arc.c revision 288518
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 24 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 25 * Copyright 2014 Nexenta Systems, Inc. All rights reserved. 26 */ 27 28/* 29 * DVA-based Adjustable Replacement Cache 30 * 31 * While much of the theory of operation used here is 32 * based on the self-tuning, low overhead replacement cache 33 * presented by Megiddo and Modha at FAST 2003, there are some 34 * significant differences: 35 * 36 * 1. The Megiddo and Modha model assumes any page is evictable. 37 * Pages in its cache cannot be "locked" into memory. This makes 38 * the eviction algorithm simple: evict the last page in the list. 39 * This also make the performance characteristics easy to reason 40 * about. Our cache is not so simple. At any given moment, some 41 * subset of the blocks in the cache are un-evictable because we 42 * have handed out a reference to them. Blocks are only evictable 43 * when there are no external references active. This makes 44 * eviction far more problematic: we choose to evict the evictable 45 * blocks that are the "lowest" in the list. 46 * 47 * There are times when it is not possible to evict the requested 48 * space. In these circumstances we are unable to adjust the cache 49 * size. To prevent the cache growing unbounded at these times we 50 * implement a "cache throttle" that slows the flow of new data 51 * into the cache until we can make space available. 52 * 53 * 2. The Megiddo and Modha model assumes a fixed cache size. 54 * Pages are evicted when the cache is full and there is a cache 55 * miss. Our model has a variable sized cache. It grows with 56 * high use, but also tries to react to memory pressure from the 57 * operating system: decreasing its size when system memory is 58 * tight. 59 * 60 * 3. The Megiddo and Modha model assumes a fixed page size. All 61 * elements of the cache are therefore exactly the same size. So 62 * when adjusting the cache size following a cache miss, its simply 63 * a matter of choosing a single page to evict. In our model, we 64 * have variable sized cache blocks (rangeing from 512 bytes to 65 * 128K bytes). We therefore choose a set of blocks to evict to make 66 * space for a cache miss that approximates as closely as possible 67 * the space used by the new block. 68 * 69 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 70 * by N. Megiddo & D. Modha, FAST 2003 71 */ 72 73/* 74 * The locking model: 75 * 76 * A new reference to a cache buffer can be obtained in two 77 * ways: 1) via a hash table lookup using the DVA as a key, 78 * or 2) via one of the ARC lists. The arc_read() interface 79 * uses method 1, while the internal arc algorithms for 80 * adjusting the cache use method 2. We therefore provide two 81 * types of locks: 1) the hash table lock array, and 2) the 82 * arc list locks. 83 * 84 * Buffers do not have their own mutexs, rather they rely on the 85 * hash table mutexs for the bulk of their protection (i.e. most 86 * fields in the arc_buf_hdr_t are protected by these mutexs). 87 * 88 * buf_hash_find() returns the appropriate mutex (held) when it 89 * locates the requested buffer in the hash table. It returns 90 * NULL for the mutex if the buffer was not in the table. 91 * 92 * buf_hash_remove() expects the appropriate hash mutex to be 93 * already held before it is invoked. 94 * 95 * Each arc state also has a mutex which is used to protect the 96 * buffer list associated with the state. When attempting to 97 * obtain a hash table lock while holding an arc list lock you 98 * must use: mutex_tryenter() to avoid deadlock. Also note that 99 * the active state mutex must be held before the ghost state mutex. 100 * 101 * Arc buffers may have an associated eviction callback function. 102 * This function will be invoked prior to removing the buffer (e.g. 103 * in arc_do_user_evicts()). Note however that the data associated 104 * with the buffer may be evicted prior to the callback. The callback 105 * must be made with *no locks held* (to prevent deadlock). Additionally, 106 * the users of callbacks must ensure that their private data is 107 * protected from simultaneous callbacks from arc_clear_callback() 108 * and arc_do_user_evicts(). 109 * 110 * Note that the majority of the performance stats are manipulated 111 * with atomic operations. 112 * 113 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following: 114 * 115 * - L2ARC buflist creation 116 * - L2ARC buflist eviction 117 * - L2ARC write completion, which walks L2ARC buflists 118 * - ARC header destruction, as it removes from L2ARC buflists 119 * - ARC header release, as it removes from L2ARC buflists 120 */ 121 122#include <sys/spa.h> 123#include <sys/zio.h> 124#include <sys/zio_compress.h> 125#include <sys/zfs_context.h> 126#include <sys/arc.h> 127#include <sys/refcount.h> 128#include <sys/vdev.h> 129#include <sys/vdev_impl.h> 130#include <sys/dsl_pool.h> 131#ifdef _KERNEL 132#include <sys/dnlc.h> 133#endif 134#include <sys/callb.h> 135#include <sys/kstat.h> 136#include <sys/trim_map.h> 137#include <zfs_fletcher.h> 138#include <sys/sdt.h> 139 140#include <vm/vm_pageout.h> 141#include <machine/vmparam.h> 142 143#ifdef illumos 144#ifndef _KERNEL 145/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 146boolean_t arc_watch = B_FALSE; 147int arc_procfd; 148#endif 149#endif /* illumos */ 150 151static kmutex_t arc_reclaim_thr_lock; 152static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 153static uint8_t arc_thread_exit; 154 155#define ARC_REDUCE_DNLC_PERCENT 3 156uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 157 158typedef enum arc_reclaim_strategy { 159 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 160 ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 161} arc_reclaim_strategy_t; 162 163/* 164 * The number of iterations through arc_evict_*() before we 165 * drop & reacquire the lock. 166 */ 167int arc_evict_iterations = 100; 168 169/* number of seconds before growing cache again */ 170static int arc_grow_retry = 60; 171 172/* shift of arc_c for calculating both min and max arc_p */ 173static int arc_p_min_shift = 4; 174 175/* log2(fraction of arc to reclaim) */ 176static int arc_shrink_shift = 5; 177 178/* 179 * minimum lifespan of a prefetch block in clock ticks 180 * (initialized in arc_init()) 181 */ 182static int arc_min_prefetch_lifespan; 183 184/* 185 * If this percent of memory is free, don't throttle. 186 */ 187int arc_lotsfree_percent = 10; 188 189static int arc_dead; 190extern int zfs_prefetch_disable; 191 192/* 193 * The arc has filled available memory and has now warmed up. 194 */ 195static boolean_t arc_warm; 196 197uint64_t zfs_arc_max; 198uint64_t zfs_arc_min; 199uint64_t zfs_arc_meta_limit = 0; 200uint64_t zfs_arc_meta_min = 0; 201int zfs_arc_grow_retry = 0; 202int zfs_arc_shrink_shift = 0; 203int zfs_arc_p_min_shift = 0; 204int zfs_disable_dup_eviction = 0; 205uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 206u_int zfs_arc_free_target = 0; 207 208static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 209static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 210 211#ifdef _KERNEL 212static void 213arc_free_target_init(void *unused __unused) 214{ 215 216 zfs_arc_free_target = vm_pageout_wakeup_thresh; 217} 218SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 219 arc_free_target_init, NULL); 220 221TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max); 222TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 223TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 224TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 225TUNABLE_QUAD("vfs.zfs.arc_average_blocksize", &zfs_arc_average_blocksize); 226TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 227SYSCTL_DECL(_vfs_zfs); 228SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 229 "Maximum ARC size"); 230SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 231 "Minimum ARC size"); 232SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 233 &zfs_arc_average_blocksize, 0, 234 "ARC average blocksize"); 235SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 236 &arc_shrink_shift, 0, 237 "log2(fraction of arc to reclaim)"); 238 239/* 240 * We don't have a tunable for arc_free_target due to the dependency on 241 * pagedaemon initialisation. 242 */ 243SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 244 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 245 sysctl_vfs_zfs_arc_free_target, "IU", 246 "Desired number of free pages below which ARC triggers reclaim"); 247 248static int 249sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 250{ 251 u_int val; 252 int err; 253 254 val = zfs_arc_free_target; 255 err = sysctl_handle_int(oidp, &val, 0, req); 256 if (err != 0 || req->newptr == NULL) 257 return (err); 258 259 if (val < minfree) 260 return (EINVAL); 261 if (val > cnt.v_page_count) 262 return (EINVAL); 263 264 zfs_arc_free_target = val; 265 266 return (0); 267} 268 269/* 270 * Must be declared here, before the definition of corresponding kstat 271 * macro which uses the same names will confuse the compiler. 272 */ 273SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 274 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 275 sysctl_vfs_zfs_arc_meta_limit, "QU", 276 "ARC metadata limit"); 277#endif 278 279/* 280 * Note that buffers can be in one of 6 states: 281 * ARC_anon - anonymous (discussed below) 282 * ARC_mru - recently used, currently cached 283 * ARC_mru_ghost - recentely used, no longer in cache 284 * ARC_mfu - frequently used, currently cached 285 * ARC_mfu_ghost - frequently used, no longer in cache 286 * ARC_l2c_only - exists in L2ARC but not other states 287 * When there are no active references to the buffer, they are 288 * are linked onto a list in one of these arc states. These are 289 * the only buffers that can be evicted or deleted. Within each 290 * state there are multiple lists, one for meta-data and one for 291 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 292 * etc.) is tracked separately so that it can be managed more 293 * explicitly: favored over data, limited explicitly. 294 * 295 * Anonymous buffers are buffers that are not associated with 296 * a DVA. These are buffers that hold dirty block copies 297 * before they are written to stable storage. By definition, 298 * they are "ref'd" and are considered part of arc_mru 299 * that cannot be freed. Generally, they will aquire a DVA 300 * as they are written and migrate onto the arc_mru list. 301 * 302 * The ARC_l2c_only state is for buffers that are in the second 303 * level ARC but no longer in any of the ARC_m* lists. The second 304 * level ARC itself may also contain buffers that are in any of 305 * the ARC_m* states - meaning that a buffer can exist in two 306 * places. The reason for the ARC_l2c_only state is to keep the 307 * buffer header in the hash table, so that reads that hit the 308 * second level ARC benefit from these fast lookups. 309 */ 310 311#define ARCS_LOCK_PAD CACHE_LINE_SIZE 312struct arcs_lock { 313 kmutex_t arcs_lock; 314#ifdef _KERNEL 315 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 316#endif 317}; 318 319/* 320 * must be power of two for mask use to work 321 * 322 */ 323#define ARC_BUFC_NUMDATALISTS 16 324#define ARC_BUFC_NUMMETADATALISTS 16 325#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 326 327typedef struct arc_state { 328 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 329 uint64_t arcs_size; /* total amount of data in this state */ 330 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 331 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 332} arc_state_t; 333 334#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 335 336/* The 6 states: */ 337static arc_state_t ARC_anon; 338static arc_state_t ARC_mru; 339static arc_state_t ARC_mru_ghost; 340static arc_state_t ARC_mfu; 341static arc_state_t ARC_mfu_ghost; 342static arc_state_t ARC_l2c_only; 343 344typedef struct arc_stats { 345 kstat_named_t arcstat_hits; 346 kstat_named_t arcstat_misses; 347 kstat_named_t arcstat_demand_data_hits; 348 kstat_named_t arcstat_demand_data_misses; 349 kstat_named_t arcstat_demand_metadata_hits; 350 kstat_named_t arcstat_demand_metadata_misses; 351 kstat_named_t arcstat_prefetch_data_hits; 352 kstat_named_t arcstat_prefetch_data_misses; 353 kstat_named_t arcstat_prefetch_metadata_hits; 354 kstat_named_t arcstat_prefetch_metadata_misses; 355 kstat_named_t arcstat_mru_hits; 356 kstat_named_t arcstat_mru_ghost_hits; 357 kstat_named_t arcstat_mfu_hits; 358 kstat_named_t arcstat_mfu_ghost_hits; 359 kstat_named_t arcstat_allocated; 360 kstat_named_t arcstat_deleted; 361 kstat_named_t arcstat_stolen; 362 kstat_named_t arcstat_recycle_miss; 363 /* 364 * Number of buffers that could not be evicted because the hash lock 365 * was held by another thread. The lock may not necessarily be held 366 * by something using the same buffer, since hash locks are shared 367 * by multiple buffers. 368 */ 369 kstat_named_t arcstat_mutex_miss; 370 /* 371 * Number of buffers skipped because they have I/O in progress, are 372 * indrect prefetch buffers that have not lived long enough, or are 373 * not from the spa we're trying to evict from. 374 */ 375 kstat_named_t arcstat_evict_skip; 376 kstat_named_t arcstat_evict_l2_cached; 377 kstat_named_t arcstat_evict_l2_eligible; 378 kstat_named_t arcstat_evict_l2_ineligible; 379 kstat_named_t arcstat_hash_elements; 380 kstat_named_t arcstat_hash_elements_max; 381 kstat_named_t arcstat_hash_collisions; 382 kstat_named_t arcstat_hash_chains; 383 kstat_named_t arcstat_hash_chain_max; 384 kstat_named_t arcstat_p; 385 kstat_named_t arcstat_c; 386 kstat_named_t arcstat_c_min; 387 kstat_named_t arcstat_c_max; 388 kstat_named_t arcstat_size; 389 kstat_named_t arcstat_hdr_size; 390 kstat_named_t arcstat_data_size; 391 kstat_named_t arcstat_other_size; 392 kstat_named_t arcstat_l2_hits; 393 kstat_named_t arcstat_l2_misses; 394 kstat_named_t arcstat_l2_feeds; 395 kstat_named_t arcstat_l2_rw_clash; 396 kstat_named_t arcstat_l2_read_bytes; 397 kstat_named_t arcstat_l2_write_bytes; 398 kstat_named_t arcstat_l2_writes_sent; 399 kstat_named_t arcstat_l2_writes_done; 400 kstat_named_t arcstat_l2_writes_error; 401 kstat_named_t arcstat_l2_writes_hdr_miss; 402 kstat_named_t arcstat_l2_evict_lock_retry; 403 kstat_named_t arcstat_l2_evict_reading; 404 kstat_named_t arcstat_l2_free_on_write; 405 kstat_named_t arcstat_l2_cdata_free_on_write; 406 kstat_named_t arcstat_l2_abort_lowmem; 407 kstat_named_t arcstat_l2_cksum_bad; 408 kstat_named_t arcstat_l2_io_error; 409 kstat_named_t arcstat_l2_size; 410 kstat_named_t arcstat_l2_asize; 411 kstat_named_t arcstat_l2_hdr_size; 412 kstat_named_t arcstat_l2_compress_successes; 413 kstat_named_t arcstat_l2_compress_zeros; 414 kstat_named_t arcstat_l2_compress_failures; 415 kstat_named_t arcstat_l2_write_trylock_fail; 416 kstat_named_t arcstat_l2_write_passed_headroom; 417 kstat_named_t arcstat_l2_write_spa_mismatch; 418 kstat_named_t arcstat_l2_write_in_l2; 419 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 420 kstat_named_t arcstat_l2_write_not_cacheable; 421 kstat_named_t arcstat_l2_write_full; 422 kstat_named_t arcstat_l2_write_buffer_iter; 423 kstat_named_t arcstat_l2_write_pios; 424 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 425 kstat_named_t arcstat_l2_write_buffer_list_iter; 426 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 427 kstat_named_t arcstat_memory_throttle_count; 428 kstat_named_t arcstat_duplicate_buffers; 429 kstat_named_t arcstat_duplicate_buffers_size; 430 kstat_named_t arcstat_duplicate_reads; 431 kstat_named_t arcstat_meta_used; 432 kstat_named_t arcstat_meta_limit; 433 kstat_named_t arcstat_meta_max; 434 kstat_named_t arcstat_meta_min; 435} arc_stats_t; 436 437static arc_stats_t arc_stats = { 438 { "hits", KSTAT_DATA_UINT64 }, 439 { "misses", KSTAT_DATA_UINT64 }, 440 { "demand_data_hits", KSTAT_DATA_UINT64 }, 441 { "demand_data_misses", KSTAT_DATA_UINT64 }, 442 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 443 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 444 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 445 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 446 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 447 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 448 { "mru_hits", KSTAT_DATA_UINT64 }, 449 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 450 { "mfu_hits", KSTAT_DATA_UINT64 }, 451 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 452 { "allocated", KSTAT_DATA_UINT64 }, 453 { "deleted", KSTAT_DATA_UINT64 }, 454 { "stolen", KSTAT_DATA_UINT64 }, 455 { "recycle_miss", KSTAT_DATA_UINT64 }, 456 { "mutex_miss", KSTAT_DATA_UINT64 }, 457 { "evict_skip", KSTAT_DATA_UINT64 }, 458 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 459 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 460 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 461 { "hash_elements", KSTAT_DATA_UINT64 }, 462 { "hash_elements_max", KSTAT_DATA_UINT64 }, 463 { "hash_collisions", KSTAT_DATA_UINT64 }, 464 { "hash_chains", KSTAT_DATA_UINT64 }, 465 { "hash_chain_max", KSTAT_DATA_UINT64 }, 466 { "p", KSTAT_DATA_UINT64 }, 467 { "c", KSTAT_DATA_UINT64 }, 468 { "c_min", KSTAT_DATA_UINT64 }, 469 { "c_max", KSTAT_DATA_UINT64 }, 470 { "size", KSTAT_DATA_UINT64 }, 471 { "hdr_size", KSTAT_DATA_UINT64 }, 472 { "data_size", KSTAT_DATA_UINT64 }, 473 { "other_size", KSTAT_DATA_UINT64 }, 474 { "l2_hits", KSTAT_DATA_UINT64 }, 475 { "l2_misses", KSTAT_DATA_UINT64 }, 476 { "l2_feeds", KSTAT_DATA_UINT64 }, 477 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 478 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 479 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 480 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 481 { "l2_writes_done", KSTAT_DATA_UINT64 }, 482 { "l2_writes_error", KSTAT_DATA_UINT64 }, 483 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 484 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 485 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 486 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 487 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, 488 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 489 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 490 { "l2_io_error", KSTAT_DATA_UINT64 }, 491 { "l2_size", KSTAT_DATA_UINT64 }, 492 { "l2_asize", KSTAT_DATA_UINT64 }, 493 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 494 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 495 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 496 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 497 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 498 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 499 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 500 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 501 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 502 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 503 { "l2_write_full", KSTAT_DATA_UINT64 }, 504 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 505 { "l2_write_pios", KSTAT_DATA_UINT64 }, 506 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 507 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 508 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 509 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 510 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 511 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 512 { "duplicate_reads", KSTAT_DATA_UINT64 }, 513 { "arc_meta_used", KSTAT_DATA_UINT64 }, 514 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 515 { "arc_meta_max", KSTAT_DATA_UINT64 }, 516 { "arc_meta_min", KSTAT_DATA_UINT64 } 517}; 518 519#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 520 521#define ARCSTAT_INCR(stat, val) \ 522 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 523 524#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 525#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 526 527#define ARCSTAT_MAX(stat, val) { \ 528 uint64_t m; \ 529 while ((val) > (m = arc_stats.stat.value.ui64) && \ 530 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 531 continue; \ 532} 533 534#define ARCSTAT_MAXSTAT(stat) \ 535 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 536 537/* 538 * We define a macro to allow ARC hits/misses to be easily broken down by 539 * two separate conditions, giving a total of four different subtypes for 540 * each of hits and misses (so eight statistics total). 541 */ 542#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 543 if (cond1) { \ 544 if (cond2) { \ 545 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 546 } else { \ 547 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 548 } \ 549 } else { \ 550 if (cond2) { \ 551 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 552 } else { \ 553 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 554 } \ 555 } 556 557kstat_t *arc_ksp; 558static arc_state_t *arc_anon; 559static arc_state_t *arc_mru; 560static arc_state_t *arc_mru_ghost; 561static arc_state_t *arc_mfu; 562static arc_state_t *arc_mfu_ghost; 563static arc_state_t *arc_l2c_only; 564 565/* 566 * There are several ARC variables that are critical to export as kstats -- 567 * but we don't want to have to grovel around in the kstat whenever we wish to 568 * manipulate them. For these variables, we therefore define them to be in 569 * terms of the statistic variable. This assures that we are not introducing 570 * the possibility of inconsistency by having shadow copies of the variables, 571 * while still allowing the code to be readable. 