buffer header in the hash table, so that reads that hit the 309 * second level ARC benefit from these fast lookups. 310 */ 311 312#define ARCS_LOCK_PAD CACHE_LINE_SIZE 313struct arcs_lock { 314 kmutex_t arcs_lock; 315#ifdef _KERNEL 316 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 317#endif 318}; 319 320/* 321 * must be power of two for mask use to work 322 * 323 */ 324#define ARC_BUFC_NUMDATALISTS 16 325#define ARC_BUFC_NUMMETADATALISTS 16 326#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 327 328typedef struct arc_state { 329 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 330 uint64_t arcs_size; /* total amount of data in this state */ 331 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 332 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 333} arc_state_t; 334 335#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 336 337/* The 6 states: */ 338static arc_state_t ARC_anon; 339static arc_state_t ARC_mru; 340static arc_state_t ARC_mru_ghost; 341static arc_state_t ARC_mfu; 342static arc_state_t ARC_mfu_ghost; 343static arc_state_t ARC_l2c_only; 344 345typedef struct arc_stats { 346 kstat_named_t arcstat_hits; 347 kstat_named_t arcstat_misses; 348 kstat_named_t arcstat_demand_data_hits; 349 kstat_named_t arcstat_demand_data_misses; 350 kstat_named_t arcstat_demand_metadata_hits; 351 kstat_named_t arcstat_demand_metadata_misses; 352 kstat_named_t arcstat_prefetch_data_hits; 353 kstat_named_t arcstat_prefetch_data_misses; 354 kstat_named_t arcstat_prefetch_metadata_hits; 355 kstat_named_t arcstat_prefetch_metadata_misses; 356 kstat_named_t arcstat_mru_hits; 357 kstat_named_t arcstat_mru_ghost_hits; 358 kstat_named_t arcstat_mfu_hits; 359 kstat_named_t arcstat_mfu_ghost_hits; 360 kstat_named_t arcstat_allocated; 361 kstat_named_t arcstat_deleted; 362 kstat_named_t arcstat_stolen; 363 kstat_named_t arcstat_recycle_miss; 364 /* 365 * Number of buffers that could not be evicted because the hash lock 366 * was held by another thread. The lock may not necessarily be held 367 * by something using the same buffer, since hash locks are shared 368 * by multiple buffers. 369 */ 370 kstat_named_t arcstat_mutex_miss; 371 /* 372 * Number of buffers skipped because they have I/O in progress, are 373 * indrect prefetch buffers that have not lived long enough, or are 374 * not from the spa we're trying to evict from. 375 */ 376 kstat_named_t arcstat_evict_skip; 377 kstat_named_t arcstat_evict_l2_cached; 378 kstat_named_t arcstat_evict_l2_eligible; 379 kstat_named_t arcstat_evict_l2_ineligible; 380 kstat_named_t arcstat_hash_elements; 381 kstat_named_t arcstat_hash_elements_max; 382 kstat_named_t arcstat_hash_collisions; 383 kstat_named_t arcstat_hash_chains; 384 kstat_named_t arcstat_hash_chain_max; 385 kstat_named_t arcstat_p; 386 kstat_named_t arcstat_c; 387 kstat_named_t arcstat_c_min; 388 kstat_named_t arcstat_c_max; 389 kstat_named_t arcstat_size; 390 kstat_named_t arcstat_hdr_size; 391 kstat_named_t arcstat_data_size; 392 kstat_named_t arcstat_other_size; 393 kstat_named_t arcstat_l2_hits; 394 kstat_named_t arcstat_l2_misses; 395 kstat_named_t arcstat_l2_feeds; 396 kstat_named_t arcstat_l2_rw_clash; 397 kstat_named_t arcstat_l2_read_bytes; 398 kstat_named_t arcstat_l2_write_bytes; 399 kstat_named_t arcstat_l2_writes_sent; 400 kstat_named_t arcstat_l2_writes_done; 401 kstat_named_t arcstat_l2_writes_error; 402 kstat_named_t arcstat_l2_writes_hdr_miss; 403 kstat_named_t arcstat_l2_evict_lock_retry; 404 kstat_named_t arcstat_l2_evict_reading; 405 kstat_named_t arcstat_l2_free_on_write; 406 kstat_named_t arcstat_l2_cdata_free_on_write; 407 kstat_named_t arcstat_l2_abort_lowmem; 408 kstat_named_t arcstat_l2_cksum_bad; 409 kstat_named_t arcstat_l2_io_error; 410 kstat_named_t arcstat_l2_size; 411 kstat_named_t arcstat_l2_asize; 412 kstat_named_t arcstat_l2_hdr_size; 413 kstat_named_t arcstat_l2_compress_successes; 414 kstat_named_t arcstat_l2_compress_zeros; 415 kstat_named_t arcstat_l2_compress_failures; 416 kstat_named_t arcstat_l2_write_trylock_fail; 417 kstat_named_t arcstat_l2_write_passed_headroom; 418 kstat_named_t arcstat_l2_write_spa_mismatch; 419 kstat_named_t arcstat_l2_write_in_l2; 420 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 421 kstat_named_t arcstat_l2_write_not_cacheable; 422 kstat_named_t arcstat_l2_write_full; 423 kstat_named_t arcstat_l2_write_buffer_iter; 424 kstat_named_t arcstat_l2_write_pios; 425 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 426 kstat_named_t arcstat_l2_write_buffer_list_iter; 427 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 428 kstat_named_t arcstat_memory_throttle_count; 429 kstat_named_t arcstat_duplicate_buffers; 430 kstat_named_t arcstat_duplicate_buffers_size; 431 kstat_named_t arcstat_duplicate_reads; 432 kstat_named_t arcstat_meta_used; 433 kstat_named_t arcstat_meta_limit; 434 kstat_named_t arcstat_meta_max; 435 kstat_named_t arcstat_meta_min; 436} arc_stats_t; 437 438static arc_stats_t arc_stats = { 439 { "hits", KSTAT_DATA_UINT64 }, 440 { "misses", KSTAT_DATA_UINT64 }, 441 { "demand_data_hits", KSTAT_DATA_UINT64 }, 442 { "demand_data_misses", KSTAT_DATA_UINT64 }, 443 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 444 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 445 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 446 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 447 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 448 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 449 { "mru_hits", KSTAT_DATA_UINT64 }, 450 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 451 { "mfu_hits", KSTAT_DATA_UINT64 }, 452 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 453 { "allocated", KSTAT_DATA_UINT64 }, 454 { "deleted", KSTAT_DATA_UINT64 }, 455 { "stolen", KSTAT_DATA_UINT64 }, 456 { "recycle_miss", KSTAT_DATA_UINT64 }, 457 { "mutex_miss", KSTAT_DATA_UINT64 }, 458 { "evict_skip", KSTAT_DATA_UINT64 }, 459 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 460 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 461 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 462 { "hash_elements", KSTAT_DATA_UINT64 }, 463 { "hash_elements_max", KSTAT_DATA_UINT64 }, 464 { "hash_collisions", KSTAT_DATA_UINT64 }, 465 { "hash_chains", KSTAT_DATA_UINT64 }, 466 { "hash_chain_max", KSTAT_DATA_UINT64 }, 467 { "p", KSTAT_DATA_UINT64 }, 468 { "c", KSTAT_DATA_UINT64 }, 469 { "c_min", KSTAT_DATA_UINT64 }, 470 { "c_max", KSTAT_DATA_UINT64 }, 471 { "size", KSTAT_DATA_UINT64 }, 472 { "hdr_size", KSTAT_DATA_UINT64 }, 473 { "data_size", KSTAT_DATA_UINT64 }, 474 { "other_size", KSTAT_DATA_UINT64 }, 475 { "l2_hits", KSTAT_DATA_UINT64 }, 476 { "l2_misses", KSTAT_DATA_UINT64 }, 477 { "l2_feeds", KSTAT_DATA_UINT64 }, 478 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 479 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 480 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 481 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 482 { "l2_writes_done", KSTAT_DATA_UINT64 }, 483 { "l2_writes_error", KSTAT_DATA_UINT64 }, 484 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 485 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 486 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 487 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 488 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, 489 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 490 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 491 { "l2_io_error", KSTAT_DATA_UINT64 }, 492 { "l2_size", KSTAT_DATA_UINT64 }, 493 { "l2_asize", KSTAT_DATA_UINT64 }, 494 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 495 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 496 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 497 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 498 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 499 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 500 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 501 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 502 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 503 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 504 { "l2_write_full", KSTAT_DATA_UINT64 }, 505 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 506 { "l2_write_pios", KSTAT_DATA_UINT64 }, 507 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 508 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 509 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 510 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 511 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 512 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 513 { "duplicate_reads", KSTAT_DATA_UINT64 }, 514 { "arc_meta_used", KSTAT_DATA_UINT64 }, 515 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 516 { "arc_meta_max", KSTAT_DATA_UINT64 }, 517 { "arc_meta_min", KSTAT_DATA_UINT64 } 518}; 519 520#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 521 522#define ARCSTAT_INCR(stat, val) \ 523 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 524 525#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 526#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 527 528#define ARCSTAT_MAX(stat, val) { \ 529 uint64_t m; \ 530 while ((val) > (m = arc_stats.stat.value.ui64) && \ 531 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 532 continue; \ 533} 534 535#define ARCSTAT_MAXSTAT(stat) \ 536 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 537 538/* 539 * We define a macro to allow ARC hits/misses to be easily broken down by 540 * two separate conditions, giving a total of four different subtypes for 541 * each of hits and misses (so eight statistics total). 542 */ 543#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 544 if (cond1) { \ 545 if (cond2) { \ 546 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 547 } else { \ 548 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 549 } \ 550 } else { \ 551 if (cond2) { \ 552 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 553 } else { \ 554 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 555 } \ 556 } 557 558kstat_t *arc_ksp; 559static arc_state_t *arc_anon; 560static arc_state_t *arc_mru; 561static arc_state_t *arc_mru_ghost; 562static arc_state_t *arc_mfu; 563static arc_state_t *arc_mfu_ghost; 564static arc_state_t *arc_l2c_only; 565 566/* 567 * There are several ARC variables that are critical to export as kstats -- 568 * but we don't want to have to grovel around in the kstat whenever we wish to 569 * manipulate them. For these variables, we therefore define them to be in 570 * terms of the statistic variable. This assures that we are not introducing 571 * the possibility of inconsistency by having shadow copies of the variables, 572 * while still allowing the code to be readable. 573 */ 574#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 575#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 576#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 577#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 578#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 579#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 580#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 581#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 582#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 583 584#define L2ARC_IS_VALID_COMPRESS(_c_) \ 585 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 586 587static int arc_no_grow; /* Don't try to grow cache size */ 588static uint64_t arc_tempreserve; 589static uint64_t arc_loaned_bytes; 590 591typedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 592 593typedef struct arc_callback arc_callback_t; 594 595struct arc_callback { 596 void *acb_private; 597 arc_done_func_t *acb_done; 598 arc_buf_t *acb_buf; 599 zio_t *acb_zio_dummy; 600 arc_callback_t *acb_next; 601}; 602 603typedef struct arc_write_callback arc_write_callback_t; 604 605struct arc_write_callback { 606 void *awcb_private; 607 arc_done_func_t *awcb_ready; 608 arc_done_func_t *awcb_physdone; 609 arc_done_func_t *awcb_done; 610 arc_buf_t *awcb_buf; 611}; 612 613struct arc_buf_hdr { 614 /* protected by hash lock */ 615 dva_t b_dva; 616 uint64_t b_birth; 617 uint64_t b_cksum0; 618 619 kmutex_t b_freeze_lock; 620 zio_cksum_t *b_freeze_cksum; 621 void *b_thawed; 622 623 arc_buf_hdr_t *b_hash_next; 624 arc_buf_t *b_buf; 625 arc_flags_t b_flags; 626 uint32_t b_datacnt; 627 628 arc_callback_t *b_acb; 629 kcondvar_t b_cv; 630 631 /* immutable */ 632 arc_buf_contents_t b_type; 633 uint64_t b_size; 634 uint64_t b_spa; 635 636 /* protected by arc state mutex */ 637 arc_state_t *b_state; 638 list_node_t b_arc_node; 639 640 /* updated atomically */ 641 clock_t b_arc_access; 642 643 /* self protecting */ 644 refcount_t b_refcnt; 645 646 l2arc_buf_hdr_t *b_l2hdr; 647 list_node_t b_l2node; 648}; 649 650#ifdef _KERNEL 651static int 652sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 653{ 654 uint64_t val; 655 int err; 656 657 val = arc_meta_limit; 658 err = sysctl_handle_64(oidp, &val, 0, req); 659 if (err != 0 || req->newptr == NULL) 660 return (err); 661 662 if (val <= 0 || val > arc_c_max) 663 return (EINVAL); 664 665 arc_meta_limit = val; 666 return (0); 667} 668#endif 669 670static arc_buf_t *arc_eviction_list; 671static kmutex_t arc_eviction_mtx; 672static arc_buf_hdr_t arc_eviction_hdr; 673 674#define GHOST_STATE(state) \ 675 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 676 (state) == arc_l2c_only) 677 678#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 679#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 680#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 681#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 682#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ) 683#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE) 684#define HDR_FREE_IN_PROGRESS(hdr) \ 685 ((hdr)->b_flags & ARC_FLAG_FREE_IN_PROGRESS) 686#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 687#define HDR_L2_READING(hdr) \ 688 ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS && \ 689 (hdr)->b_l2hdr != NULL) 690#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 691#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 692#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 693 694/* 695 * Other sizes 696 */ 697 698#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 699#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 700 701/* 702 * Hash table routines 703 */ 704 705#define HT_LOCK_PAD CACHE_LINE_SIZE 706 707struct ht_lock { 708 kmutex_t ht_lock; 709#ifdef _KERNEL 710 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 711#endif 712}; 713 714#define BUF_LOCKS 256 715typedef struct buf_hash_table { 716 uint64_t ht_mask; 717 arc_buf_hdr_t **ht_table; 718 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 719} buf_hash_table_t; 720 721static buf_hash_table_t buf_hash_table; 722 723#define BUF_HASH_INDEX(spa, dva, birth) \ 724 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 725#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 726#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 727#define HDR_LOCK(hdr) \ 728 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 729 730uint64_t zfs_crc64_table[256]; 731 732/* 733 * Level 2 ARC 734 */ 735 736#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 737#define L2ARC_HEADROOM 2 /* num of writes */ 738/* 739 * If we discover during ARC scan any buffers to be compressed, we boost 740 * our headroom for the next scanning cycle by this percentage multiple. 741 */ 742#define L2ARC_HEADROOM_BOOST 200 743#define L2ARC_FEED_SECS 1 /* caching interval secs */ 744#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 745 746#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 747#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 748 749/* L2ARC Performance Tunables */ 750uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 751uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 752uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 753uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 754uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 755uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 756boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 757boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 758boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 759 760SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 761 &l2arc_write_max, 0, "max write size"); 762SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 763 &l2arc_write_boost, 0, "extra write during warmup"); 764SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 765 &l2arc_headroom, 0, "number of dev writes"); 766SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 767 &l2arc_feed_secs, 0, "interval seconds"); 768SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 769 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 770 771SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 772 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 773SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 774 &l2arc_feed_again, 0, "turbo warmup"); 775SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 776 &l2arc_norw, 0, "no reads during writes"); 777 778SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 779 &ARC_anon.arcs_size, 0, "size of anonymous state"); 780SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 781 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 782SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 783 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 784 785SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 786 &ARC_mru.arcs_size, 0, "size of mru state"); 787SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 788 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 789SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 790 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 791 792SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 793 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 794SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 795 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 796 "size of metadata in mru ghost state"); 797SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 798 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 799 "size of data in mru ghost state"); 800 801SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 802 &ARC_mfu.arcs_size, 0, "size of mfu state"); 803SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 804 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 805SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 806 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 807 808SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 809 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 810SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 811 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 812 "size of metadata in mfu ghost state"); 813SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 814 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 815 "size of data in mfu ghost state"); 816 817SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 818 &ARC_l2c_only.arcs_size, 0, "size of mru state"); 819 820/* 821 * L2ARC Internals 822 */ 823typedef struct l2arc_dev { 824 vdev_t *l2ad_vdev; /* vdev */ 825 spa_t *l2ad_spa; /* spa */ 826 uint64_t l2ad_hand; /* next write location */ 827 uint64_t l2ad_start; /* first addr on device */ 828 uint64_t l2ad_end; /* last addr on device */ 829 uint64_t l2ad_evict; /* last addr eviction reached */ 830 boolean_t l2ad_first; /* first sweep through */ 831 boolean_t l2ad_writing; /* currently writing */ 832 list_t *l2ad_buflist; /* buffer list */ 833 list_node_t l2ad_node; /* device list node */ 834} l2arc_dev_t; 835 836static list_t L2ARC_dev_list; /* device list */ 837static list_t *l2arc_dev_list; /* device list pointer */ 838static kmutex_t l2arc_dev_mtx; /* device list mutex */ 839static l2arc_dev_t *l2arc_dev_last; /* last device used */ 840static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 841static list_t L2ARC_free_on_write; /* free after write buf list */ 842static list_t *l2arc_free_on_write; /* free after write list ptr */ 843static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 844static uint64_t l2arc_ndev; /* number of devices */ 845 846typedef struct l2arc_read_callback { 847 arc_buf_t *l2rcb_buf; /* read buffer */ 848 spa_t *l2rcb_spa; /* spa */ 849 blkptr_t l2rcb_bp; /* original blkptr */ 850 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 851 int l2rcb_flags; /* original flags */ 852 enum zio_compress l2rcb_compress; /* applied compress */ 853} l2arc_read_callback_t; 854 855typedef struct l2arc_write_callback { 856 l2arc_dev_t *l2wcb_dev; /* device info */ 857 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 858} l2arc_write_callback_t; 859 860struct l2arc_buf_hdr { 861 /* protected by arc_buf_hdr mutex */ 862 l2arc_dev_t *b_dev; /* L2ARC device */ 863 uint64_t b_daddr; /* disk address, offset byte */ 864 /* compression applied to buffer data */ 865 enum zio_compress b_compress; 866 /* real alloc'd buffer size depending on b_compress applied */ 867 int b_asize; 868 /* temporary buffer holder for in-flight compressed data */ 869 void *b_tmp_cdata; 870}; 871 872typedef struct l2arc_data_free { 873 /* protected by l2arc_free_on_write_mtx */ 874 void *l2df_data; 875 size_t l2df_size; 876 void (*l2df_func)(void *, size_t); 877 list_node_t l2df_list_node; 878} l2arc_data_free_t; 879 880static kmutex_t l2arc_feed_thr_lock; 881static kcondvar_t l2arc_feed_thr_cv; 882static uint8_t l2arc_thread_exit; 883 884static void arc_get_data_buf(arc_buf_t *); 885static void arc_access(arc_buf_hdr_t *, kmutex_t *); 886static int arc_evict_needed(arc_buf_contents_t); 887static void arc_evict_ghost(arc_state_t *, uint64_t, int64_t); 888static void arc_buf_watch(arc_buf_t *); 889 890static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 891static void l2arc_read_done(zio_t *); 892static void l2arc_hdr_stat_add(void); 893static void l2arc_hdr_stat_remove(void); 894 895static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *); 896static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress); 897static void l2arc_release_cdata_buf(arc_buf_hdr_t *); 898 899static uint64_t 900buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 901{ 902 uint8_t *vdva = (uint8_t *)dva; 903 uint64_t crc = -1ULL; 904 int i; 905 906 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 907 908 for (i = 0; i < sizeof (dva_t); i++) 909 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 910 911 crc ^= (spa>>8) ^ birth; 912 913 return (crc); 914} 915 916#define BUF_EMPTY(buf) \ 917 ((buf)->b_dva.dva_word[0] == 0 && \ 918 (buf)->b_dva.dva_word[1] == 0 && \ 919 (buf)->b_cksum0 == 0) 920 921#define BUF_EQUAL(spa, dva, birth, buf) \ 922 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 923 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 924 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 925 926static void 927buf_discard_identity(arc_buf_hdr_t *hdr) 928{ 929 hdr->b_dva.dva_word[0] = 0; 930 hdr->b_dva.dva_word[1] = 0; 931 hdr->b_birth = 0; 932 hdr->b_cksum0 = 0; 933} 934 935static arc_buf_hdr_t * 936buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 937{ 938 const dva_t *dva = BP_IDENTITY(bp); 939 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 940 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 941 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 942 arc_buf_hdr_t *hdr; 943 944 mutex_enter(hash_lock); 945 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 946 hdr = hdr->b_hash_next) { 947 if (BUF_EQUAL(spa, dva, birth, hdr)) { 948 *lockp = hash_lock; 949 return (hdr); 950 } 951 } 952 mutex_exit(hash_lock); 953 *lockp = NULL; 954 return (NULL); 955} 956 957/* 958 * Insert an entry into the hash table. If there is already an element 959 * equal to elem in the hash table, then the already existing element 960 * will be returned and the new element will not be inserted. 961 * Otherwise returns NULL. 962 */ 963static arc_buf_hdr_t * 964buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 965{ 966 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 967 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 968 arc_buf_hdr_t *fhdr; 969 uint32_t i; 970 971 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 972 ASSERT(hdr->b_birth != 0); 973 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 974 *lockp = hash_lock; 975 mutex_enter(hash_lock); 976 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 977 fhdr = fhdr->b_hash_next, i++) { 978 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 979 return (fhdr); 980 } 981 982 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 983 buf_hash_table.ht_table[idx] = hdr; 984 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 985 986 /* collect some hash table performance data */ 987 if (i > 0) { 988 ARCSTAT_BUMP(arcstat_hash_collisions); 989 if (i == 1) 990 ARCSTAT_BUMP(arcstat_hash_chains); 991 992 ARCSTAT_MAX(arcstat_hash_chain_max, i); 993 } 994 995 ARCSTAT_BUMP(arcstat_hash_elements); 996 ARCSTAT_MAXSTAT(arcstat_hash_elements); 997 998 return (NULL); 999} 1000 1001static void 1002buf_hash_remove(arc_buf_hdr_t *hdr) 1003{ 1004 arc_buf_hdr_t *fhdr, **hdrp; 1005 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1006 1007 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1008 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1009 1010 hdrp = &buf_hash_table.ht_table[idx]; 1011 while ((fhdr = *hdrp) != hdr) { 1012 ASSERT(fhdr != NULL); 1013 hdrp = &fhdr->b_hash_next; 1014 } 1015 *hdrp = hdr->b_hash_next; 1016 hdr->b_hash_next = NULL; 1017 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; 1018 1019 /* collect some hash table performance data */ 1020 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1021 1022 if (buf_hash_table.ht_table[idx] && 1023 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1024 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1025} 1026 1027/* 1028 * Global data structures and functions for the buf kmem cache. 1029 */ 1030static kmem_cache_t *hdr_cache; 1031static kmem_cache_t *buf_cache; 1032 1033static void 1034buf_fini(void) 1035{ 1036 int i; 1037 1038 kmem_free(buf_hash_table.ht_table, 1039 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1040 for (i = 0; i < BUF_LOCKS; i++) 1041 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1042 kmem_cache_destroy(hdr_cache); 1043 kmem_cache_destroy(buf_cache); 1044} 1045 1046/* 1047 * Constructor callback - called when the cache is empty 1048 * and a new buf is requested. 1049 */ 1050/* ARGSUSED */ 1051static int 1052hdr_cons(void *vbuf, void *unused, int kmflag) 1053{ 1054 arc_buf_hdr_t *hdr = vbuf; 1055 1056 bzero(hdr, sizeof (arc_buf_hdr_t)); 1057 refcount_create(&hdr->b_refcnt); 1058 cv_init(&hdr->b_cv, NULL, CV_DEFAULT, NULL); 1059 mutex_init(&hdr->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1060 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1061 1062 return (0); 1063} 1064 1065/* ARGSUSED */ 1066static int 1067buf_cons(void *vbuf, void *unused, int kmflag) 1068{ 1069 arc_buf_t *buf = vbuf; 1070 1071 bzero(buf, sizeof (arc_buf_t)); 1072 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1073 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1074 1075 return (0); 1076} 1077 1078/* 1079 * Destructor callback - called when a cached buf is 1080 * no longer required. 1081 */ 1082/* ARGSUSED */ 1083static void 1084hdr_dest(void *vbuf, void *unused) 1085{ 1086 arc_buf_hdr_t *hdr = vbuf; 1087 1088 ASSERT(BUF_EMPTY(hdr)); 1089 refcount_destroy(&hdr->b_refcnt); 1090 cv_destroy(&hdr->b_cv); 1091 mutex_destroy(&hdr->b_freeze_lock); 1092 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1093} 1094 1095/* ARGSUSED */ 1096static void 1097buf_dest(void *vbuf, void *unused) 1098{ 1099 arc_buf_t *buf = vbuf; 1100 1101 mutex_destroy(&buf->b_evict_lock); 1102 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1103} 1104 1105/* 1106 * Reclaim callback -- invoked when memory is low. 1107 */ 1108/* ARGSUSED */ 1109static void 1110hdr_recl(void *unused) 1111{ 1112 dprintf("hdr_recl called\n"); 1113 /* 1114 * umem calls the reclaim func when we destroy the buf cache, 1115 * which is after we do arc_fini(). 1116 */ 1117 if (!arc_dead) 1118 cv_signal(&arc_reclaim_thr_cv); 1119} 1120 1121static void 1122buf_init(void) 1123{ 1124 uint64_t *ct; 1125 uint64_t hsize = 1ULL << 12; 1126 int i, j; 1127 1128 /* 1129 * The hash table is big enough to fill all of physical memory 1130 * with an average block size of zfs_arc_average_blocksize (default 8K). 1131 * By default, the table will take up 1132 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1133 */ 1134 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1135 hsize <<= 1; 1136retry: 1137 buf_hash_table.ht_mask = hsize - 1; 1138 buf_hash_table.ht_table = 1139 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1140 if (buf_hash_table.ht_table == NULL) { 1141 ASSERT(hsize > (1ULL << 8)); 1142 hsize >>= 1; 1143 goto retry; 1144 } 1145 1146 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1147 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1148 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1149 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1150 1151 for (i = 0; i < 256; i++) 1152 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1153 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1154 1155 for (i = 0; i < BUF_LOCKS; i++) { 1156 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1157 NULL, MUTEX_DEFAULT, NULL); 1158 } 1159} 1160 1161#define ARC_MINTIME (hz>>4) /* 62 ms */ 1162 1163static void 1164arc_cksum_verify(arc_buf_t *buf) 1165{ 1166 zio_cksum_t zc; 1167 1168 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1169 return; 1170 1171 mutex_enter(&buf->b_hdr->b_freeze_lock); 1172 if (buf->b_hdr->b_freeze_cksum == NULL || 1173 (buf->b_hdr->b_flags & ARC_FLAG_IO_ERROR)) { 1174 mutex_exit(&buf->b_hdr->b_freeze_lock); 1175 return; 1176 } 1177 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1178 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1179 panic("buffer modified while frozen!"); 1180 mutex_exit(&buf->b_hdr->b_freeze_lock); 1181} 1182 1183static int 1184arc_cksum_equal(arc_buf_t *buf) 1185{ 1186 zio_cksum_t zc; 1187 int equal; 1188 1189 mutex_enter(&buf->b_hdr->b_freeze_lock); 1190 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1191 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1192 mutex_exit(&buf->b_hdr->b_freeze_lock); 1193 1194 return (equal); 1195} 1196 1197static void 1198arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1199{ 1200 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1201 return; 1202 1203 mutex_enter(&buf->b_hdr->b_freeze_lock); 1204 if (buf->b_hdr->b_freeze_cksum != NULL) { 1205 mutex_exit(&buf->b_hdr->b_freeze_lock); 1206 return; 1207 } 1208 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1209 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1210 buf->b_hdr->b_freeze_cksum); 1211 mutex_exit(&buf->b_hdr->b_freeze_lock); 1212#ifdef illumos 1213 arc_buf_watch(buf); 1214#endif /* illumos */ 1215} 1216 1217#ifdef illumos 1218#ifndef _KERNEL 1219typedef struct procctl { 1220 long cmd; 1221 prwatch_t prwatch; 1222} procctl_t; 1223#endif 1224 1225/* ARGSUSED */ 1226static void 1227arc_buf_unwatch(arc_buf_t *buf) 1228{ 1229#ifndef _KERNEL 1230 if (arc_watch) { 1231 int result; 1232 procctl_t ctl; 1233 ctl.cmd = PCWATCH; 1234 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1235 ctl.prwatch.pr_size = 0; 1236 ctl.prwatch.pr_wflags = 0; 1237 result = write(arc_procfd, &ctl, sizeof (ctl)); 1238 ASSERT3U(result, ==, sizeof (ctl)); 1239 } 1240#endif 1241} 1242 1243/* ARGSUSED */ 1244static void 1245arc_buf_watch(arc_buf_t *buf) 1246{ 1247#ifndef _KERNEL 1248 if (arc_watch) { 1249 int result; 1250 procctl_t ctl; 1251 ctl.cmd = PCWATCH; 1252 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1253 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1254 ctl.prwatch.pr_wflags = WA_WRITE; 1255 result = write(arc_procfd, &ctl, sizeof (ctl)); 1256 ASSERT3U(result, ==, sizeof (ctl)); 1257 } 1258#endif 1259} 1260#endif /* illumos */ 1261 1262void 1263arc_buf_thaw(arc_buf_t *buf) 1264{ 1265 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1266 if (buf->b_hdr->b_state != arc_anon) 1267 panic("modifying non-anon buffer!"); 1268 if (buf->b_hdr->b_flags & ARC_FLAG_IO_IN_PROGRESS) 1269 panic("modifying buffer while i/o in progress!"); 1270 arc_cksum_verify(buf); 1271 } 1272 1273 mutex_enter(&buf->b_hdr->b_freeze_lock); 1274 if (buf->b_hdr->b_freeze_cksum != NULL) { 1275 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1276 buf->b_hdr->b_freeze_cksum = NULL; 1277 } 1278 1279 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1280 if (buf->b_hdr->b_thawed) 1281 kmem_free(buf->b_hdr->b_thawed, 1); 1282 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1283 } 1284 1285 mutex_exit(&buf->b_hdr->b_freeze_lock); 1286 1287#ifdef illumos 1288 arc_buf_unwatch(buf); 1289#endif /* illumos */ 1290} 1291 1292void 1293arc_buf_freeze(arc_buf_t *buf) 1294{ 1295 kmutex_t *hash_lock; 1296 1297 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1298 return; 1299 1300 hash_lock = HDR_LOCK(buf->b_hdr); 1301 mutex_enter(hash_lock); 1302 1303 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1304 buf->b_hdr->b_state == arc_anon); 1305 arc_cksum_compute(buf, B_FALSE); 1306 mutex_exit(hash_lock); 1307 1308} 1309 1310static void 1311get_buf_info(arc_buf_hdr_t *hdr, arc_state_t *state, list_t **list, kmutex_t **lock) 1312{ 1313 uint64_t buf_hashid = buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1314 1315 if (hdr->b_type == ARC_BUFC_METADATA) 1316 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1317 else { 1318 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1319 buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1320 } 1321 1322 *list = &state->arcs_lists[buf_hashid]; 1323 *lock = ARCS_LOCK(state, buf_hashid); 1324} 1325 1326 1327static void 1328add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1329{ 1330 ASSERT(MUTEX_HELD(hash_lock)); 1331 1332 if ((refcount_add(&hdr->b_refcnt, tag) == 1) && 1333 (hdr->b_state != arc_anon)) { 1334 uint64_t delta = hdr->b_size * hdr->b_datacnt; 1335 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 1336 list_t *list; 1337 kmutex_t *lock; 1338 1339 get_buf_info(hdr, hdr->b_state, &list, &lock); 1340 ASSERT(!MUTEX_HELD(lock)); 1341 mutex_enter(lock); 1342 ASSERT(list_link_active(&hdr->b_arc_node)); 1343 list_remove(list, hdr); 1344 if (GHOST_STATE(hdr->b_state)) { 1345 ASSERT0(hdr->b_datacnt); 1346 ASSERT3P(hdr->b_buf, ==, NULL); 1347 delta = hdr->b_size; 1348 } 1349 ASSERT(delta > 0); 1350 ASSERT3U(*size, >=, delta); 1351 atomic_add_64(size, -delta); 1352 mutex_exit(lock); 1353 /* remove the prefetch flag if we get a reference */ 1354 if (hdr->b_flags & ARC_FLAG_PREFETCH) 1355 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 1356 } 1357} 1358 1359static int 1360remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1361{ 1362 int cnt; 1363 arc_state_t *state = hdr->b_state; 1364 1365 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1366 ASSERT(!GHOST_STATE(state)); 1367 1368 if (((cnt = refcount_remove(&hdr->b_refcnt, tag)) == 0) && 1369 (state != arc_anon)) { 1370 uint64_t *size = &state->arcs_lsize[hdr->b_type]; 1371 list_t *list; 1372 kmutex_t *lock; 1373 1374 get_buf_info(hdr, state, &list, &lock); 1375 ASSERT(!MUTEX_HELD(lock)); 1376 mutex_enter(lock); 1377 ASSERT(!list_link_active(&hdr->b_arc_node)); 1378 list_insert_head(list, hdr); 1379 ASSERT(hdr->b_datacnt > 0); 1380 atomic_add_64(size, hdr->b_size * hdr->b_datacnt); 1381 mutex_exit(lock); 1382 } 1383 return (cnt); 1384} 1385 1386/* 1387 * Move the supplied buffer to the indicated state. The mutex 1388 * for the buffer must be held by the caller. 1389 */ 1390static void 1391arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 1392 kmutex_t *hash_lock) 1393{ 1394 arc_state_t *old_state = hdr->b_state; 1395 int64_t refcnt = refcount_count(&hdr->b_refcnt); 1396 uint64_t from_delta, to_delta; 1397 list_t *list; 1398 kmutex_t *lock; 1399 1400 ASSERT(MUTEX_HELD(hash_lock)); 1401 ASSERT3P(new_state, !=, old_state); 1402 ASSERT(refcnt == 0 || hdr->b_datacnt > 0); 1403 ASSERT(hdr->b_datacnt == 0 || !GHOST_STATE(new_state)); 1404 ASSERT(hdr->b_datacnt <= 1 || old_state != arc_anon); 1405 1406 from_delta = to_delta = hdr->b_datacnt * hdr->b_size; 1407 1408 /* 1409 * If this buffer is evictable, transfer it from the 1410 * old state list to the new state list. 1411 */ 1412 if (refcnt == 0) { 1413 if (old_state != arc_anon) { 1414 int use_mutex; 1415 uint64_t *size = &old_state->arcs_lsize[hdr->b_type]; 1416 1417 get_buf_info(hdr, old_state, &list, &lock); 1418 use_mutex = !MUTEX_HELD(lock); 1419 if (use_mutex) 1420 mutex_enter(lock); 1421 1422 ASSERT(list_link_active(&hdr->b_arc_node)); 1423 list_remove(list, hdr); 1424 1425 /* 1426 * If prefetching out of the ghost cache, 1427 * we will have a non-zero datacnt. 1428 */ 1429 if (GHOST_STATE(old_state) && hdr->b_datacnt == 0) { 1430 /* ghost elements have a ghost size */ 1431 ASSERT(hdr->b_buf == NULL); 1432 from_delta = hdr->b_size; 1433 } 1434 ASSERT3U(*size, >=, from_delta); 1435 atomic_add_64(size, -from_delta); 1436 1437 if (use_mutex) 1438 mutex_exit(lock); 1439 } 1440 if (new_state != arc_anon) { 1441 int use_mutex; 1442 uint64_t *size = &new_state->arcs_lsize[hdr->b_type]; 1443 1444 get_buf_info(hdr, new_state, &list, &lock); 1445 use_mutex = !MUTEX_HELD(lock); 1446 if (use_mutex) 1447 mutex_enter(lock); 1448 1449 list_insert_head(list, hdr); 1450 1451 /* ghost elements have a ghost size */ 1452 if (GHOST_STATE(new_state)) { 1453 ASSERT(hdr->b_datacnt == 0); 1454 ASSERT(hdr->b_buf == NULL); 1455 to_delta = hdr->b_size; 1456 } 1457 atomic_add_64(size, to_delta); 1458 1459 if (use_mutex) 1460 mutex_exit(lock); 1461 } 1462 } 1463 1464 ASSERT(!BUF_EMPTY(hdr)); 1465 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 1466 buf_hash_remove(hdr); 1467 1468 /* adjust state sizes */ 1469 if (to_delta) 1470 atomic_add_64(&new_state->arcs_size, to_delta); 1471 if (from_delta) { 1472 ASSERT3U(old_state->arcs_size, >=, from_delta); 1473 atomic_add_64(&old_state->arcs_size, -from_delta); 1474 } 1475 hdr->b_state = new_state; 1476 1477 /* adjust l2arc hdr stats */ 1478 if (new_state == arc_l2c_only) 1479 l2arc_hdr_stat_add(); 1480 else if (old_state == arc_l2c_only) 1481 l2arc_hdr_stat_remove(); 1482} 1483 1484void 1485arc_space_consume(uint64_t space, arc_space_type_t type) 1486{ 1487 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1488 1489 switch (type) { 1490 case ARC_SPACE_DATA: 1491 ARCSTAT_INCR(arcstat_data_size, space); 1492 break; 1493 case ARC_SPACE_OTHER: 1494 ARCSTAT_INCR(arcstat_other_size, space); 1495 break; 1496 case ARC_SPACE_HDRS: 1497 ARCSTAT_INCR(arcstat_hdr_size, space); 1498 break; 1499 case ARC_SPACE_L2HDRS: 1500 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1501 break; 1502 } 1503 1504 ARCSTAT_INCR(arcstat_meta_used, space); 1505 atomic_add_64(&arc_size, space); 1506} 1507 1508void 1509arc_space_return(uint64_t space, arc_space_type_t type) 1510{ 1511 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1512 1513 switch (type) { 1514 case ARC_SPACE_DATA: 1515 ARCSTAT_INCR(arcstat_data_size, -space); 1516 break; 1517 case ARC_SPACE_OTHER: 1518 ARCSTAT_INCR(arcstat_other_size, -space); 1519 break; 1520 case ARC_SPACE_HDRS: 1521 ARCSTAT_INCR(arcstat_hdr_size, -space); 1522 break; 1523 case ARC_SPACE_L2HDRS: 1524 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1525 break; 1526 } 1527 1528 ASSERT(arc_meta_used >= space); 1529 if (arc_meta_max < arc_meta_used) 1530 arc_meta_max = arc_meta_used; 1531 ARCSTAT_INCR(arcstat_meta_used, -space); 1532 ASSERT(arc_size >= space); 1533 atomic_add_64(&arc_size, -space); 1534} 1535 1536arc_buf_t * 1537arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1538{ 1539 arc_buf_hdr_t *hdr; 1540 arc_buf_t *buf; 1541 1542 ASSERT3U(size, >, 0); 1543 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1544 ASSERT(BUF_EMPTY(hdr)); 1545 hdr->b_size = size; 1546 hdr->b_type = type; 1547 hdr->b_spa = spa_load_guid(spa); 1548 hdr->b_state = arc_anon; 1549 hdr->b_arc_access = 0; 1550 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1551 buf->b_hdr = hdr; 1552 buf->b_data = NULL; 1553 buf->b_efunc = NULL; 1554 buf->b_private = NULL; 1555 buf->b_next = NULL; 1556 hdr->b_buf = buf; 1557 arc_get_data_buf(buf); 1558 hdr->b_datacnt = 1; 1559 hdr->b_flags = 0; 1560 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1561 (void) refcount_add(&hdr->b_refcnt, tag); 1562 1563 return (buf); 1564} 1565 1566static char *arc_onloan_tag = "onloan"; 1567 1568/* 1569 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1570 * flight data by arc_tempreserve_space() until they are "returned". Loaned 1571 * buffers must be returned to the arc before they can be used by the DMU or 1572 * freed. 1573 */ 1574arc_buf_t * 1575arc_loan_buf(spa_t *spa, int size) 1576{ 1577 arc_buf_t *buf; 1578 1579 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1580 1581 atomic_add_64(&arc_loaned_bytes, size); 1582 return (buf); 1583} 1584 1585/* 1586 * Return a loaned arc buffer to the arc. 1587 */ 1588void 1589arc_return_buf(arc_buf_t *buf, void *tag) 1590{ 1591 arc_buf_hdr_t *hdr = buf->b_hdr; 1592 1593 ASSERT(buf->b_data != NULL); 1594 (void) refcount_add(&hdr->b_refcnt, tag); 1595 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1596 1597 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1598} 1599 1600/* Detach an arc_buf from a dbuf (tag) */ 1601void 1602arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1603{ 1604 arc_buf_hdr_t *hdr; 1605 1606 ASSERT(buf->b_data != NULL); 1607 hdr = buf->b_hdr; 1608 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1609 (void) refcount_remove(&hdr->b_refcnt, tag); 1610 buf->b_efunc = NULL; 1611 buf->b_private = NULL; 1612 1613 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1614} 1615 1616static arc_buf_t * 1617arc_buf_clone(arc_buf_t *from) 1618{ 1619 arc_buf_t *buf; 1620 arc_buf_hdr_t *hdr = from->b_hdr; 1621 uint64_t size = hdr->b_size; 1622 1623 ASSERT(hdr->b_state != arc_anon); 1624 1625 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1626 buf->b_hdr = hdr; 1627 buf->b_data = NULL; 1628 buf->b_efunc = NULL; 1629 buf->b_private = NULL; 1630 buf->b_next = hdr->b_buf; 1631 hdr->b_buf = buf; 1632 arc_get_data_buf(buf); 1633 bcopy(from->b_data, buf->b_data, size); 1634 1635 /* 1636 * This buffer already exists in the arc so create a duplicate 1637 * copy for the caller. If the buffer is associated with user data 1638 * then track the size and number of duplicates. These stats will be 1639 * updated as duplicate buffers are created and destroyed. 