572 */ 573#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 574#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 575#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 576#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 577#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 578#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 579#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 580#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 581#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 582 583#define L2ARC_IS_VALID_COMPRESS(_c_) \ 584 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 585 586static int arc_no_grow; /* Don't try to grow cache size */ 587static uint64_t arc_tempreserve; 588static uint64_t arc_loaned_bytes; 589 590typedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 591 592typedef struct arc_callback arc_callback_t; 593 594struct arc_callback { 595 void *acb_private; 596 arc_done_func_t *acb_done; 597 arc_buf_t *acb_buf; 598 zio_t *acb_zio_dummy; 599 arc_callback_t *acb_next; 600}; 601 602typedef struct arc_write_callback arc_write_callback_t; 603 604struct arc_write_callback { 605 void *awcb_private; 606 arc_done_func_t *awcb_ready; 607 arc_done_func_t *awcb_physdone; 608 arc_done_func_t *awcb_done; 609 arc_buf_t *awcb_buf; 610}; 611 612struct arc_buf_hdr { 613 /* protected by hash lock */ 614 dva_t b_dva; 615 uint64_t b_birth; 616 uint64_t b_cksum0; 617 618 kmutex_t b_freeze_lock; 619 zio_cksum_t *b_freeze_cksum; 620 void *b_thawed; 621 622 arc_buf_hdr_t *b_hash_next; 623 arc_buf_t *b_buf; 624 arc_flags_t b_flags; 625 uint32_t b_datacnt; 626 627 arc_callback_t *b_acb; 628 kcondvar_t b_cv; 629 630 /* immutable */ 631 arc_buf_contents_t b_type; 632 uint64_t b_size; 633 uint64_t b_spa; 634 635 /* protected by arc state mutex */ 636 arc_state_t *b_state; 637 list_node_t b_arc_node; 638 639 /* updated atomically */ 640 clock_t b_arc_access; 641 642 /* self protecting */ 643 refcount_t b_refcnt; 644 645 l2arc_buf_hdr_t *b_l2hdr; 646 list_node_t b_l2node; 647}; 648 649#ifdef _KERNEL 650static int 651sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 652{ 653 uint64_t val; 654 int err; 655 656 val = arc_meta_limit; 657 err = sysctl_handle_64(oidp, &val, 0, req); 658 if (err != 0 || req->newptr == NULL) 659 return (err); 660 661 if (val <= 0 || val > arc_c_max) 662 return (EINVAL); 663 664 arc_meta_limit = val; 665 return (0); 666} 667#endif 668 669static arc_buf_t *arc_eviction_list; 670static kmutex_t arc_eviction_mtx; 671static arc_buf_hdr_t arc_eviction_hdr; 672 673#define GHOST_STATE(state) \ 674 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 675 (state) == arc_l2c_only) 676 677#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 678#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 679#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 680#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 681#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ) 682#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE) 683#define HDR_FREE_IN_PROGRESS(hdr) \ 684 ((hdr)->b_flags & ARC_FLAG_FREE_IN_PROGRESS) 685#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 686#define HDR_L2_READING(hdr) \ 687 ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS && \ 688 (hdr)->b_l2hdr != NULL) 689#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 690#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 691#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 692 693/* 694 * Other sizes 695 */ 696 697#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 698#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 699 700/* 701 * Hash table routines 702 */ 703 704#define HT_LOCK_PAD CACHE_LINE_SIZE 705 706struct ht_lock { 707 kmutex_t ht_lock; 708#ifdef _KERNEL 709 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 710#endif 711}; 712 713#define BUF_LOCKS 256 714typedef struct buf_hash_table { 715 uint64_t ht_mask; 716 arc_buf_hdr_t **ht_table; 717 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 718} buf_hash_table_t; 719 720static buf_hash_table_t buf_hash_table; 721 722#define BUF_HASH_INDEX(spa, dva, birth) \ 723 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 724#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 725#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 726#define HDR_LOCK(hdr) \ 727 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 728 729uint64_t zfs_crc64_table[256]; 730 731/* 732 * Level 2 ARC 733 */ 734 735#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 736#define L2ARC_HEADROOM 2 /* num of writes */ 737/* 738 * If we discover during ARC scan any buffers to be compressed, we boost 739 * our headroom for the next scanning cycle by this percentage multiple. 740 */ 741#define L2ARC_HEADROOM_BOOST 200 742#define L2ARC_FEED_SECS 1 /* caching interval secs */ 743#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 744 745#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 746#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 747 748/* L2ARC Performance Tunables */ 749uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 750uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 751uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 752uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 753uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 754uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 755boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 756boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 757boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 758 759SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 760 &l2arc_write_max, 0, "max write size"); 761SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 762 &l2arc_write_boost, 0, "extra write during warmup"); 763SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 764 &l2arc_headroom, 0, "number of dev writes"); 765SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 766 &l2arc_feed_secs, 0, "interval seconds"); 767SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 768 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 769 770SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 771 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 772SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 773 &l2arc_feed_again, 0, "turbo warmup"); 774SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 775 &l2arc_norw, 0, "no reads during writes"); 776 777SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 778 &ARC_anon.arcs_size, 0, "size of anonymous state"); 779SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 780 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 781SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 782 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 783 784SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 785 &ARC_mru.arcs_size, 0, "size of mru state"); 786SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 787 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 788SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 789 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 790 791SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 792 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 793SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 794 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 795 "size of metadata in mru ghost state"); 796SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 797 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 798 "size of data in mru ghost state"); 799 800SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 801 &ARC_mfu.arcs_size, 0, "size of mfu state"); 802SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 803 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 804SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 805 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 806 807SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 808 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 809SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 810 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 811 "size of metadata in mfu ghost state"); 812SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 813 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 814 "size of data in mfu ghost state"); 815 816SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 817 &ARC_l2c_only.arcs_size, 0, "size of mru state"); 818 819/* 820 * L2ARC Internals 821 */ 822typedef struct l2arc_dev { 823 vdev_t *l2ad_vdev; /* vdev */ 824 spa_t *l2ad_spa; /* spa */ 825 uint64_t l2ad_hand; /* next write location */ 826 uint64_t l2ad_start; /* first addr on device */ 827 uint64_t l2ad_end; /* last addr on device */ 828 uint64_t l2ad_evict; /* last addr eviction reached */ 829 boolean_t l2ad_first; /* first sweep through */ 830 boolean_t l2ad_writing; /* currently writing */ 831 list_t *l2ad_buflist; /* buffer list */ 832 list_node_t l2ad_node; /* device list node */ 833} l2arc_dev_t; 834 835static list_t L2ARC_dev_list; /* device list */ 836static list_t *l2arc_dev_list; /* device list pointer */ 837static kmutex_t l2arc_dev_mtx; /* device list mutex */ 838static l2arc_dev_t *l2arc_dev_last; /* last device used */ 839static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 840static list_t L2ARC_free_on_write; /* free after write buf list */ 841static list_t *l2arc_free_on_write; /* free after write list ptr */ 842static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 843static uint64_t l2arc_ndev; /* number of devices */ 844 845typedef struct l2arc_read_callback { 846 arc_buf_t *l2rcb_buf; /* read buffer */ 847 spa_t *l2rcb_spa; /* spa */ 848 blkptr_t l2rcb_bp; /* original blkptr */ 849 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 850 int l2rcb_flags; /* original flags */ 851 enum zio_compress l2rcb_compress; /* applied compress */ 852} l2arc_read_callback_t; 853 854typedef struct l2arc_write_callback { 855 l2arc_dev_t *l2wcb_dev; /* device info */ 856 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 857} l2arc_write_callback_t; 858 859struct l2arc_buf_hdr { 860 /* protected by arc_buf_hdr mutex */ 861 l2arc_dev_t *b_dev; /* L2ARC device */ 862 uint64_t b_daddr; /* disk address, offset byte */ 863 /* compression applied to buffer data */ 864 enum zio_compress b_compress; 865 /* real alloc'd buffer size depending on b_compress applied */ 866 int b_asize; 867 /* temporary buffer holder for in-flight compressed data */ 868 void *b_tmp_cdata; 869}; 870 871typedef struct l2arc_data_free { 872 /* protected by l2arc_free_on_write_mtx */ 873 void *l2df_data; 874 size_t l2df_size; 875 void (*l2df_func)(void *, size_t); 876 list_node_t l2df_list_node; 877} l2arc_data_free_t; 878 879static kmutex_t l2arc_feed_thr_lock; 880static kcondvar_t l2arc_feed_thr_cv; 881static uint8_t l2arc_thread_exit; 882 883static void arc_get_data_buf(arc_buf_t *); 884static void arc_access(arc_buf_hdr_t *, kmutex_t *); 885static int arc_evict_needed(arc_buf_contents_t); 886static void arc_evict_ghost(arc_state_t *, uint64_t, int64_t); 887static void arc_buf_watch(arc_buf_t *); 888 889static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 890static void l2arc_read_done(zio_t *); 891static void l2arc_hdr_stat_add(void); 892static void l2arc_hdr_stat_remove(void); 893 894static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *); 895static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress); 896static void l2arc_release_cdata_buf(arc_buf_hdr_t *); 897 898static uint64_t 899buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 900{ 901 uint8_t *vdva = (uint8_t *)dva; 902 uint64_t crc = -1ULL; 903 int i; 904 905 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 906 907 for (i = 0; i < sizeof (dva_t); i++) 908 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 909 910 crc ^= (spa>>8) ^ birth; 911 912 return (crc); 913} 914 915#define BUF_EMPTY(buf) \ 916 ((buf)->b_dva.dva_word[0] == 0 && \ 917 (buf)->b_dva.dva_word[1] == 0 && \ 918 (buf)->b_cksum0 == 0) 919 920#define BUF_EQUAL(spa, dva, birth, buf) \ 921 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 922 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 923 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 924 925static void 926buf_discard_identity(arc_buf_hdr_t *hdr) 927{ 928 hdr->b_dva.dva_word[0] = 0; 929 hdr->b_dva.dva_word[1] = 0; 930 hdr->b_birth = 0; 931 hdr->b_cksum0 = 0; 932} 933 934static arc_buf_hdr_t * 935buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 936{ 937 const dva_t *dva = BP_IDENTITY(bp); 938 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 939 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 940 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 941 arc_buf_hdr_t *hdr; 942 943 mutex_enter(hash_lock); 944 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 945 hdr = hdr->b_hash_next) { 946 if (BUF_EQUAL(spa, dva, birth, hdr)) { 947 *lockp = hash_lock; 948 return (hdr); 949 } 950 } 951 mutex_exit(hash_lock); 952 *lockp = NULL; 953 return (NULL); 954} 955 956/* 957 * Insert an entry into the hash table. If there is already an element 958 * equal to elem in the hash table, then the already existing element 959 * will be returned and the new element will not be inserted. 960 * Otherwise returns NULL. 961 */ 962static arc_buf_hdr_t * 963buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 964{ 965 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 966 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 967 arc_buf_hdr_t *fhdr; 968 uint32_t i; 969 970 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 971 ASSERT(hdr->b_birth != 0); 972 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 973 *lockp = hash_lock; 974 mutex_enter(hash_lock); 975 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 976 fhdr = fhdr->b_hash_next, i++) { 977 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 978 return (fhdr); 979 } 980 981 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 982 buf_hash_table.ht_table[idx] = hdr; 983 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 984 985 /* collect some hash table performance data */ 986 if (i > 0) { 987 ARCSTAT_BUMP(arcstat_hash_collisions); 988 if (i == 1) 989 ARCSTAT_BUMP(arcstat_hash_chains); 990 991 ARCSTAT_MAX(arcstat_hash_chain_max, i); 992 } 993 994 ARCSTAT_BUMP(arcstat_hash_elements); 995 ARCSTAT_MAXSTAT(arcstat_hash_elements); 996 997 return (NULL); 998} 999 1000static void 1001buf_hash_remove(arc_buf_hdr_t *hdr) 1002{ 1003 arc_buf_hdr_t *fhdr, **hdrp; 1004 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1005 1006 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1007 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1008 1009 hdrp = &buf_hash_table.ht_table[idx]; 1010 while ((fhdr = *hdrp) != hdr) { 1011 ASSERT(fhdr != NULL); 1012 hdrp = &fhdr->b_hash_next; 1013 } 1014 *hdrp = hdr->b_hash_next; 1015 hdr->b_hash_next = NULL; 1016 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; 1017 1018 /* collect some hash table performance data */ 1019 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1020 1021 if (buf_hash_table.ht_table[idx] && 1022 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1023 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1024} 1025 1026/* 1027 * Global data structures and functions for the buf kmem cache. 1028 */ 1029static kmem_cache_t *hdr_cache; 1030static kmem_cache_t *buf_cache; 1031 1032static void 1033buf_fini(void) 1034{ 1035 int i; 1036 1037 kmem_free(buf_hash_table.ht_table, 1038 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1039 for (i = 0; i < BUF_LOCKS; i++) 1040 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1041 kmem_cache_destroy(hdr_cache); 1042 kmem_cache_destroy(buf_cache); 1043} 1044 1045/* 1046 * Constructor callback - called when the cache is empty 1047 * and a new buf is requested. 1048 */ 1049/* ARGSUSED */ 1050static int 1051hdr_cons(void *vbuf, void *unused, int kmflag) 1052{ 1053 arc_buf_hdr_t *hdr = vbuf; 1054 1055 bzero(hdr, sizeof (arc_buf_hdr_t)); 1056 refcount_create(&hdr->b_refcnt); 1057 cv_init(&hdr->b_cv, NULL, CV_DEFAULT, NULL); 1058 mutex_init(&hdr->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1059 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1060 1061 return (0); 1062} 1063 1064/* ARGSUSED */ 1065static int 1066buf_cons(void *vbuf, void *unused, int kmflag) 1067{ 1068 arc_buf_t *buf = vbuf; 1069 1070 bzero(buf, sizeof (arc_buf_t)); 1071 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1072 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1073 1074 return (0); 1075} 1076 1077/* 1078 * Destructor callback - called when a cached buf is 1079 * no longer required. 1080 */ 1081/* ARGSUSED */ 1082static void 1083hdr_dest(void *vbuf, void *unused) 1084{ 1085 arc_buf_hdr_t *hdr = vbuf; 1086 1087 ASSERT(BUF_EMPTY(hdr)); 1088 refcount_destroy(&hdr->b_refcnt); 1089 cv_destroy(&hdr->b_cv); 1090 mutex_destroy(&hdr->b_freeze_lock); 1091 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1092} 1093 1094/* ARGSUSED */ 1095static void 1096buf_dest(void *vbuf, void *unused) 1097{ 1098 arc_buf_t *buf = vbuf; 1099 1100 mutex_destroy(&buf->b_evict_lock); 1101 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1102} 1103 1104/* 1105 * Reclaim callback -- invoked when memory is low. 1106 */ 1107/* ARGSUSED */ 1108static void 1109hdr_recl(void *unused) 1110{ 1111 dprintf("hdr_recl called\n"); 1112 /* 1113 * umem calls the reclaim func when we destroy the buf cache, 1114 * which is after we do arc_fini(). 1115 */ 1116 if (!arc_dead) 1117 cv_signal(&arc_reclaim_thr_cv); 1118} 1119 1120static void 1121buf_init(void) 1122{ 1123 uint64_t *ct; 1124 uint64_t hsize = 1ULL << 12; 1125 int i, j; 1126 1127 /* 1128 * The hash table is big enough to fill all of physical memory 1129 * with an average block size of zfs_arc_average_blocksize (default 8K). 1130 * By default, the table will take up 1131 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1132 */ 1133 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1134 hsize <<= 1; 1135retry: 1136 buf_hash_table.ht_mask = hsize - 1; 1137 buf_hash_table.ht_table = 1138 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1139 if (buf_hash_table.ht_table == NULL) { 1140 ASSERT(hsize > (1ULL << 8)); 1141 hsize >>= 1; 1142 goto retry; 1143 } 1144 1145 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1146 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1147 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1148 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1149 1150 for (i = 0; i < 256; i++) 1151 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1152 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1153 1154 for (i = 0; i < BUF_LOCKS; i++) { 1155 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1156 NULL, MUTEX_DEFAULT, NULL); 1157 } 1158} 1159 1160#define ARC_MINTIME (hz>>4) /* 62 ms */ 1161 1162static void 1163arc_cksum_verify(arc_buf_t *buf) 1164{ 1165 zio_cksum_t zc; 1166 1167 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1168 return; 1169 1170 mutex_enter(&buf->b_hdr->b_freeze_lock); 1171 if (buf->b_hdr->b_freeze_cksum == NULL || 1172 (buf->b_hdr->b_flags & ARC_FLAG_IO_ERROR)) { 1173 mutex_exit(&buf->b_hdr->b_freeze_lock); 1174 return; 1175 } 1176 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1177 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1178 panic("buffer modified while frozen!"); 1179 mutex_exit(&buf->b_hdr->b_freeze_lock); 1180} 1181 1182static int 1183arc_cksum_equal(arc_buf_t *buf) 1184{ 1185 zio_cksum_t zc; 1186 int equal; 1187 1188 mutex_enter(&buf->b_hdr->b_freeze_lock); 1189 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1190 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1191 mutex_exit(&buf->b_hdr->b_freeze_lock); 1192 1193 return (equal); 1194} 1195 1196static void 1197arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1198{ 1199 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1200 return; 1201 1202 mutex_enter(&buf->b_hdr->b_freeze_lock); 1203 if (buf->b_hdr->b_freeze_cksum != NULL) { 1204 mutex_exit(&buf->b_hdr->b_freeze_lock); 1205 return; 1206 } 1207 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1208 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1209 buf->b_hdr->b_freeze_cksum); 1210 mutex_exit(&buf->b_hdr->b_freeze_lock); 1211#ifdef illumos 1212 arc_buf_watch(buf); 1213#endif /* illumos */ 1214} 1215 1216#ifdef illumos 1217#ifndef _KERNEL 1218typedef struct procctl { 1219 long cmd; 1220 prwatch_t prwatch; 1221} procctl_t; 1222#endif 1223 1224/* ARGSUSED */ 1225static void 1226arc_buf_unwatch(arc_buf_t *buf) 1227{ 1228#ifndef _KERNEL 1229 if (arc_watch) { 1230 int result; 1231 procctl_t ctl; 1232 ctl.cmd = PCWATCH; 1233 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1234 ctl.prwatch.pr_size = 0; 1235 ctl.prwatch.pr_wflags = 0; 1236 result = write(arc_procfd, &ctl, sizeof (ctl)); 1237 ASSERT3U(result, ==, sizeof (ctl)); 1238 } 1239#endif 1240} 1241 1242/* ARGSUSED */ 1243static void 1244arc_buf_watch(arc_buf_t *buf) 1245{ 1246#ifndef _KERNEL 1247 if (arc_watch) { 1248 int result; 1249 procctl_t ctl; 1250 ctl.cmd = PCWATCH; 1251 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1252 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1253 ctl.prwatch.pr_wflags = WA_WRITE; 1254 result = write(arc_procfd, &ctl, sizeof (ctl)); 1255 ASSERT3U(result, ==, sizeof (ctl)); 1256 } 1257#endif 1258} 1259#endif /* illumos */ 1260 1261void 1262arc_buf_thaw(arc_buf_t *buf) 1263{ 1264 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1265 if (buf->b_hdr->b_state != arc_anon) 1266 panic("modifying non-anon buffer!"); 1267 if (buf->b_hdr->b_flags & ARC_FLAG_IO_IN_PROGRESS) 1268 panic("modifying buffer while i/o in progress!"); 1269 arc_cksum_verify(buf); 1270 } 1271 1272 mutex_enter(&buf->b_hdr->b_freeze_lock); 1273 if (buf->b_hdr->b_freeze_cksum != NULL) { 1274 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1275 buf->b_hdr->b_freeze_cksum = NULL; 1276 } 1277 1278 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1279 if (buf->b_hdr->b_thawed) 1280 kmem_free(buf->b_hdr->b_thawed, 1); 1281 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1282 } 1283 1284 mutex_exit(&buf->b_hdr->b_freeze_lock); 1285 1286#ifdef illumos 1287 arc_buf_unwatch(buf); 1288#endif /* illumos */ 1289} 1290 1291void 1292arc_buf_freeze(arc_buf_t *buf) 1293{ 1294 kmutex_t *hash_lock; 1295 1296 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1297 return; 1298 1299 hash_lock = HDR_LOCK(buf->b_hdr); 1300 mutex_enter(hash_lock); 1301 1302 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1303 buf->b_hdr->b_state == arc_anon); 1304 arc_cksum_compute(buf, B_FALSE); 1305 mutex_exit(hash_lock); 1306 1307} 1308 1309static void 1310get_buf_info(arc_buf_hdr_t *hdr, arc_state_t *state, list_t **list, kmutex_t **lock) 1311{ 1312 uint64_t buf_hashid = buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1313 1314 if (hdr->b_type == ARC_BUFC_METADATA) 1315 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1316 else { 1317 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1318 buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1319 } 1320 1321 *list = &state->arcs_lists[buf_hashid]; 1322 *lock = ARCS_LOCK(state, buf_hashid); 1323} 1324 1325 1326static void 1327add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1328{ 1329 ASSERT(MUTEX_HELD(hash_lock)); 1330 1331 if ((refcount_add(&hdr->b_refcnt, tag) == 1) && 1332 (hdr->b_state != arc_anon)) { 1333 uint64_t delta = hdr->b_size * hdr->b_datacnt; 1334 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 1335 list_t *list; 1336 kmutex_t *lock; 1337 1338 get_buf_info(hdr, hdr->b_state, &list, &lock); 1339 ASSERT(!MUTEX_HELD(lock)); 1340 mutex_enter(lock); 1341 ASSERT(list_link_active(&hdr->b_arc_node)); 1342 list_remove(list, hdr); 1343 if (GHOST_STATE(hdr->b_state)) { 1344 ASSERT0(hdr->b_datacnt); 1345 ASSERT3P(hdr->b_buf, ==, NULL); 1346 delta = hdr->b_size; 1347 } 1348 ASSERT(delta > 0); 1349 ASSERT3U(*size, >=, delta); 1350 atomic_add_64(size, -delta); 1351 mutex_exit(lock); 1352 /* remove the prefetch flag if we get a reference */ 1353 if (hdr->b_flags & ARC_FLAG_PREFETCH) 1354 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 1355 } 1356} 1357 1358static int 1359remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1360{ 1361 int cnt; 1362 arc_state_t *state = hdr->b_state; 1363 1364 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1365 ASSERT(!GHOST_STATE(state)); 1366 1367 if (((cnt = refcount_remove(&hdr->b_refcnt, tag)) == 0) && 1368 (state != arc_anon)) { 1369 uint64_t *size = &state->arcs_lsize[hdr->b_type]; 1370 list_t *list; 1371 kmutex_t *lock; 1372 1373 get_buf_info(hdr, state, &list, &lock); 1374 ASSERT(!MUTEX_HELD(lock)); 1375 mutex_enter(lock); 1376 ASSERT(!list_link_active(&hdr->b_arc_node)); 1377 list_insert_head(list, hdr); 1378 ASSERT(hdr->b_datacnt > 0); 1379 atomic_add_64(size, hdr->b_size * hdr->b_datacnt); 1380 mutex_exit(lock); 1381 } 1382 return (cnt); 1383} 1384 1385/* 1386 * Move the supplied buffer to the indicated state. The mutex 1387 * for the buffer must be held by the caller. 1388 */ 1389static void 1390arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 1391 kmutex_t *hash_lock) 1392{ 1393 arc_state_t *old_state = hdr->b_state; 1394 int64_t refcnt = refcount_count(&hdr->b_refcnt); 1395 uint64_t from_delta, to_delta; 1396 list_t *list; 1397 kmutex_t *lock; 1398 1399 ASSERT(MUTEX_HELD(hash_lock)); 1400 ASSERT3P(new_state, !=, old_state); 1401 ASSERT(refcnt == 0 || hdr->b_datacnt > 0); 1402 ASSERT(hdr->b_datacnt == 0 || !GHOST_STATE(new_state)); 1403 ASSERT(hdr->b_datacnt <= 1 || old_state != arc_anon); 1404 1405 from_delta = to_delta = hdr->b_datacnt * hdr->b_size; 1406 1407 /* 1408 * If this buffer is evictable, transfer it from the 1409 * old state list to the new state list. 1410 */ 1411 if (refcnt == 0) { 1412 if (old_state != arc_anon) { 1413 int use_mutex; 1414 uint64_t *size = &old_state->arcs_lsize[hdr->b_type]; 1415 1416 get_buf_info(hdr, old_state, &list, &lock); 1417 use_mutex = !MUTEX_HELD(lock); 1418 if (use_mutex) 1419 mutex_enter(lock); 1420 1421 ASSERT(list_link_active(&hdr->b_arc_node)); 1422 list_remove(list, hdr); 1423 1424 /* 1425 * If prefetching out of the ghost cache, 1426 * we will have a non-zero datacnt. 1427 */ 1428 if (GHOST_STATE(old_state) && hdr->b_datacnt == 0) { 1429 /* ghost elements have a ghost size */ 1430 ASSERT(hdr->b_buf == NULL); 1431 from_delta = hdr->b_size; 1432 } 1433 ASSERT3U(*size, >=, from_delta); 1434 atomic_add_64(size, -from_delta); 1435 1436 if (use_mutex) 1437 mutex_exit(lock); 1438 } 1439 if (new_state != arc_anon) { 1440 int use_mutex; 1441 uint64_t *size = &new_state->arcs_lsize[hdr->b_type]; 1442 1443 get_buf_info(hdr, new_state, &list, &lock); 1444 use_mutex = !MUTEX_HELD(lock); 1445 if (use_mutex) 1446 mutex_enter(lock); 1447 1448 list_insert_head(list, hdr); 1449 1450 /* ghost elements have a ghost size */ 1451 if (GHOST_STATE(new_state)) { 1452 ASSERT(hdr->b_datacnt == 0); 1453 ASSERT(hdr->b_buf == NULL); 1454 to_delta = hdr->b_size; 1455 } 1456 atomic_add_64(size, to_delta); 1457 1458 if (use_mutex) 1459 mutex_exit(lock); 1460 } 1461 } 1462 1463 ASSERT(!BUF_EMPTY(hdr)); 1464 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 1465 buf_hash_remove(hdr); 1466 1467 /* adjust state sizes */ 1468 if (to_delta) 1469 atomic_add_64(&new_state->arcs_size, to_delta); 1470 if (from_delta) { 1471 ASSERT3U(old_state->arcs_size, >=, from_delta); 1472 atomic_add_64(&old_state->arcs_size, -from_delta); 1473 } 1474 hdr->b_state = new_state; 1475 1476 /* adjust l2arc hdr stats */ 1477 if (new_state == arc_l2c_only) 1478 l2arc_hdr_stat_add(); 1479 else if (old_state == arc_l2c_only) 1480 l2arc_hdr_stat_remove(); 1481} 1482 1483void 1484arc_space_consume(uint64_t space, arc_space_type_t type) 1485{ 1486 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1487 1488 switch (type) { 1489 case ARC_SPACE_DATA: 1490 ARCSTAT_INCR(arcstat_data_size, space); 1491 break; 1492 case ARC_SPACE_OTHER: 1493 ARCSTAT_INCR(arcstat_other_size, space); 1494 break; 1495 case ARC_SPACE_HDRS: 1496 ARCSTAT_INCR(arcstat_hdr_size, space); 1497 break; 1498 case ARC_SPACE_L2HDRS: 1499 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1500 break; 1501 } 1502 1503 ARCSTAT_INCR(arcstat_meta_used, space); 1504 atomic_add_64(&arc_size, space); 1505} 1506 1507void 1508arc_space_return(uint64_t space, arc_space_type_t type) 1509{ 1510 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1511 1512 switch (type) { 1513 case ARC_SPACE_DATA: 1514 ARCSTAT_INCR(arcstat_data_size, -space); 1515 break; 1516 case ARC_SPACE_OTHER: 1517 ARCSTAT_INCR(arcstat_other_size, -space); 1518 break; 1519 case ARC_SPACE_HDRS: 1520 ARCSTAT_INCR(arcstat_hdr_size, -space); 1521 break; 1522 case ARC_SPACE_L2HDRS: 1523 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1524 break; 1525 } 1526 1527 ASSERT(arc_meta_used >= space); 1528 if (arc_meta_max < arc_meta_used) 1529 arc_meta_max = arc_meta_used; 1530 ARCSTAT_INCR(arcstat_meta_used, -space); 1531 ASSERT(arc_size >= space); 1532 atomic_add_64(&arc_size, -space); 1533} 1534 1535arc_buf_t * 1536arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1537{ 1538 arc_buf_hdr_t *hdr; 1539 arc_buf_t *buf; 1540 1541 ASSERT3U(size, >, 0); 1542 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1543 ASSERT(BUF_EMPTY(hdr)); 1544 hdr->b_size = size; 1545 hdr->b_type = type; 1546 hdr->b_spa = spa_load_guid(spa); 1547 hdr->b_state = arc_anon; 1548 hdr->b_arc_access = 0; 1549 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1550 buf->b_hdr = hdr; 1551 buf->b_data = NULL; 1552 buf->b_efunc = NULL; 1553 buf->b_private = NULL; 1554 buf->b_next = NULL; 1555 hdr->b_buf = buf; 1556 arc_get_data_buf(buf); 1557 hdr->b_datacnt = 1; 1558 hdr->b_flags = 0; 1559 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1560 (void) refcount_add(&hdr->b_refcnt, tag); 1561 1562 return (buf); 1563} 1564 1565static char *arc_onloan_tag = "onloan"; 1566 1567/* 1568 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1569 * flight data by arc_tempreserve_space() until they are "returned". Loaned 1570 * buffers must be returned to the arc before they can be used by the DMU or 1571 * freed. 1572 */ 1573arc_buf_t * 1574arc_loan_buf(spa_t *spa, int size) 1575{ 1576 arc_buf_t *buf; 1577 1578 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1579 1580 atomic_add_64(&arc_loaned_bytes, size); 1581 return (buf); 1582} 1583 1584/* 1585 * Return a loaned arc buffer to the arc. 1586 */ 1587void 1588arc_return_buf(arc_buf_t *buf, void *tag) 1589{ 1590 arc_buf_hdr_t *hdr = buf->b_hdr; 1591 1592 ASSERT(buf->b_data != NULL); 1593 (void) refcount_add(&hdr->b_refcnt, tag); 1594 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1595 1596 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1597} 1598 1599/* Detach an arc_buf from a dbuf (tag) */ 1600void 1601arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1602{ 1603 arc_buf_hdr_t *hdr; 1604 1605 ASSERT(buf->b_data != NULL); 1606 hdr = buf->b_hdr; 1607 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1608 (void) refcount_remove(&hdr->b_refcnt, tag); 1609 buf->b_efunc = NULL; 1610 buf->b_private = NULL; 1611 1612 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1613} 1614 1615static arc_buf_t * 1616arc_buf_clone(arc_buf_t *from) 1617{ 1618 arc_buf_t *buf; 1619 arc_buf_hdr_t *hdr = from->b_hdr; 1620 uint64_t size = hdr->b_size; 1621 1622 ASSERT(hdr->b_state != arc_anon); 1623 1624 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1625 buf->b_hdr = hdr; 1626 buf->b_data = NULL; 1627 buf->b_efunc = NULL; 1628 buf->b_private = NULL; 1629 buf->b_next = hdr->b_buf; 1630 hdr->b_buf = buf; 1631 arc_get_data_buf(buf); 1632 bcopy(from->b_data, buf->b_data, size); 1633 1634 /* 1635 * This buffer already exists in the arc so create a duplicate 1636 * copy for the caller. If the buffer is associated with user data 1637 * then track the size and number of duplicates. These stats will be 1638 * updated as duplicate buffers are created and destroyed. 1639 */ 1640 if (hdr->b_type == ARC_BUFC_DATA) { 1641 ARCSTAT_BUMP(arcstat_duplicate_buffers); 1642 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1643 } 1644 hdr->b_datacnt += 1; 1645 return (buf); 1646} 1647 1648void 1649arc_buf_add_ref(arc_buf_t *buf, void* tag) 1650{ 1651 arc_buf_hdr_t *hdr; 1652 kmutex_t *hash_lock; 1653 1654 /* 1655 * Check to see if this buffer is evicted. Callers 1656 * must verify b_data != NULL to know if the add_ref 1657 * was successful. 1658 */ 1659 mutex_enter(&buf->b_evict_lock); 1660 if (buf->b_data == NULL) { 1661 mutex_exit(&buf->b_evict_lock); 1662 return; 1663 } 1664 hash_lock = HDR_LOCK(buf->b_hdr); 1665 mutex_enter(hash_lock); 1666 hdr = buf->b_hdr; 1667 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1668 mutex_exit(&buf->b_evict_lock); 1669 1670 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1671 add_reference(hdr, hash_lock, tag); 1672 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1673 arc_access(hdr, hash_lock); 1674 mutex_exit(hash_lock); 1675 ARCSTAT_BUMP(arcstat_hits); 1676 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 1677 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1678 data, metadata, hits); 1679} 1680 1681static void 1682arc_buf_free_on_write(void *data, size_t size, 1683 void (*free_func)(void *, size_t)) 1684{ 1685 l2arc_data_free_t *df; 1686 1687 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1688 df->l2df_data = data; 1689 df->l2df_size = size; 1690 df->l2df_func = free_func; 1691 mutex_enter(&l2arc_free_on_write_mtx); 1692 list_insert_head(l2arc_free_on_write, df); 1693 mutex_exit(&l2arc_free_on_write_mtx); 1694} 1695 1696/* 1697 * Free the arc data buffer. If it is an l2arc write in progress, 1698 * the buffer is placed on l2arc_free_on_write to be freed later. 1699 */ 1700static void 1701arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1702{ 1703 arc_buf_hdr_t *hdr = buf->b_hdr; 1704 1705 if (HDR_L2_WRITING(hdr)) { 1706 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 1707 ARCSTAT_BUMP(arcstat_l2_free_on_write); 1708 } else { 1709 free_func(buf->b_data, hdr->b_size); 1710 } 1711} 1712 1713/* 1714 * Free up buf->b_data and if 'remove' is set, then pull the 1715 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 1716 */ 1717static void 1718arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 1719{ 1720 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1721 1722 ASSERT(MUTEX_HELD(&l2arc_buflist_mtx)); 1723 1724 if (l2hdr->b_tmp_cdata == NULL) 1725 return; 1726 1727 ASSERT(HDR_L2_WRITING(hdr)); 1728 arc_buf_free_on_write(l2hdr->b_tmp_cdata, hdr->b_size, 1729 zio_data_buf_free); 1730 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 1731 l2hdr->b_tmp_cdata = NULL; 1732} 1733 1734static void 1735arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 1736{ 1737 arc_buf_t **bufp; 1738 1739 /* free up data associated with the buf */ 1740 if (buf->b_data) { 1741 arc_state_t *state = buf->b_hdr->b_state; 1742 uint64_t size = buf->b_hdr->b_size; 1743 arc_buf_contents_t type = buf->b_hdr->b_type; 1744 1745 arc_cksum_verify(buf); 1746#ifdef illumos 1747 arc_buf_unwatch(buf); 1748#endif /* illumos */ 1749 1750 if (!recycle) { 1751 if (type == ARC_BUFC_METADATA) { 1752 arc_buf_data_free(buf, zio_buf_free); 1753 arc_space_return(size, ARC_SPACE_DATA); 1754 } else { 1755 ASSERT(type == ARC_BUFC_DATA); 1756 arc_buf_data_free(buf, zio_data_buf_free); 1757 ARCSTAT_INCR(arcstat_data_size, -size); 1758 atomic_add_64(&arc_size, -size); 1759 } 1760 } 1761 if (list_link_active(&buf->b_hdr->b_arc_node)) { 1762 uint64_t *cnt = &state->arcs_lsize[type]; 1763 1764 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1765 ASSERT(state != arc_anon); 1766 1767 ASSERT3U(*cnt, >=, size); 1768 atomic_add_64(cnt, -size); 1769 } 1770 ASSERT3U(state->arcs_size, >=, size); 1771 atomic_add_64(&state->arcs_size, -size); 1772 buf->b_data = NULL; 1773 1774 /* 1775 * If we're destroying a duplicate buffer make sure 1776 * that the appropriate statistics are updated. 1777 */ 1778 if (buf->b_hdr->b_datacnt > 1 && 1779 buf->b_hdr->b_type == ARC_BUFC_DATA) { 1780 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1781 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1782 } 1783 ASSERT(buf->b_hdr->b_datacnt > 0); 1784 buf->b_hdr->b_datacnt -= 1; 1785 } 1786 1787 /* only remove the buf if requested */ 1788 if (!remove) 1789 return; 1790 1791 /* remove the buf from the hdr list */ 1792 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1793 continue; 1794 *bufp = buf->b_next; 1795 buf->b_next = NULL; 1796 1797 ASSERT(buf->b_efunc == NULL); 1798 1799 /* clean up the buf */ 1800 buf->b_hdr = NULL; 1801 kmem_cache_free(buf_cache, buf); 1802} 1803 1804static void 1805arc_hdr_destroy(arc_buf_hdr_t *hdr) 1806{ 1807 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1808 ASSERT3P(hdr->b_state, ==, arc_anon); 1809 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1810 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1811 1812 if (l2hdr != NULL) { 1813 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1814 /* 1815 * To prevent arc_free() and l2arc_evict() from 1816 * attempting to free the same buffer at the same time, 1817 * a FREE_IN_PROGRESS flag is given to arc_free() to 1818 * give it priority. l2arc_evict() can't destroy this 1819 * header while we are waiting on l2arc_buflist_mtx. 1820 * 1821 * The hdr may be removed from l2ad_buflist before we 1822 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1823 */ 1824 if (!buflist_held) { 1825 mutex_enter(&l2arc_buflist_mtx); 1826 l2hdr = hdr->b_l2hdr; 1827 } 1828 1829 if (l2hdr != NULL) { 1830 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1831 l2hdr->b_asize, 0); 1832 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1833 arc_buf_l2_cdata_free(hdr); 1834 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1835 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1836 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 1837 -l2hdr->b_asize, 0, 0); 1838 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1839 if (hdr->b_state == arc_l2c_only) 1840 l2arc_hdr_stat_remove(); 1841 hdr->b_l2hdr = NULL; 1842 } 1843 1844 if (!buflist_held) 1845 mutex_exit(&l2arc_buflist_mtx); 1846 } 1847 1848 if (!BUF_EMPTY(hdr)) { 1849 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1850 buf_discard_identity(hdr); 1851 } 1852 while (hdr->b_buf) { 1853 arc_buf_t *buf = hdr->b_buf; 1854 1855 if (buf->b_efunc) { 1856 mutex_enter(&arc_eviction_mtx); 1857 mutex_enter(&buf->b_evict_lock); 1858 ASSERT(buf->b_hdr != NULL); 1859 arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1860 hdr->b_buf = buf->b_next; 1861 buf->b_hdr = &arc_eviction_hdr; 1862 buf->b_next = arc_eviction_list; 1863 arc_eviction_list = buf; 1864 mutex_exit(&buf->b_evict_lock); 1865 mutex_exit(&arc_eviction_mtx); 1866 } else { 1867 arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1868 } 1869 } 1870 if (hdr->b_freeze_cksum != NULL) { 1871 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1872 hdr->b_freeze_cksum = NULL; 1873 } 1874 if (hdr->b_thawed) { 1875 kmem_free(hdr->b_thawed, 1); 1876 hdr->b_thawed = NULL; 1877 } 1878 1879 ASSERT(!list_link_active(&hdr->b_arc_node)); 1880 ASSERT3P(hdr->b_hash_next, ==, NULL); 1881 ASSERT3P(hdr->b_acb, ==, NULL); 1882 kmem_cache_free(hdr_cache, hdr); 1883} 1884 1885void 1886arc_buf_free(arc_buf_t *buf, void *tag) 1887{ 1888 arc_buf_hdr_t *hdr = buf->b_hdr; 1889 int hashed = hdr->b_state != arc_anon; 1890 1891 ASSERT(buf->b_efunc == NULL); 1892 ASSERT(buf->b_data != NULL); 1893 1894 if (hashed) { 1895 kmutex_t *hash_lock = HDR_LOCK(hdr); 1896 1897 mutex_enter(hash_lock); 1898 hdr = buf->b_hdr; 1899 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1900 1901 (void) remove_reference(hdr, hash_lock, tag); 1902 if (hdr->b_datacnt > 1) { 1903 arc_buf_destroy(buf, FALSE, TRUE); 1904 } else { 1905 ASSERT(buf == hdr->b_buf); 1906 ASSERT(buf->b_efunc == NULL); 1907 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 1908 } 1909 mutex_exit(hash_lock); 1910 } else if (HDR_IO_IN_PROGRESS(hdr)) { 1911 int destroy_hdr; 1912 /* 1913 * We are in the middle of an async write. Don't destroy 1914 * this buffer unless the write completes before we finish 1915 * decrementing the reference count. 1916 */ 1917 mutex_enter(&arc_eviction_mtx); 1918 (void) remove_reference(hdr, NULL, tag); 1919 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1920 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1921 mutex_exit(&arc_eviction_mtx); 1922 if (destroy_hdr) 1923 arc_hdr_destroy(hdr); 1924 } else { 1925 if (remove_reference(hdr, NULL, tag) > 0) 1926 arc_buf_destroy(buf, FALSE, TRUE); 1927 else 1928 arc_hdr_destroy(hdr); 1929 } 1930} 1931 1932boolean_t 1933arc_buf_remove_ref(arc_buf_t *buf, void* tag) 1934{ 1935 arc_buf_hdr_t *hdr = buf->b_hdr; 1936 kmutex_t *hash_lock = HDR_LOCK(hdr); 1937 boolean_t no_callback = (buf->b_efunc == NULL); 1938 1939 if (hdr->b_state == arc_anon) { 1940 ASSERT(hdr->b_datacnt == 1); 1941 arc_buf_free(buf, tag); 1942 return (no_callback); 1943 } 1944 1945 mutex_enter(hash_lock); 1946 hdr = buf->b_hdr; 1947 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1948 ASSERT(hdr->b_state != arc_anon); 1949 ASSERT(buf->b_data != NULL); 1950 1951 (void) remove_reference(hdr, hash_lock, tag); 1952 if (hdr->b_datacnt > 1) { 1953 if (no_callback) 1954 arc_buf_destroy(buf, FALSE, TRUE); 1955 } else if (no_callback) { 1956 ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1957 ASSERT(buf->b_efunc == NULL); 1958 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 1959 } 1960 ASSERT(no_callback || hdr->b_datacnt > 1 || 1961 refcount_is_zero(&hdr->b_refcnt)); 1962 mutex_exit(hash_lock); 1963 return (no_callback); 1964} 1965 1966int 1967arc_buf_size(arc_buf_t *buf) 1968{ 1969 return (buf->b_hdr->b_size); 1970} 1971 1972/* 1973 * Called from the DMU to determine if the current buffer should be 1974 * evicted. In order to ensure proper locking, the eviction must be initiated 1975 * from the DMU. Return true if the buffer is associated with user data and 1976 * duplicate buffers still exist. 1977 */ 1978boolean_t 1979arc_buf_eviction_needed(arc_buf_t *buf) 1980{ 1981 arc_buf_hdr_t *hdr; 1982 boolean_t evict_needed = B_FALSE; 1983 1984 if (zfs_disable_dup_eviction) 1985 return (B_FALSE); 1986 1987 mutex_enter(&buf->b_evict_lock); 1988 hdr = buf->b_hdr; 1989 if (hdr == NULL) { 1990 /* 1991 * We are in arc_do_user_evicts(); let that function 1992 * perform the eviction. 1993 */ 1994 ASSERT(buf->b_data == NULL); 1995 mutex_exit(&buf->b_evict_lock); 1996 return (B_FALSE); 1997 } else if (buf->b_data == NULL) { 1998 /* 1999 * We have already been added to the arc eviction list; 2000 * recommend eviction. 2001 */ 2002 ASSERT3P(hdr, ==, &arc_eviction_hdr); 2003 mutex_exit(&buf->b_evict_lock); 2004 return (B_TRUE); 2005 } 2006 2007 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 2008 evict_needed = B_TRUE; 2009 2010 mutex_exit(&buf->b_evict_lock); 2011 return (evict_needed); 2012} 2013 2014/* 2015 * Evict buffers from list until we've removed the specified number of 2016 * bytes. Move the removed buffers to the appropriate evict state. 2017 * If the recycle flag is set, then attempt to "recycle" a buffer: 2018 * - look for a buffer to evict that is `bytes' long. 2019 * - return the data block from this buffer rather than freeing it. 2020 * This flag is used by callers that are trying to make space for a 2021 * new buffer in a full arc cache. 2022 * 2023 * This function makes a "best effort". It skips over any buffers 2024 * it can't get a hash_lock on, and so may not catch all candidates. 2025 * It may also return without evicting as much space as requested. 2026 */ 2027static void * 2028arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 2029 arc_buf_contents_t type) 2030{ 2031 arc_state_t *evicted_state; 2032 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 2033 int64_t bytes_remaining; 2034 arc_buf_hdr_t *hdr, *hdr_prev = NULL; 2035 list_t *evicted_list, *list, *evicted_list_start, *list_start; 2036 kmutex_t *lock, *evicted_lock; 2037 kmutex_t *hash_lock; 2038 boolean_t have_lock; 2039 void *stolen = NULL; 2040 arc_buf_hdr_t marker = { 0 }; 2041 int count = 0; 2042 static int evict_metadata_offset, evict_data_offset; 2043 int i, idx, offset, list_count, lists; 2044 2045 ASSERT(state == arc_mru || state == arc_mfu); 2046 2047 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2048 2049 /* 2050 * Decide which "type" (data vs metadata) to recycle from. 2051 * 2052 * If we are over the metadata limit, recycle from metadata. 2053 * If we are under the metadata minimum, recycle from data. 2054 * Otherwise, recycle from whichever type has the oldest (least 2055 * recently accessed) header. This is not yet implemented. 2056 */ 2057 if (recycle) { 2058 arc_buf_contents_t realtype; 2059 if (state->arcs_lsize[ARC_BUFC_DATA] == 0) { 2060 realtype = ARC_BUFC_METADATA; 2061 } else if (state->arcs_lsize[ARC_BUFC_METADATA] == 0) { 2062 realtype = ARC_BUFC_DATA; 2063 } else if (arc_meta_used >= arc_meta_limit) { 2064 realtype = ARC_BUFC_METADATA; 2065 } else if (arc_meta_used <= arc_meta_min) { 2066 realtype = ARC_BUFC_DATA; 2067 } else { 2068#ifdef illumos 2069 if (data_hdr->b_arc_access < 2070 metadata_hdr->b_arc_access) { 2071 realtype = ARC_BUFC_DATA; 2072 } else { 2073 realtype = ARC_BUFC_METADATA; 2074 } 2075#else 2076 /* TODO */ 2077 realtype = type; 2078#endif 2079 } 2080 if (realtype != type) { 2081 /* 2082 * If we want to evict from a different list, 2083 * we can not recycle, because DATA vs METADATA 2084 * buffers are segregated into different kmem 2085 * caches (and vmem arenas). 