1640 */ 1641 if (hdr->b_type == ARC_BUFC_DATA) { 1642 ARCSTAT_BUMP(arcstat_duplicate_buffers); 1643 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1644 } 1645 hdr->b_datacnt += 1; 1646 return (buf); 1647} 1648 1649void 1650arc_buf_add_ref(arc_buf_t *buf, void* tag) 1651{ 1652 arc_buf_hdr_t *hdr; 1653 kmutex_t *hash_lock; 1654 1655 /* 1656 * Check to see if this buffer is evicted. Callers 1657 * must verify b_data != NULL to know if the add_ref 1658 * was successful. 1659 */ 1660 mutex_enter(&buf->b_evict_lock); 1661 if (buf->b_data == NULL) { 1662 mutex_exit(&buf->b_evict_lock); 1663 return; 1664 } 1665 hash_lock = HDR_LOCK(buf->b_hdr); 1666 mutex_enter(hash_lock); 1667 hdr = buf->b_hdr; 1668 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1669 mutex_exit(&buf->b_evict_lock); 1670 1671 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1672 add_reference(hdr, hash_lock, tag); 1673 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1674 arc_access(hdr, hash_lock); 1675 mutex_exit(hash_lock); 1676 ARCSTAT_BUMP(arcstat_hits); 1677 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 1678 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1679 data, metadata, hits); 1680} 1681 1682static void 1683arc_buf_free_on_write(void *data, size_t size, 1684 void (*free_func)(void *, size_t)) 1685{ 1686 l2arc_data_free_t *df; 1687 1688 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1689 df->l2df_data = data; 1690 df->l2df_size = size; 1691 df->l2df_func = free_func; 1692 mutex_enter(&l2arc_free_on_write_mtx); 1693 list_insert_head(l2arc_free_on_write, df); 1694 mutex_exit(&l2arc_free_on_write_mtx); 1695} 1696 1697/* 1698 * Free the arc data buffer. If it is an l2arc write in progress, 1699 * the buffer is placed on l2arc_free_on_write to be freed later. 1700 */ 1701static void 1702arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1703{ 1704 arc_buf_hdr_t *hdr = buf->b_hdr; 1705 1706 if (HDR_L2_WRITING(hdr)) { 1707 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 1708 ARCSTAT_BUMP(arcstat_l2_free_on_write); 1709 } else { 1710 free_func(buf->b_data, hdr->b_size); 1711 } 1712} 1713 1714/* 1715 * Free up buf->b_data and if 'remove' is set, then pull the 1716 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 1717 */ 1718static void 1719arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 1720{ 1721 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1722 1723 ASSERT(MUTEX_HELD(&l2arc_buflist_mtx)); 1724 1725 if (l2hdr->b_tmp_cdata == NULL) 1726 return; 1727 1728 ASSERT(HDR_L2_WRITING(hdr)); 1729 arc_buf_free_on_write(l2hdr->b_tmp_cdata, hdr->b_size, 1730 zio_data_buf_free); 1731 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 1732 l2hdr->b_tmp_cdata = NULL; 1733} 1734 1735static void 1736arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 1737{ 1738 arc_buf_t **bufp; 1739 1740 /* free up data associated with the buf */ 1741 if (buf->b_data) { 1742 arc_state_t *state = buf->b_hdr->b_state; 1743 uint64_t size = buf->b_hdr->b_size; 1744 arc_buf_contents_t type = buf->b_hdr->b_type; 1745 1746 arc_cksum_verify(buf); 1747#ifdef illumos 1748 arc_buf_unwatch(buf); 1749#endif /* illumos */ 1750 1751 if (!recycle) { 1752 if (type == ARC_BUFC_METADATA) { 1753 arc_buf_data_free(buf, zio_buf_free); 1754 arc_space_return(size, ARC_SPACE_DATA); 1755 } else { 1756 ASSERT(type == ARC_BUFC_DATA); 1757 arc_buf_data_free(buf, zio_data_buf_free); 1758 ARCSTAT_INCR(arcstat_data_size, -size); 1759 atomic_add_64(&arc_size, -size); 1760 } 1761 } 1762 if (list_link_active(&buf->b_hdr->b_arc_node)) { 1763 uint64_t *cnt = &state->arcs_lsize[type]; 1764 1765 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1766 ASSERT(state != arc_anon); 1767 1768 ASSERT3U(*cnt, >=, size); 1769 atomic_add_64(cnt, -size); 1770 } 1771 ASSERT3U(state->arcs_size, >=, size); 1772 atomic_add_64(&state->arcs_size, -size); 1773 buf->b_data = NULL; 1774 1775 /* 1776 * If we're destroying a duplicate buffer make sure 1777 * that the appropriate statistics are updated. 1778 */ 1779 if (buf->b_hdr->b_datacnt > 1 && 1780 buf->b_hdr->b_type == ARC_BUFC_DATA) { 1781 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1782 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1783 } 1784 ASSERT(buf->b_hdr->b_datacnt > 0); 1785 buf->b_hdr->b_datacnt -= 1; 1786 } 1787 1788 /* only remove the buf if requested */ 1789 if (!remove) 1790 return; 1791 1792 /* remove the buf from the hdr list */ 1793 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1794 continue; 1795 *bufp = buf->b_next; 1796 buf->b_next = NULL; 1797 1798 ASSERT(buf->b_efunc == NULL); 1799 1800 /* clean up the buf */ 1801 buf->b_hdr = NULL; 1802 kmem_cache_free(buf_cache, buf); 1803} 1804 1805static void 1806arc_hdr_destroy(arc_buf_hdr_t *hdr) 1807{ 1808 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1809 ASSERT3P(hdr->b_state, ==, arc_anon); 1810 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1811 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1812 1813 if (l2hdr != NULL) { 1814 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1815 /* 1816 * To prevent arc_free() and l2arc_evict() from 1817 * attempting to free the same buffer at the same time, 1818 * a FREE_IN_PROGRESS flag is given to arc_free() to 1819 * give it priority. l2arc_evict() can't destroy this 1820 * header while we are waiting on l2arc_buflist_mtx. 1821 * 1822 * The hdr may be removed from l2ad_buflist before we 1823 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1824 */ 1825 if (!buflist_held) { 1826 mutex_enter(&l2arc_buflist_mtx); 1827 l2hdr = hdr->b_l2hdr; 1828 } 1829 1830 if (l2hdr != NULL) { 1831 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1832 l2hdr->b_asize, 0); 1833 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1834 arc_buf_l2_cdata_free(hdr); 1835 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1836 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1837 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 1838 -l2hdr->b_asize, 0, 0); 1839 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1840 if (hdr->b_state == arc_l2c_only) 1841 l2arc_hdr_stat_remove(); 1842 hdr->b_l2hdr = NULL; 1843 } 1844 1845 if (!buflist_held) 1846 mutex_exit(&l2arc_buflist_mtx); 1847 } 1848 1849 if (!BUF_EMPTY(hdr)) { 1850 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1851 buf_discard_identity(hdr); 1852 } 1853 while (hdr->b_buf) { 1854 arc_buf_t *buf = hdr->b_buf; 1855 1856 if (buf->b_efunc) { 1857 mutex_enter(&arc_eviction_mtx); 1858 mutex_enter(&buf->b_evict_lock); 1859 ASSERT(buf->b_hdr != NULL); 1860 arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1861 hdr->b_buf = buf->b_next; 1862 buf->b_hdr = &arc_eviction_hdr; 1863 buf->b_next = arc_eviction_list; 1864 arc_eviction_list = buf; 1865 mutex_exit(&buf->b_evict_lock); 1866 mutex_exit(&arc_eviction_mtx); 1867 } else { 1868 arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1869 } 1870 } 1871 if (hdr->b_freeze_cksum != NULL) { 1872 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1873 hdr->b_freeze_cksum = NULL; 1874 } 1875 if (hdr->b_thawed) { 1876 kmem_free(hdr->b_thawed, 1); 1877 hdr->b_thawed = NULL; 1878 } 1879 1880 ASSERT(!list_link_active(&hdr->b_arc_node)); 1881 ASSERT3P(hdr->b_hash_next, ==, NULL); 1882 ASSERT3P(hdr->b_acb, ==, NULL); 1883 kmem_cache_free(hdr_cache, hdr); 1884} 1885 1886void 1887arc_buf_free(arc_buf_t *buf, void *tag) 1888{ 1889 arc_buf_hdr_t *hdr = buf->b_hdr; 1890 int hashed = hdr->b_state != arc_anon; 1891 1892 ASSERT(buf->b_efunc == NULL); 1893 ASSERT(buf->b_data != NULL); 1894 1895 if (hashed) { 1896 kmutex_t *hash_lock = HDR_LOCK(hdr); 1897 1898 mutex_enter(hash_lock); 1899 hdr = buf->b_hdr; 1900 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1901 1902 (void) remove_reference(hdr, hash_lock, tag); 1903 if (hdr->b_datacnt > 1) { 1904 arc_buf_destroy(buf, FALSE, TRUE); 1905 } else { 1906 ASSERT(buf == hdr->b_buf); 1907 ASSERT(buf->b_efunc == NULL); 1908 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 1909 } 1910 mutex_exit(hash_lock); 1911 } else if (HDR_IO_IN_PROGRESS(hdr)) { 1912 int destroy_hdr; 1913 /* 1914 * We are in the middle of an async write. Don't destroy 1915 * this buffer unless the write completes before we finish 1916 * decrementing the reference count. 1917 */ 1918 mutex_enter(&arc_eviction_mtx); 1919 (void) remove_reference(hdr, NULL, tag); 1920 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1921 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1922 mutex_exit(&arc_eviction_mtx); 1923 if (destroy_hdr) 1924 arc_hdr_destroy(hdr); 1925 } else { 1926 if (remove_reference(hdr, NULL, tag) > 0) 1927 arc_buf_destroy(buf, FALSE, TRUE); 1928 else 1929 arc_hdr_destroy(hdr); 1930 } 1931} 1932 1933boolean_t 1934arc_buf_remove_ref(arc_buf_t *buf, void* tag) 1935{ 1936 arc_buf_hdr_t *hdr = buf->b_hdr; 1937 kmutex_t *hash_lock = HDR_LOCK(hdr); 1938 boolean_t no_callback = (buf->b_efunc == NULL); 1939 1940 if (hdr->b_state == arc_anon) { 1941 ASSERT(hdr->b_datacnt == 1); 1942 arc_buf_free(buf, tag); 1943 return (no_callback); 1944 } 1945 1946 mutex_enter(hash_lock); 1947 hdr = buf->b_hdr; 1948 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1949 ASSERT(hdr->b_state != arc_anon); 1950 ASSERT(buf->b_data != NULL); 1951 1952 (void) remove_reference(hdr, hash_lock, tag); 1953 if (hdr->b_datacnt > 1) { 1954 if (no_callback) 1955 arc_buf_destroy(buf, FALSE, TRUE); 1956 } else if (no_callback) { 1957 ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1958 ASSERT(buf->b_efunc == NULL); 1959 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 1960 } 1961 ASSERT(no_callback || hdr->b_datacnt > 1 || 1962 refcount_is_zero(&hdr->b_refcnt)); 1963 mutex_exit(hash_lock); 1964 return (no_callback); 1965} 1966 1967int 1968arc_buf_size(arc_buf_t *buf) 1969{ 1970 return (buf->b_hdr->b_size); 1971} 1972 1973/* 1974 * Called from the DMU to determine if the current buffer should be 1975 * evicted. In order to ensure proper locking, the eviction must be initiated 1976 * from the DMU. Return true if the buffer is associated with user data and 1977 * duplicate buffers still exist. 1978 */ 1979boolean_t 1980arc_buf_eviction_needed(arc_buf_t *buf) 1981{ 1982 arc_buf_hdr_t *hdr; 1983 boolean_t evict_needed = B_FALSE; 1984 1985 if (zfs_disable_dup_eviction) 1986 return (B_FALSE); 1987 1988 mutex_enter(&buf->b_evict_lock); 1989 hdr = buf->b_hdr; 1990 if (hdr == NULL) { 1991 /* 1992 * We are in arc_do_user_evicts(); let that function 1993 * perform the eviction. 1994 */ 1995 ASSERT(buf->b_data == NULL); 1996 mutex_exit(&buf->b_evict_lock); 1997 return (B_FALSE); 1998 } else if (buf->b_data == NULL) { 1999 /* 2000 * We have already been added to the arc eviction list; 2001 * recommend eviction. 2002 */ 2003 ASSERT3P(hdr, ==, &arc_eviction_hdr); 2004 mutex_exit(&buf->b_evict_lock); 2005 return (B_TRUE); 2006 } 2007 2008 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 2009 evict_needed = B_TRUE; 2010 2011 mutex_exit(&buf->b_evict_lock); 2012 return (evict_needed); 2013} 2014 2015/* 2016 * Evict buffers from list until we've removed the specified number of 2017 * bytes. Move the removed buffers to the appropriate evict state. 2018 * If the recycle flag is set, then attempt to "recycle" a buffer: 2019 * - look for a buffer to evict that is `bytes' long. 2020 * - return the data block from this buffer rather than freeing it. 2021 * This flag is used by callers that are trying to make space for a 2022 * new buffer in a full arc cache. 2023 * 2024 * This function makes a "best effort". It skips over any buffers 2025 * it can't get a hash_lock on, and so may not catch all candidates. 2026 * It may also return without evicting as much space as requested. 2027 */ 2028static void * 2029arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 2030 arc_buf_contents_t type) 2031{ 2032 arc_state_t *evicted_state; 2033 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 2034 int64_t bytes_remaining; 2035 arc_buf_hdr_t *hdr, *hdr_prev = NULL; 2036 list_t *evicted_list, *list, *evicted_list_start, *list_start; 2037 kmutex_t *lock, *evicted_lock; 2038 kmutex_t *hash_lock; 2039 boolean_t have_lock; 2040 void *stolen = NULL; 2041 arc_buf_hdr_t marker = { 0 }; 2042 int count = 0; 2043 static int evict_metadata_offset, evict_data_offset; 2044 int i, idx, offset, list_count, lists; 2045 2046 ASSERT(state == arc_mru || state == arc_mfu); 2047 2048 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2049 2050 /* 2051 * Decide which "type" (data vs metadata) to recycle from. 2052 * 2053 * If we are over the metadata limit, recycle from metadata. 2054 * If we are under the metadata minimum, recycle from data. 2055 * Otherwise, recycle from whichever type has the oldest (least 2056 * recently accessed) header. This is not yet implemented. 2057 */ 2058 if (recycle) { 2059 arc_buf_contents_t realtype; 2060 if (state->arcs_lsize[ARC_BUFC_DATA] == 0) { 2061 realtype = ARC_BUFC_METADATA; 2062 } else if (state->arcs_lsize[ARC_BUFC_METADATA] == 0) { 2063 realtype = ARC_BUFC_DATA; 2064 } else if (arc_meta_used >= arc_meta_limit) { 2065 realtype = ARC_BUFC_METADATA; 2066 } else if (arc_meta_used <= arc_meta_min) { 2067 realtype = ARC_BUFC_DATA; 2068 } else { 2069#ifdef illumos 2070 if (data_hdr->b_arc_access < 2071 metadata_hdr->b_arc_access) { 2072 realtype = ARC_BUFC_DATA; 2073 } else { 2074 realtype = ARC_BUFC_METADATA; 2075 } 2076#else 2077 /* TODO */ 2078 realtype = type; 2079#endif 2080 } 2081 if (realtype != type) { 2082 /* 2083 * If we want to evict from a different list, 2084 * we can not recycle, because DATA vs METADATA 2085 * buffers are segregated into different kmem 2086 * caches (and vmem arenas). 2087 */ 2088 type = realtype; 2089 recycle = B_FALSE; 2090 } 2091 } 2092 2093 if (type == ARC_BUFC_METADATA) { 2094 offset = 0; 2095 list_count = ARC_BUFC_NUMMETADATALISTS; 2096 list_start = &state->arcs_lists[0]; 2097 evicted_list_start = &evicted_state->arcs_lists[0]; 2098 idx = evict_metadata_offset; 2099 } else { 2100 offset = ARC_BUFC_NUMMETADATALISTS; 2101 list_start = &state->arcs_lists[offset]; 2102 evicted_list_start = &evicted_state->arcs_lists[offset]; 2103 list_count = ARC_BUFC_NUMDATALISTS; 2104 idx = evict_data_offset; 2105 } 2106 bytes_remaining = evicted_state->arcs_lsize[type]; 2107 lists = 0; 2108 2109evict_start: 2110 list = &list_start[idx]; 2111 evicted_list = &evicted_list_start[idx]; 2112 lock = ARCS_LOCK(state, (offset + idx)); 2113 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 2114 2115 mutex_enter(lock); 2116 mutex_enter(evicted_lock); 2117 2118 for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2119 hdr_prev = list_prev(list, hdr); 2120 bytes_remaining -= (hdr->b_size * hdr->b_datacnt); 2121 /* prefetch buffers have a minimum lifespan */ 2122 if (HDR_IO_IN_PROGRESS(hdr) || 2123 (spa && hdr->b_spa != spa) || 2124 (hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT) && 2125 ddi_get_lbolt() - hdr->b_arc_access < 2126 arc_min_prefetch_lifespan)) { 2127 skipped++; 2128 continue; 2129 } 2130 /* "lookahead" for better eviction candidate */ 2131 if (recycle && hdr->b_size != bytes && 2132 hdr_prev && hdr_prev->b_size == bytes) 2133 continue; 2134 2135 /* ignore markers */ 2136 if (hdr->b_spa == 0) 2137 continue; 2138 2139 /* 2140 * It may take a long time to evict all the bufs requested. 2141 * To avoid blocking all arc activity, periodically drop 2142 * the arcs_mtx and give other threads a chance to run 2143 * before reacquiring the lock. 2144 * 2145 * If we are looking for a buffer to recycle, we are in 2146 * the hot code path, so don't sleep. 2147 */ 2148 if (!recycle && count++ > arc_evict_iterations) { 2149 list_insert_after(list, hdr, &marker); 2150 mutex_exit(evicted_lock); 2151 mutex_exit(lock); 2152 kpreempt(KPREEMPT_SYNC); 2153 mutex_enter(lock); 2154 mutex_enter(evicted_lock); 2155 hdr_prev = list_prev(list, &marker); 2156 list_remove(list, &marker); 2157 count = 0; 2158 continue; 2159 } 2160 2161 hash_lock = HDR_LOCK(hdr); 2162 have_lock = MUTEX_HELD(hash_lock); 2163 if (have_lock || mutex_tryenter(hash_lock)) { 2164 ASSERT0(refcount_count(&hdr->b_refcnt)); 2165 ASSERT(hdr->b_datacnt > 0); 2166 while (hdr->b_buf) { 2167 arc_buf_t *buf = hdr->b_buf; 2168 if (!mutex_tryenter(&buf->b_evict_lock)) { 2169 missed += 1; 2170 break; 2171 } 2172 if (buf->b_data) { 2173 bytes_evicted += hdr->b_size; 2174 if (recycle && hdr->b_type == type && 2175 hdr->b_size == bytes && 2176 !HDR_L2_WRITING(hdr)) { 2177 stolen = buf->b_data; 2178 recycle = FALSE; 2179 } 2180 } 2181 if (buf->b_efunc) { 2182 mutex_enter(&arc_eviction_mtx); 2183 arc_buf_destroy(buf, 2184 buf->b_data == stolen, FALSE); 2185 hdr->b_buf = buf->b_next; 2186 buf->b_hdr = &arc_eviction_hdr; 2187 buf->b_next = arc_eviction_list; 2188 arc_eviction_list = buf; 2189 mutex_exit(&arc_eviction_mtx); 2190 mutex_exit(&buf->b_evict_lock); 2191 } else { 2192 mutex_exit(&buf->b_evict_lock); 2193 arc_buf_destroy(buf, 2194 buf->b_data == stolen, TRUE); 2195 } 2196 } 2197 2198 if (hdr->b_l2hdr) { 2199 ARCSTAT_INCR(arcstat_evict_l2_cached, 2200 hdr->b_size); 2201 } else { 2202 if (l2arc_write_eligible(hdr->b_spa, hdr)) { 2203 ARCSTAT_INCR(arcstat_evict_l2_eligible, 2204 hdr->b_size); 2205 } else { 2206 ARCSTAT_INCR( 2207 arcstat_evict_l2_ineligible, 2208 hdr->b_size); 2209 } 2210 } 2211 2212 if (hdr->b_datacnt == 0) { 2213 arc_change_state(evicted_state, hdr, hash_lock); 2214 ASSERT(HDR_IN_HASH_TABLE(hdr)); 2215 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 2216 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 2217 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 2218 } 2219 if (!have_lock) 2220 mutex_exit(hash_lock); 2221 if (bytes >= 0 && bytes_evicted >= bytes) 2222 break; 2223 if (bytes_remaining > 0) { 2224 mutex_exit(evicted_lock); 2225 mutex_exit(lock); 2226 idx = ((idx + 1) & (list_count - 1)); 2227 lists++; 2228 goto evict_start; 2229 } 2230 } else { 2231 missed += 1; 2232 } 2233 } 2234 2235 mutex_exit(evicted_lock); 2236 mutex_exit(lock); 2237 2238 idx = ((idx + 1) & (list_count - 1)); 2239 lists++; 2240 2241 if (bytes_evicted < bytes) { 2242 if (lists < list_count) 2243 goto evict_start; 2244 else 2245 dprintf("only evicted %lld bytes from %x", 2246 (longlong_t)bytes_evicted, state); 2247 } 2248 if (type == ARC_BUFC_METADATA) 2249 evict_metadata_offset = idx; 2250 else 2251 evict_data_offset = idx; 2252 2253 if (skipped) 2254 ARCSTAT_INCR(arcstat_evict_skip, skipped); 2255 2256 if (missed) 2257 ARCSTAT_INCR(arcstat_mutex_miss, missed); 2258 2259 /* 2260 * Note: we have just evicted some data into the ghost state, 2261 * potentially putting the ghost size over the desired size. Rather 2262 * that evicting from the ghost list in this hot code path, leave 2263 * this chore to the arc_reclaim_thread(). 2264 */ 2265 2266 if (stolen) 2267 ARCSTAT_BUMP(arcstat_stolen); 2268 return (stolen); 2269} 2270 2271/* 2272 * Remove buffers from list until we've removed the specified number of 2273 * bytes. Destroy the buffers that are removed. 2274 */ 2275static void 2276arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2277{ 2278 arc_buf_hdr_t *hdr, *hdr_prev; 2279 arc_buf_hdr_t marker = { 0 }; 2280 list_t *list, *list_start; 2281 kmutex_t *hash_lock, *lock; 2282 uint64_t bytes_deleted = 0; 2283 uint64_t bufs_skipped = 0; 2284 int count = 0; 2285 static int evict_offset; 2286 int list_count, idx = evict_offset; 2287 int offset, lists = 0; 2288 2289 ASSERT(GHOST_STATE(state)); 2290 2291 /* 2292 * data lists come after metadata lists 2293 */ 2294 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2295 list_count = ARC_BUFC_NUMDATALISTS; 2296 offset = ARC_BUFC_NUMMETADATALISTS; 2297 2298evict_start: 2299 list = &list_start[idx]; 2300 lock = ARCS_LOCK(state, idx + offset); 2301 2302 mutex_enter(lock); 2303 for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2304 hdr_prev = list_prev(list, hdr); 2305 if (hdr->b_type > ARC_BUFC_NUMTYPES) 2306 panic("invalid hdr=%p", (void *)hdr); 2307 if (spa && hdr->b_spa != spa) 2308 continue; 2309 2310 /* ignore markers */ 2311 if (hdr->b_spa == 0) 2312 continue; 2313 2314 hash_lock = HDR_LOCK(hdr); 2315 /* caller may be trying to modify this buffer, skip it */ 2316 if (MUTEX_HELD(hash_lock)) 2317 continue; 2318 2319 /* 2320 * It may take a long time to evict all the bufs requested. 2321 * To avoid blocking all arc activity, periodically drop 2322 * the arcs_mtx and give other threads a chance to run 2323 * before reacquiring the lock. 2324 */ 2325 if (count++ > arc_evict_iterations) { 2326 list_insert_after(list, hdr, &marker); 2327 mutex_exit(lock); 2328 kpreempt(KPREEMPT_SYNC); 2329 mutex_enter(lock); 2330 hdr_prev = list_prev(list, &marker); 2331 list_remove(list, &marker); 2332 count = 0; 2333 continue; 2334 } 2335 if (mutex_tryenter(hash_lock)) { 2336 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2337 ASSERT(hdr->b_buf == NULL); 2338 ARCSTAT_BUMP(arcstat_deleted); 2339 bytes_deleted += hdr->b_size; 2340 2341 if (hdr->b_l2hdr != NULL) { 2342 /* 2343 * This buffer is cached on the 2nd Level ARC; 2344 * don't destroy the header. 