2086 */ 2087 type = realtype; 2088 recycle = B_FALSE; 2089 } 2090 } 2091 2092 if (type == ARC_BUFC_METADATA) { 2093 offset = 0; 2094 list_count = ARC_BUFC_NUMMETADATALISTS; 2095 list_start = &state->arcs_lists[0]; 2096 evicted_list_start = &evicted_state->arcs_lists[0]; 2097 idx = evict_metadata_offset; 2098 } else { 2099 offset = ARC_BUFC_NUMMETADATALISTS; 2100 list_start = &state->arcs_lists[offset]; 2101 evicted_list_start = &evicted_state->arcs_lists[offset]; 2102 list_count = ARC_BUFC_NUMDATALISTS; 2103 idx = evict_data_offset; 2104 } 2105 bytes_remaining = evicted_state->arcs_lsize[type]; 2106 lists = 0; 2107 2108evict_start: 2109 list = &list_start[idx]; 2110 evicted_list = &evicted_list_start[idx]; 2111 lock = ARCS_LOCK(state, (offset + idx)); 2112 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 2113 2114 mutex_enter(lock); 2115 mutex_enter(evicted_lock); 2116 2117 for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2118 hdr_prev = list_prev(list, hdr); 2119 bytes_remaining -= (hdr->b_size * hdr->b_datacnt); 2120 /* prefetch buffers have a minimum lifespan */ 2121 if (HDR_IO_IN_PROGRESS(hdr) || 2122 (spa && hdr->b_spa != spa) || 2123 (hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT) && 2124 ddi_get_lbolt() - hdr->b_arc_access < 2125 arc_min_prefetch_lifespan)) { 2126 skipped++; 2127 continue; 2128 } 2129 /* "lookahead" for better eviction candidate */ 2130 if (recycle && hdr->b_size != bytes && 2131 hdr_prev && hdr_prev->b_size == bytes) 2132 continue; 2133 2134 /* ignore markers */ 2135 if (hdr->b_spa == 0) 2136 continue; 2137 2138 /* 2139 * It may take a long time to evict all the bufs requested. 2140 * To avoid blocking all arc activity, periodically drop 2141 * the arcs_mtx and give other threads a chance to run 2142 * before reacquiring the lock. 2143 * 2144 * If we are looking for a buffer to recycle, we are in 2145 * the hot code path, so don't sleep. 2146 */ 2147 if (!recycle && count++ > arc_evict_iterations) { 2148 list_insert_after(list, hdr, &marker); 2149 mutex_exit(evicted_lock); 2150 mutex_exit(lock); 2151 kpreempt(KPREEMPT_SYNC); 2152 mutex_enter(lock); 2153 mutex_enter(evicted_lock); 2154 hdr_prev = list_prev(list, &marker); 2155 list_remove(list, &marker); 2156 count = 0; 2157 continue; 2158 } 2159 2160 hash_lock = HDR_LOCK(hdr); 2161 have_lock = MUTEX_HELD(hash_lock); 2162 if (have_lock || mutex_tryenter(hash_lock)) { 2163 ASSERT0(refcount_count(&hdr->b_refcnt)); 2164 ASSERT(hdr->b_datacnt > 0); 2165 while (hdr->b_buf) { 2166 arc_buf_t *buf = hdr->b_buf; 2167 if (!mutex_tryenter(&buf->b_evict_lock)) { 2168 missed += 1; 2169 break; 2170 } 2171 if (buf->b_data) { 2172 bytes_evicted += hdr->b_size; 2173 if (recycle && hdr->b_type == type && 2174 hdr->b_size == bytes && 2175 !HDR_L2_WRITING(hdr)) { 2176 stolen = buf->b_data; 2177 recycle = FALSE; 2178 } 2179 } 2180 if (buf->b_efunc) { 2181 mutex_enter(&arc_eviction_mtx); 2182 arc_buf_destroy(buf, 2183 buf->b_data == stolen, FALSE); 2184 hdr->b_buf = buf->b_next; 2185 buf->b_hdr = &arc_eviction_hdr; 2186 buf->b_next = arc_eviction_list; 2187 arc_eviction_list = buf; 2188 mutex_exit(&arc_eviction_mtx); 2189 mutex_exit(&buf->b_evict_lock); 2190 } else { 2191 mutex_exit(&buf->b_evict_lock); 2192 arc_buf_destroy(buf, 2193 buf->b_data == stolen, TRUE); 2194 } 2195 } 2196 2197 if (hdr->b_l2hdr) { 2198 ARCSTAT_INCR(arcstat_evict_l2_cached, 2199 hdr->b_size); 2200 } else { 2201 if (l2arc_write_eligible(hdr->b_spa, hdr)) { 2202 ARCSTAT_INCR(arcstat_evict_l2_eligible, 2203 hdr->b_size); 2204 } else { 2205 ARCSTAT_INCR( 2206 arcstat_evict_l2_ineligible, 2207 hdr->b_size); 2208 } 2209 } 2210 2211 if (hdr->b_datacnt == 0) { 2212 arc_change_state(evicted_state, hdr, hash_lock); 2213 ASSERT(HDR_IN_HASH_TABLE(hdr)); 2214 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 2215 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 2216 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 2217 } 2218 if (!have_lock) 2219 mutex_exit(hash_lock); 2220 if (bytes >= 0 && bytes_evicted >= bytes) 2221 break; 2222 if (bytes_remaining > 0) { 2223 mutex_exit(evicted_lock); 2224 mutex_exit(lock); 2225 idx = ((idx + 1) & (list_count - 1)); 2226 lists++; 2227 goto evict_start; 2228 } 2229 } else { 2230 missed += 1; 2231 } 2232 } 2233 2234 mutex_exit(evicted_lock); 2235 mutex_exit(lock); 2236 2237 idx = ((idx + 1) & (list_count - 1)); 2238 lists++; 2239 2240 if (bytes_evicted < bytes) { 2241 if (lists < list_count) 2242 goto evict_start; 2243 else 2244 dprintf("only evicted %lld bytes from %x", 2245 (longlong_t)bytes_evicted, state); 2246 } 2247 if (type == ARC_BUFC_METADATA) 2248 evict_metadata_offset = idx; 2249 else 2250 evict_data_offset = idx; 2251 2252 if (skipped) 2253 ARCSTAT_INCR(arcstat_evict_skip, skipped); 2254 2255 if (missed) 2256 ARCSTAT_INCR(arcstat_mutex_miss, missed); 2257 2258 /* 2259 * Note: we have just evicted some data into the ghost state, 2260 * potentially putting the ghost size over the desired size. Rather 2261 * that evicting from the ghost list in this hot code path, leave 2262 * this chore to the arc_reclaim_thread(). 2263 */ 2264 2265 if (stolen) 2266 ARCSTAT_BUMP(arcstat_stolen); 2267 return (stolen); 2268} 2269 2270/* 2271 * Remove buffers from list until we've removed the specified number of 2272 * bytes. Destroy the buffers that are removed. 2273 */ 2274static void 2275arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2276{ 2277 arc_buf_hdr_t *hdr, *hdr_prev; 2278 arc_buf_hdr_t marker = { 0 }; 2279 list_t *list, *list_start; 2280 kmutex_t *hash_lock, *lock; 2281 uint64_t bytes_deleted = 0; 2282 uint64_t bufs_skipped = 0; 2283 int count = 0; 2284 static int evict_offset; 2285 int list_count, idx = evict_offset; 2286 int offset, lists = 0; 2287 2288 ASSERT(GHOST_STATE(state)); 2289 2290 /* 2291 * data lists come after metadata lists 2292 */ 2293 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2294 list_count = ARC_BUFC_NUMDATALISTS; 2295 offset = ARC_BUFC_NUMMETADATALISTS; 2296 2297evict_start: 2298 list = &list_start[idx]; 2299 lock = ARCS_LOCK(state, idx + offset); 2300 2301 mutex_enter(lock); 2302 for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2303 hdr_prev = list_prev(list, hdr); 2304 if (hdr->b_type > ARC_BUFC_NUMTYPES) 2305 panic("invalid hdr=%p", (void *)hdr); 2306 if (spa && hdr->b_spa != spa) 2307 continue; 2308 2309 /* ignore markers */ 2310 if (hdr->b_spa == 0) 2311 continue; 2312 2313 hash_lock = HDR_LOCK(hdr); 2314 /* caller may be trying to modify this buffer, skip it */ 2315 if (MUTEX_HELD(hash_lock)) 2316 continue; 2317 2318 /* 2319 * It may take a long time to evict all the bufs requested. 2320 * To avoid blocking all arc activity, periodically drop 2321 * the arcs_mtx and give other threads a chance to run 2322 * before reacquiring the lock. 2323 */ 2324 if (count++ > arc_evict_iterations) { 2325 list_insert_after(list, hdr, &marker); 2326 mutex_exit(lock); 2327 kpreempt(KPREEMPT_SYNC); 2328 mutex_enter(lock); 2329 hdr_prev = list_prev(list, &marker); 2330 list_remove(list, &marker); 2331 count = 0; 2332 continue; 2333 } 2334 if (mutex_tryenter(hash_lock)) { 2335 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2336 ASSERT(hdr->b_buf == NULL); 2337 ARCSTAT_BUMP(arcstat_deleted); 2338 bytes_deleted += hdr->b_size; 2339 2340 if (hdr->b_l2hdr != NULL) { 2341 /* 2342 * This buffer is cached on the 2nd Level ARC; 2343 * don't destroy the header. 2344 */ 2345 arc_change_state(arc_l2c_only, hdr, hash_lock); 2346 mutex_exit(hash_lock); 2347 } else { 2348 arc_change_state(arc_anon, hdr, hash_lock); 2349 mutex_exit(hash_lock); 2350 arc_hdr_destroy(hdr); 2351 } 2352 2353 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 2354 if (bytes >= 0 && bytes_deleted >= bytes) 2355 break; 2356 } else if (bytes < 0) { 2357 /* 2358 * Insert a list marker and then wait for the 2359 * hash lock to become available. Once its 2360 * available, restart from where we left off. 2361 */ 2362 list_insert_after(list, hdr, &marker); 2363 mutex_exit(lock); 2364 mutex_enter(hash_lock); 2365 mutex_exit(hash_lock); 2366 mutex_enter(lock); 2367 hdr_prev = list_prev(list, &marker); 2368 list_remove(list, &marker); 2369 } else { 2370 bufs_skipped += 1; 2371 } 2372 2373 } 2374 mutex_exit(lock); 2375 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2376 lists++; 2377 2378 if (lists < list_count) 2379 goto evict_start; 2380 2381 evict_offset = idx; 2382 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2383 (bytes < 0 || bytes_deleted < bytes)) { 2384 list_start = &state->arcs_lists[0]; 2385 list_count = ARC_BUFC_NUMMETADATALISTS; 2386 offset = lists = 0; 2387 goto evict_start; 2388 } 2389 2390 if (bufs_skipped) { 2391 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2392 ASSERT(bytes >= 0); 2393 } 2394 2395 if (bytes_deleted < bytes) 2396 dprintf("only deleted %lld bytes from %p", 2397 (longlong_t)bytes_deleted, state); 2398} 2399 2400static void 2401arc_adjust(void) 2402{ 2403 int64_t adjustment, delta; 2404 2405 /* 2406 * Adjust MRU size 2407 */ 2408 2409 adjustment = MIN((int64_t)(arc_size - arc_c), 2410 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2411 arc_p)); 2412 2413 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2414 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2415 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2416 adjustment -= delta; 2417 } 2418 2419 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2420 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2421 (void) arc_evict(arc_mru, 0, delta, FALSE, 2422 ARC_BUFC_METADATA); 2423 } 2424 2425 /* 2426 * Adjust MFU size 2427 */ 2428 2429 adjustment = arc_size - arc_c; 2430 2431 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2432 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2433 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2434 adjustment -= delta; 2435 } 2436 2437 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2438 int64_t delta = MIN(adjustment, 2439 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2440 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2441 ARC_BUFC_METADATA); 2442 } 2443 2444 /* 2445 * Adjust ghost lists 2446 */ 2447 2448 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2449 2450 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2451 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2452 arc_evict_ghost(arc_mru_ghost, 0, delta); 2453 } 2454 2455 adjustment = 2456 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2457 2458 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2459 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2460 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2461 } 2462} 2463 2464static void 2465arc_do_user_evicts(void) 2466{ 2467 static arc_buf_t *tmp_arc_eviction_list; 2468 2469 /* 2470 * Move list over to avoid LOR 2471 */ 2472restart: 2473 mutex_enter(&arc_eviction_mtx); 2474 tmp_arc_eviction_list = arc_eviction_list; 2475 arc_eviction_list = NULL; 2476 mutex_exit(&arc_eviction_mtx); 2477 2478 while (tmp_arc_eviction_list != NULL) { 2479 arc_buf_t *buf = tmp_arc_eviction_list; 2480 tmp_arc_eviction_list = buf->b_next; 2481 mutex_enter(&buf->b_evict_lock); 2482 buf->b_hdr = NULL; 2483 mutex_exit(&buf->b_evict_lock); 2484 2485 if (buf->b_efunc != NULL) 2486 VERIFY0(buf->b_efunc(buf->b_private)); 2487 2488 buf->b_efunc = NULL; 2489 buf->b_private = NULL; 2490 kmem_cache_free(buf_cache, buf); 2491 } 2492 2493 if (arc_eviction_list != NULL) 2494 goto restart; 2495} 2496 2497/* 2498 * Flush all *evictable* data from the cache for the given spa. 2499 * NOTE: this will not touch "active" (i.e. referenced) data. 2500 */ 2501void 2502arc_flush(spa_t *spa) 2503{ 2504 uint64_t guid = 0; 2505 2506 if (spa) 2507 guid = spa_load_guid(spa); 2508 2509 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2510 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2511 if (spa) 2512 break; 2513 } 2514 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2515 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2516 if (spa) 2517 break; 2518 } 2519 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2520 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2521 if (spa) 2522 break; 2523 } 2524 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2525 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2526 if (spa) 2527 break; 2528 } 2529 2530 arc_evict_ghost(arc_mru_ghost, guid, -1); 2531 arc_evict_ghost(arc_mfu_ghost, guid, -1); 2532 2533 mutex_enter(&arc_reclaim_thr_lock); 2534 arc_do_user_evicts(); 2535 mutex_exit(&arc_reclaim_thr_lock); 2536 ASSERT(spa || arc_eviction_list == NULL); 2537} 2538 2539void 2540arc_shrink(void) 2541{ 2542 2543 if (arc_c > arc_c_min) { 2544 uint64_t to_free; 2545 2546 to_free = arc_c >> arc_shrink_shift; 2547 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 2548 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 2549 if (arc_c > arc_c_min + to_free) 2550 atomic_add_64(&arc_c, -to_free); 2551 else 2552 arc_c = arc_c_min; 2553 2554 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2555 if (arc_c > arc_size) 2556 arc_c = MAX(arc_size, arc_c_min); 2557 if (arc_p > arc_c) 2558 arc_p = (arc_c >> 1); 2559 2560 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 2561 arc_p); 2562 2563 ASSERT(arc_c >= arc_c_min); 2564 ASSERT((int64_t)arc_p >= 0); 2565 } 2566 2567 if (arc_size > arc_c) { 2568 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 2569 uint64_t, arc_c); 2570 arc_adjust(); 2571 } 2572} 2573 2574static int needfree = 0; 2575 2576static int 2577arc_reclaim_needed(void) 2578{ 2579 2580#ifdef _KERNEL 2581 2582 if (needfree) { 2583 DTRACE_PROBE(arc__reclaim_needfree); 2584 return (1); 2585 } 2586 2587 /* 2588 * Cooperate with pagedaemon when it's time for it to scan 2589 * and reclaim some pages. 2590 */ 2591 if (freemem < zfs_arc_free_target) { 2592 DTRACE_PROBE2(arc__reclaim_freemem, uint64_t, 2593 freemem, uint64_t, zfs_arc_free_target); 2594 return (1); 2595 } 2596 2597#ifdef sun 2598 /* 2599 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2600 */ 2601 extra = desfree; 2602 2603 /* 2604 * check that we're out of range of the pageout scanner. It starts to 2605 * schedule paging if freemem is less than lotsfree and needfree. 2606 * lotsfree is the high-water mark for pageout, and needfree is the 2607 * number of needed free pages. We add extra pages here to make sure 2608 * the scanner doesn't start up while we're freeing memory. 2609 */ 2610 if (freemem < lotsfree + needfree + extra) 2611 return (1); 2612 2613 /* 2614 * check to make sure that swapfs has enough space so that anon 2615 * reservations can still succeed. anon_resvmem() checks that the 2616 * availrmem is greater than swapfs_minfree, and the number of reserved 2617 * swap pages. We also add a bit of extra here just to prevent 2618 * circumstances from getting really dire. 2619 */ 2620 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2621 return (1); 2622 2623 /* 2624 * Check that we have enough availrmem that memory locking (e.g., via 2625 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 2626 * stores the number of pages that cannot be locked; when availrmem 2627 * drops below pages_pp_maximum, page locking mechanisms such as 2628 * page_pp_lock() will fail.) 2629 */ 2630 if (availrmem <= pages_pp_maximum) 2631 return (1); 2632 2633#endif /* sun */ 2634#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 2635 /* 2636 * If we're on an i386 platform, it's possible that we'll exhaust the 2637 * kernel heap space before we ever run out of available physical 2638 * memory. Most checks of the size of the heap_area compare against 2639 * tune.t_minarmem, which is the minimum available real memory that we 2640 * can have in the system. However, this is generally fixed at 25 pages 2641 * which is so low that it's useless. In this comparison, we seek to 2642 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2643 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2644 * free) 2645 */ 2646 if (vmem_size(heap_arena, VMEM_FREE) < 2647 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2)) { 2648 DTRACE_PROBE2(arc__reclaim_used, uint64_t, 2649 vmem_size(heap_arena, VMEM_FREE), uint64_t, 2650 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2); 2651 return (1); 2652 } 2653#define zio_arena NULL 2654#else 2655#define zio_arena heap_arena 2656#endif 2657 2658 /* 2659 * If zio data pages are being allocated out of a separate heap segment, 2660 * then enforce that the size of available vmem for this arena remains 2661 * above about 1/16th free. 2662 * 2663 * Note: The 1/16th arena free requirement was put in place 2664 * to aggressively evict memory from the arc in order to avoid 2665 * memory fragmentation issues. 2666 */ 2667 if (zio_arena != NULL && 2668 vmem_size(zio_arena, VMEM_FREE) < 2669 (vmem_size(zio_arena, VMEM_ALLOC) >> 4)) 2670 return (1); 2671 2672 /* 2673 * Above limits know nothing about real level of KVA fragmentation. 2674 * Start aggressive reclamation if too little sequential KVA left. 2675 */ 2676 if (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) 2677 return (1); 2678 2679#else /* _KERNEL */ 2680 if (spa_get_random(100) == 0) 2681 return (1); 2682#endif /* _KERNEL */ 2683 DTRACE_PROBE(arc__reclaim_no); 2684 2685 return (0); 2686} 2687 2688extern kmem_cache_t *zio_buf_cache[]; 2689extern kmem_cache_t *zio_data_buf_cache[]; 2690extern kmem_cache_t *range_seg_cache; 2691 2692static __noinline void 2693arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2694{ 2695 size_t i; 2696 kmem_cache_t *prev_cache = NULL; 2697 kmem_cache_t *prev_data_cache = NULL; 2698 2699 DTRACE_PROBE(arc__kmem_reap_start); 2700#ifdef _KERNEL 2701 if (arc_meta_used >= arc_meta_limit) { 2702 /* 2703 * We are exceeding our meta-data cache limit. 2704 * Purge some DNLC entries to release holds on meta-data. 2705 */ 2706 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2707 } 2708#if defined(__i386) 2709 /* 2710 * Reclaim unused memory from all kmem caches. 2711 */ 2712 kmem_reap(); 2713#endif 2714#endif 2715 2716 /* 2717 * An aggressive reclamation will shrink the cache size as well as 2718 * reap free buffers from the arc kmem caches. 2719 */ 2720 if (strat == ARC_RECLAIM_AGGR) 2721 arc_shrink(); 2722 2723 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2724 if (zio_buf_cache[i] != prev_cache) { 2725 prev_cache = zio_buf_cache[i]; 2726 kmem_cache_reap_now(zio_buf_cache[i]); 2727 } 2728 if (zio_data_buf_cache[i] != prev_data_cache) { 2729 prev_data_cache = zio_data_buf_cache[i]; 2730 kmem_cache_reap_now(zio_data_buf_cache[i]); 2731 } 2732 } 2733 kmem_cache_reap_now(buf_cache); 2734 kmem_cache_reap_now(hdr_cache); 2735 kmem_cache_reap_now(range_seg_cache); 2736 2737#ifdef sun 2738 /* 2739 * Ask the vmem arena to reclaim unused memory from its 2740 * quantum caches. 2741 */ 2742 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR) 2743 vmem_qcache_reap(zio_arena); 2744#endif 2745 DTRACE_PROBE(arc__kmem_reap_end); 2746} 2747 2748static void 2749arc_reclaim_thread(void *dummy __unused) 2750{ 2751 clock_t growtime = 0; 2752 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2753 callb_cpr_t cpr; 2754 2755 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2756 2757 mutex_enter(&arc_reclaim_thr_lock); 2758 while (arc_thread_exit == 0) { 2759 if (arc_reclaim_needed()) { 2760 2761 if (arc_no_grow) { 2762 if (last_reclaim == ARC_RECLAIM_CONS) { 2763 DTRACE_PROBE(arc__reclaim_aggr_no_grow); 2764 last_reclaim = ARC_RECLAIM_AGGR; 2765 } else { 2766 last_reclaim = ARC_RECLAIM_CONS; 2767 } 2768 } else { 2769 arc_no_grow = TRUE; 2770 last_reclaim = ARC_RECLAIM_AGGR; 2771 DTRACE_PROBE(arc__reclaim_aggr); 2772 membar_producer(); 2773 } 2774 2775 /* reset the growth delay for every reclaim */ 2776 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2777 2778 if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2779 /* 2780 * If needfree is TRUE our vm_lowmem hook 2781 * was called and in that case we must free some 2782 * memory, so switch to aggressive mode. 2783 */ 2784 arc_no_grow = TRUE; 2785 last_reclaim = ARC_RECLAIM_AGGR; 2786 } 2787 arc_kmem_reap_now(last_reclaim); 2788 arc_warm = B_TRUE; 2789 2790 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2791 arc_no_grow = FALSE; 2792 } 2793 2794 arc_adjust(); 2795 2796 if (arc_eviction_list != NULL) 2797 arc_do_user_evicts(); 2798 2799#ifdef _KERNEL 2800 if (needfree) { 2801 needfree = 0; 2802 wakeup(&needfree); 2803 } 2804#endif 2805 2806 /* block until needed, or one second, whichever is shorter */ 2807 CALLB_CPR_SAFE_BEGIN(&cpr); 2808 (void) cv_timedwait(&arc_reclaim_thr_cv, 2809 &arc_reclaim_thr_lock, hz); 2810 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2811 } 2812 2813 arc_thread_exit = 0; 2814 cv_broadcast(&arc_reclaim_thr_cv); 2815 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2816 thread_exit(); 2817} 2818 2819/* 2820 * Adapt arc info given the number of bytes we are trying to add and 2821 * the state that we are comming from. This function is only called 2822 * when we are adding new content to the cache. 2823 */ 2824static void 2825arc_adapt(int bytes, arc_state_t *state) 2826{ 2827 int mult; 2828 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2829 2830 if (state == arc_l2c_only) 2831 return; 2832 2833 ASSERT(bytes > 0); 2834 /* 2835 * Adapt the target size of the MRU list: 2836 * - if we just hit in the MRU ghost list, then increase 2837 * the target size of the MRU list. 2838 * - if we just hit in the MFU ghost list, then increase 2839 * the target size of the MFU list by decreasing the 2840 * target size of the MRU list. 2841 */ 2842 if (state == arc_mru_ghost) { 2843 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2844 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2845 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2846 2847 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2848 } else if (state == arc_mfu_ghost) { 2849 uint64_t delta; 2850 2851 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2852 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2853 mult = MIN(mult, 10); 2854 2855 delta = MIN(bytes * mult, arc_p); 2856 arc_p = MAX(arc_p_min, arc_p - delta); 2857 } 2858 ASSERT((int64_t)arc_p >= 0); 2859 2860 if (arc_reclaim_needed()) { 2861 cv_signal(&arc_reclaim_thr_cv); 2862 return; 2863 } 2864 2865 if (arc_no_grow) 2866 return; 2867 2868 if (arc_c >= arc_c_max) 2869 return; 2870 2871 /* 2872 * If we're within (2 * maxblocksize) bytes of the target 2873 * cache size, increment the target cache size 2874 */ 2875 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2876 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 2877 atomic_add_64(&arc_c, (int64_t)bytes); 2878 if (arc_c > arc_c_max) 2879 arc_c = arc_c_max; 2880 else if (state == arc_anon) 2881 atomic_add_64(&arc_p, (int64_t)bytes); 2882 if (arc_p > arc_c) 2883 arc_p = arc_c; 2884 } 2885 ASSERT((int64_t)arc_p >= 0); 2886} 2887 2888/* 2889 * Check if the cache has reached its limits and eviction is required 2890 * prior to insert. 2891 */ 2892static int 2893arc_evict_needed(arc_buf_contents_t type) 2894{ 2895 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2896 return (1); 2897 2898 if (arc_reclaim_needed()) 2899 return (1); 2900 2901 return (arc_size > arc_c); 2902} 2903 2904/* 2905 * The buffer, supplied as the first argument, needs a data block. 2906 * So, if we are at cache max, determine which cache should be victimized. 2907 * We have the following cases: 2908 * 2909 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2910 * In this situation if we're out of space, but the resident size of the MFU is 2911 * under the limit, victimize the MFU cache to satisfy this insertion request. 2912 * 2913 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2914 * Here, we've used up all of the available space for the MRU, so we need to 2915 * evict from our own cache instead. Evict from the set of resident MRU 2916 * entries. 2917 * 2918 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2919 * c minus p represents the MFU space in the cache, since p is the size of the 2920 * cache that is dedicated to the MRU. In this situation there's still space on 2921 * the MFU side, so the MRU side needs to be victimized. 2922 * 2923 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2924 * MFU's resident set is consuming more space than it has been allotted. In 2925 * this situation, we must victimize our own cache, the MFU, for this insertion. 2926 */ 2927static void 2928arc_get_data_buf(arc_buf_t *buf) 2929{ 2930 arc_state_t *state = buf->b_hdr->b_state; 2931 uint64_t size = buf->b_hdr->b_size; 2932 arc_buf_contents_t type = buf->b_hdr->b_type; 2933 2934 arc_adapt(size, state); 2935 2936 /* 2937 * We have not yet reached cache maximum size, 2938 * just allocate a new buffer. 2939 */ 2940 if (!arc_evict_needed(type)) { 2941 if (type == ARC_BUFC_METADATA) { 2942 buf->b_data = zio_buf_alloc(size); 2943 arc_space_consume(size, ARC_SPACE_DATA); 2944 } else { 2945 ASSERT(type == ARC_BUFC_DATA); 2946 buf->b_data = zio_data_buf_alloc(size); 2947 ARCSTAT_INCR(arcstat_data_size, size); 2948 atomic_add_64(&arc_size, size); 2949 } 2950 goto out; 2951 } 2952 2953 /* 2954 * If we are prefetching from the mfu ghost list, this buffer 2955 * will end up on the mru list; so steal space from there. 2956 */ 2957 if (state == arc_mfu_ghost) 2958 state = buf->b_hdr->b_flags & ARC_FLAG_PREFETCH ? 2959 arc_mru : arc_mfu; 2960 else if (state == arc_mru_ghost) 2961 state = arc_mru; 2962 2963 if (state == arc_mru || state == arc_anon) { 2964 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2965 state = (arc_mfu->arcs_lsize[type] >= size && 2966 arc_p > mru_used) ? arc_mfu : arc_mru; 2967 } else { 2968 /* MFU cases */ 2969 uint64_t mfu_space = arc_c - arc_p; 2970 state = (arc_mru->arcs_lsize[type] >= size && 2971 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2972 } 2973 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2974 if (type == ARC_BUFC_METADATA) { 2975 buf->b_data = zio_buf_alloc(size); 2976 arc_space_consume(size, ARC_SPACE_DATA); 2977 } else { 2978 ASSERT(type == ARC_BUFC_DATA); 2979 buf->b_data = zio_data_buf_alloc(size); 2980 ARCSTAT_INCR(arcstat_data_size, size); 2981 atomic_add_64(&arc_size, size); 2982 } 2983 ARCSTAT_BUMP(arcstat_recycle_miss); 2984 } 2985 ASSERT(buf->b_data != NULL); 2986out: 2987 /* 2988 * Update the state size. Note that ghost states have a 2989 * "ghost size" and so don't need to be updated. 2990 */ 2991 if (!GHOST_STATE(buf->b_hdr->b_state)) { 2992 arc_buf_hdr_t *hdr = buf->b_hdr; 2993 2994 atomic_add_64(&hdr->b_state->arcs_size, size); 2995 if (list_link_active(&hdr->b_arc_node)) { 2996 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2997 atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2998 } 2999 /* 3000 * If we are growing the cache, and we are adding anonymous 3001 * data, and we have outgrown arc_p, update arc_p 3002 */ 3003 if (arc_size < arc_c && hdr->b_state == arc_anon && 3004 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 3005 arc_p = MIN(arc_c, arc_p + size); 3006 } 3007 ARCSTAT_BUMP(arcstat_allocated); 3008} 3009 3010/* 3011 * This routine is called whenever a buffer is accessed. 3012 * NOTE: the hash lock is dropped in this function. 3013 */ 3014static void 3015arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3016{ 3017 clock_t now; 3018 3019 ASSERT(MUTEX_HELD(hash_lock)); 3020 3021 if (hdr->b_state == arc_anon) { 3022 /* 3023 * This buffer is not in the cache, and does not 3024 * appear in our "ghost" list. Add the new buffer 3025 * to the MRU state. 3026 */ 3027 3028 ASSERT(hdr->b_arc_access == 0); 3029 hdr->b_arc_access = ddi_get_lbolt(); 3030 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3031 arc_change_state(arc_mru, hdr, hash_lock); 3032 3033 } else if (hdr->b_state == arc_mru) { 3034 now = ddi_get_lbolt(); 3035 3036 /* 3037 * If this buffer is here because of a prefetch, then either: 3038 * - clear the flag if this is a "referencing" read 3039 * (any subsequent access will bump this into the MFU state). 3040 * or 3041 * - move the buffer to the head of the list if this is 3042 * another prefetch (to make it less likely to be evicted). 3043 */ 3044 if ((hdr->b_flags & ARC_FLAG_PREFETCH) != 0) { 3045 if (refcount_count(&hdr->b_refcnt) == 0) { 3046 ASSERT(list_link_active(&hdr->b_arc_node)); 3047 } else { 3048 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3049 ARCSTAT_BUMP(arcstat_mru_hits); 3050 } 3051 hdr->b_arc_access = now; 3052 return; 3053 } 3054 3055 /* 3056 * This buffer has been "accessed" only once so far, 3057 * but it is still in the cache. Move it to the MFU 3058 * state. 3059 */ 3060 if (now > hdr->b_arc_access + ARC_MINTIME) { 3061 /* 3062 * More than 125ms have passed since we 3063 * instantiated this buffer. Move it to the 3064 * most frequently used state. 3065 */ 3066 hdr->b_arc_access = now; 3067 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3068 arc_change_state(arc_mfu, hdr, hash_lock); 3069 } 3070 ARCSTAT_BUMP(arcstat_mru_hits); 3071 } else if (hdr->b_state == arc_mru_ghost) { 3072 arc_state_t *new_state; 3073 /* 3074 * This buffer has been "accessed" recently, but 3075 * was evicted from the cache. Move it to the 3076 * MFU state. 3077 */ 3078 3079 if (hdr->b_flags & ARC_FLAG_PREFETCH) { 3080 new_state = arc_mru; 3081 if (refcount_count(&hdr->b_refcnt) > 0) 3082 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3083 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3084 } else { 3085 new_state = arc_mfu; 3086 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3087 } 3088 3089 hdr->b_arc_access = ddi_get_lbolt(); 3090 arc_change_state(new_state, hdr, hash_lock); 3091 3092 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3093 } else if (hdr->b_state == arc_mfu) { 3094 /* 3095 * This buffer has been accessed more than once and is 3096 * still in the cache. Keep it in the MFU state. 3097 * 3098 * NOTE: an add_reference() that occurred when we did 3099 * the arc_read() will have kicked this off the list. 3100 * If it was a prefetch, we will explicitly move it to 3101 * the head of the list now. 3102 */ 3103 if ((hdr->b_flags & ARC_FLAG_PREFETCH) != 0) { 3104 ASSERT(refcount_count(&hdr->b_refcnt) == 0); 3105 ASSERT(list_link_active(&hdr->b_arc_node)); 3106 } 3107 ARCSTAT_BUMP(arcstat_mfu_hits); 3108 hdr->b_arc_access = ddi_get_lbolt(); 3109 } else if (hdr->b_state == arc_mfu_ghost) { 3110 arc_state_t *new_state = arc_mfu; 3111 /* 3112 * This buffer has been accessed more than once but has 3113 * been evicted from the cache. Move it back to the 3114 * MFU state. 3115 */ 3116 3117 if (hdr->b_flags & ARC_FLAG_PREFETCH) { 3118 /* 3119 * This is a prefetch access... 3120 * move this block back to the MRU state. 3121 */ 3122 ASSERT0(refcount_count(&hdr->b_refcnt)); 3123 new_state = arc_mru; 3124 } 3125 3126 hdr->b_arc_access = ddi_get_lbolt(); 3127 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3128 arc_change_state(new_state, hdr, hash_lock); 3129 3130 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 3131 } else if (hdr->b_state == arc_l2c_only) { 3132 /* 3133 * This buffer is on the 2nd Level ARC. 3134 */ 3135 3136 hdr->b_arc_access = ddi_get_lbolt(); 3137 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3138 arc_change_state(arc_mfu, hdr, hash_lock); 3139 } else { 3140 ASSERT(!"invalid arc state"); 3141 } 3142} 3143 3144/* a generic arc_done_func_t which you can use */ 3145/* ARGSUSED */ 3146void 3147arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 3148{ 3149 if (zio == NULL || zio->io_error == 0) 3150 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 3151 VERIFY(arc_buf_remove_ref(buf, arg)); 3152} 3153 3154/* a generic arc_done_func_t */ 3155void 3156arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 3157{ 3158 arc_buf_t **bufp = arg; 3159 if (zio && zio->io_error) { 3160 VERIFY(arc_buf_remove_ref(buf, arg)); 3161 *bufp = NULL; 3162 } else { 3163 *bufp = buf; 3164 ASSERT(buf->b_data); 3165 } 3166} 3167 3168static void 3169arc_read_done(zio_t *zio) 3170{ 3171 arc_buf_hdr_t *hdr; 3172 arc_buf_t *buf; 3173 arc_buf_t *abuf; /* buffer we're assigning to callback */ 3174 kmutex_t *hash_lock = NULL; 3175 arc_callback_t *callback_list, *acb; 3176 int freeable = FALSE; 3177 3178 buf = zio->io_private; 3179 hdr = buf->b_hdr; 3180 3181 /* 3182 * The hdr was inserted into hash-table and removed from lists 3183 * prior to starting I/O. We should find this header, since 3184 * it's in the hash table, and it should be legit since it's 3185 * not possible to evict it during the I/O. The only possible 3186 * reason for it not to be found is if we were freed during the 3187 * read. 3188 */ 3189 if (HDR_IN_HASH_TABLE(hdr)) { 3190 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3191 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3192 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3193 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3194 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3195 3196 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3197 &hash_lock); 3198 3199 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3200 hash_lock == NULL) || 3201 (found == hdr && 3202 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3203 (found == hdr && HDR_L2_READING(hdr))); 3204 } 3205 3206 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; 3207 if (l2arc_noprefetch && (hdr->b_flags & ARC_FLAG_PREFETCH)) 3208 hdr->b_flags &= ~ARC_FLAG_L2CACHE; 3209 3210 /* byteswap if necessary */ 3211 callback_list = hdr->b_acb; 3212 ASSERT(callback_list != NULL); 3213 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3214 dmu_object_byteswap_t bswap = 3215 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3216 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3217 byteswap_uint64_array : 3218 dmu_ot_byteswap[bswap].ob_func; 3219 func(buf->b_data, hdr->b_size); 3220 } 3221 3222 arc_cksum_compute(buf, B_FALSE); 3223#ifdef illumos 3224 arc_buf_watch(buf); 3225#endif /* illumos */ 3226 3227 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3228 /* 3229 * Only call arc_access on anonymous buffers. This is because 3230 * if we've issued an I/O for an evicted buffer, we've already 3231 * called arc_access (to prevent any simultaneous readers from 3232 * getting confused). 3233 */ 3234 arc_access(hdr, hash_lock); 3235 } 3236 3237 /* create copies of the data buffer for the callers */ 3238 abuf = buf; 3239 for (acb = callback_list; acb; acb = acb->acb_next) { 3240 if (acb->acb_done) { 3241 if (abuf == NULL) { 3242 ARCSTAT_BUMP(arcstat_duplicate_reads); 3243 abuf = arc_buf_clone(buf); 3244 } 3245 acb->acb_buf = abuf; 3246 abuf = NULL; 3247 } 3248 } 3249 hdr->b_acb = NULL; 3250 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 3251 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3252 if (abuf == buf) { 3253 ASSERT(buf->b_efunc == NULL); 3254 ASSERT(hdr->b_datacnt == 1); 3255 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 3256 } 3257 3258 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3259 3260 if (zio->io_error != 0) { 3261 hdr->b_flags |= ARC_FLAG_IO_ERROR; 3262 if (hdr->b_state != arc_anon) 3263 arc_change_state(arc_anon, hdr, hash_lock); 3264 if (HDR_IN_HASH_TABLE(hdr)) 3265 buf_hash_remove(hdr); 3266 freeable = refcount_is_zero(&hdr->b_refcnt); 3267 } 3268 3269 /* 3270 * Broadcast before we drop the hash_lock to avoid the possibility 3271 * that the hdr (and hence the cv) might be freed before we get to 3272 * the cv_broadcast(). 3273 */ 3274 cv_broadcast(&hdr->b_cv); 3275 3276 if (hash_lock) { 3277 mutex_exit(hash_lock); 3278 } else { 3279 /* 3280 * This block was freed while we waited for the read to 3281 * complete. It has been removed from the hash table and 3282 * moved to the anonymous state (so that it won't show up 3283 * in the cache). 3284 */ 3285 ASSERT3P(hdr->b_state, ==, arc_anon); 3286 freeable = refcount_is_zero(&hdr->b_refcnt); 3287 } 3288 3289 /* execute each callback and free its structure */ 3290 while ((acb = callback_list) != NULL) { 3291 if (acb->acb_done) 3292 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3293 3294 if (acb->acb_zio_dummy != NULL) { 3295 acb->acb_zio_dummy->io_error = zio->io_error; 3296 zio_nowait(acb->acb_zio_dummy); 3297 } 3298 3299 callback_list = acb->acb_next; 3300 kmem_free(acb, sizeof (arc_callback_t)); 3301 } 3302 3303 if (freeable) 3304 arc_hdr_destroy(hdr); 3305} 3306 3307/* 3308 * "Read" the block block at the specified DVA (in bp) via the 3309 * cache. If the block is found in the cache, invoke the provided 3310 * callback immediately and return. Note that the `zio' parameter 3311 * in the callback will be NULL in this case, since no IO was 3312 * required. If the block is not in the cache pass the read request 3313 * on to the spa with a substitute callback function, so that the 3314 * requested block will be added to the cache. 3315 * 3316 * If a read request arrives for a block that has a read in-progress, 3317 * either wait for the in-progress read to complete (and return the 3318 * results); or, if this is a read with a "done" func, add a record 3319 * to the read to invoke the "done" func when the read completes, 3320 * and return; or just return. 3321 * 3322 * arc_read_done() will invoke all the requested "done" functions 3323 * for readers of this block. 3324 */ 3325int 3326arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3327 void *private, zio_priority_t priority, int zio_flags, 3328 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 3329{ 3330 arc_buf_hdr_t *hdr = NULL; 3331 arc_buf_t *buf = NULL; 3332 kmutex_t *hash_lock = NULL; 3333 zio_t *rzio; 3334 uint64_t guid = spa_load_guid(spa); 3335 3336 ASSERT(!BP_IS_EMBEDDED(bp) || 3337 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3338 3339top: 3340 if (!BP_IS_EMBEDDED(bp)) { 3341 /* 3342 * Embedded BP's have no DVA and require no I/O to "read". 3343 * Create an anonymous arc buf to back it. 3344 */ 3345 hdr = buf_hash_find(guid, bp, &hash_lock); 3346 } 3347 3348 if (hdr != NULL && hdr->b_datacnt > 0) { 3349 3350 *arc_flags |= ARC_FLAG_CACHED; 3351 3352 if (HDR_IO_IN_PROGRESS(hdr)) { 3353 3354 if (*arc_flags & ARC_FLAG_WAIT) { 3355 cv_wait(&hdr->b_cv, hash_lock); 3356 mutex_exit(hash_lock); 3357 goto top; 3358 } 3359 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 3360 3361 if (done) { 3362 arc_callback_t *acb = NULL; 3363 3364 acb = kmem_zalloc(sizeof (arc_callback_t), 3365 KM_SLEEP); 3366 acb->acb_done = done; 3367 acb->acb_private = private; 3368 if (pio != NULL) 3369 acb->acb_zio_dummy = zio_null(pio, 3370 spa, NULL, NULL, NULL, zio_flags); 3371 3372 ASSERT(acb->acb_done != NULL); 3373 acb->acb_next = hdr->b_acb; 3374 hdr->b_acb = acb; 3375 add_reference(hdr, hash_lock, private); 3376 mutex_exit(hash_lock); 3377 return (0); 3378 } 3379 mutex_exit(hash_lock); 3380 return (0); 3381 } 3382 3383 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3384 3385 if (done) { 3386 add_reference(hdr, hash_lock, private); 3387 /* 3388 * If this block is already in use, create a new 3389 * copy of the data so that we will be guaranteed 3390 * that arc_release() will always succeed. 3391 */ 3392 buf = hdr->b_buf; 3393 ASSERT(buf); 3394 ASSERT(buf->b_data); 3395 if (HDR_BUF_AVAILABLE(hdr)) { 3396 ASSERT(buf->b_efunc == NULL); 3397 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 3398 } else { 3399 buf = arc_buf_clone(buf); 3400 } 3401 3402 } else if (*arc_flags & ARC_FLAG_PREFETCH && 3403 refcount_count(&hdr->b_refcnt) == 0) { 3404 hdr->b_flags |= ARC_FLAG_PREFETCH; 3405 } 3406 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3407 arc_access(hdr, hash_lock); 3408 if (*arc_flags & ARC_FLAG_L2CACHE) 3409 hdr->b_flags |= ARC_FLAG_L2CACHE; 3410 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3411 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3412 mutex_exit(hash_lock); 3413 ARCSTAT_BUMP(arcstat_hits); 3414 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 3415 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3416 data, metadata, hits); 3417 3418 if (done) 3419 done(NULL, buf, private); 3420 } else { 3421 uint64_t size = BP_GET_LSIZE(bp); 3422 arc_callback_t *acb; 3423 vdev_t *vd = NULL; 3424 uint64_t addr = 0; 3425 boolean_t devw = B_FALSE; 3426 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3427 uint64_t b_asize = 0; 3428 3429 if (hdr == NULL) { 3430 /* this block is not in the cache */ 3431 arc_buf_hdr_t *exists = NULL; 3432 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3433 buf = arc_buf_alloc(spa, size, private, type); 3434 hdr = buf->b_hdr; 3435 if (!BP_IS_EMBEDDED(bp)) { 3436 hdr->b_dva = *BP_IDENTITY(bp); 3437 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3438 hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3439 exists = buf_hash_insert(hdr, &hash_lock); 3440 } 3441 if (exists != NULL) { 3442 /* somebody beat us to the hash insert */ 3443 mutex_exit(hash_lock); 3444 buf_discard_identity(hdr); 3445 (void) arc_buf_remove_ref(buf, private); 3446 goto top; /* restart the IO request */ 3447 } 3448 3449 /* if this is a prefetch, we don't have a reference */ 3450 if (*arc_flags & ARC_FLAG_PREFETCH) { 3451 (void) remove_reference(hdr, hash_lock, 3452 private); 3453 hdr->b_flags |= ARC_FLAG_PREFETCH; 3454 } 3455 if (*arc_flags & ARC_FLAG_L2CACHE) 3456 hdr->b_flags |= ARC_FLAG_L2CACHE; 3457 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3458 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3459 if (BP_GET_LEVEL(bp) > 0) 3460 hdr->b_flags |= ARC_FLAG_INDIRECT; 3461 } else { 3462 /* this block is in the ghost cache */ 3463 ASSERT(GHOST_STATE(hdr->b_state)); 3464 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3465 ASSERT0(refcount_count(&hdr->b_refcnt)); 3466 ASSERT(hdr->b_buf == NULL); 3467 3468 /* if this is a prefetch, we don't have a reference */ 3469 if (*arc_flags & ARC_FLAG_PREFETCH) 3470 hdr->b_flags |= ARC_FLAG_PREFETCH; 3471 else 3472 add_reference(hdr, hash_lock, private); 3473 if (*arc_flags & ARC_FLAG_L2CACHE) 3474 hdr->b_flags |= ARC_FLAG_L2CACHE; 3475 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3476 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3477 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3478 buf->b_hdr = hdr; 3479 buf->b_data = NULL; 3480 buf->b_efunc = NULL; 3481 buf->b_private = NULL; 3482 buf->b_next = NULL; 3483 hdr->b_buf = buf; 3484 ASSERT(hdr->b_datacnt == 0); 3485 hdr->b_datacnt = 1; 3486 arc_get_data_buf(buf); 3487 arc_access(hdr, hash_lock); 3488 } 3489 3490 ASSERT(!GHOST_STATE(hdr->b_state)); 3491 3492 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3493 acb->acb_done = done; 3494 acb->acb_private = private; 3495 3496 ASSERT(hdr->b_acb == NULL); 3497 hdr->b_acb = acb; 3498 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 3499 3500 if (hdr->b_l2hdr != NULL && 3501 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3502 devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3503 addr = hdr->b_l2hdr->b_daddr; 3504 b_compress = hdr->b_l2hdr->b_compress; 3505 b_asize = hdr->b_l2hdr->b_asize; 3506 /* 3507 * Lock out device removal. 3508 */ 3509 if (vdev_is_dead(vd) || 3510 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3511 vd = NULL; 3512 } 3513 3514 if (hash_lock != NULL) 3515 mutex_exit(hash_lock); 3516 3517 /* 3518 * At this point, we have a level 1 cache miss. Try again in 3519 * L2ARC if possible. 3520 */ 3521 ASSERT3U(hdr->b_size, ==, size); 3522 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3523 uint64_t, size, zbookmark_phys_t *, zb); 3524 ARCSTAT_BUMP(arcstat_misses); 3525 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 3526 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3527 data, metadata, misses); 3528#ifdef _KERNEL 3529 curthread->td_ru.ru_inblock++; 3530#endif 3531 3532 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3533 /* 3534 * Read from the L2ARC if the following are true: 3535 * 1. The L2ARC vdev was previously cached. 3536 * 2. This buffer still has L2ARC metadata. 3537 * 3. This buffer isn't currently writing to the L2ARC. 3538 * 4. The L2ARC entry wasn't evicted, which may 3539 * also have invalidated the vdev. 3540 * 5. This isn't prefetch and l2arc_noprefetch is set. 3541 */ 3542 if (hdr->b_l2hdr != NULL && 3543 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3544 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3545 l2arc_read_callback_t *cb; 3546 3547 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3548 ARCSTAT_BUMP(arcstat_l2_hits); 3549 3550 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3551 KM_SLEEP); 3552 cb->l2rcb_buf = buf; 3553 cb->l2rcb_spa = spa; 3554 cb->l2rcb_bp = *bp; 3555 cb->l2rcb_zb = *zb; 3556 cb->l2rcb_flags = zio_flags; 3557 cb->l2rcb_compress = b_compress; 3558 3559 ASSERT(addr >= VDEV_LABEL_START_SIZE && 3560 addr + size < vd->vdev_psize - 3561 VDEV_LABEL_END_SIZE); 3562 3563 /* 3564 * l2arc read. The SCL_L2ARC lock will be 3565 * released by l2arc_read_done(). 3566 * Issue a null zio if the underlying buffer 3567 * was squashed to zero size by compression. 