2345 */ 2346 arc_change_state(arc_l2c_only, hdr, hash_lock); 2347 mutex_exit(hash_lock); 2348 } else { 2349 arc_change_state(arc_anon, hdr, hash_lock); 2350 mutex_exit(hash_lock); 2351 arc_hdr_destroy(hdr); 2352 } 2353 2354 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 2355 if (bytes >= 0 && bytes_deleted >= bytes) 2356 break; 2357 } else if (bytes < 0) { 2358 /* 2359 * Insert a list marker and then wait for the 2360 * hash lock to become available. Once its 2361 * available, restart from where we left off. 2362 */ 2363 list_insert_after(list, hdr, &marker); 2364 mutex_exit(lock); 2365 mutex_enter(hash_lock); 2366 mutex_exit(hash_lock); 2367 mutex_enter(lock); 2368 hdr_prev = list_prev(list, &marker); 2369 list_remove(list, &marker); 2370 } else { 2371 bufs_skipped += 1; 2372 } 2373 2374 } 2375 mutex_exit(lock); 2376 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2377 lists++; 2378 2379 if (lists < list_count) 2380 goto evict_start; 2381 2382 evict_offset = idx; 2383 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2384 (bytes < 0 || bytes_deleted < bytes)) { 2385 list_start = &state->arcs_lists[0]; 2386 list_count = ARC_BUFC_NUMMETADATALISTS; 2387 offset = lists = 0; 2388 goto evict_start; 2389 } 2390 2391 if (bufs_skipped) { 2392 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2393 ASSERT(bytes >= 0); 2394 } 2395 2396 if (bytes_deleted < bytes) 2397 dprintf("only deleted %lld bytes from %p", 2398 (longlong_t)bytes_deleted, state); 2399} 2400 2401static void 2402arc_adjust(void) 2403{ 2404 int64_t adjustment, delta; 2405 2406 /* 2407 * Adjust MRU size 2408 */ 2409 2410 adjustment = MIN((int64_t)(arc_size - arc_c), 2411 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2412 arc_p)); 2413 2414 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2415 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2416 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2417 adjustment -= delta; 2418 } 2419 2420 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2421 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2422 (void) arc_evict(arc_mru, 0, delta, FALSE, 2423 ARC_BUFC_METADATA); 2424 } 2425 2426 /* 2427 * Adjust MFU size 2428 */ 2429 2430 adjustment = arc_size - arc_c; 2431 2432 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2433 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2434 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2435 adjustment -= delta; 2436 } 2437 2438 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2439 int64_t delta = MIN(adjustment, 2440 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2441 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2442 ARC_BUFC_METADATA); 2443 } 2444 2445 /* 2446 * Adjust ghost lists 2447 */ 2448 2449 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2450 2451 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2452 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2453 arc_evict_ghost(arc_mru_ghost, 0, delta); 2454 } 2455 2456 adjustment = 2457 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2458 2459 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2460 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2461 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2462 } 2463} 2464 2465static void 2466arc_do_user_evicts(void) 2467{ 2468 static arc_buf_t *tmp_arc_eviction_list; 2469 2470 /* 2471 * Move list over to avoid LOR 2472 */ 2473restart: 2474 mutex_enter(&arc_eviction_mtx); 2475 tmp_arc_eviction_list = arc_eviction_list; 2476 arc_eviction_list = NULL; 2477 mutex_exit(&arc_eviction_mtx); 2478 2479 while (tmp_arc_eviction_list != NULL) { 2480 arc_buf_t *buf = tmp_arc_eviction_list; 2481 tmp_arc_eviction_list = buf->b_next; 2482 mutex_enter(&buf->b_evict_lock); 2483 buf->b_hdr = NULL; 2484 mutex_exit(&buf->b_evict_lock); 2485 2486 if (buf->b_efunc != NULL) 2487 VERIFY0(buf->b_efunc(buf->b_private)); 2488 2489 buf->b_efunc = NULL; 2490 buf->b_private = NULL; 2491 kmem_cache_free(buf_cache, buf); 2492 } 2493 2494 if (arc_eviction_list != NULL) 2495 goto restart; 2496} 2497 2498/* 2499 * Flush all *evictable* data from the cache for the given spa. 2500 * NOTE: this will not touch "active" (i.e. referenced) data. 2501 */ 2502void 2503arc_flush(spa_t *spa) 2504{ 2505 uint64_t guid = 0; 2506 2507 if (spa) 2508 guid = spa_load_guid(spa); 2509 2510 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2511 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2512 if (spa) 2513 break; 2514 } 2515 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2516 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2517 if (spa) 2518 break; 2519 } 2520 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2521 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2522 if (spa) 2523 break; 2524 } 2525 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2526 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2527 if (spa) 2528 break; 2529 } 2530 2531 arc_evict_ghost(arc_mru_ghost, guid, -1); 2532 arc_evict_ghost(arc_mfu_ghost, guid, -1); 2533 2534 mutex_enter(&arc_reclaim_thr_lock); 2535 arc_do_user_evicts(); 2536 mutex_exit(&arc_reclaim_thr_lock); 2537 ASSERT(spa || arc_eviction_list == NULL); 2538} 2539 2540void 2541arc_shrink(void) 2542{ 2543 2544 if (arc_c > arc_c_min) { 2545 uint64_t to_free; 2546 2547 to_free = arc_c >> arc_shrink_shift; 2548 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 2549 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 2550 if (arc_c > arc_c_min + to_free) 2551 atomic_add_64(&arc_c, -to_free); 2552 else 2553 arc_c = arc_c_min; 2554 2555 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2556 if (arc_c > arc_size) 2557 arc_c = MAX(arc_size, arc_c_min); 2558 if (arc_p > arc_c) 2559 arc_p = (arc_c >> 1); 2560 2561 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 2562 arc_p); 2563 2564 ASSERT(arc_c >= arc_c_min); 2565 ASSERT((int64_t)arc_p >= 0); 2566 } 2567 2568 if (arc_size > arc_c) { 2569 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 2570 uint64_t, arc_c); 2571 arc_adjust(); 2572 } 2573} 2574 2575static int needfree = 0; 2576 2577static int 2578arc_reclaim_needed(void) 2579{ 2580 2581#ifdef _KERNEL 2582 2583 if (needfree) { 2584 DTRACE_PROBE(arc__reclaim_needfree); 2585 return (1); 2586 } 2587 2588 /* 2589 * Cooperate with pagedaemon when it's time for it to scan 2590 * and reclaim some pages. 2591 */ 2592 if (freemem < zfs_arc_free_target) { 2593 DTRACE_PROBE2(arc__reclaim_freemem, uint64_t, 2594 freemem, uint64_t, zfs_arc_free_target); 2595 return (1); 2596 } 2597 2598#ifdef sun 2599 /* 2600 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2601 */ 2602 extra = desfree; 2603 2604 /* 2605 * check that we're out of range of the pageout scanner. It starts to 2606 * schedule paging if freemem is less than lotsfree and needfree. 2607 * lotsfree is the high-water mark for pageout, and needfree is the 2608 * number of needed free pages. We add extra pages here to make sure 2609 * the scanner doesn't start up while we're freeing memory. 2610 */ 2611 if (freemem < lotsfree + needfree + extra) 2612 return (1); 2613 2614 /* 2615 * check to make sure that swapfs has enough space so that anon 2616 * reservations can still succeed. anon_resvmem() checks that the 2617 * availrmem is greater than swapfs_minfree, and the number of reserved 2618 * swap pages. We also add a bit of extra here just to prevent 2619 * circumstances from getting really dire. 2620 */ 2621 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2622 return (1); 2623 2624 /* 2625 * Check that we have enough availrmem that memory locking (e.g., via 2626 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 2627 * stores the number of pages that cannot be locked; when availrmem 2628 * drops below pages_pp_maximum, page locking mechanisms such as 2629 * page_pp_lock() will fail.) 2630 */ 2631 if (availrmem <= pages_pp_maximum) 2632 return (1); 2633 2634#endif /* sun */ 2635#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 2636 /* 2637 * If we're on an i386 platform, it's possible that we'll exhaust the 2638 * kernel heap space before we ever run out of available physical 2639 * memory. Most checks of the size of the heap_area compare against 2640 * tune.t_minarmem, which is the minimum available real memory that we 2641 * can have in the system. However, this is generally fixed at 25 pages 2642 * which is so low that it's useless. In this comparison, we seek to 2643 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2644 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2645 * free) 2646 */ 2647 if (vmem_size(heap_arena, VMEM_FREE) < 2648 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2)) { 2649 DTRACE_PROBE2(arc__reclaim_used, uint64_t, 2650 vmem_size(heap_arena, VMEM_FREE), uint64_t, 2651 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2); 2652 return (1); 2653 } 2654#define zio_arena NULL 2655#else 2656#define zio_arena heap_arena 2657#endif 2658 2659 /* 2660 * If zio data pages are being allocated out of a separate heap segment, 2661 * then enforce that the size of available vmem for this arena remains 2662 * above about 1/16th free. 2663 * 2664 * Note: The 1/16th arena free requirement was put in place 2665 * to aggressively evict memory from the arc in order to avoid 2666 * memory fragmentation issues. 2667 */ 2668 if (zio_arena != NULL && 2669 vmem_size(zio_arena, VMEM_FREE) < 2670 (vmem_size(zio_arena, VMEM_ALLOC) >> 4)) 2671 return (1); 2672 2673 /* 2674 * Above limits know nothing about real level of KVA fragmentation. 2675 * Start aggressive reclamation if too little sequential KVA left. 2676 */ 2677 if (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) 2678 return (1); 2679 2680#else /* _KERNEL */ 2681 if (spa_get_random(100) == 0) 2682 return (1); 2683#endif /* _KERNEL */ 2684 DTRACE_PROBE(arc__reclaim_no); 2685 2686 return (0); 2687} 2688 2689extern kmem_cache_t *zio_buf_cache[]; 2690extern kmem_cache_t *zio_data_buf_cache[]; 2691extern kmem_cache_t *range_seg_cache; 2692 2693static __noinline void 2694arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2695{ 2696 size_t i; 2697 kmem_cache_t *prev_cache = NULL; 2698 kmem_cache_t *prev_data_cache = NULL; 2699 2700 DTRACE_PROBE(arc__kmem_reap_start); 2701#ifdef _KERNEL 2702 if (arc_meta_used >= arc_meta_limit) { 2703 /* 2704 * We are exceeding our meta-data cache limit. 2705 * Purge some DNLC entries to release holds on meta-data. 2706 */ 2707 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2708 } 2709#if defined(__i386) 2710 /* 2711 * Reclaim unused memory from all kmem caches. 2712 */ 2713 kmem_reap(); 2714#endif 2715#endif 2716 2717 /* 2718 * An aggressive reclamation will shrink the cache size as well as 2719 * reap free buffers from the arc kmem caches. 2720 */ 2721 if (strat == ARC_RECLAIM_AGGR) 2722 arc_shrink(); 2723 2724 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2725 if (zio_buf_cache[i] != prev_cache) { 2726 prev_cache = zio_buf_cache[i]; 2727 kmem_cache_reap_now(zio_buf_cache[i]); 2728 } 2729 if (zio_data_buf_cache[i] != prev_data_cache) { 2730 prev_data_cache = zio_data_buf_cache[i]; 2731 kmem_cache_reap_now(zio_data_buf_cache[i]); 2732 } 2733 } 2734 kmem_cache_reap_now(buf_cache); 2735 kmem_cache_reap_now(hdr_cache); 2736 kmem_cache_reap_now(range_seg_cache); 2737 2738#ifdef sun 2739 /* 2740 * Ask the vmem arena to reclaim unused memory from its 2741 * quantum caches. 2742 */ 2743 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR) 2744 vmem_qcache_reap(zio_arena); 2745#endif 2746 DTRACE_PROBE(arc__kmem_reap_end); 2747} 2748 2749static void 2750arc_reclaim_thread(void *dummy __unused) 2751{ 2752 clock_t growtime = 0; 2753 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2754 callb_cpr_t cpr; 2755 2756 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2757 2758 mutex_enter(&arc_reclaim_thr_lock); 2759 while (arc_thread_exit == 0) { 2760 if (arc_reclaim_needed()) { 2761 2762 if (arc_no_grow) { 2763 if (last_reclaim == ARC_RECLAIM_CONS) { 2764 DTRACE_PROBE(arc__reclaim_aggr_no_grow); 2765 last_reclaim = ARC_RECLAIM_AGGR; 2766 } else { 2767 last_reclaim = ARC_RECLAIM_CONS; 2768 } 2769 } else { 2770 arc_no_grow = TRUE; 2771 last_reclaim = ARC_RECLAIM_AGGR; 2772 DTRACE_PROBE(arc__reclaim_aggr); 2773 membar_producer(); 2774 } 2775 2776 /* reset the growth delay for every reclaim */ 2777 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2778 2779 if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2780 /* 2781 * If needfree is TRUE our vm_lowmem hook 2782 * was called and in that case we must free some 2783 * memory, so switch to aggressive mode. 2784 */ 2785 arc_no_grow = TRUE; 2786 last_reclaim = ARC_RECLAIM_AGGR; 2787 } 2788 arc_kmem_reap_now(last_reclaim); 2789 arc_warm = B_TRUE; 2790 2791 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2792 arc_no_grow = FALSE; 2793 } 2794 2795 arc_adjust(); 2796 2797 if (arc_eviction_list != NULL) 2798 arc_do_user_evicts(); 2799 2800#ifdef _KERNEL 2801 if (needfree) { 2802 needfree = 0; 2803 wakeup(&needfree); 2804 } 2805#endif 2806 2807 /* block until needed, or one second, whichever is shorter */ 2808 CALLB_CPR_SAFE_BEGIN(&cpr); 2809 (void) cv_timedwait(&arc_reclaim_thr_cv, 2810 &arc_reclaim_thr_lock, hz); 2811 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2812 } 2813 2814 arc_thread_exit = 0; 2815 cv_broadcast(&arc_reclaim_thr_cv); 2816 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2817 thread_exit(); 2818} 2819 2820/* 2821 * Adapt arc info given the number of bytes we are trying to add and 2822 * the state that we are comming from. This function is only called 2823 * when we are adding new content to the cache. 2824 */ 2825static void 2826arc_adapt(int bytes, arc_state_t *state) 2827{ 2828 int mult; 2829 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2830 2831 if (state == arc_l2c_only) 2832 return; 2833 2834 ASSERT(bytes > 0); 2835 /* 2836 * Adapt the target size of the MRU list: 2837 * - if we just hit in the MRU ghost list, then increase 2838 * the target size of the MRU list. 2839 * - if we just hit in the MFU ghost list, then increase 2840 * the target size of the MFU list by decreasing the 2841 * target size of the MRU list. 2842 */ 2843 if (state == arc_mru_ghost) { 2844 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2845 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2846 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2847 2848 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2849 } else if (state == arc_mfu_ghost) { 2850 uint64_t delta; 2851 2852 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2853 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2854 mult = MIN(mult, 10); 2855 2856 delta = MIN(bytes * mult, arc_p); 2857 arc_p = MAX(arc_p_min, arc_p - delta); 2858 } 2859 ASSERT((int64_t)arc_p >= 0); 2860 2861 if (arc_reclaim_needed()) { 2862 cv_signal(&arc_reclaim_thr_cv); 2863 return; 2864 } 2865 2866 if (arc_no_grow) 2867 return; 2868 2869 if (arc_c >= arc_c_max) 2870 return; 2871 2872 /* 2873 * If we're within (2 * maxblocksize) bytes of the target 2874 * cache size, increment the target cache size 2875 */ 2876 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2877 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 2878 atomic_add_64(&arc_c, (int64_t)bytes); 2879 if (arc_c > arc_c_max) 2880 arc_c = arc_c_max; 2881 else if (state == arc_anon) 2882 atomic_add_64(&arc_p, (int64_t)bytes); 2883 if (arc_p > arc_c) 2884 arc_p = arc_c; 2885 } 2886 ASSERT((int64_t)arc_p >= 0); 2887} 2888 2889/* 2890 * Check if the cache has reached its limits and eviction is required 2891 * prior to insert. 2892 */ 2893static int 2894arc_evict_needed(arc_buf_contents_t type) 2895{ 2896 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2897 return (1); 2898 2899 if (arc_reclaim_needed()) 2900 return (1); 2901 2902 return (arc_size > arc_c); 2903} 2904 2905/* 2906 * The buffer, supplied as the first argument, needs a data block. 2907 * So, if we are at cache max, determine which cache should be victimized. 2908 * We have the following cases: 2909 * 2910 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2911 * In this situation if we're out of space, but the resident size of the MFU is 2912 * under the limit, victimize the MFU cache to satisfy this insertion request. 2913 * 2914 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2915 * Here, we've used up all of the available space for the MRU, so we need to 2916 * evict from our own cache instead. Evict from the set of resident MRU 2917 * entries. 2918 * 2919 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2920 * c minus p represents the MFU space in the cache, since p is the size of the 2921 * cache that is dedicated to the MRU. In this situation there's still space on 2922 * the MFU side, so the MRU side needs to be victimized. 2923 * 2924 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2925 * MFU's resident set is consuming more space than it has been allotted. In 2926 * this situation, we must victimize our own cache, the MFU, for this insertion. 2927 */ 2928static void 2929arc_get_data_buf(arc_buf_t *buf) 2930{ 2931 arc_state_t *state = buf->b_hdr->b_state; 2932 uint64_t size = buf->b_hdr->b_size; 2933 arc_buf_contents_t type = buf->b_hdr->b_type; 2934 2935 arc_adapt(size, state); 2936 2937 /* 2938 * We have not yet reached cache maximum size, 2939 * just allocate a new buffer. 2940 */ 2941 if (!arc_evict_needed(type)) { 2942 if (type == ARC_BUFC_METADATA) { 2943 buf->b_data = zio_buf_alloc(size); 2944 arc_space_consume(size, ARC_SPACE_DATA); 2945 } else { 2946 ASSERT(type == ARC_BUFC_DATA); 2947 buf->b_data = zio_data_buf_alloc(size); 2948 ARCSTAT_INCR(arcstat_data_size, size); 2949 atomic_add_64(&arc_size, size); 2950 } 2951 goto out; 2952 } 2953 2954 /* 2955 * If we are prefetching from the mfu ghost list, this buffer 2956 * will end up on the mru list; so steal space from there. 2957 */ 2958 if (state == arc_mfu_ghost) 2959 state = buf->b_hdr->b_flags & ARC_FLAG_PREFETCH ? 2960 arc_mru : arc_mfu; 2961 else if (state == arc_mru_ghost) 2962 state = arc_mru; 2963 2964 if (state == arc_mru || state == arc_anon) { 2965 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2966 state = (arc_mfu->arcs_lsize[type] >= size && 2967 arc_p > mru_used) ? arc_mfu : arc_mru; 2968 } else { 2969 /* MFU cases */ 2970 uint64_t mfu_space = arc_c - arc_p; 2971 state = (arc_mru->arcs_lsize[type] >= size && 2972 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2973 } 2974 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2975 if (type == ARC_BUFC_METADATA) { 2976 buf->b_data = zio_buf_alloc(size); 2977 arc_space_consume(size, ARC_SPACE_DATA); 2978 } else { 2979 ASSERT(type == ARC_BUFC_DATA); 2980 buf->b_data = zio_data_buf_alloc(size); 2981 ARCSTAT_INCR(arcstat_data_size, size); 2982 atomic_add_64(&arc_size, size); 2983 } 2984 ARCSTAT_BUMP(arcstat_recycle_miss); 2985 } 2986 ASSERT(buf->b_data != NULL); 2987out: 2988 /* 2989 * Update the state size. Note that ghost states have a 2990 * "ghost size" and so don't need to be updated. 2991 */ 2992 if (!GHOST_STATE(buf->b_hdr->b_state)) { 2993 arc_buf_hdr_t *hdr = buf->b_hdr; 2994 2995 atomic_add_64(&hdr->b_state->arcs_size, size); 2996 if (list_link_active(&hdr->b_arc_node)) { 2997 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2998 atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2999 } 3000 /* 3001 * If we are growing the cache, and we are adding anonymous 3002 * data, and we have outgrown arc_p, update arc_p 3003 */ 3004 if (arc_size < arc_c && hdr->b_state == arc_anon && 3005 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 3006 arc_p = MIN(arc_c, arc_p + size); 3007 } 3008 ARCSTAT_BUMP(arcstat_allocated); 3009} 3010 3011/* 3012 * This routine is called whenever a buffer is accessed. 3013 * NOTE: the hash lock is dropped in this function. 3014 */ 3015static void 3016arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3017{ 3018 clock_t now; 3019 3020 ASSERT(MUTEX_HELD(hash_lock)); 3021 3022 if (hdr->b_state == arc_anon) { 3023 /* 3024 * This buffer is not in the cache, and does not 3025 * appear in our "ghost" list. Add the new buffer 3026 * to the MRU state. 3027 */ 3028 3029 ASSERT(hdr->b_arc_access == 0); 3030 hdr->b_arc_access = ddi_get_lbolt(); 3031 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3032 arc_change_state(arc_mru, hdr, hash_lock); 3033 3034 } else if (hdr->b_state == arc_mru) { 3035 now = ddi_get_lbolt(); 3036 3037 /* 3038 * If this buffer is here because of a prefetch, then either: 3039 * - clear the flag if this is a "referencing" read 3040 * (any subsequent access will bump this into the MFU state). 3041 * or 3042 * - move the buffer to the head of the list if this is 3043 * another prefetch (to make it less likely to be evicted). 3044 */ 3045 if ((hdr->b_flags & ARC_FLAG_PREFETCH) != 0) { 3046 if (refcount_count(&hdr->b_refcnt) == 0) { 3047 ASSERT(list_link_active(&hdr->b_arc_node)); 3048 } else { 3049 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3050 ARCSTAT_BUMP(arcstat_mru_hits); 3051 } 3052 hdr->b_arc_access = now; 3053 return; 3054 } 3055 3056 /* 3057 * This buffer has been "accessed" only once so far, 3058 * but it is still in the cache. Move it to the MFU 3059 * state. 3060 */ 3061 if (now > hdr->b_arc_access + ARC_MINTIME) { 3062 /* 3063 * More than 125ms have passed since we 3064 * instantiated this buffer. Move it to the 3065 * most frequently used state. 