3568 */ 3569 if (b_compress == ZIO_COMPRESS_EMPTY) { 3570 rzio = zio_null(pio, spa, vd, 3571 l2arc_read_done, cb, 3572 zio_flags | ZIO_FLAG_DONT_CACHE | 3573 ZIO_FLAG_CANFAIL | 3574 ZIO_FLAG_DONT_PROPAGATE | 3575 ZIO_FLAG_DONT_RETRY); 3576 } else { 3577 rzio = zio_read_phys(pio, vd, addr, 3578 b_asize, buf->b_data, 3579 ZIO_CHECKSUM_OFF, 3580 l2arc_read_done, cb, priority, 3581 zio_flags | ZIO_FLAG_DONT_CACHE | 3582 ZIO_FLAG_CANFAIL | 3583 ZIO_FLAG_DONT_PROPAGATE | 3584 ZIO_FLAG_DONT_RETRY, B_FALSE); 3585 } 3586 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3587 zio_t *, rzio); 3588 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3589 3590 if (*arc_flags & ARC_FLAG_NOWAIT) { 3591 zio_nowait(rzio); 3592 return (0); 3593 } 3594 3595 ASSERT(*arc_flags & ARC_FLAG_WAIT); 3596 if (zio_wait(rzio) == 0) 3597 return (0); 3598 3599 /* l2arc read error; goto zio_read() */ 3600 } else { 3601 DTRACE_PROBE1(l2arc__miss, 3602 arc_buf_hdr_t *, hdr); 3603 ARCSTAT_BUMP(arcstat_l2_misses); 3604 if (HDR_L2_WRITING(hdr)) 3605 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3606 spa_config_exit(spa, SCL_L2ARC, vd); 3607 } 3608 } else { 3609 if (vd != NULL) 3610 spa_config_exit(spa, SCL_L2ARC, vd); 3611 if (l2arc_ndev != 0) { 3612 DTRACE_PROBE1(l2arc__miss, 3613 arc_buf_hdr_t *, hdr); 3614 ARCSTAT_BUMP(arcstat_l2_misses); 3615 } 3616 } 3617 3618 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3619 arc_read_done, buf, priority, zio_flags, zb); 3620 3621 if (*arc_flags & ARC_FLAG_WAIT) 3622 return (zio_wait(rzio)); 3623 3624 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 3625 zio_nowait(rzio); 3626 } 3627 return (0); 3628} 3629 3630void 3631arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3632{ 3633 ASSERT(buf->b_hdr != NULL); 3634 ASSERT(buf->b_hdr->b_state != arc_anon); 3635 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3636 ASSERT(buf->b_efunc == NULL); 3637 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3638 3639 buf->b_efunc = func; 3640 buf->b_private = private; 3641} 3642 3643/* 3644 * Notify the arc that a block was freed, and thus will never be used again. 3645 */ 3646void 3647arc_freed(spa_t *spa, const blkptr_t *bp) 3648{ 3649 arc_buf_hdr_t *hdr; 3650 kmutex_t *hash_lock; 3651 uint64_t guid = spa_load_guid(spa); 3652 3653 ASSERT(!BP_IS_EMBEDDED(bp)); 3654 3655 hdr = buf_hash_find(guid, bp, &hash_lock); 3656 if (hdr == NULL) 3657 return; 3658 if (HDR_BUF_AVAILABLE(hdr)) { 3659 arc_buf_t *buf = hdr->b_buf; 3660 add_reference(hdr, hash_lock, FTAG); 3661 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 3662 mutex_exit(hash_lock); 3663 3664 arc_release(buf, FTAG); 3665 (void) arc_buf_remove_ref(buf, FTAG); 3666 } else { 3667 mutex_exit(hash_lock); 3668 } 3669 3670} 3671 3672/* 3673 * Clear the user eviction callback set by arc_set_callback(), first calling 3674 * it if it exists. Because the presence of a callback keeps an arc_buf cached 3675 * clearing the callback may result in the arc_buf being destroyed. However, 3676 * it will not result in the *last* arc_buf being destroyed, hence the data 3677 * will remain cached in the ARC. We make a copy of the arc buffer here so 3678 * that we can process the callback without holding any locks. 3679 * 3680 * It's possible that the callback is already in the process of being cleared 3681 * by another thread. In this case we can not clear the callback. 3682 * 3683 * Returns B_TRUE if the callback was successfully called and cleared. 3684 */ 3685boolean_t 3686arc_clear_callback(arc_buf_t *buf) 3687{ 3688 arc_buf_hdr_t *hdr; 3689 kmutex_t *hash_lock; 3690 arc_evict_func_t *efunc = buf->b_efunc; 3691 void *private = buf->b_private; 3692 list_t *list, *evicted_list; 3693 kmutex_t *lock, *evicted_lock; 3694 3695 mutex_enter(&buf->b_evict_lock); 3696 hdr = buf->b_hdr; 3697 if (hdr == NULL) { 3698 /* 3699 * We are in arc_do_user_evicts(). 3700 */ 3701 ASSERT(buf->b_data == NULL); 3702 mutex_exit(&buf->b_evict_lock); 3703 return (B_FALSE); 3704 } else if (buf->b_data == NULL) { 3705 /* 3706 * We are on the eviction list; process this buffer now 3707 * but let arc_do_user_evicts() do the reaping. 3708 */ 3709 buf->b_efunc = NULL; 3710 mutex_exit(&buf->b_evict_lock); 3711 VERIFY0(efunc(private)); 3712 return (B_TRUE); 3713 } 3714 hash_lock = HDR_LOCK(hdr); 3715 mutex_enter(hash_lock); 3716 hdr = buf->b_hdr; 3717 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3718 3719 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3720 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3721 3722 buf->b_efunc = NULL; 3723 buf->b_private = NULL; 3724 3725 if (hdr->b_datacnt > 1) { 3726 mutex_exit(&buf->b_evict_lock); 3727 arc_buf_destroy(buf, FALSE, TRUE); 3728 } else { 3729 ASSERT(buf == hdr->b_buf); 3730 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 3731 mutex_exit(&buf->b_evict_lock); 3732 } 3733 3734 mutex_exit(hash_lock); 3735 VERIFY0(efunc(private)); 3736 return (B_TRUE); 3737} 3738 3739/* 3740 * Release this buffer from the cache, making it an anonymous buffer. This 3741 * must be done after a read and prior to modifying the buffer contents. 3742 * If the buffer has more than one reference, we must make 3743 * a new hdr for the buffer. 3744 */ 3745void 3746arc_release(arc_buf_t *buf, void *tag) 3747{ 3748 arc_buf_hdr_t *hdr; 3749 kmutex_t *hash_lock = NULL; 3750 l2arc_buf_hdr_t *l2hdr; 3751 uint64_t buf_size; 3752 3753 /* 3754 * It would be nice to assert that if it's DMU metadata (level > 3755 * 0 || it's the dnode file), then it must be syncing context. 3756 * But we don't know that information at this level. 3757 */ 3758 3759 mutex_enter(&buf->b_evict_lock); 3760 hdr = buf->b_hdr; 3761 3762 /* this buffer is not on any list */ 3763 ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3764 3765 if (hdr->b_state == arc_anon) { 3766 /* this buffer is already released */ 3767 ASSERT(buf->b_efunc == NULL); 3768 } else { 3769 hash_lock = HDR_LOCK(hdr); 3770 mutex_enter(hash_lock); 3771 hdr = buf->b_hdr; 3772 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3773 } 3774 3775 l2hdr = hdr->b_l2hdr; 3776 if (l2hdr) { 3777 mutex_enter(&l2arc_buflist_mtx); 3778 arc_buf_l2_cdata_free(hdr); 3779 hdr->b_l2hdr = NULL; 3780 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3781 } 3782 buf_size = hdr->b_size; 3783 3784 /* 3785 * Do we have more than one buf? 3786 */ 3787 if (hdr->b_datacnt > 1) { 3788 arc_buf_hdr_t *nhdr; 3789 arc_buf_t **bufp; 3790 uint64_t blksz = hdr->b_size; 3791 uint64_t spa = hdr->b_spa; 3792 arc_buf_contents_t type = hdr->b_type; 3793 uint32_t flags = hdr->b_flags; 3794 3795 ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3796 /* 3797 * Pull the data off of this hdr and attach it to 3798 * a new anonymous hdr. 3799 */ 3800 (void) remove_reference(hdr, hash_lock, tag); 3801 bufp = &hdr->b_buf; 3802 while (*bufp != buf) 3803 bufp = &(*bufp)->b_next; 3804 *bufp = buf->b_next; 3805 buf->b_next = NULL; 3806 3807 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3808 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3809 if (refcount_is_zero(&hdr->b_refcnt)) { 3810 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3811 ASSERT3U(*size, >=, hdr->b_size); 3812 atomic_add_64(size, -hdr->b_size); 3813 } 3814 3815 /* 3816 * We're releasing a duplicate user data buffer, update 3817 * our statistics accordingly. 3818 */ 3819 if (hdr->b_type == ARC_BUFC_DATA) { 3820 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3821 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3822 -hdr->b_size); 3823 } 3824 hdr->b_datacnt -= 1; 3825 arc_cksum_verify(buf); 3826#ifdef illumos 3827 arc_buf_unwatch(buf); 3828#endif /* illumos */ 3829 3830 mutex_exit(hash_lock); 3831 3832 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3833 nhdr->b_size = blksz; 3834 nhdr->b_spa = spa; 3835 nhdr->b_type = type; 3836 nhdr->b_buf = buf; 3837 nhdr->b_state = arc_anon; 3838 nhdr->b_arc_access = 0; 3839 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; 3840 nhdr->b_l2hdr = NULL; 3841 nhdr->b_datacnt = 1; 3842 nhdr->b_freeze_cksum = NULL; 3843 (void) refcount_add(&nhdr->b_refcnt, tag); 3844 buf->b_hdr = nhdr; 3845 mutex_exit(&buf->b_evict_lock); 3846 atomic_add_64(&arc_anon->arcs_size, blksz); 3847 } else { 3848 mutex_exit(&buf->b_evict_lock); 3849 ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3850 ASSERT(!list_link_active(&hdr->b_arc_node)); 3851 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3852 if (hdr->b_state != arc_anon) 3853 arc_change_state(arc_anon, hdr, hash_lock); 3854 hdr->b_arc_access = 0; 3855 if (hash_lock) 3856 mutex_exit(hash_lock); 3857 3858 buf_discard_identity(hdr); 3859 arc_buf_thaw(buf); 3860 } 3861 buf->b_efunc = NULL; 3862 buf->b_private = NULL; 3863 3864 if (l2hdr) { 3865 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3866 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 3867 -l2hdr->b_asize, 0, 0); 3868 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3869 l2hdr->b_asize, 0); 3870 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3871 ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3872 mutex_exit(&l2arc_buflist_mtx); 3873 } 3874} 3875 3876int 3877arc_released(arc_buf_t *buf) 3878{ 3879 int released; 3880 3881 mutex_enter(&buf->b_evict_lock); 3882 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3883 mutex_exit(&buf->b_evict_lock); 3884 return (released); 3885} 3886 3887#ifdef ZFS_DEBUG 3888int 3889arc_referenced(arc_buf_t *buf) 3890{ 3891 int referenced; 3892 3893 mutex_enter(&buf->b_evict_lock); 3894 referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3895 mutex_exit(&buf->b_evict_lock); 3896 return (referenced); 3897} 3898#endif 3899 3900static void 3901arc_write_ready(zio_t *zio) 3902{ 3903 arc_write_callback_t *callback = zio->io_private; 3904 arc_buf_t *buf = callback->awcb_buf; 3905 arc_buf_hdr_t *hdr = buf->b_hdr; 3906 3907 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3908 callback->awcb_ready(zio, buf, callback->awcb_private); 3909 3910 /* 3911 * If the IO is already in progress, then this is a re-write 3912 * attempt, so we need to thaw and re-compute the cksum. 3913 * It is the responsibility of the callback to handle the 3914 * accounting for any re-write attempt. 3915 */ 3916 if (HDR_IO_IN_PROGRESS(hdr)) { 3917 mutex_enter(&hdr->b_freeze_lock); 3918 if (hdr->b_freeze_cksum != NULL) { 3919 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3920 hdr->b_freeze_cksum = NULL; 3921 } 3922 mutex_exit(&hdr->b_freeze_lock); 3923 } 3924 arc_cksum_compute(buf, B_FALSE); 3925 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 3926} 3927 3928/* 3929 * The SPA calls this callback for each physical write that happens on behalf 3930 * of a logical write. See the comment in dbuf_write_physdone() for details. 3931 */ 3932static void 3933arc_write_physdone(zio_t *zio) 3934{ 3935 arc_write_callback_t *cb = zio->io_private; 3936 if (cb->awcb_physdone != NULL) 3937 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3938} 3939 3940static void 3941arc_write_done(zio_t *zio) 3942{ 3943 arc_write_callback_t *callback = zio->io_private; 3944 arc_buf_t *buf = callback->awcb_buf; 3945 arc_buf_hdr_t *hdr = buf->b_hdr; 3946 3947 ASSERT(hdr->b_acb == NULL); 3948 3949 if (zio->io_error == 0) { 3950 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 3951 buf_discard_identity(hdr); 3952 } else { 3953 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3954 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3955 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3956 } 3957 } else { 3958 ASSERT(BUF_EMPTY(hdr)); 3959 } 3960 3961 /* 3962 * If the block to be written was all-zero or compressed enough to be 3963 * embedded in the BP, no write was performed so there will be no 3964 * dva/birth/checksum. The buffer must therefore remain anonymous 3965 * (and uncached). 3966 */ 3967 if (!BUF_EMPTY(hdr)) { 3968 arc_buf_hdr_t *exists; 3969 kmutex_t *hash_lock; 3970 3971 ASSERT(zio->io_error == 0); 3972 3973 arc_cksum_verify(buf); 3974 3975 exists = buf_hash_insert(hdr, &hash_lock); 3976 if (exists) { 3977 /* 3978 * This can only happen if we overwrite for 3979 * sync-to-convergence, because we remove 3980 * buffers from the hash table when we arc_free(). 3981 */ 3982 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3983 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3984 panic("bad overwrite, hdr=%p exists=%p", 3985 (void *)hdr, (void *)exists); 3986 ASSERT(refcount_is_zero(&exists->b_refcnt)); 3987 arc_change_state(arc_anon, exists, hash_lock); 3988 mutex_exit(hash_lock); 3989 arc_hdr_destroy(exists); 3990 exists = buf_hash_insert(hdr, &hash_lock); 3991 ASSERT3P(exists, ==, NULL); 3992 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3993 /* nopwrite */ 3994 ASSERT(zio->io_prop.zp_nopwrite); 3995 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3996 panic("bad nopwrite, hdr=%p exists=%p", 3997 (void *)hdr, (void *)exists); 3998 } else { 3999 /* Dedup */ 4000 ASSERT(hdr->b_datacnt == 1); 4001 ASSERT(hdr->b_state == arc_anon); 4002 ASSERT(BP_GET_DEDUP(zio->io_bp)); 4003 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 4004 } 4005 } 4006 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4007 /* if it's not anon, we are doing a scrub */ 4008 if (!exists && hdr->b_state == arc_anon) 4009 arc_access(hdr, hash_lock); 4010 mutex_exit(hash_lock); 4011 } else { 4012 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4013 } 4014 4015 ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 4016 callback->awcb_done(zio, buf, callback->awcb_private); 4017 4018 kmem_free(callback, sizeof (arc_write_callback_t)); 4019} 4020 4021zio_t * 4022arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 4023 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 4024 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 4025 arc_done_func_t *done, void *private, zio_priority_t priority, 4026 int zio_flags, const zbookmark_phys_t *zb) 4027{ 4028 arc_buf_hdr_t *hdr = buf->b_hdr; 4029 arc_write_callback_t *callback; 4030 zio_t *zio; 4031 4032 ASSERT(ready != NULL); 4033 ASSERT(done != NULL); 4034 ASSERT(!HDR_IO_ERROR(hdr)); 4035 ASSERT((hdr->b_flags & ARC_FLAG_IO_IN_PROGRESS) == 0); 4036 ASSERT(hdr->b_acb == NULL); 4037 if (l2arc) 4038 hdr->b_flags |= ARC_FLAG_L2CACHE; 4039 if (l2arc_compress) 4040 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4041 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 4042 callback->awcb_ready = ready; 4043 callback->awcb_physdone = physdone; 4044 callback->awcb_done = done; 4045 callback->awcb_private = private; 4046 callback->awcb_buf = buf; 4047 4048 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 4049 arc_write_ready, arc_write_physdone, arc_write_done, callback, 4050 priority, zio_flags, zb); 4051 4052 return (zio); 4053} 4054 4055static int 4056arc_memory_throttle(uint64_t reserve, uint64_t txg) 4057{ 4058#ifdef _KERNEL 4059 uint64_t available_memory = ptob(freemem); 4060 static uint64_t page_load = 0; 4061 static uint64_t last_txg = 0; 4062 4063#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 4064 available_memory = 4065 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 4066#endif 4067 4068 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 4069 return (0); 4070 4071 if (txg > last_txg) { 4072 last_txg = txg; 4073 page_load = 0; 4074 } 4075 /* 4076 * If we are in pageout, we know that memory is already tight, 4077 * the arc is already going to be evicting, so we just want to 4078 * continue to let page writes occur as quickly as possible. 4079 */ 4080 if (curproc == pageproc) { 4081 if (page_load > MAX(ptob(minfree), available_memory) / 4) 4082 return (SET_ERROR(ERESTART)); 4083 /* Note: reserve is inflated, so we deflate */ 4084 page_load += reserve / 8; 4085 return (0); 4086 } else if (page_load > 0 && arc_reclaim_needed()) { 4087 /* memory is low, delay before restarting */ 4088 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 4089 return (SET_ERROR(EAGAIN)); 4090 } 4091 page_load = 0; 4092#endif 4093 return (0); 4094} 4095 4096void 4097arc_tempreserve_clear(uint64_t reserve) 4098{ 4099 atomic_add_64(&arc_tempreserve, -reserve); 4100 ASSERT((int64_t)arc_tempreserve >= 0); 4101} 4102 4103int 4104arc_tempreserve_space(uint64_t reserve, uint64_t txg) 4105{ 4106 int error; 4107 uint64_t anon_size; 4108 4109 if (reserve > arc_c/4 && !arc_no_grow) { 4110 arc_c = MIN(arc_c_max, reserve * 4); 4111 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 4112 } 4113 if (reserve > arc_c) 4114 return (SET_ERROR(ENOMEM)); 4115 4116 /* 4117 * Don't count loaned bufs as in flight dirty data to prevent long 4118 * network delays from blocking transactions that are ready to be 4119 * assigned to a txg. 4120 */ 4121 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 4122 4123 /* 4124 * Writes will, almost always, require additional memory allocations 4125 * in order to compress/encrypt/etc the data. We therefore need to 4126 * make sure that there is sufficient available memory for this. 4127 */ 4128 error = arc_memory_throttle(reserve, txg); 4129 if (error != 0) 4130 return (error); 4131 4132 /* 4133 * Throttle writes when the amount of dirty data in the cache 4134 * gets too large. We try to keep the cache less than half full 4135 * of dirty blocks so that our sync times don't grow too large. 4136 * Note: if two requests come in concurrently, we might let them 4137 * both succeed, when one of them should fail. Not a huge deal. 4138 */ 4139 4140 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 4141 anon_size > arc_c / 4) { 4142 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 4143 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 4144 arc_tempreserve>>10, 4145 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 4146 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 4147 reserve>>10, arc_c>>10); 4148 return (SET_ERROR(ERESTART)); 4149 } 4150 atomic_add_64(&arc_tempreserve, reserve); 4151 return (0); 4152} 4153 4154static kmutex_t arc_lowmem_lock; 4155#ifdef _KERNEL 4156static eventhandler_tag arc_event_lowmem = NULL; 4157 4158static void 4159arc_lowmem(void *arg __unused, int howto __unused) 4160{ 4161 4162 /* Serialize access via arc_lowmem_lock. */ 4163 mutex_enter(&arc_lowmem_lock); 4164 mutex_enter(&arc_reclaim_thr_lock); 4165 needfree = 1; 4166 DTRACE_PROBE(arc__needfree); 4167 cv_signal(&arc_reclaim_thr_cv); 4168 4169 /* 4170 * It is unsafe to block here in arbitrary threads, because we can come 4171 * here from ARC itself and may hold ARC locks and thus risk a deadlock 4172 * with ARC reclaim thread. 4173 */ 4174 if (curproc == pageproc) { 4175 while (needfree) 4176 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4177 } 4178 mutex_exit(&arc_reclaim_thr_lock); 4179 mutex_exit(&arc_lowmem_lock); 4180} 4181#endif 4182 4183void 4184arc_init(void) 4185{ 4186 int i, prefetch_tunable_set = 0; 4187 4188 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4189 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4190 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4191 4192 /* Convert seconds to clock ticks */ 4193 arc_min_prefetch_lifespan = 1 * hz; 4194 4195 /* Start out with 1/8 of all memory */ 4196 arc_c = kmem_size() / 8; 4197 4198#ifdef sun 4199#ifdef _KERNEL 4200 /* 4201 * On architectures where the physical memory can be larger 4202 * than the addressable space (intel in 32-bit mode), we may 4203 * need to limit the cache to 1/8 of VM size. 4204 */ 4205 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4206#endif 4207#endif /* sun */ 4208 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4209 arc_c_min = MAX(arc_c / 4, 16 << 20); 4210 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4211 if (arc_c * 8 >= 1 << 30) 4212 arc_c_max = (arc_c * 8) - (1 << 30); 4213 else 4214 arc_c_max = arc_c_min; 4215 arc_c_max = MAX(arc_c * 5, arc_c_max); 4216 4217#ifdef _KERNEL 4218 /* 4219 * Allow the tunables to override our calculations if they are 4220 * reasonable (ie. over 16MB) 4221 */ 4222 if (zfs_arc_max > 16 << 20 && zfs_arc_max < kmem_size()) 4223 arc_c_max = zfs_arc_max; 4224 if (zfs_arc_min > 16 << 20 && zfs_arc_min <= arc_c_max) 4225 arc_c_min = zfs_arc_min; 4226#endif 4227 4228 arc_c = arc_c_max; 4229 arc_p = (arc_c >> 1); 4230 4231 /* limit meta-data to 1/4 of the arc capacity */ 4232 arc_meta_limit = arc_c_max / 4; 4233 4234 /* Allow the tunable to override if it is reasonable */ 4235 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4236 arc_meta_limit = zfs_arc_meta_limit; 4237 4238 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4239 arc_c_min = arc_meta_limit / 2; 4240 4241 if (zfs_arc_meta_min > 0) { 4242 arc_meta_min = zfs_arc_meta_min; 4243 } else { 4244 arc_meta_min = arc_c_min / 2; 4245 } 4246 4247 if (zfs_arc_grow_retry > 0) 4248 arc_grow_retry = zfs_arc_grow_retry; 4249 4250 if (zfs_arc_shrink_shift > 0) 4251 arc_shrink_shift = zfs_arc_shrink_shift; 4252 4253 if (zfs_arc_p_min_shift > 0) 4254 arc_p_min_shift = zfs_arc_p_min_shift; 4255 4256 /* if kmem_flags are set, lets try to use less memory */ 4257 if (kmem_debugging()) 4258 arc_c = arc_c / 2; 4259 if (arc_c < arc_c_min) 4260 arc_c = arc_c_min; 4261 4262 zfs_arc_min = arc_c_min; 4263 zfs_arc_max = arc_c_max; 4264 4265 arc_anon = &ARC_anon; 4266 arc_mru = &ARC_mru; 4267 arc_mru_ghost = &ARC_mru_ghost; 4268 arc_mfu = &ARC_mfu; 4269 arc_mfu_ghost = &ARC_mfu_ghost; 4270 arc_l2c_only = &ARC_l2c_only; 4271 arc_size = 0; 4272 4273 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4274 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4275 NULL, MUTEX_DEFAULT, NULL); 4276 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4277 NULL, MUTEX_DEFAULT, NULL); 4278 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4279 NULL, MUTEX_DEFAULT, NULL); 4280 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4281 NULL, MUTEX_DEFAULT, NULL); 4282 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4283 NULL, MUTEX_DEFAULT, NULL); 4284 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4285 NULL, MUTEX_DEFAULT, NULL); 4286 4287 list_create(&arc_mru->arcs_lists[i], 4288 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4289 list_create(&arc_mru_ghost->arcs_lists[i], 4290 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4291 list_create(&arc_mfu->arcs_lists[i], 4292 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4293 list_create(&arc_mfu_ghost->arcs_lists[i], 4294 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4295 list_create(&arc_mfu_ghost->arcs_lists[i], 4296 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4297 list_create(&arc_l2c_only->arcs_lists[i], 4298 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4299 } 4300 4301 buf_init(); 4302 4303 arc_thread_exit = 0; 4304 arc_eviction_list = NULL; 4305 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4306 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4307 4308 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4309 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4310 4311 if (arc_ksp != NULL) { 4312 arc_ksp->ks_data = &arc_stats; 4313 kstat_install(arc_ksp); 4314 } 4315 4316 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4317 TS_RUN, minclsyspri); 4318 4319#ifdef _KERNEL 4320 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4321 EVENTHANDLER_PRI_FIRST); 4322#endif 4323 4324 arc_dead = FALSE; 4325 arc_warm = B_FALSE; 4326 4327 /* 4328 * Calculate maximum amount of dirty data per pool. 4329 * 4330 * If it has been set by /etc/system, take that. 4331 * Otherwise, use a percentage of physical memory defined by 4332 * zfs_dirty_data_max_percent (default 10%) with a cap at 4333 * zfs_dirty_data_max_max (default 4GB). 