3066 */ 3067 hdr->b_arc_access = now; 3068 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3069 arc_change_state(arc_mfu, hdr, hash_lock); 3070 } 3071 ARCSTAT_BUMP(arcstat_mru_hits); 3072 } else if (hdr->b_state == arc_mru_ghost) { 3073 arc_state_t *new_state; 3074 /* 3075 * This buffer has been "accessed" recently, but 3076 * was evicted from the cache. Move it to the 3077 * MFU state. 3078 */ 3079 3080 if (hdr->b_flags & ARC_FLAG_PREFETCH) { 3081 new_state = arc_mru; 3082 if (refcount_count(&hdr->b_refcnt) > 0) 3083 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3084 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3085 } else { 3086 new_state = arc_mfu; 3087 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3088 } 3089 3090 hdr->b_arc_access = ddi_get_lbolt(); 3091 arc_change_state(new_state, hdr, hash_lock); 3092 3093 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3094 } else if (hdr->b_state == arc_mfu) { 3095 /* 3096 * This buffer has been accessed more than once and is 3097 * still in the cache. Keep it in the MFU state. 3098 * 3099 * NOTE: an add_reference() that occurred when we did 3100 * the arc_read() will have kicked this off the list. 3101 * If it was a prefetch, we will explicitly move it to 3102 * the head of the list now. 3103 */ 3104 if ((hdr->b_flags & ARC_FLAG_PREFETCH) != 0) { 3105 ASSERT(refcount_count(&hdr->b_refcnt) == 0); 3106 ASSERT(list_link_active(&hdr->b_arc_node)); 3107 } 3108 ARCSTAT_BUMP(arcstat_mfu_hits); 3109 hdr->b_arc_access = ddi_get_lbolt(); 3110 } else if (hdr->b_state == arc_mfu_ghost) { 3111 arc_state_t *new_state = arc_mfu; 3112 /* 3113 * This buffer has been accessed more than once but has 3114 * been evicted from the cache. Move it back to the 3115 * MFU state. 3116 */ 3117 3118 if (hdr->b_flags & ARC_FLAG_PREFETCH) { 3119 /* 3120 * This is a prefetch access... 3121 * move this block back to the MRU state. 3122 */ 3123 ASSERT0(refcount_count(&hdr->b_refcnt)); 3124 new_state = arc_mru; 3125 } 3126 3127 hdr->b_arc_access = ddi_get_lbolt(); 3128 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3129 arc_change_state(new_state, hdr, hash_lock); 3130 3131 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 3132 } else if (hdr->b_state == arc_l2c_only) { 3133 /* 3134 * This buffer is on the 2nd Level ARC. 3135 */ 3136 3137 hdr->b_arc_access = ddi_get_lbolt(); 3138 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3139 arc_change_state(arc_mfu, hdr, hash_lock); 3140 } else { 3141 ASSERT(!"invalid arc state"); 3142 } 3143} 3144 3145/* a generic arc_done_func_t which you can use */ 3146/* ARGSUSED */ 3147void 3148arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 3149{ 3150 if (zio == NULL || zio->io_error == 0) 3151 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 3152 VERIFY(arc_buf_remove_ref(buf, arg)); 3153} 3154 3155/* a generic arc_done_func_t */ 3156void 3157arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 3158{ 3159 arc_buf_t **bufp = arg; 3160 if (zio && zio->io_error) { 3161 VERIFY(arc_buf_remove_ref(buf, arg)); 3162 *bufp = NULL; 3163 } else { 3164 *bufp = buf; 3165 ASSERT(buf->b_data); 3166 } 3167} 3168 3169static void 3170arc_read_done(zio_t *zio) 3171{ 3172 arc_buf_hdr_t *hdr; 3173 arc_buf_t *buf; 3174 arc_buf_t *abuf; /* buffer we're assigning to callback */ 3175 kmutex_t *hash_lock = NULL; 3176 arc_callback_t *callback_list, *acb; 3177 int freeable = FALSE; 3178 3179 buf = zio->io_private; 3180 hdr = buf->b_hdr; 3181 3182 /* 3183 * The hdr was inserted into hash-table and removed from lists 3184 * prior to starting I/O. We should find this header, since 3185 * it's in the hash table, and it should be legit since it's 3186 * not possible to evict it during the I/O. The only possible 3187 * reason for it not to be found is if we were freed during the 3188 * read. 3189 */ 3190 if (HDR_IN_HASH_TABLE(hdr)) { 3191 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3192 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3193 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3194 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3195 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3196 3197 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3198 &hash_lock); 3199 3200 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3201 hash_lock == NULL) || 3202 (found == hdr && 3203 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3204 (found == hdr && HDR_L2_READING(hdr))); 3205 } 3206 3207 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; 3208 if (l2arc_noprefetch && (hdr->b_flags & ARC_FLAG_PREFETCH)) 3209 hdr->b_flags &= ~ARC_FLAG_L2CACHE; 3210 3211 /* byteswap if necessary */ 3212 callback_list = hdr->b_acb; 3213 ASSERT(callback_list != NULL); 3214 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3215 dmu_object_byteswap_t bswap = 3216 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3217 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3218 byteswap_uint64_array : 3219 dmu_ot_byteswap[bswap].ob_func; 3220 func(buf->b_data, hdr->b_size); 3221 } 3222 3223 arc_cksum_compute(buf, B_FALSE); 3224#ifdef illumos 3225 arc_buf_watch(buf); 3226#endif /* illumos */ 3227 3228 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3229 /* 3230 * Only call arc_access on anonymous buffers. This is because 3231 * if we've issued an I/O for an evicted buffer, we've already 3232 * called arc_access (to prevent any simultaneous readers from 3233 * getting confused). 3234 */ 3235 arc_access(hdr, hash_lock); 3236 } 3237 3238 /* create copies of the data buffer for the callers */ 3239 abuf = buf; 3240 for (acb = callback_list; acb; acb = acb->acb_next) { 3241 if (acb->acb_done) { 3242 if (abuf == NULL) { 3243 ARCSTAT_BUMP(arcstat_duplicate_reads); 3244 abuf = arc_buf_clone(buf); 3245 } 3246 acb->acb_buf = abuf; 3247 abuf = NULL; 3248 } 3249 } 3250 hdr->b_acb = NULL; 3251 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 3252 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3253 if (abuf == buf) { 3254 ASSERT(buf->b_efunc == NULL); 3255 ASSERT(hdr->b_datacnt == 1); 3256 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 3257 } 3258 3259 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3260 3261 if (zio->io_error != 0) { 3262 hdr->b_flags |= ARC_FLAG_IO_ERROR; 3263 if (hdr->b_state != arc_anon) 3264 arc_change_state(arc_anon, hdr, hash_lock); 3265 if (HDR_IN_HASH_TABLE(hdr)) 3266 buf_hash_remove(hdr); 3267 freeable = refcount_is_zero(&hdr->b_refcnt); 3268 } 3269 3270 /* 3271 * Broadcast before we drop the hash_lock to avoid the possibility 3272 * that the hdr (and hence the cv) might be freed before we get to 3273 * the cv_broadcast(). 3274 */ 3275 cv_broadcast(&hdr->b_cv); 3276 3277 if (hash_lock) { 3278 mutex_exit(hash_lock); 3279 } else { 3280 /* 3281 * This block was freed while we waited for the read to 3282 * complete. It has been removed from the hash table and 3283 * moved to the anonymous state (so that it won't show up 3284 * in the cache). 3285 */ 3286 ASSERT3P(hdr->b_state, ==, arc_anon); 3287 freeable = refcount_is_zero(&hdr->b_refcnt); 3288 } 3289 3290 /* execute each callback and free its structure */ 3291 while ((acb = callback_list) != NULL) { 3292 if (acb->acb_done) 3293 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3294 3295 if (acb->acb_zio_dummy != NULL) { 3296 acb->acb_zio_dummy->io_error = zio->io_error; 3297 zio_nowait(acb->acb_zio_dummy); 3298 } 3299 3300 callback_list = acb->acb_next; 3301 kmem_free(acb, sizeof (arc_callback_t)); 3302 } 3303 3304 if (freeable) 3305 arc_hdr_destroy(hdr); 3306} 3307 3308/* 3309 * "Read" the block block at the specified DVA (in bp) via the 3310 * cache. If the block is found in the cache, invoke the provided 3311 * callback immediately and return. Note that the `zio' parameter 3312 * in the callback will be NULL in this case, since no IO was 3313 * required. If the block is not in the cache pass the read request 3314 * on to the spa with a substitute callback function, so that the 3315 * requested block will be added to the cache. 3316 * 3317 * If a read request arrives for a block that has a read in-progress, 3318 * either wait for the in-progress read to complete (and return the 3319 * results); or, if this is a read with a "done" func, add a record 3320 * to the read to invoke the "done" func when the read completes, 3321 * and return; or just return. 3322 * 3323 * arc_read_done() will invoke all the requested "done" functions 3324 * for readers of this block. 3325 */ 3326int 3327arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3328 void *private, zio_priority_t priority, int zio_flags, 3329 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 3330{ 3331 arc_buf_hdr_t *hdr = NULL; 3332 arc_buf_t *buf = NULL; 3333 kmutex_t *hash_lock = NULL; 3334 zio_t *rzio; 3335 uint64_t guid = spa_load_guid(spa); 3336 3337 ASSERT(!BP_IS_EMBEDDED(bp) || 3338 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3339 3340top: 3341 if (!BP_IS_EMBEDDED(bp)) { 3342 /* 3343 * Embedded BP's have no DVA and require no I/O to "read". 3344 * Create an anonymous arc buf to back it. 3345 */ 3346 hdr = buf_hash_find(guid, bp, &hash_lock); 3347 } 3348 3349 if (hdr != NULL && hdr->b_datacnt > 0) { 3350 3351 *arc_flags |= ARC_FLAG_CACHED; 3352 3353 if (HDR_IO_IN_PROGRESS(hdr)) { 3354 3355 if (*arc_flags & ARC_FLAG_WAIT) { 3356 cv_wait(&hdr->b_cv, hash_lock); 3357 mutex_exit(hash_lock); 3358 goto top; 3359 } 3360 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 3361 3362 if (done) { 3363 arc_callback_t *acb = NULL; 3364 3365 acb = kmem_zalloc(sizeof (arc_callback_t), 3366 KM_SLEEP); 3367 acb->acb_done = done; 3368 acb->acb_private = private; 3369 if (pio != NULL) 3370 acb->acb_zio_dummy = zio_null(pio, 3371 spa, NULL, NULL, NULL, zio_flags); 3372 3373 ASSERT(acb->acb_done != NULL); 3374 acb->acb_next = hdr->b_acb; 3375 hdr->b_acb = acb; 3376 add_reference(hdr, hash_lock, private); 3377 mutex_exit(hash_lock); 3378 return (0); 3379 } 3380 mutex_exit(hash_lock); 3381 return (0); 3382 } 3383 3384 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3385 3386 if (done) { 3387 add_reference(hdr, hash_lock, private); 3388 /* 3389 * If this block is already in use, create a new 3390 * copy of the data so that we will be guaranteed 3391 * that arc_release() will always succeed. 3392 */ 3393 buf = hdr->b_buf; 3394 ASSERT(buf); 3395 ASSERT(buf->b_data); 3396 if (HDR_BUF_AVAILABLE(hdr)) { 3397 ASSERT(buf->b_efunc == NULL); 3398 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 3399 } else { 3400 buf = arc_buf_clone(buf); 3401 } 3402 3403 } else if (*arc_flags & ARC_FLAG_PREFETCH && 3404 refcount_count(&hdr->b_refcnt) == 0) { 3405 hdr->b_flags |= ARC_FLAG_PREFETCH; 3406 } 3407 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3408 arc_access(hdr, hash_lock); 3409 if (*arc_flags & ARC_FLAG_L2CACHE) 3410 hdr->b_flags |= ARC_FLAG_L2CACHE; 3411 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3412 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3413 mutex_exit(hash_lock); 3414 ARCSTAT_BUMP(arcstat_hits); 3415 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 3416 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3417 data, metadata, hits); 3418 3419 if (done) 3420 done(NULL, buf, private); 3421 } else { 3422 uint64_t size = BP_GET_LSIZE(bp); 3423 arc_callback_t *acb; 3424 vdev_t *vd = NULL; 3425 uint64_t addr = 0; 3426 boolean_t devw = B_FALSE; 3427 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3428 uint64_t b_asize = 0; 3429 3430 if (hdr == NULL) { 3431 /* this block is not in the cache */ 3432 arc_buf_hdr_t *exists = NULL; 3433 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3434 buf = arc_buf_alloc(spa, size, private, type); 3435 hdr = buf->b_hdr; 3436 if (!BP_IS_EMBEDDED(bp)) { 3437 hdr->b_dva = *BP_IDENTITY(bp); 3438 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3439 hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3440 exists = buf_hash_insert(hdr, &hash_lock); 3441 } 3442 if (exists != NULL) { 3443 /* somebody beat us to the hash insert */ 3444 mutex_exit(hash_lock); 3445 buf_discard_identity(hdr); 3446 (void) arc_buf_remove_ref(buf, private); 3447 goto top; /* restart the IO request */ 3448 } 3449 3450 /* if this is a prefetch, we don't have a reference */ 3451 if (*arc_flags & ARC_FLAG_PREFETCH) { 3452 (void) remove_reference(hdr, hash_lock, 3453 private); 3454 hdr->b_flags |= ARC_FLAG_PREFETCH; 3455 } 3456 if (*arc_flags & ARC_FLAG_L2CACHE) 3457 hdr->b_flags |= ARC_FLAG_L2CACHE; 3458 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3459 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3460 if (BP_GET_LEVEL(bp) > 0) 3461 hdr->b_flags |= ARC_FLAG_INDIRECT; 3462 } else { 3463 /* this block is in the ghost cache */ 3464 ASSERT(GHOST_STATE(hdr->b_state)); 3465 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3466 ASSERT0(refcount_count(&hdr->b_refcnt)); 3467 ASSERT(hdr->b_buf == NULL); 3468 3469 /* if this is a prefetch, we don't have a reference */ 3470 if (*arc_flags & ARC_FLAG_PREFETCH) 3471 hdr->b_flags |= ARC_FLAG_PREFETCH; 3472 else 3473 add_reference(hdr, hash_lock, private); 3474 if (*arc_flags & ARC_FLAG_L2CACHE) 3475 hdr->b_flags |= ARC_FLAG_L2CACHE; 3476 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3477 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3478 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3479 buf->b_hdr = hdr; 3480 buf->b_data = NULL; 3481 buf->b_efunc = NULL; 3482 buf->b_private = NULL; 3483 buf->b_next = NULL; 3484 hdr->b_buf = buf; 3485 ASSERT(hdr->b_datacnt == 0); 3486 hdr->b_datacnt = 1; 3487 arc_get_data_buf(buf); 3488 arc_access(hdr, hash_lock); 3489 } 3490 3491 ASSERT(!GHOST_STATE(hdr->b_state)); 3492 3493 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3494 acb->acb_done = done; 3495 acb->acb_private = private; 3496 3497 ASSERT(hdr->b_acb == NULL); 3498 hdr->b_acb = acb; 3499 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 3500 3501 if (hdr->b_l2hdr != NULL && 3502 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3503 devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3504 addr = hdr->b_l2hdr->b_daddr; 3505 b_compress = hdr->b_l2hdr->b_compress; 3506 b_asize = hdr->b_l2hdr->b_asize; 3507 /* 3508 * Lock out device removal. 3509 */ 3510 if (vdev_is_dead(vd) || 3511 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3512 vd = NULL; 3513 } 3514 3515 if (hash_lock != NULL) 3516 mutex_exit(hash_lock); 3517 3518 /* 3519 * At this point, we have a level 1 cache miss. Try again in 3520 * L2ARC if possible. 3521 */ 3522 ASSERT3U(hdr->b_size, ==, size); 3523 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3524 uint64_t, size, zbookmark_phys_t *, zb); 3525 ARCSTAT_BUMP(arcstat_misses); 3526 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 3527 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3528 data, metadata, misses); 3529#ifdef _KERNEL 3530 curthread->td_ru.ru_inblock++; 3531#endif 3532 3533 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3534 /* 3535 * Read from the L2ARC if the following are true: 3536 * 1. The L2ARC vdev was previously cached. 3537 * 2. This buffer still has L2ARC metadata. 3538 * 3. This buffer isn't currently writing to the L2ARC. 3539 * 4. The L2ARC entry wasn't evicted, which may 3540 * also have invalidated the vdev. 3541 * 5. This isn't prefetch and l2arc_noprefetch is set. 3542 */ 3543 if (hdr->b_l2hdr != NULL && 3544 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3545 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3546 l2arc_read_callback_t *cb; 3547 3548 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3549 ARCSTAT_BUMP(arcstat_l2_hits); 3550 3551 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3552 KM_SLEEP); 3553 cb->l2rcb_buf = buf; 3554 cb->l2rcb_spa = spa; 3555 cb->l2rcb_bp = *bp; 3556 cb->l2rcb_zb = *zb; 3557 cb->l2rcb_flags = zio_flags; 3558 cb->l2rcb_compress = b_compress; 3559 3560 ASSERT(addr >= VDEV_LABEL_START_SIZE && 3561 addr + size < vd->vdev_psize - 3562 VDEV_LABEL_END_SIZE); 3563 3564 /* 3565 * l2arc read. The SCL_L2ARC lock will be 3566 * released by l2arc_read_done(). 3567 * Issue a null zio if the underlying buffer 3568 * was squashed to zero size by compression. 3569 */ 3570 if (b_compress == ZIO_COMPRESS_EMPTY) { 3571 rzio = zio_null(pio, spa, vd, 3572 l2arc_read_done, cb, 3573 zio_flags | ZIO_FLAG_DONT_CACHE | 3574 ZIO_FLAG_CANFAIL | 3575 ZIO_FLAG_DONT_PROPAGATE | 3576 ZIO_FLAG_DONT_RETRY); 3577 } else { 3578 rzio = zio_read_phys(pio, vd, addr, 3579 b_asize, buf->b_data, 3580 ZIO_CHECKSUM_OFF, 3581 l2arc_read_done, cb, priority, 3582 zio_flags | ZIO_FLAG_DONT_CACHE | 3583 ZIO_FLAG_CANFAIL | 3584 ZIO_FLAG_DONT_PROPAGATE | 3585 ZIO_FLAG_DONT_RETRY, B_FALSE); 3586 } 3587 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3588 zio_t *, rzio); 3589 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3590 3591 if (*arc_flags & ARC_FLAG_NOWAIT) { 3592 zio_nowait(rzio); 3593 return (0); 3594 } 3595 3596 ASSERT(*arc_flags & ARC_FLAG_WAIT); 3597 if (zio_wait(rzio) == 0) 3598 return (0); 3599 3600 /* l2arc read error; goto zio_read() */ 3601 } else { 3602 DTRACE_PROBE1(l2arc__miss, 3603 arc_buf_hdr_t *, hdr); 3604 ARCSTAT_BUMP(arcstat_l2_misses); 3605 if (HDR_L2_WRITING(hdr)) 3606 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3607 spa_config_exit(spa, SCL_L2ARC, vd); 3608 } 3609 } else { 3610 if (vd != NULL) 3611 spa_config_exit(spa, SCL_L2ARC, vd); 3612 if (l2arc_ndev != 0) { 3613 DTRACE_PROBE1(l2arc__miss, 3614 arc_buf_hdr_t *, hdr); 3615 ARCSTAT_BUMP(arcstat_l2_misses); 3616 } 3617 } 3618 3619 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3620 arc_read_done, buf, priority, zio_flags, zb); 3621 3622 if (*arc_flags & ARC_FLAG_WAIT) 3623 return (zio_wait(rzio)); 3624 3625 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 3626 zio_nowait(rzio); 3627 } 3628 return (0); 3629} 3630 3631void 3632arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3633{ 3634 ASSERT(buf->b_hdr != NULL); 3635 ASSERT(buf->b_hdr->b_state != arc_anon); 3636 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3637 ASSERT(buf->b_efunc == NULL); 3638 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3639 3640 buf->b_efunc = func; 3641 buf->b_private = private; 3642} 3643 3644/* 3645 * Notify the arc that a block was freed, and thus will never be used again. 3646 */ 3647void 3648arc_freed(spa_t *spa, const blkptr_t *bp) 3649{ 3650 arc_buf_hdr_t *hdr; 3651 kmutex_t *hash_lock; 3652 uint64_t guid = spa_load_guid(spa); 3653 3654 ASSERT(!BP_IS_EMBEDDED(bp)); 3655 3656 hdr = buf_hash_find(guid, bp, &hash_lock); 3657 if (hdr == NULL) 3658 return; 3659 if (HDR_BUF_AVAILABLE(hdr)) { 3660 arc_buf_t *buf = hdr->b_buf; 3661 add_reference(hdr, hash_lock, FTAG); 3662 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 3663 mutex_exit(hash_lock); 3664 3665 arc_release(buf, FTAG); 3666 (void) arc_buf_remove_ref(buf, FTAG); 3667 } else { 3668 mutex_exit(hash_lock); 3669 } 3670 3671} 3672 3673/* 3674 * Clear the user eviction callback set by arc_set_callback(), first calling 3675 * it if it exists. Because the presence of a callback keeps an arc_buf cached 3676 * clearing the callback may result in the arc_buf being destroyed. However, 3677 * it will not result in the *last* arc_buf being destroyed, hence the data 3678 * will remain cached in the ARC. We make a copy of the arc buffer here so 3679 * that we can process the callback without holding any locks. 3680 * 3681 * It's possible that the callback is already in the process of being cleared 3682 * by another thread. In this case we can not clear the callback. 3683 * 3684 * Returns B_TRUE if the callback was successfully called and cleared. 3685 */ 3686boolean_t 3687arc_clear_callback(arc_buf_t *buf) 3688{ 3689 arc_buf_hdr_t *hdr; 3690 kmutex_t *hash_lock; 3691 arc_evict_func_t *efunc = buf->b_efunc; 3692 void *private = buf->b_private; 3693 list_t *list, *evicted_list; 3694 kmutex_t *lock, *evicted_lock; 3695 3696 mutex_enter(&buf->b_evict_lock); 3697 hdr = buf->b_hdr; 3698 if (hdr == NULL) { 3699 /* 3700 * We are in arc_do_user_evicts(). 3701 */ 3702 ASSERT(buf->b_data == NULL); 3703 mutex_exit(&buf->b_evict_lock); 3704 return (B_FALSE); 3705 } else if (buf->b_data == NULL) { 3706 /* 3707 * We are on the eviction list; process this buffer now 3708 * but let arc_do_user_evicts() do the reaping. 3709 */ 3710 buf->b_efunc = NULL; 3711 mutex_exit(&buf->b_evict_lock); 3712 VERIFY0(efunc(private)); 3713 return (B_TRUE); 3714 } 3715 hash_lock = HDR_LOCK(hdr); 3716 mutex_enter(hash_lock); 3717 hdr = buf->b_hdr; 3718 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3719 3720 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3721 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3722 3723 buf->b_efunc = NULL; 3724 buf->b_private = NULL; 3725 3726 if (hdr->b_datacnt > 1) { 3727 mutex_exit(&buf->b_evict_lock); 3728 arc_buf_destroy(buf, FALSE, TRUE); 3729 } else { 3730 ASSERT(buf == hdr->b_buf); 3731 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 3732 mutex_exit(&buf->b_evict_lock); 3733 } 3734 3735 mutex_exit(hash_lock); 3736 VERIFY0(efunc(private)); 3737 return (B_TRUE); 3738} 3739 3740/* 3741 * Release this buffer from the cache, making it an anonymous buffer. This 3742 * must be done after a read and prior to modifying the buffer contents. 