4334 */ 4335 if (zfs_dirty_data_max == 0) { 4336 zfs_dirty_data_max = ptob(physmem) * 4337 zfs_dirty_data_max_percent / 100; 4338 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4339 zfs_dirty_data_max_max); 4340 } 4341 4342#ifdef _KERNEL 4343 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4344 prefetch_tunable_set = 1; 4345 4346#ifdef __i386__ 4347 if (prefetch_tunable_set == 0) { 4348 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4349 "-- to enable,\n"); 4350 printf(" add \"vfs.zfs.prefetch_disable=0\" " 4351 "to /boot/loader.conf.\n"); 4352 zfs_prefetch_disable = 1; 4353 } 4354#else 4355 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4356 prefetch_tunable_set == 0) { 4357 printf("ZFS NOTICE: Prefetch is disabled by default if less " 4358 "than 4GB of RAM is present;\n" 4359 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4360 "to /boot/loader.conf.\n"); 4361 zfs_prefetch_disable = 1; 4362 } 4363#endif 4364 /* Warn about ZFS memory and address space requirements. */ 4365 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4366 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4367 "expect unstable behavior.\n"); 4368 } 4369 if (kmem_size() < 512 * (1 << 20)) { 4370 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4371 "expect unstable behavior.\n"); 4372 printf(" Consider tuning vm.kmem_size and " 4373 "vm.kmem_size_max\n"); 4374 printf(" in /boot/loader.conf.\n"); 4375 } 4376#endif 4377} 4378 4379void 4380arc_fini(void) 4381{ 4382 int i; 4383 4384 mutex_enter(&arc_reclaim_thr_lock); 4385 arc_thread_exit = 1; 4386 cv_signal(&arc_reclaim_thr_cv); 4387 while (arc_thread_exit != 0) 4388 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4389 mutex_exit(&arc_reclaim_thr_lock); 4390 4391 arc_flush(NULL); 4392 4393 arc_dead = TRUE; 4394 4395 if (arc_ksp != NULL) { 4396 kstat_delete(arc_ksp); 4397 arc_ksp = NULL; 4398 } 4399 4400 mutex_destroy(&arc_eviction_mtx); 4401 mutex_destroy(&arc_reclaim_thr_lock); 4402 cv_destroy(&arc_reclaim_thr_cv); 4403 4404 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4405 list_destroy(&arc_mru->arcs_lists[i]); 4406 list_destroy(&arc_mru_ghost->arcs_lists[i]); 4407 list_destroy(&arc_mfu->arcs_lists[i]); 4408 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4409 list_destroy(&arc_l2c_only->arcs_lists[i]); 4410 4411 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4412 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4413 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4414 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4415 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4416 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4417 } 4418 4419 buf_fini(); 4420 4421 ASSERT(arc_loaned_bytes == 0); 4422 4423 mutex_destroy(&arc_lowmem_lock); 4424#ifdef _KERNEL 4425 if (arc_event_lowmem != NULL) 4426 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4427#endif 4428} 4429 4430/* 4431 * Level 2 ARC 4432 * 4433 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4434 * It uses dedicated storage devices to hold cached data, which are populated 4435 * using large infrequent writes. The main role of this cache is to boost 4436 * the performance of random read workloads. The intended L2ARC devices 4437 * include short-stroked disks, solid state disks, and other media with 4438 * substantially faster read latency than disk. 4439 * 4440 * +-----------------------+ 4441 * | ARC | 4442 * +-----------------------+ 4443 * | ^ ^ 4444 * | | | 4445 * l2arc_feed_thread() arc_read() 4446 * | | | 4447 * | l2arc read | 4448 * V | | 4449 * +---------------+ | 4450 * | L2ARC | | 4451 * +---------------+ | 4452 * | ^ | 4453 * l2arc_write() | | 4454 * | | | 4455 * V | | 4456 * +-------+ +-------+ 4457 * | vdev | | vdev | 4458 * | cache | | cache | 4459 * +-------+ +-------+ 4460 * +=========+ .-----. 4461 * : L2ARC : |-_____-| 4462 * : devices : | Disks | 4463 * +=========+ `-_____-' 4464 * 4465 * Read requests are satisfied from the following sources, in order: 4466 * 4467 * 1) ARC 4468 * 2) vdev cache of L2ARC devices 4469 * 3) L2ARC devices 4470 * 4) vdev cache of disks 4471 * 5) disks 4472 * 4473 * Some L2ARC device types exhibit extremely slow write performance. 4474 * To accommodate for this there are some significant differences between 4475 * the L2ARC and traditional cache design: 4476 * 4477 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4478 * the ARC behave as usual, freeing buffers and placing headers on ghost 4479 * lists. The ARC does not send buffers to the L2ARC during eviction as 4480 * this would add inflated write latencies for all ARC memory pressure. 4481 * 4482 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4483 * It does this by periodically scanning buffers from the eviction-end of 4484 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4485 * not already there. It scans until a headroom of buffers is satisfied, 4486 * which itself is a buffer for ARC eviction. If a compressible buffer is 4487 * found during scanning and selected for writing to an L2ARC device, we 4488 * temporarily boost scanning headroom during the next scan cycle to make 4489 * sure we adapt to compression effects (which might significantly reduce 4490 * the data volume we write to L2ARC). The thread that does this is 4491 * l2arc_feed_thread(), illustrated below; example sizes are included to 4492 * provide a better sense of ratio than this diagram: 4493 * 4494 * head --> tail 4495 * +---------------------+----------+ 4496 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4497 * +---------------------+----------+ | o L2ARC eligible 4498 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4499 * +---------------------+----------+ | 4500 * 15.9 Gbytes ^ 32 Mbytes | 4501 * headroom | 4502 * l2arc_feed_thread() 4503 * | 4504 * l2arc write hand <--[oooo]--' 4505 * | 8 Mbyte 4506 * | write max 4507 * V 4508 * +==============================+ 4509 * L2ARC dev |####|#|###|###| |####| ... | 4510 * +==============================+ 4511 * 32 Gbytes 4512 * 4513 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4514 * evicted, then the L2ARC has cached a buffer much sooner than it probably 4515 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4516 * safe to say that this is an uncommon case, since buffers at the end of 4517 * the ARC lists have moved there due to inactivity. 4518 * 4519 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4520 * then the L2ARC simply misses copying some buffers. This serves as a 4521 * pressure valve to prevent heavy read workloads from both stalling the ARC 4522 * with waits and clogging the L2ARC with writes. This also helps prevent 4523 * the potential for the L2ARC to churn if it attempts to cache content too 4524 * quickly, such as during backups of the entire pool. 4525 * 4526 * 5. After system boot and before the ARC has filled main memory, there are 4527 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4528 * lists can remain mostly static. Instead of searching from tail of these 4529 * lists as pictured, the l2arc_feed_thread() will search from the list heads 4530 * for eligible buffers, greatly increasing its chance of finding them. 4531 * 4532 * The L2ARC device write speed is also boosted during this time so that 4533 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4534 * there are no L2ARC reads, and no fear of degrading read performance 4535 * through increased writes. 4536 * 4537 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4538 * the vdev queue can aggregate them into larger and fewer writes. Each 4539 * device is written to in a rotor fashion, sweeping writes through 4540 * available space then repeating. 4541 * 4542 * 7. The L2ARC does not store dirty content. It never needs to flush 4543 * write buffers back to disk based storage. 4544 * 4545 * 8. If an ARC buffer is written (and dirtied) which also exists in the 4546 * L2ARC, the now stale L2ARC buffer is immediately dropped. 4547 * 4548 * The performance of the L2ARC can be tweaked by a number of tunables, which 4549 * may be necessary for different workloads: 4550 * 4551 * l2arc_write_max max write bytes per interval 4552 * l2arc_write_boost extra write bytes during device warmup 4553 * l2arc_noprefetch skip caching prefetched buffers 4554 * l2arc_headroom number of max device writes to precache 4555 * l2arc_headroom_boost when we find compressed buffers during ARC 4556 * scanning, we multiply headroom by this 4557 * percentage factor for the next scan cycle, 4558 * since more compressed buffers are likely to 4559 * be present 4560 * l2arc_feed_secs seconds between L2ARC writing 4561 * 4562 * Tunables may be removed or added as future performance improvements are 4563 * integrated, and also may become zpool properties. 4564 * 4565 * There are three key functions that control how the L2ARC warms up: 4566 * 4567 * l2arc_write_eligible() check if a buffer is eligible to cache 4568 * l2arc_write_size() calculate how much to write 4569 * l2arc_write_interval() calculate sleep delay between writes 4570 * 4571 * These three functions determine what to write, how much, and how quickly 4572 * to send writes. 4573 */ 4574 4575static boolean_t 4576l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 4577{ 4578 /* 4579 * A buffer is *not* eligible for the L2ARC if it: 4580 * 1. belongs to a different spa. 4581 * 2. is already cached on the L2ARC. 4582 * 3. has an I/O in progress (it may be an incomplete read). 4583 * 4. is flagged not eligible (zfs property). 4584 */ 4585 if (hdr->b_spa != spa_guid) { 4586 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4587 return (B_FALSE); 4588 } 4589 if (hdr->b_l2hdr != NULL) { 4590 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4591 return (B_FALSE); 4592 } 4593 if (HDR_IO_IN_PROGRESS(hdr)) { 4594 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4595 return (B_FALSE); 4596 } 4597 if (!HDR_L2CACHE(hdr)) { 4598 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4599 return (B_FALSE); 4600 } 4601 4602 return (B_TRUE); 4603} 4604 4605static uint64_t 4606l2arc_write_size(void) 4607{ 4608 uint64_t size; 4609 4610 /* 4611 * Make sure our globals have meaningful values in case the user 4612 * altered them. 4613 */ 4614 size = l2arc_write_max; 4615 if (size == 0) { 4616 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4617 "be greater than zero, resetting it to the default (%d)", 4618 L2ARC_WRITE_SIZE); 4619 size = l2arc_write_max = L2ARC_WRITE_SIZE; 4620 } 4621 4622 if (arc_warm == B_FALSE) 4623 size += l2arc_write_boost; 4624 4625 return (size); 4626 4627} 4628 4629static clock_t 4630l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4631{ 4632 clock_t interval, next, now; 4633 4634 /* 4635 * If the ARC lists are busy, increase our write rate; if the 4636 * lists are stale, idle back. This is achieved by checking 4637 * how much we previously wrote - if it was more than half of 4638 * what we wanted, schedule the next write much sooner. 4639 */ 4640 if (l2arc_feed_again && wrote > (wanted / 2)) 4641 interval = (hz * l2arc_feed_min_ms) / 1000; 4642 else 4643 interval = hz * l2arc_feed_secs; 4644 4645 now = ddi_get_lbolt(); 4646 next = MAX(now, MIN(now + interval, began + interval)); 4647 4648 return (next); 4649} 4650 4651static void 4652l2arc_hdr_stat_add(void) 4653{ 4654 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4655 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4656} 4657 4658static void 4659l2arc_hdr_stat_remove(void) 4660{ 4661 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4662 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4663} 4664 4665/* 4666 * Cycle through L2ARC devices. This is how L2ARC load balances. 4667 * If a device is returned, this also returns holding the spa config lock. 4668 */ 4669static l2arc_dev_t * 4670l2arc_dev_get_next(void) 4671{ 4672 l2arc_dev_t *first, *next = NULL; 4673 4674 /* 4675 * Lock out the removal of spas (spa_namespace_lock), then removal 4676 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4677 * both locks will be dropped and a spa config lock held instead. 4678 */ 4679 mutex_enter(&spa_namespace_lock); 4680 mutex_enter(&l2arc_dev_mtx); 4681 4682 /* if there are no vdevs, there is nothing to do */ 4683 if (l2arc_ndev == 0) 4684 goto out; 4685 4686 first = NULL; 4687 next = l2arc_dev_last; 4688 do { 4689 /* loop around the list looking for a non-faulted vdev */ 4690 if (next == NULL) { 4691 next = list_head(l2arc_dev_list); 4692 } else { 4693 next = list_next(l2arc_dev_list, next); 4694 if (next == NULL) 4695 next = list_head(l2arc_dev_list); 4696 } 4697 4698 /* if we have come back to the start, bail out */ 4699 if (first == NULL) 4700 first = next; 4701 else if (next == first) 4702 break; 4703 4704 } while (vdev_is_dead(next->l2ad_vdev)); 4705 4706 /* if we were unable to find any usable vdevs, return NULL */ 4707 if (vdev_is_dead(next->l2ad_vdev)) 4708 next = NULL; 4709 4710 l2arc_dev_last = next; 4711 4712out: 4713 mutex_exit(&l2arc_dev_mtx); 4714 4715 /* 4716 * Grab the config lock to prevent the 'next' device from being 4717 * removed while we are writing to it. 4718 */ 4719 if (next != NULL) 4720 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4721 mutex_exit(&spa_namespace_lock); 4722 4723 return (next); 4724} 4725 4726/* 4727 * Free buffers that were tagged for destruction. 4728 */ 4729static void 4730l2arc_do_free_on_write() 4731{ 4732 list_t *buflist; 4733 l2arc_data_free_t *df, *df_prev; 4734 4735 mutex_enter(&l2arc_free_on_write_mtx); 4736 buflist = l2arc_free_on_write; 4737 4738 for (df = list_tail(buflist); df; df = df_prev) { 4739 df_prev = list_prev(buflist, df); 4740 ASSERT(df->l2df_data != NULL); 4741 ASSERT(df->l2df_func != NULL); 4742 df->l2df_func(df->l2df_data, df->l2df_size); 4743 list_remove(buflist, df); 4744 kmem_free(df, sizeof (l2arc_data_free_t)); 4745 } 4746 4747 mutex_exit(&l2arc_free_on_write_mtx); 4748} 4749 4750/* 4751 * A write to a cache device has completed. Update all headers to allow 4752 * reads from these buffers to begin. 4753 */ 4754static void 4755l2arc_write_done(zio_t *zio) 4756{ 4757 l2arc_write_callback_t *cb; 4758 l2arc_dev_t *dev; 4759 list_t *buflist; 4760 arc_buf_hdr_t *head, *hdr, *hdr_prev; 4761 l2arc_buf_hdr_t *abl2; 4762 kmutex_t *hash_lock; 4763 int64_t bytes_dropped = 0; 4764 4765 cb = zio->io_private; 4766 ASSERT(cb != NULL); 4767 dev = cb->l2wcb_dev; 4768 ASSERT(dev != NULL); 4769 head = cb->l2wcb_head; 4770 ASSERT(head != NULL); 4771 buflist = dev->l2ad_buflist; 4772 ASSERT(buflist != NULL); 4773 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4774 l2arc_write_callback_t *, cb); 4775 4776 if (zio->io_error != 0) 4777 ARCSTAT_BUMP(arcstat_l2_writes_error); 4778 4779 mutex_enter(&l2arc_buflist_mtx); 4780 4781 /* 4782 * All writes completed, or an error was hit. 4783 */ 4784 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 4785 hdr_prev = list_prev(buflist, hdr); 4786 abl2 = hdr->b_l2hdr; 4787 4788 /* 4789 * Release the temporary compressed buffer as soon as possible. 4790 */ 4791 if (abl2->b_compress != ZIO_COMPRESS_OFF) 4792 l2arc_release_cdata_buf(hdr); 4793 4794 hash_lock = HDR_LOCK(hdr); 4795 if (!mutex_tryenter(hash_lock)) { 4796 /* 4797 * This buffer misses out. It may be in a stage 4798 * of eviction. Its ARC_L2_WRITING flag will be 4799 * left set, denying reads to this buffer. 4800 */ 4801 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4802 continue; 4803 } 4804 4805 if (zio->io_error != 0) { 4806 /* 4807 * Error - drop L2ARC entry. 4808 */ 4809 list_remove(buflist, hdr); 4810 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4811 bytes_dropped += abl2->b_asize; 4812 hdr->b_l2hdr = NULL; 4813 trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4814 abl2->b_asize, 0); 4815 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4816 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 4817 } 4818 4819 /* 4820 * Allow ARC to begin reads to this L2ARC entry. 4821 */ 4822 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 4823 4824 mutex_exit(hash_lock); 4825 } 4826 4827 atomic_inc_64(&l2arc_writes_done); 4828 list_remove(buflist, head); 4829 kmem_cache_free(hdr_cache, head); 4830 mutex_exit(&l2arc_buflist_mtx); 4831 4832 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 4833 4834 l2arc_do_free_on_write(); 4835 4836 kmem_free(cb, sizeof (l2arc_write_callback_t)); 4837} 4838 4839/* 4840 * A read to a cache device completed. Validate buffer contents before 4841 * handing over to the regular ARC routines. 4842 */ 4843static void 4844l2arc_read_done(zio_t *zio) 4845{ 4846 l2arc_read_callback_t *cb; 4847 arc_buf_hdr_t *hdr; 4848 arc_buf_t *buf; 4849 kmutex_t *hash_lock; 4850 int equal; 4851 4852 ASSERT(zio->io_vd != NULL); 4853 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4854 4855 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4856 4857 cb = zio->io_private; 4858 ASSERT(cb != NULL); 4859 buf = cb->l2rcb_buf; 4860 ASSERT(buf != NULL); 4861 4862 hash_lock = HDR_LOCK(buf->b_hdr); 4863 mutex_enter(hash_lock); 4864 hdr = buf->b_hdr; 4865 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4866 4867 /* 4868 * If the buffer was compressed, decompress it first. 4869 */ 4870 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4871 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4872 ASSERT(zio->io_data != NULL); 4873 4874 /* 4875 * Check this survived the L2ARC journey. 4876 */ 4877 equal = arc_cksum_equal(buf); 4878 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4879 mutex_exit(hash_lock); 4880 zio->io_private = buf; 4881 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4882 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4883 arc_read_done(zio); 4884 } else { 4885 mutex_exit(hash_lock); 4886 /* 4887 * Buffer didn't survive caching. Increment stats and 4888 * reissue to the original storage device. 4889 */ 4890 if (zio->io_error != 0) { 4891 ARCSTAT_BUMP(arcstat_l2_io_error); 4892 } else { 4893 zio->io_error = SET_ERROR(EIO); 4894 } 4895 if (!equal) 4896 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4897 4898 /* 4899 * If there's no waiter, issue an async i/o to the primary 4900 * storage now. If there *is* a waiter, the caller must 4901 * issue the i/o in a context where it's OK to block. 4902 */ 4903 if (zio->io_waiter == NULL) { 4904 zio_t *pio = zio_unique_parent(zio); 4905 4906 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4907 4908 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4909 buf->b_data, zio->io_size, arc_read_done, buf, 4910 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4911 } 4912 } 4913 4914 kmem_free(cb, sizeof (l2arc_read_callback_t)); 4915} 4916 4917/* 4918 * This is the list priority from which the L2ARC will search for pages to 4919 * cache. This is used within loops (0..3) to cycle through lists in the 4920 * desired order. This order can have a significant effect on cache 4921 * performance. 4922 * 4923 * Currently the metadata lists are hit first, MFU then MRU, followed by 4924 * the data lists. This function returns a locked list, and also returns 4925 * the lock pointer. 4926 */ 4927static list_t * 4928l2arc_list_locked(int list_num, kmutex_t **lock) 4929{ 4930 list_t *list = NULL; 4931 int idx; 4932 4933 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4934 4935 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4936 idx = list_num; 4937 list = &arc_mfu->arcs_lists[idx]; 4938 *lock = ARCS_LOCK(arc_mfu, idx); 4939 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4940 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4941 list = &arc_mru->arcs_lists[idx]; 4942 *lock = ARCS_LOCK(arc_mru, idx); 4943 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4944 ARC_BUFC_NUMDATALISTS)) { 4945 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4946 list = &arc_mfu->arcs_lists[idx]; 4947 *lock = ARCS_LOCK(arc_mfu, idx); 4948 } else { 4949 idx = list_num - ARC_BUFC_NUMLISTS; 4950 list = &arc_mru->arcs_lists[idx]; 4951 *lock = ARCS_LOCK(arc_mru, idx); 4952 } 4953 4954 ASSERT(!(MUTEX_HELD(*lock))); 4955 mutex_enter(*lock); 4956 return (list); 4957} 4958 4959/* 4960 * Evict buffers from the device write hand to the distance specified in 4961 * bytes. This distance may span populated buffers, it may span nothing. 4962 * This is clearing a region on the L2ARC device ready for writing. 4963 * If the 'all' boolean is set, every buffer is evicted. 4964 */ 4965static void 4966l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4967{ 4968 list_t *buflist; 4969 l2arc_buf_hdr_t *abl2; 4970 arc_buf_hdr_t *hdr, *hdr_prev; 4971 kmutex_t *hash_lock; 4972 uint64_t taddr; 4973 int64_t bytes_evicted = 0; 4974 4975 buflist = dev->l2ad_buflist; 4976 4977 if (buflist == NULL) 4978 return; 4979 4980 if (!all && dev->l2ad_first) { 4981 /* 4982 * This is the first sweep through the device. There is 4983 * nothing to evict. 4984 */ 4985 return; 4986 } 4987 4988 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4989 /* 4990 * When nearing the end of the device, evict to the end 4991 * before the device write hand jumps to the start. 4992 */ 4993 taddr = dev->l2ad_end; 4994 } else { 4995 taddr = dev->l2ad_hand + distance; 4996 } 4997 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4998 uint64_t, taddr, boolean_t, all); 4999 5000top: 5001 mutex_enter(&l2arc_buflist_mtx); 5002 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 5003 hdr_prev = list_prev(buflist, hdr); 5004 5005 hash_lock = HDR_LOCK(hdr); 5006 if (!mutex_tryenter(hash_lock)) { 5007 /* 5008 * Missed the hash lock. Retry. 5009 */ 5010 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 5011 mutex_exit(&l2arc_buflist_mtx); 5012 mutex_enter(hash_lock); 5013 mutex_exit(hash_lock); 5014 goto top; 5015 } 5016 5017 if (HDR_L2_WRITE_HEAD(hdr)) { 5018 /* 5019 * We hit a write head node. Leave it for 5020 * l2arc_write_done(). 5021 */ 5022 list_remove(buflist, hdr); 5023 mutex_exit(hash_lock); 5024 continue; 5025 } 5026 5027 if (!all && hdr->b_l2hdr != NULL && 5028 (hdr->b_l2hdr->b_daddr > taddr || 5029 hdr->b_l2hdr->b_daddr < dev->l2ad_hand)) { 5030 /* 5031 * We've evicted to the target address, 5032 * or the end of the device. 5033 */ 5034 mutex_exit(hash_lock); 5035 break; 5036 } 5037 5038 if (HDR_FREE_IN_PROGRESS(hdr)) { 5039 /* 5040 * Already on the path to destruction. 5041 */ 5042 mutex_exit(hash_lock); 5043 continue; 5044 } 5045 5046 if (hdr->b_state == arc_l2c_only) { 5047 ASSERT(!HDR_L2_READING(hdr)); 5048 /* 5049 * This doesn't exist in the ARC. Destroy. 5050 * arc_hdr_destroy() will call list_remove() 5051 * and decrement arcstat_l2_size. 5052 */ 5053 arc_change_state(arc_anon, hdr, hash_lock); 5054 arc_hdr_destroy(hdr); 5055 } else { 5056 /* 5057 * Invalidate issued or about to be issued 5058 * reads, since we may be about to write 5059 * over this location. 