3743 * If the buffer has more than one reference, we must make 3744 * a new hdr for the buffer. 3745 */ 3746void 3747arc_release(arc_buf_t *buf, void *tag) 3748{ 3749 arc_buf_hdr_t *hdr; 3750 kmutex_t *hash_lock = NULL; 3751 l2arc_buf_hdr_t *l2hdr; 3752 uint64_t buf_size; 3753 3754 /* 3755 * It would be nice to assert that if it's DMU metadata (level > 3756 * 0 || it's the dnode file), then it must be syncing context. 3757 * But we don't know that information at this level. 3758 */ 3759 3760 mutex_enter(&buf->b_evict_lock); 3761 hdr = buf->b_hdr; 3762 3763 /* this buffer is not on any list */ 3764 ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3765 3766 if (hdr->b_state == arc_anon) { 3767 /* this buffer is already released */ 3768 ASSERT(buf->b_efunc == NULL); 3769 } else { 3770 hash_lock = HDR_LOCK(hdr); 3771 mutex_enter(hash_lock); 3772 hdr = buf->b_hdr; 3773 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3774 } 3775 3776 l2hdr = hdr->b_l2hdr; 3777 if (l2hdr) { 3778 mutex_enter(&l2arc_buflist_mtx); 3779 arc_buf_l2_cdata_free(hdr); 3780 hdr->b_l2hdr = NULL; 3781 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3782 } 3783 buf_size = hdr->b_size; 3784 3785 /* 3786 * Do we have more than one buf? 3787 */ 3788 if (hdr->b_datacnt > 1) { 3789 arc_buf_hdr_t *nhdr; 3790 arc_buf_t **bufp; 3791 uint64_t blksz = hdr->b_size; 3792 uint64_t spa = hdr->b_spa; 3793 arc_buf_contents_t type = hdr->b_type; 3794 uint32_t flags = hdr->b_flags; 3795 3796 ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3797 /* 3798 * Pull the data off of this hdr and attach it to 3799 * a new anonymous hdr. 3800 */ 3801 (void) remove_reference(hdr, hash_lock, tag); 3802 bufp = &hdr->b_buf; 3803 while (*bufp != buf) 3804 bufp = &(*bufp)->b_next; 3805 *bufp = buf->b_next; 3806 buf->b_next = NULL; 3807 3808 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3809 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3810 if (refcount_is_zero(&hdr->b_refcnt)) { 3811 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3812 ASSERT3U(*size, >=, hdr->b_size); 3813 atomic_add_64(size, -hdr->b_size); 3814 } 3815 3816 /* 3817 * We're releasing a duplicate user data buffer, update 3818 * our statistics accordingly. 3819 */ 3820 if (hdr->b_type == ARC_BUFC_DATA) { 3821 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3822 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3823 -hdr->b_size); 3824 } 3825 hdr->b_datacnt -= 1; 3826 arc_cksum_verify(buf); 3827#ifdef illumos 3828 arc_buf_unwatch(buf); 3829#endif /* illumos */ 3830 3831 mutex_exit(hash_lock); 3832 3833 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3834 nhdr->b_size = blksz; 3835 nhdr->b_spa = spa; 3836 nhdr->b_type = type; 3837 nhdr->b_buf = buf; 3838 nhdr->b_state = arc_anon; 3839 nhdr->b_arc_access = 0; 3840 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; 3841 nhdr->b_l2hdr = NULL; 3842 nhdr->b_datacnt = 1; 3843 nhdr->b_freeze_cksum = NULL; 3844 (void) refcount_add(&nhdr->b_refcnt, tag); 3845 buf->b_hdr = nhdr; 3846 mutex_exit(&buf->b_evict_lock); 3847 atomic_add_64(&arc_anon->arcs_size, blksz); 3848 } else { 3849 mutex_exit(&buf->b_evict_lock); 3850 ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3851 ASSERT(!list_link_active(&hdr->b_arc_node)); 3852 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3853 if (hdr->b_state != arc_anon) 3854 arc_change_state(arc_anon, hdr, hash_lock); 3855 hdr->b_arc_access = 0; 3856 if (hash_lock) 3857 mutex_exit(hash_lock); 3858 3859 buf_discard_identity(hdr); 3860 arc_buf_thaw(buf); 3861 } 3862 buf->b_efunc = NULL; 3863 buf->b_private = NULL; 3864 3865 if (l2hdr) { 3866 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3867 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 3868 -l2hdr->b_asize, 0, 0); 3869 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3870 l2hdr->b_asize, 0); 3871 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3872 ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3873 mutex_exit(&l2arc_buflist_mtx); 3874 } 3875} 3876 3877int 3878arc_released(arc_buf_t *buf) 3879{ 3880 int released; 3881 3882 mutex_enter(&buf->b_evict_lock); 3883 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3884 mutex_exit(&buf->b_evict_lock); 3885 return (released); 3886} 3887 3888#ifdef ZFS_DEBUG 3889int 3890arc_referenced(arc_buf_t *buf) 3891{ 3892 int referenced; 3893 3894 mutex_enter(&buf->b_evict_lock); 3895 referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3896 mutex_exit(&buf->b_evict_lock); 3897 return (referenced); 3898} 3899#endif 3900 3901static void 3902arc_write_ready(zio_t *zio) 3903{ 3904 arc_write_callback_t *callback = zio->io_private; 3905 arc_buf_t *buf = callback->awcb_buf; 3906 arc_buf_hdr_t *hdr = buf->b_hdr; 3907 3908 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3909 callback->awcb_ready(zio, buf, callback->awcb_private); 3910 3911 /* 3912 * If the IO is already in progress, then this is a re-write 3913 * attempt, so we need to thaw and re-compute the cksum. 3914 * It is the responsibility of the callback to handle the 3915 * accounting for any re-write attempt. 3916 */ 3917 if (HDR_IO_IN_PROGRESS(hdr)) { 3918 mutex_enter(&hdr->b_freeze_lock); 3919 if (hdr->b_freeze_cksum != NULL) { 3920 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3921 hdr->b_freeze_cksum = NULL; 3922 } 3923 mutex_exit(&hdr->b_freeze_lock); 3924 } 3925 arc_cksum_compute(buf, B_FALSE); 3926 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 3927} 3928 3929/* 3930 * The SPA calls this callback for each physical write that happens on behalf 3931 * of a logical write. See the comment in dbuf_write_physdone() for details. 3932 */ 3933static void 3934arc_write_physdone(zio_t *zio) 3935{ 3936 arc_write_callback_t *cb = zio->io_private; 3937 if (cb->awcb_physdone != NULL) 3938 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3939} 3940 3941static void 3942arc_write_done(zio_t *zio) 3943{ 3944 arc_write_callback_t *callback = zio->io_private; 3945 arc_buf_t *buf = callback->awcb_buf; 3946 arc_buf_hdr_t *hdr = buf->b_hdr; 3947 3948 ASSERT(hdr->b_acb == NULL); 3949 3950 if (zio->io_error == 0) { 3951 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 3952 buf_discard_identity(hdr); 3953 } else { 3954 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3955 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3956 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3957 } 3958 } else { 3959 ASSERT(BUF_EMPTY(hdr)); 3960 } 3961 3962 /* 3963 * If the block to be written was all-zero or compressed enough to be 3964 * embedded in the BP, no write was performed so there will be no 3965 * dva/birth/checksum. The buffer must therefore remain anonymous 3966 * (and uncached). 3967 */ 3968 if (!BUF_EMPTY(hdr)) { 3969 arc_buf_hdr_t *exists; 3970 kmutex_t *hash_lock; 3971 3972 ASSERT(zio->io_error == 0); 3973 3974 arc_cksum_verify(buf); 3975 3976 exists = buf_hash_insert(hdr, &hash_lock); 3977 if (exists) { 3978 /* 3979 * This can only happen if we overwrite for 3980 * sync-to-convergence, because we remove 3981 * buffers from the hash table when we arc_free(). 3982 */ 3983 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3984 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3985 panic("bad overwrite, hdr=%p exists=%p", 3986 (void *)hdr, (void *)exists); 3987 ASSERT(refcount_is_zero(&exists->b_refcnt)); 3988 arc_change_state(arc_anon, exists, hash_lock); 3989 mutex_exit(hash_lock); 3990 arc_hdr_destroy(exists); 3991 exists = buf_hash_insert(hdr, &hash_lock); 3992 ASSERT3P(exists, ==, NULL); 3993 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3994 /* nopwrite */ 3995 ASSERT(zio->io_prop.zp_nopwrite); 3996 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3997 panic("bad nopwrite, hdr=%p exists=%p", 3998 (void *)hdr, (void *)exists); 3999 } else { 4000 /* Dedup */ 4001 ASSERT(hdr->b_datacnt == 1); 4002 ASSERT(hdr->b_state == arc_anon); 4003 ASSERT(BP_GET_DEDUP(zio->io_bp)); 4004 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 4005 } 4006 } 4007 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4008 /* if it's not anon, we are doing a scrub */ 4009 if (!exists && hdr->b_state == arc_anon) 4010 arc_access(hdr, hash_lock); 4011 mutex_exit(hash_lock); 4012 } else { 4013 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4014 } 4015 4016 ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 4017 callback->awcb_done(zio, buf, callback->awcb_private); 4018 4019 kmem_free(callback, sizeof (arc_write_callback_t)); 4020} 4021 4022zio_t * 4023arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 4024 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 4025 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 4026 arc_done_func_t *done, void *private, zio_priority_t priority, 4027 int zio_flags, const zbookmark_phys_t *zb) 4028{ 4029 arc_buf_hdr_t *hdr = buf->b_hdr; 4030 arc_write_callback_t *callback; 4031 zio_t *zio; 4032 4033 ASSERT(ready != NULL); 4034 ASSERT(done != NULL); 4035 ASSERT(!HDR_IO_ERROR(hdr)); 4036 ASSERT((hdr->b_flags & ARC_FLAG_IO_IN_PROGRESS) == 0); 4037 ASSERT(hdr->b_acb == NULL); 4038 if (l2arc) 4039 hdr->b_flags |= ARC_FLAG_L2CACHE; 4040 if (l2arc_compress) 4041 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4042 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 4043 callback->awcb_ready = ready; 4044 callback->awcb_physdone = physdone; 4045 callback->awcb_done = done; 4046 callback->awcb_private = private; 4047 callback->awcb_buf = buf; 4048 4049 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 4050 arc_write_ready, arc_write_physdone, arc_write_done, callback, 4051 priority, zio_flags, zb); 4052 4053 return (zio); 4054} 4055 4056static int 4057arc_memory_throttle(uint64_t reserve, uint64_t txg) 4058{ 4059#ifdef _KERNEL 4060 uint64_t available_memory = ptob(freemem); 4061 static uint64_t page_load = 0; 4062 static uint64_t last_txg = 0; 4063 4064#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 4065 available_memory = 4066 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 4067#endif 4068 4069 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 4070 return (0); 4071 4072 if (txg > last_txg) { 4073 last_txg = txg; 4074 page_load = 0; 4075 } 4076 /* 4077 * If we are in pageout, we know that memory is already tight, 4078 * the arc is already going to be evicting, so we just want to 4079 * continue to let page writes occur as quickly as possible. 4080 */ 4081 if (curproc == pageproc) { 4082 if (page_load > MAX(ptob(minfree), available_memory) / 4) 4083 return (SET_ERROR(ERESTART)); 4084 /* Note: reserve is inflated, so we deflate */ 4085 page_load += reserve / 8; 4086 return (0); 4087 } else if (page_load > 0 && arc_reclaim_needed()) { 4088 /* memory is low, delay before restarting */ 4089 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 4090 return (SET_ERROR(EAGAIN)); 4091 } 4092 page_load = 0; 4093#endif 4094 return (0); 4095} 4096 4097void 4098arc_tempreserve_clear(uint64_t reserve) 4099{ 4100 atomic_add_64(&arc_tempreserve, -reserve); 4101 ASSERT((int64_t)arc_tempreserve >= 0); 4102} 4103 4104int 4105arc_tempreserve_space(uint64_t reserve, uint64_t txg) 4106{ 4107 int error; 4108 uint64_t anon_size; 4109 4110 if (reserve > arc_c/4 && !arc_no_grow) { 4111 arc_c = MIN(arc_c_max, reserve * 4); 4112 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 4113 } 4114 if (reserve > arc_c) 4115 return (SET_ERROR(ENOMEM)); 4116 4117 /* 4118 * Don't count loaned bufs as in flight dirty data to prevent long 4119 * network delays from blocking transactions that are ready to be 4120 * assigned to a txg. 4121 */ 4122 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 4123 4124 /* 4125 * Writes will, almost always, require additional memory allocations 4126 * in order to compress/encrypt/etc the data. We therefore need to 4127 * make sure that there is sufficient available memory for this. 4128 */ 4129 error = arc_memory_throttle(reserve, txg); 4130 if (error != 0) 4131 return (error); 4132 4133 /* 4134 * Throttle writes when the amount of dirty data in the cache 4135 * gets too large. We try to keep the cache less than half full 4136 * of dirty blocks so that our sync times don't grow too large. 4137 * Note: if two requests come in concurrently, we might let them 4138 * both succeed, when one of them should fail. Not a huge deal. 4139 */ 4140 4141 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 4142 anon_size > arc_c / 4) { 4143 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 4144 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 4145 arc_tempreserve>>10, 4146 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 4147 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 4148 reserve>>10, arc_c>>10); 4149 return (SET_ERROR(ERESTART)); 4150 } 4151 atomic_add_64(&arc_tempreserve, reserve); 4152 return (0); 4153} 4154 4155static kmutex_t arc_lowmem_lock; 4156#ifdef _KERNEL 4157static eventhandler_tag arc_event_lowmem = NULL; 4158 4159static void 4160arc_lowmem(void *arg __unused, int howto __unused) 4161{ 4162 4163 /* Serialize access via arc_lowmem_lock. */ 4164 mutex_enter(&arc_lowmem_lock); 4165 mutex_enter(&arc_reclaim_thr_lock); 4166 needfree = 1; 4167 DTRACE_PROBE(arc__needfree); 4168 cv_signal(&arc_reclaim_thr_cv); 4169 4170 /* 4171 * It is unsafe to block here in arbitrary threads, because we can come 4172 * here from ARC itself and may hold ARC locks and thus risk a deadlock 4173 * with ARC reclaim thread. 4174 */ 4175 if (curproc == pageproc) { 4176 while (needfree) 4177 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4178 } 4179 mutex_exit(&arc_reclaim_thr_lock); 4180 mutex_exit(&arc_lowmem_lock); 4181} 4182#endif 4183 4184void 4185arc_init(void) 4186{ 4187 int i, prefetch_tunable_set = 0; 4188 4189 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4190 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4191 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4192 4193 /* Convert seconds to clock ticks */ 4194 arc_min_prefetch_lifespan = 1 * hz; 4195 4196 /* Start out with 1/8 of all memory */ 4197 arc_c = kmem_size() / 8; 4198 4199#ifdef sun 4200#ifdef _KERNEL 4201 /* 4202 * On architectures where the physical memory can be larger 4203 * than the addressable space (intel in 32-bit mode), we may 4204 * need to limit the cache to 1/8 of VM size. 4205 */ 4206 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4207#endif 4208#endif /* sun */ 4209 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4210 arc_c_min = MAX(arc_c / 4, 16 << 20); 4211 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4212 if (arc_c * 8 >= 1 << 30) 4213 arc_c_max = (arc_c * 8) - (1 << 30); 4214 else 4215 arc_c_max = arc_c_min; 4216 arc_c_max = MAX(arc_c * 5, arc_c_max); 4217 4218#ifdef _KERNEL 4219 /* 4220 * Allow the tunables to override our calculations if they are 4221 * reasonable (ie. over 16MB) 4222 */ 4223 if (zfs_arc_max > 16 << 20 && zfs_arc_max < kmem_size()) 4224 arc_c_max = zfs_arc_max; 4225 if (zfs_arc_min > 16 << 20 && zfs_arc_min <= arc_c_max) 4226 arc_c_min = zfs_arc_min; 4227#endif 4228 4229 arc_c = arc_c_max; 4230 arc_p = (arc_c >> 1); 4231 4232 /* limit meta-data to 1/4 of the arc capacity */ 4233 arc_meta_limit = arc_c_max / 4; 4234 4235 /* Allow the tunable to override if it is reasonable */ 4236 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4237 arc_meta_limit = zfs_arc_meta_limit; 4238 4239 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4240 arc_c_min = arc_meta_limit / 2; 4241 4242 if (zfs_arc_meta_min > 0) { 4243 arc_meta_min = zfs_arc_meta_min; 4244 } else { 4245 arc_meta_min = arc_c_min / 2; 4246 } 4247 4248 if (zfs_arc_grow_retry > 0) 4249 arc_grow_retry = zfs_arc_grow_retry; 4250 4251 if (zfs_arc_shrink_shift > 0) 4252 arc_shrink_shift = zfs_arc_shrink_shift; 4253 4254 if (zfs_arc_p_min_shift > 0) 4255 arc_p_min_shift = zfs_arc_p_min_shift; 4256 4257 /* if kmem_flags are set, lets try to use less memory */ 4258 if (kmem_debugging()) 4259 arc_c = arc_c / 2; 4260 if (arc_c < arc_c_min) 4261 arc_c = arc_c_min; 4262 4263 zfs_arc_min = arc_c_min; 4264 zfs_arc_max = arc_c_max; 4265 4266 arc_anon = &ARC_anon; 4267 arc_mru = &ARC_mru; 4268 arc_mru_ghost = &ARC_mru_ghost; 4269 arc_mfu = &ARC_mfu; 4270 arc_mfu_ghost = &ARC_mfu_ghost; 4271 arc_l2c_only = &ARC_l2c_only; 4272 arc_size = 0; 4273 4274 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4275 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4276 NULL, MUTEX_DEFAULT, NULL); 4277 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4278 NULL, MUTEX_DEFAULT, NULL); 4279 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4280 NULL, MUTEX_DEFAULT, NULL); 4281 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4282 NULL, MUTEX_DEFAULT, NULL); 4283 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4284 NULL, MUTEX_DEFAULT, NULL); 4285 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4286 NULL, MUTEX_DEFAULT, NULL); 4287 4288 list_create(&arc_mru->arcs_lists[i], 4289 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4290 list_create(&arc_mru_ghost->arcs_lists[i], 4291 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4292 list_create(&arc_mfu->arcs_lists[i], 4293 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4294 list_create(&arc_mfu_ghost->arcs_lists[i], 4295 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4296 list_create(&arc_mfu_ghost->arcs_lists[i], 4297 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4298 list_create(&arc_l2c_only->arcs_lists[i], 4299 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4300 } 4301 4302 buf_init(); 4303 4304 arc_thread_exit = 0; 4305 arc_eviction_list = NULL; 4306 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4307 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4308 4309 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4310 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4311 4312 if (arc_ksp != NULL) { 4313 arc_ksp->ks_data = &arc_stats; 4314 kstat_install(arc_ksp); 4315 } 4316 4317 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4318 TS_RUN, minclsyspri); 4319 4320#ifdef _KERNEL 4321 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4322 EVENTHANDLER_PRI_FIRST); 4323#endif 4324 4325 arc_dead = FALSE; 4326 arc_warm = B_FALSE; 4327 4328 /* 4329 * Calculate maximum amount of dirty data per pool. 4330 * 4331 * If it has been set by /etc/system, take that. 4332 * Otherwise, use a percentage of physical memory defined by 4333 * zfs_dirty_data_max_percent (default 10%) with a cap at 4334 * zfs_dirty_data_max_max (default 4GB). 4335 */ 4336 if (zfs_dirty_data_max == 0) { 4337 zfs_dirty_data_max = ptob(physmem) * 4338 zfs_dirty_data_max_percent / 100; 4339 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4340 zfs_dirty_data_max_max); 4341 } 4342 4343#ifdef _KERNEL 4344 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4345 prefetch_tunable_set = 1; 4346 4347#ifdef __i386__ 4348 if (prefetch_tunable_set == 0) { 4349 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4350 "-- to enable,\n"); 4351 printf(" add \"vfs.zfs.prefetch_disable=0\" " 4352 "to /boot/loader.conf.\n"); 4353 zfs_prefetch_disable = 1; 4354 } 4355#else 4356 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4357 prefetch_tunable_set == 0) { 4358 printf("ZFS NOTICE: Prefetch is disabled by default if less " 4359 "than 4GB of RAM is present;\n" 4360 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4361 "to /boot/loader.conf.\n"); 4362 zfs_prefetch_disable = 1; 4363 } 4364#endif 4365 /* Warn about ZFS memory and address space requirements. */ 4366 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4367 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4368 "expect unstable behavior.\n"); 4369 } 4370 if (kmem_size() < 512 * (1 << 20)) { 4371 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4372 "expect unstable behavior.\n"); 4373 printf(" Consider tuning vm.kmem_size and " 4374 "vm.kmem_size_max\n"); 4375 printf(" in /boot/loader.conf.\n"); 4376 } 4377#endif 4378} 4379 4380void 4381arc_fini(void) 4382{ 4383 int i; 4384 4385 mutex_enter(&arc_reclaim_thr_lock); 4386 arc_thread_exit = 1; 4387 cv_signal(&arc_reclaim_thr_cv); 4388 while (arc_thread_exit != 0) 4389 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4390 mutex_exit(&arc_reclaim_thr_lock); 4391 4392 arc_flush(NULL); 4393 4394 arc_dead = TRUE; 4395 4396 if (arc_ksp != NULL) { 4397 kstat_delete(arc_ksp); 4398 arc_ksp = NULL; 4399 } 4400 4401 mutex_destroy(&arc_eviction_mtx); 4402 mutex_destroy(&arc_reclaim_thr_lock); 4403 cv_destroy(&arc_reclaim_thr_cv); 4404 4405 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4406 list_destroy(&arc_mru->arcs_lists[i]); 4407 list_destroy(&arc_mru_ghost->arcs_lists[i]); 4408 list_destroy(&arc_mfu->arcs_lists[i]); 4409 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4410 list_destroy(&arc_l2c_only->arcs_lists[i]); 4411 4412 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4413 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4414 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4415 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4416 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4417 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4418 } 4419 4420 buf_fini(); 4421 4422 ASSERT(arc_loaned_bytes == 0); 4423 4424 mutex_destroy(&arc_lowmem_lock); 4425#ifdef _KERNEL 4426 if (arc_event_lowmem != NULL) 4427 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4428#endif 4429} 4430 4431/* 4432 * Level 2 ARC 4433 * 4434 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4435 * It uses dedicated storage devices to hold cached data, which are populated 4436 * using large infrequent writes. The main role of this cache is to boost 4437 * the performance of random read workloads. The intended L2ARC devices 4438 * include short-stroked disks, solid state disks, and other media with 4439 * substantially faster read latency than disk. 4440 * 4441 * +-----------------------+ 4442 * | ARC | 4443 * +-----------------------+ 4444 * | ^ ^ 4445 * | | | 4446 * l2arc_feed_thread() arc_read() 4447 * | | | 4448 * | l2arc read | 4449 * V | | 4450 * +---------------+ | 4451 * | L2ARC | | 4452 * +---------------+ | 4453 * | ^ | 4454 * l2arc_write() | | 4455 * | | | 4456 * V | | 4457 * +-------+ +-------+ 4458 * | vdev | | vdev | 4459 * | cache | | cache | 4460 * +-------+ +-------+ 4461 * +=========+ .-----. 4462 * : L2ARC : |-_____-| 4463 * : devices : | Disks | 4464 * +=========+ `-_____-' 4465 * 4466 * Read requests are satisfied from the following sources, in order: 4467 * 4468 * 1) ARC 4469 * 2) vdev cache of L2ARC devices 4470 * 3) L2ARC devices 4471 * 4) vdev cache of disks 4472 * 5) disks 4473 * 4474 * Some L2ARC device types exhibit extremely slow write performance. 4475 * To accommodate for this there are some significant differences between 4476 * the L2ARC and traditional cache design: 4477 * 4478 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4479 * the ARC behave as usual, freeing buffers and placing headers on ghost 4480 * lists. The ARC does not send buffers to the L2ARC during eviction as 4481 * this would add inflated write latencies for all ARC memory pressure. 4482 * 4483 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4484 * It does this by periodically scanning buffers from the eviction-end of 4485 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4486 * not already there. It scans until a headroom of buffers is satisfied, 4487 * which itself is a buffer for ARC eviction. If a compressible buffer is 4488 * found during scanning and selected for writing to an L2ARC device, we 4489 * temporarily boost scanning headroom during the next scan cycle to make 4490 * sure we adapt to compression effects (which might significantly reduce 4491 * the data volume we write to L2ARC). The thread that does this is 4492 * l2arc_feed_thread(), illustrated below; example sizes are included to 4493 * provide a better sense of ratio than this diagram: 4494 * 4495 * head --> tail 4496 * +---------------------+----------+ 4497 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4498 * +---------------------+----------+ | o L2ARC eligible 4499 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4500 * +---------------------+----------+ | 4501 * 15.9 Gbytes ^ 32 Mbytes | 4502 * headroom | 4503 * l2arc_feed_thread() 4504 * | 4505 * l2arc write hand <--[oooo]--' 4506 * | 8 Mbyte 4507 * | write max 4508 * V 4509 * +==============================+ 4510 * L2ARC dev |####|#|###|###| |####| ... | 4511 * +==============================+ 4512 * 32 Gbytes 4513 * 4514 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4515 * evicted, then the L2ARC has cached a buffer much sooner than it probably 4516 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4517 * safe to say that this is an uncommon case, since buffers at the end of 4518 * the ARC lists have moved there due to inactivity. 4519 * 4520 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4521 * then the L2ARC simply misses copying some buffers. This serves as a 4522 * pressure valve to prevent heavy read workloads from both stalling the ARC 4523 * with waits and clogging the L2ARC with writes. This also helps prevent 4524 * the potential for the L2ARC to churn if it attempts to cache content too 4525 * quickly, such as during backups of the entire pool. 4526 * 4527 * 5. After system boot and before the ARC has filled main memory, there are 4528 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4529 * lists can remain mostly static. Instead of searching from tail of these 4530 * lists as pictured, the l2arc_feed_thread() will search from the list heads 4531 * for eligible buffers, greatly increasing its chance of finding them. 4532 * 4533 * The L2ARC device write speed is also boosted during this time so that 4534 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4535 * there are no L2ARC reads, and no fear of degrading read performance 4536 * through increased writes. 4537 * 4538 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4539 * the vdev queue can aggregate them into larger and fewer writes. Each 4540 * device is written to in a rotor fashion, sweeping writes through 4541 * available space then repeating. 4542 * 4543 * 7. The L2ARC does not store dirty content. It never needs to flush 4544 * write buffers back to disk based storage. 4545 * 4546 * 8. If an ARC buffer is written (and dirtied) which also exists in the 4547 * L2ARC, the now stale L2ARC buffer is immediately dropped. 4548 * 4549 * The performance of the L2ARC can be tweaked by a number of tunables, which 4550 * may be necessary for different workloads: 4551 * 4552 * l2arc_write_max max write bytes per interval 4553 * l2arc_write_boost extra write bytes during device warmup 4554 * l2arc_noprefetch skip caching prefetched buffers 4555 * l2arc_headroom number of max device writes to precache 4556 * l2arc_headroom_boost when we find compressed buffers during ARC 4557 * scanning, we multiply headroom by this 4558 * percentage factor for the next scan cycle, 4559 * since more compressed buffers are likely to 4560 * be present 4561 * l2arc_feed_secs seconds between L2ARC writing 4562 * 4563 * Tunables may be removed or added as future performance improvements are 4564 * integrated, and also may become zpool properties. 4565 * 4566 * There are three key functions that control how the L2ARC warms up: 4567 * 4568 * l2arc_write_eligible() check if a buffer is eligible to cache 4569 * l2arc_write_size() calculate how much to write 4570 * l2arc_write_interval() calculate sleep delay between writes 4571 * 4572 * These three functions determine what to write, how much, and how quickly 4573 * to send writes. 4574 */ 4575 4576static boolean_t 4577l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 4578{ 4579 /* 4580 * A buffer is *not* eligible for the L2ARC if it: 4581 * 1. belongs to a different spa. 4582 * 2. is already cached on the L2ARC. 4583 * 3. has an I/O in progress (it may be an incomplete read). 4584 * 4. is flagged not eligible (zfs property). 4585 */ 4586 if (hdr->b_spa != spa_guid) { 4587 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4588 return (B_FALSE); 4589 } 4590 if (hdr->b_l2hdr != NULL) { 4591 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4592 return (B_FALSE); 4593 } 4594 if (HDR_IO_IN_PROGRESS(hdr)) { 4595 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4596 return (B_FALSE); 4597 } 4598 if (!HDR_L2CACHE(hdr)) { 4599 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4600 return (B_FALSE); 4601 } 4602 4603 return (B_TRUE); 4604} 4605 4606static uint64_t 4607l2arc_write_size(void) 4608{ 4609 uint64_t size; 4610 4611 /* 4612 * Make sure our globals have meaningful values in case the user 4613 * altered them. 4614 */ 4615 size = l2arc_write_max; 4616 if (size == 0) { 4617 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4618 "be greater than zero, resetting it to the default (%d)", 4619 L2ARC_WRITE_SIZE); 4620 size = l2arc_write_max = L2ARC_WRITE_SIZE; 4621 } 4622 4623 if (arc_warm == B_FALSE) 4624 size += l2arc_write_boost; 4625 4626 return (size); 4627 4628} 4629 4630static clock_t 4631l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4632{ 4633 clock_t interval, next, now; 4634 4635 /* 4636 * If the ARC lists are busy, increase our write rate; if the 4637 * lists are stale, idle back. This is achieved by checking 4638 * how much we previously wrote - if it was more than half of 4639 * what we wanted, schedule the next write much sooner. 4640 */ 4641 if (l2arc_feed_again && wrote > (wanted / 2)) 4642 interval = (hz * l2arc_feed_min_ms) / 1000; 4643 else 4644 interval = hz * l2arc_feed_secs; 4645 4646 now = ddi_get_lbolt(); 4647 next = MAX(now, MIN(now + interval, began + interval)); 4648 4649 return (next); 4650} 4651 4652static void 4653l2arc_hdr_stat_add(void) 4654{ 4655 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4656 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4657} 4658 4659static void 4660l2arc_hdr_stat_remove(void) 4661{ 4662 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4663 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4664} 4665 4666/* 4667 * Cycle through L2ARC devices. This is how L2ARC load balances. 4668 * If a device is returned, this also returns holding the spa config lock. 4669 */ 4670static l2arc_dev_t * 4671l2arc_dev_get_next(void) 4672{ 4673 l2arc_dev_t *first, *next = NULL; 4674 4675 /* 4676 * Lock out the removal of spas (spa_namespace_lock), then removal 4677 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4678 * both locks will be dropped and a spa config lock held instead. 4679 */ 4680 mutex_enter(&spa_namespace_lock); 4681 mutex_enter(&l2arc_dev_mtx); 4682 4683 /* if there are no vdevs, there is nothing to do */ 4684 if (l2arc_ndev == 0) 4685 goto out; 4686 4687 first = NULL; 4688 next = l2arc_dev_last; 4689 do { 4690 /* loop around the list looking for a non-faulted vdev */ 4691 if (next == NULL) { 4692 next = list_head(l2arc_dev_list); 4693 } else { 4694 next = list_next(l2arc_dev_list, next); 4695 if (next == NULL) 4696 next = list_head(l2arc_dev_list); 4697 } 4698 4699 /* if we have come back to the start, bail out */ 4700 if (first == NULL) 4701 first = next; 4702 else if (next == first) 4703 break; 4704 4705 } while (vdev_is_dead(next->l2ad_vdev)); 4706 4707 /* if we were unable to find any usable vdevs, return NULL */ 4708 if (vdev_is_dead(next->l2ad_vdev)) 4709 next = NULL; 4710 4711 l2arc_dev_last = next; 4712 4713out: 4714 mutex_exit(&l2arc_dev_mtx); 4715 4716 /* 4717 * Grab the config lock to prevent the 'next' device from being 4718 * removed while we are writing to it. 4719 */ 4720 if (next != NULL) 4721 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4722 mutex_exit(&spa_namespace_lock); 4723 4724 return (next); 4725} 4726 4727/* 4728 * Free buffers that were tagged for destruction. 4729 */ 4730static void 4731l2arc_do_free_on_write() 4732{ 4733 list_t *buflist; 4734 l2arc_data_free_t *df, *df_prev; 4735 4736 mutex_enter(&l2arc_free_on_write_mtx); 4737 buflist = l2arc_free_on_write; 4738 4739 for (df = list_tail(buflist); df; df = df_prev) { 4740 df_prev = list_prev(buflist, df); 4741 ASSERT(df->l2df_data != NULL); 4742 ASSERT(df->l2df_func != NULL); 4743 df->l2df_func(df->l2df_data, df->l2df_size); 4744 list_remove(buflist, df); 4745 kmem_free(df, sizeof (l2arc_data_free_t)); 4746 } 4747 4748 mutex_exit(&l2arc_free_on_write_mtx); 4749} 4750 4751/* 4752 * A write to a cache device has completed. Update all headers to allow 4753 * reads from these buffers to begin. 4754 */ 4755static void 4756l2arc_write_done(zio_t *zio) 4757{ 4758 l2arc_write_callback_t *cb; 4759 l2arc_dev_t *dev; 4760 list_t *buflist; 4761 arc_buf_hdr_t *head, *hdr, *hdr_prev; 4762 l2arc_buf_hdr_t *abl2; 4763 kmutex_t *hash_lock; 4764 int64_t bytes_dropped = 0; 4765 4766 cb = zio->io_private; 4767 ASSERT(cb != NULL); 4768 dev = cb->l2wcb_dev; 4769 ASSERT(dev != NULL); 4770 head = cb->l2wcb_head; 4771 ASSERT(head != NULL); 4772 buflist = dev->l2ad_buflist; 4773 ASSERT(buflist != NULL); 4774 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4775 l2arc_write_callback_t *, cb); 4776 4777 if (zio->io_error != 0) 4778 ARCSTAT_BUMP(arcstat_l2_writes_error); 4779 4780 mutex_enter(&l2arc_buflist_mtx); 4781 4782 /* 4783 * All writes completed, or an error was hit. 4784 */ 4785 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 4786 hdr_prev = list_prev(buflist, hdr); 4787 abl2 = hdr->b_l2hdr; 4788 4789 /* 4790 * Release the temporary compressed buffer as soon as possible. 4791 */ 4792 if (abl2->b_compress != ZIO_COMPRESS_OFF) 4793 l2arc_release_cdata_buf(hdr); 4794 4795 hash_lock = HDR_LOCK(hdr); 4796 if (!mutex_tryenter(hash_lock)) { 4797 /* 4798 * This buffer misses out. It may be in a stage 4799 * of eviction. Its ARC_L2_WRITING flag will be 4800 * left set, denying reads to this buffer. 4801 */ 4802 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4803 continue; 4804 } 4805 4806 if (zio->io_error != 0) { 4807 /* 4808 * Error - drop L2ARC entry. 4809 */ 4810 list_remove(buflist, hdr); 4811 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4812 bytes_dropped += abl2->b_asize; 4813 hdr->b_l2hdr = NULL; 4814 trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4815 abl2->b_asize, 0); 4816 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4817 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 4818 } 4819 4820 /* 4821 * Allow ARC to begin reads to this L2ARC entry. 4822 */ 4823 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 4824 4825 mutex_exit(hash_lock); 4826 } 4827 4828 atomic_inc_64(&l2arc_writes_done); 4829 list_remove(buflist, head); 4830 kmem_cache_free(hdr_cache, head); 4831 mutex_exit(&l2arc_buflist_mtx); 4832 4833 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 4834 4835 l2arc_do_free_on_write(); 4836 4837 kmem_free(cb, sizeof (l2arc_write_callback_t)); 4838} 4839 4840/* 4841 * A read to a cache device completed. Validate buffer contents before 4842 * handing over to the regular ARC routines. 4843 */ 4844static void 4845l2arc_read_done(zio_t *zio) 4846{ 4847 l2arc_read_callback_t *cb; 4848 arc_buf_hdr_t *hdr; 4849 arc_buf_t *buf; 4850 kmutex_t *hash_lock; 4851 int equal; 4852 4853 ASSERT(zio->io_vd != NULL); 4854 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4855 4856 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4857 4858 cb = zio->io_private; 4859 ASSERT(cb != NULL); 4860 buf = cb->l2rcb_buf; 4861 ASSERT(buf != NULL); 4862 4863 hash_lock = HDR_LOCK(buf->b_hdr); 4864 mutex_enter(hash_lock); 4865 hdr = buf->b_hdr; 4866 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4867 4868 /* 4869 * If the buffer was compressed, decompress it first. 4870 */ 4871 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4872 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4873 ASSERT(zio->io_data != NULL); 4874 4875 /* 4876 * Check this survived the L2ARC journey. 4877 */ 4878 equal = arc_cksum_equal(buf); 4879 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4880 mutex_exit(hash_lock); 4881 zio->io_private = buf; 4882 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4883 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4884 arc_read_done(zio); 4885 } else { 4886 mutex_exit(hash_lock); 4887 /* 4888 * Buffer didn't survive caching. Increment stats and 4889 * reissue to the original storage device. 4890 */ 4891 if (zio->io_error != 0) { 4892 ARCSTAT_BUMP(arcstat_l2_io_error); 4893 } else { 4894 zio->io_error = SET_ERROR(EIO); 4895 } 4896 if (!equal) 4897 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4898 4899 /* 4900 * If there's no waiter, issue an async i/o to the primary 4901 * storage now. If there *is* a waiter, the caller must 4902 * issue the i/o in a context where it's OK to block. 4903 */ 4904 if (zio->io_waiter == NULL) { 4905 zio_t *pio = zio_unique_parent(zio); 4906 4907 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4908 4909 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4910 buf->b_data, zio->io_size, arc_read_done, buf, 4911 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4912 } 4913 } 4914 4915 kmem_free(cb, sizeof (l2arc_read_callback_t)); 4916} 4917 4918/* 4919 * This is the list priority from which the L2ARC will search for pages to 4920 * cache. This is used within loops (0..3) to cycle through lists in the 4921 * desired order. This order can have a significant effect on cache 4922 * performance. 4923 * 4924 * Currently the metadata lists are hit first, MFU then MRU, followed by 4925 * the data lists. This function returns a locked list, and also returns 4926 * the lock pointer. 4927 */ 4928static list_t * 4929l2arc_list_locked(int list_num, kmutex_t **lock) 4930{ 4931 list_t *list = NULL; 4932 int idx; 4933 4934 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4935 4936 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4937 idx = list_num; 4938 list = &arc_mfu->arcs_lists[idx]; 4939 *lock = ARCS_LOCK(arc_mfu, idx); 4940 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4941 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4942 list = &arc_mru->arcs_lists[idx]; 4943 *lock = ARCS_LOCK(arc_mru, idx); 4944 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4945 ARC_BUFC_NUMDATALISTS)) { 4946 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4947 list = &arc_mfu->arcs_lists[idx]; 4948 *lock = ARCS_LOCK(arc_mfu, idx); 4949 } else { 4950 idx = list_num - ARC_BUFC_NUMLISTS; 4951 list = &arc_mru->arcs_lists[idx]; 4952 *lock = ARCS_LOCK(arc_mru, idx); 4953 } 4954 4955 ASSERT(!(MUTEX_HELD(*lock))); 4956 mutex_enter(*lock); 4957 return (list); 4958} 4959 4960/* 4961 * Evict buffers from the device write hand to the distance specified in 4962 * bytes. This distance may span populated buffers, it may span nothing. 4963 * This is clearing a region on the L2ARC device ready for writing. 4964 * If the 'all' boolean is set, every buffer is evicted. 4965 */ 4966static void 4967l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4968{ 4969 list_t *buflist; 4970 l2arc_buf_hdr_t *abl2; 4971 arc_buf_hdr_t *hdr, *hdr_prev; 4972 kmutex_t *hash_lock; 4973 uint64_t taddr; 4974 int64_t bytes_evicted = 0; 4975 4976 buflist = dev->l2ad_buflist; 4977 4978 if (buflist == NULL) 4979 return; 4980 4981 if (!all && dev->l2ad_first) { 4982 /* 4983 * This is the first sweep through the device. There is 4984 * nothing to evict. 4985 */ 4986 return; 4987 } 4988 4989 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4990 /* 4991 * When nearing the end of the device, evict to the end 4992 * before the device write hand jumps to the start. 4993 */ 4994 taddr = dev->l2ad_end; 4995 } else { 4996 taddr = dev->l2ad_hand + distance; 4997 } 4998 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4999 uint64_t, taddr, boolean_t, all); 5000 5001top: 5002 mutex_enter(&l2arc_buflist_mtx); 5003 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 5004 hdr_prev = list_prev(buflist, hdr); 5005 5006 hash_lock = HDR_LOCK(hdr); 5007 if (!mutex_tryenter(hash_lock)) { 5008 /* 5009 * Missed the hash lock. Retry. 5010 */ 5011 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 5012 mutex_exit(&l2arc_buflist_mtx); 5013 mutex_enter(hash_lock); 5014 mutex_exit(hash_lock); 5015 goto top; 5016 } 5017 5018 if (HDR_L2_WRITE_HEAD(hdr)) { 5019 /* 5020 * We hit a write head node. Leave it for 5021 * l2arc_write_done(). 5022 */ 5023 list_remove(buflist, hdr); 5024 mutex_exit(hash_lock); 5025 continue; 5026 } 5027 5028 if (!all && hdr->b_l2hdr != NULL && 5029 (hdr->b_l2hdr->b_daddr > taddr || 5030 hdr->b_l2hdr->b_daddr < dev->l2ad_hand)) { 5031 /* 5032 * We've evicted to the target address, 5033 * or the end of the device. 5034 */ 5035 mutex_exit(hash_lock); 5036 break; 5037 } 5038 5039 if (HDR_FREE_IN_PROGRESS(hdr)) { 5040 /* 5041 * Already on the path to destruction. 5042 */ 5043 mutex_exit(hash_lock); 5044 continue; 5045 } 5046 5047 if (hdr->b_state == arc_l2c_only) { 5048 ASSERT(!HDR_L2_READING(hdr)); 5049 /* 5050 * This doesn't exist in the ARC. Destroy. 5051 * arc_hdr_destroy() will call list_remove() 5052 * and decrement arcstat_l2_size. 5053 */ 5054 arc_change_state(arc_anon, hdr, hash_lock); 5055 arc_hdr_destroy(hdr); 5056 } else { 5057 /* 5058 * Invalidate issued or about to be issued 5059 * reads, since we may be about to write 5060 * over this location. 5061 */ 5062 if (HDR_L2_READING(hdr)) { 5063 ARCSTAT_BUMP(arcstat_l2_evict_reading); 5064 hdr->b_flags |= ARC_FLAG_L2_EVICTED; 5065 } 5066 5067 /* 5068 * Tell ARC this no longer exists in L2ARC. 5069 */ 5070 if (hdr->b_l2hdr != NULL) { 5071 abl2 = hdr->b_l2hdr; 5072 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 5073 bytes_evicted += abl2->b_asize; 5074 hdr->b_l2hdr = NULL; 5075 /* 5076 * We are destroying l2hdr, so ensure that 5077 * its compressed buffer, if any, is not leaked. 5078 */ 5079 ASSERT(abl2->b_tmp_cdata == NULL); 5080 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 5081 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 5082 } 5083 list_remove(buflist, hdr); 5084 5085 /* 5086 * This may have been leftover after a 5087 * failed write. 5088 */ 5089 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 5090 } 5091 mutex_exit(hash_lock); 5092 } 5093 mutex_exit(&l2arc_buflist_mtx); 5094 5095 vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); 5096 dev->l2ad_evict = taddr; 5097} 5098 5099/* 5100 * Find and write ARC buffers to the L2ARC device. 5101 * 5102 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 5103 * for reading until they have completed writing. 5104 * The headroom_boost is an in-out parameter used to maintain headroom boost 5105 * state between calls to this function. 