5060 */ 5061 if (HDR_L2_READING(hdr)) { 5062 ARCSTAT_BUMP(arcstat_l2_evict_reading); 5063 hdr->b_flags |= ARC_FLAG_L2_EVICTED; 5064 } 5065 5066 /* 5067 * Tell ARC this no longer exists in L2ARC. 5068 */ 5069 if (hdr->b_l2hdr != NULL) { 5070 abl2 = hdr->b_l2hdr; 5071 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 5072 bytes_evicted += abl2->b_asize; 5073 hdr->b_l2hdr = NULL; 5074 /* 5075 * We are destroying l2hdr, so ensure that 5076 * its compressed buffer, if any, is not leaked. 5077 */ 5078 ASSERT(abl2->b_tmp_cdata == NULL); 5079 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 5080 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 5081 } 5082 list_remove(buflist, hdr); 5083 5084 /* 5085 * This may have been leftover after a 5086 * failed write. 5087 */ 5088 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 5089 } 5090 mutex_exit(hash_lock); 5091 } 5092 mutex_exit(&l2arc_buflist_mtx); 5093 5094 vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); 5095 dev->l2ad_evict = taddr; 5096} 5097 5098/* 5099 * Find and write ARC buffers to the L2ARC device. 5100 * 5101 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 5102 * for reading until they have completed writing. 5103 * The headroom_boost is an in-out parameter used to maintain headroom boost 5104 * state between calls to this function. 5105 * 5106 * Returns the number of bytes actually written (which may be smaller than 5107 * the delta by which the device hand has changed due to alignment). 5108 */ 5109static uint64_t 5110l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 5111 boolean_t *headroom_boost) 5112{ 5113 arc_buf_hdr_t *hdr, *hdr_prev, *head; 5114 list_t *list; 5115 uint64_t write_asize, write_sz, headroom, buf_compress_minsz; 5116 void *buf_data; 5117 kmutex_t *list_lock; 5118 boolean_t full; 5119 l2arc_write_callback_t *cb; 5120 zio_t *pio, *wzio; 5121 uint64_t guid = spa_load_guid(spa); 5122 const boolean_t do_headroom_boost = *headroom_boost; 5123 int try; 5124 5125 ASSERT(dev->l2ad_vdev != NULL); 5126 5127 /* Lower the flag now, we might want to raise it again later. */ 5128 *headroom_boost = B_FALSE; 5129 5130 pio = NULL; 5131 write_sz = write_asize = 0; 5132 full = B_FALSE; 5133 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 5134 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; 5135 5136 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 5137 /* 5138 * We will want to try to compress buffers that are at least 2x the 5139 * device sector size. 5140 */ 5141 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 5142 5143 /* 5144 * Copy buffers for L2ARC writing. 5145 */ 5146 mutex_enter(&l2arc_buflist_mtx); 5147 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 5148 uint64_t passed_sz = 0; 5149 5150 list = l2arc_list_locked(try, &list_lock); 5151 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 5152 5153 /* 5154 * L2ARC fast warmup. 5155 * 5156 * Until the ARC is warm and starts to evict, read from the 5157 * head of the ARC lists rather than the tail. 5158 */ 5159 if (arc_warm == B_FALSE) 5160 hdr = list_head(list); 5161 else 5162 hdr = list_tail(list); 5163 if (hdr == NULL) 5164 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 5165 5166 headroom = target_sz * l2arc_headroom * 2 / ARC_BUFC_NUMLISTS; 5167 if (do_headroom_boost) 5168 headroom = (headroom * l2arc_headroom_boost) / 100; 5169 5170 for (; hdr; hdr = hdr_prev) { 5171 l2arc_buf_hdr_t *l2hdr; 5172 kmutex_t *hash_lock; 5173 uint64_t buf_sz; 5174 uint64_t buf_a_sz; 5175 5176 if (arc_warm == B_FALSE) 5177 hdr_prev = list_next(list, hdr); 5178 else 5179 hdr_prev = list_prev(list, hdr); 5180 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size); 5181 5182 hash_lock = HDR_LOCK(hdr); 5183 if (!mutex_tryenter(hash_lock)) { 5184 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5185 /* 5186 * Skip this buffer rather than waiting. 5187 */ 5188 continue; 5189 } 5190 5191 passed_sz += hdr->b_size; 5192 if (passed_sz > headroom) { 5193 /* 5194 * Searched too far. 5195 */ 5196 mutex_exit(hash_lock); 5197 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5198 break; 5199 } 5200 5201 if (!l2arc_write_eligible(guid, hdr)) { 5202 mutex_exit(hash_lock); 5203 continue; 5204 } 5205 5206 /* 5207 * Assume that the buffer is not going to be compressed 5208 * and could take more space on disk because of a larger 5209 * disk block size. 5210 */ 5211 buf_sz = hdr->b_size; 5212 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5213 5214 if ((write_asize + buf_a_sz) > target_sz) { 5215 full = B_TRUE; 5216 mutex_exit(hash_lock); 5217 ARCSTAT_BUMP(arcstat_l2_write_full); 5218 break; 5219 } 5220 5221 if (pio == NULL) { 5222 /* 5223 * Insert a dummy header on the buflist so 5224 * l2arc_write_done() can find where the 5225 * write buffers begin without searching. 5226 */ 5227 list_insert_head(dev->l2ad_buflist, head); 5228 5229 cb = kmem_alloc( 5230 sizeof (l2arc_write_callback_t), KM_SLEEP); 5231 cb->l2wcb_dev = dev; 5232 cb->l2wcb_head = head; 5233 pio = zio_root(spa, l2arc_write_done, cb, 5234 ZIO_FLAG_CANFAIL); 5235 ARCSTAT_BUMP(arcstat_l2_write_pios); 5236 } 5237 5238 /* 5239 * Create and add a new L2ARC header. 5240 */ 5241 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5242 l2hdr->b_dev = dev; 5243 hdr->b_flags |= ARC_FLAG_L2_WRITING; 5244 5245 /* 5246 * Temporarily stash the data buffer in b_tmp_cdata. 5247 * The subsequent write step will pick it up from 5248 * there. This is because can't access hdr->b_buf 5249 * without holding the hash_lock, which we in turn 5250 * can't access without holding the ARC list locks 5251 * (which we want to avoid during compression/writing). 5252 */ 5253 l2hdr->b_compress = ZIO_COMPRESS_OFF; 5254 l2hdr->b_asize = hdr->b_size; 5255 l2hdr->b_tmp_cdata = hdr->b_buf->b_data; 5256 5257 hdr->b_l2hdr = l2hdr; 5258 5259 list_insert_head(dev->l2ad_buflist, hdr); 5260 5261 /* 5262 * Compute and store the buffer cksum before 5263 * writing. On debug the cksum is verified first. 5264 */ 5265 arc_cksum_verify(hdr->b_buf); 5266 arc_cksum_compute(hdr->b_buf, B_TRUE); 5267 5268 mutex_exit(hash_lock); 5269 5270 write_sz += buf_sz; 5271 write_asize += buf_a_sz; 5272 } 5273 5274 mutex_exit(list_lock); 5275 5276 if (full == B_TRUE) 5277 break; 5278 } 5279 5280 /* No buffers selected for writing? */ 5281 if (pio == NULL) { 5282 ASSERT0(write_sz); 5283 mutex_exit(&l2arc_buflist_mtx); 5284 kmem_cache_free(hdr_cache, head); 5285 return (0); 5286 } 5287 5288 /* 5289 * Note that elsewhere in this file arcstat_l2_asize 5290 * and the used space on l2ad_vdev are updated using b_asize, 5291 * which is not necessarily rounded up to the device block size. 5292 * Too keep accounting consistent we do the same here as well: 5293 * stats_size accumulates the sum of b_asize of the written buffers, 5294 * while write_asize accumulates the sum of b_asize rounded up 5295 * to the device block size. 5296 * The latter sum is used only to validate the corectness of the code. 5297 */ 5298 uint64_t stats_size = 0; 5299 write_asize = 0; 5300 5301 /* 5302 * Now start writing the buffers. We're starting at the write head 5303 * and work backwards, retracing the course of the buffer selector 5304 * loop above. 5305 */ 5306 for (hdr = list_prev(dev->l2ad_buflist, head); hdr; 5307 hdr = list_prev(dev->l2ad_buflist, hdr)) { 5308 l2arc_buf_hdr_t *l2hdr; 5309 uint64_t buf_sz; 5310 5311 /* 5312 * We shouldn't need to lock the buffer here, since we flagged 5313 * it as ARC_FLAG_L2_WRITING in the previous step, but we must 5314 * take care to only access its L2 cache parameters. In 5315 * particular, hdr->b_buf may be invalid by now due to 5316 * ARC eviction. 5317 */ 5318 l2hdr = hdr->b_l2hdr; 5319 l2hdr->b_daddr = dev->l2ad_hand; 5320 5321 if ((hdr->b_flags & ARC_FLAG_L2COMPRESS) && 5322 l2hdr->b_asize >= buf_compress_minsz) { 5323 if (l2arc_compress_buf(l2hdr)) { 5324 /* 5325 * If compression succeeded, enable headroom 5326 * boost on the next scan cycle. 5327 */ 5328 *headroom_boost = B_TRUE; 5329 } 5330 } 5331 5332 /* 5333 * Pick up the buffer data we had previously stashed away 5334 * (and now potentially also compressed). 5335 */ 5336 buf_data = l2hdr->b_tmp_cdata; 5337 buf_sz = l2hdr->b_asize; 5338 5339 /* 5340 * If the data has not been compressed, then clear b_tmp_cdata 5341 * to make sure that it points only to a temporary compression 5342 * buffer. 5343 */ 5344 if (!L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)) 5345 l2hdr->b_tmp_cdata = NULL; 5346 5347 /* Compression may have squashed the buffer to zero length. */ 5348 if (buf_sz != 0) { 5349 uint64_t buf_a_sz; 5350 5351 wzio = zio_write_phys(pio, dev->l2ad_vdev, 5352 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5353 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5354 ZIO_FLAG_CANFAIL, B_FALSE); 5355 5356 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5357 zio_t *, wzio); 5358 (void) zio_nowait(wzio); 5359 5360 stats_size += buf_sz; 5361 /* 5362 * Keep the clock hand suitably device-aligned. 5363 */ 5364 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5365 write_asize += buf_a_sz; 5366 dev->l2ad_hand += buf_a_sz; 5367 } 5368 } 5369 5370 mutex_exit(&l2arc_buflist_mtx); 5371 5372 ASSERT3U(write_asize, <=, target_sz); 5373 ARCSTAT_BUMP(arcstat_l2_writes_sent); 5374 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5375 ARCSTAT_INCR(arcstat_l2_size, write_sz); 5376 ARCSTAT_INCR(arcstat_l2_asize, stats_size); 5377 vdev_space_update(dev->l2ad_vdev, stats_size, 0, 0); 5378 5379 /* 5380 * Bump device hand to the device start if it is approaching the end. 5381 * l2arc_evict() will already have evicted ahead for this case. 5382 */ 5383 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5384 dev->l2ad_hand = dev->l2ad_start; 5385 dev->l2ad_evict = dev->l2ad_start; 5386 dev->l2ad_first = B_FALSE; 5387 } 5388 5389 dev->l2ad_writing = B_TRUE; 5390 (void) zio_wait(pio); 5391 dev->l2ad_writing = B_FALSE; 5392 5393 return (write_asize); 5394} 5395 5396/* 5397 * Compresses an L2ARC buffer. 5398 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5399 * size in l2hdr->b_asize. This routine tries to compress the data and 5400 * depending on the compression result there are three possible outcomes: 5401 * *) The buffer was incompressible. The original l2hdr contents were left 5402 * untouched and are ready for writing to an L2 device. 5403 * *) The buffer was all-zeros, so there is no need to write it to an L2 5404 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5405 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5406 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5407 * data buffer which holds the compressed data to be written, and b_asize 5408 * tells us how much data there is. b_compress is set to the appropriate 5409 * compression algorithm. Once writing is done, invoke 5410 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5411 * 5412 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5413 * buffer was incompressible). 5414 */ 5415static boolean_t 5416l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5417{ 5418 void *cdata; 5419 size_t csize, len, rounded; 5420 5421 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5422 ASSERT(l2hdr->b_tmp_cdata != NULL); 5423 5424 len = l2hdr->b_asize; 5425 cdata = zio_data_buf_alloc(len); 5426 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5427 cdata, l2hdr->b_asize); 5428 5429 if (csize == 0) { 5430 /* zero block, indicate that there's nothing to write */ 5431 zio_data_buf_free(cdata, len); 5432 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5433 l2hdr->b_asize = 0; 5434 l2hdr->b_tmp_cdata = NULL; 5435 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5436 return (B_TRUE); 5437 } 5438 5439 rounded = P2ROUNDUP(csize, 5440 (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift); 5441 if (rounded < len) { 5442 /* 5443 * Compression succeeded, we'll keep the cdata around for 5444 * writing and release it afterwards. 5445 */ 5446 if (rounded > csize) { 5447 bzero((char *)cdata + csize, rounded - csize); 5448 csize = rounded; 5449 } 5450 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5451 l2hdr->b_asize = csize; 5452 l2hdr->b_tmp_cdata = cdata; 5453 ARCSTAT_BUMP(arcstat_l2_compress_successes); 5454 return (B_TRUE); 5455 } else { 5456 /* 5457 * Compression failed, release the compressed buffer. 5458 * l2hdr will be left unmodified. 5459 */ 5460 zio_data_buf_free(cdata, len); 5461 ARCSTAT_BUMP(arcstat_l2_compress_failures); 5462 return (B_FALSE); 5463 } 5464} 5465 5466/* 5467 * Decompresses a zio read back from an l2arc device. On success, the 5468 * underlying zio's io_data buffer is overwritten by the uncompressed 5469 * version. On decompression error (corrupt compressed stream), the 5470 * zio->io_error value is set to signal an I/O error. 5471 * 5472 * Please note that the compressed data stream is not checksummed, so 5473 * if the underlying device is experiencing data corruption, we may feed 5474 * corrupt data to the decompressor, so the decompressor needs to be 5475 * able to handle this situation (LZ4 does). 5476 */ 5477static void 5478l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5479{ 5480 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5481 5482 if (zio->io_error != 0) { 5483 /* 5484 * An io error has occured, just restore the original io 5485 * size in preparation for a main pool read. 5486 */ 5487 zio->io_orig_size = zio->io_size = hdr->b_size; 5488 return; 5489 } 5490 5491 if (c == ZIO_COMPRESS_EMPTY) { 5492 /* 5493 * An empty buffer results in a null zio, which means we 5494 * need to fill its io_data after we're done restoring the 5495 * buffer's contents. 5496 */ 5497 ASSERT(hdr->b_buf != NULL); 5498 bzero(hdr->b_buf->b_data, hdr->b_size); 5499 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5500 } else { 5501 ASSERT(zio->io_data != NULL); 5502 /* 5503 * We copy the compressed data from the start of the arc buffer 5504 * (the zio_read will have pulled in only what we need, the 5505 * rest is garbage which we will overwrite at decompression) 5506 * and then decompress back to the ARC data buffer. This way we 5507 * can minimize copying by simply decompressing back over the 5508 * original compressed data (rather than decompressing to an 5509 * aux buffer and then copying back the uncompressed buffer, 5510 * which is likely to be much larger). 5511 */ 5512 uint64_t csize; 5513 void *cdata; 5514 5515 csize = zio->io_size; 5516 cdata = zio_data_buf_alloc(csize); 5517 bcopy(zio->io_data, cdata, csize); 5518 if (zio_decompress_data(c, cdata, zio->io_data, csize, 5519 hdr->b_size) != 0) 5520 zio->io_error = EIO; 5521 zio_data_buf_free(cdata, csize); 5522 } 5523 5524 /* Restore the expected uncompressed IO size. */ 5525 zio->io_orig_size = zio->io_size = hdr->b_size; 5526} 5527 5528/* 5529 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5530 * This buffer serves as a temporary holder of compressed data while 5531 * the buffer entry is being written to an l2arc device. Once that is 5532 * done, we can dispose of it. 5533 */ 5534static void 5535l2arc_release_cdata_buf(arc_buf_hdr_t *hdr) 5536{ 5537 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 5538 5539 ASSERT(L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)); 5540 if (l2hdr->b_compress != ZIO_COMPRESS_EMPTY) { 5541 /* 5542 * If the data was compressed, then we've allocated a 5543 * temporary buffer for it, so now we need to release it. 5544 */ 5545 ASSERT(l2hdr->b_tmp_cdata != NULL); 5546 zio_data_buf_free(l2hdr->b_tmp_cdata, hdr->b_size); 5547 l2hdr->b_tmp_cdata = NULL; 5548 } else { 5549 ASSERT(l2hdr->b_tmp_cdata == NULL); 5550 } 5551} 5552 5553/* 5554 * This thread feeds the L2ARC at regular intervals. This is the beating 5555 * heart of the L2ARC. 5556 */ 5557static void 5558l2arc_feed_thread(void *dummy __unused) 5559{ 5560 callb_cpr_t cpr; 5561 l2arc_dev_t *dev; 5562 spa_t *spa; 5563 uint64_t size, wrote; 5564 clock_t begin, next = ddi_get_lbolt(); 5565 boolean_t headroom_boost = B_FALSE; 5566 5567 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5568 5569 mutex_enter(&l2arc_feed_thr_lock); 5570 5571 while (l2arc_thread_exit == 0) { 5572 CALLB_CPR_SAFE_BEGIN(&cpr); 5573 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5574 next - ddi_get_lbolt()); 5575 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5576 next = ddi_get_lbolt() + hz; 5577 5578 /* 5579 * Quick check for L2ARC devices. 5580 */ 5581 mutex_enter(&l2arc_dev_mtx); 5582 if (l2arc_ndev == 0) { 5583 mutex_exit(&l2arc_dev_mtx); 5584 continue; 5585 } 5586 mutex_exit(&l2arc_dev_mtx); 5587 begin = ddi_get_lbolt(); 5588 5589 /* 5590 * This selects the next l2arc device to write to, and in 5591 * doing so the next spa to feed from: dev->l2ad_spa. This 5592 * will return NULL if there are now no l2arc devices or if 5593 * they are all faulted. 5594 * 5595 * If a device is returned, its spa's config lock is also 5596 * held to prevent device removal. l2arc_dev_get_next() 5597 * will grab and release l2arc_dev_mtx. 5598 */ 5599 if ((dev = l2arc_dev_get_next()) == NULL) 5600 continue; 5601 5602 spa = dev->l2ad_spa; 5603 ASSERT(spa != NULL); 5604 5605 /* 5606 * If the pool is read-only then force the feed thread to 5607 * sleep a little longer. 5608 */ 5609 if (!spa_writeable(spa)) { 5610 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5611 spa_config_exit(spa, SCL_L2ARC, dev); 5612 continue; 5613 } 5614 5615 /* 5616 * Avoid contributing to memory pressure. 5617 */ 5618 if (arc_reclaim_needed()) { 5619 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5620 spa_config_exit(spa, SCL_L2ARC, dev); 5621 continue; 5622 } 5623 5624 ARCSTAT_BUMP(arcstat_l2_feeds); 5625 5626 size = l2arc_write_size(); 5627 5628 /* 5629 * Evict L2ARC buffers that will be overwritten. 5630 */ 5631 l2arc_evict(dev, size, B_FALSE); 5632 5633 /* 5634 * Write ARC buffers. 5635 */ 5636 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5637 5638 /* 5639 * Calculate interval between writes. 5640 */ 5641 next = l2arc_write_interval(begin, size, wrote); 5642 spa_config_exit(spa, SCL_L2ARC, dev); 5643 } 5644 5645 l2arc_thread_exit = 0; 5646 cv_broadcast(&l2arc_feed_thr_cv); 5647 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5648 thread_exit(); 5649} 5650 5651boolean_t 5652l2arc_vdev_present(vdev_t *vd) 5653{ 5654 l2arc_dev_t *dev; 5655 5656 mutex_enter(&l2arc_dev_mtx); 5657 for (dev = list_head(l2arc_dev_list); dev != NULL; 5658 dev = list_next(l2arc_dev_list, dev)) { 5659 if (dev->l2ad_vdev == vd) 5660 break; 5661 } 5662 mutex_exit(&l2arc_dev_mtx); 5663 5664 return (dev != NULL); 5665} 5666 5667/* 5668 * Add a vdev for use by the L2ARC. By this point the spa has already 5669 * validated the vdev and opened it. 5670 */ 5671void 5672l2arc_add_vdev(spa_t *spa, vdev_t *vd) 5673{ 5674 l2arc_dev_t *adddev; 5675 5676 ASSERT(!l2arc_vdev_present(vd)); 5677 5678 vdev_ashift_optimize(vd); 5679 5680 /* 5681 * Create a new l2arc device entry. 5682 */ 5683 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5684 adddev->l2ad_spa = spa; 5685 adddev->l2ad_vdev = vd; 5686 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5687 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5688 adddev->l2ad_hand = adddev->l2ad_start; 5689 adddev->l2ad_evict = adddev->l2ad_start; 5690 adddev->l2ad_first = B_TRUE; 5691 adddev->l2ad_writing = B_FALSE; 5692 5693 /* 5694 * This is a list of all ARC buffers that are still valid on the 5695 * device. 5696 */ 5697 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5698 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5699 offsetof(arc_buf_hdr_t, b_l2node)); 5700 5701 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5702 5703 /* 5704 * Add device to global list 5705 */ 5706 mutex_enter(&l2arc_dev_mtx); 5707 list_insert_head(l2arc_dev_list, adddev); 5708 atomic_inc_64(&l2arc_ndev); 5709 mutex_exit(&l2arc_dev_mtx); 5710} 5711 5712/* 5713 * Remove a vdev from the L2ARC. 5714 */ 5715void 5716l2arc_remove_vdev(vdev_t *vd) 5717{ 5718 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5719 5720 /* 5721 * Find the device by vdev 5722 */ 5723 mutex_enter(&l2arc_dev_mtx); 5724 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5725 nextdev = list_next(l2arc_dev_list, dev); 5726 if (vd == dev->l2ad_vdev) { 5727 remdev = dev; 5728 break; 5729 } 5730 } 5731 ASSERT(remdev != NULL); 5732 5733 /* 5734 * Remove device from global list 5735 */ 5736 list_remove(l2arc_dev_list, remdev); 5737 l2arc_dev_last = NULL; /* may have been invalidated */ 5738 atomic_dec_64(&l2arc_ndev); 5739 mutex_exit(&l2arc_dev_mtx); 5740 5741 /* 5742 * Clear all buflists and ARC references. L2ARC device flush. 5743 */ 5744 l2arc_evict(remdev, 0, B_TRUE); 5745 list_destroy(remdev->l2ad_buflist); 5746 kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5747 kmem_free(remdev, sizeof (l2arc_dev_t)); 5748} 5749 5750void 5751l2arc_init(void) 5752{ 5753 l2arc_thread_exit = 0; 5754 l2arc_ndev = 0; 5755 l2arc_writes_sent = 0; 5756 l2arc_writes_done = 0; 5757 5758 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5759 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5760 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5761 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5762 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5763 5764 l2arc_dev_list = &L2ARC_dev_list; 5765 l2arc_free_on_write = &L2ARC_free_on_write; 5766 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5767 offsetof(l2arc_dev_t, l2ad_node)); 5768 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5769 offsetof(l2arc_data_free_t, l2df_list_node)); 5770} 5771 5772void 5773l2arc_fini(void) 5774{ 5775 /* 5776 * This is called from dmu_fini(), which is called from spa_fini(); 5777 * Because of this, we can assume that all l2arc devices have 5778 * already been removed when the pools themselves were removed. 5779 */ 5780 5781 l2arc_do_free_on_write(); 5782 5783 mutex_destroy(&l2arc_feed_thr_lock); 5784 cv_destroy(&l2arc_feed_thr_cv); 5785 mutex_destroy(&l2arc_dev_mtx); 5786 mutex_destroy(&l2arc_buflist_mtx); 5787 mutex_destroy(&l2arc_free_on_write_mtx); 5788 5789 list_destroy(l2arc_dev_list); 5790 list_destroy(l2arc_free_on_write); 5791} 5792 5793void 5794l2arc_start(void) 5795{ 5796 if (!(spa_mode_global & FWRITE)) 5797 return; 5798 5799 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5800 TS_RUN, minclsyspri); 5801} 5802 5803void 5804l2arc_stop(void) 5805{ 5806 if (!(spa_mode_global & FWRITE)) 5807 return; 5808 5809 mutex_enter(&l2arc_feed_thr_lock); 5810 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5811 l2arc_thread_exit = 1; 5812 while (l2arc_thread_exit != 0) 5813 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5814 mutex_exit(&l2arc_feed_thr_lock); 5815} 5816