5106 * 5107 * Returns the number of bytes actually written (which may be smaller than 5108 * the delta by which the device hand has changed due to alignment). 5109 */ 5110static uint64_t 5111l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 5112 boolean_t *headroom_boost) 5113{ 5114 arc_buf_hdr_t *hdr, *hdr_prev, *head; 5115 list_t *list; 5116 uint64_t write_asize, write_sz, headroom, buf_compress_minsz; 5117 void *buf_data; 5118 kmutex_t *list_lock; 5119 boolean_t full; 5120 l2arc_write_callback_t *cb; 5121 zio_t *pio, *wzio; 5122 uint64_t guid = spa_load_guid(spa); 5123 const boolean_t do_headroom_boost = *headroom_boost; 5124 int try; 5125 5126 ASSERT(dev->l2ad_vdev != NULL); 5127 5128 /* Lower the flag now, we might want to raise it again later. */ 5129 *headroom_boost = B_FALSE; 5130 5131 pio = NULL; 5132 write_sz = write_asize = 0; 5133 full = B_FALSE; 5134 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 5135 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; 5136 5137 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 5138 /* 5139 * We will want to try to compress buffers that are at least 2x the 5140 * device sector size. 5141 */ 5142 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 5143 5144 /* 5145 * Copy buffers for L2ARC writing. 5146 */ 5147 mutex_enter(&l2arc_buflist_mtx); 5148 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 5149 uint64_t passed_sz = 0; 5150 5151 list = l2arc_list_locked(try, &list_lock); 5152 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 5153 5154 /* 5155 * L2ARC fast warmup. 5156 * 5157 * Until the ARC is warm and starts to evict, read from the 5158 * head of the ARC lists rather than the tail. 5159 */ 5160 if (arc_warm == B_FALSE) 5161 hdr = list_head(list); 5162 else 5163 hdr = list_tail(list); 5164 if (hdr == NULL) 5165 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 5166 5167 headroom = target_sz * l2arc_headroom * 2 / ARC_BUFC_NUMLISTS; 5168 if (do_headroom_boost) 5169 headroom = (headroom * l2arc_headroom_boost) / 100; 5170 5171 for (; hdr; hdr = hdr_prev) { 5172 l2arc_buf_hdr_t *l2hdr; 5173 kmutex_t *hash_lock; 5174 uint64_t buf_sz; 5175 uint64_t buf_a_sz; 5176 5177 if (arc_warm == B_FALSE) 5178 hdr_prev = list_next(list, hdr); 5179 else 5180 hdr_prev = list_prev(list, hdr); 5181 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size); 5182 5183 hash_lock = HDR_LOCK(hdr); 5184 if (!mutex_tryenter(hash_lock)) { 5185 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5186 /* 5187 * Skip this buffer rather than waiting. 5188 */ 5189 continue; 5190 } 5191 5192 passed_sz += hdr->b_size; 5193 if (passed_sz > headroom) { 5194 /* 5195 * Searched too far. 5196 */ 5197 mutex_exit(hash_lock); 5198 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5199 break; 5200 } 5201 5202 if (!l2arc_write_eligible(guid, hdr)) { 5203 mutex_exit(hash_lock); 5204 continue; 5205 } 5206 5207 /* 5208 * Assume that the buffer is not going to be compressed 5209 * and could take more space on disk because of a larger 5210 * disk block size. 5211 */ 5212 buf_sz = hdr->b_size; 5213 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5214 5215 if ((write_asize + buf_a_sz) > target_sz) { 5216 full = B_TRUE; 5217 mutex_exit(hash_lock); 5218 ARCSTAT_BUMP(arcstat_l2_write_full); 5219 break; 5220 } 5221 5222 if (pio == NULL) { 5223 /* 5224 * Insert a dummy header on the buflist so 5225 * l2arc_write_done() can find where the 5226 * write buffers begin without searching. 5227 */ 5228 list_insert_head(dev->l2ad_buflist, head); 5229 5230 cb = kmem_alloc( 5231 sizeof (l2arc_write_callback_t), KM_SLEEP); 5232 cb->l2wcb_dev = dev; 5233 cb->l2wcb_head = head; 5234 pio = zio_root(spa, l2arc_write_done, cb, 5235 ZIO_FLAG_CANFAIL); 5236 ARCSTAT_BUMP(arcstat_l2_write_pios); 5237 } 5238 5239 /* 5240 * Create and add a new L2ARC header. 5241 */ 5242 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5243 l2hdr->b_dev = dev; 5244 hdr->b_flags |= ARC_FLAG_L2_WRITING; 5245 5246 /* 5247 * Temporarily stash the data buffer in b_tmp_cdata. 5248 * The subsequent write step will pick it up from 5249 * there. This is because can't access hdr->b_buf 5250 * without holding the hash_lock, which we in turn 5251 * can't access without holding the ARC list locks 5252 * (which we want to avoid during compression/writing). 5253 */ 5254 l2hdr->b_compress = ZIO_COMPRESS_OFF; 5255 l2hdr->b_asize = hdr->b_size; 5256 l2hdr->b_tmp_cdata = hdr->b_buf->b_data; 5257 5258 hdr->b_l2hdr = l2hdr; 5259 5260 list_insert_head(dev->l2ad_buflist, hdr); 5261 5262 /* 5263 * Compute and store the buffer cksum before 5264 * writing. On debug the cksum is verified first. 5265 */ 5266 arc_cksum_verify(hdr->b_buf); 5267 arc_cksum_compute(hdr->b_buf, B_TRUE); 5268 5269 mutex_exit(hash_lock); 5270 5271 write_sz += buf_sz; 5272 write_asize += buf_a_sz; 5273 } 5274 5275 mutex_exit(list_lock); 5276 5277 if (full == B_TRUE) 5278 break; 5279 } 5280 5281 /* No buffers selected for writing? */ 5282 if (pio == NULL) { 5283 ASSERT0(write_sz); 5284 mutex_exit(&l2arc_buflist_mtx); 5285 kmem_cache_free(hdr_cache, head); 5286 return (0); 5287 } 5288 5289 /* 5290 * Note that elsewhere in this file arcstat_l2_asize 5291 * and the used space on l2ad_vdev are updated using b_asize, 5292 * which is not necessarily rounded up to the device block size. 5293 * Too keep accounting consistent we do the same here as well: 5294 * stats_size accumulates the sum of b_asize of the written buffers, 5295 * while write_asize accumulates the sum of b_asize rounded up 5296 * to the device block size. 5297 * The latter sum is used only to validate the corectness of the code. 5298 */ 5299 uint64_t stats_size = 0; 5300 write_asize = 0; 5301 5302 /* 5303 * Now start writing the buffers. We're starting at the write head 5304 * and work backwards, retracing the course of the buffer selector 5305 * loop above. 5306 */ 5307 for (hdr = list_prev(dev->l2ad_buflist, head); hdr; 5308 hdr = list_prev(dev->l2ad_buflist, hdr)) { 5309 l2arc_buf_hdr_t *l2hdr; 5310 uint64_t buf_sz; 5311 5312 /* 5313 * We shouldn't need to lock the buffer here, since we flagged 5314 * it as ARC_FLAG_L2_WRITING in the previous step, but we must 5315 * take care to only access its L2 cache parameters. In 5316 * particular, hdr->b_buf may be invalid by now due to 5317 * ARC eviction. 5318 */ 5319 l2hdr = hdr->b_l2hdr; 5320 l2hdr->b_daddr = dev->l2ad_hand; 5321 5322 if ((hdr->b_flags & ARC_FLAG_L2COMPRESS) && 5323 l2hdr->b_asize >= buf_compress_minsz) { 5324 if (l2arc_compress_buf(l2hdr)) { 5325 /* 5326 * If compression succeeded, enable headroom 5327 * boost on the next scan cycle. 5328 */ 5329 *headroom_boost = B_TRUE; 5330 } 5331 } 5332 5333 /* 5334 * Pick up the buffer data we had previously stashed away 5335 * (and now potentially also compressed). 5336 */ 5337 buf_data = l2hdr->b_tmp_cdata; 5338 buf_sz = l2hdr->b_asize; 5339 5340 /* 5341 * If the data has not been compressed, then clear b_tmp_cdata 5342 * to make sure that it points only to a temporary compression 5343 * buffer. 5344 */ 5345 if (!L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)) 5346 l2hdr->b_tmp_cdata = NULL; 5347 5348 /* Compression may have squashed the buffer to zero length. */ 5349 if (buf_sz != 0) { 5350 uint64_t buf_a_sz; 5351 5352 wzio = zio_write_phys(pio, dev->l2ad_vdev, 5353 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5354 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5355 ZIO_FLAG_CANFAIL, B_FALSE); 5356 5357 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5358 zio_t *, wzio); 5359 (void) zio_nowait(wzio); 5360 5361 stats_size += buf_sz; 5362 /* 5363 * Keep the clock hand suitably device-aligned. 5364 */ 5365 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5366 write_asize += buf_a_sz; 5367 dev->l2ad_hand += buf_a_sz; 5368 } 5369 } 5370 5371 mutex_exit(&l2arc_buflist_mtx); 5372 5373 ASSERT3U(write_asize, <=, target_sz); 5374 ARCSTAT_BUMP(arcstat_l2_writes_sent); 5375 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5376 ARCSTAT_INCR(arcstat_l2_size, write_sz); 5377 ARCSTAT_INCR(arcstat_l2_asize, stats_size); 5378 vdev_space_update(dev->l2ad_vdev, stats_size, 0, 0); 5379 5380 /* 5381 * Bump device hand to the device start if it is approaching the end. 5382 * l2arc_evict() will already have evicted ahead for this case. 5383 */ 5384 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5385 dev->l2ad_hand = dev->l2ad_start; 5386 dev->l2ad_evict = dev->l2ad_start; 5387 dev->l2ad_first = B_FALSE; 5388 } 5389 5390 dev->l2ad_writing = B_TRUE; 5391 (void) zio_wait(pio); 5392 dev->l2ad_writing = B_FALSE; 5393 5394 return (write_asize); 5395} 5396 5397/* 5398 * Compresses an L2ARC buffer. 5399 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5400 * size in l2hdr->b_asize. This routine tries to compress the data and 5401 * depending on the compression result there are three possible outcomes: 5402 * *) The buffer was incompressible. The original l2hdr contents were left 5403 * untouched and are ready for writing to an L2 device. 5404 * *) The buffer was all-zeros, so there is no need to write it to an L2 5405 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5406 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5407 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5408 * data buffer which holds the compressed data to be written, and b_asize 5409 * tells us how much data there is. b_compress is set to the appropriate 5410 * compression algorithm. Once writing is done, invoke 5411 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5412 * 5413 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5414 * buffer was incompressible). 5415 */ 5416static boolean_t 5417l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5418{ 5419 void *cdata; 5420 size_t csize, len, rounded; 5421 5422 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5423 ASSERT(l2hdr->b_tmp_cdata != NULL); 5424 5425 len = l2hdr->b_asize; 5426 cdata = zio_data_buf_alloc(len); 5427 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5428 cdata, l2hdr->b_asize); 5429 5430 if (csize == 0) { 5431 /* zero block, indicate that there's nothing to write */ 5432 zio_data_buf_free(cdata, len); 5433 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5434 l2hdr->b_asize = 0; 5435 l2hdr->b_tmp_cdata = NULL; 5436 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5437 return (B_TRUE); 5438 } 5439 5440 rounded = P2ROUNDUP(csize, 5441 (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift); 5442 if (rounded < len) { 5443 /* 5444 * Compression succeeded, we'll keep the cdata around for 5445 * writing and release it afterwards. 5446 */ 5447 if (rounded > csize) { 5448 bzero((char *)cdata + csize, rounded - csize); 5449 csize = rounded; 5450 } 5451 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5452 l2hdr->b_asize = csize; 5453 l2hdr->b_tmp_cdata = cdata; 5454 ARCSTAT_BUMP(arcstat_l2_compress_successes); 5455 return (B_TRUE); 5456 } else { 5457 /* 5458 * Compression failed, release the compressed buffer. 5459 * l2hdr will be left unmodified. 5460 */ 5461 zio_data_buf_free(cdata, len); 5462 ARCSTAT_BUMP(arcstat_l2_compress_failures); 5463 return (B_FALSE); 5464 } 5465} 5466 5467/* 5468 * Decompresses a zio read back from an l2arc device. On success, the 5469 * underlying zio's io_data buffer is overwritten by the uncompressed 5470 * version. On decompression error (corrupt compressed stream), the 5471 * zio->io_error value is set to signal an I/O error. 5472 * 5473 * Please note that the compressed data stream is not checksummed, so 5474 * if the underlying device is experiencing data corruption, we may feed 5475 * corrupt data to the decompressor, so the decompressor needs to be 5476 * able to handle this situation (LZ4 does). 5477 */ 5478static void 5479l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5480{ 5481 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5482 5483 if (zio->io_error != 0) { 5484 /* 5485 * An io error has occured, just restore the original io 5486 * size in preparation for a main pool read. 5487 */ 5488 zio->io_orig_size = zio->io_size = hdr->b_size; 5489 return; 5490 } 5491 5492 if (c == ZIO_COMPRESS_EMPTY) { 5493 /* 5494 * An empty buffer results in a null zio, which means we 5495 * need to fill its io_data after we're done restoring the 5496 * buffer's contents. 5497 */ 5498 ASSERT(hdr->b_buf != NULL); 5499 bzero(hdr->b_buf->b_data, hdr->b_size); 5500 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5501 } else { 5502 ASSERT(zio->io_data != NULL); 5503 /* 5504 * We copy the compressed data from the start of the arc buffer 5505 * (the zio_read will have pulled in only what we need, the 5506 * rest is garbage which we will overwrite at decompression) 5507 * and then decompress back to the ARC data buffer. This way we 5508 * can minimize copying by simply decompressing back over the 5509 * original compressed data (rather than decompressing to an 5510 * aux buffer and then copying back the uncompressed buffer, 5511 * which is likely to be much larger). 5512 */ 5513 uint64_t csize; 5514 void *cdata; 5515 5516 csize = zio->io_size; 5517 cdata = zio_data_buf_alloc(csize); 5518 bcopy(zio->io_data, cdata, csize); 5519 if (zio_decompress_data(c, cdata, zio->io_data, csize, 5520 hdr->b_size) != 0) 5521 zio->io_error = EIO; 5522 zio_data_buf_free(cdata, csize); 5523 } 5524 5525 /* Restore the expected uncompressed IO size. */ 5526 zio->io_orig_size = zio->io_size = hdr->b_size; 5527} 5528 5529/* 5530 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5531 * This buffer serves as a temporary holder of compressed data while 5532 * the buffer entry is being written to an l2arc device. Once that is 5533 * done, we can dispose of it. 5534 */ 5535static void 5536l2arc_release_cdata_buf(arc_buf_hdr_t *hdr) 5537{ 5538 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 5539 5540 ASSERT(L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)); 5541 if (l2hdr->b_compress != ZIO_COMPRESS_EMPTY) { 5542 /* 5543 * If the data was compressed, then we've allocated a 5544 * temporary buffer for it, so now we need to release it. 5545 */ 5546 ASSERT(l2hdr->b_tmp_cdata != NULL); 5547 zio_data_buf_free(l2hdr->b_tmp_cdata, hdr->b_size); 5548 l2hdr->b_tmp_cdata = NULL; 5549 } else { 5550 ASSERT(l2hdr->b_tmp_cdata == NULL); 5551 } 5552} 5553 5554/* 5555 * This thread feeds the L2ARC at regular intervals. This is the beating 5556 * heart of the L2ARC. 5557 */ 5558static void 5559l2arc_feed_thread(void *dummy __unused) 5560{ 5561 callb_cpr_t cpr; 5562 l2arc_dev_t *dev; 5563 spa_t *spa; 5564 uint64_t size, wrote; 5565 clock_t begin, next = ddi_get_lbolt(); 5566 boolean_t headroom_boost = B_FALSE; 5567 5568 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5569 5570 mutex_enter(&l2arc_feed_thr_lock); 5571 5572 while (l2arc_thread_exit == 0) { 5573 CALLB_CPR_SAFE_BEGIN(&cpr); 5574 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5575 next - ddi_get_lbolt()); 5576 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5577 next = ddi_get_lbolt() + hz; 5578 5579 /* 5580 * Quick check for L2ARC devices. 5581 */ 5582 mutex_enter(&l2arc_dev_mtx); 5583 if (l2arc_ndev == 0) { 5584 mutex_exit(&l2arc_dev_mtx); 5585 continue; 5586 } 5587 mutex_exit(&l2arc_dev_mtx); 5588 begin = ddi_get_lbolt(); 5589 5590 /* 5591 * This selects the next l2arc device to write to, and in 5592 * doing so the next spa to feed from: dev->l2ad_spa. This 5593 * will return NULL if there are now no l2arc devices or if 5594 * they are all faulted. 5595 * 5596 * If a device is returned, its spa's config lock is also 5597 * held to prevent device removal. l2arc_dev_get_next() 5598 * will grab and release l2arc_dev_mtx. 5599 */ 5600 if ((dev = l2arc_dev_get_next()) == NULL) 5601 continue; 5602 5603 spa = dev->l2ad_spa; 5604 ASSERT(spa != NULL); 5605 5606 /* 5607 * If the pool is read-only then force the feed thread to 5608 * sleep a little longer. 5609 */ 5610 if (!spa_writeable(spa)) { 5611 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5612 spa_config_exit(spa, SCL_L2ARC, dev); 5613 continue; 5614 } 5615 5616 /* 5617 * Avoid contributing to memory pressure. 5618 */ 5619 if (arc_reclaim_needed()) { 5620 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5621 spa_config_exit(spa, SCL_L2ARC, dev); 5622 continue; 5623 } 5624 5625 ARCSTAT_BUMP(arcstat_l2_feeds); 5626 5627 size = l2arc_write_size(); 5628 5629 /* 5630 * Evict L2ARC buffers that will be overwritten. 5631 */ 5632 l2arc_evict(dev, size, B_FALSE); 5633 5634 /* 5635 * Write ARC buffers. 5636 */ 5637 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5638 5639 /* 5640 * Calculate interval between writes. 5641 */ 5642 next = l2arc_write_interval(begin, size, wrote); 5643 spa_config_exit(spa, SCL_L2ARC, dev); 5644 } 5645 5646 l2arc_thread_exit = 0; 5647 cv_broadcast(&l2arc_feed_thr_cv); 5648 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5649 thread_exit(); 5650} 5651 5652boolean_t 5653l2arc_vdev_present(vdev_t *vd) 5654{ 5655 l2arc_dev_t *dev; 5656 5657 mutex_enter(&l2arc_dev_mtx); 5658 for (dev = list_head(l2arc_dev_list); dev != NULL; 5659 dev = list_next(l2arc_dev_list, dev)) { 5660 if (dev->l2ad_vdev == vd) 5661 break; 5662 } 5663 mutex_exit(&l2arc_dev_mtx); 5664 5665 return (dev != NULL); 5666} 5667 5668/* 5669 * Add a vdev for use by the L2ARC. By this point the spa has already 5670 * validated the vdev and opened it. 5671 */ 5672void 5673l2arc_add_vdev(spa_t *spa, vdev_t *vd) 5674{ 5675 l2arc_dev_t *adddev; 5676 5677 ASSERT(!l2arc_vdev_present(vd)); 5678 5679 vdev_ashift_optimize(vd); 5680 5681 /* 5682 * Create a new l2arc device entry. 5683 */ 5684 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5685 adddev->l2ad_spa = spa; 5686 adddev->l2ad_vdev = vd; 5687 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5688 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5689 adddev->l2ad_hand = adddev->l2ad_start; 5690 adddev->l2ad_evict = adddev->l2ad_start; 5691 adddev->l2ad_first = B_TRUE; 5692 adddev->l2ad_writing = B_FALSE; 5693 5694 /* 5695 * This is a list of all ARC buffers that are still valid on the 5696 * device. 5697 */ 5698 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5699 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5700 offsetof(arc_buf_hdr_t, b_l2node)); 5701 5702 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5703 5704 /* 5705 * Add device to global list 5706 */ 5707 mutex_enter(&l2arc_dev_mtx); 5708 list_insert_head(l2arc_dev_list, adddev); 5709 atomic_inc_64(&l2arc_ndev); 5710 mutex_exit(&l2arc_dev_mtx); 5711} 5712 5713/* 5714 * Remove a vdev from the L2ARC. 5715 */ 5716void 5717l2arc_remove_vdev(vdev_t *vd) 5718{ 5719 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5720 5721 /* 5722 * Find the device by vdev 5723 */ 5724 mutex_enter(&l2arc_dev_mtx); 5725 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5726 nextdev = list_next(l2arc_dev_list, dev); 5727 if (vd == dev->l2ad_vdev) { 5728 remdev = dev; 5729 break; 5730 } 5731 } 5732 ASSERT(remdev != NULL); 5733 5734 /* 5735 * Remove device from global list 5736 */ 5737 list_remove(l2arc_dev_list, remdev); 5738 l2arc_dev_last = NULL; /* may have been invalidated */ 5739 atomic_dec_64(&l2arc_ndev); 5740 mutex_exit(&l2arc_dev_mtx); 5741 5742 /* 5743 * Clear all buflists and ARC references. L2ARC device flush. 5744 */ 5745 l2arc_evict(remdev, 0, B_TRUE); 5746 list_destroy(remdev->l2ad_buflist); 5747 kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5748 kmem_free(remdev, sizeof (l2arc_dev_t)); 5749} 5750 5751void 5752l2arc_init(void) 5753{ 5754 l2arc_thread_exit = 0; 5755 l2arc_ndev = 0; 5756 l2arc_writes_sent = 0; 5757 l2arc_writes_done = 0; 5758 5759 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5760 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5761 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5762 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5763 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5764 5765 l2arc_dev_list = &L2ARC_dev_list; 5766 l2arc_free_on_write = &L2ARC_free_on_write; 5767 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5768 offsetof(l2arc_dev_t, l2ad_node)); 5769 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5770 offsetof(l2arc_data_free_t, l2df_list_node)); 5771} 5772 5773void 5774l2arc_fini(void) 5775{ 5776 /* 5777 * This is called from dmu_fini(), which is called from spa_fini(); 5778 * Because of this, we can assume that all l2arc devices have 5779 * already been removed when the pools themselves were removed. 5780 */ 5781 5782 l2arc_do_free_on_write(); 5783 5784 mutex_destroy(&l2arc_feed_thr_lock); 5785 cv_destroy(&l2arc_feed_thr_cv); 5786 mutex_destroy(&l2arc_dev_mtx); 5787 mutex_destroy(&l2arc_buflist_mtx); 5788 mutex_destroy(&l2arc_free_on_write_mtx); 5789 5790 list_destroy(l2arc_dev_list); 5791 list_destroy(l2arc_free_on_write); 5792} 5793 5794void 5795l2arc_start(void) 5796{ 5797 if (!(spa_mode_global & FWRITE)) 5798 return; 5799 5800 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5801 TS_RUN, minclsyspri); 5802} 5803 5804void 5805l2arc_stop(void) 5806{ 5807 if (!(spa_mode_global & FWRITE)) 5808 return; 5809 5810 mutex_enter(&l2arc_feed_thr_lock); 5811 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5812 l2arc_thread_exit = 1; 5813 while (l2arc_thread_exit != 0) 5814 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5815 mutex_exit(&l2arc_feed_thr_lock); 5816} 5817